1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_compile.cpp.incl" 27 28 /// Support for intrinsics. 29 30 // Return the index at which m must be inserted (or already exists). 31 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 33 #ifdef ASSERT 34 for (int i = 1; i < _intrinsics->length(); i++) { 35 CallGenerator* cg1 = _intrinsics->at(i-1); 36 CallGenerator* cg2 = _intrinsics->at(i); 37 assert(cg1->method() != cg2->method() 38 ? cg1->method() < cg2->method() 39 : cg1->is_virtual() < cg2->is_virtual(), 40 "compiler intrinsics list must stay sorted"); 41 } 42 #endif 43 // Binary search sorted list, in decreasing intervals [lo, hi]. 44 int lo = 0, hi = _intrinsics->length()-1; 45 while (lo <= hi) { 46 int mid = (uint)(hi + lo) / 2; 47 ciMethod* mid_m = _intrinsics->at(mid)->method(); 48 if (m < mid_m) { 49 hi = mid-1; 50 } else if (m > mid_m) { 51 lo = mid+1; 52 } else { 53 // look at minor sort key 54 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 55 if (is_virtual < mid_virt) { 56 hi = mid-1; 57 } else if (is_virtual > mid_virt) { 58 lo = mid+1; 59 } else { 60 return mid; // exact match 61 } 62 } 63 } 64 return lo; // inexact match 65 } 66 67 void Compile::register_intrinsic(CallGenerator* cg) { 68 if (_intrinsics == NULL) { 69 _intrinsics = new GrowableArray<CallGenerator*>(60); 70 } 71 // This code is stolen from ciObjectFactory::insert. 72 // Really, GrowableArray should have methods for 73 // insert_at, remove_at, and binary_search. 74 int len = _intrinsics->length(); 75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 76 if (index == len) { 77 _intrinsics->append(cg); 78 } else { 79 #ifdef ASSERT 80 CallGenerator* oldcg = _intrinsics->at(index); 81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 82 #endif 83 _intrinsics->append(_intrinsics->at(len-1)); 84 int pos; 85 for (pos = len-2; pos >= index; pos--) { 86 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 87 } 88 _intrinsics->at_put(index, cg); 89 } 90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 91 } 92 93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 94 assert(m->is_loaded(), "don't try this on unloaded methods"); 95 if (_intrinsics != NULL) { 96 int index = intrinsic_insertion_index(m, is_virtual); 97 if (index < _intrinsics->length() 98 && _intrinsics->at(index)->method() == m 99 && _intrinsics->at(index)->is_virtual() == is_virtual) { 100 return _intrinsics->at(index); 101 } 102 } 103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 104 if (m->intrinsic_id() != vmIntrinsics::_none) { 105 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 106 if (cg != NULL) { 107 // Save it for next time: 108 register_intrinsic(cg); 109 return cg; 110 } else { 111 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 112 } 113 } 114 return NULL; 115 } 116 117 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 118 // in library_call.cpp. 119 120 121 #ifndef PRODUCT 122 // statistics gathering... 123 124 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 125 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 126 127 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 128 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 129 int oflags = _intrinsic_hist_flags[id]; 130 assert(flags != 0, "what happened?"); 131 if (is_virtual) { 132 flags |= _intrinsic_virtual; 133 } 134 bool changed = (flags != oflags); 135 if ((flags & _intrinsic_worked) != 0) { 136 juint count = (_intrinsic_hist_count[id] += 1); 137 if (count == 1) { 138 changed = true; // first time 139 } 140 // increment the overall count also: 141 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 142 } 143 if (changed) { 144 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 145 // Something changed about the intrinsic's virtuality. 146 if ((flags & _intrinsic_virtual) != 0) { 147 // This is the first use of this intrinsic as a virtual call. 148 if (oflags != 0) { 149 // We already saw it as a non-virtual, so note both cases. 150 flags |= _intrinsic_both; 151 } 152 } else if ((oflags & _intrinsic_both) == 0) { 153 // This is the first use of this intrinsic as a non-virtual 154 flags |= _intrinsic_both; 155 } 156 } 157 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 158 } 159 // update the overall flags also: 160 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 161 return changed; 162 } 163 164 static char* format_flags(int flags, char* buf) { 165 buf[0] = 0; 166 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 167 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 168 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 169 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 170 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 171 if (buf[0] == 0) strcat(buf, ","); 172 assert(buf[0] == ',', "must be"); 173 return &buf[1]; 174 } 175 176 void Compile::print_intrinsic_statistics() { 177 char flagsbuf[100]; 178 ttyLocker ttyl; 179 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 180 tty->print_cr("Compiler intrinsic usage:"); 181 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 182 if (total == 0) total = 1; // avoid div0 in case of no successes 183 #define PRINT_STAT_LINE(name, c, f) \ 184 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 185 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 186 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 187 int flags = _intrinsic_hist_flags[id]; 188 juint count = _intrinsic_hist_count[id]; 189 if ((flags | count) != 0) { 190 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 191 } 192 } 193 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 194 if (xtty != NULL) xtty->tail("statistics"); 195 } 196 197 void Compile::print_statistics() { 198 { ttyLocker ttyl; 199 if (xtty != NULL) xtty->head("statistics type='opto'"); 200 Parse::print_statistics(); 201 PhaseCCP::print_statistics(); 202 PhaseRegAlloc::print_statistics(); 203 Scheduling::print_statistics(); 204 PhasePeephole::print_statistics(); 205 PhaseIdealLoop::print_statistics(); 206 if (xtty != NULL) xtty->tail("statistics"); 207 } 208 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 209 // put this under its own <statistics> element. 210 print_intrinsic_statistics(); 211 } 212 } 213 #endif //PRODUCT 214 215 // Support for bundling info 216 Bundle* Compile::node_bundling(const Node *n) { 217 assert(valid_bundle_info(n), "oob"); 218 return &_node_bundling_base[n->_idx]; 219 } 220 221 bool Compile::valid_bundle_info(const Node *n) { 222 return (_node_bundling_limit > n->_idx); 223 } 224 225 226 // Identify all nodes that are reachable from below, useful. 227 // Use breadth-first pass that records state in a Unique_Node_List, 228 // recursive traversal is slower. 229 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 230 int estimated_worklist_size = unique(); 231 useful.map( estimated_worklist_size, NULL ); // preallocate space 232 233 // Initialize worklist 234 if (root() != NULL) { useful.push(root()); } 235 // If 'top' is cached, declare it useful to preserve cached node 236 if( cached_top_node() ) { useful.push(cached_top_node()); } 237 238 // Push all useful nodes onto the list, breadthfirst 239 for( uint next = 0; next < useful.size(); ++next ) { 240 assert( next < unique(), "Unique useful nodes < total nodes"); 241 Node *n = useful.at(next); 242 uint max = n->len(); 243 for( uint i = 0; i < max; ++i ) { 244 Node *m = n->in(i); 245 if( m == NULL ) continue; 246 useful.push(m); 247 } 248 } 249 } 250 251 // Disconnect all useless nodes by disconnecting those at the boundary. 252 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 253 uint next = 0; 254 while( next < useful.size() ) { 255 Node *n = useful.at(next++); 256 // Use raw traversal of out edges since this code removes out edges 257 int max = n->outcnt(); 258 for (int j = 0; j < max; ++j ) { 259 Node* child = n->raw_out(j); 260 if( ! useful.member(child) ) { 261 assert( !child->is_top() || child != top(), 262 "If top is cached in Compile object it is in useful list"); 263 // Only need to remove this out-edge to the useless node 264 n->raw_del_out(j); 265 --j; 266 --max; 267 } 268 } 269 if (n->outcnt() == 1 && n->has_special_unique_user()) { 270 record_for_igvn( n->unique_out() ); 271 } 272 } 273 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 274 } 275 276 //------------------------------frame_size_in_words----------------------------- 277 // frame_slots in units of words 278 int Compile::frame_size_in_words() const { 279 // shift is 0 in LP32 and 1 in LP64 280 const int shift = (LogBytesPerWord - LogBytesPerInt); 281 int words = _frame_slots >> shift; 282 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 283 return words; 284 } 285 286 // ============================================================================ 287 //------------------------------CompileWrapper--------------------------------- 288 class CompileWrapper : public StackObj { 289 Compile *const _compile; 290 public: 291 CompileWrapper(Compile* compile); 292 293 ~CompileWrapper(); 294 }; 295 296 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 297 // the Compile* pointer is stored in the current ciEnv: 298 ciEnv* env = compile->env(); 299 assert(env == ciEnv::current(), "must already be a ciEnv active"); 300 assert(env->compiler_data() == NULL, "compile already active?"); 301 env->set_compiler_data(compile); 302 assert(compile == Compile::current(), "sanity"); 303 304 compile->set_type_dict(NULL); 305 compile->set_type_hwm(NULL); 306 compile->set_type_last_size(0); 307 compile->set_last_tf(NULL, NULL); 308 compile->set_indexSet_arena(NULL); 309 compile->set_indexSet_free_block_list(NULL); 310 compile->init_type_arena(); 311 Type::Initialize(compile); 312 _compile->set_scratch_buffer_blob(NULL); 313 _compile->begin_method(); 314 } 315 CompileWrapper::~CompileWrapper() { 316 _compile->end_method(); 317 if (_compile->scratch_buffer_blob() != NULL) 318 BufferBlob::free(_compile->scratch_buffer_blob()); 319 _compile->env()->set_compiler_data(NULL); 320 } 321 322 323 //----------------------------print_compile_messages--------------------------- 324 void Compile::print_compile_messages() { 325 #ifndef PRODUCT 326 // Check if recompiling 327 if (_subsume_loads == false && PrintOpto) { 328 // Recompiling without allowing machine instructions to subsume loads 329 tty->print_cr("*********************************************************"); 330 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 331 tty->print_cr("*********************************************************"); 332 } 333 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 334 // Recompiling without escape analysis 335 tty->print_cr("*********************************************************"); 336 tty->print_cr("** Bailout: Recompile without escape analysis **"); 337 tty->print_cr("*********************************************************"); 338 } 339 if (env()->break_at_compile()) { 340 // Open the debugger when compiling this method. 341 tty->print("### Breaking when compiling: "); 342 method()->print_short_name(); 343 tty->cr(); 344 BREAKPOINT; 345 } 346 347 if( PrintOpto ) { 348 if (is_osr_compilation()) { 349 tty->print("[OSR]%3d", _compile_id); 350 } else { 351 tty->print("%3d", _compile_id); 352 } 353 } 354 #endif 355 } 356 357 358 void Compile::init_scratch_buffer_blob() { 359 if( scratch_buffer_blob() != NULL ) return; 360 361 // Construct a temporary CodeBuffer to have it construct a BufferBlob 362 // Cache this BufferBlob for this compile. 363 ResourceMark rm; 364 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 365 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 366 // Record the buffer blob for next time. 367 set_scratch_buffer_blob(blob); 368 // Have we run out of code space? 369 if (scratch_buffer_blob() == NULL) { 370 // Let CompilerBroker disable further compilations. 371 record_failure("Not enough space for scratch buffer in CodeCache"); 372 return; 373 } 374 375 // Initialize the relocation buffers 376 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 377 set_scratch_locs_memory(locs_buf); 378 } 379 380 381 //-----------------------scratch_emit_size------------------------------------- 382 // Helper function that computes size by emitting code 383 uint Compile::scratch_emit_size(const Node* n) { 384 // Emit into a trash buffer and count bytes emitted. 385 // This is a pretty expensive way to compute a size, 386 // but it works well enough if seldom used. 387 // All common fixed-size instructions are given a size 388 // method by the AD file. 389 // Note that the scratch buffer blob and locs memory are 390 // allocated at the beginning of the compile task, and 391 // may be shared by several calls to scratch_emit_size. 392 // The allocation of the scratch buffer blob is particularly 393 // expensive, since it has to grab the code cache lock. 394 BufferBlob* blob = this->scratch_buffer_blob(); 395 assert(blob != NULL, "Initialize BufferBlob at start"); 396 assert(blob->size() > MAX_inst_size, "sanity"); 397 relocInfo* locs_buf = scratch_locs_memory(); 398 address blob_begin = blob->instructions_begin(); 399 address blob_end = (address)locs_buf; 400 assert(blob->instructions_contains(blob_end), "sanity"); 401 CodeBuffer buf(blob_begin, blob_end - blob_begin); 402 buf.initialize_consts_size(MAX_const_size); 403 buf.initialize_stubs_size(MAX_stubs_size); 404 assert(locs_buf != NULL, "sanity"); 405 int lsize = MAX_locs_size / 2; 406 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 407 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 408 n->emit(buf, this->regalloc()); 409 return buf.code_size(); 410 } 411 412 413 // ============================================================================ 414 //------------------------------Compile standard------------------------------- 415 debug_only( int Compile::_debug_idx = 100000; ) 416 417 // Compile a method. entry_bci is -1 for normal compilations and indicates 418 // the continuation bci for on stack replacement. 419 420 421 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 422 : Phase(Compiler), 423 _env(ci_env), 424 _log(ci_env->log()), 425 _compile_id(ci_env->compile_id()), 426 _save_argument_registers(false), 427 _stub_name(NULL), 428 _stub_function(NULL), 429 _stub_entry_point(NULL), 430 _method(target), 431 _entry_bci(osr_bci), 432 _initial_gvn(NULL), 433 _for_igvn(NULL), 434 _warm_calls(NULL), 435 _subsume_loads(subsume_loads), 436 _do_escape_analysis(do_escape_analysis), 437 _failure_reason(NULL), 438 _code_buffer("Compile::Fill_buffer"), 439 _orig_pc_slot(0), 440 _orig_pc_slot_offset_in_bytes(0), 441 _node_bundling_limit(0), 442 _node_bundling_base(NULL), 443 _nodes_limit((uint)MaxNodeLimit), 444 #ifndef PRODUCT 445 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 446 _printer(IdealGraphPrinter::printer()), 447 #endif 448 _congraph(NULL) { 449 C = this; 450 451 CompileWrapper cw(this); 452 #ifndef PRODUCT 453 if (TimeCompiler2) { 454 tty->print(" "); 455 target->holder()->name()->print(); 456 tty->print("."); 457 target->print_short_name(); 458 tty->print(" "); 459 } 460 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 461 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 462 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 463 if (!print_opto_assembly) { 464 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 465 if (print_assembly && !Disassembler::can_decode()) { 466 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 467 print_opto_assembly = true; 468 } 469 } 470 set_print_assembly(print_opto_assembly); 471 set_parsed_irreducible_loop(false); 472 #endif 473 474 if (ProfileTraps) { 475 // Make sure the method being compiled gets its own MDO, 476 // so we can at least track the decompile_count(). 477 method()->build_method_data(); 478 } 479 480 Init(::AliasLevel); 481 482 483 print_compile_messages(); 484 485 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 486 _ilt = InlineTree::build_inline_tree_root(); 487 else 488 _ilt = NULL; 489 490 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 491 assert(num_alias_types() >= AliasIdxRaw, ""); 492 493 #define MINIMUM_NODE_HASH 1023 494 // Node list that Iterative GVN will start with 495 Unique_Node_List for_igvn(comp_arena()); 496 set_for_igvn(&for_igvn); 497 498 // GVN that will be run immediately on new nodes 499 uint estimated_size = method()->code_size()*4+64; 500 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 501 PhaseGVN gvn(node_arena(), estimated_size); 502 set_initial_gvn(&gvn); 503 504 { // Scope for timing the parser 505 TracePhase t3("parse", &_t_parser, true); 506 507 // Put top into the hash table ASAP. 508 initial_gvn()->transform_no_reclaim(top()); 509 510 // Set up tf(), start(), and find a CallGenerator. 511 CallGenerator* cg; 512 if (is_osr_compilation()) { 513 const TypeTuple *domain = StartOSRNode::osr_domain(); 514 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 515 init_tf(TypeFunc::make(domain, range)); 516 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 517 initial_gvn()->set_type_bottom(s); 518 init_start(s); 519 cg = CallGenerator::for_osr(method(), entry_bci()); 520 } else { 521 // Normal case. 522 init_tf(TypeFunc::make(method())); 523 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 524 initial_gvn()->set_type_bottom(s); 525 init_start(s); 526 float past_uses = method()->interpreter_invocation_count(); 527 float expected_uses = past_uses; 528 cg = CallGenerator::for_inline(method(), expected_uses); 529 } 530 if (failing()) return; 531 if (cg == NULL) { 532 record_method_not_compilable_all_tiers("cannot parse method"); 533 return; 534 } 535 JVMState* jvms = build_start_state(start(), tf()); 536 if ((jvms = cg->generate(jvms)) == NULL) { 537 record_method_not_compilable("method parse failed"); 538 return; 539 } 540 GraphKit kit(jvms); 541 542 if (!kit.stopped()) { 543 // Accept return values, and transfer control we know not where. 544 // This is done by a special, unique ReturnNode bound to root. 545 return_values(kit.jvms()); 546 } 547 548 if (kit.has_exceptions()) { 549 // Any exceptions that escape from this call must be rethrown 550 // to whatever caller is dynamically above us on the stack. 551 // This is done by a special, unique RethrowNode bound to root. 552 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 553 } 554 555 print_method("Before RemoveUseless", 3); 556 557 // Remove clutter produced by parsing. 558 if (!failing()) { 559 ResourceMark rm; 560 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 561 } 562 } 563 564 // Note: Large methods are capped off in do_one_bytecode(). 565 if (failing()) return; 566 567 // After parsing, node notes are no longer automagic. 568 // They must be propagated by register_new_node_with_optimizer(), 569 // clone(), or the like. 570 set_default_node_notes(NULL); 571 572 for (;;) { 573 int successes = Inline_Warm(); 574 if (failing()) return; 575 if (successes == 0) break; 576 } 577 578 // Drain the list. 579 Finish_Warm(); 580 #ifndef PRODUCT 581 if (_printer) { 582 _printer->print_inlining(this); 583 } 584 #endif 585 586 if (failing()) return; 587 NOT_PRODUCT( verify_graph_edges(); ) 588 589 // Perform escape analysis 590 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 591 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 592 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction. 593 PhaseGVN* igvn = initial_gvn(); 594 Node* oop_null = igvn->zerocon(T_OBJECT); 595 Node* noop_null = igvn->zerocon(T_NARROWOOP); 596 597 _congraph = new(comp_arena()) ConnectionGraph(this); 598 bool has_non_escaping_obj = _congraph->compute_escape(); 599 600 #ifndef PRODUCT 601 if (PrintEscapeAnalysis) { 602 _congraph->dump(); 603 } 604 #endif 605 // Cleanup. 606 if (oop_null->outcnt() == 0) 607 igvn->hash_delete(oop_null); 608 if (noop_null->outcnt() == 0) 609 igvn->hash_delete(noop_null); 610 611 if (!has_non_escaping_obj) { 612 _congraph = NULL; 613 } 614 615 if (failing()) return; 616 } 617 // Now optimize 618 Optimize(); 619 if (failing()) return; 620 NOT_PRODUCT( verify_graph_edges(); ) 621 622 #ifndef PRODUCT 623 if (PrintIdeal) { 624 ttyLocker ttyl; // keep the following output all in one block 625 // This output goes directly to the tty, not the compiler log. 626 // To enable tools to match it up with the compilation activity, 627 // be sure to tag this tty output with the compile ID. 628 if (xtty != NULL) { 629 xtty->head("ideal compile_id='%d'%s", compile_id(), 630 is_osr_compilation() ? " compile_kind='osr'" : 631 ""); 632 } 633 root()->dump(9999); 634 if (xtty != NULL) { 635 xtty->tail("ideal"); 636 } 637 } 638 #endif 639 640 // Now that we know the size of all the monitors we can add a fixed slot 641 // for the original deopt pc. 642 643 _orig_pc_slot = fixed_slots(); 644 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 645 set_fixed_slots(next_slot); 646 647 // Now generate code 648 Code_Gen(); 649 if (failing()) return; 650 651 // Check if we want to skip execution of all compiled code. 652 { 653 #ifndef PRODUCT 654 if (OptoNoExecute) { 655 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 656 return; 657 } 658 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 659 #endif 660 661 if (is_osr_compilation()) { 662 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 663 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 664 } else { 665 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 666 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 667 } 668 669 env()->register_method(_method, _entry_bci, 670 &_code_offsets, 671 _orig_pc_slot_offset_in_bytes, 672 code_buffer(), 673 frame_size_in_words(), _oop_map_set, 674 &_handler_table, &_inc_table, 675 compiler, 676 env()->comp_level(), 677 true, /*has_debug_info*/ 678 has_unsafe_access() 679 ); 680 } 681 } 682 683 //------------------------------Compile---------------------------------------- 684 // Compile a runtime stub 685 Compile::Compile( ciEnv* ci_env, 686 TypeFunc_generator generator, 687 address stub_function, 688 const char *stub_name, 689 int is_fancy_jump, 690 bool pass_tls, 691 bool save_arg_registers, 692 bool return_pc ) 693 : Phase(Compiler), 694 _env(ci_env), 695 _log(ci_env->log()), 696 _compile_id(-1), 697 _save_argument_registers(save_arg_registers), 698 _method(NULL), 699 _stub_name(stub_name), 700 _stub_function(stub_function), 701 _stub_entry_point(NULL), 702 _entry_bci(InvocationEntryBci), 703 _initial_gvn(NULL), 704 _for_igvn(NULL), 705 _warm_calls(NULL), 706 _orig_pc_slot(0), 707 _orig_pc_slot_offset_in_bytes(0), 708 _subsume_loads(true), 709 _do_escape_analysis(false), 710 _failure_reason(NULL), 711 _code_buffer("Compile::Fill_buffer"), 712 _node_bundling_limit(0), 713 _node_bundling_base(NULL), 714 _nodes_limit((uint)MaxNodeLimit), 715 #ifndef PRODUCT 716 _trace_opto_output(TraceOptoOutput), 717 _printer(NULL), 718 #endif 719 _congraph(NULL) { 720 C = this; 721 722 #ifndef PRODUCT 723 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 724 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 725 set_print_assembly(PrintFrameConverterAssembly); 726 set_parsed_irreducible_loop(false); 727 #endif 728 CompileWrapper cw(this); 729 Init(/*AliasLevel=*/ 0); 730 init_tf((*generator)()); 731 732 { 733 // The following is a dummy for the sake of GraphKit::gen_stub 734 Unique_Node_List for_igvn(comp_arena()); 735 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 736 PhaseGVN gvn(Thread::current()->resource_area(),255); 737 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 738 gvn.transform_no_reclaim(top()); 739 740 GraphKit kit; 741 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 742 } 743 744 NOT_PRODUCT( verify_graph_edges(); ) 745 Code_Gen(); 746 if (failing()) return; 747 748 749 // Entry point will be accessed using compile->stub_entry_point(); 750 if (code_buffer() == NULL) { 751 Matcher::soft_match_failure(); 752 } else { 753 if (PrintAssembly && (WizardMode || Verbose)) 754 tty->print_cr("### Stub::%s", stub_name); 755 756 if (!failing()) { 757 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 758 759 // Make the NMethod 760 // For now we mark the frame as never safe for profile stackwalking 761 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 762 code_buffer(), 763 CodeOffsets::frame_never_safe, 764 // _code_offsets.value(CodeOffsets::Frame_Complete), 765 frame_size_in_words(), 766 _oop_map_set, 767 save_arg_registers); 768 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 769 770 _stub_entry_point = rs->entry_point(); 771 } 772 } 773 } 774 775 #ifndef PRODUCT 776 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 777 if(PrintOpto && Verbose) { 778 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 779 } 780 } 781 #endif 782 783 void Compile::print_codes() { 784 } 785 786 //------------------------------Init------------------------------------------- 787 // Prepare for a single compilation 788 void Compile::Init(int aliaslevel) { 789 _unique = 0; 790 _regalloc = NULL; 791 792 _tf = NULL; // filled in later 793 _top = NULL; // cached later 794 _matcher = NULL; // filled in later 795 _cfg = NULL; // filled in later 796 797 set_24_bit_selection_and_mode(Use24BitFP, false); 798 799 _node_note_array = NULL; 800 _default_node_notes = NULL; 801 802 _immutable_memory = NULL; // filled in at first inquiry 803 804 // Globally visible Nodes 805 // First set TOP to NULL to give safe behavior during creation of RootNode 806 set_cached_top_node(NULL); 807 set_root(new (this, 3) RootNode()); 808 // Now that you have a Root to point to, create the real TOP 809 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 810 set_recent_alloc(NULL, NULL); 811 812 // Create Debug Information Recorder to record scopes, oopmaps, etc. 813 env()->set_oop_recorder(new OopRecorder(comp_arena())); 814 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 815 env()->set_dependencies(new Dependencies(env())); 816 817 _fixed_slots = 0; 818 set_has_split_ifs(false); 819 set_has_loops(has_method() && method()->has_loops()); // first approximation 820 _deopt_happens = true; // start out assuming the worst 821 _trap_can_recompile = false; // no traps emitted yet 822 _major_progress = true; // start out assuming good things will happen 823 set_has_unsafe_access(false); 824 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 825 set_decompile_count(0); 826 827 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 828 // Compilation level related initialization 829 if (env()->comp_level() == CompLevel_fast_compile) { 830 set_num_loop_opts(Tier1LoopOptsCount); 831 set_do_inlining(Tier1Inline != 0); 832 set_max_inline_size(Tier1MaxInlineSize); 833 set_freq_inline_size(Tier1FreqInlineSize); 834 set_do_scheduling(false); 835 set_do_count_invocations(Tier1CountInvocations); 836 set_do_method_data_update(Tier1UpdateMethodData); 837 } else { 838 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 839 set_num_loop_opts(LoopOptsCount); 840 set_do_inlining(Inline); 841 set_max_inline_size(MaxInlineSize); 842 set_freq_inline_size(FreqInlineSize); 843 set_do_scheduling(OptoScheduling); 844 set_do_count_invocations(false); 845 set_do_method_data_update(false); 846 } 847 848 if (debug_info()->recording_non_safepoints()) { 849 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 850 (comp_arena(), 8, 0, NULL)); 851 set_default_node_notes(Node_Notes::make(this)); 852 } 853 854 // // -- Initialize types before each compile -- 855 // // Update cached type information 856 // if( _method && _method->constants() ) 857 // Type::update_loaded_types(_method, _method->constants()); 858 859 // Init alias_type map. 860 if (!_do_escape_analysis && aliaslevel == 3) 861 aliaslevel = 2; // No unique types without escape analysis 862 _AliasLevel = aliaslevel; 863 const int grow_ats = 16; 864 _max_alias_types = grow_ats; 865 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 866 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 867 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 868 { 869 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 870 } 871 // Initialize the first few types. 872 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 873 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 874 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 875 _num_alias_types = AliasIdxRaw+1; 876 // Zero out the alias type cache. 877 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 878 // A NULL adr_type hits in the cache right away. Preload the right answer. 879 probe_alias_cache(NULL)->_index = AliasIdxTop; 880 881 _intrinsics = NULL; 882 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 883 register_library_intrinsics(); 884 } 885 886 //---------------------------init_start---------------------------------------- 887 // Install the StartNode on this compile object. 888 void Compile::init_start(StartNode* s) { 889 if (failing()) 890 return; // already failing 891 assert(s == start(), ""); 892 } 893 894 StartNode* Compile::start() const { 895 assert(!failing(), ""); 896 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 897 Node* start = root()->fast_out(i); 898 if( start->is_Start() ) 899 return start->as_Start(); 900 } 901 ShouldNotReachHere(); 902 return NULL; 903 } 904 905 //-------------------------------immutable_memory------------------------------------- 906 // Access immutable memory 907 Node* Compile::immutable_memory() { 908 if (_immutable_memory != NULL) { 909 return _immutable_memory; 910 } 911 StartNode* s = start(); 912 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 913 Node *p = s->fast_out(i); 914 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 915 _immutable_memory = p; 916 return _immutable_memory; 917 } 918 } 919 ShouldNotReachHere(); 920 return NULL; 921 } 922 923 //----------------------set_cached_top_node------------------------------------ 924 // Install the cached top node, and make sure Node::is_top works correctly. 925 void Compile::set_cached_top_node(Node* tn) { 926 if (tn != NULL) verify_top(tn); 927 Node* old_top = _top; 928 _top = tn; 929 // Calling Node::setup_is_top allows the nodes the chance to adjust 930 // their _out arrays. 931 if (_top != NULL) _top->setup_is_top(); 932 if (old_top != NULL) old_top->setup_is_top(); 933 assert(_top == NULL || top()->is_top(), ""); 934 } 935 936 #ifndef PRODUCT 937 void Compile::verify_top(Node* tn) const { 938 if (tn != NULL) { 939 assert(tn->is_Con(), "top node must be a constant"); 940 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 941 assert(tn->in(0) != NULL, "must have live top node"); 942 } 943 } 944 #endif 945 946 947 ///-------------------Managing Per-Node Debug & Profile Info------------------- 948 949 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 950 guarantee(arr != NULL, ""); 951 int num_blocks = arr->length(); 952 if (grow_by < num_blocks) grow_by = num_blocks; 953 int num_notes = grow_by * _node_notes_block_size; 954 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 955 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 956 while (num_notes > 0) { 957 arr->append(notes); 958 notes += _node_notes_block_size; 959 num_notes -= _node_notes_block_size; 960 } 961 assert(num_notes == 0, "exact multiple, please"); 962 } 963 964 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 965 if (source == NULL || dest == NULL) return false; 966 967 if (dest->is_Con()) 968 return false; // Do not push debug info onto constants. 969 970 #ifdef ASSERT 971 // Leave a bread crumb trail pointing to the original node: 972 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 973 dest->set_debug_orig(source); 974 } 975 #endif 976 977 if (node_note_array() == NULL) 978 return false; // Not collecting any notes now. 979 980 // This is a copy onto a pre-existing node, which may already have notes. 981 // If both nodes have notes, do not overwrite any pre-existing notes. 982 Node_Notes* source_notes = node_notes_at(source->_idx); 983 if (source_notes == NULL || source_notes->is_clear()) return false; 984 Node_Notes* dest_notes = node_notes_at(dest->_idx); 985 if (dest_notes == NULL || dest_notes->is_clear()) { 986 return set_node_notes_at(dest->_idx, source_notes); 987 } 988 989 Node_Notes merged_notes = (*source_notes); 990 // The order of operations here ensures that dest notes will win... 991 merged_notes.update_from(dest_notes); 992 return set_node_notes_at(dest->_idx, &merged_notes); 993 } 994 995 996 //--------------------------allow_range_check_smearing------------------------- 997 // Gating condition for coalescing similar range checks. 998 // Sometimes we try 'speculatively' replacing a series of a range checks by a 999 // single covering check that is at least as strong as any of them. 1000 // If the optimization succeeds, the simplified (strengthened) range check 1001 // will always succeed. If it fails, we will deopt, and then give up 1002 // on the optimization. 1003 bool Compile::allow_range_check_smearing() const { 1004 // If this method has already thrown a range-check, 1005 // assume it was because we already tried range smearing 1006 // and it failed. 1007 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1008 return !already_trapped; 1009 } 1010 1011 1012 //------------------------------flatten_alias_type----------------------------- 1013 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1014 int offset = tj->offset(); 1015 TypePtr::PTR ptr = tj->ptr(); 1016 1017 // Known instance (scalarizable allocation) alias only with itself. 1018 bool is_known_inst = tj->isa_oopptr() != NULL && 1019 tj->is_oopptr()->is_known_instance(); 1020 1021 // Process weird unsafe references. 1022 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1023 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1024 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1025 tj = TypeOopPtr::BOTTOM; 1026 ptr = tj->ptr(); 1027 offset = tj->offset(); 1028 } 1029 1030 // Array pointers need some flattening 1031 const TypeAryPtr *ta = tj->isa_aryptr(); 1032 if( ta && is_known_inst ) { 1033 if ( offset != Type::OffsetBot && 1034 offset > arrayOopDesc::length_offset_in_bytes() ) { 1035 offset = Type::OffsetBot; // Flatten constant access into array body only 1036 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1037 } 1038 } else if( ta && _AliasLevel >= 2 ) { 1039 // For arrays indexed by constant indices, we flatten the alias 1040 // space to include all of the array body. Only the header, klass 1041 // and array length can be accessed un-aliased. 1042 if( offset != Type::OffsetBot ) { 1043 if( ta->const_oop() ) { // methodDataOop or methodOop 1044 offset = Type::OffsetBot; // Flatten constant access into array body 1045 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1046 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1047 // range is OK as-is. 1048 tj = ta = TypeAryPtr::RANGE; 1049 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1050 tj = TypeInstPtr::KLASS; // all klass loads look alike 1051 ta = TypeAryPtr::RANGE; // generic ignored junk 1052 ptr = TypePtr::BotPTR; 1053 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1054 tj = TypeInstPtr::MARK; 1055 ta = TypeAryPtr::RANGE; // generic ignored junk 1056 ptr = TypePtr::BotPTR; 1057 } else { // Random constant offset into array body 1058 offset = Type::OffsetBot; // Flatten constant access into array body 1059 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1060 } 1061 } 1062 // Arrays of fixed size alias with arrays of unknown size. 1063 if (ta->size() != TypeInt::POS) { 1064 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1065 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1066 } 1067 // Arrays of known objects become arrays of unknown objects. 1068 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1069 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1070 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1071 } 1072 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1073 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1074 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1075 } 1076 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1077 // cannot be distinguished by bytecode alone. 1078 if (ta->elem() == TypeInt::BOOL) { 1079 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1080 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1081 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1082 } 1083 // During the 2nd round of IterGVN, NotNull castings are removed. 1084 // Make sure the Bottom and NotNull variants alias the same. 1085 // Also, make sure exact and non-exact variants alias the same. 1086 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1087 if (ta->const_oop()) { 1088 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1089 } else { 1090 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1091 } 1092 } 1093 } 1094 1095 // Oop pointers need some flattening 1096 const TypeInstPtr *to = tj->isa_instptr(); 1097 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1098 if( ptr == TypePtr::Constant ) { 1099 // No constant oop pointers (such as Strings); they alias with 1100 // unknown strings. 1101 assert(!is_known_inst, "not scalarizable allocation"); 1102 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1103 } else if( is_known_inst ) { 1104 tj = to; // Keep NotNull and klass_is_exact for instance type 1105 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1106 // During the 2nd round of IterGVN, NotNull castings are removed. 1107 // Make sure the Bottom and NotNull variants alias the same. 1108 // Also, make sure exact and non-exact variants alias the same. 1109 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1110 } 1111 // Canonicalize the holder of this field 1112 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1113 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1114 // First handle header references such as a LoadKlassNode, even if the 1115 // object's klass is unloaded at compile time (4965979). 1116 if (!is_known_inst) { // Do it only for non-instance types 1117 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1118 } 1119 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1120 to = NULL; 1121 tj = TypeOopPtr::BOTTOM; 1122 offset = tj->offset(); 1123 } else { 1124 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1125 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1126 if( is_known_inst ) { 1127 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1128 } else { 1129 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1130 } 1131 } 1132 } 1133 } 1134 1135 // Klass pointers to object array klasses need some flattening 1136 const TypeKlassPtr *tk = tj->isa_klassptr(); 1137 if( tk ) { 1138 // If we are referencing a field within a Klass, we need 1139 // to assume the worst case of an Object. Both exact and 1140 // inexact types must flatten to the same alias class. 1141 // Since the flattened result for a klass is defined to be 1142 // precisely java.lang.Object, use a constant ptr. 1143 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1144 1145 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1146 TypeKlassPtr::OBJECT->klass(), 1147 offset); 1148 } 1149 1150 ciKlass* klass = tk->klass(); 1151 if( klass->is_obj_array_klass() ) { 1152 ciKlass* k = TypeAryPtr::OOPS->klass(); 1153 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1154 k = TypeInstPtr::BOTTOM->klass(); 1155 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1156 } 1157 1158 // Check for precise loads from the primary supertype array and force them 1159 // to the supertype cache alias index. Check for generic array loads from 1160 // the primary supertype array and also force them to the supertype cache 1161 // alias index. Since the same load can reach both, we need to merge 1162 // these 2 disparate memories into the same alias class. Since the 1163 // primary supertype array is read-only, there's no chance of confusion 1164 // where we bypass an array load and an array store. 1165 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1166 if( offset == Type::OffsetBot || 1167 off2 < Klass::primary_super_limit()*wordSize ) { 1168 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1169 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1170 } 1171 } 1172 1173 // Flatten all Raw pointers together. 1174 if (tj->base() == Type::RawPtr) 1175 tj = TypeRawPtr::BOTTOM; 1176 1177 if (tj->base() == Type::AnyPtr) 1178 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1179 1180 // Flatten all to bottom for now 1181 switch( _AliasLevel ) { 1182 case 0: 1183 tj = TypePtr::BOTTOM; 1184 break; 1185 case 1: // Flatten to: oop, static, field or array 1186 switch (tj->base()) { 1187 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1188 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1189 case Type::AryPtr: // do not distinguish arrays at all 1190 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1191 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1192 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1193 default: ShouldNotReachHere(); 1194 } 1195 break; 1196 case 2: // No collapsing at level 2; keep all splits 1197 case 3: // No collapsing at level 3; keep all splits 1198 break; 1199 default: 1200 Unimplemented(); 1201 } 1202 1203 offset = tj->offset(); 1204 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1205 1206 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1207 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1208 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1209 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1210 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1211 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1212 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1213 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1214 assert( tj->ptr() != TypePtr::TopPTR && 1215 tj->ptr() != TypePtr::AnyNull && 1216 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1217 // assert( tj->ptr() != TypePtr::Constant || 1218 // tj->base() == Type::RawPtr || 1219 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1220 1221 return tj; 1222 } 1223 1224 void Compile::AliasType::Init(int i, const TypePtr* at) { 1225 _index = i; 1226 _adr_type = at; 1227 _field = NULL; 1228 _is_rewritable = true; // default 1229 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1230 if (atoop != NULL && atoop->is_known_instance()) { 1231 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1232 _general_index = Compile::current()->get_alias_index(gt); 1233 } else { 1234 _general_index = 0; 1235 } 1236 } 1237 1238 //---------------------------------print_on------------------------------------ 1239 #ifndef PRODUCT 1240 void Compile::AliasType::print_on(outputStream* st) { 1241 if (index() < 10) 1242 st->print("@ <%d> ", index()); 1243 else st->print("@ <%d>", index()); 1244 st->print(is_rewritable() ? " " : " RO"); 1245 int offset = adr_type()->offset(); 1246 if (offset == Type::OffsetBot) 1247 st->print(" +any"); 1248 else st->print(" +%-3d", offset); 1249 st->print(" in "); 1250 adr_type()->dump_on(st); 1251 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1252 if (field() != NULL && tjp) { 1253 if (tjp->klass() != field()->holder() || 1254 tjp->offset() != field()->offset_in_bytes()) { 1255 st->print(" != "); 1256 field()->print(); 1257 st->print(" ***"); 1258 } 1259 } 1260 } 1261 1262 void print_alias_types() { 1263 Compile* C = Compile::current(); 1264 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1265 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1266 C->alias_type(idx)->print_on(tty); 1267 tty->cr(); 1268 } 1269 } 1270 #endif 1271 1272 1273 //----------------------------probe_alias_cache-------------------------------- 1274 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1275 intptr_t key = (intptr_t) adr_type; 1276 key ^= key >> logAliasCacheSize; 1277 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1278 } 1279 1280 1281 //-----------------------------grow_alias_types-------------------------------- 1282 void Compile::grow_alias_types() { 1283 const int old_ats = _max_alias_types; // how many before? 1284 const int new_ats = old_ats; // how many more? 1285 const int grow_ats = old_ats+new_ats; // how many now? 1286 _max_alias_types = grow_ats; 1287 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1288 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1289 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1290 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1291 } 1292 1293 1294 //--------------------------------find_alias_type------------------------------ 1295 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1296 if (_AliasLevel == 0) 1297 return alias_type(AliasIdxBot); 1298 1299 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1300 if (ace->_adr_type == adr_type) { 1301 return alias_type(ace->_index); 1302 } 1303 1304 // Handle special cases. 1305 if (adr_type == NULL) return alias_type(AliasIdxTop); 1306 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1307 1308 // Do it the slow way. 1309 const TypePtr* flat = flatten_alias_type(adr_type); 1310 1311 #ifdef ASSERT 1312 assert(flat == flatten_alias_type(flat), "idempotent"); 1313 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1314 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1315 const TypeOopPtr* foop = flat->is_oopptr(); 1316 // Scalarizable allocations have exact klass always. 1317 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1318 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1319 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1320 } 1321 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1322 #endif 1323 1324 int idx = AliasIdxTop; 1325 for (int i = 0; i < num_alias_types(); i++) { 1326 if (alias_type(i)->adr_type() == flat) { 1327 idx = i; 1328 break; 1329 } 1330 } 1331 1332 if (idx == AliasIdxTop) { 1333 if (no_create) return NULL; 1334 // Grow the array if necessary. 1335 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1336 // Add a new alias type. 1337 idx = _num_alias_types++; 1338 _alias_types[idx]->Init(idx, flat); 1339 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1340 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1341 if (flat->isa_instptr()) { 1342 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1343 && flat->is_instptr()->klass() == env()->Class_klass()) 1344 alias_type(idx)->set_rewritable(false); 1345 } 1346 if (flat->isa_klassptr()) { 1347 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1348 alias_type(idx)->set_rewritable(false); 1349 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1350 alias_type(idx)->set_rewritable(false); 1351 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1352 alias_type(idx)->set_rewritable(false); 1353 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1354 alias_type(idx)->set_rewritable(false); 1355 } 1356 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1357 // but the base pointer type is not distinctive enough to identify 1358 // references into JavaThread.) 1359 1360 // Check for final instance fields. 1361 const TypeInstPtr* tinst = flat->isa_instptr(); 1362 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1363 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1364 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1365 // Set field() and is_rewritable() attributes. 1366 if (field != NULL) alias_type(idx)->set_field(field); 1367 } 1368 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1369 // Check for final static fields. 1370 if (tklass && tklass->klass()->is_instance_klass()) { 1371 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1372 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1373 // Set field() and is_rewritable() attributes. 1374 if (field != NULL) alias_type(idx)->set_field(field); 1375 } 1376 } 1377 1378 // Fill the cache for next time. 1379 ace->_adr_type = adr_type; 1380 ace->_index = idx; 1381 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1382 1383 // Might as well try to fill the cache for the flattened version, too. 1384 AliasCacheEntry* face = probe_alias_cache(flat); 1385 if (face->_adr_type == NULL) { 1386 face->_adr_type = flat; 1387 face->_index = idx; 1388 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1389 } 1390 1391 return alias_type(idx); 1392 } 1393 1394 1395 Compile::AliasType* Compile::alias_type(ciField* field) { 1396 const TypeOopPtr* t; 1397 if (field->is_static()) 1398 t = TypeKlassPtr::make(field->holder()); 1399 else 1400 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1401 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1402 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1403 return atp; 1404 } 1405 1406 1407 //------------------------------have_alias_type-------------------------------- 1408 bool Compile::have_alias_type(const TypePtr* adr_type) { 1409 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1410 if (ace->_adr_type == adr_type) { 1411 return true; 1412 } 1413 1414 // Handle special cases. 1415 if (adr_type == NULL) return true; 1416 if (adr_type == TypePtr::BOTTOM) return true; 1417 1418 return find_alias_type(adr_type, true) != NULL; 1419 } 1420 1421 //-----------------------------must_alias-------------------------------------- 1422 // True if all values of the given address type are in the given alias category. 1423 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1424 if (alias_idx == AliasIdxBot) return true; // the universal category 1425 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1426 if (alias_idx == AliasIdxTop) return false; // the empty category 1427 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1428 1429 // the only remaining possible overlap is identity 1430 int adr_idx = get_alias_index(adr_type); 1431 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1432 assert(adr_idx == alias_idx || 1433 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1434 && adr_type != TypeOopPtr::BOTTOM), 1435 "should not be testing for overlap with an unsafe pointer"); 1436 return adr_idx == alias_idx; 1437 } 1438 1439 //------------------------------can_alias-------------------------------------- 1440 // True if any values of the given address type are in the given alias category. 1441 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1442 if (alias_idx == AliasIdxTop) return false; // the empty category 1443 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1444 if (alias_idx == AliasIdxBot) return true; // the universal category 1445 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1446 1447 // the only remaining possible overlap is identity 1448 int adr_idx = get_alias_index(adr_type); 1449 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1450 return adr_idx == alias_idx; 1451 } 1452 1453 1454 1455 //---------------------------pop_warm_call------------------------------------- 1456 WarmCallInfo* Compile::pop_warm_call() { 1457 WarmCallInfo* wci = _warm_calls; 1458 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1459 return wci; 1460 } 1461 1462 //----------------------------Inline_Warm-------------------------------------- 1463 int Compile::Inline_Warm() { 1464 // If there is room, try to inline some more warm call sites. 1465 // %%% Do a graph index compaction pass when we think we're out of space? 1466 if (!InlineWarmCalls) return 0; 1467 1468 int calls_made_hot = 0; 1469 int room_to_grow = NodeCountInliningCutoff - unique(); 1470 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1471 int amount_grown = 0; 1472 WarmCallInfo* call; 1473 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1474 int est_size = (int)call->size(); 1475 if (est_size > (room_to_grow - amount_grown)) { 1476 // This one won't fit anyway. Get rid of it. 1477 call->make_cold(); 1478 continue; 1479 } 1480 call->make_hot(); 1481 calls_made_hot++; 1482 amount_grown += est_size; 1483 amount_to_grow -= est_size; 1484 } 1485 1486 if (calls_made_hot > 0) set_major_progress(); 1487 return calls_made_hot; 1488 } 1489 1490 1491 //----------------------------Finish_Warm-------------------------------------- 1492 void Compile::Finish_Warm() { 1493 if (!InlineWarmCalls) return; 1494 if (failing()) return; 1495 if (warm_calls() == NULL) return; 1496 1497 // Clean up loose ends, if we are out of space for inlining. 1498 WarmCallInfo* call; 1499 while ((call = pop_warm_call()) != NULL) { 1500 call->make_cold(); 1501 } 1502 } 1503 1504 1505 //------------------------------Optimize--------------------------------------- 1506 // Given a graph, optimize it. 1507 void Compile::Optimize() { 1508 TracePhase t1("optimizer", &_t_optimizer, true); 1509 1510 #ifndef PRODUCT 1511 if (env()->break_at_compile()) { 1512 BREAKPOINT; 1513 } 1514 1515 #endif 1516 1517 ResourceMark rm; 1518 int loop_opts_cnt; 1519 1520 NOT_PRODUCT( verify_graph_edges(); ) 1521 1522 print_method("After Parsing"); 1523 1524 { 1525 // Iterative Global Value Numbering, including ideal transforms 1526 // Initialize IterGVN with types and values from parse-time GVN 1527 PhaseIterGVN igvn(initial_gvn()); 1528 { 1529 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1530 igvn.optimize(); 1531 } 1532 1533 print_method("Iter GVN 1", 2); 1534 1535 if (failing()) return; 1536 1537 // Loop transforms on the ideal graph. Range Check Elimination, 1538 // peeling, unrolling, etc. 1539 1540 // Set loop opts counter 1541 loop_opts_cnt = num_loop_opts(); 1542 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1543 { 1544 TracePhase t2("idealLoop", &_t_idealLoop, true); 1545 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1546 loop_opts_cnt--; 1547 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1548 if (failing()) return; 1549 } 1550 // Loop opts pass if partial peeling occurred in previous pass 1551 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1552 TracePhase t3("idealLoop", &_t_idealLoop, true); 1553 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1554 loop_opts_cnt--; 1555 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1556 if (failing()) return; 1557 } 1558 // Loop opts pass for loop-unrolling before CCP 1559 if(major_progress() && (loop_opts_cnt > 0)) { 1560 TracePhase t4("idealLoop", &_t_idealLoop, true); 1561 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1562 loop_opts_cnt--; 1563 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1564 } 1565 } 1566 if (failing()) return; 1567 1568 // Conditional Constant Propagation; 1569 PhaseCCP ccp( &igvn ); 1570 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1571 { 1572 TracePhase t2("ccp", &_t_ccp, true); 1573 ccp.do_transform(); 1574 } 1575 print_method("PhaseCPP 1", 2); 1576 1577 assert( true, "Break here to ccp.dump_old2new_map()"); 1578 1579 // Iterative Global Value Numbering, including ideal transforms 1580 { 1581 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1582 igvn = ccp; 1583 igvn.optimize(); 1584 } 1585 1586 print_method("Iter GVN 2", 2); 1587 1588 if (failing()) return; 1589 1590 // Loop transforms on the ideal graph. Range Check Elimination, 1591 // peeling, unrolling, etc. 1592 if(loop_opts_cnt > 0) { 1593 debug_only( int cnt = 0; ); 1594 while(major_progress() && (loop_opts_cnt > 0)) { 1595 TracePhase t2("idealLoop", &_t_idealLoop, true); 1596 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1597 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1598 loop_opts_cnt--; 1599 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1600 if (failing()) return; 1601 } 1602 } 1603 { 1604 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1605 PhaseMacroExpand mex(igvn); 1606 if (mex.expand_macro_nodes()) { 1607 assert(failing(), "must bail out w/ explicit message"); 1608 return; 1609 } 1610 } 1611 1612 } // (End scope of igvn; run destructor if necessary for asserts.) 1613 1614 // A method with only infinite loops has no edges entering loops from root 1615 { 1616 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1617 if (final_graph_reshaping()) { 1618 assert(failing(), "must bail out w/ explicit message"); 1619 return; 1620 } 1621 } 1622 1623 print_method("Optimize finished", 2); 1624 } 1625 1626 1627 //------------------------------Code_Gen--------------------------------------- 1628 // Given a graph, generate code for it 1629 void Compile::Code_Gen() { 1630 if (failing()) return; 1631 1632 // Perform instruction selection. You might think we could reclaim Matcher 1633 // memory PDQ, but actually the Matcher is used in generating spill code. 1634 // Internals of the Matcher (including some VectorSets) must remain live 1635 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1636 // set a bit in reclaimed memory. 1637 1638 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1639 // nodes. Mapping is only valid at the root of each matched subtree. 1640 NOT_PRODUCT( verify_graph_edges(); ) 1641 1642 Node_List proj_list; 1643 Matcher m(proj_list); 1644 _matcher = &m; 1645 { 1646 TracePhase t2("matcher", &_t_matcher, true); 1647 m.match(); 1648 } 1649 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1650 // nodes. Mapping is only valid at the root of each matched subtree. 1651 NOT_PRODUCT( verify_graph_edges(); ) 1652 1653 // If you have too many nodes, or if matching has failed, bail out 1654 check_node_count(0, "out of nodes matching instructions"); 1655 if (failing()) return; 1656 1657 // Build a proper-looking CFG 1658 PhaseCFG cfg(node_arena(), root(), m); 1659 _cfg = &cfg; 1660 { 1661 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1662 cfg.Dominators(); 1663 if (failing()) return; 1664 1665 NOT_PRODUCT( verify_graph_edges(); ) 1666 1667 cfg.Estimate_Block_Frequency(); 1668 cfg.GlobalCodeMotion(m,unique(),proj_list); 1669 1670 print_method("Global code motion", 2); 1671 1672 if (failing()) return; 1673 NOT_PRODUCT( verify_graph_edges(); ) 1674 1675 debug_only( cfg.verify(); ) 1676 } 1677 NOT_PRODUCT( verify_graph_edges(); ) 1678 1679 PhaseChaitin regalloc(unique(),cfg,m); 1680 _regalloc = ®alloc; 1681 { 1682 TracePhase t2("regalloc", &_t_registerAllocation, true); 1683 // Perform any platform dependent preallocation actions. This is used, 1684 // for example, to avoid taking an implicit null pointer exception 1685 // using the frame pointer on win95. 1686 _regalloc->pd_preallocate_hook(); 1687 1688 // Perform register allocation. After Chaitin, use-def chains are 1689 // no longer accurate (at spill code) and so must be ignored. 1690 // Node->LRG->reg mappings are still accurate. 1691 _regalloc->Register_Allocate(); 1692 1693 // Bail out if the allocator builds too many nodes 1694 if (failing()) return; 1695 } 1696 1697 // Prior to register allocation we kept empty basic blocks in case the 1698 // the allocator needed a place to spill. After register allocation we 1699 // are not adding any new instructions. If any basic block is empty, we 1700 // can now safely remove it. 1701 { 1702 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1703 cfg.remove_empty(); 1704 if (do_freq_based_layout()) { 1705 PhaseBlockLayout layout(cfg); 1706 } else { 1707 cfg.set_loop_alignment(); 1708 } 1709 cfg.fixup_flow(); 1710 } 1711 1712 // Perform any platform dependent postallocation verifications. 1713 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1714 1715 // Apply peephole optimizations 1716 if( OptoPeephole ) { 1717 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1718 PhasePeephole peep( _regalloc, cfg); 1719 peep.do_transform(); 1720 } 1721 1722 // Convert Nodes to instruction bits in a buffer 1723 { 1724 // %%%% workspace merge brought two timers together for one job 1725 TracePhase t2a("output", &_t_output, true); 1726 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1727 Output(); 1728 } 1729 1730 print_method("Final Code"); 1731 1732 // He's dead, Jim. 1733 _cfg = (PhaseCFG*)0xdeadbeef; 1734 _regalloc = (PhaseChaitin*)0xdeadbeef; 1735 } 1736 1737 1738 //------------------------------dump_asm--------------------------------------- 1739 // Dump formatted assembly 1740 #ifndef PRODUCT 1741 void Compile::dump_asm(int *pcs, uint pc_limit) { 1742 bool cut_short = false; 1743 tty->print_cr("#"); 1744 tty->print("# "); _tf->dump(); tty->cr(); 1745 tty->print_cr("#"); 1746 1747 // For all blocks 1748 int pc = 0x0; // Program counter 1749 char starts_bundle = ' '; 1750 _regalloc->dump_frame(); 1751 1752 Node *n = NULL; 1753 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1754 if (VMThread::should_terminate()) { cut_short = true; break; } 1755 Block *b = _cfg->_blocks[i]; 1756 if (b->is_connector() && !Verbose) continue; 1757 n = b->_nodes[0]; 1758 if (pcs && n->_idx < pc_limit) 1759 tty->print("%3.3x ", pcs[n->_idx]); 1760 else 1761 tty->print(" "); 1762 b->dump_head( &_cfg->_bbs ); 1763 if (b->is_connector()) { 1764 tty->print_cr(" # Empty connector block"); 1765 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1766 tty->print_cr(" # Block is sole successor of call"); 1767 } 1768 1769 // For all instructions 1770 Node *delay = NULL; 1771 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1772 if (VMThread::should_terminate()) { cut_short = true; break; } 1773 n = b->_nodes[j]; 1774 if (valid_bundle_info(n)) { 1775 Bundle *bundle = node_bundling(n); 1776 if (bundle->used_in_unconditional_delay()) { 1777 delay = n; 1778 continue; 1779 } 1780 if (bundle->starts_bundle()) 1781 starts_bundle = '+'; 1782 } 1783 1784 if (WizardMode) n->dump(); 1785 1786 if( !n->is_Region() && // Dont print in the Assembly 1787 !n->is_Phi() && // a few noisely useless nodes 1788 !n->is_Proj() && 1789 !n->is_MachTemp() && 1790 !n->is_Catch() && // Would be nice to print exception table targets 1791 !n->is_MergeMem() && // Not very interesting 1792 !n->is_top() && // Debug info table constants 1793 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1794 ) { 1795 if (pcs && n->_idx < pc_limit) 1796 tty->print("%3.3x", pcs[n->_idx]); 1797 else 1798 tty->print(" "); 1799 tty->print(" %c ", starts_bundle); 1800 starts_bundle = ' '; 1801 tty->print("\t"); 1802 n->format(_regalloc, tty); 1803 tty->cr(); 1804 } 1805 1806 // If we have an instruction with a delay slot, and have seen a delay, 1807 // then back up and print it 1808 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1809 assert(delay != NULL, "no unconditional delay instruction"); 1810 if (WizardMode) delay->dump(); 1811 1812 if (node_bundling(delay)->starts_bundle()) 1813 starts_bundle = '+'; 1814 if (pcs && n->_idx < pc_limit) 1815 tty->print("%3.3x", pcs[n->_idx]); 1816 else 1817 tty->print(" "); 1818 tty->print(" %c ", starts_bundle); 1819 starts_bundle = ' '; 1820 tty->print("\t"); 1821 delay->format(_regalloc, tty); 1822 tty->print_cr(""); 1823 delay = NULL; 1824 } 1825 1826 // Dump the exception table as well 1827 if( n->is_Catch() && (Verbose || WizardMode) ) { 1828 // Print the exception table for this offset 1829 _handler_table.print_subtable_for(pc); 1830 } 1831 } 1832 1833 if (pcs && n->_idx < pc_limit) 1834 tty->print_cr("%3.3x", pcs[n->_idx]); 1835 else 1836 tty->print_cr(""); 1837 1838 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1839 1840 } // End of per-block dump 1841 tty->print_cr(""); 1842 1843 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1844 } 1845 #endif 1846 1847 //------------------------------Final_Reshape_Counts--------------------------- 1848 // This class defines counters to help identify when a method 1849 // may/must be executed using hardware with only 24-bit precision. 1850 struct Final_Reshape_Counts : public StackObj { 1851 int _call_count; // count non-inlined 'common' calls 1852 int _float_count; // count float ops requiring 24-bit precision 1853 int _double_count; // count double ops requiring more precision 1854 int _java_call_count; // count non-inlined 'java' calls 1855 VectorSet _visited; // Visitation flags 1856 Node_List _tests; // Set of IfNodes & PCTableNodes 1857 1858 Final_Reshape_Counts() : 1859 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), 1860 _visited( Thread::current()->resource_area() ) { } 1861 1862 void inc_call_count () { _call_count ++; } 1863 void inc_float_count () { _float_count ++; } 1864 void inc_double_count() { _double_count++; } 1865 void inc_java_call_count() { _java_call_count++; } 1866 1867 int get_call_count () const { return _call_count ; } 1868 int get_float_count () const { return _float_count ; } 1869 int get_double_count() const { return _double_count; } 1870 int get_java_call_count() const { return _java_call_count; } 1871 }; 1872 1873 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1874 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1875 // Make sure the offset goes inside the instance layout. 1876 return k->contains_field_offset(tp->offset()); 1877 // Note that OffsetBot and OffsetTop are very negative. 1878 } 1879 1880 //------------------------------final_graph_reshaping_impl---------------------- 1881 // Implement items 1-5 from final_graph_reshaping below. 1882 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { 1883 1884 if ( n->outcnt() == 0 ) return; // dead node 1885 uint nop = n->Opcode(); 1886 1887 // Check for 2-input instruction with "last use" on right input. 1888 // Swap to left input. Implements item (2). 1889 if( n->req() == 3 && // two-input instruction 1890 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1891 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1892 n->in(2)->outcnt() == 1 &&// right use IS a last use 1893 !n->in(2)->is_Con() ) { // right use is not a constant 1894 // Check for commutative opcode 1895 switch( nop ) { 1896 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1897 case Op_MaxI: case Op_MinI: 1898 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1899 case Op_AndL: case Op_XorL: case Op_OrL: 1900 case Op_AndI: case Op_XorI: case Op_OrI: { 1901 // Move "last use" input to left by swapping inputs 1902 n->swap_edges(1, 2); 1903 break; 1904 } 1905 default: 1906 break; 1907 } 1908 } 1909 1910 // Count FPU ops and common calls, implements item (3) 1911 switch( nop ) { 1912 // Count all float operations that may use FPU 1913 case Op_AddF: 1914 case Op_SubF: 1915 case Op_MulF: 1916 case Op_DivF: 1917 case Op_NegF: 1918 case Op_ModF: 1919 case Op_ConvI2F: 1920 case Op_ConF: 1921 case Op_CmpF: 1922 case Op_CmpF3: 1923 // case Op_ConvL2F: // longs are split into 32-bit halves 1924 fpu.inc_float_count(); 1925 break; 1926 1927 case Op_ConvF2D: 1928 case Op_ConvD2F: 1929 fpu.inc_float_count(); 1930 fpu.inc_double_count(); 1931 break; 1932 1933 // Count all double operations that may use FPU 1934 case Op_AddD: 1935 case Op_SubD: 1936 case Op_MulD: 1937 case Op_DivD: 1938 case Op_NegD: 1939 case Op_ModD: 1940 case Op_ConvI2D: 1941 case Op_ConvD2I: 1942 // case Op_ConvL2D: // handled by leaf call 1943 // case Op_ConvD2L: // handled by leaf call 1944 case Op_ConD: 1945 case Op_CmpD: 1946 case Op_CmpD3: 1947 fpu.inc_double_count(); 1948 break; 1949 case Op_Opaque1: // Remove Opaque Nodes before matching 1950 case Op_Opaque2: // Remove Opaque Nodes before matching 1951 n->subsume_by(n->in(1)); 1952 break; 1953 case Op_CallStaticJava: 1954 case Op_CallJava: 1955 case Op_CallDynamicJava: 1956 fpu.inc_java_call_count(); // Count java call site; 1957 case Op_CallRuntime: 1958 case Op_CallLeaf: 1959 case Op_CallLeafNoFP: { 1960 assert( n->is_Call(), "" ); 1961 CallNode *call = n->as_Call(); 1962 // Count call sites where the FP mode bit would have to be flipped. 1963 // Do not count uncommon runtime calls: 1964 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 1965 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 1966 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 1967 fpu.inc_call_count(); // Count the call site 1968 } else { // See if uncommon argument is shared 1969 Node *n = call->in(TypeFunc::Parms); 1970 int nop = n->Opcode(); 1971 // Clone shared simple arguments to uncommon calls, item (1). 1972 if( n->outcnt() > 1 && 1973 !n->is_Proj() && 1974 nop != Op_CreateEx && 1975 nop != Op_CheckCastPP && 1976 nop != Op_DecodeN && 1977 !n->is_Mem() ) { 1978 Node *x = n->clone(); 1979 call->set_req( TypeFunc::Parms, x ); 1980 } 1981 } 1982 break; 1983 } 1984 1985 case Op_StoreD: 1986 case Op_LoadD: 1987 case Op_LoadD_unaligned: 1988 fpu.inc_double_count(); 1989 goto handle_mem; 1990 case Op_StoreF: 1991 case Op_LoadF: 1992 fpu.inc_float_count(); 1993 goto handle_mem; 1994 1995 case Op_StoreB: 1996 case Op_StoreC: 1997 case Op_StoreCM: 1998 case Op_StorePConditional: 1999 case Op_StoreI: 2000 case Op_StoreL: 2001 case Op_StoreIConditional: 2002 case Op_StoreLConditional: 2003 case Op_CompareAndSwapI: 2004 case Op_CompareAndSwapL: 2005 case Op_CompareAndSwapP: 2006 case Op_CompareAndSwapN: 2007 case Op_StoreP: 2008 case Op_StoreN: 2009 case Op_LoadB: 2010 case Op_LoadUB: 2011 case Op_LoadUS: 2012 case Op_LoadI: 2013 case Op_LoadUI2L: 2014 case Op_LoadKlass: 2015 case Op_LoadNKlass: 2016 case Op_LoadL: 2017 case Op_LoadL_unaligned: 2018 case Op_LoadPLocked: 2019 case Op_LoadLLocked: 2020 case Op_LoadP: 2021 case Op_LoadN: 2022 case Op_LoadRange: 2023 case Op_LoadS: { 2024 handle_mem: 2025 #ifdef ASSERT 2026 if( VerifyOptoOopOffsets ) { 2027 assert( n->is_Mem(), "" ); 2028 MemNode *mem = (MemNode*)n; 2029 // Check to see if address types have grounded out somehow. 2030 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2031 assert( !tp || oop_offset_is_sane(tp), "" ); 2032 } 2033 #endif 2034 break; 2035 } 2036 2037 case Op_AddP: { // Assert sane base pointers 2038 Node *addp = n->in(AddPNode::Address); 2039 assert( !addp->is_AddP() || 2040 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2041 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2042 "Base pointers must match" ); 2043 #ifdef _LP64 2044 if (UseCompressedOops && 2045 addp->Opcode() == Op_ConP && 2046 addp == n->in(AddPNode::Base) && 2047 n->in(AddPNode::Offset)->is_Con()) { 2048 // Use addressing with narrow klass to load with offset on x86. 2049 // On sparc loading 32-bits constant and decoding it have less 2050 // instructions (4) then load 64-bits constant (7). 2051 // Do this transformation here since IGVN will convert ConN back to ConP. 2052 const Type* t = addp->bottom_type(); 2053 if (t->isa_oopptr()) { 2054 Node* nn = NULL; 2055 2056 // Look for existing ConN node of the same exact type. 2057 Compile* C = Compile::current(); 2058 Node* r = C->root(); 2059 uint cnt = r->outcnt(); 2060 for (uint i = 0; i < cnt; i++) { 2061 Node* m = r->raw_out(i); 2062 if (m!= NULL && m->Opcode() == Op_ConN && 2063 m->bottom_type()->make_ptr() == t) { 2064 nn = m; 2065 break; 2066 } 2067 } 2068 if (nn != NULL) { 2069 // Decode a narrow oop to match address 2070 // [R12 + narrow_oop_reg<<3 + offset] 2071 nn = new (C, 2) DecodeNNode(nn, t); 2072 n->set_req(AddPNode::Base, nn); 2073 n->set_req(AddPNode::Address, nn); 2074 if (addp->outcnt() == 0) { 2075 addp->disconnect_inputs(NULL); 2076 } 2077 } 2078 } 2079 } 2080 #endif 2081 break; 2082 } 2083 2084 #ifdef _LP64 2085 case Op_CastPP: 2086 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) { 2087 Compile* C = Compile::current(); 2088 Node* in1 = n->in(1); 2089 const Type* t = n->bottom_type(); 2090 Node* new_in1 = in1->clone(); 2091 new_in1->as_DecodeN()->set_type(t); 2092 2093 if (!Matcher::clone_shift_expressions) { 2094 // 2095 // x86, ARM and friends can handle 2 adds in addressing mode 2096 // and Matcher can fold a DecodeN node into address by using 2097 // a narrow oop directly and do implicit NULL check in address: 2098 // 2099 // [R12 + narrow_oop_reg<<3 + offset] 2100 // NullCheck narrow_oop_reg 2101 // 2102 // On other platforms (Sparc) we have to keep new DecodeN node and 2103 // use it to do implicit NULL check in address: 2104 // 2105 // decode_not_null narrow_oop_reg, base_reg 2106 // [base_reg + offset] 2107 // NullCheck base_reg 2108 // 2109 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2110 // to keep the information to which NULL check the new DecodeN node 2111 // corresponds to use it as value in implicit_null_check(). 2112 // 2113 new_in1->set_req(0, n->in(0)); 2114 } 2115 2116 n->subsume_by(new_in1); 2117 if (in1->outcnt() == 0) { 2118 in1->disconnect_inputs(NULL); 2119 } 2120 } 2121 break; 2122 2123 case Op_CmpP: 2124 // Do this transformation here to preserve CmpPNode::sub() and 2125 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2126 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2127 Node* in1 = n->in(1); 2128 Node* in2 = n->in(2); 2129 if (!in1->is_DecodeN()) { 2130 in2 = in1; 2131 in1 = n->in(2); 2132 } 2133 assert(in1->is_DecodeN(), "sanity"); 2134 2135 Compile* C = Compile::current(); 2136 Node* new_in2 = NULL; 2137 if (in2->is_DecodeN()) { 2138 new_in2 = in2->in(1); 2139 } else if (in2->Opcode() == Op_ConP) { 2140 const Type* t = in2->bottom_type(); 2141 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) { 2142 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2143 // 2144 // This transformation together with CastPP transformation above 2145 // will generated code for implicit NULL checks for compressed oops. 2146 // 2147 // The original code after Optimize() 2148 // 2149 // LoadN memory, narrow_oop_reg 2150 // decode narrow_oop_reg, base_reg 2151 // CmpP base_reg, NULL 2152 // CastPP base_reg // NotNull 2153 // Load [base_reg + offset], val_reg 2154 // 2155 // after these transformations will be 2156 // 2157 // LoadN memory, narrow_oop_reg 2158 // CmpN narrow_oop_reg, NULL 2159 // decode_not_null narrow_oop_reg, base_reg 2160 // Load [base_reg + offset], val_reg 2161 // 2162 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2163 // since narrow oops can be used in debug info now (see the code in 2164 // final_graph_reshaping_walk()). 2165 // 2166 // At the end the code will be matched to 2167 // on x86: 2168 // 2169 // Load_narrow_oop memory, narrow_oop_reg 2170 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2171 // NullCheck narrow_oop_reg 2172 // 2173 // and on sparc: 2174 // 2175 // Load_narrow_oop memory, narrow_oop_reg 2176 // decode_not_null narrow_oop_reg, base_reg 2177 // Load [base_reg + offset], val_reg 2178 // NullCheck base_reg 2179 // 2180 } else if (t->isa_oopptr()) { 2181 new_in2 = ConNode::make(C, t->make_narrowoop()); 2182 } 2183 } 2184 if (new_in2 != NULL) { 2185 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2186 n->subsume_by( cmpN ); 2187 if (in1->outcnt() == 0) { 2188 in1->disconnect_inputs(NULL); 2189 } 2190 if (in2->outcnt() == 0) { 2191 in2->disconnect_inputs(NULL); 2192 } 2193 } 2194 } 2195 break; 2196 2197 case Op_DecodeN: 2198 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2199 // DecodeN could be pinned on Sparc where it can't be fold into 2200 // an address expression, see the code for Op_CastPP above. 2201 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc"); 2202 break; 2203 2204 case Op_EncodeP: { 2205 Node* in1 = n->in(1); 2206 if (in1->is_DecodeN()) { 2207 n->subsume_by(in1->in(1)); 2208 } else if (in1->Opcode() == Op_ConP) { 2209 Compile* C = Compile::current(); 2210 const Type* t = in1->bottom_type(); 2211 if (t == TypePtr::NULL_PTR) { 2212 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2213 } else if (t->isa_oopptr()) { 2214 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2215 } 2216 } 2217 if (in1->outcnt() == 0) { 2218 in1->disconnect_inputs(NULL); 2219 } 2220 break; 2221 } 2222 2223 case Op_Phi: 2224 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2225 // The EncodeP optimization may create Phi with the same edges 2226 // for all paths. It is not handled well by Register Allocator. 2227 Node* unique_in = n->in(1); 2228 assert(unique_in != NULL, ""); 2229 uint cnt = n->req(); 2230 for (uint i = 2; i < cnt; i++) { 2231 Node* m = n->in(i); 2232 assert(m != NULL, ""); 2233 if (unique_in != m) 2234 unique_in = NULL; 2235 } 2236 if (unique_in != NULL) { 2237 n->subsume_by(unique_in); 2238 } 2239 } 2240 break; 2241 2242 #endif 2243 2244 case Op_ModI: 2245 if (UseDivMod) { 2246 // Check if a%b and a/b both exist 2247 Node* d = n->find_similar(Op_DivI); 2248 if (d) { 2249 // Replace them with a fused divmod if supported 2250 Compile* C = Compile::current(); 2251 if (Matcher::has_match_rule(Op_DivModI)) { 2252 DivModINode* divmod = DivModINode::make(C, n); 2253 d->subsume_by(divmod->div_proj()); 2254 n->subsume_by(divmod->mod_proj()); 2255 } else { 2256 // replace a%b with a-((a/b)*b) 2257 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2258 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2259 n->subsume_by( sub ); 2260 } 2261 } 2262 } 2263 break; 2264 2265 case Op_ModL: 2266 if (UseDivMod) { 2267 // Check if a%b and a/b both exist 2268 Node* d = n->find_similar(Op_DivL); 2269 if (d) { 2270 // Replace them with a fused divmod if supported 2271 Compile* C = Compile::current(); 2272 if (Matcher::has_match_rule(Op_DivModL)) { 2273 DivModLNode* divmod = DivModLNode::make(C, n); 2274 d->subsume_by(divmod->div_proj()); 2275 n->subsume_by(divmod->mod_proj()); 2276 } else { 2277 // replace a%b with a-((a/b)*b) 2278 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2279 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2280 n->subsume_by( sub ); 2281 } 2282 } 2283 } 2284 break; 2285 2286 case Op_Load16B: 2287 case Op_Load8B: 2288 case Op_Load4B: 2289 case Op_Load8S: 2290 case Op_Load4S: 2291 case Op_Load2S: 2292 case Op_Load8C: 2293 case Op_Load4C: 2294 case Op_Load2C: 2295 case Op_Load4I: 2296 case Op_Load2I: 2297 case Op_Load2L: 2298 case Op_Load4F: 2299 case Op_Load2F: 2300 case Op_Load2D: 2301 case Op_Store16B: 2302 case Op_Store8B: 2303 case Op_Store4B: 2304 case Op_Store8C: 2305 case Op_Store4C: 2306 case Op_Store2C: 2307 case Op_Store4I: 2308 case Op_Store2I: 2309 case Op_Store2L: 2310 case Op_Store4F: 2311 case Op_Store2F: 2312 case Op_Store2D: 2313 break; 2314 2315 case Op_PackB: 2316 case Op_PackS: 2317 case Op_PackC: 2318 case Op_PackI: 2319 case Op_PackF: 2320 case Op_PackL: 2321 case Op_PackD: 2322 if (n->req()-1 > 2) { 2323 // Replace many operand PackNodes with a binary tree for matching 2324 PackNode* p = (PackNode*) n; 2325 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2326 n->subsume_by(btp); 2327 } 2328 break; 2329 default: 2330 assert( !n->is_Call(), "" ); 2331 assert( !n->is_Mem(), "" ); 2332 break; 2333 } 2334 2335 // Collect CFG split points 2336 if (n->is_MultiBranch()) 2337 fpu._tests.push(n); 2338 } 2339 2340 //------------------------------final_graph_reshaping_walk--------------------- 2341 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2342 // requires that the walk visits a node's inputs before visiting the node. 2343 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { 2344 ResourceArea *area = Thread::current()->resource_area(); 2345 Unique_Node_List sfpt(area); 2346 2347 fpu._visited.set(root->_idx); // first, mark node as visited 2348 uint cnt = root->req(); 2349 Node *n = root; 2350 uint i = 0; 2351 while (true) { 2352 if (i < cnt) { 2353 // Place all non-visited non-null inputs onto stack 2354 Node* m = n->in(i); 2355 ++i; 2356 if (m != NULL && !fpu._visited.test_set(m->_idx)) { 2357 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2358 sfpt.push(m); 2359 cnt = m->req(); 2360 nstack.push(n, i); // put on stack parent and next input's index 2361 n = m; 2362 i = 0; 2363 } 2364 } else { 2365 // Now do post-visit work 2366 final_graph_reshaping_impl( n, fpu ); 2367 if (nstack.is_empty()) 2368 break; // finished 2369 n = nstack.node(); // Get node from stack 2370 cnt = n->req(); 2371 i = nstack.index(); 2372 nstack.pop(); // Shift to the next node on stack 2373 } 2374 } 2375 2376 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2377 // It could be done for an uncommon traps or any safepoints/calls 2378 // if the DecodeN node is referenced only in a debug info. 2379 while (sfpt.size() > 0) { 2380 n = sfpt.pop(); 2381 JVMState *jvms = n->as_SafePoint()->jvms(); 2382 assert(jvms != NULL, "sanity"); 2383 int start = jvms->debug_start(); 2384 int end = n->req(); 2385 bool is_uncommon = (n->is_CallStaticJava() && 2386 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2387 for (int j = start; j < end; j++) { 2388 Node* in = n->in(j); 2389 if (in->is_DecodeN()) { 2390 bool safe_to_skip = true; 2391 if (!is_uncommon ) { 2392 // Is it safe to skip? 2393 for (uint i = 0; i < in->outcnt(); i++) { 2394 Node* u = in->raw_out(i); 2395 if (!u->is_SafePoint() || 2396 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2397 safe_to_skip = false; 2398 } 2399 } 2400 } 2401 if (safe_to_skip) { 2402 n->set_req(j, in->in(1)); 2403 } 2404 if (in->outcnt() == 0) { 2405 in->disconnect_inputs(NULL); 2406 } 2407 } 2408 } 2409 } 2410 } 2411 2412 //------------------------------final_graph_reshaping-------------------------- 2413 // Final Graph Reshaping. 2414 // 2415 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2416 // and not commoned up and forced early. Must come after regular 2417 // optimizations to avoid GVN undoing the cloning. Clone constant 2418 // inputs to Loop Phis; these will be split by the allocator anyways. 2419 // Remove Opaque nodes. 2420 // (2) Move last-uses by commutative operations to the left input to encourage 2421 // Intel update-in-place two-address operations and better register usage 2422 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2423 // calls canonicalizing them back. 2424 // (3) Count the number of double-precision FP ops, single-precision FP ops 2425 // and call sites. On Intel, we can get correct rounding either by 2426 // forcing singles to memory (requires extra stores and loads after each 2427 // FP bytecode) or we can set a rounding mode bit (requires setting and 2428 // clearing the mode bit around call sites). The mode bit is only used 2429 // if the relative frequency of single FP ops to calls is low enough. 2430 // This is a key transform for SPEC mpeg_audio. 2431 // (4) Detect infinite loops; blobs of code reachable from above but not 2432 // below. Several of the Code_Gen algorithms fail on such code shapes, 2433 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2434 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2435 // Detection is by looking for IfNodes where only 1 projection is 2436 // reachable from below or CatchNodes missing some targets. 2437 // (5) Assert for insane oop offsets in debug mode. 2438 2439 bool Compile::final_graph_reshaping() { 2440 // an infinite loop may have been eliminated by the optimizer, 2441 // in which case the graph will be empty. 2442 if (root()->req() == 1) { 2443 record_method_not_compilable("trivial infinite loop"); 2444 return true; 2445 } 2446 2447 Final_Reshape_Counts fpu; 2448 2449 // Visit everybody reachable! 2450 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2451 Node_Stack nstack(unique() >> 1); 2452 final_graph_reshaping_walk(nstack, root(), fpu); 2453 2454 // Check for unreachable (from below) code (i.e., infinite loops). 2455 for( uint i = 0; i < fpu._tests.size(); i++ ) { 2456 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch(); 2457 // Get number of CFG targets. 2458 // Note that PCTables include exception targets after calls. 2459 uint required_outcnt = n->required_outcnt(); 2460 if (n->outcnt() != required_outcnt) { 2461 // Check for a few special cases. Rethrow Nodes never take the 2462 // 'fall-thru' path, so expected kids is 1 less. 2463 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2464 if (n->in(0)->in(0)->is_Call()) { 2465 CallNode *call = n->in(0)->in(0)->as_Call(); 2466 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2467 required_outcnt--; // Rethrow always has 1 less kid 2468 } else if (call->req() > TypeFunc::Parms && 2469 call->is_CallDynamicJava()) { 2470 // Check for null receiver. In such case, the optimizer has 2471 // detected that the virtual call will always result in a null 2472 // pointer exception. The fall-through projection of this CatchNode 2473 // will not be populated. 2474 Node *arg0 = call->in(TypeFunc::Parms); 2475 if (arg0->is_Type() && 2476 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2477 required_outcnt--; 2478 } 2479 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2480 call->req() > TypeFunc::Parms+1 && 2481 call->is_CallStaticJava()) { 2482 // Check for negative array length. In such case, the optimizer has 2483 // detected that the allocation attempt will always result in an 2484 // exception. There is no fall-through projection of this CatchNode . 2485 Node *arg1 = call->in(TypeFunc::Parms+1); 2486 if (arg1->is_Type() && 2487 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2488 required_outcnt--; 2489 } 2490 } 2491 } 2492 } 2493 // Recheck with a better notion of 'required_outcnt' 2494 if (n->outcnt() != required_outcnt) { 2495 record_method_not_compilable("malformed control flow"); 2496 return true; // Not all targets reachable! 2497 } 2498 } 2499 // Check that I actually visited all kids. Unreached kids 2500 // must be infinite loops. 2501 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2502 if (!fpu._visited.test(n->fast_out(j)->_idx)) { 2503 record_method_not_compilable("infinite loop"); 2504 return true; // Found unvisited kid; must be unreach 2505 } 2506 } 2507 2508 // If original bytecodes contained a mixture of floats and doubles 2509 // check if the optimizer has made it homogenous, item (3). 2510 if( Use24BitFPMode && Use24BitFP && 2511 fpu.get_float_count() > 32 && 2512 fpu.get_double_count() == 0 && 2513 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { 2514 set_24_bit_selection_and_mode( false, true ); 2515 } 2516 2517 set_has_java_calls(fpu.get_java_call_count() > 0); 2518 2519 // No infinite loops, no reason to bail out. 2520 return false; 2521 } 2522 2523 //-----------------------------too_many_traps---------------------------------- 2524 // Report if there are too many traps at the current method and bci. 2525 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2526 bool Compile::too_many_traps(ciMethod* method, 2527 int bci, 2528 Deoptimization::DeoptReason reason) { 2529 ciMethodData* md = method->method_data(); 2530 if (md->is_empty()) { 2531 // Assume the trap has not occurred, or that it occurred only 2532 // because of a transient condition during start-up in the interpreter. 2533 return false; 2534 } 2535 if (md->has_trap_at(bci, reason) != 0) { 2536 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2537 // Also, if there are multiple reasons, or if there is no per-BCI record, 2538 // assume the worst. 2539 if (log()) 2540 log()->elem("observe trap='%s' count='%d'", 2541 Deoptimization::trap_reason_name(reason), 2542 md->trap_count(reason)); 2543 return true; 2544 } else { 2545 // Ignore method/bci and see if there have been too many globally. 2546 return too_many_traps(reason, md); 2547 } 2548 } 2549 2550 // Less-accurate variant which does not require a method and bci. 2551 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2552 ciMethodData* logmd) { 2553 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2554 // Too many traps globally. 2555 // Note that we use cumulative trap_count, not just md->trap_count. 2556 if (log()) { 2557 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2558 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2559 Deoptimization::trap_reason_name(reason), 2560 mcount, trap_count(reason)); 2561 } 2562 return true; 2563 } else { 2564 // The coast is clear. 2565 return false; 2566 } 2567 } 2568 2569 //--------------------------too_many_recompiles-------------------------------- 2570 // Report if there are too many recompiles at the current method and bci. 2571 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2572 // Is not eager to return true, since this will cause the compiler to use 2573 // Action_none for a trap point, to avoid too many recompilations. 2574 bool Compile::too_many_recompiles(ciMethod* method, 2575 int bci, 2576 Deoptimization::DeoptReason reason) { 2577 ciMethodData* md = method->method_data(); 2578 if (md->is_empty()) { 2579 // Assume the trap has not occurred, or that it occurred only 2580 // because of a transient condition during start-up in the interpreter. 2581 return false; 2582 } 2583 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2584 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2585 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2586 Deoptimization::DeoptReason per_bc_reason 2587 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2588 if ((per_bc_reason == Deoptimization::Reason_none 2589 || md->has_trap_at(bci, reason) != 0) 2590 // The trap frequency measure we care about is the recompile count: 2591 && md->trap_recompiled_at(bci) 2592 && md->overflow_recompile_count() >= bc_cutoff) { 2593 // Do not emit a trap here if it has already caused recompilations. 2594 // Also, if there are multiple reasons, or if there is no per-BCI record, 2595 // assume the worst. 2596 if (log()) 2597 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2598 Deoptimization::trap_reason_name(reason), 2599 md->trap_count(reason), 2600 md->overflow_recompile_count()); 2601 return true; 2602 } else if (trap_count(reason) != 0 2603 && decompile_count() >= m_cutoff) { 2604 // Too many recompiles globally, and we have seen this sort of trap. 2605 // Use cumulative decompile_count, not just md->decompile_count. 2606 if (log()) 2607 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2608 Deoptimization::trap_reason_name(reason), 2609 md->trap_count(reason), trap_count(reason), 2610 md->decompile_count(), decompile_count()); 2611 return true; 2612 } else { 2613 // The coast is clear. 2614 return false; 2615 } 2616 } 2617 2618 2619 #ifndef PRODUCT 2620 //------------------------------verify_graph_edges--------------------------- 2621 // Walk the Graph and verify that there is a one-to-one correspondence 2622 // between Use-Def edges and Def-Use edges in the graph. 2623 void Compile::verify_graph_edges(bool no_dead_code) { 2624 if (VerifyGraphEdges) { 2625 ResourceArea *area = Thread::current()->resource_area(); 2626 Unique_Node_List visited(area); 2627 // Call recursive graph walk to check edges 2628 _root->verify_edges(visited); 2629 if (no_dead_code) { 2630 // Now make sure that no visited node is used by an unvisited node. 2631 bool dead_nodes = 0; 2632 Unique_Node_List checked(area); 2633 while (visited.size() > 0) { 2634 Node* n = visited.pop(); 2635 checked.push(n); 2636 for (uint i = 0; i < n->outcnt(); i++) { 2637 Node* use = n->raw_out(i); 2638 if (checked.member(use)) continue; // already checked 2639 if (visited.member(use)) continue; // already in the graph 2640 if (use->is_Con()) continue; // a dead ConNode is OK 2641 // At this point, we have found a dead node which is DU-reachable. 2642 if (dead_nodes++ == 0) 2643 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2644 use->dump(2); 2645 tty->print_cr("---"); 2646 checked.push(use); // No repeats; pretend it is now checked. 2647 } 2648 } 2649 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2650 } 2651 } 2652 } 2653 #endif 2654 2655 // The Compile object keeps track of failure reasons separately from the ciEnv. 2656 // This is required because there is not quite a 1-1 relation between the 2657 // ciEnv and its compilation task and the Compile object. Note that one 2658 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2659 // to backtrack and retry without subsuming loads. Other than this backtracking 2660 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2661 // by the logic in C2Compiler. 2662 void Compile::record_failure(const char* reason) { 2663 if (log() != NULL) { 2664 log()->elem("failure reason='%s' phase='compile'", reason); 2665 } 2666 if (_failure_reason == NULL) { 2667 // Record the first failure reason. 2668 _failure_reason = reason; 2669 } 2670 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2671 C->print_method(_failure_reason); 2672 } 2673 _root = NULL; // flush the graph, too 2674 } 2675 2676 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2677 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2678 { 2679 if (dolog) { 2680 C = Compile::current(); 2681 _log = C->log(); 2682 } else { 2683 C = NULL; 2684 _log = NULL; 2685 } 2686 if (_log != NULL) { 2687 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2688 _log->stamp(); 2689 _log->end_head(); 2690 } 2691 } 2692 2693 Compile::TracePhase::~TracePhase() { 2694 if (_log != NULL) { 2695 _log->done("phase nodes='%d'", C->unique()); 2696 } 2697 }