1 /* 2 * Copyright 1997-2010 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 _has_method_handle_invokes(false), 469 _node_bundling_limit(0), 470 _node_bundling_base(NULL), 471 _java_calls(0), 472 _inner_loops(0), 473 #ifndef PRODUCT 474 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 475 _printer(IdealGraphPrinter::printer()), 476 #endif 477 _congraph(NULL) { 478 C = this; 479 480 CompileWrapper cw(this); 481 #ifndef PRODUCT 482 if (TimeCompiler2) { 483 tty->print(" "); 484 target->holder()->name()->print(); 485 tty->print("."); 486 target->print_short_name(); 487 tty->print(" "); 488 } 489 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 490 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 491 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 492 if (!print_opto_assembly) { 493 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 494 if (print_assembly && !Disassembler::can_decode()) { 495 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 496 print_opto_assembly = true; 497 } 498 } 499 set_print_assembly(print_opto_assembly); 500 set_parsed_irreducible_loop(false); 501 #endif 502 503 if (ProfileTraps) { 504 // Make sure the method being compiled gets its own MDO, 505 // so we can at least track the decompile_count(). 506 method()->build_method_data(); 507 } 508 509 Init(::AliasLevel); 510 511 512 print_compile_messages(); 513 514 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 515 _ilt = InlineTree::build_inline_tree_root(); 516 else 517 _ilt = NULL; 518 519 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 520 assert(num_alias_types() >= AliasIdxRaw, ""); 521 522 #define MINIMUM_NODE_HASH 1023 523 // Node list that Iterative GVN will start with 524 Unique_Node_List for_igvn(comp_arena()); 525 set_for_igvn(&for_igvn); 526 527 // GVN that will be run immediately on new nodes 528 uint estimated_size = method()->code_size()*4+64; 529 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 530 PhaseGVN gvn(node_arena(), estimated_size); 531 set_initial_gvn(&gvn); 532 533 { // Scope for timing the parser 534 TracePhase t3("parse", &_t_parser, true); 535 536 // Put top into the hash table ASAP. 537 initial_gvn()->transform_no_reclaim(top()); 538 539 // Set up tf(), start(), and find a CallGenerator. 540 CallGenerator* cg; 541 if (is_osr_compilation()) { 542 const TypeTuple *domain = StartOSRNode::osr_domain(); 543 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 544 init_tf(TypeFunc::make(domain, range)); 545 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 546 initial_gvn()->set_type_bottom(s); 547 init_start(s); 548 cg = CallGenerator::for_osr(method(), entry_bci()); 549 } else { 550 // Normal case. 551 init_tf(TypeFunc::make(method())); 552 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 553 initial_gvn()->set_type_bottom(s); 554 init_start(s); 555 float past_uses = method()->interpreter_invocation_count(); 556 float expected_uses = past_uses; 557 cg = CallGenerator::for_inline(method(), expected_uses); 558 } 559 if (failing()) return; 560 if (cg == NULL) { 561 record_method_not_compilable_all_tiers("cannot parse method"); 562 return; 563 } 564 JVMState* jvms = build_start_state(start(), tf()); 565 if ((jvms = cg->generate(jvms)) == NULL) { 566 record_method_not_compilable("method parse failed"); 567 return; 568 } 569 GraphKit kit(jvms); 570 571 if (!kit.stopped()) { 572 // Accept return values, and transfer control we know not where. 573 // This is done by a special, unique ReturnNode bound to root. 574 return_values(kit.jvms()); 575 } 576 577 if (kit.has_exceptions()) { 578 // Any exceptions that escape from this call must be rethrown 579 // to whatever caller is dynamically above us on the stack. 580 // This is done by a special, unique RethrowNode bound to root. 581 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 582 } 583 584 if (!failing() && has_stringbuilder()) { 585 { 586 // remove useless nodes to make the usage analysis simpler 587 ResourceMark rm; 588 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 589 } 590 591 { 592 ResourceMark rm; 593 print_method("Before StringOpts", 3); 594 PhaseStringOpts pso(initial_gvn(), &for_igvn); 595 print_method("After StringOpts", 3); 596 } 597 598 // now inline anything that we skipped the first time around 599 while (_late_inlines.length() > 0) { 600 CallGenerator* cg = _late_inlines.pop(); 601 cg->do_late_inline(); 602 } 603 } 604 assert(_late_inlines.length() == 0, "should have been processed"); 605 606 print_method("Before RemoveUseless", 3); 607 608 // Remove clutter produced by parsing. 609 if (!failing()) { 610 ResourceMark rm; 611 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 612 } 613 } 614 615 // Note: Large methods are capped off in do_one_bytecode(). 616 if (failing()) return; 617 618 // After parsing, node notes are no longer automagic. 619 // They must be propagated by register_new_node_with_optimizer(), 620 // clone(), or the like. 621 set_default_node_notes(NULL); 622 623 for (;;) { 624 int successes = Inline_Warm(); 625 if (failing()) return; 626 if (successes == 0) break; 627 } 628 629 // Drain the list. 630 Finish_Warm(); 631 #ifndef PRODUCT 632 if (_printer) { 633 _printer->print_inlining(this); 634 } 635 #endif 636 637 if (failing()) return; 638 NOT_PRODUCT( verify_graph_edges(); ) 639 640 // Perform escape analysis 641 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 642 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 643 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction. 644 PhaseGVN* igvn = initial_gvn(); 645 Node* oop_null = igvn->zerocon(T_OBJECT); 646 Node* noop_null = igvn->zerocon(T_NARROWOOP); 647 648 _congraph = new(comp_arena()) ConnectionGraph(this); 649 bool has_non_escaping_obj = _congraph->compute_escape(); 650 651 #ifndef PRODUCT 652 if (PrintEscapeAnalysis) { 653 _congraph->dump(); 654 } 655 #endif 656 // Cleanup. 657 if (oop_null->outcnt() == 0) 658 igvn->hash_delete(oop_null); 659 if (noop_null->outcnt() == 0) 660 igvn->hash_delete(noop_null); 661 662 if (!has_non_escaping_obj) { 663 _congraph = NULL; 664 } 665 666 if (failing()) return; 667 } 668 // Now optimize 669 Optimize(); 670 if (failing()) return; 671 NOT_PRODUCT( verify_graph_edges(); ) 672 673 #ifndef PRODUCT 674 if (PrintIdeal) { 675 ttyLocker ttyl; // keep the following output all in one block 676 // This output goes directly to the tty, not the compiler log. 677 // To enable tools to match it up with the compilation activity, 678 // be sure to tag this tty output with the compile ID. 679 if (xtty != NULL) { 680 xtty->head("ideal compile_id='%d'%s", compile_id(), 681 is_osr_compilation() ? " compile_kind='osr'" : 682 ""); 683 } 684 root()->dump(9999); 685 if (xtty != NULL) { 686 xtty->tail("ideal"); 687 } 688 } 689 #endif 690 691 // Now that we know the size of all the monitors we can add a fixed slot 692 // for the original deopt pc. 693 694 _orig_pc_slot = fixed_slots(); 695 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 696 set_fixed_slots(next_slot); 697 698 // Now generate code 699 Code_Gen(); 700 if (failing()) return; 701 702 // Check if we want to skip execution of all compiled code. 703 { 704 #ifndef PRODUCT 705 if (OptoNoExecute) { 706 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 707 return; 708 } 709 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 710 #endif 711 712 if (is_osr_compilation()) { 713 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 714 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 715 } else { 716 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 717 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 718 } 719 720 env()->register_method(_method, _entry_bci, 721 &_code_offsets, 722 _orig_pc_slot_offset_in_bytes, 723 code_buffer(), 724 frame_size_in_words(), _oop_map_set, 725 &_handler_table, &_inc_table, 726 compiler, 727 env()->comp_level(), 728 true, /*has_debug_info*/ 729 has_unsafe_access() 730 ); 731 } 732 } 733 734 //------------------------------Compile---------------------------------------- 735 // Compile a runtime stub 736 Compile::Compile( ciEnv* ci_env, 737 TypeFunc_generator generator, 738 address stub_function, 739 const char *stub_name, 740 int is_fancy_jump, 741 bool pass_tls, 742 bool save_arg_registers, 743 bool return_pc ) 744 : Phase(Compiler), 745 _env(ci_env), 746 _log(ci_env->log()), 747 _compile_id(-1), 748 _save_argument_registers(save_arg_registers), 749 _method(NULL), 750 _stub_name(stub_name), 751 _stub_function(stub_function), 752 _stub_entry_point(NULL), 753 _entry_bci(InvocationEntryBci), 754 _initial_gvn(NULL), 755 _for_igvn(NULL), 756 _warm_calls(NULL), 757 _orig_pc_slot(0), 758 _orig_pc_slot_offset_in_bytes(0), 759 _subsume_loads(true), 760 _do_escape_analysis(false), 761 _failure_reason(NULL), 762 _code_buffer("Compile::Fill_buffer"), 763 _has_method_handle_invokes(false), 764 _node_bundling_limit(0), 765 _node_bundling_base(NULL), 766 _java_calls(0), 767 _inner_loops(0), 768 #ifndef PRODUCT 769 _trace_opto_output(TraceOptoOutput), 770 _printer(NULL), 771 #endif 772 _congraph(NULL) { 773 C = this; 774 775 #ifndef PRODUCT 776 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 777 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 778 set_print_assembly(PrintFrameConverterAssembly); 779 set_parsed_irreducible_loop(false); 780 #endif 781 CompileWrapper cw(this); 782 Init(/*AliasLevel=*/ 0); 783 init_tf((*generator)()); 784 785 { 786 // The following is a dummy for the sake of GraphKit::gen_stub 787 Unique_Node_List for_igvn(comp_arena()); 788 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 789 PhaseGVN gvn(Thread::current()->resource_area(),255); 790 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 791 gvn.transform_no_reclaim(top()); 792 793 GraphKit kit; 794 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 795 } 796 797 NOT_PRODUCT( verify_graph_edges(); ) 798 Code_Gen(); 799 if (failing()) return; 800 801 802 // Entry point will be accessed using compile->stub_entry_point(); 803 if (code_buffer() == NULL) { 804 Matcher::soft_match_failure(); 805 } else { 806 if (PrintAssembly && (WizardMode || Verbose)) 807 tty->print_cr("### Stub::%s", stub_name); 808 809 if (!failing()) { 810 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 811 812 // Make the NMethod 813 // For now we mark the frame as never safe for profile stackwalking 814 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 815 code_buffer(), 816 CodeOffsets::frame_never_safe, 817 // _code_offsets.value(CodeOffsets::Frame_Complete), 818 frame_size_in_words(), 819 _oop_map_set, 820 save_arg_registers); 821 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 822 823 _stub_entry_point = rs->entry_point(); 824 } 825 } 826 } 827 828 #ifndef PRODUCT 829 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 830 if(PrintOpto && Verbose) { 831 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 832 } 833 } 834 #endif 835 836 void Compile::print_codes() { 837 } 838 839 //------------------------------Init------------------------------------------- 840 // Prepare for a single compilation 841 void Compile::Init(int aliaslevel) { 842 _unique = 0; 843 _regalloc = NULL; 844 845 _tf = NULL; // filled in later 846 _top = NULL; // cached later 847 _matcher = NULL; // filled in later 848 _cfg = NULL; // filled in later 849 850 set_24_bit_selection_and_mode(Use24BitFP, false); 851 852 _node_note_array = NULL; 853 _default_node_notes = NULL; 854 855 _immutable_memory = NULL; // filled in at first inquiry 856 857 // Globally visible Nodes 858 // First set TOP to NULL to give safe behavior during creation of RootNode 859 set_cached_top_node(NULL); 860 set_root(new (this, 3) RootNode()); 861 // Now that you have a Root to point to, create the real TOP 862 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 863 set_recent_alloc(NULL, NULL); 864 865 // Create Debug Information Recorder to record scopes, oopmaps, etc. 866 env()->set_oop_recorder(new OopRecorder(comp_arena())); 867 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 868 env()->set_dependencies(new Dependencies(env())); 869 870 _fixed_slots = 0; 871 set_has_split_ifs(false); 872 set_has_loops(has_method() && method()->has_loops()); // first approximation 873 set_has_stringbuilder(false); 874 _trap_can_recompile = false; // no traps emitted yet 875 _major_progress = true; // start out assuming good things will happen 876 set_has_unsafe_access(false); 877 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 878 set_decompile_count(0); 879 880 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 881 // Compilation level related initialization 882 if (env()->comp_level() == CompLevel_fast_compile) { 883 set_num_loop_opts(Tier1LoopOptsCount); 884 set_do_inlining(Tier1Inline != 0); 885 set_max_inline_size(Tier1MaxInlineSize); 886 set_freq_inline_size(Tier1FreqInlineSize); 887 set_do_scheduling(false); 888 set_do_count_invocations(Tier1CountInvocations); 889 set_do_method_data_update(Tier1UpdateMethodData); 890 } else { 891 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 892 set_num_loop_opts(LoopOptsCount); 893 set_do_inlining(Inline); 894 set_max_inline_size(MaxInlineSize); 895 set_freq_inline_size(FreqInlineSize); 896 set_do_scheduling(OptoScheduling); 897 set_do_count_invocations(false); 898 set_do_method_data_update(false); 899 } 900 901 if (debug_info()->recording_non_safepoints()) { 902 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 903 (comp_arena(), 8, 0, NULL)); 904 set_default_node_notes(Node_Notes::make(this)); 905 } 906 907 // // -- Initialize types before each compile -- 908 // // Update cached type information 909 // if( _method && _method->constants() ) 910 // Type::update_loaded_types(_method, _method->constants()); 911 912 // Init alias_type map. 913 if (!_do_escape_analysis && aliaslevel == 3) 914 aliaslevel = 2; // No unique types without escape analysis 915 _AliasLevel = aliaslevel; 916 const int grow_ats = 16; 917 _max_alias_types = grow_ats; 918 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 919 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 920 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 921 { 922 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 923 } 924 // Initialize the first few types. 925 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 926 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 927 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 928 _num_alias_types = AliasIdxRaw+1; 929 // Zero out the alias type cache. 930 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 931 // A NULL adr_type hits in the cache right away. Preload the right answer. 932 probe_alias_cache(NULL)->_index = AliasIdxTop; 933 934 _intrinsics = NULL; 935 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 936 _predicate_opaqs = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 937 register_library_intrinsics(); 938 } 939 940 //---------------------------init_start---------------------------------------- 941 // Install the StartNode on this compile object. 942 void Compile::init_start(StartNode* s) { 943 if (failing()) 944 return; // already failing 945 assert(s == start(), ""); 946 } 947 948 StartNode* Compile::start() const { 949 assert(!failing(), ""); 950 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 951 Node* start = root()->fast_out(i); 952 if( start->is_Start() ) 953 return start->as_Start(); 954 } 955 ShouldNotReachHere(); 956 return NULL; 957 } 958 959 //-------------------------------immutable_memory------------------------------------- 960 // Access immutable memory 961 Node* Compile::immutable_memory() { 962 if (_immutable_memory != NULL) { 963 return _immutable_memory; 964 } 965 StartNode* s = start(); 966 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 967 Node *p = s->fast_out(i); 968 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 969 _immutable_memory = p; 970 return _immutable_memory; 971 } 972 } 973 ShouldNotReachHere(); 974 return NULL; 975 } 976 977 //----------------------set_cached_top_node------------------------------------ 978 // Install the cached top node, and make sure Node::is_top works correctly. 979 void Compile::set_cached_top_node(Node* tn) { 980 if (tn != NULL) verify_top(tn); 981 Node* old_top = _top; 982 _top = tn; 983 // Calling Node::setup_is_top allows the nodes the chance to adjust 984 // their _out arrays. 985 if (_top != NULL) _top->setup_is_top(); 986 if (old_top != NULL) old_top->setup_is_top(); 987 assert(_top == NULL || top()->is_top(), ""); 988 } 989 990 #ifndef PRODUCT 991 void Compile::verify_top(Node* tn) const { 992 if (tn != NULL) { 993 assert(tn->is_Con(), "top node must be a constant"); 994 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 995 assert(tn->in(0) != NULL, "must have live top node"); 996 } 997 } 998 #endif 999 1000 1001 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1002 1003 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1004 guarantee(arr != NULL, ""); 1005 int num_blocks = arr->length(); 1006 if (grow_by < num_blocks) grow_by = num_blocks; 1007 int num_notes = grow_by * _node_notes_block_size; 1008 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1009 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1010 while (num_notes > 0) { 1011 arr->append(notes); 1012 notes += _node_notes_block_size; 1013 num_notes -= _node_notes_block_size; 1014 } 1015 assert(num_notes == 0, "exact multiple, please"); 1016 } 1017 1018 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1019 if (source == NULL || dest == NULL) return false; 1020 1021 if (dest->is_Con()) 1022 return false; // Do not push debug info onto constants. 1023 1024 #ifdef ASSERT 1025 // Leave a bread crumb trail pointing to the original node: 1026 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1027 dest->set_debug_orig(source); 1028 } 1029 #endif 1030 1031 if (node_note_array() == NULL) 1032 return false; // Not collecting any notes now. 1033 1034 // This is a copy onto a pre-existing node, which may already have notes. 1035 // If both nodes have notes, do not overwrite any pre-existing notes. 1036 Node_Notes* source_notes = node_notes_at(source->_idx); 1037 if (source_notes == NULL || source_notes->is_clear()) return false; 1038 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1039 if (dest_notes == NULL || dest_notes->is_clear()) { 1040 return set_node_notes_at(dest->_idx, source_notes); 1041 } 1042 1043 Node_Notes merged_notes = (*source_notes); 1044 // The order of operations here ensures that dest notes will win... 1045 merged_notes.update_from(dest_notes); 1046 return set_node_notes_at(dest->_idx, &merged_notes); 1047 } 1048 1049 1050 //--------------------------allow_range_check_smearing------------------------- 1051 // Gating condition for coalescing similar range checks. 1052 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1053 // single covering check that is at least as strong as any of them. 1054 // If the optimization succeeds, the simplified (strengthened) range check 1055 // will always succeed. If it fails, we will deopt, and then give up 1056 // on the optimization. 1057 bool Compile::allow_range_check_smearing() const { 1058 // If this method has already thrown a range-check, 1059 // assume it was because we already tried range smearing 1060 // and it failed. 1061 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1062 return !already_trapped; 1063 } 1064 1065 1066 //------------------------------flatten_alias_type----------------------------- 1067 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1068 int offset = tj->offset(); 1069 TypePtr::PTR ptr = tj->ptr(); 1070 1071 // Known instance (scalarizable allocation) alias only with itself. 1072 bool is_known_inst = tj->isa_oopptr() != NULL && 1073 tj->is_oopptr()->is_known_instance(); 1074 1075 // Process weird unsafe references. 1076 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1077 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1078 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1079 tj = TypeOopPtr::BOTTOM; 1080 ptr = tj->ptr(); 1081 offset = tj->offset(); 1082 } 1083 1084 // Array pointers need some flattening 1085 const TypeAryPtr *ta = tj->isa_aryptr(); 1086 if( ta && is_known_inst ) { 1087 if ( offset != Type::OffsetBot && 1088 offset > arrayOopDesc::length_offset_in_bytes() ) { 1089 offset = Type::OffsetBot; // Flatten constant access into array body only 1090 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1091 } 1092 } else if( ta && _AliasLevel >= 2 ) { 1093 // For arrays indexed by constant indices, we flatten the alias 1094 // space to include all of the array body. Only the header, klass 1095 // and array length can be accessed un-aliased. 1096 if( offset != Type::OffsetBot ) { 1097 if( ta->const_oop() ) { // methodDataOop or methodOop 1098 offset = Type::OffsetBot; // Flatten constant access into array body 1099 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1100 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1101 // range is OK as-is. 1102 tj = ta = TypeAryPtr::RANGE; 1103 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1104 tj = TypeInstPtr::KLASS; // all klass loads look alike 1105 ta = TypeAryPtr::RANGE; // generic ignored junk 1106 ptr = TypePtr::BotPTR; 1107 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1108 tj = TypeInstPtr::MARK; 1109 ta = TypeAryPtr::RANGE; // generic ignored junk 1110 ptr = TypePtr::BotPTR; 1111 } else { // Random constant offset into array body 1112 offset = Type::OffsetBot; // Flatten constant access into array body 1113 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1114 } 1115 } 1116 // Arrays of fixed size alias with arrays of unknown size. 1117 if (ta->size() != TypeInt::POS) { 1118 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1119 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1120 } 1121 // Arrays of known objects become arrays of unknown objects. 1122 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1123 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1124 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1125 } 1126 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1127 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1128 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1129 } 1130 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1131 // cannot be distinguished by bytecode alone. 1132 if (ta->elem() == TypeInt::BOOL) { 1133 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1134 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1135 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1136 } 1137 // During the 2nd round of IterGVN, NotNull castings are removed. 1138 // Make sure the Bottom and NotNull variants alias the same. 1139 // Also, make sure exact and non-exact variants alias the same. 1140 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1141 if (ta->const_oop()) { 1142 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1143 } else { 1144 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1145 } 1146 } 1147 } 1148 1149 // Oop pointers need some flattening 1150 const TypeInstPtr *to = tj->isa_instptr(); 1151 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1152 if( ptr == TypePtr::Constant ) { 1153 // No constant oop pointers (such as Strings); they alias with 1154 // unknown strings. 1155 assert(!is_known_inst, "not scalarizable allocation"); 1156 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1157 } else if( is_known_inst ) { 1158 tj = to; // Keep NotNull and klass_is_exact for instance type 1159 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1160 // During the 2nd round of IterGVN, NotNull castings are removed. 1161 // Make sure the Bottom and NotNull variants alias the same. 1162 // Also, make sure exact and non-exact variants alias the same. 1163 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1164 } 1165 // Canonicalize the holder of this field 1166 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1167 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1168 // First handle header references such as a LoadKlassNode, even if the 1169 // object's klass is unloaded at compile time (4965979). 1170 if (!is_known_inst) { // Do it only for non-instance types 1171 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1172 } 1173 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1174 to = NULL; 1175 tj = TypeOopPtr::BOTTOM; 1176 offset = tj->offset(); 1177 } else { 1178 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1179 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1180 if( is_known_inst ) { 1181 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1182 } else { 1183 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1184 } 1185 } 1186 } 1187 } 1188 1189 // Klass pointers to object array klasses need some flattening 1190 const TypeKlassPtr *tk = tj->isa_klassptr(); 1191 if( tk ) { 1192 // If we are referencing a field within a Klass, we need 1193 // to assume the worst case of an Object. Both exact and 1194 // inexact types must flatten to the same alias class. 1195 // Since the flattened result for a klass is defined to be 1196 // precisely java.lang.Object, use a constant ptr. 1197 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1198 1199 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1200 TypeKlassPtr::OBJECT->klass(), 1201 offset); 1202 } 1203 1204 ciKlass* klass = tk->klass(); 1205 if( klass->is_obj_array_klass() ) { 1206 ciKlass* k = TypeAryPtr::OOPS->klass(); 1207 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1208 k = TypeInstPtr::BOTTOM->klass(); 1209 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1210 } 1211 1212 // Check for precise loads from the primary supertype array and force them 1213 // to the supertype cache alias index. Check for generic array loads from 1214 // the primary supertype array and also force them to the supertype cache 1215 // alias index. Since the same load can reach both, we need to merge 1216 // these 2 disparate memories into the same alias class. Since the 1217 // primary supertype array is read-only, there's no chance of confusion 1218 // where we bypass an array load and an array store. 1219 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1220 if( offset == Type::OffsetBot || 1221 off2 < Klass::primary_super_limit()*wordSize ) { 1222 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1223 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1224 } 1225 } 1226 1227 // Flatten all Raw pointers together. 1228 if (tj->base() == Type::RawPtr) 1229 tj = TypeRawPtr::BOTTOM; 1230 1231 if (tj->base() == Type::AnyPtr) 1232 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1233 1234 // Flatten all to bottom for now 1235 switch( _AliasLevel ) { 1236 case 0: 1237 tj = TypePtr::BOTTOM; 1238 break; 1239 case 1: // Flatten to: oop, static, field or array 1240 switch (tj->base()) { 1241 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1242 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1243 case Type::AryPtr: // do not distinguish arrays at all 1244 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1245 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1246 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1247 default: ShouldNotReachHere(); 1248 } 1249 break; 1250 case 2: // No collapsing at level 2; keep all splits 1251 case 3: // No collapsing at level 3; keep all splits 1252 break; 1253 default: 1254 Unimplemented(); 1255 } 1256 1257 offset = tj->offset(); 1258 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1259 1260 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1261 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1262 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1263 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1264 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1265 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1266 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1267 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1268 assert( tj->ptr() != TypePtr::TopPTR && 1269 tj->ptr() != TypePtr::AnyNull && 1270 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1271 // assert( tj->ptr() != TypePtr::Constant || 1272 // tj->base() == Type::RawPtr || 1273 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1274 1275 return tj; 1276 } 1277 1278 void Compile::AliasType::Init(int i, const TypePtr* at) { 1279 _index = i; 1280 _adr_type = at; 1281 _field = NULL; 1282 _is_rewritable = true; // default 1283 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1284 if (atoop != NULL && atoop->is_known_instance()) { 1285 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1286 _general_index = Compile::current()->get_alias_index(gt); 1287 } else { 1288 _general_index = 0; 1289 } 1290 } 1291 1292 //---------------------------------print_on------------------------------------ 1293 #ifndef PRODUCT 1294 void Compile::AliasType::print_on(outputStream* st) { 1295 if (index() < 10) 1296 st->print("@ <%d> ", index()); 1297 else st->print("@ <%d>", index()); 1298 st->print(is_rewritable() ? " " : " RO"); 1299 int offset = adr_type()->offset(); 1300 if (offset == Type::OffsetBot) 1301 st->print(" +any"); 1302 else st->print(" +%-3d", offset); 1303 st->print(" in "); 1304 adr_type()->dump_on(st); 1305 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1306 if (field() != NULL && tjp) { 1307 if (tjp->klass() != field()->holder() || 1308 tjp->offset() != field()->offset_in_bytes()) { 1309 st->print(" != "); 1310 field()->print(); 1311 st->print(" ***"); 1312 } 1313 } 1314 } 1315 1316 void print_alias_types() { 1317 Compile* C = Compile::current(); 1318 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1319 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1320 C->alias_type(idx)->print_on(tty); 1321 tty->cr(); 1322 } 1323 } 1324 #endif 1325 1326 1327 //----------------------------probe_alias_cache-------------------------------- 1328 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1329 intptr_t key = (intptr_t) adr_type; 1330 key ^= key >> logAliasCacheSize; 1331 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1332 } 1333 1334 1335 //-----------------------------grow_alias_types-------------------------------- 1336 void Compile::grow_alias_types() { 1337 const int old_ats = _max_alias_types; // how many before? 1338 const int new_ats = old_ats; // how many more? 1339 const int grow_ats = old_ats+new_ats; // how many now? 1340 _max_alias_types = grow_ats; 1341 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1342 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1343 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1344 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1345 } 1346 1347 1348 //--------------------------------find_alias_type------------------------------ 1349 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1350 if (_AliasLevel == 0) 1351 return alias_type(AliasIdxBot); 1352 1353 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1354 if (ace->_adr_type == adr_type) { 1355 return alias_type(ace->_index); 1356 } 1357 1358 // Handle special cases. 1359 if (adr_type == NULL) return alias_type(AliasIdxTop); 1360 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1361 1362 // Do it the slow way. 1363 const TypePtr* flat = flatten_alias_type(adr_type); 1364 1365 #ifdef ASSERT 1366 assert(flat == flatten_alias_type(flat), "idempotent"); 1367 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1368 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1369 const TypeOopPtr* foop = flat->is_oopptr(); 1370 // Scalarizable allocations have exact klass always. 1371 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1372 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1373 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1374 } 1375 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1376 #endif 1377 1378 int idx = AliasIdxTop; 1379 for (int i = 0; i < num_alias_types(); i++) { 1380 if (alias_type(i)->adr_type() == flat) { 1381 idx = i; 1382 break; 1383 } 1384 } 1385 1386 if (idx == AliasIdxTop) { 1387 if (no_create) return NULL; 1388 // Grow the array if necessary. 1389 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1390 // Add a new alias type. 1391 idx = _num_alias_types++; 1392 _alias_types[idx]->Init(idx, flat); 1393 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1394 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1395 if (flat->isa_instptr()) { 1396 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1397 && flat->is_instptr()->klass() == env()->Class_klass()) 1398 alias_type(idx)->set_rewritable(false); 1399 } 1400 if (flat->isa_klassptr()) { 1401 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1402 alias_type(idx)->set_rewritable(false); 1403 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1404 alias_type(idx)->set_rewritable(false); 1405 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1406 alias_type(idx)->set_rewritable(false); 1407 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1408 alias_type(idx)->set_rewritable(false); 1409 } 1410 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1411 // but the base pointer type is not distinctive enough to identify 1412 // references into JavaThread.) 1413 1414 // Check for final instance fields. 1415 const TypeInstPtr* tinst = flat->isa_instptr(); 1416 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1417 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1418 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1419 // Set field() and is_rewritable() attributes. 1420 if (field != NULL) alias_type(idx)->set_field(field); 1421 } 1422 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1423 // Check for final static fields. 1424 if (tklass && tklass->klass()->is_instance_klass()) { 1425 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1426 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1427 // Set field() and is_rewritable() attributes. 1428 if (field != NULL) alias_type(idx)->set_field(field); 1429 } 1430 } 1431 1432 // Fill the cache for next time. 1433 ace->_adr_type = adr_type; 1434 ace->_index = idx; 1435 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1436 1437 // Might as well try to fill the cache for the flattened version, too. 1438 AliasCacheEntry* face = probe_alias_cache(flat); 1439 if (face->_adr_type == NULL) { 1440 face->_adr_type = flat; 1441 face->_index = idx; 1442 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1443 } 1444 1445 return alias_type(idx); 1446 } 1447 1448 1449 Compile::AliasType* Compile::alias_type(ciField* field) { 1450 const TypeOopPtr* t; 1451 if (field->is_static()) 1452 t = TypeKlassPtr::make(field->holder()); 1453 else 1454 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1455 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1456 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1457 return atp; 1458 } 1459 1460 1461 //------------------------------have_alias_type-------------------------------- 1462 bool Compile::have_alias_type(const TypePtr* adr_type) { 1463 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1464 if (ace->_adr_type == adr_type) { 1465 return true; 1466 } 1467 1468 // Handle special cases. 1469 if (adr_type == NULL) return true; 1470 if (adr_type == TypePtr::BOTTOM) return true; 1471 1472 return find_alias_type(adr_type, true) != NULL; 1473 } 1474 1475 //-----------------------------must_alias-------------------------------------- 1476 // True if all values of the given address type are in the given alias category. 1477 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1478 if (alias_idx == AliasIdxBot) return true; // the universal category 1479 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1480 if (alias_idx == AliasIdxTop) return false; // the empty category 1481 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1482 1483 // the only remaining possible overlap is identity 1484 int adr_idx = get_alias_index(adr_type); 1485 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1486 assert(adr_idx == alias_idx || 1487 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1488 && adr_type != TypeOopPtr::BOTTOM), 1489 "should not be testing for overlap with an unsafe pointer"); 1490 return adr_idx == alias_idx; 1491 } 1492 1493 //------------------------------can_alias-------------------------------------- 1494 // True if any values of the given address type are in the given alias category. 1495 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1496 if (alias_idx == AliasIdxTop) return false; // the empty category 1497 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1498 if (alias_idx == AliasIdxBot) return true; // the universal category 1499 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1500 1501 // the only remaining possible overlap is identity 1502 int adr_idx = get_alias_index(adr_type); 1503 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1504 return adr_idx == alias_idx; 1505 } 1506 1507 1508 1509 //---------------------------pop_warm_call------------------------------------- 1510 WarmCallInfo* Compile::pop_warm_call() { 1511 WarmCallInfo* wci = _warm_calls; 1512 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1513 return wci; 1514 } 1515 1516 //----------------------------Inline_Warm-------------------------------------- 1517 int Compile::Inline_Warm() { 1518 // If there is room, try to inline some more warm call sites. 1519 // %%% Do a graph index compaction pass when we think we're out of space? 1520 if (!InlineWarmCalls) return 0; 1521 1522 int calls_made_hot = 0; 1523 int room_to_grow = NodeCountInliningCutoff - unique(); 1524 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1525 int amount_grown = 0; 1526 WarmCallInfo* call; 1527 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1528 int est_size = (int)call->size(); 1529 if (est_size > (room_to_grow - amount_grown)) { 1530 // This one won't fit anyway. Get rid of it. 1531 call->make_cold(); 1532 continue; 1533 } 1534 call->make_hot(); 1535 calls_made_hot++; 1536 amount_grown += est_size; 1537 amount_to_grow -= est_size; 1538 } 1539 1540 if (calls_made_hot > 0) set_major_progress(); 1541 return calls_made_hot; 1542 } 1543 1544 1545 //----------------------------Finish_Warm-------------------------------------- 1546 void Compile::Finish_Warm() { 1547 if (!InlineWarmCalls) return; 1548 if (failing()) return; 1549 if (warm_calls() == NULL) return; 1550 1551 // Clean up loose ends, if we are out of space for inlining. 1552 WarmCallInfo* call; 1553 while ((call = pop_warm_call()) != NULL) { 1554 call->make_cold(); 1555 } 1556 } 1557 1558 //---------------------cleanup_loop_predicates----------------------- 1559 // Remove the opaque nodes that protect the predicates so that all unused 1560 // checks and uncommon_traps will be eliminated from the ideal graph 1561 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1562 if (predicate_count()==0) return; 1563 for (int i = predicate_count(); i > 0; i--) { 1564 Node * n = predicate_opaque1_node(i-1); 1565 assert(n->Opcode() == Op_Opaque1, "must be"); 1566 igvn.replace_node(n, n->in(1)); 1567 } 1568 assert(predicate_count()==0, "should be clean!"); 1569 igvn.optimize(); 1570 } 1571 1572 //------------------------------Optimize--------------------------------------- 1573 // Given a graph, optimize it. 1574 void Compile::Optimize() { 1575 TracePhase t1("optimizer", &_t_optimizer, true); 1576 1577 #ifndef PRODUCT 1578 if (env()->break_at_compile()) { 1579 BREAKPOINT; 1580 } 1581 1582 #endif 1583 1584 ResourceMark rm; 1585 int loop_opts_cnt; 1586 1587 NOT_PRODUCT( verify_graph_edges(); ) 1588 1589 print_method("After Parsing"); 1590 1591 { 1592 // Iterative Global Value Numbering, including ideal transforms 1593 // Initialize IterGVN with types and values from parse-time GVN 1594 PhaseIterGVN igvn(initial_gvn()); 1595 { 1596 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1597 igvn.optimize(); 1598 } 1599 1600 print_method("Iter GVN 1", 2); 1601 1602 if (failing()) return; 1603 1604 // Loop transforms on the ideal graph. Range Check Elimination, 1605 // peeling, unrolling, etc. 1606 1607 // Set loop opts counter 1608 loop_opts_cnt = num_loop_opts(); 1609 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1610 { 1611 TracePhase t2("idealLoop", &_t_idealLoop, true); 1612 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate); 1613 loop_opts_cnt--; 1614 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1615 if (failing()) return; 1616 } 1617 // Loop opts pass if partial peeling occurred in previous pass 1618 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1619 TracePhase t3("idealLoop", &_t_idealLoop, true); 1620 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1621 loop_opts_cnt--; 1622 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1623 if (failing()) return; 1624 } 1625 // Loop opts pass for loop-unrolling before CCP 1626 if(major_progress() && (loop_opts_cnt > 0)) { 1627 TracePhase t4("idealLoop", &_t_idealLoop, true); 1628 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1629 loop_opts_cnt--; 1630 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1631 } 1632 if (!failing()) { 1633 // Verify that last round of loop opts produced a valid graph 1634 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1635 PhaseIdealLoop::verify(igvn); 1636 } 1637 } 1638 if (failing()) return; 1639 1640 // Conditional Constant Propagation; 1641 PhaseCCP ccp( &igvn ); 1642 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1643 { 1644 TracePhase t2("ccp", &_t_ccp, true); 1645 ccp.do_transform(); 1646 } 1647 print_method("PhaseCPP 1", 2); 1648 1649 assert( true, "Break here to ccp.dump_old2new_map()"); 1650 1651 // Iterative Global Value Numbering, including ideal transforms 1652 { 1653 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1654 igvn = ccp; 1655 igvn.optimize(); 1656 } 1657 1658 print_method("Iter GVN 2", 2); 1659 1660 if (failing()) return; 1661 1662 // Loop transforms on the ideal graph. Range Check Elimination, 1663 // peeling, unrolling, etc. 1664 if(loop_opts_cnt > 0) { 1665 debug_only( int cnt = 0; ); 1666 bool loop_predication = UseLoopPredicate; 1667 while(major_progress() && (loop_opts_cnt > 0)) { 1668 TracePhase t2("idealLoop", &_t_idealLoop, true); 1669 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1670 PhaseIdealLoop ideal_loop( igvn, true, loop_predication); 1671 loop_opts_cnt--; 1672 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1673 if (failing()) return; 1674 // Perform loop predication optimization during first iteration after CCP. 1675 // After that switch it off and cleanup unused loop predicates. 1676 if (loop_predication) { 1677 loop_predication = false; 1678 cleanup_loop_predicates(igvn); 1679 if (failing()) return; 1680 } 1681 } 1682 } 1683 1684 { 1685 // Verify that all previous optimizations produced a valid graph 1686 // at least to this point, even if no loop optimizations were done. 1687 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1688 PhaseIdealLoop::verify(igvn); 1689 } 1690 1691 { 1692 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1693 PhaseMacroExpand mex(igvn); 1694 if (mex.expand_macro_nodes()) { 1695 assert(failing(), "must bail out w/ explicit message"); 1696 return; 1697 } 1698 } 1699 1700 } // (End scope of igvn; run destructor if necessary for asserts.) 1701 1702 // A method with only infinite loops has no edges entering loops from root 1703 { 1704 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1705 if (final_graph_reshaping()) { 1706 assert(failing(), "must bail out w/ explicit message"); 1707 return; 1708 } 1709 } 1710 1711 print_method("Optimize finished", 2); 1712 } 1713 1714 1715 //------------------------------Code_Gen--------------------------------------- 1716 // Given a graph, generate code for it 1717 void Compile::Code_Gen() { 1718 if (failing()) return; 1719 1720 // Perform instruction selection. You might think we could reclaim Matcher 1721 // memory PDQ, but actually the Matcher is used in generating spill code. 1722 // Internals of the Matcher (including some VectorSets) must remain live 1723 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1724 // set a bit in reclaimed memory. 1725 1726 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1727 // nodes. Mapping is only valid at the root of each matched subtree. 1728 NOT_PRODUCT( verify_graph_edges(); ) 1729 1730 Node_List proj_list; 1731 Matcher m(proj_list); 1732 _matcher = &m; 1733 { 1734 TracePhase t2("matcher", &_t_matcher, true); 1735 m.match(); 1736 } 1737 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1738 // nodes. Mapping is only valid at the root of each matched subtree. 1739 NOT_PRODUCT( verify_graph_edges(); ) 1740 1741 // If you have too many nodes, or if matching has failed, bail out 1742 check_node_count(0, "out of nodes matching instructions"); 1743 if (failing()) return; 1744 1745 // Build a proper-looking CFG 1746 PhaseCFG cfg(node_arena(), root(), m); 1747 _cfg = &cfg; 1748 { 1749 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1750 cfg.Dominators(); 1751 if (failing()) return; 1752 1753 NOT_PRODUCT( verify_graph_edges(); ) 1754 1755 cfg.Estimate_Block_Frequency(); 1756 cfg.GlobalCodeMotion(m,unique(),proj_list); 1757 1758 print_method("Global code motion", 2); 1759 1760 if (failing()) return; 1761 NOT_PRODUCT( verify_graph_edges(); ) 1762 1763 debug_only( cfg.verify(); ) 1764 } 1765 NOT_PRODUCT( verify_graph_edges(); ) 1766 1767 PhaseChaitin regalloc(unique(),cfg,m); 1768 _regalloc = ®alloc; 1769 { 1770 TracePhase t2("regalloc", &_t_registerAllocation, true); 1771 // Perform any platform dependent preallocation actions. This is used, 1772 // for example, to avoid taking an implicit null pointer exception 1773 // using the frame pointer on win95. 1774 _regalloc->pd_preallocate_hook(); 1775 1776 // Perform register allocation. After Chaitin, use-def chains are 1777 // no longer accurate (at spill code) and so must be ignored. 1778 // Node->LRG->reg mappings are still accurate. 1779 _regalloc->Register_Allocate(); 1780 1781 // Bail out if the allocator builds too many nodes 1782 if (failing()) return; 1783 } 1784 1785 // Prior to register allocation we kept empty basic blocks in case the 1786 // the allocator needed a place to spill. After register allocation we 1787 // are not adding any new instructions. If any basic block is empty, we 1788 // can now safely remove it. 1789 { 1790 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1791 cfg.remove_empty(); 1792 if (do_freq_based_layout()) { 1793 PhaseBlockLayout layout(cfg); 1794 } else { 1795 cfg.set_loop_alignment(); 1796 } 1797 cfg.fixup_flow(); 1798 } 1799 1800 // Perform any platform dependent postallocation verifications. 1801 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1802 1803 // Apply peephole optimizations 1804 if( OptoPeephole ) { 1805 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1806 PhasePeephole peep( _regalloc, cfg); 1807 peep.do_transform(); 1808 } 1809 1810 // Convert Nodes to instruction bits in a buffer 1811 { 1812 // %%%% workspace merge brought two timers together for one job 1813 TracePhase t2a("output", &_t_output, true); 1814 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1815 Output(); 1816 } 1817 1818 print_method("Final Code"); 1819 1820 // He's dead, Jim. 1821 _cfg = (PhaseCFG*)0xdeadbeef; 1822 _regalloc = (PhaseChaitin*)0xdeadbeef; 1823 } 1824 1825 1826 //------------------------------dump_asm--------------------------------------- 1827 // Dump formatted assembly 1828 #ifndef PRODUCT 1829 void Compile::dump_asm(int *pcs, uint pc_limit) { 1830 bool cut_short = false; 1831 tty->print_cr("#"); 1832 tty->print("# "); _tf->dump(); tty->cr(); 1833 tty->print_cr("#"); 1834 1835 // For all blocks 1836 int pc = 0x0; // Program counter 1837 char starts_bundle = ' '; 1838 _regalloc->dump_frame(); 1839 1840 Node *n = NULL; 1841 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1842 if (VMThread::should_terminate()) { cut_short = true; break; } 1843 Block *b = _cfg->_blocks[i]; 1844 if (b->is_connector() && !Verbose) continue; 1845 n = b->_nodes[0]; 1846 if (pcs && n->_idx < pc_limit) 1847 tty->print("%3.3x ", pcs[n->_idx]); 1848 else 1849 tty->print(" "); 1850 b->dump_head( &_cfg->_bbs ); 1851 if (b->is_connector()) { 1852 tty->print_cr(" # Empty connector block"); 1853 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1854 tty->print_cr(" # Block is sole successor of call"); 1855 } 1856 1857 // For all instructions 1858 Node *delay = NULL; 1859 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1860 if (VMThread::should_terminate()) { cut_short = true; break; } 1861 n = b->_nodes[j]; 1862 if (valid_bundle_info(n)) { 1863 Bundle *bundle = node_bundling(n); 1864 if (bundle->used_in_unconditional_delay()) { 1865 delay = n; 1866 continue; 1867 } 1868 if (bundle->starts_bundle()) 1869 starts_bundle = '+'; 1870 } 1871 1872 if (WizardMode) n->dump(); 1873 1874 if( !n->is_Region() && // Dont print in the Assembly 1875 !n->is_Phi() && // a few noisely useless nodes 1876 !n->is_Proj() && 1877 !n->is_MachTemp() && 1878 !n->is_SafePointScalarObject() && 1879 !n->is_Catch() && // Would be nice to print exception table targets 1880 !n->is_MergeMem() && // Not very interesting 1881 !n->is_top() && // Debug info table constants 1882 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1883 ) { 1884 if (pcs && n->_idx < pc_limit) 1885 tty->print("%3.3x", pcs[n->_idx]); 1886 else 1887 tty->print(" "); 1888 tty->print(" %c ", starts_bundle); 1889 starts_bundle = ' '; 1890 tty->print("\t"); 1891 n->format(_regalloc, tty); 1892 tty->cr(); 1893 } 1894 1895 // If we have an instruction with a delay slot, and have seen a delay, 1896 // then back up and print it 1897 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1898 assert(delay != NULL, "no unconditional delay instruction"); 1899 if (WizardMode) delay->dump(); 1900 1901 if (node_bundling(delay)->starts_bundle()) 1902 starts_bundle = '+'; 1903 if (pcs && n->_idx < pc_limit) 1904 tty->print("%3.3x", pcs[n->_idx]); 1905 else 1906 tty->print(" "); 1907 tty->print(" %c ", starts_bundle); 1908 starts_bundle = ' '; 1909 tty->print("\t"); 1910 delay->format(_regalloc, tty); 1911 tty->print_cr(""); 1912 delay = NULL; 1913 } 1914 1915 // Dump the exception table as well 1916 if( n->is_Catch() && (Verbose || WizardMode) ) { 1917 // Print the exception table for this offset 1918 _handler_table.print_subtable_for(pc); 1919 } 1920 } 1921 1922 if (pcs && n->_idx < pc_limit) 1923 tty->print_cr("%3.3x", pcs[n->_idx]); 1924 else 1925 tty->print_cr(""); 1926 1927 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1928 1929 } // End of per-block dump 1930 tty->print_cr(""); 1931 1932 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1933 } 1934 #endif 1935 1936 //------------------------------Final_Reshape_Counts--------------------------- 1937 // This class defines counters to help identify when a method 1938 // may/must be executed using hardware with only 24-bit precision. 1939 struct Final_Reshape_Counts : public StackObj { 1940 int _call_count; // count non-inlined 'common' calls 1941 int _float_count; // count float ops requiring 24-bit precision 1942 int _double_count; // count double ops requiring more precision 1943 int _java_call_count; // count non-inlined 'java' calls 1944 int _inner_loop_count; // count loops which need alignment 1945 VectorSet _visited; // Visitation flags 1946 Node_List _tests; // Set of IfNodes & PCTableNodes 1947 1948 Final_Reshape_Counts() : 1949 _call_count(0), _float_count(0), _double_count(0), 1950 _java_call_count(0), _inner_loop_count(0), 1951 _visited( Thread::current()->resource_area() ) { } 1952 1953 void inc_call_count () { _call_count ++; } 1954 void inc_float_count () { _float_count ++; } 1955 void inc_double_count() { _double_count++; } 1956 void inc_java_call_count() { _java_call_count++; } 1957 void inc_inner_loop_count() { _inner_loop_count++; } 1958 1959 int get_call_count () const { return _call_count ; } 1960 int get_float_count () const { return _float_count ; } 1961 int get_double_count() const { return _double_count; } 1962 int get_java_call_count() const { return _java_call_count; } 1963 int get_inner_loop_count() const { return _inner_loop_count; } 1964 }; 1965 1966 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1967 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1968 // Make sure the offset goes inside the instance layout. 1969 return k->contains_field_offset(tp->offset()); 1970 // Note that OffsetBot and OffsetTop are very negative. 1971 } 1972 1973 //------------------------------final_graph_reshaping_impl---------------------- 1974 // Implement items 1-5 from final_graph_reshaping below. 1975 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 1976 1977 if ( n->outcnt() == 0 ) return; // dead node 1978 uint nop = n->Opcode(); 1979 1980 // Check for 2-input instruction with "last use" on right input. 1981 // Swap to left input. Implements item (2). 1982 if( n->req() == 3 && // two-input instruction 1983 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1984 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1985 n->in(2)->outcnt() == 1 &&// right use IS a last use 1986 !n->in(2)->is_Con() ) { // right use is not a constant 1987 // Check for commutative opcode 1988 switch( nop ) { 1989 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1990 case Op_MaxI: case Op_MinI: 1991 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1992 case Op_AndL: case Op_XorL: case Op_OrL: 1993 case Op_AndI: case Op_XorI: case Op_OrI: { 1994 // Move "last use" input to left by swapping inputs 1995 n->swap_edges(1, 2); 1996 break; 1997 } 1998 default: 1999 break; 2000 } 2001 } 2002 2003 // Count FPU ops and common calls, implements item (3) 2004 switch( nop ) { 2005 // Count all float operations that may use FPU 2006 case Op_AddF: 2007 case Op_SubF: 2008 case Op_MulF: 2009 case Op_DivF: 2010 case Op_NegF: 2011 case Op_ModF: 2012 case Op_ConvI2F: 2013 case Op_ConF: 2014 case Op_CmpF: 2015 case Op_CmpF3: 2016 // case Op_ConvL2F: // longs are split into 32-bit halves 2017 frc.inc_float_count(); 2018 break; 2019 2020 case Op_ConvF2D: 2021 case Op_ConvD2F: 2022 frc.inc_float_count(); 2023 frc.inc_double_count(); 2024 break; 2025 2026 // Count all double operations that may use FPU 2027 case Op_AddD: 2028 case Op_SubD: 2029 case Op_MulD: 2030 case Op_DivD: 2031 case Op_NegD: 2032 case Op_ModD: 2033 case Op_ConvI2D: 2034 case Op_ConvD2I: 2035 // case Op_ConvL2D: // handled by leaf call 2036 // case Op_ConvD2L: // handled by leaf call 2037 case Op_ConD: 2038 case Op_CmpD: 2039 case Op_CmpD3: 2040 frc.inc_double_count(); 2041 break; 2042 case Op_Opaque1: // Remove Opaque Nodes before matching 2043 case Op_Opaque2: // Remove Opaque Nodes before matching 2044 n->subsume_by(n->in(1)); 2045 break; 2046 case Op_CallStaticJava: 2047 case Op_CallJava: 2048 case Op_CallDynamicJava: 2049 frc.inc_java_call_count(); // Count java call site; 2050 case Op_CallRuntime: 2051 case Op_CallLeaf: 2052 case Op_CallLeafNoFP: { 2053 assert( n->is_Call(), "" ); 2054 CallNode *call = n->as_Call(); 2055 // Count call sites where the FP mode bit would have to be flipped. 2056 // Do not count uncommon runtime calls: 2057 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2058 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2059 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2060 frc.inc_call_count(); // Count the call site 2061 } else { // See if uncommon argument is shared 2062 Node *n = call->in(TypeFunc::Parms); 2063 int nop = n->Opcode(); 2064 // Clone shared simple arguments to uncommon calls, item (1). 2065 if( n->outcnt() > 1 && 2066 !n->is_Proj() && 2067 nop != Op_CreateEx && 2068 nop != Op_CheckCastPP && 2069 nop != Op_DecodeN && 2070 !n->is_Mem() ) { 2071 Node *x = n->clone(); 2072 call->set_req( TypeFunc::Parms, x ); 2073 } 2074 } 2075 break; 2076 } 2077 2078 case Op_StoreD: 2079 case Op_LoadD: 2080 case Op_LoadD_unaligned: 2081 frc.inc_double_count(); 2082 goto handle_mem; 2083 case Op_StoreF: 2084 case Op_LoadF: 2085 frc.inc_float_count(); 2086 goto handle_mem; 2087 2088 case Op_StoreB: 2089 case Op_StoreC: 2090 case Op_StoreCM: 2091 case Op_StorePConditional: 2092 case Op_StoreI: 2093 case Op_StoreL: 2094 case Op_StoreIConditional: 2095 case Op_StoreLConditional: 2096 case Op_CompareAndSwapI: 2097 case Op_CompareAndSwapL: 2098 case Op_CompareAndSwapP: 2099 case Op_CompareAndSwapN: 2100 case Op_StoreP: 2101 case Op_StoreN: 2102 case Op_LoadB: 2103 case Op_LoadUB: 2104 case Op_LoadUS: 2105 case Op_LoadI: 2106 case Op_LoadUI2L: 2107 case Op_LoadKlass: 2108 case Op_LoadNKlass: 2109 case Op_LoadL: 2110 case Op_LoadL_unaligned: 2111 case Op_LoadPLocked: 2112 case Op_LoadLLocked: 2113 case Op_LoadP: 2114 case Op_LoadN: 2115 case Op_LoadRange: 2116 case Op_LoadS: { 2117 handle_mem: 2118 #ifdef ASSERT 2119 if( VerifyOptoOopOffsets ) { 2120 assert( n->is_Mem(), "" ); 2121 MemNode *mem = (MemNode*)n; 2122 // Check to see if address types have grounded out somehow. 2123 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2124 assert( !tp || oop_offset_is_sane(tp), "" ); 2125 } 2126 #endif 2127 break; 2128 } 2129 2130 case Op_AddP: { // Assert sane base pointers 2131 Node *addp = n->in(AddPNode::Address); 2132 assert( !addp->is_AddP() || 2133 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2134 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2135 "Base pointers must match" ); 2136 #ifdef _LP64 2137 if (UseCompressedOops && 2138 addp->Opcode() == Op_ConP && 2139 addp == n->in(AddPNode::Base) && 2140 n->in(AddPNode::Offset)->is_Con()) { 2141 // Use addressing with narrow klass to load with offset on x86. 2142 // On sparc loading 32-bits constant and decoding it have less 2143 // instructions (4) then load 64-bits constant (7). 2144 // Do this transformation here since IGVN will convert ConN back to ConP. 2145 const Type* t = addp->bottom_type(); 2146 if (t->isa_oopptr()) { 2147 Node* nn = NULL; 2148 2149 // Look for existing ConN node of the same exact type. 2150 Compile* C = Compile::current(); 2151 Node* r = C->root(); 2152 uint cnt = r->outcnt(); 2153 for (uint i = 0; i < cnt; i++) { 2154 Node* m = r->raw_out(i); 2155 if (m!= NULL && m->Opcode() == Op_ConN && 2156 m->bottom_type()->make_ptr() == t) { 2157 nn = m; 2158 break; 2159 } 2160 } 2161 if (nn != NULL) { 2162 // Decode a narrow oop to match address 2163 // [R12 + narrow_oop_reg<<3 + offset] 2164 nn = new (C, 2) DecodeNNode(nn, t); 2165 n->set_req(AddPNode::Base, nn); 2166 n->set_req(AddPNode::Address, nn); 2167 if (addp->outcnt() == 0) { 2168 addp->disconnect_inputs(NULL); 2169 } 2170 } 2171 } 2172 } 2173 #endif 2174 break; 2175 } 2176 2177 #ifdef _LP64 2178 case Op_CastPP: 2179 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) { 2180 Compile* C = Compile::current(); 2181 Node* in1 = n->in(1); 2182 const Type* t = n->bottom_type(); 2183 Node* new_in1 = in1->clone(); 2184 new_in1->as_DecodeN()->set_type(t); 2185 2186 if (!Matcher::narrow_oop_use_complex_address()) { 2187 // 2188 // x86, ARM and friends can handle 2 adds in addressing mode 2189 // and Matcher can fold a DecodeN node into address by using 2190 // a narrow oop directly and do implicit NULL check in address: 2191 // 2192 // [R12 + narrow_oop_reg<<3 + offset] 2193 // NullCheck narrow_oop_reg 2194 // 2195 // On other platforms (Sparc) we have to keep new DecodeN node and 2196 // use it to do implicit NULL check in address: 2197 // 2198 // decode_not_null narrow_oop_reg, base_reg 2199 // [base_reg + offset] 2200 // NullCheck base_reg 2201 // 2202 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2203 // to keep the information to which NULL check the new DecodeN node 2204 // corresponds to use it as value in implicit_null_check(). 2205 // 2206 new_in1->set_req(0, n->in(0)); 2207 } 2208 2209 n->subsume_by(new_in1); 2210 if (in1->outcnt() == 0) { 2211 in1->disconnect_inputs(NULL); 2212 } 2213 } 2214 break; 2215 2216 case Op_CmpP: 2217 // Do this transformation here to preserve CmpPNode::sub() and 2218 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2219 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2220 Node* in1 = n->in(1); 2221 Node* in2 = n->in(2); 2222 if (!in1->is_DecodeN()) { 2223 in2 = in1; 2224 in1 = n->in(2); 2225 } 2226 assert(in1->is_DecodeN(), "sanity"); 2227 2228 Compile* C = Compile::current(); 2229 Node* new_in2 = NULL; 2230 if (in2->is_DecodeN()) { 2231 new_in2 = in2->in(1); 2232 } else if (in2->Opcode() == Op_ConP) { 2233 const Type* t = in2->bottom_type(); 2234 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) { 2235 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2236 // 2237 // This transformation together with CastPP transformation above 2238 // will generated code for implicit NULL checks for compressed oops. 2239 // 2240 // The original code after Optimize() 2241 // 2242 // LoadN memory, narrow_oop_reg 2243 // decode narrow_oop_reg, base_reg 2244 // CmpP base_reg, NULL 2245 // CastPP base_reg // NotNull 2246 // Load [base_reg + offset], val_reg 2247 // 2248 // after these transformations will be 2249 // 2250 // LoadN memory, narrow_oop_reg 2251 // CmpN narrow_oop_reg, NULL 2252 // decode_not_null narrow_oop_reg, base_reg 2253 // Load [base_reg + offset], val_reg 2254 // 2255 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2256 // since narrow oops can be used in debug info now (see the code in 2257 // final_graph_reshaping_walk()). 2258 // 2259 // At the end the code will be matched to 2260 // on x86: 2261 // 2262 // Load_narrow_oop memory, narrow_oop_reg 2263 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2264 // NullCheck narrow_oop_reg 2265 // 2266 // and on sparc: 2267 // 2268 // Load_narrow_oop memory, narrow_oop_reg 2269 // decode_not_null narrow_oop_reg, base_reg 2270 // Load [base_reg + offset], val_reg 2271 // NullCheck base_reg 2272 // 2273 } else if (t->isa_oopptr()) { 2274 new_in2 = ConNode::make(C, t->make_narrowoop()); 2275 } 2276 } 2277 if (new_in2 != NULL) { 2278 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2279 n->subsume_by( cmpN ); 2280 if (in1->outcnt() == 0) { 2281 in1->disconnect_inputs(NULL); 2282 } 2283 if (in2->outcnt() == 0) { 2284 in2->disconnect_inputs(NULL); 2285 } 2286 } 2287 } 2288 break; 2289 2290 case Op_DecodeN: 2291 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2292 // DecodeN could be pinned on Sparc where it can't be fold into 2293 // an address expression, see the code for Op_CastPP above. 2294 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control except on sparc"); 2295 break; 2296 2297 case Op_EncodeP: { 2298 Node* in1 = n->in(1); 2299 if (in1->is_DecodeN()) { 2300 n->subsume_by(in1->in(1)); 2301 } else if (in1->Opcode() == Op_ConP) { 2302 Compile* C = Compile::current(); 2303 const Type* t = in1->bottom_type(); 2304 if (t == TypePtr::NULL_PTR) { 2305 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2306 } else if (t->isa_oopptr()) { 2307 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2308 } 2309 } 2310 if (in1->outcnt() == 0) { 2311 in1->disconnect_inputs(NULL); 2312 } 2313 break; 2314 } 2315 2316 case Op_Proj: { 2317 if (OptimizeStringConcat) { 2318 ProjNode* p = n->as_Proj(); 2319 if (p->_is_io_use) { 2320 // Separate projections were used for the exception path which 2321 // are normally removed by a late inline. If it wasn't inlined 2322 // then they will hang around and should just be replaced with 2323 // the original one. 2324 Node* proj = NULL; 2325 // Replace with just one 2326 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2327 Node *use = i.get(); 2328 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2329 proj = use; 2330 break; 2331 } 2332 } 2333 assert(p != NULL, "must be found"); 2334 p->subsume_by(proj); 2335 } 2336 } 2337 break; 2338 } 2339 2340 case Op_Phi: 2341 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2342 // The EncodeP optimization may create Phi with the same edges 2343 // for all paths. It is not handled well by Register Allocator. 2344 Node* unique_in = n->in(1); 2345 assert(unique_in != NULL, ""); 2346 uint cnt = n->req(); 2347 for (uint i = 2; i < cnt; i++) { 2348 Node* m = n->in(i); 2349 assert(m != NULL, ""); 2350 if (unique_in != m) 2351 unique_in = NULL; 2352 } 2353 if (unique_in != NULL) { 2354 n->subsume_by(unique_in); 2355 } 2356 } 2357 break; 2358 2359 #endif 2360 2361 case Op_ModI: 2362 if (UseDivMod) { 2363 // Check if a%b and a/b both exist 2364 Node* d = n->find_similar(Op_DivI); 2365 if (d) { 2366 // Replace them with a fused divmod if supported 2367 Compile* C = Compile::current(); 2368 if (Matcher::has_match_rule(Op_DivModI)) { 2369 DivModINode* divmod = DivModINode::make(C, n); 2370 d->subsume_by(divmod->div_proj()); 2371 n->subsume_by(divmod->mod_proj()); 2372 } else { 2373 // replace a%b with a-((a/b)*b) 2374 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2375 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2376 n->subsume_by( sub ); 2377 } 2378 } 2379 } 2380 break; 2381 2382 case Op_ModL: 2383 if (UseDivMod) { 2384 // Check if a%b and a/b both exist 2385 Node* d = n->find_similar(Op_DivL); 2386 if (d) { 2387 // Replace them with a fused divmod if supported 2388 Compile* C = Compile::current(); 2389 if (Matcher::has_match_rule(Op_DivModL)) { 2390 DivModLNode* divmod = DivModLNode::make(C, n); 2391 d->subsume_by(divmod->div_proj()); 2392 n->subsume_by(divmod->mod_proj()); 2393 } else { 2394 // replace a%b with a-((a/b)*b) 2395 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2396 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2397 n->subsume_by( sub ); 2398 } 2399 } 2400 } 2401 break; 2402 2403 case Op_Load16B: 2404 case Op_Load8B: 2405 case Op_Load4B: 2406 case Op_Load8S: 2407 case Op_Load4S: 2408 case Op_Load2S: 2409 case Op_Load8C: 2410 case Op_Load4C: 2411 case Op_Load2C: 2412 case Op_Load4I: 2413 case Op_Load2I: 2414 case Op_Load2L: 2415 case Op_Load4F: 2416 case Op_Load2F: 2417 case Op_Load2D: 2418 case Op_Store16B: 2419 case Op_Store8B: 2420 case Op_Store4B: 2421 case Op_Store8C: 2422 case Op_Store4C: 2423 case Op_Store2C: 2424 case Op_Store4I: 2425 case Op_Store2I: 2426 case Op_Store2L: 2427 case Op_Store4F: 2428 case Op_Store2F: 2429 case Op_Store2D: 2430 break; 2431 2432 case Op_PackB: 2433 case Op_PackS: 2434 case Op_PackC: 2435 case Op_PackI: 2436 case Op_PackF: 2437 case Op_PackL: 2438 case Op_PackD: 2439 if (n->req()-1 > 2) { 2440 // Replace many operand PackNodes with a binary tree for matching 2441 PackNode* p = (PackNode*) n; 2442 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2443 n->subsume_by(btp); 2444 } 2445 break; 2446 case Op_Loop: 2447 case Op_CountedLoop: 2448 if (n->as_Loop()->is_inner_loop()) { 2449 frc.inc_inner_loop_count(); 2450 } 2451 break; 2452 default: 2453 assert( !n->is_Call(), "" ); 2454 assert( !n->is_Mem(), "" ); 2455 break; 2456 } 2457 2458 // Collect CFG split points 2459 if (n->is_MultiBranch()) 2460 frc._tests.push(n); 2461 } 2462 2463 //------------------------------final_graph_reshaping_walk--------------------- 2464 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2465 // requires that the walk visits a node's inputs before visiting the node. 2466 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2467 ResourceArea *area = Thread::current()->resource_area(); 2468 Unique_Node_List sfpt(area); 2469 2470 frc._visited.set(root->_idx); // first, mark node as visited 2471 uint cnt = root->req(); 2472 Node *n = root; 2473 uint i = 0; 2474 while (true) { 2475 if (i < cnt) { 2476 // Place all non-visited non-null inputs onto stack 2477 Node* m = n->in(i); 2478 ++i; 2479 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2480 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2481 sfpt.push(m); 2482 cnt = m->req(); 2483 nstack.push(n, i); // put on stack parent and next input's index 2484 n = m; 2485 i = 0; 2486 } 2487 } else { 2488 // Now do post-visit work 2489 final_graph_reshaping_impl( n, frc ); 2490 if (nstack.is_empty()) 2491 break; // finished 2492 n = nstack.node(); // Get node from stack 2493 cnt = n->req(); 2494 i = nstack.index(); 2495 nstack.pop(); // Shift to the next node on stack 2496 } 2497 } 2498 2499 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2500 // It could be done for an uncommon traps or any safepoints/calls 2501 // if the DecodeN node is referenced only in a debug info. 2502 while (sfpt.size() > 0) { 2503 n = sfpt.pop(); 2504 JVMState *jvms = n->as_SafePoint()->jvms(); 2505 assert(jvms != NULL, "sanity"); 2506 int start = jvms->debug_start(); 2507 int end = n->req(); 2508 bool is_uncommon = (n->is_CallStaticJava() && 2509 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2510 for (int j = start; j < end; j++) { 2511 Node* in = n->in(j); 2512 if (in->is_DecodeN()) { 2513 bool safe_to_skip = true; 2514 if (!is_uncommon ) { 2515 // Is it safe to skip? 2516 for (uint i = 0; i < in->outcnt(); i++) { 2517 Node* u = in->raw_out(i); 2518 if (!u->is_SafePoint() || 2519 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2520 safe_to_skip = false; 2521 } 2522 } 2523 } 2524 if (safe_to_skip) { 2525 n->set_req(j, in->in(1)); 2526 } 2527 if (in->outcnt() == 0) { 2528 in->disconnect_inputs(NULL); 2529 } 2530 } 2531 } 2532 } 2533 } 2534 2535 //------------------------------final_graph_reshaping-------------------------- 2536 // Final Graph Reshaping. 2537 // 2538 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2539 // and not commoned up and forced early. Must come after regular 2540 // optimizations to avoid GVN undoing the cloning. Clone constant 2541 // inputs to Loop Phis; these will be split by the allocator anyways. 2542 // Remove Opaque nodes. 2543 // (2) Move last-uses by commutative operations to the left input to encourage 2544 // Intel update-in-place two-address operations and better register usage 2545 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2546 // calls canonicalizing them back. 2547 // (3) Count the number of double-precision FP ops, single-precision FP ops 2548 // and call sites. On Intel, we can get correct rounding either by 2549 // forcing singles to memory (requires extra stores and loads after each 2550 // FP bytecode) or we can set a rounding mode bit (requires setting and 2551 // clearing the mode bit around call sites). The mode bit is only used 2552 // if the relative frequency of single FP ops to calls is low enough. 2553 // This is a key transform for SPEC mpeg_audio. 2554 // (4) Detect infinite loops; blobs of code reachable from above but not 2555 // below. Several of the Code_Gen algorithms fail on such code shapes, 2556 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2557 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2558 // Detection is by looking for IfNodes where only 1 projection is 2559 // reachable from below or CatchNodes missing some targets. 2560 // (5) Assert for insane oop offsets in debug mode. 2561 2562 bool Compile::final_graph_reshaping() { 2563 // an infinite loop may have been eliminated by the optimizer, 2564 // in which case the graph will be empty. 2565 if (root()->req() == 1) { 2566 record_method_not_compilable("trivial infinite loop"); 2567 return true; 2568 } 2569 2570 Final_Reshape_Counts frc; 2571 2572 // Visit everybody reachable! 2573 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2574 Node_Stack nstack(unique() >> 1); 2575 final_graph_reshaping_walk(nstack, root(), frc); 2576 2577 // Check for unreachable (from below) code (i.e., infinite loops). 2578 for( uint i = 0; i < frc._tests.size(); i++ ) { 2579 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2580 // Get number of CFG targets. 2581 // Note that PCTables include exception targets after calls. 2582 uint required_outcnt = n->required_outcnt(); 2583 if (n->outcnt() != required_outcnt) { 2584 // Check for a few special cases. Rethrow Nodes never take the 2585 // 'fall-thru' path, so expected kids is 1 less. 2586 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2587 if (n->in(0)->in(0)->is_Call()) { 2588 CallNode *call = n->in(0)->in(0)->as_Call(); 2589 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2590 required_outcnt--; // Rethrow always has 1 less kid 2591 } else if (call->req() > TypeFunc::Parms && 2592 call->is_CallDynamicJava()) { 2593 // Check for null receiver. In such case, the optimizer has 2594 // detected that the virtual call will always result in a null 2595 // pointer exception. The fall-through projection of this CatchNode 2596 // will not be populated. 2597 Node *arg0 = call->in(TypeFunc::Parms); 2598 if (arg0->is_Type() && 2599 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2600 required_outcnt--; 2601 } 2602 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2603 call->req() > TypeFunc::Parms+1 && 2604 call->is_CallStaticJava()) { 2605 // Check for negative array length. In such case, the optimizer has 2606 // detected that the allocation attempt will always result in an 2607 // exception. There is no fall-through projection of this CatchNode . 2608 Node *arg1 = call->in(TypeFunc::Parms+1); 2609 if (arg1->is_Type() && 2610 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2611 required_outcnt--; 2612 } 2613 } 2614 } 2615 } 2616 // Recheck with a better notion of 'required_outcnt' 2617 if (n->outcnt() != required_outcnt) { 2618 record_method_not_compilable("malformed control flow"); 2619 return true; // Not all targets reachable! 2620 } 2621 } 2622 // Check that I actually visited all kids. Unreached kids 2623 // must be infinite loops. 2624 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2625 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2626 record_method_not_compilable("infinite loop"); 2627 return true; // Found unvisited kid; must be unreach 2628 } 2629 } 2630 2631 // If original bytecodes contained a mixture of floats and doubles 2632 // check if the optimizer has made it homogenous, item (3). 2633 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2634 frc.get_float_count() > 32 && 2635 frc.get_double_count() == 0 && 2636 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2637 set_24_bit_selection_and_mode( false, true ); 2638 } 2639 2640 set_java_calls(frc.get_java_call_count()); 2641 set_inner_loops(frc.get_inner_loop_count()); 2642 2643 // No infinite loops, no reason to bail out. 2644 return false; 2645 } 2646 2647 //-----------------------------too_many_traps---------------------------------- 2648 // Report if there are too many traps at the current method and bci. 2649 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2650 bool Compile::too_many_traps(ciMethod* method, 2651 int bci, 2652 Deoptimization::DeoptReason reason) { 2653 ciMethodData* md = method->method_data(); 2654 if (md->is_empty()) { 2655 // Assume the trap has not occurred, or that it occurred only 2656 // because of a transient condition during start-up in the interpreter. 2657 return false; 2658 } 2659 if (md->has_trap_at(bci, reason) != 0) { 2660 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2661 // Also, if there are multiple reasons, or if there is no per-BCI record, 2662 // assume the worst. 2663 if (log()) 2664 log()->elem("observe trap='%s' count='%d'", 2665 Deoptimization::trap_reason_name(reason), 2666 md->trap_count(reason)); 2667 return true; 2668 } else { 2669 // Ignore method/bci and see if there have been too many globally. 2670 return too_many_traps(reason, md); 2671 } 2672 } 2673 2674 // Less-accurate variant which does not require a method and bci. 2675 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2676 ciMethodData* logmd) { 2677 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2678 // Too many traps globally. 2679 // Note that we use cumulative trap_count, not just md->trap_count. 2680 if (log()) { 2681 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2682 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2683 Deoptimization::trap_reason_name(reason), 2684 mcount, trap_count(reason)); 2685 } 2686 return true; 2687 } else { 2688 // The coast is clear. 2689 return false; 2690 } 2691 } 2692 2693 //--------------------------too_many_recompiles-------------------------------- 2694 // Report if there are too many recompiles at the current method and bci. 2695 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2696 // Is not eager to return true, since this will cause the compiler to use 2697 // Action_none for a trap point, to avoid too many recompilations. 2698 bool Compile::too_many_recompiles(ciMethod* method, 2699 int bci, 2700 Deoptimization::DeoptReason reason) { 2701 ciMethodData* md = method->method_data(); 2702 if (md->is_empty()) { 2703 // Assume the trap has not occurred, or that it occurred only 2704 // because of a transient condition during start-up in the interpreter. 2705 return false; 2706 } 2707 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2708 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2709 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2710 Deoptimization::DeoptReason per_bc_reason 2711 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2712 if ((per_bc_reason == Deoptimization::Reason_none 2713 || md->has_trap_at(bci, reason) != 0) 2714 // The trap frequency measure we care about is the recompile count: 2715 && md->trap_recompiled_at(bci) 2716 && md->overflow_recompile_count() >= bc_cutoff) { 2717 // Do not emit a trap here if it has already caused recompilations. 2718 // Also, if there are multiple reasons, or if there is no per-BCI record, 2719 // assume the worst. 2720 if (log()) 2721 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2722 Deoptimization::trap_reason_name(reason), 2723 md->trap_count(reason), 2724 md->overflow_recompile_count()); 2725 return true; 2726 } else if (trap_count(reason) != 0 2727 && decompile_count() >= m_cutoff) { 2728 // Too many recompiles globally, and we have seen this sort of trap. 2729 // Use cumulative decompile_count, not just md->decompile_count. 2730 if (log()) 2731 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2732 Deoptimization::trap_reason_name(reason), 2733 md->trap_count(reason), trap_count(reason), 2734 md->decompile_count(), decompile_count()); 2735 return true; 2736 } else { 2737 // The coast is clear. 2738 return false; 2739 } 2740 } 2741 2742 2743 #ifndef PRODUCT 2744 //------------------------------verify_graph_edges--------------------------- 2745 // Walk the Graph and verify that there is a one-to-one correspondence 2746 // between Use-Def edges and Def-Use edges in the graph. 2747 void Compile::verify_graph_edges(bool no_dead_code) { 2748 if (VerifyGraphEdges) { 2749 ResourceArea *area = Thread::current()->resource_area(); 2750 Unique_Node_List visited(area); 2751 // Call recursive graph walk to check edges 2752 _root->verify_edges(visited); 2753 if (no_dead_code) { 2754 // Now make sure that no visited node is used by an unvisited node. 2755 bool dead_nodes = 0; 2756 Unique_Node_List checked(area); 2757 while (visited.size() > 0) { 2758 Node* n = visited.pop(); 2759 checked.push(n); 2760 for (uint i = 0; i < n->outcnt(); i++) { 2761 Node* use = n->raw_out(i); 2762 if (checked.member(use)) continue; // already checked 2763 if (visited.member(use)) continue; // already in the graph 2764 if (use->is_Con()) continue; // a dead ConNode is OK 2765 // At this point, we have found a dead node which is DU-reachable. 2766 if (dead_nodes++ == 0) 2767 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2768 use->dump(2); 2769 tty->print_cr("---"); 2770 checked.push(use); // No repeats; pretend it is now checked. 2771 } 2772 } 2773 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2774 } 2775 } 2776 } 2777 #endif 2778 2779 // The Compile object keeps track of failure reasons separately from the ciEnv. 2780 // This is required because there is not quite a 1-1 relation between the 2781 // ciEnv and its compilation task and the Compile object. Note that one 2782 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2783 // to backtrack and retry without subsuming loads. Other than this backtracking 2784 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2785 // by the logic in C2Compiler. 2786 void Compile::record_failure(const char* reason) { 2787 if (log() != NULL) { 2788 log()->elem("failure reason='%s' phase='compile'", reason); 2789 } 2790 if (_failure_reason == NULL) { 2791 // Record the first failure reason. 2792 _failure_reason = reason; 2793 } 2794 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2795 C->print_method(_failure_reason); 2796 } 2797 _root = NULL; // flush the graph, too 2798 } 2799 2800 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2801 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2802 { 2803 if (dolog) { 2804 C = Compile::current(); 2805 _log = C->log(); 2806 } else { 2807 C = NULL; 2808 _log = NULL; 2809 } 2810 if (_log != NULL) { 2811 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2812 _log->stamp(); 2813 _log->end_head(); 2814 } 2815 } 2816 2817 Compile::TracePhase::~TracePhase() { 2818 if (_log != NULL) { 2819 _log->done("phase nodes='%d'", C->unique()); 2820 } 2821 }