1 /* 2 * Copyright (c) 1997, 2020, 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 "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "jfr/jfrEvents.hpp" 39 #include "memory/resourceArea.hpp" 40 #include "opto/addnode.hpp" 41 #include "opto/block.hpp" 42 #include "opto/c2compiler.hpp" 43 #include "opto/callGenerator.hpp" 44 #include "opto/callnode.hpp" 45 #include "opto/castnode.hpp" 46 #include "opto/cfgnode.hpp" 47 #include "opto/chaitin.hpp" 48 #include "opto/compile.hpp" 49 #include "opto/connode.hpp" 50 #include "opto/convertnode.hpp" 51 #include "opto/divnode.hpp" 52 #include "opto/escape.hpp" 53 #include "opto/idealGraphPrinter.hpp" 54 #include "opto/loopnode.hpp" 55 #include "opto/machnode.hpp" 56 #include "opto/macro.hpp" 57 #include "opto/matcher.hpp" 58 #include "opto/mathexactnode.hpp" 59 #include "opto/memnode.hpp" 60 #include "opto/mulnode.hpp" 61 #include "opto/narrowptrnode.hpp" 62 #include "opto/node.hpp" 63 #include "opto/opcodes.hpp" 64 #include "opto/output.hpp" 65 #include "opto/parse.hpp" 66 #include "opto/phaseX.hpp" 67 #include "opto/rootnode.hpp" 68 #include "opto/runtime.hpp" 69 #include "opto/stringopts.hpp" 70 #include "opto/type.hpp" 71 #include "opto/vectornode.hpp" 72 #include "runtime/arguments.hpp" 73 #include "runtime/sharedRuntime.hpp" 74 #include "runtime/signature.hpp" 75 #include "runtime/stubRoutines.hpp" 76 #include "runtime/timer.hpp" 77 #include "utilities/align.hpp" 78 #include "utilities/copy.hpp" 79 #include "utilities/macros.hpp" 80 #include "utilities/resourceHash.hpp" 81 82 83 // -------------------- Compile::mach_constant_base_node ----------------------- 84 // Constant table base node singleton. 85 MachConstantBaseNode* Compile::mach_constant_base_node() { 86 if (_mach_constant_base_node == NULL) { 87 _mach_constant_base_node = new MachConstantBaseNode(); 88 _mach_constant_base_node->add_req(C->root()); 89 } 90 return _mach_constant_base_node; 91 } 92 93 94 /// Support for intrinsics. 95 96 // Return the index at which m must be inserted (or already exists). 97 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 98 class IntrinsicDescPair { 99 private: 100 ciMethod* _m; 101 bool _is_virtual; 102 public: 103 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 104 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 105 ciMethod* m= elt->method(); 106 ciMethod* key_m = key->_m; 107 if (key_m < m) return -1; 108 else if (key_m > m) return 1; 109 else { 110 bool is_virtual = elt->is_virtual(); 111 bool key_virtual = key->_is_virtual; 112 if (key_virtual < is_virtual) return -1; 113 else if (key_virtual > is_virtual) return 1; 114 else return 0; 115 } 116 } 117 }; 118 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 119 #ifdef ASSERT 120 for (int i = 1; i < _intrinsics->length(); i++) { 121 CallGenerator* cg1 = _intrinsics->at(i-1); 122 CallGenerator* cg2 = _intrinsics->at(i); 123 assert(cg1->method() != cg2->method() 124 ? cg1->method() < cg2->method() 125 : cg1->is_virtual() < cg2->is_virtual(), 126 "compiler intrinsics list must stay sorted"); 127 } 128 #endif 129 IntrinsicDescPair pair(m, is_virtual); 130 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 131 } 132 133 void Compile::register_intrinsic(CallGenerator* cg) { 134 if (_intrinsics == NULL) { 135 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 136 } 137 int len = _intrinsics->length(); 138 bool found = false; 139 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 140 assert(!found, "registering twice"); 141 _intrinsics->insert_before(index, cg); 142 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 143 } 144 145 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 146 assert(m->is_loaded(), "don't try this on unloaded methods"); 147 if (_intrinsics != NULL) { 148 bool found = false; 149 int index = intrinsic_insertion_index(m, is_virtual, found); 150 if (found) { 151 return _intrinsics->at(index); 152 } 153 } 154 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 155 if (m->intrinsic_id() != vmIntrinsics::_none && 156 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 157 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 158 if (cg != NULL) { 159 // Save it for next time: 160 register_intrinsic(cg); 161 return cg; 162 } else { 163 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 164 } 165 } 166 return NULL; 167 } 168 169 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 170 // in library_call.cpp. 171 172 173 #ifndef PRODUCT 174 // statistics gathering... 175 176 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 177 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 178 179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 181 int oflags = _intrinsic_hist_flags[id]; 182 assert(flags != 0, "what happened?"); 183 if (is_virtual) { 184 flags |= _intrinsic_virtual; 185 } 186 bool changed = (flags != oflags); 187 if ((flags & _intrinsic_worked) != 0) { 188 juint count = (_intrinsic_hist_count[id] += 1); 189 if (count == 1) { 190 changed = true; // first time 191 } 192 // increment the overall count also: 193 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 194 } 195 if (changed) { 196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 197 // Something changed about the intrinsic's virtuality. 198 if ((flags & _intrinsic_virtual) != 0) { 199 // This is the first use of this intrinsic as a virtual call. 200 if (oflags != 0) { 201 // We already saw it as a non-virtual, so note both cases. 202 flags |= _intrinsic_both; 203 } 204 } else if ((oflags & _intrinsic_both) == 0) { 205 // This is the first use of this intrinsic as a non-virtual 206 flags |= _intrinsic_both; 207 } 208 } 209 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 210 } 211 // update the overall flags also: 212 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 213 return changed; 214 } 215 216 static char* format_flags(int flags, char* buf) { 217 buf[0] = 0; 218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 223 if (buf[0] == 0) strcat(buf, ","); 224 assert(buf[0] == ',', "must be"); 225 return &buf[1]; 226 } 227 228 void Compile::print_intrinsic_statistics() { 229 char flagsbuf[100]; 230 ttyLocker ttyl; 231 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 232 tty->print_cr("Compiler intrinsic usage:"); 233 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 234 if (total == 0) total = 1; // avoid div0 in case of no successes 235 #define PRINT_STAT_LINE(name, c, f) \ 236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 237 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 238 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 239 int flags = _intrinsic_hist_flags[id]; 240 juint count = _intrinsic_hist_count[id]; 241 if ((flags | count) != 0) { 242 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 243 } 244 } 245 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 246 if (xtty != NULL) xtty->tail("statistics"); 247 } 248 249 void Compile::print_statistics() { 250 { ttyLocker ttyl; 251 if (xtty != NULL) xtty->head("statistics type='opto'"); 252 Parse::print_statistics(); 253 PhaseCCP::print_statistics(); 254 PhaseRegAlloc::print_statistics(); 255 PhaseOutput::print_statistics(); 256 PhasePeephole::print_statistics(); 257 PhaseIdealLoop::print_statistics(); 258 if (xtty != NULL) xtty->tail("statistics"); 259 } 260 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 261 // put this under its own <statistics> element. 262 print_intrinsic_statistics(); 263 } 264 } 265 #endif //PRODUCT 266 267 void Compile::gvn_replace_by(Node* n, Node* nn) { 268 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 269 Node* use = n->last_out(i); 270 bool is_in_table = initial_gvn()->hash_delete(use); 271 uint uses_found = 0; 272 for (uint j = 0; j < use->len(); j++) { 273 if (use->in(j) == n) { 274 if (j < use->req()) 275 use->set_req(j, nn); 276 else 277 use->set_prec(j, nn); 278 uses_found++; 279 } 280 } 281 if (is_in_table) { 282 // reinsert into table 283 initial_gvn()->hash_find_insert(use); 284 } 285 record_for_igvn(use); 286 i -= uses_found; // we deleted 1 or more copies of this edge 287 } 288 } 289 290 291 static inline bool not_a_node(const Node* n) { 292 if (n == NULL) return true; 293 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 294 if (*(address*)n == badAddress) return true; // kill by Node::destruct 295 return false; 296 } 297 298 // Identify all nodes that are reachable from below, useful. 299 // Use breadth-first pass that records state in a Unique_Node_List, 300 // recursive traversal is slower. 301 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 302 int estimated_worklist_size = live_nodes(); 303 useful.map( estimated_worklist_size, NULL ); // preallocate space 304 305 // Initialize worklist 306 if (root() != NULL) { useful.push(root()); } 307 // If 'top' is cached, declare it useful to preserve cached node 308 if( cached_top_node() ) { useful.push(cached_top_node()); } 309 310 // Push all useful nodes onto the list, breadthfirst 311 for( uint next = 0; next < useful.size(); ++next ) { 312 assert( next < unique(), "Unique useful nodes < total nodes"); 313 Node *n = useful.at(next); 314 uint max = n->len(); 315 for( uint i = 0; i < max; ++i ) { 316 Node *m = n->in(i); 317 if (not_a_node(m)) continue; 318 useful.push(m); 319 } 320 } 321 } 322 323 // Update dead_node_list with any missing dead nodes using useful 324 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 325 void Compile::update_dead_node_list(Unique_Node_List &useful) { 326 uint max_idx = unique(); 327 VectorSet& useful_node_set = useful.member_set(); 328 329 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 330 // If node with index node_idx is not in useful set, 331 // mark it as dead in dead node list. 332 if (!useful_node_set.test(node_idx)) { 333 record_dead_node(node_idx); 334 } 335 } 336 } 337 338 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 339 int shift = 0; 340 for (int i = 0; i < inlines->length(); i++) { 341 CallGenerator* cg = inlines->at(i); 342 CallNode* call = cg->call_node(); 343 if (shift > 0) { 344 inlines->at_put(i-shift, cg); 345 } 346 if (!useful.member(call)) { 347 shift++; 348 } 349 } 350 inlines->trunc_to(inlines->length()-shift); 351 } 352 353 // Disconnect all useless nodes by disconnecting those at the boundary. 354 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 355 uint next = 0; 356 while (next < useful.size()) { 357 Node *n = useful.at(next++); 358 if (n->is_SafePoint()) { 359 // We're done with a parsing phase. Replaced nodes are not valid 360 // beyond that point. 361 n->as_SafePoint()->delete_replaced_nodes(); 362 } 363 // Use raw traversal of out edges since this code removes out edges 364 int max = n->outcnt(); 365 for (int j = 0; j < max; ++j) { 366 Node* child = n->raw_out(j); 367 if (! useful.member(child)) { 368 assert(!child->is_top() || child != top(), 369 "If top is cached in Compile object it is in useful list"); 370 // Only need to remove this out-edge to the useless node 371 n->raw_del_out(j); 372 --j; 373 --max; 374 } 375 } 376 if (n->outcnt() == 1 && n->has_special_unique_user()) { 377 record_for_igvn(n->unique_out()); 378 } 379 } 380 // Remove useless macro and predicate opaq nodes 381 for (int i = C->macro_count()-1; i >= 0; i--) { 382 Node* n = C->macro_node(i); 383 if (!useful.member(n)) { 384 remove_macro_node(n); 385 } 386 } 387 // Remove useless CastII nodes with range check dependency 388 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 389 Node* cast = range_check_cast_node(i); 390 if (!useful.member(cast)) { 391 remove_range_check_cast(cast); 392 } 393 } 394 // Remove useless expensive nodes 395 for (int i = C->expensive_count()-1; i >= 0; i--) { 396 Node* n = C->expensive_node(i); 397 if (!useful.member(n)) { 398 remove_expensive_node(n); 399 } 400 } 401 // Remove useless Opaque4 nodes 402 for (int i = opaque4_count() - 1; i >= 0; i--) { 403 Node* opaq = opaque4_node(i); 404 if (!useful.member(opaq)) { 405 remove_opaque4_node(opaq); 406 } 407 } 408 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 409 bs->eliminate_useless_gc_barriers(useful, this); 410 // clean up the late inline lists 411 remove_useless_late_inlines(&_string_late_inlines, useful); 412 remove_useless_late_inlines(&_boxing_late_inlines, useful); 413 remove_useless_late_inlines(&_late_inlines, useful); 414 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 415 } 416 417 // ============================================================================ 418 //------------------------------CompileWrapper--------------------------------- 419 class CompileWrapper : public StackObj { 420 Compile *const _compile; 421 public: 422 CompileWrapper(Compile* compile); 423 424 ~CompileWrapper(); 425 }; 426 427 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 428 // the Compile* pointer is stored in the current ciEnv: 429 ciEnv* env = compile->env(); 430 assert(env == ciEnv::current(), "must already be a ciEnv active"); 431 assert(env->compiler_data() == NULL, "compile already active?"); 432 env->set_compiler_data(compile); 433 assert(compile == Compile::current(), "sanity"); 434 435 compile->set_type_dict(NULL); 436 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 437 compile->clone_map().set_clone_idx(0); 438 compile->set_type_last_size(0); 439 compile->set_last_tf(NULL, NULL); 440 compile->set_indexSet_arena(NULL); 441 compile->set_indexSet_free_block_list(NULL); 442 compile->init_type_arena(); 443 Type::Initialize(compile); 444 _compile->begin_method(); 445 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 446 } 447 CompileWrapper::~CompileWrapper() { 448 _compile->end_method(); 449 _compile->env()->set_compiler_data(NULL); 450 } 451 452 453 //----------------------------print_compile_messages--------------------------- 454 void Compile::print_compile_messages() { 455 #ifndef PRODUCT 456 // Check if recompiling 457 if (_subsume_loads == false && PrintOpto) { 458 // Recompiling without allowing machine instructions to subsume loads 459 tty->print_cr("*********************************************************"); 460 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 461 tty->print_cr("*********************************************************"); 462 } 463 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 464 // Recompiling without escape analysis 465 tty->print_cr("*********************************************************"); 466 tty->print_cr("** Bailout: Recompile without escape analysis **"); 467 tty->print_cr("*********************************************************"); 468 } 469 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 470 // Recompiling without boxing elimination 471 tty->print_cr("*********************************************************"); 472 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 473 tty->print_cr("*********************************************************"); 474 } 475 if (C->directive()->BreakAtCompileOption) { 476 // Open the debugger when compiling this method. 477 tty->print("### Breaking when compiling: "); 478 method()->print_short_name(); 479 tty->cr(); 480 BREAKPOINT; 481 } 482 483 if( PrintOpto ) { 484 if (is_osr_compilation()) { 485 tty->print("[OSR]%3d", _compile_id); 486 } else { 487 tty->print("%3d", _compile_id); 488 } 489 } 490 #endif 491 } 492 493 // ============================================================================ 494 //------------------------------Compile standard------------------------------- 495 debug_only( int Compile::_debug_idx = 100000; ) 496 497 // Compile a method. entry_bci is -1 for normal compilations and indicates 498 // the continuation bci for on stack replacement. 499 500 501 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 502 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 503 : Phase(Compiler), 504 _compile_id(ci_env->compile_id()), 505 _save_argument_registers(false), 506 _subsume_loads(subsume_loads), 507 _do_escape_analysis(do_escape_analysis), 508 _eliminate_boxing(eliminate_boxing), 509 _method(target), 510 _entry_bci(osr_bci), 511 _stub_function(NULL), 512 _stub_name(NULL), 513 _stub_entry_point(NULL), 514 _max_node_limit(MaxNodeLimit), 515 _inlining_progress(false), 516 _inlining_incrementally(false), 517 _do_cleanup(false), 518 _has_reserved_stack_access(target->has_reserved_stack_access()), 519 #ifndef PRODUCT 520 _trace_opto_output(directive->TraceOptoOutputOption), 521 _print_ideal(directive->PrintIdealOption), 522 #endif 523 _has_method_handle_invokes(false), 524 _clinit_barrier_on_entry(false), 525 _comp_arena(mtCompiler), 526 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 527 _env(ci_env), 528 _directive(directive), 529 _log(ci_env->log()), 530 _failure_reason(NULL), 531 _congraph(NULL), 532 #ifndef PRODUCT 533 _printer(IdealGraphPrinter::printer()), 534 #endif 535 _dead_node_list(comp_arena()), 536 _dead_node_count(0), 537 _node_arena(mtCompiler), 538 _old_arena(mtCompiler), 539 _mach_constant_base_node(NULL), 540 _Compile_types(mtCompiler), 541 _initial_gvn(NULL), 542 _for_igvn(NULL), 543 _warm_calls(NULL), 544 _late_inlines(comp_arena(), 2, 0, NULL), 545 _string_late_inlines(comp_arena(), 2, 0, NULL), 546 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 547 _late_inlines_pos(0), 548 _number_of_mh_late_inlines(0), 549 _print_inlining_stream(NULL), 550 _print_inlining_list(NULL), 551 _print_inlining_idx(0), 552 _print_inlining_output(NULL), 553 _replay_inline_data(NULL), 554 _java_calls(0), 555 _inner_loops(0), 556 _interpreter_frame_size(0) 557 #ifndef PRODUCT 558 , _in_dump_cnt(0) 559 #endif 560 { 561 C = this; 562 #ifndef PRODUCT 563 if (_printer != NULL) { 564 _printer->set_compile(this); 565 } 566 #endif 567 CompileWrapper cw(this); 568 569 if (CITimeVerbose) { 570 tty->print(" "); 571 target->holder()->name()->print(); 572 tty->print("."); 573 target->print_short_name(); 574 tty->print(" "); 575 } 576 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 577 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 578 579 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 580 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 581 // We can always print a disassembly, either abstract (hex dump) or 582 // with the help of a suitable hsdis library. Thus, we should not 583 // couple print_assembly and print_opto_assembly controls. 584 // But: always print opto and regular assembly on compile command 'print'. 585 bool print_assembly = directive->PrintAssemblyOption; 586 set_print_assembly(print_opto_assembly || print_assembly); 587 #else 588 set_print_assembly(false); // must initialize. 589 #endif 590 591 #ifndef PRODUCT 592 set_parsed_irreducible_loop(false); 593 594 if (directive->ReplayInlineOption) { 595 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 596 } 597 #endif 598 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 599 set_print_intrinsics(directive->PrintIntrinsicsOption); 600 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 601 602 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 603 // Make sure the method being compiled gets its own MDO, 604 // so we can at least track the decompile_count(). 605 // Need MDO to record RTM code generation state. 606 method()->ensure_method_data(); 607 } 608 609 Init(::AliasLevel); 610 611 612 print_compile_messages(); 613 614 _ilt = InlineTree::build_inline_tree_root(); 615 616 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 617 assert(num_alias_types() >= AliasIdxRaw, ""); 618 619 #define MINIMUM_NODE_HASH 1023 620 // Node list that Iterative GVN will start with 621 Unique_Node_List for_igvn(comp_arena()); 622 set_for_igvn(&for_igvn); 623 624 // GVN that will be run immediately on new nodes 625 uint estimated_size = method()->code_size()*4+64; 626 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 627 PhaseGVN gvn(node_arena(), estimated_size); 628 set_initial_gvn(&gvn); 629 630 print_inlining_init(); 631 { // Scope for timing the parser 632 TracePhase tp("parse", &timers[_t_parser]); 633 634 // Put top into the hash table ASAP. 635 initial_gvn()->transform_no_reclaim(top()); 636 637 // Set up tf(), start(), and find a CallGenerator. 638 CallGenerator* cg = NULL; 639 if (is_osr_compilation()) { 640 const TypeTuple *domain = StartOSRNode::osr_domain(); 641 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 642 init_tf(TypeFunc::make(domain, range)); 643 StartNode* s = new StartOSRNode(root(), domain); 644 initial_gvn()->set_type_bottom(s); 645 init_start(s); 646 cg = CallGenerator::for_osr(method(), entry_bci()); 647 } else { 648 // Normal case. 649 init_tf(TypeFunc::make(method())); 650 StartNode* s = new StartNode(root(), tf()->domain()); 651 initial_gvn()->set_type_bottom(s); 652 init_start(s); 653 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 654 // With java.lang.ref.reference.get() we must go through the 655 // intrinsic - even when get() is the root 656 // method of the compile - so that, if necessary, the value in 657 // the referent field of the reference object gets recorded by 658 // the pre-barrier code. 659 cg = find_intrinsic(method(), false); 660 } 661 if (cg == NULL) { 662 float past_uses = method()->interpreter_invocation_count(); 663 float expected_uses = past_uses; 664 cg = CallGenerator::for_inline(method(), expected_uses); 665 } 666 } 667 if (failing()) return; 668 if (cg == NULL) { 669 record_method_not_compilable("cannot parse method"); 670 return; 671 } 672 JVMState* jvms = build_start_state(start(), tf()); 673 if ((jvms = cg->generate(jvms)) == NULL) { 674 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 675 record_method_not_compilable("method parse failed"); 676 } 677 return; 678 } 679 GraphKit kit(jvms); 680 681 if (!kit.stopped()) { 682 // Accept return values, and transfer control we know not where. 683 // This is done by a special, unique ReturnNode bound to root. 684 return_values(kit.jvms()); 685 } 686 687 if (kit.has_exceptions()) { 688 // Any exceptions that escape from this call must be rethrown 689 // to whatever caller is dynamically above us on the stack. 690 // This is done by a special, unique RethrowNode bound to root. 691 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 692 } 693 694 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 695 696 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 697 inline_string_calls(true); 698 } 699 700 if (failing()) return; 701 702 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 703 704 // Remove clutter produced by parsing. 705 if (!failing()) { 706 ResourceMark rm; 707 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 708 } 709 } 710 711 // Note: Large methods are capped off in do_one_bytecode(). 712 if (failing()) return; 713 714 // After parsing, node notes are no longer automagic. 715 // They must be propagated by register_new_node_with_optimizer(), 716 // clone(), or the like. 717 set_default_node_notes(NULL); 718 719 for (;;) { 720 int successes = Inline_Warm(); 721 if (failing()) return; 722 if (successes == 0) break; 723 } 724 725 // Drain the list. 726 Finish_Warm(); 727 #ifndef PRODUCT 728 if (_printer && _printer->should_print(1)) { 729 _printer->print_inlining(); 730 } 731 #endif 732 733 if (failing()) return; 734 NOT_PRODUCT( verify_graph_edges(); ) 735 736 // Now optimize 737 Optimize(); 738 if (failing()) return; 739 NOT_PRODUCT( verify_graph_edges(); ) 740 741 #ifndef PRODUCT 742 if (print_ideal()) { 743 ttyLocker ttyl; // keep the following output all in one block 744 // This output goes directly to the tty, not the compiler log. 745 // To enable tools to match it up with the compilation activity, 746 // be sure to tag this tty output with the compile ID. 747 if (xtty != NULL) { 748 xtty->head("ideal compile_id='%d'%s", compile_id(), 749 is_osr_compilation() ? " compile_kind='osr'" : 750 ""); 751 } 752 root()->dump(9999); 753 if (xtty != NULL) { 754 xtty->tail("ideal"); 755 } 756 } 757 #endif 758 759 #ifdef ASSERT 760 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 761 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 762 #endif 763 764 // Dump compilation data to replay it. 765 if (directive->DumpReplayOption) { 766 env()->dump_replay_data(_compile_id); 767 } 768 if (directive->DumpInlineOption && (ilt() != NULL)) { 769 env()->dump_inline_data(_compile_id); 770 } 771 772 // Now that we know the size of all the monitors we can add a fixed slot 773 // for the original deopt pc. 774 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 775 set_fixed_slots(next_slot); 776 777 // Compute when to use implicit null checks. Used by matching trap based 778 // nodes and NullCheck optimization. 779 set_allowed_deopt_reasons(); 780 781 // Now generate code 782 Code_Gen(); 783 } 784 785 //------------------------------Compile---------------------------------------- 786 // Compile a runtime stub 787 Compile::Compile( ciEnv* ci_env, 788 TypeFunc_generator generator, 789 address stub_function, 790 const char *stub_name, 791 int is_fancy_jump, 792 bool pass_tls, 793 bool save_arg_registers, 794 bool return_pc, 795 DirectiveSet* directive) 796 : Phase(Compiler), 797 _compile_id(0), 798 _save_argument_registers(save_arg_registers), 799 _subsume_loads(true), 800 _do_escape_analysis(false), 801 _eliminate_boxing(false), 802 _method(NULL), 803 _entry_bci(InvocationEntryBci), 804 _stub_function(stub_function), 805 _stub_name(stub_name), 806 _stub_entry_point(NULL), 807 _max_node_limit(MaxNodeLimit), 808 _inlining_progress(false), 809 _inlining_incrementally(false), 810 _has_reserved_stack_access(false), 811 #ifndef PRODUCT 812 _trace_opto_output(directive->TraceOptoOutputOption), 813 _print_ideal(directive->PrintIdealOption), 814 #endif 815 _has_method_handle_invokes(false), 816 _clinit_barrier_on_entry(false), 817 _comp_arena(mtCompiler), 818 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 819 _env(ci_env), 820 _directive(directive), 821 _log(ci_env->log()), 822 _failure_reason(NULL), 823 _congraph(NULL), 824 #ifndef PRODUCT 825 _printer(NULL), 826 #endif 827 _dead_node_list(comp_arena()), 828 _dead_node_count(0), 829 _node_arena(mtCompiler), 830 _old_arena(mtCompiler), 831 _mach_constant_base_node(NULL), 832 _Compile_types(mtCompiler), 833 _initial_gvn(NULL), 834 _for_igvn(NULL), 835 _warm_calls(NULL), 836 _number_of_mh_late_inlines(0), 837 _print_inlining_stream(NULL), 838 _print_inlining_list(NULL), 839 _print_inlining_idx(0), 840 _print_inlining_output(NULL), 841 _replay_inline_data(NULL), 842 _java_calls(0), 843 _inner_loops(0), 844 _interpreter_frame_size(0), 845 #ifndef PRODUCT 846 _in_dump_cnt(0), 847 #endif 848 _allowed_reasons(0) { 849 C = this; 850 851 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 852 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 853 854 #ifndef PRODUCT 855 set_print_assembly(PrintFrameConverterAssembly); 856 set_parsed_irreducible_loop(false); 857 #else 858 set_print_assembly(false); // Must initialize. 859 #endif 860 set_has_irreducible_loop(false); // no loops 861 862 CompileWrapper cw(this); 863 Init(/*AliasLevel=*/ 0); 864 init_tf((*generator)()); 865 866 { 867 // The following is a dummy for the sake of GraphKit::gen_stub 868 Unique_Node_List for_igvn(comp_arena()); 869 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 870 PhaseGVN gvn(Thread::current()->resource_area(),255); 871 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 872 gvn.transform_no_reclaim(top()); 873 874 GraphKit kit; 875 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 876 } 877 878 NOT_PRODUCT( verify_graph_edges(); ) 879 880 Code_Gen(); 881 } 882 883 //------------------------------Init------------------------------------------- 884 // Prepare for a single compilation 885 void Compile::Init(int aliaslevel) { 886 _unique = 0; 887 _regalloc = NULL; 888 889 _tf = NULL; // filled in later 890 _top = NULL; // cached later 891 _matcher = NULL; // filled in later 892 _cfg = NULL; // filled in later 893 894 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 895 896 _node_note_array = NULL; 897 _default_node_notes = NULL; 898 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 899 900 _immutable_memory = NULL; // filled in at first inquiry 901 902 // Globally visible Nodes 903 // First set TOP to NULL to give safe behavior during creation of RootNode 904 set_cached_top_node(NULL); 905 set_root(new RootNode()); 906 // Now that you have a Root to point to, create the real TOP 907 set_cached_top_node( new ConNode(Type::TOP) ); 908 set_recent_alloc(NULL, NULL); 909 910 // Create Debug Information Recorder to record scopes, oopmaps, etc. 911 env()->set_oop_recorder(new OopRecorder(env()->arena())); 912 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 913 env()->set_dependencies(new Dependencies(env())); 914 915 _fixed_slots = 0; 916 set_has_split_ifs(false); 917 set_has_loops(has_method() && method()->has_loops()); // first approximation 918 set_has_stringbuilder(false); 919 set_has_boxed_value(false); 920 _trap_can_recompile = false; // no traps emitted yet 921 _major_progress = true; // start out assuming good things will happen 922 set_has_unsafe_access(false); 923 set_max_vector_size(0); 924 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 925 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 926 set_decompile_count(0); 927 928 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 929 _loop_opts_cnt = LoopOptsCount; 930 set_do_inlining(Inline); 931 set_max_inline_size(MaxInlineSize); 932 set_freq_inline_size(FreqInlineSize); 933 set_do_scheduling(OptoScheduling); 934 set_do_count_invocations(false); 935 set_do_method_data_update(false); 936 937 set_do_vector_loop(false); 938 939 if (AllowVectorizeOnDemand) { 940 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 941 set_do_vector_loop(true); 942 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());}) 943 } else if (has_method() && method()->name() != 0 && 944 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 945 set_do_vector_loop(true); 946 } 947 } 948 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 949 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());}) 950 951 set_age_code(has_method() && method()->profile_aging()); 952 set_rtm_state(NoRTM); // No RTM lock eliding by default 953 _max_node_limit = _directive->MaxNodeLimitOption; 954 955 #if INCLUDE_RTM_OPT 956 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 957 int rtm_state = method()->method_data()->rtm_state(); 958 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 959 // Don't generate RTM lock eliding code. 960 set_rtm_state(NoRTM); 961 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 962 // Generate RTM lock eliding code without abort ratio calculation code. 963 set_rtm_state(UseRTM); 964 } else if (UseRTMDeopt) { 965 // Generate RTM lock eliding code and include abort ratio calculation 966 // code if UseRTMDeopt is on. 967 set_rtm_state(ProfileRTM); 968 } 969 } 970 #endif 971 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 972 set_clinit_barrier_on_entry(true); 973 } 974 if (debug_info()->recording_non_safepoints()) { 975 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 976 (comp_arena(), 8, 0, NULL)); 977 set_default_node_notes(Node_Notes::make(this)); 978 } 979 980 // // -- Initialize types before each compile -- 981 // // Update cached type information 982 // if( _method && _method->constants() ) 983 // Type::update_loaded_types(_method, _method->constants()); 984 985 // Init alias_type map. 986 if (!_do_escape_analysis && aliaslevel == 3) 987 aliaslevel = 2; // No unique types without escape analysis 988 _AliasLevel = aliaslevel; 989 const int grow_ats = 16; 990 _max_alias_types = grow_ats; 991 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 992 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 993 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 994 { 995 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 996 } 997 // Initialize the first few types. 998 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 999 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1000 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1001 _num_alias_types = AliasIdxRaw+1; 1002 // Zero out the alias type cache. 1003 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1004 // A NULL adr_type hits in the cache right away. Preload the right answer. 1005 probe_alias_cache(NULL)->_index = AliasIdxTop; 1006 1007 _intrinsics = NULL; 1008 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1009 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1010 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1011 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1012 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1013 register_library_intrinsics(); 1014 #ifdef ASSERT 1015 _type_verify_symmetry = true; 1016 #endif 1017 } 1018 1019 //---------------------------init_start---------------------------------------- 1020 // Install the StartNode on this compile object. 1021 void Compile::init_start(StartNode* s) { 1022 if (failing()) 1023 return; // already failing 1024 assert(s == start(), ""); 1025 } 1026 1027 /** 1028 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1029 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1030 * the ideal graph. 1031 */ 1032 StartNode* Compile::start() const { 1033 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1034 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1035 Node* start = root()->fast_out(i); 1036 if (start->is_Start()) { 1037 return start->as_Start(); 1038 } 1039 } 1040 fatal("Did not find Start node!"); 1041 return NULL; 1042 } 1043 1044 //-------------------------------immutable_memory------------------------------------- 1045 // Access immutable memory 1046 Node* Compile::immutable_memory() { 1047 if (_immutable_memory != NULL) { 1048 return _immutable_memory; 1049 } 1050 StartNode* s = start(); 1051 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1052 Node *p = s->fast_out(i); 1053 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1054 _immutable_memory = p; 1055 return _immutable_memory; 1056 } 1057 } 1058 ShouldNotReachHere(); 1059 return NULL; 1060 } 1061 1062 //----------------------set_cached_top_node------------------------------------ 1063 // Install the cached top node, and make sure Node::is_top works correctly. 1064 void Compile::set_cached_top_node(Node* tn) { 1065 if (tn != NULL) verify_top(tn); 1066 Node* old_top = _top; 1067 _top = tn; 1068 // Calling Node::setup_is_top allows the nodes the chance to adjust 1069 // their _out arrays. 1070 if (_top != NULL) _top->setup_is_top(); 1071 if (old_top != NULL) old_top->setup_is_top(); 1072 assert(_top == NULL || top()->is_top(), ""); 1073 } 1074 1075 #ifdef ASSERT 1076 uint Compile::count_live_nodes_by_graph_walk() { 1077 Unique_Node_List useful(comp_arena()); 1078 // Get useful node list by walking the graph. 1079 identify_useful_nodes(useful); 1080 return useful.size(); 1081 } 1082 1083 void Compile::print_missing_nodes() { 1084 1085 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1086 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1087 return; 1088 } 1089 1090 // This is an expensive function. It is executed only when the user 1091 // specifies VerifyIdealNodeCount option or otherwise knows the 1092 // additional work that needs to be done to identify reachable nodes 1093 // by walking the flow graph and find the missing ones using 1094 // _dead_node_list. 1095 1096 Unique_Node_List useful(comp_arena()); 1097 // Get useful node list by walking the graph. 1098 identify_useful_nodes(useful); 1099 1100 uint l_nodes = C->live_nodes(); 1101 uint l_nodes_by_walk = useful.size(); 1102 1103 if (l_nodes != l_nodes_by_walk) { 1104 if (_log != NULL) { 1105 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1106 _log->stamp(); 1107 _log->end_head(); 1108 } 1109 VectorSet& useful_member_set = useful.member_set(); 1110 int last_idx = l_nodes_by_walk; 1111 for (int i = 0; i < last_idx; i++) { 1112 if (useful_member_set.test(i)) { 1113 if (_dead_node_list.test(i)) { 1114 if (_log != NULL) { 1115 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1116 } 1117 if (PrintIdealNodeCount) { 1118 // Print the log message to tty 1119 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1120 useful.at(i)->dump(); 1121 } 1122 } 1123 } 1124 else if (! _dead_node_list.test(i)) { 1125 if (_log != NULL) { 1126 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1127 } 1128 if (PrintIdealNodeCount) { 1129 // Print the log message to tty 1130 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1131 } 1132 } 1133 } 1134 if (_log != NULL) { 1135 _log->tail("mismatched_nodes"); 1136 } 1137 } 1138 } 1139 void Compile::record_modified_node(Node* n) { 1140 if (_modified_nodes != NULL && !_inlining_incrementally && 1141 n->outcnt() != 0 && !n->is_Con()) { 1142 _modified_nodes->push(n); 1143 } 1144 } 1145 1146 void Compile::remove_modified_node(Node* n) { 1147 if (_modified_nodes != NULL) { 1148 _modified_nodes->remove(n); 1149 } 1150 } 1151 #endif 1152 1153 #ifndef PRODUCT 1154 void Compile::verify_top(Node* tn) const { 1155 if (tn != NULL) { 1156 assert(tn->is_Con(), "top node must be a constant"); 1157 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1158 assert(tn->in(0) != NULL, "must have live top node"); 1159 } 1160 } 1161 #endif 1162 1163 1164 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1165 1166 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1167 guarantee(arr != NULL, ""); 1168 int num_blocks = arr->length(); 1169 if (grow_by < num_blocks) grow_by = num_blocks; 1170 int num_notes = grow_by * _node_notes_block_size; 1171 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1172 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1173 while (num_notes > 0) { 1174 arr->append(notes); 1175 notes += _node_notes_block_size; 1176 num_notes -= _node_notes_block_size; 1177 } 1178 assert(num_notes == 0, "exact multiple, please"); 1179 } 1180 1181 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1182 if (source == NULL || dest == NULL) return false; 1183 1184 if (dest->is_Con()) 1185 return false; // Do not push debug info onto constants. 1186 1187 #ifdef ASSERT 1188 // Leave a bread crumb trail pointing to the original node: 1189 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1190 dest->set_debug_orig(source); 1191 } 1192 #endif 1193 1194 if (node_note_array() == NULL) 1195 return false; // Not collecting any notes now. 1196 1197 // This is a copy onto a pre-existing node, which may already have notes. 1198 // If both nodes have notes, do not overwrite any pre-existing notes. 1199 Node_Notes* source_notes = node_notes_at(source->_idx); 1200 if (source_notes == NULL || source_notes->is_clear()) return false; 1201 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1202 if (dest_notes == NULL || dest_notes->is_clear()) { 1203 return set_node_notes_at(dest->_idx, source_notes); 1204 } 1205 1206 Node_Notes merged_notes = (*source_notes); 1207 // The order of operations here ensures that dest notes will win... 1208 merged_notes.update_from(dest_notes); 1209 return set_node_notes_at(dest->_idx, &merged_notes); 1210 } 1211 1212 1213 //--------------------------allow_range_check_smearing------------------------- 1214 // Gating condition for coalescing similar range checks. 1215 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1216 // single covering check that is at least as strong as any of them. 1217 // If the optimization succeeds, the simplified (strengthened) range check 1218 // will always succeed. If it fails, we will deopt, and then give up 1219 // on the optimization. 1220 bool Compile::allow_range_check_smearing() const { 1221 // If this method has already thrown a range-check, 1222 // assume it was because we already tried range smearing 1223 // and it failed. 1224 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1225 return !already_trapped; 1226 } 1227 1228 1229 //------------------------------flatten_alias_type----------------------------- 1230 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1231 int offset = tj->offset(); 1232 TypePtr::PTR ptr = tj->ptr(); 1233 1234 // Known instance (scalarizable allocation) alias only with itself. 1235 bool is_known_inst = tj->isa_oopptr() != NULL && 1236 tj->is_oopptr()->is_known_instance(); 1237 1238 // Process weird unsafe references. 1239 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1240 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1241 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1242 tj = TypeOopPtr::BOTTOM; 1243 ptr = tj->ptr(); 1244 offset = tj->offset(); 1245 } 1246 1247 // Array pointers need some flattening 1248 const TypeAryPtr *ta = tj->isa_aryptr(); 1249 if (ta && ta->is_stable()) { 1250 // Erase stability property for alias analysis. 1251 tj = ta = ta->cast_to_stable(false); 1252 } 1253 if( ta && is_known_inst ) { 1254 if ( offset != Type::OffsetBot && 1255 offset > arrayOopDesc::length_offset_in_bytes() ) { 1256 offset = Type::OffsetBot; // Flatten constant access into array body only 1257 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1258 } 1259 } else if( ta && _AliasLevel >= 2 ) { 1260 // For arrays indexed by constant indices, we flatten the alias 1261 // space to include all of the array body. Only the header, klass 1262 // and array length can be accessed un-aliased. 1263 if( offset != Type::OffsetBot ) { 1264 if( ta->const_oop() ) { // MethodData* or Method* 1265 offset = Type::OffsetBot; // Flatten constant access into array body 1266 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1267 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1268 // range is OK as-is. 1269 tj = ta = TypeAryPtr::RANGE; 1270 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1271 tj = TypeInstPtr::KLASS; // all klass loads look alike 1272 ta = TypeAryPtr::RANGE; // generic ignored junk 1273 ptr = TypePtr::BotPTR; 1274 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1275 tj = TypeInstPtr::MARK; 1276 ta = TypeAryPtr::RANGE; // generic ignored junk 1277 ptr = TypePtr::BotPTR; 1278 } else { // Random constant offset into array body 1279 offset = Type::OffsetBot; // Flatten constant access into array body 1280 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1281 } 1282 } 1283 // Arrays of fixed size alias with arrays of unknown size. 1284 if (ta->size() != TypeInt::POS) { 1285 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1286 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1287 } 1288 // Arrays of known objects become arrays of unknown objects. 1289 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1290 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1291 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1292 } 1293 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1294 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1295 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1296 } 1297 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1298 // cannot be distinguished by bytecode alone. 1299 if (ta->elem() == TypeInt::BOOL) { 1300 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1301 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1302 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1303 } 1304 // During the 2nd round of IterGVN, NotNull castings are removed. 1305 // Make sure the Bottom and NotNull variants alias the same. 1306 // Also, make sure exact and non-exact variants alias the same. 1307 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1308 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1309 } 1310 } 1311 1312 // Oop pointers need some flattening 1313 const TypeInstPtr *to = tj->isa_instptr(); 1314 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1315 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1316 if( ptr == TypePtr::Constant ) { 1317 if (to->klass() != ciEnv::current()->Class_klass() || 1318 offset < k->size_helper() * wordSize) { 1319 // No constant oop pointers (such as Strings); they alias with 1320 // unknown strings. 1321 assert(!is_known_inst, "not scalarizable allocation"); 1322 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1323 } 1324 } else if( is_known_inst ) { 1325 tj = to; // Keep NotNull and klass_is_exact for instance type 1326 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1327 // During the 2nd round of IterGVN, NotNull castings are removed. 1328 // Make sure the Bottom and NotNull variants alias the same. 1329 // Also, make sure exact and non-exact variants alias the same. 1330 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1331 } 1332 if (to->speculative() != NULL) { 1333 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1334 } 1335 // Canonicalize the holder of this field 1336 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1337 // First handle header references such as a LoadKlassNode, even if the 1338 // object's klass is unloaded at compile time (4965979). 1339 if (!is_known_inst) { // Do it only for non-instance types 1340 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1341 } 1342 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1343 // Static fields are in the space above the normal instance 1344 // fields in the java.lang.Class instance. 1345 if (to->klass() != ciEnv::current()->Class_klass()) { 1346 to = NULL; 1347 tj = TypeOopPtr::BOTTOM; 1348 offset = tj->offset(); 1349 } 1350 } else { 1351 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1352 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1353 if( is_known_inst ) { 1354 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1355 } else { 1356 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1357 } 1358 } 1359 } 1360 } 1361 1362 // Klass pointers to object array klasses need some flattening 1363 const TypeKlassPtr *tk = tj->isa_klassptr(); 1364 if( tk ) { 1365 // If we are referencing a field within a Klass, we need 1366 // to assume the worst case of an Object. Both exact and 1367 // inexact types must flatten to the same alias class so 1368 // use NotNull as the PTR. 1369 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1370 1371 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1372 TypeKlassPtr::OBJECT->klass(), 1373 offset); 1374 } 1375 1376 ciKlass* klass = tk->klass(); 1377 if( klass->is_obj_array_klass() ) { 1378 ciKlass* k = TypeAryPtr::OOPS->klass(); 1379 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1380 k = TypeInstPtr::BOTTOM->klass(); 1381 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1382 } 1383 1384 // Check for precise loads from the primary supertype array and force them 1385 // to the supertype cache alias index. Check for generic array loads from 1386 // the primary supertype array and also force them to the supertype cache 1387 // alias index. Since the same load can reach both, we need to merge 1388 // these 2 disparate memories into the same alias class. Since the 1389 // primary supertype array is read-only, there's no chance of confusion 1390 // where we bypass an array load and an array store. 1391 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1392 if (offset == Type::OffsetBot || 1393 (offset >= primary_supers_offset && 1394 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1395 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1396 offset = in_bytes(Klass::secondary_super_cache_offset()); 1397 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1398 } 1399 } 1400 1401 // Flatten all Raw pointers together. 1402 if (tj->base() == Type::RawPtr) 1403 tj = TypeRawPtr::BOTTOM; 1404 1405 if (tj->base() == Type::AnyPtr) 1406 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1407 1408 // Flatten all to bottom for now 1409 switch( _AliasLevel ) { 1410 case 0: 1411 tj = TypePtr::BOTTOM; 1412 break; 1413 case 1: // Flatten to: oop, static, field or array 1414 switch (tj->base()) { 1415 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1416 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1417 case Type::AryPtr: // do not distinguish arrays at all 1418 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1419 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1420 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1421 default: ShouldNotReachHere(); 1422 } 1423 break; 1424 case 2: // No collapsing at level 2; keep all splits 1425 case 3: // No collapsing at level 3; keep all splits 1426 break; 1427 default: 1428 Unimplemented(); 1429 } 1430 1431 offset = tj->offset(); 1432 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1433 1434 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1435 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1436 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1437 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1438 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1439 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1440 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1441 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1442 assert( tj->ptr() != TypePtr::TopPTR && 1443 tj->ptr() != TypePtr::AnyNull && 1444 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1445 // assert( tj->ptr() != TypePtr::Constant || 1446 // tj->base() == Type::RawPtr || 1447 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1448 1449 return tj; 1450 } 1451 1452 void Compile::AliasType::Init(int i, const TypePtr* at) { 1453 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1454 _index = i; 1455 _adr_type = at; 1456 _field = NULL; 1457 _element = NULL; 1458 _is_rewritable = true; // default 1459 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1460 if (atoop != NULL && atoop->is_known_instance()) { 1461 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1462 _general_index = Compile::current()->get_alias_index(gt); 1463 } else { 1464 _general_index = 0; 1465 } 1466 } 1467 1468 BasicType Compile::AliasType::basic_type() const { 1469 if (element() != NULL) { 1470 const Type* element = adr_type()->is_aryptr()->elem(); 1471 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1472 } if (field() != NULL) { 1473 return field()->layout_type(); 1474 } else { 1475 return T_ILLEGAL; // unknown 1476 } 1477 } 1478 1479 //---------------------------------print_on------------------------------------ 1480 #ifndef PRODUCT 1481 void Compile::AliasType::print_on(outputStream* st) { 1482 if (index() < 10) 1483 st->print("@ <%d> ", index()); 1484 else st->print("@ <%d>", index()); 1485 st->print(is_rewritable() ? " " : " RO"); 1486 int offset = adr_type()->offset(); 1487 if (offset == Type::OffsetBot) 1488 st->print(" +any"); 1489 else st->print(" +%-3d", offset); 1490 st->print(" in "); 1491 adr_type()->dump_on(st); 1492 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1493 if (field() != NULL && tjp) { 1494 if (tjp->klass() != field()->holder() || 1495 tjp->offset() != field()->offset_in_bytes()) { 1496 st->print(" != "); 1497 field()->print(); 1498 st->print(" ***"); 1499 } 1500 } 1501 } 1502 1503 void print_alias_types() { 1504 Compile* C = Compile::current(); 1505 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1506 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1507 C->alias_type(idx)->print_on(tty); 1508 tty->cr(); 1509 } 1510 } 1511 #endif 1512 1513 1514 //----------------------------probe_alias_cache-------------------------------- 1515 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1516 intptr_t key = (intptr_t) adr_type; 1517 key ^= key >> logAliasCacheSize; 1518 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1519 } 1520 1521 1522 //-----------------------------grow_alias_types-------------------------------- 1523 void Compile::grow_alias_types() { 1524 const int old_ats = _max_alias_types; // how many before? 1525 const int new_ats = old_ats; // how many more? 1526 const int grow_ats = old_ats+new_ats; // how many now? 1527 _max_alias_types = grow_ats; 1528 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1529 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1530 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1531 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1532 } 1533 1534 1535 //--------------------------------find_alias_type------------------------------ 1536 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1537 if (_AliasLevel == 0) 1538 return alias_type(AliasIdxBot); 1539 1540 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1541 if (ace->_adr_type == adr_type) { 1542 return alias_type(ace->_index); 1543 } 1544 1545 // Handle special cases. 1546 if (adr_type == NULL) return alias_type(AliasIdxTop); 1547 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1548 1549 // Do it the slow way. 1550 const TypePtr* flat = flatten_alias_type(adr_type); 1551 1552 #ifdef ASSERT 1553 { 1554 ResourceMark rm; 1555 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1556 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1557 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1558 Type::str(adr_type)); 1559 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1560 const TypeOopPtr* foop = flat->is_oopptr(); 1561 // Scalarizable allocations have exact klass always. 1562 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1563 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1564 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1565 Type::str(foop), Type::str(xoop)); 1566 } 1567 } 1568 #endif 1569 1570 int idx = AliasIdxTop; 1571 for (int i = 0; i < num_alias_types(); i++) { 1572 if (alias_type(i)->adr_type() == flat) { 1573 idx = i; 1574 break; 1575 } 1576 } 1577 1578 if (idx == AliasIdxTop) { 1579 if (no_create) return NULL; 1580 // Grow the array if necessary. 1581 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1582 // Add a new alias type. 1583 idx = _num_alias_types++; 1584 _alias_types[idx]->Init(idx, flat); 1585 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1586 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1587 if (flat->isa_instptr()) { 1588 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1589 && flat->is_instptr()->klass() == env()->Class_klass()) 1590 alias_type(idx)->set_rewritable(false); 1591 } 1592 if (flat->isa_aryptr()) { 1593 #ifdef ASSERT 1594 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1595 // (T_BYTE has the weakest alignment and size restrictions...) 1596 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1597 #endif 1598 if (flat->offset() == TypePtr::OffsetBot) { 1599 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1600 } 1601 } 1602 if (flat->isa_klassptr()) { 1603 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1604 alias_type(idx)->set_rewritable(false); 1605 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1606 alias_type(idx)->set_rewritable(false); 1607 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1608 alias_type(idx)->set_rewritable(false); 1609 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1610 alias_type(idx)->set_rewritable(false); 1611 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1612 alias_type(idx)->set_rewritable(false); 1613 } 1614 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1615 // but the base pointer type is not distinctive enough to identify 1616 // references into JavaThread.) 1617 1618 // Check for final fields. 1619 const TypeInstPtr* tinst = flat->isa_instptr(); 1620 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1621 ciField* field; 1622 if (tinst->const_oop() != NULL && 1623 tinst->klass() == ciEnv::current()->Class_klass() && 1624 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1625 // static field 1626 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1627 field = k->get_field_by_offset(tinst->offset(), true); 1628 } else { 1629 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1630 field = k->get_field_by_offset(tinst->offset(), false); 1631 } 1632 assert(field == NULL || 1633 original_field == NULL || 1634 (field->holder() == original_field->holder() && 1635 field->offset() == original_field->offset() && 1636 field->is_static() == original_field->is_static()), "wrong field?"); 1637 // Set field() and is_rewritable() attributes. 1638 if (field != NULL) alias_type(idx)->set_field(field); 1639 } 1640 } 1641 1642 // Fill the cache for next time. 1643 ace->_adr_type = adr_type; 1644 ace->_index = idx; 1645 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1646 1647 // Might as well try to fill the cache for the flattened version, too. 1648 AliasCacheEntry* face = probe_alias_cache(flat); 1649 if (face->_adr_type == NULL) { 1650 face->_adr_type = flat; 1651 face->_index = idx; 1652 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1653 } 1654 1655 return alias_type(idx); 1656 } 1657 1658 1659 Compile::AliasType* Compile::alias_type(ciField* field) { 1660 const TypeOopPtr* t; 1661 if (field->is_static()) 1662 t = TypeInstPtr::make(field->holder()->java_mirror()); 1663 else 1664 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1665 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1666 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1667 return atp; 1668 } 1669 1670 1671 //------------------------------have_alias_type-------------------------------- 1672 bool Compile::have_alias_type(const TypePtr* adr_type) { 1673 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1674 if (ace->_adr_type == adr_type) { 1675 return true; 1676 } 1677 1678 // Handle special cases. 1679 if (adr_type == NULL) return true; 1680 if (adr_type == TypePtr::BOTTOM) return true; 1681 1682 return find_alias_type(adr_type, true, NULL) != NULL; 1683 } 1684 1685 //-----------------------------must_alias-------------------------------------- 1686 // True if all values of the given address type are in the given alias category. 1687 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1688 if (alias_idx == AliasIdxBot) return true; // the universal category 1689 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1690 if (alias_idx == AliasIdxTop) return false; // the empty category 1691 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1692 1693 // the only remaining possible overlap is identity 1694 int adr_idx = get_alias_index(adr_type); 1695 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1696 assert(adr_idx == alias_idx || 1697 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1698 && adr_type != TypeOopPtr::BOTTOM), 1699 "should not be testing for overlap with an unsafe pointer"); 1700 return adr_idx == alias_idx; 1701 } 1702 1703 //------------------------------can_alias-------------------------------------- 1704 // True if any values of the given address type are in the given alias category. 1705 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1706 if (alias_idx == AliasIdxTop) return false; // the empty category 1707 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1708 // Known instance doesn't alias with bottom memory 1709 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1710 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1711 1712 // the only remaining possible overlap is identity 1713 int adr_idx = get_alias_index(adr_type); 1714 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1715 return adr_idx == alias_idx; 1716 } 1717 1718 1719 1720 //---------------------------pop_warm_call------------------------------------- 1721 WarmCallInfo* Compile::pop_warm_call() { 1722 WarmCallInfo* wci = _warm_calls; 1723 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1724 return wci; 1725 } 1726 1727 //----------------------------Inline_Warm-------------------------------------- 1728 int Compile::Inline_Warm() { 1729 // If there is room, try to inline some more warm call sites. 1730 // %%% Do a graph index compaction pass when we think we're out of space? 1731 if (!InlineWarmCalls) return 0; 1732 1733 int calls_made_hot = 0; 1734 int room_to_grow = NodeCountInliningCutoff - unique(); 1735 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1736 int amount_grown = 0; 1737 WarmCallInfo* call; 1738 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1739 int est_size = (int)call->size(); 1740 if (est_size > (room_to_grow - amount_grown)) { 1741 // This one won't fit anyway. Get rid of it. 1742 call->make_cold(); 1743 continue; 1744 } 1745 call->make_hot(); 1746 calls_made_hot++; 1747 amount_grown += est_size; 1748 amount_to_grow -= est_size; 1749 } 1750 1751 if (calls_made_hot > 0) set_major_progress(); 1752 return calls_made_hot; 1753 } 1754 1755 1756 //----------------------------Finish_Warm-------------------------------------- 1757 void Compile::Finish_Warm() { 1758 if (!InlineWarmCalls) return; 1759 if (failing()) return; 1760 if (warm_calls() == NULL) return; 1761 1762 // Clean up loose ends, if we are out of space for inlining. 1763 WarmCallInfo* call; 1764 while ((call = pop_warm_call()) != NULL) { 1765 call->make_cold(); 1766 } 1767 } 1768 1769 //---------------------cleanup_loop_predicates----------------------- 1770 // Remove the opaque nodes that protect the predicates so that all unused 1771 // checks and uncommon_traps will be eliminated from the ideal graph 1772 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1773 if (predicate_count()==0) return; 1774 for (int i = predicate_count(); i > 0; i--) { 1775 Node * n = predicate_opaque1_node(i-1); 1776 assert(n->Opcode() == Op_Opaque1, "must be"); 1777 igvn.replace_node(n, n->in(1)); 1778 } 1779 assert(predicate_count()==0, "should be clean!"); 1780 } 1781 1782 void Compile::add_range_check_cast(Node* n) { 1783 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1784 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1785 _range_check_casts->append(n); 1786 } 1787 1788 // Remove all range check dependent CastIINodes. 1789 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1790 for (int i = range_check_cast_count(); i > 0; i--) { 1791 Node* cast = range_check_cast_node(i-1); 1792 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1793 igvn.replace_node(cast, cast->in(1)); 1794 } 1795 assert(range_check_cast_count() == 0, "should be empty"); 1796 } 1797 1798 void Compile::add_opaque4_node(Node* n) { 1799 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1800 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1801 _opaque4_nodes->append(n); 1802 } 1803 1804 // Remove all Opaque4 nodes. 1805 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1806 for (int i = opaque4_count(); i > 0; i--) { 1807 Node* opaq = opaque4_node(i-1); 1808 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 1809 igvn.replace_node(opaq, opaq->in(2)); 1810 } 1811 assert(opaque4_count() == 0, "should be empty"); 1812 } 1813 1814 // StringOpts and late inlining of string methods 1815 void Compile::inline_string_calls(bool parse_time) { 1816 { 1817 // remove useless nodes to make the usage analysis simpler 1818 ResourceMark rm; 1819 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1820 } 1821 1822 { 1823 ResourceMark rm; 1824 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1825 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1826 print_method(PHASE_AFTER_STRINGOPTS, 3); 1827 } 1828 1829 // now inline anything that we skipped the first time around 1830 if (!parse_time) { 1831 _late_inlines_pos = _late_inlines.length(); 1832 } 1833 1834 while (_string_late_inlines.length() > 0) { 1835 CallGenerator* cg = _string_late_inlines.pop(); 1836 cg->do_late_inline(); 1837 if (failing()) return; 1838 } 1839 _string_late_inlines.trunc_to(0); 1840 } 1841 1842 // Late inlining of boxing methods 1843 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1844 if (_boxing_late_inlines.length() > 0) { 1845 assert(has_boxed_value(), "inconsistent"); 1846 1847 PhaseGVN* gvn = initial_gvn(); 1848 set_inlining_incrementally(true); 1849 1850 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1851 for_igvn()->clear(); 1852 gvn->replace_with(&igvn); 1853 1854 _late_inlines_pos = _late_inlines.length(); 1855 1856 while (_boxing_late_inlines.length() > 0) { 1857 CallGenerator* cg = _boxing_late_inlines.pop(); 1858 cg->do_late_inline(); 1859 if (failing()) return; 1860 } 1861 _boxing_late_inlines.trunc_to(0); 1862 1863 inline_incrementally_cleanup(igvn); 1864 1865 set_inlining_incrementally(false); 1866 } 1867 } 1868 1869 bool Compile::inline_incrementally_one() { 1870 assert(IncrementalInline, "incremental inlining should be on"); 1871 1872 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1873 set_inlining_progress(false); 1874 set_do_cleanup(false); 1875 int i = 0; 1876 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1877 CallGenerator* cg = _late_inlines.at(i); 1878 _late_inlines_pos = i+1; 1879 cg->do_late_inline(); 1880 if (failing()) return false; 1881 } 1882 int j = 0; 1883 for (; i < _late_inlines.length(); i++, j++) { 1884 _late_inlines.at_put(j, _late_inlines.at(i)); 1885 } 1886 _late_inlines.trunc_to(j); 1887 assert(inlining_progress() || _late_inlines.length() == 0, ""); 1888 1889 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 1890 1891 set_inlining_progress(false); 1892 set_do_cleanup(false); 1893 return (_late_inlines.length() > 0) && !needs_cleanup; 1894 } 1895 1896 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 1897 { 1898 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1899 ResourceMark rm; 1900 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1901 } 1902 { 1903 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1904 igvn = PhaseIterGVN(initial_gvn()); 1905 igvn.optimize(); 1906 } 1907 } 1908 1909 // Perform incremental inlining until bound on number of live nodes is reached 1910 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1911 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1912 1913 set_inlining_incrementally(true); 1914 uint low_live_nodes = 0; 1915 1916 while (_late_inlines.length() > 0) { 1917 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1918 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1919 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 1920 // PhaseIdealLoop is expensive so we only try it once we are 1921 // out of live nodes and we only try it again if the previous 1922 // helped got the number of nodes down significantly 1923 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 1924 if (failing()) return; 1925 low_live_nodes = live_nodes(); 1926 _major_progress = true; 1927 } 1928 1929 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1930 break; // finish 1931 } 1932 } 1933 1934 for_igvn()->clear(); 1935 initial_gvn()->replace_with(&igvn); 1936 1937 while (inline_incrementally_one()) { 1938 assert(!failing(), "inconsistent"); 1939 } 1940 1941 if (failing()) return; 1942 1943 inline_incrementally_cleanup(igvn); 1944 1945 if (failing()) return; 1946 } 1947 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1948 1949 if (_string_late_inlines.length() > 0) { 1950 assert(has_stringbuilder(), "inconsistent"); 1951 for_igvn()->clear(); 1952 initial_gvn()->replace_with(&igvn); 1953 1954 inline_string_calls(false); 1955 1956 if (failing()) return; 1957 1958 inline_incrementally_cleanup(igvn); 1959 } 1960 1961 set_inlining_incrementally(false); 1962 } 1963 1964 1965 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 1966 if(_loop_opts_cnt > 0) { 1967 debug_only( int cnt = 0; ); 1968 while(major_progress() && (_loop_opts_cnt > 0)) { 1969 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 1970 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1971 PhaseIdealLoop::optimize(igvn, mode); 1972 _loop_opts_cnt--; 1973 if (failing()) return false; 1974 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 1975 } 1976 } 1977 return true; 1978 } 1979 1980 // Remove edges from "root" to each SafePoint at a backward branch. 1981 // They were inserted during parsing (see add_safepoint()) to make 1982 // infinite loops without calls or exceptions visible to root, i.e., 1983 // useful. 1984 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 1985 Node *r = root(); 1986 if (r != NULL) { 1987 for (uint i = r->req(); i < r->len(); ++i) { 1988 Node *n = r->in(i); 1989 if (n != NULL && n->is_SafePoint()) { 1990 r->rm_prec(i); 1991 if (n->outcnt() == 0) { 1992 igvn.remove_dead_node(n); 1993 } 1994 --i; 1995 } 1996 } 1997 // Parsing may have added top inputs to the root node (Path 1998 // leading to the Halt node proven dead). Make sure we get a 1999 // chance to clean them up. 2000 igvn._worklist.push(r); 2001 igvn.optimize(); 2002 } 2003 } 2004 2005 //------------------------------Optimize--------------------------------------- 2006 // Given a graph, optimize it. 2007 void Compile::Optimize() { 2008 TracePhase tp("optimizer", &timers[_t_optimizer]); 2009 2010 #ifndef PRODUCT 2011 if (_directive->BreakAtCompileOption) { 2012 BREAKPOINT; 2013 } 2014 2015 #endif 2016 2017 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2018 #ifdef ASSERT 2019 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2020 #endif 2021 2022 ResourceMark rm; 2023 2024 print_inlining_reinit(); 2025 2026 NOT_PRODUCT( verify_graph_edges(); ) 2027 2028 print_method(PHASE_AFTER_PARSING); 2029 2030 { 2031 // Iterative Global Value Numbering, including ideal transforms 2032 // Initialize IterGVN with types and values from parse-time GVN 2033 PhaseIterGVN igvn(initial_gvn()); 2034 #ifdef ASSERT 2035 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2036 #endif 2037 { 2038 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2039 igvn.optimize(); 2040 } 2041 2042 if (failing()) return; 2043 2044 print_method(PHASE_ITER_GVN1, 2); 2045 2046 inline_incrementally(igvn); 2047 2048 print_method(PHASE_INCREMENTAL_INLINE, 2); 2049 2050 if (failing()) return; 2051 2052 if (eliminate_boxing()) { 2053 // Inline valueOf() methods now. 2054 inline_boxing_calls(igvn); 2055 2056 if (AlwaysIncrementalInline) { 2057 inline_incrementally(igvn); 2058 } 2059 2060 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2061 2062 if (failing()) return; 2063 } 2064 2065 // Now that all inlining is over, cut edge from root to loop 2066 // safepoints 2067 remove_root_to_sfpts_edges(igvn); 2068 2069 // Remove the speculative part of types and clean up the graph from 2070 // the extra CastPP nodes whose only purpose is to carry them. Do 2071 // that early so that optimizations are not disrupted by the extra 2072 // CastPP nodes. 2073 remove_speculative_types(igvn); 2074 2075 // No more new expensive nodes will be added to the list from here 2076 // so keep only the actual candidates for optimizations. 2077 cleanup_expensive_nodes(igvn); 2078 2079 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2080 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2081 initial_gvn()->replace_with(&igvn); 2082 for_igvn()->clear(); 2083 Unique_Node_List new_worklist(C->comp_arena()); 2084 { 2085 ResourceMark rm; 2086 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2087 } 2088 set_for_igvn(&new_worklist); 2089 igvn = PhaseIterGVN(initial_gvn()); 2090 igvn.optimize(); 2091 } 2092 2093 // Perform escape analysis 2094 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2095 if (has_loops()) { 2096 // Cleanup graph (remove dead nodes). 2097 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2098 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2099 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2100 if (failing()) return; 2101 } 2102 ConnectionGraph::do_analysis(this, &igvn); 2103 2104 if (failing()) return; 2105 2106 // Optimize out fields loads from scalar replaceable allocations. 2107 igvn.optimize(); 2108 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2109 2110 if (failing()) return; 2111 2112 if (congraph() != NULL && macro_count() > 0) { 2113 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2114 PhaseMacroExpand mexp(igvn); 2115 mexp.eliminate_macro_nodes(); 2116 igvn.set_delay_transform(false); 2117 2118 igvn.optimize(); 2119 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2120 2121 if (failing()) return; 2122 } 2123 } 2124 2125 // Loop transforms on the ideal graph. Range Check Elimination, 2126 // peeling, unrolling, etc. 2127 2128 // Set loop opts counter 2129 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2130 { 2131 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2132 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2133 _loop_opts_cnt--; 2134 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2135 if (failing()) return; 2136 } 2137 // Loop opts pass if partial peeling occurred in previous pass 2138 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2139 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2140 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2141 _loop_opts_cnt--; 2142 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2143 if (failing()) return; 2144 } 2145 // Loop opts pass for loop-unrolling before CCP 2146 if(major_progress() && (_loop_opts_cnt > 0)) { 2147 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2148 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2149 _loop_opts_cnt--; 2150 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2151 } 2152 if (!failing()) { 2153 // Verify that last round of loop opts produced a valid graph 2154 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2155 PhaseIdealLoop::verify(igvn); 2156 } 2157 } 2158 if (failing()) return; 2159 2160 // Conditional Constant Propagation; 2161 PhaseCCP ccp( &igvn ); 2162 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2163 { 2164 TracePhase tp("ccp", &timers[_t_ccp]); 2165 ccp.do_transform(); 2166 } 2167 print_method(PHASE_CPP1, 2); 2168 2169 assert( true, "Break here to ccp.dump_old2new_map()"); 2170 2171 // Iterative Global Value Numbering, including ideal transforms 2172 { 2173 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2174 igvn = ccp; 2175 igvn.optimize(); 2176 } 2177 print_method(PHASE_ITER_GVN2, 2); 2178 2179 if (failing()) return; 2180 2181 // Loop transforms on the ideal graph. Range Check Elimination, 2182 // peeling, unrolling, etc. 2183 if (!optimize_loops(igvn, LoopOptsDefault)) { 2184 return; 2185 } 2186 2187 if (failing()) return; 2188 2189 // Ensure that major progress is now clear 2190 C->clear_major_progress(); 2191 2192 { 2193 // Verify that all previous optimizations produced a valid graph 2194 // at least to this point, even if no loop optimizations were done. 2195 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2196 PhaseIdealLoop::verify(igvn); 2197 } 2198 2199 if (range_check_cast_count() > 0) { 2200 // No more loop optimizations. Remove all range check dependent CastIINodes. 2201 C->remove_range_check_casts(igvn); 2202 igvn.optimize(); 2203 } 2204 2205 #ifdef ASSERT 2206 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2207 #endif 2208 2209 { 2210 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2211 PhaseMacroExpand mex(igvn); 2212 if (mex.expand_macro_nodes()) { 2213 assert(failing(), "must bail out w/ explicit message"); 2214 return; 2215 } 2216 print_method(PHASE_MACRO_EXPANSION, 2); 2217 } 2218 2219 { 2220 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2221 if (bs->expand_barriers(this, igvn)) { 2222 assert(failing(), "must bail out w/ explicit message"); 2223 return; 2224 } 2225 print_method(PHASE_BARRIER_EXPANSION, 2); 2226 } 2227 2228 if (opaque4_count() > 0) { 2229 C->remove_opaque4_nodes(igvn); 2230 igvn.optimize(); 2231 } 2232 2233 if (C->max_vector_size() > 0) { 2234 C->optimize_logic_cones(igvn); 2235 igvn.optimize(); 2236 } 2237 2238 DEBUG_ONLY( _modified_nodes = NULL; ) 2239 } // (End scope of igvn; run destructor if necessary for asserts.) 2240 2241 process_print_inlining(); 2242 // A method with only infinite loops has no edges entering loops from root 2243 { 2244 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2245 if (final_graph_reshaping()) { 2246 assert(failing(), "must bail out w/ explicit message"); 2247 return; 2248 } 2249 } 2250 2251 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2252 } 2253 2254 //---------------------------- Bitwise operation packing optimization --------------------------- 2255 2256 static bool is_vector_unary_bitwise_op(Node* n) { 2257 return n->Opcode() == Op_XorV && 2258 VectorNode::is_vector_bitwise_not_pattern(n); 2259 } 2260 2261 static bool is_vector_binary_bitwise_op(Node* n) { 2262 switch (n->Opcode()) { 2263 case Op_AndV: 2264 case Op_OrV: 2265 return true; 2266 2267 case Op_XorV: 2268 return !is_vector_unary_bitwise_op(n); 2269 2270 default: 2271 return false; 2272 } 2273 } 2274 2275 static bool is_vector_ternary_bitwise_op(Node* n) { 2276 return n->Opcode() == Op_MacroLogicV; 2277 } 2278 2279 static bool is_vector_bitwise_op(Node* n) { 2280 return is_vector_unary_bitwise_op(n) || 2281 is_vector_binary_bitwise_op(n) || 2282 is_vector_ternary_bitwise_op(n); 2283 } 2284 2285 static bool is_vector_bitwise_cone_root(Node* n) { 2286 if (!is_vector_bitwise_op(n)) { 2287 return false; 2288 } 2289 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2290 if (is_vector_bitwise_op(n->fast_out(i))) { 2291 return false; 2292 } 2293 } 2294 return true; 2295 } 2296 2297 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2298 uint cnt = 0; 2299 if (is_vector_bitwise_op(n)) { 2300 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2301 for (uint i = 1; i < n->req(); i++) { 2302 Node* in = n->in(i); 2303 bool skip = VectorNode::is_all_ones_vector(in); 2304 if (!skip && !inputs.member(in)) { 2305 inputs.push(in); 2306 cnt++; 2307 } 2308 } 2309 assert(cnt <= 1, "not unary"); 2310 } else { 2311 uint last_req = n->req(); 2312 if (is_vector_ternary_bitwise_op(n)) { 2313 last_req = n->req() - 1; // skip last input 2314 } 2315 for (uint i = 1; i < last_req; i++) { 2316 Node* def = n->in(i); 2317 if (!inputs.member(def)) { 2318 inputs.push(def); 2319 cnt++; 2320 } 2321 } 2322 } 2323 partition.push(n); 2324 } else { // not a bitwise operations 2325 if (!inputs.member(n)) { 2326 inputs.push(n); 2327 cnt++; 2328 } 2329 } 2330 return cnt; 2331 } 2332 2333 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2334 Unique_Node_List useful_nodes; 2335 C->identify_useful_nodes(useful_nodes); 2336 2337 for (uint i = 0; i < useful_nodes.size(); i++) { 2338 Node* n = useful_nodes.at(i); 2339 if (is_vector_bitwise_cone_root(n)) { 2340 list.push(n); 2341 } 2342 } 2343 } 2344 2345 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2346 const TypeVect* vt, 2347 Unique_Node_List& partition, 2348 Unique_Node_List& inputs) { 2349 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2350 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2351 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2352 2353 Node* in1 = inputs.at(0); 2354 Node* in2 = inputs.at(1); 2355 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2356 2357 uint func = compute_truth_table(partition, inputs); 2358 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt)); 2359 } 2360 2361 static uint extract_bit(uint func, uint pos) { 2362 return (func & (1 << pos)) >> pos; 2363 } 2364 2365 // 2366 // A macro logic node represents a truth table. It has 4 inputs, 2367 // First three inputs corresponds to 3 columns of a truth table 2368 // and fourth input captures the logic function. 2369 // 2370 // eg. fn = (in1 AND in2) OR in3; 2371 // 2372 // MacroNode(in1,in2,in3,fn) 2373 // 2374 // ----------------- 2375 // in1 in2 in3 fn 2376 // ----------------- 2377 // 0 0 0 0 2378 // 0 0 1 1 2379 // 0 1 0 0 2380 // 0 1 1 1 2381 // 1 0 0 0 2382 // 1 0 1 1 2383 // 1 1 0 1 2384 // 1 1 1 1 2385 // 2386 2387 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2388 int res = 0; 2389 for (int i = 0; i < 8; i++) { 2390 int bit1 = extract_bit(in1, i); 2391 int bit2 = extract_bit(in2, i); 2392 int bit3 = extract_bit(in3, i); 2393 2394 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2395 int func_bit = extract_bit(func, func_bit_pos); 2396 2397 res |= func_bit << i; 2398 } 2399 return res; 2400 } 2401 2402 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2403 assert(n != NULL, ""); 2404 assert(eval_map.contains(n), "absent"); 2405 return *(eval_map.get(n)); 2406 } 2407 2408 static void eval_operands(Node* n, 2409 uint& func1, uint& func2, uint& func3, 2410 ResourceHashtable<Node*,uint>& eval_map) { 2411 assert(is_vector_bitwise_op(n), ""); 2412 func1 = eval_operand(n->in(1), eval_map); 2413 2414 if (is_vector_binary_bitwise_op(n)) { 2415 func2 = eval_operand(n->in(2), eval_map); 2416 } else if (is_vector_ternary_bitwise_op(n)) { 2417 func2 = eval_operand(n->in(2), eval_map); 2418 func3 = eval_operand(n->in(3), eval_map); 2419 } else { 2420 assert(is_vector_unary_bitwise_op(n), "not unary"); 2421 } 2422 } 2423 2424 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2425 assert(inputs.size() <= 3, "sanity"); 2426 ResourceMark rm; 2427 uint res = 0; 2428 ResourceHashtable<Node*,uint> eval_map; 2429 2430 // Populate precomputed functions for inputs. 2431 // Each input corresponds to one column of 3 input truth-table. 2432 uint input_funcs[] = { 0xAA, // (_, _, a) -> a 2433 0xCC, // (_, b, _) -> b 2434 0xF0 }; // (c, _, _) -> c 2435 for (uint i = 0; i < inputs.size(); i++) { 2436 eval_map.put(inputs.at(i), input_funcs[i]); 2437 } 2438 2439 for (uint i = 0; i < partition.size(); i++) { 2440 Node* n = partition.at(i); 2441 2442 uint func1 = 0, func2 = 0, func3 = 0; 2443 eval_operands(n, func1, func2, func3, eval_map); 2444 2445 switch (n->Opcode()) { 2446 case Op_OrV: 2447 assert(func3 == 0, "not binary"); 2448 res = func1 | func2; 2449 break; 2450 case Op_AndV: 2451 assert(func3 == 0, "not binary"); 2452 res = func1 & func2; 2453 break; 2454 case Op_XorV: 2455 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2456 assert(func2 == 0 && func3 == 0, "not unary"); 2457 res = (~func1) & 0xFF; 2458 } else { 2459 assert(func3 == 0, "not binary"); 2460 res = func1 ^ func2; 2461 } 2462 break; 2463 case Op_MacroLogicV: 2464 // Ordering of inputs may change during evaluation of sub-tree 2465 // containing MacroLogic node as a child node, thus a re-evaluation 2466 // makes sure that function is evaluated in context of current 2467 // inputs. 2468 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2469 break; 2470 2471 default: assert(false, "not supported: %s", n->Name()); 2472 } 2473 assert(res <= 0xFF, "invalid"); 2474 eval_map.put(n, res); 2475 } 2476 return res; 2477 } 2478 2479 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2480 assert(partition.size() == 0, "not empty"); 2481 assert(inputs.size() == 0, "not empty"); 2482 if (is_vector_ternary_bitwise_op(n)) { 2483 return false; 2484 } 2485 2486 bool is_unary_op = is_vector_unary_bitwise_op(n); 2487 if (is_unary_op) { 2488 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary"); 2489 return false; // too few inputs 2490 } 2491 2492 assert(is_vector_binary_bitwise_op(n), "not binary"); 2493 Node* in1 = n->in(1); 2494 Node* in2 = n->in(2); 2495 2496 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs); 2497 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs); 2498 partition.push(n); 2499 2500 // Too many inputs? 2501 if (inputs.size() > 3) { 2502 partition.clear(); 2503 inputs.clear(); 2504 { // Recompute in2 inputs 2505 Unique_Node_List not_used; 2506 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used); 2507 } 2508 // Pick the node with minimum number of inputs. 2509 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) { 2510 return false; // still too many inputs 2511 } 2512 // Recompute partition & inputs. 2513 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2); 2514 collect_unique_inputs(child, partition, inputs); 2515 2516 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1); 2517 inputs.push(other_input); 2518 2519 partition.push(n); 2520 } 2521 2522 return (partition.size() == 2 || partition.size() == 3) && 2523 (inputs.size() == 2 || inputs.size() == 3); 2524 } 2525 2526 2527 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2528 assert(is_vector_bitwise_op(n), "not a root"); 2529 2530 visited.set(n->_idx); 2531 2532 // 1) Do a DFS walk over the logic cone. 2533 for (uint i = 1; i < n->req(); i++) { 2534 Node* in = n->in(i); 2535 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2536 process_logic_cone_root(igvn, in, visited); 2537 } 2538 } 2539 2540 // 2) Bottom up traversal: Merge node[s] with 2541 // the parent to form macro logic node. 2542 Unique_Node_List partition; 2543 Unique_Node_List inputs; 2544 if (compute_logic_cone(n, partition, inputs)) { 2545 const TypeVect* vt = n->bottom_type()->is_vect(); 2546 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2547 igvn.replace_node(n, macro_logic); 2548 } 2549 } 2550 2551 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2552 ResourceMark rm; 2553 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2554 Unique_Node_List list; 2555 collect_logic_cone_roots(list); 2556 2557 while (list.size() > 0) { 2558 Node* n = list.pop(); 2559 const TypeVect* vt = n->bottom_type()->is_vect(); 2560 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2561 if (supported) { 2562 VectorSet visited(comp_arena()); 2563 process_logic_cone_root(igvn, n, visited); 2564 } 2565 } 2566 } 2567 } 2568 2569 //------------------------------Code_Gen--------------------------------------- 2570 // Given a graph, generate code for it 2571 void Compile::Code_Gen() { 2572 if (failing()) { 2573 return; 2574 } 2575 2576 // Perform instruction selection. You might think we could reclaim Matcher 2577 // memory PDQ, but actually the Matcher is used in generating spill code. 2578 // Internals of the Matcher (including some VectorSets) must remain live 2579 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2580 // set a bit in reclaimed memory. 2581 2582 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2583 // nodes. Mapping is only valid at the root of each matched subtree. 2584 NOT_PRODUCT( verify_graph_edges(); ) 2585 2586 Matcher matcher; 2587 _matcher = &matcher; 2588 { 2589 TracePhase tp("matcher", &timers[_t_matcher]); 2590 matcher.match(); 2591 if (failing()) { 2592 return; 2593 } 2594 } 2595 2596 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2597 // nodes. Mapping is only valid at the root of each matched subtree. 2598 NOT_PRODUCT( verify_graph_edges(); ) 2599 2600 // If you have too many nodes, or if matching has failed, bail out 2601 check_node_count(0, "out of nodes matching instructions"); 2602 if (failing()) { 2603 return; 2604 } 2605 2606 print_method(PHASE_MATCHING, 2); 2607 2608 // Build a proper-looking CFG 2609 PhaseCFG cfg(node_arena(), root(), matcher); 2610 _cfg = &cfg; 2611 { 2612 TracePhase tp("scheduler", &timers[_t_scheduler]); 2613 bool success = cfg.do_global_code_motion(); 2614 if (!success) { 2615 return; 2616 } 2617 2618 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2619 NOT_PRODUCT( verify_graph_edges(); ) 2620 debug_only( cfg.verify(); ) 2621 } 2622 2623 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2624 _regalloc = ®alloc; 2625 { 2626 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2627 // Perform register allocation. After Chaitin, use-def chains are 2628 // no longer accurate (at spill code) and so must be ignored. 2629 // Node->LRG->reg mappings are still accurate. 2630 _regalloc->Register_Allocate(); 2631 2632 // Bail out if the allocator builds too many nodes 2633 if (failing()) { 2634 return; 2635 } 2636 } 2637 2638 // Prior to register allocation we kept empty basic blocks in case the 2639 // the allocator needed a place to spill. After register allocation we 2640 // are not adding any new instructions. If any basic block is empty, we 2641 // can now safely remove it. 2642 { 2643 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2644 cfg.remove_empty_blocks(); 2645 if (do_freq_based_layout()) { 2646 PhaseBlockLayout layout(cfg); 2647 } else { 2648 cfg.set_loop_alignment(); 2649 } 2650 cfg.fixup_flow(); 2651 } 2652 2653 // Apply peephole optimizations 2654 if( OptoPeephole ) { 2655 TracePhase tp("peephole", &timers[_t_peephole]); 2656 PhasePeephole peep( _regalloc, cfg); 2657 peep.do_transform(); 2658 } 2659 2660 // Do late expand if CPU requires this. 2661 if (Matcher::require_postalloc_expand) { 2662 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2663 cfg.postalloc_expand(_regalloc); 2664 } 2665 2666 // Convert Nodes to instruction bits in a buffer 2667 { 2668 TracePhase tp("output", &timers[_t_output]); 2669 PhaseOutput output; 2670 output.Output(); 2671 if (failing()) return; 2672 output.install(); 2673 } 2674 2675 print_method(PHASE_FINAL_CODE); 2676 2677 // He's dead, Jim. 2678 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2679 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2680 } 2681 2682 //------------------------------Final_Reshape_Counts--------------------------- 2683 // This class defines counters to help identify when a method 2684 // may/must be executed using hardware with only 24-bit precision. 2685 struct Final_Reshape_Counts : public StackObj { 2686 int _call_count; // count non-inlined 'common' calls 2687 int _float_count; // count float ops requiring 24-bit precision 2688 int _double_count; // count double ops requiring more precision 2689 int _java_call_count; // count non-inlined 'java' calls 2690 int _inner_loop_count; // count loops which need alignment 2691 VectorSet _visited; // Visitation flags 2692 Node_List _tests; // Set of IfNodes & PCTableNodes 2693 2694 Final_Reshape_Counts() : 2695 _call_count(0), _float_count(0), _double_count(0), 2696 _java_call_count(0), _inner_loop_count(0), 2697 _visited( Thread::current()->resource_area() ) { } 2698 2699 void inc_call_count () { _call_count ++; } 2700 void inc_float_count () { _float_count ++; } 2701 void inc_double_count() { _double_count++; } 2702 void inc_java_call_count() { _java_call_count++; } 2703 void inc_inner_loop_count() { _inner_loop_count++; } 2704 2705 int get_call_count () const { return _call_count ; } 2706 int get_float_count () const { return _float_count ; } 2707 int get_double_count() const { return _double_count; } 2708 int get_java_call_count() const { return _java_call_count; } 2709 int get_inner_loop_count() const { return _inner_loop_count; } 2710 }; 2711 2712 #ifdef ASSERT 2713 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2714 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2715 // Make sure the offset goes inside the instance layout. 2716 return k->contains_field_offset(tp->offset()); 2717 // Note that OffsetBot and OffsetTop are very negative. 2718 } 2719 #endif 2720 2721 // Eliminate trivially redundant StoreCMs and accumulate their 2722 // precedence edges. 2723 void Compile::eliminate_redundant_card_marks(Node* n) { 2724 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2725 if (n->in(MemNode::Address)->outcnt() > 1) { 2726 // There are multiple users of the same address so it might be 2727 // possible to eliminate some of the StoreCMs 2728 Node* mem = n->in(MemNode::Memory); 2729 Node* adr = n->in(MemNode::Address); 2730 Node* val = n->in(MemNode::ValueIn); 2731 Node* prev = n; 2732 bool done = false; 2733 // Walk the chain of StoreCMs eliminating ones that match. As 2734 // long as it's a chain of single users then the optimization is 2735 // safe. Eliminating partially redundant StoreCMs would require 2736 // cloning copies down the other paths. 2737 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2738 if (adr == mem->in(MemNode::Address) && 2739 val == mem->in(MemNode::ValueIn)) { 2740 // redundant StoreCM 2741 if (mem->req() > MemNode::OopStore) { 2742 // Hasn't been processed by this code yet. 2743 n->add_prec(mem->in(MemNode::OopStore)); 2744 } else { 2745 // Already converted to precedence edge 2746 for (uint i = mem->req(); i < mem->len(); i++) { 2747 // Accumulate any precedence edges 2748 if (mem->in(i) != NULL) { 2749 n->add_prec(mem->in(i)); 2750 } 2751 } 2752 // Everything above this point has been processed. 2753 done = true; 2754 } 2755 // Eliminate the previous StoreCM 2756 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2757 assert(mem->outcnt() == 0, "should be dead"); 2758 mem->disconnect_inputs(NULL, this); 2759 } else { 2760 prev = mem; 2761 } 2762 mem = prev->in(MemNode::Memory); 2763 } 2764 } 2765 } 2766 2767 //------------------------------final_graph_reshaping_impl---------------------- 2768 // Implement items 1-5 from final_graph_reshaping below. 2769 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2770 2771 if ( n->outcnt() == 0 ) return; // dead node 2772 uint nop = n->Opcode(); 2773 2774 // Check for 2-input instruction with "last use" on right input. 2775 // Swap to left input. Implements item (2). 2776 if( n->req() == 3 && // two-input instruction 2777 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2778 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2779 n->in(2)->outcnt() == 1 &&// right use IS a last use 2780 !n->in(2)->is_Con() ) { // right use is not a constant 2781 // Check for commutative opcode 2782 switch( nop ) { 2783 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2784 case Op_MaxI: case Op_MinI: 2785 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2786 case Op_AndL: case Op_XorL: case Op_OrL: 2787 case Op_AndI: case Op_XorI: case Op_OrI: { 2788 // Move "last use" input to left by swapping inputs 2789 n->swap_edges(1, 2); 2790 break; 2791 } 2792 default: 2793 break; 2794 } 2795 } 2796 2797 #ifdef ASSERT 2798 if( n->is_Mem() ) { 2799 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2800 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2801 // oop will be recorded in oop map if load crosses safepoint 2802 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2803 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2804 "raw memory operations should have control edge"); 2805 } 2806 if (n->is_MemBar()) { 2807 MemBarNode* mb = n->as_MemBar(); 2808 if (mb->trailing_store() || mb->trailing_load_store()) { 2809 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2810 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 2811 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2812 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2813 } else if (mb->leading()) { 2814 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2815 } 2816 } 2817 #endif 2818 // Count FPU ops and common calls, implements item (3) 2819 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2820 if (!gc_handled) { 2821 final_graph_reshaping_main_switch(n, frc, nop); 2822 } 2823 2824 // Collect CFG split points 2825 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2826 frc._tests.push(n); 2827 } 2828 } 2829 2830 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2831 switch( nop ) { 2832 // Count all float operations that may use FPU 2833 case Op_AddF: 2834 case Op_SubF: 2835 case Op_MulF: 2836 case Op_DivF: 2837 case Op_NegF: 2838 case Op_ModF: 2839 case Op_ConvI2F: 2840 case Op_ConF: 2841 case Op_CmpF: 2842 case Op_CmpF3: 2843 // case Op_ConvL2F: // longs are split into 32-bit halves 2844 frc.inc_float_count(); 2845 break; 2846 2847 case Op_ConvF2D: 2848 case Op_ConvD2F: 2849 frc.inc_float_count(); 2850 frc.inc_double_count(); 2851 break; 2852 2853 // Count all double operations that may use FPU 2854 case Op_AddD: 2855 case Op_SubD: 2856 case Op_MulD: 2857 case Op_DivD: 2858 case Op_NegD: 2859 case Op_ModD: 2860 case Op_ConvI2D: 2861 case Op_ConvD2I: 2862 // case Op_ConvL2D: // handled by leaf call 2863 // case Op_ConvD2L: // handled by leaf call 2864 case Op_ConD: 2865 case Op_CmpD: 2866 case Op_CmpD3: 2867 frc.inc_double_count(); 2868 break; 2869 case Op_Opaque1: // Remove Opaque Nodes before matching 2870 case Op_Opaque2: // Remove Opaque Nodes before matching 2871 case Op_Opaque3: 2872 n->subsume_by(n->in(1), this); 2873 break; 2874 case Op_CallStaticJava: 2875 case Op_CallJava: 2876 case Op_CallDynamicJava: 2877 frc.inc_java_call_count(); // Count java call site; 2878 case Op_CallRuntime: 2879 case Op_CallLeaf: 2880 case Op_CallLeafNoFP: { 2881 assert (n->is_Call(), ""); 2882 CallNode *call = n->as_Call(); 2883 // Count call sites where the FP mode bit would have to be flipped. 2884 // Do not count uncommon runtime calls: 2885 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2886 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2887 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2888 frc.inc_call_count(); // Count the call site 2889 } else { // See if uncommon argument is shared 2890 Node *n = call->in(TypeFunc::Parms); 2891 int nop = n->Opcode(); 2892 // Clone shared simple arguments to uncommon calls, item (1). 2893 if (n->outcnt() > 1 && 2894 !n->is_Proj() && 2895 nop != Op_CreateEx && 2896 nop != Op_CheckCastPP && 2897 nop != Op_DecodeN && 2898 nop != Op_DecodeNKlass && 2899 !n->is_Mem() && 2900 !n->is_Phi()) { 2901 Node *x = n->clone(); 2902 call->set_req(TypeFunc::Parms, x); 2903 } 2904 } 2905 break; 2906 } 2907 2908 case Op_StoreD: 2909 case Op_LoadD: 2910 case Op_LoadD_unaligned: 2911 frc.inc_double_count(); 2912 goto handle_mem; 2913 case Op_StoreF: 2914 case Op_LoadF: 2915 frc.inc_float_count(); 2916 goto handle_mem; 2917 2918 case Op_StoreCM: 2919 { 2920 // Convert OopStore dependence into precedence edge 2921 Node* prec = n->in(MemNode::OopStore); 2922 n->del_req(MemNode::OopStore); 2923 n->add_prec(prec); 2924 eliminate_redundant_card_marks(n); 2925 } 2926 2927 // fall through 2928 2929 case Op_StoreB: 2930 case Op_StoreC: 2931 case Op_StorePConditional: 2932 case Op_StoreI: 2933 case Op_StoreL: 2934 case Op_StoreIConditional: 2935 case Op_StoreLConditional: 2936 case Op_CompareAndSwapB: 2937 case Op_CompareAndSwapS: 2938 case Op_CompareAndSwapI: 2939 case Op_CompareAndSwapL: 2940 case Op_CompareAndSwapP: 2941 case Op_CompareAndSwapN: 2942 case Op_WeakCompareAndSwapB: 2943 case Op_WeakCompareAndSwapS: 2944 case Op_WeakCompareAndSwapI: 2945 case Op_WeakCompareAndSwapL: 2946 case Op_WeakCompareAndSwapP: 2947 case Op_WeakCompareAndSwapN: 2948 case Op_CompareAndExchangeB: 2949 case Op_CompareAndExchangeS: 2950 case Op_CompareAndExchangeI: 2951 case Op_CompareAndExchangeL: 2952 case Op_CompareAndExchangeP: 2953 case Op_CompareAndExchangeN: 2954 case Op_GetAndAddS: 2955 case Op_GetAndAddB: 2956 case Op_GetAndAddI: 2957 case Op_GetAndAddL: 2958 case Op_GetAndSetS: 2959 case Op_GetAndSetB: 2960 case Op_GetAndSetI: 2961 case Op_GetAndSetL: 2962 case Op_GetAndSetP: 2963 case Op_GetAndSetN: 2964 case Op_StoreP: 2965 case Op_StoreN: 2966 case Op_StoreNKlass: 2967 case Op_LoadB: 2968 case Op_LoadUB: 2969 case Op_LoadUS: 2970 case Op_LoadI: 2971 case Op_LoadKlass: 2972 case Op_LoadNKlass: 2973 case Op_LoadL: 2974 case Op_LoadL_unaligned: 2975 case Op_LoadPLocked: 2976 case Op_LoadP: 2977 case Op_LoadN: 2978 case Op_LoadRange: 2979 case Op_LoadS: { 2980 handle_mem: 2981 #ifdef ASSERT 2982 if( VerifyOptoOopOffsets ) { 2983 MemNode* mem = n->as_Mem(); 2984 // Check to see if address types have grounded out somehow. 2985 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2986 assert( !tp || oop_offset_is_sane(tp), "" ); 2987 } 2988 #endif 2989 break; 2990 } 2991 2992 case Op_AddP: { // Assert sane base pointers 2993 Node *addp = n->in(AddPNode::Address); 2994 assert( !addp->is_AddP() || 2995 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2996 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2997 "Base pointers must match (addp %u)", addp->_idx ); 2998 #ifdef _LP64 2999 if ((UseCompressedOops || UseCompressedClassPointers) && 3000 addp->Opcode() == Op_ConP && 3001 addp == n->in(AddPNode::Base) && 3002 n->in(AddPNode::Offset)->is_Con()) { 3003 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3004 // on the platform and on the compressed oops mode. 3005 // Use addressing with narrow klass to load with offset on x86. 3006 // Some platforms can use the constant pool to load ConP. 3007 // Do this transformation here since IGVN will convert ConN back to ConP. 3008 const Type* t = addp->bottom_type(); 3009 bool is_oop = t->isa_oopptr() != NULL; 3010 bool is_klass = t->isa_klassptr() != NULL; 3011 3012 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3013 (is_klass && Matcher::const_klass_prefer_decode())) { 3014 Node* nn = NULL; 3015 3016 int op = is_oop ? Op_ConN : Op_ConNKlass; 3017 3018 // Look for existing ConN node of the same exact type. 3019 Node* r = root(); 3020 uint cnt = r->outcnt(); 3021 for (uint i = 0; i < cnt; i++) { 3022 Node* m = r->raw_out(i); 3023 if (m!= NULL && m->Opcode() == op && 3024 m->bottom_type()->make_ptr() == t) { 3025 nn = m; 3026 break; 3027 } 3028 } 3029 if (nn != NULL) { 3030 // Decode a narrow oop to match address 3031 // [R12 + narrow_oop_reg<<3 + offset] 3032 if (is_oop) { 3033 nn = new DecodeNNode(nn, t); 3034 } else { 3035 nn = new DecodeNKlassNode(nn, t); 3036 } 3037 // Check for succeeding AddP which uses the same Base. 3038 // Otherwise we will run into the assertion above when visiting that guy. 3039 for (uint i = 0; i < n->outcnt(); ++i) { 3040 Node *out_i = n->raw_out(i); 3041 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3042 out_i->set_req(AddPNode::Base, nn); 3043 #ifdef ASSERT 3044 for (uint j = 0; j < out_i->outcnt(); ++j) { 3045 Node *out_j = out_i->raw_out(j); 3046 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3047 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3048 } 3049 #endif 3050 } 3051 } 3052 n->set_req(AddPNode::Base, nn); 3053 n->set_req(AddPNode::Address, nn); 3054 if (addp->outcnt() == 0) { 3055 addp->disconnect_inputs(NULL, this); 3056 } 3057 } 3058 } 3059 } 3060 #endif 3061 // platform dependent reshaping of the address expression 3062 reshape_address(n->as_AddP()); 3063 break; 3064 } 3065 3066 case Op_CastPP: { 3067 // Remove CastPP nodes to gain more freedom during scheduling but 3068 // keep the dependency they encode as control or precedence edges 3069 // (if control is set already) on memory operations. Some CastPP 3070 // nodes don't have a control (don't carry a dependency): skip 3071 // those. 3072 if (n->in(0) != NULL) { 3073 ResourceMark rm; 3074 Unique_Node_List wq; 3075 wq.push(n); 3076 for (uint next = 0; next < wq.size(); ++next) { 3077 Node *m = wq.at(next); 3078 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3079 Node* use = m->fast_out(i); 3080 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3081 use->ensure_control_or_add_prec(n->in(0)); 3082 } else { 3083 switch(use->Opcode()) { 3084 case Op_AddP: 3085 case Op_DecodeN: 3086 case Op_DecodeNKlass: 3087 case Op_CheckCastPP: 3088 case Op_CastPP: 3089 wq.push(use); 3090 break; 3091 } 3092 } 3093 } 3094 } 3095 } 3096 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3097 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3098 Node* in1 = n->in(1); 3099 const Type* t = n->bottom_type(); 3100 Node* new_in1 = in1->clone(); 3101 new_in1->as_DecodeN()->set_type(t); 3102 3103 if (!Matcher::narrow_oop_use_complex_address()) { 3104 // 3105 // x86, ARM and friends can handle 2 adds in addressing mode 3106 // and Matcher can fold a DecodeN node into address by using 3107 // a narrow oop directly and do implicit NULL check in address: 3108 // 3109 // [R12 + narrow_oop_reg<<3 + offset] 3110 // NullCheck narrow_oop_reg 3111 // 3112 // On other platforms (Sparc) we have to keep new DecodeN node and 3113 // use it to do implicit NULL check in address: 3114 // 3115 // decode_not_null narrow_oop_reg, base_reg 3116 // [base_reg + offset] 3117 // NullCheck base_reg 3118 // 3119 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3120 // to keep the information to which NULL check the new DecodeN node 3121 // corresponds to use it as value in implicit_null_check(). 3122 // 3123 new_in1->set_req(0, n->in(0)); 3124 } 3125 3126 n->subsume_by(new_in1, this); 3127 if (in1->outcnt() == 0) { 3128 in1->disconnect_inputs(NULL, this); 3129 } 3130 } else { 3131 n->subsume_by(n->in(1), this); 3132 if (n->outcnt() == 0) { 3133 n->disconnect_inputs(NULL, this); 3134 } 3135 } 3136 break; 3137 } 3138 #ifdef _LP64 3139 case Op_CmpP: 3140 // Do this transformation here to preserve CmpPNode::sub() and 3141 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3142 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3143 Node* in1 = n->in(1); 3144 Node* in2 = n->in(2); 3145 if (!in1->is_DecodeNarrowPtr()) { 3146 in2 = in1; 3147 in1 = n->in(2); 3148 } 3149 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3150 3151 Node* new_in2 = NULL; 3152 if (in2->is_DecodeNarrowPtr()) { 3153 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3154 new_in2 = in2->in(1); 3155 } else if (in2->Opcode() == Op_ConP) { 3156 const Type* t = in2->bottom_type(); 3157 if (t == TypePtr::NULL_PTR) { 3158 assert(in1->is_DecodeN(), "compare klass to null?"); 3159 // Don't convert CmpP null check into CmpN if compressed 3160 // oops implicit null check is not generated. 3161 // This will allow to generate normal oop implicit null check. 3162 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3163 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3164 // 3165 // This transformation together with CastPP transformation above 3166 // will generated code for implicit NULL checks for compressed oops. 3167 // 3168 // The original code after Optimize() 3169 // 3170 // LoadN memory, narrow_oop_reg 3171 // decode narrow_oop_reg, base_reg 3172 // CmpP base_reg, NULL 3173 // CastPP base_reg // NotNull 3174 // Load [base_reg + offset], val_reg 3175 // 3176 // after these transformations will be 3177 // 3178 // LoadN memory, narrow_oop_reg 3179 // CmpN narrow_oop_reg, NULL 3180 // decode_not_null narrow_oop_reg, base_reg 3181 // Load [base_reg + offset], val_reg 3182 // 3183 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3184 // since narrow oops can be used in debug info now (see the code in 3185 // final_graph_reshaping_walk()). 3186 // 3187 // At the end the code will be matched to 3188 // on x86: 3189 // 3190 // Load_narrow_oop memory, narrow_oop_reg 3191 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3192 // NullCheck narrow_oop_reg 3193 // 3194 // and on sparc: 3195 // 3196 // Load_narrow_oop memory, narrow_oop_reg 3197 // decode_not_null narrow_oop_reg, base_reg 3198 // Load [base_reg + offset], val_reg 3199 // NullCheck base_reg 3200 // 3201 } else if (t->isa_oopptr()) { 3202 new_in2 = ConNode::make(t->make_narrowoop()); 3203 } else if (t->isa_klassptr()) { 3204 new_in2 = ConNode::make(t->make_narrowklass()); 3205 } 3206 } 3207 if (new_in2 != NULL) { 3208 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3209 n->subsume_by(cmpN, this); 3210 if (in1->outcnt() == 0) { 3211 in1->disconnect_inputs(NULL, this); 3212 } 3213 if (in2->outcnt() == 0) { 3214 in2->disconnect_inputs(NULL, this); 3215 } 3216 } 3217 } 3218 break; 3219 3220 case Op_DecodeN: 3221 case Op_DecodeNKlass: 3222 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3223 // DecodeN could be pinned when it can't be fold into 3224 // an address expression, see the code for Op_CastPP above. 3225 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3226 break; 3227 3228 case Op_EncodeP: 3229 case Op_EncodePKlass: { 3230 Node* in1 = n->in(1); 3231 if (in1->is_DecodeNarrowPtr()) { 3232 n->subsume_by(in1->in(1), this); 3233 } else if (in1->Opcode() == Op_ConP) { 3234 const Type* t = in1->bottom_type(); 3235 if (t == TypePtr::NULL_PTR) { 3236 assert(t->isa_oopptr(), "null klass?"); 3237 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3238 } else if (t->isa_oopptr()) { 3239 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3240 } else if (t->isa_klassptr()) { 3241 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3242 } 3243 } 3244 if (in1->outcnt() == 0) { 3245 in1->disconnect_inputs(NULL, this); 3246 } 3247 break; 3248 } 3249 3250 case Op_Proj: { 3251 if (OptimizeStringConcat) { 3252 ProjNode* p = n->as_Proj(); 3253 if (p->_is_io_use) { 3254 // Separate projections were used for the exception path which 3255 // are normally removed by a late inline. If it wasn't inlined 3256 // then they will hang around and should just be replaced with 3257 // the original one. 3258 Node* proj = NULL; 3259 // Replace with just one 3260 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3261 Node *use = i.get(); 3262 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3263 proj = use; 3264 break; 3265 } 3266 } 3267 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3268 if (proj != NULL) { 3269 p->subsume_by(proj, this); 3270 } 3271 } 3272 } 3273 break; 3274 } 3275 3276 case Op_Phi: 3277 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3278 // The EncodeP optimization may create Phi with the same edges 3279 // for all paths. It is not handled well by Register Allocator. 3280 Node* unique_in = n->in(1); 3281 assert(unique_in != NULL, ""); 3282 uint cnt = n->req(); 3283 for (uint i = 2; i < cnt; i++) { 3284 Node* m = n->in(i); 3285 assert(m != NULL, ""); 3286 if (unique_in != m) 3287 unique_in = NULL; 3288 } 3289 if (unique_in != NULL) { 3290 n->subsume_by(unique_in, this); 3291 } 3292 } 3293 break; 3294 3295 #endif 3296 3297 #ifdef ASSERT 3298 case Op_CastII: 3299 // Verify that all range check dependent CastII nodes were removed. 3300 if (n->isa_CastII()->has_range_check()) { 3301 n->dump(3); 3302 assert(false, "Range check dependent CastII node was not removed"); 3303 } 3304 break; 3305 #endif 3306 3307 case Op_ModI: 3308 if (UseDivMod) { 3309 // Check if a%b and a/b both exist 3310 Node* d = n->find_similar(Op_DivI); 3311 if (d) { 3312 // Replace them with a fused divmod if supported 3313 if (Matcher::has_match_rule(Op_DivModI)) { 3314 DivModINode* divmod = DivModINode::make(n); 3315 d->subsume_by(divmod->div_proj(), this); 3316 n->subsume_by(divmod->mod_proj(), this); 3317 } else { 3318 // replace a%b with a-((a/b)*b) 3319 Node* mult = new MulINode(d, d->in(2)); 3320 Node* sub = new SubINode(d->in(1), mult); 3321 n->subsume_by(sub, this); 3322 } 3323 } 3324 } 3325 break; 3326 3327 case Op_ModL: 3328 if (UseDivMod) { 3329 // Check if a%b and a/b both exist 3330 Node* d = n->find_similar(Op_DivL); 3331 if (d) { 3332 // Replace them with a fused divmod if supported 3333 if (Matcher::has_match_rule(Op_DivModL)) { 3334 DivModLNode* divmod = DivModLNode::make(n); 3335 d->subsume_by(divmod->div_proj(), this); 3336 n->subsume_by(divmod->mod_proj(), this); 3337 } else { 3338 // replace a%b with a-((a/b)*b) 3339 Node* mult = new MulLNode(d, d->in(2)); 3340 Node* sub = new SubLNode(d->in(1), mult); 3341 n->subsume_by(sub, this); 3342 } 3343 } 3344 } 3345 break; 3346 3347 case Op_LoadVector: 3348 case Op_StoreVector: 3349 break; 3350 3351 case Op_AddReductionVI: 3352 case Op_AddReductionVL: 3353 case Op_AddReductionVF: 3354 case Op_AddReductionVD: 3355 case Op_MulReductionVI: 3356 case Op_MulReductionVL: 3357 case Op_MulReductionVF: 3358 case Op_MulReductionVD: 3359 case Op_MinReductionV: 3360 case Op_MaxReductionV: 3361 case Op_AndReductionV: 3362 case Op_OrReductionV: 3363 case Op_XorReductionV: 3364 break; 3365 3366 case Op_PackB: 3367 case Op_PackS: 3368 case Op_PackI: 3369 case Op_PackF: 3370 case Op_PackL: 3371 case Op_PackD: 3372 if (n->req()-1 > 2) { 3373 // Replace many operand PackNodes with a binary tree for matching 3374 PackNode* p = (PackNode*) n; 3375 Node* btp = p->binary_tree_pack(1, n->req()); 3376 n->subsume_by(btp, this); 3377 } 3378 break; 3379 case Op_Loop: 3380 case Op_CountedLoop: 3381 case Op_OuterStripMinedLoop: 3382 if (n->as_Loop()->is_inner_loop()) { 3383 frc.inc_inner_loop_count(); 3384 } 3385 n->as_Loop()->verify_strip_mined(0); 3386 break; 3387 case Op_LShiftI: 3388 case Op_RShiftI: 3389 case Op_URShiftI: 3390 case Op_LShiftL: 3391 case Op_RShiftL: 3392 case Op_URShiftL: 3393 if (Matcher::need_masked_shift_count) { 3394 // The cpu's shift instructions don't restrict the count to the 3395 // lower 5/6 bits. We need to do the masking ourselves. 3396 Node* in2 = n->in(2); 3397 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3398 const TypeInt* t = in2->find_int_type(); 3399 if (t != NULL && t->is_con()) { 3400 juint shift = t->get_con(); 3401 if (shift > mask) { // Unsigned cmp 3402 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3403 } 3404 } else { 3405 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3406 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3407 n->set_req(2, shift); 3408 } 3409 } 3410 if (in2->outcnt() == 0) { // Remove dead node 3411 in2->disconnect_inputs(NULL, this); 3412 } 3413 } 3414 break; 3415 case Op_MemBarStoreStore: 3416 case Op_MemBarRelease: 3417 // Break the link with AllocateNode: it is no longer useful and 3418 // confuses register allocation. 3419 if (n->req() > MemBarNode::Precedent) { 3420 n->set_req(MemBarNode::Precedent, top()); 3421 } 3422 break; 3423 case Op_MemBarAcquire: { 3424 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3425 // At parse time, the trailing MemBarAcquire for a volatile load 3426 // is created with an edge to the load. After optimizations, 3427 // that input may be a chain of Phis. If those phis have no 3428 // other use, then the MemBarAcquire keeps them alive and 3429 // register allocation can be confused. 3430 ResourceMark rm; 3431 Unique_Node_List wq; 3432 wq.push(n->in(MemBarNode::Precedent)); 3433 n->set_req(MemBarNode::Precedent, top()); 3434 while (wq.size() > 0) { 3435 Node* m = wq.pop(); 3436 if (m->outcnt() == 0) { 3437 for (uint j = 0; j < m->req(); j++) { 3438 Node* in = m->in(j); 3439 if (in != NULL) { 3440 wq.push(in); 3441 } 3442 } 3443 m->disconnect_inputs(NULL, this); 3444 } 3445 } 3446 } 3447 break; 3448 } 3449 case Op_RangeCheck: { 3450 RangeCheckNode* rc = n->as_RangeCheck(); 3451 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3452 n->subsume_by(iff, this); 3453 frc._tests.push(iff); 3454 break; 3455 } 3456 case Op_ConvI2L: { 3457 if (!Matcher::convi2l_type_required) { 3458 // Code generation on some platforms doesn't need accurate 3459 // ConvI2L types. Widening the type can help remove redundant 3460 // address computations. 3461 n->as_Type()->set_type(TypeLong::INT); 3462 ResourceMark rm; 3463 Unique_Node_List wq; 3464 wq.push(n); 3465 for (uint next = 0; next < wq.size(); next++) { 3466 Node *m = wq.at(next); 3467 3468 for(;;) { 3469 // Loop over all nodes with identical inputs edges as m 3470 Node* k = m->find_similar(m->Opcode()); 3471 if (k == NULL) { 3472 break; 3473 } 3474 // Push their uses so we get a chance to remove node made 3475 // redundant 3476 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3477 Node* u = k->fast_out(i); 3478 if (u->Opcode() == Op_LShiftL || 3479 u->Opcode() == Op_AddL || 3480 u->Opcode() == Op_SubL || 3481 u->Opcode() == Op_AddP) { 3482 wq.push(u); 3483 } 3484 } 3485 // Replace all nodes with identical edges as m with m 3486 k->subsume_by(m, this); 3487 } 3488 } 3489 } 3490 break; 3491 } 3492 case Op_CmpUL: { 3493 if (!Matcher::has_match_rule(Op_CmpUL)) { 3494 // No support for unsigned long comparisons 3495 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3496 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3497 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3498 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3499 Node* andl = new AndLNode(orl, remove_sign_mask); 3500 Node* cmp = new CmpLNode(andl, n->in(2)); 3501 n->subsume_by(cmp, this); 3502 } 3503 break; 3504 } 3505 default: 3506 assert(!n->is_Call(), ""); 3507 assert(!n->is_Mem(), ""); 3508 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3509 break; 3510 } 3511 } 3512 3513 //------------------------------final_graph_reshaping_walk--------------------- 3514 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3515 // requires that the walk visits a node's inputs before visiting the node. 3516 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3517 ResourceArea *area = Thread::current()->resource_area(); 3518 Unique_Node_List sfpt(area); 3519 3520 frc._visited.set(root->_idx); // first, mark node as visited 3521 uint cnt = root->req(); 3522 Node *n = root; 3523 uint i = 0; 3524 while (true) { 3525 if (i < cnt) { 3526 // Place all non-visited non-null inputs onto stack 3527 Node* m = n->in(i); 3528 ++i; 3529 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3530 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3531 // compute worst case interpreter size in case of a deoptimization 3532 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3533 3534 sfpt.push(m); 3535 } 3536 cnt = m->req(); 3537 nstack.push(n, i); // put on stack parent and next input's index 3538 n = m; 3539 i = 0; 3540 } 3541 } else { 3542 // Now do post-visit work 3543 final_graph_reshaping_impl( n, frc ); 3544 if (nstack.is_empty()) 3545 break; // finished 3546 n = nstack.node(); // Get node from stack 3547 cnt = n->req(); 3548 i = nstack.index(); 3549 nstack.pop(); // Shift to the next node on stack 3550 } 3551 } 3552 3553 // Skip next transformation if compressed oops are not used. 3554 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3555 (!UseCompressedOops && !UseCompressedClassPointers)) 3556 return; 3557 3558 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3559 // It could be done for an uncommon traps or any safepoints/calls 3560 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3561 while (sfpt.size() > 0) { 3562 n = sfpt.pop(); 3563 JVMState *jvms = n->as_SafePoint()->jvms(); 3564 assert(jvms != NULL, "sanity"); 3565 int start = jvms->debug_start(); 3566 int end = n->req(); 3567 bool is_uncommon = (n->is_CallStaticJava() && 3568 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3569 for (int j = start; j < end; j++) { 3570 Node* in = n->in(j); 3571 if (in->is_DecodeNarrowPtr()) { 3572 bool safe_to_skip = true; 3573 if (!is_uncommon ) { 3574 // Is it safe to skip? 3575 for (uint i = 0; i < in->outcnt(); i++) { 3576 Node* u = in->raw_out(i); 3577 if (!u->is_SafePoint() || 3578 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3579 safe_to_skip = false; 3580 } 3581 } 3582 } 3583 if (safe_to_skip) { 3584 n->set_req(j, in->in(1)); 3585 } 3586 if (in->outcnt() == 0) { 3587 in->disconnect_inputs(NULL, this); 3588 } 3589 } 3590 } 3591 } 3592 } 3593 3594 //------------------------------final_graph_reshaping-------------------------- 3595 // Final Graph Reshaping. 3596 // 3597 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3598 // and not commoned up and forced early. Must come after regular 3599 // optimizations to avoid GVN undoing the cloning. Clone constant 3600 // inputs to Loop Phis; these will be split by the allocator anyways. 3601 // Remove Opaque nodes. 3602 // (2) Move last-uses by commutative operations to the left input to encourage 3603 // Intel update-in-place two-address operations and better register usage 3604 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3605 // calls canonicalizing them back. 3606 // (3) Count the number of double-precision FP ops, single-precision FP ops 3607 // and call sites. On Intel, we can get correct rounding either by 3608 // forcing singles to memory (requires extra stores and loads after each 3609 // FP bytecode) or we can set a rounding mode bit (requires setting and 3610 // clearing the mode bit around call sites). The mode bit is only used 3611 // if the relative frequency of single FP ops to calls is low enough. 3612 // This is a key transform for SPEC mpeg_audio. 3613 // (4) Detect infinite loops; blobs of code reachable from above but not 3614 // below. Several of the Code_Gen algorithms fail on such code shapes, 3615 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3616 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3617 // Detection is by looking for IfNodes where only 1 projection is 3618 // reachable from below or CatchNodes missing some targets. 3619 // (5) Assert for insane oop offsets in debug mode. 3620 3621 bool Compile::final_graph_reshaping() { 3622 // an infinite loop may have been eliminated by the optimizer, 3623 // in which case the graph will be empty. 3624 if (root()->req() == 1) { 3625 record_method_not_compilable("trivial infinite loop"); 3626 return true; 3627 } 3628 3629 // Expensive nodes have their control input set to prevent the GVN 3630 // from freely commoning them. There's no GVN beyond this point so 3631 // no need to keep the control input. We want the expensive nodes to 3632 // be freely moved to the least frequent code path by gcm. 3633 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3634 for (int i = 0; i < expensive_count(); i++) { 3635 _expensive_nodes->at(i)->set_req(0, NULL); 3636 } 3637 3638 Final_Reshape_Counts frc; 3639 3640 // Visit everybody reachable! 3641 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3642 Node_Stack nstack(live_nodes() >> 1); 3643 final_graph_reshaping_walk(nstack, root(), frc); 3644 3645 // Check for unreachable (from below) code (i.e., infinite loops). 3646 for( uint i = 0; i < frc._tests.size(); i++ ) { 3647 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3648 // Get number of CFG targets. 3649 // Note that PCTables include exception targets after calls. 3650 uint required_outcnt = n->required_outcnt(); 3651 if (n->outcnt() != required_outcnt) { 3652 // Check for a few special cases. Rethrow Nodes never take the 3653 // 'fall-thru' path, so expected kids is 1 less. 3654 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3655 if (n->in(0)->in(0)->is_Call()) { 3656 CallNode *call = n->in(0)->in(0)->as_Call(); 3657 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3658 required_outcnt--; // Rethrow always has 1 less kid 3659 } else if (call->req() > TypeFunc::Parms && 3660 call->is_CallDynamicJava()) { 3661 // Check for null receiver. In such case, the optimizer has 3662 // detected that the virtual call will always result in a null 3663 // pointer exception. The fall-through projection of this CatchNode 3664 // will not be populated. 3665 Node *arg0 = call->in(TypeFunc::Parms); 3666 if (arg0->is_Type() && 3667 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3668 required_outcnt--; 3669 } 3670 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3671 call->req() > TypeFunc::Parms+1 && 3672 call->is_CallStaticJava()) { 3673 // Check for negative array length. In such case, the optimizer has 3674 // detected that the allocation attempt will always result in an 3675 // exception. There is no fall-through projection of this CatchNode . 3676 Node *arg1 = call->in(TypeFunc::Parms+1); 3677 if (arg1->is_Type() && 3678 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3679 required_outcnt--; 3680 } 3681 } 3682 } 3683 } 3684 // Recheck with a better notion of 'required_outcnt' 3685 if (n->outcnt() != required_outcnt) { 3686 record_method_not_compilable("malformed control flow"); 3687 return true; // Not all targets reachable! 3688 } 3689 } 3690 // Check that I actually visited all kids. Unreached kids 3691 // must be infinite loops. 3692 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3693 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3694 record_method_not_compilable("infinite loop"); 3695 return true; // Found unvisited kid; must be unreach 3696 } 3697 3698 // Here so verification code in final_graph_reshaping_walk() 3699 // always see an OuterStripMinedLoopEnd 3700 if (n->is_OuterStripMinedLoopEnd()) { 3701 IfNode* init_iff = n->as_If(); 3702 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3703 n->subsume_by(iff, this); 3704 } 3705 } 3706 3707 #ifdef IA32 3708 // If original bytecodes contained a mixture of floats and doubles 3709 // check if the optimizer has made it homogenous, item (3). 3710 if (UseSSE == 0 && 3711 frc.get_float_count() > 32 && 3712 frc.get_double_count() == 0 && 3713 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3714 set_24_bit_selection_and_mode(false, true); 3715 } 3716 #endif // IA32 3717 3718 set_java_calls(frc.get_java_call_count()); 3719 set_inner_loops(frc.get_inner_loop_count()); 3720 3721 // No infinite loops, no reason to bail out. 3722 return false; 3723 } 3724 3725 //-----------------------------too_many_traps---------------------------------- 3726 // Report if there are too many traps at the current method and bci. 3727 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3728 bool Compile::too_many_traps(ciMethod* method, 3729 int bci, 3730 Deoptimization::DeoptReason reason) { 3731 ciMethodData* md = method->method_data(); 3732 if (md->is_empty()) { 3733 // Assume the trap has not occurred, or that it occurred only 3734 // because of a transient condition during start-up in the interpreter. 3735 return false; 3736 } 3737 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3738 if (md->has_trap_at(bci, m, reason) != 0) { 3739 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3740 // Also, if there are multiple reasons, or if there is no per-BCI record, 3741 // assume the worst. 3742 if (log()) 3743 log()->elem("observe trap='%s' count='%d'", 3744 Deoptimization::trap_reason_name(reason), 3745 md->trap_count(reason)); 3746 return true; 3747 } else { 3748 // Ignore method/bci and see if there have been too many globally. 3749 return too_many_traps(reason, md); 3750 } 3751 } 3752 3753 // Less-accurate variant which does not require a method and bci. 3754 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3755 ciMethodData* logmd) { 3756 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3757 // Too many traps globally. 3758 // Note that we use cumulative trap_count, not just md->trap_count. 3759 if (log()) { 3760 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3761 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3762 Deoptimization::trap_reason_name(reason), 3763 mcount, trap_count(reason)); 3764 } 3765 return true; 3766 } else { 3767 // The coast is clear. 3768 return false; 3769 } 3770 } 3771 3772 //--------------------------too_many_recompiles-------------------------------- 3773 // Report if there are too many recompiles at the current method and bci. 3774 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3775 // Is not eager to return true, since this will cause the compiler to use 3776 // Action_none for a trap point, to avoid too many recompilations. 3777 bool Compile::too_many_recompiles(ciMethod* method, 3778 int bci, 3779 Deoptimization::DeoptReason reason) { 3780 ciMethodData* md = method->method_data(); 3781 if (md->is_empty()) { 3782 // Assume the trap has not occurred, or that it occurred only 3783 // because of a transient condition during start-up in the interpreter. 3784 return false; 3785 } 3786 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3787 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3788 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3789 Deoptimization::DeoptReason per_bc_reason 3790 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3791 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3792 if ((per_bc_reason == Deoptimization::Reason_none 3793 || md->has_trap_at(bci, m, reason) != 0) 3794 // The trap frequency measure we care about is the recompile count: 3795 && md->trap_recompiled_at(bci, m) 3796 && md->overflow_recompile_count() >= bc_cutoff) { 3797 // Do not emit a trap here if it has already caused recompilations. 3798 // Also, if there are multiple reasons, or if there is no per-BCI record, 3799 // assume the worst. 3800 if (log()) 3801 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3802 Deoptimization::trap_reason_name(reason), 3803 md->trap_count(reason), 3804 md->overflow_recompile_count()); 3805 return true; 3806 } else if (trap_count(reason) != 0 3807 && decompile_count() >= m_cutoff) { 3808 // Too many recompiles globally, and we have seen this sort of trap. 3809 // Use cumulative decompile_count, not just md->decompile_count. 3810 if (log()) 3811 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3812 Deoptimization::trap_reason_name(reason), 3813 md->trap_count(reason), trap_count(reason), 3814 md->decompile_count(), decompile_count()); 3815 return true; 3816 } else { 3817 // The coast is clear. 3818 return false; 3819 } 3820 } 3821 3822 // Compute when not to trap. Used by matching trap based nodes and 3823 // NullCheck optimization. 3824 void Compile::set_allowed_deopt_reasons() { 3825 _allowed_reasons = 0; 3826 if (is_method_compilation()) { 3827 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3828 assert(rs < BitsPerInt, "recode bit map"); 3829 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3830 _allowed_reasons |= nth_bit(rs); 3831 } 3832 } 3833 } 3834 } 3835 3836 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 3837 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 3838 } 3839 3840 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 3841 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 3842 } 3843 3844 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 3845 if (holder->is_initialized()) { 3846 return false; 3847 } 3848 if (holder->is_being_initialized()) { 3849 if (accessing_method->holder() == holder) { 3850 // Access inside a class. The barrier can be elided when access happens in <clinit>, 3851 // <init>, or a static method. In all those cases, there was an initialization 3852 // barrier on the holder klass passed. 3853 if (accessing_method->is_static_initializer() || 3854 accessing_method->is_object_initializer() || 3855 accessing_method->is_static()) { 3856 return false; 3857 } 3858 } else if (accessing_method->holder()->is_subclass_of(holder)) { 3859 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 3860 // In case of <init> or a static method, the barrier is on the subclass is not enough: 3861 // child class can become fully initialized while its parent class is still being initialized. 3862 if (accessing_method->is_static_initializer()) { 3863 return false; 3864 } 3865 } 3866 ciMethod* root = method(); // the root method of compilation 3867 if (root != accessing_method) { 3868 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 3869 } 3870 } 3871 return true; 3872 } 3873 3874 #ifndef PRODUCT 3875 //------------------------------verify_graph_edges--------------------------- 3876 // Walk the Graph and verify that there is a one-to-one correspondence 3877 // between Use-Def edges and Def-Use edges in the graph. 3878 void Compile::verify_graph_edges(bool no_dead_code) { 3879 if (VerifyGraphEdges) { 3880 ResourceArea *area = Thread::current()->resource_area(); 3881 Unique_Node_List visited(area); 3882 // Call recursive graph walk to check edges 3883 _root->verify_edges(visited); 3884 if (no_dead_code) { 3885 // Now make sure that no visited node is used by an unvisited node. 3886 bool dead_nodes = false; 3887 Unique_Node_List checked(area); 3888 while (visited.size() > 0) { 3889 Node* n = visited.pop(); 3890 checked.push(n); 3891 for (uint i = 0; i < n->outcnt(); i++) { 3892 Node* use = n->raw_out(i); 3893 if (checked.member(use)) continue; // already checked 3894 if (visited.member(use)) continue; // already in the graph 3895 if (use->is_Con()) continue; // a dead ConNode is OK 3896 // At this point, we have found a dead node which is DU-reachable. 3897 if (!dead_nodes) { 3898 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3899 dead_nodes = true; 3900 } 3901 use->dump(2); 3902 tty->print_cr("---"); 3903 checked.push(use); // No repeats; pretend it is now checked. 3904 } 3905 } 3906 assert(!dead_nodes, "using nodes must be reachable from root"); 3907 } 3908 } 3909 } 3910 #endif 3911 3912 // The Compile object keeps track of failure reasons separately from the ciEnv. 3913 // This is required because there is not quite a 1-1 relation between the 3914 // ciEnv and its compilation task and the Compile object. Note that one 3915 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3916 // to backtrack and retry without subsuming loads. Other than this backtracking 3917 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3918 // by the logic in C2Compiler. 3919 void Compile::record_failure(const char* reason) { 3920 if (log() != NULL) { 3921 log()->elem("failure reason='%s' phase='compile'", reason); 3922 } 3923 if (_failure_reason == NULL) { 3924 // Record the first failure reason. 3925 _failure_reason = reason; 3926 } 3927 3928 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3929 C->print_method(PHASE_FAILURE); 3930 } 3931 _root = NULL; // flush the graph, too 3932 } 3933 3934 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3935 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3936 _phase_name(name), _dolog(CITimeVerbose) 3937 { 3938 if (_dolog) { 3939 C = Compile::current(); 3940 _log = C->log(); 3941 } else { 3942 C = NULL; 3943 _log = NULL; 3944 } 3945 if (_log != NULL) { 3946 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3947 _log->stamp(); 3948 _log->end_head(); 3949 } 3950 } 3951 3952 Compile::TracePhase::~TracePhase() { 3953 3954 C = Compile::current(); 3955 if (_dolog) { 3956 _log = C->log(); 3957 } else { 3958 _log = NULL; 3959 } 3960 3961 #ifdef ASSERT 3962 if (PrintIdealNodeCount) { 3963 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3964 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3965 } 3966 3967 if (VerifyIdealNodeCount) { 3968 Compile::current()->print_missing_nodes(); 3969 } 3970 #endif 3971 3972 if (_log != NULL) { 3973 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3974 } 3975 } 3976 3977 //----------------------------static_subtype_check----------------------------- 3978 // Shortcut important common cases when superklass is exact: 3979 // (0) superklass is java.lang.Object (can occur in reflective code) 3980 // (1) subklass is already limited to a subtype of superklass => always ok 3981 // (2) subklass does not overlap with superklass => always fail 3982 // (3) superklass has NO subtypes and we can check with a simple compare. 3983 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3984 if (StressReflectiveCode) { 3985 return SSC_full_test; // Let caller generate the general case. 3986 } 3987 3988 if (superk == env()->Object_klass()) { 3989 return SSC_always_true; // (0) this test cannot fail 3990 } 3991 3992 ciType* superelem = superk; 3993 if (superelem->is_array_klass()) 3994 superelem = superelem->as_array_klass()->base_element_type(); 3995 3996 if (!subk->is_interface()) { // cannot trust static interface types yet 3997 if (subk->is_subtype_of(superk)) { 3998 return SSC_always_true; // (1) false path dead; no dynamic test needed 3999 } 4000 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4001 !superk->is_subtype_of(subk)) { 4002 return SSC_always_false; 4003 } 4004 } 4005 4006 // If casting to an instance klass, it must have no subtypes 4007 if (superk->is_interface()) { 4008 // Cannot trust interfaces yet. 4009 // %%% S.B. superk->nof_implementors() == 1 4010 } else if (superelem->is_instance_klass()) { 4011 ciInstanceKlass* ik = superelem->as_instance_klass(); 4012 if (!ik->has_subklass() && !ik->is_interface()) { 4013 if (!ik->is_final()) { 4014 // Add a dependency if there is a chance of a later subclass. 4015 dependencies()->assert_leaf_type(ik); 4016 } 4017 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4018 } 4019 } else { 4020 // A primitive array type has no subtypes. 4021 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4022 } 4023 4024 return SSC_full_test; 4025 } 4026 4027 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4028 #ifdef _LP64 4029 // The scaled index operand to AddP must be a clean 64-bit value. 4030 // Java allows a 32-bit int to be incremented to a negative 4031 // value, which appears in a 64-bit register as a large 4032 // positive number. Using that large positive number as an 4033 // operand in pointer arithmetic has bad consequences. 4034 // On the other hand, 32-bit overflow is rare, and the possibility 4035 // can often be excluded, if we annotate the ConvI2L node with 4036 // a type assertion that its value is known to be a small positive 4037 // number. (The prior range check has ensured this.) 4038 // This assertion is used by ConvI2LNode::Ideal. 4039 int index_max = max_jint - 1; // array size is max_jint, index is one less 4040 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4041 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4042 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4043 #endif 4044 return idx; 4045 } 4046 4047 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4048 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4049 if (ctrl != NULL) { 4050 // Express control dependency by a CastII node with a narrow type. 4051 value = new CastIINode(value, itype, false, true /* range check dependency */); 4052 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4053 // node from floating above the range check during loop optimizations. Otherwise, the 4054 // ConvI2L node may be eliminated independently of the range check, causing the data path 4055 // to become TOP while the control path is still there (although it's unreachable). 4056 value->set_req(0, ctrl); 4057 // Save CastII node to remove it after loop optimizations. 4058 phase->C->add_range_check_cast(value); 4059 value = phase->transform(value); 4060 } 4061 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4062 return phase->transform(new ConvI2LNode(value, ltype)); 4063 } 4064 4065 void Compile::print_inlining_stream_free() { 4066 if (_print_inlining_stream != NULL) { 4067 _print_inlining_stream->~stringStream(); 4068 _print_inlining_stream = NULL; 4069 } 4070 } 4071 4072 // The message about the current inlining is accumulated in 4073 // _print_inlining_stream and transfered into the _print_inlining_list 4074 // once we know whether inlining succeeds or not. For regular 4075 // inlining, messages are appended to the buffer pointed by 4076 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4077 // a new buffer is added after _print_inlining_idx in the list. This 4078 // way we can update the inlining message for late inlining call site 4079 // when the inlining is attempted again. 4080 void Compile::print_inlining_init() { 4081 if (print_inlining() || print_intrinsics()) { 4082 // print_inlining_init is actually called several times. 4083 print_inlining_stream_free(); 4084 _print_inlining_stream = new stringStream(); 4085 // Watch out: The memory initialized by the constructor call PrintInliningBuffer() 4086 // will be copied into the only initial element. The default destructor of 4087 // PrintInliningBuffer will be called when leaving the scope here. If it 4088 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss 4089 // would be destructed, too! 4090 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4091 } 4092 } 4093 4094 void Compile::print_inlining_reinit() { 4095 if (print_inlining() || print_intrinsics()) { 4096 print_inlining_stream_free(); 4097 // Re allocate buffer when we change ResourceMark 4098 _print_inlining_stream = new stringStream(); 4099 } 4100 } 4101 4102 void Compile::print_inlining_reset() { 4103 _print_inlining_stream->reset(); 4104 } 4105 4106 void Compile::print_inlining_commit() { 4107 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4108 // Transfer the message from _print_inlining_stream to the current 4109 // _print_inlining_list buffer and clear _print_inlining_stream. 4110 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4111 print_inlining_reset(); 4112 } 4113 4114 void Compile::print_inlining_push() { 4115 // Add new buffer to the _print_inlining_list at current position 4116 _print_inlining_idx++; 4117 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4118 } 4119 4120 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4121 return _print_inlining_list->at(_print_inlining_idx); 4122 } 4123 4124 void Compile::print_inlining_update(CallGenerator* cg) { 4125 if (print_inlining() || print_intrinsics()) { 4126 if (!cg->is_late_inline()) { 4127 if (print_inlining_current().cg() != NULL) { 4128 print_inlining_push(); 4129 } 4130 print_inlining_commit(); 4131 } else { 4132 if (print_inlining_current().cg() != cg && 4133 (print_inlining_current().cg() != NULL || 4134 print_inlining_current().ss()->size() != 0)) { 4135 print_inlining_push(); 4136 } 4137 print_inlining_commit(); 4138 print_inlining_current().set_cg(cg); 4139 } 4140 } 4141 } 4142 4143 void Compile::print_inlining_move_to(CallGenerator* cg) { 4144 // We resume inlining at a late inlining call site. Locate the 4145 // corresponding inlining buffer so that we can update it. 4146 if (print_inlining()) { 4147 for (int i = 0; i < _print_inlining_list->length(); i++) { 4148 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4149 _print_inlining_idx = i; 4150 return; 4151 } 4152 } 4153 ShouldNotReachHere(); 4154 } 4155 } 4156 4157 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4158 if (print_inlining()) { 4159 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4160 assert(print_inlining_current().cg() == cg, "wrong entry"); 4161 // replace message with new message 4162 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4163 print_inlining_commit(); 4164 print_inlining_current().set_cg(cg); 4165 } 4166 } 4167 4168 void Compile::print_inlining_assert_ready() { 4169 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4170 } 4171 4172 void Compile::process_print_inlining() { 4173 bool do_print_inlining = print_inlining() || print_intrinsics(); 4174 if (do_print_inlining || log() != NULL) { 4175 // Print inlining message for candidates that we couldn't inline 4176 // for lack of space 4177 for (int i = 0; i < _late_inlines.length(); i++) { 4178 CallGenerator* cg = _late_inlines.at(i); 4179 if (!cg->is_mh_late_inline()) { 4180 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4181 if (do_print_inlining) { 4182 cg->print_inlining_late(msg); 4183 } 4184 log_late_inline_failure(cg, msg); 4185 } 4186 } 4187 } 4188 if (do_print_inlining) { 4189 ResourceMark rm; 4190 stringStream ss; 4191 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4192 for (int i = 0; i < _print_inlining_list->length(); i++) { 4193 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4194 _print_inlining_list->at(i).freeStream(); 4195 } 4196 // Reset _print_inlining_list, it only contains destructed objects. 4197 // It is on the arena, so it will be freed when the arena is reset. 4198 _print_inlining_list = NULL; 4199 // _print_inlining_stream won't be used anymore, either. 4200 print_inlining_stream_free(); 4201 size_t end = ss.size(); 4202 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4203 strncpy(_print_inlining_output, ss.base(), end+1); 4204 _print_inlining_output[end] = 0; 4205 } 4206 } 4207 4208 void Compile::dump_print_inlining() { 4209 if (_print_inlining_output != NULL) { 4210 tty->print_raw(_print_inlining_output); 4211 } 4212 } 4213 4214 void Compile::log_late_inline(CallGenerator* cg) { 4215 if (log() != NULL) { 4216 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4217 cg->unique_id()); 4218 JVMState* p = cg->call_node()->jvms(); 4219 while (p != NULL) { 4220 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4221 p = p->caller(); 4222 } 4223 log()->tail("late_inline"); 4224 } 4225 } 4226 4227 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4228 log_late_inline(cg); 4229 if (log() != NULL) { 4230 log()->inline_fail(msg); 4231 } 4232 } 4233 4234 void Compile::log_inline_id(CallGenerator* cg) { 4235 if (log() != NULL) { 4236 // The LogCompilation tool needs a unique way to identify late 4237 // inline call sites. This id must be unique for this call site in 4238 // this compilation. Try to have it unique across compilations as 4239 // well because it can be convenient when grepping through the log 4240 // file. 4241 // Distinguish OSR compilations from others in case CICountOSR is 4242 // on. 4243 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4244 cg->set_unique_id(id); 4245 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4246 } 4247 } 4248 4249 void Compile::log_inline_failure(const char* msg) { 4250 if (C->log() != NULL) { 4251 C->log()->inline_fail(msg); 4252 } 4253 } 4254 4255 4256 // Dump inlining replay data to the stream. 4257 // Don't change thread state and acquire any locks. 4258 void Compile::dump_inline_data(outputStream* out) { 4259 InlineTree* inl_tree = ilt(); 4260 if (inl_tree != NULL) { 4261 out->print(" inline %d", inl_tree->count()); 4262 inl_tree->dump_replay_data(out); 4263 } 4264 } 4265 4266 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4267 if (n1->Opcode() < n2->Opcode()) return -1; 4268 else if (n1->Opcode() > n2->Opcode()) return 1; 4269 4270 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4271 for (uint i = 1; i < n1->req(); i++) { 4272 if (n1->in(i) < n2->in(i)) return -1; 4273 else if (n1->in(i) > n2->in(i)) return 1; 4274 } 4275 4276 return 0; 4277 } 4278 4279 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4280 Node* n1 = *n1p; 4281 Node* n2 = *n2p; 4282 4283 return cmp_expensive_nodes(n1, n2); 4284 } 4285 4286 void Compile::sort_expensive_nodes() { 4287 if (!expensive_nodes_sorted()) { 4288 _expensive_nodes->sort(cmp_expensive_nodes); 4289 } 4290 } 4291 4292 bool Compile::expensive_nodes_sorted() const { 4293 for (int i = 1; i < _expensive_nodes->length(); i++) { 4294 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4295 return false; 4296 } 4297 } 4298 return true; 4299 } 4300 4301 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4302 if (_expensive_nodes->length() == 0) { 4303 return false; 4304 } 4305 4306 assert(OptimizeExpensiveOps, "optimization off?"); 4307 4308 // Take this opportunity to remove dead nodes from the list 4309 int j = 0; 4310 for (int i = 0; i < _expensive_nodes->length(); i++) { 4311 Node* n = _expensive_nodes->at(i); 4312 if (!n->is_unreachable(igvn)) { 4313 assert(n->is_expensive(), "should be expensive"); 4314 _expensive_nodes->at_put(j, n); 4315 j++; 4316 } 4317 } 4318 _expensive_nodes->trunc_to(j); 4319 4320 // Then sort the list so that similar nodes are next to each other 4321 // and check for at least two nodes of identical kind with same data 4322 // inputs. 4323 sort_expensive_nodes(); 4324 4325 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4326 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4327 return true; 4328 } 4329 } 4330 4331 return false; 4332 } 4333 4334 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4335 if (_expensive_nodes->length() == 0) { 4336 return; 4337 } 4338 4339 assert(OptimizeExpensiveOps, "optimization off?"); 4340 4341 // Sort to bring similar nodes next to each other and clear the 4342 // control input of nodes for which there's only a single copy. 4343 sort_expensive_nodes(); 4344 4345 int j = 0; 4346 int identical = 0; 4347 int i = 0; 4348 bool modified = false; 4349 for (; i < _expensive_nodes->length()-1; i++) { 4350 assert(j <= i, "can't write beyond current index"); 4351 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4352 identical++; 4353 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4354 continue; 4355 } 4356 if (identical > 0) { 4357 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4358 identical = 0; 4359 } else { 4360 Node* n = _expensive_nodes->at(i); 4361 igvn.replace_input_of(n, 0, NULL); 4362 igvn.hash_insert(n); 4363 modified = true; 4364 } 4365 } 4366 if (identical > 0) { 4367 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4368 } else if (_expensive_nodes->length() >= 1) { 4369 Node* n = _expensive_nodes->at(i); 4370 igvn.replace_input_of(n, 0, NULL); 4371 igvn.hash_insert(n); 4372 modified = true; 4373 } 4374 _expensive_nodes->trunc_to(j); 4375 if (modified) { 4376 igvn.optimize(); 4377 } 4378 } 4379 4380 void Compile::add_expensive_node(Node * n) { 4381 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4382 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4383 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4384 if (OptimizeExpensiveOps) { 4385 _expensive_nodes->append(n); 4386 } else { 4387 // Clear control input and let IGVN optimize expensive nodes if 4388 // OptimizeExpensiveOps is off. 4389 n->set_req(0, NULL); 4390 } 4391 } 4392 4393 /** 4394 * Remove the speculative part of types and clean up the graph 4395 */ 4396 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4397 if (UseTypeSpeculation) { 4398 Unique_Node_List worklist; 4399 worklist.push(root()); 4400 int modified = 0; 4401 // Go over all type nodes that carry a speculative type, drop the 4402 // speculative part of the type and enqueue the node for an igvn 4403 // which may optimize it out. 4404 for (uint next = 0; next < worklist.size(); ++next) { 4405 Node *n = worklist.at(next); 4406 if (n->is_Type()) { 4407 TypeNode* tn = n->as_Type(); 4408 const Type* t = tn->type(); 4409 const Type* t_no_spec = t->remove_speculative(); 4410 if (t_no_spec != t) { 4411 bool in_hash = igvn.hash_delete(n); 4412 assert(in_hash, "node should be in igvn hash table"); 4413 tn->set_type(t_no_spec); 4414 igvn.hash_insert(n); 4415 igvn._worklist.push(n); // give it a chance to go away 4416 modified++; 4417 } 4418 } 4419 uint max = n->len(); 4420 for( uint i = 0; i < max; ++i ) { 4421 Node *m = n->in(i); 4422 if (not_a_node(m)) continue; 4423 worklist.push(m); 4424 } 4425 } 4426 // Drop the speculative part of all types in the igvn's type table 4427 igvn.remove_speculative_types(); 4428 if (modified > 0) { 4429 igvn.optimize(); 4430 } 4431 #ifdef ASSERT 4432 // Verify that after the IGVN is over no speculative type has resurfaced 4433 worklist.clear(); 4434 worklist.push(root()); 4435 for (uint next = 0; next < worklist.size(); ++next) { 4436 Node *n = worklist.at(next); 4437 const Type* t = igvn.type_or_null(n); 4438 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4439 if (n->is_Type()) { 4440 t = n->as_Type()->type(); 4441 assert(t == t->remove_speculative(), "no more speculative types"); 4442 } 4443 uint max = n->len(); 4444 for( uint i = 0; i < max; ++i ) { 4445 Node *m = n->in(i); 4446 if (not_a_node(m)) continue; 4447 worklist.push(m); 4448 } 4449 } 4450 igvn.check_no_speculative_types(); 4451 #endif 4452 } 4453 } 4454 4455 // Auxiliary method to support randomized stressing/fuzzing. 4456 // 4457 // This method can be called the arbitrary number of times, with current count 4458 // as the argument. The logic allows selecting a single candidate from the 4459 // running list of candidates as follows: 4460 // int count = 0; 4461 // Cand* selected = null; 4462 // while(cand = cand->next()) { 4463 // if (randomized_select(++count)) { 4464 // selected = cand; 4465 // } 4466 // } 4467 // 4468 // Including count equalizes the chances any candidate is "selected". 4469 // This is useful when we don't have the complete list of candidates to choose 4470 // from uniformly. In this case, we need to adjust the randomicity of the 4471 // selection, or else we will end up biasing the selection towards the latter 4472 // candidates. 4473 // 4474 // Quick back-envelope calculation shows that for the list of n candidates 4475 // the equal probability for the candidate to persist as "best" can be 4476 // achieved by replacing it with "next" k-th candidate with the probability 4477 // of 1/k. It can be easily shown that by the end of the run, the 4478 // probability for any candidate is converged to 1/n, thus giving the 4479 // uniform distribution among all the candidates. 4480 // 4481 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4482 #define RANDOMIZED_DOMAIN_POW 29 4483 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4484 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4485 bool Compile::randomized_select(int count) { 4486 assert(count > 0, "only positive"); 4487 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4488 } 4489 4490 CloneMap& Compile::clone_map() { return _clone_map; } 4491 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4492 4493 void NodeCloneInfo::dump() const { 4494 tty->print(" {%d:%d} ", idx(), gen()); 4495 } 4496 4497 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4498 uint64_t val = value(old->_idx); 4499 NodeCloneInfo cio(val); 4500 assert(val != 0, "old node should be in the map"); 4501 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4502 insert(nnn->_idx, cin.get()); 4503 #ifndef PRODUCT 4504 if (is_debug()) { 4505 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4506 } 4507 #endif 4508 } 4509 4510 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4511 NodeCloneInfo cio(value(old->_idx)); 4512 if (cio.get() == 0) { 4513 cio.set(old->_idx, 0); 4514 insert(old->_idx, cio.get()); 4515 #ifndef PRODUCT 4516 if (is_debug()) { 4517 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4518 } 4519 #endif 4520 } 4521 clone(old, nnn, gen); 4522 } 4523 4524 int CloneMap::max_gen() const { 4525 int g = 0; 4526 DictI di(_dict); 4527 for(; di.test(); ++di) { 4528 int t = gen(di._key); 4529 if (g < t) { 4530 g = t; 4531 #ifndef PRODUCT 4532 if (is_debug()) { 4533 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4534 } 4535 #endif 4536 } 4537 } 4538 return g; 4539 } 4540 4541 void CloneMap::dump(node_idx_t key) const { 4542 uint64_t val = value(key); 4543 if (val != 0) { 4544 NodeCloneInfo ni(val); 4545 ni.dump(); 4546 } 4547 } 4548 4549 // Move Allocate nodes to the start of the list 4550 void Compile::sort_macro_nodes() { 4551 int count = macro_count(); 4552 int allocates = 0; 4553 for (int i = 0; i < count; i++) { 4554 Node* n = macro_node(i); 4555 if (n->is_Allocate()) { 4556 if (i != allocates) { 4557 Node* tmp = macro_node(allocates); 4558 _macro_nodes->at_put(allocates, n); 4559 _macro_nodes->at_put(i, tmp); 4560 } 4561 allocates++; 4562 } 4563 } 4564 } 4565 4566 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) { 4567 EventCompilerPhase event; 4568 if (event.should_commit()) { 4569 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 4570 } 4571 4572 #ifndef PRODUCT 4573 if (should_print(level)) { 4574 char output[1024]; 4575 if (idx != 0) { 4576 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx); 4577 } else { 4578 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt)); 4579 } 4580 _printer->print_method(output, level); 4581 } 4582 #endif 4583 C->_latest_stage_start_counter.stamp(); 4584 } 4585 4586 void Compile::end_method(int level) { 4587 EventCompilerPhase event; 4588 if (event.should_commit()) { 4589 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level); 4590 } 4591 4592 #ifndef PRODUCT 4593 if (_printer && _printer->should_print(level)) { 4594 _printer->end_method(); 4595 } 4596 #endif 4597 } 4598 4599 4600 #ifndef PRODUCT 4601 IdealGraphPrinter* Compile::_debug_file_printer = NULL; 4602 IdealGraphPrinter* Compile::_debug_network_printer = NULL; 4603 4604 // Called from debugger. Prints method to the default file with the default phase name. 4605 // This works regardless of any Ideal Graph Visualizer flags set or not. 4606 void igv_print() { 4607 Compile::current()->igv_print_method_to_file(); 4608 } 4609 4610 // Same as igv_print() above but with a specified phase name. 4611 void igv_print(const char* phase_name) { 4612 Compile::current()->igv_print_method_to_file(phase_name); 4613 } 4614 4615 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 4616 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 4617 // This works regardless of any Ideal Graph Visualizer flags set or not. 4618 void igv_print(bool network) { 4619 if (network) { 4620 Compile::current()->igv_print_method_to_network(); 4621 } else { 4622 Compile::current()->igv_print_method_to_file(); 4623 } 4624 } 4625 4626 // Same as igv_print(bool network) above but with a specified phase name. 4627 void igv_print(bool network, const char* phase_name) { 4628 if (network) { 4629 Compile::current()->igv_print_method_to_network(phase_name); 4630 } else { 4631 Compile::current()->igv_print_method_to_file(phase_name); 4632 } 4633 } 4634 4635 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 4636 void igv_print_default() { 4637 Compile::current()->print_method(PHASE_DEBUG, 0, 0); 4638 } 4639 4640 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 4641 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 4642 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 4643 void igv_append() { 4644 Compile::current()->igv_print_method_to_file("Debug", true); 4645 } 4646 4647 // Same as igv_append() above but with a specified phase name. 4648 void igv_append(const char* phase_name) { 4649 Compile::current()->igv_print_method_to_file(phase_name, true); 4650 } 4651 4652 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 4653 const char* file_name = "custom_debug.xml"; 4654 if (_debug_file_printer == NULL) { 4655 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 4656 } else { 4657 _debug_file_printer->update_compiled_method(C->method()); 4658 } 4659 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 4660 _debug_file_printer->print_method(phase_name, 0); 4661 } 4662 4663 void Compile::igv_print_method_to_network(const char* phase_name) { 4664 if (_debug_network_printer == NULL) { 4665 _debug_network_printer = new IdealGraphPrinter(C); 4666 } else { 4667 _debug_network_printer->update_compiled_method(C->method()); 4668 } 4669 tty->print_cr("Method printed over network stream to IGV"); 4670 _debug_network_printer->print_method(phase_name, 0); 4671 } 4672 #endif 4673