compile.cpp revision 196:d1605aabd0a1
1/* 2 * Copyright 1997-2008 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. 32int 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 67void 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 93CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 94 assert(m->is_loaded(), "don't try this on unloaded methods"); 95 if (_intrinsics != NULL) { 96 int index = intrinsic_insertion_index(m, is_virtual); 97 if (index < _intrinsics->length() 98 && _intrinsics->at(index)->method() == m 99 && _intrinsics->at(index)->is_virtual() == is_virtual) { 100 return _intrinsics->at(index); 101 } 102 } 103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 104 if (m->intrinsic_id() != vmIntrinsics::_none) { 105 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 106 if (cg != NULL) { 107 // Save it for next time: 108 register_intrinsic(cg); 109 return cg; 110 } else { 111 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 112 } 113 } 114 return NULL; 115} 116 117// Compile:: register_library_intrinsics and make_vm_intrinsic are defined 118// in library_call.cpp. 119 120 121#ifndef PRODUCT 122// statistics gathering... 123 124juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 125jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 126 127bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 128 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 129 int oflags = _intrinsic_hist_flags[id]; 130 assert(flags != 0, "what happened?"); 131 if (is_virtual) { 132 flags |= _intrinsic_virtual; 133 } 134 bool changed = (flags != oflags); 135 if ((flags & _intrinsic_worked) != 0) { 136 juint count = (_intrinsic_hist_count[id] += 1); 137 if (count == 1) { 138 changed = true; // first time 139 } 140 // increment the overall count also: 141 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 142 } 143 if (changed) { 144 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 145 // Something changed about the intrinsic's virtuality. 146 if ((flags & _intrinsic_virtual) != 0) { 147 // This is the first use of this intrinsic as a virtual call. 148 if (oflags != 0) { 149 // We already saw it as a non-virtual, so note both cases. 150 flags |= _intrinsic_both; 151 } 152 } else if ((oflags & _intrinsic_both) == 0) { 153 // This is the first use of this intrinsic as a non-virtual 154 flags |= _intrinsic_both; 155 } 156 } 157 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 158 } 159 // update the overall flags also: 160 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 161 return changed; 162} 163 164static char* format_flags(int flags, char* buf) { 165 buf[0] = 0; 166 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 167 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 168 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 169 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 170 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 171 if (buf[0] == 0) strcat(buf, ","); 172 assert(buf[0] == ',', "must be"); 173 return &buf[1]; 174} 175 176void Compile::print_intrinsic_statistics() { 177 char flagsbuf[100]; 178 ttyLocker ttyl; 179 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 180 tty->print_cr("Compiler intrinsic usage:"); 181 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 182 if (total == 0) total = 1; // avoid div0 in case of no successes 183 #define PRINT_STAT_LINE(name, c, f) \ 184 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 185 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 186 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 187 int flags = _intrinsic_hist_flags[id]; 188 juint count = _intrinsic_hist_count[id]; 189 if ((flags | count) != 0) { 190 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 191 } 192 } 193 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 194 if (xtty != NULL) xtty->tail("statistics"); 195} 196 197void Compile::print_statistics() { 198 { ttyLocker ttyl; 199 if (xtty != NULL) xtty->head("statistics type='opto'"); 200 Parse::print_statistics(); 201 PhaseCCP::print_statistics(); 202 PhaseRegAlloc::print_statistics(); 203 Scheduling::print_statistics(); 204 PhasePeephole::print_statistics(); 205 PhaseIdealLoop::print_statistics(); 206 if (xtty != NULL) xtty->tail("statistics"); 207 } 208 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 209 // put this under its own <statistics> element. 210 print_intrinsic_statistics(); 211 } 212} 213#endif //PRODUCT 214 215// Support for bundling info 216Bundle* Compile::node_bundling(const Node *n) { 217 assert(valid_bundle_info(n), "oob"); 218 return &_node_bundling_base[n->_idx]; 219} 220 221bool Compile::valid_bundle_info(const Node *n) { 222 return (_node_bundling_limit > n->_idx); 223} 224 225 226// Identify all nodes that are reachable from below, useful. 227// Use breadth-first pass that records state in a Unique_Node_List, 228// recursive traversal is slower. 229void Compile::identify_useful_nodes(Unique_Node_List &useful) { 230 int estimated_worklist_size = unique(); 231 useful.map( estimated_worklist_size, NULL ); // preallocate space 232 233 // Initialize worklist 234 if (root() != NULL) { useful.push(root()); } 235 // If 'top' is cached, declare it useful to preserve cached node 236 if( cached_top_node() ) { useful.push(cached_top_node()); } 237 238 // Push all useful nodes onto the list, breadthfirst 239 for( uint next = 0; next < useful.size(); ++next ) { 240 assert( next < unique(), "Unique useful nodes < total nodes"); 241 Node *n = useful.at(next); 242 uint max = n->len(); 243 for( uint i = 0; i < max; ++i ) { 244 Node *m = n->in(i); 245 if( m == NULL ) continue; 246 useful.push(m); 247 } 248 } 249} 250 251// Disconnect all useless nodes by disconnecting those at the boundary. 252void Compile::remove_useless_nodes(Unique_Node_List &useful) { 253 uint next = 0; 254 while( next < useful.size() ) { 255 Node *n = useful.at(next++); 256 // Use raw traversal of out edges since this code removes out edges 257 int max = n->outcnt(); 258 for (int j = 0; j < max; ++j ) { 259 Node* child = n->raw_out(j); 260 if( ! useful.member(child) ) { 261 assert( !child->is_top() || child != top(), 262 "If top is cached in Compile object it is in useful list"); 263 // Only need to remove this out-edge to the useless node 264 n->raw_del_out(j); 265 --j; 266 --max; 267 } 268 } 269 if (n->outcnt() == 1 && n->has_special_unique_user()) { 270 record_for_igvn( n->unique_out() ); 271 } 272 } 273 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 274} 275 276//------------------------------frame_size_in_words----------------------------- 277// frame_slots in units of words 278int Compile::frame_size_in_words() const { 279 // shift is 0 in LP32 and 1 in LP64 280 const int shift = (LogBytesPerWord - LogBytesPerInt); 281 int words = _frame_slots >> shift; 282 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 283 return words; 284} 285 286// ============================================================================ 287//------------------------------CompileWrapper--------------------------------- 288class CompileWrapper : public StackObj { 289 Compile *const _compile; 290 public: 291 CompileWrapper(Compile* compile); 292 293 ~CompileWrapper(); 294}; 295 296CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 297 // the Compile* pointer is stored in the current ciEnv: 298 ciEnv* env = compile->env(); 299 assert(env == ciEnv::current(), "must already be a ciEnv active"); 300 assert(env->compiler_data() == NULL, "compile already active?"); 301 env->set_compiler_data(compile); 302 assert(compile == Compile::current(), "sanity"); 303 304 compile->set_type_dict(NULL); 305 compile->set_type_hwm(NULL); 306 compile->set_type_last_size(0); 307 compile->set_last_tf(NULL, NULL); 308 compile->set_indexSet_arena(NULL); 309 compile->set_indexSet_free_block_list(NULL); 310 compile->init_type_arena(); 311 Type::Initialize(compile); 312 _compile->set_scratch_buffer_blob(NULL); 313 _compile->begin_method(); 314} 315CompileWrapper::~CompileWrapper() { 316 if (_compile->failing()) { 317 _compile->print_method("Failed"); 318 } 319 _compile->end_method(); 320 if (_compile->scratch_buffer_blob() != NULL) 321 BufferBlob::free(_compile->scratch_buffer_blob()); 322 _compile->env()->set_compiler_data(NULL); 323} 324 325 326//----------------------------print_compile_messages--------------------------- 327void Compile::print_compile_messages() { 328#ifndef PRODUCT 329 // Check if recompiling 330 if (_subsume_loads == false && PrintOpto) { 331 // Recompiling without allowing machine instructions to subsume loads 332 tty->print_cr("*********************************************************"); 333 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 334 tty->print_cr("*********************************************************"); 335 } 336 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 337 // Recompiling without escape analysis 338 tty->print_cr("*********************************************************"); 339 tty->print_cr("** Bailout: Recompile without escape analysis **"); 340 tty->print_cr("*********************************************************"); 341 } 342 if (env()->break_at_compile()) { 343 // Open the debugger when compiing this method. 344 tty->print("### Breaking when compiling: "); 345 method()->print_short_name(); 346 tty->cr(); 347 BREAKPOINT; 348 } 349 350 if( PrintOpto ) { 351 if (is_osr_compilation()) { 352 tty->print("[OSR]%3d", _compile_id); 353 } else { 354 tty->print("%3d", _compile_id); 355 } 356 } 357#endif 358} 359 360 361void Compile::init_scratch_buffer_blob() { 362 if( scratch_buffer_blob() != NULL ) return; 363 364 // Construct a temporary CodeBuffer to have it construct a BufferBlob 365 // Cache this BufferBlob for this compile. 366 ResourceMark rm; 367 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 368 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 369 // Record the buffer blob for next time. 370 set_scratch_buffer_blob(blob); 371 // Have we run out of code space? 372 if (scratch_buffer_blob() == NULL) { 373 // Let CompilerBroker disable further compilations. 374 record_failure("Not enough space for scratch buffer in CodeCache"); 375 return; 376 } 377 378 // Initialize the relocation buffers 379 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 380 set_scratch_locs_memory(locs_buf); 381} 382 383 384//-----------------------scratch_emit_size------------------------------------- 385// Helper function that computes size by emitting code 386uint Compile::scratch_emit_size(const Node* n) { 387 // Emit into a trash buffer and count bytes emitted. 388 // This is a pretty expensive way to compute a size, 389 // but it works well enough if seldom used. 390 // All common fixed-size instructions are given a size 391 // method by the AD file. 392 // Note that the scratch buffer blob and locs memory are 393 // allocated at the beginning of the compile task, and 394 // may be shared by several calls to scratch_emit_size. 395 // The allocation of the scratch buffer blob is particularly 396 // expensive, since it has to grab the code cache lock. 397 BufferBlob* blob = this->scratch_buffer_blob(); 398 assert(blob != NULL, "Initialize BufferBlob at start"); 399 assert(blob->size() > MAX_inst_size, "sanity"); 400 relocInfo* locs_buf = scratch_locs_memory(); 401 address blob_begin = blob->instructions_begin(); 402 address blob_end = (address)locs_buf; 403 assert(blob->instructions_contains(blob_end), "sanity"); 404 CodeBuffer buf(blob_begin, blob_end - blob_begin); 405 buf.initialize_consts_size(MAX_const_size); 406 buf.initialize_stubs_size(MAX_stubs_size); 407 assert(locs_buf != NULL, "sanity"); 408 int lsize = MAX_locs_size / 2; 409 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 410 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 411 n->emit(buf, this->regalloc()); 412 return buf.code_size(); 413} 414 415 416// ============================================================================ 417//------------------------------Compile standard------------------------------- 418debug_only( int Compile::_debug_idx = 100000; ) 419 420// Compile a method. entry_bci is -1 for normal compilations and indicates 421// the continuation bci for on stack replacement. 422 423 424Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 425 : Phase(Compiler), 426 _env(ci_env), 427 _log(ci_env->log()), 428 _compile_id(ci_env->compile_id()), 429 _save_argument_registers(false), 430 _stub_name(NULL), 431 _stub_function(NULL), 432 _stub_entry_point(NULL), 433 _method(target), 434 _entry_bci(osr_bci), 435 _initial_gvn(NULL), 436 _for_igvn(NULL), 437 _warm_calls(NULL), 438 _subsume_loads(subsume_loads), 439 _do_escape_analysis(do_escape_analysis), 440 _failure_reason(NULL), 441 _code_buffer("Compile::Fill_buffer"), 442 _orig_pc_slot(0), 443 _orig_pc_slot_offset_in_bytes(0), 444 _node_bundling_limit(0), 445 _node_bundling_base(NULL), 446#ifndef PRODUCT 447 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 448 _printer(IdealGraphPrinter::printer()), 449#endif 450 _congraph(NULL) { 451 C = this; 452 453 CompileWrapper cw(this); 454#ifndef PRODUCT 455 if (TimeCompiler2) { 456 tty->print(" "); 457 target->holder()->name()->print(); 458 tty->print("."); 459 target->print_short_name(); 460 tty->print(" "); 461 } 462 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 463 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 464 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 465 if (!print_opto_assembly) { 466 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 467 if (print_assembly && !Disassembler::can_decode()) { 468 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 469 print_opto_assembly = true; 470 } 471 } 472 set_print_assembly(print_opto_assembly); 473#endif 474 475 if (ProfileTraps) { 476 // Make sure the method being compiled gets its own MDO, 477 // so we can at least track the decompile_count(). 478 method()->build_method_data(); 479 } 480 481 Init(::AliasLevel); 482 483 484 print_compile_messages(); 485 486 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 487 _ilt = InlineTree::build_inline_tree_root(); 488 else 489 _ilt = NULL; 490 491 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 492 assert(num_alias_types() >= AliasIdxRaw, ""); 493 494#define MINIMUM_NODE_HASH 1023 495 // Node list that Iterative GVN will start with 496 Unique_Node_List for_igvn(comp_arena()); 497 set_for_igvn(&for_igvn); 498 499 // GVN that will be run immediately on new nodes 500 uint estimated_size = method()->code_size()*4+64; 501 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 502 PhaseGVN gvn(node_arena(), estimated_size); 503 set_initial_gvn(&gvn); 504 505 { // Scope for timing the parser 506 TracePhase t3("parse", &_t_parser, true); 507 508 // Put top into the hash table ASAP. 509 initial_gvn()->transform_no_reclaim(top()); 510 511 // Set up tf(), start(), and find a CallGenerator. 512 CallGenerator* cg; 513 if (is_osr_compilation()) { 514 const TypeTuple *domain = StartOSRNode::osr_domain(); 515 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 516 init_tf(TypeFunc::make(domain, range)); 517 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 518 initial_gvn()->set_type_bottom(s); 519 init_start(s); 520 cg = CallGenerator::for_osr(method(), entry_bci()); 521 } else { 522 // Normal case. 523 init_tf(TypeFunc::make(method())); 524 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 525 initial_gvn()->set_type_bottom(s); 526 init_start(s); 527 float past_uses = method()->interpreter_invocation_count(); 528 float expected_uses = past_uses; 529 cg = CallGenerator::for_inline(method(), expected_uses); 530 } 531 if (failing()) return; 532 if (cg == NULL) { 533 record_method_not_compilable_all_tiers("cannot parse method"); 534 return; 535 } 536 JVMState* jvms = build_start_state(start(), tf()); 537 if ((jvms = cg->generate(jvms)) == NULL) { 538 record_method_not_compilable("method parse failed"); 539 return; 540 } 541 GraphKit kit(jvms); 542 543 if (!kit.stopped()) { 544 // Accept return values, and transfer control we know not where. 545 // This is done by a special, unique ReturnNode bound to root. 546 return_values(kit.jvms()); 547 } 548 549 if (kit.has_exceptions()) { 550 // Any exceptions that escape from this call must be rethrown 551 // to whatever caller is dynamically above us on the stack. 552 // This is done by a special, unique RethrowNode bound to root. 553 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 554 } 555 556 // Remove clutter produced by parsing. 557 if (!failing()) { 558 ResourceMark rm; 559 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 560 } 561 } 562 563 // Note: Large methods are capped off in do_one_bytecode(). 564 if (failing()) return; 565 566 // After parsing, node notes are no longer automagic. 567 // They must be propagated by register_new_node_with_optimizer(), 568 // clone(), or the like. 569 set_default_node_notes(NULL); 570 571 for (;;) { 572 int successes = Inline_Warm(); 573 if (failing()) return; 574 if (successes == 0) break; 575 } 576 577 // Drain the list. 578 Finish_Warm(); 579#ifndef PRODUCT 580 if (_printer) { 581 _printer->print_inlining(this); 582 } 583#endif 584 585 if (failing()) return; 586 NOT_PRODUCT( verify_graph_edges(); ) 587 588 // Perform escape analysis 589 if (_do_escape_analysis) 590 _congraph = new ConnectionGraph(this); 591 if (_congraph != NULL) { 592 NOT_PRODUCT( TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, TimeCompiler); ) 593 _congraph->compute_escape(); 594 if (failing()) return; 595 596#ifndef PRODUCT 597 if (PrintEscapeAnalysis) { 598 _congraph->dump(); 599 } 600#endif 601 } 602 // Now optimize 603 Optimize(); 604 if (failing()) return; 605 NOT_PRODUCT( verify_graph_edges(); ) 606 607#ifndef PRODUCT 608 if (PrintIdeal) { 609 ttyLocker ttyl; // keep the following output all in one block 610 // This output goes directly to the tty, not the compiler log. 611 // To enable tools to match it up with the compilation activity, 612 // be sure to tag this tty output with the compile ID. 613 if (xtty != NULL) { 614 xtty->head("ideal compile_id='%d'%s", compile_id(), 615 is_osr_compilation() ? " compile_kind='osr'" : 616 ""); 617 } 618 root()->dump(9999); 619 if (xtty != NULL) { 620 xtty->tail("ideal"); 621 } 622 } 623#endif 624 625 // Now that we know the size of all the monitors we can add a fixed slot 626 // for the original deopt pc. 627 628 _orig_pc_slot = fixed_slots(); 629 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 630 set_fixed_slots(next_slot); 631 632 // Now generate code 633 Code_Gen(); 634 if (failing()) return; 635 636 // Check if we want to skip execution of all compiled code. 637 { 638#ifndef PRODUCT 639 if (OptoNoExecute) { 640 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 641 return; 642 } 643 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 644#endif 645 646 if (is_osr_compilation()) { 647 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 648 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 649 } else { 650 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 651 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 652 } 653 654 env()->register_method(_method, _entry_bci, 655 &_code_offsets, 656 _orig_pc_slot_offset_in_bytes, 657 code_buffer(), 658 frame_size_in_words(), _oop_map_set, 659 &_handler_table, &_inc_table, 660 compiler, 661 env()->comp_level(), 662 true, /*has_debug_info*/ 663 has_unsafe_access() 664 ); 665 } 666} 667 668//------------------------------Compile---------------------------------------- 669// Compile a runtime stub 670Compile::Compile( ciEnv* ci_env, 671 TypeFunc_generator generator, 672 address stub_function, 673 const char *stub_name, 674 int is_fancy_jump, 675 bool pass_tls, 676 bool save_arg_registers, 677 bool return_pc ) 678 : Phase(Compiler), 679 _env(ci_env), 680 _log(ci_env->log()), 681 _compile_id(-1), 682 _save_argument_registers(save_arg_registers), 683 _method(NULL), 684 _stub_name(stub_name), 685 _stub_function(stub_function), 686 _stub_entry_point(NULL), 687 _entry_bci(InvocationEntryBci), 688 _initial_gvn(NULL), 689 _for_igvn(NULL), 690 _warm_calls(NULL), 691 _orig_pc_slot(0), 692 _orig_pc_slot_offset_in_bytes(0), 693 _subsume_loads(true), 694 _do_escape_analysis(false), 695 _failure_reason(NULL), 696 _code_buffer("Compile::Fill_buffer"), 697 _node_bundling_limit(0), 698 _node_bundling_base(NULL), 699#ifndef PRODUCT 700 _trace_opto_output(TraceOptoOutput), 701 _printer(NULL), 702#endif 703 _congraph(NULL) { 704 C = this; 705 706#ifndef PRODUCT 707 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 708 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 709 set_print_assembly(PrintFrameConverterAssembly); 710#endif 711 CompileWrapper cw(this); 712 Init(/*AliasLevel=*/ 0); 713 init_tf((*generator)()); 714 715 { 716 // The following is a dummy for the sake of GraphKit::gen_stub 717 Unique_Node_List for_igvn(comp_arena()); 718 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 719 PhaseGVN gvn(Thread::current()->resource_area(),255); 720 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 721 gvn.transform_no_reclaim(top()); 722 723 GraphKit kit; 724 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 725 } 726 727 NOT_PRODUCT( verify_graph_edges(); ) 728 Code_Gen(); 729 if (failing()) return; 730 731 732 // Entry point will be accessed using compile->stub_entry_point(); 733 if (code_buffer() == NULL) { 734 Matcher::soft_match_failure(); 735 } else { 736 if (PrintAssembly && (WizardMode || Verbose)) 737 tty->print_cr("### Stub::%s", stub_name); 738 739 if (!failing()) { 740 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 741 742 // Make the NMethod 743 // For now we mark the frame as never safe for profile stackwalking 744 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 745 code_buffer(), 746 CodeOffsets::frame_never_safe, 747 // _code_offsets.value(CodeOffsets::Frame_Complete), 748 frame_size_in_words(), 749 _oop_map_set, 750 save_arg_registers); 751 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 752 753 _stub_entry_point = rs->entry_point(); 754 } 755 } 756} 757 758#ifndef PRODUCT 759void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 760 if(PrintOpto && Verbose) { 761 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 762 } 763} 764#endif 765 766void Compile::print_codes() { 767} 768 769//------------------------------Init------------------------------------------- 770// Prepare for a single compilation 771void Compile::Init(int aliaslevel) { 772 _unique = 0; 773 _regalloc = NULL; 774 775 _tf = NULL; // filled in later 776 _top = NULL; // cached later 777 _matcher = NULL; // filled in later 778 _cfg = NULL; // filled in later 779 780 set_24_bit_selection_and_mode(Use24BitFP, false); 781 782 _node_note_array = NULL; 783 _default_node_notes = NULL; 784 785 _immutable_memory = NULL; // filled in at first inquiry 786 787 // Globally visible Nodes 788 // First set TOP to NULL to give safe behavior during creation of RootNode 789 set_cached_top_node(NULL); 790 set_root(new (this, 3) RootNode()); 791 // Now that you have a Root to point to, create the real TOP 792 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 793 set_recent_alloc(NULL, NULL); 794 795 // Create Debug Information Recorder to record scopes, oopmaps, etc. 796 env()->set_oop_recorder(new OopRecorder(comp_arena())); 797 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 798 env()->set_dependencies(new Dependencies(env())); 799 800 _fixed_slots = 0; 801 set_has_split_ifs(false); 802 set_has_loops(has_method() && method()->has_loops()); // first approximation 803 _deopt_happens = true; // start out assuming the worst 804 _trap_can_recompile = false; // no traps emitted yet 805 _major_progress = true; // start out assuming good things will happen 806 set_has_unsafe_access(false); 807 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 808 set_decompile_count(0); 809 810 // Compilation level related initialization 811 if (env()->comp_level() == CompLevel_fast_compile) { 812 set_num_loop_opts(Tier1LoopOptsCount); 813 set_do_inlining(Tier1Inline != 0); 814 set_max_inline_size(Tier1MaxInlineSize); 815 set_freq_inline_size(Tier1FreqInlineSize); 816 set_do_scheduling(false); 817 set_do_count_invocations(Tier1CountInvocations); 818 set_do_method_data_update(Tier1UpdateMethodData); 819 } else { 820 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 821 set_num_loop_opts(LoopOptsCount); 822 set_do_inlining(Inline); 823 set_max_inline_size(MaxInlineSize); 824 set_freq_inline_size(FreqInlineSize); 825 set_do_scheduling(OptoScheduling); 826 set_do_count_invocations(false); 827 set_do_method_data_update(false); 828 } 829 830 if (debug_info()->recording_non_safepoints()) { 831 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 832 (comp_arena(), 8, 0, NULL)); 833 set_default_node_notes(Node_Notes::make(this)); 834 } 835 836 // // -- Initialize types before each compile -- 837 // // Update cached type information 838 // if( _method && _method->constants() ) 839 // Type::update_loaded_types(_method, _method->constants()); 840 841 // Init alias_type map. 842 if (!_do_escape_analysis && aliaslevel == 3) 843 aliaslevel = 2; // No unique types without escape analysis 844 _AliasLevel = aliaslevel; 845 const int grow_ats = 16; 846 _max_alias_types = grow_ats; 847 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 848 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 849 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 850 { 851 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 852 } 853 // Initialize the first few types. 854 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 855 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 856 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 857 _num_alias_types = AliasIdxRaw+1; 858 // Zero out the alias type cache. 859 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 860 // A NULL adr_type hits in the cache right away. Preload the right answer. 861 probe_alias_cache(NULL)->_index = AliasIdxTop; 862 863 _intrinsics = NULL; 864 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 865 register_library_intrinsics(); 866} 867 868//---------------------------init_start---------------------------------------- 869// Install the StartNode on this compile object. 870void Compile::init_start(StartNode* s) { 871 if (failing()) 872 return; // already failing 873 assert(s == start(), ""); 874} 875 876StartNode* Compile::start() const { 877 assert(!failing(), ""); 878 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 879 Node* start = root()->fast_out(i); 880 if( start->is_Start() ) 881 return start->as_Start(); 882 } 883 ShouldNotReachHere(); 884 return NULL; 885} 886 887//-------------------------------immutable_memory------------------------------------- 888// Access immutable memory 889Node* Compile::immutable_memory() { 890 if (_immutable_memory != NULL) { 891 return _immutable_memory; 892 } 893 StartNode* s = start(); 894 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 895 Node *p = s->fast_out(i); 896 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 897 _immutable_memory = p; 898 return _immutable_memory; 899 } 900 } 901 ShouldNotReachHere(); 902 return NULL; 903} 904 905//----------------------set_cached_top_node------------------------------------ 906// Install the cached top node, and make sure Node::is_top works correctly. 907void Compile::set_cached_top_node(Node* tn) { 908 if (tn != NULL) verify_top(tn); 909 Node* old_top = _top; 910 _top = tn; 911 // Calling Node::setup_is_top allows the nodes the chance to adjust 912 // their _out arrays. 913 if (_top != NULL) _top->setup_is_top(); 914 if (old_top != NULL) old_top->setup_is_top(); 915 assert(_top == NULL || top()->is_top(), ""); 916} 917 918#ifndef PRODUCT 919void Compile::verify_top(Node* tn) const { 920 if (tn != NULL) { 921 assert(tn->is_Con(), "top node must be a constant"); 922 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 923 assert(tn->in(0) != NULL, "must have live top node"); 924 } 925} 926#endif 927 928 929///-------------------Managing Per-Node Debug & Profile Info------------------- 930 931void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 932 guarantee(arr != NULL, ""); 933 int num_blocks = arr->length(); 934 if (grow_by < num_blocks) grow_by = num_blocks; 935 int num_notes = grow_by * _node_notes_block_size; 936 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 937 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 938 while (num_notes > 0) { 939 arr->append(notes); 940 notes += _node_notes_block_size; 941 num_notes -= _node_notes_block_size; 942 } 943 assert(num_notes == 0, "exact multiple, please"); 944} 945 946bool Compile::copy_node_notes_to(Node* dest, Node* source) { 947 if (source == NULL || dest == NULL) return false; 948 949 if (dest->is_Con()) 950 return false; // Do not push debug info onto constants. 951 952#ifdef ASSERT 953 // Leave a bread crumb trail pointing to the original node: 954 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 955 dest->set_debug_orig(source); 956 } 957#endif 958 959 if (node_note_array() == NULL) 960 return false; // Not collecting any notes now. 961 962 // This is a copy onto a pre-existing node, which may already have notes. 963 // If both nodes have notes, do not overwrite any pre-existing notes. 964 Node_Notes* source_notes = node_notes_at(source->_idx); 965 if (source_notes == NULL || source_notes->is_clear()) return false; 966 Node_Notes* dest_notes = node_notes_at(dest->_idx); 967 if (dest_notes == NULL || dest_notes->is_clear()) { 968 return set_node_notes_at(dest->_idx, source_notes); 969 } 970 971 Node_Notes merged_notes = (*source_notes); 972 // The order of operations here ensures that dest notes will win... 973 merged_notes.update_from(dest_notes); 974 return set_node_notes_at(dest->_idx, &merged_notes); 975} 976 977 978//--------------------------allow_range_check_smearing------------------------- 979// Gating condition for coalescing similar range checks. 980// Sometimes we try 'speculatively' replacing a series of a range checks by a 981// single covering check that is at least as strong as any of them. 982// If the optimization succeeds, the simplified (strengthened) range check 983// will always succeed. If it fails, we will deopt, and then give up 984// on the optimization. 985bool Compile::allow_range_check_smearing() const { 986 // If this method has already thrown a range-check, 987 // assume it was because we already tried range smearing 988 // and it failed. 989 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 990 return !already_trapped; 991} 992 993 994//------------------------------flatten_alias_type----------------------------- 995const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 996 int offset = tj->offset(); 997 TypePtr::PTR ptr = tj->ptr(); 998 999 // Process weird unsafe references. 1000 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1001 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1002 tj = TypeOopPtr::BOTTOM; 1003 ptr = tj->ptr(); 1004 offset = tj->offset(); 1005 } 1006 1007 // Array pointers need some flattening 1008 const TypeAryPtr *ta = tj->isa_aryptr(); 1009 if( ta && _AliasLevel >= 2 ) { 1010 // For arrays indexed by constant indices, we flatten the alias 1011 // space to include all of the array body. Only the header, klass 1012 // and array length can be accessed un-aliased. 1013 if( offset != Type::OffsetBot ) { 1014 if( ta->const_oop() ) { // methodDataOop or methodOop 1015 offset = Type::OffsetBot; // Flatten constant access into array body 1016 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); 1017 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1018 // range is OK as-is. 1019 tj = ta = TypeAryPtr::RANGE; 1020 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1021 tj = TypeInstPtr::KLASS; // all klass loads look alike 1022 ta = TypeAryPtr::RANGE; // generic ignored junk 1023 ptr = TypePtr::BotPTR; 1024 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1025 tj = TypeInstPtr::MARK; 1026 ta = TypeAryPtr::RANGE; // generic ignored junk 1027 ptr = TypePtr::BotPTR; 1028 } else { // Random constant offset into array body 1029 offset = Type::OffsetBot; // Flatten constant access into array body 1030 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); 1031 } 1032 } 1033 // Arrays of fixed size alias with arrays of unknown size. 1034 if (ta->size() != TypeInt::POS) { 1035 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1036 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset, ta->instance_id()); 1037 } 1038 // Arrays of known objects become arrays of unknown objects. 1039 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1040 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1041 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id()); 1042 } 1043 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1044 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1045 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id()); 1046 } 1047 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1048 // cannot be distinguished by bytecode alone. 1049 if (ta->elem() == TypeInt::BOOL) { 1050 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1051 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1052 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset, ta->instance_id()); 1053 } 1054 // During the 2nd round of IterGVN, NotNull castings are removed. 1055 // Make sure the Bottom and NotNull variants alias the same. 1056 // Also, make sure exact and non-exact variants alias the same. 1057 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1058 if (ta->const_oop()) { 1059 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1060 } else { 1061 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1062 } 1063 } 1064 } 1065 1066 // Oop pointers need some flattening 1067 const TypeInstPtr *to = tj->isa_instptr(); 1068 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1069 if( ptr == TypePtr::Constant ) { 1070 // No constant oop pointers (such as Strings); they alias with 1071 // unknown strings. 1072 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1073 } else if( to->is_instance_field() ) { 1074 tj = to; // Keep NotNull and klass_is_exact for instance type 1075 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1076 // During the 2nd round of IterGVN, NotNull castings are removed. 1077 // Make sure the Bottom and NotNull variants alias the same. 1078 // Also, make sure exact and non-exact variants alias the same. 1079 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset, to->instance_id()); 1080 } 1081 // Canonicalize the holder of this field 1082 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1083 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1084 // First handle header references such as a LoadKlassNode, even if the 1085 // object's klass is unloaded at compile time (4965979). 1086 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset, to->instance_id()); 1087 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1088 to = NULL; 1089 tj = TypeOopPtr::BOTTOM; 1090 offset = tj->offset(); 1091 } else { 1092 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1093 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1094 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset, to->instance_id()); 1095 } 1096 } 1097 } 1098 1099 // Klass pointers to object array klasses need some flattening 1100 const TypeKlassPtr *tk = tj->isa_klassptr(); 1101 if( tk ) { 1102 // If we are referencing a field within a Klass, we need 1103 // to assume the worst case of an Object. Both exact and 1104 // inexact types must flatten to the same alias class. 1105 // Since the flattened result for a klass is defined to be 1106 // precisely java.lang.Object, use a constant ptr. 1107 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1108 1109 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1110 TypeKlassPtr::OBJECT->klass(), 1111 offset); 1112 } 1113 1114 ciKlass* klass = tk->klass(); 1115 if( klass->is_obj_array_klass() ) { 1116 ciKlass* k = TypeAryPtr::OOPS->klass(); 1117 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1118 k = TypeInstPtr::BOTTOM->klass(); 1119 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1120 } 1121 1122 // Check for precise loads from the primary supertype array and force them 1123 // to the supertype cache alias index. Check for generic array loads from 1124 // the primary supertype array and also force them to the supertype cache 1125 // alias index. Since the same load can reach both, we need to merge 1126 // these 2 disparate memories into the same alias class. Since the 1127 // primary supertype array is read-only, there's no chance of confusion 1128 // where we bypass an array load and an array store. 1129 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1130 if( offset == Type::OffsetBot || 1131 off2 < Klass::primary_super_limit()*wordSize ) { 1132 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1133 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1134 } 1135 } 1136 1137 // Flatten all Raw pointers together. 1138 if (tj->base() == Type::RawPtr) 1139 tj = TypeRawPtr::BOTTOM; 1140 1141 if (tj->base() == Type::AnyPtr) 1142 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1143 1144 // Flatten all to bottom for now 1145 switch( _AliasLevel ) { 1146 case 0: 1147 tj = TypePtr::BOTTOM; 1148 break; 1149 case 1: // Flatten to: oop, static, field or array 1150 switch (tj->base()) { 1151 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1152 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1153 case Type::AryPtr: // do not distinguish arrays at all 1154 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1155 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1156 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1157 default: ShouldNotReachHere(); 1158 } 1159 break; 1160 case 2: // No collasping at level 2; keep all splits 1161 case 3: // No collasping at level 3; keep all splits 1162 break; 1163 default: 1164 Unimplemented(); 1165 } 1166 1167 offset = tj->offset(); 1168 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1169 1170 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1171 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1172 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1173 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1174 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1175 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1176 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1177 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1178 assert( tj->ptr() != TypePtr::TopPTR && 1179 tj->ptr() != TypePtr::AnyNull && 1180 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1181// assert( tj->ptr() != TypePtr::Constant || 1182// tj->base() == Type::RawPtr || 1183// tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1184 1185 return tj; 1186} 1187 1188void Compile::AliasType::Init(int i, const TypePtr* at) { 1189 _index = i; 1190 _adr_type = at; 1191 _field = NULL; 1192 _is_rewritable = true; // default 1193 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1194 if (atoop != NULL && atoop->is_instance()) { 1195 const TypeOopPtr *gt = atoop->cast_to_instance(TypeOopPtr::UNKNOWN_INSTANCE); 1196 _general_index = Compile::current()->get_alias_index(gt); 1197 } else { 1198 _general_index = 0; 1199 } 1200} 1201 1202//---------------------------------print_on------------------------------------ 1203#ifndef PRODUCT 1204void Compile::AliasType::print_on(outputStream* st) { 1205 if (index() < 10) 1206 st->print("@ <%d> ", index()); 1207 else st->print("@ <%d>", index()); 1208 st->print(is_rewritable() ? " " : " RO"); 1209 int offset = adr_type()->offset(); 1210 if (offset == Type::OffsetBot) 1211 st->print(" +any"); 1212 else st->print(" +%-3d", offset); 1213 st->print(" in "); 1214 adr_type()->dump_on(st); 1215 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1216 if (field() != NULL && tjp) { 1217 if (tjp->klass() != field()->holder() || 1218 tjp->offset() != field()->offset_in_bytes()) { 1219 st->print(" != "); 1220 field()->print(); 1221 st->print(" ***"); 1222 } 1223 } 1224} 1225 1226void print_alias_types() { 1227 Compile* C = Compile::current(); 1228 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1229 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1230 C->alias_type(idx)->print_on(tty); 1231 tty->cr(); 1232 } 1233} 1234#endif 1235 1236 1237//----------------------------probe_alias_cache-------------------------------- 1238Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1239 intptr_t key = (intptr_t) adr_type; 1240 key ^= key >> logAliasCacheSize; 1241 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1242} 1243 1244 1245//-----------------------------grow_alias_types-------------------------------- 1246void Compile::grow_alias_types() { 1247 const int old_ats = _max_alias_types; // how many before? 1248 const int new_ats = old_ats; // how many more? 1249 const int grow_ats = old_ats+new_ats; // how many now? 1250 _max_alias_types = grow_ats; 1251 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1252 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1253 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1254 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1255} 1256 1257 1258//--------------------------------find_alias_type------------------------------ 1259Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1260 if (_AliasLevel == 0) 1261 return alias_type(AliasIdxBot); 1262 1263 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1264 if (ace->_adr_type == adr_type) { 1265 return alias_type(ace->_index); 1266 } 1267 1268 // Handle special cases. 1269 if (adr_type == NULL) return alias_type(AliasIdxTop); 1270 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1271 1272 // Do it the slow way. 1273 const TypePtr* flat = flatten_alias_type(adr_type); 1274 1275#ifdef ASSERT 1276 assert(flat == flatten_alias_type(flat), "idempotent"); 1277 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1278 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1279 const TypeOopPtr* foop = flat->is_oopptr(); 1280 const TypePtr* xoop = foop->cast_to_exactness(!foop->klass_is_exact())->is_ptr(); 1281 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1282 } 1283 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1284#endif 1285 1286 int idx = AliasIdxTop; 1287 for (int i = 0; i < num_alias_types(); i++) { 1288 if (alias_type(i)->adr_type() == flat) { 1289 idx = i; 1290 break; 1291 } 1292 } 1293 1294 if (idx == AliasIdxTop) { 1295 if (no_create) return NULL; 1296 // Grow the array if necessary. 1297 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1298 // Add a new alias type. 1299 idx = _num_alias_types++; 1300 _alias_types[idx]->Init(idx, flat); 1301 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1302 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1303 if (flat->isa_instptr()) { 1304 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1305 && flat->is_instptr()->klass() == env()->Class_klass()) 1306 alias_type(idx)->set_rewritable(false); 1307 } 1308 if (flat->isa_klassptr()) { 1309 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1310 alias_type(idx)->set_rewritable(false); 1311 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1312 alias_type(idx)->set_rewritable(false); 1313 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1314 alias_type(idx)->set_rewritable(false); 1315 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1316 alias_type(idx)->set_rewritable(false); 1317 } 1318 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1319 // but the base pointer type is not distinctive enough to identify 1320 // references into JavaThread.) 1321 1322 // Check for final instance fields. 1323 const TypeInstPtr* tinst = flat->isa_instptr(); 1324 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1325 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1326 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1327 // Set field() and is_rewritable() attributes. 1328 if (field != NULL) alias_type(idx)->set_field(field); 1329 } 1330 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1331 // Check for final static fields. 1332 if (tklass && tklass->klass()->is_instance_klass()) { 1333 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1334 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1335 // Set field() and is_rewritable() attributes. 1336 if (field != NULL) alias_type(idx)->set_field(field); 1337 } 1338 } 1339 1340 // Fill the cache for next time. 1341 ace->_adr_type = adr_type; 1342 ace->_index = idx; 1343 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1344 1345 // Might as well try to fill the cache for the flattened version, too. 1346 AliasCacheEntry* face = probe_alias_cache(flat); 1347 if (face->_adr_type == NULL) { 1348 face->_adr_type = flat; 1349 face->_index = idx; 1350 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1351 } 1352 1353 return alias_type(idx); 1354} 1355 1356 1357Compile::AliasType* Compile::alias_type(ciField* field) { 1358 const TypeOopPtr* t; 1359 if (field->is_static()) 1360 t = TypeKlassPtr::make(field->holder()); 1361 else 1362 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1363 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1364 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1365 return atp; 1366} 1367 1368 1369//------------------------------have_alias_type-------------------------------- 1370bool Compile::have_alias_type(const TypePtr* adr_type) { 1371 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1372 if (ace->_adr_type == adr_type) { 1373 return true; 1374 } 1375 1376 // Handle special cases. 1377 if (adr_type == NULL) return true; 1378 if (adr_type == TypePtr::BOTTOM) return true; 1379 1380 return find_alias_type(adr_type, true) != NULL; 1381} 1382 1383//-----------------------------must_alias-------------------------------------- 1384// True if all values of the given address type are in the given alias category. 1385bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1386 if (alias_idx == AliasIdxBot) return true; // the universal category 1387 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1388 if (alias_idx == AliasIdxTop) return false; // the empty category 1389 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1390 1391 // the only remaining possible overlap is identity 1392 int adr_idx = get_alias_index(adr_type); 1393 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1394 assert(adr_idx == alias_idx || 1395 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1396 && adr_type != TypeOopPtr::BOTTOM), 1397 "should not be testing for overlap with an unsafe pointer"); 1398 return adr_idx == alias_idx; 1399} 1400 1401//------------------------------can_alias-------------------------------------- 1402// True if any values of the given address type are in the given alias category. 1403bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1404 if (alias_idx == AliasIdxTop) return false; // the empty category 1405 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1406 if (alias_idx == AliasIdxBot) return true; // the universal category 1407 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1408 1409 // the only remaining possible overlap is identity 1410 int adr_idx = get_alias_index(adr_type); 1411 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1412 return adr_idx == alias_idx; 1413} 1414 1415 1416 1417//---------------------------pop_warm_call------------------------------------- 1418WarmCallInfo* Compile::pop_warm_call() { 1419 WarmCallInfo* wci = _warm_calls; 1420 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1421 return wci; 1422} 1423 1424//----------------------------Inline_Warm-------------------------------------- 1425int Compile::Inline_Warm() { 1426 // If there is room, try to inline some more warm call sites. 1427 // %%% Do a graph index compaction pass when we think we're out of space? 1428 if (!InlineWarmCalls) return 0; 1429 1430 int calls_made_hot = 0; 1431 int room_to_grow = NodeCountInliningCutoff - unique(); 1432 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1433 int amount_grown = 0; 1434 WarmCallInfo* call; 1435 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1436 int est_size = (int)call->size(); 1437 if (est_size > (room_to_grow - amount_grown)) { 1438 // This one won't fit anyway. Get rid of it. 1439 call->make_cold(); 1440 continue; 1441 } 1442 call->make_hot(); 1443 calls_made_hot++; 1444 amount_grown += est_size; 1445 amount_to_grow -= est_size; 1446 } 1447 1448 if (calls_made_hot > 0) set_major_progress(); 1449 return calls_made_hot; 1450} 1451 1452 1453//----------------------------Finish_Warm-------------------------------------- 1454void Compile::Finish_Warm() { 1455 if (!InlineWarmCalls) return; 1456 if (failing()) return; 1457 if (warm_calls() == NULL) return; 1458 1459 // Clean up loose ends, if we are out of space for inlining. 1460 WarmCallInfo* call; 1461 while ((call = pop_warm_call()) != NULL) { 1462 call->make_cold(); 1463 } 1464} 1465 1466 1467//------------------------------Optimize--------------------------------------- 1468// Given a graph, optimize it. 1469void Compile::Optimize() { 1470 TracePhase t1("optimizer", &_t_optimizer, true); 1471 1472#ifndef PRODUCT 1473 if (env()->break_at_compile()) { 1474 BREAKPOINT; 1475 } 1476 1477#endif 1478 1479 ResourceMark rm; 1480 int loop_opts_cnt; 1481 1482 NOT_PRODUCT( verify_graph_edges(); ) 1483 1484 print_method("Start"); 1485 1486 { 1487 // Iterative Global Value Numbering, including ideal transforms 1488 // Initialize IterGVN with types and values from parse-time GVN 1489 PhaseIterGVN igvn(initial_gvn()); 1490 { 1491 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1492 igvn.optimize(); 1493 } 1494 1495 print_method("Iter GVN 1", 2); 1496 1497 if (failing()) return; 1498 1499 // get rid of the connection graph since it's information is not 1500 // updated by optimizations 1501 _congraph = NULL; 1502 1503 1504 // Loop transforms on the ideal graph. Range Check Elimination, 1505 // peeling, unrolling, etc. 1506 1507 // Set loop opts counter 1508 loop_opts_cnt = num_loop_opts(); 1509 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1510 { 1511 TracePhase t2("idealLoop", &_t_idealLoop, true); 1512 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1513 loop_opts_cnt--; 1514 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1515 if (failing()) return; 1516 } 1517 // Loop opts pass if partial peeling occurred in previous pass 1518 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1519 TracePhase t3("idealLoop", &_t_idealLoop, true); 1520 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1521 loop_opts_cnt--; 1522 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1523 if (failing()) return; 1524 } 1525 // Loop opts pass for loop-unrolling before CCP 1526 if(major_progress() && (loop_opts_cnt > 0)) { 1527 TracePhase t4("idealLoop", &_t_idealLoop, true); 1528 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1529 loop_opts_cnt--; 1530 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1531 } 1532 } 1533 if (failing()) return; 1534 1535 // Conditional Constant Propagation; 1536 PhaseCCP ccp( &igvn ); 1537 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1538 { 1539 TracePhase t2("ccp", &_t_ccp, true); 1540 ccp.do_transform(); 1541 } 1542 print_method("PhaseCPP 1", 2); 1543 1544 assert( true, "Break here to ccp.dump_old2new_map()"); 1545 1546 // Iterative Global Value Numbering, including ideal transforms 1547 { 1548 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1549 igvn = ccp; 1550 igvn.optimize(); 1551 } 1552 1553 print_method("Iter GVN 2", 2); 1554 1555 if (failing()) return; 1556 1557 // Loop transforms on the ideal graph. Range Check Elimination, 1558 // peeling, unrolling, etc. 1559 if(loop_opts_cnt > 0) { 1560 debug_only( int cnt = 0; ); 1561 while(major_progress() && (loop_opts_cnt > 0)) { 1562 TracePhase t2("idealLoop", &_t_idealLoop, true); 1563 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1564 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1565 loop_opts_cnt--; 1566 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1567 if (failing()) return; 1568 } 1569 } 1570 { 1571 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1572 PhaseMacroExpand mex(igvn); 1573 if (mex.expand_macro_nodes()) { 1574 assert(failing(), "must bail out w/ explicit message"); 1575 return; 1576 } 1577 } 1578 1579 } // (End scope of igvn; run destructor if necessary for asserts.) 1580 1581 // A method with only infinite loops has no edges entering loops from root 1582 { 1583 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1584 if (final_graph_reshaping()) { 1585 assert(failing(), "must bail out w/ explicit message"); 1586 return; 1587 } 1588 } 1589 1590 print_method("Optimize finished", 2); 1591} 1592 1593 1594//------------------------------Code_Gen--------------------------------------- 1595// Given a graph, generate code for it 1596void Compile::Code_Gen() { 1597 if (failing()) return; 1598 1599 // Perform instruction selection. You might think we could reclaim Matcher 1600 // memory PDQ, but actually the Matcher is used in generating spill code. 1601 // Internals of the Matcher (including some VectorSets) must remain live 1602 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1603 // set a bit in reclaimed memory. 1604 1605 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1606 // nodes. Mapping is only valid at the root of each matched subtree. 1607 NOT_PRODUCT( verify_graph_edges(); ) 1608 1609 Node_List proj_list; 1610 Matcher m(proj_list); 1611 _matcher = &m; 1612 { 1613 TracePhase t2("matcher", &_t_matcher, true); 1614 m.match(); 1615 } 1616 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1617 // nodes. Mapping is only valid at the root of each matched subtree. 1618 NOT_PRODUCT( verify_graph_edges(); ) 1619 1620 // If you have too many nodes, or if matching has failed, bail out 1621 check_node_count(0, "out of nodes matching instructions"); 1622 if (failing()) return; 1623 1624 // Build a proper-looking CFG 1625 PhaseCFG cfg(node_arena(), root(), m); 1626 _cfg = &cfg; 1627 { 1628 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1629 cfg.Dominators(); 1630 if (failing()) return; 1631 1632 NOT_PRODUCT( verify_graph_edges(); ) 1633 1634 cfg.Estimate_Block_Frequency(); 1635 cfg.GlobalCodeMotion(m,unique(),proj_list); 1636 1637 print_method("Global code motion", 2); 1638 1639 if (failing()) return; 1640 NOT_PRODUCT( verify_graph_edges(); ) 1641 1642 debug_only( cfg.verify(); ) 1643 } 1644 NOT_PRODUCT( verify_graph_edges(); ) 1645 1646 PhaseChaitin regalloc(unique(),cfg,m); 1647 _regalloc = ®alloc; 1648 { 1649 TracePhase t2("regalloc", &_t_registerAllocation, true); 1650 // Perform any platform dependent preallocation actions. This is used, 1651 // for example, to avoid taking an implicit null pointer exception 1652 // using the frame pointer on win95. 1653 _regalloc->pd_preallocate_hook(); 1654 1655 // Perform register allocation. After Chaitin, use-def chains are 1656 // no longer accurate (at spill code) and so must be ignored. 1657 // Node->LRG->reg mappings are still accurate. 1658 _regalloc->Register_Allocate(); 1659 1660 // Bail out if the allocator builds too many nodes 1661 if (failing()) return; 1662 } 1663 1664 // Prior to register allocation we kept empty basic blocks in case the 1665 // the allocator needed a place to spill. After register allocation we 1666 // are not adding any new instructions. If any basic block is empty, we 1667 // can now safely remove it. 1668 { 1669 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); ) 1670 cfg.RemoveEmpty(); 1671 } 1672 1673 // Perform any platform dependent postallocation verifications. 1674 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1675 1676 // Apply peephole optimizations 1677 if( OptoPeephole ) { 1678 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1679 PhasePeephole peep( _regalloc, cfg); 1680 peep.do_transform(); 1681 } 1682 1683 // Convert Nodes to instruction bits in a buffer 1684 { 1685 // %%%% workspace merge brought two timers together for one job 1686 TracePhase t2a("output", &_t_output, true); 1687 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1688 Output(); 1689 } 1690 1691 print_method("End"); 1692 1693 // He's dead, Jim. 1694 _cfg = (PhaseCFG*)0xdeadbeef; 1695 _regalloc = (PhaseChaitin*)0xdeadbeef; 1696} 1697 1698 1699//------------------------------dump_asm--------------------------------------- 1700// Dump formatted assembly 1701#ifndef PRODUCT 1702void Compile::dump_asm(int *pcs, uint pc_limit) { 1703 bool cut_short = false; 1704 tty->print_cr("#"); 1705 tty->print("# "); _tf->dump(); tty->cr(); 1706 tty->print_cr("#"); 1707 1708 // For all blocks 1709 int pc = 0x0; // Program counter 1710 char starts_bundle = ' '; 1711 _regalloc->dump_frame(); 1712 1713 Node *n = NULL; 1714 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1715 if (VMThread::should_terminate()) { cut_short = true; break; } 1716 Block *b = _cfg->_blocks[i]; 1717 if (b->is_connector() && !Verbose) continue; 1718 n = b->_nodes[0]; 1719 if (pcs && n->_idx < pc_limit) 1720 tty->print("%3.3x ", pcs[n->_idx]); 1721 else 1722 tty->print(" "); 1723 b->dump_head( &_cfg->_bbs ); 1724 if (b->is_connector()) { 1725 tty->print_cr(" # Empty connector block"); 1726 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1727 tty->print_cr(" # Block is sole successor of call"); 1728 } 1729 1730 // For all instructions 1731 Node *delay = NULL; 1732 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1733 if (VMThread::should_terminate()) { cut_short = true; break; } 1734 n = b->_nodes[j]; 1735 if (valid_bundle_info(n)) { 1736 Bundle *bundle = node_bundling(n); 1737 if (bundle->used_in_unconditional_delay()) { 1738 delay = n; 1739 continue; 1740 } 1741 if (bundle->starts_bundle()) 1742 starts_bundle = '+'; 1743 } 1744 1745 if (WizardMode) n->dump(); 1746 1747 if( !n->is_Region() && // Dont print in the Assembly 1748 !n->is_Phi() && // a few noisely useless nodes 1749 !n->is_Proj() && 1750 !n->is_MachTemp() && 1751 !n->is_Catch() && // Would be nice to print exception table targets 1752 !n->is_MergeMem() && // Not very interesting 1753 !n->is_top() && // Debug info table constants 1754 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1755 ) { 1756 if (pcs && n->_idx < pc_limit) 1757 tty->print("%3.3x", pcs[n->_idx]); 1758 else 1759 tty->print(" "); 1760 tty->print(" %c ", starts_bundle); 1761 starts_bundle = ' '; 1762 tty->print("\t"); 1763 n->format(_regalloc, tty); 1764 tty->cr(); 1765 } 1766 1767 // If we have an instruction with a delay slot, and have seen a delay, 1768 // then back up and print it 1769 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1770 assert(delay != NULL, "no unconditional delay instruction"); 1771 if (WizardMode) delay->dump(); 1772 1773 if (node_bundling(delay)->starts_bundle()) 1774 starts_bundle = '+'; 1775 if (pcs && n->_idx < pc_limit) 1776 tty->print("%3.3x", pcs[n->_idx]); 1777 else 1778 tty->print(" "); 1779 tty->print(" %c ", starts_bundle); 1780 starts_bundle = ' '; 1781 tty->print("\t"); 1782 delay->format(_regalloc, tty); 1783 tty->print_cr(""); 1784 delay = NULL; 1785 } 1786 1787 // Dump the exception table as well 1788 if( n->is_Catch() && (Verbose || WizardMode) ) { 1789 // Print the exception table for this offset 1790 _handler_table.print_subtable_for(pc); 1791 } 1792 } 1793 1794 if (pcs && n->_idx < pc_limit) 1795 tty->print_cr("%3.3x", pcs[n->_idx]); 1796 else 1797 tty->print_cr(""); 1798 1799 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1800 1801 } // End of per-block dump 1802 tty->print_cr(""); 1803 1804 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1805} 1806#endif 1807 1808//------------------------------Final_Reshape_Counts--------------------------- 1809// This class defines counters to help identify when a method 1810// may/must be executed using hardware with only 24-bit precision. 1811struct Final_Reshape_Counts : public StackObj { 1812 int _call_count; // count non-inlined 'common' calls 1813 int _float_count; // count float ops requiring 24-bit precision 1814 int _double_count; // count double ops requiring more precision 1815 int _java_call_count; // count non-inlined 'java' calls 1816 VectorSet _visited; // Visitation flags 1817 Node_List _tests; // Set of IfNodes & PCTableNodes 1818 1819 Final_Reshape_Counts() : 1820 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), 1821 _visited( Thread::current()->resource_area() ) { } 1822 1823 void inc_call_count () { _call_count ++; } 1824 void inc_float_count () { _float_count ++; } 1825 void inc_double_count() { _double_count++; } 1826 void inc_java_call_count() { _java_call_count++; } 1827 1828 int get_call_count () const { return _call_count ; } 1829 int get_float_count () const { return _float_count ; } 1830 int get_double_count() const { return _double_count; } 1831 int get_java_call_count() const { return _java_call_count; } 1832}; 1833 1834static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1835 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1836 // Make sure the offset goes inside the instance layout. 1837 return k->contains_field_offset(tp->offset()); 1838 // Note that OffsetBot and OffsetTop are very negative. 1839} 1840 1841//------------------------------final_graph_reshaping_impl---------------------- 1842// Implement items 1-5 from final_graph_reshaping below. 1843static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { 1844 1845 if ( n->outcnt() == 0 ) return; // dead node 1846 uint nop = n->Opcode(); 1847 1848 // Check for 2-input instruction with "last use" on right input. 1849 // Swap to left input. Implements item (2). 1850 if( n->req() == 3 && // two-input instruction 1851 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1852 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1853 n->in(2)->outcnt() == 1 &&// right use IS a last use 1854 !n->in(2)->is_Con() ) { // right use is not a constant 1855 // Check for commutative opcode 1856 switch( nop ) { 1857 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1858 case Op_MaxI: case Op_MinI: 1859 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1860 case Op_AndL: case Op_XorL: case Op_OrL: 1861 case Op_AndI: case Op_XorI: case Op_OrI: { 1862 // Move "last use" input to left by swapping inputs 1863 n->swap_edges(1, 2); 1864 break; 1865 } 1866 default: 1867 break; 1868 } 1869 } 1870 1871 // Count FPU ops and common calls, implements item (3) 1872 switch( nop ) { 1873 // Count all float operations that may use FPU 1874 case Op_AddF: 1875 case Op_SubF: 1876 case Op_MulF: 1877 case Op_DivF: 1878 case Op_NegF: 1879 case Op_ModF: 1880 case Op_ConvI2F: 1881 case Op_ConF: 1882 case Op_CmpF: 1883 case Op_CmpF3: 1884 // case Op_ConvL2F: // longs are split into 32-bit halves 1885 fpu.inc_float_count(); 1886 break; 1887 1888 case Op_ConvF2D: 1889 case Op_ConvD2F: 1890 fpu.inc_float_count(); 1891 fpu.inc_double_count(); 1892 break; 1893 1894 // Count all double operations that may use FPU 1895 case Op_AddD: 1896 case Op_SubD: 1897 case Op_MulD: 1898 case Op_DivD: 1899 case Op_NegD: 1900 case Op_ModD: 1901 case Op_ConvI2D: 1902 case Op_ConvD2I: 1903 // case Op_ConvL2D: // handled by leaf call 1904 // case Op_ConvD2L: // handled by leaf call 1905 case Op_ConD: 1906 case Op_CmpD: 1907 case Op_CmpD3: 1908 fpu.inc_double_count(); 1909 break; 1910 case Op_Opaque1: // Remove Opaque Nodes before matching 1911 case Op_Opaque2: // Remove Opaque Nodes before matching 1912 n->subsume_by(n->in(1)); 1913 break; 1914 case Op_CallStaticJava: 1915 case Op_CallJava: 1916 case Op_CallDynamicJava: 1917 fpu.inc_java_call_count(); // Count java call site; 1918 case Op_CallRuntime: 1919 case Op_CallLeaf: 1920 case Op_CallLeafNoFP: { 1921 assert( n->is_Call(), "" ); 1922 CallNode *call = n->as_Call(); 1923 // Count call sites where the FP mode bit would have to be flipped. 1924 // Do not count uncommon runtime calls: 1925 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 1926 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 1927 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 1928 fpu.inc_call_count(); // Count the call site 1929 } else { // See if uncommon argument is shared 1930 Node *n = call->in(TypeFunc::Parms); 1931 int nop = n->Opcode(); 1932 // Clone shared simple arguments to uncommon calls, item (1). 1933 if( n->outcnt() > 1 && 1934 !n->is_Proj() && 1935 nop != Op_CreateEx && 1936 nop != Op_CheckCastPP && 1937 !n->is_Mem() ) { 1938 Node *x = n->clone(); 1939 call->set_req( TypeFunc::Parms, x ); 1940 } 1941 } 1942 break; 1943 } 1944 1945 case Op_StoreD: 1946 case Op_LoadD: 1947 case Op_LoadD_unaligned: 1948 fpu.inc_double_count(); 1949 goto handle_mem; 1950 case Op_StoreF: 1951 case Op_LoadF: 1952 fpu.inc_float_count(); 1953 goto handle_mem; 1954 1955 case Op_StoreB: 1956 case Op_StoreC: 1957 case Op_StoreCM: 1958 case Op_StorePConditional: 1959 case Op_StoreI: 1960 case Op_StoreL: 1961 case Op_StoreLConditional: 1962 case Op_CompareAndSwapI: 1963 case Op_CompareAndSwapL: 1964 case Op_CompareAndSwapP: 1965 case Op_CompareAndSwapN: 1966 case Op_StoreP: 1967 case Op_StoreN: 1968 case Op_LoadB: 1969 case Op_LoadC: 1970 case Op_LoadI: 1971 case Op_LoadKlass: 1972 case Op_LoadNKlass: 1973 case Op_LoadL: 1974 case Op_LoadL_unaligned: 1975 case Op_LoadPLocked: 1976 case Op_LoadLLocked: 1977 case Op_LoadP: 1978 case Op_LoadN: 1979 case Op_LoadRange: 1980 case Op_LoadS: { 1981 handle_mem: 1982#ifdef ASSERT 1983 if( VerifyOptoOopOffsets ) { 1984 assert( n->is_Mem(), "" ); 1985 MemNode *mem = (MemNode*)n; 1986 // Check to see if address types have grounded out somehow. 1987 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 1988 assert( !tp || oop_offset_is_sane(tp), "" ); 1989 } 1990#endif 1991 break; 1992 } 1993 1994 case Op_AddP: { // Assert sane base pointers 1995 Node *addp = n->in(AddPNode::Address); 1996 assert( !addp->is_AddP() || 1997 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 1998 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 1999 "Base pointers must match" ); 2000#ifdef _LP64 2001 if (UseCompressedOops && 2002 addp->Opcode() == Op_ConP && 2003 addp == n->in(AddPNode::Base) && 2004 n->in(AddPNode::Offset)->is_Con()) { 2005 // Use addressing with narrow klass to load with offset on x86. 2006 // On sparc loading 32-bits constant and decoding it have less 2007 // instructions (4) then load 64-bits constant (7). 2008 // Do this transformation here since IGVN will convert ConN back to ConP. 2009 const Type* t = addp->bottom_type(); 2010 if (t->isa_oopptr()) { 2011 Node* nn = NULL; 2012 2013 // Look for existing ConN node of the same exact type. 2014 Compile* C = Compile::current(); 2015 Node* r = C->root(); 2016 uint cnt = r->outcnt(); 2017 for (uint i = 0; i < cnt; i++) { 2018 Node* m = r->raw_out(i); 2019 if (m!= NULL && m->Opcode() == Op_ConN && 2020 m->bottom_type()->is_narrowoop()->make_oopptr() == t) { 2021 nn = m; 2022 break; 2023 } 2024 } 2025 if (nn != NULL) { 2026 // Decode a narrow oop to match address 2027 // [R12 + narrow_oop_reg<<3 + offset] 2028 nn = new (C, 2) DecodeNNode(nn, t); 2029 n->set_req(AddPNode::Base, nn); 2030 n->set_req(AddPNode::Address, nn); 2031 if (addp->outcnt() == 0) { 2032 addp->disconnect_inputs(NULL); 2033 } 2034 } 2035 } 2036 } 2037#endif 2038 break; 2039 } 2040 2041#ifdef _LP64 2042 case Op_CmpP: 2043 // Do this transformation here to preserve CmpPNode::sub() and 2044 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2045 if( n->in(1)->is_DecodeN() ) { 2046 Compile* C = Compile::current(); 2047 Node* in2 = NULL; 2048 if( n->in(2)->is_DecodeN() ) { 2049 in2 = n->in(2)->in(1); 2050 } else if ( n->in(2)->Opcode() == Op_ConP ) { 2051 const Type* t = n->in(2)->bottom_type(); 2052 if (t == TypePtr::NULL_PTR) { 2053 Node *in1 = n->in(1); 2054 if (Matcher::clone_shift_expressions) { 2055 // x86, ARM and friends can handle 2 adds in addressing mode. 2056 // Decode a narrow oop and do implicit NULL check in address 2057 // [R12 + narrow_oop_reg<<3 + offset] 2058 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2059 } else { 2060 // Don't replace CmpP(o ,null) if 'o' is used in AddP 2061 // to generate implicit NULL check on Sparc where 2062 // narrow oops can't be used in address. 2063 uint i = 0; 2064 for (; i < in1->outcnt(); i++) { 2065 if (in1->raw_out(i)->is_AddP()) 2066 break; 2067 } 2068 if (i >= in1->outcnt()) { 2069 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2070 } 2071 } 2072 } else if (t->isa_oopptr()) { 2073 in2 = ConNode::make(C, t->is_oopptr()->make_narrowoop()); 2074 } 2075 } 2076 if( in2 != NULL ) { 2077 Node* cmpN = new (C, 3) CmpNNode(n->in(1)->in(1), in2); 2078 n->subsume_by( cmpN ); 2079 } 2080 } 2081#endif 2082 2083 case Op_ModI: 2084 if (UseDivMod) { 2085 // Check if a%b and a/b both exist 2086 Node* d = n->find_similar(Op_DivI); 2087 if (d) { 2088 // Replace them with a fused divmod if supported 2089 Compile* C = Compile::current(); 2090 if (Matcher::has_match_rule(Op_DivModI)) { 2091 DivModINode* divmod = DivModINode::make(C, n); 2092 d->subsume_by(divmod->div_proj()); 2093 n->subsume_by(divmod->mod_proj()); 2094 } else { 2095 // replace a%b with a-((a/b)*b) 2096 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2097 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2098 n->subsume_by( sub ); 2099 } 2100 } 2101 } 2102 break; 2103 2104 case Op_ModL: 2105 if (UseDivMod) { 2106 // Check if a%b and a/b both exist 2107 Node* d = n->find_similar(Op_DivL); 2108 if (d) { 2109 // Replace them with a fused divmod if supported 2110 Compile* C = Compile::current(); 2111 if (Matcher::has_match_rule(Op_DivModL)) { 2112 DivModLNode* divmod = DivModLNode::make(C, n); 2113 d->subsume_by(divmod->div_proj()); 2114 n->subsume_by(divmod->mod_proj()); 2115 } else { 2116 // replace a%b with a-((a/b)*b) 2117 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2118 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2119 n->subsume_by( sub ); 2120 } 2121 } 2122 } 2123 break; 2124 2125 case Op_Load16B: 2126 case Op_Load8B: 2127 case Op_Load4B: 2128 case Op_Load8S: 2129 case Op_Load4S: 2130 case Op_Load2S: 2131 case Op_Load8C: 2132 case Op_Load4C: 2133 case Op_Load2C: 2134 case Op_Load4I: 2135 case Op_Load2I: 2136 case Op_Load2L: 2137 case Op_Load4F: 2138 case Op_Load2F: 2139 case Op_Load2D: 2140 case Op_Store16B: 2141 case Op_Store8B: 2142 case Op_Store4B: 2143 case Op_Store8C: 2144 case Op_Store4C: 2145 case Op_Store2C: 2146 case Op_Store4I: 2147 case Op_Store2I: 2148 case Op_Store2L: 2149 case Op_Store4F: 2150 case Op_Store2F: 2151 case Op_Store2D: 2152 break; 2153 2154 case Op_PackB: 2155 case Op_PackS: 2156 case Op_PackC: 2157 case Op_PackI: 2158 case Op_PackF: 2159 case Op_PackL: 2160 case Op_PackD: 2161 if (n->req()-1 > 2) { 2162 // Replace many operand PackNodes with a binary tree for matching 2163 PackNode* p = (PackNode*) n; 2164 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2165 n->subsume_by(btp); 2166 } 2167 break; 2168 default: 2169 assert( !n->is_Call(), "" ); 2170 assert( !n->is_Mem(), "" ); 2171 break; 2172 } 2173 2174 // Collect CFG split points 2175 if (n->is_MultiBranch()) 2176 fpu._tests.push(n); 2177} 2178 2179//------------------------------final_graph_reshaping_walk--------------------- 2180// Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2181// requires that the walk visits a node's inputs before visiting the node. 2182static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { 2183 fpu._visited.set(root->_idx); // first, mark node as visited 2184 uint cnt = root->req(); 2185 Node *n = root; 2186 uint i = 0; 2187 while (true) { 2188 if (i < cnt) { 2189 // Place all non-visited non-null inputs onto stack 2190 Node* m = n->in(i); 2191 ++i; 2192 if (m != NULL && !fpu._visited.test_set(m->_idx)) { 2193 cnt = m->req(); 2194 nstack.push(n, i); // put on stack parent and next input's index 2195 n = m; 2196 i = 0; 2197 } 2198 } else { 2199 // Now do post-visit work 2200 final_graph_reshaping_impl( n, fpu ); 2201 if (nstack.is_empty()) 2202 break; // finished 2203 n = nstack.node(); // Get node from stack 2204 cnt = n->req(); 2205 i = nstack.index(); 2206 nstack.pop(); // Shift to the next node on stack 2207 } 2208 } 2209} 2210 2211//------------------------------final_graph_reshaping-------------------------- 2212// Final Graph Reshaping. 2213// 2214// (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2215// and not commoned up and forced early. Must come after regular 2216// optimizations to avoid GVN undoing the cloning. Clone constant 2217// inputs to Loop Phis; these will be split by the allocator anyways. 2218// Remove Opaque nodes. 2219// (2) Move last-uses by commutative operations to the left input to encourage 2220// Intel update-in-place two-address operations and better register usage 2221// on RISCs. Must come after regular optimizations to avoid GVN Ideal 2222// calls canonicalizing them back. 2223// (3) Count the number of double-precision FP ops, single-precision FP ops 2224// and call sites. On Intel, we can get correct rounding either by 2225// forcing singles to memory (requires extra stores and loads after each 2226// FP bytecode) or we can set a rounding mode bit (requires setting and 2227// clearing the mode bit around call sites). The mode bit is only used 2228// if the relative frequency of single FP ops to calls is low enough. 2229// This is a key transform for SPEC mpeg_audio. 2230// (4) Detect infinite loops; blobs of code reachable from above but not 2231// below. Several of the Code_Gen algorithms fail on such code shapes, 2232// so we simply bail out. Happens a lot in ZKM.jar, but also happens 2233// from time to time in other codes (such as -Xcomp finalizer loops, etc). 2234// Detection is by looking for IfNodes where only 1 projection is 2235// reachable from below or CatchNodes missing some targets. 2236// (5) Assert for insane oop offsets in debug mode. 2237 2238bool Compile::final_graph_reshaping() { 2239 // an infinite loop may have been eliminated by the optimizer, 2240 // in which case the graph will be empty. 2241 if (root()->req() == 1) { 2242 record_method_not_compilable("trivial infinite loop"); 2243 return true; 2244 } 2245 2246 Final_Reshape_Counts fpu; 2247 2248 // Visit everybody reachable! 2249 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2250 Node_Stack nstack(unique() >> 1); 2251 final_graph_reshaping_walk(nstack, root(), fpu); 2252 2253 // Check for unreachable (from below) code (i.e., infinite loops). 2254 for( uint i = 0; i < fpu._tests.size(); i++ ) { 2255 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch(); 2256 // Get number of CFG targets. 2257 // Note that PCTables include exception targets after calls. 2258 uint required_outcnt = n->required_outcnt(); 2259 if (n->outcnt() != required_outcnt) { 2260 // Check for a few special cases. Rethrow Nodes never take the 2261 // 'fall-thru' path, so expected kids is 1 less. 2262 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2263 if (n->in(0)->in(0)->is_Call()) { 2264 CallNode *call = n->in(0)->in(0)->as_Call(); 2265 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2266 required_outcnt--; // Rethrow always has 1 less kid 2267 } else if (call->req() > TypeFunc::Parms && 2268 call->is_CallDynamicJava()) { 2269 // Check for null receiver. In such case, the optimizer has 2270 // detected that the virtual call will always result in a null 2271 // pointer exception. The fall-through projection of this CatchNode 2272 // will not be populated. 2273 Node *arg0 = call->in(TypeFunc::Parms); 2274 if (arg0->is_Type() && 2275 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2276 required_outcnt--; 2277 } 2278 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2279 call->req() > TypeFunc::Parms+1 && 2280 call->is_CallStaticJava()) { 2281 // Check for negative array length. In such case, the optimizer has 2282 // detected that the allocation attempt will always result in an 2283 // exception. There is no fall-through projection of this CatchNode . 2284 Node *arg1 = call->in(TypeFunc::Parms+1); 2285 if (arg1->is_Type() && 2286 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2287 required_outcnt--; 2288 } 2289 } 2290 } 2291 } 2292 // Recheck with a better notion of 'required_outcnt' 2293 if (n->outcnt() != required_outcnt) { 2294 record_method_not_compilable("malformed control flow"); 2295 return true; // Not all targets reachable! 2296 } 2297 } 2298 // Check that I actually visited all kids. Unreached kids 2299 // must be infinite loops. 2300 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2301 if (!fpu._visited.test(n->fast_out(j)->_idx)) { 2302 record_method_not_compilable("infinite loop"); 2303 return true; // Found unvisited kid; must be unreach 2304 } 2305 } 2306 2307 // If original bytecodes contained a mixture of floats and doubles 2308 // check if the optimizer has made it homogenous, item (3). 2309 if( Use24BitFPMode && Use24BitFP && 2310 fpu.get_float_count() > 32 && 2311 fpu.get_double_count() == 0 && 2312 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { 2313 set_24_bit_selection_and_mode( false, true ); 2314 } 2315 2316 set_has_java_calls(fpu.get_java_call_count() > 0); 2317 2318 // No infinite loops, no reason to bail out. 2319 return false; 2320} 2321 2322//-----------------------------too_many_traps---------------------------------- 2323// Report if there are too many traps at the current method and bci. 2324// Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2325bool Compile::too_many_traps(ciMethod* method, 2326 int bci, 2327 Deoptimization::DeoptReason reason) { 2328 ciMethodData* md = method->method_data(); 2329 if (md->is_empty()) { 2330 // Assume the trap has not occurred, or that it occurred only 2331 // because of a transient condition during start-up in the interpreter. 2332 return false; 2333 } 2334 if (md->has_trap_at(bci, reason) != 0) { 2335 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2336 // Also, if there are multiple reasons, or if there is no per-BCI record, 2337 // assume the worst. 2338 if (log()) 2339 log()->elem("observe trap='%s' count='%d'", 2340 Deoptimization::trap_reason_name(reason), 2341 md->trap_count(reason)); 2342 return true; 2343 } else { 2344 // Ignore method/bci and see if there have been too many globally. 2345 return too_many_traps(reason, md); 2346 } 2347} 2348 2349// Less-accurate variant which does not require a method and bci. 2350bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2351 ciMethodData* logmd) { 2352 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2353 // Too many traps globally. 2354 // Note that we use cumulative trap_count, not just md->trap_count. 2355 if (log()) { 2356 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2357 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2358 Deoptimization::trap_reason_name(reason), 2359 mcount, trap_count(reason)); 2360 } 2361 return true; 2362 } else { 2363 // The coast is clear. 2364 return false; 2365 } 2366} 2367 2368//--------------------------too_many_recompiles-------------------------------- 2369// Report if there are too many recompiles at the current method and bci. 2370// Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2371// Is not eager to return true, since this will cause the compiler to use 2372// Action_none for a trap point, to avoid too many recompilations. 2373bool Compile::too_many_recompiles(ciMethod* method, 2374 int bci, 2375 Deoptimization::DeoptReason reason) { 2376 ciMethodData* md = method->method_data(); 2377 if (md->is_empty()) { 2378 // Assume the trap has not occurred, or that it occurred only 2379 // because of a transient condition during start-up in the interpreter. 2380 return false; 2381 } 2382 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2383 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2384 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2385 Deoptimization::DeoptReason per_bc_reason 2386 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2387 if ((per_bc_reason == Deoptimization::Reason_none 2388 || md->has_trap_at(bci, reason) != 0) 2389 // The trap frequency measure we care about is the recompile count: 2390 && md->trap_recompiled_at(bci) 2391 && md->overflow_recompile_count() >= bc_cutoff) { 2392 // Do not emit a trap here if it has already caused recompilations. 2393 // Also, if there are multiple reasons, or if there is no per-BCI record, 2394 // assume the worst. 2395 if (log()) 2396 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2397 Deoptimization::trap_reason_name(reason), 2398 md->trap_count(reason), 2399 md->overflow_recompile_count()); 2400 return true; 2401 } else if (trap_count(reason) != 0 2402 && decompile_count() >= m_cutoff) { 2403 // Too many recompiles globally, and we have seen this sort of trap. 2404 // Use cumulative decompile_count, not just md->decompile_count. 2405 if (log()) 2406 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2407 Deoptimization::trap_reason_name(reason), 2408 md->trap_count(reason), trap_count(reason), 2409 md->decompile_count(), decompile_count()); 2410 return true; 2411 } else { 2412 // The coast is clear. 2413 return false; 2414 } 2415} 2416 2417 2418#ifndef PRODUCT 2419//------------------------------verify_graph_edges--------------------------- 2420// Walk the Graph and verify that there is a one-to-one correspondence 2421// between Use-Def edges and Def-Use edges in the graph. 2422void Compile::verify_graph_edges(bool no_dead_code) { 2423 if (VerifyGraphEdges) { 2424 ResourceArea *area = Thread::current()->resource_area(); 2425 Unique_Node_List visited(area); 2426 // Call recursive graph walk to check edges 2427 _root->verify_edges(visited); 2428 if (no_dead_code) { 2429 // Now make sure that no visited node is used by an unvisited node. 2430 bool dead_nodes = 0; 2431 Unique_Node_List checked(area); 2432 while (visited.size() > 0) { 2433 Node* n = visited.pop(); 2434 checked.push(n); 2435 for (uint i = 0; i < n->outcnt(); i++) { 2436 Node* use = n->raw_out(i); 2437 if (checked.member(use)) continue; // already checked 2438 if (visited.member(use)) continue; // already in the graph 2439 if (use->is_Con()) continue; // a dead ConNode is OK 2440 // At this point, we have found a dead node which is DU-reachable. 2441 if (dead_nodes++ == 0) 2442 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2443 use->dump(2); 2444 tty->print_cr("---"); 2445 checked.push(use); // No repeats; pretend it is now checked. 2446 } 2447 } 2448 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2449 } 2450 } 2451} 2452#endif 2453 2454// The Compile object keeps track of failure reasons separately from the ciEnv. 2455// This is required because there is not quite a 1-1 relation between the 2456// ciEnv and its compilation task and the Compile object. Note that one 2457// ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2458// to backtrack and retry without subsuming loads. Other than this backtracking 2459// behavior, the Compile's failure reason is quietly copied up to the ciEnv 2460// by the logic in C2Compiler. 2461void Compile::record_failure(const char* reason) { 2462 if (log() != NULL) { 2463 log()->elem("failure reason='%s' phase='compile'", reason); 2464 } 2465 if (_failure_reason == NULL) { 2466 // Record the first failure reason. 2467 _failure_reason = reason; 2468 } 2469 _root = NULL; // flush the graph, too 2470} 2471 2472Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2473 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2474{ 2475 if (dolog) { 2476 C = Compile::current(); 2477 _log = C->log(); 2478 } else { 2479 C = NULL; 2480 _log = NULL; 2481 } 2482 if (_log != NULL) { 2483 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2484 _log->stamp(); 2485 _log->end_head(); 2486 } 2487} 2488 2489Compile::TracePhase::~TracePhase() { 2490 if (_log != NULL) { 2491 _log->done("phase nodes='%d'", C->unique()); 2492 } 2493} 2494