compile.cpp revision 222:2a1a77d3458f
1/* 2 * Copyright 1997-2007 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 _compile->end_method(); 317 if (_compile->scratch_buffer_blob() != NULL) 318 BufferBlob::free(_compile->scratch_buffer_blob()); 319 _compile->env()->set_compiler_data(NULL); 320} 321 322 323//----------------------------print_compile_messages--------------------------- 324void Compile::print_compile_messages() { 325#ifndef PRODUCT 326 // Check if recompiling 327 if (_subsume_loads == false && PrintOpto) { 328 // Recompiling without allowing machine instructions to subsume loads 329 tty->print_cr("*********************************************************"); 330 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 331 tty->print_cr("*********************************************************"); 332 } 333 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 334 // Recompiling without escape analysis 335 tty->print_cr("*********************************************************"); 336 tty->print_cr("** Bailout: Recompile without escape analysis **"); 337 tty->print_cr("*********************************************************"); 338 } 339 if (env()->break_at_compile()) { 340 // Open the debugger when compiing this method. 341 tty->print("### Breaking when compiling: "); 342 method()->print_short_name(); 343 tty->cr(); 344 BREAKPOINT; 345 } 346 347 if( PrintOpto ) { 348 if (is_osr_compilation()) { 349 tty->print("[OSR]%3d", _compile_id); 350 } else { 351 tty->print("%3d", _compile_id); 352 } 353 } 354#endif 355} 356 357 358void Compile::init_scratch_buffer_blob() { 359 if( scratch_buffer_blob() != NULL ) return; 360 361 // Construct a temporary CodeBuffer to have it construct a BufferBlob 362 // Cache this BufferBlob for this compile. 363 ResourceMark rm; 364 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 365 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 366 // Record the buffer blob for next time. 367 set_scratch_buffer_blob(blob); 368 // Have we run out of code space? 369 if (scratch_buffer_blob() == NULL) { 370 // Let CompilerBroker disable further compilations. 371 record_failure("Not enough space for scratch buffer in CodeCache"); 372 return; 373 } 374 375 // Initialize the relocation buffers 376 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 377 set_scratch_locs_memory(locs_buf); 378} 379 380 381//-----------------------scratch_emit_size------------------------------------- 382// Helper function that computes size by emitting code 383uint Compile::scratch_emit_size(const Node* n) { 384 // Emit into a trash buffer and count bytes emitted. 385 // This is a pretty expensive way to compute a size, 386 // but it works well enough if seldom used. 387 // All common fixed-size instructions are given a size 388 // method by the AD file. 389 // Note that the scratch buffer blob and locs memory are 390 // allocated at the beginning of the compile task, and 391 // may be shared by several calls to scratch_emit_size. 392 // The allocation of the scratch buffer blob is particularly 393 // expensive, since it has to grab the code cache lock. 394 BufferBlob* blob = this->scratch_buffer_blob(); 395 assert(blob != NULL, "Initialize BufferBlob at start"); 396 assert(blob->size() > MAX_inst_size, "sanity"); 397 relocInfo* locs_buf = scratch_locs_memory(); 398 address blob_begin = blob->instructions_begin(); 399 address blob_end = (address)locs_buf; 400 assert(blob->instructions_contains(blob_end), "sanity"); 401 CodeBuffer buf(blob_begin, blob_end - blob_begin); 402 buf.initialize_consts_size(MAX_const_size); 403 buf.initialize_stubs_size(MAX_stubs_size); 404 assert(locs_buf != NULL, "sanity"); 405 int lsize = MAX_locs_size / 2; 406 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 407 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 408 n->emit(buf, this->regalloc()); 409 return buf.code_size(); 410} 411 412 413// ============================================================================ 414//------------------------------Compile standard------------------------------- 415debug_only( int Compile::_debug_idx = 100000; ) 416 417// Compile a method. entry_bci is -1 for normal compilations and indicates 418// the continuation bci for on stack replacement. 419 420 421Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 422 : Phase(Compiler), 423 _env(ci_env), 424 _log(ci_env->log()), 425 _compile_id(ci_env->compile_id()), 426 _save_argument_registers(false), 427 _stub_name(NULL), 428 _stub_function(NULL), 429 _stub_entry_point(NULL), 430 _method(target), 431 _entry_bci(osr_bci), 432 _initial_gvn(NULL), 433 _for_igvn(NULL), 434 _warm_calls(NULL), 435 _subsume_loads(subsume_loads), 436 _do_escape_analysis(do_escape_analysis), 437 _failure_reason(NULL), 438 _code_buffer("Compile::Fill_buffer"), 439 _orig_pc_slot(0), 440 _orig_pc_slot_offset_in_bytes(0), 441 _node_bundling_limit(0), 442 _node_bundling_base(NULL), 443#ifndef PRODUCT 444 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 445 _printer(IdealGraphPrinter::printer()), 446#endif 447 _congraph(NULL) { 448 C = this; 449 450 CompileWrapper cw(this); 451#ifndef PRODUCT 452 if (TimeCompiler2) { 453 tty->print(" "); 454 target->holder()->name()->print(); 455 tty->print("."); 456 target->print_short_name(); 457 tty->print(" "); 458 } 459 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 460 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 461 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 462 if (!print_opto_assembly) { 463 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 464 if (print_assembly && !Disassembler::can_decode()) { 465 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 466 print_opto_assembly = true; 467 } 468 } 469 set_print_assembly(print_opto_assembly); 470#endif 471 472 if (ProfileTraps) { 473 // Make sure the method being compiled gets its own MDO, 474 // so we can at least track the decompile_count(). 475 method()->build_method_data(); 476 } 477 478 Init(::AliasLevel); 479 480 481 print_compile_messages(); 482 483 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 484 _ilt = InlineTree::build_inline_tree_root(); 485 else 486 _ilt = NULL; 487 488 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 489 assert(num_alias_types() >= AliasIdxRaw, ""); 490 491#define MINIMUM_NODE_HASH 1023 492 // Node list that Iterative GVN will start with 493 Unique_Node_List for_igvn(comp_arena()); 494 set_for_igvn(&for_igvn); 495 496 // GVN that will be run immediately on new nodes 497 uint estimated_size = method()->code_size()*4+64; 498 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 499 PhaseGVN gvn(node_arena(), estimated_size); 500 set_initial_gvn(&gvn); 501 502 { // Scope for timing the parser 503 TracePhase t3("parse", &_t_parser, true); 504 505 // Put top into the hash table ASAP. 506 initial_gvn()->transform_no_reclaim(top()); 507 508 // Set up tf(), start(), and find a CallGenerator. 509 CallGenerator* cg; 510 if (is_osr_compilation()) { 511 const TypeTuple *domain = StartOSRNode::osr_domain(); 512 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 513 init_tf(TypeFunc::make(domain, range)); 514 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 515 initial_gvn()->set_type_bottom(s); 516 init_start(s); 517 cg = CallGenerator::for_osr(method(), entry_bci()); 518 } else { 519 // Normal case. 520 init_tf(TypeFunc::make(method())); 521 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 522 initial_gvn()->set_type_bottom(s); 523 init_start(s); 524 float past_uses = method()->interpreter_invocation_count(); 525 float expected_uses = past_uses; 526 cg = CallGenerator::for_inline(method(), expected_uses); 527 } 528 if (failing()) return; 529 if (cg == NULL) { 530 record_method_not_compilable_all_tiers("cannot parse method"); 531 return; 532 } 533 JVMState* jvms = build_start_state(start(), tf()); 534 if ((jvms = cg->generate(jvms)) == NULL) { 535 record_method_not_compilable("method parse failed"); 536 return; 537 } 538 GraphKit kit(jvms); 539 540 if (!kit.stopped()) { 541 // Accept return values, and transfer control we know not where. 542 // This is done by a special, unique ReturnNode bound to root. 543 return_values(kit.jvms()); 544 } 545 546 if (kit.has_exceptions()) { 547 // Any exceptions that escape from this call must be rethrown 548 // to whatever caller is dynamically above us on the stack. 549 // This is done by a special, unique RethrowNode bound to root. 550 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 551 } 552 553 // Remove clutter produced by parsing. 554 if (!failing()) { 555 ResourceMark rm; 556 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 557 } 558 } 559 560 // Note: Large methods are capped off in do_one_bytecode(). 561 if (failing()) return; 562 563 // After parsing, node notes are no longer automagic. 564 // They must be propagated by register_new_node_with_optimizer(), 565 // clone(), or the like. 566 set_default_node_notes(NULL); 567 568 for (;;) { 569 int successes = Inline_Warm(); 570 if (failing()) return; 571 if (successes == 0) break; 572 } 573 574 // Drain the list. 575 Finish_Warm(); 576#ifndef PRODUCT 577 if (_printer) { 578 _printer->print_inlining(this); 579 } 580#endif 581 582 if (failing()) return; 583 NOT_PRODUCT( verify_graph_edges(); ) 584 585 // Perform escape analysis 586 if (_do_escape_analysis) 587 _congraph = new ConnectionGraph(this); 588 if (_congraph != NULL) { 589 NOT_PRODUCT( TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, TimeCompiler); ) 590 _congraph->compute_escape(); 591 if (failing()) return; 592 593#ifndef PRODUCT 594 if (PrintEscapeAnalysis) { 595 _congraph->dump(); 596 } 597#endif 598 } 599 // Now optimize 600 Optimize(); 601 if (failing()) return; 602 NOT_PRODUCT( verify_graph_edges(); ) 603 604 print_method("Before Matching"); 605 606#ifndef PRODUCT 607 if (PrintIdeal) { 608 ttyLocker ttyl; // keep the following output all in one block 609 // This output goes directly to the tty, not the compiler log. 610 // To enable tools to match it up with the compilation activity, 611 // be sure to tag this tty output with the compile ID. 612 if (xtty != NULL) { 613 xtty->head("ideal compile_id='%d'%s", compile_id(), 614 is_osr_compilation() ? " compile_kind='osr'" : 615 ""); 616 } 617 root()->dump(9999); 618 if (xtty != NULL) { 619 xtty->tail("ideal"); 620 } 621 } 622#endif 623 624 // Now that we know the size of all the monitors we can add a fixed slot 625 // for the original deopt pc. 626 627 _orig_pc_slot = fixed_slots(); 628 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 629 set_fixed_slots(next_slot); 630 631 // Now generate code 632 Code_Gen(); 633 if (failing()) return; 634 635 // Check if we want to skip execution of all compiled code. 636 { 637#ifndef PRODUCT 638 if (OptoNoExecute) { 639 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 640 return; 641 } 642 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 643#endif 644 645 if (is_osr_compilation()) { 646 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 647 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 648 } else { 649 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 650 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 651 } 652 653 env()->register_method(_method, _entry_bci, 654 &_code_offsets, 655 _orig_pc_slot_offset_in_bytes, 656 code_buffer(), 657 frame_size_in_words(), _oop_map_set, 658 &_handler_table, &_inc_table, 659 compiler, 660 env()->comp_level(), 661 true, /*has_debug_info*/ 662 has_unsafe_access() 663 ); 664 } 665} 666 667//------------------------------Compile---------------------------------------- 668// Compile a runtime stub 669Compile::Compile( ciEnv* ci_env, 670 TypeFunc_generator generator, 671 address stub_function, 672 const char *stub_name, 673 int is_fancy_jump, 674 bool pass_tls, 675 bool save_arg_registers, 676 bool return_pc ) 677 : Phase(Compiler), 678 _env(ci_env), 679 _log(ci_env->log()), 680 _compile_id(-1), 681 _save_argument_registers(save_arg_registers), 682 _method(NULL), 683 _stub_name(stub_name), 684 _stub_function(stub_function), 685 _stub_entry_point(NULL), 686 _entry_bci(InvocationEntryBci), 687 _initial_gvn(NULL), 688 _for_igvn(NULL), 689 _warm_calls(NULL), 690 _orig_pc_slot(0), 691 _orig_pc_slot_offset_in_bytes(0), 692 _subsume_loads(true), 693 _do_escape_analysis(false), 694 _failure_reason(NULL), 695 _code_buffer("Compile::Fill_buffer"), 696 _node_bundling_limit(0), 697 _node_bundling_base(NULL), 698#ifndef PRODUCT 699 _trace_opto_output(TraceOptoOutput), 700 _printer(NULL), 701#endif 702 _congraph(NULL) { 703 C = this; 704 705#ifndef PRODUCT 706 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 707 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 708 set_print_assembly(PrintFrameConverterAssembly); 709#endif 710 CompileWrapper cw(this); 711 Init(/*AliasLevel=*/ 0); 712 init_tf((*generator)()); 713 714 { 715 // The following is a dummy for the sake of GraphKit::gen_stub 716 Unique_Node_List for_igvn(comp_arena()); 717 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 718 PhaseGVN gvn(Thread::current()->resource_area(),255); 719 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 720 gvn.transform_no_reclaim(top()); 721 722 GraphKit kit; 723 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 724 } 725 726 NOT_PRODUCT( verify_graph_edges(); ) 727 Code_Gen(); 728 if (failing()) return; 729 730 731 // Entry point will be accessed using compile->stub_entry_point(); 732 if (code_buffer() == NULL) { 733 Matcher::soft_match_failure(); 734 } else { 735 if (PrintAssembly && (WizardMode || Verbose)) 736 tty->print_cr("### Stub::%s", stub_name); 737 738 if (!failing()) { 739 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 740 741 // Make the NMethod 742 // For now we mark the frame as never safe for profile stackwalking 743 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 744 code_buffer(), 745 CodeOffsets::frame_never_safe, 746 // _code_offsets.value(CodeOffsets::Frame_Complete), 747 frame_size_in_words(), 748 _oop_map_set, 749 save_arg_registers); 750 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 751 752 _stub_entry_point = rs->entry_point(); 753 } 754 } 755} 756 757#ifndef PRODUCT 758void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 759 if(PrintOpto && Verbose) { 760 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 761 } 762} 763#endif 764 765void Compile::print_codes() { 766} 767 768//------------------------------Init------------------------------------------- 769// Prepare for a single compilation 770void Compile::Init(int aliaslevel) { 771 _unique = 0; 772 _regalloc = NULL; 773 774 _tf = NULL; // filled in later 775 _top = NULL; // cached later 776 _matcher = NULL; // filled in later 777 _cfg = NULL; // filled in later 778 779 set_24_bit_selection_and_mode(Use24BitFP, false); 780 781 _node_note_array = NULL; 782 _default_node_notes = NULL; 783 784 _immutable_memory = NULL; // filled in at first inquiry 785 786 // Globally visible Nodes 787 // First set TOP to NULL to give safe behavior during creation of RootNode 788 set_cached_top_node(NULL); 789 set_root(new (this, 3) RootNode()); 790 // Now that you have a Root to point to, create the real TOP 791 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 792 set_recent_alloc(NULL, NULL); 793 794 // Create Debug Information Recorder to record scopes, oopmaps, etc. 795 env()->set_oop_recorder(new OopRecorder(comp_arena())); 796 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 797 env()->set_dependencies(new Dependencies(env())); 798 799 _fixed_slots = 0; 800 set_has_split_ifs(false); 801 set_has_loops(has_method() && method()->has_loops()); // first approximation 802 _deopt_happens = true; // start out assuming the worst 803 _trap_can_recompile = false; // no traps emitted yet 804 _major_progress = true; // start out assuming good things will happen 805 set_has_unsafe_access(false); 806 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 807 set_decompile_count(0); 808 809 // Compilation level related initialization 810 if (env()->comp_level() == CompLevel_fast_compile) { 811 set_num_loop_opts(Tier1LoopOptsCount); 812 set_do_inlining(Tier1Inline != 0); 813 set_max_inline_size(Tier1MaxInlineSize); 814 set_freq_inline_size(Tier1FreqInlineSize); 815 set_do_scheduling(false); 816 set_do_count_invocations(Tier1CountInvocations); 817 set_do_method_data_update(Tier1UpdateMethodData); 818 } else { 819 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 820 set_num_loop_opts(LoopOptsCount); 821 set_do_inlining(Inline); 822 set_max_inline_size(MaxInlineSize); 823 set_freq_inline_size(FreqInlineSize); 824 set_do_scheduling(OptoScheduling); 825 set_do_count_invocations(false); 826 set_do_method_data_update(false); 827 } 828 829 if (debug_info()->recording_non_safepoints()) { 830 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 831 (comp_arena(), 8, 0, NULL)); 832 set_default_node_notes(Node_Notes::make(this)); 833 } 834 835 // // -- Initialize types before each compile -- 836 // // Update cached type information 837 // if( _method && _method->constants() ) 838 // Type::update_loaded_types(_method, _method->constants()); 839 840 // Init alias_type map. 841 if (!_do_escape_analysis && aliaslevel == 3) 842 aliaslevel = 2; // No unique types without escape analysis 843 _AliasLevel = aliaslevel; 844 const int grow_ats = 16; 845 _max_alias_types = grow_ats; 846 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 847 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 848 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 849 { 850 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 851 } 852 // Initialize the first few types. 853 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 854 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 855 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 856 _num_alias_types = AliasIdxRaw+1; 857 // Zero out the alias type cache. 858 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 859 // A NULL adr_type hits in the cache right away. Preload the right answer. 860 probe_alias_cache(NULL)->_index = AliasIdxTop; 861 862 _intrinsics = NULL; 863 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 864 register_library_intrinsics(); 865} 866 867//---------------------------init_start---------------------------------------- 868// Install the StartNode on this compile object. 869void Compile::init_start(StartNode* s) { 870 if (failing()) 871 return; // already failing 872 assert(s == start(), ""); 873} 874 875StartNode* Compile::start() const { 876 assert(!failing(), ""); 877 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 878 Node* start = root()->fast_out(i); 879 if( start->is_Start() ) 880 return start->as_Start(); 881 } 882 ShouldNotReachHere(); 883 return NULL; 884} 885 886//-------------------------------immutable_memory------------------------------------- 887// Access immutable memory 888Node* Compile::immutable_memory() { 889 if (_immutable_memory != NULL) { 890 return _immutable_memory; 891 } 892 StartNode* s = start(); 893 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 894 Node *p = s->fast_out(i); 895 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 896 _immutable_memory = p; 897 return _immutable_memory; 898 } 899 } 900 ShouldNotReachHere(); 901 return NULL; 902} 903 904//----------------------set_cached_top_node------------------------------------ 905// Install the cached top node, and make sure Node::is_top works correctly. 906void Compile::set_cached_top_node(Node* tn) { 907 if (tn != NULL) verify_top(tn); 908 Node* old_top = _top; 909 _top = tn; 910 // Calling Node::setup_is_top allows the nodes the chance to adjust 911 // their _out arrays. 912 if (_top != NULL) _top->setup_is_top(); 913 if (old_top != NULL) old_top->setup_is_top(); 914 assert(_top == NULL || top()->is_top(), ""); 915} 916 917#ifndef PRODUCT 918void Compile::verify_top(Node* tn) const { 919 if (tn != NULL) { 920 assert(tn->is_Con(), "top node must be a constant"); 921 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 922 assert(tn->in(0) != NULL, "must have live top node"); 923 } 924} 925#endif 926 927 928///-------------------Managing Per-Node Debug & Profile Info------------------- 929 930void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 931 guarantee(arr != NULL, ""); 932 int num_blocks = arr->length(); 933 if (grow_by < num_blocks) grow_by = num_blocks; 934 int num_notes = grow_by * _node_notes_block_size; 935 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 936 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 937 while (num_notes > 0) { 938 arr->append(notes); 939 notes += _node_notes_block_size; 940 num_notes -= _node_notes_block_size; 941 } 942 assert(num_notes == 0, "exact multiple, please"); 943} 944 945bool Compile::copy_node_notes_to(Node* dest, Node* source) { 946 if (source == NULL || dest == NULL) return false; 947 948 if (dest->is_Con()) 949 return false; // Do not push debug info onto constants. 950 951#ifdef ASSERT 952 // Leave a bread crumb trail pointing to the original node: 953 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 954 dest->set_debug_orig(source); 955 } 956#endif 957 958 if (node_note_array() == NULL) 959 return false; // Not collecting any notes now. 960 961 // This is a copy onto a pre-existing node, which may already have notes. 962 // If both nodes have notes, do not overwrite any pre-existing notes. 963 Node_Notes* source_notes = node_notes_at(source->_idx); 964 if (source_notes == NULL || source_notes->is_clear()) return false; 965 Node_Notes* dest_notes = node_notes_at(dest->_idx); 966 if (dest_notes == NULL || dest_notes->is_clear()) { 967 return set_node_notes_at(dest->_idx, source_notes); 968 } 969 970 Node_Notes merged_notes = (*source_notes); 971 // The order of operations here ensures that dest notes will win... 972 merged_notes.update_from(dest_notes); 973 return set_node_notes_at(dest->_idx, &merged_notes); 974} 975 976 977//--------------------------allow_range_check_smearing------------------------- 978// Gating condition for coalescing similar range checks. 979// Sometimes we try 'speculatively' replacing a series of a range checks by a 980// single covering check that is at least as strong as any of them. 981// If the optimization succeeds, the simplified (strengthened) range check 982// will always succeed. If it fails, we will deopt, and then give up 983// on the optimization. 984bool Compile::allow_range_check_smearing() const { 985 // If this method has already thrown a range-check, 986 // assume it was because we already tried range smearing 987 // and it failed. 988 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 989 return !already_trapped; 990} 991 992 993//------------------------------flatten_alias_type----------------------------- 994const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 995 int offset = tj->offset(); 996 TypePtr::PTR ptr = tj->ptr(); 997 998 // Process weird unsafe references. 999 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1000 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1001 tj = TypeOopPtr::BOTTOM; 1002 ptr = tj->ptr(); 1003 offset = tj->offset(); 1004 } 1005 1006 // Array pointers need some flattening 1007 const TypeAryPtr *ta = tj->isa_aryptr(); 1008 if( ta && _AliasLevel >= 2 ) { 1009 // For arrays indexed by constant indices, we flatten the alias 1010 // space to include all of the array body. Only the header, klass 1011 // and array length can be accessed un-aliased. 1012 if( offset != Type::OffsetBot ) { 1013 if( ta->const_oop() ) { // methodDataOop or methodOop 1014 offset = Type::OffsetBot; // Flatten constant access into array body 1015 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); 1016 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1017 // range is OK as-is. 1018 tj = ta = TypeAryPtr::RANGE; 1019 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1020 tj = TypeInstPtr::KLASS; // all klass loads look alike 1021 ta = TypeAryPtr::RANGE; // generic ignored junk 1022 ptr = TypePtr::BotPTR; 1023 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1024 tj = TypeInstPtr::MARK; 1025 ta = TypeAryPtr::RANGE; // generic ignored junk 1026 ptr = TypePtr::BotPTR; 1027 } else { // Random constant offset into array body 1028 offset = Type::OffsetBot; // Flatten constant access into array body 1029 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::OffsetBot, ta->instance_id()); 1030 } 1031 } 1032 // Arrays of fixed size alias with arrays of unknown size. 1033 if (ta->size() != TypeInt::POS) { 1034 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1035 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset, ta->instance_id()); 1036 } 1037 // Arrays of known objects become arrays of unknown objects. 1038 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1039 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1040 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id()); 1041 } 1042 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1043 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1044 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset, ta->instance_id()); 1045 } 1046 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1047 // cannot be distinguished by bytecode alone. 1048 if (ta->elem() == TypeInt::BOOL) { 1049 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1050 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1051 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset, ta->instance_id()); 1052 } 1053 // During the 2nd round of IterGVN, NotNull castings are removed. 1054 // Make sure the Bottom and NotNull variants alias the same. 1055 // Also, make sure exact and non-exact variants alias the same. 1056 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1057 if (ta->const_oop()) { 1058 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1059 } else { 1060 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1061 } 1062 } 1063 } 1064 1065 // Oop pointers need some flattening 1066 const TypeInstPtr *to = tj->isa_instptr(); 1067 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1068 if( ptr == TypePtr::Constant ) { 1069 // No constant oop pointers (such as Strings); they alias with 1070 // unknown strings. 1071 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1072 } else if( to->is_instance_field() ) { 1073 tj = to; // Keep NotNull and klass_is_exact for instance type 1074 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1075 // During the 2nd round of IterGVN, NotNull castings are removed. 1076 // Make sure the Bottom and NotNull variants alias the same. 1077 // Also, make sure exact and non-exact variants alias the same. 1078 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset, to->instance_id()); 1079 } 1080 // Canonicalize the holder of this field 1081 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1082 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1083 // First handle header references such as a LoadKlassNode, even if the 1084 // object's klass is unloaded at compile time (4965979). 1085 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset, to->instance_id()); 1086 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1087 to = NULL; 1088 tj = TypeOopPtr::BOTTOM; 1089 offset = tj->offset(); 1090 } else { 1091 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1092 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1093 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset, to->instance_id()); 1094 } 1095 } 1096 } 1097 1098 // Klass pointers to object array klasses need some flattening 1099 const TypeKlassPtr *tk = tj->isa_klassptr(); 1100 if( tk ) { 1101 // If we are referencing a field within a Klass, we need 1102 // to assume the worst case of an Object. Both exact and 1103 // inexact types must flatten to the same alias class. 1104 // Since the flattened result for a klass is defined to be 1105 // precisely java.lang.Object, use a constant ptr. 1106 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1107 1108 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1109 TypeKlassPtr::OBJECT->klass(), 1110 offset); 1111 } 1112 1113 ciKlass* klass = tk->klass(); 1114 if( klass->is_obj_array_klass() ) { 1115 ciKlass* k = TypeAryPtr::OOPS->klass(); 1116 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1117 k = TypeInstPtr::BOTTOM->klass(); 1118 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1119 } 1120 1121 // Check for precise loads from the primary supertype array and force them 1122 // to the supertype cache alias index. Check for generic array loads from 1123 // the primary supertype array and also force them to the supertype cache 1124 // alias index. Since the same load can reach both, we need to merge 1125 // these 2 disparate memories into the same alias class. Since the 1126 // primary supertype array is read-only, there's no chance of confusion 1127 // where we bypass an array load and an array store. 1128 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1129 if( offset == Type::OffsetBot || 1130 off2 < Klass::primary_super_limit()*wordSize ) { 1131 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1132 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1133 } 1134 } 1135 1136 // Flatten all Raw pointers together. 1137 if (tj->base() == Type::RawPtr) 1138 tj = TypeRawPtr::BOTTOM; 1139 1140 if (tj->base() == Type::AnyPtr) 1141 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1142 1143 // Flatten all to bottom for now 1144 switch( _AliasLevel ) { 1145 case 0: 1146 tj = TypePtr::BOTTOM; 1147 break; 1148 case 1: // Flatten to: oop, static, field or array 1149 switch (tj->base()) { 1150 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1151 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1152 case Type::AryPtr: // do not distinguish arrays at all 1153 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1154 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1155 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1156 default: ShouldNotReachHere(); 1157 } 1158 break; 1159 case 2: // No collasping at level 2; keep all splits 1160 case 3: // No collasping at level 3; keep all splits 1161 break; 1162 default: 1163 Unimplemented(); 1164 } 1165 1166 offset = tj->offset(); 1167 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1168 1169 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1170 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1171 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1172 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1173 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1174 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1175 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1176 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1177 assert( tj->ptr() != TypePtr::TopPTR && 1178 tj->ptr() != TypePtr::AnyNull && 1179 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1180// assert( tj->ptr() != TypePtr::Constant || 1181// tj->base() == Type::RawPtr || 1182// tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1183 1184 return tj; 1185} 1186 1187void Compile::AliasType::Init(int i, const TypePtr* at) { 1188 _index = i; 1189 _adr_type = at; 1190 _field = NULL; 1191 _is_rewritable = true; // default 1192 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1193 if (atoop != NULL && atoop->is_instance()) { 1194 const TypeOopPtr *gt = atoop->cast_to_instance(TypeOopPtr::UNKNOWN_INSTANCE); 1195 _general_index = Compile::current()->get_alias_index(gt); 1196 } else { 1197 _general_index = 0; 1198 } 1199} 1200 1201//---------------------------------print_on------------------------------------ 1202#ifndef PRODUCT 1203void Compile::AliasType::print_on(outputStream* st) { 1204 if (index() < 10) 1205 st->print("@ <%d> ", index()); 1206 else st->print("@ <%d>", index()); 1207 st->print(is_rewritable() ? " " : " RO"); 1208 int offset = adr_type()->offset(); 1209 if (offset == Type::OffsetBot) 1210 st->print(" +any"); 1211 else st->print(" +%-3d", offset); 1212 st->print(" in "); 1213 adr_type()->dump_on(st); 1214 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1215 if (field() != NULL && tjp) { 1216 if (tjp->klass() != field()->holder() || 1217 tjp->offset() != field()->offset_in_bytes()) { 1218 st->print(" != "); 1219 field()->print(); 1220 st->print(" ***"); 1221 } 1222 } 1223} 1224 1225void print_alias_types() { 1226 Compile* C = Compile::current(); 1227 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1228 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1229 C->alias_type(idx)->print_on(tty); 1230 tty->cr(); 1231 } 1232} 1233#endif 1234 1235 1236//----------------------------probe_alias_cache-------------------------------- 1237Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1238 intptr_t key = (intptr_t) adr_type; 1239 key ^= key >> logAliasCacheSize; 1240 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1241} 1242 1243 1244//-----------------------------grow_alias_types-------------------------------- 1245void Compile::grow_alias_types() { 1246 const int old_ats = _max_alias_types; // how many before? 1247 const int new_ats = old_ats; // how many more? 1248 const int grow_ats = old_ats+new_ats; // how many now? 1249 _max_alias_types = grow_ats; 1250 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1251 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1252 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1253 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1254} 1255 1256 1257//--------------------------------find_alias_type------------------------------ 1258Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1259 if (_AliasLevel == 0) 1260 return alias_type(AliasIdxBot); 1261 1262 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1263 if (ace->_adr_type == adr_type) { 1264 return alias_type(ace->_index); 1265 } 1266 1267 // Handle special cases. 1268 if (adr_type == NULL) return alias_type(AliasIdxTop); 1269 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1270 1271 // Do it the slow way. 1272 const TypePtr* flat = flatten_alias_type(adr_type); 1273 1274#ifdef ASSERT 1275 assert(flat == flatten_alias_type(flat), "idempotent"); 1276 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1277 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1278 const TypeOopPtr* foop = flat->is_oopptr(); 1279 const TypePtr* xoop = foop->cast_to_exactness(!foop->klass_is_exact())->is_ptr(); 1280 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1281 } 1282 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1283#endif 1284 1285 int idx = AliasIdxTop; 1286 for (int i = 0; i < num_alias_types(); i++) { 1287 if (alias_type(i)->adr_type() == flat) { 1288 idx = i; 1289 break; 1290 } 1291 } 1292 1293 if (idx == AliasIdxTop) { 1294 if (no_create) return NULL; 1295 // Grow the array if necessary. 1296 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1297 // Add a new alias type. 1298 idx = _num_alias_types++; 1299 _alias_types[idx]->Init(idx, flat); 1300 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1301 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1302 if (flat->isa_instptr()) { 1303 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1304 && flat->is_instptr()->klass() == env()->Class_klass()) 1305 alias_type(idx)->set_rewritable(false); 1306 } 1307 if (flat->isa_klassptr()) { 1308 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1309 alias_type(idx)->set_rewritable(false); 1310 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1311 alias_type(idx)->set_rewritable(false); 1312 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1313 alias_type(idx)->set_rewritable(false); 1314 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1315 alias_type(idx)->set_rewritable(false); 1316 } 1317 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1318 // but the base pointer type is not distinctive enough to identify 1319 // references into JavaThread.) 1320 1321 // Check for final instance fields. 1322 const TypeInstPtr* tinst = flat->isa_instptr(); 1323 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1324 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1325 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1326 // Set field() and is_rewritable() attributes. 1327 if (field != NULL) alias_type(idx)->set_field(field); 1328 } 1329 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1330 // Check for final static fields. 1331 if (tklass && tklass->klass()->is_instance_klass()) { 1332 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1333 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1334 // Set field() and is_rewritable() attributes. 1335 if (field != NULL) alias_type(idx)->set_field(field); 1336 } 1337 } 1338 1339 // Fill the cache for next time. 1340 ace->_adr_type = adr_type; 1341 ace->_index = idx; 1342 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1343 1344 // Might as well try to fill the cache for the flattened version, too. 1345 AliasCacheEntry* face = probe_alias_cache(flat); 1346 if (face->_adr_type == NULL) { 1347 face->_adr_type = flat; 1348 face->_index = idx; 1349 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1350 } 1351 1352 return alias_type(idx); 1353} 1354 1355 1356Compile::AliasType* Compile::alias_type(ciField* field) { 1357 const TypeOopPtr* t; 1358 if (field->is_static()) 1359 t = TypeKlassPtr::make(field->holder()); 1360 else 1361 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1362 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1363 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1364 return atp; 1365} 1366 1367 1368//------------------------------have_alias_type-------------------------------- 1369bool Compile::have_alias_type(const TypePtr* adr_type) { 1370 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1371 if (ace->_adr_type == adr_type) { 1372 return true; 1373 } 1374 1375 // Handle special cases. 1376 if (adr_type == NULL) return true; 1377 if (adr_type == TypePtr::BOTTOM) return true; 1378 1379 return find_alias_type(adr_type, true) != NULL; 1380} 1381 1382//-----------------------------must_alias-------------------------------------- 1383// True if all values of the given address type are in the given alias category. 1384bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1385 if (alias_idx == AliasIdxBot) return true; // the universal category 1386 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1387 if (alias_idx == AliasIdxTop) return false; // the empty category 1388 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1389 1390 // the only remaining possible overlap is identity 1391 int adr_idx = get_alias_index(adr_type); 1392 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1393 assert(adr_idx == alias_idx || 1394 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1395 && adr_type != TypeOopPtr::BOTTOM), 1396 "should not be testing for overlap with an unsafe pointer"); 1397 return adr_idx == alias_idx; 1398} 1399 1400//------------------------------can_alias-------------------------------------- 1401// True if any values of the given address type are in the given alias category. 1402bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1403 if (alias_idx == AliasIdxTop) return false; // the empty category 1404 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1405 if (alias_idx == AliasIdxBot) return true; // the universal category 1406 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1407 1408 // the only remaining possible overlap is identity 1409 int adr_idx = get_alias_index(adr_type); 1410 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1411 return adr_idx == alias_idx; 1412} 1413 1414 1415 1416//---------------------------pop_warm_call------------------------------------- 1417WarmCallInfo* Compile::pop_warm_call() { 1418 WarmCallInfo* wci = _warm_calls; 1419 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1420 return wci; 1421} 1422 1423//----------------------------Inline_Warm-------------------------------------- 1424int Compile::Inline_Warm() { 1425 // If there is room, try to inline some more warm call sites. 1426 // %%% Do a graph index compaction pass when we think we're out of space? 1427 if (!InlineWarmCalls) return 0; 1428 1429 int calls_made_hot = 0; 1430 int room_to_grow = NodeCountInliningCutoff - unique(); 1431 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1432 int amount_grown = 0; 1433 WarmCallInfo* call; 1434 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1435 int est_size = (int)call->size(); 1436 if (est_size > (room_to_grow - amount_grown)) { 1437 // This one won't fit anyway. Get rid of it. 1438 call->make_cold(); 1439 continue; 1440 } 1441 call->make_hot(); 1442 calls_made_hot++; 1443 amount_grown += est_size; 1444 amount_to_grow -= est_size; 1445 } 1446 1447 if (calls_made_hot > 0) set_major_progress(); 1448 return calls_made_hot; 1449} 1450 1451 1452//----------------------------Finish_Warm-------------------------------------- 1453void Compile::Finish_Warm() { 1454 if (!InlineWarmCalls) return; 1455 if (failing()) return; 1456 if (warm_calls() == NULL) return; 1457 1458 // Clean up loose ends, if we are out of space for inlining. 1459 WarmCallInfo* call; 1460 while ((call = pop_warm_call()) != NULL) { 1461 call->make_cold(); 1462 } 1463} 1464 1465 1466//------------------------------Optimize--------------------------------------- 1467// Given a graph, optimize it. 1468void Compile::Optimize() { 1469 TracePhase t1("optimizer", &_t_optimizer, true); 1470 1471#ifndef PRODUCT 1472 if (env()->break_at_compile()) { 1473 BREAKPOINT; 1474 } 1475 1476#endif 1477 1478 ResourceMark rm; 1479 int loop_opts_cnt; 1480 1481 NOT_PRODUCT( verify_graph_edges(); ) 1482 1483 print_method("After Parsing"); 1484 1485 { 1486 // Iterative Global Value Numbering, including ideal transforms 1487 // Initialize IterGVN with types and values from parse-time GVN 1488 PhaseIterGVN igvn(initial_gvn()); 1489 { 1490 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1491 igvn.optimize(); 1492 } 1493 1494 print_method("Iter GVN 1", 2); 1495 1496 if (failing()) return; 1497 1498 // get rid of the connection graph since it's information is not 1499 // updated by optimizations 1500 _congraph = NULL; 1501 1502 1503 // Loop transforms on the ideal graph. Range Check Elimination, 1504 // peeling, unrolling, etc. 1505 1506 // Set loop opts counter 1507 loop_opts_cnt = num_loop_opts(); 1508 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1509 { 1510 TracePhase t2("idealLoop", &_t_idealLoop, true); 1511 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1512 loop_opts_cnt--; 1513 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1514 if (failing()) return; 1515 } 1516 // Loop opts pass if partial peeling occurred in previous pass 1517 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1518 TracePhase t3("idealLoop", &_t_idealLoop, true); 1519 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1520 loop_opts_cnt--; 1521 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1522 if (failing()) return; 1523 } 1524 // Loop opts pass for loop-unrolling before CCP 1525 if(major_progress() && (loop_opts_cnt > 0)) { 1526 TracePhase t4("idealLoop", &_t_idealLoop, true); 1527 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1528 loop_opts_cnt--; 1529 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1530 } 1531 } 1532 if (failing()) return; 1533 1534 // Conditional Constant Propagation; 1535 PhaseCCP ccp( &igvn ); 1536 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1537 { 1538 TracePhase t2("ccp", &_t_ccp, true); 1539 ccp.do_transform(); 1540 } 1541 print_method("PhaseCPP 1", 2); 1542 1543 assert( true, "Break here to ccp.dump_old2new_map()"); 1544 1545 // Iterative Global Value Numbering, including ideal transforms 1546 { 1547 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1548 igvn = ccp; 1549 igvn.optimize(); 1550 } 1551 1552 print_method("Iter GVN 2", 2); 1553 1554 if (failing()) return; 1555 1556 // Loop transforms on the ideal graph. Range Check Elimination, 1557 // peeling, unrolling, etc. 1558 if(loop_opts_cnt > 0) { 1559 debug_only( int cnt = 0; ); 1560 while(major_progress() && (loop_opts_cnt > 0)) { 1561 TracePhase t2("idealLoop", &_t_idealLoop, true); 1562 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1563 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1564 loop_opts_cnt--; 1565 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1566 if (failing()) return; 1567 } 1568 } 1569 { 1570 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1571 PhaseMacroExpand mex(igvn); 1572 if (mex.expand_macro_nodes()) { 1573 assert(failing(), "must bail out w/ explicit message"); 1574 return; 1575 } 1576 } 1577 1578 } // (End scope of igvn; run destructor if necessary for asserts.) 1579 1580 // A method with only infinite loops has no edges entering loops from root 1581 { 1582 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1583 if (final_graph_reshaping()) { 1584 assert(failing(), "must bail out w/ explicit message"); 1585 return; 1586 } 1587 } 1588 1589 print_method("Optimize finished", 2); 1590} 1591 1592 1593//------------------------------Code_Gen--------------------------------------- 1594// Given a graph, generate code for it 1595void Compile::Code_Gen() { 1596 if (failing()) return; 1597 1598 // Perform instruction selection. You might think we could reclaim Matcher 1599 // memory PDQ, but actually the Matcher is used in generating spill code. 1600 // Internals of the Matcher (including some VectorSets) must remain live 1601 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1602 // set a bit in reclaimed memory. 1603 1604 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1605 // nodes. Mapping is only valid at the root of each matched subtree. 1606 NOT_PRODUCT( verify_graph_edges(); ) 1607 1608 Node_List proj_list; 1609 Matcher m(proj_list); 1610 _matcher = &m; 1611 { 1612 TracePhase t2("matcher", &_t_matcher, true); 1613 m.match(); 1614 } 1615 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1616 // nodes. Mapping is only valid at the root of each matched subtree. 1617 NOT_PRODUCT( verify_graph_edges(); ) 1618 1619 // If you have too many nodes, or if matching has failed, bail out 1620 check_node_count(0, "out of nodes matching instructions"); 1621 if (failing()) return; 1622 1623 // Build a proper-looking CFG 1624 PhaseCFG cfg(node_arena(), root(), m); 1625 _cfg = &cfg; 1626 { 1627 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1628 cfg.Dominators(); 1629 if (failing()) return; 1630 1631 NOT_PRODUCT( verify_graph_edges(); ) 1632 1633 cfg.Estimate_Block_Frequency(); 1634 cfg.GlobalCodeMotion(m,unique(),proj_list); 1635 1636 print_method("Global code motion", 2); 1637 1638 if (failing()) return; 1639 NOT_PRODUCT( verify_graph_edges(); ) 1640 1641 debug_only( cfg.verify(); ) 1642 } 1643 NOT_PRODUCT( verify_graph_edges(); ) 1644 1645 PhaseChaitin regalloc(unique(),cfg,m); 1646 _regalloc = ®alloc; 1647 { 1648 TracePhase t2("regalloc", &_t_registerAllocation, true); 1649 // Perform any platform dependent preallocation actions. This is used, 1650 // for example, to avoid taking an implicit null pointer exception 1651 // using the frame pointer on win95. 1652 _regalloc->pd_preallocate_hook(); 1653 1654 // Perform register allocation. After Chaitin, use-def chains are 1655 // no longer accurate (at spill code) and so must be ignored. 1656 // Node->LRG->reg mappings are still accurate. 1657 _regalloc->Register_Allocate(); 1658 1659 // Bail out if the allocator builds too many nodes 1660 if (failing()) return; 1661 } 1662 1663 // Prior to register allocation we kept empty basic blocks in case the 1664 // the allocator needed a place to spill. After register allocation we 1665 // are not adding any new instructions. If any basic block is empty, we 1666 // can now safely remove it. 1667 { 1668 NOT_PRODUCT( TracePhase t2("removeEmpty", &_t_removeEmptyBlocks, TimeCompiler); ) 1669 cfg.RemoveEmpty(); 1670 } 1671 1672 // Perform any platform dependent postallocation verifications. 1673 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1674 1675 // Apply peephole optimizations 1676 if( OptoPeephole ) { 1677 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1678 PhasePeephole peep( _regalloc, cfg); 1679 peep.do_transform(); 1680 } 1681 1682 // Convert Nodes to instruction bits in a buffer 1683 { 1684 // %%%% workspace merge brought two timers together for one job 1685 TracePhase t2a("output", &_t_output, true); 1686 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1687 Output(); 1688 } 1689 1690 print_method("Final Code"); 1691 1692 // He's dead, Jim. 1693 _cfg = (PhaseCFG*)0xdeadbeef; 1694 _regalloc = (PhaseChaitin*)0xdeadbeef; 1695} 1696 1697 1698//------------------------------dump_asm--------------------------------------- 1699// Dump formatted assembly 1700#ifndef PRODUCT 1701void Compile::dump_asm(int *pcs, uint pc_limit) { 1702 bool cut_short = false; 1703 tty->print_cr("#"); 1704 tty->print("# "); _tf->dump(); tty->cr(); 1705 tty->print_cr("#"); 1706 1707 // For all blocks 1708 int pc = 0x0; // Program counter 1709 char starts_bundle = ' '; 1710 _regalloc->dump_frame(); 1711 1712 Node *n = NULL; 1713 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1714 if (VMThread::should_terminate()) { cut_short = true; break; } 1715 Block *b = _cfg->_blocks[i]; 1716 if (b->is_connector() && !Verbose) continue; 1717 n = b->_nodes[0]; 1718 if (pcs && n->_idx < pc_limit) 1719 tty->print("%3.3x ", pcs[n->_idx]); 1720 else 1721 tty->print(" "); 1722 b->dump_head( &_cfg->_bbs ); 1723 if (b->is_connector()) { 1724 tty->print_cr(" # Empty connector block"); 1725 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1726 tty->print_cr(" # Block is sole successor of call"); 1727 } 1728 1729 // For all instructions 1730 Node *delay = NULL; 1731 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1732 if (VMThread::should_terminate()) { cut_short = true; break; } 1733 n = b->_nodes[j]; 1734 if (valid_bundle_info(n)) { 1735 Bundle *bundle = node_bundling(n); 1736 if (bundle->used_in_unconditional_delay()) { 1737 delay = n; 1738 continue; 1739 } 1740 if (bundle->starts_bundle()) 1741 starts_bundle = '+'; 1742 } 1743 1744 if (WizardMode) n->dump(); 1745 1746 if( !n->is_Region() && // Dont print in the Assembly 1747 !n->is_Phi() && // a few noisely useless nodes 1748 !n->is_Proj() && 1749 !n->is_MachTemp() && 1750 !n->is_Catch() && // Would be nice to print exception table targets 1751 !n->is_MergeMem() && // Not very interesting 1752 !n->is_top() && // Debug info table constants 1753 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1754 ) { 1755 if (pcs && n->_idx < pc_limit) 1756 tty->print("%3.3x", pcs[n->_idx]); 1757 else 1758 tty->print(" "); 1759 tty->print(" %c ", starts_bundle); 1760 starts_bundle = ' '; 1761 tty->print("\t"); 1762 n->format(_regalloc, tty); 1763 tty->cr(); 1764 } 1765 1766 // If we have an instruction with a delay slot, and have seen a delay, 1767 // then back up and print it 1768 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1769 assert(delay != NULL, "no unconditional delay instruction"); 1770 if (WizardMode) delay->dump(); 1771 1772 if (node_bundling(delay)->starts_bundle()) 1773 starts_bundle = '+'; 1774 if (pcs && n->_idx < pc_limit) 1775 tty->print("%3.3x", pcs[n->_idx]); 1776 else 1777 tty->print(" "); 1778 tty->print(" %c ", starts_bundle); 1779 starts_bundle = ' '; 1780 tty->print("\t"); 1781 delay->format(_regalloc, tty); 1782 tty->print_cr(""); 1783 delay = NULL; 1784 } 1785 1786 // Dump the exception table as well 1787 if( n->is_Catch() && (Verbose || WizardMode) ) { 1788 // Print the exception table for this offset 1789 _handler_table.print_subtable_for(pc); 1790 } 1791 } 1792 1793 if (pcs && n->_idx < pc_limit) 1794 tty->print_cr("%3.3x", pcs[n->_idx]); 1795 else 1796 tty->print_cr(""); 1797 1798 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1799 1800 } // End of per-block dump 1801 tty->print_cr(""); 1802 1803 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1804} 1805#endif 1806 1807//------------------------------Final_Reshape_Counts--------------------------- 1808// This class defines counters to help identify when a method 1809// may/must be executed using hardware with only 24-bit precision. 1810struct Final_Reshape_Counts : public StackObj { 1811 int _call_count; // count non-inlined 'common' calls 1812 int _float_count; // count float ops requiring 24-bit precision 1813 int _double_count; // count double ops requiring more precision 1814 int _java_call_count; // count non-inlined 'java' calls 1815 VectorSet _visited; // Visitation flags 1816 Node_List _tests; // Set of IfNodes & PCTableNodes 1817 1818 Final_Reshape_Counts() : 1819 _call_count(0), _float_count(0), _double_count(0), _java_call_count(0), 1820 _visited( Thread::current()->resource_area() ) { } 1821 1822 void inc_call_count () { _call_count ++; } 1823 void inc_float_count () { _float_count ++; } 1824 void inc_double_count() { _double_count++; } 1825 void inc_java_call_count() { _java_call_count++; } 1826 1827 int get_call_count () const { return _call_count ; } 1828 int get_float_count () const { return _float_count ; } 1829 int get_double_count() const { return _double_count; } 1830 int get_java_call_count() const { return _java_call_count; } 1831}; 1832 1833static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1834 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1835 // Make sure the offset goes inside the instance layout. 1836 return k->contains_field_offset(tp->offset()); 1837 // Note that OffsetBot and OffsetTop are very negative. 1838} 1839 1840//------------------------------final_graph_reshaping_impl---------------------- 1841// Implement items 1-5 from final_graph_reshaping below. 1842static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &fpu ) { 1843 1844 if ( n->outcnt() == 0 ) return; // dead node 1845 uint nop = n->Opcode(); 1846 1847 // Check for 2-input instruction with "last use" on right input. 1848 // Swap to left input. Implements item (2). 1849 if( n->req() == 3 && // two-input instruction 1850 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1851 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1852 n->in(2)->outcnt() == 1 &&// right use IS a last use 1853 !n->in(2)->is_Con() ) { // right use is not a constant 1854 // Check for commutative opcode 1855 switch( nop ) { 1856 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1857 case Op_MaxI: case Op_MinI: 1858 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1859 case Op_AndL: case Op_XorL: case Op_OrL: 1860 case Op_AndI: case Op_XorI: case Op_OrI: { 1861 // Move "last use" input to left by swapping inputs 1862 n->swap_edges(1, 2); 1863 break; 1864 } 1865 default: 1866 break; 1867 } 1868 } 1869 1870 // Count FPU ops and common calls, implements item (3) 1871 switch( nop ) { 1872 // Count all float operations that may use FPU 1873 case Op_AddF: 1874 case Op_SubF: 1875 case Op_MulF: 1876 case Op_DivF: 1877 case Op_NegF: 1878 case Op_ModF: 1879 case Op_ConvI2F: 1880 case Op_ConF: 1881 case Op_CmpF: 1882 case Op_CmpF3: 1883 // case Op_ConvL2F: // longs are split into 32-bit halves 1884 fpu.inc_float_count(); 1885 break; 1886 1887 case Op_ConvF2D: 1888 case Op_ConvD2F: 1889 fpu.inc_float_count(); 1890 fpu.inc_double_count(); 1891 break; 1892 1893 // Count all double operations that may use FPU 1894 case Op_AddD: 1895 case Op_SubD: 1896 case Op_MulD: 1897 case Op_DivD: 1898 case Op_NegD: 1899 case Op_ModD: 1900 case Op_ConvI2D: 1901 case Op_ConvD2I: 1902 // case Op_ConvL2D: // handled by leaf call 1903 // case Op_ConvD2L: // handled by leaf call 1904 case Op_ConD: 1905 case Op_CmpD: 1906 case Op_CmpD3: 1907 fpu.inc_double_count(); 1908 break; 1909 case Op_Opaque1: // Remove Opaque Nodes before matching 1910 case Op_Opaque2: // Remove Opaque Nodes before matching 1911 n->subsume_by(n->in(1)); 1912 break; 1913 case Op_CallStaticJava: 1914 case Op_CallJava: 1915 case Op_CallDynamicJava: 1916 fpu.inc_java_call_count(); // Count java call site; 1917 case Op_CallRuntime: 1918 case Op_CallLeaf: 1919 case Op_CallLeafNoFP: { 1920 assert( n->is_Call(), "" ); 1921 CallNode *call = n->as_Call(); 1922 // Count call sites where the FP mode bit would have to be flipped. 1923 // Do not count uncommon runtime calls: 1924 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 1925 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 1926 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 1927 fpu.inc_call_count(); // Count the call site 1928 } else { // See if uncommon argument is shared 1929 Node *n = call->in(TypeFunc::Parms); 1930 int nop = n->Opcode(); 1931 // Clone shared simple arguments to uncommon calls, item (1). 1932 if( n->outcnt() > 1 && 1933 !n->is_Proj() && 1934 nop != Op_CreateEx && 1935 nop != Op_CheckCastPP && 1936 !n->is_Mem() ) { 1937 Node *x = n->clone(); 1938 call->set_req( TypeFunc::Parms, x ); 1939 } 1940 } 1941 break; 1942 } 1943 1944 case Op_StoreD: 1945 case Op_LoadD: 1946 case Op_LoadD_unaligned: 1947 fpu.inc_double_count(); 1948 goto handle_mem; 1949 case Op_StoreF: 1950 case Op_LoadF: 1951 fpu.inc_float_count(); 1952 goto handle_mem; 1953 1954 case Op_StoreB: 1955 case Op_StoreC: 1956 case Op_StoreCM: 1957 case Op_StorePConditional: 1958 case Op_StoreI: 1959 case Op_StoreL: 1960 case Op_StoreLConditional: 1961 case Op_CompareAndSwapI: 1962 case Op_CompareAndSwapL: 1963 case Op_CompareAndSwapP: 1964 case Op_CompareAndSwapN: 1965 case Op_StoreP: 1966 case Op_StoreN: 1967 case Op_LoadB: 1968 case Op_LoadC: 1969 case Op_LoadI: 1970 case Op_LoadKlass: 1971 case Op_LoadNKlass: 1972 case Op_LoadL: 1973 case Op_LoadL_unaligned: 1974 case Op_LoadPLocked: 1975 case Op_LoadLLocked: 1976 case Op_LoadP: 1977 case Op_LoadN: 1978 case Op_LoadRange: 1979 case Op_LoadS: { 1980 handle_mem: 1981#ifdef ASSERT 1982 if( VerifyOptoOopOffsets ) { 1983 assert( n->is_Mem(), "" ); 1984 MemNode *mem = (MemNode*)n; 1985 // Check to see if address types have grounded out somehow. 1986 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 1987 assert( !tp || oop_offset_is_sane(tp), "" ); 1988 } 1989#endif 1990 break; 1991 } 1992 1993 case Op_AddP: { // Assert sane base pointers 1994 Node *addp = n->in(AddPNode::Address); 1995 assert( !addp->is_AddP() || 1996 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 1997 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 1998 "Base pointers must match" ); 1999#ifdef _LP64 2000 if (UseCompressedOops && 2001 addp->Opcode() == Op_ConP && 2002 addp == n->in(AddPNode::Base) && 2003 n->in(AddPNode::Offset)->is_Con()) { 2004 // Use addressing with narrow klass to load with offset on x86. 2005 // On sparc loading 32-bits constant and decoding it have less 2006 // instructions (4) then load 64-bits constant (7). 2007 // Do this transformation here since IGVN will convert ConN back to ConP. 2008 const Type* t = addp->bottom_type(); 2009 if (t->isa_oopptr()) { 2010 Node* nn = NULL; 2011 2012 // Look for existing ConN node of the same exact type. 2013 Compile* C = Compile::current(); 2014 Node* r = C->root(); 2015 uint cnt = r->outcnt(); 2016 for (uint i = 0; i < cnt; i++) { 2017 Node* m = r->raw_out(i); 2018 if (m!= NULL && m->Opcode() == Op_ConN && 2019 m->bottom_type()->make_ptr() == t) { 2020 nn = m; 2021 break; 2022 } 2023 } 2024 if (nn != NULL) { 2025 // Decode a narrow oop to match address 2026 // [R12 + narrow_oop_reg<<3 + offset] 2027 nn = new (C, 2) DecodeNNode(nn, t); 2028 n->set_req(AddPNode::Base, nn); 2029 n->set_req(AddPNode::Address, nn); 2030 if (addp->outcnt() == 0) { 2031 addp->disconnect_inputs(NULL); 2032 } 2033 } 2034 } 2035 } 2036#endif 2037 break; 2038 } 2039 2040#ifdef _LP64 2041 case Op_CmpP: 2042 // Do this transformation here to preserve CmpPNode::sub() and 2043 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2044 if( n->in(1)->is_DecodeN() ) { 2045 Compile* C = Compile::current(); 2046 Node* in2 = NULL; 2047 if( n->in(2)->is_DecodeN() ) { 2048 in2 = n->in(2)->in(1); 2049 } else if ( n->in(2)->Opcode() == Op_ConP ) { 2050 const Type* t = n->in(2)->bottom_type(); 2051 if (t == TypePtr::NULL_PTR) { 2052 Node *in1 = n->in(1); 2053 if (Matcher::clone_shift_expressions) { 2054 // x86, ARM and friends can handle 2 adds in addressing mode. 2055 // Decode a narrow oop and do implicit NULL check in address 2056 // [R12 + narrow_oop_reg<<3 + offset] 2057 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2058 } else { 2059 // Don't replace CmpP(o ,null) if 'o' is used in AddP 2060 // to generate implicit NULL check on Sparc where 2061 // narrow oops can't be used in address. 2062 uint i = 0; 2063 for (; i < in1->outcnt(); i++) { 2064 if (in1->raw_out(i)->is_AddP()) 2065 break; 2066 } 2067 if (i >= in1->outcnt()) { 2068 in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2069 } 2070 } 2071 } else if (t->isa_oopptr()) { 2072 in2 = ConNode::make(C, t->make_narrowoop()); 2073 } 2074 } 2075 if( in2 != NULL ) { 2076 Node* cmpN = new (C, 3) CmpNNode(n->in(1)->in(1), in2); 2077 n->subsume_by( cmpN ); 2078 } 2079 } 2080#endif 2081 2082 case Op_ModI: 2083 if (UseDivMod) { 2084 // Check if a%b and a/b both exist 2085 Node* d = n->find_similar(Op_DivI); 2086 if (d) { 2087 // Replace them with a fused divmod if supported 2088 Compile* C = Compile::current(); 2089 if (Matcher::has_match_rule(Op_DivModI)) { 2090 DivModINode* divmod = DivModINode::make(C, n); 2091 d->subsume_by(divmod->div_proj()); 2092 n->subsume_by(divmod->mod_proj()); 2093 } else { 2094 // replace a%b with a-((a/b)*b) 2095 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2096 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2097 n->subsume_by( sub ); 2098 } 2099 } 2100 } 2101 break; 2102 2103 case Op_ModL: 2104 if (UseDivMod) { 2105 // Check if a%b and a/b both exist 2106 Node* d = n->find_similar(Op_DivL); 2107 if (d) { 2108 // Replace them with a fused divmod if supported 2109 Compile* C = Compile::current(); 2110 if (Matcher::has_match_rule(Op_DivModL)) { 2111 DivModLNode* divmod = DivModLNode::make(C, n); 2112 d->subsume_by(divmod->div_proj()); 2113 n->subsume_by(divmod->mod_proj()); 2114 } else { 2115 // replace a%b with a-((a/b)*b) 2116 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2117 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2118 n->subsume_by( sub ); 2119 } 2120 } 2121 } 2122 break; 2123 2124 case Op_Load16B: 2125 case Op_Load8B: 2126 case Op_Load4B: 2127 case Op_Load8S: 2128 case Op_Load4S: 2129 case Op_Load2S: 2130 case Op_Load8C: 2131 case Op_Load4C: 2132 case Op_Load2C: 2133 case Op_Load4I: 2134 case Op_Load2I: 2135 case Op_Load2L: 2136 case Op_Load4F: 2137 case Op_Load2F: 2138 case Op_Load2D: 2139 case Op_Store16B: 2140 case Op_Store8B: 2141 case Op_Store4B: 2142 case Op_Store8C: 2143 case Op_Store4C: 2144 case Op_Store2C: 2145 case Op_Store4I: 2146 case Op_Store2I: 2147 case Op_Store2L: 2148 case Op_Store4F: 2149 case Op_Store2F: 2150 case Op_Store2D: 2151 break; 2152 2153 case Op_PackB: 2154 case Op_PackS: 2155 case Op_PackC: 2156 case Op_PackI: 2157 case Op_PackF: 2158 case Op_PackL: 2159 case Op_PackD: 2160 if (n->req()-1 > 2) { 2161 // Replace many operand PackNodes with a binary tree for matching 2162 PackNode* p = (PackNode*) n; 2163 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2164 n->subsume_by(btp); 2165 } 2166 break; 2167 default: 2168 assert( !n->is_Call(), "" ); 2169 assert( !n->is_Mem(), "" ); 2170 break; 2171 } 2172 2173 // Collect CFG split points 2174 if (n->is_MultiBranch()) 2175 fpu._tests.push(n); 2176} 2177 2178//------------------------------final_graph_reshaping_walk--------------------- 2179// Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2180// requires that the walk visits a node's inputs before visiting the node. 2181static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &fpu ) { 2182 fpu._visited.set(root->_idx); // first, mark node as visited 2183 uint cnt = root->req(); 2184 Node *n = root; 2185 uint i = 0; 2186 while (true) { 2187 if (i < cnt) { 2188 // Place all non-visited non-null inputs onto stack 2189 Node* m = n->in(i); 2190 ++i; 2191 if (m != NULL && !fpu._visited.test_set(m->_idx)) { 2192 cnt = m->req(); 2193 nstack.push(n, i); // put on stack parent and next input's index 2194 n = m; 2195 i = 0; 2196 } 2197 } else { 2198 // Now do post-visit work 2199 final_graph_reshaping_impl( n, fpu ); 2200 if (nstack.is_empty()) 2201 break; // finished 2202 n = nstack.node(); // Get node from stack 2203 cnt = n->req(); 2204 i = nstack.index(); 2205 nstack.pop(); // Shift to the next node on stack 2206 } 2207 } 2208} 2209 2210//------------------------------final_graph_reshaping-------------------------- 2211// Final Graph Reshaping. 2212// 2213// (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2214// and not commoned up and forced early. Must come after regular 2215// optimizations to avoid GVN undoing the cloning. Clone constant 2216// inputs to Loop Phis; these will be split by the allocator anyways. 2217// Remove Opaque nodes. 2218// (2) Move last-uses by commutative operations to the left input to encourage 2219// Intel update-in-place two-address operations and better register usage 2220// on RISCs. Must come after regular optimizations to avoid GVN Ideal 2221// calls canonicalizing them back. 2222// (3) Count the number of double-precision FP ops, single-precision FP ops 2223// and call sites. On Intel, we can get correct rounding either by 2224// forcing singles to memory (requires extra stores and loads after each 2225// FP bytecode) or we can set a rounding mode bit (requires setting and 2226// clearing the mode bit around call sites). The mode bit is only used 2227// if the relative frequency of single FP ops to calls is low enough. 2228// This is a key transform for SPEC mpeg_audio. 2229// (4) Detect infinite loops; blobs of code reachable from above but not 2230// below. Several of the Code_Gen algorithms fail on such code shapes, 2231// so we simply bail out. Happens a lot in ZKM.jar, but also happens 2232// from time to time in other codes (such as -Xcomp finalizer loops, etc). 2233// Detection is by looking for IfNodes where only 1 projection is 2234// reachable from below or CatchNodes missing some targets. 2235// (5) Assert for insane oop offsets in debug mode. 2236 2237bool Compile::final_graph_reshaping() { 2238 // an infinite loop may have been eliminated by the optimizer, 2239 // in which case the graph will be empty. 2240 if (root()->req() == 1) { 2241 record_method_not_compilable("trivial infinite loop"); 2242 return true; 2243 } 2244 2245 Final_Reshape_Counts fpu; 2246 2247 // Visit everybody reachable! 2248 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2249 Node_Stack nstack(unique() >> 1); 2250 final_graph_reshaping_walk(nstack, root(), fpu); 2251 2252 // Check for unreachable (from below) code (i.e., infinite loops). 2253 for( uint i = 0; i < fpu._tests.size(); i++ ) { 2254 MultiBranchNode *n = fpu._tests[i]->as_MultiBranch(); 2255 // Get number of CFG targets. 2256 // Note that PCTables include exception targets after calls. 2257 uint required_outcnt = n->required_outcnt(); 2258 if (n->outcnt() != required_outcnt) { 2259 // Check for a few special cases. Rethrow Nodes never take the 2260 // 'fall-thru' path, so expected kids is 1 less. 2261 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2262 if (n->in(0)->in(0)->is_Call()) { 2263 CallNode *call = n->in(0)->in(0)->as_Call(); 2264 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2265 required_outcnt--; // Rethrow always has 1 less kid 2266 } else if (call->req() > TypeFunc::Parms && 2267 call->is_CallDynamicJava()) { 2268 // Check for null receiver. In such case, the optimizer has 2269 // detected that the virtual call will always result in a null 2270 // pointer exception. The fall-through projection of this CatchNode 2271 // will not be populated. 2272 Node *arg0 = call->in(TypeFunc::Parms); 2273 if (arg0->is_Type() && 2274 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2275 required_outcnt--; 2276 } 2277 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2278 call->req() > TypeFunc::Parms+1 && 2279 call->is_CallStaticJava()) { 2280 // Check for negative array length. In such case, the optimizer has 2281 // detected that the allocation attempt will always result in an 2282 // exception. There is no fall-through projection of this CatchNode . 2283 Node *arg1 = call->in(TypeFunc::Parms+1); 2284 if (arg1->is_Type() && 2285 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2286 required_outcnt--; 2287 } 2288 } 2289 } 2290 } 2291 // Recheck with a better notion of 'required_outcnt' 2292 if (n->outcnt() != required_outcnt) { 2293 record_method_not_compilable("malformed control flow"); 2294 return true; // Not all targets reachable! 2295 } 2296 } 2297 // Check that I actually visited all kids. Unreached kids 2298 // must be infinite loops. 2299 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2300 if (!fpu._visited.test(n->fast_out(j)->_idx)) { 2301 record_method_not_compilable("infinite loop"); 2302 return true; // Found unvisited kid; must be unreach 2303 } 2304 } 2305 2306 // If original bytecodes contained a mixture of floats and doubles 2307 // check if the optimizer has made it homogenous, item (3). 2308 if( Use24BitFPMode && Use24BitFP && 2309 fpu.get_float_count() > 32 && 2310 fpu.get_double_count() == 0 && 2311 (10 * fpu.get_call_count() < fpu.get_float_count()) ) { 2312 set_24_bit_selection_and_mode( false, true ); 2313 } 2314 2315 set_has_java_calls(fpu.get_java_call_count() > 0); 2316 2317 // No infinite loops, no reason to bail out. 2318 return false; 2319} 2320 2321//-----------------------------too_many_traps---------------------------------- 2322// Report if there are too many traps at the current method and bci. 2323// Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2324bool Compile::too_many_traps(ciMethod* method, 2325 int bci, 2326 Deoptimization::DeoptReason reason) { 2327 ciMethodData* md = method->method_data(); 2328 if (md->is_empty()) { 2329 // Assume the trap has not occurred, or that it occurred only 2330 // because of a transient condition during start-up in the interpreter. 2331 return false; 2332 } 2333 if (md->has_trap_at(bci, reason) != 0) { 2334 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2335 // Also, if there are multiple reasons, or if there is no per-BCI record, 2336 // assume the worst. 2337 if (log()) 2338 log()->elem("observe trap='%s' count='%d'", 2339 Deoptimization::trap_reason_name(reason), 2340 md->trap_count(reason)); 2341 return true; 2342 } else { 2343 // Ignore method/bci and see if there have been too many globally. 2344 return too_many_traps(reason, md); 2345 } 2346} 2347 2348// Less-accurate variant which does not require a method and bci. 2349bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2350 ciMethodData* logmd) { 2351 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2352 // Too many traps globally. 2353 // Note that we use cumulative trap_count, not just md->trap_count. 2354 if (log()) { 2355 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2356 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2357 Deoptimization::trap_reason_name(reason), 2358 mcount, trap_count(reason)); 2359 } 2360 return true; 2361 } else { 2362 // The coast is clear. 2363 return false; 2364 } 2365} 2366 2367//--------------------------too_many_recompiles-------------------------------- 2368// Report if there are too many recompiles at the current method and bci. 2369// Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2370// Is not eager to return true, since this will cause the compiler to use 2371// Action_none for a trap point, to avoid too many recompilations. 2372bool Compile::too_many_recompiles(ciMethod* method, 2373 int bci, 2374 Deoptimization::DeoptReason reason) { 2375 ciMethodData* md = method->method_data(); 2376 if (md->is_empty()) { 2377 // Assume the trap has not occurred, or that it occurred only 2378 // because of a transient condition during start-up in the interpreter. 2379 return false; 2380 } 2381 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2382 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2383 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2384 Deoptimization::DeoptReason per_bc_reason 2385 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2386 if ((per_bc_reason == Deoptimization::Reason_none 2387 || md->has_trap_at(bci, reason) != 0) 2388 // The trap frequency measure we care about is the recompile count: 2389 && md->trap_recompiled_at(bci) 2390 && md->overflow_recompile_count() >= bc_cutoff) { 2391 // Do not emit a trap here if it has already caused recompilations. 2392 // Also, if there are multiple reasons, or if there is no per-BCI record, 2393 // assume the worst. 2394 if (log()) 2395 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2396 Deoptimization::trap_reason_name(reason), 2397 md->trap_count(reason), 2398 md->overflow_recompile_count()); 2399 return true; 2400 } else if (trap_count(reason) != 0 2401 && decompile_count() >= m_cutoff) { 2402 // Too many recompiles globally, and we have seen this sort of trap. 2403 // Use cumulative decompile_count, not just md->decompile_count. 2404 if (log()) 2405 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2406 Deoptimization::trap_reason_name(reason), 2407 md->trap_count(reason), trap_count(reason), 2408 md->decompile_count(), decompile_count()); 2409 return true; 2410 } else { 2411 // The coast is clear. 2412 return false; 2413 } 2414} 2415 2416 2417#ifndef PRODUCT 2418//------------------------------verify_graph_edges--------------------------- 2419// Walk the Graph and verify that there is a one-to-one correspondence 2420// between Use-Def edges and Def-Use edges in the graph. 2421void Compile::verify_graph_edges(bool no_dead_code) { 2422 if (VerifyGraphEdges) { 2423 ResourceArea *area = Thread::current()->resource_area(); 2424 Unique_Node_List visited(area); 2425 // Call recursive graph walk to check edges 2426 _root->verify_edges(visited); 2427 if (no_dead_code) { 2428 // Now make sure that no visited node is used by an unvisited node. 2429 bool dead_nodes = 0; 2430 Unique_Node_List checked(area); 2431 while (visited.size() > 0) { 2432 Node* n = visited.pop(); 2433 checked.push(n); 2434 for (uint i = 0; i < n->outcnt(); i++) { 2435 Node* use = n->raw_out(i); 2436 if (checked.member(use)) continue; // already checked 2437 if (visited.member(use)) continue; // already in the graph 2438 if (use->is_Con()) continue; // a dead ConNode is OK 2439 // At this point, we have found a dead node which is DU-reachable. 2440 if (dead_nodes++ == 0) 2441 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2442 use->dump(2); 2443 tty->print_cr("---"); 2444 checked.push(use); // No repeats; pretend it is now checked. 2445 } 2446 } 2447 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2448 } 2449 } 2450} 2451#endif 2452 2453// The Compile object keeps track of failure reasons separately from the ciEnv. 2454// This is required because there is not quite a 1-1 relation between the 2455// ciEnv and its compilation task and the Compile object. Note that one 2456// ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2457// to backtrack and retry without subsuming loads. Other than this backtracking 2458// behavior, the Compile's failure reason is quietly copied up to the ciEnv 2459// by the logic in C2Compiler. 2460void Compile::record_failure(const char* reason) { 2461 if (log() != NULL) { 2462 log()->elem("failure reason='%s' phase='compile'", reason); 2463 } 2464 if (_failure_reason == NULL) { 2465 // Record the first failure reason. 2466 _failure_reason = reason; 2467 } 2468 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2469 C->print_method(_failure_reason); 2470 } 2471 _root = NULL; // flush the graph, too 2472} 2473 2474Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2475 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2476{ 2477 if (dolog) { 2478 C = Compile::current(); 2479 _log = C->log(); 2480 } else { 2481 C = NULL; 2482 _log = NULL; 2483 } 2484 if (_log != NULL) { 2485 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2486 _log->stamp(); 2487 _log->end_head(); 2488 } 2489} 2490 2491Compile::TracePhase::~TracePhase() { 2492 if (_log != NULL) { 2493 _log->done("phase nodes='%d'", C->unique()); 2494 } 2495} 2496