1 /* 2 * Copyright (c) 1997, 2010, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * 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 void Compile::gvn_replace_by(Node* n, Node* nn) { 228 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 229 Node* use = n->last_out(i); 230 bool is_in_table = initial_gvn()->hash_delete(use); 231 uint uses_found = 0; 232 for (uint j = 0; j < use->len(); j++) { 233 if (use->in(j) == n) { 234 if (j < use->req()) 235 use->set_req(j, nn); 236 else 237 use->set_prec(j, nn); 238 uses_found++; 239 } 240 } 241 if (is_in_table) { 242 // reinsert into table 243 initial_gvn()->hash_find_insert(use); 244 } 245 record_for_igvn(use); 246 i -= uses_found; // we deleted 1 or more copies of this edge 247 } 248 } 249 250 251 252 253 // Identify all nodes that are reachable from below, useful. 254 // Use breadth-first pass that records state in a Unique_Node_List, 255 // recursive traversal is slower. 256 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 257 int estimated_worklist_size = unique(); 258 useful.map( estimated_worklist_size, NULL ); // preallocate space 259 260 // Initialize worklist 261 if (root() != NULL) { useful.push(root()); } 262 // If 'top' is cached, declare it useful to preserve cached node 263 if( cached_top_node() ) { useful.push(cached_top_node()); } 264 265 // Push all useful nodes onto the list, breadthfirst 266 for( uint next = 0; next < useful.size(); ++next ) { 267 assert( next < unique(), "Unique useful nodes < total nodes"); 268 Node *n = useful.at(next); 269 uint max = n->len(); 270 for( uint i = 0; i < max; ++i ) { 271 Node *m = n->in(i); 272 if( m == NULL ) continue; 273 useful.push(m); 274 } 275 } 276 } 277 278 // Disconnect all useless nodes by disconnecting those at the boundary. 279 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 280 uint next = 0; 281 while( next < useful.size() ) { 282 Node *n = useful.at(next++); 283 // Use raw traversal of out edges since this code removes out edges 284 int max = n->outcnt(); 285 for (int j = 0; j < max; ++j ) { 286 Node* child = n->raw_out(j); 287 if( ! useful.member(child) ) { 288 assert( !child->is_top() || child != top(), 289 "If top is cached in Compile object it is in useful list"); 290 // Only need to remove this out-edge to the useless node 291 n->raw_del_out(j); 292 --j; 293 --max; 294 } 295 } 296 if (n->outcnt() == 1 && n->has_special_unique_user()) { 297 record_for_igvn( n->unique_out() ); 298 } 299 } 300 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 301 } 302 303 //------------------------------frame_size_in_words----------------------------- 304 // frame_slots in units of words 305 int Compile::frame_size_in_words() const { 306 // shift is 0 in LP32 and 1 in LP64 307 const int shift = (LogBytesPerWord - LogBytesPerInt); 308 int words = _frame_slots >> shift; 309 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 310 return words; 311 } 312 313 // ============================================================================ 314 //------------------------------CompileWrapper--------------------------------- 315 class CompileWrapper : public StackObj { 316 Compile *const _compile; 317 public: 318 CompileWrapper(Compile* compile); 319 320 ~CompileWrapper(); 321 }; 322 323 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 324 // the Compile* pointer is stored in the current ciEnv: 325 ciEnv* env = compile->env(); 326 assert(env == ciEnv::current(), "must already be a ciEnv active"); 327 assert(env->compiler_data() == NULL, "compile already active?"); 328 env->set_compiler_data(compile); 329 assert(compile == Compile::current(), "sanity"); 330 331 compile->set_type_dict(NULL); 332 compile->set_type_hwm(NULL); 333 compile->set_type_last_size(0); 334 compile->set_last_tf(NULL, NULL); 335 compile->set_indexSet_arena(NULL); 336 compile->set_indexSet_free_block_list(NULL); 337 compile->init_type_arena(); 338 Type::Initialize(compile); 339 _compile->set_scratch_buffer_blob(NULL); 340 _compile->begin_method(); 341 } 342 CompileWrapper::~CompileWrapper() { 343 _compile->end_method(); 344 if (_compile->scratch_buffer_blob() != NULL) 345 BufferBlob::free(_compile->scratch_buffer_blob()); 346 _compile->env()->set_compiler_data(NULL); 347 } 348 349 350 //----------------------------print_compile_messages--------------------------- 351 void Compile::print_compile_messages() { 352 #ifndef PRODUCT 353 // Check if recompiling 354 if (_subsume_loads == false && PrintOpto) { 355 // Recompiling without allowing machine instructions to subsume loads 356 tty->print_cr("*********************************************************"); 357 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 358 tty->print_cr("*********************************************************"); 359 } 360 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 361 // Recompiling without escape analysis 362 tty->print_cr("*********************************************************"); 363 tty->print_cr("** Bailout: Recompile without escape analysis **"); 364 tty->print_cr("*********************************************************"); 365 } 366 if (env()->break_at_compile()) { 367 // Open the debugger when compiling this method. 368 tty->print("### Breaking when compiling: "); 369 method()->print_short_name(); 370 tty->cr(); 371 BREAKPOINT; 372 } 373 374 if( PrintOpto ) { 375 if (is_osr_compilation()) { 376 tty->print("[OSR]%3d", _compile_id); 377 } else { 378 tty->print("%3d", _compile_id); 379 } 380 } 381 #endif 382 } 383 384 385 void Compile::init_scratch_buffer_blob() { 386 if( scratch_buffer_blob() != NULL ) return; 387 388 // Construct a temporary CodeBuffer to have it construct a BufferBlob 389 // Cache this BufferBlob for this compile. 390 ResourceMark rm; 391 int size = (MAX_inst_size + MAX_stubs_size + MAX_const_size); 392 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 393 // Record the buffer blob for next time. 394 set_scratch_buffer_blob(blob); 395 // Have we run out of code space? 396 if (scratch_buffer_blob() == NULL) { 397 // Let CompilerBroker disable further compilations. 398 record_failure("Not enough space for scratch buffer in CodeCache"); 399 return; 400 } 401 402 // Initialize the relocation buffers 403 relocInfo* locs_buf = (relocInfo*) blob->instructions_end() - MAX_locs_size; 404 set_scratch_locs_memory(locs_buf); 405 } 406 407 408 //-----------------------scratch_emit_size------------------------------------- 409 // Helper function that computes size by emitting code 410 uint Compile::scratch_emit_size(const Node* n) { 411 // Emit into a trash buffer and count bytes emitted. 412 // This is a pretty expensive way to compute a size, 413 // but it works well enough if seldom used. 414 // All common fixed-size instructions are given a size 415 // method by the AD file. 416 // Note that the scratch buffer blob and locs memory are 417 // allocated at the beginning of the compile task, and 418 // may be shared by several calls to scratch_emit_size. 419 // The allocation of the scratch buffer blob is particularly 420 // expensive, since it has to grab the code cache lock. 421 BufferBlob* blob = this->scratch_buffer_blob(); 422 assert(blob != NULL, "Initialize BufferBlob at start"); 423 assert(blob->size() > MAX_inst_size, "sanity"); 424 relocInfo* locs_buf = scratch_locs_memory(); 425 address blob_begin = blob->instructions_begin(); 426 address blob_end = (address)locs_buf; 427 assert(blob->instructions_contains(blob_end), "sanity"); 428 CodeBuffer buf(blob_begin, blob_end - blob_begin); 429 buf.initialize_consts_size(MAX_const_size); 430 buf.initialize_stubs_size(MAX_stubs_size); 431 assert(locs_buf != NULL, "sanity"); 432 int lsize = MAX_locs_size / 2; 433 buf.insts()->initialize_shared_locs(&locs_buf[0], lsize); 434 buf.stubs()->initialize_shared_locs(&locs_buf[lsize], lsize); 435 n->emit(buf, this->regalloc()); 436 return buf.code_size(); 437 } 438 439 440 // ============================================================================ 441 //------------------------------Compile standard------------------------------- 442 debug_only( int Compile::_debug_idx = 100000; ) 443 444 // Compile a method. entry_bci is -1 for normal compilations and indicates 445 // the continuation bci for on stack replacement. 446 447 448 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 449 : Phase(Compiler), 450 _env(ci_env), 451 _log(ci_env->log()), 452 _compile_id(ci_env->compile_id()), 453 _save_argument_registers(false), 454 _stub_name(NULL), 455 _stub_function(NULL), 456 _stub_entry_point(NULL), 457 _method(target), 458 _entry_bci(osr_bci), 459 _initial_gvn(NULL), 460 _for_igvn(NULL), 461 _warm_calls(NULL), 462 _subsume_loads(subsume_loads), 463 _do_escape_analysis(do_escape_analysis), 464 _failure_reason(NULL), 465 _code_buffer("Compile::Fill_buffer"), 466 _orig_pc_slot(0), 467 _orig_pc_slot_offset_in_bytes(0), 468 _has_method_handle_invokes(false), 469 _node_bundling_limit(0), 470 _node_bundling_base(NULL), 471 _java_calls(0), 472 _inner_loops(0), 473 #ifndef PRODUCT 474 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 475 _printer(IdealGraphPrinter::printer()), 476 #endif 477 _congraph(NULL) { 478 C = this; 479 480 CompileWrapper cw(this); 481 #ifndef PRODUCT 482 if (TimeCompiler2) { 483 tty->print(" "); 484 target->holder()->name()->print(); 485 tty->print("."); 486 target->print_short_name(); 487 tty->print(" "); 488 } 489 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 490 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 491 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 492 if (!print_opto_assembly) { 493 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 494 if (print_assembly && !Disassembler::can_decode()) { 495 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 496 print_opto_assembly = true; 497 } 498 } 499 set_print_assembly(print_opto_assembly); 500 set_parsed_irreducible_loop(false); 501 #endif 502 503 if (ProfileTraps) { 504 // Make sure the method being compiled gets its own MDO, 505 // so we can at least track the decompile_count(). 506 method()->build_method_data(); 507 } 508 509 Init(::AliasLevel); 510 511 512 print_compile_messages(); 513 514 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 515 _ilt = InlineTree::build_inline_tree_root(); 516 else 517 _ilt = NULL; 518 519 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 520 assert(num_alias_types() >= AliasIdxRaw, ""); 521 522 #define MINIMUM_NODE_HASH 1023 523 // Node list that Iterative GVN will start with 524 Unique_Node_List for_igvn(comp_arena()); 525 set_for_igvn(&for_igvn); 526 527 // GVN that will be run immediately on new nodes 528 uint estimated_size = method()->code_size()*4+64; 529 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 530 PhaseGVN gvn(node_arena(), estimated_size); 531 set_initial_gvn(&gvn); 532 533 { // Scope for timing the parser 534 TracePhase t3("parse", &_t_parser, true); 535 536 // Put top into the hash table ASAP. 537 initial_gvn()->transform_no_reclaim(top()); 538 539 // Set up tf(), start(), and find a CallGenerator. 540 CallGenerator* cg; 541 if (is_osr_compilation()) { 542 const TypeTuple *domain = StartOSRNode::osr_domain(); 543 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 544 init_tf(TypeFunc::make(domain, range)); 545 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 546 initial_gvn()->set_type_bottom(s); 547 init_start(s); 548 cg = CallGenerator::for_osr(method(), entry_bci()); 549 } else { 550 // Normal case. 551 init_tf(TypeFunc::make(method())); 552 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 553 initial_gvn()->set_type_bottom(s); 554 init_start(s); 555 float past_uses = method()->interpreter_invocation_count(); 556 float expected_uses = past_uses; 557 cg = CallGenerator::for_inline(method(), expected_uses); 558 } 559 if (failing()) return; 560 if (cg == NULL) { 561 record_method_not_compilable_all_tiers("cannot parse method"); 562 return; 563 } 564 JVMState* jvms = build_start_state(start(), tf()); 565 if ((jvms = cg->generate(jvms)) == NULL) { 566 record_method_not_compilable("method parse failed"); 567 return; 568 } 569 GraphKit kit(jvms); 570 571 if (!kit.stopped()) { 572 // Accept return values, and transfer control we know not where. 573 // This is done by a special, unique ReturnNode bound to root. 574 return_values(kit.jvms()); 575 } 576 577 if (kit.has_exceptions()) { 578 // Any exceptions that escape from this call must be rethrown 579 // to whatever caller is dynamically above us on the stack. 580 // This is done by a special, unique RethrowNode bound to root. 581 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 582 } 583 584 if (!failing() && has_stringbuilder()) { 585 { 586 // remove useless nodes to make the usage analysis simpler 587 ResourceMark rm; 588 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 589 } 590 591 { 592 ResourceMark rm; 593 print_method("Before StringOpts", 3); 594 PhaseStringOpts pso(initial_gvn(), &for_igvn); 595 print_method("After StringOpts", 3); 596 } 597 598 // now inline anything that we skipped the first time around 599 while (_late_inlines.length() > 0) { 600 CallGenerator* cg = _late_inlines.pop(); 601 cg->do_late_inline(); 602 } 603 } 604 assert(_late_inlines.length() == 0, "should have been processed"); 605 606 print_method("Before RemoveUseless", 3); 607 608 // Remove clutter produced by parsing. 609 if (!failing()) { 610 ResourceMark rm; 611 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 612 } 613 } 614 615 // Note: Large methods are capped off in do_one_bytecode(). 616 if (failing()) return; 617 618 // After parsing, node notes are no longer automagic. 619 // They must be propagated by register_new_node_with_optimizer(), 620 // clone(), or the like. 621 set_default_node_notes(NULL); 622 623 for (;;) { 624 int successes = Inline_Warm(); 625 if (failing()) return; 626 if (successes == 0) break; 627 } 628 629 // Drain the list. 630 Finish_Warm(); 631 #ifndef PRODUCT 632 if (_printer) { 633 _printer->print_inlining(this); 634 } 635 #endif 636 637 if (failing()) return; 638 NOT_PRODUCT( verify_graph_edges(); ) 639 640 // Now optimize 641 Optimize(); 642 if (failing()) return; 643 NOT_PRODUCT( verify_graph_edges(); ) 644 645 #ifndef PRODUCT 646 if (PrintIdeal) { 647 ttyLocker ttyl; // keep the following output all in one block 648 // This output goes directly to the tty, not the compiler log. 649 // To enable tools to match it up with the compilation activity, 650 // be sure to tag this tty output with the compile ID. 651 if (xtty != NULL) { 652 xtty->head("ideal compile_id='%d'%s", compile_id(), 653 is_osr_compilation() ? " compile_kind='osr'" : 654 ""); 655 } 656 root()->dump(9999); 657 if (xtty != NULL) { 658 xtty->tail("ideal"); 659 } 660 } 661 #endif 662 663 // Now that we know the size of all the monitors we can add a fixed slot 664 // for the original deopt pc. 665 666 _orig_pc_slot = fixed_slots(); 667 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 668 set_fixed_slots(next_slot); 669 670 // Now generate code 671 Code_Gen(); 672 if (failing()) return; 673 674 // Check if we want to skip execution of all compiled code. 675 { 676 #ifndef PRODUCT 677 if (OptoNoExecute) { 678 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 679 return; 680 } 681 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 682 #endif 683 684 if (is_osr_compilation()) { 685 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 686 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 687 } else { 688 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 689 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 690 } 691 692 env()->register_method(_method, _entry_bci, 693 &_code_offsets, 694 _orig_pc_slot_offset_in_bytes, 695 code_buffer(), 696 frame_size_in_words(), _oop_map_set, 697 &_handler_table, &_inc_table, 698 compiler, 699 env()->comp_level(), 700 true, /*has_debug_info*/ 701 has_unsafe_access() 702 ); 703 } 704 } 705 706 //------------------------------Compile---------------------------------------- 707 // Compile a runtime stub 708 Compile::Compile( ciEnv* ci_env, 709 TypeFunc_generator generator, 710 address stub_function, 711 const char *stub_name, 712 int is_fancy_jump, 713 bool pass_tls, 714 bool save_arg_registers, 715 bool return_pc ) 716 : Phase(Compiler), 717 _env(ci_env), 718 _log(ci_env->log()), 719 _compile_id(-1), 720 _save_argument_registers(save_arg_registers), 721 _method(NULL), 722 _stub_name(stub_name), 723 _stub_function(stub_function), 724 _stub_entry_point(NULL), 725 _entry_bci(InvocationEntryBci), 726 _initial_gvn(NULL), 727 _for_igvn(NULL), 728 _warm_calls(NULL), 729 _orig_pc_slot(0), 730 _orig_pc_slot_offset_in_bytes(0), 731 _subsume_loads(true), 732 _do_escape_analysis(false), 733 _failure_reason(NULL), 734 _code_buffer("Compile::Fill_buffer"), 735 _has_method_handle_invokes(false), 736 _node_bundling_limit(0), 737 _node_bundling_base(NULL), 738 _java_calls(0), 739 _inner_loops(0), 740 #ifndef PRODUCT 741 _trace_opto_output(TraceOptoOutput), 742 _printer(NULL), 743 #endif 744 _congraph(NULL) { 745 C = this; 746 747 #ifndef PRODUCT 748 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 749 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 750 set_print_assembly(PrintFrameConverterAssembly); 751 set_parsed_irreducible_loop(false); 752 #endif 753 CompileWrapper cw(this); 754 Init(/*AliasLevel=*/ 0); 755 init_tf((*generator)()); 756 757 { 758 // The following is a dummy for the sake of GraphKit::gen_stub 759 Unique_Node_List for_igvn(comp_arena()); 760 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 761 PhaseGVN gvn(Thread::current()->resource_area(),255); 762 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 763 gvn.transform_no_reclaim(top()); 764 765 GraphKit kit; 766 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 767 } 768 769 NOT_PRODUCT( verify_graph_edges(); ) 770 Code_Gen(); 771 if (failing()) return; 772 773 774 // Entry point will be accessed using compile->stub_entry_point(); 775 if (code_buffer() == NULL) { 776 Matcher::soft_match_failure(); 777 } else { 778 if (PrintAssembly && (WizardMode || Verbose)) 779 tty->print_cr("### Stub::%s", stub_name); 780 781 if (!failing()) { 782 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 783 784 // Make the NMethod 785 // For now we mark the frame as never safe for profile stackwalking 786 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 787 code_buffer(), 788 CodeOffsets::frame_never_safe, 789 // _code_offsets.value(CodeOffsets::Frame_Complete), 790 frame_size_in_words(), 791 _oop_map_set, 792 save_arg_registers); 793 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 794 795 _stub_entry_point = rs->entry_point(); 796 } 797 } 798 } 799 800 #ifndef PRODUCT 801 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 802 if(PrintOpto && Verbose) { 803 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 804 } 805 } 806 #endif 807 808 void Compile::print_codes() { 809 } 810 811 //------------------------------Init------------------------------------------- 812 // Prepare for a single compilation 813 void Compile::Init(int aliaslevel) { 814 _unique = 0; 815 _regalloc = NULL; 816 817 _tf = NULL; // filled in later 818 _top = NULL; // cached later 819 _matcher = NULL; // filled in later 820 _cfg = NULL; // filled in later 821 822 set_24_bit_selection_and_mode(Use24BitFP, false); 823 824 _node_note_array = NULL; 825 _default_node_notes = NULL; 826 827 _immutable_memory = NULL; // filled in at first inquiry 828 829 // Globally visible Nodes 830 // First set TOP to NULL to give safe behavior during creation of RootNode 831 set_cached_top_node(NULL); 832 set_root(new (this, 3) RootNode()); 833 // Now that you have a Root to point to, create the real TOP 834 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 835 set_recent_alloc(NULL, NULL); 836 837 // Create Debug Information Recorder to record scopes, oopmaps, etc. 838 env()->set_oop_recorder(new OopRecorder(comp_arena())); 839 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 840 env()->set_dependencies(new Dependencies(env())); 841 842 _fixed_slots = 0; 843 set_has_split_ifs(false); 844 set_has_loops(has_method() && method()->has_loops()); // first approximation 845 set_has_stringbuilder(false); 846 _trap_can_recompile = false; // no traps emitted yet 847 _major_progress = true; // start out assuming good things will happen 848 set_has_unsafe_access(false); 849 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 850 set_decompile_count(0); 851 852 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 853 // Compilation level related initialization 854 if (env()->comp_level() == CompLevel_fast_compile) { 855 set_num_loop_opts(Tier1LoopOptsCount); 856 set_do_inlining(Tier1Inline != 0); 857 set_max_inline_size(Tier1MaxInlineSize); 858 set_freq_inline_size(Tier1FreqInlineSize); 859 set_do_scheduling(false); 860 set_do_count_invocations(Tier1CountInvocations); 861 set_do_method_data_update(Tier1UpdateMethodData); 862 } else { 863 assert(env()->comp_level() == CompLevel_full_optimization, "unknown comp level"); 864 set_num_loop_opts(LoopOptsCount); 865 set_do_inlining(Inline); 866 set_max_inline_size(MaxInlineSize); 867 set_freq_inline_size(FreqInlineSize); 868 set_do_scheduling(OptoScheduling); 869 set_do_count_invocations(false); 870 set_do_method_data_update(false); 871 } 872 873 if (debug_info()->recording_non_safepoints()) { 874 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 875 (comp_arena(), 8, 0, NULL)); 876 set_default_node_notes(Node_Notes::make(this)); 877 } 878 879 // // -- Initialize types before each compile -- 880 // // Update cached type information 881 // if( _method && _method->constants() ) 882 // Type::update_loaded_types(_method, _method->constants()); 883 884 // Init alias_type map. 885 if (!_do_escape_analysis && aliaslevel == 3) 886 aliaslevel = 2; // No unique types without escape analysis 887 _AliasLevel = aliaslevel; 888 const int grow_ats = 16; 889 _max_alias_types = grow_ats; 890 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 891 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 892 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 893 { 894 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 895 } 896 // Initialize the first few types. 897 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 898 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 899 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 900 _num_alias_types = AliasIdxRaw+1; 901 // Zero out the alias type cache. 902 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 903 // A NULL adr_type hits in the cache right away. Preload the right answer. 904 probe_alias_cache(NULL)->_index = AliasIdxTop; 905 906 _intrinsics = NULL; 907 _macro_nodes = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 908 _predicate_opaqs = new GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 909 register_library_intrinsics(); 910 } 911 912 //---------------------------init_start---------------------------------------- 913 // Install the StartNode on this compile object. 914 void Compile::init_start(StartNode* s) { 915 if (failing()) 916 return; // already failing 917 assert(s == start(), ""); 918 } 919 920 StartNode* Compile::start() const { 921 assert(!failing(), ""); 922 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 923 Node* start = root()->fast_out(i); 924 if( start->is_Start() ) 925 return start->as_Start(); 926 } 927 ShouldNotReachHere(); 928 return NULL; 929 } 930 931 //-------------------------------immutable_memory------------------------------------- 932 // Access immutable memory 933 Node* Compile::immutable_memory() { 934 if (_immutable_memory != NULL) { 935 return _immutable_memory; 936 } 937 StartNode* s = start(); 938 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 939 Node *p = s->fast_out(i); 940 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 941 _immutable_memory = p; 942 return _immutable_memory; 943 } 944 } 945 ShouldNotReachHere(); 946 return NULL; 947 } 948 949 //----------------------set_cached_top_node------------------------------------ 950 // Install the cached top node, and make sure Node::is_top works correctly. 951 void Compile::set_cached_top_node(Node* tn) { 952 if (tn != NULL) verify_top(tn); 953 Node* old_top = _top; 954 _top = tn; 955 // Calling Node::setup_is_top allows the nodes the chance to adjust 956 // their _out arrays. 957 if (_top != NULL) _top->setup_is_top(); 958 if (old_top != NULL) old_top->setup_is_top(); 959 assert(_top == NULL || top()->is_top(), ""); 960 } 961 962 #ifndef PRODUCT 963 void Compile::verify_top(Node* tn) const { 964 if (tn != NULL) { 965 assert(tn->is_Con(), "top node must be a constant"); 966 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 967 assert(tn->in(0) != NULL, "must have live top node"); 968 } 969 } 970 #endif 971 972 973 ///-------------------Managing Per-Node Debug & Profile Info------------------- 974 975 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 976 guarantee(arr != NULL, ""); 977 int num_blocks = arr->length(); 978 if (grow_by < num_blocks) grow_by = num_blocks; 979 int num_notes = grow_by * _node_notes_block_size; 980 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 981 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 982 while (num_notes > 0) { 983 arr->append(notes); 984 notes += _node_notes_block_size; 985 num_notes -= _node_notes_block_size; 986 } 987 assert(num_notes == 0, "exact multiple, please"); 988 } 989 990 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 991 if (source == NULL || dest == NULL) return false; 992 993 if (dest->is_Con()) 994 return false; // Do not push debug info onto constants. 995 996 #ifdef ASSERT 997 // Leave a bread crumb trail pointing to the original node: 998 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 999 dest->set_debug_orig(source); 1000 } 1001 #endif 1002 1003 if (node_note_array() == NULL) 1004 return false; // Not collecting any notes now. 1005 1006 // This is a copy onto a pre-existing node, which may already have notes. 1007 // If both nodes have notes, do not overwrite any pre-existing notes. 1008 Node_Notes* source_notes = node_notes_at(source->_idx); 1009 if (source_notes == NULL || source_notes->is_clear()) return false; 1010 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1011 if (dest_notes == NULL || dest_notes->is_clear()) { 1012 return set_node_notes_at(dest->_idx, source_notes); 1013 } 1014 1015 Node_Notes merged_notes = (*source_notes); 1016 // The order of operations here ensures that dest notes will win... 1017 merged_notes.update_from(dest_notes); 1018 return set_node_notes_at(dest->_idx, &merged_notes); 1019 } 1020 1021 1022 //--------------------------allow_range_check_smearing------------------------- 1023 // Gating condition for coalescing similar range checks. 1024 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1025 // single covering check that is at least as strong as any of them. 1026 // If the optimization succeeds, the simplified (strengthened) range check 1027 // will always succeed. If it fails, we will deopt, and then give up 1028 // on the optimization. 1029 bool Compile::allow_range_check_smearing() const { 1030 // If this method has already thrown a range-check, 1031 // assume it was because we already tried range smearing 1032 // and it failed. 1033 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1034 return !already_trapped; 1035 } 1036 1037 1038 //------------------------------flatten_alias_type----------------------------- 1039 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1040 int offset = tj->offset(); 1041 TypePtr::PTR ptr = tj->ptr(); 1042 1043 // Known instance (scalarizable allocation) alias only with itself. 1044 bool is_known_inst = tj->isa_oopptr() != NULL && 1045 tj->is_oopptr()->is_known_instance(); 1046 1047 // Process weird unsafe references. 1048 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1049 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1050 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1051 tj = TypeOopPtr::BOTTOM; 1052 ptr = tj->ptr(); 1053 offset = tj->offset(); 1054 } 1055 1056 // Array pointers need some flattening 1057 const TypeAryPtr *ta = tj->isa_aryptr(); 1058 if( ta && is_known_inst ) { 1059 if ( offset != Type::OffsetBot && 1060 offset > arrayOopDesc::length_offset_in_bytes() ) { 1061 offset = Type::OffsetBot; // Flatten constant access into array body only 1062 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1063 } 1064 } else if( ta && _AliasLevel >= 2 ) { 1065 // For arrays indexed by constant indices, we flatten the alias 1066 // space to include all of the array body. Only the header, klass 1067 // and array length can be accessed un-aliased. 1068 if( offset != Type::OffsetBot ) { 1069 if( ta->const_oop() ) { // methodDataOop or methodOop 1070 offset = Type::OffsetBot; // Flatten constant access into array body 1071 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1072 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1073 // range is OK as-is. 1074 tj = ta = TypeAryPtr::RANGE; 1075 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1076 tj = TypeInstPtr::KLASS; // all klass loads look alike 1077 ta = TypeAryPtr::RANGE; // generic ignored junk 1078 ptr = TypePtr::BotPTR; 1079 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1080 tj = TypeInstPtr::MARK; 1081 ta = TypeAryPtr::RANGE; // generic ignored junk 1082 ptr = TypePtr::BotPTR; 1083 } else { // Random constant offset into array body 1084 offset = Type::OffsetBot; // Flatten constant access into array body 1085 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1086 } 1087 } 1088 // Arrays of fixed size alias with arrays of unknown size. 1089 if (ta->size() != TypeInt::POS) { 1090 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1091 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1092 } 1093 // Arrays of known objects become arrays of unknown objects. 1094 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1095 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1096 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1097 } 1098 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1099 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1100 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1101 } 1102 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1103 // cannot be distinguished by bytecode alone. 1104 if (ta->elem() == TypeInt::BOOL) { 1105 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1106 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1107 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1108 } 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 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1113 if (ta->const_oop()) { 1114 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1115 } else { 1116 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1117 } 1118 } 1119 } 1120 1121 // Oop pointers need some flattening 1122 const TypeInstPtr *to = tj->isa_instptr(); 1123 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1124 if( ptr == TypePtr::Constant ) { 1125 // No constant oop pointers (such as Strings); they alias with 1126 // unknown strings. 1127 assert(!is_known_inst, "not scalarizable allocation"); 1128 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1129 } else if( is_known_inst ) { 1130 tj = to; // Keep NotNull and klass_is_exact for instance type 1131 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1132 // During the 2nd round of IterGVN, NotNull castings are removed. 1133 // Make sure the Bottom and NotNull variants alias the same. 1134 // Also, make sure exact and non-exact variants alias the same. 1135 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1136 } 1137 // Canonicalize the holder of this field 1138 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1139 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1140 // First handle header references such as a LoadKlassNode, even if the 1141 // object's klass is unloaded at compile time (4965979). 1142 if (!is_known_inst) { // Do it only for non-instance types 1143 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1144 } 1145 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1146 to = NULL; 1147 tj = TypeOopPtr::BOTTOM; 1148 offset = tj->offset(); 1149 } else { 1150 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1151 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1152 if( is_known_inst ) { 1153 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1154 } else { 1155 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1156 } 1157 } 1158 } 1159 } 1160 1161 // Klass pointers to object array klasses need some flattening 1162 const TypeKlassPtr *tk = tj->isa_klassptr(); 1163 if( tk ) { 1164 // If we are referencing a field within a Klass, we need 1165 // to assume the worst case of an Object. Both exact and 1166 // inexact types must flatten to the same alias class. 1167 // Since the flattened result for a klass is defined to be 1168 // precisely java.lang.Object, use a constant ptr. 1169 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1170 1171 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1172 TypeKlassPtr::OBJECT->klass(), 1173 offset); 1174 } 1175 1176 ciKlass* klass = tk->klass(); 1177 if( klass->is_obj_array_klass() ) { 1178 ciKlass* k = TypeAryPtr::OOPS->klass(); 1179 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1180 k = TypeInstPtr::BOTTOM->klass(); 1181 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1182 } 1183 1184 // Check for precise loads from the primary supertype array and force them 1185 // to the supertype cache alias index. Check for generic array loads from 1186 // the primary supertype array and also force them to the supertype cache 1187 // alias index. Since the same load can reach both, we need to merge 1188 // these 2 disparate memories into the same alias class. Since the 1189 // primary supertype array is read-only, there's no chance of confusion 1190 // where we bypass an array load and an array store. 1191 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1192 if( offset == Type::OffsetBot || 1193 off2 < Klass::primary_super_limit()*wordSize ) { 1194 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1195 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1196 } 1197 } 1198 1199 // Flatten all Raw pointers together. 1200 if (tj->base() == Type::RawPtr) 1201 tj = TypeRawPtr::BOTTOM; 1202 1203 if (tj->base() == Type::AnyPtr) 1204 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1205 1206 // Flatten all to bottom for now 1207 switch( _AliasLevel ) { 1208 case 0: 1209 tj = TypePtr::BOTTOM; 1210 break; 1211 case 1: // Flatten to: oop, static, field or array 1212 switch (tj->base()) { 1213 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1214 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1215 case Type::AryPtr: // do not distinguish arrays at all 1216 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1217 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1218 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1219 default: ShouldNotReachHere(); 1220 } 1221 break; 1222 case 2: // No collapsing at level 2; keep all splits 1223 case 3: // No collapsing at level 3; keep all splits 1224 break; 1225 default: 1226 Unimplemented(); 1227 } 1228 1229 offset = tj->offset(); 1230 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1231 1232 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1233 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1234 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1235 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1236 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1237 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1238 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1239 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1240 assert( tj->ptr() != TypePtr::TopPTR && 1241 tj->ptr() != TypePtr::AnyNull && 1242 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1243 // assert( tj->ptr() != TypePtr::Constant || 1244 // tj->base() == Type::RawPtr || 1245 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1246 1247 return tj; 1248 } 1249 1250 void Compile::AliasType::Init(int i, const TypePtr* at) { 1251 _index = i; 1252 _adr_type = at; 1253 _field = NULL; 1254 _is_rewritable = true; // default 1255 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1256 if (atoop != NULL && atoop->is_known_instance()) { 1257 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1258 _general_index = Compile::current()->get_alias_index(gt); 1259 } else { 1260 _general_index = 0; 1261 } 1262 } 1263 1264 //---------------------------------print_on------------------------------------ 1265 #ifndef PRODUCT 1266 void Compile::AliasType::print_on(outputStream* st) { 1267 if (index() < 10) 1268 st->print("@ <%d> ", index()); 1269 else st->print("@ <%d>", index()); 1270 st->print(is_rewritable() ? " " : " RO"); 1271 int offset = adr_type()->offset(); 1272 if (offset == Type::OffsetBot) 1273 st->print(" +any"); 1274 else st->print(" +%-3d", offset); 1275 st->print(" in "); 1276 adr_type()->dump_on(st); 1277 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1278 if (field() != NULL && tjp) { 1279 if (tjp->klass() != field()->holder() || 1280 tjp->offset() != field()->offset_in_bytes()) { 1281 st->print(" != "); 1282 field()->print(); 1283 st->print(" ***"); 1284 } 1285 } 1286 } 1287 1288 void print_alias_types() { 1289 Compile* C = Compile::current(); 1290 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1291 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1292 C->alias_type(idx)->print_on(tty); 1293 tty->cr(); 1294 } 1295 } 1296 #endif 1297 1298 1299 //----------------------------probe_alias_cache-------------------------------- 1300 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1301 intptr_t key = (intptr_t) adr_type; 1302 key ^= key >> logAliasCacheSize; 1303 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1304 } 1305 1306 1307 //-----------------------------grow_alias_types-------------------------------- 1308 void Compile::grow_alias_types() { 1309 const int old_ats = _max_alias_types; // how many before? 1310 const int new_ats = old_ats; // how many more? 1311 const int grow_ats = old_ats+new_ats; // how many now? 1312 _max_alias_types = grow_ats; 1313 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1314 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1315 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1316 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1317 } 1318 1319 1320 //--------------------------------find_alias_type------------------------------ 1321 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1322 if (_AliasLevel == 0) 1323 return alias_type(AliasIdxBot); 1324 1325 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1326 if (ace->_adr_type == adr_type) { 1327 return alias_type(ace->_index); 1328 } 1329 1330 // Handle special cases. 1331 if (adr_type == NULL) return alias_type(AliasIdxTop); 1332 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1333 1334 // Do it the slow way. 1335 const TypePtr* flat = flatten_alias_type(adr_type); 1336 1337 #ifdef ASSERT 1338 assert(flat == flatten_alias_type(flat), "idempotent"); 1339 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1340 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1341 const TypeOopPtr* foop = flat->is_oopptr(); 1342 // Scalarizable allocations have exact klass always. 1343 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1344 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1345 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1346 } 1347 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1348 #endif 1349 1350 int idx = AliasIdxTop; 1351 for (int i = 0; i < num_alias_types(); i++) { 1352 if (alias_type(i)->adr_type() == flat) { 1353 idx = i; 1354 break; 1355 } 1356 } 1357 1358 if (idx == AliasIdxTop) { 1359 if (no_create) return NULL; 1360 // Grow the array if necessary. 1361 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1362 // Add a new alias type. 1363 idx = _num_alias_types++; 1364 _alias_types[idx]->Init(idx, flat); 1365 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1366 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1367 if (flat->isa_instptr()) { 1368 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1369 && flat->is_instptr()->klass() == env()->Class_klass()) 1370 alias_type(idx)->set_rewritable(false); 1371 } 1372 if (flat->isa_klassptr()) { 1373 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1374 alias_type(idx)->set_rewritable(false); 1375 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1376 alias_type(idx)->set_rewritable(false); 1377 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1378 alias_type(idx)->set_rewritable(false); 1379 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1380 alias_type(idx)->set_rewritable(false); 1381 } 1382 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1383 // but the base pointer type is not distinctive enough to identify 1384 // references into JavaThread.) 1385 1386 // Check for final instance fields. 1387 const TypeInstPtr* tinst = flat->isa_instptr(); 1388 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1389 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1390 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1391 // Set field() and is_rewritable() attributes. 1392 if (field != NULL) alias_type(idx)->set_field(field); 1393 } 1394 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1395 // Check for final static fields. 1396 if (tklass && tklass->klass()->is_instance_klass()) { 1397 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1398 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1399 // Set field() and is_rewritable() attributes. 1400 if (field != NULL) alias_type(idx)->set_field(field); 1401 } 1402 } 1403 1404 // Fill the cache for next time. 1405 ace->_adr_type = adr_type; 1406 ace->_index = idx; 1407 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1408 1409 // Might as well try to fill the cache for the flattened version, too. 1410 AliasCacheEntry* face = probe_alias_cache(flat); 1411 if (face->_adr_type == NULL) { 1412 face->_adr_type = flat; 1413 face->_index = idx; 1414 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1415 } 1416 1417 return alias_type(idx); 1418 } 1419 1420 1421 Compile::AliasType* Compile::alias_type(ciField* field) { 1422 const TypeOopPtr* t; 1423 if (field->is_static()) 1424 t = TypeKlassPtr::make(field->holder()); 1425 else 1426 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1427 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1428 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1429 return atp; 1430 } 1431 1432 1433 //------------------------------have_alias_type-------------------------------- 1434 bool Compile::have_alias_type(const TypePtr* adr_type) { 1435 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1436 if (ace->_adr_type == adr_type) { 1437 return true; 1438 } 1439 1440 // Handle special cases. 1441 if (adr_type == NULL) return true; 1442 if (adr_type == TypePtr::BOTTOM) return true; 1443 1444 return find_alias_type(adr_type, true) != NULL; 1445 } 1446 1447 //-----------------------------must_alias-------------------------------------- 1448 // True if all values of the given address type are in the given alias category. 1449 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1450 if (alias_idx == AliasIdxBot) return true; // the universal category 1451 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1452 if (alias_idx == AliasIdxTop) return false; // the empty category 1453 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1454 1455 // the only remaining possible overlap is identity 1456 int adr_idx = get_alias_index(adr_type); 1457 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1458 assert(adr_idx == alias_idx || 1459 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1460 && adr_type != TypeOopPtr::BOTTOM), 1461 "should not be testing for overlap with an unsafe pointer"); 1462 return adr_idx == alias_idx; 1463 } 1464 1465 //------------------------------can_alias-------------------------------------- 1466 // True if any values of the given address type are in the given alias category. 1467 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1468 if (alias_idx == AliasIdxTop) return false; // the empty category 1469 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1470 if (alias_idx == AliasIdxBot) return true; // the universal category 1471 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1472 1473 // the only remaining possible overlap is identity 1474 int adr_idx = get_alias_index(adr_type); 1475 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1476 return adr_idx == alias_idx; 1477 } 1478 1479 1480 1481 //---------------------------pop_warm_call------------------------------------- 1482 WarmCallInfo* Compile::pop_warm_call() { 1483 WarmCallInfo* wci = _warm_calls; 1484 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1485 return wci; 1486 } 1487 1488 //----------------------------Inline_Warm-------------------------------------- 1489 int Compile::Inline_Warm() { 1490 // If there is room, try to inline some more warm call sites. 1491 // %%% Do a graph index compaction pass when we think we're out of space? 1492 if (!InlineWarmCalls) return 0; 1493 1494 int calls_made_hot = 0; 1495 int room_to_grow = NodeCountInliningCutoff - unique(); 1496 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1497 int amount_grown = 0; 1498 WarmCallInfo* call; 1499 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1500 int est_size = (int)call->size(); 1501 if (est_size > (room_to_grow - amount_grown)) { 1502 // This one won't fit anyway. Get rid of it. 1503 call->make_cold(); 1504 continue; 1505 } 1506 call->make_hot(); 1507 calls_made_hot++; 1508 amount_grown += est_size; 1509 amount_to_grow -= est_size; 1510 } 1511 1512 if (calls_made_hot > 0) set_major_progress(); 1513 return calls_made_hot; 1514 } 1515 1516 1517 //----------------------------Finish_Warm-------------------------------------- 1518 void Compile::Finish_Warm() { 1519 if (!InlineWarmCalls) return; 1520 if (failing()) return; 1521 if (warm_calls() == NULL) return; 1522 1523 // Clean up loose ends, if we are out of space for inlining. 1524 WarmCallInfo* call; 1525 while ((call = pop_warm_call()) != NULL) { 1526 call->make_cold(); 1527 } 1528 } 1529 1530 //---------------------cleanup_loop_predicates----------------------- 1531 // Remove the opaque nodes that protect the predicates so that all unused 1532 // checks and uncommon_traps will be eliminated from the ideal graph 1533 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1534 if (predicate_count()==0) return; 1535 for (int i = predicate_count(); i > 0; i--) { 1536 Node * n = predicate_opaque1_node(i-1); 1537 assert(n->Opcode() == Op_Opaque1, "must be"); 1538 igvn.replace_node(n, n->in(1)); 1539 } 1540 assert(predicate_count()==0, "should be clean!"); 1541 igvn.optimize(); 1542 } 1543 1544 //------------------------------Optimize--------------------------------------- 1545 // Given a graph, optimize it. 1546 void Compile::Optimize() { 1547 TracePhase t1("optimizer", &_t_optimizer, true); 1548 1549 #ifndef PRODUCT 1550 if (env()->break_at_compile()) { 1551 BREAKPOINT; 1552 } 1553 1554 #endif 1555 1556 ResourceMark rm; 1557 int loop_opts_cnt; 1558 1559 NOT_PRODUCT( verify_graph_edges(); ) 1560 1561 print_method("After Parsing"); 1562 1563 { 1564 // Iterative Global Value Numbering, including ideal transforms 1565 // Initialize IterGVN with types and values from parse-time GVN 1566 PhaseIterGVN igvn(initial_gvn()); 1567 { 1568 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1569 igvn.optimize(); 1570 } 1571 1572 print_method("Iter GVN 1", 2); 1573 1574 if (failing()) return; 1575 1576 // Perform escape analysis 1577 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1578 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 1579 ConnectionGraph::do_analysis(this, &igvn); 1580 1581 if (failing()) return; 1582 1583 igvn.optimize(); 1584 print_method("Iter GVN 3", 2); 1585 1586 if (failing()) return; 1587 1588 } 1589 1590 // Loop transforms on the ideal graph. Range Check Elimination, 1591 // peeling, unrolling, etc. 1592 1593 // Set loop opts counter 1594 loop_opts_cnt = num_loop_opts(); 1595 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1596 { 1597 TracePhase t2("idealLoop", &_t_idealLoop, true); 1598 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate); 1599 loop_opts_cnt--; 1600 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1601 if (failing()) return; 1602 } 1603 // Loop opts pass if partial peeling occurred in previous pass 1604 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1605 TracePhase t3("idealLoop", &_t_idealLoop, true); 1606 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1607 loop_opts_cnt--; 1608 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1609 if (failing()) return; 1610 } 1611 // Loop opts pass for loop-unrolling before CCP 1612 if(major_progress() && (loop_opts_cnt > 0)) { 1613 TracePhase t4("idealLoop", &_t_idealLoop, true); 1614 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1615 loop_opts_cnt--; 1616 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1617 } 1618 if (!failing()) { 1619 // Verify that last round of loop opts produced a valid graph 1620 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1621 PhaseIdealLoop::verify(igvn); 1622 } 1623 } 1624 if (failing()) return; 1625 1626 // Conditional Constant Propagation; 1627 PhaseCCP ccp( &igvn ); 1628 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1629 { 1630 TracePhase t2("ccp", &_t_ccp, true); 1631 ccp.do_transform(); 1632 } 1633 print_method("PhaseCPP 1", 2); 1634 1635 assert( true, "Break here to ccp.dump_old2new_map()"); 1636 1637 // Iterative Global Value Numbering, including ideal transforms 1638 { 1639 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1640 igvn = ccp; 1641 igvn.optimize(); 1642 } 1643 1644 print_method("Iter GVN 2", 2); 1645 1646 if (failing()) return; 1647 1648 // Loop transforms on the ideal graph. Range Check Elimination, 1649 // peeling, unrolling, etc. 1650 if(loop_opts_cnt > 0) { 1651 debug_only( int cnt = 0; ); 1652 bool loop_predication = UseLoopPredicate; 1653 while(major_progress() && (loop_opts_cnt > 0)) { 1654 TracePhase t2("idealLoop", &_t_idealLoop, true); 1655 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1656 PhaseIdealLoop ideal_loop( igvn, true, loop_predication); 1657 loop_opts_cnt--; 1658 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1659 if (failing()) return; 1660 // Perform loop predication optimization during first iteration after CCP. 1661 // After that switch it off and cleanup unused loop predicates. 1662 if (loop_predication) { 1663 loop_predication = false; 1664 cleanup_loop_predicates(igvn); 1665 if (failing()) return; 1666 } 1667 } 1668 } 1669 1670 { 1671 // Verify that all previous optimizations produced a valid graph 1672 // at least to this point, even if no loop optimizations were done. 1673 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1674 PhaseIdealLoop::verify(igvn); 1675 } 1676 1677 { 1678 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1679 PhaseMacroExpand mex(igvn); 1680 if (mex.expand_macro_nodes()) { 1681 assert(failing(), "must bail out w/ explicit message"); 1682 return; 1683 } 1684 } 1685 1686 } // (End scope of igvn; run destructor if necessary for asserts.) 1687 1688 // A method with only infinite loops has no edges entering loops from root 1689 { 1690 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1691 if (final_graph_reshaping()) { 1692 assert(failing(), "must bail out w/ explicit message"); 1693 return; 1694 } 1695 } 1696 1697 print_method("Optimize finished", 2); 1698 } 1699 1700 1701 //------------------------------Code_Gen--------------------------------------- 1702 // Given a graph, generate code for it 1703 void Compile::Code_Gen() { 1704 if (failing()) return; 1705 1706 // Perform instruction selection. You might think we could reclaim Matcher 1707 // memory PDQ, but actually the Matcher is used in generating spill code. 1708 // Internals of the Matcher (including some VectorSets) must remain live 1709 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1710 // set a bit in reclaimed memory. 1711 1712 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1713 // nodes. Mapping is only valid at the root of each matched subtree. 1714 NOT_PRODUCT( verify_graph_edges(); ) 1715 1716 Node_List proj_list; 1717 Matcher m(proj_list); 1718 _matcher = &m; 1719 { 1720 TracePhase t2("matcher", &_t_matcher, true); 1721 m.match(); 1722 } 1723 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1724 // nodes. Mapping is only valid at the root of each matched subtree. 1725 NOT_PRODUCT( verify_graph_edges(); ) 1726 1727 // If you have too many nodes, or if matching has failed, bail out 1728 check_node_count(0, "out of nodes matching instructions"); 1729 if (failing()) return; 1730 1731 // Build a proper-looking CFG 1732 PhaseCFG cfg(node_arena(), root(), m); 1733 _cfg = &cfg; 1734 { 1735 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1736 cfg.Dominators(); 1737 if (failing()) return; 1738 1739 NOT_PRODUCT( verify_graph_edges(); ) 1740 1741 cfg.Estimate_Block_Frequency(); 1742 cfg.GlobalCodeMotion(m,unique(),proj_list); 1743 1744 print_method("Global code motion", 2); 1745 1746 if (failing()) return; 1747 NOT_PRODUCT( verify_graph_edges(); ) 1748 1749 debug_only( cfg.verify(); ) 1750 } 1751 NOT_PRODUCT( verify_graph_edges(); ) 1752 1753 PhaseChaitin regalloc(unique(),cfg,m); 1754 _regalloc = ®alloc; 1755 { 1756 TracePhase t2("regalloc", &_t_registerAllocation, true); 1757 // Perform any platform dependent preallocation actions. This is used, 1758 // for example, to avoid taking an implicit null pointer exception 1759 // using the frame pointer on win95. 1760 _regalloc->pd_preallocate_hook(); 1761 1762 // Perform register allocation. After Chaitin, use-def chains are 1763 // no longer accurate (at spill code) and so must be ignored. 1764 // Node->LRG->reg mappings are still accurate. 1765 _regalloc->Register_Allocate(); 1766 1767 // Bail out if the allocator builds too many nodes 1768 if (failing()) return; 1769 } 1770 1771 // Prior to register allocation we kept empty basic blocks in case the 1772 // the allocator needed a place to spill. After register allocation we 1773 // are not adding any new instructions. If any basic block is empty, we 1774 // can now safely remove it. 1775 { 1776 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1777 cfg.remove_empty(); 1778 if (do_freq_based_layout()) { 1779 PhaseBlockLayout layout(cfg); 1780 } else { 1781 cfg.set_loop_alignment(); 1782 } 1783 cfg.fixup_flow(); 1784 } 1785 1786 // Perform any platform dependent postallocation verifications. 1787 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1788 1789 // Apply peephole optimizations 1790 if( OptoPeephole ) { 1791 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1792 PhasePeephole peep( _regalloc, cfg); 1793 peep.do_transform(); 1794 } 1795 1796 // Convert Nodes to instruction bits in a buffer 1797 { 1798 // %%%% workspace merge brought two timers together for one job 1799 TracePhase t2a("output", &_t_output, true); 1800 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1801 Output(); 1802 } 1803 1804 print_method("Final Code"); 1805 1806 // He's dead, Jim. 1807 _cfg = (PhaseCFG*)0xdeadbeef; 1808 _regalloc = (PhaseChaitin*)0xdeadbeef; 1809 } 1810 1811 1812 //------------------------------dump_asm--------------------------------------- 1813 // Dump formatted assembly 1814 #ifndef PRODUCT 1815 void Compile::dump_asm(int *pcs, uint pc_limit) { 1816 bool cut_short = false; 1817 tty->print_cr("#"); 1818 tty->print("# "); _tf->dump(); tty->cr(); 1819 tty->print_cr("#"); 1820 1821 // For all blocks 1822 int pc = 0x0; // Program counter 1823 char starts_bundle = ' '; 1824 _regalloc->dump_frame(); 1825 1826 Node *n = NULL; 1827 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1828 if (VMThread::should_terminate()) { cut_short = true; break; } 1829 Block *b = _cfg->_blocks[i]; 1830 if (b->is_connector() && !Verbose) continue; 1831 n = b->_nodes[0]; 1832 if (pcs && n->_idx < pc_limit) 1833 tty->print("%3.3x ", pcs[n->_idx]); 1834 else 1835 tty->print(" "); 1836 b->dump_head( &_cfg->_bbs ); 1837 if (b->is_connector()) { 1838 tty->print_cr(" # Empty connector block"); 1839 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1840 tty->print_cr(" # Block is sole successor of call"); 1841 } 1842 1843 // For all instructions 1844 Node *delay = NULL; 1845 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1846 if (VMThread::should_terminate()) { cut_short = true; break; } 1847 n = b->_nodes[j]; 1848 if (valid_bundle_info(n)) { 1849 Bundle *bundle = node_bundling(n); 1850 if (bundle->used_in_unconditional_delay()) { 1851 delay = n; 1852 continue; 1853 } 1854 if (bundle->starts_bundle()) 1855 starts_bundle = '+'; 1856 } 1857 1858 if (WizardMode) n->dump(); 1859 1860 if( !n->is_Region() && // Dont print in the Assembly 1861 !n->is_Phi() && // a few noisely useless nodes 1862 !n->is_Proj() && 1863 !n->is_MachTemp() && 1864 !n->is_SafePointScalarObject() && 1865 !n->is_Catch() && // Would be nice to print exception table targets 1866 !n->is_MergeMem() && // Not very interesting 1867 !n->is_top() && // Debug info table constants 1868 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1869 ) { 1870 if (pcs && n->_idx < pc_limit) 1871 tty->print("%3.3x", pcs[n->_idx]); 1872 else 1873 tty->print(" "); 1874 tty->print(" %c ", starts_bundle); 1875 starts_bundle = ' '; 1876 tty->print("\t"); 1877 n->format(_regalloc, tty); 1878 tty->cr(); 1879 } 1880 1881 // If we have an instruction with a delay slot, and have seen a delay, 1882 // then back up and print it 1883 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1884 assert(delay != NULL, "no unconditional delay instruction"); 1885 if (WizardMode) delay->dump(); 1886 1887 if (node_bundling(delay)->starts_bundle()) 1888 starts_bundle = '+'; 1889 if (pcs && n->_idx < pc_limit) 1890 tty->print("%3.3x", pcs[n->_idx]); 1891 else 1892 tty->print(" "); 1893 tty->print(" %c ", starts_bundle); 1894 starts_bundle = ' '; 1895 tty->print("\t"); 1896 delay->format(_regalloc, tty); 1897 tty->print_cr(""); 1898 delay = NULL; 1899 } 1900 1901 // Dump the exception table as well 1902 if( n->is_Catch() && (Verbose || WizardMode) ) { 1903 // Print the exception table for this offset 1904 _handler_table.print_subtable_for(pc); 1905 } 1906 } 1907 1908 if (pcs && n->_idx < pc_limit) 1909 tty->print_cr("%3.3x", pcs[n->_idx]); 1910 else 1911 tty->print_cr(""); 1912 1913 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1914 1915 } // End of per-block dump 1916 tty->print_cr(""); 1917 1918 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1919 } 1920 #endif 1921 1922 //------------------------------Final_Reshape_Counts--------------------------- 1923 // This class defines counters to help identify when a method 1924 // may/must be executed using hardware with only 24-bit precision. 1925 struct Final_Reshape_Counts : public StackObj { 1926 int _call_count; // count non-inlined 'common' calls 1927 int _float_count; // count float ops requiring 24-bit precision 1928 int _double_count; // count double ops requiring more precision 1929 int _java_call_count; // count non-inlined 'java' calls 1930 int _inner_loop_count; // count loops which need alignment 1931 VectorSet _visited; // Visitation flags 1932 Node_List _tests; // Set of IfNodes & PCTableNodes 1933 1934 Final_Reshape_Counts() : 1935 _call_count(0), _float_count(0), _double_count(0), 1936 _java_call_count(0), _inner_loop_count(0), 1937 _visited( Thread::current()->resource_area() ) { } 1938 1939 void inc_call_count () { _call_count ++; } 1940 void inc_float_count () { _float_count ++; } 1941 void inc_double_count() { _double_count++; } 1942 void inc_java_call_count() { _java_call_count++; } 1943 void inc_inner_loop_count() { _inner_loop_count++; } 1944 1945 int get_call_count () const { return _call_count ; } 1946 int get_float_count () const { return _float_count ; } 1947 int get_double_count() const { return _double_count; } 1948 int get_java_call_count() const { return _java_call_count; } 1949 int get_inner_loop_count() const { return _inner_loop_count; } 1950 }; 1951 1952 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1953 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1954 // Make sure the offset goes inside the instance layout. 1955 return k->contains_field_offset(tp->offset()); 1956 // Note that OffsetBot and OffsetTop are very negative. 1957 } 1958 1959 //------------------------------final_graph_reshaping_impl---------------------- 1960 // Implement items 1-5 from final_graph_reshaping below. 1961 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 1962 1963 if ( n->outcnt() == 0 ) return; // dead node 1964 uint nop = n->Opcode(); 1965 1966 // Check for 2-input instruction with "last use" on right input. 1967 // Swap to left input. Implements item (2). 1968 if( n->req() == 3 && // two-input instruction 1969 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1970 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1971 n->in(2)->outcnt() == 1 &&// right use IS a last use 1972 !n->in(2)->is_Con() ) { // right use is not a constant 1973 // Check for commutative opcode 1974 switch( nop ) { 1975 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1976 case Op_MaxI: case Op_MinI: 1977 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 1978 case Op_AndL: case Op_XorL: case Op_OrL: 1979 case Op_AndI: case Op_XorI: case Op_OrI: { 1980 // Move "last use" input to left by swapping inputs 1981 n->swap_edges(1, 2); 1982 break; 1983 } 1984 default: 1985 break; 1986 } 1987 } 1988 1989 #ifdef ASSERT 1990 if( n->is_Mem() ) { 1991 Compile* C = Compile::current(); 1992 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 1993 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 1994 // oop will be recorded in oop map if load crosses safepoint 1995 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 1996 LoadNode::is_immutable_value(n->in(MemNode::Address))), 1997 "raw memory operations should have control edge"); 1998 } 1999 #endif 2000 // Count FPU ops and common calls, implements item (3) 2001 switch( nop ) { 2002 // Count all float operations that may use FPU 2003 case Op_AddF: 2004 case Op_SubF: 2005 case Op_MulF: 2006 case Op_DivF: 2007 case Op_NegF: 2008 case Op_ModF: 2009 case Op_ConvI2F: 2010 case Op_ConF: 2011 case Op_CmpF: 2012 case Op_CmpF3: 2013 // case Op_ConvL2F: // longs are split into 32-bit halves 2014 frc.inc_float_count(); 2015 break; 2016 2017 case Op_ConvF2D: 2018 case Op_ConvD2F: 2019 frc.inc_float_count(); 2020 frc.inc_double_count(); 2021 break; 2022 2023 // Count all double operations that may use FPU 2024 case Op_AddD: 2025 case Op_SubD: 2026 case Op_MulD: 2027 case Op_DivD: 2028 case Op_NegD: 2029 case Op_ModD: 2030 case Op_ConvI2D: 2031 case Op_ConvD2I: 2032 // case Op_ConvL2D: // handled by leaf call 2033 // case Op_ConvD2L: // handled by leaf call 2034 case Op_ConD: 2035 case Op_CmpD: 2036 case Op_CmpD3: 2037 frc.inc_double_count(); 2038 break; 2039 case Op_Opaque1: // Remove Opaque Nodes before matching 2040 case Op_Opaque2: // Remove Opaque Nodes before matching 2041 n->subsume_by(n->in(1)); 2042 break; 2043 case Op_CallStaticJava: 2044 case Op_CallJava: 2045 case Op_CallDynamicJava: 2046 frc.inc_java_call_count(); // Count java call site; 2047 case Op_CallRuntime: 2048 case Op_CallLeaf: 2049 case Op_CallLeafNoFP: { 2050 assert( n->is_Call(), "" ); 2051 CallNode *call = n->as_Call(); 2052 // Count call sites where the FP mode bit would have to be flipped. 2053 // Do not count uncommon runtime calls: 2054 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2055 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2056 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2057 frc.inc_call_count(); // Count the call site 2058 } else { // See if uncommon argument is shared 2059 Node *n = call->in(TypeFunc::Parms); 2060 int nop = n->Opcode(); 2061 // Clone shared simple arguments to uncommon calls, item (1). 2062 if( n->outcnt() > 1 && 2063 !n->is_Proj() && 2064 nop != Op_CreateEx && 2065 nop != Op_CheckCastPP && 2066 nop != Op_DecodeN && 2067 !n->is_Mem() ) { 2068 Node *x = n->clone(); 2069 call->set_req( TypeFunc::Parms, x ); 2070 } 2071 } 2072 break; 2073 } 2074 2075 case Op_StoreD: 2076 case Op_LoadD: 2077 case Op_LoadD_unaligned: 2078 frc.inc_double_count(); 2079 goto handle_mem; 2080 case Op_StoreF: 2081 case Op_LoadF: 2082 frc.inc_float_count(); 2083 goto handle_mem; 2084 2085 case Op_StoreB: 2086 case Op_StoreC: 2087 case Op_StoreCM: 2088 case Op_StorePConditional: 2089 case Op_StoreI: 2090 case Op_StoreL: 2091 case Op_StoreIConditional: 2092 case Op_StoreLConditional: 2093 case Op_CompareAndSwapI: 2094 case Op_CompareAndSwapL: 2095 case Op_CompareAndSwapP: 2096 case Op_CompareAndSwapN: 2097 case Op_StoreP: 2098 case Op_StoreN: 2099 case Op_LoadB: 2100 case Op_LoadUB: 2101 case Op_LoadUS: 2102 case Op_LoadI: 2103 case Op_LoadUI2L: 2104 case Op_LoadKlass: 2105 case Op_LoadNKlass: 2106 case Op_LoadL: 2107 case Op_LoadL_unaligned: 2108 case Op_LoadPLocked: 2109 case Op_LoadLLocked: 2110 case Op_LoadP: 2111 case Op_LoadN: 2112 case Op_LoadRange: 2113 case Op_LoadS: { 2114 handle_mem: 2115 #ifdef ASSERT 2116 if( VerifyOptoOopOffsets ) { 2117 assert( n->is_Mem(), "" ); 2118 MemNode *mem = (MemNode*)n; 2119 // Check to see if address types have grounded out somehow. 2120 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2121 assert( !tp || oop_offset_is_sane(tp), "" ); 2122 } 2123 #endif 2124 break; 2125 } 2126 2127 case Op_AddP: { // Assert sane base pointers 2128 Node *addp = n->in(AddPNode::Address); 2129 assert( !addp->is_AddP() || 2130 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2131 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2132 "Base pointers must match" ); 2133 #ifdef _LP64 2134 if (UseCompressedOops && 2135 addp->Opcode() == Op_ConP && 2136 addp == n->in(AddPNode::Base) && 2137 n->in(AddPNode::Offset)->is_Con()) { 2138 // Use addressing with narrow klass to load with offset on x86. 2139 // On sparc loading 32-bits constant and decoding it have less 2140 // instructions (4) then load 64-bits constant (7). 2141 // Do this transformation here since IGVN will convert ConN back to ConP. 2142 const Type* t = addp->bottom_type(); 2143 if (t->isa_oopptr()) { 2144 Node* nn = NULL; 2145 2146 // Look for existing ConN node of the same exact type. 2147 Compile* C = Compile::current(); 2148 Node* r = C->root(); 2149 uint cnt = r->outcnt(); 2150 for (uint i = 0; i < cnt; i++) { 2151 Node* m = r->raw_out(i); 2152 if (m!= NULL && m->Opcode() == Op_ConN && 2153 m->bottom_type()->make_ptr() == t) { 2154 nn = m; 2155 break; 2156 } 2157 } 2158 if (nn != NULL) { 2159 // Decode a narrow oop to match address 2160 // [R12 + narrow_oop_reg<<3 + offset] 2161 nn = new (C, 2) DecodeNNode(nn, t); 2162 n->set_req(AddPNode::Base, nn); 2163 n->set_req(AddPNode::Address, nn); 2164 if (addp->outcnt() == 0) { 2165 addp->disconnect_inputs(NULL); 2166 } 2167 } 2168 } 2169 } 2170 #endif 2171 break; 2172 } 2173 2174 #ifdef _LP64 2175 case Op_CastPP: 2176 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2177 Compile* C = Compile::current(); 2178 Node* in1 = n->in(1); 2179 const Type* t = n->bottom_type(); 2180 Node* new_in1 = in1->clone(); 2181 new_in1->as_DecodeN()->set_type(t); 2182 2183 if (!Matcher::narrow_oop_use_complex_address()) { 2184 // 2185 // x86, ARM and friends can handle 2 adds in addressing mode 2186 // and Matcher can fold a DecodeN node into address by using 2187 // a narrow oop directly and do implicit NULL check in address: 2188 // 2189 // [R12 + narrow_oop_reg<<3 + offset] 2190 // NullCheck narrow_oop_reg 2191 // 2192 // On other platforms (Sparc) we have to keep new DecodeN node and 2193 // use it to do implicit NULL check in address: 2194 // 2195 // decode_not_null narrow_oop_reg, base_reg 2196 // [base_reg + offset] 2197 // NullCheck base_reg 2198 // 2199 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2200 // to keep the information to which NULL check the new DecodeN node 2201 // corresponds to use it as value in implicit_null_check(). 2202 // 2203 new_in1->set_req(0, n->in(0)); 2204 } 2205 2206 n->subsume_by(new_in1); 2207 if (in1->outcnt() == 0) { 2208 in1->disconnect_inputs(NULL); 2209 } 2210 } 2211 break; 2212 2213 case Op_CmpP: 2214 // Do this transformation here to preserve CmpPNode::sub() and 2215 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2216 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2217 Node* in1 = n->in(1); 2218 Node* in2 = n->in(2); 2219 if (!in1->is_DecodeN()) { 2220 in2 = in1; 2221 in1 = n->in(2); 2222 } 2223 assert(in1->is_DecodeN(), "sanity"); 2224 2225 Compile* C = Compile::current(); 2226 Node* new_in2 = NULL; 2227 if (in2->is_DecodeN()) { 2228 new_in2 = in2->in(1); 2229 } else if (in2->Opcode() == Op_ConP) { 2230 const Type* t = in2->bottom_type(); 2231 if (t == TypePtr::NULL_PTR) { 2232 // Don't convert CmpP null check into CmpN if compressed 2233 // oops implicit null check is not generated. 2234 // This will allow to generate normal oop implicit null check. 2235 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2236 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2237 // 2238 // This transformation together with CastPP transformation above 2239 // will generated code for implicit NULL checks for compressed oops. 2240 // 2241 // The original code after Optimize() 2242 // 2243 // LoadN memory, narrow_oop_reg 2244 // decode narrow_oop_reg, base_reg 2245 // CmpP base_reg, NULL 2246 // CastPP base_reg // NotNull 2247 // Load [base_reg + offset], val_reg 2248 // 2249 // after these transformations will be 2250 // 2251 // LoadN memory, narrow_oop_reg 2252 // CmpN narrow_oop_reg, NULL 2253 // decode_not_null narrow_oop_reg, base_reg 2254 // Load [base_reg + offset], val_reg 2255 // 2256 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2257 // since narrow oops can be used in debug info now (see the code in 2258 // final_graph_reshaping_walk()). 2259 // 2260 // At the end the code will be matched to 2261 // on x86: 2262 // 2263 // Load_narrow_oop memory, narrow_oop_reg 2264 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2265 // NullCheck narrow_oop_reg 2266 // 2267 // and on sparc: 2268 // 2269 // Load_narrow_oop memory, narrow_oop_reg 2270 // decode_not_null narrow_oop_reg, base_reg 2271 // Load [base_reg + offset], val_reg 2272 // NullCheck base_reg 2273 // 2274 } else if (t->isa_oopptr()) { 2275 new_in2 = ConNode::make(C, t->make_narrowoop()); 2276 } 2277 } 2278 if (new_in2 != NULL) { 2279 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2280 n->subsume_by( cmpN ); 2281 if (in1->outcnt() == 0) { 2282 in1->disconnect_inputs(NULL); 2283 } 2284 if (in2->outcnt() == 0) { 2285 in2->disconnect_inputs(NULL); 2286 } 2287 } 2288 } 2289 break; 2290 2291 case Op_DecodeN: 2292 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2293 // DecodeN could be pinned when it can't be fold into 2294 // an address expression, see the code for Op_CastPP above. 2295 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2296 break; 2297 2298 case Op_EncodeP: { 2299 Node* in1 = n->in(1); 2300 if (in1->is_DecodeN()) { 2301 n->subsume_by(in1->in(1)); 2302 } else if (in1->Opcode() == Op_ConP) { 2303 Compile* C = Compile::current(); 2304 const Type* t = in1->bottom_type(); 2305 if (t == TypePtr::NULL_PTR) { 2306 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2307 } else if (t->isa_oopptr()) { 2308 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2309 } 2310 } 2311 if (in1->outcnt() == 0) { 2312 in1->disconnect_inputs(NULL); 2313 } 2314 break; 2315 } 2316 2317 case Op_Proj: { 2318 if (OptimizeStringConcat) { 2319 ProjNode* p = n->as_Proj(); 2320 if (p->_is_io_use) { 2321 // Separate projections were used for the exception path which 2322 // are normally removed by a late inline. If it wasn't inlined 2323 // then they will hang around and should just be replaced with 2324 // the original one. 2325 Node* proj = NULL; 2326 // Replace with just one 2327 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2328 Node *use = i.get(); 2329 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2330 proj = use; 2331 break; 2332 } 2333 } 2334 assert(p != NULL, "must be found"); 2335 p->subsume_by(proj); 2336 } 2337 } 2338 break; 2339 } 2340 2341 case Op_Phi: 2342 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2343 // The EncodeP optimization may create Phi with the same edges 2344 // for all paths. It is not handled well by Register Allocator. 2345 Node* unique_in = n->in(1); 2346 assert(unique_in != NULL, ""); 2347 uint cnt = n->req(); 2348 for (uint i = 2; i < cnt; i++) { 2349 Node* m = n->in(i); 2350 assert(m != NULL, ""); 2351 if (unique_in != m) 2352 unique_in = NULL; 2353 } 2354 if (unique_in != NULL) { 2355 n->subsume_by(unique_in); 2356 } 2357 } 2358 break; 2359 2360 #endif 2361 2362 case Op_ModI: 2363 if (UseDivMod) { 2364 // Check if a%b and a/b both exist 2365 Node* d = n->find_similar(Op_DivI); 2366 if (d) { 2367 // Replace them with a fused divmod if supported 2368 Compile* C = Compile::current(); 2369 if (Matcher::has_match_rule(Op_DivModI)) { 2370 DivModINode* divmod = DivModINode::make(C, n); 2371 d->subsume_by(divmod->div_proj()); 2372 n->subsume_by(divmod->mod_proj()); 2373 } else { 2374 // replace a%b with a-((a/b)*b) 2375 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2376 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2377 n->subsume_by( sub ); 2378 } 2379 } 2380 } 2381 break; 2382 2383 case Op_ModL: 2384 if (UseDivMod) { 2385 // Check if a%b and a/b both exist 2386 Node* d = n->find_similar(Op_DivL); 2387 if (d) { 2388 // Replace them with a fused divmod if supported 2389 Compile* C = Compile::current(); 2390 if (Matcher::has_match_rule(Op_DivModL)) { 2391 DivModLNode* divmod = DivModLNode::make(C, n); 2392 d->subsume_by(divmod->div_proj()); 2393 n->subsume_by(divmod->mod_proj()); 2394 } else { 2395 // replace a%b with a-((a/b)*b) 2396 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2397 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2398 n->subsume_by( sub ); 2399 } 2400 } 2401 } 2402 break; 2403 2404 case Op_Load16B: 2405 case Op_Load8B: 2406 case Op_Load4B: 2407 case Op_Load8S: 2408 case Op_Load4S: 2409 case Op_Load2S: 2410 case Op_Load8C: 2411 case Op_Load4C: 2412 case Op_Load2C: 2413 case Op_Load4I: 2414 case Op_Load2I: 2415 case Op_Load2L: 2416 case Op_Load4F: 2417 case Op_Load2F: 2418 case Op_Load2D: 2419 case Op_Store16B: 2420 case Op_Store8B: 2421 case Op_Store4B: 2422 case Op_Store8C: 2423 case Op_Store4C: 2424 case Op_Store2C: 2425 case Op_Store4I: 2426 case Op_Store2I: 2427 case Op_Store2L: 2428 case Op_Store4F: 2429 case Op_Store2F: 2430 case Op_Store2D: 2431 break; 2432 2433 case Op_PackB: 2434 case Op_PackS: 2435 case Op_PackC: 2436 case Op_PackI: 2437 case Op_PackF: 2438 case Op_PackL: 2439 case Op_PackD: 2440 if (n->req()-1 > 2) { 2441 // Replace many operand PackNodes with a binary tree for matching 2442 PackNode* p = (PackNode*) n; 2443 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2444 n->subsume_by(btp); 2445 } 2446 break; 2447 case Op_Loop: 2448 case Op_CountedLoop: 2449 if (n->as_Loop()->is_inner_loop()) { 2450 frc.inc_inner_loop_count(); 2451 } 2452 break; 2453 default: 2454 assert( !n->is_Call(), "" ); 2455 assert( !n->is_Mem(), "" ); 2456 break; 2457 } 2458 2459 // Collect CFG split points 2460 if (n->is_MultiBranch()) 2461 frc._tests.push(n); 2462 } 2463 2464 //------------------------------final_graph_reshaping_walk--------------------- 2465 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2466 // requires that the walk visits a node's inputs before visiting the node. 2467 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2468 ResourceArea *area = Thread::current()->resource_area(); 2469 Unique_Node_List sfpt(area); 2470 2471 frc._visited.set(root->_idx); // first, mark node as visited 2472 uint cnt = root->req(); 2473 Node *n = root; 2474 uint i = 0; 2475 while (true) { 2476 if (i < cnt) { 2477 // Place all non-visited non-null inputs onto stack 2478 Node* m = n->in(i); 2479 ++i; 2480 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2481 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2482 sfpt.push(m); 2483 cnt = m->req(); 2484 nstack.push(n, i); // put on stack parent and next input's index 2485 n = m; 2486 i = 0; 2487 } 2488 } else { 2489 // Now do post-visit work 2490 final_graph_reshaping_impl( n, frc ); 2491 if (nstack.is_empty()) 2492 break; // finished 2493 n = nstack.node(); // Get node from stack 2494 cnt = n->req(); 2495 i = nstack.index(); 2496 nstack.pop(); // Shift to the next node on stack 2497 } 2498 } 2499 2500 // Skip next transformation if compressed oops are not used. 2501 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) 2502 return; 2503 2504 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2505 // It could be done for an uncommon traps or any safepoints/calls 2506 // if the DecodeN node is referenced only in a debug info. 2507 while (sfpt.size() > 0) { 2508 n = sfpt.pop(); 2509 JVMState *jvms = n->as_SafePoint()->jvms(); 2510 assert(jvms != NULL, "sanity"); 2511 int start = jvms->debug_start(); 2512 int end = n->req(); 2513 bool is_uncommon = (n->is_CallStaticJava() && 2514 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2515 for (int j = start; j < end; j++) { 2516 Node* in = n->in(j); 2517 if (in->is_DecodeN()) { 2518 bool safe_to_skip = true; 2519 if (!is_uncommon ) { 2520 // Is it safe to skip? 2521 for (uint i = 0; i < in->outcnt(); i++) { 2522 Node* u = in->raw_out(i); 2523 if (!u->is_SafePoint() || 2524 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2525 safe_to_skip = false; 2526 } 2527 } 2528 } 2529 if (safe_to_skip) { 2530 n->set_req(j, in->in(1)); 2531 } 2532 if (in->outcnt() == 0) { 2533 in->disconnect_inputs(NULL); 2534 } 2535 } 2536 } 2537 } 2538 } 2539 2540 //------------------------------final_graph_reshaping-------------------------- 2541 // Final Graph Reshaping. 2542 // 2543 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2544 // and not commoned up and forced early. Must come after regular 2545 // optimizations to avoid GVN undoing the cloning. Clone constant 2546 // inputs to Loop Phis; these will be split by the allocator anyways. 2547 // Remove Opaque nodes. 2548 // (2) Move last-uses by commutative operations to the left input to encourage 2549 // Intel update-in-place two-address operations and better register usage 2550 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2551 // calls canonicalizing them back. 2552 // (3) Count the number of double-precision FP ops, single-precision FP ops 2553 // and call sites. On Intel, we can get correct rounding either by 2554 // forcing singles to memory (requires extra stores and loads after each 2555 // FP bytecode) or we can set a rounding mode bit (requires setting and 2556 // clearing the mode bit around call sites). The mode bit is only used 2557 // if the relative frequency of single FP ops to calls is low enough. 2558 // This is a key transform for SPEC mpeg_audio. 2559 // (4) Detect infinite loops; blobs of code reachable from above but not 2560 // below. Several of the Code_Gen algorithms fail on such code shapes, 2561 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2562 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2563 // Detection is by looking for IfNodes where only 1 projection is 2564 // reachable from below or CatchNodes missing some targets. 2565 // (5) Assert for insane oop offsets in debug mode. 2566 2567 bool Compile::final_graph_reshaping() { 2568 // an infinite loop may have been eliminated by the optimizer, 2569 // in which case the graph will be empty. 2570 if (root()->req() == 1) { 2571 record_method_not_compilable("trivial infinite loop"); 2572 return true; 2573 } 2574 2575 Final_Reshape_Counts frc; 2576 2577 // Visit everybody reachable! 2578 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2579 Node_Stack nstack(unique() >> 1); 2580 final_graph_reshaping_walk(nstack, root(), frc); 2581 2582 // Check for unreachable (from below) code (i.e., infinite loops). 2583 for( uint i = 0; i < frc._tests.size(); i++ ) { 2584 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2585 // Get number of CFG targets. 2586 // Note that PCTables include exception targets after calls. 2587 uint required_outcnt = n->required_outcnt(); 2588 if (n->outcnt() != required_outcnt) { 2589 // Check for a few special cases. Rethrow Nodes never take the 2590 // 'fall-thru' path, so expected kids is 1 less. 2591 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2592 if (n->in(0)->in(0)->is_Call()) { 2593 CallNode *call = n->in(0)->in(0)->as_Call(); 2594 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2595 required_outcnt--; // Rethrow always has 1 less kid 2596 } else if (call->req() > TypeFunc::Parms && 2597 call->is_CallDynamicJava()) { 2598 // Check for null receiver. In such case, the optimizer has 2599 // detected that the virtual call will always result in a null 2600 // pointer exception. The fall-through projection of this CatchNode 2601 // will not be populated. 2602 Node *arg0 = call->in(TypeFunc::Parms); 2603 if (arg0->is_Type() && 2604 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2605 required_outcnt--; 2606 } 2607 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2608 call->req() > TypeFunc::Parms+1 && 2609 call->is_CallStaticJava()) { 2610 // Check for negative array length. In such case, the optimizer has 2611 // detected that the allocation attempt will always result in an 2612 // exception. There is no fall-through projection of this CatchNode . 2613 Node *arg1 = call->in(TypeFunc::Parms+1); 2614 if (arg1->is_Type() && 2615 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2616 required_outcnt--; 2617 } 2618 } 2619 } 2620 } 2621 // Recheck with a better notion of 'required_outcnt' 2622 if (n->outcnt() != required_outcnt) { 2623 record_method_not_compilable("malformed control flow"); 2624 return true; // Not all targets reachable! 2625 } 2626 } 2627 // Check that I actually visited all kids. Unreached kids 2628 // must be infinite loops. 2629 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2630 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2631 record_method_not_compilable("infinite loop"); 2632 return true; // Found unvisited kid; must be unreach 2633 } 2634 } 2635 2636 // If original bytecodes contained a mixture of floats and doubles 2637 // check if the optimizer has made it homogenous, item (3). 2638 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2639 frc.get_float_count() > 32 && 2640 frc.get_double_count() == 0 && 2641 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2642 set_24_bit_selection_and_mode( false, true ); 2643 } 2644 2645 set_java_calls(frc.get_java_call_count()); 2646 set_inner_loops(frc.get_inner_loop_count()); 2647 2648 // No infinite loops, no reason to bail out. 2649 return false; 2650 } 2651 2652 //-----------------------------too_many_traps---------------------------------- 2653 // Report if there are too many traps at the current method and bci. 2654 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2655 bool Compile::too_many_traps(ciMethod* method, 2656 int bci, 2657 Deoptimization::DeoptReason reason) { 2658 ciMethodData* md = method->method_data(); 2659 if (md->is_empty()) { 2660 // Assume the trap has not occurred, or that it occurred only 2661 // because of a transient condition during start-up in the interpreter. 2662 return false; 2663 } 2664 if (md->has_trap_at(bci, reason) != 0) { 2665 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2666 // Also, if there are multiple reasons, or if there is no per-BCI record, 2667 // assume the worst. 2668 if (log()) 2669 log()->elem("observe trap='%s' count='%d'", 2670 Deoptimization::trap_reason_name(reason), 2671 md->trap_count(reason)); 2672 return true; 2673 } else { 2674 // Ignore method/bci and see if there have been too many globally. 2675 return too_many_traps(reason, md); 2676 } 2677 } 2678 2679 // Less-accurate variant which does not require a method and bci. 2680 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2681 ciMethodData* logmd) { 2682 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2683 // Too many traps globally. 2684 // Note that we use cumulative trap_count, not just md->trap_count. 2685 if (log()) { 2686 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2687 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2688 Deoptimization::trap_reason_name(reason), 2689 mcount, trap_count(reason)); 2690 } 2691 return true; 2692 } else { 2693 // The coast is clear. 2694 return false; 2695 } 2696 } 2697 2698 //--------------------------too_many_recompiles-------------------------------- 2699 // Report if there are too many recompiles at the current method and bci. 2700 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2701 // Is not eager to return true, since this will cause the compiler to use 2702 // Action_none for a trap point, to avoid too many recompilations. 2703 bool Compile::too_many_recompiles(ciMethod* method, 2704 int bci, 2705 Deoptimization::DeoptReason reason) { 2706 ciMethodData* md = method->method_data(); 2707 if (md->is_empty()) { 2708 // Assume the trap has not occurred, or that it occurred only 2709 // because of a transient condition during start-up in the interpreter. 2710 return false; 2711 } 2712 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2713 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2714 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2715 Deoptimization::DeoptReason per_bc_reason 2716 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2717 if ((per_bc_reason == Deoptimization::Reason_none 2718 || md->has_trap_at(bci, reason) != 0) 2719 // The trap frequency measure we care about is the recompile count: 2720 && md->trap_recompiled_at(bci) 2721 && md->overflow_recompile_count() >= bc_cutoff) { 2722 // Do not emit a trap here if it has already caused recompilations. 2723 // Also, if there are multiple reasons, or if there is no per-BCI record, 2724 // assume the worst. 2725 if (log()) 2726 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2727 Deoptimization::trap_reason_name(reason), 2728 md->trap_count(reason), 2729 md->overflow_recompile_count()); 2730 return true; 2731 } else if (trap_count(reason) != 0 2732 && decompile_count() >= m_cutoff) { 2733 // Too many recompiles globally, and we have seen this sort of trap. 2734 // Use cumulative decompile_count, not just md->decompile_count. 2735 if (log()) 2736 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2737 Deoptimization::trap_reason_name(reason), 2738 md->trap_count(reason), trap_count(reason), 2739 md->decompile_count(), decompile_count()); 2740 return true; 2741 } else { 2742 // The coast is clear. 2743 return false; 2744 } 2745 } 2746 2747 2748 #ifndef PRODUCT 2749 //------------------------------verify_graph_edges--------------------------- 2750 // Walk the Graph and verify that there is a one-to-one correspondence 2751 // between Use-Def edges and Def-Use edges in the graph. 2752 void Compile::verify_graph_edges(bool no_dead_code) { 2753 if (VerifyGraphEdges) { 2754 ResourceArea *area = Thread::current()->resource_area(); 2755 Unique_Node_List visited(area); 2756 // Call recursive graph walk to check edges 2757 _root->verify_edges(visited); 2758 if (no_dead_code) { 2759 // Now make sure that no visited node is used by an unvisited node. 2760 bool dead_nodes = 0; 2761 Unique_Node_List checked(area); 2762 while (visited.size() > 0) { 2763 Node* n = visited.pop(); 2764 checked.push(n); 2765 for (uint i = 0; i < n->outcnt(); i++) { 2766 Node* use = n->raw_out(i); 2767 if (checked.member(use)) continue; // already checked 2768 if (visited.member(use)) continue; // already in the graph 2769 if (use->is_Con()) continue; // a dead ConNode is OK 2770 // At this point, we have found a dead node which is DU-reachable. 2771 if (dead_nodes++ == 0) 2772 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2773 use->dump(2); 2774 tty->print_cr("---"); 2775 checked.push(use); // No repeats; pretend it is now checked. 2776 } 2777 } 2778 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2779 } 2780 } 2781 } 2782 #endif 2783 2784 // The Compile object keeps track of failure reasons separately from the ciEnv. 2785 // This is required because there is not quite a 1-1 relation between the 2786 // ciEnv and its compilation task and the Compile object. Note that one 2787 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2788 // to backtrack and retry without subsuming loads. Other than this backtracking 2789 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2790 // by the logic in C2Compiler. 2791 void Compile::record_failure(const char* reason) { 2792 if (log() != NULL) { 2793 log()->elem("failure reason='%s' phase='compile'", reason); 2794 } 2795 if (_failure_reason == NULL) { 2796 // Record the first failure reason. 2797 _failure_reason = reason; 2798 } 2799 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2800 C->print_method(_failure_reason); 2801 } 2802 _root = NULL; // flush the graph, too 2803 } 2804 2805 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2806 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2807 { 2808 if (dolog) { 2809 C = Compile::current(); 2810 _log = C->log(); 2811 } else { 2812 C = NULL; 2813 _log = NULL; 2814 } 2815 if (_log != NULL) { 2816 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2817 _log->stamp(); 2818 _log->end_head(); 2819 } 2820 } 2821 2822 Compile::TracePhase::~TracePhase() { 2823 if (_log != NULL) { 2824 _log->done("phase nodes='%d'", C->unique()); 2825 } 2826 }