1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_compile.cpp.incl" 27 28 /// Support for intrinsics. 29 30 // Return the index at which m must be inserted (or already exists). 31 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 33 #ifdef ASSERT 34 for (int i = 1; i < _intrinsics->length(); i++) { 35 CallGenerator* cg1 = _intrinsics->at(i-1); 36 CallGenerator* cg2 = _intrinsics->at(i); 37 assert(cg1->method() != cg2->method() 38 ? cg1->method() < cg2->method() 39 : cg1->is_virtual() < cg2->is_virtual(), 40 "compiler intrinsics list must stay sorted"); 41 } 42 #endif 43 // Binary search sorted list, in decreasing intervals [lo, hi]. 44 int lo = 0, hi = _intrinsics->length()-1; 45 while (lo <= hi) { 46 int mid = (uint)(hi + lo) / 2; 47 ciMethod* mid_m = _intrinsics->at(mid)->method(); 48 if (m < mid_m) { 49 hi = mid-1; 50 } else if (m > mid_m) { 51 lo = mid+1; 52 } else { 53 // look at minor sort key 54 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 55 if (is_virtual < mid_virt) { 56 hi = mid-1; 57 } else if (is_virtual > mid_virt) { 58 lo = mid+1; 59 } else { 60 return mid; // exact match 61 } 62 } 63 } 64 return lo; // inexact match 65 } 66 67 void Compile::register_intrinsic(CallGenerator* cg) { 68 if (_intrinsics == NULL) { 69 _intrinsics = new GrowableArray<CallGenerator*>(60); 70 } 71 // This code is stolen from ciObjectFactory::insert. 72 // Really, GrowableArray should have methods for 73 // insert_at, remove_at, and binary_search. 74 int len = _intrinsics->length(); 75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 76 if (index == len) { 77 _intrinsics->append(cg); 78 } else { 79 #ifdef ASSERT 80 CallGenerator* oldcg = _intrinsics->at(index); 81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 82 #endif 83 _intrinsics->append(_intrinsics->at(len-1)); 84 int pos; 85 for (pos = len-2; pos >= index; pos--) { 86 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 87 } 88 _intrinsics->at_put(index, cg); 89 } 90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 91 } 92 93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 94 assert(m->is_loaded(), "don't try this on unloaded methods"); 95 if (_intrinsics != NULL) { 96 int index = intrinsic_insertion_index(m, is_virtual); 97 if (index < _intrinsics->length() 98 && _intrinsics->at(index)->method() == m 99 && _intrinsics->at(index)->is_virtual() == is_virtual) { 100 return _intrinsics->at(index); 101 } 102 } 103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 104 if (m->intrinsic_id() != vmIntrinsics::_none) { 105 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 106 if (cg != NULL) { 107 // Save it for next time: 108 register_intrinsic(cg); 109 return cg; 110 } else { 111 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 112 } 113 } 114 return NULL; 115 } 116 117 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 118 // in library_call.cpp. 119 120 121 #ifndef PRODUCT 122 // statistics gathering... 123 124 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 125 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 126 127 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 128 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 129 int oflags = _intrinsic_hist_flags[id]; 130 assert(flags != 0, "what happened?"); 131 if (is_virtual) { 132 flags |= _intrinsic_virtual; 133 } 134 bool changed = (flags != oflags); 135 if ((flags & _intrinsic_worked) != 0) { 136 juint count = (_intrinsic_hist_count[id] += 1); 137 if (count == 1) { 138 changed = true; // first time 139 } 140 // increment the overall count also: 141 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 142 } 143 if (changed) { 144 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 145 // Something changed about the intrinsic's virtuality. 146 if ((flags & _intrinsic_virtual) != 0) { 147 // This is the first use of this intrinsic as a virtual call. 148 if (oflags != 0) { 149 // We already saw it as a non-virtual, so note both cases. 150 flags |= _intrinsic_both; 151 } 152 } else if ((oflags & _intrinsic_both) == 0) { 153 // This is the first use of this intrinsic as a non-virtual 154 flags |= _intrinsic_both; 155 } 156 } 157 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 158 } 159 // update the overall flags also: 160 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 161 return changed; 162 } 163 164 static char* format_flags(int flags, char* buf) { 165 buf[0] = 0; 166 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 167 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 168 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 169 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 170 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 171 if (buf[0] == 0) strcat(buf, ","); 172 assert(buf[0] == ',', "must be"); 173 return &buf[1]; 174 } 175 176 void Compile::print_intrinsic_statistics() { 177 char flagsbuf[100]; 178 ttyLocker ttyl; 179 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 180 tty->print_cr("Compiler intrinsic usage:"); 181 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 182 if (total == 0) total = 1; // avoid div0 in case of no successes 183 #define PRINT_STAT_LINE(name, c, f) \ 184 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 185 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 186 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 187 int flags = _intrinsic_hist_flags[id]; 188 juint count = _intrinsic_hist_count[id]; 189 if ((flags | count) != 0) { 190 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 191 } 192 } 193 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 194 if (xtty != NULL) xtty->tail("statistics"); 195 } 196 197 void Compile::print_statistics() { 198 { ttyLocker ttyl; 199 if (xtty != NULL) xtty->head("statistics type='opto'"); 200 Parse::print_statistics(); 201 PhaseCCP::print_statistics(); 202 PhaseRegAlloc::print_statistics(); 203 Scheduling::print_statistics(); 204 PhasePeephole::print_statistics(); 205 PhaseIdealLoop::print_statistics(); 206 if (xtty != NULL) xtty->tail("statistics"); 207 } 208 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 209 // put this under its own <statistics> element. 210 print_intrinsic_statistics(); 211 } 212 } 213 #endif //PRODUCT 214 215 // Support for bundling info 216 Bundle* Compile::node_bundling(const Node *n) { 217 assert(valid_bundle_info(n), "oob"); 218 return &_node_bundling_base[n->_idx]; 219 } 220 221 bool Compile::valid_bundle_info(const Node *n) { 222 return (_node_bundling_limit > n->_idx); 223 } 224 225 226 // Identify all nodes that are reachable from below, useful. 227 // Use breadth-first pass that records state in a Unique_Node_List, 228 // recursive traversal is slower. 229 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 230 int estimated_worklist_size = unique(); 231 useful.map( estimated_worklist_size, NULL ); // preallocate space 232 233 // Initialize worklist 234 if (root() != NULL) { useful.push(root()); } 235 // If 'top' is cached, declare it useful to preserve cached node 236 if( cached_top_node() ) { useful.push(cached_top_node()); } 237 238 // Push all useful nodes onto the list, breadthfirst 239 for( uint next = 0; next < useful.size(); ++next ) { 240 assert( next < unique(), "Unique useful nodes < total nodes"); 241 Node *n = useful.at(next); 242 uint max = n->len(); 243 for( uint i = 0; i < max; ++i ) { 244 Node *m = n->in(i); 245 if( m == NULL ) continue; 246 useful.push(m); 247 } 248 } 249 } 250 251 // Disconnect all useless nodes by disconnecting those at the boundary. 252 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 253 uint next = 0; 254 while( next < useful.size() ) { 255 Node *n = useful.at(next++); 256 // Use raw traversal of out edges since this code removes out edges 257 int max = n->outcnt(); 258 for (int j = 0; j < max; ++j ) { 259 Node* child = n->raw_out(j); 260 if( ! useful.member(child) ) { 261 assert( !child->is_top() || child != top(), 262 "If top is cached in Compile object it is in useful list"); 263 // Only need to remove this out-edge to the useless node 264 n->raw_del_out(j); 265 --j; 266 --max; 267 } 268 } 269 if (n->outcnt() == 1 && n->has_special_unique_user()) { 270 record_for_igvn( n->unique_out() ); 271 } 272 } 273 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 274 } 275 276 //------------------------------frame_size_in_words----------------------------- 277 // frame_slots in units of words 278 int Compile::frame_size_in_words() const { 279 // shift is 0 in LP32 and 1 in LP64 280 const int shift = (LogBytesPerWord - LogBytesPerInt); 281 int words = _frame_slots >> shift; 282 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 283 return words; 284 } 285 286 // ============================================================================ 287 //------------------------------CompileWrapper--------------------------------- 288 class CompileWrapper : public StackObj { 289 Compile *const _compile; 290 public: 291 CompileWrapper(Compile* compile); 292 293 ~CompileWrapper(); 294 }; 295 296 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 297 // the Compile* pointer is stored in the current ciEnv: 298 ciEnv* env = compile->env(); 299 assert(env == ciEnv::current(), "must already be a ciEnv active"); 300 assert(env->compiler_data() == NULL, "compile already active?"); 301 env->set_compiler_data(compile); 302 assert(compile == Compile::current(), "sanity"); 303 304 compile->set_type_dict(NULL); 305 compile->set_type_hwm(NULL); 306 compile->set_type_last_size(0); 307 compile->set_last_tf(NULL, NULL); 308 compile->set_indexSet_arena(NULL); 309 compile->set_indexSet_free_block_list(NULL); 310 compile->init_type_arena(); 311 Type::Initialize(compile); 312 _compile->set_scratch_buffer_blob(NULL); 313 _compile->begin_method(); 314 } 315 CompileWrapper::~CompileWrapper() { 316 _compile->end_method(); 317 if (_compile->scratch_buffer_blob() != NULL) 318 BufferBlob::free(_compile->scratch_buffer_blob()); 319 _compile->env()->set_compiler_data(NULL); 320 } 321 322 323 //----------------------------print_compile_messages--------------------------- 324 void Compile::print_compile_messages() { 325 #ifndef PRODUCT 326 // Check if recompiling 327 if (_subsume_loads == false && PrintOpto) { 328 // Recompiling without allowing machine instructions to subsume loads 329 tty->print_cr("*********************************************************"); 330 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 331 tty->print_cr("*********************************************************"); 332 } 333 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 334 // Recompiling without escape analysis 335 tty->print_cr("*********************************************************"); 336 tty->print_cr("** Bailout: Recompile without escape analysis **"); 337 tty->print_cr("*********************************************************"); 338 } 339 if (env()->break_at_compile()) { 340 // Open the debugger when compiling this method. 341 tty->print("### Breaking when compiling: "); 342 method()->print_short_name(); 343 tty->cr(); 344 BREAKPOINT; 345 } 346 347 if( PrintOpto ) { 348 if (is_osr_compilation()) { 349 tty->print("[OSR]%3d", _compile_id); 350 } else { 351 tty->print("%3d", _compile_id); 352 } 353 } 354 #endif 355 } 356 357 358 void Compile::init_scratch_buffer_blob() { 359 if( scratch_buffer_blob() != NULL ) return; 360 361 // Construct a temporary CodeBuffer to have it construct a BufferBlob 362 // Cache this BufferBlob for this compile. 363 ResourceMark rm; 364 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 365 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 366 // Record the buffer blob for next time. 367 set_scratch_buffer_blob(blob); 368 // Have we run out of code space? 369 if (scratch_buffer_blob() == NULL) { 370 // Let CompilerBroker disable further compilations. 371 record_failure("Not enough space for scratch buffer in CodeCache"); 372 return; 373 } 374 375 // Initialize the relocation buffers 376 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 377 set_scratch_locs_memory(locs_buf); 378 } 379 380 381 //-----------------------scratch_emit_size------------------------------------- 382 // Helper function that computes size by emitting code 383 uint Compile::scratch_emit_size(const Node* n) { 384 // Emit into a trash buffer and count bytes emitted. 385 // This is a pretty expensive way to compute a size, 386 // but it works well enough if seldom used. 387 // All common fixed-size instructions are given a size 388 // method by the AD file. 389 // Note that the scratch buffer blob and locs memory are 390 // allocated at the beginning of the compile task, and 391 // may be shared by several calls to scratch_emit_size. 392 // The allocation of the scratch buffer blob is particularly 393 // expensive, since it has to grab the code cache lock. 394 BufferBlob* blob = this->scratch_buffer_blob(); 395 assert(blob != NULL, "Initialize BufferBlob at start"); 396 assert(blob->size() > MAX_inst_size, "sanity"); 397 relocInfo* locs_buf = scratch_locs_memory(); 398 address blob_begin = blob->instructions_begin(); 399 address blob_end = (address)locs_buf; 400 assert(blob->instructions_contains(blob_end), "sanity"); 401 CodeBuffer buf(blob_begin, blob_end - blob_begin); 402 buf.initialize_consts_size(MAX_const_size); 403 buf.initialize_stubs_size(MAX_stubs_size); 404 assert(locs_buf != NULL, "sanity"); 405 int lsize = MAX_locs_size / 2; 406 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 407 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 408 n->emit(buf, this->regalloc()); 409 return buf.code_size(); 410 } 411 412 413 // ============================================================================ 414 //------------------------------Compile standard------------------------------- 415 debug_only( int Compile::_debug_idx = 100000; ) 416 417 // Compile a method. entry_bci is -1 for normal compilations and indicates 418 // the continuation bci for on stack replacement. 419 420 421 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 422 : Phase(Compiler), 423 _env(ci_env), 424 _log(ci_env->log()), 425 _compile_id(ci_env->compile_id()), 426 _save_argument_registers(false), 427 _stub_name(NULL), 428 _stub_function(NULL), 429 _stub_entry_point(NULL), 430 _method(target), 431 _entry_bci(osr_bci), 432 _initial_gvn(NULL), 433 _for_igvn(NULL), 434 _warm_calls(NULL), 435 _subsume_loads(subsume_loads), 436 _do_escape_analysis(do_escape_analysis), 437 _failure_reason(NULL), 438 _code_buffer("Compile::Fill_buffer"), 439 _orig_pc_slot(0), 440 _orig_pc_slot_offset_in_bytes(0), 441 _node_bundling_limit(0), 442 _node_bundling_base(NULL), 443 #ifndef PRODUCT 444 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 445 _printer(IdealGraphPrinter::printer()), 446 #endif 447 _congraph(NULL) { 448 C = this; 449 450 CompileWrapper cw(this); 451 #ifndef PRODUCT 452 if (TimeCompiler2) { 453 tty->print(" "); 454 target->holder()->name()->print(); 455 tty->print("."); 456 target->print_short_name(); 457 tty->print(" "); 458 } 459 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 460 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 461 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 462 if (!print_opto_assembly) { 463 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 464 if (print_assembly && !Disassembler::can_decode()) { 465 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 466 print_opto_assembly = true; 467 } 468 } 469 set_print_assembly(print_opto_assembly); 470 set_parsed_irreducible_loop(false); 471 #endif 472 473 if (ProfileTraps) { 474 // Make sure the method being compiled gets its own MDO, 475 // so we can at least track the decompile_count(). 476 method()->build_method_data(); 477 } 478 479 Init(::AliasLevel); 480 481 482 print_compile_messages(); 483 484 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 485 _ilt = InlineTree::build_inline_tree_root(); 486 else 487 _ilt = NULL; 488 489 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 490 assert(num_alias_types() >= AliasIdxRaw, ""); 491 492 #define MINIMUM_NODE_HASH 1023 493 // Node list that Iterative GVN will start with 494 Unique_Node_List for_igvn(comp_arena()); 495 set_for_igvn(&for_igvn); 496 497 // GVN that will be run immediately on new nodes 498 uint estimated_size = method()->code_size()*4+64; 499 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 500 PhaseGVN gvn(node_arena(), estimated_size); 501 set_initial_gvn(&gvn); 502 503 { // Scope for timing the parser 504 TracePhase t3("parse", &_t_parser, true); 505 506 // Put top into the hash table ASAP. 507 initial_gvn()->transform_no_reclaim(top()); 508 509 // Set up tf(), start(), and find a CallGenerator. 510 CallGenerator* cg; 511 if (is_osr_compilation()) { 512 const TypeTuple *domain = StartOSRNode::osr_domain(); 513 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 514 init_tf(TypeFunc::make(domain, range)); 515 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 516 initial_gvn()->set_type_bottom(s); 517 init_start(s); 518 cg = CallGenerator::for_osr(method(), entry_bci()); 519 } else { 520 // Normal case. 521 init_tf(TypeFunc::make(method())); 522 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 523 initial_gvn()->set_type_bottom(s); 524 init_start(s); 525 float past_uses = method()->interpreter_invocation_count(); 526 float expected_uses = past_uses; 527 cg = CallGenerator::for_inline(method(), expected_uses); 528 } 529 if (failing()) return; 530 if (cg == NULL) { 531 record_method_not_compilable_all_tiers("cannot parse method"); 532 return; 533 } 534 JVMState* jvms = build_start_state(start(), tf()); 535 if ((jvms = cg->generate(jvms)) == NULL) { 536 record_method_not_compilable("method parse failed"); 537 return; 538 } 539 GraphKit kit(jvms); 540 541 if (!kit.stopped()) { 542 // Accept return values, and transfer control we know not where. 543 // This is done by a special, unique ReturnNode bound to root. 544 return_values(kit.jvms()); 545 } 546 547 if (kit.has_exceptions()) { 548 // Any exceptions that escape from this call must be rethrown 549 // to whatever caller is dynamically above us on the stack. 550 // This is done by a special, unique RethrowNode bound to root. 551 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 552 } 553 554 print_method("Before RemoveUseless", 3); 555 556 // Remove clutter produced by parsing. 557 if (!failing()) { 558 ResourceMark rm; 559 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 560 } 561 } 562 563 // Note: Large methods are capped off in do_one_bytecode(). 564 if (failing()) return; 565 566 // After parsing, node notes are no longer automagic. 567 // They must be propagated by register_new_node_with_optimizer(), 568 // clone(), or the like. 569 set_default_node_notes(NULL); 570 571 for (;;) { 572 int successes = Inline_Warm(); 573 if (failing()) return; 574 if (successes == 0) break; 575 } 576 577 // Drain the list. 578 Finish_Warm(); 579 #ifndef PRODUCT 580 if (_printer) { 581 _printer->print_inlining(this); 582 } 583 #endif 584 585 if (failing()) return; 586 NOT_PRODUCT( verify_graph_edges(); ) 587 588 // Perform escape analysis 589 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 590 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 591 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction. 592 PhaseGVN* igvn = initial_gvn(); 593 Node* oop_null = igvn->zerocon(T_OBJECT); 594 Node* noop_null = igvn->zerocon(T_NARROWOOP); 595 596 _congraph = new(comp_arena()) ConnectionGraph(this); 597 bool has_non_escaping_obj = _congraph->compute_escape(); 598 599 #ifndef PRODUCT 600 if (PrintEscapeAnalysis) { 601 _congraph->dump(); 602 } 603 #endif 604 // Cleanup. 605 if (oop_null->outcnt() == 0) 606 igvn->hash_delete(oop_null); 607 if (noop_null->outcnt() == 0) 608 igvn->hash_delete(noop_null); 609 610 if (!has_non_escaping_obj) { 611 _congraph = NULL; 612 } 613 614 if (failing()) return; 615 } 616 // Now optimize 617 Optimize(); 618 if (failing()) return; 619 NOT_PRODUCT( verify_graph_edges(); ) 620 621 #ifndef PRODUCT 622 if (PrintIdeal) { 623 ttyLocker ttyl; // keep the following output all in one block 624 // This output goes directly to the tty, not the compiler log. 625 // To enable tools to match it up with the compilation activity, 626 // be sure to tag this tty output with the compile ID. 627 if (xtty != NULL) { 628 xtty->head("ideal compile_id='%d'%s", compile_id(), 629 is_osr_compilation() ? " compile_kind='osr'" : 630 ""); 631 } 632 root()->dump(9999); 633 if (xtty != NULL) { 634 xtty->tail("ideal"); 635 } 636 } 637 #endif 638 639 // Now that we know the size of all the monitors we can add a fixed slot 640 // for the original deopt pc. 641 642 _orig_pc_slot = fixed_slots(); 643 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 644 set_fixed_slots(next_slot); 645 646 // Now generate code 647 Code_Gen(); 648 if (failing()) return; 649 650 // Check if we want to skip execution of all compiled code. 651 { 652 #ifndef PRODUCT 653 if (OptoNoExecute) { 654 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 655 return; 656 } 657 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 658 #endif 659 660 if (is_osr_compilation()) { 661 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 662 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 663 } else { 664 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 665 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 666 } 667 668 env()->register_method(_method, _entry_bci, 669 &_code_offsets, 670 _orig_pc_slot_offset_in_bytes, 671 code_buffer(), 672 frame_size_in_words(), _oop_map_set, 673 &_handler_table, &_inc_table, 674 compiler, 675 env()->comp_level(), 676 true, /*has_debug_info*/ 677 has_unsafe_access() 678 ); 679 } 680 } 681 682 //------------------------------Compile---------------------------------------- 683 // Compile a runtime stub 684 Compile::Compile( ciEnv* ci_env, 685 TypeFunc_generator generator, 686 address stub_function, 687 const char *stub_name, 688 int is_fancy_jump, 689 bool pass_tls, 690 bool save_arg_registers, 691 bool return_pc ) 692 : Phase(Compiler), 693 _env(ci_env), 694 _log(ci_env->log()), 695 _compile_id(-1), 696 _save_argument_registers(save_arg_registers), 697 _method(NULL), 698 _stub_name(stub_name), 699 _stub_function(stub_function), 700 _stub_entry_point(NULL), 701 _entry_bci(InvocationEntryBci), 702 _initial_gvn(NULL), 703 _for_igvn(NULL), 704 _warm_calls(NULL), 705 _orig_pc_slot(0), 706 _orig_pc_slot_offset_in_bytes(0), 707 _subsume_loads(true), 708 _do_escape_analysis(false), 709 _failure_reason(NULL), 710 _code_buffer("Compile::Fill_buffer"), 711 _node_bundling_limit(0), 712 _node_bundling_base(NULL), 713 #ifndef PRODUCT 714 _trace_opto_output(TraceOptoOutput), 715 _printer(NULL), 716 #endif 717 _congraph(NULL) { 718 C = this; 719 720 #ifndef PRODUCT 721 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 722 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 723 set_print_assembly(PrintFrameConverterAssembly); 724 set_parsed_irreducible_loop(false); 725 #endif 726 CompileWrapper cw(this); 727 Init(/*AliasLevel=*/ 0); 728 init_tf((*generator)()); 729 730 { 731 // The following is a dummy for the sake of GraphKit::gen_stub 732 Unique_Node_List for_igvn(comp_arena()); 733 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 734 PhaseGVN gvn(Thread::current()->resource_area(),255); 735 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 736 gvn.transform_no_reclaim(top()); 737 738 GraphKit kit; 739 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 740 } 741 742 NOT_PRODUCT( verify_graph_edges(); ) 743 Code_Gen(); 744 if (failing()) return; 745 746 747 // Entry point will be accessed using compile->stub_entry_point(); 748 if (code_buffer() == NULL) { 749 Matcher::soft_match_failure(); 750 } else { 751 if (PrintAssembly && (WizardMode || Verbose)) 752 tty->print_cr("### Stub::%s", stub_name); 753 754 if (!failing()) { 755 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 756 757 // Make the NMethod 758 // For now we mark the frame as never safe for profile stackwalking 759 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 760 code_buffer(), 761 CodeOffsets::frame_never_safe, 762 // _code_offsets.value(CodeOffsets::Frame_Complete), 763 frame_size_in_words(), 764 _oop_map_set, 765 save_arg_registers); 766 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 767 768 _stub_entry_point = rs->entry_point(); 769 } 770 } 771 } 772 773 #ifndef PRODUCT 774 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 775 if(PrintOpto && Verbose) { 776 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 777 } 778 } 779 #endif 780 781 void Compile::print_codes() { 782 } 783 784 //------------------------------Init------------------------------------------- 785 // Prepare for a single compilation 786 void Compile::Init(int aliaslevel) { 787 _unique = 0; 788 _regalloc = NULL; 789 790 _tf = NULL; // filled in later 791 _top = NULL; // cached later 792 _matcher = NULL; // filled in later 793 _cfg = NULL; // filled in later 794 795 set_24_bit_selection_and_mode(Use24BitFP, false); 796 797 _node_note_array = NULL; 798 _default_node_notes = NULL; 799 800 _immutable_memory = NULL; // filled in at first inquiry 801 802 // Globally visible Nodes 803 // First set TOP to NULL to give safe behavior during creation of RootNode 804 set_cached_top_node(NULL); 805 set_root(new (this, 3) RootNode()); 806 // Now that you have a Root to point to, create the real TOP 807 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 808 set_recent_alloc(NULL, NULL); 809 810 // Create Debug Information Recorder to record scopes, oopmaps, etc. 811 env()->set_oop_recorder(new OopRecorder(comp_arena())); 812 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 813 env()->set_dependencies(new Dependencies(env())); 814 815 _fixed_slots = 0; 816 set_has_split_ifs(false); 817 set_has_loops(has_method() && method()->has_loops()); // first approximation 818 _deopt_happens = true; // start out assuming the worst 819 _trap_can_recompile = false; // no traps emitted yet 820 _major_progress = true; // start out assuming good things will happen 821 set_has_unsafe_access(false); 822 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 823 set_decompile_count(0); 824 825 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 826 // Compilation level related initialization 827 if (env()->comp_level() == CompLevel_fast_compile) { 828 set_num_loop_opts(Tier1LoopOptsCount); 829 set_do_inlining(Tier1Inline != 0); 830 set_max_inline_size(Tier1MaxInlineSize); 831 set_freq_inline_size(Tier1FreqInlineSize); 832 set_do_scheduling(false); 833 set_do_count_invocations(Tier1CountInvocations); 834 set_do_method_data_update(Tier1UpdateMethodData); 835 } else { 836 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 837 set_num_loop_opts(LoopOptsCount); 838 set_do_inlining(Inline); 839 set_max_inline_size(MaxInlineSize); 840 set_freq_inline_size(FreqInlineSize); 841 set_do_scheduling(OptoScheduling); 842 set_do_count_invocations(false); 843 set_do_method_data_update(false); 844 } 845 846 if (debug_info()->recording_non_safepoints()) { 847 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 848 (comp_arena(), 8, 0, NULL)); 849 set_default_node_notes(Node_Notes::make(this)); 850 } 851 852 // // -- Initialize types before each compile -- 853 // // Update cached type information 854 // if( _method && _method->constants() ) 855 // Type::update_loaded_types(_method, _method->constants()); 856 857 // Init alias_type map. 858 if (!_do_escape_analysis && aliaslevel == 3) 859 aliaslevel = 2; // No unique types without escape analysis 860 _AliasLevel = aliaslevel; 861 const int grow_ats = 16; 862 _max_alias_types = grow_ats; 863 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 864 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 865 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 866 { 867 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 868 } 869 // Initialize the first few types. 870 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 871 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 872 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 873 _num_alias_types = AliasIdxRaw+1; 874 // Zero out the alias type cache. 875 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 876 // A NULL adr_type hits in the cache right away. Preload the right answer. 877 probe_alias_cache(NULL)->_index = AliasIdxTop; 878 879 _intrinsics = NULL; 880 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 881 register_library_intrinsics(); 882 } 883 884 //---------------------------init_start---------------------------------------- 885 // Install the StartNode on this compile object. 886 void Compile::init_start(StartNode* s) { 887 if (failing()) 888 return; // already failing 889 assert(s == start(), ""); 890 } 891 892 StartNode* Compile::start() const { 893 assert(!failing(), ""); 894 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 895 Node* start = root()->fast_out(i); 896 if( start->is_Start() ) 897 return start->as_Start(); 898 } 899 ShouldNotReachHere(); 900 return NULL; 901 } 902 903 //-------------------------------immutable_memory------------------------------------- 904 // Access immutable memory 905 Node* Compile::immutable_memory() { 906 if (_immutable_memory != NULL) { 907 return _immutable_memory; 908 } 909 StartNode* s = start(); 910 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 911 Node *p = s->fast_out(i); 912 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 913 _immutable_memory = p; 914 return _immutable_memory; 915 } 916 } 917 ShouldNotReachHere(); 918 return NULL; 919 } 920 921 //----------------------set_cached_top_node------------------------------------ 922 // Install the cached top node, and make sure Node::is_top works correctly. 923 void Compile::set_cached_top_node(Node* tn) { 924 if (tn != NULL) verify_top(tn); 925 Node* old_top = _top; 926 _top = tn; 927 // Calling Node::setup_is_top allows the nodes the chance to adjust 928 // their _out arrays. 929 if (_top != NULL) _top->setup_is_top(); 930 if (old_top != NULL) old_top->setup_is_top(); 931 assert(_top == NULL || top()->is_top(), ""); 932 } 933 934 #ifndef PRODUCT 935 void Compile::verify_top(Node* tn) const { 936 if (tn != NULL) { 937 assert(tn->is_Con(), "top node must be a constant"); 938 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 939 assert(tn->in(0) != NULL, "must have live top node"); 940 } 941 } 942 #endif 943 944 945 ///-------------------Managing Per-Node Debug & Profile Info------------------- 946 947 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 948 guarantee(arr != NULL, ""); 949 int num_blocks = arr->length(); 950 if (grow_by < num_blocks) grow_by = num_blocks; 951 int num_notes = grow_by * _node_notes_block_size; 952 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 953 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 954 while (num_notes > 0) { 955 arr->append(notes); 956 notes += _node_notes_block_size; 957 num_notes -= _node_notes_block_size; 958 } 959 assert(num_notes == 0, "exact multiple, please"); 960 } 961 962 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 963 if (source == NULL || dest == NULL) return false; 964 965 if (dest->is_Con()) 966 return false; // Do not push debug info onto constants. 967 968 #ifdef ASSERT 969 // Leave a bread crumb trail pointing to the original node: 970 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 971 dest->set_debug_orig(source); 972 } 973 #endif 974 975 if (node_note_array() == NULL) 976 return false; // Not collecting any notes now. 977 978 // This is a copy onto a pre-existing node, which may already have notes. 979 // If both nodes have notes, do not overwrite any pre-existing notes. 980 Node_Notes* source_notes = node_notes_at(source->_idx); 981 if (source_notes == NULL || source_notes->is_clear()) return false; 982 Node_Notes* dest_notes = node_notes_at(dest->_idx); 983 if (dest_notes == NULL || dest_notes->is_clear()) { 984 return set_node_notes_at(dest->_idx, source_notes); 985 } 986 987 Node_Notes merged_notes = (*source_notes); 988 // The order of operations here ensures that dest notes will win... 989 merged_notes.update_from(dest_notes); 990 return set_node_notes_at(dest->_idx, &merged_notes); 991 } 992 993 994 //--------------------------allow_range_check_smearing------------------------- 995 // Gating condition for coalescing similar range checks. 996 // Sometimes we try 'speculatively' replacing a series of a range checks by a 997 // single covering check that is at least as strong as any of them. 998 // If the optimization succeeds, the simplified (strengthened) range check 999 // will always succeed. If it fails, we will deopt, and then give up 1000 // on the optimization. 1001 bool Compile::allow_range_check_smearing() const { 1002 // If this method has already thrown a range-check, 1003 // assume it was because we already tried range smearing 1004 // and it failed. 1005 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1006 return !already_trapped; 1007 } 1008 1009 1010 //------------------------------flatten_alias_type----------------------------- 1011 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1012 int offset = tj->offset(); 1013 TypePtr::PTR ptr = tj->ptr(); 1014 1015 // Known instance (scalarizable allocation) alias only with itself. 1016 bool is_known_inst = tj->isa_oopptr() != NULL && 1017 tj->is_oopptr()->is_known_instance(); 1018 1019 // Process weird unsafe references. 1020 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1021 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1022 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1023 tj = TypeOopPtr::BOTTOM; 1024 ptr = tj->ptr(); 1025 offset = tj->offset(); 1026 } 1027 1028 // Array pointers need some flattening 1029 const TypeAryPtr *ta = tj->isa_aryptr(); 1030 if( ta && is_known_inst ) { 1031 if ( offset != Type::OffsetBot && 1032 offset > arrayOopDesc::length_offset_in_bytes() ) { 1033 offset = Type::OffsetBot; // Flatten constant access into array body only 1034 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1035 } 1036 } else if( ta && _AliasLevel >= 2 ) { 1037 // For arrays indexed by constant indices, we flatten the alias 1038 // space to include all of the array body. Only the header, klass 1039 // and array length can be accessed un-aliased. 1040 if( offset != Type::OffsetBot ) { 1041 if( ta->const_oop() ) { // methodDataOop or methodOop 1042 offset = Type::OffsetBot; // Flatten constant access into array body 1043 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1044 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1045 // range is OK as-is. 1046 tj = ta = TypeAryPtr::RANGE; 1047 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1048 tj = TypeInstPtr::KLASS; // all klass loads look alike 1049 ta = TypeAryPtr::RANGE; // generic ignored junk 1050 ptr = TypePtr::BotPTR; 1051 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1052 tj = TypeInstPtr::MARK; 1053 ta = TypeAryPtr::RANGE; // generic ignored junk 1054 ptr = TypePtr::BotPTR; 1055 } else { // Random constant offset into array body 1056 offset = Type::OffsetBot; // Flatten constant access into array body 1057 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1058 } 1059 } 1060 // Arrays of fixed size alias with arrays of unknown size. 1061 if (ta->size() != TypeInt::POS) { 1062 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1063 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1064 } 1065 // Arrays of known objects become arrays of unknown objects. 1066 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1067 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1068 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1069 } 1070 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1071 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1072 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1073 } 1074 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1075 // cannot be distinguished by bytecode alone. 1076 if (ta->elem() == TypeInt::BOOL) { 1077 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1078 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1079 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1080 } 1081 // During the 2nd round of IterGVN, NotNull castings are removed. 1082 // Make sure the Bottom and NotNull variants alias the same. 1083 // Also, make sure exact and non-exact variants alias the same. 1084 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1085 if (ta->const_oop()) { 1086 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1087 } else { 1088 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1089 } 1090 } 1091 } 1092 1093 // Oop pointers need some flattening 1094 const TypeInstPtr *to = tj->isa_instptr(); 1095 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1096 if( ptr == TypePtr::Constant ) { 1097 // No constant oop pointers (such as Strings); they alias with 1098 // unknown strings. 1099 assert(!is_known_inst, "not scalarizable allocation"); 1100 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1101 } else if( is_known_inst ) { 1102 tj = to; // Keep NotNull and klass_is_exact for instance type 1103 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1104 // During the 2nd round of IterGVN, NotNull castings are removed. 1105 // Make sure the Bottom and NotNull variants alias the same. 1106 // Also, make sure exact and non-exact variants alias the same. 1107 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1108 } 1109 // Canonicalize the holder of this field 1110 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1111 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1112 // First handle header references such as a LoadKlassNode, even if the 1113 // object's klass is unloaded at compile time (4965979). 1114 if (!is_known_inst) { // Do it only for non-instance types 1115 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1116 } 1117 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1118 to = NULL; 1119 tj = TypeOopPtr::BOTTOM; 1120 offset = tj->offset(); 1121 } else { 1122 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1123 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1124 if( is_known_inst ) { 1125 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1126 } else { 1127 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1128 } 1129 } 1130 } 1131 } 1132 1133 // Klass pointers to object array klasses need some flattening 1134 const TypeKlassPtr *tk = tj->isa_klassptr(); 1135 if( tk ) { 1136 // If we are referencing a field within a Klass, we need 1137 // to assume the worst case of an Object. Both exact and 1138 // inexact types must flatten to the same alias class. 1139 // Since the flattened result for a klass is defined to be 1140 // precisely java.lang.Object, use a constant ptr. 1141 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1142 1143 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1144 TypeKlassPtr::OBJECT->klass(), 1145 offset); 1146 } 1147 1148 ciKlass* klass = tk->klass(); 1149 if( klass->is_obj_array_klass() ) { 1150 ciKlass* k = TypeAryPtr::OOPS->klass(); 1151 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1152 k = TypeInstPtr::BOTTOM->klass(); 1153 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1154 } 1155 1156 // Check for precise loads from the primary supertype array and force them 1157 // to the supertype cache alias index. Check for generic array loads from 1158 // the primary supertype array and also force them to the supertype cache 1159 // alias index. Since the same load can reach both, we need to merge 1160 // these 2 disparate memories into the same alias class. Since the 1161 // primary supertype array is read-only, there's no chance of confusion 1162 // where we bypass an array load and an array store. 1163 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1164 if( offset == Type::OffsetBot || 1165 off2 < Klass::primary_super_limit()*wordSize ) { 1166 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1167 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1168 } 1169 } 1170 1171 // Flatten all Raw pointers together. 1172 if (tj->base() == Type::RawPtr) 1173 tj = TypeRawPtr::BOTTOM; 1174 1175 if (tj->base() == Type::AnyPtr) 1176 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1177 1178 // Flatten all to bottom for now 1179 switch( _AliasLevel ) { 1180 case 0: 1181 tj = TypePtr::BOTTOM; 1182 break; 1183 case 1: // Flatten to: oop, static, field or array 1184 switch (tj->base()) { 1185 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1186 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1187 case Type::AryPtr: // do not distinguish arrays at all 1188 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1189 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1190 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1191 default: ShouldNotReachHere(); 1192 } 1193 break; 1194 case 2: // No collapsing at level 2; keep all splits 1195 case 3: // No collapsing at level 3; keep all splits 1196 break; 1197 default: 1198 Unimplemented(); 1199 } 1200 1201 offset = tj->offset(); 1202 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1203 1204 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1205 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1206 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1207 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1208 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1209 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1210 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1211 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1212 assert( tj->ptr() != TypePtr::TopPTR && 1213 tj->ptr() != TypePtr::AnyNull && 1214 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1215 // assert( tj->ptr() != TypePtr::Constant || 1216 // tj->base() == Type::RawPtr || 1217 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1218 1219 return tj; 1220 } 1221 1222 void Compile::AliasType::Init(int i, const TypePtr* at) { 1223 _index = i; 1224 _adr_type = at; 1225 _field = NULL; 1226 _is_rewritable = true; // default 1227 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1228 if (atoop != NULL && atoop->is_known_instance()) { 1229 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1230 _general_index = Compile::current()->get_alias_index(gt); 1231 } else { 1232 _general_index = 0; 1233 } 1234 } 1235 1236 //---------------------------------print_on------------------------------------ 1237 #ifndef PRODUCT 1238 void Compile::AliasType::print_on(outputStream* st) { 1239 if (index() < 10) 1240 st->print("@ <%d> ", index()); 1241 else st->print("@ <%d>", index()); 1242 st->print(is_rewritable() ? " " : " RO"); 1243 int offset = adr_type()->offset(); 1244 if (offset == Type::OffsetBot) 1245 st->print(" +any"); 1246 else st->print(" +%-3d", offset); 1247 st->print(" in "); 1248 adr_type()->dump_on(st); 1249 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1250 if (field() != NULL && tjp) { 1251 if (tjp->klass() != field()->holder() || 1252 tjp->offset() != field()->offset_in_bytes()) { 1253 st->print(" != "); 1254 field()->print(); 1255 st->print(" ***"); 1256 } 1257 } 1258 } 1259 1260 void print_alias_types() { 1261 Compile* C = Compile::current(); 1262 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1263 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1264 C->alias_type(idx)->print_on(tty); 1265 tty->cr(); 1266 } 1267 } 1268 #endif 1269 1270 1271 //----------------------------probe_alias_cache-------------------------------- 1272 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1273 intptr_t key = (intptr_t) adr_type; 1274 key ^= key >> logAliasCacheSize; 1275 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1276 } 1277 1278 1279 //-----------------------------grow_alias_types-------------------------------- 1280 void Compile::grow_alias_types() { 1281 const int old_ats = _max_alias_types; // how many before? 1282 const int new_ats = old_ats; // how many more? 1283 const int grow_ats = old_ats+new_ats; // how many now? 1284 _max_alias_types = grow_ats; 1285 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1286 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1287 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1288 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1289 } 1290 1291 1292 //--------------------------------find_alias_type------------------------------ 1293 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1294 if (_AliasLevel == 0) 1295 return alias_type(AliasIdxBot); 1296 1297 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1298 if (ace->_adr_type == adr_type) { 1299 return alias_type(ace->_index); 1300 } 1301 1302 // Handle special cases. 1303 if (adr_type == NULL) return alias_type(AliasIdxTop); 1304 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1305 1306 // Do it the slow way. 1307 const TypePtr* flat = flatten_alias_type(adr_type); 1308 1309 #ifdef ASSERT 1310 assert(flat == flatten_alias_type(flat), "idempotent"); 1311 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1312 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1313 const TypeOopPtr* foop = flat->is_oopptr(); 1314 // Scalarizable allocations have exact klass always. 1315 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1316 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1317 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1318 } 1319 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1320 #endif 1321 1322 int idx = AliasIdxTop; 1323 for (int i = 0; i < num_alias_types(); i++) { 1324 if (alias_type(i)->adr_type() == flat) { 1325 idx = i; 1326 break; 1327 } 1328 } 1329 1330 if (idx == AliasIdxTop) { 1331 if (no_create) return NULL; 1332 // Grow the array if necessary. 1333 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1334 // Add a new alias type. 1335 idx = _num_alias_types++; 1336 _alias_types[idx]->Init(idx, flat); 1337 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1338 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1339 if (flat->isa_instptr()) { 1340 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1341 && flat->is_instptr()->klass() == env()->Class_klass()) 1342 alias_type(idx)->set_rewritable(false); 1343 } 1344 if (flat->isa_klassptr()) { 1345 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1346 alias_type(idx)->set_rewritable(false); 1347 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1348 alias_type(idx)->set_rewritable(false); 1349 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1350 alias_type(idx)->set_rewritable(false); 1351 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1352 alias_type(idx)->set_rewritable(false); 1353 } 1354 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1355 // but the base pointer type is not distinctive enough to identify 1356 // references into JavaThread.) 1357 1358 // Check for final instance fields. 1359 const TypeInstPtr* tinst = flat->isa_instptr(); 1360 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1361 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1362 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1363 // Set field() and is_rewritable() attributes. 1364 if (field != NULL) alias_type(idx)->set_field(field); 1365 } 1366 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1367 // Check for final static fields. 1368 if (tklass && tklass->klass()->is_instance_klass()) { 1369 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1370 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1371 // Set field() and is_rewritable() attributes. 1372 if (field != NULL) alias_type(idx)->set_field(field); 1373 } 1374 } 1375 1376 // Fill the cache for next time. 1377 ace->_adr_type = adr_type; 1378 ace->_index = idx; 1379 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1380 1381 // Might as well try to fill the cache for the flattened version, too. 1382 AliasCacheEntry* face = probe_alias_cache(flat); 1383 if (face->_adr_type == NULL) { 1384 face->_adr_type = flat; 1385 face->_index = idx; 1386 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1387 } 1388 1389 return alias_type(idx); 1390 } 1391 1392 1393 Compile::AliasType* Compile::alias_type(ciField* field) { 1394 const TypeOopPtr* t; 1395 if (field->is_static()) 1396 t = TypeKlassPtr::make(field->holder()); 1397 else 1398 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1399 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1400 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1401 return atp; 1402 } 1403 1404 1405 //------------------------------have_alias_type-------------------------------- 1406 bool Compile::have_alias_type(const TypePtr* adr_type) { 1407 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1408 if (ace->_adr_type == adr_type) { 1409 return true; 1410 } 1411 1412 // Handle special cases. 1413 if (adr_type == NULL) return true; 1414 if (adr_type == TypePtr::BOTTOM) return true; 1415 1416 return find_alias_type(adr_type, true) != NULL; 1417 } 1418 1419 //-----------------------------must_alias-------------------------------------- 1420 // True if all values of the given address type are in the given alias category. 1421 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1422 if (alias_idx == AliasIdxBot) return true; // the universal category 1423 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1424 if (alias_idx == AliasIdxTop) return false; // the empty category 1425 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1426 1427 // the only remaining possible overlap is identity 1428 int adr_idx = get_alias_index(adr_type); 1429 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1430 assert(adr_idx == alias_idx || 1431 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1432 && adr_type != TypeOopPtr::BOTTOM), 1433 "should not be testing for overlap with an unsafe pointer"); 1434 return adr_idx == alias_idx; 1435 } 1436 1437 //------------------------------can_alias-------------------------------------- 1438 // True if any values of the given address type are in the given alias category. 1439 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1440 if (alias_idx == AliasIdxTop) return false; // the empty category 1441 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1442 if (alias_idx == AliasIdxBot) return true; // the universal category 1443 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1444 1445 // the only remaining possible overlap is identity 1446 int adr_idx = get_alias_index(adr_type); 1447 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1448 return adr_idx == alias_idx; 1449 } 1450 1451 1452 1453 //---------------------------pop_warm_call------------------------------------- 1454 WarmCallInfo* Compile::pop_warm_call() { 1455 WarmCallInfo* wci = _warm_calls; 1456 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1457 return wci; 1458 } 1459 1460 //----------------------------Inline_Warm-------------------------------------- 1461 int Compile::Inline_Warm() { 1462 // If there is room, try to inline some more warm call sites. 1463 // %%% Do a graph index compaction pass when we think we're out of space? 1464 if (!InlineWarmCalls) return 0; 1465 1466 int calls_made_hot = 0; 1467 int room_to_grow = NodeCountInliningCutoff - unique(); 1468 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1469 int amount_grown = 0; 1470 WarmCallInfo* call; 1471 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1472 int est_size = (int)call->size(); 1473 if (est_size > (room_to_grow - amount_grown)) { 1474 // This one won't fit anyway. Get rid of it. 1475 call->make_cold(); 1476 continue; 1477 } 1478 call->make_hot(); 1479 calls_made_hot++; 1480 amount_grown += est_size; 1481 amount_to_grow -= est_size; 1482 } 1483 1484 if (calls_made_hot > 0) set_major_progress(); 1485 return calls_made_hot; 1486 } 1487 1488 1489 //----------------------------Finish_Warm-------------------------------------- 1490 void Compile::Finish_Warm() { 1491 if (!InlineWarmCalls) return; 1492 if (failing()) return; 1493 if (warm_calls() == NULL) return; 1494 1495 // Clean up loose ends, if we are out of space for inlining. 1496 WarmCallInfo* call; 1497 while ((call = pop_warm_call()) != NULL) { 1498 call->make_cold(); 1499 } 1500 } 1501 1502 1503 //------------------------------Optimize--------------------------------------- 1504 // Given a graph, optimize it. 1505 void Compile::Optimize() { 1506 TracePhase t1("optimizer", &_t_optimizer, true); 1507 1508 #ifndef PRODUCT 1509 if (env()->break_at_compile()) { 1510 BREAKPOINT; 1511 } 1512 1513 #endif 1514 1515 ResourceMark rm; 1516 int loop_opts_cnt; 1517 1518 NOT_PRODUCT( verify_graph_edges(); ) 1519 1520 print_method("After Parsing"); 1521 1522 { 1523 // Iterative Global Value Numbering, including ideal transforms 1524 // Initialize IterGVN with types and values from parse-time GVN 1525 PhaseIterGVN igvn(initial_gvn()); 1526 { 1527 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1528 igvn.optimize(); 1529 } 1530 1531 print_method("Iter GVN 1", 2); 1532 1533 if (failing()) return; 1534 1535 // Loop transforms on the ideal graph. Range Check Elimination, 1536 // peeling, unrolling, etc. 1537 1538 // Set loop opts counter 1539 loop_opts_cnt = num_loop_opts(); 1540 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1541 { 1542 TracePhase t2("idealLoop", &_t_idealLoop, true); 1543 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1544 loop_opts_cnt--; 1545 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1546 if (failing()) return; 1547 } 1548 // Loop opts pass if partial peeling occurred in previous pass 1549 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1550 TracePhase t3("idealLoop", &_t_idealLoop, true); 1551 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1552 loop_opts_cnt--; 1553 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1554 if (failing()) return; 1555 } 1556 // Loop opts pass for loop-unrolling before CCP 1557 if(major_progress() && (loop_opts_cnt > 0)) { 1558 TracePhase t4("idealLoop", &_t_idealLoop, true); 1559 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1560 loop_opts_cnt--; 1561 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1562 } 1563 } 1564 if (failing()) return; 1565 1566 // Conditional Constant Propagation; 1567 PhaseCCP ccp( &igvn ); 1568 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1569 { 1570 TracePhase t2("ccp", &_t_ccp, true); 1571 ccp.do_transform(); 1572 } 1573 print_method("PhaseCPP 1", 2); 1574 1575 assert( true, "Break here to ccp.dump_old2new_map()"); 1576 1577 // Iterative Global Value Numbering, including ideal transforms 1578 { 1579 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1580 igvn = ccp; 1581 igvn.optimize(); 1582 } 1583 1584 print_method("Iter GVN 2", 2); 1585 1586 if (failing()) return; 1587 1588 // Loop transforms on the ideal graph. Range Check Elimination, 1589 // peeling, unrolling, etc. 1590 if(loop_opts_cnt > 0) { 1591 debug_only( int cnt = 0; ); 1592 while(major_progress() && (loop_opts_cnt > 0)) { 1593 TracePhase t2("idealLoop", &_t_idealLoop, true); 1594 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1595 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1596 loop_opts_cnt--; 1597 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1598 if (failing()) return; 1599 } 1600 } 1601 { 1602 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1603 PhaseMacroExpand mex(igvn); 1604 if (mex.expand_macro_nodes()) { 1605 assert(failing(), "must bail out w/ explicit message"); 1606 return; 1607 } 1608 } 1609 1610 } // (End scope of igvn; run destructor if necessary for asserts.) 1611 1612 // A method with only infinite loops has no edges entering loops from root 1613 { 1614 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1615 if (final_graph_reshaping()) { 1616 assert(failing(), "must bail out w/ explicit message"); 1617 return; 1618 } 1619 } 1620 1621 print_method("Optimize finished", 2); 1622 } 1623 1624 1625 //------------------------------Code_Gen--------------------------------------- 1626 // Given a graph, generate code for it 1627 void Compile::Code_Gen() { 1628 if (failing()) return; 1629 1630 // Perform instruction selection. You might think we could reclaim Matcher 1631 // memory PDQ, but actually the Matcher is used in generating spill code. 1632 // Internals of the Matcher (including some VectorSets) must remain live 1633 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1634 // set a bit in reclaimed memory. 1635 1636 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1637 // nodes. Mapping is only valid at the root of each matched subtree. 1638 NOT_PRODUCT( verify_graph_edges(); ) 1639 1640 Node_List proj_list; 1641 Matcher m(proj_list); 1642 _matcher = &m; 1643 { 1644 TracePhase t2("matcher", &_t_matcher, true); 1645 m.match(); 1646 } 1647 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1648 // nodes. Mapping is only valid at the root of each matched subtree. 1649 NOT_PRODUCT( verify_graph_edges(); ) 1650 1651 // If you have too many nodes, or if matching has failed, bail out 1652 check_node_count(0, "out of nodes matching instructions"); 1653 if (failing()) return; 1654 1655 // Build a proper-looking CFG 1656 PhaseCFG cfg(node_arena(), root(), m); 1657 _cfg = &cfg; 1658 { 1659 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1660 cfg.Dominators(); 1661 if (failing()) return; 1662 1663 NOT_PRODUCT( verify_graph_edges(); ) 1664 1665 cfg.Estimate_Block_Frequency(); 1666 cfg.GlobalCodeMotion(m,unique(),proj_list); 1667 1668 print_method("Global code motion", 2); 1669 1670 if (failing()) return; 1671 NOT_PRODUCT( verify_graph_edges(); ) 1672 1673 debug_only( cfg.verify(); ) 1674 } 1675 NOT_PRODUCT( verify_graph_edges(); ) 1676 1677 PhaseChaitin regalloc(unique(),cfg,m); 1678 _regalloc = ®alloc; 1679 { 1680 TracePhase t2("regalloc", &_t_registerAllocation, true); 1681 // Perform any platform dependent preallocation actions. This is used, 1682 // for example, to avoid taking an implicit null pointer exception 1683 // using the frame pointer on win95. 1684 _regalloc->pd_preallocate_hook(); 1685 1686 // Perform register allocation. After Chaitin, use-def chains are 1687 // no longer accurate (at spill code) and so must be ignored. 1688 // Node->LRG->reg mappings are still accurate. 1689 _regalloc->Register_Allocate(); 1690 1691 // Bail out if the allocator builds too many nodes 1692 if (failing()) return; 1693 } 1694 1695 // Prior to register allocation we kept empty basic blocks in case the 1696 // the allocator needed a place to spill. After register allocation we 1697 // are not adding any new instructions. If any basic block is empty, we 1698 // can now safely remove it. 1699 { 1700 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1701 cfg.remove_empty(); 1702 if (do_freq_based_layout()) { 1703 PhaseBlockLayout layout(cfg); 1704 } else { 1705 cfg.set_loop_alignment(); 1706 } 1707 cfg.fixup_flow(); 1708 } 1709 1710 // Perform any platform dependent postallocation verifications. 1711 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1712 1713 // Apply peephole optimizations 1714 if( OptoPeephole ) { 1715 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1716 PhasePeephole peep( _regalloc, cfg); 1717 peep.do_transform(); 1718 } 1719 1720 // Convert Nodes to instruction bits in a buffer 1721 { 1722 // %%%% workspace merge brought two timers together for one job 1723 TracePhase t2a("output", &_t_output, true); 1724 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1725 Output(); 1726 } 1727 1728 print_method("Final Code"); 1729 1730 // He's dead, Jim. 1731 _cfg = (PhaseCFG*)0xdeadbeef; 1732 _regalloc = (PhaseChaitin*)0xdeadbeef; 1733 } 1734 1735 1736 //------------------------------dump_asm--------------------------------------- 1737 // Dump formatted assembly 1738 #ifndef PRODUCT 1739 void Compile::dump_asm(int *pcs, uint pc_limit) { 1740 bool cut_short = false; 1741 tty->print_cr("#"); 1742 tty->print("# "); _tf->dump(); tty->cr(); 1743 tty->print_cr("#"); 1744 1745 // For all blocks 1746 int pc = 0x0; // Program counter 1747 char starts_bundle = ' '; 1748 _regalloc->dump_frame(); 1749 1750 Node *n = NULL; 1751 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1752 if (VMThread::should_terminate()) { cut_short = true; break; } 1753 Block *b = _cfg->_blocks[i]; 1754 if (b->is_connector() && !Verbose) continue; 1755 n = b->_nodes[0]; 1756 if (pcs && n->_idx < pc_limit) 1757 tty->print("%3.3x ", pcs[n->_idx]); 1758 else 1759 tty->print(" "); 1760 b->dump_head( &_cfg->_bbs ); 1761 if (b->is_connector()) { 1762 tty->print_cr(" # Empty connector block"); 1763 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1764 tty->print_cr(" # Block is sole successor of call"); 1765 } 1766 1767 // For all instructions 1768 Node *delay = NULL; 1769 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1770 if (VMThread::should_terminate()) { cut_short = true; break; } 1771 n = b->_nodes[j]; 1772 if (valid_bundle_info(n)) { 1773 Bundle *bundle = node_bundling(n); 1774 if (bundle->used_in_unconditional_delay()) { 1775 delay = n; 1776 continue; 1777 } 1778 if (bundle->starts_bundle()) 1779 starts_bundle = '+'; 1780 } 1781 1782 if (WizardMode) n->dump(); 1783 1784 if( !n->is_Region() && // Dont print in the Assembly 1785 !n->is_Phi() && // a few noisely useless nodes 1786 !n->is_Proj() && 1787 !n->is_MachTemp() && 1788 !n->is_Catch() && // Would be nice to print exception table targets 1789 !n->is_MergeMem() && // Not very interesting 1790 !n->is_top() && // Debug info table constants 1791 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1792 ) { 1793 if (pcs && n->_idx < pc_limit) 1794 tty->print("%3.3x", pcs[n->_idx]); 1795 else 1796 tty->print(" "); 1797 tty->print(" %c ", starts_bundle); 1798 starts_bundle = ' '; 1799 tty->print("\t"); 1800 n->format(_regalloc, tty); 1801 tty->cr(); 1802 } 1803 1804 // If we have an instruction with a delay slot, and have seen a delay, 1805 // then back up and print it 1806 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1807 assert(delay != NULL, "no unconditional delay instruction"); 1808 if (WizardMode) delay->dump(); 1809 1810 if (node_bundling(delay)->starts_bundle()) 1811 starts_bundle = '+'; 1812 if (pcs && n->_idx < pc_limit) 1813 tty->print("%3.3x", pcs[n->_idx]); 1814 else 1815 tty->print(" "); 1816 tty->print(" %c ", starts_bundle); 1817 starts_bundle = ' '; 1818 tty->print("\t"); 1819 delay->format(_regalloc, tty); 1820 tty->print_cr(""); 1821 delay = NULL; 1822 } 1823 1824 // Dump the exception table as well 1825 if( n->is_Catch() && (Verbose || WizardMode) ) { 1826 // Print the exception table for this offset 1827 _handler_table.print_subtable_for(pc); 1828 } 1829 } 1830 1831 if (pcs && n->_idx < pc_limit) 1832 tty->print_cr("%3.3x", pcs[n->_idx]); 1833 else 1834 tty->print_cr(""); 1835 1836 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1837 1838 } // End of per-block dump 1839 tty->print_cr(""); 1840 1841 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1842 } 1843 #endif 1844 1845 //------------------------------Final_Reshape_Counts--------------------------- 1846 // This class defines counters to help identify when a method 1847 // may/must be executed using hardware with only 24-bit precision. 1848 struct Final_Reshape_Counts : public StackObj { 1849 int _call_count; // count non-inlined 'common' calls 1850 int _float_count; // count float ops requiring 24-bit precision 1851 int _double_count; // count double ops requiring more precision 1852 int _java_call_count; // count non-inlined 'java' calls 1853 VectorSet _visited; // Visitation flags 1854 Node_List _tests; // Set of IfNodes & PCTableNodes 1855 1856 Final_Reshape_Counts() : 1857 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), 1858 _visited( Thread::current()->resource_area() ) { } 1859 1860 void inc_call_count () { _call_count ++; } 1861 void inc_float_count () { _float_count ++; } 1862 void inc_double_count() { _double_count++; } 1863 void inc_java_call_count() { _java_call_count++; } 1864 1865 int get_call_count () const { return _call_count ; } 1866 int get_float_count () const { return _float_count ; } 1867 int get_double_count() const { return _double_count; } 1868 int get_java_call_count() const { return _java_call_count; } 1869 }; 1870 1871 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1872 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1873 // Make sure the offset goes inside the instance layout. 1874 return k->contains_field_offset(tp->offset()); 1875 // Note that OffsetBot and OffsetTop are very negative. 1876 } 1877 1878 //------------------------------final_graph_reshaping_impl---------------------- 1879 // Implement items 1-5 from final_graph_reshaping below. 1880 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { 1881 1882 if ( n->outcnt() == 0 ) return; // dead node 1883 uint nop = n->Opcode(); 1884 1885 // Check for 2-input instruction with "last use" on right input. 1886 // Swap to left input. Implements item (2). 1887 if( n->req() == 3 && // two-input instruction 1888 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1889 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1890 n->in(2)->outcnt() == 1 &&// right use IS a last use 1891 !n->in(2)->is_Con() ) { // right use is not a constant 1892 // Check for commutative opcode 1893 switch( nop ) { 1894 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1895 case Op_MaxI: case Op_MinI: 1896 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1897 case Op_AndL: case Op_XorL: case Op_OrL: 1898 case Op_AndI: case Op_XorI: case Op_OrI: { 1899 // Move "last use" input to left by swapping inputs 1900 n->swap_edges(1, 2); 1901 break; 1902 } 1903 default: 1904 break; 1905 } 1906 } 1907 1908 // Count FPU ops and common calls, implements item (3) 1909 switch( nop ) { 1910 // Count all float operations that may use FPU 1911 case Op_AddF: 1912 case Op_SubF: 1913 case Op_MulF: 1914 case Op_DivF: 1915 case Op_NegF: 1916 case Op_ModF: 1917 case Op_ConvI2F: 1918 case Op_ConF: 1919 case Op_CmpF: 1920 case Op_CmpF3: 1921 // case Op_ConvL2F: // longs are split into 32-bit halves 1922 fpu.inc_float_count(); 1923 break; 1924 1925 case Op_ConvF2D: 1926 case Op_ConvD2F: 1927 fpu.inc_float_count(); 1928 fpu.inc_double_count(); 1929 break; 1930 1931 // Count all double operations that may use FPU 1932 case Op_AddD: 1933 case Op_SubD: 1934 case Op_MulD: 1935 case Op_DivD: 1936 case Op_NegD: 1937 case Op_ModD: 1938 case Op_ConvI2D: 1939 case Op_ConvD2I: 1940 // case Op_ConvL2D: // handled by leaf call 1941 // case Op_ConvD2L: // handled by leaf call 1942 case Op_ConD: 1943 case Op_CmpD: 1944 case Op_CmpD3: 1945 fpu.inc_double_count(); 1946 break; 1947 case Op_Opaque1: // Remove Opaque Nodes before matching 1948 case Op_Opaque2: // Remove Opaque Nodes before matching 1949 n->subsume_by(n->in(1)); 1950 break; 1951 case Op_CallStaticJava: 1952 case Op_CallJava: 1953 case Op_CallDynamicJava: 1954 fpu.inc_java_call_count(); // Count java call site; 1955 case Op_CallRuntime: 1956 case Op_CallLeaf: 1957 case Op_CallLeafNoFP: { 1958 assert( n->is_Call(), "" ); 1959 CallNode *call = n->as_Call(); 1960 // Count call sites where the FP mode bit would have to be flipped. 1961 // Do not count uncommon runtime calls: 1962 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 1963 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 1964 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 1965 fpu.inc_call_count(); // Count the call site 1966 } else { // See if uncommon argument is shared 1967 Node *n = call->in(TypeFunc::Parms); 1968 int nop = n->Opcode(); 1969 // Clone shared simple arguments to uncommon calls, item (1). 1970 if( n->outcnt() > 1 && 1971 !n->is_Proj() && 1972 nop != Op_CreateEx && 1973 nop != Op_CheckCastPP && 1974 nop != Op_DecodeN && 1975 !n->is_Mem() ) { 1976 Node *x = n->clone(); 1977 call->set_req( TypeFunc::Parms, x ); 1978 } 1979 } 1980 break; 1981 } 1982 1983 case Op_StoreD: 1984 case Op_LoadD: 1985 case Op_LoadD_unaligned: 1986 fpu.inc_double_count(); 1987 goto handle_mem; 1988 case Op_StoreF: 1989 case Op_LoadF: 1990 fpu.inc_float_count(); 1991 goto handle_mem; 1992 1993 case Op_StoreB: 1994 case Op_StoreC: 1995 case Op_StoreCM: 1996 case Op_StorePConditional: 1997 case Op_StoreI: 1998 case Op_StoreL: 1999 case Op_StoreIConditional: 2000 case Op_StoreLConditional: 2001 case Op_CompareAndSwapI: 2002 case Op_CompareAndSwapL: 2003 case Op_CompareAndSwapP: 2004 case Op_CompareAndSwapN: 2005 case Op_StoreP: 2006 case Op_StoreN: 2007 case Op_LoadB: 2008 case Op_LoadUB: 2009 case Op_LoadUS: 2010 case Op_LoadI: 2011 case Op_LoadUI2L: 2012 case Op_LoadKlass: 2013 case Op_LoadNKlass: 2014 case Op_LoadL: 2015 case Op_LoadL_unaligned: 2016 case Op_LoadPLocked: 2017 case Op_LoadLLocked: 2018 case Op_LoadP: 2019 case Op_LoadN: 2020 case Op_LoadRange: 2021 case Op_LoadS: { 2022 handle_mem: 2023 #ifdef ASSERT 2024 if( VerifyOptoOopOffsets ) { 2025 assert( n->is_Mem(), "" ); 2026 MemNode *mem = (MemNode*)n; 2027 // Check to see if address types have grounded out somehow. 2028 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2029 assert( !tp || oop_offset_is_sane(tp), "" ); 2030 } 2031 #endif 2032 break; 2033 } 2034 2035 case Op_AddP: { // Assert sane base pointers 2036 Node *addp = n->in(AddPNode::Address); 2037 assert( !addp->is_AddP() || 2038 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2039 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2040 "Base pointers must match" ); 2041 #ifdef _LP64 2042 if (UseCompressedOops && 2043 addp->Opcode() == Op_ConP && 2044 addp == n->in(AddPNode::Base) && 2045 n->in(AddPNode::Offset)->is_Con()) { 2046 // Use addressing with narrow klass to load with offset on x86. 2047 // On sparc loading 32-bits constant and decoding it have less 2048 // instructions (4) then load 64-bits constant (7). 2049 // Do this transformation here since IGVN will convert ConN back to ConP. 2050 const Type* t = addp->bottom_type(); 2051 if (t->isa_oopptr()) { 2052 Node* nn = NULL; 2053 2054 // Look for existing ConN node of the same exact type. 2055 Compile* C = Compile::current(); 2056 Node* r = C->root(); 2057 uint cnt = r->outcnt(); 2058 for (uint i = 0; i < cnt; i++) { 2059 Node* m = r->raw_out(i); 2060 if (m!= NULL && m->Opcode() == Op_ConN && 2061 m->bottom_type()->make_ptr() == t) { 2062 nn = m; 2063 break; 2064 } 2065 } 2066 if (nn != NULL) { 2067 // Decode a narrow oop to match address 2068 // [R12 + narrow_oop_reg<<3 + offset] 2069 nn = new (C, 2) DecodeNNode(nn, t); 2070 n->set_req(AddPNode::Base, nn); 2071 n->set_req(AddPNode::Address, nn); 2072 if (addp->outcnt() == 0) { 2073 addp->disconnect_inputs(NULL); 2074 } 2075 } 2076 } 2077 } 2078 #endif 2079 break; 2080 } 2081 2082 #ifdef _LP64 2083 case Op_CastPP: 2084 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) { 2085 Compile* C = Compile::current(); 2086 Node* in1 = n->in(1); 2087 const Type* t = n->bottom_type(); 2088 Node* new_in1 = in1->clone(); 2089 new_in1->as_DecodeN()->set_type(t); 2090 2091 if (!Matcher::clone_shift_expressions) { 2092 // 2093 // x86, ARM and friends can handle 2 adds in addressing mode 2094 // and Matcher can fold a DecodeN node into address by using 2095 // a narrow oop directly and do implicit NULL check in address: 2096 // 2097 // [R12 + narrow_oop_reg<<3 + offset] 2098 // NullCheck narrow_oop_reg 2099 // 2100 // On other platforms (Sparc) we have to keep new DecodeN node and 2101 // use it to do implicit NULL check in address: 2102 // 2103 // decode_not_null narrow_oop_reg, base_reg 2104 // [base_reg + offset] 2105 // NullCheck base_reg 2106 // 2107 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2108 // to keep the information to which NULL check the new DecodeN node 2109 // corresponds to use it as value in implicit_null_check(). 2110 // 2111 new_in1->set_req(0, n->in(0)); 2112 } 2113 2114 n->subsume_by(new_in1); 2115 if (in1->outcnt() == 0) { 2116 in1->disconnect_inputs(NULL); 2117 } 2118 } 2119 break; 2120 2121 case Op_CmpP: 2122 // Do this transformation here to preserve CmpPNode::sub() and 2123 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2124 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2125 Node* in1 = n->in(1); 2126 Node* in2 = n->in(2); 2127 if (!in1->is_DecodeN()) { 2128 in2 = in1; 2129 in1 = n->in(2); 2130 } 2131 assert(in1->is_DecodeN(), "sanity"); 2132 2133 Compile* C = Compile::current(); 2134 Node* new_in2 = NULL; 2135 if (in2->is_DecodeN()) { 2136 new_in2 = in2->in(1); 2137 } else if (in2->Opcode() == Op_ConP) { 2138 const Type* t = in2->bottom_type(); 2139 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) { 2140 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2141 // 2142 // This transformation together with CastPP transformation above 2143 // will generated code for implicit NULL checks for compressed oops. 2144 // 2145 // The original code after Optimize() 2146 // 2147 // LoadN memory, narrow_oop_reg 2148 // decode narrow_oop_reg, base_reg 2149 // CmpP base_reg, NULL 2150 // CastPP base_reg // NotNull 2151 // Load [base_reg + offset], val_reg 2152 // 2153 // after these transformations will be 2154 // 2155 // LoadN memory, narrow_oop_reg 2156 // CmpN narrow_oop_reg, NULL 2157 // decode_not_null narrow_oop_reg, base_reg 2158 // Load [base_reg + offset], val_reg 2159 // 2160 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2161 // since narrow oops can be used in debug info now (see the code in 2162 // final_graph_reshaping_walk()). 2163 // 2164 // At the end the code will be matched to 2165 // on x86: 2166 // 2167 // Load_narrow_oop memory, narrow_oop_reg 2168 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2169 // NullCheck narrow_oop_reg 2170 // 2171 // and on sparc: 2172 // 2173 // Load_narrow_oop memory, narrow_oop_reg 2174 // decode_not_null narrow_oop_reg, base_reg 2175 // Load [base_reg + offset], val_reg 2176 // NullCheck base_reg 2177 // 2178 } else if (t->isa_oopptr()) { 2179 new_in2 = ConNode::make(C, t->make_narrowoop()); 2180 } 2181 } 2182 if (new_in2 != NULL) { 2183 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2184 n->subsume_by( cmpN ); 2185 if (in1->outcnt() == 0) { 2186 in1->disconnect_inputs(NULL); 2187 } 2188 if (in2->outcnt() == 0) { 2189 in2->disconnect_inputs(NULL); 2190 } 2191 } 2192 } 2193 break; 2194 2195 case Op_DecodeN: 2196 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2197 // DecodeN could be pinned on Sparc where it can't be fold into 2198 // an address expression, see the code for Op_CastPP above. 2199 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc"); 2200 break; 2201 2202 case Op_EncodeP: { 2203 Node* in1 = n->in(1); 2204 if (in1->is_DecodeN()) { 2205 n->subsume_by(in1->in(1)); 2206 } else if (in1->Opcode() == Op_ConP) { 2207 Compile* C = Compile::current(); 2208 const Type* t = in1->bottom_type(); 2209 if (t == TypePtr::NULL_PTR) { 2210 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2211 } else if (t->isa_oopptr()) { 2212 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2213 } 2214 } 2215 if (in1->outcnt() == 0) { 2216 in1->disconnect_inputs(NULL); 2217 } 2218 break; 2219 } 2220 2221 case Op_Phi: 2222 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2223 // The EncodeP optimization may create Phi with the same edges 2224 // for all paths. It is not handled well by Register Allocator. 2225 Node* unique_in = n->in(1); 2226 assert(unique_in != NULL, ""); 2227 uint cnt = n->req(); 2228 for (uint i = 2; i < cnt; i++) { 2229 Node* m = n->in(i); 2230 assert(m != NULL, ""); 2231 if (unique_in != m) 2232 unique_in = NULL; 2233 } 2234 if (unique_in != NULL) { 2235 n->subsume_by(unique_in); 2236 } 2237 } 2238 break; 2239 2240 #endif 2241 2242 case Op_ModI: 2243 if (UseDivMod) { 2244 // Check if a%b and a/b both exist 2245 Node* d = n->find_similar(Op_DivI); 2246 if (d) { 2247 // Replace them with a fused divmod if supported 2248 Compile* C = Compile::current(); 2249 if (Matcher::has_match_rule(Op_DivModI)) { 2250 DivModINode* divmod = DivModINode::make(C, n); 2251 d->subsume_by(divmod->div_proj()); 2252 n->subsume_by(divmod->mod_proj()); 2253 } else { 2254 // replace a%b with a-((a/b)*b) 2255 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2256 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2257 n->subsume_by( sub ); 2258 } 2259 } 2260 } 2261 break; 2262 2263 case Op_ModL: 2264 if (UseDivMod) { 2265 // Check if a%b and a/b both exist 2266 Node* d = n->find_similar(Op_DivL); 2267 if (d) { 2268 // Replace them with a fused divmod if supported 2269 Compile* C = Compile::current(); 2270 if (Matcher::has_match_rule(Op_DivModL)) { 2271 DivModLNode* divmod = DivModLNode::make(C, n); 2272 d->subsume_by(divmod->div_proj()); 2273 n->subsume_by(divmod->mod_proj()); 2274 } else { 2275 // replace a%b with a-((a/b)*b) 2276 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2277 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2278 n->subsume_by( sub ); 2279 } 2280 } 2281 } 2282 break; 2283 2284 case Op_Load16B: 2285 case Op_Load8B: 2286 case Op_Load4B: 2287 case Op_Load8S: 2288 case Op_Load4S: 2289 case Op_Load2S: 2290 case Op_Load8C: 2291 case Op_Load4C: 2292 case Op_Load2C: 2293 case Op_Load4I: 2294 case Op_Load2I: 2295 case Op_Load2L: 2296 case Op_Load4F: 2297 case Op_Load2F: 2298 case Op_Load2D: 2299 case Op_Store16B: 2300 case Op_Store8B: 2301 case Op_Store4B: 2302 case Op_Store8C: 2303 case Op_Store4C: 2304 case Op_Store2C: 2305 case Op_Store4I: 2306 case Op_Store2I: 2307 case Op_Store2L: 2308 case Op_Store4F: 2309 case Op_Store2F: 2310 case Op_Store2D: 2311 break; 2312 2313 case Op_PackB: 2314 case Op_PackS: 2315 case Op_PackC: 2316 case Op_PackI: 2317 case Op_PackF: 2318 case Op_PackL: 2319 case Op_PackD: 2320 if (n->req()-1 > 2) { 2321 // Replace many operand PackNodes with a binary tree for matching 2322 PackNode* p = (PackNode*) n; 2323 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2324 n->subsume_by(btp); 2325 } 2326 break; 2327 default: 2328 assert( !n->is_Call(), "" ); 2329 assert( !n->is_Mem(), "" ); 2330 break; 2331 } 2332 2333 // Collect CFG split points 2334 if (n->is_MultiBranch()) 2335 fpu._tests.push(n); 2336 } 2337 2338 //------------------------------final_graph_reshaping_walk--------------------- 2339 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2340 // requires that the walk visits a node's inputs before visiting the node. 2341 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { 2342 ResourceArea *area = Thread::current()->resource_area(); 2343 Unique_Node_List sfpt(area); 2344 2345 fpu._visited.set(root->_idx); // first, mark node as visited 2346 uint cnt = root->req(); 2347 Node *n = root; 2348 uint i = 0; 2349 while (true) { 2350 if (i < cnt) { 2351 // Place all non-visited non-null inputs onto stack 2352 Node* m = n->in(i); 2353 ++i; 2354 if (m != NULL && !fpu._visited.test_set(m->_idx)) { 2355 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2356 sfpt.push(m); 2357 cnt = m->req(); 2358 nstack.push(n, i); // put on stack parent and next input's index 2359 n = m; 2360 i = 0; 2361 } 2362 } else { 2363 // Now do post-visit work 2364 final_graph_reshaping_impl( n, fpu ); 2365 if (nstack.is_empty()) 2366 break; // finished 2367 n = nstack.node(); // Get node from stack 2368 cnt = n->req(); 2369 i = nstack.index(); 2370 nstack.pop(); // Shift to the next node on stack 2371 } 2372 } 2373 2374 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2375 // It could be done for an uncommon traps or any safepoints/calls 2376 // if the DecodeN node is referenced only in a debug info. 2377 while (sfpt.size() > 0) { 2378 n = sfpt.pop(); 2379 JVMState *jvms = n->as_SafePoint()->jvms(); 2380 assert(jvms != NULL, "sanity"); 2381 int start = jvms->debug_start(); 2382 int end = n->req(); 2383 bool is_uncommon = (n->is_CallStaticJava() && 2384 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2385 for (int j = start; j < end; j++) { 2386 Node* in = n->in(j); 2387 if (in->is_DecodeN()) { 2388 bool safe_to_skip = true; 2389 if (!is_uncommon ) { 2390 // Is it safe to skip? 2391 for (uint i = 0; i < in->outcnt(); i++) { 2392 Node* u = in->raw_out(i); 2393 if (!u->is_SafePoint() || 2394 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2395 safe_to_skip = false; 2396 } 2397 } 2398 } 2399 if (safe_to_skip) { 2400 n->set_req(j, in->in(1)); 2401 } 2402 if (in->outcnt() == 0) { 2403 in->disconnect_inputs(NULL); 2404 } 2405 } 2406 } 2407 } 2408 } 2409 2410 //------------------------------final_graph_reshaping-------------------------- 2411 // Final Graph Reshaping. 2412 // 2413 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2414 // and not commoned up and forced early. Must come after regular 2415 // optimizations to avoid GVN undoing the cloning. Clone constant 2416 // inputs to Loop Phis; these will be split by the allocator anyways. 2417 // Remove Opaque nodes. 2418 // (2) Move last-uses by commutative operations to the left input to encourage 2419 // Intel update-in-place two-address operations and better register usage 2420 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2421 // calls canonicalizing them back. 2422 // (3) Count the number of double-precision FP ops, single-precision FP ops 2423 // and call sites. On Intel, we can get correct rounding either by 2424 // forcing singles to memory (requires extra stores and loads after each 2425 // FP bytecode) or we can set a rounding mode bit (requires setting and 2426 // clearing the mode bit around call sites). The mode bit is only used 2427 // if the relative frequency of single FP ops to calls is low enough. 2428 // This is a key transform for SPEC mpeg_audio. 2429 // (4) Detect infinite loops; blobs of code reachable from above but not 2430 // below. Several of the Code_Gen algorithms fail on such code shapes, 2431 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2432 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2433 // Detection is by looking for IfNodes where only 1 projection is 2434 // reachable from below or CatchNodes missing some targets. 2435 // (5) Assert for insane oop offsets in debug mode. 2436 2437 bool Compile::final_graph_reshaping() { 2438 // an infinite loop may have been eliminated by the optimizer, 2439 // in which case the graph will be empty. 2440 if (root()->req() == 1) { 2441 record_method_not_compilable("trivial infinite loop"); 2442 return true; 2443 } 2444 2445 Final_Reshape_Counts fpu; 2446 2447 // Visit everybody reachable! 2448 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2449 Node_Stack nstack(unique() >> 1); 2450 final_graph_reshaping_walk(nstack, root(), fpu); 2451 2452 // Check for unreachable (from below) code (i.e., infinite loops). 2453 for( uint i = 0; i < fpu._tests.size(); i++ ) { 2454 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch(); 2455 // Get number of CFG targets. 2456 // Note that PCTables include exception targets after calls. 2457 uint required_outcnt = n->required_outcnt(); 2458 if (n->outcnt() != required_outcnt) { 2459 // Check for a few special cases. Rethrow Nodes never take the 2460 // 'fall-thru' path, so expected kids is 1 less. 2461 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2462 if (n->in(0)->in(0)->is_Call()) { 2463 CallNode *call = n->in(0)->in(0)->as_Call(); 2464 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2465 required_outcnt--; // Rethrow always has 1 less kid 2466 } else if (call->req() > TypeFunc::Parms && 2467 call->is_CallDynamicJava()) { 2468 // Check for null receiver. In such case, the optimizer has 2469 // detected that the virtual call will always result in a null 2470 // pointer exception. The fall-through projection of this CatchNode 2471 // will not be populated. 2472 Node *arg0 = call->in(TypeFunc::Parms); 2473 if (arg0->is_Type() && 2474 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2475 required_outcnt--; 2476 } 2477 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2478 call->req() > TypeFunc::Parms+1 && 2479 call->is_CallStaticJava()) { 2480 // Check for negative array length. In such case, the optimizer has 2481 // detected that the allocation attempt will always result in an 2482 // exception. There is no fall-through projection of this CatchNode . 2483 Node *arg1 = call->in(TypeFunc::Parms+1); 2484 if (arg1->is_Type() && 2485 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2486 required_outcnt--; 2487 } 2488 } 2489 } 2490 } 2491 // Recheck with a better notion of 'required_outcnt' 2492 if (n->outcnt() != required_outcnt) { 2493 record_method_not_compilable("malformed control flow"); 2494 return true; // Not all targets reachable! 2495 } 2496 } 2497 // Check that I actually visited all kids. Unreached kids 2498 // must be infinite loops. 2499 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2500 if (!fpu._visited.test(n->fast_out(j)->_idx)) { 2501 record_method_not_compilable("infinite loop"); 2502 return true; // Found unvisited kid; must be unreach 2503 } 2504 } 2505 2506 // If original bytecodes contained a mixture of floats and doubles 2507 // check if the optimizer has made it homogenous, item (3). 2508 if( Use24BitFPMode && Use24BitFP && 2509 fpu.get_float_count() > 32 && 2510 fpu.get_double_count() == 0 && 2511 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { 2512 set_24_bit_selection_and_mode( false, true ); 2513 } 2514 2515 set_has_java_calls(fpu.get_java_call_count() > 0); 2516 2517 // No infinite loops, no reason to bail out. 2518 return false; 2519 } 2520 2521 //-----------------------------too_many_traps---------------------------------- 2522 // Report if there are too many traps at the current method and bci. 2523 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2524 bool Compile::too_many_traps(ciMethod* method, 2525 int bci, 2526 Deoptimization::DeoptReason reason) { 2527 ciMethodData* md = method->method_data(); 2528 if (md->is_empty()) { 2529 // Assume the trap has not occurred, or that it occurred only 2530 // because of a transient condition during start-up in the interpreter. 2531 return false; 2532 } 2533 if (md->has_trap_at(bci, reason) != 0) { 2534 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2535 // Also, if there are multiple reasons, or if there is no per-BCI record, 2536 // assume the worst. 2537 if (log()) 2538 log()->elem("observe trap='%s' count='%d'", 2539 Deoptimization::trap_reason_name(reason), 2540 md->trap_count(reason)); 2541 return true; 2542 } else { 2543 // Ignore method/bci and see if there have been too many globally. 2544 return too_many_traps(reason, md); 2545 } 2546 } 2547 2548 // Less-accurate variant which does not require a method and bci. 2549 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2550 ciMethodData* logmd) { 2551 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2552 // Too many traps globally. 2553 // Note that we use cumulative trap_count, not just md->trap_count. 2554 if (log()) { 2555 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2556 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2557 Deoptimization::trap_reason_name(reason), 2558 mcount, trap_count(reason)); 2559 } 2560 return true; 2561 } else { 2562 // The coast is clear. 2563 return false; 2564 } 2565 } 2566 2567 //--------------------------too_many_recompiles-------------------------------- 2568 // Report if there are too many recompiles at the current method and bci. 2569 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2570 // Is not eager to return true, since this will cause the compiler to use 2571 // Action_none for a trap point, to avoid too many recompilations. 2572 bool Compile::too_many_recompiles(ciMethod* method, 2573 int bci, 2574 Deoptimization::DeoptReason reason) { 2575 ciMethodData* md = method->method_data(); 2576 if (md->is_empty()) { 2577 // Assume the trap has not occurred, or that it occurred only 2578 // because of a transient condition during start-up in the interpreter. 2579 return false; 2580 } 2581 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2582 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2583 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2584 Deoptimization::DeoptReason per_bc_reason 2585 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2586 if ((per_bc_reason == Deoptimization::Reason_none 2587 || md->has_trap_at(bci, reason) != 0) 2588 // The trap frequency measure we care about is the recompile count: 2589 && md->trap_recompiled_at(bci) 2590 && md->overflow_recompile_count() >= bc_cutoff) { 2591 // Do not emit a trap here if it has already caused recompilations. 2592 // Also, if there are multiple reasons, or if there is no per-BCI record, 2593 // assume the worst. 2594 if (log()) 2595 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2596 Deoptimization::trap_reason_name(reason), 2597 md->trap_count(reason), 2598 md->overflow_recompile_count()); 2599 return true; 2600 } else if (trap_count(reason) != 0 2601 && decompile_count() >= m_cutoff) { 2602 // Too many recompiles globally, and we have seen this sort of trap. 2603 // Use cumulative decompile_count, not just md->decompile_count. 2604 if (log()) 2605 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2606 Deoptimization::trap_reason_name(reason), 2607 md->trap_count(reason), trap_count(reason), 2608 md->decompile_count(), decompile_count()); 2609 return true; 2610 } else { 2611 // The coast is clear. 2612 return false; 2613 } 2614 } 2615 2616 2617 #ifndef PRODUCT 2618 //------------------------------verify_graph_edges--------------------------- 2619 // Walk the Graph and verify that there is a one-to-one correspondence 2620 // between Use-Def edges and Def-Use edges in the graph. 2621 void Compile::verify_graph_edges(bool no_dead_code) { 2622 if (VerifyGraphEdges) { 2623 ResourceArea *area = Thread::current()->resource_area(); 2624 Unique_Node_List visited(area); 2625 // Call recursive graph walk to check edges 2626 _root->verify_edges(visited); 2627 if (no_dead_code) { 2628 // Now make sure that no visited node is used by an unvisited node. 2629 bool dead_nodes = 0; 2630 Unique_Node_List checked(area); 2631 while (visited.size() > 0) { 2632 Node* n = visited.pop(); 2633 checked.push(n); 2634 for (uint i = 0; i < n->outcnt(); i++) { 2635 Node* use = n->raw_out(i); 2636 if (checked.member(use)) continue; // already checked 2637 if (visited.member(use)) continue; // already in the graph 2638 if (use->is_Con()) continue; // a dead ConNode is OK 2639 // At this point, we have found a dead node which is DU-reachable. 2640 if (dead_nodes++ == 0) 2641 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2642 use->dump(2); 2643 tty->print_cr("---"); 2644 checked.push(use); // No repeats; pretend it is now checked. 2645 } 2646 } 2647 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2648 } 2649 } 2650 } 2651 #endif 2652 2653 // The Compile object keeps track of failure reasons separately from the ciEnv. 2654 // This is required because there is not quite a 1-1 relation between the 2655 // ciEnv and its compilation task and the Compile object. Note that one 2656 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2657 // to backtrack and retry without subsuming loads. Other than this backtracking 2658 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2659 // by the logic in C2Compiler. 2660 void Compile::record_failure(const char* reason) { 2661 if (log() != NULL) { 2662 log()->elem("failure reason='%s' phase='compile'", reason); 2663 } 2664 if (_failure_reason == NULL) { 2665 // Record the first failure reason. 2666 _failure_reason = reason; 2667 } 2668 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2669 C->print_method(_failure_reason); 2670 } 2671 _root = NULL; // flush the graph, too 2672 } 2673 2674 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2675 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2676 { 2677 if (dolog) { 2678 C = Compile::current(); 2679 _log = C->log(); 2680 } else { 2681 C = NULL; 2682 _log = NULL; 2683 } 2684 if (_log != NULL) { 2685 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2686 _log->stamp(); 2687 _log->end_head(); 2688 } 2689 } 2690 2691 Compile::TracePhase::~TracePhase() { 2692 if (_log != NULL) { 2693 _log->done("phase nodes='%d'", C->unique()); 2694 } 2695 }