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