1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. All Rights Reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 #include "incls/_precompiled.incl" 26 #include "incls/_compile.cpp.incl" 27 28 /// Support for intrinsics. 29 30 // Return the index at which m must be inserted (or already exists). 31 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 32 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 33 #ifdef ASSERT 34 for (int i = 1; i < _intrinsics->length(); i++) { 35 CallGenerator* cg1 = _intrinsics->at(i-1); 36 CallGenerator* cg2 = _intrinsics->at(i); 37 assert(cg1->method() != cg2->method() 38 ? cg1->method() < cg2->method() 39 : cg1->is_virtual() < cg2->is_virtual(), 40 "compiler intrinsics list must stay sorted"); 41 } 42 #endif 43 // Binary search sorted list, in decreasing intervals [lo, hi]. 44 int lo = 0, hi = _intrinsics->length()-1; 45 while (lo <= hi) { 46 int mid = (uint)(hi + lo) / 2; 47 ciMethod* mid_m = _intrinsics->at(mid)->method(); 48 if (m < mid_m) { 49 hi = mid-1; 50 } else if (m > mid_m) { 51 lo = mid+1; 52 } else { 53 // look at minor sort key 54 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 55 if (is_virtual < mid_virt) { 56 hi = mid-1; 57 } else if (is_virtual > mid_virt) { 58 lo = mid+1; 59 } else { 60 return mid; // exact match 61 } 62 } 63 } 64 return lo; // inexact match 65 } 66 67 void Compile::register_intrinsic(CallGenerator* cg) { 68 if (_intrinsics == NULL) { 69 _intrinsics = new GrowableArray<CallGenerator*>(60); 70 } 71 // This code is stolen from ciObjectFactory::insert. 72 // Really, GrowableArray should have methods for 73 // insert_at, remove_at, and binary_search. 74 int len = _intrinsics->length(); 75 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 76 if (index == len) { 77 _intrinsics->append(cg); 78 } else { 79 #ifdef ASSERT 80 CallGenerator* oldcg = _intrinsics->at(index); 81 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 82 #endif 83 _intrinsics->append(_intrinsics->at(len-1)); 84 int pos; 85 for (pos = len-2; pos >= index; pos--) { 86 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 87 } 88 _intrinsics->at_put(index, cg); 89 } 90 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 91 } 92 93 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 94 assert(m->is_loaded(), "don't try this on unloaded methods"); 95 if (_intrinsics != NULL) { 96 int index = intrinsic_insertion_index(m, is_virtual); 97 if (index < _intrinsics->length() 98 && _intrinsics->at(index)->method() == m 99 && _intrinsics->at(index)->is_virtual() == is_virtual) { 100 return _intrinsics->at(index); 101 } 102 } 103 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 104 if (m->intrinsic_id() != vmIntrinsics::_none && 105 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 106 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 107 if (cg != NULL) { 108 // Save it for next time: 109 register_intrinsic(cg); 110 return cg; 111 } else { 112 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 113 } 114 } 115 return NULL; 116 } 117 118 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 119 // in library_call.cpp. 120 121 122 #ifndef PRODUCT 123 // statistics gathering... 124 125 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 126 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 127 128 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 129 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 130 int oflags = _intrinsic_hist_flags[id]; 131 assert(flags != 0, "what happened?"); 132 if (is_virtual) { 133 flags |= _intrinsic_virtual; 134 } 135 bool changed = (flags != oflags); 136 if ((flags & _intrinsic_worked) != 0) { 137 juint count = (_intrinsic_hist_count[id] += 1); 138 if (count == 1) { 139 changed = true; // first time 140 } 141 // increment the overall count also: 142 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 143 } 144 if (changed) { 145 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 146 // Something changed about the intrinsic's virtuality. 147 if ((flags & _intrinsic_virtual) != 0) { 148 // This is the first use of this intrinsic as a virtual call. 149 if (oflags != 0) { 150 // We already saw it as a non-virtual, so note both cases. 151 flags |= _intrinsic_both; 152 } 153 } else if ((oflags & _intrinsic_both) == 0) { 154 // This is the first use of this intrinsic as a non-virtual 155 flags |= _intrinsic_both; 156 } 157 } 158 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 159 } 160 // update the overall flags also: 161 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 162 return changed; 163 } 164 165 static char* format_flags(int flags, char* buf) { 166 buf[0] = 0; 167 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 168 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 169 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 170 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 171 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 172 if (buf[0] == 0) strcat(buf, ","); 173 assert(buf[0] == ',', "must be"); 174 return &buf[1]; 175 } 176 177 void Compile::print_intrinsic_statistics() { 178 char flagsbuf[100]; 179 ttyLocker ttyl; 180 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 181 tty->print_cr("Compiler intrinsic usage:"); 182 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 183 if (total == 0) total = 1; // avoid div0 in case of no successes 184 #define PRINT_STAT_LINE(name, c, f) \ 185 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 186 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 187 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 188 int flags = _intrinsic_hist_flags[id]; 189 juint count = _intrinsic_hist_count[id]; 190 if ((flags | count) != 0) { 191 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 192 } 193 } 194 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 195 if (xtty != NULL) xtty->tail("statistics"); 196 } 197 198 void Compile::print_statistics() { 199 { ttyLocker ttyl; 200 if (xtty != NULL) xtty->head("statistics type='opto'"); 201 Parse::print_statistics(); 202 PhaseCCP::print_statistics(); 203 PhaseRegAlloc::print_statistics(); 204 Scheduling::print_statistics(); 205 PhasePeephole::print_statistics(); 206 PhaseIdealLoop::print_statistics(); 207 if (xtty != NULL) xtty->tail("statistics"); 208 } 209 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 210 // put this under its own <statistics> element. 211 print_intrinsic_statistics(); 212 } 213 } 214 #endif //PRODUCT 215 216 // Support for bundling info 217 Bundle* Compile::node_bundling(const Node *n) { 218 assert(valid_bundle_info(n), "oob"); 219 return &_node_bundling_base[n->_idx]; 220 } 221 222 bool Compile::valid_bundle_info(const Node *n) { 223 return (_node_bundling_limit > n->_idx); 224 } 225 226 227 // Identify all nodes that are reachable from below, useful. 228 // Use breadth-first pass that records state in a Unique_Node_List, 229 // recursive traversal is slower. 230 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 231 int estimated_worklist_size = unique(); 232 useful.map( estimated_worklist_size, NULL ); // preallocate space 233 234 // Initialize worklist 235 if (root() != NULL) { useful.push(root()); } 236 // If 'top' is cached, declare it useful to preserve cached node 237 if( cached_top_node() ) { useful.push(cached_top_node()); } 238 239 // Push all useful nodes onto the list, breadthfirst 240 for( uint next = 0; next < useful.size(); ++next ) { 241 assert( next < unique(), "Unique useful nodes < total nodes"); 242 Node *n = useful.at(next); 243 uint max = n->len(); 244 for( uint i = 0; i < max; ++i ) { 245 Node *m = n->in(i); 246 if( m == NULL ) continue; 247 useful.push(m); 248 } 249 } 250 } 251 252 // Disconnect all useless nodes by disconnecting those at the boundary. 253 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 254 uint next = 0; 255 while( next < useful.size() ) { 256 Node *n = useful.at(next++); 257 // Use raw traversal of out edges since this code removes out edges 258 int max = n->outcnt(); 259 for (int j = 0; j < max; ++j ) { 260 Node* child = n->raw_out(j); 261 if( ! useful.member(child) ) { 262 assert( !child->is_top() || child != top(), 263 "If top is cached in Compile object it is in useful list"); 264 // Only need to remove this out-edge to the useless node 265 n->raw_del_out(j); 266 --j; 267 --max; 268 } 269 } 270 if (n->outcnt() == 1 && n->has_special_unique_user()) { 271 record_for_igvn( n->unique_out() ); 272 } 273 } 274 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 275 } 276 277 //------------------------------frame_size_in_words----------------------------- 278 // frame_slots in units of words 279 int Compile::frame_size_in_words() const { 280 // shift is 0 in LP32 and 1 in LP64 281 const int shift = (LogBytesPerWord - LogBytesPerInt); 282 int words = _frame_slots >> shift; 283 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 284 return words; 285 } 286 287 // ============================================================================ 288 //------------------------------CompileWrapper--------------------------------- 289 class CompileWrapper : public StackObj { 290 Compile *const _compile; 291 public: 292 CompileWrapper(Compile* compile); 293 294 ~CompileWrapper(); 295 }; 296 297 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 298 // the Compile* pointer is stored in the current ciEnv: 299 ciEnv* env = compile->env(); 300 assert(env == ciEnv::current(), "must already be a ciEnv active"); 301 assert(env->compiler_data() == NULL, "compile already active?"); 302 env->set_compiler_data(compile); 303 assert(compile == Compile::current(), "sanity"); 304 305 compile->set_type_dict(NULL); 306 compile->set_type_hwm(NULL); 307 compile->set_type_last_size(0); 308 compile->set_last_tf(NULL, NULL); 309 compile->set_indexSet_arena(NULL); 310 compile->set_indexSet_free_block_list(NULL); 311 compile->init_type_arena(); 312 Type::Initialize(compile); 313 _compile->set_scratch_buffer_blob(NULL); 314 _compile->begin_method(); 315 } 316 CompileWrapper::~CompileWrapper() { 317 _compile->end_method(); 318 if (_compile->scratch_buffer_blob() != NULL) 319 BufferBlob::free(_compile->scratch_buffer_blob()); 320 _compile->env()->set_compiler_data(NULL); 321 } 322 323 324 //----------------------------print_compile_messages--------------------------- 325 void Compile::print_compile_messages() { 326 #ifndef PRODUCT 327 // Check if recompiling 328 if (_subsume_loads == false && PrintOpto) { 329 // Recompiling without allowing machine instructions to subsume loads 330 tty->print_cr("*********************************************************"); 331 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 332 tty->print_cr("*********************************************************"); 333 } 334 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 335 // Recompiling without escape analysis 336 tty->print_cr("*********************************************************"); 337 tty->print_cr("** Bailout: Recompile without escape analysis **"); 338 tty->print_cr("*********************************************************"); 339 } 340 if (env()->break_at_compile()) { 341 // Open the debugger when compiling this method. 342 tty->print("### Breaking when compiling: "); 343 method()->print_short_name(); 344 tty->cr(); 345 BREAKPOINT; 346 } 347 348 if( PrintOpto ) { 349 if (is_osr_compilation()) { 350 tty->print("[OSR]%3d", _compile_id); 351 } else { 352 tty->print("%3d", _compile_id); 353 } 354 } 355 #endif 356 } 357 358 359 void Compile::init_scratch_buffer_blob() { 360 if( scratch_buffer_blob() != NULL ) return; 361 362 // Construct a temporary CodeBuffer to have it construct a BufferBlob 363 // Cache this BufferBlob for this compile. 364 ResourceMark rm; 365 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 366 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 367 // Record the buffer blob for next time. 368 set_scratch_buffer_blob(blob); 369 // Have we run out of code space? 370 if (scratch_buffer_blob() == NULL) { 371 // Let CompilerBroker disable further compilations. 372 record_failure("Not enough space for scratch buffer in CodeCache"); 373 return; 374 } 375 376 // Initialize the relocation buffers 377 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 378 set_scratch_locs_memory(locs_buf); 379 } 380 381 382 //-----------------------scratch_emit_size------------------------------------- 383 // Helper function that computes size by emitting code 384 uint Compile::scratch_emit_size(const Node* n) { 385 // Emit into a trash buffer and count bytes emitted. 386 // This is a pretty expensive way to compute a size, 387 // but it works well enough if seldom used. 388 // All common fixed-size instructions are given a size 389 // method by the AD file. 390 // Note that the scratch buffer blob and locs memory are 391 // allocated at the beginning of the compile task, and 392 // may be shared by several calls to scratch_emit_size. 393 // The allocation of the scratch buffer blob is particularly 394 // expensive, since it has to grab the code cache lock. 395 BufferBlob* blob = this->scratch_buffer_blob(); 396 assert(blob != NULL, "Initialize BufferBlob at start"); 397 assert(blob->size() > MAX_inst_size, "sanity"); 398 relocInfo* locs_buf = scratch_locs_memory(); 399 address blob_begin = blob->instructions_begin(); 400 address blob_end = (address)locs_buf; 401 assert(blob->instructions_contains(blob_end), "sanity"); 402 CodeBuffer buf(blob_begin, blob_end - blob_begin); 403 buf.initialize_consts_size(MAX_const_size); 404 buf.initialize_stubs_size(MAX_stubs_size); 405 assert(locs_buf != NULL, "sanity"); 406 int lsize = MAX_locs_size / 2; 407 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 408 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 409 n->emit(buf, this->regalloc()); 410 return buf.code_size(); 411 } 412 413 414 // ============================================================================ 415 //------------------------------Compile standard------------------------------- 416 debug_only( int Compile::_debug_idx = 100000; ) 417 418 // Compile a method. entry_bci is -1 for normal compilations and indicates 419 // the continuation bci for on stack replacement. 420 421 422 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 423 : Phase(Compiler), 424 _env(ci_env), 425 _log(ci_env->log()), 426 _compile_id(ci_env->compile_id()), 427 _save_argument_registers(false), 428 _stub_name(NULL), 429 _stub_function(NULL), 430 _stub_entry_point(NULL), 431 _method(target), 432 _entry_bci(osr_bci), 433 _initial_gvn(NULL), 434 _for_igvn(NULL), 435 _warm_calls(NULL), 436 _subsume_loads(subsume_loads), 437 _do_escape_analysis(do_escape_analysis), 438 _failure_reason(NULL), 439 _code_buffer("Compile::Fill_buffer"), 440 _orig_pc_slot(0), 441 _orig_pc_slot_offset_in_bytes(0), 442 _node_bundling_limit(0), 443 _node_bundling_base(NULL), 444 _java_calls(0), 445 _inner_loops(0), 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 set_parsed_irreducible_loop(false); 474 #endif 475 476 if (ProfileTraps) { 477 // Make sure the method being compiled gets its own MDO, 478 // so we can at least track the decompile_count(). 479 method()->build_method_data(); 480 } 481 482 Init(::AliasLevel); 483 484 485 print_compile_messages(); 486 487 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 488 _ilt = InlineTree::build_inline_tree_root(); 489 else 490 _ilt = NULL; 491 492 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 493 assert(num_alias_types() >= AliasIdxRaw, ""); 494 495 #define MINIMUM_NODE_HASH 1023 496 // Node list that Iterative GVN will start with 497 Unique_Node_List for_igvn(comp_arena()); 498 set_for_igvn(&for_igvn); 499 500 // GVN that will be run immediately on new nodes 501 uint estimated_size = method()->code_size()*4+64; 502 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 503 PhaseGVN gvn(node_arena(), estimated_size); 504 set_initial_gvn(&gvn); 505 506 { // Scope for timing the parser 507 TracePhase t3("parse", &_t_parser, true); 508 509 // Put top into the hash table ASAP. 510 initial_gvn()->transform_no_reclaim(top()); 511 512 // Set up tf(), start(), and find a CallGenerator. 513 CallGenerator* cg; 514 if (is_osr_compilation()) { 515 const TypeTuple *domain = StartOSRNode::osr_domain(); 516 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 517 init_tf(TypeFunc::make(domain, range)); 518 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 519 initial_gvn()->set_type_bottom(s); 520 init_start(s); 521 cg = CallGenerator::for_osr(method(), entry_bci()); 522 } else { 523 // Normal case. 524 init_tf(TypeFunc::make(method())); 525 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 526 initial_gvn()->set_type_bottom(s); 527 init_start(s); 528 float past_uses = method()->interpreter_invocation_count(); 529 float expected_uses = past_uses; 530 cg = CallGenerator::for_inline(method(), expected_uses); 531 } 532 if (failing()) return; 533 if (cg == NULL) { 534 record_method_not_compilable_all_tiers("cannot parse method"); 535 return; 536 } 537 JVMState* jvms = build_start_state(start(), tf()); 538 if ((jvms = cg->generate(jvms)) == NULL) { 539 record_method_not_compilable("method parse failed"); 540 return; 541 } 542 GraphKit kit(jvms); 543 544 if (!kit.stopped()) { 545 // Accept return values, and transfer control we know not where. 546 // This is done by a special, unique ReturnNode bound to root. 547 return_values(kit.jvms()); 548 } 549 550 if (kit.has_exceptions()) { 551 // Any exceptions that escape from this call must be rethrown 552 // to whatever caller is dynamically above us on the stack. 553 // This is done by a special, unique RethrowNode bound to root. 554 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 555 } 556 557 print_method("Before RemoveUseless", 3); 558 559 // Remove clutter produced by parsing. 560 if (!failing()) { 561 ResourceMark rm; 562 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 563 } 564 } 565 566 // Note: Large methods are capped off in do_one_bytecode(). 567 if (failing()) return; 568 569 // After parsing, node notes are no longer automagic. 570 // They must be propagated by register_new_node_with_optimizer(), 571 // clone(), or the like. 572 set_default_node_notes(NULL); 573 574 for (;;) { 575 int successes = Inline_Warm(); 576 if (failing()) return; 577 if (successes == 0) break; 578 } 579 580 // Drain the list. 581 Finish_Warm(); 582 #ifndef PRODUCT 583 if (_printer) { 584 _printer->print_inlining(this); 585 } 586 #endif 587 588 if (failing()) return; 589 NOT_PRODUCT( verify_graph_edges(); ) 590 591 // Perform escape analysis 592 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 593 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 594 // Add ConP#NULL and ConN#NULL nodes before ConnectionGraph construction. 595 PhaseGVN* igvn = initial_gvn(); 596 Node* oop_null = igvn->zerocon(T_OBJECT); 597 Node* noop_null = igvn->zerocon(T_NARROWOOP); 598 599 _congraph = new(comp_arena()) ConnectionGraph(this); 600 bool has_non_escaping_obj = _congraph->compute_escape(); 601 602 #ifndef PRODUCT 603 if (PrintEscapeAnalysis) { 604 _congraph->dump(); 605 } 606 #endif 607 // Cleanup. 608 if (oop_null->outcnt() == 0) 609 igvn->hash_delete(oop_null); 610 if (noop_null->outcnt() == 0) 611 igvn->hash_delete(noop_null); 612 613 if (!has_non_escaping_obj) { 614 _congraph = NULL; 615 } 616 617 if (failing()) return; 618 } 619 // Now optimize 620 Optimize(); 621 if (failing()) return; 622 NOT_PRODUCT( verify_graph_edges(); ) 623 624 #ifndef PRODUCT 625 if (PrintIdeal) { 626 ttyLocker ttyl; // keep the following output all in one block 627 // This output goes directly to the tty, not the compiler log. 628 // To enable tools to match it up with the compilation activity, 629 // be sure to tag this tty output with the compile ID. 630 if (xtty != NULL) { 631 xtty->head("ideal compile_id='%d'%s", compile_id(), 632 is_osr_compilation() ? " compile_kind='osr'" : 633 ""); 634 } 635 root()->dump(9999); 636 if (xtty != NULL) { 637 xtty->tail("ideal"); 638 } 639 } 640 #endif 641 642 // Now that we know the size of all the monitors we can add a fixed slot 643 // for the original deopt pc. 644 645 _orig_pc_slot = fixed_slots(); 646 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 647 set_fixed_slots(next_slot); 648 649 // Now generate code 650 Code_Gen(); 651 if (failing()) return; 652 653 // Check if we want to skip execution of all compiled code. 654 { 655 #ifndef PRODUCT 656 if (OptoNoExecute) { 657 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 658 return; 659 } 660 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 661 #endif 662 663 if (is_osr_compilation()) { 664 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 665 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 666 } else { 667 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 668 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 669 } 670 671 env()->register_method(_method, _entry_bci, 672 &_code_offsets, 673 _orig_pc_slot_offset_in_bytes, 674 code_buffer(), 675 frame_size_in_words(), _oop_map_set, 676 &_handler_table, &_inc_table, 677 compiler, 678 env()->comp_level(), 679 true, /*has_debug_info*/ 680 has_unsafe_access() 681 ); 682 } 683 } 684 685 //------------------------------Compile---------------------------------------- 686 // Compile a runtime stub 687 Compile::Compile( ciEnv* ci_env, 688 TypeFunc_generator generator, 689 address stub_function, 690 const char *stub_name, 691 int is_fancy_jump, 692 bool pass_tls, 693 bool save_arg_registers, 694 bool return_pc ) 695 : Phase(Compiler), 696 _env(ci_env), 697 _log(ci_env->log()), 698 _compile_id(-1), 699 _save_argument_registers(save_arg_registers), 700 _method(NULL), 701 _stub_name(stub_name), 702 _stub_function(stub_function), 703 _stub_entry_point(NULL), 704 _entry_bci(InvocationEntryBci), 705 _initial_gvn(NULL), 706 _for_igvn(NULL), 707 _warm_calls(NULL), 708 _orig_pc_slot(0), 709 _orig_pc_slot_offset_in_bytes(0), 710 _subsume_loads(true), 711 _do_escape_analysis(false), 712 _failure_reason(NULL), 713 _code_buffer("Compile::Fill_buffer"), 714 _node_bundling_limit(0), 715 _node_bundling_base(NULL), 716 _java_calls(0), 717 _inner_loops(0), 718 #ifndef PRODUCT 719 _trace_opto_output(TraceOptoOutput), 720 _printer(NULL), 721 #endif 722 _congraph(NULL) { 723 C = this; 724 725 #ifndef PRODUCT 726 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 727 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 728 set_print_assembly(PrintFrameConverterAssembly); 729 set_parsed_irreducible_loop(false); 730 #endif 731 CompileWrapper cw(this); 732 Init(/*AliasLevel=*/ 0); 733 init_tf((*generator)()); 734 735 { 736 // The following is a dummy for the sake of GraphKit::gen_stub 737 Unique_Node_List for_igvn(comp_arena()); 738 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 739 PhaseGVN gvn(Thread::current()->resource_area(),255); 740 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 741 gvn.transform_no_reclaim(top()); 742 743 GraphKit kit; 744 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 745 } 746 747 NOT_PRODUCT( verify_graph_edges(); ) 748 Code_Gen(); 749 if (failing()) return; 750 751 752 // Entry point will be accessed using compile->stub_entry_point(); 753 if (code_buffer() == NULL) { 754 Matcher::soft_match_failure(); 755 } else { 756 if (PrintAssembly && (WizardMode || Verbose)) 757 tty->print_cr("### Stub::%s", stub_name); 758 759 if (!failing()) { 760 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 761 762 // Make the NMethod 763 // For now we mark the frame as never safe for profile stackwalking 764 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 765 code_buffer(), 766 CodeOffsets::frame_never_safe, 767 // _code_offsets.value(CodeOffsets::Frame_Complete), 768 frame_size_in_words(), 769 _oop_map_set, 770 save_arg_registers); 771 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 772 773 _stub_entry_point = rs->entry_point(); 774 } 775 } 776 } 777 778 #ifndef PRODUCT 779 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 780 if(PrintOpto && Verbose) { 781 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 782 } 783 } 784 #endif 785 786 void Compile::print_codes() { 787 } 788 789 //------------------------------Init------------------------------------------- 790 // Prepare for a single compilation 791 void Compile::Init(int aliaslevel) { 792 _unique = 0; 793 _regalloc = NULL; 794 795 _tf = NULL; // filled in later 796 _top = NULL; // cached later 797 _matcher = NULL; // filled in later 798 _cfg = NULL; // filled in later 799 800 set_24_bit_selection_and_mode(Use24BitFP, false); 801 802 _node_note_array = NULL; 803 _default_node_notes = NULL; 804 805 _immutable_memory = NULL; // filled in at first inquiry 806 807 // Globally visible Nodes 808 // First set TOP to NULL to give safe behavior during creation of RootNode 809 set_cached_top_node(NULL); 810 set_root(new (this, 3) RootNode()); 811 // Now that you have a Root to point to, create the real TOP 812 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 813 set_recent_alloc(NULL, NULL); 814 815 // Create Debug Information Recorder to record scopes, oopmaps, etc. 816 env()->set_oop_recorder(new OopRecorder(comp_arena())); 817 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 818 env()->set_dependencies(new Dependencies(env())); 819 820 _fixed_slots = 0; 821 set_has_split_ifs(false); 822 set_has_loops(has_method() && method()->has_loops()); // first approximation 823 _deopt_happens = true; // start out assuming the worst 824 _trap_can_recompile = false; // no traps emitted yet 825 _major_progress = true; // start out assuming good things will happen 826 set_has_unsafe_access(false); 827 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 828 set_decompile_count(0); 829 830 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 831 // Compilation level related initialization 832 if (env()->comp_level() == CompLevel_fast_compile) { 833 set_num_loop_opts(Tier1LoopOptsCount); 834 set_do_inlining(Tier1Inline != 0); 835 set_max_inline_size(Tier1MaxInlineSize); 836 set_freq_inline_size(Tier1FreqInlineSize); 837 set_do_scheduling(false); 838 set_do_count_invocations(Tier1CountInvocations); 839 set_do_method_data_update(Tier1UpdateMethodData); 840 } else { 841 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 842 set_num_loop_opts(LoopOptsCount); 843 set_do_inlining(Inline); 844 set_max_inline_size(MaxInlineSize); 845 set_freq_inline_size(FreqInlineSize); 846 set_do_scheduling(OptoScheduling); 847 set_do_count_invocations(false); 848 set_do_method_data_update(false); 849 } 850 851 if (debug_info()->recording_non_safepoints()) { 852 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 853 (comp_arena(), 8, 0, NULL)); 854 set_default_node_notes(Node_Notes::make(this)); 855 } 856 857 // // -- Initialize types before each compile -- 858 // // Update cached type information 859 // if( _method && _method->constants() ) 860 // Type::update_loaded_types(_method, _method->constants()); 861 862 // Init alias_type map. 863 if (!_do_escape_analysis && aliaslevel == 3) 864 aliaslevel = 2; // No unique types without escape analysis 865 _AliasLevel = aliaslevel; 866 const int grow_ats = 16; 867 _max_alias_types = grow_ats; 868 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 869 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 870 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 871 { 872 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 873 } 874 // Initialize the first few types. 875 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 876 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 877 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 878 _num_alias_types = AliasIdxRaw+1; 879 // Zero out the alias type cache. 880 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 881 // A NULL adr_type hits in the cache right away. Preload the right answer. 882 probe_alias_cache(NULL)->_index = AliasIdxTop; 883 884 _intrinsics = NULL; 885 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 886 register_library_intrinsics(); 887 } 888 889 //---------------------------init_start---------------------------------------- 890 // Install the StartNode on this compile object. 891 void Compile::init_start(StartNode* s) { 892 if (failing()) 893 return; // already failing 894 assert(s == start(), ""); 895 } 896 897 StartNode* Compile::start() const { 898 assert(!failing(), ""); 899 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 900 Node* start = root()->fast_out(i); 901 if( start->is_Start() ) 902 return start->as_Start(); 903 } 904 ShouldNotReachHere(); 905 return NULL; 906 } 907 908 //-------------------------------immutable_memory------------------------------------- 909 // Access immutable memory 910 Node* Compile::immutable_memory() { 911 if (_immutable_memory != NULL) { 912 return _immutable_memory; 913 } 914 StartNode* s = start(); 915 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 916 Node *p = s->fast_out(i); 917 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 918 _immutable_memory = p; 919 return _immutable_memory; 920 } 921 } 922 ShouldNotReachHere(); 923 return NULL; 924 } 925 926 //----------------------set_cached_top_node------------------------------------ 927 // Install the cached top node, and make sure Node::is_top works correctly. 928 void Compile::set_cached_top_node(Node* tn) { 929 if (tn != NULL) verify_top(tn); 930 Node* old_top = _top; 931 _top = tn; 932 // Calling Node::setup_is_top allows the nodes the chance to adjust 933 // their _out arrays. 934 if (_top != NULL) _top->setup_is_top(); 935 if (old_top != NULL) old_top->setup_is_top(); 936 assert(_top == NULL || top()->is_top(), ""); 937 } 938 939 #ifndef PRODUCT 940 void Compile::verify_top(Node* tn) const { 941 if (tn != NULL) { 942 assert(tn->is_Con(), "top node must be a constant"); 943 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 944 assert(tn->in(0) != NULL, "must have live top node"); 945 } 946 } 947 #endif 948 949 950 ///-------------------Managing Per-Node Debug & Profile Info------------------- 951 952 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 953 guarantee(arr != NULL, ""); 954 int num_blocks = arr->length(); 955 if (grow_by < num_blocks) grow_by = num_blocks; 956 int num_notes = grow_by * _node_notes_block_size; 957 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 958 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 959 while (num_notes > 0) { 960 arr->append(notes); 961 notes += _node_notes_block_size; 962 num_notes -= _node_notes_block_size; 963 } 964 assert(num_notes == 0, "exact multiple, please"); 965 } 966 967 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 968 if (source == NULL || dest == NULL) return false; 969 970 if (dest->is_Con()) 971 return false; // Do not push debug info onto constants. 972 973 #ifdef ASSERT 974 // Leave a bread crumb trail pointing to the original node: 975 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 976 dest->set_debug_orig(source); 977 } 978 #endif 979 980 if (node_note_array() == NULL) 981 return false; // Not collecting any notes now. 982 983 // This is a copy onto a pre-existing node, which may already have notes. 984 // If both nodes have notes, do not overwrite any pre-existing notes. 985 Node_Notes* source_notes = node_notes_at(source->_idx); 986 if (source_notes == NULL || source_notes->is_clear()) return false; 987 Node_Notes* dest_notes = node_notes_at(dest->_idx); 988 if (dest_notes == NULL || dest_notes->is_clear()) { 989 return set_node_notes_at(dest->_idx, source_notes); 990 } 991 992 Node_Notes merged_notes = (*source_notes); 993 // The order of operations here ensures that dest notes will win... 994 merged_notes.update_from(dest_notes); 995 return set_node_notes_at(dest->_idx, &merged_notes); 996 } 997 998 999 //--------------------------allow_range_check_smearing------------------------- 1000 // Gating condition for coalescing similar range checks. 1001 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1002 // single covering check that is at least as strong as any of them. 1003 // If the optimization succeeds, the simplified (strengthened) range check 1004 // will always succeed. If it fails, we will deopt, and then give up 1005 // on the optimization. 1006 bool Compile::allow_range_check_smearing() const { 1007 // If this method has already thrown a range-check, 1008 // assume it was because we already tried range smearing 1009 // and it failed. 1010 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1011 return !already_trapped; 1012 } 1013 1014 1015 //------------------------------flatten_alias_type----------------------------- 1016 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1017 int offset = tj->offset(); 1018 TypePtr::PTR ptr = tj->ptr(); 1019 1020 // Known instance (scalarizable allocation) alias only with itself. 1021 bool is_known_inst = tj->isa_oopptr() != NULL && 1022 tj->is_oopptr()->is_known_instance(); 1023 1024 // Process weird unsafe references. 1025 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1026 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1027 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1028 tj = TypeOopPtr::BOTTOM; 1029 ptr = tj->ptr(); 1030 offset = tj->offset(); 1031 } 1032 1033 // Array pointers need some flattening 1034 const TypeAryPtr *ta = tj->isa_aryptr(); 1035 if( ta && is_known_inst ) { 1036 if ( offset != Type::OffsetBot && 1037 offset > arrayOopDesc::length_offset_in_bytes() ) { 1038 offset = Type::OffsetBot; // Flatten constant access into array body only 1039 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1040 } 1041 } else if( ta && _AliasLevel >= 2 ) { 1042 // For arrays indexed by constant indices, we flatten the alias 1043 // space to include all of the array body. Only the header, klass 1044 // and array length can be accessed un-aliased. 1045 if( offset != Type::OffsetBot ) { 1046 if( ta->const_oop() ) { // methodDataOop or methodOop 1047 offset = Type::OffsetBot; // Flatten constant access into array body 1048 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1049 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1050 // range is OK as-is. 1051 tj = ta = TypeAryPtr::RANGE; 1052 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1053 tj = TypeInstPtr::KLASS; // all klass loads look alike 1054 ta = TypeAryPtr::RANGE; // generic ignored junk 1055 ptr = TypePtr::BotPTR; 1056 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1057 tj = TypeInstPtr::MARK; 1058 ta = TypeAryPtr::RANGE; // generic ignored junk 1059 ptr = TypePtr::BotPTR; 1060 } else { // Random constant offset into array body 1061 offset = Type::OffsetBot; // Flatten constant access into array body 1062 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1063 } 1064 } 1065 // Arrays of fixed size alias with arrays of unknown size. 1066 if (ta->size() != TypeInt::POS) { 1067 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1068 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1069 } 1070 // Arrays of known objects become arrays of unknown objects. 1071 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1072 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1073 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1074 } 1075 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1076 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1077 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1078 } 1079 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1080 // cannot be distinguished by bytecode alone. 1081 if (ta->elem() == TypeInt::BOOL) { 1082 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1083 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1084 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1085 } 1086 // During the 2nd round of IterGVN, NotNull castings are removed. 1087 // Make sure the Bottom and NotNull variants alias the same. 1088 // Also, make sure exact and non-exact variants alias the same. 1089 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1090 if (ta->const_oop()) { 1091 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1092 } else { 1093 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1094 } 1095 } 1096 } 1097 1098 // Oop pointers need some flattening 1099 const TypeInstPtr *to = tj->isa_instptr(); 1100 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1101 if( ptr == TypePtr::Constant ) { 1102 // No constant oop pointers (such as Strings); they alias with 1103 // unknown strings. 1104 assert(!is_known_inst, "not scalarizable allocation"); 1105 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1106 } else if( is_known_inst ) { 1107 tj = to; // Keep NotNull and klass_is_exact for instance type 1108 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1109 // During the 2nd round of IterGVN, NotNull castings are removed. 1110 // Make sure the Bottom and NotNull variants alias the same. 1111 // Also, make sure exact and non-exact variants alias the same. 1112 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1113 } 1114 // Canonicalize the holder of this field 1115 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1116 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1117 // First handle header references such as a LoadKlassNode, even if the 1118 // object's klass is unloaded at compile time (4965979). 1119 if (!is_known_inst) { // Do it only for non-instance types 1120 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1121 } 1122 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1123 to = NULL; 1124 tj = TypeOopPtr::BOTTOM; 1125 offset = tj->offset(); 1126 } else { 1127 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1128 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1129 if( is_known_inst ) { 1130 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1131 } else { 1132 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1133 } 1134 } 1135 } 1136 } 1137 1138 // Klass pointers to object array klasses need some flattening 1139 const TypeKlassPtr *tk = tj->isa_klassptr(); 1140 if( tk ) { 1141 // If we are referencing a field within a Klass, we need 1142 // to assume the worst case of an Object. Both exact and 1143 // inexact types must flatten to the same alias class. 1144 // Since the flattened result for a klass is defined to be 1145 // precisely java.lang.Object, use a constant ptr. 1146 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1147 1148 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1149 TypeKlassPtr::OBJECT->klass(), 1150 offset); 1151 } 1152 1153 ciKlass* klass = tk->klass(); 1154 if( klass->is_obj_array_klass() ) { 1155 ciKlass* k = TypeAryPtr::OOPS->klass(); 1156 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1157 k = TypeInstPtr::BOTTOM->klass(); 1158 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1159 } 1160 1161 // Check for precise loads from the primary supertype array and force them 1162 // to the supertype cache alias index. Check for generic array loads from 1163 // the primary supertype array and also force them to the supertype cache 1164 // alias index. Since the same load can reach both, we need to merge 1165 // these 2 disparate memories into the same alias class. Since the 1166 // primary supertype array is read-only, there's no chance of confusion 1167 // where we bypass an array load and an array store. 1168 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1169 if( offset == Type::OffsetBot || 1170 off2 < Klass::primary_super_limit()*wordSize ) { 1171 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1172 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1173 } 1174 } 1175 1176 // Flatten all Raw pointers together. 1177 if (tj->base() == Type::RawPtr) 1178 tj = TypeRawPtr::BOTTOM; 1179 1180 if (tj->base() == Type::AnyPtr) 1181 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1182 1183 // Flatten all to bottom for now 1184 switch( _AliasLevel ) { 1185 case 0: 1186 tj = TypePtr::BOTTOM; 1187 break; 1188 case 1: // Flatten to: oop, static, field or array 1189 switch (tj->base()) { 1190 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1191 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1192 case Type::AryPtr: // do not distinguish arrays at all 1193 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1194 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1195 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1196 default: ShouldNotReachHere(); 1197 } 1198 break; 1199 case 2: // No collapsing at level 2; keep all splits 1200 case 3: // No collapsing at level 3; keep all splits 1201 break; 1202 default: 1203 Unimplemented(); 1204 } 1205 1206 offset = tj->offset(); 1207 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1208 1209 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1210 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1211 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1212 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1213 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1214 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1215 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1216 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1217 assert( tj->ptr() != TypePtr::TopPTR && 1218 tj->ptr() != TypePtr::AnyNull && 1219 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1220 // assert( tj->ptr() != TypePtr::Constant || 1221 // tj->base() == Type::RawPtr || 1222 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1223 1224 return tj; 1225 } 1226 1227 void Compile::AliasType::Init(int i, const TypePtr* at) { 1228 _index = i; 1229 _adr_type = at; 1230 _field = NULL; 1231 _is_rewritable = true; // default 1232 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1233 if (atoop != NULL && atoop->is_known_instance()) { 1234 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1235 _general_index = Compile::current()->get_alias_index(gt); 1236 } else { 1237 _general_index = 0; 1238 } 1239 } 1240 1241 //---------------------------------print_on------------------------------------ 1242 #ifndef PRODUCT 1243 void Compile::AliasType::print_on(outputStream* st) { 1244 if (index() < 10) 1245 st->print("@ <%d> ", index()); 1246 else st->print("@ <%d>", index()); 1247 st->print(is_rewritable() ? " " : " RO"); 1248 int offset = adr_type()->offset(); 1249 if (offset == Type::OffsetBot) 1250 st->print(" +any"); 1251 else st->print(" +%-3d", offset); 1252 st->print(" in "); 1253 adr_type()->dump_on(st); 1254 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1255 if (field() != NULL && tjp) { 1256 if (tjp->klass() != field()->holder() || 1257 tjp->offset() != field()->offset_in_bytes()) { 1258 st->print(" != "); 1259 field()->print(); 1260 st->print(" ***"); 1261 } 1262 } 1263 } 1264 1265 void print_alias_types() { 1266 Compile* C = Compile::current(); 1267 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1268 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1269 C->alias_type(idx)->print_on(tty); 1270 tty->cr(); 1271 } 1272 } 1273 #endif 1274 1275 1276 //----------------------------probe_alias_cache-------------------------------- 1277 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1278 intptr_t key = (intptr_t) adr_type; 1279 key ^= key >> logAliasCacheSize; 1280 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1281 } 1282 1283 1284 //-----------------------------grow_alias_types-------------------------------- 1285 void Compile::grow_alias_types() { 1286 const int old_ats = _max_alias_types; // how many before? 1287 const int new_ats = old_ats; // how many more? 1288 const int grow_ats = old_ats+new_ats; // how many now? 1289 _max_alias_types = grow_ats; 1290 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1291 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1292 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1293 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1294 } 1295 1296 1297 //--------------------------------find_alias_type------------------------------ 1298 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1299 if (_AliasLevel == 0) 1300 return alias_type(AliasIdxBot); 1301 1302 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1303 if (ace->_adr_type == adr_type) { 1304 return alias_type(ace->_index); 1305 } 1306 1307 // Handle special cases. 1308 if (adr_type == NULL) return alias_type(AliasIdxTop); 1309 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1310 1311 // Do it the slow way. 1312 const TypePtr* flat = flatten_alias_type(adr_type); 1313 1314 #ifdef ASSERT 1315 assert(flat == flatten_alias_type(flat), "idempotent"); 1316 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1317 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1318 const TypeOopPtr* foop = flat->is_oopptr(); 1319 // Scalarizable allocations have exact klass always. 1320 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1321 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1322 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1323 } 1324 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1325 #endif 1326 1327 int idx = AliasIdxTop; 1328 for (int i = 0; i < num_alias_types(); i++) { 1329 if (alias_type(i)->adr_type() == flat) { 1330 idx = i; 1331 break; 1332 } 1333 } 1334 1335 if (idx == AliasIdxTop) { 1336 if (no_create) return NULL; 1337 // Grow the array if necessary. 1338 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1339 // Add a new alias type. 1340 idx = _num_alias_types++; 1341 _alias_types[idx]->Init(idx, flat); 1342 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1343 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1344 if (flat->isa_instptr()) { 1345 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1346 && flat->is_instptr()->klass() == env()->Class_klass()) 1347 alias_type(idx)->set_rewritable(false); 1348 } 1349 if (flat->isa_klassptr()) { 1350 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1351 alias_type(idx)->set_rewritable(false); 1352 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1353 alias_type(idx)->set_rewritable(false); 1354 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1355 alias_type(idx)->set_rewritable(false); 1356 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1357 alias_type(idx)->set_rewritable(false); 1358 } 1359 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1360 // but the base pointer type is not distinctive enough to identify 1361 // references into JavaThread.) 1362 1363 // Check for final instance fields. 1364 const TypeInstPtr* tinst = flat->isa_instptr(); 1365 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1366 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1367 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1368 // Set field() and is_rewritable() attributes. 1369 if (field != NULL) alias_type(idx)->set_field(field); 1370 } 1371 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1372 // Check for final static fields. 1373 if (tklass && tklass->klass()->is_instance_klass()) { 1374 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1375 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1376 // Set field() and is_rewritable() attributes. 1377 if (field != NULL) alias_type(idx)->set_field(field); 1378 } 1379 } 1380 1381 // Fill the cache for next time. 1382 ace->_adr_type = adr_type; 1383 ace->_index = idx; 1384 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1385 1386 // Might as well try to fill the cache for the flattened version, too. 1387 AliasCacheEntry* face = probe_alias_cache(flat); 1388 if (face->_adr_type == NULL) { 1389 face->_adr_type = flat; 1390 face->_index = idx; 1391 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1392 } 1393 1394 return alias_type(idx); 1395 } 1396 1397 1398 Compile::AliasType* Compile::alias_type(ciField* field) { 1399 const TypeOopPtr* t; 1400 if (field->is_static()) 1401 t = TypeKlassPtr::make(field->holder()); 1402 else 1403 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1404 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1405 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1406 return atp; 1407 } 1408 1409 1410 //------------------------------have_alias_type-------------------------------- 1411 bool Compile::have_alias_type(const TypePtr* adr_type) { 1412 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1413 if (ace->_adr_type == adr_type) { 1414 return true; 1415 } 1416 1417 // Handle special cases. 1418 if (adr_type == NULL) return true; 1419 if (adr_type == TypePtr::BOTTOM) return true; 1420 1421 return find_alias_type(adr_type, true) != NULL; 1422 } 1423 1424 //-----------------------------must_alias-------------------------------------- 1425 // True if all values of the given address type are in the given alias category. 1426 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1427 if (alias_idx == AliasIdxBot) return true; // the universal category 1428 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1429 if (alias_idx == AliasIdxTop) return false; // the empty category 1430 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1431 1432 // the only remaining possible overlap is identity 1433 int adr_idx = get_alias_index(adr_type); 1434 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1435 assert(adr_idx == alias_idx || 1436 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1437 && adr_type != TypeOopPtr::BOTTOM), 1438 "should not be testing for overlap with an unsafe pointer"); 1439 return adr_idx == alias_idx; 1440 } 1441 1442 //------------------------------can_alias-------------------------------------- 1443 // True if any values of the given address type are in the given alias category. 1444 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1445 if (alias_idx == AliasIdxTop) return false; // the empty category 1446 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1447 if (alias_idx == AliasIdxBot) return true; // the universal category 1448 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1449 1450 // the only remaining possible overlap is identity 1451 int adr_idx = get_alias_index(adr_type); 1452 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1453 return adr_idx == alias_idx; 1454 } 1455 1456 1457 1458 //---------------------------pop_warm_call------------------------------------- 1459 WarmCallInfo* Compile::pop_warm_call() { 1460 WarmCallInfo* wci = _warm_calls; 1461 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1462 return wci; 1463 } 1464 1465 //----------------------------Inline_Warm-------------------------------------- 1466 int Compile::Inline_Warm() { 1467 // If there is room, try to inline some more warm call sites. 1468 // %%% Do a graph index compaction pass when we think we're out of space? 1469 if (!InlineWarmCalls) return 0; 1470 1471 int calls_made_hot = 0; 1472 int room_to_grow = NodeCountInliningCutoff - unique(); 1473 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1474 int amount_grown = 0; 1475 WarmCallInfo* call; 1476 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1477 int est_size = (int)call->size(); 1478 if (est_size > (room_to_grow - amount_grown)) { 1479 // This one won't fit anyway. Get rid of it. 1480 call->make_cold(); 1481 continue; 1482 } 1483 call->make_hot(); 1484 calls_made_hot++; 1485 amount_grown += est_size; 1486 amount_to_grow -= est_size; 1487 } 1488 1489 if (calls_made_hot > 0) set_major_progress(); 1490 return calls_made_hot; 1491 } 1492 1493 1494 //----------------------------Finish_Warm-------------------------------------- 1495 void Compile::Finish_Warm() { 1496 if (!InlineWarmCalls) return; 1497 if (failing()) return; 1498 if (warm_calls() == NULL) return; 1499 1500 // Clean up loose ends, if we are out of space for inlining. 1501 WarmCallInfo* call; 1502 while ((call = pop_warm_call()) != NULL) { 1503 call->make_cold(); 1504 } 1505 } 1506 1507 1508 //------------------------------Optimize--------------------------------------- 1509 // Given a graph, optimize it. 1510 void Compile::Optimize() { 1511 TracePhase t1("optimizer", &_t_optimizer, true); 1512 1513 #ifndef PRODUCT 1514 if (env()->break_at_compile()) { 1515 BREAKPOINT; 1516 } 1517 1518 #endif 1519 1520 ResourceMark rm; 1521 int loop_opts_cnt; 1522 1523 NOT_PRODUCT( verify_graph_edges(); ) 1524 1525 print_method("After Parsing"); 1526 1527 { 1528 // Iterative Global Value Numbering, including ideal transforms 1529 // Initialize IterGVN with types and values from parse-time GVN 1530 PhaseIterGVN igvn(initial_gvn()); 1531 { 1532 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1533 igvn.optimize(); 1534 } 1535 1536 print_method("Iter GVN 1", 2); 1537 1538 if (failing()) return; 1539 1540 // Loop transforms on the ideal graph. Range Check Elimination, 1541 // peeling, unrolling, etc. 1542 1543 // Set loop opts counter 1544 loop_opts_cnt = num_loop_opts(); 1545 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1546 { 1547 TracePhase t2("idealLoop", &_t_idealLoop, true); 1548 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1549 loop_opts_cnt--; 1550 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1551 if (failing()) return; 1552 } 1553 // Loop opts pass if partial peeling occurred in previous pass 1554 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1555 TracePhase t3("idealLoop", &_t_idealLoop, true); 1556 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1557 loop_opts_cnt--; 1558 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1559 if (failing()) return; 1560 } 1561 // Loop opts pass for loop-unrolling before CCP 1562 if(major_progress() && (loop_opts_cnt > 0)) { 1563 TracePhase t4("idealLoop", &_t_idealLoop, true); 1564 PhaseIdealLoop ideal_loop( igvn, NULL, false ); 1565 loop_opts_cnt--; 1566 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1567 } 1568 } 1569 if (failing()) return; 1570 1571 // Conditional Constant Propagation; 1572 PhaseCCP ccp( &igvn ); 1573 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1574 { 1575 TracePhase t2("ccp", &_t_ccp, true); 1576 ccp.do_transform(); 1577 } 1578 print_method("PhaseCPP 1", 2); 1579 1580 assert( true, "Break here to ccp.dump_old2new_map()"); 1581 1582 // Iterative Global Value Numbering, including ideal transforms 1583 { 1584 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1585 igvn = ccp; 1586 igvn.optimize(); 1587 } 1588 1589 print_method("Iter GVN 2", 2); 1590 1591 if (failing()) return; 1592 1593 // Loop transforms on the ideal graph. Range Check Elimination, 1594 // peeling, unrolling, etc. 1595 if(loop_opts_cnt > 0) { 1596 debug_only( int cnt = 0; ); 1597 while(major_progress() && (loop_opts_cnt > 0)) { 1598 TracePhase t2("idealLoop", &_t_idealLoop, true); 1599 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1600 PhaseIdealLoop ideal_loop( igvn, NULL, true ); 1601 loop_opts_cnt--; 1602 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1603 if (failing()) return; 1604 } 1605 } 1606 { 1607 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1608 PhaseMacroExpand mex(igvn); 1609 if (mex.expand_macro_nodes()) { 1610 assert(failing(), "must bail out w/ explicit message"); 1611 return; 1612 } 1613 } 1614 1615 } // (End scope of igvn; run destructor if necessary for asserts.) 1616 1617 // A method with only infinite loops has no edges entering loops from root 1618 { 1619 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1620 if (final_graph_reshaping()) { 1621 assert(failing(), "must bail out w/ explicit message"); 1622 return; 1623 } 1624 } 1625 1626 print_method("Optimize finished", 2); 1627 } 1628 1629 1630 //------------------------------Code_Gen--------------------------------------- 1631 // Given a graph, generate code for it 1632 void Compile::Code_Gen() { 1633 if (failing()) return; 1634 1635 // Perform instruction selection. You might think we could reclaim Matcher 1636 // memory PDQ, but actually the Matcher is used in generating spill code. 1637 // Internals of the Matcher (including some VectorSets) must remain live 1638 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1639 // set a bit in reclaimed memory. 1640 1641 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1642 // nodes. Mapping is only valid at the root of each matched subtree. 1643 NOT_PRODUCT( verify_graph_edges(); ) 1644 1645 Node_List proj_list; 1646 Matcher m(proj_list); 1647 _matcher = &m; 1648 { 1649 TracePhase t2("matcher", &_t_matcher, true); 1650 m.match(); 1651 } 1652 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1653 // nodes. Mapping is only valid at the root of each matched subtree. 1654 NOT_PRODUCT( verify_graph_edges(); ) 1655 1656 // If you have too many nodes, or if matching has failed, bail out 1657 check_node_count(0, "out of nodes matching instructions"); 1658 if (failing()) return; 1659 1660 // Build a proper-looking CFG 1661 PhaseCFG cfg(node_arena(), root(), m); 1662 _cfg = &cfg; 1663 { 1664 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1665 cfg.Dominators(); 1666 if (failing()) return; 1667 1668 NOT_PRODUCT( verify_graph_edges(); ) 1669 1670 cfg.Estimate_Block_Frequency(); 1671 cfg.GlobalCodeMotion(m,unique(),proj_list); 1672 1673 print_method("Global code motion", 2); 1674 1675 if (failing()) return; 1676 NOT_PRODUCT( verify_graph_edges(); ) 1677 1678 debug_only( cfg.verify(); ) 1679 } 1680 NOT_PRODUCT( verify_graph_edges(); ) 1681 1682 PhaseChaitin regalloc(unique(),cfg,m); 1683 _regalloc = ®alloc; 1684 { 1685 TracePhase t2("regalloc", &_t_registerAllocation, true); 1686 // Perform any platform dependent preallocation actions. This is used, 1687 // for example, to avoid taking an implicit null pointer exception 1688 // using the frame pointer on win95. 1689 _regalloc->pd_preallocate_hook(); 1690 1691 // Perform register allocation. After Chaitin, use-def chains are 1692 // no longer accurate (at spill code) and so must be ignored. 1693 // Node->LRG->reg mappings are still accurate. 1694 _regalloc->Register_Allocate(); 1695 1696 // Bail out if the allocator builds too many nodes 1697 if (failing()) return; 1698 } 1699 1700 // Prior to register allocation we kept empty basic blocks in case the 1701 // the allocator needed a place to spill. After register allocation we 1702 // are not adding any new instructions. If any basic block is empty, we 1703 // can now safely remove it. 1704 { 1705 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1706 cfg.remove_empty(); 1707 if (do_freq_based_layout()) { 1708 PhaseBlockLayout layout(cfg); 1709 } else { 1710 cfg.set_loop_alignment(); 1711 } 1712 cfg.fixup_flow(); 1713 } 1714 1715 // Perform any platform dependent postallocation verifications. 1716 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1717 1718 // Apply peephole optimizations 1719 if( OptoPeephole ) { 1720 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1721 PhasePeephole peep( _regalloc, cfg); 1722 peep.do_transform(); 1723 } 1724 1725 // Convert Nodes to instruction bits in a buffer 1726 { 1727 // %%%% workspace merge brought two timers together for one job 1728 TracePhase t2a("output", &_t_output, true); 1729 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1730 Output(); 1731 } 1732 1733 print_method("Final Code"); 1734 1735 // He's dead, Jim. 1736 _cfg = (PhaseCFG*)0xdeadbeef; 1737 _regalloc = (PhaseChaitin*)0xdeadbeef; 1738 } 1739 1740 1741 //------------------------------dump_asm--------------------------------------- 1742 // Dump formatted assembly 1743 #ifndef PRODUCT 1744 void Compile::dump_asm(int *pcs, uint pc_limit) { 1745 bool cut_short = false; 1746 tty->print_cr("#"); 1747 tty->print("# "); _tf->dump(); tty->cr(); 1748 tty->print_cr("#"); 1749 1750 // For all blocks 1751 int pc = 0x0; // Program counter 1752 char starts_bundle = ' '; 1753 _regalloc->dump_frame(); 1754 1755 Node *n = NULL; 1756 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1757 if (VMThread::should_terminate()) { cut_short = true; break; } 1758 Block *b = _cfg->_blocks[i]; 1759 if (b->is_connector() && !Verbose) continue; 1760 n = b->_nodes[0]; 1761 if (pcs && n->_idx < pc_limit) 1762 tty->print("%3.3x ", pcs[n->_idx]); 1763 else 1764 tty->print(" "); 1765 b->dump_head( &_cfg->_bbs ); 1766 if (b->is_connector()) { 1767 tty->print_cr(" # Empty connector block"); 1768 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1769 tty->print_cr(" # Block is sole successor of call"); 1770 } 1771 1772 // For all instructions 1773 Node *delay = NULL; 1774 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1775 if (VMThread::should_terminate()) { cut_short = true; break; } 1776 n = b->_nodes[j]; 1777 if (valid_bundle_info(n)) { 1778 Bundle *bundle = node_bundling(n); 1779 if (bundle->used_in_unconditional_delay()) { 1780 delay = n; 1781 continue; 1782 } 1783 if (bundle->starts_bundle()) 1784 starts_bundle = '+'; 1785 } 1786 1787 if (WizardMode) n->dump(); 1788 1789 if( !n->is_Region() && // Dont print in the Assembly 1790 !n->is_Phi() && // a few noisely useless nodes 1791 !n->is_Proj() && 1792 !n->is_MachTemp() && 1793 !n->is_Catch() && // Would be nice to print exception table targets 1794 !n->is_MergeMem() && // Not very interesting 1795 !n->is_top() && // Debug info table constants 1796 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1797 ) { 1798 if (pcs && n->_idx < pc_limit) 1799 tty->print("%3.3x", pcs[n->_idx]); 1800 else 1801 tty->print(" "); 1802 tty->print(" %c ", starts_bundle); 1803 starts_bundle = ' '; 1804 tty->print("\t"); 1805 n->format(_regalloc, tty); 1806 tty->cr(); 1807 } 1808 1809 // If we have an instruction with a delay slot, and have seen a delay, 1810 // then back up and print it 1811 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1812 assert(delay != NULL, "no unconditional delay instruction"); 1813 if (WizardMode) delay->dump(); 1814 1815 if (node_bundling(delay)->starts_bundle()) 1816 starts_bundle = '+'; 1817 if (pcs && n->_idx < pc_limit) 1818 tty->print("%3.3x", pcs[n->_idx]); 1819 else 1820 tty->print(" "); 1821 tty->print(" %c ", starts_bundle); 1822 starts_bundle = ' '; 1823 tty->print("\t"); 1824 delay->format(_regalloc, tty); 1825 tty->print_cr(""); 1826 delay = NULL; 1827 } 1828 1829 // Dump the exception table as well 1830 if( n->is_Catch() && (Verbose || WizardMode) ) { 1831 // Print the exception table for this offset 1832 _handler_table.print_subtable_for(pc); 1833 } 1834 } 1835 1836 if (pcs && n->_idx < pc_limit) 1837 tty->print_cr("%3.3x", pcs[n->_idx]); 1838 else 1839 tty->print_cr(""); 1840 1841 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1842 1843 } // End of per-block dump 1844 tty->print_cr(""); 1845 1846 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1847 } 1848 #endif 1849 1850 //------------------------------Final_Reshape_Counts--------------------------- 1851 // This class defines counters to help identify when a method 1852 // may/must be executed using hardware with only 24-bit precision. 1853 struct Final_Reshape_Counts : public StackObj { 1854 int _call_count; // count non-inlined 'common' calls 1855 int _float_count; // count float ops requiring 24-bit precision 1856 int _double_count; // count double ops requiring more precision 1857 int _java_call_count; // count non-inlined 'java' calls 1858 int _inner_loop_count; // count loops which need alignment 1859 VectorSet _visited; // Visitation flags 1860 Node_List _tests; // Set of IfNodes & PCTableNodes 1861 1862 Final_Reshape_Counts() : 1863 _call_count(0), _float_count(0), _double_count(0), 1864 _java_call_count(0), _inner_loop_count(0), 1865 _visited( Thread::current()->resource_area() ) { } 1866 1867 void inc_call_count () { _call_count ++; } 1868 void inc_float_count () { _float_count ++; } 1869 void inc_double_count() { _double_count++; } 1870 void inc_java_call_count() { _java_call_count++; } 1871 void inc_inner_loop_count() { _inner_loop_count++; } 1872 1873 int get_call_count () const { return _call_count ; } 1874 int get_float_count () const { return _float_count ; } 1875 int get_double_count() const { return _double_count; } 1876 int get_java_call_count() const { return _java_call_count; } 1877 int get_inner_loop_count() const { return _inner_loop_count; } 1878 }; 1879 1880 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1881 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1882 // Make sure the offset goes inside the instance layout. 1883 return k->contains_field_offset(tp->offset()); 1884 // Note that OffsetBot and OffsetTop are very negative. 1885 } 1886 1887 //------------------------------final_graph_reshaping_impl---------------------- 1888 // Implement items 1-5 from final_graph_reshaping below. 1889 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 1890 1891 if ( n->outcnt() == 0 ) return; // dead node 1892 uint nop = n->Opcode(); 1893 1894 // Check for 2-input instruction with "last use" on right input. 1895 // Swap to left input. Implements item (2). 1896 if( n->req() == 3 && // two-input instruction 1897 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1898 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1899 n->in(2)->outcnt() == 1 &&// right use IS a last use 1900 !n->in(2)->is_Con() ) { // right use is not a constant 1901 // Check for commutative opcode 1902 switch( nop ) { 1903 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1904 case Op_MaxI: case Op_MinI: 1905 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1906 case Op_AndL: case Op_XorL: case Op_OrL: 1907 case Op_AndI: case Op_XorI: case Op_OrI: { 1908 // Move "last use" input to left by swapping inputs 1909 n->swap_edges(1, 2); 1910 break; 1911 } 1912 default: 1913 break; 1914 } 1915 } 1916 1917 // Count FPU ops and common calls, implements item (3) 1918 switch( nop ) { 1919 // Count all float operations that may use FPU 1920 case Op_AddF: 1921 case Op_SubF: 1922 case Op_MulF: 1923 case Op_DivF: 1924 case Op_NegF: 1925 case Op_ModF: 1926 case Op_ConvI2F: 1927 case Op_ConF: 1928 case Op_CmpF: 1929 case Op_CmpF3: 1930 // case Op_ConvL2F: // longs are split into 32-bit halves 1931 frc.inc_float_count(); 1932 break; 1933 1934 case Op_ConvF2D: 1935 case Op_ConvD2F: 1936 frc.inc_float_count(); 1937 frc.inc_double_count(); 1938 break; 1939 1940 // Count all double operations that may use FPU 1941 case Op_AddD: 1942 case Op_SubD: 1943 case Op_MulD: 1944 case Op_DivD: 1945 case Op_NegD: 1946 case Op_ModD: 1947 case Op_ConvI2D: 1948 case Op_ConvD2I: 1949 // case Op_ConvL2D: // handled by leaf call 1950 // case Op_ConvD2L: // handled by leaf call 1951 case Op_ConD: 1952 case Op_CmpD: 1953 case Op_CmpD3: 1954 frc.inc_double_count(); 1955 break; 1956 case Op_Opaque1: // Remove Opaque Nodes before matching 1957 case Op_Opaque2: // Remove Opaque Nodes before matching 1958 n->subsume_by(n->in(1)); 1959 break; 1960 case Op_CallStaticJava: 1961 case Op_CallJava: 1962 case Op_CallDynamicJava: 1963 frc.inc_java_call_count(); // Count java call site; 1964 case Op_CallRuntime: 1965 case Op_CallLeaf: 1966 case Op_CallLeafNoFP: { 1967 assert( n->is_Call(), "" ); 1968 CallNode *call = n->as_Call(); 1969 // Count call sites where the FP mode bit would have to be flipped. 1970 // Do not count uncommon runtime calls: 1971 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 1972 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 1973 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 1974 frc.inc_call_count(); // Count the call site 1975 } else { // See if uncommon argument is shared 1976 Node *n = call->in(TypeFunc::Parms); 1977 int nop = n->Opcode(); 1978 // Clone shared simple arguments to uncommon calls, item (1). 1979 if( n->outcnt() > 1 && 1980 !n->is_Proj() && 1981 nop != Op_CreateEx && 1982 nop != Op_CheckCastPP && 1983 nop != Op_DecodeN && 1984 !n->is_Mem() ) { 1985 Node *x = n->clone(); 1986 call->set_req( TypeFunc::Parms, x ); 1987 } 1988 } 1989 break; 1990 } 1991 1992 case Op_StoreD: 1993 case Op_LoadD: 1994 case Op_LoadD_unaligned: 1995 frc.inc_double_count(); 1996 goto handle_mem; 1997 case Op_StoreF: 1998 case Op_LoadF: 1999 frc.inc_float_count(); 2000 goto handle_mem; 2001 2002 case Op_StoreB: 2003 case Op_StoreC: 2004 case Op_StoreCM: 2005 case Op_StorePConditional: 2006 case Op_StoreI: 2007 case Op_StoreL: 2008 case Op_StoreIConditional: 2009 case Op_StoreLConditional: 2010 case Op_CompareAndSwapI: 2011 case Op_CompareAndSwapL: 2012 case Op_CompareAndSwapP: 2013 case Op_CompareAndSwapN: 2014 case Op_StoreP: 2015 case Op_StoreN: 2016 case Op_LoadB: 2017 case Op_LoadUB: 2018 case Op_LoadUS: 2019 case Op_LoadI: 2020 case Op_LoadUI2L: 2021 case Op_LoadKlass: 2022 case Op_LoadNKlass: 2023 case Op_LoadL: 2024 case Op_LoadL_unaligned: 2025 case Op_LoadPLocked: 2026 case Op_LoadLLocked: 2027 case Op_LoadP: 2028 case Op_LoadN: 2029 case Op_LoadRange: 2030 case Op_LoadS: { 2031 handle_mem: 2032 #ifdef ASSERT 2033 if( VerifyOptoOopOffsets ) { 2034 assert( n->is_Mem(), "" ); 2035 MemNode *mem = (MemNode*)n; 2036 // Check to see if address types have grounded out somehow. 2037 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2038 assert( !tp || oop_offset_is_sane(tp), "" ); 2039 } 2040 #endif 2041 break; 2042 } 2043 2044 case Op_AddP: { // Assert sane base pointers 2045 Node *addp = n->in(AddPNode::Address); 2046 assert( !addp->is_AddP() || 2047 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2048 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2049 "Base pointers must match" ); 2050 #ifdef _LP64 2051 if (UseCompressedOops && 2052 addp->Opcode() == Op_ConP && 2053 addp == n->in(AddPNode::Base) && 2054 n->in(AddPNode::Offset)->is_Con()) { 2055 // Use addressing with narrow klass to load with offset on x86. 2056 // On sparc loading 32-bits constant and decoding it have less 2057 // instructions (4) then load 64-bits constant (7). 2058 // Do this transformation here since IGVN will convert ConN back to ConP. 2059 const Type* t = addp->bottom_type(); 2060 if (t->isa_oopptr()) { 2061 Node* nn = NULL; 2062 2063 // Look for existing ConN node of the same exact type. 2064 Compile* C = Compile::current(); 2065 Node* r = C->root(); 2066 uint cnt = r->outcnt(); 2067 for (uint i = 0; i < cnt; i++) { 2068 Node* m = r->raw_out(i); 2069 if (m!= NULL && m->Opcode() == Op_ConN && 2070 m->bottom_type()->make_ptr() == t) { 2071 nn = m; 2072 break; 2073 } 2074 } 2075 if (nn != NULL) { 2076 // Decode a narrow oop to match address 2077 // [R12 + narrow_oop_reg<<3 + offset] 2078 nn = new (C, 2) DecodeNNode(nn, t); 2079 n->set_req(AddPNode::Base, nn); 2080 n->set_req(AddPNode::Address, nn); 2081 if (addp->outcnt() == 0) { 2082 addp->disconnect_inputs(NULL); 2083 } 2084 } 2085 } 2086 } 2087 #endif 2088 break; 2089 } 2090 2091 #ifdef _LP64 2092 case Op_CastPP: 2093 if (n->in(1)->is_DecodeN() && Universe::narrow_oop_use_implicit_null_checks()) { 2094 Compile* C = Compile::current(); 2095 Node* in1 = n->in(1); 2096 const Type* t = n->bottom_type(); 2097 Node* new_in1 = in1->clone(); 2098 new_in1->as_DecodeN()->set_type(t); 2099 2100 if (!Matcher::clone_shift_expressions) { 2101 // 2102 // x86, ARM and friends can handle 2 adds in addressing mode 2103 // and Matcher can fold a DecodeN node into address by using 2104 // a narrow oop directly and do implicit NULL check in address: 2105 // 2106 // [R12 + narrow_oop_reg<<3 + offset] 2107 // NullCheck narrow_oop_reg 2108 // 2109 // On other platforms (Sparc) we have to keep new DecodeN node and 2110 // use it to do implicit NULL check in address: 2111 // 2112 // decode_not_null narrow_oop_reg, base_reg 2113 // [base_reg + offset] 2114 // NullCheck base_reg 2115 // 2116 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2117 // to keep the information to which NULL check the new DecodeN node 2118 // corresponds to use it as value in implicit_null_check(). 2119 // 2120 new_in1->set_req(0, n->in(0)); 2121 } 2122 2123 n->subsume_by(new_in1); 2124 if (in1->outcnt() == 0) { 2125 in1->disconnect_inputs(NULL); 2126 } 2127 } 2128 break; 2129 2130 case Op_CmpP: 2131 // Do this transformation here to preserve CmpPNode::sub() and 2132 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2133 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2134 Node* in1 = n->in(1); 2135 Node* in2 = n->in(2); 2136 if (!in1->is_DecodeN()) { 2137 in2 = in1; 2138 in1 = n->in(2); 2139 } 2140 assert(in1->is_DecodeN(), "sanity"); 2141 2142 Compile* C = Compile::current(); 2143 Node* new_in2 = NULL; 2144 if (in2->is_DecodeN()) { 2145 new_in2 = in2->in(1); 2146 } else if (in2->Opcode() == Op_ConP) { 2147 const Type* t = in2->bottom_type(); 2148 if (t == TypePtr::NULL_PTR && Universe::narrow_oop_use_implicit_null_checks()) { 2149 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2150 // 2151 // This transformation together with CastPP transformation above 2152 // will generated code for implicit NULL checks for compressed oops. 2153 // 2154 // The original code after Optimize() 2155 // 2156 // LoadN memory, narrow_oop_reg 2157 // decode narrow_oop_reg, base_reg 2158 // CmpP base_reg, NULL 2159 // CastPP base_reg // NotNull 2160 // Load [base_reg + offset], val_reg 2161 // 2162 // after these transformations will be 2163 // 2164 // LoadN memory, narrow_oop_reg 2165 // CmpN narrow_oop_reg, NULL 2166 // decode_not_null narrow_oop_reg, base_reg 2167 // Load [base_reg + offset], val_reg 2168 // 2169 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2170 // since narrow oops can be used in debug info now (see the code in 2171 // final_graph_reshaping_walk()). 2172 // 2173 // At the end the code will be matched to 2174 // on x86: 2175 // 2176 // Load_narrow_oop memory, narrow_oop_reg 2177 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2178 // NullCheck narrow_oop_reg 2179 // 2180 // and on sparc: 2181 // 2182 // Load_narrow_oop memory, narrow_oop_reg 2183 // decode_not_null narrow_oop_reg, base_reg 2184 // Load [base_reg + offset], val_reg 2185 // NullCheck base_reg 2186 // 2187 } else if (t->isa_oopptr()) { 2188 new_in2 = ConNode::make(C, t->make_narrowoop()); 2189 } 2190 } 2191 if (new_in2 != NULL) { 2192 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2193 n->subsume_by( cmpN ); 2194 if (in1->outcnt() == 0) { 2195 in1->disconnect_inputs(NULL); 2196 } 2197 if (in2->outcnt() == 0) { 2198 in2->disconnect_inputs(NULL); 2199 } 2200 } 2201 } 2202 break; 2203 2204 case Op_DecodeN: 2205 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2206 // DecodeN could be pinned on Sparc where it can't be fold into 2207 // an address expression, see the code for Op_CastPP above. 2208 assert(n->in(0) == NULL || !Matcher::clone_shift_expressions, "no control except on sparc"); 2209 break; 2210 2211 case Op_EncodeP: { 2212 Node* in1 = n->in(1); 2213 if (in1->is_DecodeN()) { 2214 n->subsume_by(in1->in(1)); 2215 } else if (in1->Opcode() == Op_ConP) { 2216 Compile* C = Compile::current(); 2217 const Type* t = in1->bottom_type(); 2218 if (t == TypePtr::NULL_PTR) { 2219 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2220 } else if (t->isa_oopptr()) { 2221 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2222 } 2223 } 2224 if (in1->outcnt() == 0) { 2225 in1->disconnect_inputs(NULL); 2226 } 2227 break; 2228 } 2229 2230 case Op_Phi: 2231 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2232 // The EncodeP optimization may create Phi with the same edges 2233 // for all paths. It is not handled well by Register Allocator. 2234 Node* unique_in = n->in(1); 2235 assert(unique_in != NULL, ""); 2236 uint cnt = n->req(); 2237 for (uint i = 2; i < cnt; i++) { 2238 Node* m = n->in(i); 2239 assert(m != NULL, ""); 2240 if (unique_in != m) 2241 unique_in = NULL; 2242 } 2243 if (unique_in != NULL) { 2244 n->subsume_by(unique_in); 2245 } 2246 } 2247 break; 2248 2249 #endif 2250 2251 case Op_ModI: 2252 if (UseDivMod) { 2253 // Check if a%b and a/b both exist 2254 Node* d = n->find_similar(Op_DivI); 2255 if (d) { 2256 // Replace them with a fused divmod if supported 2257 Compile* C = Compile::current(); 2258 if (Matcher::has_match_rule(Op_DivModI)) { 2259 DivModINode* divmod = DivModINode::make(C, n); 2260 d->subsume_by(divmod->div_proj()); 2261 n->subsume_by(divmod->mod_proj()); 2262 } else { 2263 // replace a%b with a-((a/b)*b) 2264 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2265 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2266 n->subsume_by( sub ); 2267 } 2268 } 2269 } 2270 break; 2271 2272 case Op_ModL: 2273 if (UseDivMod) { 2274 // Check if a%b and a/b both exist 2275 Node* d = n->find_similar(Op_DivL); 2276 if (d) { 2277 // Replace them with a fused divmod if supported 2278 Compile* C = Compile::current(); 2279 if (Matcher::has_match_rule(Op_DivModL)) { 2280 DivModLNode* divmod = DivModLNode::make(C, n); 2281 d->subsume_by(divmod->div_proj()); 2282 n->subsume_by(divmod->mod_proj()); 2283 } else { 2284 // replace a%b with a-((a/b)*b) 2285 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2286 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2287 n->subsume_by( sub ); 2288 } 2289 } 2290 } 2291 break; 2292 2293 case Op_Load16B: 2294 case Op_Load8B: 2295 case Op_Load4B: 2296 case Op_Load8S: 2297 case Op_Load4S: 2298 case Op_Load2S: 2299 case Op_Load8C: 2300 case Op_Load4C: 2301 case Op_Load2C: 2302 case Op_Load4I: 2303 case Op_Load2I: 2304 case Op_Load2L: 2305 case Op_Load4F: 2306 case Op_Load2F: 2307 case Op_Load2D: 2308 case Op_Store16B: 2309 case Op_Store8B: 2310 case Op_Store4B: 2311 case Op_Store8C: 2312 case Op_Store4C: 2313 case Op_Store2C: 2314 case Op_Store4I: 2315 case Op_Store2I: 2316 case Op_Store2L: 2317 case Op_Store4F: 2318 case Op_Store2F: 2319 case Op_Store2D: 2320 break; 2321 2322 case Op_PackB: 2323 case Op_PackS: 2324 case Op_PackC: 2325 case Op_PackI: 2326 case Op_PackF: 2327 case Op_PackL: 2328 case Op_PackD: 2329 if (n->req()-1 > 2) { 2330 // Replace many operand PackNodes with a binary tree for matching 2331 PackNode* p = (PackNode*) n; 2332 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2333 n->subsume_by(btp); 2334 } 2335 break; 2336 case Op_Loop: 2337 case Op_CountedLoop: 2338 if (n->as_Loop()->is_inner_loop()) { 2339 frc.inc_inner_loop_count(); 2340 } 2341 break; 2342 default: 2343 assert( !n->is_Call(), "" ); 2344 assert( !n->is_Mem(), "" ); 2345 break; 2346 } 2347 2348 // Collect CFG split points 2349 if (n->is_MultiBranch()) 2350 frc._tests.push(n); 2351 } 2352 2353 //------------------------------final_graph_reshaping_walk--------------------- 2354 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2355 // requires that the walk visits a node's inputs before visiting the node. 2356 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2357 ResourceArea *area = Thread::current()->resource_area(); 2358 Unique_Node_List sfpt(area); 2359 2360 frc._visited.set(root->_idx); // first, mark node as visited 2361 uint cnt = root->req(); 2362 Node *n = root; 2363 uint i = 0; 2364 while (true) { 2365 if (i < cnt) { 2366 // Place all non-visited non-null inputs onto stack 2367 Node* m = n->in(i); 2368 ++i; 2369 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2370 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2371 sfpt.push(m); 2372 cnt = m->req(); 2373 nstack.push(n, i); // put on stack parent and next input's index 2374 n = m; 2375 i = 0; 2376 } 2377 } else { 2378 // Now do post-visit work 2379 final_graph_reshaping_impl( n, frc ); 2380 if (nstack.is_empty()) 2381 break; // finished 2382 n = nstack.node(); // Get node from stack 2383 cnt = n->req(); 2384 i = nstack.index(); 2385 nstack.pop(); // Shift to the next node on stack 2386 } 2387 } 2388 2389 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2390 // It could be done for an uncommon traps or any safepoints/calls 2391 // if the DecodeN node is referenced only in a debug info. 2392 while (sfpt.size() > 0) { 2393 n = sfpt.pop(); 2394 JVMState *jvms = n->as_SafePoint()->jvms(); 2395 assert(jvms != NULL, "sanity"); 2396 int start = jvms->debug_start(); 2397 int end = n->req(); 2398 bool is_uncommon = (n->is_CallStaticJava() && 2399 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2400 for (int j = start; j < end; j++) { 2401 Node* in = n->in(j); 2402 if (in->is_DecodeN()) { 2403 bool safe_to_skip = true; 2404 if (!is_uncommon ) { 2405 // Is it safe to skip? 2406 for (uint i = 0; i < in->outcnt(); i++) { 2407 Node* u = in->raw_out(i); 2408 if (!u->is_SafePoint() || 2409 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2410 safe_to_skip = false; 2411 } 2412 } 2413 } 2414 if (safe_to_skip) { 2415 n->set_req(j, in->in(1)); 2416 } 2417 if (in->outcnt() == 0) { 2418 in->disconnect_inputs(NULL); 2419 } 2420 } 2421 } 2422 } 2423 } 2424 2425 //------------------------------final_graph_reshaping-------------------------- 2426 // Final Graph Reshaping. 2427 // 2428 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2429 // and not commoned up and forced early. Must come after regular 2430 // optimizations to avoid GVN undoing the cloning. Clone constant 2431 // inputs to Loop Phis; these will be split by the allocator anyways. 2432 // Remove Opaque nodes. 2433 // (2) Move last-uses by commutative operations to the left input to encourage 2434 // Intel update-in-place two-address operations and better register usage 2435 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2436 // calls canonicalizing them back. 2437 // (3) Count the number of double-precision FP ops, single-precision FP ops 2438 // and call sites. On Intel, we can get correct rounding either by 2439 // forcing singles to memory (requires extra stores and loads after each 2440 // FP bytecode) or we can set a rounding mode bit (requires setting and 2441 // clearing the mode bit around call sites). The mode bit is only used 2442 // if the relative frequency of single FP ops to calls is low enough. 2443 // This is a key transform for SPEC mpeg_audio. 2444 // (4) Detect infinite loops; blobs of code reachable from above but not 2445 // below. Several of the Code_Gen algorithms fail on such code shapes, 2446 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2447 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2448 // Detection is by looking for IfNodes where only 1 projection is 2449 // reachable from below or CatchNodes missing some targets. 2450 // (5) Assert for insane oop offsets in debug mode. 2451 2452 bool Compile::final_graph_reshaping() { 2453 // an infinite loop may have been eliminated by the optimizer, 2454 // in which case the graph will be empty. 2455 if (root()->req() == 1) { 2456 record_method_not_compilable("trivial infinite loop"); 2457 return true; 2458 } 2459 2460 Final_Reshape_Counts frc; 2461 2462 // Visit everybody reachable! 2463 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2464 Node_Stack nstack(unique() >> 1); 2465 final_graph_reshaping_walk(nstack, root(), frc); 2466 2467 // Check for unreachable (from below) code (i.e., infinite loops). 2468 for( uint i = 0; i < frc._tests.size(); i++ ) { 2469 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2470 // Get number of CFG targets. 2471 // Note that PCTables include exception targets after calls. 2472 uint required_outcnt = n->required_outcnt(); 2473 if (n->outcnt() != required_outcnt) { 2474 // Check for a few special cases. Rethrow Nodes never take the 2475 // 'fall-thru' path, so expected kids is 1 less. 2476 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2477 if (n->in(0)->in(0)->is_Call()) { 2478 CallNode *call = n->in(0)->in(0)->as_Call(); 2479 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2480 required_outcnt--; // Rethrow always has 1 less kid 2481 } else if (call->req() > TypeFunc::Parms && 2482 call->is_CallDynamicJava()) { 2483 // Check for null receiver. In such case, the optimizer has 2484 // detected that the virtual call will always result in a null 2485 // pointer exception. The fall-through projection of this CatchNode 2486 // will not be populated. 2487 Node *arg0 = call->in(TypeFunc::Parms); 2488 if (arg0->is_Type() && 2489 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2490 required_outcnt--; 2491 } 2492 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2493 call->req() > TypeFunc::Parms+1 && 2494 call->is_CallStaticJava()) { 2495 // Check for negative array length. In such case, the optimizer has 2496 // detected that the allocation attempt will always result in an 2497 // exception. There is no fall-through projection of this CatchNode . 2498 Node *arg1 = call->in(TypeFunc::Parms+1); 2499 if (arg1->is_Type() && 2500 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2501 required_outcnt--; 2502 } 2503 } 2504 } 2505 } 2506 // Recheck with a better notion of 'required_outcnt' 2507 if (n->outcnt() != required_outcnt) { 2508 record_method_not_compilable("malformed control flow"); 2509 return true; // Not all targets reachable! 2510 } 2511 } 2512 // Check that I actually visited all kids. Unreached kids 2513 // must be infinite loops. 2514 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2515 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2516 record_method_not_compilable("infinite loop"); 2517 return true; // Found unvisited kid; must be unreach 2518 } 2519 } 2520 2521 // If original bytecodes contained a mixture of floats and doubles 2522 // check if the optimizer has made it homogenous, item (3). 2523 if( Use24BitFPMode && Use24BitFP && 2524 frc.get_float_count() > 32 && 2525 frc.get_double_count() == 0 && 2526 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2527 set_24_bit_selection_and_mode( false, true ); 2528 } 2529 2530 set_java_calls(frc.get_java_call_count()); 2531 set_inner_loops(frc.get_inner_loop_count()); 2532 2533 // No infinite loops, no reason to bail out. 2534 return false; 2535 } 2536 2537 //-----------------------------too_many_traps---------------------------------- 2538 // Report if there are too many traps at the current method and bci. 2539 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2540 bool Compile::too_many_traps(ciMethod* method, 2541 int bci, 2542 Deoptimization::DeoptReason reason) { 2543 ciMethodData* md = method->method_data(); 2544 if (md->is_empty()) { 2545 // Assume the trap has not occurred, or that it occurred only 2546 // because of a transient condition during start-up in the interpreter. 2547 return false; 2548 } 2549 if (md->has_trap_at(bci, reason) != 0) { 2550 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2551 // Also, if there are multiple reasons, or if there is no per-BCI record, 2552 // assume the worst. 2553 if (log()) 2554 log()->elem("observe trap='%s' count='%d'", 2555 Deoptimization::trap_reason_name(reason), 2556 md->trap_count(reason)); 2557 return true; 2558 } else { 2559 // Ignore method/bci and see if there have been too many globally. 2560 return too_many_traps(reason, md); 2561 } 2562 } 2563 2564 // Less-accurate variant which does not require a method and bci. 2565 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2566 ciMethodData* logmd) { 2567 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2568 // Too many traps globally. 2569 // Note that we use cumulative trap_count, not just md->trap_count. 2570 if (log()) { 2571 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2572 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2573 Deoptimization::trap_reason_name(reason), 2574 mcount, trap_count(reason)); 2575 } 2576 return true; 2577 } else { 2578 // The coast is clear. 2579 return false; 2580 } 2581 } 2582 2583 //--------------------------too_many_recompiles-------------------------------- 2584 // Report if there are too many recompiles at the current method and bci. 2585 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2586 // Is not eager to return true, since this will cause the compiler to use 2587 // Action_none for a trap point, to avoid too many recompilations. 2588 bool Compile::too_many_recompiles(ciMethod* method, 2589 int bci, 2590 Deoptimization::DeoptReason reason) { 2591 ciMethodData* md = method->method_data(); 2592 if (md->is_empty()) { 2593 // Assume the trap has not occurred, or that it occurred only 2594 // because of a transient condition during start-up in the interpreter. 2595 return false; 2596 } 2597 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2598 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2599 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2600 Deoptimization::DeoptReason per_bc_reason 2601 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2602 if ((per_bc_reason == Deoptimization::Reason_none 2603 || md->has_trap_at(bci, reason) != 0) 2604 // The trap frequency measure we care about is the recompile count: 2605 && md->trap_recompiled_at(bci) 2606 && md->overflow_recompile_count() >= bc_cutoff) { 2607 // Do not emit a trap here if it has already caused recompilations. 2608 // Also, if there are multiple reasons, or if there is no per-BCI record, 2609 // assume the worst. 2610 if (log()) 2611 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2612 Deoptimization::trap_reason_name(reason), 2613 md->trap_count(reason), 2614 md->overflow_recompile_count()); 2615 return true; 2616 } else if (trap_count(reason) != 0 2617 && decompile_count() >= m_cutoff) { 2618 // Too many recompiles globally, and we have seen this sort of trap. 2619 // Use cumulative decompile_count, not just md->decompile_count. 2620 if (log()) 2621 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2622 Deoptimization::trap_reason_name(reason), 2623 md->trap_count(reason), trap_count(reason), 2624 md->decompile_count(), decompile_count()); 2625 return true; 2626 } else { 2627 // The coast is clear. 2628 return false; 2629 } 2630 } 2631 2632 2633 #ifndef PRODUCT 2634 //------------------------------verify_graph_edges--------------------------- 2635 // Walk the Graph and verify that there is a one-to-one correspondence 2636 // between Use-Def edges and Def-Use edges in the graph. 2637 void Compile::verify_graph_edges(bool no_dead_code) { 2638 if (VerifyGraphEdges) { 2639 ResourceArea *area = Thread::current()->resource_area(); 2640 Unique_Node_List visited(area); 2641 // Call recursive graph walk to check edges 2642 _root->verify_edges(visited); 2643 if (no_dead_code) { 2644 // Now make sure that no visited node is used by an unvisited node. 2645 bool dead_nodes = 0; 2646 Unique_Node_List checked(area); 2647 while (visited.size() > 0) { 2648 Node* n = visited.pop(); 2649 checked.push(n); 2650 for (uint i = 0; i < n->outcnt(); i++) { 2651 Node* use = n->raw_out(i); 2652 if (checked.member(use)) continue; // already checked 2653 if (visited.member(use)) continue; // already in the graph 2654 if (use->is_Con()) continue; // a dead ConNode is OK 2655 // At this point, we have found a dead node which is DU-reachable. 2656 if (dead_nodes++ == 0) 2657 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2658 use->dump(2); 2659 tty->print_cr("---"); 2660 checked.push(use); // No repeats; pretend it is now checked. 2661 } 2662 } 2663 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2664 } 2665 } 2666 } 2667 #endif 2668 2669 // The Compile object keeps track of failure reasons separately from the ciEnv. 2670 // This is required because there is not quite a 1-1 relation between the 2671 // ciEnv and its compilation task and the Compile object. Note that one 2672 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2673 // to backtrack and retry without subsuming loads. Other than this backtracking 2674 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2675 // by the logic in C2Compiler. 2676 void Compile::record_failure(const char* reason) { 2677 if (log() != NULL) { 2678 log()->elem("failure reason='%s' phase='compile'", reason); 2679 } 2680 if (_failure_reason == NULL) { 2681 // Record the first failure reason. 2682 _failure_reason = reason; 2683 } 2684 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2685 C->print_method(_failure_reason); 2686 } 2687 _root = NULL; // flush the graph, too 2688 } 2689 2690 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2691 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2692 { 2693 if (dolog) { 2694 C = Compile::current(); 2695 _log = C->log(); 2696 } else { 2697 C = NULL; 2698 _log = NULL; 2699 } 2700 if (_log != NULL) { 2701 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2702 _log->stamp(); 2703 _log->end_head(); 2704 } 2705 } 2706 2707 Compile::TracePhase::~TracePhase() { 2708 if (_log != NULL) { 2709 _log->done("phase nodes='%d'", C->unique()); 2710 } 2711 }