1 /* 2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "jfr/jfrEvents.hpp" 39 #include "memory/resourceArea.hpp" 40 #include "opto/addnode.hpp" 41 #include "opto/block.hpp" 42 #include "opto/c2compiler.hpp" 43 #include "opto/callGenerator.hpp" 44 #include "opto/callnode.hpp" 45 #include "opto/castnode.hpp" 46 #include "opto/cfgnode.hpp" 47 #include "opto/chaitin.hpp" 48 #include "opto/compile.hpp" 49 #include "opto/connode.hpp" 50 #include "opto/convertnode.hpp" 51 #include "opto/divnode.hpp" 52 #include "opto/escape.hpp" 53 #include "opto/idealGraphPrinter.hpp" 54 #include "opto/loopnode.hpp" 55 #include "opto/machnode.hpp" 56 #include "opto/macro.hpp" 57 #include "opto/matcher.hpp" 58 #include "opto/mathexactnode.hpp" 59 #include "opto/memnode.hpp" 60 #include "opto/mulnode.hpp" 61 #include "opto/narrowptrnode.hpp" 62 #include "opto/node.hpp" 63 #include "opto/opcodes.hpp" 64 #include "opto/output.hpp" 65 #include "opto/parse.hpp" 66 #include "opto/phaseX.hpp" 67 #include "opto/rootnode.hpp" 68 #include "opto/runtime.hpp" 69 #include "opto/stringopts.hpp" 70 #include "opto/type.hpp" 71 #include "opto/vector.hpp" 72 #include "opto/vectornode.hpp" 73 #include "runtime/arguments.hpp" 74 #include "runtime/sharedRuntime.hpp" 75 #include "runtime/signature.hpp" 76 #include "runtime/stubRoutines.hpp" 77 #include "runtime/timer.hpp" 78 #include "utilities/align.hpp" 79 #include "utilities/copy.hpp" 80 #include "utilities/macros.hpp" 81 #include "utilities/resourceHash.hpp" 82 83 84 // -------------------- Compile::mach_constant_base_node ----------------------- 85 // Constant table base node singleton. 86 MachConstantBaseNode* Compile::mach_constant_base_node() { 87 if (_mach_constant_base_node == NULL) { 88 _mach_constant_base_node = new MachConstantBaseNode(); 89 _mach_constant_base_node->add_req(C->root()); 90 } 91 return _mach_constant_base_node; 92 } 93 94 95 /// Support for intrinsics. 96 97 // Return the index at which m must be inserted (or already exists). 98 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 99 class IntrinsicDescPair { 100 private: 101 ciMethod* _m; 102 bool _is_virtual; 103 public: 104 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 105 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 106 ciMethod* m= elt->method(); 107 ciMethod* key_m = key->_m; 108 if (key_m < m) return -1; 109 else if (key_m > m) return 1; 110 else { 111 bool is_virtual = elt->is_virtual(); 112 bool key_virtual = key->_is_virtual; 113 if (key_virtual < is_virtual) return -1; 114 else if (key_virtual > is_virtual) return 1; 115 else return 0; 116 } 117 } 118 }; 119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 120 #ifdef ASSERT 121 for (int i = 1; i < _intrinsics->length(); i++) { 122 CallGenerator* cg1 = _intrinsics->at(i-1); 123 CallGenerator* cg2 = _intrinsics->at(i); 124 assert(cg1->method() != cg2->method() 125 ? cg1->method() < cg2->method() 126 : cg1->is_virtual() < cg2->is_virtual(), 127 "compiler intrinsics list must stay sorted"); 128 } 129 #endif 130 IntrinsicDescPair pair(m, is_virtual); 131 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 132 } 133 134 void Compile::register_intrinsic(CallGenerator* cg) { 135 if (_intrinsics == NULL) { 136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 137 } 138 int len = _intrinsics->length(); 139 bool found = false; 140 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 141 assert(!found, "registering twice"); 142 _intrinsics->insert_before(index, cg); 143 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 144 } 145 146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 147 assert(m->is_loaded(), "don't try this on unloaded methods"); 148 if (_intrinsics != NULL) { 149 bool found = false; 150 int index = intrinsic_insertion_index(m, is_virtual, found); 151 if (found) { 152 return _intrinsics->at(index); 153 } 154 } 155 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 156 if (m->intrinsic_id() != vmIntrinsics::_none && 157 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 158 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 159 if (cg != NULL) { 160 // Save it for next time: 161 register_intrinsic(cg); 162 return cg; 163 } else { 164 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 165 } 166 } 167 return NULL; 168 } 169 170 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 171 // in library_call.cpp. 172 173 174 #ifndef PRODUCT 175 // statistics gathering... 176 177 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 179 180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 182 int oflags = _intrinsic_hist_flags[id]; 183 assert(flags != 0, "what happened?"); 184 if (is_virtual) { 185 flags |= _intrinsic_virtual; 186 } 187 bool changed = (flags != oflags); 188 if ((flags & _intrinsic_worked) != 0) { 189 juint count = (_intrinsic_hist_count[id] += 1); 190 if (count == 1) { 191 changed = true; // first time 192 } 193 // increment the overall count also: 194 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 195 } 196 if (changed) { 197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 198 // Something changed about the intrinsic's virtuality. 199 if ((flags & _intrinsic_virtual) != 0) { 200 // This is the first use of this intrinsic as a virtual call. 201 if (oflags != 0) { 202 // We already saw it as a non-virtual, so note both cases. 203 flags |= _intrinsic_both; 204 } 205 } else if ((oflags & _intrinsic_both) == 0) { 206 // This is the first use of this intrinsic as a non-virtual 207 flags |= _intrinsic_both; 208 } 209 } 210 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 211 } 212 // update the overall flags also: 213 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 214 return changed; 215 } 216 217 static char* format_flags(int flags, char* buf) { 218 buf[0] = 0; 219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 224 if (buf[0] == 0) strcat(buf, ","); 225 assert(buf[0] == ',', "must be"); 226 return &buf[1]; 227 } 228 229 void Compile::print_intrinsic_statistics() { 230 char flagsbuf[100]; 231 ttyLocker ttyl; 232 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 233 tty->print_cr("Compiler intrinsic usage:"); 234 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 235 if (total == 0) total = 1; // avoid div0 in case of no successes 236 #define PRINT_STAT_LINE(name, c, f) \ 237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 238 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 239 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 240 int flags = _intrinsic_hist_flags[id]; 241 juint count = _intrinsic_hist_count[id]; 242 if ((flags | count) != 0) { 243 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 244 } 245 } 246 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 247 if (xtty != NULL) xtty->tail("statistics"); 248 } 249 250 void Compile::print_statistics() { 251 { ttyLocker ttyl; 252 if (xtty != NULL) xtty->head("statistics type='opto'"); 253 Parse::print_statistics(); 254 PhaseCCP::print_statistics(); 255 PhaseRegAlloc::print_statistics(); 256 PhaseOutput::print_statistics(); 257 PhasePeephole::print_statistics(); 258 PhaseIdealLoop::print_statistics(); 259 if (xtty != NULL) xtty->tail("statistics"); 260 } 261 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 262 // put this under its own <statistics> element. 263 print_intrinsic_statistics(); 264 } 265 } 266 #endif //PRODUCT 267 268 void Compile::gvn_replace_by(Node* n, Node* nn) { 269 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 270 Node* use = n->last_out(i); 271 bool is_in_table = initial_gvn()->hash_delete(use); 272 uint uses_found = 0; 273 for (uint j = 0; j < use->len(); j++) { 274 if (use->in(j) == n) { 275 if (j < use->req()) 276 use->set_req(j, nn); 277 else 278 use->set_prec(j, nn); 279 uses_found++; 280 } 281 } 282 if (is_in_table) { 283 // reinsert into table 284 initial_gvn()->hash_find_insert(use); 285 } 286 record_for_igvn(use); 287 i -= uses_found; // we deleted 1 or more copies of this edge 288 } 289 } 290 291 292 static inline bool not_a_node(const Node* n) { 293 if (n == NULL) return true; 294 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 295 if (*(address*)n == badAddress) return true; // kill by Node::destruct 296 return false; 297 } 298 299 // Identify all nodes that are reachable from below, useful. 300 // Use breadth-first pass that records state in a Unique_Node_List, 301 // recursive traversal is slower. 302 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 303 int estimated_worklist_size = live_nodes(); 304 useful.map( estimated_worklist_size, NULL ); // preallocate space 305 306 // Initialize worklist 307 if (root() != NULL) { useful.push(root()); } 308 // If 'top' is cached, declare it useful to preserve cached node 309 if( cached_top_node() ) { useful.push(cached_top_node()); } 310 311 // Push all useful nodes onto the list, breadthfirst 312 for( uint next = 0; next < useful.size(); ++next ) { 313 assert( next < unique(), "Unique useful nodes < total nodes"); 314 Node *n = useful.at(next); 315 uint max = n->len(); 316 for( uint i = 0; i < max; ++i ) { 317 Node *m = n->in(i); 318 if (not_a_node(m)) continue; 319 useful.push(m); 320 } 321 } 322 } 323 324 // Update dead_node_list with any missing dead nodes using useful 325 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 326 void Compile::update_dead_node_list(Unique_Node_List &useful) { 327 uint max_idx = unique(); 328 VectorSet& useful_node_set = useful.member_set(); 329 330 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 331 // If node with index node_idx is not in useful set, 332 // mark it as dead in dead node list. 333 if (!useful_node_set.test(node_idx)) { 334 record_dead_node(node_idx); 335 } 336 } 337 } 338 339 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 340 int shift = 0; 341 for (int i = 0; i < inlines->length(); i++) { 342 CallGenerator* cg = inlines->at(i); 343 CallNode* call = cg->call_node(); 344 if (shift > 0) { 345 inlines->at_put(i-shift, cg); 346 } 347 if (!useful.member(call)) { 348 shift++; 349 } 350 } 351 inlines->trunc_to(inlines->length()-shift); 352 } 353 354 // Disconnect all useless nodes by disconnecting those at the boundary. 355 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 356 uint next = 0; 357 while (next < useful.size()) { 358 Node *n = useful.at(next++); 359 if (n->is_SafePoint()) { 360 // We're done with a parsing phase. Replaced nodes are not valid 361 // beyond that point. 362 n->as_SafePoint()->delete_replaced_nodes(); 363 } 364 // Use raw traversal of out edges since this code removes out edges 365 int max = n->outcnt(); 366 for (int j = 0; j < max; ++j) { 367 Node* child = n->raw_out(j); 368 if (! useful.member(child)) { 369 assert(!child->is_top() || child != top(), 370 "If top is cached in Compile object it is in useful list"); 371 // Only need to remove this out-edge to the useless node 372 n->raw_del_out(j); 373 --j; 374 --max; 375 } 376 } 377 if (n->outcnt() == 1 && n->has_special_unique_user()) { 378 record_for_igvn(n->unique_out()); 379 } 380 } 381 // Remove useless macro and predicate opaq nodes 382 for (int i = C->macro_count()-1; i >= 0; i--) { 383 Node* n = C->macro_node(i); 384 if (!useful.member(n)) { 385 remove_macro_node(n); 386 } 387 } 388 // Remove useless CastII nodes with range check dependency 389 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 390 Node* cast = range_check_cast_node(i); 391 if (!useful.member(cast)) { 392 remove_range_check_cast(cast); 393 } 394 } 395 // Remove useless expensive nodes 396 for (int i = C->expensive_count()-1; i >= 0; i--) { 397 Node* n = C->expensive_node(i); 398 if (!useful.member(n)) { 399 remove_expensive_node(n); 400 } 401 } 402 // Remove useless Opaque4 nodes 403 for (int i = opaque4_count() - 1; i >= 0; i--) { 404 Node* opaq = opaque4_node(i); 405 if (!useful.member(opaq)) { 406 remove_opaque4_node(opaq); 407 } 408 } 409 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 410 bs->eliminate_useless_gc_barriers(useful, this); 411 // clean up the late inline lists 412 remove_useless_late_inlines(&_string_late_inlines, useful); 413 remove_useless_late_inlines(&_boxing_late_inlines, useful); 414 remove_useless_late_inlines(&_late_inlines, useful); 415 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 // ============================================================================ 420 //------------------------------CompileWrapper--------------------------------- 421 class CompileWrapper : public StackObj { 422 Compile *const _compile; 423 public: 424 CompileWrapper(Compile* compile); 425 426 ~CompileWrapper(); 427 }; 428 429 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 430 // the Compile* pointer is stored in the current ciEnv: 431 ciEnv* env = compile->env(); 432 assert(env == ciEnv::current(), "must already be a ciEnv active"); 433 assert(env->compiler_data() == NULL, "compile already active?"); 434 env->set_compiler_data(compile); 435 assert(compile == Compile::current(), "sanity"); 436 437 compile->set_type_dict(NULL); 438 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 439 compile->clone_map().set_clone_idx(0); 440 compile->set_type_last_size(0); 441 compile->set_last_tf(NULL, NULL); 442 compile->set_indexSet_arena(NULL); 443 compile->set_indexSet_free_block_list(NULL); 444 compile->init_type_arena(); 445 Type::Initialize(compile); 446 _compile->begin_method(); 447 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 448 } 449 CompileWrapper::~CompileWrapper() { 450 _compile->end_method(); 451 _compile->env()->set_compiler_data(NULL); 452 } 453 454 455 //----------------------------print_compile_messages--------------------------- 456 void Compile::print_compile_messages() { 457 #ifndef PRODUCT 458 // Check if recompiling 459 if (_subsume_loads == false && PrintOpto) { 460 // Recompiling without allowing machine instructions to subsume loads 461 tty->print_cr("*********************************************************"); 462 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 463 tty->print_cr("*********************************************************"); 464 } 465 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 466 // Recompiling without escape analysis 467 tty->print_cr("*********************************************************"); 468 tty->print_cr("** Bailout: Recompile without escape analysis **"); 469 tty->print_cr("*********************************************************"); 470 } 471 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 472 // Recompiling without boxing elimination 473 tty->print_cr("*********************************************************"); 474 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 475 tty->print_cr("*********************************************************"); 476 } 477 if (C->directive()->BreakAtCompileOption) { 478 // Open the debugger when compiling this method. 479 tty->print("### Breaking when compiling: "); 480 method()->print_short_name(); 481 tty->cr(); 482 BREAKPOINT; 483 } 484 485 if( PrintOpto ) { 486 if (is_osr_compilation()) { 487 tty->print("[OSR]%3d", _compile_id); 488 } else { 489 tty->print("%3d", _compile_id); 490 } 491 } 492 #endif 493 } 494 495 // ============================================================================ 496 //------------------------------Compile standard------------------------------- 497 debug_only( int Compile::_debug_idx = 100000; ) 498 499 // Compile a method. entry_bci is -1 for normal compilations and indicates 500 // the continuation bci for on stack replacement. 501 502 503 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 504 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, bool install_code, DirectiveSet* directive) 505 : Phase(Compiler), 506 _compile_id(ci_env->compile_id()), 507 _save_argument_registers(false), 508 _subsume_loads(subsume_loads), 509 _do_escape_analysis(do_escape_analysis), 510 _install_code(install_code), 511 _eliminate_boxing(eliminate_boxing), 512 _method(target), 513 _entry_bci(osr_bci), 514 _stub_function(NULL), 515 _stub_name(NULL), 516 _stub_entry_point(NULL), 517 _max_node_limit(MaxNodeLimit), 518 _inlining_progress(false), 519 _inlining_incrementally(false), 520 _do_cleanup(false), 521 _has_reserved_stack_access(target->has_reserved_stack_access()), 522 #ifndef PRODUCT 523 _trace_opto_output(directive->TraceOptoOutputOption), 524 _print_ideal(directive->PrintIdealOption), 525 #endif 526 _has_method_handle_invokes(false), 527 _clinit_barrier_on_entry(false), 528 _comp_arena(mtCompiler), 529 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 530 _env(ci_env), 531 _directive(directive), 532 _log(ci_env->log()), 533 _failure_reason(NULL), 534 _congraph(NULL), 535 NOT_PRODUCT(_printer(NULL) COMMA) 536 _dead_node_list(comp_arena()), 537 _dead_node_count(0), 538 _node_arena(mtCompiler), 539 _old_arena(mtCompiler), 540 _mach_constant_base_node(NULL), 541 _Compile_types(mtCompiler), 542 _initial_gvn(NULL), 543 _for_igvn(NULL), 544 _warm_calls(NULL), 545 _late_inlines(comp_arena(), 2, 0, NULL), 546 _string_late_inlines(comp_arena(), 2, 0, NULL), 547 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 548 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL), 549 _late_inlines_pos(0), 550 _number_of_mh_late_inlines(0), 551 _print_inlining_stream(NULL), 552 _print_inlining_list(NULL), 553 _print_inlining_idx(0), 554 _print_inlining_output(NULL), 555 _replay_inline_data(NULL), 556 _java_calls(0), 557 _inner_loops(0), 558 _interpreter_frame_size(0) 559 #ifndef PRODUCT 560 , _in_dump_cnt(0) 561 #endif 562 { 563 C = this; 564 CompileWrapper cw(this); 565 566 if (CITimeVerbose) { 567 tty->print(" "); 568 target->holder()->name()->print(); 569 tty->print("."); 570 target->print_short_name(); 571 tty->print(" "); 572 } 573 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 574 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 575 576 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 577 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 578 // We can always print a disassembly, either abstract (hex dump) or 579 // with the help of a suitable hsdis library. Thus, we should not 580 // couple print_assembly and print_opto_assembly controls. 581 // But: always print opto and regular assembly on compile command 'print'. 582 bool print_assembly = directive->PrintAssemblyOption; 583 set_print_assembly(print_opto_assembly || print_assembly); 584 #else 585 set_print_assembly(false); // must initialize. 586 #endif 587 588 #ifndef PRODUCT 589 set_parsed_irreducible_loop(false); 590 591 if (directive->ReplayInlineOption) { 592 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 593 } 594 #endif 595 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 596 set_print_intrinsics(directive->PrintIntrinsicsOption); 597 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 598 599 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 600 // Make sure the method being compiled gets its own MDO, 601 // so we can at least track the decompile_count(). 602 // Need MDO to record RTM code generation state. 603 method()->ensure_method_data(); 604 } 605 606 Init(::AliasLevel); 607 608 609 print_compile_messages(); 610 611 _ilt = InlineTree::build_inline_tree_root(); 612 613 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 614 assert(num_alias_types() >= AliasIdxRaw, ""); 615 616 #define MINIMUM_NODE_HASH 1023 617 // Node list that Iterative GVN will start with 618 Unique_Node_List for_igvn(comp_arena()); 619 set_for_igvn(&for_igvn); 620 621 // GVN that will be run immediately on new nodes 622 uint estimated_size = method()->code_size()*4+64; 623 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 624 PhaseGVN gvn(node_arena(), estimated_size); 625 set_initial_gvn(&gvn); 626 627 print_inlining_init(); 628 { // Scope for timing the parser 629 TracePhase tp("parse", &timers[_t_parser]); 630 631 // Put top into the hash table ASAP. 632 initial_gvn()->transform_no_reclaim(top()); 633 634 // Set up tf(), start(), and find a CallGenerator. 635 CallGenerator* cg = NULL; 636 if (is_osr_compilation()) { 637 const TypeTuple *domain = StartOSRNode::osr_domain(); 638 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 639 init_tf(TypeFunc::make(domain, range)); 640 StartNode* s = new StartOSRNode(root(), domain); 641 initial_gvn()->set_type_bottom(s); 642 init_start(s); 643 cg = CallGenerator::for_osr(method(), entry_bci()); 644 } else { 645 // Normal case. 646 init_tf(TypeFunc::make(method())); 647 StartNode* s = new StartNode(root(), tf()->domain()); 648 initial_gvn()->set_type_bottom(s); 649 init_start(s); 650 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 651 // With java.lang.ref.reference.get() we must go through the 652 // intrinsic - even when get() is the root 653 // method of the compile - so that, if necessary, the value in 654 // the referent field of the reference object gets recorded by 655 // the pre-barrier code. 656 cg = find_intrinsic(method(), false); 657 } 658 if (cg == NULL) { 659 float past_uses = method()->interpreter_invocation_count(); 660 float expected_uses = past_uses; 661 cg = CallGenerator::for_inline(method(), expected_uses); 662 } 663 } 664 if (failing()) return; 665 if (cg == NULL) { 666 record_method_not_compilable("cannot parse method"); 667 return; 668 } 669 JVMState* jvms = build_start_state(start(), tf()); 670 if ((jvms = cg->generate(jvms)) == NULL) { 671 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 672 record_method_not_compilable("method parse failed"); 673 } 674 return; 675 } 676 GraphKit kit(jvms); 677 678 if (!kit.stopped()) { 679 // Accept return values, and transfer control we know not where. 680 // This is done by a special, unique ReturnNode bound to root. 681 return_values(kit.jvms()); 682 } 683 684 if (kit.has_exceptions()) { 685 // Any exceptions that escape from this call must be rethrown 686 // to whatever caller is dynamically above us on the stack. 687 // This is done by a special, unique RethrowNode bound to root. 688 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 689 } 690 691 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 692 693 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 694 inline_string_calls(true); 695 } 696 697 if (failing()) return; 698 699 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 700 701 // Remove clutter produced by parsing. 702 if (!failing()) { 703 ResourceMark rm; 704 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 705 } 706 } 707 708 // Note: Large methods are capped off in do_one_bytecode(). 709 if (failing()) return; 710 711 // After parsing, node notes are no longer automagic. 712 // They must be propagated by register_new_node_with_optimizer(), 713 // clone(), or the like. 714 set_default_node_notes(NULL); 715 716 for (;;) { 717 int successes = Inline_Warm(); 718 if (failing()) return; 719 if (successes == 0) break; 720 } 721 722 // Drain the list. 723 Finish_Warm(); 724 #ifndef PRODUCT 725 if (should_print(1)) { 726 _printer->print_inlining(); 727 } 728 #endif 729 730 if (failing()) return; 731 NOT_PRODUCT( verify_graph_edges(); ) 732 733 // Now optimize 734 Optimize(); 735 if (failing()) return; 736 NOT_PRODUCT( verify_graph_edges(); ) 737 738 #ifndef PRODUCT 739 if (print_ideal()) { 740 ttyLocker ttyl; // keep the following output all in one block 741 // This output goes directly to the tty, not the compiler log. 742 // To enable tools to match it up with the compilation activity, 743 // be sure to tag this tty output with the compile ID. 744 if (xtty != NULL) { 745 xtty->head("ideal compile_id='%d'%s", compile_id(), 746 is_osr_compilation() ? " compile_kind='osr'" : 747 ""); 748 } 749 root()->dump(9999); 750 if (xtty != NULL) { 751 xtty->tail("ideal"); 752 } 753 } 754 #endif 755 756 #ifdef ASSERT 757 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 758 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 759 #endif 760 761 // Dump compilation data to replay it. 762 if (directive->DumpReplayOption) { 763 env()->dump_replay_data(_compile_id); 764 } 765 if (directive->DumpInlineOption && (ilt() != NULL)) { 766 env()->dump_inline_data(_compile_id); 767 } 768 769 // Now that we know the size of all the monitors we can add a fixed slot 770 // for the original deopt pc. 771 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 772 set_fixed_slots(next_slot); 773 774 // Compute when to use implicit null checks. Used by matching trap based 775 // nodes and NullCheck optimization. 776 set_allowed_deopt_reasons(); 777 778 // Now generate code 779 Code_Gen(); 780 } 781 782 //------------------------------Compile---------------------------------------- 783 // Compile a runtime stub 784 Compile::Compile( ciEnv* ci_env, 785 TypeFunc_generator generator, 786 address stub_function, 787 const char *stub_name, 788 int is_fancy_jump, 789 bool pass_tls, 790 bool save_arg_registers, 791 bool return_pc, 792 DirectiveSet* directive) 793 : Phase(Compiler), 794 _compile_id(0), 795 _save_argument_registers(save_arg_registers), 796 _subsume_loads(true), 797 _do_escape_analysis(false), 798 _install_code(true), 799 _eliminate_boxing(false), 800 _method(NULL), 801 _entry_bci(InvocationEntryBci), 802 _stub_function(stub_function), 803 _stub_name(stub_name), 804 _stub_entry_point(NULL), 805 _max_node_limit(MaxNodeLimit), 806 _inlining_progress(false), 807 _inlining_incrementally(false), 808 _has_reserved_stack_access(false), 809 #ifndef PRODUCT 810 _trace_opto_output(directive->TraceOptoOutputOption), 811 _print_ideal(directive->PrintIdealOption), 812 #endif 813 _has_method_handle_invokes(false), 814 _clinit_barrier_on_entry(false), 815 _comp_arena(mtCompiler), 816 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 817 _env(ci_env), 818 _directive(directive), 819 _log(ci_env->log()), 820 _failure_reason(NULL), 821 _congraph(NULL), 822 NOT_PRODUCT(_printer(NULL) COMMA) 823 _dead_node_list(comp_arena()), 824 _dead_node_count(0), 825 _node_arena(mtCompiler), 826 _old_arena(mtCompiler), 827 _mach_constant_base_node(NULL), 828 _Compile_types(mtCompiler), 829 _initial_gvn(NULL), 830 _for_igvn(NULL), 831 _warm_calls(NULL), 832 _number_of_mh_late_inlines(0), 833 _print_inlining_stream(NULL), 834 _print_inlining_list(NULL), 835 _print_inlining_idx(0), 836 _print_inlining_output(NULL), 837 _replay_inline_data(NULL), 838 _java_calls(0), 839 _inner_loops(0), 840 _interpreter_frame_size(0), 841 #ifndef PRODUCT 842 _in_dump_cnt(0), 843 #endif 844 _allowed_reasons(0) { 845 C = this; 846 847 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 848 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 849 850 #ifndef PRODUCT 851 set_print_assembly(PrintFrameConverterAssembly); 852 set_parsed_irreducible_loop(false); 853 #else 854 set_print_assembly(false); // Must initialize. 855 #endif 856 set_has_irreducible_loop(false); // no loops 857 858 CompileWrapper cw(this); 859 Init(/*AliasLevel=*/ 0); 860 init_tf((*generator)()); 861 862 { 863 // The following is a dummy for the sake of GraphKit::gen_stub 864 Unique_Node_List for_igvn(comp_arena()); 865 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 866 PhaseGVN gvn(Thread::current()->resource_area(),255); 867 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 868 gvn.transform_no_reclaim(top()); 869 870 GraphKit kit; 871 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 872 } 873 874 NOT_PRODUCT( verify_graph_edges(); ) 875 876 Code_Gen(); 877 } 878 879 //------------------------------Init------------------------------------------- 880 // Prepare for a single compilation 881 void Compile::Init(int aliaslevel) { 882 _unique = 0; 883 _regalloc = NULL; 884 885 _tf = NULL; // filled in later 886 _top = NULL; // cached later 887 _matcher = NULL; // filled in later 888 _cfg = NULL; // filled in later 889 890 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 891 892 _node_note_array = NULL; 893 _default_node_notes = NULL; 894 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 895 896 _immutable_memory = NULL; // filled in at first inquiry 897 898 // Globally visible Nodes 899 // First set TOP to NULL to give safe behavior during creation of RootNode 900 set_cached_top_node(NULL); 901 set_root(new RootNode()); 902 // Now that you have a Root to point to, create the real TOP 903 set_cached_top_node( new ConNode(Type::TOP) ); 904 set_recent_alloc(NULL, NULL); 905 906 // Create Debug Information Recorder to record scopes, oopmaps, etc. 907 env()->set_oop_recorder(new OopRecorder(env()->arena())); 908 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 909 env()->set_dependencies(new Dependencies(env())); 910 911 _fixed_slots = 0; 912 set_has_split_ifs(false); 913 set_has_loops(has_method() && method()->has_loops()); // first approximation 914 set_has_stringbuilder(false); 915 set_has_boxed_value(false); 916 _trap_can_recompile = false; // no traps emitted yet 917 _major_progress = true; // start out assuming good things will happen 918 set_has_unsafe_access(false); 919 set_max_vector_size(0); 920 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 921 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 922 set_decompile_count(0); 923 924 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 925 _loop_opts_cnt = LoopOptsCount; 926 set_do_inlining(Inline); 927 set_max_inline_size(MaxInlineSize); 928 set_freq_inline_size(FreqInlineSize); 929 set_do_scheduling(OptoScheduling); 930 set_do_count_invocations(false); 931 set_do_method_data_update(false); 932 933 set_do_vector_loop(false); 934 935 if (AllowVectorizeOnDemand) { 936 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 937 set_do_vector_loop(true); 938 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());}) 939 } else if (has_method() && method()->name() != 0 && 940 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 941 set_do_vector_loop(true); 942 } 943 } 944 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 945 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());}) 946 947 set_age_code(has_method() && method()->profile_aging()); 948 set_rtm_state(NoRTM); // No RTM lock eliding by default 949 _max_node_limit = _directive->MaxNodeLimitOption; 950 951 #if INCLUDE_RTM_OPT 952 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 953 int rtm_state = method()->method_data()->rtm_state(); 954 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 955 // Don't generate RTM lock eliding code. 956 set_rtm_state(NoRTM); 957 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 958 // Generate RTM lock eliding code without abort ratio calculation code. 959 set_rtm_state(UseRTM); 960 } else if (UseRTMDeopt) { 961 // Generate RTM lock eliding code and include abort ratio calculation 962 // code if UseRTMDeopt is on. 963 set_rtm_state(ProfileRTM); 964 } 965 } 966 #endif 967 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 968 set_clinit_barrier_on_entry(true); 969 } 970 if (debug_info()->recording_non_safepoints()) { 971 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 972 (comp_arena(), 8, 0, NULL)); 973 set_default_node_notes(Node_Notes::make(this)); 974 } 975 976 // // -- Initialize types before each compile -- 977 // // Update cached type information 978 // if( _method && _method->constants() ) 979 // Type::update_loaded_types(_method, _method->constants()); 980 981 // Init alias_type map. 982 if (!_do_escape_analysis && aliaslevel == 3) 983 aliaslevel = 2; // No unique types without escape analysis 984 _AliasLevel = aliaslevel; 985 const int grow_ats = 16; 986 _max_alias_types = grow_ats; 987 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 988 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 989 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 990 { 991 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 992 } 993 // Initialize the first few types. 994 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 995 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 996 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 997 _num_alias_types = AliasIdxRaw+1; 998 // Zero out the alias type cache. 999 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1000 // A NULL adr_type hits in the cache right away. Preload the right answer. 1001 probe_alias_cache(NULL)->_index = AliasIdxTop; 1002 1003 _intrinsics = NULL; 1004 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1005 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1006 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1007 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1008 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1009 register_library_intrinsics(); 1010 #ifdef ASSERT 1011 _type_verify_symmetry = true; 1012 _phase_optimize_finished = false; 1013 #endif 1014 } 1015 1016 //---------------------------init_start---------------------------------------- 1017 // Install the StartNode on this compile object. 1018 void Compile::init_start(StartNode* s) { 1019 if (failing()) 1020 return; // already failing 1021 assert(s == start(), ""); 1022 } 1023 1024 /** 1025 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1026 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1027 * the ideal graph. 1028 */ 1029 StartNode* Compile::start() const { 1030 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1031 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1032 Node* start = root()->fast_out(i); 1033 if (start->is_Start()) { 1034 return start->as_Start(); 1035 } 1036 } 1037 fatal("Did not find Start node!"); 1038 return NULL; 1039 } 1040 1041 //-------------------------------immutable_memory------------------------------------- 1042 // Access immutable memory 1043 Node* Compile::immutable_memory() { 1044 if (_immutable_memory != NULL) { 1045 return _immutable_memory; 1046 } 1047 StartNode* s = start(); 1048 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1049 Node *p = s->fast_out(i); 1050 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1051 _immutable_memory = p; 1052 return _immutable_memory; 1053 } 1054 } 1055 ShouldNotReachHere(); 1056 return NULL; 1057 } 1058 1059 //----------------------set_cached_top_node------------------------------------ 1060 // Install the cached top node, and make sure Node::is_top works correctly. 1061 void Compile::set_cached_top_node(Node* tn) { 1062 if (tn != NULL) verify_top(tn); 1063 Node* old_top = _top; 1064 _top = tn; 1065 // Calling Node::setup_is_top allows the nodes the chance to adjust 1066 // their _out arrays. 1067 if (_top != NULL) _top->setup_is_top(); 1068 if (old_top != NULL) old_top->setup_is_top(); 1069 assert(_top == NULL || top()->is_top(), ""); 1070 } 1071 1072 #ifdef ASSERT 1073 uint Compile::count_live_nodes_by_graph_walk() { 1074 Unique_Node_List useful(comp_arena()); 1075 // Get useful node list by walking the graph. 1076 identify_useful_nodes(useful); 1077 return useful.size(); 1078 } 1079 1080 void Compile::print_missing_nodes() { 1081 1082 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1083 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1084 return; 1085 } 1086 1087 // This is an expensive function. It is executed only when the user 1088 // specifies VerifyIdealNodeCount option or otherwise knows the 1089 // additional work that needs to be done to identify reachable nodes 1090 // by walking the flow graph and find the missing ones using 1091 // _dead_node_list. 1092 1093 Unique_Node_List useful(comp_arena()); 1094 // Get useful node list by walking the graph. 1095 identify_useful_nodes(useful); 1096 1097 uint l_nodes = C->live_nodes(); 1098 uint l_nodes_by_walk = useful.size(); 1099 1100 if (l_nodes != l_nodes_by_walk) { 1101 if (_log != NULL) { 1102 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1103 _log->stamp(); 1104 _log->end_head(); 1105 } 1106 VectorSet& useful_member_set = useful.member_set(); 1107 int last_idx = l_nodes_by_walk; 1108 for (int i = 0; i < last_idx; i++) { 1109 if (useful_member_set.test(i)) { 1110 if (_dead_node_list.test(i)) { 1111 if (_log != NULL) { 1112 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1113 } 1114 if (PrintIdealNodeCount) { 1115 // Print the log message to tty 1116 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1117 useful.at(i)->dump(); 1118 } 1119 } 1120 } 1121 else if (! _dead_node_list.test(i)) { 1122 if (_log != NULL) { 1123 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1124 } 1125 if (PrintIdealNodeCount) { 1126 // Print the log message to tty 1127 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1128 } 1129 } 1130 } 1131 if (_log != NULL) { 1132 _log->tail("mismatched_nodes"); 1133 } 1134 } 1135 } 1136 void Compile::record_modified_node(Node* n) { 1137 if (_modified_nodes != NULL && !_inlining_incrementally && 1138 n->outcnt() != 0 && !n->is_Con()) { 1139 _modified_nodes->push(n); 1140 } 1141 } 1142 1143 void Compile::remove_modified_node(Node* n) { 1144 if (_modified_nodes != NULL) { 1145 _modified_nodes->remove(n); 1146 } 1147 } 1148 #endif 1149 1150 #ifndef PRODUCT 1151 void Compile::verify_top(Node* tn) const { 1152 if (tn != NULL) { 1153 assert(tn->is_Con(), "top node must be a constant"); 1154 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1155 assert(tn->in(0) != NULL, "must have live top node"); 1156 } 1157 } 1158 #endif 1159 1160 1161 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1162 1163 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1164 guarantee(arr != NULL, ""); 1165 int num_blocks = arr->length(); 1166 if (grow_by < num_blocks) grow_by = num_blocks; 1167 int num_notes = grow_by * _node_notes_block_size; 1168 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1169 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1170 while (num_notes > 0) { 1171 arr->append(notes); 1172 notes += _node_notes_block_size; 1173 num_notes -= _node_notes_block_size; 1174 } 1175 assert(num_notes == 0, "exact multiple, please"); 1176 } 1177 1178 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1179 if (source == NULL || dest == NULL) return false; 1180 1181 if (dest->is_Con()) 1182 return false; // Do not push debug info onto constants. 1183 1184 #ifdef ASSERT 1185 // Leave a bread crumb trail pointing to the original node: 1186 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1187 dest->set_debug_orig(source); 1188 } 1189 #endif 1190 1191 if (node_note_array() == NULL) 1192 return false; // Not collecting any notes now. 1193 1194 // This is a copy onto a pre-existing node, which may already have notes. 1195 // If both nodes have notes, do not overwrite any pre-existing notes. 1196 Node_Notes* source_notes = node_notes_at(source->_idx); 1197 if (source_notes == NULL || source_notes->is_clear()) return false; 1198 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1199 if (dest_notes == NULL || dest_notes->is_clear()) { 1200 return set_node_notes_at(dest->_idx, source_notes); 1201 } 1202 1203 Node_Notes merged_notes = (*source_notes); 1204 // The order of operations here ensures that dest notes will win... 1205 merged_notes.update_from(dest_notes); 1206 return set_node_notes_at(dest->_idx, &merged_notes); 1207 } 1208 1209 1210 //--------------------------allow_range_check_smearing------------------------- 1211 // Gating condition for coalescing similar range checks. 1212 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1213 // single covering check that is at least as strong as any of them. 1214 // If the optimization succeeds, the simplified (strengthened) range check 1215 // will always succeed. If it fails, we will deopt, and then give up 1216 // on the optimization. 1217 bool Compile::allow_range_check_smearing() const { 1218 // If this method has already thrown a range-check, 1219 // assume it was because we already tried range smearing 1220 // and it failed. 1221 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1222 return !already_trapped; 1223 } 1224 1225 1226 //------------------------------flatten_alias_type----------------------------- 1227 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1228 int offset = tj->offset(); 1229 TypePtr::PTR ptr = tj->ptr(); 1230 1231 // Known instance (scalarizable allocation) alias only with itself. 1232 bool is_known_inst = tj->isa_oopptr() != NULL && 1233 tj->is_oopptr()->is_known_instance(); 1234 1235 // Process weird unsafe references. 1236 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1237 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1238 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1239 tj = TypeOopPtr::BOTTOM; 1240 ptr = tj->ptr(); 1241 offset = tj->offset(); 1242 } 1243 1244 // Array pointers need some flattening 1245 const TypeAryPtr *ta = tj->isa_aryptr(); 1246 if (ta && ta->is_stable()) { 1247 // Erase stability property for alias analysis. 1248 tj = ta = ta->cast_to_stable(false); 1249 } 1250 if( ta && is_known_inst ) { 1251 if ( offset != Type::OffsetBot && 1252 offset > arrayOopDesc::length_offset_in_bytes() ) { 1253 offset = Type::OffsetBot; // Flatten constant access into array body only 1254 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1255 } 1256 } else if( ta && _AliasLevel >= 2 ) { 1257 // For arrays indexed by constant indices, we flatten the alias 1258 // space to include all of the array body. Only the header, klass 1259 // and array length can be accessed un-aliased. 1260 if( offset != Type::OffsetBot ) { 1261 if( ta->const_oop() ) { // MethodData* or Method* 1262 offset = Type::OffsetBot; // Flatten constant access into array body 1263 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1264 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1265 // range is OK as-is. 1266 tj = ta = TypeAryPtr::RANGE; 1267 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1268 tj = TypeInstPtr::KLASS; // all klass loads look alike 1269 ta = TypeAryPtr::RANGE; // generic ignored junk 1270 ptr = TypePtr::BotPTR; 1271 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1272 tj = TypeInstPtr::MARK; 1273 ta = TypeAryPtr::RANGE; // generic ignored junk 1274 ptr = TypePtr::BotPTR; 1275 } else { // Random constant offset into array body 1276 offset = Type::OffsetBot; // Flatten constant access into array body 1277 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1278 } 1279 } 1280 // Arrays of fixed size alias with arrays of unknown size. 1281 if (ta->size() != TypeInt::POS) { 1282 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1283 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1284 } 1285 // Arrays of known objects become arrays of unknown objects. 1286 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1287 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1288 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1289 } 1290 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1291 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1292 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1293 } 1294 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1295 // cannot be distinguished by bytecode alone. 1296 if (ta->elem() == TypeInt::BOOL) { 1297 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1298 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1299 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1300 } 1301 // During the 2nd round of IterGVN, NotNull castings are removed. 1302 // Make sure the Bottom and NotNull variants alias the same. 1303 // Also, make sure exact and non-exact variants alias the same. 1304 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1305 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1306 } 1307 } 1308 1309 // Oop pointers need some flattening 1310 const TypeInstPtr *to = tj->isa_instptr(); 1311 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1312 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1313 if( ptr == TypePtr::Constant ) { 1314 if (to->klass() != ciEnv::current()->Class_klass() || 1315 offset < k->size_helper() * wordSize) { 1316 // No constant oop pointers (such as Strings); they alias with 1317 // unknown strings. 1318 assert(!is_known_inst, "not scalarizable allocation"); 1319 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1320 } 1321 } else if( is_known_inst ) { 1322 tj = to; // Keep NotNull and klass_is_exact for instance type 1323 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1324 // During the 2nd round of IterGVN, NotNull castings are removed. 1325 // Make sure the Bottom and NotNull variants alias the same. 1326 // Also, make sure exact and non-exact variants alias the same. 1327 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1328 } 1329 if (to->speculative() != NULL) { 1330 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1331 } 1332 // Canonicalize the holder of this field 1333 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1334 // First handle header references such as a LoadKlassNode, even if the 1335 // object's klass is unloaded at compile time (4965979). 1336 if (!is_known_inst) { // Do it only for non-instance types 1337 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1338 } 1339 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1340 // Static fields are in the space above the normal instance 1341 // fields in the java.lang.Class instance. 1342 if (to->klass() != ciEnv::current()->Class_klass()) { 1343 to = NULL; 1344 tj = TypeOopPtr::BOTTOM; 1345 offset = tj->offset(); 1346 } 1347 } else { 1348 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1349 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1350 if( is_known_inst ) { 1351 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1352 } else { 1353 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1354 } 1355 } 1356 } 1357 } 1358 1359 // Klass pointers to object array klasses need some flattening 1360 const TypeKlassPtr *tk = tj->isa_klassptr(); 1361 if( tk ) { 1362 // If we are referencing a field within a Klass, we need 1363 // to assume the worst case of an Object. Both exact and 1364 // inexact types must flatten to the same alias class so 1365 // use NotNull as the PTR. 1366 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1367 1368 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1369 TypeKlassPtr::OBJECT->klass(), 1370 offset); 1371 } 1372 1373 ciKlass* klass = tk->klass(); 1374 if( klass->is_obj_array_klass() ) { 1375 ciKlass* k = TypeAryPtr::OOPS->klass(); 1376 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1377 k = TypeInstPtr::BOTTOM->klass(); 1378 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1379 } 1380 1381 // Check for precise loads from the primary supertype array and force them 1382 // to the supertype cache alias index. Check for generic array loads from 1383 // the primary supertype array and also force them to the supertype cache 1384 // alias index. Since the same load can reach both, we need to merge 1385 // these 2 disparate memories into the same alias class. Since the 1386 // primary supertype array is read-only, there's no chance of confusion 1387 // where we bypass an array load and an array store. 1388 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1389 if (offset == Type::OffsetBot || 1390 (offset >= primary_supers_offset && 1391 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1392 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1393 offset = in_bytes(Klass::secondary_super_cache_offset()); 1394 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1395 } 1396 } 1397 1398 // Flatten all Raw pointers together. 1399 if (tj->base() == Type::RawPtr) 1400 tj = TypeRawPtr::BOTTOM; 1401 1402 if (tj->base() == Type::AnyPtr) 1403 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1404 1405 // Flatten all to bottom for now 1406 switch( _AliasLevel ) { 1407 case 0: 1408 tj = TypePtr::BOTTOM; 1409 break; 1410 case 1: // Flatten to: oop, static, field or array 1411 switch (tj->base()) { 1412 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1413 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1414 case Type::AryPtr: // do not distinguish arrays at all 1415 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1416 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1417 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1418 default: ShouldNotReachHere(); 1419 } 1420 break; 1421 case 2: // No collapsing at level 2; keep all splits 1422 case 3: // No collapsing at level 3; keep all splits 1423 break; 1424 default: 1425 Unimplemented(); 1426 } 1427 1428 offset = tj->offset(); 1429 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1430 1431 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1432 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1433 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1434 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1435 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1436 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1437 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1438 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1439 assert( tj->ptr() != TypePtr::TopPTR && 1440 tj->ptr() != TypePtr::AnyNull && 1441 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1442 // assert( tj->ptr() != TypePtr::Constant || 1443 // tj->base() == Type::RawPtr || 1444 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1445 1446 return tj; 1447 } 1448 1449 void Compile::AliasType::Init(int i, const TypePtr* at) { 1450 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1451 _index = i; 1452 _adr_type = at; 1453 _field = NULL; 1454 _element = NULL; 1455 _is_rewritable = true; // default 1456 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1457 if (atoop != NULL && atoop->is_known_instance()) { 1458 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1459 _general_index = Compile::current()->get_alias_index(gt); 1460 } else { 1461 _general_index = 0; 1462 } 1463 } 1464 1465 BasicType Compile::AliasType::basic_type() const { 1466 if (element() != NULL) { 1467 const Type* element = adr_type()->is_aryptr()->elem(); 1468 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1469 } if (field() != NULL) { 1470 return field()->layout_type(); 1471 } else { 1472 return T_ILLEGAL; // unknown 1473 } 1474 } 1475 1476 //---------------------------------print_on------------------------------------ 1477 #ifndef PRODUCT 1478 void Compile::AliasType::print_on(outputStream* st) { 1479 if (index() < 10) 1480 st->print("@ <%d> ", index()); 1481 else st->print("@ <%d>", index()); 1482 st->print(is_rewritable() ? " " : " RO"); 1483 int offset = adr_type()->offset(); 1484 if (offset == Type::OffsetBot) 1485 st->print(" +any"); 1486 else st->print(" +%-3d", offset); 1487 st->print(" in "); 1488 adr_type()->dump_on(st); 1489 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1490 if (field() != NULL && tjp) { 1491 if (tjp->klass() != field()->holder() || 1492 tjp->offset() != field()->offset_in_bytes()) { 1493 st->print(" != "); 1494 field()->print(); 1495 st->print(" ***"); 1496 } 1497 } 1498 } 1499 1500 void print_alias_types() { 1501 Compile* C = Compile::current(); 1502 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1503 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1504 C->alias_type(idx)->print_on(tty); 1505 tty->cr(); 1506 } 1507 } 1508 #endif 1509 1510 1511 //----------------------------probe_alias_cache-------------------------------- 1512 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1513 intptr_t key = (intptr_t) adr_type; 1514 key ^= key >> logAliasCacheSize; 1515 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1516 } 1517 1518 1519 //-----------------------------grow_alias_types-------------------------------- 1520 void Compile::grow_alias_types() { 1521 const int old_ats = _max_alias_types; // how many before? 1522 const int new_ats = old_ats; // how many more? 1523 const int grow_ats = old_ats+new_ats; // how many now? 1524 _max_alias_types = grow_ats; 1525 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1526 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1527 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1528 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1529 } 1530 1531 1532 //--------------------------------find_alias_type------------------------------ 1533 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1534 if (_AliasLevel == 0) 1535 return alias_type(AliasIdxBot); 1536 1537 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1538 if (ace->_adr_type == adr_type) { 1539 return alias_type(ace->_index); 1540 } 1541 1542 // Handle special cases. 1543 if (adr_type == NULL) return alias_type(AliasIdxTop); 1544 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1545 1546 // Do it the slow way. 1547 const TypePtr* flat = flatten_alias_type(adr_type); 1548 1549 #ifdef ASSERT 1550 { 1551 ResourceMark rm; 1552 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1553 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1554 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1555 Type::str(adr_type)); 1556 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1557 const TypeOopPtr* foop = flat->is_oopptr(); 1558 // Scalarizable allocations have exact klass always. 1559 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1560 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1561 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1562 Type::str(foop), Type::str(xoop)); 1563 } 1564 } 1565 #endif 1566 1567 int idx = AliasIdxTop; 1568 for (int i = 0; i < num_alias_types(); i++) { 1569 if (alias_type(i)->adr_type() == flat) { 1570 idx = i; 1571 break; 1572 } 1573 } 1574 1575 if (idx == AliasIdxTop) { 1576 if (no_create) return NULL; 1577 // Grow the array if necessary. 1578 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1579 // Add a new alias type. 1580 idx = _num_alias_types++; 1581 _alias_types[idx]->Init(idx, flat); 1582 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1583 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1584 if (flat->isa_instptr()) { 1585 if (flat->offset() == java_lang_Class::klass_offset() 1586 && flat->is_instptr()->klass() == env()->Class_klass()) 1587 alias_type(idx)->set_rewritable(false); 1588 } 1589 if (flat->isa_aryptr()) { 1590 #ifdef ASSERT 1591 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1592 // (T_BYTE has the weakest alignment and size restrictions...) 1593 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1594 #endif 1595 if (flat->offset() == TypePtr::OffsetBot) { 1596 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1597 } 1598 } 1599 if (flat->isa_klassptr()) { 1600 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1601 alias_type(idx)->set_rewritable(false); 1602 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1603 alias_type(idx)->set_rewritable(false); 1604 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1605 alias_type(idx)->set_rewritable(false); 1606 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1607 alias_type(idx)->set_rewritable(false); 1608 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1609 alias_type(idx)->set_rewritable(false); 1610 } 1611 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1612 // but the base pointer type is not distinctive enough to identify 1613 // references into JavaThread.) 1614 1615 // Check for final fields. 1616 const TypeInstPtr* tinst = flat->isa_instptr(); 1617 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1618 ciField* field; 1619 if (tinst->const_oop() != NULL && 1620 tinst->klass() == ciEnv::current()->Class_klass() && 1621 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1622 // static field 1623 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1624 field = k->get_field_by_offset(tinst->offset(), true); 1625 } else { 1626 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1627 field = k->get_field_by_offset(tinst->offset(), false); 1628 } 1629 assert(field == NULL || 1630 original_field == NULL || 1631 (field->holder() == original_field->holder() && 1632 field->offset() == original_field->offset() && 1633 field->is_static() == original_field->is_static()), "wrong field?"); 1634 // Set field() and is_rewritable() attributes. 1635 if (field != NULL) alias_type(idx)->set_field(field); 1636 } 1637 } 1638 1639 // Fill the cache for next time. 1640 ace->_adr_type = adr_type; 1641 ace->_index = idx; 1642 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1643 1644 // Might as well try to fill the cache for the flattened version, too. 1645 AliasCacheEntry* face = probe_alias_cache(flat); 1646 if (face->_adr_type == NULL) { 1647 face->_adr_type = flat; 1648 face->_index = idx; 1649 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1650 } 1651 1652 return alias_type(idx); 1653 } 1654 1655 1656 Compile::AliasType* Compile::alias_type(ciField* field) { 1657 const TypeOopPtr* t; 1658 if (field->is_static()) 1659 t = TypeInstPtr::make(field->holder()->java_mirror()); 1660 else 1661 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1662 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1663 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1664 return atp; 1665 } 1666 1667 1668 //------------------------------have_alias_type-------------------------------- 1669 bool Compile::have_alias_type(const TypePtr* adr_type) { 1670 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1671 if (ace->_adr_type == adr_type) { 1672 return true; 1673 } 1674 1675 // Handle special cases. 1676 if (adr_type == NULL) return true; 1677 if (adr_type == TypePtr::BOTTOM) return true; 1678 1679 return find_alias_type(adr_type, true, NULL) != NULL; 1680 } 1681 1682 //-----------------------------must_alias-------------------------------------- 1683 // True if all values of the given address type are in the given alias category. 1684 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1685 if (alias_idx == AliasIdxBot) return true; // the universal category 1686 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1687 if (alias_idx == AliasIdxTop) return false; // the empty category 1688 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1689 1690 // the only remaining possible overlap is identity 1691 int adr_idx = get_alias_index(adr_type); 1692 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1693 assert(adr_idx == alias_idx || 1694 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1695 && adr_type != TypeOopPtr::BOTTOM), 1696 "should not be testing for overlap with an unsafe pointer"); 1697 return adr_idx == alias_idx; 1698 } 1699 1700 //------------------------------can_alias-------------------------------------- 1701 // True if any values of the given address type are in the given alias category. 1702 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1703 if (alias_idx == AliasIdxTop) return false; // the empty category 1704 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1705 // Known instance doesn't alias with bottom memory 1706 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1707 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1708 1709 // the only remaining possible overlap is identity 1710 int adr_idx = get_alias_index(adr_type); 1711 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1712 return adr_idx == alias_idx; 1713 } 1714 1715 1716 1717 //---------------------------pop_warm_call------------------------------------- 1718 WarmCallInfo* Compile::pop_warm_call() { 1719 WarmCallInfo* wci = _warm_calls; 1720 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1721 return wci; 1722 } 1723 1724 //----------------------------Inline_Warm-------------------------------------- 1725 int Compile::Inline_Warm() { 1726 // If there is room, try to inline some more warm call sites. 1727 // %%% Do a graph index compaction pass when we think we're out of space? 1728 if (!InlineWarmCalls) return 0; 1729 1730 int calls_made_hot = 0; 1731 int room_to_grow = NodeCountInliningCutoff - unique(); 1732 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1733 int amount_grown = 0; 1734 WarmCallInfo* call; 1735 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1736 int est_size = (int)call->size(); 1737 if (est_size > (room_to_grow - amount_grown)) { 1738 // This one won't fit anyway. Get rid of it. 1739 call->make_cold(); 1740 continue; 1741 } 1742 call->make_hot(); 1743 calls_made_hot++; 1744 amount_grown += est_size; 1745 amount_to_grow -= est_size; 1746 } 1747 1748 if (calls_made_hot > 0) set_major_progress(); 1749 return calls_made_hot; 1750 } 1751 1752 1753 //----------------------------Finish_Warm-------------------------------------- 1754 void Compile::Finish_Warm() { 1755 if (!InlineWarmCalls) return; 1756 if (failing()) return; 1757 if (warm_calls() == NULL) return; 1758 1759 // Clean up loose ends, if we are out of space for inlining. 1760 WarmCallInfo* call; 1761 while ((call = pop_warm_call()) != NULL) { 1762 call->make_cold(); 1763 } 1764 } 1765 1766 //---------------------cleanup_loop_predicates----------------------- 1767 // Remove the opaque nodes that protect the predicates so that all unused 1768 // checks and uncommon_traps will be eliminated from the ideal graph 1769 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1770 if (predicate_count()==0) return; 1771 for (int i = predicate_count(); i > 0; i--) { 1772 Node * n = predicate_opaque1_node(i-1); 1773 assert(n->Opcode() == Op_Opaque1, "must be"); 1774 igvn.replace_node(n, n->in(1)); 1775 } 1776 assert(predicate_count()==0, "should be clean!"); 1777 } 1778 1779 void Compile::add_range_check_cast(Node* n) { 1780 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1781 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1782 _range_check_casts->append(n); 1783 } 1784 1785 // Remove all range check dependent CastIINodes. 1786 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1787 for (int i = range_check_cast_count(); i > 0; i--) { 1788 Node* cast = range_check_cast_node(i-1); 1789 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1790 igvn.replace_node(cast, cast->in(1)); 1791 } 1792 assert(range_check_cast_count() == 0, "should be empty"); 1793 } 1794 1795 void Compile::add_opaque4_node(Node* n) { 1796 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1797 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1798 _opaque4_nodes->append(n); 1799 } 1800 1801 // Remove all Opaque4 nodes. 1802 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1803 for (int i = opaque4_count(); i > 0; i--) { 1804 Node* opaq = opaque4_node(i-1); 1805 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 1806 // With Opaque4 nodes, the expectation is that the test of input 1 1807 // is always equal to the constant value of input 2. So we can 1808 // remove the Opaque4 and replace it by input 2. In debug builds, 1809 // leave the non constant test in instead to sanity check that it 1810 // never fails (if it does, that subgraph was constructed so, at 1811 // runtime, a Halt node is executed). 1812 #ifdef ASSERT 1813 igvn.replace_node(opaq, opaq->in(1)); 1814 #else 1815 igvn.replace_node(opaq, opaq->in(2)); 1816 #endif 1817 } 1818 assert(opaque4_count() == 0, "should be empty"); 1819 } 1820 1821 // StringOpts and late inlining of string methods 1822 void Compile::inline_string_calls(bool parse_time) { 1823 { 1824 // remove useless nodes to make the usage analysis simpler 1825 ResourceMark rm; 1826 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1827 } 1828 1829 { 1830 ResourceMark rm; 1831 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1832 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1833 print_method(PHASE_AFTER_STRINGOPTS, 3); 1834 } 1835 1836 // now inline anything that we skipped the first time around 1837 if (!parse_time) { 1838 _late_inlines_pos = _late_inlines.length(); 1839 } 1840 1841 while (_string_late_inlines.length() > 0) { 1842 CallGenerator* cg = _string_late_inlines.pop(); 1843 cg->do_late_inline(); 1844 if (failing()) return; 1845 } 1846 _string_late_inlines.trunc_to(0); 1847 } 1848 1849 // Late inlining of boxing methods 1850 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1851 if (_boxing_late_inlines.length() > 0) { 1852 assert(has_boxed_value(), "inconsistent"); 1853 1854 PhaseGVN* gvn = initial_gvn(); 1855 set_inlining_incrementally(true); 1856 1857 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1858 for_igvn()->clear(); 1859 gvn->replace_with(&igvn); 1860 1861 _late_inlines_pos = _late_inlines.length(); 1862 1863 while (_boxing_late_inlines.length() > 0) { 1864 CallGenerator* cg = _boxing_late_inlines.pop(); 1865 cg->do_late_inline(); 1866 if (failing()) return; 1867 } 1868 _boxing_late_inlines.trunc_to(0); 1869 1870 inline_incrementally_cleanup(igvn); 1871 1872 set_inlining_incrementally(false); 1873 } 1874 } 1875 1876 bool Compile::inline_incrementally_one() { 1877 assert(IncrementalInline, "incremental inlining should be on"); 1878 1879 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1880 set_inlining_progress(false); 1881 set_do_cleanup(false); 1882 int i = 0; 1883 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1884 CallGenerator* cg = _late_inlines.at(i); 1885 _late_inlines_pos = i+1; 1886 cg->do_late_inline(); 1887 if (failing()) return false; 1888 } 1889 int j = 0; 1890 for (; i < _late_inlines.length(); i++, j++) { 1891 _late_inlines.at_put(j, _late_inlines.at(i)); 1892 } 1893 _late_inlines.trunc_to(j); 1894 assert(inlining_progress() || _late_inlines.length() == 0, ""); 1895 1896 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 1897 1898 set_inlining_progress(false); 1899 set_do_cleanup(false); 1900 return (_late_inlines.length() > 0) && !needs_cleanup; 1901 } 1902 1903 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 1904 { 1905 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1906 ResourceMark rm; 1907 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1908 } 1909 { 1910 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1911 igvn = PhaseIterGVN(initial_gvn()); 1912 igvn.optimize(); 1913 } 1914 } 1915 1916 // Perform incremental inlining until bound on number of live nodes is reached 1917 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1918 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1919 1920 set_inlining_incrementally(true); 1921 uint low_live_nodes = 0; 1922 1923 while (_late_inlines.length() > 0) { 1924 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1925 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1926 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 1927 // PhaseIdealLoop is expensive so we only try it once we are 1928 // out of live nodes and we only try it again if the previous 1929 // helped got the number of nodes down significantly 1930 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 1931 if (failing()) return; 1932 low_live_nodes = live_nodes(); 1933 _major_progress = true; 1934 } 1935 1936 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1937 break; // finish 1938 } 1939 } 1940 1941 for_igvn()->clear(); 1942 initial_gvn()->replace_with(&igvn); 1943 1944 while (inline_incrementally_one()) { 1945 assert(!failing(), "inconsistent"); 1946 } 1947 1948 if (failing()) return; 1949 1950 inline_incrementally_cleanup(igvn); 1951 1952 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 1953 1954 if (failing()) return; 1955 } 1956 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1957 1958 if (_string_late_inlines.length() > 0) { 1959 assert(has_stringbuilder(), "inconsistent"); 1960 for_igvn()->clear(); 1961 initial_gvn()->replace_with(&igvn); 1962 1963 inline_string_calls(false); 1964 1965 if (failing()) return; 1966 1967 inline_incrementally_cleanup(igvn); 1968 } 1969 1970 set_inlining_incrementally(false); 1971 } 1972 1973 1974 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 1975 if(_loop_opts_cnt > 0) { 1976 debug_only( int cnt = 0; ); 1977 while(major_progress() && (_loop_opts_cnt > 0)) { 1978 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 1979 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1980 PhaseIdealLoop::optimize(igvn, mode); 1981 _loop_opts_cnt--; 1982 if (failing()) return false; 1983 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 1984 } 1985 } 1986 return true; 1987 } 1988 1989 // Remove edges from "root" to each SafePoint at a backward branch. 1990 // They were inserted during parsing (see add_safepoint()) to make 1991 // infinite loops without calls or exceptions visible to root, i.e., 1992 // useful. 1993 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 1994 Node *r = root(); 1995 if (r != NULL) { 1996 for (uint i = r->req(); i < r->len(); ++i) { 1997 Node *n = r->in(i); 1998 if (n != NULL && n->is_SafePoint()) { 1999 r->rm_prec(i); 2000 if (n->outcnt() == 0) { 2001 igvn.remove_dead_node(n); 2002 } 2003 --i; 2004 } 2005 } 2006 // Parsing may have added top inputs to the root node (Path 2007 // leading to the Halt node proven dead). Make sure we get a 2008 // chance to clean them up. 2009 igvn._worklist.push(r); 2010 igvn.optimize(); 2011 } 2012 } 2013 2014 //------------------------------Optimize--------------------------------------- 2015 // Given a graph, optimize it. 2016 void Compile::Optimize() { 2017 TracePhase tp("optimizer", &timers[_t_optimizer]); 2018 2019 #ifndef PRODUCT 2020 if (_directive->BreakAtCompileOption) { 2021 BREAKPOINT; 2022 } 2023 2024 #endif 2025 2026 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2027 #ifdef ASSERT 2028 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2029 #endif 2030 2031 ResourceMark rm; 2032 2033 print_inlining_reinit(); 2034 2035 NOT_PRODUCT( verify_graph_edges(); ) 2036 2037 print_method(PHASE_AFTER_PARSING); 2038 2039 { 2040 // Iterative Global Value Numbering, including ideal transforms 2041 // Initialize IterGVN with types and values from parse-time GVN 2042 PhaseIterGVN igvn(initial_gvn()); 2043 #ifdef ASSERT 2044 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2045 #endif 2046 { 2047 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2048 igvn.optimize(); 2049 } 2050 2051 if (failing()) return; 2052 2053 print_method(PHASE_ITER_GVN1, 2); 2054 2055 inline_incrementally(igvn); 2056 2057 print_method(PHASE_INCREMENTAL_INLINE, 2); 2058 2059 if (failing()) return; 2060 2061 if (eliminate_boxing()) { 2062 // Inline valueOf() methods now. 2063 inline_boxing_calls(igvn); 2064 2065 if (AlwaysIncrementalInline) { 2066 inline_incrementally(igvn); 2067 } 2068 2069 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2070 2071 if (failing()) return; 2072 } 2073 2074 // Now that all inlining is over, cut edge from root to loop 2075 // safepoints 2076 remove_root_to_sfpts_edges(igvn); 2077 2078 // Remove the speculative part of types and clean up the graph from 2079 // the extra CastPP nodes whose only purpose is to carry them. Do 2080 // that early so that optimizations are not disrupted by the extra 2081 // CastPP nodes. 2082 remove_speculative_types(igvn); 2083 2084 // No more new expensive nodes will be added to the list from here 2085 // so keep only the actual candidates for optimizations. 2086 cleanup_expensive_nodes(igvn); 2087 2088 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2089 if (EnableVectorSupport && has_vbox_nodes()) { 2090 TracePhase tp("", &timers[_t_vector]); 2091 PhaseVector pv(igvn); 2092 pv.optimize_vector_boxes(); 2093 2094 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2095 } 2096 assert(!has_vbox_nodes(), "sanity"); 2097 2098 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2099 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2100 initial_gvn()->replace_with(&igvn); 2101 for_igvn()->clear(); 2102 Unique_Node_List new_worklist(C->comp_arena()); 2103 { 2104 ResourceMark rm; 2105 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2106 } 2107 set_for_igvn(&new_worklist); 2108 igvn = PhaseIterGVN(initial_gvn()); 2109 igvn.optimize(); 2110 } 2111 2112 // Perform escape analysis 2113 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2114 if (has_loops()) { 2115 // Cleanup graph (remove dead nodes). 2116 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2117 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2118 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2119 if (failing()) return; 2120 } 2121 ConnectionGraph::do_analysis(this, &igvn); 2122 2123 if (failing()) return; 2124 2125 // Optimize out fields loads from scalar replaceable allocations. 2126 igvn.optimize(); 2127 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2128 2129 if (failing()) return; 2130 2131 if (congraph() != NULL && macro_count() > 0) { 2132 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2133 PhaseMacroExpand mexp(igvn); 2134 mexp.eliminate_macro_nodes(); 2135 igvn.set_delay_transform(false); 2136 2137 igvn.optimize(); 2138 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2139 2140 if (failing()) return; 2141 } 2142 } 2143 2144 // Loop transforms on the ideal graph. Range Check Elimination, 2145 // peeling, unrolling, etc. 2146 2147 // Set loop opts counter 2148 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2149 { 2150 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2151 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2152 _loop_opts_cnt--; 2153 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2154 if (failing()) return; 2155 } 2156 // Loop opts pass if partial peeling occurred in previous pass 2157 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2158 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2159 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2160 _loop_opts_cnt--; 2161 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2162 if (failing()) return; 2163 } 2164 // Loop opts pass for loop-unrolling before CCP 2165 if(major_progress() && (_loop_opts_cnt > 0)) { 2166 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2167 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2168 _loop_opts_cnt--; 2169 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2170 } 2171 if (!failing()) { 2172 // Verify that last round of loop opts produced a valid graph 2173 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2174 PhaseIdealLoop::verify(igvn); 2175 } 2176 } 2177 if (failing()) return; 2178 2179 // Conditional Constant Propagation; 2180 PhaseCCP ccp( &igvn ); 2181 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2182 { 2183 TracePhase tp("ccp", &timers[_t_ccp]); 2184 ccp.do_transform(); 2185 } 2186 print_method(PHASE_CPP1, 2); 2187 2188 assert( true, "Break here to ccp.dump_old2new_map()"); 2189 2190 // Iterative Global Value Numbering, including ideal transforms 2191 { 2192 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2193 igvn = ccp; 2194 igvn.optimize(); 2195 } 2196 print_method(PHASE_ITER_GVN2, 2); 2197 2198 if (failing()) return; 2199 2200 // Loop transforms on the ideal graph. Range Check Elimination, 2201 // peeling, unrolling, etc. 2202 if (!optimize_loops(igvn, LoopOptsDefault)) { 2203 return; 2204 } 2205 2206 if (failing()) return; 2207 2208 // Ensure that major progress is now clear 2209 C->clear_major_progress(); 2210 2211 { 2212 // Verify that all previous optimizations produced a valid graph 2213 // at least to this point, even if no loop optimizations were done. 2214 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2215 PhaseIdealLoop::verify(igvn); 2216 } 2217 2218 if (range_check_cast_count() > 0) { 2219 // No more loop optimizations. Remove all range check dependent CastIINodes. 2220 C->remove_range_check_casts(igvn); 2221 igvn.optimize(); 2222 } 2223 2224 #ifdef ASSERT 2225 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2226 #endif 2227 2228 { 2229 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2230 PhaseMacroExpand mex(igvn); 2231 if (mex.expand_macro_nodes()) { 2232 assert(failing(), "must bail out w/ explicit message"); 2233 return; 2234 } 2235 print_method(PHASE_MACRO_EXPANSION, 2); 2236 } 2237 2238 { 2239 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2240 if (bs->expand_barriers(this, igvn)) { 2241 assert(failing(), "must bail out w/ explicit message"); 2242 return; 2243 } 2244 print_method(PHASE_BARRIER_EXPANSION, 2); 2245 } 2246 2247 if (opaque4_count() > 0) { 2248 C->remove_opaque4_nodes(igvn); 2249 igvn.optimize(); 2250 } 2251 2252 if (C->max_vector_size() > 0) { 2253 C->optimize_logic_cones(igvn); 2254 igvn.optimize(); 2255 } 2256 2257 DEBUG_ONLY( _modified_nodes = NULL; ) 2258 } // (End scope of igvn; run destructor if necessary for asserts.) 2259 2260 process_print_inlining(); 2261 // A method with only infinite loops has no edges entering loops from root 2262 { 2263 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2264 if (final_graph_reshaping()) { 2265 assert(failing(), "must bail out w/ explicit message"); 2266 return; 2267 } 2268 } 2269 2270 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2271 DEBUG_ONLY(set_phase_optimize_finished();) 2272 } 2273 2274 //---------------------------- Bitwise operation packing optimization --------------------------- 2275 2276 static bool is_vector_unary_bitwise_op(Node* n) { 2277 return n->Opcode() == Op_XorV && 2278 VectorNode::is_vector_bitwise_not_pattern(n); 2279 } 2280 2281 static bool is_vector_binary_bitwise_op(Node* n) { 2282 switch (n->Opcode()) { 2283 case Op_AndV: 2284 case Op_OrV: 2285 return true; 2286 2287 case Op_XorV: 2288 return !is_vector_unary_bitwise_op(n); 2289 2290 default: 2291 return false; 2292 } 2293 } 2294 2295 static bool is_vector_ternary_bitwise_op(Node* n) { 2296 return n->Opcode() == Op_MacroLogicV; 2297 } 2298 2299 static bool is_vector_bitwise_op(Node* n) { 2300 return is_vector_unary_bitwise_op(n) || 2301 is_vector_binary_bitwise_op(n) || 2302 is_vector_ternary_bitwise_op(n); 2303 } 2304 2305 static bool is_vector_bitwise_cone_root(Node* n) { 2306 if (!is_vector_bitwise_op(n)) { 2307 return false; 2308 } 2309 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2310 if (is_vector_bitwise_op(n->fast_out(i))) { 2311 return false; 2312 } 2313 } 2314 return true; 2315 } 2316 2317 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2318 uint cnt = 0; 2319 if (is_vector_bitwise_op(n)) { 2320 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2321 for (uint i = 1; i < n->req(); i++) { 2322 Node* in = n->in(i); 2323 bool skip = VectorNode::is_all_ones_vector(in); 2324 if (!skip && !inputs.member(in)) { 2325 inputs.push(in); 2326 cnt++; 2327 } 2328 } 2329 assert(cnt <= 1, "not unary"); 2330 } else { 2331 uint last_req = n->req(); 2332 if (is_vector_ternary_bitwise_op(n)) { 2333 last_req = n->req() - 1; // skip last input 2334 } 2335 for (uint i = 1; i < last_req; i++) { 2336 Node* def = n->in(i); 2337 if (!inputs.member(def)) { 2338 inputs.push(def); 2339 cnt++; 2340 } 2341 } 2342 } 2343 partition.push(n); 2344 } else { // not a bitwise operations 2345 if (!inputs.member(n)) { 2346 inputs.push(n); 2347 cnt++; 2348 } 2349 } 2350 return cnt; 2351 } 2352 2353 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2354 Unique_Node_List useful_nodes; 2355 C->identify_useful_nodes(useful_nodes); 2356 2357 for (uint i = 0; i < useful_nodes.size(); i++) { 2358 Node* n = useful_nodes.at(i); 2359 if (is_vector_bitwise_cone_root(n)) { 2360 list.push(n); 2361 } 2362 } 2363 } 2364 2365 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2366 const TypeVect* vt, 2367 Unique_Node_List& partition, 2368 Unique_Node_List& inputs) { 2369 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2370 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2371 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2372 2373 Node* in1 = inputs.at(0); 2374 Node* in2 = inputs.at(1); 2375 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2376 2377 uint func = compute_truth_table(partition, inputs); 2378 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt)); 2379 } 2380 2381 static uint extract_bit(uint func, uint pos) { 2382 return (func & (1 << pos)) >> pos; 2383 } 2384 2385 // 2386 // A macro logic node represents a truth table. It has 4 inputs, 2387 // First three inputs corresponds to 3 columns of a truth table 2388 // and fourth input captures the logic function. 2389 // 2390 // eg. fn = (in1 AND in2) OR in3; 2391 // 2392 // MacroNode(in1,in2,in3,fn) 2393 // 2394 // ----------------- 2395 // in1 in2 in3 fn 2396 // ----------------- 2397 // 0 0 0 0 2398 // 0 0 1 1 2399 // 0 1 0 0 2400 // 0 1 1 1 2401 // 1 0 0 0 2402 // 1 0 1 1 2403 // 1 1 0 1 2404 // 1 1 1 1 2405 // 2406 2407 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2408 int res = 0; 2409 for (int i = 0; i < 8; i++) { 2410 int bit1 = extract_bit(in1, i); 2411 int bit2 = extract_bit(in2, i); 2412 int bit3 = extract_bit(in3, i); 2413 2414 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2415 int func_bit = extract_bit(func, func_bit_pos); 2416 2417 res |= func_bit << i; 2418 } 2419 return res; 2420 } 2421 2422 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2423 assert(n != NULL, ""); 2424 assert(eval_map.contains(n), "absent"); 2425 return *(eval_map.get(n)); 2426 } 2427 2428 static void eval_operands(Node* n, 2429 uint& func1, uint& func2, uint& func3, 2430 ResourceHashtable<Node*,uint>& eval_map) { 2431 assert(is_vector_bitwise_op(n), ""); 2432 2433 if (is_vector_unary_bitwise_op(n)) { 2434 Node* opnd = n->in(1); 2435 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2436 opnd = n->in(2); 2437 } 2438 func1 = eval_operand(opnd, eval_map); 2439 } else if (is_vector_binary_bitwise_op(n)) { 2440 func1 = eval_operand(n->in(1), eval_map); 2441 func2 = eval_operand(n->in(2), eval_map); 2442 } else { 2443 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2444 func1 = eval_operand(n->in(1), eval_map); 2445 func2 = eval_operand(n->in(2), eval_map); 2446 func3 = eval_operand(n->in(3), eval_map); 2447 } 2448 } 2449 2450 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2451 assert(inputs.size() <= 3, "sanity"); 2452 ResourceMark rm; 2453 uint res = 0; 2454 ResourceHashtable<Node*,uint> eval_map; 2455 2456 // Populate precomputed functions for inputs. 2457 // Each input corresponds to one column of 3 input truth-table. 2458 uint input_funcs[] = { 0xAA, // (_, _, a) -> a 2459 0xCC, // (_, b, _) -> b 2460 0xF0 }; // (c, _, _) -> c 2461 for (uint i = 0; i < inputs.size(); i++) { 2462 eval_map.put(inputs.at(i), input_funcs[i]); 2463 } 2464 2465 for (uint i = 0; i < partition.size(); i++) { 2466 Node* n = partition.at(i); 2467 2468 uint func1 = 0, func2 = 0, func3 = 0; 2469 eval_operands(n, func1, func2, func3, eval_map); 2470 2471 switch (n->Opcode()) { 2472 case Op_OrV: 2473 assert(func3 == 0, "not binary"); 2474 res = func1 | func2; 2475 break; 2476 case Op_AndV: 2477 assert(func3 == 0, "not binary"); 2478 res = func1 & func2; 2479 break; 2480 case Op_XorV: 2481 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2482 assert(func2 == 0 && func3 == 0, "not unary"); 2483 res = (~func1) & 0xFF; 2484 } else { 2485 assert(func3 == 0, "not binary"); 2486 res = func1 ^ func2; 2487 } 2488 break; 2489 case Op_MacroLogicV: 2490 // Ordering of inputs may change during evaluation of sub-tree 2491 // containing MacroLogic node as a child node, thus a re-evaluation 2492 // makes sure that function is evaluated in context of current 2493 // inputs. 2494 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2495 break; 2496 2497 default: assert(false, "not supported: %s", n->Name()); 2498 } 2499 assert(res <= 0xFF, "invalid"); 2500 eval_map.put(n, res); 2501 } 2502 return res; 2503 } 2504 2505 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2506 assert(partition.size() == 0, "not empty"); 2507 assert(inputs.size() == 0, "not empty"); 2508 if (is_vector_ternary_bitwise_op(n)) { 2509 return false; 2510 } 2511 2512 bool is_unary_op = is_vector_unary_bitwise_op(n); 2513 if (is_unary_op) { 2514 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary"); 2515 return false; // too few inputs 2516 } 2517 2518 assert(is_vector_binary_bitwise_op(n), "not binary"); 2519 Node* in1 = n->in(1); 2520 Node* in2 = n->in(2); 2521 2522 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs); 2523 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs); 2524 partition.push(n); 2525 2526 // Too many inputs? 2527 if (inputs.size() > 3) { 2528 partition.clear(); 2529 inputs.clear(); 2530 { // Recompute in2 inputs 2531 Unique_Node_List not_used; 2532 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used); 2533 } 2534 // Pick the node with minimum number of inputs. 2535 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) { 2536 return false; // still too many inputs 2537 } 2538 // Recompute partition & inputs. 2539 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2); 2540 collect_unique_inputs(child, partition, inputs); 2541 2542 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1); 2543 inputs.push(other_input); 2544 2545 partition.push(n); 2546 } 2547 2548 return (partition.size() == 2 || partition.size() == 3) && 2549 (inputs.size() == 2 || inputs.size() == 3); 2550 } 2551 2552 void Compile::inline_vector_reboxing_calls() { 2553 if (C->_vector_reboxing_late_inlines.length() > 0) { 2554 PhaseGVN* gvn = C->initial_gvn(); 2555 2556 _late_inlines_pos = C->_late_inlines.length(); 2557 while (_vector_reboxing_late_inlines.length() > 0) { 2558 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2559 cg->do_late_inline(); 2560 if (failing()) return; 2561 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node()); 2562 } 2563 _vector_reboxing_late_inlines.trunc_to(0); 2564 } 2565 } 2566 2567 bool Compile::has_vbox_nodes() { 2568 if (C->_vector_reboxing_late_inlines.length() > 0) { 2569 return true; 2570 } 2571 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2572 Node * n = C->macro_node(macro_idx); 2573 assert(n->is_macro(), "only macro nodes expected here"); 2574 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2575 return true; 2576 } 2577 } 2578 return false; 2579 } 2580 2581 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2582 assert(is_vector_bitwise_op(n), "not a root"); 2583 2584 visited.set(n->_idx); 2585 2586 // 1) Do a DFS walk over the logic cone. 2587 for (uint i = 1; i < n->req(); i++) { 2588 Node* in = n->in(i); 2589 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2590 process_logic_cone_root(igvn, in, visited); 2591 } 2592 } 2593 2594 // 2) Bottom up traversal: Merge node[s] with 2595 // the parent to form macro logic node. 2596 Unique_Node_List partition; 2597 Unique_Node_List inputs; 2598 if (compute_logic_cone(n, partition, inputs)) { 2599 const TypeVect* vt = n->bottom_type()->is_vect(); 2600 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2601 igvn.replace_node(n, macro_logic); 2602 } 2603 } 2604 2605 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2606 ResourceMark rm; 2607 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2608 Unique_Node_List list; 2609 collect_logic_cone_roots(list); 2610 2611 while (list.size() > 0) { 2612 Node* n = list.pop(); 2613 const TypeVect* vt = n->bottom_type()->is_vect(); 2614 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2615 if (supported) { 2616 VectorSet visited(comp_arena()); 2617 process_logic_cone_root(igvn, n, visited); 2618 } 2619 } 2620 } 2621 } 2622 2623 //------------------------------Code_Gen--------------------------------------- 2624 // Given a graph, generate code for it 2625 void Compile::Code_Gen() { 2626 if (failing()) { 2627 return; 2628 } 2629 2630 // Perform instruction selection. You might think we could reclaim Matcher 2631 // memory PDQ, but actually the Matcher is used in generating spill code. 2632 // Internals of the Matcher (including some VectorSets) must remain live 2633 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2634 // set a bit in reclaimed memory. 2635 2636 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2637 // nodes. Mapping is only valid at the root of each matched subtree. 2638 NOT_PRODUCT( verify_graph_edges(); ) 2639 2640 Matcher matcher; 2641 _matcher = &matcher; 2642 { 2643 TracePhase tp("matcher", &timers[_t_matcher]); 2644 matcher.match(); 2645 if (failing()) { 2646 return; 2647 } 2648 print_method(PHASE_AFTER_MATCHING, 3); 2649 } 2650 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2651 // nodes. Mapping is only valid at the root of each matched subtree. 2652 NOT_PRODUCT( verify_graph_edges(); ) 2653 2654 // If you have too many nodes, or if matching has failed, bail out 2655 check_node_count(0, "out of nodes matching instructions"); 2656 if (failing()) { 2657 return; 2658 } 2659 2660 print_method(PHASE_MATCHING, 2); 2661 2662 // Build a proper-looking CFG 2663 PhaseCFG cfg(node_arena(), root(), matcher); 2664 _cfg = &cfg; 2665 { 2666 TracePhase tp("scheduler", &timers[_t_scheduler]); 2667 bool success = cfg.do_global_code_motion(); 2668 if (!success) { 2669 return; 2670 } 2671 2672 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2673 NOT_PRODUCT( verify_graph_edges(); ) 2674 debug_only( cfg.verify(); ) 2675 } 2676 2677 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2678 _regalloc = ®alloc; 2679 { 2680 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2681 // Perform register allocation. After Chaitin, use-def chains are 2682 // no longer accurate (at spill code) and so must be ignored. 2683 // Node->LRG->reg mappings are still accurate. 2684 _regalloc->Register_Allocate(); 2685 2686 // Bail out if the allocator builds too many nodes 2687 if (failing()) { 2688 return; 2689 } 2690 } 2691 2692 // Prior to register allocation we kept empty basic blocks in case the 2693 // the allocator needed a place to spill. After register allocation we 2694 // are not adding any new instructions. If any basic block is empty, we 2695 // can now safely remove it. 2696 { 2697 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2698 cfg.remove_empty_blocks(); 2699 if (do_freq_based_layout()) { 2700 PhaseBlockLayout layout(cfg); 2701 } else { 2702 cfg.set_loop_alignment(); 2703 } 2704 cfg.fixup_flow(); 2705 } 2706 2707 // Apply peephole optimizations 2708 if( OptoPeephole ) { 2709 TracePhase tp("peephole", &timers[_t_peephole]); 2710 PhasePeephole peep( _regalloc, cfg); 2711 peep.do_transform(); 2712 } 2713 2714 // Do late expand if CPU requires this. 2715 if (Matcher::require_postalloc_expand) { 2716 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2717 cfg.postalloc_expand(_regalloc); 2718 } 2719 2720 // Convert Nodes to instruction bits in a buffer 2721 { 2722 TracePhase tp("output", &timers[_t_output]); 2723 PhaseOutput output; 2724 output.Output(); 2725 if (failing()) return; 2726 output.install(); 2727 } 2728 2729 print_method(PHASE_FINAL_CODE); 2730 2731 // He's dead, Jim. 2732 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2733 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2734 } 2735 2736 //------------------------------Final_Reshape_Counts--------------------------- 2737 // This class defines counters to help identify when a method 2738 // may/must be executed using hardware with only 24-bit precision. 2739 struct Final_Reshape_Counts : public StackObj { 2740 int _call_count; // count non-inlined 'common' calls 2741 int _float_count; // count float ops requiring 24-bit precision 2742 int _double_count; // count double ops requiring more precision 2743 int _java_call_count; // count non-inlined 'java' calls 2744 int _inner_loop_count; // count loops which need alignment 2745 VectorSet _visited; // Visitation flags 2746 Node_List _tests; // Set of IfNodes & PCTableNodes 2747 2748 Final_Reshape_Counts() : 2749 _call_count(0), _float_count(0), _double_count(0), 2750 _java_call_count(0), _inner_loop_count(0) { } 2751 2752 void inc_call_count () { _call_count ++; } 2753 void inc_float_count () { _float_count ++; } 2754 void inc_double_count() { _double_count++; } 2755 void inc_java_call_count() { _java_call_count++; } 2756 void inc_inner_loop_count() { _inner_loop_count++; } 2757 2758 int get_call_count () const { return _call_count ; } 2759 int get_float_count () const { return _float_count ; } 2760 int get_double_count() const { return _double_count; } 2761 int get_java_call_count() const { return _java_call_count; } 2762 int get_inner_loop_count() const { return _inner_loop_count; } 2763 }; 2764 2765 #ifdef ASSERT 2766 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2767 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2768 // Make sure the offset goes inside the instance layout. 2769 return k->contains_field_offset(tp->offset()); 2770 // Note that OffsetBot and OffsetTop are very negative. 2771 } 2772 #endif 2773 2774 // Eliminate trivially redundant StoreCMs and accumulate their 2775 // precedence edges. 2776 void Compile::eliminate_redundant_card_marks(Node* n) { 2777 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2778 if (n->in(MemNode::Address)->outcnt() > 1) { 2779 // There are multiple users of the same address so it might be 2780 // possible to eliminate some of the StoreCMs 2781 Node* mem = n->in(MemNode::Memory); 2782 Node* adr = n->in(MemNode::Address); 2783 Node* val = n->in(MemNode::ValueIn); 2784 Node* prev = n; 2785 bool done = false; 2786 // Walk the chain of StoreCMs eliminating ones that match. As 2787 // long as it's a chain of single users then the optimization is 2788 // safe. Eliminating partially redundant StoreCMs would require 2789 // cloning copies down the other paths. 2790 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2791 if (adr == mem->in(MemNode::Address) && 2792 val == mem->in(MemNode::ValueIn)) { 2793 // redundant StoreCM 2794 if (mem->req() > MemNode::OopStore) { 2795 // Hasn't been processed by this code yet. 2796 n->add_prec(mem->in(MemNode::OopStore)); 2797 } else { 2798 // Already converted to precedence edge 2799 for (uint i = mem->req(); i < mem->len(); i++) { 2800 // Accumulate any precedence edges 2801 if (mem->in(i) != NULL) { 2802 n->add_prec(mem->in(i)); 2803 } 2804 } 2805 // Everything above this point has been processed. 2806 done = true; 2807 } 2808 // Eliminate the previous StoreCM 2809 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2810 assert(mem->outcnt() == 0, "should be dead"); 2811 mem->disconnect_inputs(NULL, this); 2812 } else { 2813 prev = mem; 2814 } 2815 mem = prev->in(MemNode::Memory); 2816 } 2817 } 2818 } 2819 2820 //------------------------------final_graph_reshaping_impl---------------------- 2821 // Implement items 1-5 from final_graph_reshaping below. 2822 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2823 2824 if ( n->outcnt() == 0 ) return; // dead node 2825 uint nop = n->Opcode(); 2826 2827 // Check for 2-input instruction with "last use" on right input. 2828 // Swap to left input. Implements item (2). 2829 if( n->req() == 3 && // two-input instruction 2830 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2831 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2832 n->in(2)->outcnt() == 1 &&// right use IS a last use 2833 !n->in(2)->is_Con() ) { // right use is not a constant 2834 // Check for commutative opcode 2835 switch( nop ) { 2836 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2837 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 2838 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 2839 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2840 case Op_AndL: case Op_XorL: case Op_OrL: 2841 case Op_AndI: case Op_XorI: case Op_OrI: { 2842 // Move "last use" input to left by swapping inputs 2843 n->swap_edges(1, 2); 2844 break; 2845 } 2846 default: 2847 break; 2848 } 2849 } 2850 2851 #ifdef ASSERT 2852 if( n->is_Mem() ) { 2853 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2854 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2855 // oop will be recorded in oop map if load crosses safepoint 2856 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2857 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2858 "raw memory operations should have control edge"); 2859 } 2860 if (n->is_MemBar()) { 2861 MemBarNode* mb = n->as_MemBar(); 2862 if (mb->trailing_store() || mb->trailing_load_store()) { 2863 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2864 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 2865 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2866 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2867 } else if (mb->leading()) { 2868 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2869 } 2870 } 2871 #endif 2872 // Count FPU ops and common calls, implements item (3) 2873 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2874 if (!gc_handled) { 2875 final_graph_reshaping_main_switch(n, frc, nop); 2876 } 2877 2878 // Collect CFG split points 2879 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2880 frc._tests.push(n); 2881 } 2882 } 2883 2884 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2885 switch( nop ) { 2886 // Count all float operations that may use FPU 2887 case Op_AddF: 2888 case Op_SubF: 2889 case Op_MulF: 2890 case Op_DivF: 2891 case Op_NegF: 2892 case Op_ModF: 2893 case Op_ConvI2F: 2894 case Op_ConF: 2895 case Op_CmpF: 2896 case Op_CmpF3: 2897 // case Op_ConvL2F: // longs are split into 32-bit halves 2898 frc.inc_float_count(); 2899 break; 2900 2901 case Op_ConvF2D: 2902 case Op_ConvD2F: 2903 frc.inc_float_count(); 2904 frc.inc_double_count(); 2905 break; 2906 2907 // Count all double operations that may use FPU 2908 case Op_AddD: 2909 case Op_SubD: 2910 case Op_MulD: 2911 case Op_DivD: 2912 case Op_NegD: 2913 case Op_ModD: 2914 case Op_ConvI2D: 2915 case Op_ConvD2I: 2916 // case Op_ConvL2D: // handled by leaf call 2917 // case Op_ConvD2L: // handled by leaf call 2918 case Op_ConD: 2919 case Op_CmpD: 2920 case Op_CmpD3: 2921 frc.inc_double_count(); 2922 break; 2923 case Op_Opaque1: // Remove Opaque Nodes before matching 2924 case Op_Opaque2: // Remove Opaque Nodes before matching 2925 case Op_Opaque3: 2926 n->subsume_by(n->in(1), this); 2927 break; 2928 case Op_CallStaticJava: 2929 case Op_CallJava: 2930 case Op_CallDynamicJava: 2931 frc.inc_java_call_count(); // Count java call site; 2932 case Op_CallRuntime: 2933 case Op_CallLeaf: 2934 case Op_CallLeafNoFP: { 2935 assert (n->is_Call(), ""); 2936 CallNode *call = n->as_Call(); 2937 // Count call sites where the FP mode bit would have to be flipped. 2938 // Do not count uncommon runtime calls: 2939 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2940 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2941 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2942 frc.inc_call_count(); // Count the call site 2943 } else { // See if uncommon argument is shared 2944 Node *n = call->in(TypeFunc::Parms); 2945 int nop = n->Opcode(); 2946 // Clone shared simple arguments to uncommon calls, item (1). 2947 if (n->outcnt() > 1 && 2948 !n->is_Proj() && 2949 nop != Op_CreateEx && 2950 nop != Op_CheckCastPP && 2951 nop != Op_DecodeN && 2952 nop != Op_DecodeNKlass && 2953 !n->is_Mem() && 2954 !n->is_Phi()) { 2955 Node *x = n->clone(); 2956 call->set_req(TypeFunc::Parms, x); 2957 } 2958 } 2959 break; 2960 } 2961 2962 case Op_StoreD: 2963 case Op_LoadD: 2964 case Op_LoadD_unaligned: 2965 frc.inc_double_count(); 2966 goto handle_mem; 2967 case Op_StoreF: 2968 case Op_LoadF: 2969 frc.inc_float_count(); 2970 goto handle_mem; 2971 2972 case Op_StoreCM: 2973 { 2974 // Convert OopStore dependence into precedence edge 2975 Node* prec = n->in(MemNode::OopStore); 2976 n->del_req(MemNode::OopStore); 2977 n->add_prec(prec); 2978 eliminate_redundant_card_marks(n); 2979 } 2980 2981 // fall through 2982 2983 case Op_StoreB: 2984 case Op_StoreC: 2985 case Op_StorePConditional: 2986 case Op_StoreI: 2987 case Op_StoreL: 2988 case Op_StoreIConditional: 2989 case Op_StoreLConditional: 2990 case Op_CompareAndSwapB: 2991 case Op_CompareAndSwapS: 2992 case Op_CompareAndSwapI: 2993 case Op_CompareAndSwapL: 2994 case Op_CompareAndSwapP: 2995 case Op_CompareAndSwapN: 2996 case Op_WeakCompareAndSwapB: 2997 case Op_WeakCompareAndSwapS: 2998 case Op_WeakCompareAndSwapI: 2999 case Op_WeakCompareAndSwapL: 3000 case Op_WeakCompareAndSwapP: 3001 case Op_WeakCompareAndSwapN: 3002 case Op_CompareAndExchangeB: 3003 case Op_CompareAndExchangeS: 3004 case Op_CompareAndExchangeI: 3005 case Op_CompareAndExchangeL: 3006 case Op_CompareAndExchangeP: 3007 case Op_CompareAndExchangeN: 3008 case Op_GetAndAddS: 3009 case Op_GetAndAddB: 3010 case Op_GetAndAddI: 3011 case Op_GetAndAddL: 3012 case Op_GetAndSetS: 3013 case Op_GetAndSetB: 3014 case Op_GetAndSetI: 3015 case Op_GetAndSetL: 3016 case Op_GetAndSetP: 3017 case Op_GetAndSetN: 3018 case Op_StoreP: 3019 case Op_StoreN: 3020 case Op_StoreNKlass: 3021 case Op_LoadB: 3022 case Op_LoadUB: 3023 case Op_LoadUS: 3024 case Op_LoadI: 3025 case Op_LoadKlass: 3026 case Op_LoadNKlass: 3027 case Op_LoadL: 3028 case Op_LoadL_unaligned: 3029 case Op_LoadPLocked: 3030 case Op_LoadP: 3031 case Op_LoadN: 3032 case Op_LoadRange: 3033 case Op_LoadS: { 3034 handle_mem: 3035 #ifdef ASSERT 3036 if( VerifyOptoOopOffsets ) { 3037 MemNode* mem = n->as_Mem(); 3038 // Check to see if address types have grounded out somehow. 3039 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3040 assert( !tp || oop_offset_is_sane(tp), "" ); 3041 } 3042 #endif 3043 break; 3044 } 3045 3046 case Op_AddP: { // Assert sane base pointers 3047 Node *addp = n->in(AddPNode::Address); 3048 assert( !addp->is_AddP() || 3049 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3050 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3051 "Base pointers must match (addp %u)", addp->_idx ); 3052 #ifdef _LP64 3053 if ((UseCompressedOops || UseCompressedClassPointers) && 3054 addp->Opcode() == Op_ConP && 3055 addp == n->in(AddPNode::Base) && 3056 n->in(AddPNode::Offset)->is_Con()) { 3057 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3058 // on the platform and on the compressed oops mode. 3059 // Use addressing with narrow klass to load with offset on x86. 3060 // Some platforms can use the constant pool to load ConP. 3061 // Do this transformation here since IGVN will convert ConN back to ConP. 3062 const Type* t = addp->bottom_type(); 3063 bool is_oop = t->isa_oopptr() != NULL; 3064 bool is_klass = t->isa_klassptr() != NULL; 3065 3066 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3067 (is_klass && Matcher::const_klass_prefer_decode())) { 3068 Node* nn = NULL; 3069 3070 int op = is_oop ? Op_ConN : Op_ConNKlass; 3071 3072 // Look for existing ConN node of the same exact type. 3073 Node* r = root(); 3074 uint cnt = r->outcnt(); 3075 for (uint i = 0; i < cnt; i++) { 3076 Node* m = r->raw_out(i); 3077 if (m!= NULL && m->Opcode() == op && 3078 m->bottom_type()->make_ptr() == t) { 3079 nn = m; 3080 break; 3081 } 3082 } 3083 if (nn != NULL) { 3084 // Decode a narrow oop to match address 3085 // [R12 + narrow_oop_reg<<3 + offset] 3086 if (is_oop) { 3087 nn = new DecodeNNode(nn, t); 3088 } else { 3089 nn = new DecodeNKlassNode(nn, t); 3090 } 3091 // Check for succeeding AddP which uses the same Base. 3092 // Otherwise we will run into the assertion above when visiting that guy. 3093 for (uint i = 0; i < n->outcnt(); ++i) { 3094 Node *out_i = n->raw_out(i); 3095 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3096 out_i->set_req(AddPNode::Base, nn); 3097 #ifdef ASSERT 3098 for (uint j = 0; j < out_i->outcnt(); ++j) { 3099 Node *out_j = out_i->raw_out(j); 3100 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3101 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3102 } 3103 #endif 3104 } 3105 } 3106 n->set_req(AddPNode::Base, nn); 3107 n->set_req(AddPNode::Address, nn); 3108 if (addp->outcnt() == 0) { 3109 addp->disconnect_inputs(NULL, this); 3110 } 3111 } 3112 } 3113 } 3114 #endif 3115 // platform dependent reshaping of the address expression 3116 reshape_address(n->as_AddP()); 3117 break; 3118 } 3119 3120 case Op_CastPP: { 3121 // Remove CastPP nodes to gain more freedom during scheduling but 3122 // keep the dependency they encode as control or precedence edges 3123 // (if control is set already) on memory operations. Some CastPP 3124 // nodes don't have a control (don't carry a dependency): skip 3125 // those. 3126 if (n->in(0) != NULL) { 3127 ResourceMark rm; 3128 Unique_Node_List wq; 3129 wq.push(n); 3130 for (uint next = 0; next < wq.size(); ++next) { 3131 Node *m = wq.at(next); 3132 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3133 Node* use = m->fast_out(i); 3134 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3135 use->ensure_control_or_add_prec(n->in(0)); 3136 } else { 3137 switch(use->Opcode()) { 3138 case Op_AddP: 3139 case Op_DecodeN: 3140 case Op_DecodeNKlass: 3141 case Op_CheckCastPP: 3142 case Op_CastPP: 3143 wq.push(use); 3144 break; 3145 } 3146 } 3147 } 3148 } 3149 } 3150 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3151 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3152 Node* in1 = n->in(1); 3153 const Type* t = n->bottom_type(); 3154 Node* new_in1 = in1->clone(); 3155 new_in1->as_DecodeN()->set_type(t); 3156 3157 if (!Matcher::narrow_oop_use_complex_address()) { 3158 // 3159 // x86, ARM and friends can handle 2 adds in addressing mode 3160 // and Matcher can fold a DecodeN node into address by using 3161 // a narrow oop directly and do implicit NULL check in address: 3162 // 3163 // [R12 + narrow_oop_reg<<3 + offset] 3164 // NullCheck narrow_oop_reg 3165 // 3166 // On other platforms (Sparc) we have to keep new DecodeN node and 3167 // use it to do implicit NULL check in address: 3168 // 3169 // decode_not_null narrow_oop_reg, base_reg 3170 // [base_reg + offset] 3171 // NullCheck base_reg 3172 // 3173 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3174 // to keep the information to which NULL check the new DecodeN node 3175 // corresponds to use it as value in implicit_null_check(). 3176 // 3177 new_in1->set_req(0, n->in(0)); 3178 } 3179 3180 n->subsume_by(new_in1, this); 3181 if (in1->outcnt() == 0) { 3182 in1->disconnect_inputs(NULL, this); 3183 } 3184 } else { 3185 n->subsume_by(n->in(1), this); 3186 if (n->outcnt() == 0) { 3187 n->disconnect_inputs(NULL, this); 3188 } 3189 } 3190 break; 3191 } 3192 #ifdef _LP64 3193 case Op_CmpP: 3194 // Do this transformation here to preserve CmpPNode::sub() and 3195 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3196 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3197 Node* in1 = n->in(1); 3198 Node* in2 = n->in(2); 3199 if (!in1->is_DecodeNarrowPtr()) { 3200 in2 = in1; 3201 in1 = n->in(2); 3202 } 3203 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3204 3205 Node* new_in2 = NULL; 3206 if (in2->is_DecodeNarrowPtr()) { 3207 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3208 new_in2 = in2->in(1); 3209 } else if (in2->Opcode() == Op_ConP) { 3210 const Type* t = in2->bottom_type(); 3211 if (t == TypePtr::NULL_PTR) { 3212 assert(in1->is_DecodeN(), "compare klass to null?"); 3213 // Don't convert CmpP null check into CmpN if compressed 3214 // oops implicit null check is not generated. 3215 // This will allow to generate normal oop implicit null check. 3216 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3217 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3218 // 3219 // This transformation together with CastPP transformation above 3220 // will generated code for implicit NULL checks for compressed oops. 3221 // 3222 // The original code after Optimize() 3223 // 3224 // LoadN memory, narrow_oop_reg 3225 // decode narrow_oop_reg, base_reg 3226 // CmpP base_reg, NULL 3227 // CastPP base_reg // NotNull 3228 // Load [base_reg + offset], val_reg 3229 // 3230 // after these transformations will be 3231 // 3232 // LoadN memory, narrow_oop_reg 3233 // CmpN narrow_oop_reg, NULL 3234 // decode_not_null narrow_oop_reg, base_reg 3235 // Load [base_reg + offset], val_reg 3236 // 3237 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3238 // since narrow oops can be used in debug info now (see the code in 3239 // final_graph_reshaping_walk()). 3240 // 3241 // At the end the code will be matched to 3242 // on x86: 3243 // 3244 // Load_narrow_oop memory, narrow_oop_reg 3245 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3246 // NullCheck narrow_oop_reg 3247 // 3248 // and on sparc: 3249 // 3250 // Load_narrow_oop memory, narrow_oop_reg 3251 // decode_not_null narrow_oop_reg, base_reg 3252 // Load [base_reg + offset], val_reg 3253 // NullCheck base_reg 3254 // 3255 } else if (t->isa_oopptr()) { 3256 new_in2 = ConNode::make(t->make_narrowoop()); 3257 } else if (t->isa_klassptr()) { 3258 new_in2 = ConNode::make(t->make_narrowklass()); 3259 } 3260 } 3261 if (new_in2 != NULL) { 3262 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3263 n->subsume_by(cmpN, this); 3264 if (in1->outcnt() == 0) { 3265 in1->disconnect_inputs(NULL, this); 3266 } 3267 if (in2->outcnt() == 0) { 3268 in2->disconnect_inputs(NULL, this); 3269 } 3270 } 3271 } 3272 break; 3273 3274 case Op_DecodeN: 3275 case Op_DecodeNKlass: 3276 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3277 // DecodeN could be pinned when it can't be fold into 3278 // an address expression, see the code for Op_CastPP above. 3279 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3280 break; 3281 3282 case Op_EncodeP: 3283 case Op_EncodePKlass: { 3284 Node* in1 = n->in(1); 3285 if (in1->is_DecodeNarrowPtr()) { 3286 n->subsume_by(in1->in(1), this); 3287 } else if (in1->Opcode() == Op_ConP) { 3288 const Type* t = in1->bottom_type(); 3289 if (t == TypePtr::NULL_PTR) { 3290 assert(t->isa_oopptr(), "null klass?"); 3291 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3292 } else if (t->isa_oopptr()) { 3293 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3294 } else if (t->isa_klassptr()) { 3295 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3296 } 3297 } 3298 if (in1->outcnt() == 0) { 3299 in1->disconnect_inputs(NULL, this); 3300 } 3301 break; 3302 } 3303 3304 case Op_Proj: { 3305 if (OptimizeStringConcat) { 3306 ProjNode* p = n->as_Proj(); 3307 if (p->_is_io_use) { 3308 // Separate projections were used for the exception path which 3309 // are normally removed by a late inline. If it wasn't inlined 3310 // then they will hang around and should just be replaced with 3311 // the original one. 3312 Node* proj = NULL; 3313 // Replace with just one 3314 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3315 Node *use = i.get(); 3316 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3317 proj = use; 3318 break; 3319 } 3320 } 3321 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3322 if (proj != NULL) { 3323 p->subsume_by(proj, this); 3324 } 3325 } 3326 } 3327 break; 3328 } 3329 3330 case Op_Phi: 3331 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3332 // The EncodeP optimization may create Phi with the same edges 3333 // for all paths. It is not handled well by Register Allocator. 3334 Node* unique_in = n->in(1); 3335 assert(unique_in != NULL, ""); 3336 uint cnt = n->req(); 3337 for (uint i = 2; i < cnt; i++) { 3338 Node* m = n->in(i); 3339 assert(m != NULL, ""); 3340 if (unique_in != m) 3341 unique_in = NULL; 3342 } 3343 if (unique_in != NULL) { 3344 n->subsume_by(unique_in, this); 3345 } 3346 } 3347 break; 3348 3349 #endif 3350 3351 #ifdef ASSERT 3352 case Op_CastII: 3353 // Verify that all range check dependent CastII nodes were removed. 3354 if (n->isa_CastII()->has_range_check()) { 3355 n->dump(3); 3356 assert(false, "Range check dependent CastII node was not removed"); 3357 } 3358 break; 3359 #endif 3360 3361 case Op_ModI: 3362 if (UseDivMod) { 3363 // Check if a%b and a/b both exist 3364 Node* d = n->find_similar(Op_DivI); 3365 if (d) { 3366 // Replace them with a fused divmod if supported 3367 if (Matcher::has_match_rule(Op_DivModI)) { 3368 DivModINode* divmod = DivModINode::make(n); 3369 d->subsume_by(divmod->div_proj(), this); 3370 n->subsume_by(divmod->mod_proj(), this); 3371 } else { 3372 // replace a%b with a-((a/b)*b) 3373 Node* mult = new MulINode(d, d->in(2)); 3374 Node* sub = new SubINode(d->in(1), mult); 3375 n->subsume_by(sub, this); 3376 } 3377 } 3378 } 3379 break; 3380 3381 case Op_ModL: 3382 if (UseDivMod) { 3383 // Check if a%b and a/b both exist 3384 Node* d = n->find_similar(Op_DivL); 3385 if (d) { 3386 // Replace them with a fused divmod if supported 3387 if (Matcher::has_match_rule(Op_DivModL)) { 3388 DivModLNode* divmod = DivModLNode::make(n); 3389 d->subsume_by(divmod->div_proj(), this); 3390 n->subsume_by(divmod->mod_proj(), this); 3391 } else { 3392 // replace a%b with a-((a/b)*b) 3393 Node* mult = new MulLNode(d, d->in(2)); 3394 Node* sub = new SubLNode(d->in(1), mult); 3395 n->subsume_by(sub, this); 3396 } 3397 } 3398 } 3399 break; 3400 3401 case Op_LoadVector: 3402 case Op_StoreVector: 3403 case Op_LoadVectorGather: 3404 case Op_StoreVectorScatter: 3405 break; 3406 3407 case Op_AddReductionVI: 3408 case Op_AddReductionVL: 3409 case Op_AddReductionVF: 3410 case Op_AddReductionVD: 3411 case Op_MulReductionVI: 3412 case Op_MulReductionVL: 3413 case Op_MulReductionVF: 3414 case Op_MulReductionVD: 3415 case Op_MinReductionV: 3416 case Op_MaxReductionV: 3417 case Op_AndReductionV: 3418 case Op_OrReductionV: 3419 case Op_XorReductionV: 3420 break; 3421 3422 case Op_PackB: 3423 case Op_PackS: 3424 case Op_PackI: 3425 case Op_PackF: 3426 case Op_PackL: 3427 case Op_PackD: 3428 if (n->req()-1 > 2) { 3429 // Replace many operand PackNodes with a binary tree for matching 3430 PackNode* p = (PackNode*) n; 3431 Node* btp = p->binary_tree_pack(1, n->req()); 3432 n->subsume_by(btp, this); 3433 } 3434 break; 3435 case Op_Loop: 3436 case Op_CountedLoop: 3437 case Op_OuterStripMinedLoop: 3438 if (n->as_Loop()->is_inner_loop()) { 3439 frc.inc_inner_loop_count(); 3440 } 3441 n->as_Loop()->verify_strip_mined(0); 3442 break; 3443 case Op_LShiftI: 3444 case Op_RShiftI: 3445 case Op_URShiftI: 3446 case Op_LShiftL: 3447 case Op_RShiftL: 3448 case Op_URShiftL: 3449 if (Matcher::need_masked_shift_count) { 3450 // The cpu's shift instructions don't restrict the count to the 3451 // lower 5/6 bits. We need to do the masking ourselves. 3452 Node* in2 = n->in(2); 3453 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3454 const TypeInt* t = in2->find_int_type(); 3455 if (t != NULL && t->is_con()) { 3456 juint shift = t->get_con(); 3457 if (shift > mask) { // Unsigned cmp 3458 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3459 } 3460 } else { 3461 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3462 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3463 n->set_req(2, shift); 3464 } 3465 } 3466 if (in2->outcnt() == 0) { // Remove dead node 3467 in2->disconnect_inputs(NULL, this); 3468 } 3469 } 3470 break; 3471 case Op_MemBarStoreStore: 3472 case Op_MemBarRelease: 3473 // Break the link with AllocateNode: it is no longer useful and 3474 // confuses register allocation. 3475 if (n->req() > MemBarNode::Precedent) { 3476 n->set_req(MemBarNode::Precedent, top()); 3477 } 3478 break; 3479 case Op_MemBarAcquire: { 3480 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3481 // At parse time, the trailing MemBarAcquire for a volatile load 3482 // is created with an edge to the load. After optimizations, 3483 // that input may be a chain of Phis. If those phis have no 3484 // other use, then the MemBarAcquire keeps them alive and 3485 // register allocation can be confused. 3486 ResourceMark rm; 3487 Unique_Node_List wq; 3488 wq.push(n->in(MemBarNode::Precedent)); 3489 n->set_req(MemBarNode::Precedent, top()); 3490 while (wq.size() > 0) { 3491 Node* m = wq.pop(); 3492 if (m->outcnt() == 0) { 3493 for (uint j = 0; j < m->req(); j++) { 3494 Node* in = m->in(j); 3495 if (in != NULL) { 3496 wq.push(in); 3497 } 3498 } 3499 m->disconnect_inputs(NULL, this); 3500 } 3501 } 3502 } 3503 break; 3504 } 3505 case Op_RangeCheck: { 3506 RangeCheckNode* rc = n->as_RangeCheck(); 3507 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3508 n->subsume_by(iff, this); 3509 frc._tests.push(iff); 3510 break; 3511 } 3512 case Op_ConvI2L: { 3513 if (!Matcher::convi2l_type_required) { 3514 // Code generation on some platforms doesn't need accurate 3515 // ConvI2L types. Widening the type can help remove redundant 3516 // address computations. 3517 n->as_Type()->set_type(TypeLong::INT); 3518 ResourceMark rm; 3519 Unique_Node_List wq; 3520 wq.push(n); 3521 for (uint next = 0; next < wq.size(); next++) { 3522 Node *m = wq.at(next); 3523 3524 for(;;) { 3525 // Loop over all nodes with identical inputs edges as m 3526 Node* k = m->find_similar(m->Opcode()); 3527 if (k == NULL) { 3528 break; 3529 } 3530 // Push their uses so we get a chance to remove node made 3531 // redundant 3532 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3533 Node* u = k->fast_out(i); 3534 if (u->Opcode() == Op_LShiftL || 3535 u->Opcode() == Op_AddL || 3536 u->Opcode() == Op_SubL || 3537 u->Opcode() == Op_AddP) { 3538 wq.push(u); 3539 } 3540 } 3541 // Replace all nodes with identical edges as m with m 3542 k->subsume_by(m, this); 3543 } 3544 } 3545 } 3546 break; 3547 } 3548 case Op_CmpUL: { 3549 if (!Matcher::has_match_rule(Op_CmpUL)) { 3550 // No support for unsigned long comparisons 3551 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3552 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3553 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3554 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3555 Node* andl = new AndLNode(orl, remove_sign_mask); 3556 Node* cmp = new CmpLNode(andl, n->in(2)); 3557 n->subsume_by(cmp, this); 3558 } 3559 break; 3560 } 3561 default: 3562 assert(!n->is_Call(), ""); 3563 assert(!n->is_Mem(), ""); 3564 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3565 break; 3566 } 3567 } 3568 3569 //------------------------------final_graph_reshaping_walk--------------------- 3570 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3571 // requires that the walk visits a node's inputs before visiting the node. 3572 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3573 Unique_Node_List sfpt; 3574 3575 frc._visited.set(root->_idx); // first, mark node as visited 3576 uint cnt = root->req(); 3577 Node *n = root; 3578 uint i = 0; 3579 while (true) { 3580 if (i < cnt) { 3581 // Place all non-visited non-null inputs onto stack 3582 Node* m = n->in(i); 3583 ++i; 3584 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3585 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3586 // compute worst case interpreter size in case of a deoptimization 3587 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3588 3589 sfpt.push(m); 3590 } 3591 cnt = m->req(); 3592 nstack.push(n, i); // put on stack parent and next input's index 3593 n = m; 3594 i = 0; 3595 } 3596 } else { 3597 // Now do post-visit work 3598 final_graph_reshaping_impl( n, frc ); 3599 if (nstack.is_empty()) 3600 break; // finished 3601 n = nstack.node(); // Get node from stack 3602 cnt = n->req(); 3603 i = nstack.index(); 3604 nstack.pop(); // Shift to the next node on stack 3605 } 3606 } 3607 3608 // Skip next transformation if compressed oops are not used. 3609 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3610 (!UseCompressedOops && !UseCompressedClassPointers)) 3611 return; 3612 3613 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3614 // It could be done for an uncommon traps or any safepoints/calls 3615 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3616 while (sfpt.size() > 0) { 3617 n = sfpt.pop(); 3618 JVMState *jvms = n->as_SafePoint()->jvms(); 3619 assert(jvms != NULL, "sanity"); 3620 int start = jvms->debug_start(); 3621 int end = n->req(); 3622 bool is_uncommon = (n->is_CallStaticJava() && 3623 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3624 for (int j = start; j < end; j++) { 3625 Node* in = n->in(j); 3626 if (in->is_DecodeNarrowPtr()) { 3627 bool safe_to_skip = true; 3628 if (!is_uncommon ) { 3629 // Is it safe to skip? 3630 for (uint i = 0; i < in->outcnt(); i++) { 3631 Node* u = in->raw_out(i); 3632 if (!u->is_SafePoint() || 3633 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3634 safe_to_skip = false; 3635 } 3636 } 3637 } 3638 if (safe_to_skip) { 3639 n->set_req(j, in->in(1)); 3640 } 3641 if (in->outcnt() == 0) { 3642 in->disconnect_inputs(NULL, this); 3643 } 3644 } 3645 } 3646 } 3647 } 3648 3649 //------------------------------final_graph_reshaping-------------------------- 3650 // Final Graph Reshaping. 3651 // 3652 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3653 // and not commoned up and forced early. Must come after regular 3654 // optimizations to avoid GVN undoing the cloning. Clone constant 3655 // inputs to Loop Phis; these will be split by the allocator anyways. 3656 // Remove Opaque nodes. 3657 // (2) Move last-uses by commutative operations to the left input to encourage 3658 // Intel update-in-place two-address operations and better register usage 3659 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3660 // calls canonicalizing them back. 3661 // (3) Count the number of double-precision FP ops, single-precision FP ops 3662 // and call sites. On Intel, we can get correct rounding either by 3663 // forcing singles to memory (requires extra stores and loads after each 3664 // FP bytecode) or we can set a rounding mode bit (requires setting and 3665 // clearing the mode bit around call sites). The mode bit is only used 3666 // if the relative frequency of single FP ops to calls is low enough. 3667 // This is a key transform for SPEC mpeg_audio. 3668 // (4) Detect infinite loops; blobs of code reachable from above but not 3669 // below. Several of the Code_Gen algorithms fail on such code shapes, 3670 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3671 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3672 // Detection is by looking for IfNodes where only 1 projection is 3673 // reachable from below or CatchNodes missing some targets. 3674 // (5) Assert for insane oop offsets in debug mode. 3675 3676 bool Compile::final_graph_reshaping() { 3677 // an infinite loop may have been eliminated by the optimizer, 3678 // in which case the graph will be empty. 3679 if (root()->req() == 1) { 3680 record_method_not_compilable("trivial infinite loop"); 3681 return true; 3682 } 3683 3684 // Expensive nodes have their control input set to prevent the GVN 3685 // from freely commoning them. There's no GVN beyond this point so 3686 // no need to keep the control input. We want the expensive nodes to 3687 // be freely moved to the least frequent code path by gcm. 3688 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3689 for (int i = 0; i < expensive_count(); i++) { 3690 _expensive_nodes->at(i)->set_req(0, NULL); 3691 } 3692 3693 Final_Reshape_Counts frc; 3694 3695 // Visit everybody reachable! 3696 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3697 Node_Stack nstack(live_nodes() >> 1); 3698 final_graph_reshaping_walk(nstack, root(), frc); 3699 3700 // Check for unreachable (from below) code (i.e., infinite loops). 3701 for( uint i = 0; i < frc._tests.size(); i++ ) { 3702 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3703 // Get number of CFG targets. 3704 // Note that PCTables include exception targets after calls. 3705 uint required_outcnt = n->required_outcnt(); 3706 if (n->outcnt() != required_outcnt) { 3707 // Check for a few special cases. Rethrow Nodes never take the 3708 // 'fall-thru' path, so expected kids is 1 less. 3709 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3710 if (n->in(0)->in(0)->is_Call()) { 3711 CallNode *call = n->in(0)->in(0)->as_Call(); 3712 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3713 required_outcnt--; // Rethrow always has 1 less kid 3714 } else if (call->req() > TypeFunc::Parms && 3715 call->is_CallDynamicJava()) { 3716 // Check for null receiver. In such case, the optimizer has 3717 // detected that the virtual call will always result in a null 3718 // pointer exception. The fall-through projection of this CatchNode 3719 // will not be populated. 3720 Node *arg0 = call->in(TypeFunc::Parms); 3721 if (arg0->is_Type() && 3722 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3723 required_outcnt--; 3724 } 3725 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3726 call->req() > TypeFunc::Parms+1 && 3727 call->is_CallStaticJava()) { 3728 // Check for negative array length. In such case, the optimizer has 3729 // detected that the allocation attempt will always result in an 3730 // exception. There is no fall-through projection of this CatchNode . 3731 Node *arg1 = call->in(TypeFunc::Parms+1); 3732 if (arg1->is_Type() && 3733 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3734 required_outcnt--; 3735 } 3736 } 3737 } 3738 } 3739 // Recheck with a better notion of 'required_outcnt' 3740 if (n->outcnt() != required_outcnt) { 3741 record_method_not_compilable("malformed control flow"); 3742 return true; // Not all targets reachable! 3743 } 3744 } 3745 // Check that I actually visited all kids. Unreached kids 3746 // must be infinite loops. 3747 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3748 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3749 record_method_not_compilable("infinite loop"); 3750 return true; // Found unvisited kid; must be unreach 3751 } 3752 3753 // Here so verification code in final_graph_reshaping_walk() 3754 // always see an OuterStripMinedLoopEnd 3755 if (n->is_OuterStripMinedLoopEnd()) { 3756 IfNode* init_iff = n->as_If(); 3757 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3758 n->subsume_by(iff, this); 3759 } 3760 } 3761 3762 #ifdef IA32 3763 // If original bytecodes contained a mixture of floats and doubles 3764 // check if the optimizer has made it homogenous, item (3). 3765 if (UseSSE == 0 && 3766 frc.get_float_count() > 32 && 3767 frc.get_double_count() == 0 && 3768 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3769 set_24_bit_selection_and_mode(false, true); 3770 } 3771 #endif // IA32 3772 3773 set_java_calls(frc.get_java_call_count()); 3774 set_inner_loops(frc.get_inner_loop_count()); 3775 3776 // No infinite loops, no reason to bail out. 3777 return false; 3778 } 3779 3780 //-----------------------------too_many_traps---------------------------------- 3781 // Report if there are too many traps at the current method and bci. 3782 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3783 bool Compile::too_many_traps(ciMethod* method, 3784 int bci, 3785 Deoptimization::DeoptReason reason) { 3786 ciMethodData* md = method->method_data(); 3787 if (md->is_empty()) { 3788 // Assume the trap has not occurred, or that it occurred only 3789 // because of a transient condition during start-up in the interpreter. 3790 return false; 3791 } 3792 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3793 if (md->has_trap_at(bci, m, reason) != 0) { 3794 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3795 // Also, if there are multiple reasons, or if there is no per-BCI record, 3796 // assume the worst. 3797 if (log()) 3798 log()->elem("observe trap='%s' count='%d'", 3799 Deoptimization::trap_reason_name(reason), 3800 md->trap_count(reason)); 3801 return true; 3802 } else { 3803 // Ignore method/bci and see if there have been too many globally. 3804 return too_many_traps(reason, md); 3805 } 3806 } 3807 3808 // Less-accurate variant which does not require a method and bci. 3809 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3810 ciMethodData* logmd) { 3811 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3812 // Too many traps globally. 3813 // Note that we use cumulative trap_count, not just md->trap_count. 3814 if (log()) { 3815 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3816 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3817 Deoptimization::trap_reason_name(reason), 3818 mcount, trap_count(reason)); 3819 } 3820 return true; 3821 } else { 3822 // The coast is clear. 3823 return false; 3824 } 3825 } 3826 3827 //--------------------------too_many_recompiles-------------------------------- 3828 // Report if there are too many recompiles at the current method and bci. 3829 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3830 // Is not eager to return true, since this will cause the compiler to use 3831 // Action_none for a trap point, to avoid too many recompilations. 3832 bool Compile::too_many_recompiles(ciMethod* method, 3833 int bci, 3834 Deoptimization::DeoptReason reason) { 3835 ciMethodData* md = method->method_data(); 3836 if (md->is_empty()) { 3837 // Assume the trap has not occurred, or that it occurred only 3838 // because of a transient condition during start-up in the interpreter. 3839 return false; 3840 } 3841 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3842 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3843 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3844 Deoptimization::DeoptReason per_bc_reason 3845 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3846 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3847 if ((per_bc_reason == Deoptimization::Reason_none 3848 || md->has_trap_at(bci, m, reason) != 0) 3849 // The trap frequency measure we care about is the recompile count: 3850 && md->trap_recompiled_at(bci, m) 3851 && md->overflow_recompile_count() >= bc_cutoff) { 3852 // Do not emit a trap here if it has already caused recompilations. 3853 // Also, if there are multiple reasons, or if there is no per-BCI record, 3854 // assume the worst. 3855 if (log()) 3856 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3857 Deoptimization::trap_reason_name(reason), 3858 md->trap_count(reason), 3859 md->overflow_recompile_count()); 3860 return true; 3861 } else if (trap_count(reason) != 0 3862 && decompile_count() >= m_cutoff) { 3863 // Too many recompiles globally, and we have seen this sort of trap. 3864 // Use cumulative decompile_count, not just md->decompile_count. 3865 if (log()) 3866 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3867 Deoptimization::trap_reason_name(reason), 3868 md->trap_count(reason), trap_count(reason), 3869 md->decompile_count(), decompile_count()); 3870 return true; 3871 } else { 3872 // The coast is clear. 3873 return false; 3874 } 3875 } 3876 3877 // Compute when not to trap. Used by matching trap based nodes and 3878 // NullCheck optimization. 3879 void Compile::set_allowed_deopt_reasons() { 3880 _allowed_reasons = 0; 3881 if (is_method_compilation()) { 3882 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3883 assert(rs < BitsPerInt, "recode bit map"); 3884 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3885 _allowed_reasons |= nth_bit(rs); 3886 } 3887 } 3888 } 3889 } 3890 3891 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 3892 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 3893 } 3894 3895 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 3896 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 3897 } 3898 3899 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 3900 if (holder->is_initialized()) { 3901 return false; 3902 } 3903 if (holder->is_being_initialized()) { 3904 if (accessing_method->holder() == holder) { 3905 // Access inside a class. The barrier can be elided when access happens in <clinit>, 3906 // <init>, or a static method. In all those cases, there was an initialization 3907 // barrier on the holder klass passed. 3908 if (accessing_method->is_static_initializer() || 3909 accessing_method->is_object_initializer() || 3910 accessing_method->is_static()) { 3911 return false; 3912 } 3913 } else if (accessing_method->holder()->is_subclass_of(holder)) { 3914 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 3915 // In case of <init> or a static method, the barrier is on the subclass is not enough: 3916 // child class can become fully initialized while its parent class is still being initialized. 3917 if (accessing_method->is_static_initializer()) { 3918 return false; 3919 } 3920 } 3921 ciMethod* root = method(); // the root method of compilation 3922 if (root != accessing_method) { 3923 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 3924 } 3925 } 3926 return true; 3927 } 3928 3929 #ifndef PRODUCT 3930 //------------------------------verify_graph_edges--------------------------- 3931 // Walk the Graph and verify that there is a one-to-one correspondence 3932 // between Use-Def edges and Def-Use edges in the graph. 3933 void Compile::verify_graph_edges(bool no_dead_code) { 3934 if (VerifyGraphEdges) { 3935 Unique_Node_List visited; 3936 // Call recursive graph walk to check edges 3937 _root->verify_edges(visited); 3938 if (no_dead_code) { 3939 // Now make sure that no visited node is used by an unvisited node. 3940 bool dead_nodes = false; 3941 Unique_Node_List checked; 3942 while (visited.size() > 0) { 3943 Node* n = visited.pop(); 3944 checked.push(n); 3945 for (uint i = 0; i < n->outcnt(); i++) { 3946 Node* use = n->raw_out(i); 3947 if (checked.member(use)) continue; // already checked 3948 if (visited.member(use)) continue; // already in the graph 3949 if (use->is_Con()) continue; // a dead ConNode is OK 3950 // At this point, we have found a dead node which is DU-reachable. 3951 if (!dead_nodes) { 3952 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3953 dead_nodes = true; 3954 } 3955 use->dump(2); 3956 tty->print_cr("---"); 3957 checked.push(use); // No repeats; pretend it is now checked. 3958 } 3959 } 3960 assert(!dead_nodes, "using nodes must be reachable from root"); 3961 } 3962 } 3963 } 3964 #endif 3965 3966 // The Compile object keeps track of failure reasons separately from the ciEnv. 3967 // This is required because there is not quite a 1-1 relation between the 3968 // ciEnv and its compilation task and the Compile object. Note that one 3969 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3970 // to backtrack and retry without subsuming loads. Other than this backtracking 3971 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3972 // by the logic in C2Compiler. 3973 void Compile::record_failure(const char* reason) { 3974 if (log() != NULL) { 3975 log()->elem("failure reason='%s' phase='compile'", reason); 3976 } 3977 if (_failure_reason == NULL) { 3978 // Record the first failure reason. 3979 _failure_reason = reason; 3980 } 3981 3982 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3983 C->print_method(PHASE_FAILURE); 3984 } 3985 _root = NULL; // flush the graph, too 3986 } 3987 3988 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3989 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3990 _phase_name(name), _dolog(CITimeVerbose) 3991 { 3992 if (_dolog) { 3993 C = Compile::current(); 3994 _log = C->log(); 3995 } else { 3996 C = NULL; 3997 _log = NULL; 3998 } 3999 if (_log != NULL) { 4000 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4001 _log->stamp(); 4002 _log->end_head(); 4003 } 4004 } 4005 4006 Compile::TracePhase::~TracePhase() { 4007 4008 C = Compile::current(); 4009 if (_dolog) { 4010 _log = C->log(); 4011 } else { 4012 _log = NULL; 4013 } 4014 4015 #ifdef ASSERT 4016 if (PrintIdealNodeCount) { 4017 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4018 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4019 } 4020 4021 if (VerifyIdealNodeCount) { 4022 Compile::current()->print_missing_nodes(); 4023 } 4024 #endif 4025 4026 if (_log != NULL) { 4027 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4028 } 4029 } 4030 4031 //----------------------------static_subtype_check----------------------------- 4032 // Shortcut important common cases when superklass is exact: 4033 // (0) superklass is java.lang.Object (can occur in reflective code) 4034 // (1) subklass is already limited to a subtype of superklass => always ok 4035 // (2) subklass does not overlap with superklass => always fail 4036 // (3) superklass has NO subtypes and we can check with a simple compare. 4037 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4038 if (StressReflectiveCode) { 4039 return SSC_full_test; // Let caller generate the general case. 4040 } 4041 4042 if (superk == env()->Object_klass()) { 4043 return SSC_always_true; // (0) this test cannot fail 4044 } 4045 4046 ciType* superelem = superk; 4047 if (superelem->is_array_klass()) 4048 superelem = superelem->as_array_klass()->base_element_type(); 4049 4050 if (!subk->is_interface()) { // cannot trust static interface types yet 4051 if (subk->is_subtype_of(superk)) { 4052 return SSC_always_true; // (1) false path dead; no dynamic test needed 4053 } 4054 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4055 !superk->is_subtype_of(subk)) { 4056 return SSC_always_false; 4057 } 4058 } 4059 4060 // If casting to an instance klass, it must have no subtypes 4061 if (superk->is_interface()) { 4062 // Cannot trust interfaces yet. 4063 // %%% S.B. superk->nof_implementors() == 1 4064 } else if (superelem->is_instance_klass()) { 4065 ciInstanceKlass* ik = superelem->as_instance_klass(); 4066 if (!ik->has_subklass() && !ik->is_interface()) { 4067 if (!ik->is_final()) { 4068 // Add a dependency if there is a chance of a later subclass. 4069 dependencies()->assert_leaf_type(ik); 4070 } 4071 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4072 } 4073 } else { 4074 // A primitive array type has no subtypes. 4075 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4076 } 4077 4078 return SSC_full_test; 4079 } 4080 4081 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4082 #ifdef _LP64 4083 // The scaled index operand to AddP must be a clean 64-bit value. 4084 // Java allows a 32-bit int to be incremented to a negative 4085 // value, which appears in a 64-bit register as a large 4086 // positive number. Using that large positive number as an 4087 // operand in pointer arithmetic has bad consequences. 4088 // On the other hand, 32-bit overflow is rare, and the possibility 4089 // can often be excluded, if we annotate the ConvI2L node with 4090 // a type assertion that its value is known to be a small positive 4091 // number. (The prior range check has ensured this.) 4092 // This assertion is used by ConvI2LNode::Ideal. 4093 int index_max = max_jint - 1; // array size is max_jint, index is one less 4094 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4095 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4096 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4097 #endif 4098 return idx; 4099 } 4100 4101 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4102 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4103 if (ctrl != NULL) { 4104 // Express control dependency by a CastII node with a narrow type. 4105 value = new CastIINode(value, itype, false, true /* range check dependency */); 4106 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4107 // node from floating above the range check during loop optimizations. Otherwise, the 4108 // ConvI2L node may be eliminated independently of the range check, causing the data path 4109 // to become TOP while the control path is still there (although it's unreachable). 4110 value->set_req(0, ctrl); 4111 // Save CastII node to remove it after loop optimizations. 4112 phase->C->add_range_check_cast(value); 4113 value = phase->transform(value); 4114 } 4115 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4116 return phase->transform(new ConvI2LNode(value, ltype)); 4117 } 4118 4119 void Compile::print_inlining_stream_free() { 4120 if (_print_inlining_stream != NULL) { 4121 _print_inlining_stream->~stringStream(); 4122 _print_inlining_stream = NULL; 4123 } 4124 } 4125 4126 // The message about the current inlining is accumulated in 4127 // _print_inlining_stream and transfered into the _print_inlining_list 4128 // once we know whether inlining succeeds or not. For regular 4129 // inlining, messages are appended to the buffer pointed by 4130 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4131 // a new buffer is added after _print_inlining_idx in the list. This 4132 // way we can update the inlining message for late inlining call site 4133 // when the inlining is attempted again. 4134 void Compile::print_inlining_init() { 4135 if (print_inlining() || print_intrinsics()) { 4136 // print_inlining_init is actually called several times. 4137 print_inlining_stream_free(); 4138 _print_inlining_stream = new stringStream(); 4139 // Watch out: The memory initialized by the constructor call PrintInliningBuffer() 4140 // will be copied into the only initial element. The default destructor of 4141 // PrintInliningBuffer will be called when leaving the scope here. If it 4142 // would destuct the enclosed stringStream _print_inlining_list[0]->_ss 4143 // would be destructed, too! 4144 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4145 } 4146 } 4147 4148 void Compile::print_inlining_reinit() { 4149 if (print_inlining() || print_intrinsics()) { 4150 print_inlining_stream_free(); 4151 // Re allocate buffer when we change ResourceMark 4152 _print_inlining_stream = new stringStream(); 4153 } 4154 } 4155 4156 void Compile::print_inlining_reset() { 4157 _print_inlining_stream->reset(); 4158 } 4159 4160 void Compile::print_inlining_commit() { 4161 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4162 // Transfer the message from _print_inlining_stream to the current 4163 // _print_inlining_list buffer and clear _print_inlining_stream. 4164 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4165 print_inlining_reset(); 4166 } 4167 4168 void Compile::print_inlining_push() { 4169 // Add new buffer to the _print_inlining_list at current position 4170 _print_inlining_idx++; 4171 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4172 } 4173 4174 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4175 return _print_inlining_list->at(_print_inlining_idx); 4176 } 4177 4178 void Compile::print_inlining_update(CallGenerator* cg) { 4179 if (print_inlining() || print_intrinsics()) { 4180 if (!cg->is_late_inline()) { 4181 if (print_inlining_current().cg() != NULL) { 4182 print_inlining_push(); 4183 } 4184 print_inlining_commit(); 4185 } else { 4186 if (print_inlining_current().cg() != cg && 4187 (print_inlining_current().cg() != NULL || 4188 print_inlining_current().ss()->size() != 0)) { 4189 print_inlining_push(); 4190 } 4191 print_inlining_commit(); 4192 print_inlining_current().set_cg(cg); 4193 } 4194 } 4195 } 4196 4197 void Compile::print_inlining_move_to(CallGenerator* cg) { 4198 // We resume inlining at a late inlining call site. Locate the 4199 // corresponding inlining buffer so that we can update it. 4200 if (print_inlining()) { 4201 for (int i = 0; i < _print_inlining_list->length(); i++) { 4202 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4203 _print_inlining_idx = i; 4204 return; 4205 } 4206 } 4207 ShouldNotReachHere(); 4208 } 4209 } 4210 4211 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4212 if (print_inlining()) { 4213 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4214 assert(print_inlining_current().cg() == cg, "wrong entry"); 4215 // replace message with new message 4216 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4217 print_inlining_commit(); 4218 print_inlining_current().set_cg(cg); 4219 } 4220 } 4221 4222 void Compile::print_inlining_assert_ready() { 4223 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4224 } 4225 4226 void Compile::process_print_inlining() { 4227 bool do_print_inlining = print_inlining() || print_intrinsics(); 4228 if (do_print_inlining || log() != NULL) { 4229 // Print inlining message for candidates that we couldn't inline 4230 // for lack of space 4231 for (int i = 0; i < _late_inlines.length(); i++) { 4232 CallGenerator* cg = _late_inlines.at(i); 4233 if (!cg->is_mh_late_inline()) { 4234 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4235 if (do_print_inlining) { 4236 cg->print_inlining_late(msg); 4237 } 4238 log_late_inline_failure(cg, msg); 4239 } 4240 } 4241 } 4242 if (do_print_inlining) { 4243 ResourceMark rm; 4244 stringStream ss; 4245 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4246 for (int i = 0; i < _print_inlining_list->length(); i++) { 4247 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4248 _print_inlining_list->at(i).freeStream(); 4249 } 4250 // Reset _print_inlining_list, it only contains destructed objects. 4251 // It is on the arena, so it will be freed when the arena is reset. 4252 _print_inlining_list = NULL; 4253 // _print_inlining_stream won't be used anymore, either. 4254 print_inlining_stream_free(); 4255 size_t end = ss.size(); 4256 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4257 strncpy(_print_inlining_output, ss.base(), end+1); 4258 _print_inlining_output[end] = 0; 4259 } 4260 } 4261 4262 void Compile::dump_print_inlining() { 4263 if (_print_inlining_output != NULL) { 4264 tty->print_raw(_print_inlining_output); 4265 } 4266 } 4267 4268 void Compile::log_late_inline(CallGenerator* cg) { 4269 if (log() != NULL) { 4270 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4271 cg->unique_id()); 4272 JVMState* p = cg->call_node()->jvms(); 4273 while (p != NULL) { 4274 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4275 p = p->caller(); 4276 } 4277 log()->tail("late_inline"); 4278 } 4279 } 4280 4281 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4282 log_late_inline(cg); 4283 if (log() != NULL) { 4284 log()->inline_fail(msg); 4285 } 4286 } 4287 4288 void Compile::log_inline_id(CallGenerator* cg) { 4289 if (log() != NULL) { 4290 // The LogCompilation tool needs a unique way to identify late 4291 // inline call sites. This id must be unique for this call site in 4292 // this compilation. Try to have it unique across compilations as 4293 // well because it can be convenient when grepping through the log 4294 // file. 4295 // Distinguish OSR compilations from others in case CICountOSR is 4296 // on. 4297 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4298 cg->set_unique_id(id); 4299 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4300 } 4301 } 4302 4303 void Compile::log_inline_failure(const char* msg) { 4304 if (C->log() != NULL) { 4305 C->log()->inline_fail(msg); 4306 } 4307 } 4308 4309 4310 // Dump inlining replay data to the stream. 4311 // Don't change thread state and acquire any locks. 4312 void Compile::dump_inline_data(outputStream* out) { 4313 InlineTree* inl_tree = ilt(); 4314 if (inl_tree != NULL) { 4315 out->print(" inline %d", inl_tree->count()); 4316 inl_tree->dump_replay_data(out); 4317 } 4318 } 4319 4320 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4321 if (n1->Opcode() < n2->Opcode()) return -1; 4322 else if (n1->Opcode() > n2->Opcode()) return 1; 4323 4324 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4325 for (uint i = 1; i < n1->req(); i++) { 4326 if (n1->in(i) < n2->in(i)) return -1; 4327 else if (n1->in(i) > n2->in(i)) return 1; 4328 } 4329 4330 return 0; 4331 } 4332 4333 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4334 Node* n1 = *n1p; 4335 Node* n2 = *n2p; 4336 4337 return cmp_expensive_nodes(n1, n2); 4338 } 4339 4340 void Compile::sort_expensive_nodes() { 4341 if (!expensive_nodes_sorted()) { 4342 _expensive_nodes->sort(cmp_expensive_nodes); 4343 } 4344 } 4345 4346 bool Compile::expensive_nodes_sorted() const { 4347 for (int i = 1; i < _expensive_nodes->length(); i++) { 4348 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4349 return false; 4350 } 4351 } 4352 return true; 4353 } 4354 4355 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4356 if (_expensive_nodes->length() == 0) { 4357 return false; 4358 } 4359 4360 assert(OptimizeExpensiveOps, "optimization off?"); 4361 4362 // Take this opportunity to remove dead nodes from the list 4363 int j = 0; 4364 for (int i = 0; i < _expensive_nodes->length(); i++) { 4365 Node* n = _expensive_nodes->at(i); 4366 if (!n->is_unreachable(igvn)) { 4367 assert(n->is_expensive(), "should be expensive"); 4368 _expensive_nodes->at_put(j, n); 4369 j++; 4370 } 4371 } 4372 _expensive_nodes->trunc_to(j); 4373 4374 // Then sort the list so that similar nodes are next to each other 4375 // and check for at least two nodes of identical kind with same data 4376 // inputs. 4377 sort_expensive_nodes(); 4378 4379 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4380 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4381 return true; 4382 } 4383 } 4384 4385 return false; 4386 } 4387 4388 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4389 if (_expensive_nodes->length() == 0) { 4390 return; 4391 } 4392 4393 assert(OptimizeExpensiveOps, "optimization off?"); 4394 4395 // Sort to bring similar nodes next to each other and clear the 4396 // control input of nodes for which there's only a single copy. 4397 sort_expensive_nodes(); 4398 4399 int j = 0; 4400 int identical = 0; 4401 int i = 0; 4402 bool modified = false; 4403 for (; i < _expensive_nodes->length()-1; i++) { 4404 assert(j <= i, "can't write beyond current index"); 4405 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4406 identical++; 4407 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4408 continue; 4409 } 4410 if (identical > 0) { 4411 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4412 identical = 0; 4413 } else { 4414 Node* n = _expensive_nodes->at(i); 4415 igvn.replace_input_of(n, 0, NULL); 4416 igvn.hash_insert(n); 4417 modified = true; 4418 } 4419 } 4420 if (identical > 0) { 4421 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4422 } else if (_expensive_nodes->length() >= 1) { 4423 Node* n = _expensive_nodes->at(i); 4424 igvn.replace_input_of(n, 0, NULL); 4425 igvn.hash_insert(n); 4426 modified = true; 4427 } 4428 _expensive_nodes->trunc_to(j); 4429 if (modified) { 4430 igvn.optimize(); 4431 } 4432 } 4433 4434 void Compile::add_expensive_node(Node * n) { 4435 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4436 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4437 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4438 if (OptimizeExpensiveOps) { 4439 _expensive_nodes->append(n); 4440 } else { 4441 // Clear control input and let IGVN optimize expensive nodes if 4442 // OptimizeExpensiveOps is off. 4443 n->set_req(0, NULL); 4444 } 4445 } 4446 4447 /** 4448 * Remove the speculative part of types and clean up the graph 4449 */ 4450 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4451 if (UseTypeSpeculation) { 4452 Unique_Node_List worklist; 4453 worklist.push(root()); 4454 int modified = 0; 4455 // Go over all type nodes that carry a speculative type, drop the 4456 // speculative part of the type and enqueue the node for an igvn 4457 // which may optimize it out. 4458 for (uint next = 0; next < worklist.size(); ++next) { 4459 Node *n = worklist.at(next); 4460 if (n->is_Type()) { 4461 TypeNode* tn = n->as_Type(); 4462 const Type* t = tn->type(); 4463 const Type* t_no_spec = t->remove_speculative(); 4464 if (t_no_spec != t) { 4465 bool in_hash = igvn.hash_delete(n); 4466 assert(in_hash, "node should be in igvn hash table"); 4467 tn->set_type(t_no_spec); 4468 igvn.hash_insert(n); 4469 igvn._worklist.push(n); // give it a chance to go away 4470 modified++; 4471 } 4472 } 4473 uint max = n->len(); 4474 for( uint i = 0; i < max; ++i ) { 4475 Node *m = n->in(i); 4476 if (not_a_node(m)) continue; 4477 worklist.push(m); 4478 } 4479 } 4480 // Drop the speculative part of all types in the igvn's type table 4481 igvn.remove_speculative_types(); 4482 if (modified > 0) { 4483 igvn.optimize(); 4484 } 4485 #ifdef ASSERT 4486 // Verify that after the IGVN is over no speculative type has resurfaced 4487 worklist.clear(); 4488 worklist.push(root()); 4489 for (uint next = 0; next < worklist.size(); ++next) { 4490 Node *n = worklist.at(next); 4491 const Type* t = igvn.type_or_null(n); 4492 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4493 if (n->is_Type()) { 4494 t = n->as_Type()->type(); 4495 assert(t == t->remove_speculative(), "no more speculative types"); 4496 } 4497 uint max = n->len(); 4498 for( uint i = 0; i < max; ++i ) { 4499 Node *m = n->in(i); 4500 if (not_a_node(m)) continue; 4501 worklist.push(m); 4502 } 4503 } 4504 igvn.check_no_speculative_types(); 4505 #endif 4506 } 4507 } 4508 4509 // Auxiliary method to support randomized stressing/fuzzing. 4510 // 4511 // This method can be called the arbitrary number of times, with current count 4512 // as the argument. The logic allows selecting a single candidate from the 4513 // running list of candidates as follows: 4514 // int count = 0; 4515 // Cand* selected = null; 4516 // while(cand = cand->next()) { 4517 // if (randomized_select(++count)) { 4518 // selected = cand; 4519 // } 4520 // } 4521 // 4522 // Including count equalizes the chances any candidate is "selected". 4523 // This is useful when we don't have the complete list of candidates to choose 4524 // from uniformly. In this case, we need to adjust the randomicity of the 4525 // selection, or else we will end up biasing the selection towards the latter 4526 // candidates. 4527 // 4528 // Quick back-envelope calculation shows that for the list of n candidates 4529 // the equal probability for the candidate to persist as "best" can be 4530 // achieved by replacing it with "next" k-th candidate with the probability 4531 // of 1/k. It can be easily shown that by the end of the run, the 4532 // probability for any candidate is converged to 1/n, thus giving the 4533 // uniform distribution among all the candidates. 4534 // 4535 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4536 #define RANDOMIZED_DOMAIN_POW 29 4537 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4538 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4539 bool Compile::randomized_select(int count) { 4540 assert(count > 0, "only positive"); 4541 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4542 } 4543 4544 CloneMap& Compile::clone_map() { return _clone_map; } 4545 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4546 4547 void NodeCloneInfo::dump() const { 4548 tty->print(" {%d:%d} ", idx(), gen()); 4549 } 4550 4551 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4552 uint64_t val = value(old->_idx); 4553 NodeCloneInfo cio(val); 4554 assert(val != 0, "old node should be in the map"); 4555 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4556 insert(nnn->_idx, cin.get()); 4557 #ifndef PRODUCT 4558 if (is_debug()) { 4559 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4560 } 4561 #endif 4562 } 4563 4564 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4565 NodeCloneInfo cio(value(old->_idx)); 4566 if (cio.get() == 0) { 4567 cio.set(old->_idx, 0); 4568 insert(old->_idx, cio.get()); 4569 #ifndef PRODUCT 4570 if (is_debug()) { 4571 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4572 } 4573 #endif 4574 } 4575 clone(old, nnn, gen); 4576 } 4577 4578 int CloneMap::max_gen() const { 4579 int g = 0; 4580 DictI di(_dict); 4581 for(; di.test(); ++di) { 4582 int t = gen(di._key); 4583 if (g < t) { 4584 g = t; 4585 #ifndef PRODUCT 4586 if (is_debug()) { 4587 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4588 } 4589 #endif 4590 } 4591 } 4592 return g; 4593 } 4594 4595 void CloneMap::dump(node_idx_t key) const { 4596 uint64_t val = value(key); 4597 if (val != 0) { 4598 NodeCloneInfo ni(val); 4599 ni.dump(); 4600 } 4601 } 4602 4603 // Move Allocate nodes to the start of the list 4604 void Compile::sort_macro_nodes() { 4605 int count = macro_count(); 4606 int allocates = 0; 4607 for (int i = 0; i < count; i++) { 4608 Node* n = macro_node(i); 4609 if (n->is_Allocate()) { 4610 if (i != allocates) { 4611 Node* tmp = macro_node(allocates); 4612 _macro_nodes->at_put(allocates, n); 4613 _macro_nodes->at_put(i, tmp); 4614 } 4615 allocates++; 4616 } 4617 } 4618 } 4619 4620 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level, int idx) { 4621 EventCompilerPhase event; 4622 if (event.should_commit()) { 4623 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 4624 } 4625 #ifndef PRODUCT 4626 if (should_print(level)) { 4627 _printer->print_method(name, level); 4628 } 4629 #endif 4630 C->_latest_stage_start_counter.stamp(); 4631 } 4632 4633 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) { 4634 char output[1024]; 4635 #ifndef PRODUCT 4636 if (idx != 0) { 4637 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx); 4638 } else { 4639 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt)); 4640 } 4641 #endif 4642 print_method(cpt, output, level, idx); 4643 } 4644 4645 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) { 4646 ResourceMark rm; 4647 stringStream ss; 4648 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt)); 4649 if (n != NULL) { 4650 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 4651 } else { 4652 ss.print_raw(": NULL"); 4653 } 4654 C->print_method(cpt, ss.as_string(), level); 4655 } 4656 4657 void Compile::end_method(int level) { 4658 EventCompilerPhase event; 4659 if (event.should_commit()) { 4660 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level); 4661 } 4662 4663 #ifndef PRODUCT 4664 if (_method != NULL && should_print(level)) { 4665 _printer->end_method(); 4666 } 4667 #endif 4668 } 4669 4670 4671 #ifndef PRODUCT 4672 IdealGraphPrinter* Compile::_debug_file_printer = NULL; 4673 IdealGraphPrinter* Compile::_debug_network_printer = NULL; 4674 4675 // Called from debugger. Prints method to the default file with the default phase name. 4676 // This works regardless of any Ideal Graph Visualizer flags set or not. 4677 void igv_print() { 4678 Compile::current()->igv_print_method_to_file(); 4679 } 4680 4681 // Same as igv_print() above but with a specified phase name. 4682 void igv_print(const char* phase_name) { 4683 Compile::current()->igv_print_method_to_file(phase_name); 4684 } 4685 4686 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 4687 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 4688 // This works regardless of any Ideal Graph Visualizer flags set or not. 4689 void igv_print(bool network) { 4690 if (network) { 4691 Compile::current()->igv_print_method_to_network(); 4692 } else { 4693 Compile::current()->igv_print_method_to_file(); 4694 } 4695 } 4696 4697 // Same as igv_print(bool network) above but with a specified phase name. 4698 void igv_print(bool network, const char* phase_name) { 4699 if (network) { 4700 Compile::current()->igv_print_method_to_network(phase_name); 4701 } else { 4702 Compile::current()->igv_print_method_to_file(phase_name); 4703 } 4704 } 4705 4706 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 4707 void igv_print_default() { 4708 Compile::current()->print_method(PHASE_DEBUG, 0, 0); 4709 } 4710 4711 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 4712 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 4713 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 4714 void igv_append() { 4715 Compile::current()->igv_print_method_to_file("Debug", true); 4716 } 4717 4718 // Same as igv_append() above but with a specified phase name. 4719 void igv_append(const char* phase_name) { 4720 Compile::current()->igv_print_method_to_file(phase_name, true); 4721 } 4722 4723 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 4724 const char* file_name = "custom_debug.xml"; 4725 if (_debug_file_printer == NULL) { 4726 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 4727 } else { 4728 _debug_file_printer->update_compiled_method(C->method()); 4729 } 4730 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 4731 _debug_file_printer->print_method(phase_name, 0); 4732 } 4733 4734 void Compile::igv_print_method_to_network(const char* phase_name) { 4735 if (_debug_network_printer == NULL) { 4736 _debug_network_printer = new IdealGraphPrinter(C); 4737 } else { 4738 _debug_network_printer->update_compiled_method(C->method()); 4739 } 4740 tty->print_cr("Method printed over network stream to IGV"); 4741 _debug_network_printer->print_method(phase_name, 0); 4742 } 4743 #endif 4744