1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_compile.cpp.incl" 27 28 /// Support for intrinsics. 29 30 // Return the index at which m must be inserted (or already exists). 31 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 33 #ifdef ASSERT 34 for (int i = 1; i < _intrinsics->length(); i++) { 35 CallGenerator* cg1 = _intrinsics->at(i-1); 36 CallGenerator* cg2 = _intrinsics->at(i); 37 assert(cg1->method() != cg2->method() 38 ? cg1->method() < cg2->method() 39 : cg1->is_virtual() < cg2->is_virtual(), 40 "compiler intrinsics list must stay sorted"); 41 } 42 #endif 43 // Binary search sorted list, in decreasing intervals [lo, hi]. 44 int lo = 0, hi = _intrinsics->length()-1; 45 while (lo <= hi) { 46 int mid = (uint)(hi + lo) / 2; 47 ciMethod* mid_m = _intrinsics->at(mid)->method(); 48 if (m < mid_m) { 49 hi = mid-1; 50 } else if (m > mid_m) { 51 lo = mid+1; 52 } else { 53 // look at minor sort key 54 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 55 if (is_virtual < mid_virt) { 56 hi = mid-1; 57 } else if (is_virtual > mid_virt) { 58 lo = mid+1; 59 } else { 60 return mid; // exact match 61 } 62 } 63 } 64 return lo; // inexact match 65 } 66 67 void Compile::register_intrinsic(CallGenerator* cg) { 68 if (_intrinsics == NULL) { 69 _intrinsics = new GrowableArray<CallGenerator*>(60); 70 } 71 // This code is stolen from ciObjectFactory::insert. 72 // Really, GrowableArray should have methods for 73 // insert_at, remove_at, and binary_search. 74 int len = _intrinsics->length(); 75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 76 if (index == len) { 77 _intrinsics->append(cg); 78 } else { 79 #ifdef ASSERT 80 CallGenerator* oldcg = _intrinsics->at(index); 81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 82 #endif 83 _intrinsics->append(_intrinsics->at(len-1)); 84 int pos; 85 for (pos = len-2; pos >= index; pos--) { 86 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 87 } 88 _intrinsics->at_put(index, cg); 89 } 90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 91 } 92 93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 94 assert(m->is_loaded(), "don't try this on unloaded methods"); 95 if (_intrinsics != NULL) { 96 int index = intrinsic_insertion_index(m, is_virtual); 97 if (index < _intrinsics->length() 98 && _intrinsics->at(index)->method() == m 99 && _intrinsics->at(index)->is_virtual() == is_virtual) { 100 return _intrinsics->at(index); 101 } 102 } 103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 104 if (m->intrinsic_id() != vmIntrinsics::_none && 105 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 106 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 107 if (cg != NULL) { 108 // Save it for next time: 109 register_intrinsic(cg); 110 return cg; 111 } else { 112 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 113 } 114 } 115 return NULL; 116 } 117 118 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 119 // in library_call.cpp. 120 121 122 #ifndef PRODUCT 123 // statistics gathering... 124 125 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 126 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 127 128 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 129 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 130 int oflags = _intrinsic_hist_flags[id]; 131 assert(flags != 0, "what happened?"); 132 if (is_virtual) { 133 flags |= _intrinsic_virtual; 134 } 135 bool changed = (flags != oflags); 136 if ((flags & _intrinsic_worked) != 0) { 137 juint count = (_intrinsic_hist_count[id] += 1); 138 if (count == 1) { 139 changed = true; // first time 140 } 141 // increment the overall count also: 142 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 143 } 144 if (changed) { 145 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 146 // Something changed about the intrinsic's virtuality. 147 if ((flags & _intrinsic_virtual) != 0) { 148 // This is the first use of this intrinsic as a virtual call. 149 if (oflags != 0) { 150 // We already saw it as a non-virtual, so note both cases. 151 flags |= _intrinsic_both; 152 } 153 } else if ((oflags & _intrinsic_both) == 0) { 154 // This is the first use of this intrinsic as a non-virtual 155 flags |= _intrinsic_both; 156 } 157 } 158 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 159 } 160 // update the overall flags also: 161 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 162 return changed; 163 } 164 165 static char* format_flags(int flags, char* buf) { 166 buf[0] = 0; 167 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 168 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 169 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 170 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 171 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 172 if (buf[0] == 0) strcat(buf, ","); 173 assert(buf[0] == ',', "must be"); 174 return &buf[1]; 175 } 176 177 void Compile::print_intrinsic_statistics() { 178 char flagsbuf[100]; 179 ttyLocker ttyl; 180 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 181 tty->print_cr("Compiler intrinsic usage:"); 182 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 183 if (total == 0) total = 1; // avoid div0 in case of no successes 184 #define PRINT_STAT_LINE(name, c, f) \ 185 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 186 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 187 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 188 int flags = _intrinsic_hist_flags[id]; 189 juint count = _intrinsic_hist_count[id]; 190 if ((flags | count) != 0) { 191 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 192 } 193 } 194 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 195 if (xtty != NULL) xtty->tail("statistics"); 196 } 197 198 void Compile::print_statistics() { 199 { ttyLocker ttyl; 200 if (xtty != NULL) xtty->head("statistics type='opto'"); 201 Parse::print_statistics(); 202 PhaseCCP::print_statistics(); 203 PhaseRegAlloc::print_statistics(); 204 Scheduling::print_statistics(); 205 PhasePeephole::print_statistics(); 206 PhaseIdealLoop::print_statistics(); 207 if (xtty != NULL) xtty->tail("statistics"); 208 } 209 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 210 // put this under its own <statistics> element. 211 print_intrinsic_statistics(); 212 } 213 } 214 #endif //PRODUCT 215 216 // Support for bundling info 217 Bundle* Compile::node_bundling(const Node *n) { 218 assert(valid_bundle_info(n), "oob"); 219 return &_node_bundling_base[n->_idx]; 220 } 221 222 bool Compile::valid_bundle_info(const Node *n) { 223 return (_node_bundling_limit > n->_idx); 224 } 225 226 227 // Identify all nodes that are reachable from below, useful. 228 // Use breadth-first pass that records state in a Unique_Node_List, 229 // recursive traversal is slower. 230 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 231 int estimated_worklist_size = unique(); 232 useful.map( estimated_worklist_size, NULL ); // preallocate space 233 234 // Initialize worklist 235 if (root() != NULL) { useful.push(root()); } 236 // If 'top' is cached, declare it useful to preserve cached node 237 if( cached_top_node() ) { useful.push(cached_top_node()); } 238 239 // Push all useful nodes onto the list, breadthfirst 240 for( uint next = 0; next < useful.size(); ++next ) { 241 assert( next < unique(), "Unique useful nodes < total nodes"); 242 Node *n = useful.at(next); 243 uint max = n->len(); 244 for( uint i = 0; i < max; ++i ) { 245 Node *m = n->in(i); 246 if( m == NULL ) continue; 247 useful.push(m); 248 } 249 } 250 } 251 252 // Disconnect all useless nodes by disconnecting those at the boundary. 253 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 254 uint next = 0; 255 while( next < useful.size() ) { 256 Node *n = useful.at(next++); 257 // Use raw traversal of out edges since this code removes out edges 258 int max = n->outcnt(); 259 for (int j = 0; j < max; ++j ) { 260 Node* child = n->raw_out(j); 261 if( ! useful.member(child) ) { 262 assert( !child->is_top() || child != top(), 263 "If top is cached in Compile object it is in useful list"); 264 // Only need to remove this out-edge to the useless node 265 n->raw_del_out(j); 266 --j; 267 --max; 268 } 269 } 270 if (n->outcnt() == 1 && n->has_special_unique_user()) { 271 record_for_igvn( n->unique_out() ); 272 } 273 } 274 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 275 } 276 277 //------------------------------frame_size_in_words----------------------------- 278 // frame_slots in units of words 279 int Compile::frame_size_in_words() const { 280 // shift is 0 in LP32 and 1 in LP64 281 const int shift = (LogBytesPerWord - LogBytesPerInt); 282 int words = _frame_slots >> shift; 283 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 284 return words; 285 } 286 287 // ============================================================================ 288 //------------------------------CompileWrapper--------------------------------- 289 class CompileWrapper : public StackObj { 290 Compile *const _compile; 291 public: 292 CompileWrapper(Compile* compile); 293 294 ~CompileWrapper(); 295 }; 296 297 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 298 // the Compile* pointer is stored in the current ciEnv: 299 ciEnv* env = compile->env(); 300 assert(env == ciEnv::current(), "must already be a ciEnv active"); 301 assert(env->compiler_data() == NULL, "compile already active?"); 302 env->set_compiler_data(compile); 303 assert(compile == Compile::current(), "sanity"); 304 305 compile->set_type_dict(NULL); 306 compile->set_type_hwm(NULL); 307 compile->set_type_last_size(0); 308 compile->set_last_tf(NULL, NULL); 309 compile->set_indexSet_arena(NULL); 310 compile->set_indexSet_free_block_list(NULL); 311 compile->init_type_arena(); 312 Type::Initialize(compile); 313 _compile->set_scratch_buffer_blob(NULL); 314 _compile->begin_method(); 315 } 316 CompileWrapper::~CompileWrapper() { 317 _compile->end_method(); 318 if (_compile->scratch_buffer_blob() != NULL) 319 BufferBlob::free(_compile->scratch_buffer_blob()); 320 _compile->env()->set_compiler_data(NULL); 321 } 322 323 324 //----------------------------print_compile_messages--------------------------- 325 void Compile::print_compile_messages() { 326 #ifndef PRODUCT 327 // Check if recompiling 328 if (_subsume_loads == false && PrintOpto) { 329 // Recompiling without allowing machine instructions to subsume loads 330 tty->print_cr("*********************************************************"); 331 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 332 tty->print_cr("*********************************************************"); 333 } 334 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 335 // Recompiling without escape analysis 336 tty->print_cr("*********************************************************"); 337 tty->print_cr("** Bailout: Recompile without escape analysis **"); 338 tty->print_cr("*********************************************************"); 339 } 340 if (env()->break_at_compile()) { 341 // Open the debugger when compiling this method. 342 tty->print("### Breaking when compiling: "); 343 method()->print_short_name(); 344 tty->cr(); 345 BREAKPOINT; 346 } 347 348 if( PrintOpto ) { 349 if (is_osr_compilation()) { 350 tty->print("[OSR]%3d", _compile_id); 351 } else { 352 tty->print("%3d", _compile_id); 353 } 354 } 355 #endif 356 } 357 358 359 void Compile::init_scratch_buffer_blob() { 360 if( scratch_buffer_blob() != NULL ) return; 361 362 // Construct a temporary CodeBuffer to have it construct a BufferBlob 363 // Cache this BufferBlob for this compile. 364 ResourceMark rm; 365 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 366 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 367 // Record the buffer blob for next time. 368 set_scratch_buffer_blob(blob); 369 // Have we run out of code space? 370 if (scratch_buffer_blob() == NULL) { 371 // Let CompilerBroker disable further compilations. 372 record_failure("Not enough space for scratch buffer in CodeCache"); 373 return; 374 } 375 376 // Initialize the relocation buffers 377 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 378 set_scratch_locs_memory(locs_buf); 379 } 380 381 382 //-----------------------scratch_emit_size------------------------------------- 383 // Helper function that computes size by emitting code 384 uint Compile::scratch_emit_size(const Node* n) { 385 // Emit into a trash buffer and count bytes emitted. 386 // This is a pretty expensive way to compute a size, 387 // but it works well enough if seldom used. 388 // All common fixed-size instructions are given a size 389 // method by the AD file. 390 // Note that the scratch buffer blob and locs memory are 391 // allocated at the beginning of the compile task, and 392 // may be shared by several calls to scratch_emit_size. 393 // The allocation of the scratch buffer blob is particularly 394 // expensive, since it has to grab the code cache lock. 395 BufferBlob* blob = this->scratch_buffer_blob(); 396 assert(blob != NULL, "Initialize BufferBlob at start"); 397 assert(blob->size() > MAX_inst_size, "sanity"); 398 relocInfo* locs_buf = scratch_locs_memory(); 399 address blob_begin = blob->instructions_begin(); 400 address blob_end = (address)locs_buf; 401 assert(blob->instructions_contains(blob_end), "sanity"); 402 CodeBuffer buf(blob_begin, blob_end - blob_begin); 403 buf.initialize_consts_size(MAX_const_size); 404 buf.initialize_stubs_size(MAX_stubs_size); 405 assert(locs_buf != NULL, "sanity"); 406 int lsize = MAX_locs_size / 2; 407 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 408 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 409 n->emit(buf, this->regalloc()); 410 return buf.code_size(); 411 } 412 413 414 // ============================================================================ 415 //------------------------------Compile standard------------------------------- 416 debug_only( int Compile::_debug_idx = 100000; ) 417 418 // Compile a method. entry_bci is -1 for normal compilations and indicates 419 // the continuation bci for on stack replacement. 420 421 422 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 423 : Phase(Compiler), 424 _env(ci_env), 425 _log(ci_env->log()), 426 _compile_id(ci_env->compile_id()), 427 _save_argument_registers(false), 428 _stub_name(NULL), 429 _stub_function(NULL), 430 _stub_entry_point(NULL), 431 _method(target), 432 _entry_bci(osr_bci), 433 _initial_gvn(NULL), 434 _for_igvn(NULL), 435 _warm_calls(NULL), 436 _subsume_loads(subsume_loads), 437 _do_escape_analysis(do_escape_analysis), 438 _failure_reason(NULL), 439 _code_buffer("Compile::Fill_buffer"), 440 _orig_pc_slot(0), 441 _orig_pc_slot_offset_in_bytes(0), 442 _node_bundling_limit(0), 443 _node_bundling_base(NULL), 444 #ifndef PRODUCT 445 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 446 _printer(IdealGraphPrinter::printer()), 447 #endif 448 _congraph(NULL) { 449 C = this; 450 451 CompileWrapper cw(this); 452 #ifndef PRODUCT 453 if (TimeCompiler2) { 454 tty->print(" "); 455 target->holder()->name()->print(); 456 tty->print("."); 457 target->print_short_name(); 458 tty->print(" "); 459 } 460 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 461 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 462 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 463 if (!print_opto_assembly) { 464 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 465 if (print_assembly && !Disassembler::can_decode()) { 466 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 467 print_opto_assembly = true; 468 } 469 } 470 set_print_assembly(print_opto_assembly); 471 set_parsed_irreducible_loop(false); 472 #endif 473 474 if (ProfileTraps) { 475 // Make sure the method being compiled gets its own MDO, 476 // so we can at least track the decompile_count(). 477 method()->build_method_data(); 478 } 479 480 Init(::AliasLevel); 481 482 483 print_compile_messages(); 484 485 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 486 _ilt = InlineTree::build_inline_tree_root(); 487 else 488 _ilt = NULL; 489 490 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 491 assert(num_alias_types() >= AliasIdxRaw, ""); 492 493 #define MINIMUM_NODE_HASH 1023 494 // Node list that Iterative GVN will start with 495 Unique_Node_List for_igvn(comp_arena()); 496 set_for_igvn(&for_igvn); 497 498 // GVN that will be run immediately on new nodes 499 uint estimated_size = method()->code_size()*4+64; 500 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 501 PhaseGVN gvn(node_arena(), estimated_size); 502 set_initial_gvn(&gvn); 503 504 { // Scope for timing the parser 505 TracePhase t3("parse", &_t_parser, true); 506 507 // Put top into the hash table ASAP. 508 initial_gvn()->transform_no_reclaim(top()); 509 510 // Set up tf(), start(), and find a CallGenerator. 511 CallGenerator* cg; 512 if (is_osr_compilation()) { 513 const TypeTuple *domain = StartOSRNode::osr_domain(); 514 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 515 init_tf(TypeFunc::make(domain, range)); 516 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 517 initial_gvn()->set_type_bottom(s); 518 init_start(s); 519 cg = CallGenerator::for_osr(method(), entry_bci()); 520 } else { 521 // Normal case. 522 init_tf(TypeFunc::make(method())); 523 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 524 initial_gvn()->set_type_bottom(s); 525 init_start(s); 526 float past_uses = method()->interpreter_invocation_count(); 527 float expected_uses = past_uses; 528 cg = CallGenerator::for_inline(method(), expected_uses); 529 } 530 if (failing()) return; 531 if (cg == NULL) { 532 record_method_not_compilable_all_tiers("cannot parse method"); 533 return; 534 } 535 JVMState* jvms = build_start_state(start(), tf()); 536 if ((jvms = cg->generate(jvms)) == NULL) { 537 record_method_not_compilable("method parse failed"); 538 return; 539 } 540 GraphKit kit(jvms); 541 542 if (!kit.stopped()) { 543 // Accept return values, and transfer control we know not where. 544 // This is done by a special, unique ReturnNode bound to root. 545 return_values(kit.jvms()); 546 } 547 548 if (kit.has_exceptions()) { 549 // Any exceptions that escape from this call must be rethrown 550 // to whatever caller is dynamically above us on the stack. 551 // This is done by a special, unique RethrowNode bound to root. 552 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 553 } 554 555 print_method("Before RemoveUseless", 3); 556 557 // Remove clutter produced by parsing. 558 if (!failing()) { 559 ResourceMark rm; 560 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 561 } 562 } 563 564 // Note: Large methods are capped off in do_one_bytecode(). 565 if (failing()) return; 566 567 // After parsing, node notes are no longer automagic. 568 // They must be propagated by register_new_node_with_optimizer(), 569 // clone(), or the like. 570 set_default_node_notes(NULL); 571 572 for (;;) { 573 int successes = Inline_Warm(); 574 if (failing()) return; 575 if (successes == 0) break; 576 } 577 578 // Drain the list. 579 Finish_Warm(); 580 #ifndef PRODUCT 581 if (_printer) { 582 _printer->print_inlining(this); 583 } 584 #endif 585 586 if (failing()) return; 587 NOT_PRODUCT( verify_graph_edges(); ) 588 589 // Perform escape analysis 590 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 591 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 592 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction. 593 PhaseGVN* igvn = initial_gvn(); 594 Node* oop_null = igvn->zerocon(T_OBJECT); 595 Node* noop_null = igvn->zerocon(T_NARROWOOP); 596 597 _congraph = new(comp_arena()) ConnectionGraph(this); 598 bool has_non_escaping_obj = _congraph->compute_escape(); 599 600 #ifndef PRODUCT 601 if (PrintEscapeAnalysis) { 602 _congraph->dump(); 603 } 604 #endif 605 // Cleanup. 606 if (oop_null->outcnt() == 0) 607 igvn->hash_delete(oop_null); 608 if (noop_null->outcnt() == 0) 609 igvn->hash_delete(noop_null); 610 611 if (!has_non_escaping_obj) { 612 _congraph = NULL; 613 } 614 615 if (failing()) return; 616 } 617 // Now optimize 618 Optimize(); 619 if (failing()) return; 620 NOT_PRODUCT( verify_graph_edges(); ) 621 622 #ifndef PRODUCT 623 if (PrintIdeal) { 624 ttyLocker ttyl; // keep the following output all in one block 625 // This output goes directly to the tty, not the compiler log. 626 // To enable tools to match it up with the compilation activity, 627 // be sure to tag this tty output with the compile ID. 628 if (xtty != NULL) { 629 xtty->head("ideal compile_id='%d'%s", compile_id(), 630 is_osr_compilation() ? " compile_kind='osr'" : 631 ""); 632 } 633 root()->dump(9999); 634 if (xtty != NULL) { 635 xtty->tail("ideal"); 636 } 637 } 638 #endif 639 640 // Now that we know the size of all the monitors we can add a fixed slot 641 // for the original deopt pc. 642 643 _orig_pc_slot = fixed_slots(); 644 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 645 set_fixed_slots(next_slot); 646 647 // Now generate code 648 Code_Gen(); 649 if (failing()) return; 650 651 // Check if we want to skip execution of all compiled code. 652 { 653 #ifndef PRODUCT 654 if (OptoNoExecute) { 655 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 656 return; 657 } 658 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 659 #endif 660 661 if (is_osr_compilation()) { 662 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 663 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 664 } else { 665 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 666 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 667 } 668 669 env()->register_method(_method, _entry_bci, 670 &_code_offsets, 671 _orig_pc_slot_offset_in_bytes, 672 code_buffer(), 673 frame_size_in_words(), _oop_map_set, 674 &_handler_table, &_inc_table, 675 compiler, 676 env()->comp_level(), 677 true, /*has_debug_info*/ 678 has_unsafe_access() 679 ); 680 } 681 } 682 683 //------------------------------Compile---------------------------------------- 684 // Compile a runtime stub 685 Compile::Compile( ciEnv* ci_env, 686 TypeFunc_generator generator, 687 address stub_function, 688 const char *stub_name, 689 int is_fancy_jump, 690 bool pass_tls, 691 bool save_arg_registers, 692 bool return_pc ) 693 : Phase(Compiler), 694 _env(ci_env), 695 _log(ci_env->log()), 696 _compile_id(-1), 697 _save_argument_registers(save_arg_registers), 698 _method(NULL), 699 _stub_name(stub_name), 700 _stub_function(stub_function), 701 _stub_entry_point(NULL), 702 _entry_bci(InvocationEntryBci), 703 _initial_gvn(NULL), 704 _for_igvn(NULL), 705 _warm_calls(NULL), 706 _orig_pc_slot(0), 707 _orig_pc_slot_offset_in_bytes(0), 708 _subsume_loads(true), 709 _do_escape_analysis(false), 710 _failure_reason(NULL), 711 _code_buffer("Compile::Fill_buffer"), 712 _node_bundling_limit(0), 713 _node_bundling_base(NULL), 714 #ifndef PRODUCT 715 _trace_opto_output(TraceOptoOutput), 716 _printer(NULL), 717 #endif 718 _congraph(NULL) { 719 C = this; 720 721 #ifndef PRODUCT 722 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 723 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 724 set_print_assembly(PrintFrameConverterAssembly); 725 set_parsed_irreducible_loop(false); 726 #endif 727 CompileWrapper cw(this); 728 Init(/*AliasLevel=*/ 0); 729 init_tf((*generator)()); 730 731 { 732 // The following is a dummy for the sake of GraphKit::gen_stub 733 Unique_Node_List for_igvn(comp_arena()); 734 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 735 PhaseGVN gvn(Thread::current()->resource_area(),255); 736 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 737 gvn.transform_no_reclaim(top()); 738 739 GraphKit kit; 740 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 741 } 742 743 NOT_PRODUCT( verify_graph_edges(); ) 744 Code_Gen(); 745 if (failing()) return; 746 747 748 // Entry point will be accessed using compile->stub_entry_point(); 749 if (code_buffer() == NULL) { 750 Matcher::soft_match_failure(); 751 } else { 752 if (PrintAssembly && (WizardMode || Verbose)) 753 tty->print_cr("### Stub::%s", stub_name); 754 755 if (!failing()) { 756 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 757 758 // Make the NMethod 759 // For now we mark the frame as never safe for profile stackwalking 760 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 761 code_buffer(), 762 CodeOffsets::frame_never_safe, 763 // _code_offsets.value(CodeOffsets::Frame_Complete), 764 frame_size_in_words(), 765 _oop_map_set, 766 save_arg_registers); 767 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 768 769 _stub_entry_point = rs->entry_point(); 770 } 771 } 772 } 773 774 #ifndef PRODUCT 775 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 776 if(PrintOpto && Verbose) { 777 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 778 } 779 } 780 #endif 781 782 void Compile::print_codes() { 783 } 784 785 //------------------------------Init------------------------------------------- 786 // Prepare for a single compilation 787 void Compile::Init(int aliaslevel) { 788 _unique = 0; 789 _regalloc = NULL; 790 791 _tf = NULL; // filled in later 792 _top = NULL; // cached later 793 _matcher = NULL; // filled in later 794 _cfg = NULL; // filled in later 795 796 set_24_bit_selection_and_mode(Use24BitFP, false); 797 798 _node_note_array = NULL; 799 _default_node_notes = NULL; 800 801 _immutable_memory = NULL; // filled in at first inquiry 802 803 // Globally visible Nodes 804 // First set TOP to NULL to give safe behavior during creation of RootNode 805 set_cached_top_node(NULL); 806 set_root(new (this, 3) RootNode()); 807 // Now that you have a Root to point to, create the real TOP 808 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 809 set_recent_alloc(NULL, NULL); 810 811 // Create Debug Information Recorder to record scopes, oopmaps, etc. 812 env()->set_oop_recorder(new OopRecorder(comp_arena())); 813 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 814 env()->set_dependencies(new Dependencies(env())); 815 816 _fixed_slots = 0; 817 set_has_split_ifs(false); 818 set_has_loops(has_method() && method()->has_loops()); // first approximation 819 _deopt_happens = true; // start out assuming the worst 820 _trap_can_recompile = false; // no traps emitted yet 821 _major_progress = true; // start out assuming good things will happen 822 set_has_unsafe_access(false); 823 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 824 set_decompile_count(0); 825 826 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 827 // Compilation level related initialization 828 if (env()->comp_level() == CompLevel_fast_compile) { 829 set_num_loop_opts(Tier1LoopOptsCount); 830 set_do_inlining(Tier1Inline != 0); 831 set_max_inline_size(Tier1MaxInlineSize); 832 set_freq_inline_size(Tier1FreqInlineSize); 833 set_do_scheduling(false); 834 set_do_count_invocations(Tier1CountInvocations); 835 set_do_method_data_update(Tier1UpdateMethodData); 836 } else { 837 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 838 set_num_loop_opts(LoopOptsCount); 839 set_do_inlining(Inline); 840 set_max_inline_size(MaxInlineSize); 841 set_freq_inline_size(FreqInlineSize); 842 set_do_scheduling(OptoScheduling); 843 set_do_count_invocations(false); 844 set_do_method_data_update(false); 845 } 846 847 if (debug_info()->recording_non_safepoints()) { 848 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 849 (comp_arena(), 8, 0, NULL)); 850 set_default_node_notes(Node_Notes::make(this)); 851 } 852 853 // // -- Initialize types before each compile -- 854 // // Update cached type information 855 // if( _method && _method->constants() ) 856 // Type::update_loaded_types(_method, _method->constants()); 857 858 // Init alias_type map. 859 if (!_do_escape_analysis && aliaslevel == 3) 860 aliaslevel = 2; // No unique types without escape analysis 861 _AliasLevel = aliaslevel; 862 const int grow_ats = 16; 863 _max_alias_types = grow_ats; 864 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 865 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 866 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 867 { 868 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 869 } 870 // Initialize the first few types. 871 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 872 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 873 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 874 _num_alias_types = AliasIdxRaw+1; 875 // Zero out the alias type cache. 876 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 877 // A NULL adr_type hits in the cache right away. Preload the right answer. 878 probe_alias_cache(NULL)->_index = AliasIdxTop; 879 880 _intrinsics = NULL; 881 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 882 register_library_intrinsics(); 883 } 884 885 //---------------------------init_start---------------------------------------- 886 // Install the StartNode on this compile object. 887 void Compile::init_start(StartNode* s) { 888 if (failing()) 889 return; // already failing 890 assert(s == start(), ""); 891 } 892 893 StartNode* Compile::start() const { 894 assert(!failing(), ""); 895 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 896 Node* start = root()->fast_out(i); 897 if( start->is_Start() ) 898 return start->as_Start(); 899 } 900 ShouldNotReachHere(); 901 return NULL; 902 } 903 904 //-------------------------------immutable_memory------------------------------------- 905 // Access immutable memory 906 Node* Compile::immutable_memory() { 907 if (_immutable_memory != NULL) { 908 return _immutable_memory; 909 } 910 StartNode* s = start(); 911 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 912 Node *p = s->fast_out(i); 913 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 914 _immutable_memory = p; 915 return _immutable_memory; 916 } 917 } 918 ShouldNotReachHere(); 919 return NULL; 920 } 921 922 //----------------------set_cached_top_node------------------------------------ 923 // Install the cached top node, and make sure Node::is_top works correctly. 924 void Compile::set_cached_top_node(Node* tn) { 925 if (tn != NULL) verify_top(tn); 926 Node* old_top = _top; 927 _top = tn; 928 // Calling Node::setup_is_top allows the nodes the chance to adjust 929 // their _out arrays. 930 if (_top != NULL) _top->setup_is_top(); 931 if (old_top != NULL) old_top->setup_is_top(); 932 assert(_top == NULL || top()->is_top(), ""); 933 } 934 935 #ifndef PRODUCT 936 void Compile::verify_top(Node* tn) const { 937 if (tn != NULL) { 938 assert(tn->is_Con(), "top node must be a constant"); 939 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 940 assert(tn->in(0) != NULL, "must have live top node"); 941 } 942 } 943 #endif 944 945 946 ///-------------------Managing Per-Node Debug & Profile Info------------------- 947 948 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 949 guarantee(arr != NULL, ""); 950 int num_blocks = arr->length(); 951 if (grow_by < num_blocks) grow_by = num_blocks; 952 int num_notes = grow_by * _node_notes_block_size; 953 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 954 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 955 while (num_notes > 0) { 956 arr->append(notes); 957 notes += _node_notes_block_size; 958 num_notes -= _node_notes_block_size; 959 } 960 assert(num_notes == 0, "exact multiple, please"); 961 } 962 963 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 964 if (source == NULL || dest == NULL) return false; 965 966 if (dest->is_Con()) 967 return false; // Do not push debug info onto constants. 968 969 #ifdef ASSERT 970 // Leave a bread crumb trail pointing to the original node: 971 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 972 dest->set_debug_orig(source); 973 } 974 #endif 975 976 if (node_note_array() == NULL) 977 return false; // Not collecting any notes now. 978 979 // This is a copy onto a pre-existing node, which may already have notes. 980 // If both nodes have notes, do not overwrite any pre-existing notes. 981 Node_Notes* source_notes = node_notes_at(source->_idx); 982 if (source_notes == NULL || source_notes->is_clear()) return false; 983 Node_Notes* dest_notes = node_notes_at(dest->_idx); 984 if (dest_notes == NULL || dest_notes->is_clear()) { 985 return set_node_notes_at(dest->_idx, source_notes); 986 } 987 988 Node_Notes merged_notes = (*source_notes); 989 // The order of operations here ensures that dest notes will win... 990 merged_notes.update_from(dest_notes); 991 return set_node_notes_at(dest->_idx, &merged_notes); 992 } 993 994 995 //--------------------------allow_range_check_smearing------------------------- 996 // Gating condition for coalescing similar range checks. 997 // Sometimes we try 'speculatively' replacing a series of a range checks by a 998 // single covering check that is at least as strong as any of them. 999 // If the optimization succeeds, the simplified (strengthened) range check 1000 // will always succeed. If it fails, we will deopt, and then give up 1001 // on the optimization. 1002 bool Compile::allow_range_check_smearing() const { 1003 // If this method has already thrown a range-check, 1004 // assume it was because we already tried range smearing 1005 // and it failed. 1006 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1007 return !already_trapped; 1008 } 1009 1010 1011 //------------------------------flatten_alias_type----------------------------- 1012 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1013 int offset = tj->offset(); 1014 TypePtr::PTR ptr = tj->ptr(); 1015 1016 // Known instance (scalarizable allocation) alias only with itself. 1017 bool is_known_inst = tj->isa_oopptr() != NULL && 1018 tj->is_oopptr()->is_known_instance(); 1019 1020 // Process weird unsafe references. 1021 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1022 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1023 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1024 tj = TypeOopPtr::BOTTOM; 1025 ptr = tj->ptr(); 1026 offset = tj->offset(); 1027 } 1028 1029 // Array pointers need some flattening 1030 const TypeAryPtr *ta = tj->isa_aryptr(); 1031 if( ta && is_known_inst ) { 1032 if ( offset != Type::OffsetBot && 1033 offset > arrayOopDesc::length_offset_in_bytes() ) { 1034 offset = Type::OffsetBot; // Flatten constant access into array body only 1035 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1036 } 1037 } else if( ta && _AliasLevel >= 2 ) { 1038 // For arrays indexed by constant indices, we flatten the alias 1039 // space to include all of the array body. Only the header, klass 1040 // and array length can be accessed un-aliased. 1041 if( offset != Type::OffsetBot ) { 1042 if( ta->const_oop() ) { // methodDataOop or methodOop 1043 offset = Type::OffsetBot; // Flatten constant access into array body 1044 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1045 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1046 // range is OK as-is. 1047 tj = ta = TypeAryPtr::RANGE; 1048 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1049 tj = TypeInstPtr::KLASS; // all klass loads look alike 1050 ta = TypeAryPtr::RANGE; // generic ignored junk 1051 ptr = TypePtr::BotPTR; 1052 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1053 tj = TypeInstPtr::MARK; 1054 ta = TypeAryPtr::RANGE; // generic ignored junk 1055 ptr = TypePtr::BotPTR; 1056 } else { // Random constant offset into array body 1057 offset = Type::OffsetBot; // Flatten constant access into array body 1058 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1059 } 1060 } 1061 // Arrays of fixed size alias with arrays of unknown size. 1062 if (ta->size() != TypeInt::POS) { 1063 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1064 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1065 } 1066 // Arrays of known objects become arrays of unknown objects. 1067 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1068 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1069 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1070 } 1071 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1072 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1073 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1074 } 1075 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1076 // cannot be distinguished by bytecode alone. 1077 if (ta->elem() == TypeInt::BOOL) { 1078 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1079 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1080 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1081 } 1082 // During the 2nd round of IterGVN, NotNull castings are removed. 1083 // Make sure the Bottom and NotNull variants alias the same. 1084 // Also, make sure exact and non-exact variants alias the same. 1085 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1086 if (ta->const_oop()) { 1087 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1088 } else { 1089 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1090 } 1091 } 1092 } 1093 1094 // Oop pointers need some flattening 1095 const TypeInstPtr *to = tj->isa_instptr(); 1096 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1097 if( ptr == TypePtr::Constant ) { 1098 // No constant oop pointers (such as Strings); they alias with 1099 // unknown strings. 1100 assert(!is_known_inst, "not scalarizable allocation"); 1101 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1102 } else if( is_known_inst ) { 1103 tj = to; // Keep NotNull and klass_is_exact for instance type 1104 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1105 // During the 2nd round of IterGVN, NotNull castings are removed. 1106 // Make sure the Bottom and NotNull variants alias the same. 1107 // Also, make sure exact and non-exact variants alias the same. 1108 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1109 } 1110 // Canonicalize the holder of this field 1111 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1112 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1113 // First handle header references such as a LoadKlassNode, even if the 1114 // object's klass is unloaded at compile time (4965979). 1115 if (!is_known_inst) { // Do it only for non-instance types 1116 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1117 } 1118 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1119 to = NULL; 1120 tj = TypeOopPtr::BOTTOM; 1121 offset = tj->offset(); 1122 } else { 1123 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1124 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1125 if( is_known_inst ) { 1126 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1127 } else { 1128 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1129 } 1130 } 1131 } 1132 } 1133 1134 // Klass pointers to object array klasses need some flattening 1135 const TypeKlassPtr *tk = tj->isa_klassptr(); 1136 if( tk ) { 1137 // If we are referencing a field within a Klass, we need 1138 // to assume the worst case of an Object. Both exact and 1139 // inexact types must flatten to the same alias class. 1140 // Since the flattened result for a klass is defined to be 1141 // precisely java.lang.Object, use a constant ptr. 1142 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1143 1144 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1145 TypeKlassPtr::OBJECT->klass(), 1146 offset); 1147 } 1148 1149 ciKlass* klass = tk->klass(); 1150 if( klass->is_obj_array_klass() ) { 1151 ciKlass* k = TypeAryPtr::OOPS->klass(); 1152 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1153 k = TypeInstPtr::BOTTOM->klass(); 1154 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1155 } 1156 1157 // Check for precise loads from the primary supertype array and force them 1158 // to the supertype cache alias index. Check for generic array loads from 1159 // the primary supertype array and also force them to the supertype cache 1160 // alias index. Since the same load can reach both, we need to merge 1161 // these 2 disparate memories into the same alias class. Since the 1162 // primary supertype array is read-only, there's no chance of confusion 1163 // where we bypass an array load and an array store. 1164 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1165 if( offset == Type::OffsetBot || 1166 off2 < Klass::primary_super_limit()*wordSize ) { 1167 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1168 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1169 } 1170 } 1171 1172 // Flatten all Raw pointers together. 1173 if (tj->base() == Type::RawPtr) 1174 tj = TypeRawPtr::BOTTOM; 1175 1176 if (tj->base() == Type::AnyPtr) 1177 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1178 1179 // Flatten all to bottom for now 1180 switch( _AliasLevel ) { 1181 case 0: 1182 tj = TypePtr::BOTTOM; 1183 break; 1184 case 1: // Flatten to: oop, static, field or array 1185 switch (tj->base()) { 1186 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1187 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1188 case Type::AryPtr: // do not distinguish arrays at all 1189 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1190 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1191 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1192 default: ShouldNotReachHere(); 1193 } 1194 break; 1195 case 2: // No collapsing at level 2; keep all splits 1196 case 3: // No collapsing at level 3; keep all splits 1197 break; 1198 default: 1199 Unimplemented(); 1200 } 1201 1202 offset = tj->offset(); 1203 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1204 1205 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1206 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1207 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1208 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1209 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1210 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1211 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1212 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1213 assert( tj->ptr() != TypePtr::TopPTR && 1214 tj->ptr() != TypePtr::AnyNull && 1215 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1216 // assert( tj->ptr() != TypePtr::Constant || 1217 // tj->base() == Type::RawPtr || 1218 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1219 1220 return tj; 1221 } 1222 1223 void Compile::AliasType::Init(int i, const TypePtr* at) { 1224 _index = i; 1225 _adr_type = at; 1226 _field = NULL; 1227 _is_rewritable = true; // default 1228 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1229 if (atoop != NULL && atoop->is_known_instance()) { 1230 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1231 _general_index = Compile::current()->get_alias_index(gt); 1232 } else { 1233 _general_index = 0; 1234 } 1235 } 1236 1237 //---------------------------------print_on------------------------------------ 1238 #ifndef PRODUCT 1239 void Compile::AliasType::print_on(outputStream* st) { 1240 if (index() < 10) 1241 st->print("@ <%d> ", index()); 1242 else st->print("@ <%d>", index()); 1243 st->print(is_rewritable() ? " " : " RO"); 1244 int offset = adr_type()->offset(); 1245 if (offset == Type::OffsetBot) 1246 st->print(" +any"); 1247 else st->print(" +%-3d", offset); 1248 st->print(" in "); 1249 adr_type()->dump_on(st); 1250 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1251 if (field() != NULL && tjp) { 1252 if (tjp->klass() != field()->holder() || 1253 tjp->offset() != field()->offset_in_bytes()) { 1254 st->print(" != "); 1255 field()->print(); 1256 st->print(" ***"); 1257 } 1258 } 1259 } 1260 1261 void print_alias_types() { 1262 Compile* C = Compile::current(); 1263 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1264 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1265 C->alias_type(idx)->print_on(tty); 1266 tty->cr(); 1267 } 1268 } 1269 #endif 1270 1271 1272 //----------------------------probe_alias_cache-------------------------------- 1273 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1274 intptr_t key = (intptr_t) adr_type; 1275 key ^= key >> logAliasCacheSize; 1276 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1277 } 1278 1279 1280 //-----------------------------grow_alias_types-------------------------------- 1281 void Compile::grow_alias_types() { 1282 const int old_ats = _max_alias_types; // how many before? 1283 const int new_ats = old_ats; // how many more? 1284 const int grow_ats = old_ats+new_ats; // how many now? 1285 _max_alias_types = grow_ats; 1286 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1287 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1288 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1289 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1290 } 1291 1292 1293 //--------------------------------find_alias_type------------------------------ 1294 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1295 if (_AliasLevel == 0) 1296 return alias_type(AliasIdxBot); 1297 1298 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1299 if (ace->_adr_type == adr_type) { 1300 return alias_type(ace->_index); 1301 } 1302 1303 // Handle special cases. 1304 if (adr_type == NULL) return alias_type(AliasIdxTop); 1305 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1306 1307 // Do it the slow way. 1308 const TypePtr* flat = flatten_alias_type(adr_type); 1309 1310 #ifdef ASSERT 1311 assert(flat == flatten_alias_type(flat), "idempotent"); 1312 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1313 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1314 const TypeOopPtr* foop = flat->is_oopptr(); 1315 // Scalarizable allocations have exact klass always. 1316 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1317 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1318 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1319 } 1320 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1321 #endif 1322 1323 int idx = AliasIdxTop; 1324 for (int i = 0; i < num_alias_types(); i++) { 1325 if (alias_type(i)->adr_type() == flat) { 1326 idx = i; 1327 break; 1328 } 1329 } 1330 1331 if (idx == AliasIdxTop) { 1332 if (no_create) return NULL; 1333 // Grow the array if necessary. 1334 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1335 // Add a new alias type. 1336 idx = _num_alias_types++; 1337 _alias_types[idx]->Init(idx, flat); 1338 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1339 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1340 if (flat->isa_instptr()) { 1341 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1342 && flat->is_instptr()->klass() == env()->Class_klass()) 1343 alias_type(idx)->set_rewritable(false); 1344 } 1345 if (flat->isa_klassptr()) { 1346 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1347 alias_type(idx)->set_rewritable(false); 1348 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1349 alias_type(idx)->set_rewritable(false); 1350 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1351 alias_type(idx)->set_rewritable(false); 1352 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1353 alias_type(idx)->set_rewritable(false); 1354 } 1355 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1356 // but the base pointer type is not distinctive enough to identify 1357 // references into JavaThread.) 1358 1359 // Check for final instance fields. 1360 const TypeInstPtr* tinst = flat->isa_instptr(); 1361 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1362 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1363 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1364 // Set field() and is_rewritable() attributes. 1365 if (field != NULL) alias_type(idx)->set_field(field); 1366 } 1367 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1368 // Check for final static fields. 1369 if (tklass && tklass->klass()->is_instance_klass()) { 1370 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1371 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1372 // Set field() and is_rewritable() attributes. 1373 if (field != NULL) alias_type(idx)->set_field(field); 1374 } 1375 } 1376 1377 // Fill the cache for next time. 1378 ace->_adr_type = adr_type; 1379 ace->_index = idx; 1380 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1381 1382 // Might as well try to fill the cache for the flattened version, too. 1383 AliasCacheEntry* face = probe_alias_cache(flat); 1384 if (face->_adr_type == NULL) { 1385 face->_adr_type = flat; 1386 face->_index = idx; 1387 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1388 } 1389 1390 return alias_type(idx); 1391 } 1392 1393 1394 Compile::AliasType* Compile::alias_type(ciField* field) { 1395 const TypeOopPtr* t; 1396 if (field->is_static()) 1397 t = TypeKlassPtr::make(field->holder()); 1398 else 1399 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1400 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1401 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1402 return atp; 1403 } 1404 1405 1406 //------------------------------have_alias_type-------------------------------- 1407 bool Compile::have_alias_type(const TypePtr* adr_type) { 1408 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1409 if (ace->_adr_type == adr_type) { 1410 return true; 1411 } 1412 1413 // Handle special cases. 1414 if (adr_type == NULL) return true; 1415 if (adr_type == TypePtr::BOTTOM) return true; 1416 1417 return find_alias_type(adr_type, true) != NULL; 1418 } 1419 1420 //-----------------------------must_alias-------------------------------------- 1421 // True if all values of the given address type are in the given alias category. 1422 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1423 if (alias_idx == AliasIdxBot) return true; // the universal category 1424 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1425 if (alias_idx == AliasIdxTop) return false; // the empty category 1426 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1427 1428 // the only remaining possible overlap is identity 1429 int adr_idx = get_alias_index(adr_type); 1430 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1431 assert(adr_idx == alias_idx || 1432 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1433 && adr_type != TypeOopPtr::BOTTOM), 1434 "should not be testing for overlap with an unsafe pointer"); 1435 return adr_idx == alias_idx; 1436 } 1437 1438 //------------------------------can_alias-------------------------------------- 1439 // True if any values of the given address type are in the given alias category. 1440 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1441 if (alias_idx == AliasIdxTop) return false; // the empty category 1442 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1443 if (alias_idx == AliasIdxBot) return true; // the universal category 1444 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1445 1446 // the only remaining possible overlap is identity 1447 int adr_idx = get_alias_index(adr_type); 1448 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1449 return adr_idx == alias_idx; 1450 } 1451 1452 1453 1454 //---------------------------pop_warm_call------------------------------------- 1455 WarmCallInfo* Compile::pop_warm_call() { 1456 WarmCallInfo* wci = _warm_calls; 1457 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1458 return wci; 1459 } 1460 1461 //----------------------------Inline_Warm-------------------------------------- 1462 int Compile::Inline_Warm() { 1463 // If there is room, try to inline some more warm call sites. 1464 // %%% Do a graph index compaction pass when we think we're out of space? 1465 if (!InlineWarmCalls) return 0; 1466 1467 int calls_made_hot = 0; 1468 int room_to_grow = NodeCountInliningCutoff - unique(); 1469 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1470 int amount_grown = 0; 1471 WarmCallInfo* call; 1472 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1473 int est_size = (int)call->size(); 1474 if (est_size > (room_to_grow - amount_grown)) { 1475 // This one won't fit anyway. Get rid of it. 1476 call->make_cold(); 1477 continue; 1478 } 1479 call->make_hot(); 1480 calls_made_hot++; 1481 amount_grown += est_size; 1482 amount_to_grow -= est_size; 1483 } 1484 1485 if (calls_made_hot > 0) set_major_progress(); 1486 return calls_made_hot; 1487 } 1488 1489 1490 //----------------------------Finish_Warm-------------------------------------- 1491 void Compile::Finish_Warm() { 1492 if (!InlineWarmCalls) return; 1493 if (failing()) return; 1494 if (warm_calls() == NULL) return; 1495 1496 // Clean up loose ends, if we are out of space for inlining. 1497 WarmCallInfo* call; 1498 while ((call = pop_warm_call()) != NULL) { 1499 call->make_cold(); 1500 } 1501 } 1502 1503 1504 //------------------------------Optimize--------------------------------------- 1505 // Given a graph, optimize it. 1506 void Compile::Optimize() { 1507 TracePhase t1("optimizer", &_t_optimizer, true); 1508 1509 #ifndef PRODUCT 1510 if (env()->break_at_compile()) { 1511 BREAKPOINT; 1512 } 1513 1514 #endif 1515 1516 ResourceMark rm; 1517 int loop_opts_cnt; 1518 1519 NOT_PRODUCT( verify_graph_edges(); ) 1520 1521 print_method("After Parsing"); 1522 1523 { 1524 // Iterative Global Value Numbering, including ideal transforms 1525 // Initialize IterGVN with types and values from parse-time GVN 1526 PhaseIterGVN igvn(initial_gvn()); 1527 { 1528 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1529 igvn.optimize(); 1530 } 1531 1532 print_method("Iter GVN 1", 2); 1533 1534 if (failing()) return; 1535 1536 // Loop transforms on the ideal graph. Range Check Elimination, 1537 // peeling, unrolling, etc. 1538 1539 // Set loop opts counter 1540 loop_opts_cnt = num_loop_opts(); 1541 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1542 { 1543 TracePhase t2("idealLoop", &_t_idealLoop, true); 1544 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1545 loop_opts_cnt--; 1546 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1547 if (failing()) return; 1548 } 1549 // Loop opts pass if partial peeling occurred in previous pass 1550 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1551 TracePhase t3("idealLoop", &_t_idealLoop, true); 1552 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1553 loop_opts_cnt--; 1554 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1555 if (failing()) return; 1556 } 1557 // Loop opts pass for loop-unrolling before CCP 1558 if(major_progress() && (loop_opts_cnt > 0)) { 1559 TracePhase t4("idealLoop", &_t_idealLoop, true); 1560 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1561 loop_opts_cnt--; 1562 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1563 } 1564 } 1565 if (failing()) return; 1566 1567 // Conditional Constant Propagation; 1568 PhaseCCP ccp( &igvn ); 1569 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1570 { 1571 TracePhase t2("ccp", &_t_ccp, true); 1572 ccp.do_transform(); 1573 } 1574 print_method("PhaseCPP 1", 2); 1575 1576 assert( true, "Break here to ccp.dump_old2new_map()"); 1577 1578 // Iterative Global Value Numbering, including ideal transforms 1579 { 1580 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1581 igvn = ccp; 1582 igvn.optimize(); 1583 } 1584 1585 print_method("Iter GVN 2", 2); 1586 1587 if (failing()) return; 1588 1589 // Loop transforms on the ideal graph. Range Check Elimination, 1590 // peeling, unrolling, etc. 1591 if(loop_opts_cnt > 0) { 1592 debug_only( int cnt = 0; ); 1593 while(major_progress() && (loop_opts_cnt > 0)) { 1594 TracePhase t2("idealLoop", &_t_idealLoop, true); 1595 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1596 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1597 loop_opts_cnt--; 1598 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1599 if (failing()) return; 1600 } 1601 } 1602 { 1603 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1604 PhaseMacroExpand mex(igvn); 1605 if (mex.expand_macro_nodes()) { 1606 assert(failing(), "must bail out w/ explicit message"); 1607 return; 1608 } 1609 } 1610 1611 } // (End scope of igvn; run destructor if necessary for asserts.) 1612 1613 // A method with only infinite loops has no edges entering loops from root 1614 { 1615 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1616 if (final_graph_reshaping()) { 1617 assert(failing(), "must bail out w/ explicit message"); 1618 return; 1619 } 1620 } 1621 1622 print_method("Optimize finished", 2); 1623 } 1624 1625 1626 //------------------------------Code_Gen--------------------------------------- 1627 // Given a graph, generate code for it 1628 void Compile::Code_Gen() { 1629 if (failing()) return; 1630 1631 // Perform instruction selection. You might think we could reclaim Matcher 1632 // memory PDQ, but actually the Matcher is used in generating spill code. 1633 // Internals of the Matcher (including some VectorSets) must remain live 1634 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1635 // set a bit in reclaimed memory. 1636 1637 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1638 // nodes. Mapping is only valid at the root of each matched subtree. 1639 NOT_PRODUCT( verify_graph_edges(); ) 1640 1641 Node_List proj_list; 1642 Matcher m(proj_list); 1643 _matcher = &m; 1644 { 1645 TracePhase t2("matcher", &_t_matcher, true); 1646 m.match(); 1647 } 1648 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1649 // nodes. Mapping is only valid at the root of each matched subtree. 1650 NOT_PRODUCT( verify_graph_edges(); ) 1651 1652 // If you have too many nodes, or if matching has failed, bail out 1653 check_node_count(0, "out of nodes matching instructions"); 1654 if (failing()) return; 1655 1656 // Build a proper-looking CFG 1657 PhaseCFG cfg(node_arena(), root(), m); 1658 _cfg = &cfg; 1659 { 1660 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1661 cfg.Dominators(); 1662 if (failing()) return; 1663 1664 NOT_PRODUCT( verify_graph_edges(); ) 1665 1666 cfg.Estimate_Block_Frequency(); 1667 cfg.GlobalCodeMotion(m,unique(),proj_list); 1668 1669 print_method("Global code motion", 2); 1670 1671 if (failing()) return; 1672 NOT_PRODUCT( verify_graph_edges(); ) 1673 1674 debug_only( cfg.verify(); ) 1675 } 1676 NOT_PRODUCT( verify_graph_edges(); ) 1677 1678 PhaseChaitin regalloc(unique(),cfg,m); 1679 _regalloc = ®alloc; 1680 { 1681 TracePhase t2("regalloc", &_t_registerAllocation, true); 1682 // Perform any platform dependent preallocation actions. This is used, 1683 // for example, to avoid taking an implicit null pointer exception 1684 // using the frame pointer on win95. 1685 _regalloc->pd_preallocate_hook(); 1686 1687 // Perform register allocation. After Chaitin, use-def chains are 1688 // no longer accurate (at spill code) and so must be ignored. 1689 // Node->LRG->reg mappings are still accurate. 1690 _regalloc->Register_Allocate(); 1691 1692 // Bail out if the allocator builds too many nodes 1693 if (failing()) return; 1694 } 1695 1696 // Prior to register allocation we kept empty basic blocks in case the 1697 // the allocator needed a place to spill. After register allocation we 1698 // are not adding any new instructions. If any basic block is empty, we 1699 // can now safely remove it. 1700 { 1701 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1702 cfg.remove_empty(); 1703 if (do_freq_based_layout()) { 1704 PhaseBlockLayout layout(cfg); 1705 } else { 1706 cfg.set_loop_alignment(); 1707 } 1708 cfg.fixup_flow(); 1709 } 1710 1711 // Perform any platform dependent postallocation verifications. 1712 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1713 1714 // Apply peephole optimizations 1715 if( OptoPeephole ) { 1716 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1717 PhasePeephole peep( _regalloc, cfg); 1718 peep.do_transform(); 1719 } 1720 1721 // Convert Nodes to instruction bits in a buffer 1722 { 1723 // %%%% workspace merge brought two timers together for one job 1724 TracePhase t2a("output", &_t_output, true); 1725 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1726 Output(); 1727 } 1728 1729 print_method("Final Code"); 1730 1731 // He's dead, Jim. 1732 _cfg = (PhaseCFG*)0xdeadbeef; 1733 _regalloc = (PhaseChaitin*)0xdeadbeef; 1734 } 1735 1736 1737 //------------------------------dump_asm--------------------------------------- 1738 // Dump formatted assembly 1739 #ifndef PRODUCT 1740 void Compile::dump_asm(int *pcs, uint pc_limit) { 1741 bool cut_short = false; 1742 tty->print_cr("#"); 1743 tty->print("# "); _tf->dump(); tty->cr(); 1744 tty->print_cr("#"); 1745 1746 // For all blocks 1747 int pc = 0x0; // Program counter 1748 char starts_bundle = ' '; 1749 _regalloc->dump_frame(); 1750 1751 Node *n = NULL; 1752 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1753 if (VMThread::should_terminate()) { cut_short = true; break; } 1754 Block *b = _cfg->_blocks[i]; 1755 if (b->is_connector() && !Verbose) continue; 1756 n = b->_nodes[0]; 1757 if (pcs && n->_idx < pc_limit) 1758 tty->print("%3.3x ", pcs[n->_idx]); 1759 else 1760 tty->print(" "); 1761 b->dump_head( &_cfg->_bbs ); 1762 if (b->is_connector()) { 1763 tty->print_cr(" # Empty connector block"); 1764 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1765 tty->print_cr(" # Block is sole successor of call"); 1766 } 1767 1768 // For all instructions 1769 Node *delay = NULL; 1770 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1771 if (VMThread::should_terminate()) { cut_short = true; break; } 1772 n = b->_nodes[j]; 1773 if (valid_bundle_info(n)) { 1774 Bundle *bundle = node_bundling(n); 1775 if (bundle->used_in_unconditional_delay()) { 1776 delay = n; 1777 continue; 1778 } 1779 if (bundle->starts_bundle()) 1780 starts_bundle = '+'; 1781 } 1782 1783 if (WizardMode) n->dump(); 1784 1785 if( !n->is_Region() && // Dont print in the Assembly 1786 !n->is_Phi() && // a few noisely useless nodes 1787 !n->is_Proj() && 1788 !n->is_MachTemp() && 1789 !n->is_Catch() && // Would be nice to print exception table targets 1790 !n->is_MergeMem() && // Not very interesting 1791 !n->is_top() && // Debug info table constants 1792 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1793 ) { 1794 if (pcs && n->_idx < pc_limit) 1795 tty->print("%3.3x", pcs[n->_idx]); 1796 else 1797 tty->print(" "); 1798 tty->print(" %c ", starts_bundle); 1799 starts_bundle = ' '; 1800 tty->print("\t"); 1801 n->format(_regalloc, tty); 1802 tty->cr(); 1803 } 1804 1805 // If we have an instruction with a delay slot, and have seen a delay, 1806 // then back up and print it 1807 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1808 assert(delay != NULL, "no unconditional delay instruction"); 1809 if (WizardMode) delay->dump(); 1810 1811 if (node_bundling(delay)->starts_bundle()) 1812 starts_bundle = '+'; 1813 if (pcs && n->_idx < pc_limit) 1814 tty->print("%3.3x", pcs[n->_idx]); 1815 else 1816 tty->print(" "); 1817 tty->print(" %c ", starts_bundle); 1818 starts_bundle = ' '; 1819 tty->print("\t"); 1820 delay->format(_regalloc, tty); 1821 tty->print_cr(""); 1822 delay = NULL; 1823 } 1824 1825 // Dump the exception table as well 1826 if( n->is_Catch() && (Verbose || WizardMode) ) { 1827 // Print the exception table for this offset 1828 _handler_table.print_subtable_for(pc); 1829 } 1830 } 1831 1832 if (pcs && n->_idx < pc_limit) 1833 tty->print_cr("%3.3x", pcs[n->_idx]); 1834 else 1835 tty->print_cr(""); 1836 1837 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1838 1839 } // End of per-block dump 1840 tty->print_cr(""); 1841 1842 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1843 } 1844 #endif 1845 1846 //------------------------------Final_Reshape_Counts--------------------------- 1847 // This class defines counters to help identify when a method 1848 // may/must be executed using hardware with only 24-bit precision. 1849 struct Final_Reshape_Counts : public StackObj { 1850 int _call_count; // count non-inlined 'common' calls 1851 int _float_count; // count float ops requiring 24-bit precision 1852 int _double_count; // count double ops requiring more precision 1853 int _java_call_count; // count non-inlined 'java' calls 1854 VectorSet _visited; // Visitation flags 1855 Node_List _tests; // Set of IfNodes & PCTableNodes 1856 1857 Final_Reshape_Counts() : 1858 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), 1859 _visited( Thread::current()->resource_area() ) { } 1860 1861 void inc_call_count () { _call_count ++; } 1862 void inc_float_count () { _float_count ++; } 1863 void inc_double_count() { _double_count++; } 1864 void inc_java_call_count() { _java_call_count++; } 1865 1866 int get_call_count () const { return _call_count ; } 1867 int get_float_count () const { return _float_count ; } 1868 int get_double_count() const { return _double_count; } 1869 int get_java_call_count() const { return _java_call_count; } 1870 }; 1871 1872 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1873 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1874 // Make sure the offset goes inside the instance layout. 1875 return k->contains_field_offset(tp->offset()); 1876 // Note that OffsetBot and OffsetTop are very negative. 1877 } 1878 1879 //------------------------------final_graph_reshaping_impl---------------------- 1880 // Implement items 1-5 from final_graph_reshaping below. 1881 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { 1882 1883 if ( n->outcnt() == 0 ) return; // dead node 1884 uint nop = n->Opcode(); 1885 1886 // Check for 2-input instruction with "last use" on right input. 1887 // Swap to left input. Implements item (2). 1888 if( n->req() == 3 && // two-input instruction 1889 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1890 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1891 n->in(2)->outcnt() == 1 &&// right use IS a last use 1892 !n->in(2)->is_Con() ) { // right use is not a constant 1893 // Check for commutative opcode 1894 switch( nop ) { 1895 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1896 case Op_MaxI: case Op_MinI: 1897 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1898 case Op_AndL: case Op_XorL: case Op_OrL: 1899 case Op_AndI: case Op_XorI: case Op_OrI: { 1900 // Move "last use" input to left by swapping inputs 1901 n->swap_edges(1, 2); 1902 break; 1903 } 1904 default: 1905 break; 1906 } 1907 } 1908 1909 // Count FPU ops and common calls, implements item (3) 1910 switch( nop ) { 1911 // Count all float operations that may use FPU 1912 case Op_AddF: 1913 case Op_SubF: 1914 case Op_MulF: 1915 case Op_DivF: 1916 case Op_NegF: 1917 case Op_ModF: 1918 case Op_ConvI2F: 1919 case Op_ConF: 1920 case Op_CmpF: 1921 case Op_CmpF3: 1922 // case Op_ConvL2F: // longs are split into 32-bit halves 1923 fpu.inc_float_count(); 1924 break; 1925 1926 case Op_ConvF2D: 1927 case Op_ConvD2F: 1928 fpu.inc_float_count(); 1929 fpu.inc_double_count(); 1930 break; 1931 1932 // Count all double operations that may use FPU 1933 case Op_AddD: 1934 case Op_SubD: 1935 case Op_MulD: 1936 case Op_DivD: 1937 case Op_NegD: 1938 case Op_ModD: 1939 case Op_ConvI2D: 1940 case Op_ConvD2I: 1941 // case Op_ConvL2D: // handled by leaf call 1942 // case Op_ConvD2L: // handled by leaf call 1943 case Op_ConD: 1944 case Op_CmpD: 1945 case Op_CmpD3: 1946 fpu.inc_double_count(); 1947 break; 1948 case Op_Opaque1: // Remove Opaque Nodes before matching 1949 case Op_Opaque2: // Remove Opaque Nodes before matching 1950 n->subsume_by(n->in(1)); 1951 break; 1952 case Op_CallStaticJava: 1953 case Op_CallJava: 1954 case Op_CallDynamicJava: 1955 fpu.inc_java_call_count(); // Count java call site; 1956 case Op_CallRuntime: 1957 case Op_CallLeaf: 1958 case Op_CallLeafNoFP: { 1959 assert( n->is_Call(), "" ); 1960 CallNode *call = n->as_Call(); 1961 // Count call sites where the FP mode bit would have to be flipped. 1962 // Do not count uncommon runtime calls: 1963 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 1964 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 1965 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 1966 fpu.inc_call_count(); // Count the call site 1967 } else { // See if uncommon argument is shared 1968 Node *n = call->in(TypeFunc::Parms); 1969 int nop = n->Opcode(); 1970 // Clone shared simple arguments to uncommon calls, item (1). 1971 if( n->outcnt() > 1 && 1972 !n->is_Proj() && 1973 nop != Op_CreateEx && 1974 nop != Op_CheckCastPP && 1975 nop != Op_DecodeN && 1976 !n->is_Mem() ) { 1977 Node *x = n->clone(); 1978 call->set_req( TypeFunc::Parms, x ); 1979 } 1980 } 1981 break; 1982 } 1983 1984 case Op_StoreD: 1985 case Op_LoadD: 1986 case Op_LoadD_unaligned: 1987 fpu.inc_double_count(); 1988 goto handle_mem; 1989 case Op_StoreF: 1990 case Op_LoadF: 1991 fpu.inc_float_count(); 1992 goto handle_mem; 1993 1994 case Op_StoreB: 1995 case Op_StoreC: 1996 case Op_StoreCM: 1997 case Op_StorePConditional: 1998 case Op_StoreI: 1999 case Op_StoreL: 2000 case Op_StoreIConditional: 2001 case Op_StoreLConditional: 2002 case Op_CompareAndSwapI: 2003 case Op_CompareAndSwapL: 2004 case Op_CompareAndSwapP: 2005 case Op_CompareAndSwapN: 2006 case Op_StoreP: 2007 case Op_StoreN: 2008 case Op_LoadB: 2009 case Op_LoadUB: 2010 case Op_LoadUS: 2011 case Op_LoadI: 2012 case Op_LoadUI2L: 2013 case Op_LoadKlass: 2014 case Op_LoadNKlass: 2015 case Op_LoadL: 2016 case Op_LoadL_unaligned: 2017 case Op_LoadPLocked: 2018 case Op_LoadLLocked: 2019 case Op_LoadP: 2020 case Op_LoadN: 2021 case Op_LoadRange: 2022 case Op_LoadS: { 2023 handle_mem: 2024 #ifdef ASSERT 2025 if( VerifyOptoOopOffsets ) { 2026 assert( n->is_Mem(), "" ); 2027 MemNode *mem = (MemNode*)n; 2028 // Check to see if address types have grounded out somehow. 2029 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2030 assert( !tp || oop_offset_is_sane(tp), "" ); 2031 } 2032 #endif 2033 break; 2034 } 2035 2036 case Op_AddP: { // Assert sane base pointers 2037 Node *addp = n->in(AddPNode::Address); 2038 assert( !addp->is_AddP() || 2039 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2040 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2041 "Base pointers must match" ); 2042 #ifdef _LP64 2043 if (UseCompressedOops && 2044 addp->Opcode() == Op_ConP && 2045 addp == n->in(AddPNode::Base) && 2046 n->in(AddPNode::Offset)->is_Con()) { 2047 // Use addressing with narrow klass to load with offset on x86. 2048 // On sparc loading 32-bits constant and decoding it have less 2049 // instructions (4) then load 64-bits constant (7). 2050 // Do this transformation here since IGVN will convert ConN back to ConP. 2051 const Type* t = addp->bottom_type(); 2052 if (t->isa_oopptr()) { 2053 Node* nn = NULL; 2054 2055 // Look for existing ConN node of the same exact type. 2056 Compile* C = Compile::current(); 2057 Node* r = C->root(); 2058 uint cnt = r->outcnt(); 2059 for (uint i = 0; i < cnt; i++) { 2060 Node* m = r->raw_out(i); 2061 if (m!= NULL && m->Opcode() == Op_ConN && 2062 m->bottom_type()->make_ptr() == t) { 2063 nn = m; 2064 break; 2065 } 2066 } 2067 if (nn != NULL) { 2068 // Decode a narrow oop to match address 2069 // [R12 + narrow_oop_reg<<3 + offset] 2070 nn = new (C, 2) DecodeNNode(nn, t); 2071 n->set_req(AddPNode::Base, nn); 2072 n->set_req(AddPNode::Address, nn); 2073 if (addp->outcnt() == 0) { 2074 addp->disconnect_inputs(NULL); 2075 } 2076 } 2077 } 2078 } 2079 #endif 2080 break; 2081 } 2082 2083 #ifdef _LP64 2084 case Op_CastPP: 2085 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) { 2086 Compile* C = Compile::current(); 2087 Node* in1 = n->in(1); 2088 const Type* t = n->bottom_type(); 2089 Node* new_in1 = in1->clone(); 2090 new_in1->as_DecodeN()->set_type(t); 2091 2092 if (!Matcher::clone_shift_expressions) { 2093 // 2094 // x86, ARM and friends can handle 2 adds in addressing mode 2095 // and Matcher can fold a DecodeN node into address by using 2096 // a narrow oop directly and do implicit NULL check in address: 2097 // 2098 // [R12 + narrow_oop_reg<<3 + offset] 2099 // NullCheck narrow_oop_reg 2100 // 2101 // On other platforms (Sparc) we have to keep new DecodeN node and 2102 // use it to do implicit NULL check in address: 2103 // 2104 // decode_not_null narrow_oop_reg, base_reg 2105 // [base_reg + offset] 2106 // NullCheck base_reg 2107 // 2108 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2109 // to keep the information to which NULL check the new DecodeN node 2110 // corresponds to use it as value in implicit_null_check(). 2111 // 2112 new_in1->set_req(0, n->in(0)); 2113 } 2114 2115 n->subsume_by(new_in1); 2116 if (in1->outcnt() == 0) { 2117 in1->disconnect_inputs(NULL); 2118 } 2119 } 2120 break; 2121 2122 case Op_CmpP: 2123 // Do this transformation here to preserve CmpPNode::sub() and 2124 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2125 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2126 Node* in1 = n->in(1); 2127 Node* in2 = n->in(2); 2128 if (!in1->is_DecodeN()) { 2129 in2 = in1; 2130 in1 = n->in(2); 2131 } 2132 assert(in1->is_DecodeN(), "sanity"); 2133 2134 Compile* C = Compile::current(); 2135 Node* new_in2 = NULL; 2136 if (in2->is_DecodeN()) { 2137 new_in2 = in2->in(1); 2138 } else if (in2->Opcode() == Op_ConP) { 2139 const Type* t = in2->bottom_type(); 2140 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) { 2141 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2142 // 2143 // This transformation together with CastPP transformation above 2144 // will generated code for implicit NULL checks for compressed oops. 2145 // 2146 // The original code after Optimize() 2147 // 2148 // LoadN memory, narrow_oop_reg 2149 // decode narrow_oop_reg, base_reg 2150 // CmpP base_reg, NULL 2151 // CastPP base_reg // NotNull 2152 // Load [base_reg + offset], val_reg 2153 // 2154 // after these transformations will be 2155 // 2156 // LoadN memory, narrow_oop_reg 2157 // CmpN narrow_oop_reg, NULL 2158 // decode_not_null narrow_oop_reg, base_reg 2159 // Load [base_reg + offset], val_reg 2160 // 2161 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2162 // since narrow oops can be used in debug info now (see the code in 2163 // final_graph_reshaping_walk()). 2164 // 2165 // At the end the code will be matched to 2166 // on x86: 2167 // 2168 // Load_narrow_oop memory, narrow_oop_reg 2169 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2170 // NullCheck narrow_oop_reg 2171 // 2172 // and on sparc: 2173 // 2174 // Load_narrow_oop memory, narrow_oop_reg 2175 // decode_not_null narrow_oop_reg, base_reg 2176 // Load [base_reg + offset], val_reg 2177 // NullCheck base_reg 2178 // 2179 } else if (t->isa_oopptr()) { 2180 new_in2 = ConNode::make(C, t->make_narrowoop()); 2181 } 2182 } 2183 if (new_in2 != NULL) { 2184 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2185 n->subsume_by( cmpN ); 2186 if (in1->outcnt() == 0) { 2187 in1->disconnect_inputs(NULL); 2188 } 2189 if (in2->outcnt() == 0) { 2190 in2->disconnect_inputs(NULL); 2191 } 2192 } 2193 } 2194 break; 2195 2196 case Op_DecodeN: 2197 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2198 // DecodeN could be pinned on Sparc where it can't be fold into 2199 // an address expression, see the code for Op_CastPP above. 2200 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc"); 2201 break; 2202 2203 case Op_EncodeP: { 2204 Node* in1 = n->in(1); 2205 if (in1->is_DecodeN()) { 2206 n->subsume_by(in1->in(1)); 2207 } else if (in1->Opcode() == Op_ConP) { 2208 Compile* C = Compile::current(); 2209 const Type* t = in1->bottom_type(); 2210 if (t == TypePtr::NULL_PTR) { 2211 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2212 } else if (t->isa_oopptr()) { 2213 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2214 } 2215 } 2216 if (in1->outcnt() == 0) { 2217 in1->disconnect_inputs(NULL); 2218 } 2219 break; 2220 } 2221 2222 case Op_Phi: 2223 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2224 // The EncodeP optimization may create Phi with the same edges 2225 // for all paths. It is not handled well by Register Allocator. 2226 Node* unique_in = n->in(1); 2227 assert(unique_in != NULL, ""); 2228 uint cnt = n->req(); 2229 for (uint i = 2; i < cnt; i++) { 2230 Node* m = n->in(i); 2231 assert(m != NULL, ""); 2232 if (unique_in != m) 2233 unique_in = NULL; 2234 } 2235 if (unique_in != NULL) { 2236 n->subsume_by(unique_in); 2237 } 2238 } 2239 break; 2240 2241 #endif 2242 2243 case Op_ModI: 2244 if (UseDivMod) { 2245 // Check if a%b and a/b both exist 2246 Node* d = n->find_similar(Op_DivI); 2247 if (d) { 2248 // Replace them with a fused divmod if supported 2249 Compile* C = Compile::current(); 2250 if (Matcher::has_match_rule(Op_DivModI)) { 2251 DivModINode* divmod = DivModINode::make(C, n); 2252 d->subsume_by(divmod->div_proj()); 2253 n->subsume_by(divmod->mod_proj()); 2254 } else { 2255 // replace a%b with a-((a/b)*b) 2256 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2257 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2258 n->subsume_by( sub ); 2259 } 2260 } 2261 } 2262 break; 2263 2264 case Op_ModL: 2265 if (UseDivMod) { 2266 // Check if a%b and a/b both exist 2267 Node* d = n->find_similar(Op_DivL); 2268 if (d) { 2269 // Replace them with a fused divmod if supported 2270 Compile* C = Compile::current(); 2271 if (Matcher::has_match_rule(Op_DivModL)) { 2272 DivModLNode* divmod = DivModLNode::make(C, n); 2273 d->subsume_by(divmod->div_proj()); 2274 n->subsume_by(divmod->mod_proj()); 2275 } else { 2276 // replace a%b with a-((a/b)*b) 2277 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2278 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2279 n->subsume_by( sub ); 2280 } 2281 } 2282 } 2283 break; 2284 2285 case Op_Load16B: 2286 case Op_Load8B: 2287 case Op_Load4B: 2288 case Op_Load8S: 2289 case Op_Load4S: 2290 case Op_Load2S: 2291 case Op_Load8C: 2292 case Op_Load4C: 2293 case Op_Load2C: 2294 case Op_Load4I: 2295 case Op_Load2I: 2296 case Op_Load2L: 2297 case Op_Load4F: 2298 case Op_Load2F: 2299 case Op_Load2D: 2300 case Op_Store16B: 2301 case Op_Store8B: 2302 case Op_Store4B: 2303 case Op_Store8C: 2304 case Op_Store4C: 2305 case Op_Store2C: 2306 case Op_Store4I: 2307 case Op_Store2I: 2308 case Op_Store2L: 2309 case Op_Store4F: 2310 case Op_Store2F: 2311 case Op_Store2D: 2312 break; 2313 2314 case Op_PackB: 2315 case Op_PackS: 2316 case Op_PackC: 2317 case Op_PackI: 2318 case Op_PackF: 2319 case Op_PackL: 2320 case Op_PackD: 2321 if (n->req()-1 > 2) { 2322 // Replace many operand PackNodes with a binary tree for matching 2323 PackNode* p = (PackNode*) n; 2324 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2325 n->subsume_by(btp); 2326 } 2327 break; 2328 default: 2329 assert( !n->is_Call(), "" ); 2330 assert( !n->is_Mem(), "" ); 2331 break; 2332 } 2333 2334 // Collect CFG split points 2335 if (n->is_MultiBranch()) 2336 fpu._tests.push(n); 2337 } 2338 2339 //------------------------------final_graph_reshaping_walk--------------------- 2340 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2341 // requires that the walk visits a node's inputs before visiting the node. 2342 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { 2343 ResourceArea *area = Thread::current()->resource_area(); 2344 Unique_Node_List sfpt(area); 2345 2346 fpu._visited.set(root->_idx); // first, mark node as visited 2347 uint cnt = root->req(); 2348 Node *n = root; 2349 uint i = 0; 2350 while (true) { 2351 if (i < cnt) { 2352 // Place all non-visited non-null inputs onto stack 2353 Node* m = n->in(i); 2354 ++i; 2355 if (m != NULL && !fpu._visited.test_set(m->_idx)) { 2356 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2357 sfpt.push(m); 2358 cnt = m->req(); 2359 nstack.push(n, i); // put on stack parent and next input's index 2360 n = m; 2361 i = 0; 2362 } 2363 } else { 2364 // Now do post-visit work 2365 final_graph_reshaping_impl( n, fpu ); 2366 if (nstack.is_empty()) 2367 break; // finished 2368 n = nstack.node(); // Get node from stack 2369 cnt = n->req(); 2370 i = nstack.index(); 2371 nstack.pop(); // Shift to the next node on stack 2372 } 2373 } 2374 2375 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2376 // It could be done for an uncommon traps or any safepoints/calls 2377 // if the DecodeN node is referenced only in a debug info. 2378 while (sfpt.size() > 0) { 2379 n = sfpt.pop(); 2380 JVMState *jvms = n->as_SafePoint()->jvms(); 2381 assert(jvms != NULL, "sanity"); 2382 int start = jvms->debug_start(); 2383 int end = n->req(); 2384 bool is_uncommon = (n->is_CallStaticJava() && 2385 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2386 for (int j = start; j < end; j++) { 2387 Node* in = n->in(j); 2388 if (in->is_DecodeN()) { 2389 bool safe_to_skip = true; 2390 if (!is_uncommon ) { 2391 // Is it safe to skip? 2392 for (uint i = 0; i < in->outcnt(); i++) { 2393 Node* u = in->raw_out(i); 2394 if (!u->is_SafePoint() || 2395 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2396 safe_to_skip = false; 2397 } 2398 } 2399 } 2400 if (safe_to_skip) { 2401 n->set_req(j, in->in(1)); 2402 } 2403 if (in->outcnt() == 0) { 2404 in->disconnect_inputs(NULL); 2405 } 2406 } 2407 } 2408 } 2409 } 2410 2411 //------------------------------final_graph_reshaping-------------------------- 2412 // Final Graph Reshaping. 2413 // 2414 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2415 // and not commoned up and forced early. Must come after regular 2416 // optimizations to avoid GVN undoing the cloning. Clone constant 2417 // inputs to Loop Phis; these will be split by the allocator anyways. 2418 // Remove Opaque nodes. 2419 // (2) Move last-uses by commutative operations to the left input to encourage 2420 // Intel update-in-place two-address operations and better register usage 2421 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2422 // calls canonicalizing them back. 2423 // (3) Count the number of double-precision FP ops, single-precision FP ops 2424 // and call sites. On Intel, we can get correct rounding either by 2425 // forcing singles to memory (requires extra stores and loads after each 2426 // FP bytecode) or we can set a rounding mode bit (requires setting and 2427 // clearing the mode bit around call sites). The mode bit is only used 2428 // if the relative frequency of single FP ops to calls is low enough. 2429 // This is a key transform for SPEC mpeg_audio. 2430 // (4) Detect infinite loops; blobs of code reachable from above but not 2431 // below. Several of the Code_Gen algorithms fail on such code shapes, 2432 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2433 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2434 // Detection is by looking for IfNodes where only 1 projection is 2435 // reachable from below or CatchNodes missing some targets. 2436 // (5) Assert for insane oop offsets in debug mode. 2437 2438 bool Compile::final_graph_reshaping() { 2439 // an infinite loop may have been eliminated by the optimizer, 2440 // in which case the graph will be empty. 2441 if (root()->req() == 1) { 2442 record_method_not_compilable("trivial infinite loop"); 2443 return true; 2444 } 2445 2446 Final_Reshape_Counts fpu; 2447 2448 // Visit everybody reachable! 2449 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2450 Node_Stack nstack(unique() >> 1); 2451 final_graph_reshaping_walk(nstack, root(), fpu); 2452 2453 // Check for unreachable (from below) code (i.e., infinite loops). 2454 for( uint i = 0; i < fpu._tests.size(); i++ ) { 2455 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch(); 2456 // Get number of CFG targets. 2457 // Note that PCTables include exception targets after calls. 2458 uint required_outcnt = n->required_outcnt(); 2459 if (n->outcnt() != required_outcnt) { 2460 // Check for a few special cases. Rethrow Nodes never take the 2461 // 'fall-thru' path, so expected kids is 1 less. 2462 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2463 if (n->in(0)->in(0)->is_Call()) { 2464 CallNode *call = n->in(0)->in(0)->as_Call(); 2465 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2466 required_outcnt--; // Rethrow always has 1 less kid 2467 } else if (call->req() > TypeFunc::Parms && 2468 call->is_CallDynamicJava()) { 2469 // Check for null receiver. In such case, the optimizer has 2470 // detected that the virtual call will always result in a null 2471 // pointer exception. The fall-through projection of this CatchNode 2472 // will not be populated. 2473 Node *arg0 = call->in(TypeFunc::Parms); 2474 if (arg0->is_Type() && 2475 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2476 required_outcnt--; 2477 } 2478 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2479 call->req() > TypeFunc::Parms+1 && 2480 call->is_CallStaticJava()) { 2481 // Check for negative array length. In such case, the optimizer has 2482 // detected that the allocation attempt will always result in an 2483 // exception. There is no fall-through projection of this CatchNode . 2484 Node *arg1 = call->in(TypeFunc::Parms+1); 2485 if (arg1->is_Type() && 2486 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2487 required_outcnt--; 2488 } 2489 } 2490 } 2491 } 2492 // Recheck with a better notion of 'required_outcnt' 2493 if (n->outcnt() != required_outcnt) { 2494 record_method_not_compilable("malformed control flow"); 2495 return true; // Not all targets reachable! 2496 } 2497 } 2498 // Check that I actually visited all kids. Unreached kids 2499 // must be infinite loops. 2500 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2501 if (!fpu._visited.test(n->fast_out(j)->_idx)) { 2502 record_method_not_compilable("infinite loop"); 2503 return true; // Found unvisited kid; must be unreach 2504 } 2505 } 2506 2507 // If original bytecodes contained a mixture of floats and doubles 2508 // check if the optimizer has made it homogenous, item (3). 2509 if( Use24BitFPMode && Use24BitFP && 2510 fpu.get_float_count() > 32 && 2511 fpu.get_double_count() == 0 && 2512 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { 2513 set_24_bit_selection_and_mode( false, true ); 2514 } 2515 2516 set_has_java_calls(fpu.get_java_call_count() > 0); 2517 2518 // No infinite loops, no reason to bail out. 2519 return false; 2520 } 2521 2522 //-----------------------------too_many_traps---------------------------------- 2523 // Report if there are too many traps at the current method and bci. 2524 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2525 bool Compile::too_many_traps(ciMethod* method, 2526 int bci, 2527 Deoptimization::DeoptReason reason) { 2528 ciMethodData* md = method->method_data(); 2529 if (md->is_empty()) { 2530 // Assume the trap has not occurred, or that it occurred only 2531 // because of a transient condition during start-up in the interpreter. 2532 return false; 2533 } 2534 if (md->has_trap_at(bci, reason) != 0) { 2535 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2536 // Also, if there are multiple reasons, or if there is no per-BCI record, 2537 // assume the worst. 2538 if (log()) 2539 log()->elem("observe trap='%s' count='%d'", 2540 Deoptimization::trap_reason_name(reason), 2541 md->trap_count(reason)); 2542 return true; 2543 } else { 2544 // Ignore method/bci and see if there have been too many globally. 2545 return too_many_traps(reason, md); 2546 } 2547 } 2548 2549 // Less-accurate variant which does not require a method and bci. 2550 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2551 ciMethodData* logmd) { 2552 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2553 // Too many traps globally. 2554 // Note that we use cumulative trap_count, not just md->trap_count. 2555 if (log()) { 2556 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2557 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2558 Deoptimization::trap_reason_name(reason), 2559 mcount, trap_count(reason)); 2560 } 2561 return true; 2562 } else { 2563 // The coast is clear. 2564 return false; 2565 } 2566 } 2567 2568 //--------------------------too_many_recompiles-------------------------------- 2569 // Report if there are too many recompiles at the current method and bci. 2570 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2571 // Is not eager to return true, since this will cause the compiler to use 2572 // Action_none for a trap point, to avoid too many recompilations. 2573 bool Compile::too_many_recompiles(ciMethod* method, 2574 int bci, 2575 Deoptimization::DeoptReason reason) { 2576 ciMethodData* md = method->method_data(); 2577 if (md->is_empty()) { 2578 // Assume the trap has not occurred, or that it occurred only 2579 // because of a transient condition during start-up in the interpreter. 2580 return false; 2581 } 2582 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2583 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2584 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2585 Deoptimization::DeoptReason per_bc_reason 2586 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2587 if ((per_bc_reason == Deoptimization::Reason_none 2588 || md->has_trap_at(bci, reason) != 0) 2589 // The trap frequency measure we care about is the recompile count: 2590 && md->trap_recompiled_at(bci) 2591 && md->overflow_recompile_count() >= bc_cutoff) { 2592 // Do not emit a trap here if it has already caused recompilations. 2593 // Also, if there are multiple reasons, or if there is no per-BCI record, 2594 // assume the worst. 2595 if (log()) 2596 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2597 Deoptimization::trap_reason_name(reason), 2598 md->trap_count(reason), 2599 md->overflow_recompile_count()); 2600 return true; 2601 } else if (trap_count(reason) != 0 2602 && decompile_count() >= m_cutoff) { 2603 // Too many recompiles globally, and we have seen this sort of trap. 2604 // Use cumulative decompile_count, not just md->decompile_count. 2605 if (log()) 2606 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2607 Deoptimization::trap_reason_name(reason), 2608 md->trap_count(reason), trap_count(reason), 2609 md->decompile_count(), decompile_count()); 2610 return true; 2611 } else { 2612 // The coast is clear. 2613 return false; 2614 } 2615 } 2616 2617 2618 #ifndef PRODUCT 2619 //------------------------------verify_graph_edges--------------------------- 2620 // Walk the Graph and verify that there is a one-to-one correspondence 2621 // between Use-Def edges and Def-Use edges in the graph. 2622 void Compile::verify_graph_edges(bool no_dead_code) { 2623 if (VerifyGraphEdges) { 2624 ResourceArea *area = Thread::current()->resource_area(); 2625 Unique_Node_List visited(area); 2626 // Call recursive graph walk to check edges 2627 _root->verify_edges(visited); 2628 if (no_dead_code) { 2629 // Now make sure that no visited node is used by an unvisited node. 2630 bool dead_nodes = 0; 2631 Unique_Node_List checked(area); 2632 while (visited.size() > 0) { 2633 Node* n = visited.pop(); 2634 checked.push(n); 2635 for (uint i = 0; i < n->outcnt(); i++) { 2636 Node* use = n->raw_out(i); 2637 if (checked.member(use)) continue; // already checked 2638 if (visited.member(use)) continue; // already in the graph 2639 if (use->is_Con()) continue; // a dead ConNode is OK 2640 // At this point, we have found a dead node which is DU-reachable. 2641 if (dead_nodes++ == 0) 2642 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2643 use->dump(2); 2644 tty->print_cr("---"); 2645 checked.push(use); // No repeats; pretend it is now checked. 2646 } 2647 } 2648 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2649 } 2650 } 2651 } 2652 #endif 2653 2654 // The Compile object keeps track of failure reasons separately from the ciEnv. 2655 // This is required because there is not quite a 1-1 relation between the 2656 // ciEnv and its compilation task and the Compile object. Note that one 2657 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2658 // to backtrack and retry without subsuming loads. Other than this backtracking 2659 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2660 // by the logic in C2Compiler. 2661 void Compile::record_failure(const char* reason) { 2662 if (log() != NULL) { 2663 log()->elem("failure reason='%s' phase='compile'", reason); 2664 } 2665 if (_failure_reason == NULL) { 2666 // Record the first failure reason. 2667 _failure_reason = reason; 2668 } 2669 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2670 C->print_method(_failure_reason); 2671 } 2672 _root = NULL; // flush the graph, too 2673 } 2674 2675 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2676 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2677 { 2678 if (dolog) { 2679 C = Compile::current(); 2680 _log = C->log(); 2681 } else { 2682 C = NULL; 2683 _log = NULL; 2684 } 2685 if (_log != NULL) { 2686 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2687 _log->stamp(); 2688 _log->end_head(); 2689 } 2690 } 2691 2692 Compile::TracePhase::~TracePhase() { 2693 if (_log != NULL) { 2694 _log->done("phase nodes='%d'", C->unique()); 2695 } 2696 }