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 29 // -------------------- Compile::mach_constant_base_node ----------------------- 30 // Constant table base node singleton. 31 MachConstantBaseNode* Compile::mach_constant_base_node() { 32 if (_mach_constant_base_node == NULL) { 33 _mach_constant_base_node = new (C) MachConstantBaseNode(); 34 _mach_constant_base_node->set_req(0, C->root()); 35 } 36 return _mach_constant_base_node; 37 } 38 39 40 /// Support for intrinsics. 41 42 // Return the index at which m must be inserted (or already exists). 43 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 44 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 45 #ifdef ASSERT 46 for (int i = 1; i < _intrinsics->length(); i++) { 47 CallGenerator* cg1 = _intrinsics->at(i-1); 48 CallGenerator* cg2 = _intrinsics->at(i); 49 assert(cg1->method() != cg2->method() 50 ? cg1->method() < cg2->method() 51 : cg1->is_virtual() < cg2->is_virtual(), 52 "compiler intrinsics list must stay sorted"); 53 } 54 #endif 55 // Binary search sorted list, in decreasing intervals [lo, hi]. 56 int lo = 0, hi = _intrinsics->length()-1; 57 while (lo <= hi) { 58 int mid = (uint)(hi + lo) / 2; 59 ciMethod* mid_m = _intrinsics->at(mid)->method(); 60 if (m < mid_m) { 61 hi = mid-1; 62 } else if (m > mid_m) { 63 lo = mid+1; 64 } else { 65 // look at minor sort key 66 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 67 if (is_virtual < mid_virt) { 68 hi = mid-1; 69 } else if (is_virtual > mid_virt) { 70 lo = mid+1; 71 } else { 72 return mid; // exact match 73 } 74 } 75 } 76 return lo; // inexact match 77 } 78 79 void Compile::register_intrinsic(CallGenerator* cg) { 80 if (_intrinsics == NULL) { 81 _intrinsics = new GrowableArray<CallGenerator*>(60); 82 } 83 // This code is stolen from ciObjectFactory::insert. 84 // Really, GrowableArray should have methods for 85 // insert_at, remove_at, and binary_search. 86 int len = _intrinsics->length(); 87 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 88 if (index == len) { 89 _intrinsics->append(cg); 90 } else { 91 #ifdef ASSERT 92 CallGenerator* oldcg = _intrinsics->at(index); 93 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 94 #endif 95 _intrinsics->append(_intrinsics->at(len-1)); 96 int pos; 97 for (pos = len-2; pos >= index; pos--) { 98 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 99 } 100 _intrinsics->at_put(index, cg); 101 } 102 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 103 } 104 105 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 106 assert(m->is_loaded(), "don't try this on unloaded methods"); 107 if (_intrinsics != NULL) { 108 int index = intrinsic_insertion_index(m, is_virtual); 109 if (index < _intrinsics->length() 110 && _intrinsics->at(index)->method() == m 111 && _intrinsics->at(index)->is_virtual() == is_virtual) { 112 return _intrinsics->at(index); 113 } 114 } 115 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 116 if (m->intrinsic_id() != vmIntrinsics::_none && 117 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 118 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 119 if (cg != NULL) { 120 // Save it for next time: 121 register_intrinsic(cg); 122 return cg; 123 } else { 124 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 125 } 126 } 127 return NULL; 128 } 129 130 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 131 // in library_call.cpp. 132 133 134 #ifndef PRODUCT 135 // statistics gathering... 136 137 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 138 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 139 140 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 141 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 142 int oflags = _intrinsic_hist_flags[id]; 143 assert(flags != 0, "what happened?"); 144 if (is_virtual) { 145 flags |= _intrinsic_virtual; 146 } 147 bool changed = (flags != oflags); 148 if ((flags & _intrinsic_worked) != 0) { 149 juint count = (_intrinsic_hist_count[id] += 1); 150 if (count == 1) { 151 changed = true; // first time 152 } 153 // increment the overall count also: 154 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 155 } 156 if (changed) { 157 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 158 // Something changed about the intrinsic's virtuality. 159 if ((flags & _intrinsic_virtual) != 0) { 160 // This is the first use of this intrinsic as a virtual call. 161 if (oflags != 0) { 162 // We already saw it as a non-virtual, so note both cases. 163 flags |= _intrinsic_both; 164 } 165 } else if ((oflags & _intrinsic_both) == 0) { 166 // This is the first use of this intrinsic as a non-virtual 167 flags |= _intrinsic_both; 168 } 169 } 170 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 171 } 172 // update the overall flags also: 173 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 174 return changed; 175 } 176 177 static char* format_flags(int flags, char* buf) { 178 buf[0] = 0; 179 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 180 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 181 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 182 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 183 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 184 if (buf[0] == 0) strcat(buf, ","); 185 assert(buf[0] == ',', "must be"); 186 return &buf[1]; 187 } 188 189 void Compile::print_intrinsic_statistics() { 190 char flagsbuf[100]; 191 ttyLocker ttyl; 192 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 193 tty->print_cr("Compiler intrinsic usage:"); 194 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 195 if (total == 0) total = 1; // avoid div0 in case of no successes 196 #define PRINT_STAT_LINE(name, c, f) \ 197 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 198 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 199 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 200 int flags = _intrinsic_hist_flags[id]; 201 juint count = _intrinsic_hist_count[id]; 202 if ((flags | count) != 0) { 203 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 204 } 205 } 206 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 207 if (xtty != NULL) xtty->tail("statistics"); 208 } 209 210 void Compile::print_statistics() { 211 { ttyLocker ttyl; 212 if (xtty != NULL) xtty->head("statistics type='opto'"); 213 Parse::print_statistics(); 214 PhaseCCP::print_statistics(); 215 PhaseRegAlloc::print_statistics(); 216 Scheduling::print_statistics(); 217 PhasePeephole::print_statistics(); 218 PhaseIdealLoop::print_statistics(); 219 if (xtty != NULL) xtty->tail("statistics"); 220 } 221 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 222 // put this under its own <statistics> element. 223 print_intrinsic_statistics(); 224 } 225 } 226 #endif //PRODUCT 227 228 // Support for bundling info 229 Bundle* Compile::node_bundling(const Node *n) { 230 assert(valid_bundle_info(n), "oob"); 231 return &_node_bundling_base[n->_idx]; 232 } 233 234 bool Compile::valid_bundle_info(const Node *n) { 235 return (_node_bundling_limit > n->_idx); 236 } 237 238 239 void Compile::gvn_replace_by(Node* n, Node* nn) { 240 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 241 Node* use = n->last_out(i); 242 bool is_in_table = initial_gvn()->hash_delete(use); 243 uint uses_found = 0; 244 for (uint j = 0; j < use->len(); j++) { 245 if (use->in(j) == n) { 246 if (j < use->req()) 247 use->set_req(j, nn); 248 else 249 use->set_prec(j, nn); 250 uses_found++; 251 } 252 } 253 if (is_in_table) { 254 // reinsert into table 255 initial_gvn()->hash_find_insert(use); 256 } 257 record_for_igvn(use); 258 i -= uses_found; // we deleted 1 or more copies of this edge 259 } 260 } 261 262 263 264 265 // Identify all nodes that are reachable from below, useful. 266 // Use breadth-first pass that records state in a Unique_Node_List, 267 // recursive traversal is slower. 268 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 269 int estimated_worklist_size = unique(); 270 useful.map( estimated_worklist_size, NULL ); // preallocate space 271 272 // Initialize worklist 273 if (root() != NULL) { useful.push(root()); } 274 // If 'top' is cached, declare it useful to preserve cached node 275 if( cached_top_node() ) { useful.push(cached_top_node()); } 276 277 // Push all useful nodes onto the list, breadthfirst 278 for( uint next = 0; next < useful.size(); ++next ) { 279 assert( next < unique(), "Unique useful nodes < total nodes"); 280 Node *n = useful.at(next); 281 uint max = n->len(); 282 for( uint i = 0; i < max; ++i ) { 283 Node *m = n->in(i); 284 if( m == NULL ) continue; 285 useful.push(m); 286 } 287 } 288 } 289 290 // Disconnect all useless nodes by disconnecting those at the boundary. 291 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 292 uint next = 0; 293 while( next < useful.size() ) { 294 Node *n = useful.at(next++); 295 // Use raw traversal of out edges since this code removes out edges 296 int max = n->outcnt(); 297 for (int j = 0; j < max; ++j ) { 298 Node* child = n->raw_out(j); 299 if( ! useful.member(child) ) { 300 assert( !child->is_top() || child != top(), 301 "If top is cached in Compile object it is in useful list"); 302 // Only need to remove this out-edge to the useless node 303 n->raw_del_out(j); 304 --j; 305 --max; 306 } 307 } 308 if (n->outcnt() == 1 && n->has_special_unique_user()) { 309 record_for_igvn( n->unique_out() ); 310 } 311 } 312 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 313 } 314 315 //------------------------------frame_size_in_words----------------------------- 316 // frame_slots in units of words 317 int Compile::frame_size_in_words() const { 318 // shift is 0 in LP32 and 1 in LP64 319 const int shift = (LogBytesPerWord - LogBytesPerInt); 320 int words = _frame_slots >> shift; 321 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 322 return words; 323 } 324 325 // ============================================================================ 326 //------------------------------CompileWrapper--------------------------------- 327 class CompileWrapper : public StackObj { 328 Compile *const _compile; 329 public: 330 CompileWrapper(Compile* compile); 331 332 ~CompileWrapper(); 333 }; 334 335 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 336 // the Compile* pointer is stored in the current ciEnv: 337 ciEnv* env = compile->env(); 338 assert(env == ciEnv::current(), "must already be a ciEnv active"); 339 assert(env->compiler_data() == NULL, "compile already active?"); 340 env->set_compiler_data(compile); 341 assert(compile == Compile::current(), "sanity"); 342 343 compile->set_type_dict(NULL); 344 compile->set_type_hwm(NULL); 345 compile->set_type_last_size(0); 346 compile->set_last_tf(NULL, NULL); 347 compile->set_indexSet_arena(NULL); 348 compile->set_indexSet_free_block_list(NULL); 349 compile->init_type_arena(); 350 Type::Initialize(compile); 351 _compile->set_scratch_buffer_blob(NULL); 352 _compile->begin_method(); 353 } 354 CompileWrapper::~CompileWrapper() { 355 _compile->end_method(); 356 if (_compile->scratch_buffer_blob() != NULL) 357 BufferBlob::free(_compile->scratch_buffer_blob()); 358 _compile->env()->set_compiler_data(NULL); 359 } 360 361 362 //----------------------------print_compile_messages--------------------------- 363 void Compile::print_compile_messages() { 364 #ifndef PRODUCT 365 // Check if recompiling 366 if (_subsume_loads == false && PrintOpto) { 367 // Recompiling without allowing machine instructions to subsume loads 368 tty->print_cr("*********************************************************"); 369 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 370 tty->print_cr("*********************************************************"); 371 } 372 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 373 // Recompiling without escape analysis 374 tty->print_cr("*********************************************************"); 375 tty->print_cr("** Bailout: Recompile without escape analysis **"); 376 tty->print_cr("*********************************************************"); 377 } 378 if (env()->break_at_compile()) { 379 // Open the debugger when compiling this method. 380 tty->print("### Breaking when compiling: "); 381 method()->print_short_name(); 382 tty->cr(); 383 BREAKPOINT; 384 } 385 386 if( PrintOpto ) { 387 if (is_osr_compilation()) { 388 tty->print("[OSR]%3d", _compile_id); 389 } else { 390 tty->print("%3d", _compile_id); 391 } 392 } 393 #endif 394 } 395 396 397 void Compile::init_scratch_buffer_blob(int const_size) { 398 if (scratch_buffer_blob() != NULL) return; 399 400 // Construct a temporary CodeBuffer to have it construct a BufferBlob 401 // Cache this BufferBlob for this compile. 402 ResourceMark rm; 403 _scratch_const_size = const_size; 404 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 405 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 406 // Record the buffer blob for next time. 407 set_scratch_buffer_blob(blob); 408 // Have we run out of code space? 409 if (scratch_buffer_blob() == NULL) { 410 // Let CompilerBroker disable further compilations. 411 record_failure("Not enough space for scratch buffer in CodeCache"); 412 return; 413 } 414 415 // Initialize the relocation buffers 416 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 417 set_scratch_locs_memory(locs_buf); 418 } 419 420 421 void Compile::clear_scratch_buffer_blob() { 422 assert(scratch_buffer_blob(), "no BufferBlob set"); 423 set_scratch_buffer_blob(NULL); 424 set_scratch_locs_memory(NULL); 425 } 426 427 428 //-----------------------scratch_emit_size------------------------------------- 429 // Helper function that computes size by emitting code 430 uint Compile::scratch_emit_size(const Node* n) { 431 // Start scratch_emit_size section. 432 set_in_scratch_emit_size(true); 433 434 // Emit into a trash buffer and count bytes emitted. 435 // This is a pretty expensive way to compute a size, 436 // but it works well enough if seldom used. 437 // All common fixed-size instructions are given a size 438 // method by the AD file. 439 // Note that the scratch buffer blob and locs memory are 440 // allocated at the beginning of the compile task, and 441 // may be shared by several calls to scratch_emit_size. 442 // The allocation of the scratch buffer blob is particularly 443 // expensive, since it has to grab the code cache lock. 444 BufferBlob* blob = this->scratch_buffer_blob(); 445 assert(blob != NULL, "Initialize BufferBlob at start"); 446 assert(blob->size() > MAX_inst_size, "sanity"); 447 relocInfo* locs_buf = scratch_locs_memory(); 448 address blob_begin = blob->content_begin(); 449 address blob_end = (address)locs_buf; 450 assert(blob->content_contains(blob_end), "sanity"); 451 CodeBuffer buf(blob_begin, blob_end - blob_begin); 452 buf.initialize_consts_size(_scratch_const_size); 453 buf.initialize_stubs_size(MAX_stubs_size); 454 assert(locs_buf != NULL, "sanity"); 455 int lsize = MAX_locs_size / 3; 456 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 457 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 458 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 459 460 // Do the emission. 461 n->emit(buf, this->regalloc()); 462 463 // End scratch_emit_size section. 464 set_in_scratch_emit_size(false); 465 466 return buf.insts_size(); 467 } 468 469 470 // ============================================================================ 471 //------------------------------Compile standard------------------------------- 472 debug_only( int Compile::_debug_idx = 100000; ) 473 474 // Compile a method. entry_bci is -1 for normal compilations and indicates 475 // the continuation bci for on stack replacement. 476 477 478 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 479 : Phase(Compiler), 480 _env(ci_env), 481 _log(ci_env->log()), 482 _compile_id(ci_env->compile_id()), 483 _save_argument_registers(false), 484 _stub_name(NULL), 485 _stub_function(NULL), 486 _stub_entry_point(NULL), 487 _method(target), 488 _entry_bci(osr_bci), 489 _initial_gvn(NULL), 490 _for_igvn(NULL), 491 _warm_calls(NULL), 492 _subsume_loads(subsume_loads), 493 _do_escape_analysis(do_escape_analysis), 494 _failure_reason(NULL), 495 _code_buffer("Compile::Fill_buffer"), 496 _orig_pc_slot(0), 497 _orig_pc_slot_offset_in_bytes(0), 498 _has_method_handle_invokes(false), 499 _mach_constant_base_node(NULL), 500 _node_bundling_limit(0), 501 _node_bundling_base(NULL), 502 _java_calls(0), 503 _inner_loops(0), 504 _scratch_const_size(-1), 505 _in_scratch_emit_size(false), 506 #ifndef PRODUCT 507 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 508 _printer(IdealGraphPrinter::printer()), 509 #endif 510 _congraph(NULL) { 511 C = this; 512 513 CompileWrapper cw(this); 514 #ifndef PRODUCT 515 if (TimeCompiler2) { 516 tty->print(" "); 517 target->holder()->name()->print(); 518 tty->print("."); 519 target->print_short_name(); 520 tty->print(" "); 521 } 522 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 523 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 524 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 525 if (!print_opto_assembly) { 526 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 527 if (print_assembly && !Disassembler::can_decode()) { 528 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 529 print_opto_assembly = true; 530 } 531 } 532 set_print_assembly(print_opto_assembly); 533 set_parsed_irreducible_loop(false); 534 #endif 535 536 if (ProfileTraps) { 537 // Make sure the method being compiled gets its own MDO, 538 // so we can at least track the decompile_count(). 539 method()->build_method_data(); 540 } 541 542 Init(::AliasLevel); 543 544 545 print_compile_messages(); 546 547 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 548 _ilt = InlineTree::build_inline_tree_root(); 549 else 550 _ilt = NULL; 551 552 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 553 assert(num_alias_types() >= AliasIdxRaw, ""); 554 555 #define MINIMUM_NODE_HASH 1023 556 // Node list that Iterative GVN will start with 557 Unique_Node_List for_igvn(comp_arena()); 558 set_for_igvn(&for_igvn); 559 560 // GVN that will be run immediately on new nodes 561 uint estimated_size = method()->code_size()*4+64; 562 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 563 PhaseGVN gvn(node_arena(), estimated_size); 564 set_initial_gvn(&gvn); 565 566 { // Scope for timing the parser 567 TracePhase t3("parse", &_t_parser, true); 568 569 // Put top into the hash table ASAP. 570 initial_gvn()->transform_no_reclaim(top()); 571 572 // Set up tf(), start(), and find a CallGenerator. 573 CallGenerator* cg; 574 if (is_osr_compilation()) { 575 const TypeTuple *domain = StartOSRNode::osr_domain(); 576 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 577 init_tf(TypeFunc::make(domain, range)); 578 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 579 initial_gvn()->set_type_bottom(s); 580 init_start(s); 581 cg = CallGenerator::for_osr(method(), entry_bci()); 582 } else { 583 // Normal case. 584 init_tf(TypeFunc::make(method())); 585 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 586 initial_gvn()->set_type_bottom(s); 587 init_start(s); 588 float past_uses = method()->interpreter_invocation_count(); 589 float expected_uses = past_uses; 590 cg = CallGenerator::for_inline(method(), expected_uses); 591 } 592 if (failing()) return; 593 if (cg == NULL) { 594 record_method_not_compilable_all_tiers("cannot parse method"); 595 return; 596 } 597 JVMState* jvms = build_start_state(start(), tf()); 598 if ((jvms = cg->generate(jvms)) == NULL) { 599 record_method_not_compilable("method parse failed"); 600 return; 601 } 602 GraphKit kit(jvms); 603 604 if (!kit.stopped()) { 605 // Accept return values, and transfer control we know not where. 606 // This is done by a special, unique ReturnNode bound to root. 607 return_values(kit.jvms()); 608 } 609 610 if (kit.has_exceptions()) { 611 // Any exceptions that escape from this call must be rethrown 612 // to whatever caller is dynamically above us on the stack. 613 // This is done by a special, unique RethrowNode bound to root. 614 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 615 } 616 617 if (!failing() && has_stringbuilder()) { 618 { 619 // remove useless nodes to make the usage analysis simpler 620 ResourceMark rm; 621 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 622 } 623 624 { 625 ResourceMark rm; 626 print_method("Before StringOpts", 3); 627 PhaseStringOpts pso(initial_gvn(), &for_igvn); 628 print_method("After StringOpts", 3); 629 } 630 631 // now inline anything that we skipped the first time around 632 while (_late_inlines.length() > 0) { 633 CallGenerator* cg = _late_inlines.pop(); 634 cg->do_late_inline(); 635 } 636 } 637 assert(_late_inlines.length() == 0, "should have been processed"); 638 639 print_method("Before RemoveUseless", 3); 640 641 // Remove clutter produced by parsing. 642 if (!failing()) { 643 ResourceMark rm; 644 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 645 } 646 } 647 648 // Note: Large methods are capped off in do_one_bytecode(). 649 if (failing()) return; 650 651 // After parsing, node notes are no longer automagic. 652 // They must be propagated by register_new_node_with_optimizer(), 653 // clone(), or the like. 654 set_default_node_notes(NULL); 655 656 for (;;) { 657 int successes = Inline_Warm(); 658 if (failing()) return; 659 if (successes == 0) break; 660 } 661 662 // Drain the list. 663 Finish_Warm(); 664 #ifndef PRODUCT 665 if (_printer) { 666 _printer->print_inlining(this); 667 } 668 #endif 669 670 if (failing()) return; 671 NOT_PRODUCT( verify_graph_edges(); ) 672 673 // Now optimize 674 Optimize(); 675 if (failing()) return; 676 NOT_PRODUCT( verify_graph_edges(); ) 677 678 #ifndef PRODUCT 679 if (PrintIdeal) { 680 ttyLocker ttyl; // keep the following output all in one block 681 // This output goes directly to the tty, not the compiler log. 682 // To enable tools to match it up with the compilation activity, 683 // be sure to tag this tty output with the compile ID. 684 if (xtty != NULL) { 685 xtty->head("ideal compile_id='%d'%s", compile_id(), 686 is_osr_compilation() ? " compile_kind='osr'" : 687 ""); 688 } 689 root()->dump(9999); 690 if (xtty != NULL) { 691 xtty->tail("ideal"); 692 } 693 } 694 #endif 695 696 // Now that we know the size of all the monitors we can add a fixed slot 697 // for the original deopt pc. 698 699 _orig_pc_slot = fixed_slots(); 700 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 701 set_fixed_slots(next_slot); 702 703 // Now generate code 704 Code_Gen(); 705 if (failing()) return; 706 707 // Check if we want to skip execution of all compiled code. 708 { 709 #ifndef PRODUCT 710 if (OptoNoExecute) { 711 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 712 return; 713 } 714 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 715 #endif 716 717 if (is_osr_compilation()) { 718 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 719 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 720 } else { 721 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 722 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 723 } 724 725 env()->register_method(_method, _entry_bci, 726 &_code_offsets, 727 _orig_pc_slot_offset_in_bytes, 728 code_buffer(), 729 frame_size_in_words(), _oop_map_set, 730 &_handler_table, &_inc_table, 731 compiler, 732 env()->comp_level(), 733 true, /*has_debug_info*/ 734 has_unsafe_access() 735 ); 736 } 737 } 738 739 //------------------------------Compile---------------------------------------- 740 // Compile a runtime stub 741 Compile::Compile( ciEnv* ci_env, 742 TypeFunc_generator generator, 743 address stub_function, 744 const char *stub_name, 745 int is_fancy_jump, 746 bool pass_tls, 747 bool save_arg_registers, 748 bool return_pc ) 749 : Phase(Compiler), 750 _env(ci_env), 751 _log(ci_env->log()), 752 _compile_id(-1), 753 _save_argument_registers(save_arg_registers), 754 _method(NULL), 755 _stub_name(stub_name), 756 _stub_function(stub_function), 757 _stub_entry_point(NULL), 758 _entry_bci(InvocationEntryBci), 759 _initial_gvn(NULL), 760 _for_igvn(NULL), 761 _warm_calls(NULL), 762 _orig_pc_slot(0), 763 _orig_pc_slot_offset_in_bytes(0), 764 _subsume_loads(true), 765 _do_escape_analysis(false), 766 _failure_reason(NULL), 767 _code_buffer("Compile::Fill_buffer"), 768 _has_method_handle_invokes(false), 769 _mach_constant_base_node(NULL), 770 _node_bundling_limit(0), 771 _node_bundling_base(NULL), 772 _java_calls(0), 773 _inner_loops(0), 774 #ifndef PRODUCT 775 _trace_opto_output(TraceOptoOutput), 776 _printer(NULL), 777 #endif 778 _congraph(NULL) { 779 C = this; 780 781 #ifndef PRODUCT 782 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 783 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 784 set_print_assembly(PrintFrameConverterAssembly); 785 set_parsed_irreducible_loop(false); 786 #endif 787 CompileWrapper cw(this); 788 Init(/*AliasLevel=*/ 0); 789 init_tf((*generator)()); 790 791 { 792 // The following is a dummy for the sake of GraphKit::gen_stub 793 Unique_Node_List for_igvn(comp_arena()); 794 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 795 PhaseGVN gvn(Thread::current()->resource_area(),255); 796 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 797 gvn.transform_no_reclaim(top()); 798 799 GraphKit kit; 800 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 801 } 802 803 NOT_PRODUCT( verify_graph_edges(); ) 804 Code_Gen(); 805 if (failing()) return; 806 807 808 // Entry point will be accessed using compile->stub_entry_point(); 809 if (code_buffer() == NULL) { 810 Matcher::soft_match_failure(); 811 } else { 812 if (PrintAssembly && (WizardMode || Verbose)) 813 tty->print_cr("### Stub::%s", stub_name); 814 815 if (!failing()) { 816 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 817 818 // Make the NMethod 819 // For now we mark the frame as never safe for profile stackwalking 820 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 821 code_buffer(), 822 CodeOffsets::frame_never_safe, 823 // _code_offsets.value(CodeOffsets::Frame_Complete), 824 frame_size_in_words(), 825 _oop_map_set, 826 save_arg_registers); 827 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 828 829 _stub_entry_point = rs->entry_point(); 830 } 831 } 832 } 833 834 #ifndef PRODUCT 835 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 836 if(PrintOpto && Verbose) { 837 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 838 } 839 } 840 #endif 841 842 void Compile::print_codes() { 843 } 844 845 //------------------------------Init------------------------------------------- 846 // Prepare for a single compilation 847 void Compile::Init(int aliaslevel) { 848 _unique = 0; 849 _regalloc = NULL; 850 851 _tf = NULL; // filled in later 852 _top = NULL; // cached later 853 _matcher = NULL; // filled in later 854 _cfg = NULL; // filled in later 855 856 set_24_bit_selection_and_mode(Use24BitFP, false); 857 858 _node_note_array = NULL; 859 _default_node_notes = NULL; 860 861 _immutable_memory = NULL; // filled in at first inquiry 862 863 // Globally visible Nodes 864 // First set TOP to NULL to give safe behavior during creation of RootNode 865 set_cached_top_node(NULL); 866 set_root(new (this, 3) RootNode()); 867 // Now that you have a Root to point to, create the real TOP 868 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 869 set_recent_alloc(NULL, NULL); 870 871 // Create Debug Information Recorder to record scopes, oopmaps, etc. 872 env()->set_oop_recorder(new OopRecorder(comp_arena())); 873 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 874 env()->set_dependencies(new Dependencies(env())); 875 876 _fixed_slots = 0; 877 set_has_split_ifs(false); 878 set_has_loops(has_method() && method()->has_loops()); // first approximation 879 set_has_stringbuilder(false); 880 _trap_can_recompile = false; // no traps emitted yet 881 _major_progress = true; // start out assuming good things will happen 882 set_has_unsafe_access(false); 883 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 884 set_decompile_count(0); 885 886 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 887 set_num_loop_opts(LoopOptsCount); 888 set_do_inlining(Inline); 889 set_max_inline_size(MaxInlineSize); 890 set_freq_inline_size(FreqInlineSize); 891 set_do_scheduling(OptoScheduling); 892 set_do_count_invocations(false); 893 set_do_method_data_update(false); 894 895 if (debug_info()->recording_non_safepoints()) { 896 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 897 (comp_arena(), 8, 0, NULL)); 898 set_default_node_notes(Node_Notes::make(this)); 899 } 900 901 // // -- Initialize types before each compile -- 902 // // Update cached type information 903 // if( _method && _method->constants() ) 904 // Type::update_loaded_types(_method, _method->constants()); 905 906 // Init alias_type map. 907 if (!_do_escape_analysis && aliaslevel == 3) 908 aliaslevel = 2; // No unique types without escape analysis 909 _AliasLevel = aliaslevel; 910 const int grow_ats = 16; 911 _max_alias_types = grow_ats; 912 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 913 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 914 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 915 { 916 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 917 } 918 // Initialize the first few types. 919 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 920 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 921 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 922 _num_alias_types = AliasIdxRaw+1; 923 // Zero out the alias type cache. 924 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 925 // A NULL adr_type hits in the cache right away. Preload the right answer. 926 probe_alias_cache(NULL)->_index = AliasIdxTop; 927 928 _intrinsics = NULL; 929 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 930 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 931 register_library_intrinsics(); 932 } 933 934 //---------------------------init_start---------------------------------------- 935 // Install the StartNode on this compile object. 936 void Compile::init_start(StartNode* s) { 937 if (failing()) 938 return; // already failing 939 assert(s == start(), ""); 940 } 941 942 StartNode* Compile::start() const { 943 assert(!failing(), ""); 944 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 945 Node* start = root()->fast_out(i); 946 if( start->is_Start() ) 947 return start->as_Start(); 948 } 949 ShouldNotReachHere(); 950 return NULL; 951 } 952 953 //-------------------------------immutable_memory------------------------------------- 954 // Access immutable memory 955 Node* Compile::immutable_memory() { 956 if (_immutable_memory != NULL) { 957 return _immutable_memory; 958 } 959 StartNode* s = start(); 960 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 961 Node *p = s->fast_out(i); 962 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 963 _immutable_memory = p; 964 return _immutable_memory; 965 } 966 } 967 ShouldNotReachHere(); 968 return NULL; 969 } 970 971 //----------------------set_cached_top_node------------------------------------ 972 // Install the cached top node, and make sure Node::is_top works correctly. 973 void Compile::set_cached_top_node(Node* tn) { 974 if (tn != NULL) verify_top(tn); 975 Node* old_top = _top; 976 _top = tn; 977 // Calling Node::setup_is_top allows the nodes the chance to adjust 978 // their _out arrays. 979 if (_top != NULL) _top->setup_is_top(); 980 if (old_top != NULL) old_top->setup_is_top(); 981 assert(_top == NULL || top()->is_top(), ""); 982 } 983 984 #ifndef PRODUCT 985 void Compile::verify_top(Node* tn) const { 986 if (tn != NULL) { 987 assert(tn->is_Con(), "top node must be a constant"); 988 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 989 assert(tn->in(0) != NULL, "must have live top node"); 990 } 991 } 992 #endif 993 994 995 ///-------------------Managing Per-Node Debug & Profile Info------------------- 996 997 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 998 guarantee(arr != NULL, ""); 999 int num_blocks = arr->length(); 1000 if (grow_by < num_blocks) grow_by = num_blocks; 1001 int num_notes = grow_by * _node_notes_block_size; 1002 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1003 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1004 while (num_notes > 0) { 1005 arr->append(notes); 1006 notes += _node_notes_block_size; 1007 num_notes -= _node_notes_block_size; 1008 } 1009 assert(num_notes == 0, "exact multiple, please"); 1010 } 1011 1012 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1013 if (source == NULL || dest == NULL) return false; 1014 1015 if (dest->is_Con()) 1016 return false; // Do not push debug info onto constants. 1017 1018 #ifdef ASSERT 1019 // Leave a bread crumb trail pointing to the original node: 1020 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1021 dest->set_debug_orig(source); 1022 } 1023 #endif 1024 1025 if (node_note_array() == NULL) 1026 return false; // Not collecting any notes now. 1027 1028 // This is a copy onto a pre-existing node, which may already have notes. 1029 // If both nodes have notes, do not overwrite any pre-existing notes. 1030 Node_Notes* source_notes = node_notes_at(source->_idx); 1031 if (source_notes == NULL || source_notes->is_clear()) return false; 1032 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1033 if (dest_notes == NULL || dest_notes->is_clear()) { 1034 return set_node_notes_at(dest->_idx, source_notes); 1035 } 1036 1037 Node_Notes merged_notes = (*source_notes); 1038 // The order of operations here ensures that dest notes will win... 1039 merged_notes.update_from(dest_notes); 1040 return set_node_notes_at(dest->_idx, &merged_notes); 1041 } 1042 1043 1044 //--------------------------allow_range_check_smearing------------------------- 1045 // Gating condition for coalescing similar range checks. 1046 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1047 // single covering check that is at least as strong as any of them. 1048 // If the optimization succeeds, the simplified (strengthened) range check 1049 // will always succeed. If it fails, we will deopt, and then give up 1050 // on the optimization. 1051 bool Compile::allow_range_check_smearing() const { 1052 // If this method has already thrown a range-check, 1053 // assume it was because we already tried range smearing 1054 // and it failed. 1055 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1056 return !already_trapped; 1057 } 1058 1059 1060 //------------------------------flatten_alias_type----------------------------- 1061 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1062 int offset = tj->offset(); 1063 TypePtr::PTR ptr = tj->ptr(); 1064 1065 // Known instance (scalarizable allocation) alias only with itself. 1066 bool is_known_inst = tj->isa_oopptr() != NULL && 1067 tj->is_oopptr()->is_known_instance(); 1068 1069 // Process weird unsafe references. 1070 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1071 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1072 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1073 tj = TypeOopPtr::BOTTOM; 1074 ptr = tj->ptr(); 1075 offset = tj->offset(); 1076 } 1077 1078 // Array pointers need some flattening 1079 const TypeAryPtr *ta = tj->isa_aryptr(); 1080 if( ta && is_known_inst ) { 1081 if ( offset != Type::OffsetBot && 1082 offset > arrayOopDesc::length_offset_in_bytes() ) { 1083 offset = Type::OffsetBot; // Flatten constant access into array body only 1084 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1085 } 1086 } else if( ta && _AliasLevel >= 2 ) { 1087 // For arrays indexed by constant indices, we flatten the alias 1088 // space to include all of the array body. Only the header, klass 1089 // and array length can be accessed un-aliased. 1090 if( offset != Type::OffsetBot ) { 1091 if( ta->const_oop() ) { // methodDataOop or methodOop 1092 offset = Type::OffsetBot; // Flatten constant access into array body 1093 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1094 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1095 // range is OK as-is. 1096 tj = ta = TypeAryPtr::RANGE; 1097 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1098 tj = TypeInstPtr::KLASS; // all klass loads look alike 1099 ta = TypeAryPtr::RANGE; // generic ignored junk 1100 ptr = TypePtr::BotPTR; 1101 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1102 tj = TypeInstPtr::MARK; 1103 ta = TypeAryPtr::RANGE; // generic ignored junk 1104 ptr = TypePtr::BotPTR; 1105 } else { // Random constant offset into array body 1106 offset = Type::OffsetBot; // Flatten constant access into array body 1107 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1108 } 1109 } 1110 // Arrays of fixed size alias with arrays of unknown size. 1111 if (ta->size() != TypeInt::POS) { 1112 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1113 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1114 } 1115 // Arrays of known objects become arrays of unknown objects. 1116 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1117 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1118 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1119 } 1120 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1121 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1122 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1123 } 1124 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1125 // cannot be distinguished by bytecode alone. 1126 if (ta->elem() == TypeInt::BOOL) { 1127 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1128 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1129 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1130 } 1131 // During the 2nd round of IterGVN, NotNull castings are removed. 1132 // Make sure the Bottom and NotNull variants alias the same. 1133 // Also, make sure exact and non-exact variants alias the same. 1134 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1135 if (ta->const_oop()) { 1136 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1137 } else { 1138 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1139 } 1140 } 1141 } 1142 1143 // Oop pointers need some flattening 1144 const TypeInstPtr *to = tj->isa_instptr(); 1145 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1146 if( ptr == TypePtr::Constant ) { 1147 // No constant oop pointers (such as Strings); they alias with 1148 // unknown strings. 1149 assert(!is_known_inst, "not scalarizable allocation"); 1150 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1151 } else if( is_known_inst ) { 1152 tj = to; // Keep NotNull and klass_is_exact for instance type 1153 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1154 // During the 2nd round of IterGVN, NotNull castings are removed. 1155 // Make sure the Bottom and NotNull variants alias the same. 1156 // Also, make sure exact and non-exact variants alias the same. 1157 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1158 } 1159 // Canonicalize the holder of this field 1160 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1161 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1162 // First handle header references such as a LoadKlassNode, even if the 1163 // object's klass is unloaded at compile time (4965979). 1164 if (!is_known_inst) { // Do it only for non-instance types 1165 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1166 } 1167 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1168 to = NULL; 1169 tj = TypeOopPtr::BOTTOM; 1170 offset = tj->offset(); 1171 } else { 1172 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1173 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1174 if( is_known_inst ) { 1175 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1176 } else { 1177 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1178 } 1179 } 1180 } 1181 } 1182 1183 // Klass pointers to object array klasses need some flattening 1184 const TypeKlassPtr *tk = tj->isa_klassptr(); 1185 if( tk ) { 1186 // If we are referencing a field within a Klass, we need 1187 // to assume the worst case of an Object. Both exact and 1188 // inexact types must flatten to the same alias class. 1189 // Since the flattened result for a klass is defined to be 1190 // precisely java.lang.Object, use a constant ptr. 1191 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1192 1193 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1194 TypeKlassPtr::OBJECT->klass(), 1195 offset); 1196 } 1197 1198 ciKlass* klass = tk->klass(); 1199 if( klass->is_obj_array_klass() ) { 1200 ciKlass* k = TypeAryPtr::OOPS->klass(); 1201 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1202 k = TypeInstPtr::BOTTOM->klass(); 1203 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1204 } 1205 1206 // Check for precise loads from the primary supertype array and force them 1207 // to the supertype cache alias index. Check for generic array loads from 1208 // the primary supertype array and also force them to the supertype cache 1209 // alias index. Since the same load can reach both, we need to merge 1210 // these 2 disparate memories into the same alias class. Since the 1211 // primary supertype array is read-only, there's no chance of confusion 1212 // where we bypass an array load and an array store. 1213 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1214 if( offset == Type::OffsetBot || 1215 off2 < Klass::primary_super_limit()*wordSize ) { 1216 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1217 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1218 } 1219 } 1220 1221 // Flatten all Raw pointers together. 1222 if (tj->base() == Type::RawPtr) 1223 tj = TypeRawPtr::BOTTOM; 1224 1225 if (tj->base() == Type::AnyPtr) 1226 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1227 1228 // Flatten all to bottom for now 1229 switch( _AliasLevel ) { 1230 case 0: 1231 tj = TypePtr::BOTTOM; 1232 break; 1233 case 1: // Flatten to: oop, static, field or array 1234 switch (tj->base()) { 1235 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1236 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1237 case Type::AryPtr: // do not distinguish arrays at all 1238 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1239 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1240 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1241 default: ShouldNotReachHere(); 1242 } 1243 break; 1244 case 2: // No collapsing at level 2; keep all splits 1245 case 3: // No collapsing at level 3; keep all splits 1246 break; 1247 default: 1248 Unimplemented(); 1249 } 1250 1251 offset = tj->offset(); 1252 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1253 1254 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1255 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1256 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1257 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1258 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1259 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1260 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1261 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1262 assert( tj->ptr() != TypePtr::TopPTR && 1263 tj->ptr() != TypePtr::AnyNull && 1264 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1265 // assert( tj->ptr() != TypePtr::Constant || 1266 // tj->base() == Type::RawPtr || 1267 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1268 1269 return tj; 1270 } 1271 1272 void Compile::AliasType::Init(int i, const TypePtr* at) { 1273 _index = i; 1274 _adr_type = at; 1275 _field = NULL; 1276 _is_rewritable = true; // default 1277 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1278 if (atoop != NULL && atoop->is_known_instance()) { 1279 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1280 _general_index = Compile::current()->get_alias_index(gt); 1281 } else { 1282 _general_index = 0; 1283 } 1284 } 1285 1286 //---------------------------------print_on------------------------------------ 1287 #ifndef PRODUCT 1288 void Compile::AliasType::print_on(outputStream* st) { 1289 if (index() < 10) 1290 st->print("@ <%d> ", index()); 1291 else st->print("@ <%d>", index()); 1292 st->print(is_rewritable() ? " " : " RO"); 1293 int offset = adr_type()->offset(); 1294 if (offset == Type::OffsetBot) 1295 st->print(" +any"); 1296 else st->print(" +%-3d", offset); 1297 st->print(" in "); 1298 adr_type()->dump_on(st); 1299 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1300 if (field() != NULL && tjp) { 1301 if (tjp->klass() != field()->holder() || 1302 tjp->offset() != field()->offset_in_bytes()) { 1303 st->print(" != "); 1304 field()->print(); 1305 st->print(" ***"); 1306 } 1307 } 1308 } 1309 1310 void print_alias_types() { 1311 Compile* C = Compile::current(); 1312 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1313 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1314 C->alias_type(idx)->print_on(tty); 1315 tty->cr(); 1316 } 1317 } 1318 #endif 1319 1320 1321 //----------------------------probe_alias_cache-------------------------------- 1322 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1323 intptr_t key = (intptr_t) adr_type; 1324 key ^= key >> logAliasCacheSize; 1325 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1326 } 1327 1328 1329 //-----------------------------grow_alias_types-------------------------------- 1330 void Compile::grow_alias_types() { 1331 const int old_ats = _max_alias_types; // how many before? 1332 const int new_ats = old_ats; // how many more? 1333 const int grow_ats = old_ats+new_ats; // how many now? 1334 _max_alias_types = grow_ats; 1335 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1336 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1337 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1338 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1339 } 1340 1341 1342 //--------------------------------find_alias_type------------------------------ 1343 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1344 if (_AliasLevel == 0) 1345 return alias_type(AliasIdxBot); 1346 1347 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1348 if (ace->_adr_type == adr_type) { 1349 return alias_type(ace->_index); 1350 } 1351 1352 // Handle special cases. 1353 if (adr_type == NULL) return alias_type(AliasIdxTop); 1354 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1355 1356 // Do it the slow way. 1357 const TypePtr* flat = flatten_alias_type(adr_type); 1358 1359 #ifdef ASSERT 1360 assert(flat == flatten_alias_type(flat), "idempotent"); 1361 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1362 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1363 const TypeOopPtr* foop = flat->is_oopptr(); 1364 // Scalarizable allocations have exact klass always. 1365 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1366 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1367 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1368 } 1369 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1370 #endif 1371 1372 int idx = AliasIdxTop; 1373 for (int i = 0; i < num_alias_types(); i++) { 1374 if (alias_type(i)->adr_type() == flat) { 1375 idx = i; 1376 break; 1377 } 1378 } 1379 1380 if (idx == AliasIdxTop) { 1381 if (no_create) return NULL; 1382 // Grow the array if necessary. 1383 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1384 // Add a new alias type. 1385 idx = _num_alias_types++; 1386 _alias_types[idx]->Init(idx, flat); 1387 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1388 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1389 if (flat->isa_instptr()) { 1390 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1391 && flat->is_instptr()->klass() == env()->Class_klass()) 1392 alias_type(idx)->set_rewritable(false); 1393 } 1394 if (flat->isa_klassptr()) { 1395 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1396 alias_type(idx)->set_rewritable(false); 1397 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1398 alias_type(idx)->set_rewritable(false); 1399 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1400 alias_type(idx)->set_rewritable(false); 1401 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1402 alias_type(idx)->set_rewritable(false); 1403 } 1404 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1405 // but the base pointer type is not distinctive enough to identify 1406 // references into JavaThread.) 1407 1408 // Check for final instance fields. 1409 const TypeInstPtr* tinst = flat->isa_instptr(); 1410 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1411 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1412 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1413 // Set field() and is_rewritable() attributes. 1414 if (field != NULL) alias_type(idx)->set_field(field); 1415 } 1416 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1417 // Check for final static fields. 1418 if (tklass && tklass->klass()->is_instance_klass()) { 1419 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1420 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1421 // Set field() and is_rewritable() attributes. 1422 if (field != NULL) alias_type(idx)->set_field(field); 1423 } 1424 } 1425 1426 // Fill the cache for next time. 1427 ace->_adr_type = adr_type; 1428 ace->_index = idx; 1429 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1430 1431 // Might as well try to fill the cache for the flattened version, too. 1432 AliasCacheEntry* face = probe_alias_cache(flat); 1433 if (face->_adr_type == NULL) { 1434 face->_adr_type = flat; 1435 face->_index = idx; 1436 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1437 } 1438 1439 return alias_type(idx); 1440 } 1441 1442 1443 Compile::AliasType* Compile::alias_type(ciField* field) { 1444 const TypeOopPtr* t; 1445 if (field->is_static()) 1446 t = TypeKlassPtr::make(field->holder()); 1447 else 1448 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1449 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1450 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1451 return atp; 1452 } 1453 1454 1455 //------------------------------have_alias_type-------------------------------- 1456 bool Compile::have_alias_type(const TypePtr* adr_type) { 1457 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1458 if (ace->_adr_type == adr_type) { 1459 return true; 1460 } 1461 1462 // Handle special cases. 1463 if (adr_type == NULL) return true; 1464 if (adr_type == TypePtr::BOTTOM) return true; 1465 1466 return find_alias_type(adr_type, true) != NULL; 1467 } 1468 1469 //-----------------------------must_alias-------------------------------------- 1470 // True if all values of the given address type are in the given alias category. 1471 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1472 if (alias_idx == AliasIdxBot) return true; // the universal category 1473 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1474 if (alias_idx == AliasIdxTop) return false; // the empty category 1475 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1476 1477 // the only remaining possible overlap is identity 1478 int adr_idx = get_alias_index(adr_type); 1479 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1480 assert(adr_idx == alias_idx || 1481 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1482 && adr_type != TypeOopPtr::BOTTOM), 1483 "should not be testing for overlap with an unsafe pointer"); 1484 return adr_idx == alias_idx; 1485 } 1486 1487 //------------------------------can_alias-------------------------------------- 1488 // True if any values of the given address type are in the given alias category. 1489 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1490 if (alias_idx == AliasIdxTop) return false; // the empty category 1491 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1492 if (alias_idx == AliasIdxBot) return true; // the universal category 1493 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1494 1495 // the only remaining possible overlap is identity 1496 int adr_idx = get_alias_index(adr_type); 1497 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1498 return adr_idx == alias_idx; 1499 } 1500 1501 1502 1503 //---------------------------pop_warm_call------------------------------------- 1504 WarmCallInfo* Compile::pop_warm_call() { 1505 WarmCallInfo* wci = _warm_calls; 1506 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1507 return wci; 1508 } 1509 1510 //----------------------------Inline_Warm-------------------------------------- 1511 int Compile::Inline_Warm() { 1512 // If there is room, try to inline some more warm call sites. 1513 // %%% Do a graph index compaction pass when we think we're out of space? 1514 if (!InlineWarmCalls) return 0; 1515 1516 int calls_made_hot = 0; 1517 int room_to_grow = NodeCountInliningCutoff - unique(); 1518 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1519 int amount_grown = 0; 1520 WarmCallInfo* call; 1521 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1522 int est_size = (int)call->size(); 1523 if (est_size > (room_to_grow - amount_grown)) { 1524 // This one won't fit anyway. Get rid of it. 1525 call->make_cold(); 1526 continue; 1527 } 1528 call->make_hot(); 1529 calls_made_hot++; 1530 amount_grown += est_size; 1531 amount_to_grow -= est_size; 1532 } 1533 1534 if (calls_made_hot > 0) set_major_progress(); 1535 return calls_made_hot; 1536 } 1537 1538 1539 //----------------------------Finish_Warm-------------------------------------- 1540 void Compile::Finish_Warm() { 1541 if (!InlineWarmCalls) return; 1542 if (failing()) return; 1543 if (warm_calls() == NULL) return; 1544 1545 // Clean up loose ends, if we are out of space for inlining. 1546 WarmCallInfo* call; 1547 while ((call = pop_warm_call()) != NULL) { 1548 call->make_cold(); 1549 } 1550 } 1551 1552 //---------------------cleanup_loop_predicates----------------------- 1553 // Remove the opaque nodes that protect the predicates so that all unused 1554 // checks and uncommon_traps will be eliminated from the ideal graph 1555 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1556 if (predicate_count()==0) return; 1557 for (int i = predicate_count(); i > 0; i--) { 1558 Node * n = predicate_opaque1_node(i-1); 1559 assert(n->Opcode() == Op_Opaque1, "must be"); 1560 igvn.replace_node(n, n->in(1)); 1561 } 1562 assert(predicate_count()==0, "should be clean!"); 1563 igvn.optimize(); 1564 } 1565 1566 //------------------------------Optimize--------------------------------------- 1567 // Given a graph, optimize it. 1568 void Compile::Optimize() { 1569 TracePhase t1("optimizer", &_t_optimizer, true); 1570 1571 #ifndef PRODUCT 1572 if (env()->break_at_compile()) { 1573 BREAKPOINT; 1574 } 1575 1576 #endif 1577 1578 ResourceMark rm; 1579 int loop_opts_cnt; 1580 1581 NOT_PRODUCT( verify_graph_edges(); ) 1582 1583 print_method("After Parsing"); 1584 1585 { 1586 // Iterative Global Value Numbering, including ideal transforms 1587 // Initialize IterGVN with types and values from parse-time GVN 1588 PhaseIterGVN igvn(initial_gvn()); 1589 { 1590 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1591 igvn.optimize(); 1592 } 1593 1594 print_method("Iter GVN 1", 2); 1595 1596 if (failing()) return; 1597 1598 // Perform escape analysis 1599 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1600 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 1601 ConnectionGraph::do_analysis(this, &igvn); 1602 1603 if (failing()) return; 1604 1605 igvn.optimize(); 1606 print_method("Iter GVN 3", 2); 1607 1608 if (failing()) return; 1609 1610 } 1611 1612 // Loop transforms on the ideal graph. Range Check Elimination, 1613 // peeling, unrolling, etc. 1614 1615 // Set loop opts counter 1616 loop_opts_cnt = num_loop_opts(); 1617 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1618 { 1619 TracePhase t2("idealLoop", &_t_idealLoop, true); 1620 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate); 1621 loop_opts_cnt--; 1622 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1623 if (failing()) return; 1624 } 1625 // Loop opts pass if partial peeling occurred in previous pass 1626 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1627 TracePhase t3("idealLoop", &_t_idealLoop, true); 1628 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1629 loop_opts_cnt--; 1630 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1631 if (failing()) return; 1632 } 1633 // Loop opts pass for loop-unrolling before CCP 1634 if(major_progress() && (loop_opts_cnt > 0)) { 1635 TracePhase t4("idealLoop", &_t_idealLoop, true); 1636 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1637 loop_opts_cnt--; 1638 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1639 } 1640 if (!failing()) { 1641 // Verify that last round of loop opts produced a valid graph 1642 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1643 PhaseIdealLoop::verify(igvn); 1644 } 1645 } 1646 if (failing()) return; 1647 1648 // Conditional Constant Propagation; 1649 PhaseCCP ccp( &igvn ); 1650 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1651 { 1652 TracePhase t2("ccp", &_t_ccp, true); 1653 ccp.do_transform(); 1654 } 1655 print_method("PhaseCPP 1", 2); 1656 1657 assert( true, "Break here to ccp.dump_old2new_map()"); 1658 1659 // Iterative Global Value Numbering, including ideal transforms 1660 { 1661 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1662 igvn = ccp; 1663 igvn.optimize(); 1664 } 1665 1666 print_method("Iter GVN 2", 2); 1667 1668 if (failing()) return; 1669 1670 // Loop transforms on the ideal graph. Range Check Elimination, 1671 // peeling, unrolling, etc. 1672 if(loop_opts_cnt > 0) { 1673 debug_only( int cnt = 0; ); 1674 bool loop_predication = UseLoopPredicate; 1675 while(major_progress() && (loop_opts_cnt > 0)) { 1676 TracePhase t2("idealLoop", &_t_idealLoop, true); 1677 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1678 PhaseIdealLoop ideal_loop( igvn, true, loop_predication); 1679 loop_opts_cnt--; 1680 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1681 if (failing()) return; 1682 // Perform loop predication optimization during first iteration after CCP. 1683 // After that switch it off and cleanup unused loop predicates. 1684 if (loop_predication) { 1685 loop_predication = false; 1686 cleanup_loop_predicates(igvn); 1687 if (failing()) return; 1688 } 1689 } 1690 } 1691 1692 { 1693 // Verify that all previous optimizations produced a valid graph 1694 // at least to this point, even if no loop optimizations were done. 1695 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1696 PhaseIdealLoop::verify(igvn); 1697 } 1698 1699 { 1700 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1701 PhaseMacroExpand mex(igvn); 1702 if (mex.expand_macro_nodes()) { 1703 assert(failing(), "must bail out w/ explicit message"); 1704 return; 1705 } 1706 } 1707 1708 } // (End scope of igvn; run destructor if necessary for asserts.) 1709 1710 // A method with only infinite loops has no edges entering loops from root 1711 { 1712 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1713 if (final_graph_reshaping()) { 1714 assert(failing(), "must bail out w/ explicit message"); 1715 return; 1716 } 1717 } 1718 1719 print_method("Optimize finished", 2); 1720 } 1721 1722 1723 //------------------------------Code_Gen--------------------------------------- 1724 // Given a graph, generate code for it 1725 void Compile::Code_Gen() { 1726 if (failing()) return; 1727 1728 // Perform instruction selection. You might think we could reclaim Matcher 1729 // memory PDQ, but actually the Matcher is used in generating spill code. 1730 // Internals of the Matcher (including some VectorSets) must remain live 1731 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1732 // set a bit in reclaimed memory. 1733 1734 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1735 // nodes. Mapping is only valid at the root of each matched subtree. 1736 NOT_PRODUCT( verify_graph_edges(); ) 1737 1738 Node_List proj_list; 1739 Matcher m(proj_list); 1740 _matcher = &m; 1741 { 1742 TracePhase t2("matcher", &_t_matcher, true); 1743 m.match(); 1744 } 1745 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1746 // nodes. Mapping is only valid at the root of each matched subtree. 1747 NOT_PRODUCT( verify_graph_edges(); ) 1748 1749 // If you have too many nodes, or if matching has failed, bail out 1750 check_node_count(0, "out of nodes matching instructions"); 1751 if (failing()) return; 1752 1753 // Build a proper-looking CFG 1754 PhaseCFG cfg(node_arena(), root(), m); 1755 _cfg = &cfg; 1756 { 1757 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1758 cfg.Dominators(); 1759 if (failing()) return; 1760 1761 NOT_PRODUCT( verify_graph_edges(); ) 1762 1763 cfg.Estimate_Block_Frequency(); 1764 cfg.GlobalCodeMotion(m,unique(),proj_list); 1765 1766 print_method("Global code motion", 2); 1767 1768 if (failing()) return; 1769 NOT_PRODUCT( verify_graph_edges(); ) 1770 1771 debug_only( cfg.verify(); ) 1772 } 1773 NOT_PRODUCT( verify_graph_edges(); ) 1774 1775 PhaseChaitin regalloc(unique(),cfg,m); 1776 _regalloc = ®alloc; 1777 { 1778 TracePhase t2("regalloc", &_t_registerAllocation, true); 1779 // Perform any platform dependent preallocation actions. This is used, 1780 // for example, to avoid taking an implicit null pointer exception 1781 // using the frame pointer on win95. 1782 _regalloc->pd_preallocate_hook(); 1783 1784 // Perform register allocation. After Chaitin, use-def chains are 1785 // no longer accurate (at spill code) and so must be ignored. 1786 // Node->LRG->reg mappings are still accurate. 1787 _regalloc->Register_Allocate(); 1788 1789 // Bail out if the allocator builds too many nodes 1790 if (failing()) return; 1791 } 1792 1793 // Prior to register allocation we kept empty basic blocks in case the 1794 // the allocator needed a place to spill. After register allocation we 1795 // are not adding any new instructions. If any basic block is empty, we 1796 // can now safely remove it. 1797 { 1798 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1799 cfg.remove_empty(); 1800 if (do_freq_based_layout()) { 1801 PhaseBlockLayout layout(cfg); 1802 } else { 1803 cfg.set_loop_alignment(); 1804 } 1805 cfg.fixup_flow(); 1806 } 1807 1808 // Perform any platform dependent postallocation verifications. 1809 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1810 1811 // Apply peephole optimizations 1812 if( OptoPeephole ) { 1813 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1814 PhasePeephole peep( _regalloc, cfg); 1815 peep.do_transform(); 1816 } 1817 1818 // Convert Nodes to instruction bits in a buffer 1819 { 1820 // %%%% workspace merge brought two timers together for one job 1821 TracePhase t2a("output", &_t_output, true); 1822 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1823 Output(); 1824 } 1825 1826 print_method("Final Code"); 1827 1828 // He's dead, Jim. 1829 _cfg = (PhaseCFG*)0xdeadbeef; 1830 _regalloc = (PhaseChaitin*)0xdeadbeef; 1831 } 1832 1833 1834 //------------------------------dump_asm--------------------------------------- 1835 // Dump formatted assembly 1836 #ifndef PRODUCT 1837 void Compile::dump_asm(int *pcs, uint pc_limit) { 1838 bool cut_short = false; 1839 tty->print_cr("#"); 1840 tty->print("# "); _tf->dump(); tty->cr(); 1841 tty->print_cr("#"); 1842 1843 // For all blocks 1844 int pc = 0x0; // Program counter 1845 char starts_bundle = ' '; 1846 _regalloc->dump_frame(); 1847 1848 Node *n = NULL; 1849 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1850 if (VMThread::should_terminate()) { cut_short = true; break; } 1851 Block *b = _cfg->_blocks[i]; 1852 if (b->is_connector() && !Verbose) continue; 1853 n = b->_nodes[0]; 1854 if (pcs && n->_idx < pc_limit) 1855 tty->print("%3.3x ", pcs[n->_idx]); 1856 else 1857 tty->print(" "); 1858 b->dump_head( &_cfg->_bbs ); 1859 if (b->is_connector()) { 1860 tty->print_cr(" # Empty connector block"); 1861 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1862 tty->print_cr(" # Block is sole successor of call"); 1863 } 1864 1865 // For all instructions 1866 Node *delay = NULL; 1867 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1868 if (VMThread::should_terminate()) { cut_short = true; break; } 1869 n = b->_nodes[j]; 1870 if (valid_bundle_info(n)) { 1871 Bundle *bundle = node_bundling(n); 1872 if (bundle->used_in_unconditional_delay()) { 1873 delay = n; 1874 continue; 1875 } 1876 if (bundle->starts_bundle()) 1877 starts_bundle = '+'; 1878 } 1879 1880 if (WizardMode) n->dump(); 1881 1882 if( !n->is_Region() && // Dont print in the Assembly 1883 !n->is_Phi() && // a few noisely useless nodes 1884 !n->is_Proj() && 1885 !n->is_MachTemp() && 1886 !n->is_SafePointScalarObject() && 1887 !n->is_Catch() && // Would be nice to print exception table targets 1888 !n->is_MergeMem() && // Not very interesting 1889 !n->is_top() && // Debug info table constants 1890 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1891 ) { 1892 if (pcs && n->_idx < pc_limit) 1893 tty->print("%3.3x", pcs[n->_idx]); 1894 else 1895 tty->print(" "); 1896 tty->print(" %c ", starts_bundle); 1897 starts_bundle = ' '; 1898 tty->print("\t"); 1899 n->format(_regalloc, tty); 1900 tty->cr(); 1901 } 1902 1903 // If we have an instruction with a delay slot, and have seen a delay, 1904 // then back up and print it 1905 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1906 assert(delay != NULL, "no unconditional delay instruction"); 1907 if (WizardMode) delay->dump(); 1908 1909 if (node_bundling(delay)->starts_bundle()) 1910 starts_bundle = '+'; 1911 if (pcs && n->_idx < pc_limit) 1912 tty->print("%3.3x", pcs[n->_idx]); 1913 else 1914 tty->print(" "); 1915 tty->print(" %c ", starts_bundle); 1916 starts_bundle = ' '; 1917 tty->print("\t"); 1918 delay->format(_regalloc, tty); 1919 tty->print_cr(""); 1920 delay = NULL; 1921 } 1922 1923 // Dump the exception table as well 1924 if( n->is_Catch() && (Verbose || WizardMode) ) { 1925 // Print the exception table for this offset 1926 _handler_table.print_subtable_for(pc); 1927 } 1928 } 1929 1930 if (pcs && n->_idx < pc_limit) 1931 tty->print_cr("%3.3x", pcs[n->_idx]); 1932 else 1933 tty->print_cr(""); 1934 1935 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1936 1937 } // End of per-block dump 1938 tty->print_cr(""); 1939 1940 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1941 } 1942 #endif 1943 1944 //------------------------------Final_Reshape_Counts--------------------------- 1945 // This class defines counters to help identify when a method 1946 // may/must be executed using hardware with only 24-bit precision. 1947 struct Final_Reshape_Counts : public StackObj { 1948 int _call_count; // count non-inlined 'common' calls 1949 int _float_count; // count float ops requiring 24-bit precision 1950 int _double_count; // count double ops requiring more precision 1951 int _java_call_count; // count non-inlined 'java' calls 1952 int _inner_loop_count; // count loops which need alignment 1953 VectorSet _visited; // Visitation flags 1954 Node_List _tests; // Set of IfNodes & PCTableNodes 1955 1956 Final_Reshape_Counts() : 1957 _call_count(0), _float_count(0), _double_count(0), 1958 _java_call_count(0), _inner_loop_count(0), 1959 _visited( Thread::current()->resource_area() ) { } 1960 1961 void inc_call_count () { _call_count ++; } 1962 void inc_float_count () { _float_count ++; } 1963 void inc_double_count() { _double_count++; } 1964 void inc_java_call_count() { _java_call_count++; } 1965 void inc_inner_loop_count() { _inner_loop_count++; } 1966 1967 int get_call_count () const { return _call_count ; } 1968 int get_float_count () const { return _float_count ; } 1969 int get_double_count() const { return _double_count; } 1970 int get_java_call_count() const { return _java_call_count; } 1971 int get_inner_loop_count() const { return _inner_loop_count; } 1972 }; 1973 1974 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 1975 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 1976 // Make sure the offset goes inside the instance layout. 1977 return k->contains_field_offset(tp->offset()); 1978 // Note that OffsetBot and OffsetTop are very negative. 1979 } 1980 1981 //------------------------------final_graph_reshaping_impl---------------------- 1982 // Implement items 1-5 from final_graph_reshaping below. 1983 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 1984 1985 if ( n->outcnt() == 0 ) return; // dead node 1986 uint nop = n->Opcode(); 1987 1988 // Check for 2-input instruction with "last use" on right input. 1989 // Swap to left input. Implements item (2). 1990 if( n->req() == 3 && // two-input instruction 1991 n->in(1)->outcnt() > 1 && // left use is NOT a last use 1992 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 1993 n->in(2)->outcnt() == 1 &&// right use IS a last use 1994 !n->in(2)->is_Con() ) { // right use is not a constant 1995 // Check for commutative opcode 1996 switch( nop ) { 1997 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 1998 case Op_MaxI: case Op_MinI: 1999 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2000 case Op_AndL: case Op_XorL: case Op_OrL: 2001 case Op_AndI: case Op_XorI: case Op_OrI: { 2002 // Move "last use" input to left by swapping inputs 2003 n->swap_edges(1, 2); 2004 break; 2005 } 2006 default: 2007 break; 2008 } 2009 } 2010 2011 #ifdef ASSERT 2012 if( n->is_Mem() ) { 2013 Compile* C = Compile::current(); 2014 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2015 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2016 // oop will be recorded in oop map if load crosses safepoint 2017 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2018 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2019 "raw memory operations should have control edge"); 2020 } 2021 #endif 2022 // Count FPU ops and common calls, implements item (3) 2023 switch( nop ) { 2024 // Count all float operations that may use FPU 2025 case Op_AddF: 2026 case Op_SubF: 2027 case Op_MulF: 2028 case Op_DivF: 2029 case Op_NegF: 2030 case Op_ModF: 2031 case Op_ConvI2F: 2032 case Op_ConF: 2033 case Op_CmpF: 2034 case Op_CmpF3: 2035 // case Op_ConvL2F: // longs are split into 32-bit halves 2036 frc.inc_float_count(); 2037 break; 2038 2039 case Op_ConvF2D: 2040 case Op_ConvD2F: 2041 frc.inc_float_count(); 2042 frc.inc_double_count(); 2043 break; 2044 2045 // Count all double operations that may use FPU 2046 case Op_AddD: 2047 case Op_SubD: 2048 case Op_MulD: 2049 case Op_DivD: 2050 case Op_NegD: 2051 case Op_ModD: 2052 case Op_ConvI2D: 2053 case Op_ConvD2I: 2054 // case Op_ConvL2D: // handled by leaf call 2055 // case Op_ConvD2L: // handled by leaf call 2056 case Op_ConD: 2057 case Op_CmpD: 2058 case Op_CmpD3: 2059 frc.inc_double_count(); 2060 break; 2061 case Op_Opaque1: // Remove Opaque Nodes before matching 2062 case Op_Opaque2: // Remove Opaque Nodes before matching 2063 n->subsume_by(n->in(1)); 2064 break; 2065 case Op_CallStaticJava: 2066 case Op_CallJava: 2067 case Op_CallDynamicJava: 2068 frc.inc_java_call_count(); // Count java call site; 2069 case Op_CallRuntime: 2070 case Op_CallLeaf: 2071 case Op_CallLeafNoFP: { 2072 assert( n->is_Call(), "" ); 2073 CallNode *call = n->as_Call(); 2074 // Count call sites where the FP mode bit would have to be flipped. 2075 // Do not count uncommon runtime calls: 2076 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2077 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2078 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2079 frc.inc_call_count(); // Count the call site 2080 } else { // See if uncommon argument is shared 2081 Node *n = call->in(TypeFunc::Parms); 2082 int nop = n->Opcode(); 2083 // Clone shared simple arguments to uncommon calls, item (1). 2084 if( n->outcnt() > 1 && 2085 !n->is_Proj() && 2086 nop != Op_CreateEx && 2087 nop != Op_CheckCastPP && 2088 nop != Op_DecodeN && 2089 !n->is_Mem() ) { 2090 Node *x = n->clone(); 2091 call->set_req( TypeFunc::Parms, x ); 2092 } 2093 } 2094 break; 2095 } 2096 2097 case Op_StoreD: 2098 case Op_LoadD: 2099 case Op_LoadD_unaligned: 2100 frc.inc_double_count(); 2101 goto handle_mem; 2102 case Op_StoreF: 2103 case Op_LoadF: 2104 frc.inc_float_count(); 2105 goto handle_mem; 2106 2107 case Op_StoreB: 2108 case Op_StoreC: 2109 case Op_StoreCM: 2110 case Op_StorePConditional: 2111 case Op_StoreI: 2112 case Op_StoreL: 2113 case Op_StoreIConditional: 2114 case Op_StoreLConditional: 2115 case Op_CompareAndSwapI: 2116 case Op_CompareAndSwapL: 2117 case Op_CompareAndSwapP: 2118 case Op_CompareAndSwapN: 2119 case Op_StoreP: 2120 case Op_StoreN: 2121 case Op_LoadB: 2122 case Op_LoadUB: 2123 case Op_LoadUS: 2124 case Op_LoadI: 2125 case Op_LoadUI2L: 2126 case Op_LoadKlass: 2127 case Op_LoadNKlass: 2128 case Op_LoadL: 2129 case Op_LoadL_unaligned: 2130 case Op_LoadPLocked: 2131 case Op_LoadLLocked: 2132 case Op_LoadP: 2133 case Op_LoadN: 2134 case Op_LoadRange: 2135 case Op_LoadS: { 2136 handle_mem: 2137 #ifdef ASSERT 2138 if( VerifyOptoOopOffsets ) { 2139 assert( n->is_Mem(), "" ); 2140 MemNode *mem = (MemNode*)n; 2141 // Check to see if address types have grounded out somehow. 2142 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2143 assert( !tp || oop_offset_is_sane(tp), "" ); 2144 } 2145 #endif 2146 break; 2147 } 2148 2149 case Op_AddP: { // Assert sane base pointers 2150 Node *addp = n->in(AddPNode::Address); 2151 assert( !addp->is_AddP() || 2152 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2153 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2154 "Base pointers must match" ); 2155 #ifdef _LP64 2156 if (UseCompressedOops && 2157 addp->Opcode() == Op_ConP && 2158 addp == n->in(AddPNode::Base) && 2159 n->in(AddPNode::Offset)->is_Con()) { 2160 // Use addressing with narrow klass to load with offset on x86. 2161 // On sparc loading 32-bits constant and decoding it have less 2162 // instructions (4) then load 64-bits constant (7). 2163 // Do this transformation here since IGVN will convert ConN back to ConP. 2164 const Type* t = addp->bottom_type(); 2165 if (t->isa_oopptr()) { 2166 Node* nn = NULL; 2167 2168 // Look for existing ConN node of the same exact type. 2169 Compile* C = Compile::current(); 2170 Node* r = C->root(); 2171 uint cnt = r->outcnt(); 2172 for (uint i = 0; i < cnt; i++) { 2173 Node* m = r->raw_out(i); 2174 if (m!= NULL && m->Opcode() == Op_ConN && 2175 m->bottom_type()->make_ptr() == t) { 2176 nn = m; 2177 break; 2178 } 2179 } 2180 if (nn != NULL) { 2181 // Decode a narrow oop to match address 2182 // [R12 + narrow_oop_reg<<3 + offset] 2183 nn = new (C, 2) DecodeNNode(nn, t); 2184 n->set_req(AddPNode::Base, nn); 2185 n->set_req(AddPNode::Address, nn); 2186 if (addp->outcnt() == 0) { 2187 addp->disconnect_inputs(NULL); 2188 } 2189 } 2190 } 2191 } 2192 #endif 2193 break; 2194 } 2195 2196 #ifdef _LP64 2197 case Op_CastPP: 2198 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2199 Compile* C = Compile::current(); 2200 Node* in1 = n->in(1); 2201 const Type* t = n->bottom_type(); 2202 Node* new_in1 = in1->clone(); 2203 new_in1->as_DecodeN()->set_type(t); 2204 2205 if (!Matcher::narrow_oop_use_complex_address()) { 2206 // 2207 // x86, ARM and friends can handle 2 adds in addressing mode 2208 // and Matcher can fold a DecodeN node into address by using 2209 // a narrow oop directly and do implicit NULL check in address: 2210 // 2211 // [R12 + narrow_oop_reg<<3 + offset] 2212 // NullCheck narrow_oop_reg 2213 // 2214 // On other platforms (Sparc) we have to keep new DecodeN node and 2215 // use it to do implicit NULL check in address: 2216 // 2217 // decode_not_null narrow_oop_reg, base_reg 2218 // [base_reg + offset] 2219 // NullCheck base_reg 2220 // 2221 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2222 // to keep the information to which NULL check the new DecodeN node 2223 // corresponds to use it as value in implicit_null_check(). 2224 // 2225 new_in1->set_req(0, n->in(0)); 2226 } 2227 2228 n->subsume_by(new_in1); 2229 if (in1->outcnt() == 0) { 2230 in1->disconnect_inputs(NULL); 2231 } 2232 } 2233 break; 2234 2235 case Op_CmpP: 2236 // Do this transformation here to preserve CmpPNode::sub() and 2237 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2238 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2239 Node* in1 = n->in(1); 2240 Node* in2 = n->in(2); 2241 if (!in1->is_DecodeN()) { 2242 in2 = in1; 2243 in1 = n->in(2); 2244 } 2245 assert(in1->is_DecodeN(), "sanity"); 2246 2247 Compile* C = Compile::current(); 2248 Node* new_in2 = NULL; 2249 if (in2->is_DecodeN()) { 2250 new_in2 = in2->in(1); 2251 } else if (in2->Opcode() == Op_ConP) { 2252 const Type* t = in2->bottom_type(); 2253 if (t == TypePtr::NULL_PTR) { 2254 // Don't convert CmpP null check into CmpN if compressed 2255 // oops implicit null check is not generated. 2256 // This will allow to generate normal oop implicit null check. 2257 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2258 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2259 // 2260 // This transformation together with CastPP transformation above 2261 // will generated code for implicit NULL checks for compressed oops. 2262 // 2263 // The original code after Optimize() 2264 // 2265 // LoadN memory, narrow_oop_reg 2266 // decode narrow_oop_reg, base_reg 2267 // CmpP base_reg, NULL 2268 // CastPP base_reg // NotNull 2269 // Load [base_reg + offset], val_reg 2270 // 2271 // after these transformations will be 2272 // 2273 // LoadN memory, narrow_oop_reg 2274 // CmpN narrow_oop_reg, NULL 2275 // decode_not_null narrow_oop_reg, base_reg 2276 // Load [base_reg + offset], val_reg 2277 // 2278 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2279 // since narrow oops can be used in debug info now (see the code in 2280 // final_graph_reshaping_walk()). 2281 // 2282 // At the end the code will be matched to 2283 // on x86: 2284 // 2285 // Load_narrow_oop memory, narrow_oop_reg 2286 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2287 // NullCheck narrow_oop_reg 2288 // 2289 // and on sparc: 2290 // 2291 // Load_narrow_oop memory, narrow_oop_reg 2292 // decode_not_null narrow_oop_reg, base_reg 2293 // Load [base_reg + offset], val_reg 2294 // NullCheck base_reg 2295 // 2296 } else if (t->isa_oopptr()) { 2297 new_in2 = ConNode::make(C, t->make_narrowoop()); 2298 } 2299 } 2300 if (new_in2 != NULL) { 2301 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2302 n->subsume_by( cmpN ); 2303 if (in1->outcnt() == 0) { 2304 in1->disconnect_inputs(NULL); 2305 } 2306 if (in2->outcnt() == 0) { 2307 in2->disconnect_inputs(NULL); 2308 } 2309 } 2310 } 2311 break; 2312 2313 case Op_DecodeN: 2314 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2315 // DecodeN could be pinned when it can't be fold into 2316 // an address expression, see the code for Op_CastPP above. 2317 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2318 break; 2319 2320 case Op_EncodeP: { 2321 Node* in1 = n->in(1); 2322 if (in1->is_DecodeN()) { 2323 n->subsume_by(in1->in(1)); 2324 } else if (in1->Opcode() == Op_ConP) { 2325 Compile* C = Compile::current(); 2326 const Type* t = in1->bottom_type(); 2327 if (t == TypePtr::NULL_PTR) { 2328 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2329 } else if (t->isa_oopptr()) { 2330 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2331 } 2332 } 2333 if (in1->outcnt() == 0) { 2334 in1->disconnect_inputs(NULL); 2335 } 2336 break; 2337 } 2338 2339 case Op_Proj: { 2340 if (OptimizeStringConcat) { 2341 ProjNode* p = n->as_Proj(); 2342 if (p->_is_io_use) { 2343 // Separate projections were used for the exception path which 2344 // are normally removed by a late inline. If it wasn't inlined 2345 // then they will hang around and should just be replaced with 2346 // the original one. 2347 Node* proj = NULL; 2348 // Replace with just one 2349 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2350 Node *use = i.get(); 2351 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2352 proj = use; 2353 break; 2354 } 2355 } 2356 assert(p != NULL, "must be found"); 2357 p->subsume_by(proj); 2358 } 2359 } 2360 break; 2361 } 2362 2363 case Op_Phi: 2364 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2365 // The EncodeP optimization may create Phi with the same edges 2366 // for all paths. It is not handled well by Register Allocator. 2367 Node* unique_in = n->in(1); 2368 assert(unique_in != NULL, ""); 2369 uint cnt = n->req(); 2370 for (uint i = 2; i < cnt; i++) { 2371 Node* m = n->in(i); 2372 assert(m != NULL, ""); 2373 if (unique_in != m) 2374 unique_in = NULL; 2375 } 2376 if (unique_in != NULL) { 2377 n->subsume_by(unique_in); 2378 } 2379 } 2380 break; 2381 2382 #endif 2383 2384 case Op_ModI: 2385 if (UseDivMod) { 2386 // Check if a%b and a/b both exist 2387 Node* d = n->find_similar(Op_DivI); 2388 if (d) { 2389 // Replace them with a fused divmod if supported 2390 Compile* C = Compile::current(); 2391 if (Matcher::has_match_rule(Op_DivModI)) { 2392 DivModINode* divmod = DivModINode::make(C, n); 2393 d->subsume_by(divmod->div_proj()); 2394 n->subsume_by(divmod->mod_proj()); 2395 } else { 2396 // replace a%b with a-((a/b)*b) 2397 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2398 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2399 n->subsume_by( sub ); 2400 } 2401 } 2402 } 2403 break; 2404 2405 case Op_ModL: 2406 if (UseDivMod) { 2407 // Check if a%b and a/b both exist 2408 Node* d = n->find_similar(Op_DivL); 2409 if (d) { 2410 // Replace them with a fused divmod if supported 2411 Compile* C = Compile::current(); 2412 if (Matcher::has_match_rule(Op_DivModL)) { 2413 DivModLNode* divmod = DivModLNode::make(C, n); 2414 d->subsume_by(divmod->div_proj()); 2415 n->subsume_by(divmod->mod_proj()); 2416 } else { 2417 // replace a%b with a-((a/b)*b) 2418 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2419 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2420 n->subsume_by( sub ); 2421 } 2422 } 2423 } 2424 break; 2425 2426 case Op_Load16B: 2427 case Op_Load8B: 2428 case Op_Load4B: 2429 case Op_Load8S: 2430 case Op_Load4S: 2431 case Op_Load2S: 2432 case Op_Load8C: 2433 case Op_Load4C: 2434 case Op_Load2C: 2435 case Op_Load4I: 2436 case Op_Load2I: 2437 case Op_Load2L: 2438 case Op_Load4F: 2439 case Op_Load2F: 2440 case Op_Load2D: 2441 case Op_Store16B: 2442 case Op_Store8B: 2443 case Op_Store4B: 2444 case Op_Store8C: 2445 case Op_Store4C: 2446 case Op_Store2C: 2447 case Op_Store4I: 2448 case Op_Store2I: 2449 case Op_Store2L: 2450 case Op_Store4F: 2451 case Op_Store2F: 2452 case Op_Store2D: 2453 break; 2454 2455 case Op_PackB: 2456 case Op_PackS: 2457 case Op_PackC: 2458 case Op_PackI: 2459 case Op_PackF: 2460 case Op_PackL: 2461 case Op_PackD: 2462 if (n->req()-1 > 2) { 2463 // Replace many operand PackNodes with a binary tree for matching 2464 PackNode* p = (PackNode*) n; 2465 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2466 n->subsume_by(btp); 2467 } 2468 break; 2469 case Op_Loop: 2470 case Op_CountedLoop: 2471 if (n->as_Loop()->is_inner_loop()) { 2472 frc.inc_inner_loop_count(); 2473 } 2474 break; 2475 default: 2476 assert( !n->is_Call(), "" ); 2477 assert( !n->is_Mem(), "" ); 2478 break; 2479 } 2480 2481 // Collect CFG split points 2482 if (n->is_MultiBranch()) 2483 frc._tests.push(n); 2484 } 2485 2486 //------------------------------final_graph_reshaping_walk--------------------- 2487 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2488 // requires that the walk visits a node's inputs before visiting the node. 2489 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2490 ResourceArea *area = Thread::current()->resource_area(); 2491 Unique_Node_List sfpt(area); 2492 2493 frc._visited.set(root->_idx); // first, mark node as visited 2494 uint cnt = root->req(); 2495 Node *n = root; 2496 uint i = 0; 2497 while (true) { 2498 if (i < cnt) { 2499 // Place all non-visited non-null inputs onto stack 2500 Node* m = n->in(i); 2501 ++i; 2502 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2503 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2504 sfpt.push(m); 2505 cnt = m->req(); 2506 nstack.push(n, i); // put on stack parent and next input's index 2507 n = m; 2508 i = 0; 2509 } 2510 } else { 2511 // Now do post-visit work 2512 final_graph_reshaping_impl( n, frc ); 2513 if (nstack.is_empty()) 2514 break; // finished 2515 n = nstack.node(); // Get node from stack 2516 cnt = n->req(); 2517 i = nstack.index(); 2518 nstack.pop(); // Shift to the next node on stack 2519 } 2520 } 2521 2522 // Skip next transformation if compressed oops are not used. 2523 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) 2524 return; 2525 2526 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2527 // It could be done for an uncommon traps or any safepoints/calls 2528 // if the DecodeN node is referenced only in a debug info. 2529 while (sfpt.size() > 0) { 2530 n = sfpt.pop(); 2531 JVMState *jvms = n->as_SafePoint()->jvms(); 2532 assert(jvms != NULL, "sanity"); 2533 int start = jvms->debug_start(); 2534 int end = n->req(); 2535 bool is_uncommon = (n->is_CallStaticJava() && 2536 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2537 for (int j = start; j < end; j++) { 2538 Node* in = n->in(j); 2539 if (in->is_DecodeN()) { 2540 bool safe_to_skip = true; 2541 if (!is_uncommon ) { 2542 // Is it safe to skip? 2543 for (uint i = 0; i < in->outcnt(); i++) { 2544 Node* u = in->raw_out(i); 2545 if (!u->is_SafePoint() || 2546 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2547 safe_to_skip = false; 2548 } 2549 } 2550 } 2551 if (safe_to_skip) { 2552 n->set_req(j, in->in(1)); 2553 } 2554 if (in->outcnt() == 0) { 2555 in->disconnect_inputs(NULL); 2556 } 2557 } 2558 } 2559 } 2560 } 2561 2562 //------------------------------final_graph_reshaping-------------------------- 2563 // Final Graph Reshaping. 2564 // 2565 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2566 // and not commoned up and forced early. Must come after regular 2567 // optimizations to avoid GVN undoing the cloning. Clone constant 2568 // inputs to Loop Phis; these will be split by the allocator anyways. 2569 // Remove Opaque nodes. 2570 // (2) Move last-uses by commutative operations to the left input to encourage 2571 // Intel update-in-place two-address operations and better register usage 2572 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2573 // calls canonicalizing them back. 2574 // (3) Count the number of double-precision FP ops, single-precision FP ops 2575 // and call sites. On Intel, we can get correct rounding either by 2576 // forcing singles to memory (requires extra stores and loads after each 2577 // FP bytecode) or we can set a rounding mode bit (requires setting and 2578 // clearing the mode bit around call sites). The mode bit is only used 2579 // if the relative frequency of single FP ops to calls is low enough. 2580 // This is a key transform for SPEC mpeg_audio. 2581 // (4) Detect infinite loops; blobs of code reachable from above but not 2582 // below. Several of the Code_Gen algorithms fail on such code shapes, 2583 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2584 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2585 // Detection is by looking for IfNodes where only 1 projection is 2586 // reachable from below or CatchNodes missing some targets. 2587 // (5) Assert for insane oop offsets in debug mode. 2588 2589 bool Compile::final_graph_reshaping() { 2590 // an infinite loop may have been eliminated by the optimizer, 2591 // in which case the graph will be empty. 2592 if (root()->req() == 1) { 2593 record_method_not_compilable("trivial infinite loop"); 2594 return true; 2595 } 2596 2597 Final_Reshape_Counts frc; 2598 2599 // Visit everybody reachable! 2600 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2601 Node_Stack nstack(unique() >> 1); 2602 final_graph_reshaping_walk(nstack, root(), frc); 2603 2604 // Check for unreachable (from below) code (i.e., infinite loops). 2605 for( uint i = 0; i < frc._tests.size(); i++ ) { 2606 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2607 // Get number of CFG targets. 2608 // Note that PCTables include exception targets after calls. 2609 uint required_outcnt = n->required_outcnt(); 2610 if (n->outcnt() != required_outcnt) { 2611 // Check for a few special cases. Rethrow Nodes never take the 2612 // 'fall-thru' path, so expected kids is 1 less. 2613 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2614 if (n->in(0)->in(0)->is_Call()) { 2615 CallNode *call = n->in(0)->in(0)->as_Call(); 2616 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2617 required_outcnt--; // Rethrow always has 1 less kid 2618 } else if (call->req() > TypeFunc::Parms && 2619 call->is_CallDynamicJava()) { 2620 // Check for null receiver. In such case, the optimizer has 2621 // detected that the virtual call will always result in a null 2622 // pointer exception. The fall-through projection of this CatchNode 2623 // will not be populated. 2624 Node *arg0 = call->in(TypeFunc::Parms); 2625 if (arg0->is_Type() && 2626 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2627 required_outcnt--; 2628 } 2629 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2630 call->req() > TypeFunc::Parms+1 && 2631 call->is_CallStaticJava()) { 2632 // Check for negative array length. In such case, the optimizer has 2633 // detected that the allocation attempt will always result in an 2634 // exception. There is no fall-through projection of this CatchNode . 2635 Node *arg1 = call->in(TypeFunc::Parms+1); 2636 if (arg1->is_Type() && 2637 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2638 required_outcnt--; 2639 } 2640 } 2641 } 2642 } 2643 // Recheck with a better notion of 'required_outcnt' 2644 if (n->outcnt() != required_outcnt) { 2645 record_method_not_compilable("malformed control flow"); 2646 return true; // Not all targets reachable! 2647 } 2648 } 2649 // Check that I actually visited all kids. Unreached kids 2650 // must be infinite loops. 2651 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2652 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2653 record_method_not_compilable("infinite loop"); 2654 return true; // Found unvisited kid; must be unreach 2655 } 2656 } 2657 2658 // If original bytecodes contained a mixture of floats and doubles 2659 // check if the optimizer has made it homogenous, item (3). 2660 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2661 frc.get_float_count() > 32 && 2662 frc.get_double_count() == 0 && 2663 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2664 set_24_bit_selection_and_mode( false, true ); 2665 } 2666 2667 set_java_calls(frc.get_java_call_count()); 2668 set_inner_loops(frc.get_inner_loop_count()); 2669 2670 // No infinite loops, no reason to bail out. 2671 return false; 2672 } 2673 2674 //-----------------------------too_many_traps---------------------------------- 2675 // Report if there are too many traps at the current method and bci. 2676 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2677 bool Compile::too_many_traps(ciMethod* method, 2678 int bci, 2679 Deoptimization::DeoptReason reason) { 2680 ciMethodData* md = method->method_data(); 2681 if (md->is_empty()) { 2682 // Assume the trap has not occurred, or that it occurred only 2683 // because of a transient condition during start-up in the interpreter. 2684 return false; 2685 } 2686 if (md->has_trap_at(bci, reason) != 0) { 2687 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2688 // Also, if there are multiple reasons, or if there is no per-BCI record, 2689 // assume the worst. 2690 if (log()) 2691 log()->elem("observe trap='%s' count='%d'", 2692 Deoptimization::trap_reason_name(reason), 2693 md->trap_count(reason)); 2694 return true; 2695 } else { 2696 // Ignore method/bci and see if there have been too many globally. 2697 return too_many_traps(reason, md); 2698 } 2699 } 2700 2701 // Less-accurate variant which does not require a method and bci. 2702 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2703 ciMethodData* logmd) { 2704 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2705 // Too many traps globally. 2706 // Note that we use cumulative trap_count, not just md->trap_count. 2707 if (log()) { 2708 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2709 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2710 Deoptimization::trap_reason_name(reason), 2711 mcount, trap_count(reason)); 2712 } 2713 return true; 2714 } else { 2715 // The coast is clear. 2716 return false; 2717 } 2718 } 2719 2720 //--------------------------too_many_recompiles-------------------------------- 2721 // Report if there are too many recompiles at the current method and bci. 2722 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2723 // Is not eager to return true, since this will cause the compiler to use 2724 // Action_none for a trap point, to avoid too many recompilations. 2725 bool Compile::too_many_recompiles(ciMethod* method, 2726 int bci, 2727 Deoptimization::DeoptReason reason) { 2728 ciMethodData* md = method->method_data(); 2729 if (md->is_empty()) { 2730 // Assume the trap has not occurred, or that it occurred only 2731 // because of a transient condition during start-up in the interpreter. 2732 return false; 2733 } 2734 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2735 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2736 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2737 Deoptimization::DeoptReason per_bc_reason 2738 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2739 if ((per_bc_reason == Deoptimization::Reason_none 2740 || md->has_trap_at(bci, reason) != 0) 2741 // The trap frequency measure we care about is the recompile count: 2742 && md->trap_recompiled_at(bci) 2743 && md->overflow_recompile_count() >= bc_cutoff) { 2744 // Do not emit a trap here if it has already caused recompilations. 2745 // Also, if there are multiple reasons, or if there is no per-BCI record, 2746 // assume the worst. 2747 if (log()) 2748 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2749 Deoptimization::trap_reason_name(reason), 2750 md->trap_count(reason), 2751 md->overflow_recompile_count()); 2752 return true; 2753 } else if (trap_count(reason) != 0 2754 && decompile_count() >= m_cutoff) { 2755 // Too many recompiles globally, and we have seen this sort of trap. 2756 // Use cumulative decompile_count, not just md->decompile_count. 2757 if (log()) 2758 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2759 Deoptimization::trap_reason_name(reason), 2760 md->trap_count(reason), trap_count(reason), 2761 md->decompile_count(), decompile_count()); 2762 return true; 2763 } else { 2764 // The coast is clear. 2765 return false; 2766 } 2767 } 2768 2769 2770 #ifndef PRODUCT 2771 //------------------------------verify_graph_edges--------------------------- 2772 // Walk the Graph and verify that there is a one-to-one correspondence 2773 // between Use-Def edges and Def-Use edges in the graph. 2774 void Compile::verify_graph_edges(bool no_dead_code) { 2775 if (VerifyGraphEdges) { 2776 ResourceArea *area = Thread::current()->resource_area(); 2777 Unique_Node_List visited(area); 2778 // Call recursive graph walk to check edges 2779 _root->verify_edges(visited); 2780 if (no_dead_code) { 2781 // Now make sure that no visited node is used by an unvisited node. 2782 bool dead_nodes = 0; 2783 Unique_Node_List checked(area); 2784 while (visited.size() > 0) { 2785 Node* n = visited.pop(); 2786 checked.push(n); 2787 for (uint i = 0; i < n->outcnt(); i++) { 2788 Node* use = n->raw_out(i); 2789 if (checked.member(use)) continue; // already checked 2790 if (visited.member(use)) continue; // already in the graph 2791 if (use->is_Con()) continue; // a dead ConNode is OK 2792 // At this point, we have found a dead node which is DU-reachable. 2793 if (dead_nodes++ == 0) 2794 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2795 use->dump(2); 2796 tty->print_cr("---"); 2797 checked.push(use); // No repeats; pretend it is now checked. 2798 } 2799 } 2800 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2801 } 2802 } 2803 } 2804 #endif 2805 2806 // The Compile object keeps track of failure reasons separately from the ciEnv. 2807 // This is required because there is not quite a 1-1 relation between the 2808 // ciEnv and its compilation task and the Compile object. Note that one 2809 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2810 // to backtrack and retry without subsuming loads. Other than this backtracking 2811 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2812 // by the logic in C2Compiler. 2813 void Compile::record_failure(const char* reason) { 2814 if (log() != NULL) { 2815 log()->elem("failure reason='%s' phase='compile'", reason); 2816 } 2817 if (_failure_reason == NULL) { 2818 // Record the first failure reason. 2819 _failure_reason = reason; 2820 } 2821 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2822 C->print_method(_failure_reason); 2823 } 2824 _root = NULL; // flush the graph, too 2825 } 2826 2827 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2828 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2829 { 2830 if (dolog) { 2831 C = Compile::current(); 2832 _log = C->log(); 2833 } else { 2834 C = NULL; 2835 _log = NULL; 2836 } 2837 if (_log != NULL) { 2838 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2839 _log->stamp(); 2840 _log->end_head(); 2841 } 2842 } 2843 2844 Compile::TracePhase::~TracePhase() { 2845 if (_log != NULL) { 2846 _log->done("phase nodes='%d'", C->unique()); 2847 } 2848 }