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 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 106 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 107 if (cg != NULL) { 108 // Save it for next time: 109 register_intrinsic(cg); 110 return cg; 111 } else { 112 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 113 } 114 } 115 return NULL; 116 } 117 118 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 119 // in library_call.cpp. 120 121 122 #ifndef PRODUCT 123 // statistics gathering... 124 125 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 126 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 127 128 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 129 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 130 int oflags = _intrinsic_hist_flags[id]; 131 assert(flags != 0, "what happened?"); 132 if (is_virtual) { 133 flags |= _intrinsic_virtual; 134 } 135 bool changed = (flags != oflags); 136 if ((flags & _intrinsic_worked) != 0) { 137 juint count = (_intrinsic_hist_count[id] += 1); 138 if (count == 1) { 139 changed = true; // first time 140 } 141 // increment the overall count also: 142 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 143 } 144 if (changed) { 145 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 146 // Something changed about the intrinsic's virtuality. 147 if ((flags & _intrinsic_virtual) != 0) { 148 // This is the first use of this intrinsic as a virtual call. 149 if (oflags != 0) { 150 // We already saw it as a non-virtual, so note both cases. 151 flags |= _intrinsic_both; 152 } 153 } else if ((oflags & _intrinsic_both) == 0) { 154 // This is the first use of this intrinsic as a non-virtual 155 flags |= _intrinsic_both; 156 } 157 } 158 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 159 } 160 // update the overall flags also: 161 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 162 return changed; 163 } 164 165 static char* format_flags(int flags, char* buf) { 166 buf[0] = 0; 167 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 168 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 169 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 170 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 171 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 172 if (buf[0] == 0) strcat(buf, ","); 173 assert(buf[0] == ',', "must be"); 174 return &buf[1]; 175 } 176 177 void Compile::print_intrinsic_statistics() { 178 char flagsbuf[100]; 179 ttyLocker ttyl; 180 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 181 tty->print_cr("Compiler intrinsic usage:"); 182 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 183 if (total == 0) total = 1; // avoid div0 in case of no successes 184 #define PRINT_STAT_LINE(name, c, f) \ 185 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 186 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 187 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 188 int flags = _intrinsic_hist_flags[id]; 189 juint count = _intrinsic_hist_count[id]; 190 if ((flags | count) != 0) { 191 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 192 } 193 } 194 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 195 if (xtty != NULL) xtty->tail("statistics"); 196 } 197 198 void Compile::print_statistics() { 199 { ttyLocker ttyl; 200 if (xtty != NULL) xtty->head("statistics type='opto'"); 201 Parse::print_statistics(); 202 PhaseCCP::print_statistics(); 203 PhaseRegAlloc::print_statistics(); 204 Scheduling::print_statistics(); 205 PhasePeephole::print_statistics(); 206 PhaseIdealLoop::print_statistics(); 207 if (xtty != NULL) xtty->tail("statistics"); 208 } 209 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 210 // put this under its own <statistics> element. 211 print_intrinsic_statistics(); 212 } 213 } 214 #endif //PRODUCT 215 216 // Support for bundling info 217 Bundle* Compile::node_bundling(const Node *n) { 218 assert(valid_bundle_info(n), "oob"); 219 return &_node_bundling_base[n->_idx]; 220 } 221 222 bool Compile::valid_bundle_info(const Node *n) { 223 return (_node_bundling_limit > n->_idx); 224 } 225 226 227 void Compile::gvn_replace_by(Node* n, Node* nn) { 228 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 229 Node* use = n->last_out(i); 230 bool is_in_table = initial_gvn()->hash_delete(use); 231 uint uses_found = 0; 232 for (uint j = 0; j < use->len(); j++) { 233 if (use->in(j) == n) { 234 if (j < use->req()) 235 use->set_req(j, nn); 236 else 237 use->set_prec(j, nn); 238 uses_found++; 239 } 240 } 241 if (is_in_table) { 242 // reinsert into table 243 initial_gvn()->hash_find_insert(use); 244 } 245 record_for_igvn(use); 246 i -= uses_found; // we deleted 1 or more copies of this edge 247 } 248 } 249 250 251 252 253 // Identify all nodes that are reachable from below, useful. 254 // Use breadth-first pass that records state in a Unique_Node_List, 255 // recursive traversal is slower. 256 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 257 int estimated_worklist_size = unique(); 258 useful.map( estimated_worklist_size, NULL ); // preallocate space 259 260 // Initialize worklist 261 if (root() != NULL) { useful.push(root()); } 262 // If 'top' is cached, declare it useful to preserve cached node 263 if( cached_top_node() ) { useful.push(cached_top_node()); } 264 265 // Push all useful nodes onto the list, breadthfirst 266 for( uint next = 0; next < useful.size(); ++next ) { 267 assert( next < unique(), "Unique useful nodes < total nodes"); 268 Node *n = useful.at(next); 269 uint max = n->len(); 270 for( uint i = 0; i < max; ++i ) { 271 Node *m = n->in(i); 272 if( m == NULL ) continue; 273 useful.push(m); 274 } 275 } 276 } 277 278 // Disconnect all useless nodes by disconnecting those at the boundary. 279 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 280 uint next = 0; 281 while( next < useful.size() ) { 282 Node *n = useful.at(next++); 283 // Use raw traversal of out edges since this code removes out edges 284 int max = n->outcnt(); 285 for (int j = 0; j < max; ++j ) { 286 Node* child = n->raw_out(j); 287 if( ! useful.member(child) ) { 288 assert( !child->is_top() || child != top(), 289 "If top is cached in Compile object it is in useful list"); 290 // Only need to remove this out-edge to the useless node 291 n->raw_del_out(j); 292 --j; 293 --max; 294 } 295 } 296 if (n->outcnt() == 1 && n->has_special_unique_user()) { 297 record_for_igvn( n->unique_out() ); 298 } 299 } 300 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 301 } 302 303 //------------------------------frame_size_in_words----------------------------- 304 // frame_slots in units of words 305 int Compile::frame_size_in_words() const { 306 // shift is 0 in LP32 and 1 in LP64 307 const int shift = (LogBytesPerWord - LogBytesPerInt); 308 int words = _frame_slots >> shift; 309 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 310 return words; 311 } 312 313 // ============================================================================ 314 //------------------------------CompileWrapper--------------------------------- 315 class CompileWrapper : public StackObj { 316 Compile *const _compile; 317 public: 318 CompileWrapper(Compile* compile); 319 320 ~CompileWrapper(); 321 }; 322 323 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 324 // the Compile* pointer is stored in the current ciEnv: 325 ciEnv* env = compile->env(); 326 assert(env == ciEnv::current(), "must already be a ciEnv active"); 327 assert(env->compiler_data() == NULL, "compile already active?"); 328 env->set_compiler_data(compile); 329 assert(compile == Compile::current(), "sanity"); 330 331 compile->set_type_dict(NULL); 332 compile->set_type_hwm(NULL); 333 compile->set_type_last_size(0); 334 compile->set_last_tf(NULL, NULL); 335 compile->set_indexSet_arena(NULL); 336 compile->set_indexSet_free_block_list(NULL); 337 compile->init_type_arena(); 338 Type::Initialize(compile); 339 _compile->set_scratch_buffer_blob(NULL); 340 _compile->begin_method(); 341 } 342 CompileWrapper::~CompileWrapper() { 343 _compile->end_method(); 344 if (_compile->scratch_buffer_blob() != NULL) 345 BufferBlob::free(_compile->scratch_buffer_blob()); 346 _compile->env()->set_compiler_data(NULL); 347 } 348 349 350 //----------------------------print_compile_messages--------------------------- 351 void Compile::print_compile_messages() { 352 #ifndef PRODUCT 353 // Check if recompiling 354 if (_subsume_loads == false && PrintOpto) { 355 // Recompiling without allowing machine instructions to subsume loads 356 tty->print_cr("*********************************************************"); 357 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 358 tty->print_cr("*********************************************************"); 359 } 360 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 361 // Recompiling without escape analysis 362 tty->print_cr("*********************************************************"); 363 tty->print_cr("** Bailout: Recompile without escape analysis **"); 364 tty->print_cr("*********************************************************"); 365 } 366 if (env()->break_at_compile()) { 367 // Open the debugger when compiling this method. 368 tty->print("### Breaking when compiling: "); 369 method()->print_short_name(); 370 tty->cr(); 371 BREAKPOINT; 372 } 373 374 if( PrintOpto ) { 375 if (is_osr_compilation()) { 376 tty->print("[OSR]%3d", _compile_id); 377 } else { 378 tty->print("%3d", _compile_id); 379 } 380 } 381 #endif 382 } 383 384 385 void Compile::init_scratch_buffer_blob() { 386 if( scratch_buffer_blob() != NULL ) return; 387 388 // Construct a temporary CodeBuffer to have it construct a BufferBlob 389 // Cache this BufferBlob for this compile. 390 ResourceMark rm; 391 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 392 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 393 // Record the buffer blob for next time. 394 set_scratch_buffer_blob(blob); 395 // Have we run out of code space? 396 if (scratch_buffer_blob() == NULL) { 397 // Let CompilerBroker disable further compilations. 398 record_failure("Not enough space for scratch buffer in CodeCache"); 399 return; 400 } 401 402 // Initialize the relocation buffers 403 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 404 set_scratch_locs_memory(locs_buf); 405 } 406 407 408 //-----------------------scratch_emit_size------------------------------------- 409 // Helper function that computes size by emitting code 410 uint Compile::scratch_emit_size(const Node* n) { 411 // Emit into a trash buffer and count bytes emitted. 412 // This is a pretty expensive way to compute a size, 413 // but it works well enough if seldom used. 414 // All common fixed-size instructions are given a size 415 // method by the AD file. 416 // Note that the scratch buffer blob and locs memory are 417 // allocated at the beginning of the compile task, and 418 // may be shared by several calls to scratch_emit_size. 419 // The allocation of the scratch buffer blob is particularly 420 // expensive, since it has to grab the code cache lock. 421 BufferBlob* blob = this->scratch_buffer_blob(); 422 assert(blob != NULL, "Initialize BufferBlob at start"); 423 assert(blob->size() > MAX_inst_size, "sanity"); 424 relocInfo* locs_buf = scratch_locs_memory(); 425 address blob_begin = blob->instructions_begin(); 426 address blob_end = (address)locs_buf; 427 assert(blob->instructions_contains(blob_end), "sanity"); 428 CodeBuffer buf(blob_begin, blob_end - blob_begin); 429 buf.initialize_consts_size(MAX_const_size); 430 buf.initialize_stubs_size(MAX_stubs_size); 431 assert(locs_buf != NULL, "sanity"); 432 int lsize = MAX_locs_size / 2; 433 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 434 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 435 n->emit(buf, this->regalloc()); 436 return buf.code_size(); 437 } 438 439 440 // ============================================================================ 441 //------------------------------Compile standard------------------------------- 442 debug_only( int Compile::_debug_idx = 100000; ) 443 444 // Compile a method. entry_bci is -1 for normal compilations and indicates 445 // the continuation bci for on stack replacement. 446 447 448 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 449 : Phase(Compiler), 450 _env(ci_env), 451 _log(ci_env->log()), 452 _compile_id(ci_env->compile_id()), 453 _save_argument_registers(false), 454 _stub_name(NULL), 455 _stub_function(NULL), 456 _stub_entry_point(NULL), 457 _method(target), 458 _entry_bci(osr_bci), 459 _initial_gvn(NULL), 460 _for_igvn(NULL), 461 _warm_calls(NULL), 462 _subsume_loads(subsume_loads), 463 _do_escape_analysis(do_escape_analysis), 464 _failure_reason(NULL), 465 _code_buffer("Compile::Fill_buffer"), 466 _orig_pc_slot(0), 467 _orig_pc_slot_offset_in_bytes(0), 468 _node_bundling_limit(0), 469 _node_bundling_base(NULL), 470 _java_calls(0), 471 _inner_loops(0), 472 #ifndef PRODUCT 473 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 474 _printer(IdealGraphPrinter::printer()), 475 #endif 476 _congraph(NULL) { 477 C = this; 478 479 CompileWrapper cw(this); 480 #ifndef PRODUCT 481 if (TimeCompiler2) { 482 tty->print(" "); 483 target->holder()->name()->print(); 484 tty->print("."); 485 target->print_short_name(); 486 tty->print(" "); 487 } 488 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 489 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 490 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 491 if (!print_opto_assembly) { 492 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 493 if (print_assembly && !Disassembler::can_decode()) { 494 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 495 print_opto_assembly = true; 496 } 497 } 498 set_print_assembly(print_opto_assembly); 499 set_parsed_irreducible_loop(false); 500 #endif 501 502 if (ProfileTraps) { 503 // Make sure the method being compiled gets its own MDO, 504 // so we can at least track the decompile_count(). 505 method()->build_method_data(); 506 } 507 508 Init(::AliasLevel); 509 510 511 print_compile_messages(); 512 513 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 514 _ilt = InlineTree::build_inline_tree_root(); 515 else 516 _ilt = NULL; 517 518 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 519 assert(num_alias_types() >= AliasIdxRaw, ""); 520 521 #define MINIMUM_NODE_HASH 1023 522 // Node list that Iterative GVN will start with 523 Unique_Node_List for_igvn(comp_arena()); 524 set_for_igvn(&for_igvn); 525 526 // GVN that will be run immediately on new nodes 527 uint estimated_size = method()->code_size()*4+64; 528 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 529 PhaseGVN gvn(node_arena(), estimated_size); 530 set_initial_gvn(&gvn); 531 532 { // Scope for timing the parser 533 TracePhase t3("parse", &_t_parser, true); 534 535 // Put top into the hash table ASAP. 536 initial_gvn()->transform_no_reclaim(top()); 537 538 // Set up tf(), start(), and find a CallGenerator. 539 CallGenerator* cg; 540 if (is_osr_compilation()) { 541 const TypeTuple *domain = StartOSRNode::osr_domain(); 542 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 543 init_tf(TypeFunc::make(domain, range)); 544 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 545 initial_gvn()->set_type_bottom(s); 546 init_start(s); 547 cg = CallGenerator::for_osr(method(), entry_bci()); 548 } else { 549 // Normal case. 550 init_tf(TypeFunc::make(method())); 551 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 552 initial_gvn()->set_type_bottom(s); 553 init_start(s); 554 float past_uses = method()->interpreter_invocation_count(); 555 float expected_uses = past_uses; 556 cg = CallGenerator::for_inline(method(), expected_uses); 557 } 558 if (failing()) return; 559 if (cg == NULL) { 560 record_method_not_compilable_all_tiers("cannot parse method"); 561 return; 562 } 563 JVMState* jvms = build_start_state(start(), tf()); 564 if ((jvms = cg->generate(jvms)) == NULL) { 565 record_method_not_compilable("method parse failed"); 566 return; 567 } 568 GraphKit kit(jvms); 569 570 if (!kit.stopped()) { 571 // Accept return values, and transfer control we know not where. 572 // This is done by a special, unique ReturnNode bound to root. 573 return_values(kit.jvms()); 574 } 575 576 if (kit.has_exceptions()) { 577 // Any exceptions that escape from this call must be rethrown 578 // to whatever caller is dynamically above us on the stack. 579 // This is done by a special, unique RethrowNode bound to root. 580 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 581 } 582 583 if (!failing() && has_stringbuilder()) { 584 { 585 // remove useless nodes to make the usage analysis simpler 586 ResourceMark rm; 587 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 588 } 589 590 { 591 ResourceMark rm; 592 print_method("Before StringOpts", 3); 593 PhaseStringOpts pso(initial_gvn(), &for_igvn); 594 print_method("After StringOpts", 3); 595 } 596 597 // now inline anything that we skipped the first time around 598 while (_late_inlines.length() > 0) { 599 CallGenerator* cg = _late_inlines.pop(); 600 cg->do_late_inline(); 601 } 602 } 603 assert(_late_inlines.length() == 0, "should have been processed"); 604 605 print_method("Before RemoveUseless", 3); 606 607 // Remove clutter produced by parsing. 608 if (!failing()) { 609 ResourceMark rm; 610 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 611 } 612 } 613 614 // Note: Large methods are capped off in do_one_bytecode(). 615 if (failing()) return; 616 617 // After parsing, node notes are no longer automagic. 618 // They must be propagated by register_new_node_with_optimizer(), 619 // clone(), or the like. 620 set_default_node_notes(NULL); 621 622 for (;;) { 623 int successes = Inline_Warm(); 624 if (failing()) return; 625 if (successes == 0) break; 626 } 627 628 // Drain the list. 629 Finish_Warm(); 630 #ifndef PRODUCT 631 if (_printer) { 632 _printer->print_inlining(this); 633 } 634 #endif 635 636 if (failing()) return; 637 NOT_PRODUCT( verify_graph_edges(); ) 638 639 // Perform escape analysis 640 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 641 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 642 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction. 643 PhaseGVN* igvn = initial_gvn(); 644 Node* oop_null = igvn->zerocon(T_OBJECT); 645 Node* noop_null = igvn->zerocon(T_NARROWOOP); 646 647 _congraph = new(comp_arena()) ConnectionGraph(this); 648 bool has_non_escaping_obj = _congraph->compute_escape(); 649 650 #ifndef PRODUCT 651 if (PrintEscapeAnalysis) { 652 _congraph->dump(); 653 } 654 #endif 655 // Cleanup. 656 if (oop_null->outcnt() == 0) 657 igvn->hash_delete(oop_null); 658 if (noop_null->outcnt() == 0) 659 igvn->hash_delete(noop_null); 660 661 if (!has_non_escaping_obj) { 662 _congraph = NULL; 663 } 664 665 if (failing()) return; 666 } 667 // Now optimize 668 Optimize(); 669 if (failing()) return; 670 NOT_PRODUCT( verify_graph_edges(); ) 671 672 #ifndef PRODUCT 673 if (PrintIdeal) { 674 ttyLocker ttyl; // keep the following output all in one block 675 // This output goes directly to the tty, not the compiler log. 676 // To enable tools to match it up with the compilation activity, 677 // be sure to tag this tty output with the compile ID. 678 if (xtty != NULL) { 679 xtty->head("ideal compile_id='%d'%s", compile_id(), 680 is_osr_compilation() ? " compile_kind='osr'" : 681 ""); 682 } 683 root()->dump(9999); 684 if (xtty != NULL) { 685 xtty->tail("ideal"); 686 } 687 } 688 #endif 689 690 // Now that we know the size of all the monitors we can add a fixed slot 691 // for the original deopt pc. 692 693 _orig_pc_slot = fixed_slots(); 694 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 695 set_fixed_slots(next_slot); 696 697 // Now generate code 698 Code_Gen(); 699 if (failing()) return; 700 701 // Check if we want to skip execution of all compiled code. 702 { 703 #ifndef PRODUCT 704 if (OptoNoExecute) { 705 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 706 return; 707 } 708 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 709 #endif 710 711 if (is_osr_compilation()) { 712 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 713 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 714 } else { 715 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 716 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 717 } 718 719 env()->register_method(_method, _entry_bci, 720 &_code_offsets, 721 _orig_pc_slot_offset_in_bytes, 722 code_buffer(), 723 frame_size_in_words(), _oop_map_set, 724 &_handler_table, &_inc_table, 725 compiler, 726 env()->comp_level(), 727 true, /*has_debug_info*/ 728 has_unsafe_access() 729 ); 730 } 731 } 732 733 //------------------------------Compile---------------------------------------- 734 // Compile a runtime stub 735 Compile::Compile( ciEnv* ci_env, 736 TypeFunc_generator generator, 737 address stub_function, 738 const char *stub_name, 739 int is_fancy_jump, 740 bool pass_tls, 741 bool save_arg_registers, 742 bool return_pc ) 743 : Phase(Compiler), 744 _env(ci_env), 745 _log(ci_env->log()), 746 _compile_id(-1), 747 _save_argument_registers(save_arg_registers), 748 _method(NULL), 749 _stub_name(stub_name), 750 _stub_function(stub_function), 751 _stub_entry_point(NULL), 752 _entry_bci(InvocationEntryBci), 753 _initial_gvn(NULL), 754 _for_igvn(NULL), 755 _warm_calls(NULL), 756 _orig_pc_slot(0), 757 _orig_pc_slot_offset_in_bytes(0), 758 _subsume_loads(true), 759 _do_escape_analysis(false), 760 _failure_reason(NULL), 761 _code_buffer("Compile::Fill_buffer"), 762 _node_bundling_limit(0), 763 _node_bundling_base(NULL), 764 _java_calls(0), 765 _inner_loops(0), 766 #ifndef PRODUCT 767 _trace_opto_output(TraceOptoOutput), 768 _printer(NULL), 769 #endif 770 _congraph(NULL) { 771 C = this; 772 773 #ifndef PRODUCT 774 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 775 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 776 set_print_assembly(PrintFrameConverterAssembly); 777 set_parsed_irreducible_loop(false); 778 #endif 779 CompileWrapper cw(this); 780 Init(/*AliasLevel=*/ 0); 781 init_tf((*generator)()); 782 783 { 784 // The following is a dummy for the sake of GraphKit::gen_stub 785 Unique_Node_List for_igvn(comp_arena()); 786 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 787 PhaseGVN gvn(Thread::current()->resource_area(),255); 788 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 789 gvn.transform_no_reclaim(top()); 790 791 GraphKit kit; 792 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 793 } 794 795 NOT_PRODUCT( verify_graph_edges(); ) 796 Code_Gen(); 797 if (failing()) return; 798 799 800 // Entry point will be accessed using compile->stub_entry_point(); 801 if (code_buffer() == NULL) { 802 Matcher::soft_match_failure(); 803 } else { 804 if (PrintAssembly && (WizardMode || Verbose)) 805 tty->print_cr("### Stub::%s", stub_name); 806 807 if (!failing()) { 808 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 809 810 // Make the NMethod 811 // For now we mark the frame as never safe for profile stackwalking 812 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 813 code_buffer(), 814 CodeOffsets::frame_never_safe, 815 // _code_offsets.value(CodeOffsets::Frame_Complete), 816 frame_size_in_words(), 817 _oop_map_set, 818 save_arg_registers); 819 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 820 821 _stub_entry_point = rs->entry_point(); 822 } 823 } 824 } 825 826 #ifndef PRODUCT 827 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 828 if(PrintOpto && Verbose) { 829 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 830 } 831 } 832 #endif 833 834 void Compile::print_codes() { 835 } 836 837 //------------------------------Init------------------------------------------- 838 // Prepare for a single compilation 839 void Compile::Init(int aliaslevel) { 840 _unique = 0; 841 _regalloc = NULL; 842 843 _tf = NULL; // filled in later 844 _top = NULL; // cached later 845 _matcher = NULL; // filled in later 846 _cfg = NULL; // filled in later 847 848 set_24_bit_selection_and_mode(Use24BitFP, false); 849 850 _node_note_array = NULL; 851 _default_node_notes = NULL; 852 853 _immutable_memory = NULL; // filled in at first inquiry 854 855 // Globally visible Nodes 856 // First set TOP to NULL to give safe behavior during creation of RootNode 857 set_cached_top_node(NULL); 858 set_root(new (this, 3) RootNode()); 859 // Now that you have a Root to point to, create the real TOP 860 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 861 set_recent_alloc(NULL, NULL); 862 863 // Create Debug Information Recorder to record scopes, oopmaps, etc. 864 env()->set_oop_recorder(new OopRecorder(comp_arena())); 865 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 866 env()->set_dependencies(new Dependencies(env())); 867 868 _fixed_slots = 0; 869 set_has_split_ifs(false); 870 set_has_loops(has_method() && method()->has_loops()); // first approximation 871 set_has_stringbuilder(false); 872 _deopt_happens = true; // start out assuming the worst 873 _trap_can_recompile = false; // no traps emitted yet 874 _major_progress = true; // start out assuming good things will happen 875 set_has_unsafe_access(false); 876 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 877 set_decompile_count(0); 878 879 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 880 // Compilation level related initialization 881 if (env()->comp_level() == CompLevel_fast_compile) { 882 set_num_loop_opts(Tier1LoopOptsCount); 883 set_do_inlining(Tier1Inline != 0); 884 set_max_inline_size(Tier1MaxInlineSize); 885 set_freq_inline_size(Tier1FreqInlineSize); 886 set_do_scheduling(false); 887 set_do_count_invocations(Tier1CountInvocations); 888 set_do_method_data_update(Tier1UpdateMethodData); 889 } else { 890 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 891 set_num_loop_opts(LoopOptsCount); 892 set_do_inlining(Inline); 893 set_max_inline_size(MaxInlineSize); 894 set_freq_inline_size(FreqInlineSize); 895 set_do_scheduling(OptoScheduling); 896 set_do_count_invocations(false); 897 set_do_method_data_update(false); 898 } 899 900 if (debug_info()->recording_non_safepoints()) { 901 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 902 (comp_arena(), 8, 0, NULL)); 903 set_default_node_notes(Node_Notes::make(this)); 904 } 905 906 // // -- Initialize types before each compile -- 907 // // Update cached type information 908 // if( _method && _method->constants() ) 909 // Type::update_loaded_types(_method, _method->constants()); 910 911 // Init alias_type map. 912 if (!_do_escape_analysis && aliaslevel == 3) 913 aliaslevel = 2; // No unique types without escape analysis 914 _AliasLevel = aliaslevel; 915 const int grow_ats = 16; 916 _max_alias_types = grow_ats; 917 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 918 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 919 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 920 { 921 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 922 } 923 // Initialize the first few types. 924 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 925 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 926 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 927 _num_alias_types = AliasIdxRaw+1; 928 // Zero out the alias type cache. 929 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 930 // A NULL adr_type hits in the cache right away. Preload the right answer. 931 probe_alias_cache(NULL)->_index = AliasIdxTop; 932 933 _intrinsics = NULL; 934 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 935 _predicate_opaqs = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 936 register_library_intrinsics(); 937 } 938 939 //---------------------------init_start---------------------------------------- 940 // Install the StartNode on this compile object. 941 void Compile::init_start(StartNode* s) { 942 if (failing()) 943 return; // already failing 944 assert(s == start(), ""); 945 } 946 947 StartNode* Compile::start() const { 948 assert(!failing(), ""); 949 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 950 Node* start = root()->fast_out(i); 951 if( start->is_Start() ) 952 return start->as_Start(); 953 } 954 ShouldNotReachHere(); 955 return NULL; 956 } 957 958 //-------------------------------immutable_memory------------------------------------- 959 // Access immutable memory 960 Node* Compile::immutable_memory() { 961 if (_immutable_memory != NULL) { 962 return _immutable_memory; 963 } 964 StartNode* s = start(); 965 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 966 Node *p = s->fast_out(i); 967 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 968 _immutable_memory = p; 969 return _immutable_memory; 970 } 971 } 972 ShouldNotReachHere(); 973 return NULL; 974 } 975 976 //----------------------set_cached_top_node------------------------------------ 977 // Install the cached top node, and make sure Node::is_top works correctly. 978 void Compile::set_cached_top_node(Node* tn) { 979 if (tn != NULL) verify_top(tn); 980 Node* old_top = _top; 981 _top = tn; 982 // Calling Node::setup_is_top allows the nodes the chance to adjust 983 // their _out arrays. 984 if (_top != NULL) _top->setup_is_top(); 985 if (old_top != NULL) old_top->setup_is_top(); 986 assert(_top == NULL || top()->is_top(), ""); 987 } 988 989 #ifndef PRODUCT 990 void Compile::verify_top(Node* tn) const { 991 if (tn != NULL) { 992 assert(tn->is_Con(), "top node must be a constant"); 993 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 994 assert(tn->in(0) != NULL, "must have live top node"); 995 } 996 } 997 #endif 998 999 1000 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1001 1002 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1003 guarantee(arr != NULL, ""); 1004 int num_blocks = arr->length(); 1005 if (grow_by < num_blocks) grow_by = num_blocks; 1006 int num_notes = grow_by * _node_notes_block_size; 1007 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1008 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1009 while (num_notes > 0) { 1010 arr->append(notes); 1011 notes += _node_notes_block_size; 1012 num_notes -= _node_notes_block_size; 1013 } 1014 assert(num_notes == 0, "exact multiple, please"); 1015 } 1016 1017 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1018 if (source == NULL || dest == NULL) return false; 1019 1020 if (dest->is_Con()) 1021 return false; // Do not push debug info onto constants. 1022 1023 #ifdef ASSERT 1024 // Leave a bread crumb trail pointing to the original node: 1025 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1026 dest->set_debug_orig(source); 1027 } 1028 #endif 1029 1030 if (node_note_array() == NULL) 1031 return false; // Not collecting any notes now. 1032 1033 // This is a copy onto a pre-existing node, which may already have notes. 1034 // If both nodes have notes, do not overwrite any pre-existing notes. 1035 Node_Notes* source_notes = node_notes_at(source->_idx); 1036 if (source_notes == NULL || source_notes->is_clear()) return false; 1037 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1038 if (dest_notes == NULL || dest_notes->is_clear()) { 1039 return set_node_notes_at(dest->_idx, source_notes); 1040 } 1041 1042 Node_Notes merged_notes = (*source_notes); 1043 // The order of operations here ensures that dest notes will win... 1044 merged_notes.update_from(dest_notes); 1045 return set_node_notes_at(dest->_idx, &merged_notes); 1046 } 1047 1048 1049 //--------------------------allow_range_check_smearing------------------------- 1050 // Gating condition for coalescing similar range checks. 1051 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1052 // single covering check that is at least as strong as any of them. 1053 // If the optimization succeeds, the simplified (strengthened) range check 1054 // will always succeed. If it fails, we will deopt, and then give up 1055 // on the optimization. 1056 bool Compile::allow_range_check_smearing() const { 1057 // If this method has already thrown a range-check, 1058 // assume it was because we already tried range smearing 1059 // and it failed. 1060 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1061 return !already_trapped; 1062 } 1063 1064 1065 //------------------------------flatten_alias_type----------------------------- 1066 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1067 int offset = tj->offset(); 1068 TypePtr::PTR ptr = tj->ptr(); 1069 1070 // Known instance (scalarizable allocation) alias only with itself. 1071 bool is_known_inst = tj->isa_oopptr() != NULL && 1072 tj->is_oopptr()->is_known_instance(); 1073 1074 // Process weird unsafe references. 1075 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1076 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1077 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1078 tj = TypeOopPtr::BOTTOM; 1079 ptr = tj->ptr(); 1080 offset = tj->offset(); 1081 } 1082 1083 // Array pointers need some flattening 1084 const TypeAryPtr *ta = tj->isa_aryptr(); 1085 if( ta && is_known_inst ) { 1086 if ( offset != Type::OffsetBot && 1087 offset > arrayOopDesc::length_offset_in_bytes() ) { 1088 offset = Type::OffsetBot; // Flatten constant access into array body only 1089 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1090 } 1091 } else if( ta && _AliasLevel >= 2 ) { 1092 // For arrays indexed by constant indices, we flatten the alias 1093 // space to include all of the array body. Only the header, klass 1094 // and array length can be accessed un-aliased. 1095 if( offset != Type::OffsetBot ) { 1096 if( ta->const_oop() ) { // methodDataOop or methodOop 1097 offset = Type::OffsetBot; // Flatten constant access into array body 1098 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1099 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1100 // range is OK as-is. 1101 tj = ta = TypeAryPtr::RANGE; 1102 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1103 tj = TypeInstPtr::KLASS; // all klass loads look alike 1104 ta = TypeAryPtr::RANGE; // generic ignored junk 1105 ptr = TypePtr::BotPTR; 1106 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1107 tj = TypeInstPtr::MARK; 1108 ta = TypeAryPtr::RANGE; // generic ignored junk 1109 ptr = TypePtr::BotPTR; 1110 } else { // Random constant offset into array body 1111 offset = Type::OffsetBot; // Flatten constant access into array body 1112 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1113 } 1114 } 1115 // Arrays of fixed size alias with arrays of unknown size. 1116 if (ta->size() != TypeInt::POS) { 1117 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1118 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1119 } 1120 // Arrays of known objects become arrays of unknown objects. 1121 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1122 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1123 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1124 } 1125 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1126 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1127 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1128 } 1129 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1130 // cannot be distinguished by bytecode alone. 1131 if (ta->elem() == TypeInt::BOOL) { 1132 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1133 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1134 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1135 } 1136 // During the 2nd round of IterGVN, NotNull castings are removed. 1137 // Make sure the Bottom and NotNull variants alias the same. 1138 // Also, make sure exact and non-exact variants alias the same. 1139 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1140 if (ta->const_oop()) { 1141 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1142 } else { 1143 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1144 } 1145 } 1146 } 1147 1148 // Oop pointers need some flattening 1149 const TypeInstPtr *to = tj->isa_instptr(); 1150 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1151 if( ptr == TypePtr::Constant ) { 1152 // No constant oop pointers (such as Strings); they alias with 1153 // unknown strings. 1154 assert(!is_known_inst, "not scalarizable allocation"); 1155 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1156 } else if( is_known_inst ) { 1157 tj = to; // Keep NotNull and klass_is_exact for instance type 1158 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1159 // During the 2nd round of IterGVN, NotNull castings are removed. 1160 // Make sure the Bottom and NotNull variants alias the same. 1161 // Also, make sure exact and non-exact variants alias the same. 1162 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1163 } 1164 // Canonicalize the holder of this field 1165 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1166 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1167 // First handle header references such as a LoadKlassNode, even if the 1168 // object's klass is unloaded at compile time (4965979). 1169 if (!is_known_inst) { // Do it only for non-instance types 1170 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1171 } 1172 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1173 to = NULL; 1174 tj = TypeOopPtr::BOTTOM; 1175 offset = tj->offset(); 1176 } else { 1177 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1178 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1179 if( is_known_inst ) { 1180 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1181 } else { 1182 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1183 } 1184 } 1185 } 1186 } 1187 1188 // Klass pointers to object array klasses need some flattening 1189 const TypeKlassPtr *tk = tj->isa_klassptr(); 1190 if( tk ) { 1191 // If we are referencing a field within a Klass, we need 1192 // to assume the worst case of an Object. Both exact and 1193 // inexact types must flatten to the same alias class. 1194 // Since the flattened result for a klass is defined to be 1195 // precisely java.lang.Object, use a constant ptr. 1196 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1197 1198 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1199 TypeKlassPtr::OBJECT->klass(), 1200 offset); 1201 } 1202 1203 ciKlass* klass = tk->klass(); 1204 if( klass->is_obj_array_klass() ) { 1205 ciKlass* k = TypeAryPtr::OOPS->klass(); 1206 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1207 k = TypeInstPtr::BOTTOM->klass(); 1208 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1209 } 1210 1211 // Check for precise loads from the primary supertype array and force them 1212 // to the supertype cache alias index. Check for generic array loads from 1213 // the primary supertype array and also force them to the supertype cache 1214 // alias index. Since the same load can reach both, we need to merge 1215 // these 2 disparate memories into the same alias class. Since the 1216 // primary supertype array is read-only, there's no chance of confusion 1217 // where we bypass an array load and an array store. 1218 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1219 if( offset == Type::OffsetBot || 1220 off2 < Klass::primary_super_limit()*wordSize ) { 1221 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1222 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1223 } 1224 } 1225 1226 // Flatten all Raw pointers together. 1227 if (tj->base() == Type::RawPtr) 1228 tj = TypeRawPtr::BOTTOM; 1229 1230 if (tj->base() == Type::AnyPtr) 1231 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1232 1233 // Flatten all to bottom for now 1234 switch( _AliasLevel ) { 1235 case 0: 1236 tj = TypePtr::BOTTOM; 1237 break; 1238 case 1: // Flatten to: oop, static, field or array 1239 switch (tj->base()) { 1240 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1241 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1242 case Type::AryPtr: // do not distinguish arrays at all 1243 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1244 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1245 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1246 default: ShouldNotReachHere(); 1247 } 1248 break; 1249 case 2: // No collapsing at level 2; keep all splits 1250 case 3: // No collapsing at level 3; keep all splits 1251 break; 1252 default: 1253 Unimplemented(); 1254 } 1255 1256 offset = tj->offset(); 1257 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1258 1259 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1260 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1261 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1262 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1263 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1264 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1265 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1266 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1267 assert( tj->ptr() != TypePtr::TopPTR && 1268 tj->ptr() != TypePtr::AnyNull && 1269 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1270 // assert( tj->ptr() != TypePtr::Constant || 1271 // tj->base() == Type::RawPtr || 1272 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1273 1274 return tj; 1275 } 1276 1277 void Compile::AliasType::Init(int i, const TypePtr* at) { 1278 _index = i; 1279 _adr_type = at; 1280 _field = NULL; 1281 _is_rewritable = true; // default 1282 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1283 if (atoop != NULL && atoop->is_known_instance()) { 1284 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1285 _general_index = Compile::current()->get_alias_index(gt); 1286 } else { 1287 _general_index = 0; 1288 } 1289 } 1290 1291 //---------------------------------print_on------------------------------------ 1292 #ifndef PRODUCT 1293 void Compile::AliasType::print_on(outputStream* st) { 1294 if (index() < 10) 1295 st->print("@ <%d> ", index()); 1296 else st->print("@ <%d>", index()); 1297 st->print(is_rewritable() ? " " : " RO"); 1298 int offset = adr_type()->offset(); 1299 if (offset == Type::OffsetBot) 1300 st->print(" +any"); 1301 else st->print(" +%-3d", offset); 1302 st->print(" in "); 1303 adr_type()->dump_on(st); 1304 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1305 if (field() != NULL && tjp) { 1306 if (tjp->klass() != field()->holder() || 1307 tjp->offset() != field()->offset_in_bytes()) { 1308 st->print(" != "); 1309 field()->print(); 1310 st->print(" ***"); 1311 } 1312 } 1313 } 1314 1315 void print_alias_types() { 1316 Compile* C = Compile::current(); 1317 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1318 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1319 C->alias_type(idx)->print_on(tty); 1320 tty->cr(); 1321 } 1322 } 1323 #endif 1324 1325 1326 //----------------------------probe_alias_cache-------------------------------- 1327 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1328 intptr_t key = (intptr_t) adr_type; 1329 key ^= key >> logAliasCacheSize; 1330 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1331 } 1332 1333 1334 //-----------------------------grow_alias_types-------------------------------- 1335 void Compile::grow_alias_types() { 1336 const int old_ats = _max_alias_types; // how many before? 1337 const int new_ats = old_ats; // how many more? 1338 const int grow_ats = old_ats+new_ats; // how many now? 1339 _max_alias_types = grow_ats; 1340 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1341 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1342 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1343 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1344 } 1345 1346 1347 //--------------------------------find_alias_type------------------------------ 1348 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1349 if (_AliasLevel == 0) 1350 return alias_type(AliasIdxBot); 1351 1352 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1353 if (ace->_adr_type == adr_type) { 1354 return alias_type(ace->_index); 1355 } 1356 1357 // Handle special cases. 1358 if (adr_type == NULL) return alias_type(AliasIdxTop); 1359 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1360 1361 // Do it the slow way. 1362 const TypePtr* flat = flatten_alias_type(adr_type); 1363 1364 #ifdef ASSERT 1365 assert(flat == flatten_alias_type(flat), "idempotent"); 1366 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1367 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1368 const TypeOopPtr* foop = flat->is_oopptr(); 1369 // Scalarizable allocations have exact klass always. 1370 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1371 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1372 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1373 } 1374 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1375 #endif 1376 1377 int idx = AliasIdxTop; 1378 for (int i = 0; i < num_alias_types(); i++) { 1379 if (alias_type(i)->adr_type() == flat) { 1380 idx = i; 1381 break; 1382 } 1383 } 1384 1385 if (idx == AliasIdxTop) { 1386 if (no_create) return NULL; 1387 // Grow the array if necessary. 1388 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1389 // Add a new alias type. 1390 idx = _num_alias_types++; 1391 _alias_types[idx]->Init(idx, flat); 1392 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1393 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1394 if (flat->isa_instptr()) { 1395 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1396 && flat->is_instptr()->klass() == env()->Class_klass()) 1397 alias_type(idx)->set_rewritable(false); 1398 } 1399 if (flat->isa_klassptr()) { 1400 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1401 alias_type(idx)->set_rewritable(false); 1402 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1403 alias_type(idx)->set_rewritable(false); 1404 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1405 alias_type(idx)->set_rewritable(false); 1406 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1407 alias_type(idx)->set_rewritable(false); 1408 } 1409 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1410 // but the base pointer type is not distinctive enough to identify 1411 // references into JavaThread.) 1412 1413 // Check for final instance fields. 1414 const TypeInstPtr* tinst = flat->isa_instptr(); 1415 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1416 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1417 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1418 // Set field() and is_rewritable() attributes. 1419 if (field != NULL) alias_type(idx)->set_field(field); 1420 } 1421 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1422 // Check for final static fields. 1423 if (tklass && tklass->klass()->is_instance_klass()) { 1424 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1425 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1426 // Set field() and is_rewritable() attributes. 1427 if (field != NULL) alias_type(idx)->set_field(field); 1428 } 1429 } 1430 1431 // Fill the cache for next time. 1432 ace->_adr_type = adr_type; 1433 ace->_index = idx; 1434 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1435 1436 // Might as well try to fill the cache for the flattened version, too. 1437 AliasCacheEntry* face = probe_alias_cache(flat); 1438 if (face->_adr_type == NULL) { 1439 face->_adr_type = flat; 1440 face->_index = idx; 1441 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1442 } 1443 1444 return alias_type(idx); 1445 } 1446 1447 1448 Compile::AliasType* Compile::alias_type(ciField* field) { 1449 const TypeOopPtr* t; 1450 if (field->is_static()) 1451 t = TypeKlassPtr::make(field->holder()); 1452 else 1453 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1454 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1455 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1456 return atp; 1457 } 1458 1459 1460 //------------------------------have_alias_type-------------------------------- 1461 bool Compile::have_alias_type(const TypePtr* adr_type) { 1462 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1463 if (ace->_adr_type == adr_type) { 1464 return true; 1465 } 1466 1467 // Handle special cases. 1468 if (adr_type == NULL) return true; 1469 if (adr_type == TypePtr::BOTTOM) return true; 1470 1471 return find_alias_type(adr_type, true) != NULL; 1472 } 1473 1474 //-----------------------------must_alias-------------------------------------- 1475 // True if all values of the given address type are in the given alias category. 1476 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1477 if (alias_idx == AliasIdxBot) return true; // the universal category 1478 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1479 if (alias_idx == AliasIdxTop) return false; // the empty category 1480 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1481 1482 // the only remaining possible overlap is identity 1483 int adr_idx = get_alias_index(adr_type); 1484 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1485 assert(adr_idx == alias_idx || 1486 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1487 && adr_type != TypeOopPtr::BOTTOM), 1488 "should not be testing for overlap with an unsafe pointer"); 1489 return adr_idx == alias_idx; 1490 } 1491 1492 //------------------------------can_alias-------------------------------------- 1493 // True if any values of the given address type are in the given alias category. 1494 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1495 if (alias_idx == AliasIdxTop) return false; // the empty category 1496 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1497 if (alias_idx == AliasIdxBot) return true; // the universal category 1498 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1499 1500 // the only remaining possible overlap is identity 1501 int adr_idx = get_alias_index(adr_type); 1502 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1503 return adr_idx == alias_idx; 1504 } 1505 1506 1507 1508 //---------------------------pop_warm_call------------------------------------- 1509 WarmCallInfo* Compile::pop_warm_call() { 1510 WarmCallInfo* wci = _warm_calls; 1511 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1512 return wci; 1513 } 1514 1515 //----------------------------Inline_Warm-------------------------------------- 1516 int Compile::Inline_Warm() { 1517 // If there is room, try to inline some more warm call sites. 1518 // %%% Do a graph index compaction pass when we think we're out of space? 1519 if (!InlineWarmCalls) return 0; 1520 1521 int calls_made_hot = 0; 1522 int room_to_grow = NodeCountInliningCutoff - unique(); 1523 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1524 int amount_grown = 0; 1525 WarmCallInfo* call; 1526 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1527 int est_size = (int)call->size(); 1528 if (est_size > (room_to_grow - amount_grown)) { 1529 // This one won't fit anyway. Get rid of it. 1530 call->make_cold(); 1531 continue; 1532 } 1533 call->make_hot(); 1534 calls_made_hot++; 1535 amount_grown += est_size; 1536 amount_to_grow -= est_size; 1537 } 1538 1539 if (calls_made_hot > 0) set_major_progress(); 1540 return calls_made_hot; 1541 } 1542 1543 1544 //----------------------------Finish_Warm-------------------------------------- 1545 void Compile::Finish_Warm() { 1546 if (!InlineWarmCalls) return; 1547 if (failing()) return; 1548 if (warm_calls() == NULL) return; 1549 1550 // Clean up loose ends, if we are out of space for inlining. 1551 WarmCallInfo* call; 1552 while ((call = pop_warm_call()) != NULL) { 1553 call->make_cold(); 1554 } 1555 } 1556 1557 //---------------------cleanup_loop_predicates----------------------- 1558 // Remove the opaque nodes that protect the predicates so that all unused 1559 // checks and uncommon_traps will be eliminated from the ideal graph 1560 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1561 if (predicate_count()==0 ) return; 1562 for (int i = predicate_count(); i > 0; i--) { 1563 Node * n = predicate_opaque1_node(i-1); 1564 assert(n->Opcode() == Op_Opaque1, "must be"); 1565 igvn.replace_node(n, n->in(1)); 1566 } 1567 assert(predicate_count()==0, "should be clean!"); 1568 igvn.optimize(); 1569 } 1570 1571 //------------------------------Optimize--------------------------------------- 1572 // Given a graph, optimize it. 1573 void Compile::Optimize() { 1574 TracePhase t1("optimizer", &_t_optimizer, true); 1575 1576 #ifndef PRODUCT 1577 if (env()->break_at_compile()) { 1578 BREAKPOINT; 1579 } 1580 1581 #endif 1582 1583 ResourceMark rm; 1584 int loop_opts_cnt; 1585 1586 NOT_PRODUCT( verify_graph_edges(); ) 1587 1588 print_method("After Parsing"); 1589 1590 { 1591 // Iterative Global Value Numbering, including ideal transforms 1592 // Initialize IterGVN with types and values from parse-time GVN 1593 PhaseIterGVN igvn(initial_gvn()); 1594 { 1595 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1596 igvn.optimize(); 1597 } 1598 1599 print_method("Iter GVN 1", 2); 1600 1601 if (failing()) return; 1602 1603 // Loop transforms on the ideal graph. Range Check Elimination, 1604 // peeling, unrolling, etc. 1605 1606 // Set loop opts counter 1607 loop_opts_cnt = num_loop_opts(); 1608 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1609 { 1610 TracePhase t2("idealLoop", &_t_idealLoop, true); 1611 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate); 1612 loop_opts_cnt--; 1613 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1614 if (failing()) return; 1615 } 1616 // Loop opts pass if partial peeling occurred in previous pass 1617 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1618 TracePhase t3("idealLoop", &_t_idealLoop, true); 1619 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1620 loop_opts_cnt--; 1621 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1622 if (failing()) return; 1623 } 1624 // Loop opts pass for loop-unrolling before CCP 1625 if(major_progress() && (loop_opts_cnt > 0)) { 1626 TracePhase t4("idealLoop", &_t_idealLoop, true); 1627 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1628 loop_opts_cnt--; 1629 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1630 } 1631 if (!failing()) { 1632 // Verify that last round of loop opts produced a valid graph 1633 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1634 PhaseIdealLoop::verify(igvn); 1635 } 1636 } 1637 if (failing()) return; 1638 1639 // Conditional Constant Propagation; 1640 PhaseCCP ccp( &igvn ); 1641 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1642 { 1643 TracePhase t2("ccp", &_t_ccp, true); 1644 ccp.do_transform(); 1645 } 1646 print_method("PhaseCPP 1", 2); 1647 1648 assert( true, "Break here to ccp.dump_old2new_map()"); 1649 1650 // Iterative Global Value Numbering, including ideal transforms 1651 { 1652 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1653 igvn = ccp; 1654 igvn.optimize(); 1655 } 1656 1657 print_method("Iter GVN 2", 2); 1658 1659 if (failing()) return; 1660 1661 // Loop transforms on the ideal graph. Range Check Elimination, 1662 // peeling, unrolling, etc. 1663 if(loop_opts_cnt > 0) { 1664 debug_only( int cnt = 0; ); 1665 bool loop_predication = UseLoopPredicate; 1666 while(major_progress() && (loop_opts_cnt > 0)) { 1667 TracePhase t2("idealLoop", &_t_idealLoop, true); 1668 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1669 PhaseIdealLoop ideal_loop( igvn, true, loop_predication); 1670 loop_opts_cnt--; 1671 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1672 if (failing()) return; 1673 // Perform loop predication optimization during first iteration after CCP. 1674 // After that switch it off and cleanup unused loop predicates. 1675 if (loop_predication) { 1676 loop_predication = false; 1677 cleanup_loop_predicates(igvn); 1678 if (failing()) return; 1679 } 1680 } 1681 } 1682 1683 { 1684 // Verify that all previous optimizations produced a valid graph 1685 // at least to this point, even if no loop optimizations were done. 1686 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1687 PhaseIdealLoop::verify(igvn); 1688 } 1689 1690 { 1691 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1692 PhaseMacroExpand mex(igvn); 1693 if (mex.expand_macro_nodes()) { 1694 assert(failing(), "must bail out w/ explicit message"); 1695 return; 1696 } 1697 } 1698 1699 } // (End scope of igvn; run destructor if necessary for asserts.) 1700 1701 // A method with only infinite loops has no edges entering loops from root 1702 { 1703 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1704 if (final_graph_reshaping()) { 1705 assert(failing(), "must bail out w/ explicit message"); 1706 return; 1707 } 1708 } 1709 1710 print_method("Optimize finished", 2); 1711 } 1712 1713 1714 //------------------------------Code_Gen--------------------------------------- 1715 // Given a graph, generate code for it 1716 void Compile::Code_Gen() { 1717 if (failing()) return; 1718 1719 // Perform instruction selection. You might think we could reclaim Matcher 1720 // memory PDQ, but actually the Matcher is used in generating spill code. 1721 // Internals of the Matcher (including some VectorSets) must remain live 1722 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1723 // set a bit in reclaimed memory. 1724 1725 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1726 // nodes. Mapping is only valid at the root of each matched subtree. 1727 NOT_PRODUCT( verify_graph_edges(); ) 1728 1729 Node_List proj_list; 1730 Matcher m(proj_list); 1731 _matcher = &m; 1732 { 1733 TracePhase t2("matcher", &_t_matcher, true); 1734 m.match(); 1735 } 1736 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1737 // nodes. Mapping is only valid at the root of each matched subtree. 1738 NOT_PRODUCT( verify_graph_edges(); ) 1739 1740 // If you have too many nodes, or if matching has failed, bail out 1741 check_node_count(0, "out of nodes matching instructions"); 1742 if (failing()) return; 1743 1744 // Build a proper-looking CFG 1745 PhaseCFG cfg(node_arena(), root(), m); 1746 _cfg = &cfg; 1747 { 1748 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1749 cfg.Dominators(); 1750 if (failing()) return; 1751 1752 NOT_PRODUCT( verify_graph_edges(); ) 1753 1754 cfg.Estimate_Block_Frequency(); 1755 cfg.GlobalCodeMotion(m,unique(),proj_list); 1756 1757 print_method("Global code motion", 2); 1758 1759 if (failing()) return; 1760 NOT_PRODUCT( verify_graph_edges(); ) 1761 1762 debug_only( cfg.verify(); ) 1763 } 1764 NOT_PRODUCT( verify_graph_edges(); ) 1765 1766 PhaseChaitin regalloc(unique(),cfg,m); 1767 _regalloc = ®alloc; 1768 { 1769 TracePhase t2("regalloc", &_t_registerAllocation, true); 1770 // Perform any platform dependent preallocation actions. This is used, 1771 // for example, to avoid taking an implicit null pointer exception 1772 // using the frame pointer on win95. 1773 _regalloc->pd_preallocate_hook(); 1774 1775 // Perform register allocation. After Chaitin, use-def chains are 1776 // no longer accurate (at spill code) and so must be ignored. 1777 // Node->LRG->reg mappings are still accurate. 1778 _regalloc->Register_Allocate(); 1779 1780 // Bail out if the allocator builds too many nodes 1781 if (failing()) return; 1782 } 1783 1784 // Prior to register allocation we kept empty basic blocks in case the 1785 // the allocator needed a place to spill. After register allocation we 1786 // are not adding any new instructions. If any basic block is empty, we 1787 // can now safely remove it. 1788 { 1789 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1790 cfg.remove_empty(); 1791 if (do_freq_based_layout()) { 1792 PhaseBlockLayout layout(cfg); 1793 } else { 1794 cfg.set_loop_alignment(); 1795 } 1796 cfg.fixup_flow(); 1797 } 1798 1799 // Perform any platform dependent postallocation verifications. 1800 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1801 1802 // Apply peephole optimizations 1803 if( OptoPeephole ) { 1804 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1805 PhasePeephole peep( _regalloc, cfg); 1806 peep.do_transform(); 1807 } 1808 1809 // Convert Nodes to instruction bits in a buffer 1810 { 1811 // %%%% workspace merge brought two timers together for one job 1812 TracePhase t2a("output", &_t_output, true); 1813 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1814 Output(); 1815 } 1816 1817 print_method("Final Code"); 1818 1819 // He's dead, Jim. 1820 _cfg = (PhaseCFG*)0xdeadbeef; 1821 _regalloc = (PhaseChaitin*)0xdeadbeef; 1822 } 1823 1824 1825 //------------------------------dump_asm--------------------------------------- 1826 // Dump formatted assembly 1827 #ifndef PRODUCT 1828 void Compile::dump_asm(int *pcs, uint pc_limit) { 1829 bool cut_short = false; 1830 tty->print_cr("#"); 1831 tty->print("# "); _tf->dump(); tty->cr(); 1832 tty->print_cr("#"); 1833 1834 // For all blocks 1835 int pc = 0x0; // Program counter 1836 char starts_bundle = ' '; 1837 _regalloc->dump_frame(); 1838 1839 Node *n = NULL; 1840 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1841 if (VMThread::should_terminate()) { cut_short = true; break; } 1842 Block *b = _cfg->_blocks[i]; 1843 if (b->is_connector() && !Verbose) continue; 1844 n = b->_nodes[0]; 1845 if (pcs && n->_idx < pc_limit) 1846 tty->print("%3.3x ", pcs[n->_idx]); 1847 else 1848 tty->print(" "); 1849 b->dump_head( &_cfg->_bbs ); 1850 if (b->is_connector()) { 1851 tty->print_cr(" # Empty connector block"); 1852 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1853 tty->print_cr(" # Block is sole successor of call"); 1854 } 1855 1856 // For all instructions 1857 Node *delay = NULL; 1858 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1859 if (VMThread::should_terminate()) { cut_short = true; break; } 1860 n = b->_nodes[j]; 1861 if (valid_bundle_info(n)) { 1862 Bundle *bundle = node_bundling(n); 1863 if (bundle->used_in_unconditional_delay()) { 1864 delay = n; 1865 continue; 1866 } 1867 if (bundle->starts_bundle()) 1868 starts_bundle = '+'; 1869 } 1870 1871 if (WizardMode) n->dump(); 1872 1873 if( !n->is_Region() && // Dont print in the Assembly 1874 !n->is_Phi() && // a few noisely useless nodes 1875 !n->is_Proj() && 1876 !n->is_MachTemp() && 1877 !n->is_SafePointScalarObject() && 1878 !n->is_Catch() && // Would be nice to print exception table targets 1879 !n->is_MergeMem() && // Not very interesting 1880 !n->is_top() && // Debug info table constants 1881 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1882 ) { 1883 if (pcs && n->_idx < pc_limit) 1884 tty->print("%3.3x", pcs[n->_idx]); 1885 else 1886 tty->print(" "); 1887 tty->print(" %c ", starts_bundle); 1888 starts_bundle = ' '; 1889 tty->print("\t"); 1890 n->format(_regalloc, tty); 1891 tty->cr(); 1892 } 1893 1894 // If we have an instruction with a delay slot, and have seen a delay, 1895 // then back up and print it 1896 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1897 assert(delay != NULL, "no unconditional delay instruction"); 1898 if (WizardMode) delay->dump(); 1899 1900 if (node_bundling(delay)->starts_bundle()) 1901 starts_bundle = '+'; 1902 if (pcs && n->_idx < pc_limit) 1903 tty->print("%3.3x", pcs[n->_idx]); 1904 else 1905 tty->print(" "); 1906 tty->print(" %c ", starts_bundle); 1907 starts_bundle = ' '; 1908 tty->print("\t"); 1909 delay->format(_regalloc, tty); 1910 tty->print_cr(""); 1911 delay = NULL; 1912 } 1913 1914 // Dump the exception table as well 1915 if( n->is_Catch() && (Verbose || WizardMode) ) { 1916 // Print the exception table for this offset 1917 _handler_table.print_subtable_for(pc); 1918 } 1919 } 1920 1921 if (pcs && n->_idx < pc_limit) 1922 tty->print_cr("%3.3x", pcs[n->_idx]); 1923 else 1924 tty->print_cr(""); 1925 1926 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1927 1928 } // End of per-block dump 1929 tty->print_cr(""); 1930 1931 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1932 } 1933 #endif 1934 1935 //------------------------------Final_Reshape_Counts--------------------------- 1936 // This class defines counters to help identify when a method 1937 // may/must be executed using hardware with only 24-bit precision. 1938 struct Final_Reshape_Counts : public StackObj { 1939 int _call_count; // count non-inlined 'common' calls 1940 int _float_count; // count float ops requiring 24-bit precision 1941 int _double_count; // count double ops requiring more precision 1942 int _java_call_count; // count non-inlined 'java' calls 1943 int _inner_loop_count; // count loops which need alignment 1944 VectorSet _visited; // Visitation flags 1945 Node_List _tests; // Set of IfNodes & PCTableNodes 1946 1947 Final_Reshape_Counts() : 1948 _call_count(0), _float_count(0), _double_count(0), 1949 _java_call_count(0), _inner_loop_count(0), 1950 _visited( Thread::current()->resource_area() ) { } 1951 1952 void inc_call_count () { _call_count ++; } 1953 void inc_float_count () { _float_count ++; } 1954 void inc_double_count() { _double_count++; } 1955 void inc_java_call_count() { _java_call_count++; } 1956 void inc_inner_loop_count() { _inner_loop_count++; } 1957 1958 int get_call_count () const { return _call_count ; } 1959 int get_float_count () const { return _float_count ; } 1960 int get_double_count() const { return _double_count; } 1961 int get_java_call_count() const { return _java_call_count; } 1962 int get_inner_loop_count() const { return _inner_loop_count; } 1963 }; 1964 1965 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1966 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1967 // Make sure the offset goes inside the instance layout. 1968 return k->contains_field_offset(tp->offset()); 1969 // Note that OffsetBot and OffsetTop are very negative. 1970 } 1971 1972 //------------------------------final_graph_reshaping_impl---------------------- 1973 // Implement items 1-5 from final_graph_reshaping below. 1974 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 1975 1976 if ( n->outcnt() == 0 ) return; // dead node 1977 uint nop = n->Opcode(); 1978 1979 // Check for 2-input instruction with "last use" on right input. 1980 // Swap to left input. Implements item (2). 1981 if( n->req() == 3 && // two-input instruction 1982 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1983 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1984 n->in(2)->outcnt() == 1 &&// right use IS a last use 1985 !n->in(2)->is_Con() ) { // right use is not a constant 1986 // Check for commutative opcode 1987 switch( nop ) { 1988 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1989 case Op_MaxI: case Op_MinI: 1990 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1991 case Op_AndL: case Op_XorL: case Op_OrL: 1992 case Op_AndI: case Op_XorI: case Op_OrI: { 1993 // Move "last use" input to left by swapping inputs 1994 n->swap_edges(1, 2); 1995 break; 1996 } 1997 default: 1998 break; 1999 } 2000 } 2001 2002 // Count FPU ops and common calls, implements item (3) 2003 switch( nop ) { 2004 // Count all float operations that may use FPU 2005 case Op_AddF: 2006 case Op_SubF: 2007 case Op_MulF: 2008 case Op_DivF: 2009 case Op_NegF: 2010 case Op_ModF: 2011 case Op_ConvI2F: 2012 case Op_ConF: 2013 case Op_CmpF: 2014 case Op_CmpF3: 2015 // case Op_ConvL2F: // longs are split into 32-bit halves 2016 frc.inc_float_count(); 2017 break; 2018 2019 case Op_ConvF2D: 2020 case Op_ConvD2F: 2021 frc.inc_float_count(); 2022 frc.inc_double_count(); 2023 break; 2024 2025 // Count all double operations that may use FPU 2026 case Op_AddD: 2027 case Op_SubD: 2028 case Op_MulD: 2029 case Op_DivD: 2030 case Op_NegD: 2031 case Op_ModD: 2032 case Op_ConvI2D: 2033 case Op_ConvD2I: 2034 // case Op_ConvL2D: // handled by leaf call 2035 // case Op_ConvD2L: // handled by leaf call 2036 case Op_ConD: 2037 case Op_CmpD: 2038 case Op_CmpD3: 2039 frc.inc_double_count(); 2040 break; 2041 case Op_Opaque1: // Remove Opaque Nodes before matching 2042 case Op_Opaque2: // Remove Opaque Nodes before matching 2043 n->subsume_by(n->in(1)); 2044 break; 2045 case Op_CallStaticJava: 2046 case Op_CallJava: 2047 case Op_CallDynamicJava: 2048 frc.inc_java_call_count(); // Count java call site; 2049 case Op_CallRuntime: 2050 case Op_CallLeaf: 2051 case Op_CallLeafNoFP: { 2052 assert( n->is_Call(), "" ); 2053 CallNode *call = n->as_Call(); 2054 // Count call sites where the FP mode bit would have to be flipped. 2055 // Do not count uncommon runtime calls: 2056 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2057 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2058 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2059 frc.inc_call_count(); // Count the call site 2060 } else { // See if uncommon argument is shared 2061 Node *n = call->in(TypeFunc::Parms); 2062 int nop = n->Opcode(); 2063 // Clone shared simple arguments to uncommon calls, item (1). 2064 if( n->outcnt() > 1 && 2065 !n->is_Proj() && 2066 nop != Op_CreateEx && 2067 nop != Op_CheckCastPP && 2068 nop != Op_DecodeN && 2069 !n->is_Mem() ) { 2070 Node *x = n->clone(); 2071 call->set_req( TypeFunc::Parms, x ); 2072 } 2073 } 2074 break; 2075 } 2076 2077 case Op_StoreD: 2078 case Op_LoadD: 2079 case Op_LoadD_unaligned: 2080 frc.inc_double_count(); 2081 goto handle_mem; 2082 case Op_StoreF: 2083 case Op_LoadF: 2084 frc.inc_float_count(); 2085 goto handle_mem; 2086 2087 case Op_StoreB: 2088 case Op_StoreC: 2089 case Op_StoreCM: 2090 case Op_StorePConditional: 2091 case Op_StoreI: 2092 case Op_StoreL: 2093 case Op_StoreIConditional: 2094 case Op_StoreLConditional: 2095 case Op_CompareAndSwapI: 2096 case Op_CompareAndSwapL: 2097 case Op_CompareAndSwapP: 2098 case Op_CompareAndSwapN: 2099 case Op_StoreP: 2100 case Op_StoreN: 2101 case Op_LoadB: 2102 case Op_LoadUB: 2103 case Op_LoadUS: 2104 case Op_LoadI: 2105 case Op_LoadUI2L: 2106 case Op_LoadKlass: 2107 case Op_LoadNKlass: 2108 case Op_LoadL: 2109 case Op_LoadL_unaligned: 2110 case Op_LoadPLocked: 2111 case Op_LoadLLocked: 2112 case Op_LoadP: 2113 case Op_LoadN: 2114 case Op_LoadRange: 2115 case Op_LoadS: { 2116 handle_mem: 2117 #ifdef ASSERT 2118 if( VerifyOptoOopOffsets ) { 2119 assert( n->is_Mem(), "" ); 2120 MemNode *mem = (MemNode*)n; 2121 // Check to see if address types have grounded out somehow. 2122 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2123 assert( !tp || oop_offset_is_sane(tp), "" ); 2124 } 2125 #endif 2126 break; 2127 } 2128 2129 case Op_AddP: { // Assert sane base pointers 2130 Node *addp = n->in(AddPNode::Address); 2131 assert( !addp->is_AddP() || 2132 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2133 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2134 "Base pointers must match" ); 2135 #ifdef _LP64 2136 if (UseCompressedOops && 2137 addp->Opcode() == Op_ConP && 2138 addp == n->in(AddPNode::Base) && 2139 n->in(AddPNode::Offset)->is_Con()) { 2140 // Use addressing with narrow klass to load with offset on x86. 2141 // On sparc loading 32-bits constant and decoding it have less 2142 // instructions (4) then load 64-bits constant (7). 2143 // Do this transformation here since IGVN will convert ConN back to ConP. 2144 const Type* t = addp->bottom_type(); 2145 if (t->isa_oopptr()) { 2146 Node* nn = NULL; 2147 2148 // Look for existing ConN node of the same exact type. 2149 Compile* C = Compile::current(); 2150 Node* r = C->root(); 2151 uint cnt = r->outcnt(); 2152 for (uint i = 0; i < cnt; i++) { 2153 Node* m = r->raw_out(i); 2154 if (m!= NULL && m->Opcode() == Op_ConN && 2155 m->bottom_type()->make_ptr() == t) { 2156 nn = m; 2157 break; 2158 } 2159 } 2160 if (nn != NULL) { 2161 // Decode a narrow oop to match address 2162 // [R12 + narrow_oop_reg<<3 + offset] 2163 nn = new (C, 2) DecodeNNode(nn, t); 2164 n->set_req(AddPNode::Base, nn); 2165 n->set_req(AddPNode::Address, nn); 2166 if (addp->outcnt() == 0) { 2167 addp->disconnect_inputs(NULL); 2168 } 2169 } 2170 } 2171 } 2172 #endif 2173 break; 2174 } 2175 2176 #ifdef _LP64 2177 case Op_CastPP: 2178 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) { 2179 Compile* C = Compile::current(); 2180 Node* in1 = n->in(1); 2181 const Type* t = n->bottom_type(); 2182 Node* new_in1 = in1->clone(); 2183 new_in1->as_DecodeN()->set_type(t); 2184 2185 if (!Matcher::clone_shift_expressions) { 2186 // 2187 // x86, ARM and friends can handle 2 adds in addressing mode 2188 // and Matcher can fold a DecodeN node into address by using 2189 // a narrow oop directly and do implicit NULL check in address: 2190 // 2191 // [R12 + narrow_oop_reg<<3 + offset] 2192 // NullCheck narrow_oop_reg 2193 // 2194 // On other platforms (Sparc) we have to keep new DecodeN node and 2195 // use it to do implicit NULL check in address: 2196 // 2197 // decode_not_null narrow_oop_reg, base_reg 2198 // [base_reg + offset] 2199 // NullCheck base_reg 2200 // 2201 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2202 // to keep the information to which NULL check the new DecodeN node 2203 // corresponds to use it as value in implicit_null_check(). 2204 // 2205 new_in1->set_req(0, n->in(0)); 2206 } 2207 2208 n->subsume_by(new_in1); 2209 if (in1->outcnt() == 0) { 2210 in1->disconnect_inputs(NULL); 2211 } 2212 } 2213 break; 2214 2215 case Op_CmpP: 2216 // Do this transformation here to preserve CmpPNode::sub() and 2217 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2218 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2219 Node* in1 = n->in(1); 2220 Node* in2 = n->in(2); 2221 if (!in1->is_DecodeN()) { 2222 in2 = in1; 2223 in1 = n->in(2); 2224 } 2225 assert(in1->is_DecodeN(), "sanity"); 2226 2227 Compile* C = Compile::current(); 2228 Node* new_in2 = NULL; 2229 if (in2->is_DecodeN()) { 2230 new_in2 = in2->in(1); 2231 } else if (in2->Opcode() == Op_ConP) { 2232 const Type* t = in2->bottom_type(); 2233 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) { 2234 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2235 // 2236 // This transformation together with CastPP transformation above 2237 // will generated code for implicit NULL checks for compressed oops. 2238 // 2239 // The original code after Optimize() 2240 // 2241 // LoadN memory, narrow_oop_reg 2242 // decode narrow_oop_reg, base_reg 2243 // CmpP base_reg, NULL 2244 // CastPP base_reg // NotNull 2245 // Load [base_reg + offset], val_reg 2246 // 2247 // after these transformations will be 2248 // 2249 // LoadN memory, narrow_oop_reg 2250 // CmpN narrow_oop_reg, NULL 2251 // decode_not_null narrow_oop_reg, base_reg 2252 // Load [base_reg + offset], val_reg 2253 // 2254 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2255 // since narrow oops can be used in debug info now (see the code in 2256 // final_graph_reshaping_walk()). 2257 // 2258 // At the end the code will be matched to 2259 // on x86: 2260 // 2261 // Load_narrow_oop memory, narrow_oop_reg 2262 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2263 // NullCheck narrow_oop_reg 2264 // 2265 // and on sparc: 2266 // 2267 // Load_narrow_oop memory, narrow_oop_reg 2268 // decode_not_null narrow_oop_reg, base_reg 2269 // Load [base_reg + offset], val_reg 2270 // NullCheck base_reg 2271 // 2272 } else if (t->isa_oopptr()) { 2273 new_in2 = ConNode::make(C, t->make_narrowoop()); 2274 } 2275 } 2276 if (new_in2 != NULL) { 2277 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2278 n->subsume_by( cmpN ); 2279 if (in1->outcnt() == 0) { 2280 in1->disconnect_inputs(NULL); 2281 } 2282 if (in2->outcnt() == 0) { 2283 in2->disconnect_inputs(NULL); 2284 } 2285 } 2286 } 2287 break; 2288 2289 case Op_DecodeN: 2290 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2291 // DecodeN could be pinned on Sparc where it can't be fold into 2292 // an address expression, see the code for Op_CastPP above. 2293 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc"); 2294 break; 2295 2296 case Op_EncodeP: { 2297 Node* in1 = n->in(1); 2298 if (in1->is_DecodeN()) { 2299 n->subsume_by(in1->in(1)); 2300 } else if (in1->Opcode() == Op_ConP) { 2301 Compile* C = Compile::current(); 2302 const Type* t = in1->bottom_type(); 2303 if (t == TypePtr::NULL_PTR) { 2304 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2305 } else if (t->isa_oopptr()) { 2306 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2307 } 2308 } 2309 if (in1->outcnt() == 0) { 2310 in1->disconnect_inputs(NULL); 2311 } 2312 break; 2313 } 2314 2315 case Op_Proj: { 2316 if (OptimizeStringConcat) { 2317 ProjNode* p = n->as_Proj(); 2318 if (p->_is_io_use) { 2319 // Separate projections were used for the exception path which 2320 // are normally removed by a late inline. If it wasn't inlined 2321 // then they will hang around and should just be replaced with 2322 // the original one. 2323 Node* proj = NULL; 2324 // Replace with just one 2325 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2326 Node *use = i.get(); 2327 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2328 proj = use; 2329 break; 2330 } 2331 } 2332 assert(p != NULL, "must be found"); 2333 p->subsume_by(proj); 2334 } 2335 } 2336 break; 2337 } 2338 2339 case Op_Phi: 2340 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2341 // The EncodeP optimization may create Phi with the same edges 2342 // for all paths. It is not handled well by Register Allocator. 2343 Node* unique_in = n->in(1); 2344 assert(unique_in != NULL, ""); 2345 uint cnt = n->req(); 2346 for (uint i = 2; i < cnt; i++) { 2347 Node* m = n->in(i); 2348 assert(m != NULL, ""); 2349 if (unique_in != m) 2350 unique_in = NULL; 2351 } 2352 if (unique_in != NULL) { 2353 n->subsume_by(unique_in); 2354 } 2355 } 2356 break; 2357 2358 #endif 2359 2360 case Op_ModI: 2361 if (UseDivMod) { 2362 // Check if a%b and a/b both exist 2363 Node* d = n->find_similar(Op_DivI); 2364 if (d) { 2365 // Replace them with a fused divmod if supported 2366 Compile* C = Compile::current(); 2367 if (Matcher::has_match_rule(Op_DivModI)) { 2368 DivModINode* divmod = DivModINode::make(C, n); 2369 d->subsume_by(divmod->div_proj()); 2370 n->subsume_by(divmod->mod_proj()); 2371 } else { 2372 // replace a%b with a-((a/b)*b) 2373 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2374 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2375 n->subsume_by( sub ); 2376 } 2377 } 2378 } 2379 break; 2380 2381 case Op_ModL: 2382 if (UseDivMod) { 2383 // Check if a%b and a/b both exist 2384 Node* d = n->find_similar(Op_DivL); 2385 if (d) { 2386 // Replace them with a fused divmod if supported 2387 Compile* C = Compile::current(); 2388 if (Matcher::has_match_rule(Op_DivModL)) { 2389 DivModLNode* divmod = DivModLNode::make(C, n); 2390 d->subsume_by(divmod->div_proj()); 2391 n->subsume_by(divmod->mod_proj()); 2392 } else { 2393 // replace a%b with a-((a/b)*b) 2394 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2395 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2396 n->subsume_by( sub ); 2397 } 2398 } 2399 } 2400 break; 2401 2402 case Op_Load16B: 2403 case Op_Load8B: 2404 case Op_Load4B: 2405 case Op_Load8S: 2406 case Op_Load4S: 2407 case Op_Load2S: 2408 case Op_Load8C: 2409 case Op_Load4C: 2410 case Op_Load2C: 2411 case Op_Load4I: 2412 case Op_Load2I: 2413 case Op_Load2L: 2414 case Op_Load4F: 2415 case Op_Load2F: 2416 case Op_Load2D: 2417 case Op_Store16B: 2418 case Op_Store8B: 2419 case Op_Store4B: 2420 case Op_Store8C: 2421 case Op_Store4C: 2422 case Op_Store2C: 2423 case Op_Store4I: 2424 case Op_Store2I: 2425 case Op_Store2L: 2426 case Op_Store4F: 2427 case Op_Store2F: 2428 case Op_Store2D: 2429 break; 2430 2431 case Op_PackB: 2432 case Op_PackS: 2433 case Op_PackC: 2434 case Op_PackI: 2435 case Op_PackF: 2436 case Op_PackL: 2437 case Op_PackD: 2438 if (n->req()-1 > 2) { 2439 // Replace many operand PackNodes with a binary tree for matching 2440 PackNode* p = (PackNode*) n; 2441 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2442 n->subsume_by(btp); 2443 } 2444 break; 2445 case Op_Loop: 2446 case Op_CountedLoop: 2447 if (n->as_Loop()->is_inner_loop()) { 2448 frc.inc_inner_loop_count(); 2449 } 2450 break; 2451 default: 2452 assert( !n->is_Call(), "" ); 2453 assert( !n->is_Mem(), "" ); 2454 break; 2455 } 2456 2457 // Collect CFG split points 2458 if (n->is_MultiBranch()) 2459 frc._tests.push(n); 2460 } 2461 2462 //------------------------------final_graph_reshaping_walk--------------------- 2463 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2464 // requires that the walk visits a node's inputs before visiting the node. 2465 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2466 ResourceArea *area = Thread::current()->resource_area(); 2467 Unique_Node_List sfpt(area); 2468 2469 frc._visited.set(root->_idx); // first, mark node as visited 2470 uint cnt = root->req(); 2471 Node *n = root; 2472 uint i = 0; 2473 while (true) { 2474 if (i < cnt) { 2475 // Place all non-visited non-null inputs onto stack 2476 Node* m = n->in(i); 2477 ++i; 2478 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2479 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2480 sfpt.push(m); 2481 cnt = m->req(); 2482 nstack.push(n, i); // put on stack parent and next input's index 2483 n = m; 2484 i = 0; 2485 } 2486 } else { 2487 // Now do post-visit work 2488 final_graph_reshaping_impl( n, frc ); 2489 if (nstack.is_empty()) 2490 break; // finished 2491 n = nstack.node(); // Get node from stack 2492 cnt = n->req(); 2493 i = nstack.index(); 2494 nstack.pop(); // Shift to the next node on stack 2495 } 2496 } 2497 2498 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2499 // It could be done for an uncommon traps or any safepoints/calls 2500 // if the DecodeN node is referenced only in a debug info. 2501 while (sfpt.size() > 0) { 2502 n = sfpt.pop(); 2503 JVMState *jvms = n->as_SafePoint()->jvms(); 2504 assert(jvms != NULL, "sanity"); 2505 int start = jvms->debug_start(); 2506 int end = n->req(); 2507 bool is_uncommon = (n->is_CallStaticJava() && 2508 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2509 for (int j = start; j < end; j++) { 2510 Node* in = n->in(j); 2511 if (in->is_DecodeN()) { 2512 bool safe_to_skip = true; 2513 if (!is_uncommon ) { 2514 // Is it safe to skip? 2515 for (uint i = 0; i < in->outcnt(); i++) { 2516 Node* u = in->raw_out(i); 2517 if (!u->is_SafePoint() || 2518 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2519 safe_to_skip = false; 2520 } 2521 } 2522 } 2523 if (safe_to_skip) { 2524 n->set_req(j, in->in(1)); 2525 } 2526 if (in->outcnt() == 0) { 2527 in->disconnect_inputs(NULL); 2528 } 2529 } 2530 } 2531 } 2532 } 2533 2534 //------------------------------final_graph_reshaping-------------------------- 2535 // Final Graph Reshaping. 2536 // 2537 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2538 // and not commoned up and forced early. Must come after regular 2539 // optimizations to avoid GVN undoing the cloning. Clone constant 2540 // inputs to Loop Phis; these will be split by the allocator anyways. 2541 // Remove Opaque nodes. 2542 // (2) Move last-uses by commutative operations to the left input to encourage 2543 // Intel update-in-place two-address operations and better register usage 2544 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2545 // calls canonicalizing them back. 2546 // (3) Count the number of double-precision FP ops, single-precision FP ops 2547 // and call sites. On Intel, we can get correct rounding either by 2548 // forcing singles to memory (requires extra stores and loads after each 2549 // FP bytecode) or we can set a rounding mode bit (requires setting and 2550 // clearing the mode bit around call sites). The mode bit is only used 2551 // if the relative frequency of single FP ops to calls is low enough. 2552 // This is a key transform for SPEC mpeg_audio. 2553 // (4) Detect infinite loops; blobs of code reachable from above but not 2554 // below. Several of the Code_Gen algorithms fail on such code shapes, 2555 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2556 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2557 // Detection is by looking for IfNodes where only 1 projection is 2558 // reachable from below or CatchNodes missing some targets. 2559 // (5) Assert for insane oop offsets in debug mode. 2560 2561 bool Compile::final_graph_reshaping() { 2562 // an infinite loop may have been eliminated by the optimizer, 2563 // in which case the graph will be empty. 2564 if (root()->req() == 1) { 2565 record_method_not_compilable("trivial infinite loop"); 2566 return true; 2567 } 2568 2569 Final_Reshape_Counts frc; 2570 2571 // Visit everybody reachable! 2572 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2573 Node_Stack nstack(unique() >> 1); 2574 final_graph_reshaping_walk(nstack, root(), frc); 2575 2576 // Check for unreachable (from below) code (i.e., infinite loops). 2577 for( uint i = 0; i < frc._tests.size(); i++ ) { 2578 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2579 // Get number of CFG targets. 2580 // Note that PCTables include exception targets after calls. 2581 uint required_outcnt = n->required_outcnt(); 2582 if (n->outcnt() != required_outcnt) { 2583 // Check for a few special cases. Rethrow Nodes never take the 2584 // 'fall-thru' path, so expected kids is 1 less. 2585 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2586 if (n->in(0)->in(0)->is_Call()) { 2587 CallNode *call = n->in(0)->in(0)->as_Call(); 2588 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2589 required_outcnt--; // Rethrow always has 1 less kid 2590 } else if (call->req() > TypeFunc::Parms && 2591 call->is_CallDynamicJava()) { 2592 // Check for null receiver. In such case, the optimizer has 2593 // detected that the virtual call will always result in a null 2594 // pointer exception. The fall-through projection of this CatchNode 2595 // will not be populated. 2596 Node *arg0 = call->in(TypeFunc::Parms); 2597 if (arg0->is_Type() && 2598 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2599 required_outcnt--; 2600 } 2601 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2602 call->req() > TypeFunc::Parms+1 && 2603 call->is_CallStaticJava()) { 2604 // Check for negative array length. In such case, the optimizer has 2605 // detected that the allocation attempt will always result in an 2606 // exception. There is no fall-through projection of this CatchNode . 2607 Node *arg1 = call->in(TypeFunc::Parms+1); 2608 if (arg1->is_Type() && 2609 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2610 required_outcnt--; 2611 } 2612 } 2613 } 2614 } 2615 // Recheck with a better notion of 'required_outcnt' 2616 if (n->outcnt() != required_outcnt) { 2617 record_method_not_compilable("malformed control flow"); 2618 return true; // Not all targets reachable! 2619 } 2620 } 2621 // Check that I actually visited all kids. Unreached kids 2622 // must be infinite loops. 2623 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2624 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2625 record_method_not_compilable("infinite loop"); 2626 return true; // Found unvisited kid; must be unreach 2627 } 2628 } 2629 2630 // If original bytecodes contained a mixture of floats and doubles 2631 // check if the optimizer has made it homogenous, item (3). 2632 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2633 frc.get_float_count() > 32 && 2634 frc.get_double_count() == 0 && 2635 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2636 set_24_bit_selection_and_mode( false, true ); 2637 } 2638 2639 set_java_calls(frc.get_java_call_count()); 2640 set_inner_loops(frc.get_inner_loop_count()); 2641 2642 // No infinite loops, no reason to bail out. 2643 return false; 2644 } 2645 2646 //-----------------------------too_many_traps---------------------------------- 2647 // Report if there are too many traps at the current method and bci. 2648 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2649 bool Compile::too_many_traps(ciMethod* method, 2650 int bci, 2651 Deoptimization::DeoptReason reason) { 2652 ciMethodData* md = method->method_data(); 2653 if (md->is_empty()) { 2654 // Assume the trap has not occurred, or that it occurred only 2655 // because of a transient condition during start-up in the interpreter. 2656 return false; 2657 } 2658 if (md->has_trap_at(bci, reason) != 0) { 2659 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2660 // Also, if there are multiple reasons, or if there is no per-BCI record, 2661 // assume the worst. 2662 if (log()) 2663 log()->elem("observe trap='%s' count='%d'", 2664 Deoptimization::trap_reason_name(reason), 2665 md->trap_count(reason)); 2666 return true; 2667 } else { 2668 // Ignore method/bci and see if there have been too many globally. 2669 return too_many_traps(reason, md); 2670 } 2671 } 2672 2673 // Less-accurate variant which does not require a method and bci. 2674 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2675 ciMethodData* logmd) { 2676 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2677 // Too many traps globally. 2678 // Note that we use cumulative trap_count, not just md->trap_count. 2679 if (log()) { 2680 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2681 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2682 Deoptimization::trap_reason_name(reason), 2683 mcount, trap_count(reason)); 2684 } 2685 return true; 2686 } else { 2687 // The coast is clear. 2688 return false; 2689 } 2690 } 2691 2692 //--------------------------too_many_recompiles-------------------------------- 2693 // Report if there are too many recompiles at the current method and bci. 2694 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2695 // Is not eager to return true, since this will cause the compiler to use 2696 // Action_none for a trap point, to avoid too many recompilations. 2697 bool Compile::too_many_recompiles(ciMethod* method, 2698 int bci, 2699 Deoptimization::DeoptReason reason) { 2700 ciMethodData* md = method->method_data(); 2701 if (md->is_empty()) { 2702 // Assume the trap has not occurred, or that it occurred only 2703 // because of a transient condition during start-up in the interpreter. 2704 return false; 2705 } 2706 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2707 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2708 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2709 Deoptimization::DeoptReason per_bc_reason 2710 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2711 if ((per_bc_reason == Deoptimization::Reason_none 2712 || md->has_trap_at(bci, reason) != 0) 2713 // The trap frequency measure we care about is the recompile count: 2714 && md->trap_recompiled_at(bci) 2715 && md->overflow_recompile_count() >= bc_cutoff) { 2716 // Do not emit a trap here if it has already caused recompilations. 2717 // Also, if there are multiple reasons, or if there is no per-BCI record, 2718 // assume the worst. 2719 if (log()) 2720 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2721 Deoptimization::trap_reason_name(reason), 2722 md->trap_count(reason), 2723 md->overflow_recompile_count()); 2724 return true; 2725 } else if (trap_count(reason) != 0 2726 && decompile_count() >= m_cutoff) { 2727 // Too many recompiles globally, and we have seen this sort of trap. 2728 // Use cumulative decompile_count, not just md->decompile_count. 2729 if (log()) 2730 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2731 Deoptimization::trap_reason_name(reason), 2732 md->trap_count(reason), trap_count(reason), 2733 md->decompile_count(), decompile_count()); 2734 return true; 2735 } else { 2736 // The coast is clear. 2737 return false; 2738 } 2739 } 2740 2741 2742 #ifndef PRODUCT 2743 //------------------------------verify_graph_edges--------------------------- 2744 // Walk the Graph and verify that there is a one-to-one correspondence 2745 // between Use-Def edges and Def-Use edges in the graph. 2746 void Compile::verify_graph_edges(bool no_dead_code) { 2747 if (VerifyGraphEdges) { 2748 ResourceArea *area = Thread::current()->resource_area(); 2749 Unique_Node_List visited(area); 2750 // Call recursive graph walk to check edges 2751 _root->verify_edges(visited); 2752 if (no_dead_code) { 2753 // Now make sure that no visited node is used by an unvisited node. 2754 bool dead_nodes = 0; 2755 Unique_Node_List checked(area); 2756 while (visited.size() > 0) { 2757 Node* n = visited.pop(); 2758 checked.push(n); 2759 for (uint i = 0; i < n->outcnt(); i++) { 2760 Node* use = n->raw_out(i); 2761 if (checked.member(use)) continue; // already checked 2762 if (visited.member(use)) continue; // already in the graph 2763 if (use->is_Con()) continue; // a dead ConNode is OK 2764 // At this point, we have found a dead node which is DU-reachable. 2765 if (dead_nodes++ == 0) 2766 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2767 use->dump(2); 2768 tty->print_cr("---"); 2769 checked.push(use); // No repeats; pretend it is now checked. 2770 } 2771 } 2772 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2773 } 2774 } 2775 } 2776 #endif 2777 2778 // The Compile object keeps track of failure reasons separately from the ciEnv. 2779 // This is required because there is not quite a 1-1 relation between the 2780 // ciEnv and its compilation task and the Compile object. Note that one 2781 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2782 // to backtrack and retry without subsuming loads. Other than this backtracking 2783 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2784 // by the logic in C2Compiler. 2785 void Compile::record_failure(const char* reason) { 2786 if (log() != NULL) { 2787 log()->elem("failure reason='%s' phase='compile'", reason); 2788 } 2789 if (_failure_reason == NULL) { 2790 // Record the first failure reason. 2791 _failure_reason = reason; 2792 } 2793 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2794 C->print_method(_failure_reason); 2795 } 2796 _root = NULL; // flush the graph, too 2797 } 2798 2799 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2800 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2801 { 2802 if (dolog) { 2803 C = Compile::current(); 2804 _log = C->log(); 2805 } else { 2806 C = NULL; 2807 _log = NULL; 2808 } 2809 if (_log != NULL) { 2810 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2811 _log->stamp(); 2812 _log->end_head(); 2813 } 2814 } 2815 2816 Compile::TracePhase::~TracePhase() { 2817 if (_log != NULL) { 2818 _log->done("phase nodes='%d'", C->unique()); 2819 } 2820 }