compile.cpp revision 1915:2f644f85485d
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 "precompiled.hpp" 26#include "asm/assembler.hpp" 27#include "classfile/systemDictionary.hpp" 28#include "code/exceptionHandlerTable.hpp" 29#include "code/nmethod.hpp" 30#include "compiler/compileLog.hpp" 31#include "compiler/oopMap.hpp" 32#include "opto/addnode.hpp" 33#include "opto/block.hpp" 34#include "opto/c2compiler.hpp" 35#include "opto/callGenerator.hpp" 36#include "opto/callnode.hpp" 37#include "opto/cfgnode.hpp" 38#include "opto/chaitin.hpp" 39#include "opto/compile.hpp" 40#include "opto/connode.hpp" 41#include "opto/divnode.hpp" 42#include "opto/escape.hpp" 43#include "opto/idealGraphPrinter.hpp" 44#include "opto/loopnode.hpp" 45#include "opto/machnode.hpp" 46#include "opto/macro.hpp" 47#include "opto/matcher.hpp" 48#include "opto/memnode.hpp" 49#include "opto/mulnode.hpp" 50#include "opto/node.hpp" 51#include "opto/opcodes.hpp" 52#include "opto/output.hpp" 53#include "opto/parse.hpp" 54#include "opto/phaseX.hpp" 55#include "opto/rootnode.hpp" 56#include "opto/runtime.hpp" 57#include "opto/stringopts.hpp" 58#include "opto/type.hpp" 59#include "opto/vectornode.hpp" 60#include "runtime/arguments.hpp" 61#include "runtime/signature.hpp" 62#include "runtime/stubRoutines.hpp" 63#include "runtime/timer.hpp" 64#include "utilities/copy.hpp" 65#ifdef TARGET_ARCH_MODEL_x86_32 66# include "adfiles/ad_x86_32.hpp" 67#endif 68#ifdef TARGET_ARCH_MODEL_x86_64 69# include "adfiles/ad_x86_64.hpp" 70#endif 71#ifdef TARGET_ARCH_MODEL_sparc 72# include "adfiles/ad_sparc.hpp" 73#endif 74#ifdef TARGET_ARCH_MODEL_zero 75# include "adfiles/ad_zero.hpp" 76#endif 77 78 79// -------------------- Compile::mach_constant_base_node ----------------------- 80// Constant table base node singleton. 81MachConstantBaseNode* Compile::mach_constant_base_node() { 82 if (_mach_constant_base_node == NULL) { 83 _mach_constant_base_node = new (C) MachConstantBaseNode(); 84 _mach_constant_base_node->add_req(C->root()); 85 } 86 return _mach_constant_base_node; 87} 88 89 90/// Support for intrinsics. 91 92// Return the index at which m must be inserted (or already exists). 93// The sort order is by the address of the ciMethod, with is_virtual as minor key. 94int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 95#ifdef ASSERT 96 for (int i = 1; i < _intrinsics->length(); i++) { 97 CallGenerator* cg1 = _intrinsics->at(i-1); 98 CallGenerator* cg2 = _intrinsics->at(i); 99 assert(cg1->method() != cg2->method() 100 ? cg1->method() < cg2->method() 101 : cg1->is_virtual() < cg2->is_virtual(), 102 "compiler intrinsics list must stay sorted"); 103 } 104#endif 105 // Binary search sorted list, in decreasing intervals [lo, hi]. 106 int lo = 0, hi = _intrinsics->length()-1; 107 while (lo <= hi) { 108 int mid = (uint)(hi + lo) / 2; 109 ciMethod* mid_m = _intrinsics->at(mid)->method(); 110 if (m < mid_m) { 111 hi = mid-1; 112 } else if (m > mid_m) { 113 lo = mid+1; 114 } else { 115 // look at minor sort key 116 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 117 if (is_virtual < mid_virt) { 118 hi = mid-1; 119 } else if (is_virtual > mid_virt) { 120 lo = mid+1; 121 } else { 122 return mid; // exact match 123 } 124 } 125 } 126 return lo; // inexact match 127} 128 129void Compile::register_intrinsic(CallGenerator* cg) { 130 if (_intrinsics == NULL) { 131 _intrinsics = new GrowableArray<CallGenerator*>(60); 132 } 133 // This code is stolen from ciObjectFactory::insert. 134 // Really, GrowableArray should have methods for 135 // insert_at, remove_at, and binary_search. 136 int len = _intrinsics->length(); 137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 138 if (index == len) { 139 _intrinsics->append(cg); 140 } else { 141#ifdef ASSERT 142 CallGenerator* oldcg = _intrinsics->at(index); 143 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 144#endif 145 _intrinsics->append(_intrinsics->at(len-1)); 146 int pos; 147 for (pos = len-2; pos >= index; pos--) { 148 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 149 } 150 _intrinsics->at_put(index, cg); 151 } 152 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 153} 154 155CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 156 assert(m->is_loaded(), "don't try this on unloaded methods"); 157 if (_intrinsics != NULL) { 158 int index = intrinsic_insertion_index(m, is_virtual); 159 if (index < _intrinsics->length() 160 && _intrinsics->at(index)->method() == m 161 && _intrinsics->at(index)->is_virtual() == is_virtual) { 162 return _intrinsics->at(index); 163 } 164 } 165 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 166 if (m->intrinsic_id() != vmIntrinsics::_none && 167 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 168 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 169 if (cg != NULL) { 170 // Save it for next time: 171 register_intrinsic(cg); 172 return cg; 173 } else { 174 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 175 } 176 } 177 return NULL; 178} 179 180// Compile:: register_library_intrinsics and make_vm_intrinsic are defined 181// in library_call.cpp. 182 183 184#ifndef PRODUCT 185// statistics gathering... 186 187juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 188jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 189 190bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 191 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 192 int oflags = _intrinsic_hist_flags[id]; 193 assert(flags != 0, "what happened?"); 194 if (is_virtual) { 195 flags |= _intrinsic_virtual; 196 } 197 bool changed = (flags != oflags); 198 if ((flags & _intrinsic_worked) != 0) { 199 juint count = (_intrinsic_hist_count[id] += 1); 200 if (count == 1) { 201 changed = true; // first time 202 } 203 // increment the overall count also: 204 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 205 } 206 if (changed) { 207 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 208 // Something changed about the intrinsic's virtuality. 209 if ((flags & _intrinsic_virtual) != 0) { 210 // This is the first use of this intrinsic as a virtual call. 211 if (oflags != 0) { 212 // We already saw it as a non-virtual, so note both cases. 213 flags |= _intrinsic_both; 214 } 215 } else if ((oflags & _intrinsic_both) == 0) { 216 // This is the first use of this intrinsic as a non-virtual 217 flags |= _intrinsic_both; 218 } 219 } 220 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 221 } 222 // update the overall flags also: 223 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 224 return changed; 225} 226 227static char* format_flags(int flags, char* buf) { 228 buf[0] = 0; 229 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 230 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 231 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 232 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 233 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 234 if (buf[0] == 0) strcat(buf, ","); 235 assert(buf[0] == ',', "must be"); 236 return &buf[1]; 237} 238 239void Compile::print_intrinsic_statistics() { 240 char flagsbuf[100]; 241 ttyLocker ttyl; 242 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 243 tty->print_cr("Compiler intrinsic usage:"); 244 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 245 if (total == 0) total = 1; // avoid div0 in case of no successes 246 #define PRINT_STAT_LINE(name, c, f) \ 247 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 248 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 249 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 250 int flags = _intrinsic_hist_flags[id]; 251 juint count = _intrinsic_hist_count[id]; 252 if ((flags | count) != 0) { 253 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 254 } 255 } 256 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 257 if (xtty != NULL) xtty->tail("statistics"); 258} 259 260void Compile::print_statistics() { 261 { ttyLocker ttyl; 262 if (xtty != NULL) xtty->head("statistics type='opto'"); 263 Parse::print_statistics(); 264 PhaseCCP::print_statistics(); 265 PhaseRegAlloc::print_statistics(); 266 Scheduling::print_statistics(); 267 PhasePeephole::print_statistics(); 268 PhaseIdealLoop::print_statistics(); 269 if (xtty != NULL) xtty->tail("statistics"); 270 } 271 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 272 // put this under its own <statistics> element. 273 print_intrinsic_statistics(); 274 } 275} 276#endif //PRODUCT 277 278// Support for bundling info 279Bundle* Compile::node_bundling(const Node *n) { 280 assert(valid_bundle_info(n), "oob"); 281 return &_node_bundling_base[n->_idx]; 282} 283 284bool Compile::valid_bundle_info(const Node *n) { 285 return (_node_bundling_limit > n->_idx); 286} 287 288 289void Compile::gvn_replace_by(Node* n, Node* nn) { 290 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 291 Node* use = n->last_out(i); 292 bool is_in_table = initial_gvn()->hash_delete(use); 293 uint uses_found = 0; 294 for (uint j = 0; j < use->len(); j++) { 295 if (use->in(j) == n) { 296 if (j < use->req()) 297 use->set_req(j, nn); 298 else 299 use->set_prec(j, nn); 300 uses_found++; 301 } 302 } 303 if (is_in_table) { 304 // reinsert into table 305 initial_gvn()->hash_find_insert(use); 306 } 307 record_for_igvn(use); 308 i -= uses_found; // we deleted 1 or more copies of this edge 309 } 310} 311 312 313 314 315// Identify all nodes that are reachable from below, useful. 316// Use breadth-first pass that records state in a Unique_Node_List, 317// recursive traversal is slower. 318void Compile::identify_useful_nodes(Unique_Node_List &useful) { 319 int estimated_worklist_size = unique(); 320 useful.map( estimated_worklist_size, NULL ); // preallocate space 321 322 // Initialize worklist 323 if (root() != NULL) { useful.push(root()); } 324 // If 'top' is cached, declare it useful to preserve cached node 325 if( cached_top_node() ) { useful.push(cached_top_node()); } 326 327 // Push all useful nodes onto the list, breadthfirst 328 for( uint next = 0; next < useful.size(); ++next ) { 329 assert( next < unique(), "Unique useful nodes < total nodes"); 330 Node *n = useful.at(next); 331 uint max = n->len(); 332 for( uint i = 0; i < max; ++i ) { 333 Node *m = n->in(i); 334 if( m == NULL ) continue; 335 useful.push(m); 336 } 337 } 338} 339 340// Disconnect all useless nodes by disconnecting those at the boundary. 341void Compile::remove_useless_nodes(Unique_Node_List &useful) { 342 uint next = 0; 343 while( next < useful.size() ) { 344 Node *n = useful.at(next++); 345 // Use raw traversal of out edges since this code removes out edges 346 int max = n->outcnt(); 347 for (int j = 0; j < max; ++j ) { 348 Node* child = n->raw_out(j); 349 if( ! useful.member(child) ) { 350 assert( !child->is_top() || child != top(), 351 "If top is cached in Compile object it is in useful list"); 352 // Only need to remove this out-edge to the useless node 353 n->raw_del_out(j); 354 --j; 355 --max; 356 } 357 } 358 if (n->outcnt() == 1 && n->has_special_unique_user()) { 359 record_for_igvn( n->unique_out() ); 360 } 361 } 362 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 363} 364 365//------------------------------frame_size_in_words----------------------------- 366// frame_slots in units of words 367int Compile::frame_size_in_words() const { 368 // shift is 0 in LP32 and 1 in LP64 369 const int shift = (LogBytesPerWord - LogBytesPerInt); 370 int words = _frame_slots >> shift; 371 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 372 return words; 373} 374 375// ============================================================================ 376//------------------------------CompileWrapper--------------------------------- 377class CompileWrapper : public StackObj { 378 Compile *const _compile; 379 public: 380 CompileWrapper(Compile* compile); 381 382 ~CompileWrapper(); 383}; 384 385CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 386 // the Compile* pointer is stored in the current ciEnv: 387 ciEnv* env = compile->env(); 388 assert(env == ciEnv::current(), "must already be a ciEnv active"); 389 assert(env->compiler_data() == NULL, "compile already active?"); 390 env->set_compiler_data(compile); 391 assert(compile == Compile::current(), "sanity"); 392 393 compile->set_type_dict(NULL); 394 compile->set_type_hwm(NULL); 395 compile->set_type_last_size(0); 396 compile->set_last_tf(NULL, NULL); 397 compile->set_indexSet_arena(NULL); 398 compile->set_indexSet_free_block_list(NULL); 399 compile->init_type_arena(); 400 Type::Initialize(compile); 401 _compile->set_scratch_buffer_blob(NULL); 402 _compile->begin_method(); 403} 404CompileWrapper::~CompileWrapper() { 405 _compile->end_method(); 406 if (_compile->scratch_buffer_blob() != NULL) 407 BufferBlob::free(_compile->scratch_buffer_blob()); 408 _compile->env()->set_compiler_data(NULL); 409} 410 411 412//----------------------------print_compile_messages--------------------------- 413void Compile::print_compile_messages() { 414#ifndef PRODUCT 415 // Check if recompiling 416 if (_subsume_loads == false && PrintOpto) { 417 // Recompiling without allowing machine instructions to subsume loads 418 tty->print_cr("*********************************************************"); 419 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 420 tty->print_cr("*********************************************************"); 421 } 422 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 423 // Recompiling without escape analysis 424 tty->print_cr("*********************************************************"); 425 tty->print_cr("** Bailout: Recompile without escape analysis **"); 426 tty->print_cr("*********************************************************"); 427 } 428 if (env()->break_at_compile()) { 429 // Open the debugger when compiling this method. 430 tty->print("### Breaking when compiling: "); 431 method()->print_short_name(); 432 tty->cr(); 433 BREAKPOINT; 434 } 435 436 if( PrintOpto ) { 437 if (is_osr_compilation()) { 438 tty->print("[OSR]%3d", _compile_id); 439 } else { 440 tty->print("%3d", _compile_id); 441 } 442 } 443#endif 444} 445 446 447void Compile::init_scratch_buffer_blob(int const_size) { 448 if (scratch_buffer_blob() != NULL) return; 449 450 // Construct a temporary CodeBuffer to have it construct a BufferBlob 451 // Cache this BufferBlob for this compile. 452 ResourceMark rm; 453 _scratch_const_size = const_size; 454 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 455 BufferBlob* blob = BufferBlob::create("Compile::scratch_buffer", size); 456 // Record the buffer blob for next time. 457 set_scratch_buffer_blob(blob); 458 // Have we run out of code space? 459 if (scratch_buffer_blob() == NULL) { 460 // Let CompilerBroker disable further compilations. 461 record_failure("Not enough space for scratch buffer in CodeCache"); 462 return; 463 } 464 465 // Initialize the relocation buffers 466 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 467 set_scratch_locs_memory(locs_buf); 468} 469 470 471void Compile::clear_scratch_buffer_blob() { 472 assert(scratch_buffer_blob(), "no BufferBlob set"); 473 set_scratch_buffer_blob(NULL); 474 set_scratch_locs_memory(NULL); 475} 476 477 478//-----------------------scratch_emit_size------------------------------------- 479// Helper function that computes size by emitting code 480uint Compile::scratch_emit_size(const Node* n) { 481 // Start scratch_emit_size section. 482 set_in_scratch_emit_size(true); 483 484 // Emit into a trash buffer and count bytes emitted. 485 // This is a pretty expensive way to compute a size, 486 // but it works well enough if seldom used. 487 // All common fixed-size instructions are given a size 488 // method by the AD file. 489 // Note that the scratch buffer blob and locs memory are 490 // allocated at the beginning of the compile task, and 491 // may be shared by several calls to scratch_emit_size. 492 // The allocation of the scratch buffer blob is particularly 493 // expensive, since it has to grab the code cache lock. 494 BufferBlob* blob = this->scratch_buffer_blob(); 495 assert(blob != NULL, "Initialize BufferBlob at start"); 496 assert(blob->size() > MAX_inst_size, "sanity"); 497 relocInfo* locs_buf = scratch_locs_memory(); 498 address blob_begin = blob->content_begin(); 499 address blob_end = (address)locs_buf; 500 assert(blob->content_contains(blob_end), "sanity"); 501 CodeBuffer buf(blob_begin, blob_end - blob_begin); 502 buf.initialize_consts_size(_scratch_const_size); 503 buf.initialize_stubs_size(MAX_stubs_size); 504 assert(locs_buf != NULL, "sanity"); 505 int lsize = MAX_locs_size / 3; 506 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 507 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 508 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 509 510 // Do the emission. 511 n->emit(buf, this->regalloc()); 512 513 // End scratch_emit_size section. 514 set_in_scratch_emit_size(false); 515 516 return buf.insts_size(); 517} 518 519 520// ============================================================================ 521//------------------------------Compile standard------------------------------- 522debug_only( int Compile::_debug_idx = 100000; ) 523 524// Compile a method. entry_bci is -1 for normal compilations and indicates 525// the continuation bci for on stack replacement. 526 527 528Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 529 : Phase(Compiler), 530 _env(ci_env), 531 _log(ci_env->log()), 532 _compile_id(ci_env->compile_id()), 533 _save_argument_registers(false), 534 _stub_name(NULL), 535 _stub_function(NULL), 536 _stub_entry_point(NULL), 537 _method(target), 538 _entry_bci(osr_bci), 539 _initial_gvn(NULL), 540 _for_igvn(NULL), 541 _warm_calls(NULL), 542 _subsume_loads(subsume_loads), 543 _do_escape_analysis(do_escape_analysis), 544 _failure_reason(NULL), 545 _code_buffer("Compile::Fill_buffer"), 546 _orig_pc_slot(0), 547 _orig_pc_slot_offset_in_bytes(0), 548 _has_method_handle_invokes(false), 549 _mach_constant_base_node(NULL), 550 _node_bundling_limit(0), 551 _node_bundling_base(NULL), 552 _java_calls(0), 553 _inner_loops(0), 554 _scratch_const_size(-1), 555 _in_scratch_emit_size(false), 556#ifndef PRODUCT 557 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 558 _printer(IdealGraphPrinter::printer()), 559#endif 560 _congraph(NULL) { 561 C = this; 562 563 CompileWrapper cw(this); 564#ifndef PRODUCT 565 if (TimeCompiler2) { 566 tty->print(" "); 567 target->holder()->name()->print(); 568 tty->print("."); 569 target->print_short_name(); 570 tty->print(" "); 571 } 572 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 573 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 574 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 575 if (!print_opto_assembly) { 576 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 577 if (print_assembly && !Disassembler::can_decode()) { 578 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 579 print_opto_assembly = true; 580 } 581 } 582 set_print_assembly(print_opto_assembly); 583 set_parsed_irreducible_loop(false); 584#endif 585 586 if (ProfileTraps) { 587 // Make sure the method being compiled gets its own MDO, 588 // so we can at least track the decompile_count(). 589 method()->ensure_method_data(); 590 } 591 592 Init(::AliasLevel); 593 594 595 print_compile_messages(); 596 597 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 598 _ilt = InlineTree::build_inline_tree_root(); 599 else 600 _ilt = NULL; 601 602 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 603 assert(num_alias_types() >= AliasIdxRaw, ""); 604 605#define MINIMUM_NODE_HASH 1023 606 // Node list that Iterative GVN will start with 607 Unique_Node_List for_igvn(comp_arena()); 608 set_for_igvn(&for_igvn); 609 610 // GVN that will be run immediately on new nodes 611 uint estimated_size = method()->code_size()*4+64; 612 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 613 PhaseGVN gvn(node_arena(), estimated_size); 614 set_initial_gvn(&gvn); 615 616 { // Scope for timing the parser 617 TracePhase t3("parse", &_t_parser, true); 618 619 // Put top into the hash table ASAP. 620 initial_gvn()->transform_no_reclaim(top()); 621 622 // Set up tf(), start(), and find a CallGenerator. 623 CallGenerator* cg; 624 if (is_osr_compilation()) { 625 const TypeTuple *domain = StartOSRNode::osr_domain(); 626 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 627 init_tf(TypeFunc::make(domain, range)); 628 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 629 initial_gvn()->set_type_bottom(s); 630 init_start(s); 631 cg = CallGenerator::for_osr(method(), entry_bci()); 632 } else { 633 // Normal case. 634 init_tf(TypeFunc::make(method())); 635 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 636 initial_gvn()->set_type_bottom(s); 637 init_start(s); 638 float past_uses = method()->interpreter_invocation_count(); 639 float expected_uses = past_uses; 640 cg = CallGenerator::for_inline(method(), expected_uses); 641 } 642 if (failing()) return; 643 if (cg == NULL) { 644 record_method_not_compilable_all_tiers("cannot parse method"); 645 return; 646 } 647 JVMState* jvms = build_start_state(start(), tf()); 648 if ((jvms = cg->generate(jvms)) == NULL) { 649 record_method_not_compilable("method parse failed"); 650 return; 651 } 652 GraphKit kit(jvms); 653 654 if (!kit.stopped()) { 655 // Accept return values, and transfer control we know not where. 656 // This is done by a special, unique ReturnNode bound to root. 657 return_values(kit.jvms()); 658 } 659 660 if (kit.has_exceptions()) { 661 // Any exceptions that escape from this call must be rethrown 662 // to whatever caller is dynamically above us on the stack. 663 // This is done by a special, unique RethrowNode bound to root. 664 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 665 } 666 667 if (!failing() && has_stringbuilder()) { 668 { 669 // remove useless nodes to make the usage analysis simpler 670 ResourceMark rm; 671 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 672 } 673 674 { 675 ResourceMark rm; 676 print_method("Before StringOpts", 3); 677 PhaseStringOpts pso(initial_gvn(), &for_igvn); 678 print_method("After StringOpts", 3); 679 } 680 681 // now inline anything that we skipped the first time around 682 while (_late_inlines.length() > 0) { 683 CallGenerator* cg = _late_inlines.pop(); 684 cg->do_late_inline(); 685 } 686 } 687 assert(_late_inlines.length() == 0, "should have been processed"); 688 689 print_method("Before RemoveUseless", 3); 690 691 // Remove clutter produced by parsing. 692 if (!failing()) { 693 ResourceMark rm; 694 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 695 } 696 } 697 698 // Note: Large methods are capped off in do_one_bytecode(). 699 if (failing()) return; 700 701 // After parsing, node notes are no longer automagic. 702 // They must be propagated by register_new_node_with_optimizer(), 703 // clone(), or the like. 704 set_default_node_notes(NULL); 705 706 for (;;) { 707 int successes = Inline_Warm(); 708 if (failing()) return; 709 if (successes == 0) break; 710 } 711 712 // Drain the list. 713 Finish_Warm(); 714#ifndef PRODUCT 715 if (_printer) { 716 _printer->print_inlining(this); 717 } 718#endif 719 720 if (failing()) return; 721 NOT_PRODUCT( verify_graph_edges(); ) 722 723 // Now optimize 724 Optimize(); 725 if (failing()) return; 726 NOT_PRODUCT( verify_graph_edges(); ) 727 728#ifndef PRODUCT 729 if (PrintIdeal) { 730 ttyLocker ttyl; // keep the following output all in one block 731 // This output goes directly to the tty, not the compiler log. 732 // To enable tools to match it up with the compilation activity, 733 // be sure to tag this tty output with the compile ID. 734 if (xtty != NULL) { 735 xtty->head("ideal compile_id='%d'%s", compile_id(), 736 is_osr_compilation() ? " compile_kind='osr'" : 737 ""); 738 } 739 root()->dump(9999); 740 if (xtty != NULL) { 741 xtty->tail("ideal"); 742 } 743 } 744#endif 745 746 // Now that we know the size of all the monitors we can add a fixed slot 747 // for the original deopt pc. 748 749 _orig_pc_slot = fixed_slots(); 750 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 751 set_fixed_slots(next_slot); 752 753 // Now generate code 754 Code_Gen(); 755 if (failing()) return; 756 757 // Check if we want to skip execution of all compiled code. 758 { 759#ifndef PRODUCT 760 if (OptoNoExecute) { 761 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 762 return; 763 } 764 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 765#endif 766 767 if (is_osr_compilation()) { 768 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 769 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 770 } else { 771 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 772 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 773 } 774 775 env()->register_method(_method, _entry_bci, 776 &_code_offsets, 777 _orig_pc_slot_offset_in_bytes, 778 code_buffer(), 779 frame_size_in_words(), _oop_map_set, 780 &_handler_table, &_inc_table, 781 compiler, 782 env()->comp_level(), 783 true, /*has_debug_info*/ 784 has_unsafe_access() 785 ); 786 } 787} 788 789//------------------------------Compile---------------------------------------- 790// Compile a runtime stub 791Compile::Compile( ciEnv* ci_env, 792 TypeFunc_generator generator, 793 address stub_function, 794 const char *stub_name, 795 int is_fancy_jump, 796 bool pass_tls, 797 bool save_arg_registers, 798 bool return_pc ) 799 : Phase(Compiler), 800 _env(ci_env), 801 _log(ci_env->log()), 802 _compile_id(-1), 803 _save_argument_registers(save_arg_registers), 804 _method(NULL), 805 _stub_name(stub_name), 806 _stub_function(stub_function), 807 _stub_entry_point(NULL), 808 _entry_bci(InvocationEntryBci), 809 _initial_gvn(NULL), 810 _for_igvn(NULL), 811 _warm_calls(NULL), 812 _orig_pc_slot(0), 813 _orig_pc_slot_offset_in_bytes(0), 814 _subsume_loads(true), 815 _do_escape_analysis(false), 816 _failure_reason(NULL), 817 _code_buffer("Compile::Fill_buffer"), 818 _has_method_handle_invokes(false), 819 _mach_constant_base_node(NULL), 820 _node_bundling_limit(0), 821 _node_bundling_base(NULL), 822 _java_calls(0), 823 _inner_loops(0), 824#ifndef PRODUCT 825 _trace_opto_output(TraceOptoOutput), 826 _printer(NULL), 827#endif 828 _congraph(NULL) { 829 C = this; 830 831#ifndef PRODUCT 832 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 833 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 834 set_print_assembly(PrintFrameConverterAssembly); 835 set_parsed_irreducible_loop(false); 836#endif 837 CompileWrapper cw(this); 838 Init(/*AliasLevel=*/ 0); 839 init_tf((*generator)()); 840 841 { 842 // The following is a dummy for the sake of GraphKit::gen_stub 843 Unique_Node_List for_igvn(comp_arena()); 844 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 845 PhaseGVN gvn(Thread::current()->resource_area(),255); 846 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 847 gvn.transform_no_reclaim(top()); 848 849 GraphKit kit; 850 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 851 } 852 853 NOT_PRODUCT( verify_graph_edges(); ) 854 Code_Gen(); 855 if (failing()) return; 856 857 858 // Entry point will be accessed using compile->stub_entry_point(); 859 if (code_buffer() == NULL) { 860 Matcher::soft_match_failure(); 861 } else { 862 if (PrintAssembly && (WizardMode || Verbose)) 863 tty->print_cr("### Stub::%s", stub_name); 864 865 if (!failing()) { 866 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 867 868 // Make the NMethod 869 // For now we mark the frame as never safe for profile stackwalking 870 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 871 code_buffer(), 872 CodeOffsets::frame_never_safe, 873 // _code_offsets.value(CodeOffsets::Frame_Complete), 874 frame_size_in_words(), 875 _oop_map_set, 876 save_arg_registers); 877 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 878 879 _stub_entry_point = rs->entry_point(); 880 } 881 } 882} 883 884#ifndef PRODUCT 885void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 886 if(PrintOpto && Verbose) { 887 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 888 } 889} 890#endif 891 892void Compile::print_codes() { 893} 894 895//------------------------------Init------------------------------------------- 896// Prepare for a single compilation 897void Compile::Init(int aliaslevel) { 898 _unique = 0; 899 _regalloc = NULL; 900 901 _tf = NULL; // filled in later 902 _top = NULL; // cached later 903 _matcher = NULL; // filled in later 904 _cfg = NULL; // filled in later 905 906 set_24_bit_selection_and_mode(Use24BitFP, false); 907 908 _node_note_array = NULL; 909 _default_node_notes = NULL; 910 911 _immutable_memory = NULL; // filled in at first inquiry 912 913 // Globally visible Nodes 914 // First set TOP to NULL to give safe behavior during creation of RootNode 915 set_cached_top_node(NULL); 916 set_root(new (this, 3) RootNode()); 917 // Now that you have a Root to point to, create the real TOP 918 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 919 set_recent_alloc(NULL, NULL); 920 921 // Create Debug Information Recorder to record scopes, oopmaps, etc. 922 env()->set_oop_recorder(new OopRecorder(comp_arena())); 923 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 924 env()->set_dependencies(new Dependencies(env())); 925 926 _fixed_slots = 0; 927 set_has_split_ifs(false); 928 set_has_loops(has_method() && method()->has_loops()); // first approximation 929 set_has_stringbuilder(false); 930 _trap_can_recompile = false; // no traps emitted yet 931 _major_progress = true; // start out assuming good things will happen 932 set_has_unsafe_access(false); 933 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 934 set_decompile_count(0); 935 936 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 937 set_num_loop_opts(LoopOptsCount); 938 set_do_inlining(Inline); 939 set_max_inline_size(MaxInlineSize); 940 set_freq_inline_size(FreqInlineSize); 941 set_do_scheduling(OptoScheduling); 942 set_do_count_invocations(false); 943 set_do_method_data_update(false); 944 945 if (debug_info()->recording_non_safepoints()) { 946 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 947 (comp_arena(), 8, 0, NULL)); 948 set_default_node_notes(Node_Notes::make(this)); 949 } 950 951 // // -- Initialize types before each compile -- 952 // // Update cached type information 953 // if( _method && _method->constants() ) 954 // Type::update_loaded_types(_method, _method->constants()); 955 956 // Init alias_type map. 957 if (!_do_escape_analysis && aliaslevel == 3) 958 aliaslevel = 2; // No unique types without escape analysis 959 _AliasLevel = aliaslevel; 960 const int grow_ats = 16; 961 _max_alias_types = grow_ats; 962 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 963 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 964 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 965 { 966 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 967 } 968 // Initialize the first few types. 969 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 970 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 971 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 972 _num_alias_types = AliasIdxRaw+1; 973 // Zero out the alias type cache. 974 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 975 // A NULL adr_type hits in the cache right away. Preload the right answer. 976 probe_alias_cache(NULL)->_index = AliasIdxTop; 977 978 _intrinsics = NULL; 979 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 980 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 981 register_library_intrinsics(); 982} 983 984//---------------------------init_start---------------------------------------- 985// Install the StartNode on this compile object. 986void Compile::init_start(StartNode* s) { 987 if (failing()) 988 return; // already failing 989 assert(s == start(), ""); 990} 991 992StartNode* Compile::start() const { 993 assert(!failing(), ""); 994 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 995 Node* start = root()->fast_out(i); 996 if( start->is_Start() ) 997 return start->as_Start(); 998 } 999 ShouldNotReachHere(); 1000 return NULL; 1001} 1002 1003//-------------------------------immutable_memory------------------------------------- 1004// Access immutable memory 1005Node* Compile::immutable_memory() { 1006 if (_immutable_memory != NULL) { 1007 return _immutable_memory; 1008 } 1009 StartNode* s = start(); 1010 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1011 Node *p = s->fast_out(i); 1012 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1013 _immutable_memory = p; 1014 return _immutable_memory; 1015 } 1016 } 1017 ShouldNotReachHere(); 1018 return NULL; 1019} 1020 1021//----------------------set_cached_top_node------------------------------------ 1022// Install the cached top node, and make sure Node::is_top works correctly. 1023void Compile::set_cached_top_node(Node* tn) { 1024 if (tn != NULL) verify_top(tn); 1025 Node* old_top = _top; 1026 _top = tn; 1027 // Calling Node::setup_is_top allows the nodes the chance to adjust 1028 // their _out arrays. 1029 if (_top != NULL) _top->setup_is_top(); 1030 if (old_top != NULL) old_top->setup_is_top(); 1031 assert(_top == NULL || top()->is_top(), ""); 1032} 1033 1034#ifndef PRODUCT 1035void Compile::verify_top(Node* tn) const { 1036 if (tn != NULL) { 1037 assert(tn->is_Con(), "top node must be a constant"); 1038 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1039 assert(tn->in(0) != NULL, "must have live top node"); 1040 } 1041} 1042#endif 1043 1044 1045///-------------------Managing Per-Node Debug & Profile Info------------------- 1046 1047void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1048 guarantee(arr != NULL, ""); 1049 int num_blocks = arr->length(); 1050 if (grow_by < num_blocks) grow_by = num_blocks; 1051 int num_notes = grow_by * _node_notes_block_size; 1052 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1053 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1054 while (num_notes > 0) { 1055 arr->append(notes); 1056 notes += _node_notes_block_size; 1057 num_notes -= _node_notes_block_size; 1058 } 1059 assert(num_notes == 0, "exact multiple, please"); 1060} 1061 1062bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1063 if (source == NULL || dest == NULL) return false; 1064 1065 if (dest->is_Con()) 1066 return false; // Do not push debug info onto constants. 1067 1068#ifdef ASSERT 1069 // Leave a bread crumb trail pointing to the original node: 1070 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1071 dest->set_debug_orig(source); 1072 } 1073#endif 1074 1075 if (node_note_array() == NULL) 1076 return false; // Not collecting any notes now. 1077 1078 // This is a copy onto a pre-existing node, which may already have notes. 1079 // If both nodes have notes, do not overwrite any pre-existing notes. 1080 Node_Notes* source_notes = node_notes_at(source->_idx); 1081 if (source_notes == NULL || source_notes->is_clear()) return false; 1082 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1083 if (dest_notes == NULL || dest_notes->is_clear()) { 1084 return set_node_notes_at(dest->_idx, source_notes); 1085 } 1086 1087 Node_Notes merged_notes = (*source_notes); 1088 // The order of operations here ensures that dest notes will win... 1089 merged_notes.update_from(dest_notes); 1090 return set_node_notes_at(dest->_idx, &merged_notes); 1091} 1092 1093 1094//--------------------------allow_range_check_smearing------------------------- 1095// Gating condition for coalescing similar range checks. 1096// Sometimes we try 'speculatively' replacing a series of a range checks by a 1097// single covering check that is at least as strong as any of them. 1098// If the optimization succeeds, the simplified (strengthened) range check 1099// will always succeed. If it fails, we will deopt, and then give up 1100// on the optimization. 1101bool Compile::allow_range_check_smearing() const { 1102 // If this method has already thrown a range-check, 1103 // assume it was because we already tried range smearing 1104 // and it failed. 1105 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1106 return !already_trapped; 1107} 1108 1109 1110//------------------------------flatten_alias_type----------------------------- 1111const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1112 int offset = tj->offset(); 1113 TypePtr::PTR ptr = tj->ptr(); 1114 1115 // Known instance (scalarizable allocation) alias only with itself. 1116 bool is_known_inst = tj->isa_oopptr() != NULL && 1117 tj->is_oopptr()->is_known_instance(); 1118 1119 // Process weird unsafe references. 1120 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1121 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1122 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1123 tj = TypeOopPtr::BOTTOM; 1124 ptr = tj->ptr(); 1125 offset = tj->offset(); 1126 } 1127 1128 // Array pointers need some flattening 1129 const TypeAryPtr *ta = tj->isa_aryptr(); 1130 if( ta && is_known_inst ) { 1131 if ( offset != Type::OffsetBot && 1132 offset > arrayOopDesc::length_offset_in_bytes() ) { 1133 offset = Type::OffsetBot; // Flatten constant access into array body only 1134 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1135 } 1136 } else if( ta && _AliasLevel >= 2 ) { 1137 // For arrays indexed by constant indices, we flatten the alias 1138 // space to include all of the array body. Only the header, klass 1139 // and array length can be accessed un-aliased. 1140 if( offset != Type::OffsetBot ) { 1141 if( ta->const_oop() ) { // methodDataOop or methodOop 1142 offset = Type::OffsetBot; // Flatten constant access into array body 1143 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1144 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1145 // range is OK as-is. 1146 tj = ta = TypeAryPtr::RANGE; 1147 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1148 tj = TypeInstPtr::KLASS; // all klass loads look alike 1149 ta = TypeAryPtr::RANGE; // generic ignored junk 1150 ptr = TypePtr::BotPTR; 1151 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1152 tj = TypeInstPtr::MARK; 1153 ta = TypeAryPtr::RANGE; // generic ignored junk 1154 ptr = TypePtr::BotPTR; 1155 } else { // Random constant offset into array body 1156 offset = Type::OffsetBot; // Flatten constant access into array body 1157 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1158 } 1159 } 1160 // Arrays of fixed size alias with arrays of unknown size. 1161 if (ta->size() != TypeInt::POS) { 1162 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1163 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1164 } 1165 // Arrays of known objects become arrays of unknown objects. 1166 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1167 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1168 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1169 } 1170 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1171 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1172 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1173 } 1174 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1175 // cannot be distinguished by bytecode alone. 1176 if (ta->elem() == TypeInt::BOOL) { 1177 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1178 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1179 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1180 } 1181 // During the 2nd round of IterGVN, NotNull castings are removed. 1182 // Make sure the Bottom and NotNull variants alias the same. 1183 // Also, make sure exact and non-exact variants alias the same. 1184 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1185 if (ta->const_oop()) { 1186 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1187 } else { 1188 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1189 } 1190 } 1191 } 1192 1193 // Oop pointers need some flattening 1194 const TypeInstPtr *to = tj->isa_instptr(); 1195 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1196 if( ptr == TypePtr::Constant ) { 1197 // No constant oop pointers (such as Strings); they alias with 1198 // unknown strings. 1199 assert(!is_known_inst, "not scalarizable allocation"); 1200 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1201 } else if( is_known_inst ) { 1202 tj = to; // Keep NotNull and klass_is_exact for instance type 1203 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1204 // During the 2nd round of IterGVN, NotNull castings are removed. 1205 // Make sure the Bottom and NotNull variants alias the same. 1206 // Also, make sure exact and non-exact variants alias the same. 1207 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1208 } 1209 // Canonicalize the holder of this field 1210 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1211 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1212 // First handle header references such as a LoadKlassNode, even if the 1213 // object's klass is unloaded at compile time (4965979). 1214 if (!is_known_inst) { // Do it only for non-instance types 1215 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1216 } 1217 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1218 to = NULL; 1219 tj = TypeOopPtr::BOTTOM; 1220 offset = tj->offset(); 1221 } else { 1222 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1223 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1224 if( is_known_inst ) { 1225 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1226 } else { 1227 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1228 } 1229 } 1230 } 1231 } 1232 1233 // Klass pointers to object array klasses need some flattening 1234 const TypeKlassPtr *tk = tj->isa_klassptr(); 1235 if( tk ) { 1236 // If we are referencing a field within a Klass, we need 1237 // to assume the worst case of an Object. Both exact and 1238 // inexact types must flatten to the same alias class. 1239 // Since the flattened result for a klass is defined to be 1240 // precisely java.lang.Object, use a constant ptr. 1241 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1242 1243 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1244 TypeKlassPtr::OBJECT->klass(), 1245 offset); 1246 } 1247 1248 ciKlass* klass = tk->klass(); 1249 if( klass->is_obj_array_klass() ) { 1250 ciKlass* k = TypeAryPtr::OOPS->klass(); 1251 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1252 k = TypeInstPtr::BOTTOM->klass(); 1253 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1254 } 1255 1256 // Check for precise loads from the primary supertype array and force them 1257 // to the supertype cache alias index. Check for generic array loads from 1258 // the primary supertype array and also force them to the supertype cache 1259 // alias index. Since the same load can reach both, we need to merge 1260 // these 2 disparate memories into the same alias class. Since the 1261 // primary supertype array is read-only, there's no chance of confusion 1262 // where we bypass an array load and an array store. 1263 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1264 if( offset == Type::OffsetBot || 1265 off2 < Klass::primary_super_limit()*wordSize ) { 1266 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1267 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1268 } 1269 } 1270 1271 // Flatten all Raw pointers together. 1272 if (tj->base() == Type::RawPtr) 1273 tj = TypeRawPtr::BOTTOM; 1274 1275 if (tj->base() == Type::AnyPtr) 1276 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1277 1278 // Flatten all to bottom for now 1279 switch( _AliasLevel ) { 1280 case 0: 1281 tj = TypePtr::BOTTOM; 1282 break; 1283 case 1: // Flatten to: oop, static, field or array 1284 switch (tj->base()) { 1285 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1286 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1287 case Type::AryPtr: // do not distinguish arrays at all 1288 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1289 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1290 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1291 default: ShouldNotReachHere(); 1292 } 1293 break; 1294 case 2: // No collapsing at level 2; keep all splits 1295 case 3: // No collapsing at level 3; keep all splits 1296 break; 1297 default: 1298 Unimplemented(); 1299 } 1300 1301 offset = tj->offset(); 1302 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1303 1304 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1305 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1306 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1307 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1308 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1309 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1310 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1311 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1312 assert( tj->ptr() != TypePtr::TopPTR && 1313 tj->ptr() != TypePtr::AnyNull && 1314 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1315// assert( tj->ptr() != TypePtr::Constant || 1316// tj->base() == Type::RawPtr || 1317// tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1318 1319 return tj; 1320} 1321 1322void Compile::AliasType::Init(int i, const TypePtr* at) { 1323 _index = i; 1324 _adr_type = at; 1325 _field = NULL; 1326 _is_rewritable = true; // default 1327 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1328 if (atoop != NULL && atoop->is_known_instance()) { 1329 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1330 _general_index = Compile::current()->get_alias_index(gt); 1331 } else { 1332 _general_index = 0; 1333 } 1334} 1335 1336//---------------------------------print_on------------------------------------ 1337#ifndef PRODUCT 1338void Compile::AliasType::print_on(outputStream* st) { 1339 if (index() < 10) 1340 st->print("@ <%d> ", index()); 1341 else st->print("@ <%d>", index()); 1342 st->print(is_rewritable() ? " " : " RO"); 1343 int offset = adr_type()->offset(); 1344 if (offset == Type::OffsetBot) 1345 st->print(" +any"); 1346 else st->print(" +%-3d", offset); 1347 st->print(" in "); 1348 adr_type()->dump_on(st); 1349 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1350 if (field() != NULL && tjp) { 1351 if (tjp->klass() != field()->holder() || 1352 tjp->offset() != field()->offset_in_bytes()) { 1353 st->print(" != "); 1354 field()->print(); 1355 st->print(" ***"); 1356 } 1357 } 1358} 1359 1360void print_alias_types() { 1361 Compile* C = Compile::current(); 1362 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1363 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1364 C->alias_type(idx)->print_on(tty); 1365 tty->cr(); 1366 } 1367} 1368#endif 1369 1370 1371//----------------------------probe_alias_cache-------------------------------- 1372Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1373 intptr_t key = (intptr_t) adr_type; 1374 key ^= key >> logAliasCacheSize; 1375 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1376} 1377 1378 1379//-----------------------------grow_alias_types-------------------------------- 1380void Compile::grow_alias_types() { 1381 const int old_ats = _max_alias_types; // how many before? 1382 const int new_ats = old_ats; // how many more? 1383 const int grow_ats = old_ats+new_ats; // how many now? 1384 _max_alias_types = grow_ats; 1385 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1386 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1387 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1388 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1389} 1390 1391 1392//--------------------------------find_alias_type------------------------------ 1393Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create) { 1394 if (_AliasLevel == 0) 1395 return alias_type(AliasIdxBot); 1396 1397 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1398 if (ace->_adr_type == adr_type) { 1399 return alias_type(ace->_index); 1400 } 1401 1402 // Handle special cases. 1403 if (adr_type == NULL) return alias_type(AliasIdxTop); 1404 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1405 1406 // Do it the slow way. 1407 const TypePtr* flat = flatten_alias_type(adr_type); 1408 1409#ifdef ASSERT 1410 assert(flat == flatten_alias_type(flat), "idempotent"); 1411 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1412 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1413 const TypeOopPtr* foop = flat->is_oopptr(); 1414 // Scalarizable allocations have exact klass always. 1415 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1416 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1417 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1418 } 1419 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1420#endif 1421 1422 int idx = AliasIdxTop; 1423 for (int i = 0; i < num_alias_types(); i++) { 1424 if (alias_type(i)->adr_type() == flat) { 1425 idx = i; 1426 break; 1427 } 1428 } 1429 1430 if (idx == AliasIdxTop) { 1431 if (no_create) return NULL; 1432 // Grow the array if necessary. 1433 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1434 // Add a new alias type. 1435 idx = _num_alias_types++; 1436 _alias_types[idx]->Init(idx, flat); 1437 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1438 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1439 if (flat->isa_instptr()) { 1440 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1441 && flat->is_instptr()->klass() == env()->Class_klass()) 1442 alias_type(idx)->set_rewritable(false); 1443 } 1444 if (flat->isa_klassptr()) { 1445 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1446 alias_type(idx)->set_rewritable(false); 1447 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1448 alias_type(idx)->set_rewritable(false); 1449 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1450 alias_type(idx)->set_rewritable(false); 1451 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1452 alias_type(idx)->set_rewritable(false); 1453 } 1454 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1455 // but the base pointer type is not distinctive enough to identify 1456 // references into JavaThread.) 1457 1458 // Check for final instance fields. 1459 const TypeInstPtr* tinst = flat->isa_instptr(); 1460 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1461 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1462 ciField* field = k->get_field_by_offset(tinst->offset(), false); 1463 // Set field() and is_rewritable() attributes. 1464 if (field != NULL) alias_type(idx)->set_field(field); 1465 } 1466 const TypeKlassPtr* tklass = flat->isa_klassptr(); 1467 // Check for final static fields. 1468 if (tklass && tklass->klass()->is_instance_klass()) { 1469 ciInstanceKlass *k = tklass->klass()->as_instance_klass(); 1470 ciField* field = k->get_field_by_offset(tklass->offset(), true); 1471 // Set field() and is_rewritable() attributes. 1472 if (field != NULL) alias_type(idx)->set_field(field); 1473 } 1474 } 1475 1476 // Fill the cache for next time. 1477 ace->_adr_type = adr_type; 1478 ace->_index = idx; 1479 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1480 1481 // Might as well try to fill the cache for the flattened version, too. 1482 AliasCacheEntry* face = probe_alias_cache(flat); 1483 if (face->_adr_type == NULL) { 1484 face->_adr_type = flat; 1485 face->_index = idx; 1486 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1487 } 1488 1489 return alias_type(idx); 1490} 1491 1492 1493Compile::AliasType* Compile::alias_type(ciField* field) { 1494 const TypeOopPtr* t; 1495 if (field->is_static()) 1496 t = TypeKlassPtr::make(field->holder()); 1497 else 1498 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1499 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes())); 1500 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1501 return atp; 1502} 1503 1504 1505//------------------------------have_alias_type-------------------------------- 1506bool Compile::have_alias_type(const TypePtr* adr_type) { 1507 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1508 if (ace->_adr_type == adr_type) { 1509 return true; 1510 } 1511 1512 // Handle special cases. 1513 if (adr_type == NULL) return true; 1514 if (adr_type == TypePtr::BOTTOM) return true; 1515 1516 return find_alias_type(adr_type, true) != NULL; 1517} 1518 1519//-----------------------------must_alias-------------------------------------- 1520// True if all values of the given address type are in the given alias category. 1521bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1522 if (alias_idx == AliasIdxBot) return true; // the universal category 1523 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1524 if (alias_idx == AliasIdxTop) return false; // the empty category 1525 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1526 1527 // the only remaining possible overlap is identity 1528 int adr_idx = get_alias_index(adr_type); 1529 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1530 assert(adr_idx == alias_idx || 1531 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1532 && adr_type != TypeOopPtr::BOTTOM), 1533 "should not be testing for overlap with an unsafe pointer"); 1534 return adr_idx == alias_idx; 1535} 1536 1537//------------------------------can_alias-------------------------------------- 1538// True if any values of the given address type are in the given alias category. 1539bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1540 if (alias_idx == AliasIdxTop) return false; // the empty category 1541 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1542 if (alias_idx == AliasIdxBot) return true; // the universal category 1543 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1544 1545 // the only remaining possible overlap is identity 1546 int adr_idx = get_alias_index(adr_type); 1547 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1548 return adr_idx == alias_idx; 1549} 1550 1551 1552 1553//---------------------------pop_warm_call------------------------------------- 1554WarmCallInfo* Compile::pop_warm_call() { 1555 WarmCallInfo* wci = _warm_calls; 1556 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1557 return wci; 1558} 1559 1560//----------------------------Inline_Warm-------------------------------------- 1561int Compile::Inline_Warm() { 1562 // If there is room, try to inline some more warm call sites. 1563 // %%% Do a graph index compaction pass when we think we're out of space? 1564 if (!InlineWarmCalls) return 0; 1565 1566 int calls_made_hot = 0; 1567 int room_to_grow = NodeCountInliningCutoff - unique(); 1568 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1569 int amount_grown = 0; 1570 WarmCallInfo* call; 1571 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1572 int est_size = (int)call->size(); 1573 if (est_size > (room_to_grow - amount_grown)) { 1574 // This one won't fit anyway. Get rid of it. 1575 call->make_cold(); 1576 continue; 1577 } 1578 call->make_hot(); 1579 calls_made_hot++; 1580 amount_grown += est_size; 1581 amount_to_grow -= est_size; 1582 } 1583 1584 if (calls_made_hot > 0) set_major_progress(); 1585 return calls_made_hot; 1586} 1587 1588 1589//----------------------------Finish_Warm-------------------------------------- 1590void Compile::Finish_Warm() { 1591 if (!InlineWarmCalls) return; 1592 if (failing()) return; 1593 if (warm_calls() == NULL) return; 1594 1595 // Clean up loose ends, if we are out of space for inlining. 1596 WarmCallInfo* call; 1597 while ((call = pop_warm_call()) != NULL) { 1598 call->make_cold(); 1599 } 1600} 1601 1602//---------------------cleanup_loop_predicates----------------------- 1603// Remove the opaque nodes that protect the predicates so that all unused 1604// checks and uncommon_traps will be eliminated from the ideal graph 1605void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1606 if (predicate_count()==0) return; 1607 for (int i = predicate_count(); i > 0; i--) { 1608 Node * n = predicate_opaque1_node(i-1); 1609 assert(n->Opcode() == Op_Opaque1, "must be"); 1610 igvn.replace_node(n, n->in(1)); 1611 } 1612 assert(predicate_count()==0, "should be clean!"); 1613 igvn.optimize(); 1614} 1615 1616//------------------------------Optimize--------------------------------------- 1617// Given a graph, optimize it. 1618void Compile::Optimize() { 1619 TracePhase t1("optimizer", &_t_optimizer, true); 1620 1621#ifndef PRODUCT 1622 if (env()->break_at_compile()) { 1623 BREAKPOINT; 1624 } 1625 1626#endif 1627 1628 ResourceMark rm; 1629 int loop_opts_cnt; 1630 1631 NOT_PRODUCT( verify_graph_edges(); ) 1632 1633 print_method("After Parsing"); 1634 1635 { 1636 // Iterative Global Value Numbering, including ideal transforms 1637 // Initialize IterGVN with types and values from parse-time GVN 1638 PhaseIterGVN igvn(initial_gvn()); 1639 { 1640 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1641 igvn.optimize(); 1642 } 1643 1644 print_method("Iter GVN 1", 2); 1645 1646 if (failing()) return; 1647 1648 // Perform escape analysis 1649 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1650 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 1651 ConnectionGraph::do_analysis(this, &igvn); 1652 1653 if (failing()) return; 1654 1655 igvn.optimize(); 1656 print_method("Iter GVN 3", 2); 1657 1658 if (failing()) return; 1659 1660 } 1661 1662 // Loop transforms on the ideal graph. Range Check Elimination, 1663 // peeling, unrolling, etc. 1664 1665 // Set loop opts counter 1666 loop_opts_cnt = num_loop_opts(); 1667 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1668 { 1669 TracePhase t2("idealLoop", &_t_idealLoop, true); 1670 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate); 1671 loop_opts_cnt--; 1672 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1673 if (failing()) return; 1674 } 1675 // Loop opts pass if partial peeling occurred in previous pass 1676 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1677 TracePhase t3("idealLoop", &_t_idealLoop, true); 1678 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1679 loop_opts_cnt--; 1680 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1681 if (failing()) return; 1682 } 1683 // Loop opts pass for loop-unrolling before CCP 1684 if(major_progress() && (loop_opts_cnt > 0)) { 1685 TracePhase t4("idealLoop", &_t_idealLoop, true); 1686 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1687 loop_opts_cnt--; 1688 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1689 } 1690 if (!failing()) { 1691 // Verify that last round of loop opts produced a valid graph 1692 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1693 PhaseIdealLoop::verify(igvn); 1694 } 1695 } 1696 if (failing()) return; 1697 1698 // Conditional Constant Propagation; 1699 PhaseCCP ccp( &igvn ); 1700 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1701 { 1702 TracePhase t2("ccp", &_t_ccp, true); 1703 ccp.do_transform(); 1704 } 1705 print_method("PhaseCPP 1", 2); 1706 1707 assert( true, "Break here to ccp.dump_old2new_map()"); 1708 1709 // Iterative Global Value Numbering, including ideal transforms 1710 { 1711 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1712 igvn = ccp; 1713 igvn.optimize(); 1714 } 1715 1716 print_method("Iter GVN 2", 2); 1717 1718 if (failing()) return; 1719 1720 // Loop transforms on the ideal graph. Range Check Elimination, 1721 // peeling, unrolling, etc. 1722 if(loop_opts_cnt > 0) { 1723 debug_only( int cnt = 0; ); 1724 bool loop_predication = UseLoopPredicate; 1725 while(major_progress() && (loop_opts_cnt > 0)) { 1726 TracePhase t2("idealLoop", &_t_idealLoop, true); 1727 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1728 PhaseIdealLoop ideal_loop( igvn, true, loop_predication); 1729 loop_opts_cnt--; 1730 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1731 if (failing()) return; 1732 // Perform loop predication optimization during first iteration after CCP. 1733 // After that switch it off and cleanup unused loop predicates. 1734 if (loop_predication) { 1735 loop_predication = false; 1736 cleanup_loop_predicates(igvn); 1737 if (failing()) return; 1738 } 1739 } 1740 } 1741 1742 { 1743 // Verify that all previous optimizations produced a valid graph 1744 // at least to this point, even if no loop optimizations were done. 1745 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1746 PhaseIdealLoop::verify(igvn); 1747 } 1748 1749 { 1750 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1751 PhaseMacroExpand mex(igvn); 1752 if (mex.expand_macro_nodes()) { 1753 assert(failing(), "must bail out w/ explicit message"); 1754 return; 1755 } 1756 } 1757 1758 } // (End scope of igvn; run destructor if necessary for asserts.) 1759 1760 // A method with only infinite loops has no edges entering loops from root 1761 { 1762 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1763 if (final_graph_reshaping()) { 1764 assert(failing(), "must bail out w/ explicit message"); 1765 return; 1766 } 1767 } 1768 1769 print_method("Optimize finished", 2); 1770} 1771 1772 1773//------------------------------Code_Gen--------------------------------------- 1774// Given a graph, generate code for it 1775void Compile::Code_Gen() { 1776 if (failing()) return; 1777 1778 // Perform instruction selection. You might think we could reclaim Matcher 1779 // memory PDQ, but actually the Matcher is used in generating spill code. 1780 // Internals of the Matcher (including some VectorSets) must remain live 1781 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1782 // set a bit in reclaimed memory. 1783 1784 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1785 // nodes. Mapping is only valid at the root of each matched subtree. 1786 NOT_PRODUCT( verify_graph_edges(); ) 1787 1788 Node_List proj_list; 1789 Matcher m(proj_list); 1790 _matcher = &m; 1791 { 1792 TracePhase t2("matcher", &_t_matcher, true); 1793 m.match(); 1794 } 1795 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1796 // nodes. Mapping is only valid at the root of each matched subtree. 1797 NOT_PRODUCT( verify_graph_edges(); ) 1798 1799 // If you have too many nodes, or if matching has failed, bail out 1800 check_node_count(0, "out of nodes matching instructions"); 1801 if (failing()) return; 1802 1803 // Build a proper-looking CFG 1804 PhaseCFG cfg(node_arena(), root(), m); 1805 _cfg = &cfg; 1806 { 1807 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1808 cfg.Dominators(); 1809 if (failing()) return; 1810 1811 NOT_PRODUCT( verify_graph_edges(); ) 1812 1813 cfg.Estimate_Block_Frequency(); 1814 cfg.GlobalCodeMotion(m,unique(),proj_list); 1815 1816 print_method("Global code motion", 2); 1817 1818 if (failing()) return; 1819 NOT_PRODUCT( verify_graph_edges(); ) 1820 1821 debug_only( cfg.verify(); ) 1822 } 1823 NOT_PRODUCT( verify_graph_edges(); ) 1824 1825 PhaseChaitin regalloc(unique(),cfg,m); 1826 _regalloc = ®alloc; 1827 { 1828 TracePhase t2("regalloc", &_t_registerAllocation, true); 1829 // Perform any platform dependent preallocation actions. This is used, 1830 // for example, to avoid taking an implicit null pointer exception 1831 // using the frame pointer on win95. 1832 _regalloc->pd_preallocate_hook(); 1833 1834 // Perform register allocation. After Chaitin, use-def chains are 1835 // no longer accurate (at spill code) and so must be ignored. 1836 // Node->LRG->reg mappings are still accurate. 1837 _regalloc->Register_Allocate(); 1838 1839 // Bail out if the allocator builds too many nodes 1840 if (failing()) return; 1841 } 1842 1843 // Prior to register allocation we kept empty basic blocks in case the 1844 // the allocator needed a place to spill. After register allocation we 1845 // are not adding any new instructions. If any basic block is empty, we 1846 // can now safely remove it. 1847 { 1848 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1849 cfg.remove_empty(); 1850 if (do_freq_based_layout()) { 1851 PhaseBlockLayout layout(cfg); 1852 } else { 1853 cfg.set_loop_alignment(); 1854 } 1855 cfg.fixup_flow(); 1856 } 1857 1858 // Perform any platform dependent postallocation verifications. 1859 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1860 1861 // Apply peephole optimizations 1862 if( OptoPeephole ) { 1863 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1864 PhasePeephole peep( _regalloc, cfg); 1865 peep.do_transform(); 1866 } 1867 1868 // Convert Nodes to instruction bits in a buffer 1869 { 1870 // %%%% workspace merge brought two timers together for one job 1871 TracePhase t2a("output", &_t_output, true); 1872 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1873 Output(); 1874 } 1875 1876 print_method("Final Code"); 1877 1878 // He's dead, Jim. 1879 _cfg = (PhaseCFG*)0xdeadbeef; 1880 _regalloc = (PhaseChaitin*)0xdeadbeef; 1881} 1882 1883 1884//------------------------------dump_asm--------------------------------------- 1885// Dump formatted assembly 1886#ifndef PRODUCT 1887void Compile::dump_asm(int *pcs, uint pc_limit) { 1888 bool cut_short = false; 1889 tty->print_cr("#"); 1890 tty->print("# "); _tf->dump(); tty->cr(); 1891 tty->print_cr("#"); 1892 1893 // For all blocks 1894 int pc = 0x0; // Program counter 1895 char starts_bundle = ' '; 1896 _regalloc->dump_frame(); 1897 1898 Node *n = NULL; 1899 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1900 if (VMThread::should_terminate()) { cut_short = true; break; } 1901 Block *b = _cfg->_blocks[i]; 1902 if (b->is_connector() && !Verbose) continue; 1903 n = b->_nodes[0]; 1904 if (pcs && n->_idx < pc_limit) 1905 tty->print("%3.3x ", pcs[n->_idx]); 1906 else 1907 tty->print(" "); 1908 b->dump_head( &_cfg->_bbs ); 1909 if (b->is_connector()) { 1910 tty->print_cr(" # Empty connector block"); 1911 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1912 tty->print_cr(" # Block is sole successor of call"); 1913 } 1914 1915 // For all instructions 1916 Node *delay = NULL; 1917 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1918 if (VMThread::should_terminate()) { cut_short = true; break; } 1919 n = b->_nodes[j]; 1920 if (valid_bundle_info(n)) { 1921 Bundle *bundle = node_bundling(n); 1922 if (bundle->used_in_unconditional_delay()) { 1923 delay = n; 1924 continue; 1925 } 1926 if (bundle->starts_bundle()) 1927 starts_bundle = '+'; 1928 } 1929 1930 if (WizardMode) n->dump(); 1931 1932 if( !n->is_Region() && // Dont print in the Assembly 1933 !n->is_Phi() && // a few noisely useless nodes 1934 !n->is_Proj() && 1935 !n->is_MachTemp() && 1936 !n->is_SafePointScalarObject() && 1937 !n->is_Catch() && // Would be nice to print exception table targets 1938 !n->is_MergeMem() && // Not very interesting 1939 !n->is_top() && // Debug info table constants 1940 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1941 ) { 1942 if (pcs && n->_idx < pc_limit) 1943 tty->print("%3.3x", pcs[n->_idx]); 1944 else 1945 tty->print(" "); 1946 tty->print(" %c ", starts_bundle); 1947 starts_bundle = ' '; 1948 tty->print("\t"); 1949 n->format(_regalloc, tty); 1950 tty->cr(); 1951 } 1952 1953 // If we have an instruction with a delay slot, and have seen a delay, 1954 // then back up and print it 1955 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1956 assert(delay != NULL, "no unconditional delay instruction"); 1957 if (WizardMode) delay->dump(); 1958 1959 if (node_bundling(delay)->starts_bundle()) 1960 starts_bundle = '+'; 1961 if (pcs && n->_idx < pc_limit) 1962 tty->print("%3.3x", pcs[n->_idx]); 1963 else 1964 tty->print(" "); 1965 tty->print(" %c ", starts_bundle); 1966 starts_bundle = ' '; 1967 tty->print("\t"); 1968 delay->format(_regalloc, tty); 1969 tty->print_cr(""); 1970 delay = NULL; 1971 } 1972 1973 // Dump the exception table as well 1974 if( n->is_Catch() && (Verbose || WizardMode) ) { 1975 // Print the exception table for this offset 1976 _handler_table.print_subtable_for(pc); 1977 } 1978 } 1979 1980 if (pcs && n->_idx < pc_limit) 1981 tty->print_cr("%3.3x", pcs[n->_idx]); 1982 else 1983 tty->print_cr(""); 1984 1985 assert(cut_short || delay == NULL, "no unconditional delay branch"); 1986 1987 } // End of per-block dump 1988 tty->print_cr(""); 1989 1990 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 1991} 1992#endif 1993 1994//------------------------------Final_Reshape_Counts--------------------------- 1995// This class defines counters to help identify when a method 1996// may/must be executed using hardware with only 24-bit precision. 1997struct Final_Reshape_Counts : public StackObj { 1998 int _call_count; // count non-inlined 'common' calls 1999 int _float_count; // count float ops requiring 24-bit precision 2000 int _double_count; // count double ops requiring more precision 2001 int _java_call_count; // count non-inlined 'java' calls 2002 int _inner_loop_count; // count loops which need alignment 2003 VectorSet _visited; // Visitation flags 2004 Node_List _tests; // Set of IfNodes & PCTableNodes 2005 2006 Final_Reshape_Counts() : 2007 _call_count(0), _float_count(0), _double_count(0), 2008 _java_call_count(0), _inner_loop_count(0), 2009 _visited( Thread::current()->resource_area() ) { } 2010 2011 void inc_call_count () { _call_count ++; } 2012 void inc_float_count () { _float_count ++; } 2013 void inc_double_count() { _double_count++; } 2014 void inc_java_call_count() { _java_call_count++; } 2015 void inc_inner_loop_count() { _inner_loop_count++; } 2016 2017 int get_call_count () const { return _call_count ; } 2018 int get_float_count () const { return _float_count ; } 2019 int get_double_count() const { return _double_count; } 2020 int get_java_call_count() const { return _java_call_count; } 2021 int get_inner_loop_count() const { return _inner_loop_count; } 2022}; 2023 2024static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2025 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2026 // Make sure the offset goes inside the instance layout. 2027 return k->contains_field_offset(tp->offset()); 2028 // Note that OffsetBot and OffsetTop are very negative. 2029} 2030 2031//------------------------------final_graph_reshaping_impl---------------------- 2032// Implement items 1-5 from final_graph_reshaping below. 2033static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 2034 2035 if ( n->outcnt() == 0 ) return; // dead node 2036 uint nop = n->Opcode(); 2037 2038 // Check for 2-input instruction with "last use" on right input. 2039 // Swap to left input. Implements item (2). 2040 if( n->req() == 3 && // two-input instruction 2041 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2042 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2043 n->in(2)->outcnt() == 1 &&// right use IS a last use 2044 !n->in(2)->is_Con() ) { // right use is not a constant 2045 // Check for commutative opcode 2046 switch( nop ) { 2047 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2048 case Op_MaxI: case Op_MinI: 2049 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2050 case Op_AndL: case Op_XorL: case Op_OrL: 2051 case Op_AndI: case Op_XorI: case Op_OrI: { 2052 // Move "last use" input to left by swapping inputs 2053 n->swap_edges(1, 2); 2054 break; 2055 } 2056 default: 2057 break; 2058 } 2059 } 2060 2061#ifdef ASSERT 2062 if( n->is_Mem() ) { 2063 Compile* C = Compile::current(); 2064 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2065 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2066 // oop will be recorded in oop map if load crosses safepoint 2067 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2068 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2069 "raw memory operations should have control edge"); 2070 } 2071#endif 2072 // Count FPU ops and common calls, implements item (3) 2073 switch( nop ) { 2074 // Count all float operations that may use FPU 2075 case Op_AddF: 2076 case Op_SubF: 2077 case Op_MulF: 2078 case Op_DivF: 2079 case Op_NegF: 2080 case Op_ModF: 2081 case Op_ConvI2F: 2082 case Op_ConF: 2083 case Op_CmpF: 2084 case Op_CmpF3: 2085 // case Op_ConvL2F: // longs are split into 32-bit halves 2086 frc.inc_float_count(); 2087 break; 2088 2089 case Op_ConvF2D: 2090 case Op_ConvD2F: 2091 frc.inc_float_count(); 2092 frc.inc_double_count(); 2093 break; 2094 2095 // Count all double operations that may use FPU 2096 case Op_AddD: 2097 case Op_SubD: 2098 case Op_MulD: 2099 case Op_DivD: 2100 case Op_NegD: 2101 case Op_ModD: 2102 case Op_ConvI2D: 2103 case Op_ConvD2I: 2104 // case Op_ConvL2D: // handled by leaf call 2105 // case Op_ConvD2L: // handled by leaf call 2106 case Op_ConD: 2107 case Op_CmpD: 2108 case Op_CmpD3: 2109 frc.inc_double_count(); 2110 break; 2111 case Op_Opaque1: // Remove Opaque Nodes before matching 2112 case Op_Opaque2: // Remove Opaque Nodes before matching 2113 n->subsume_by(n->in(1)); 2114 break; 2115 case Op_CallStaticJava: 2116 case Op_CallJava: 2117 case Op_CallDynamicJava: 2118 frc.inc_java_call_count(); // Count java call site; 2119 case Op_CallRuntime: 2120 case Op_CallLeaf: 2121 case Op_CallLeafNoFP: { 2122 assert( n->is_Call(), "" ); 2123 CallNode *call = n->as_Call(); 2124 // Count call sites where the FP mode bit would have to be flipped. 2125 // Do not count uncommon runtime calls: 2126 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2127 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2128 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2129 frc.inc_call_count(); // Count the call site 2130 } else { // See if uncommon argument is shared 2131 Node *n = call->in(TypeFunc::Parms); 2132 int nop = n->Opcode(); 2133 // Clone shared simple arguments to uncommon calls, item (1). 2134 if( n->outcnt() > 1 && 2135 !n->is_Proj() && 2136 nop != Op_CreateEx && 2137 nop != Op_CheckCastPP && 2138 nop != Op_DecodeN && 2139 !n->is_Mem() ) { 2140 Node *x = n->clone(); 2141 call->set_req( TypeFunc::Parms, x ); 2142 } 2143 } 2144 break; 2145 } 2146 2147 case Op_StoreD: 2148 case Op_LoadD: 2149 case Op_LoadD_unaligned: 2150 frc.inc_double_count(); 2151 goto handle_mem; 2152 case Op_StoreF: 2153 case Op_LoadF: 2154 frc.inc_float_count(); 2155 goto handle_mem; 2156 2157 case Op_StoreB: 2158 case Op_StoreC: 2159 case Op_StoreCM: 2160 case Op_StorePConditional: 2161 case Op_StoreI: 2162 case Op_StoreL: 2163 case Op_StoreIConditional: 2164 case Op_StoreLConditional: 2165 case Op_CompareAndSwapI: 2166 case Op_CompareAndSwapL: 2167 case Op_CompareAndSwapP: 2168 case Op_CompareAndSwapN: 2169 case Op_StoreP: 2170 case Op_StoreN: 2171 case Op_LoadB: 2172 case Op_LoadUB: 2173 case Op_LoadUS: 2174 case Op_LoadI: 2175 case Op_LoadUI2L: 2176 case Op_LoadKlass: 2177 case Op_LoadNKlass: 2178 case Op_LoadL: 2179 case Op_LoadL_unaligned: 2180 case Op_LoadPLocked: 2181 case Op_LoadLLocked: 2182 case Op_LoadP: 2183 case Op_LoadN: 2184 case Op_LoadRange: 2185 case Op_LoadS: { 2186 handle_mem: 2187#ifdef ASSERT 2188 if( VerifyOptoOopOffsets ) { 2189 assert( n->is_Mem(), "" ); 2190 MemNode *mem = (MemNode*)n; 2191 // Check to see if address types have grounded out somehow. 2192 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2193 assert( !tp || oop_offset_is_sane(tp), "" ); 2194 } 2195#endif 2196 break; 2197 } 2198 2199 case Op_AddP: { // Assert sane base pointers 2200 Node *addp = n->in(AddPNode::Address); 2201 assert( !addp->is_AddP() || 2202 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2203 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2204 "Base pointers must match" ); 2205#ifdef _LP64 2206 if (UseCompressedOops && 2207 addp->Opcode() == Op_ConP && 2208 addp == n->in(AddPNode::Base) && 2209 n->in(AddPNode::Offset)->is_Con()) { 2210 // Use addressing with narrow klass to load with offset on x86. 2211 // On sparc loading 32-bits constant and decoding it have less 2212 // instructions (4) then load 64-bits constant (7). 2213 // Do this transformation here since IGVN will convert ConN back to ConP. 2214 const Type* t = addp->bottom_type(); 2215 if (t->isa_oopptr()) { 2216 Node* nn = NULL; 2217 2218 // Look for existing ConN node of the same exact type. 2219 Compile* C = Compile::current(); 2220 Node* r = C->root(); 2221 uint cnt = r->outcnt(); 2222 for (uint i = 0; i < cnt; i++) { 2223 Node* m = r->raw_out(i); 2224 if (m!= NULL && m->Opcode() == Op_ConN && 2225 m->bottom_type()->make_ptr() == t) { 2226 nn = m; 2227 break; 2228 } 2229 } 2230 if (nn != NULL) { 2231 // Decode a narrow oop to match address 2232 // [R12 + narrow_oop_reg<<3 + offset] 2233 nn = new (C, 2) DecodeNNode(nn, t); 2234 n->set_req(AddPNode::Base, nn); 2235 n->set_req(AddPNode::Address, nn); 2236 if (addp->outcnt() == 0) { 2237 addp->disconnect_inputs(NULL); 2238 } 2239 } 2240 } 2241 } 2242#endif 2243 break; 2244 } 2245 2246#ifdef _LP64 2247 case Op_CastPP: 2248 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2249 Compile* C = Compile::current(); 2250 Node* in1 = n->in(1); 2251 const Type* t = n->bottom_type(); 2252 Node* new_in1 = in1->clone(); 2253 new_in1->as_DecodeN()->set_type(t); 2254 2255 if (!Matcher::narrow_oop_use_complex_address()) { 2256 // 2257 // x86, ARM and friends can handle 2 adds in addressing mode 2258 // and Matcher can fold a DecodeN node into address by using 2259 // a narrow oop directly and do implicit NULL check in address: 2260 // 2261 // [R12 + narrow_oop_reg<<3 + offset] 2262 // NullCheck narrow_oop_reg 2263 // 2264 // On other platforms (Sparc) we have to keep new DecodeN node and 2265 // use it to do implicit NULL check in address: 2266 // 2267 // decode_not_null narrow_oop_reg, base_reg 2268 // [base_reg + offset] 2269 // NullCheck base_reg 2270 // 2271 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2272 // to keep the information to which NULL check the new DecodeN node 2273 // corresponds to use it as value in implicit_null_check(). 2274 // 2275 new_in1->set_req(0, n->in(0)); 2276 } 2277 2278 n->subsume_by(new_in1); 2279 if (in1->outcnt() == 0) { 2280 in1->disconnect_inputs(NULL); 2281 } 2282 } 2283 break; 2284 2285 case Op_CmpP: 2286 // Do this transformation here to preserve CmpPNode::sub() and 2287 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2288 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2289 Node* in1 = n->in(1); 2290 Node* in2 = n->in(2); 2291 if (!in1->is_DecodeN()) { 2292 in2 = in1; 2293 in1 = n->in(2); 2294 } 2295 assert(in1->is_DecodeN(), "sanity"); 2296 2297 Compile* C = Compile::current(); 2298 Node* new_in2 = NULL; 2299 if (in2->is_DecodeN()) { 2300 new_in2 = in2->in(1); 2301 } else if (in2->Opcode() == Op_ConP) { 2302 const Type* t = in2->bottom_type(); 2303 if (t == TypePtr::NULL_PTR) { 2304 // Don't convert CmpP null check into CmpN if compressed 2305 // oops implicit null check is not generated. 2306 // This will allow to generate normal oop implicit null check. 2307 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2308 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2309 // 2310 // This transformation together with CastPP transformation above 2311 // will generated code for implicit NULL checks for compressed oops. 2312 // 2313 // The original code after Optimize() 2314 // 2315 // LoadN memory, narrow_oop_reg 2316 // decode narrow_oop_reg, base_reg 2317 // CmpP base_reg, NULL 2318 // CastPP base_reg // NotNull 2319 // Load [base_reg + offset], val_reg 2320 // 2321 // after these transformations will be 2322 // 2323 // LoadN memory, narrow_oop_reg 2324 // CmpN narrow_oop_reg, NULL 2325 // decode_not_null narrow_oop_reg, base_reg 2326 // Load [base_reg + offset], val_reg 2327 // 2328 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2329 // since narrow oops can be used in debug info now (see the code in 2330 // final_graph_reshaping_walk()). 2331 // 2332 // At the end the code will be matched to 2333 // on x86: 2334 // 2335 // Load_narrow_oop memory, narrow_oop_reg 2336 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2337 // NullCheck narrow_oop_reg 2338 // 2339 // and on sparc: 2340 // 2341 // Load_narrow_oop memory, narrow_oop_reg 2342 // decode_not_null narrow_oop_reg, base_reg 2343 // Load [base_reg + offset], val_reg 2344 // NullCheck base_reg 2345 // 2346 } else if (t->isa_oopptr()) { 2347 new_in2 = ConNode::make(C, t->make_narrowoop()); 2348 } 2349 } 2350 if (new_in2 != NULL) { 2351 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2352 n->subsume_by( cmpN ); 2353 if (in1->outcnt() == 0) { 2354 in1->disconnect_inputs(NULL); 2355 } 2356 if (in2->outcnt() == 0) { 2357 in2->disconnect_inputs(NULL); 2358 } 2359 } 2360 } 2361 break; 2362 2363 case Op_DecodeN: 2364 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2365 // DecodeN could be pinned when it can't be fold into 2366 // an address expression, see the code for Op_CastPP above. 2367 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2368 break; 2369 2370 case Op_EncodeP: { 2371 Node* in1 = n->in(1); 2372 if (in1->is_DecodeN()) { 2373 n->subsume_by(in1->in(1)); 2374 } else if (in1->Opcode() == Op_ConP) { 2375 Compile* C = Compile::current(); 2376 const Type* t = in1->bottom_type(); 2377 if (t == TypePtr::NULL_PTR) { 2378 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2379 } else if (t->isa_oopptr()) { 2380 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2381 } 2382 } 2383 if (in1->outcnt() == 0) { 2384 in1->disconnect_inputs(NULL); 2385 } 2386 break; 2387 } 2388 2389 case Op_Proj: { 2390 if (OptimizeStringConcat) { 2391 ProjNode* p = n->as_Proj(); 2392 if (p->_is_io_use) { 2393 // Separate projections were used for the exception path which 2394 // are normally removed by a late inline. If it wasn't inlined 2395 // then they will hang around and should just be replaced with 2396 // the original one. 2397 Node* proj = NULL; 2398 // Replace with just one 2399 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2400 Node *use = i.get(); 2401 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2402 proj = use; 2403 break; 2404 } 2405 } 2406 assert(p != NULL, "must be found"); 2407 p->subsume_by(proj); 2408 } 2409 } 2410 break; 2411 } 2412 2413 case Op_Phi: 2414 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2415 // The EncodeP optimization may create Phi with the same edges 2416 // for all paths. It is not handled well by Register Allocator. 2417 Node* unique_in = n->in(1); 2418 assert(unique_in != NULL, ""); 2419 uint cnt = n->req(); 2420 for (uint i = 2; i < cnt; i++) { 2421 Node* m = n->in(i); 2422 assert(m != NULL, ""); 2423 if (unique_in != m) 2424 unique_in = NULL; 2425 } 2426 if (unique_in != NULL) { 2427 n->subsume_by(unique_in); 2428 } 2429 } 2430 break; 2431 2432#endif 2433 2434 case Op_ModI: 2435 if (UseDivMod) { 2436 // Check if a%b and a/b both exist 2437 Node* d = n->find_similar(Op_DivI); 2438 if (d) { 2439 // Replace them with a fused divmod if supported 2440 Compile* C = Compile::current(); 2441 if (Matcher::has_match_rule(Op_DivModI)) { 2442 DivModINode* divmod = DivModINode::make(C, n); 2443 d->subsume_by(divmod->div_proj()); 2444 n->subsume_by(divmod->mod_proj()); 2445 } else { 2446 // replace a%b with a-((a/b)*b) 2447 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2448 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2449 n->subsume_by( sub ); 2450 } 2451 } 2452 } 2453 break; 2454 2455 case Op_ModL: 2456 if (UseDivMod) { 2457 // Check if a%b and a/b both exist 2458 Node* d = n->find_similar(Op_DivL); 2459 if (d) { 2460 // Replace them with a fused divmod if supported 2461 Compile* C = Compile::current(); 2462 if (Matcher::has_match_rule(Op_DivModL)) { 2463 DivModLNode* divmod = DivModLNode::make(C, n); 2464 d->subsume_by(divmod->div_proj()); 2465 n->subsume_by(divmod->mod_proj()); 2466 } else { 2467 // replace a%b with a-((a/b)*b) 2468 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2469 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2470 n->subsume_by( sub ); 2471 } 2472 } 2473 } 2474 break; 2475 2476 case Op_Load16B: 2477 case Op_Load8B: 2478 case Op_Load4B: 2479 case Op_Load8S: 2480 case Op_Load4S: 2481 case Op_Load2S: 2482 case Op_Load8C: 2483 case Op_Load4C: 2484 case Op_Load2C: 2485 case Op_Load4I: 2486 case Op_Load2I: 2487 case Op_Load2L: 2488 case Op_Load4F: 2489 case Op_Load2F: 2490 case Op_Load2D: 2491 case Op_Store16B: 2492 case Op_Store8B: 2493 case Op_Store4B: 2494 case Op_Store8C: 2495 case Op_Store4C: 2496 case Op_Store2C: 2497 case Op_Store4I: 2498 case Op_Store2I: 2499 case Op_Store2L: 2500 case Op_Store4F: 2501 case Op_Store2F: 2502 case Op_Store2D: 2503 break; 2504 2505 case Op_PackB: 2506 case Op_PackS: 2507 case Op_PackC: 2508 case Op_PackI: 2509 case Op_PackF: 2510 case Op_PackL: 2511 case Op_PackD: 2512 if (n->req()-1 > 2) { 2513 // Replace many operand PackNodes with a binary tree for matching 2514 PackNode* p = (PackNode*) n; 2515 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2516 n->subsume_by(btp); 2517 } 2518 break; 2519 case Op_Loop: 2520 case Op_CountedLoop: 2521 if (n->as_Loop()->is_inner_loop()) { 2522 frc.inc_inner_loop_count(); 2523 } 2524 break; 2525 default: 2526 assert( !n->is_Call(), "" ); 2527 assert( !n->is_Mem(), "" ); 2528 break; 2529 } 2530 2531 // Collect CFG split points 2532 if (n->is_MultiBranch()) 2533 frc._tests.push(n); 2534} 2535 2536//------------------------------final_graph_reshaping_walk--------------------- 2537// Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2538// requires that the walk visits a node's inputs before visiting the node. 2539static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2540 ResourceArea *area = Thread::current()->resource_area(); 2541 Unique_Node_List sfpt(area); 2542 2543 frc._visited.set(root->_idx); // first, mark node as visited 2544 uint cnt = root->req(); 2545 Node *n = root; 2546 uint i = 0; 2547 while (true) { 2548 if (i < cnt) { 2549 // Place all non-visited non-null inputs onto stack 2550 Node* m = n->in(i); 2551 ++i; 2552 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2553 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2554 sfpt.push(m); 2555 cnt = m->req(); 2556 nstack.push(n, i); // put on stack parent and next input's index 2557 n = m; 2558 i = 0; 2559 } 2560 } else { 2561 // Now do post-visit work 2562 final_graph_reshaping_impl( n, frc ); 2563 if (nstack.is_empty()) 2564 break; // finished 2565 n = nstack.node(); // Get node from stack 2566 cnt = n->req(); 2567 i = nstack.index(); 2568 nstack.pop(); // Shift to the next node on stack 2569 } 2570 } 2571 2572 // Skip next transformation if compressed oops are not used. 2573 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) 2574 return; 2575 2576 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2577 // It could be done for an uncommon traps or any safepoints/calls 2578 // if the DecodeN node is referenced only in a debug info. 2579 while (sfpt.size() > 0) { 2580 n = sfpt.pop(); 2581 JVMState *jvms = n->as_SafePoint()->jvms(); 2582 assert(jvms != NULL, "sanity"); 2583 int start = jvms->debug_start(); 2584 int end = n->req(); 2585 bool is_uncommon = (n->is_CallStaticJava() && 2586 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2587 for (int j = start; j < end; j++) { 2588 Node* in = n->in(j); 2589 if (in->is_DecodeN()) { 2590 bool safe_to_skip = true; 2591 if (!is_uncommon ) { 2592 // Is it safe to skip? 2593 for (uint i = 0; i < in->outcnt(); i++) { 2594 Node* u = in->raw_out(i); 2595 if (!u->is_SafePoint() || 2596 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2597 safe_to_skip = false; 2598 } 2599 } 2600 } 2601 if (safe_to_skip) { 2602 n->set_req(j, in->in(1)); 2603 } 2604 if (in->outcnt() == 0) { 2605 in->disconnect_inputs(NULL); 2606 } 2607 } 2608 } 2609 } 2610} 2611 2612//------------------------------final_graph_reshaping-------------------------- 2613// Final Graph Reshaping. 2614// 2615// (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2616// and not commoned up and forced early. Must come after regular 2617// optimizations to avoid GVN undoing the cloning. Clone constant 2618// inputs to Loop Phis; these will be split by the allocator anyways. 2619// Remove Opaque nodes. 2620// (2) Move last-uses by commutative operations to the left input to encourage 2621// Intel update-in-place two-address operations and better register usage 2622// on RISCs. Must come after regular optimizations to avoid GVN Ideal 2623// calls canonicalizing them back. 2624// (3) Count the number of double-precision FP ops, single-precision FP ops 2625// and call sites. On Intel, we can get correct rounding either by 2626// forcing singles to memory (requires extra stores and loads after each 2627// FP bytecode) or we can set a rounding mode bit (requires setting and 2628// clearing the mode bit around call sites). The mode bit is only used 2629// if the relative frequency of single FP ops to calls is low enough. 2630// This is a key transform for SPEC mpeg_audio. 2631// (4) Detect infinite loops; blobs of code reachable from above but not 2632// below. Several of the Code_Gen algorithms fail on such code shapes, 2633// so we simply bail out. Happens a lot in ZKM.jar, but also happens 2634// from time to time in other codes (such as -Xcomp finalizer loops, etc). 2635// Detection is by looking for IfNodes where only 1 projection is 2636// reachable from below or CatchNodes missing some targets. 2637// (5) Assert for insane oop offsets in debug mode. 2638 2639bool Compile::final_graph_reshaping() { 2640 // an infinite loop may have been eliminated by the optimizer, 2641 // in which case the graph will be empty. 2642 if (root()->req() == 1) { 2643 record_method_not_compilable("trivial infinite loop"); 2644 return true; 2645 } 2646 2647 Final_Reshape_Counts frc; 2648 2649 // Visit everybody reachable! 2650 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2651 Node_Stack nstack(unique() >> 1); 2652 final_graph_reshaping_walk(nstack, root(), frc); 2653 2654 // Check for unreachable (from below) code (i.e., infinite loops). 2655 for( uint i = 0; i < frc._tests.size(); i++ ) { 2656 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2657 // Get number of CFG targets. 2658 // Note that PCTables include exception targets after calls. 2659 uint required_outcnt = n->required_outcnt(); 2660 if (n->outcnt() != required_outcnt) { 2661 // Check for a few special cases. Rethrow Nodes never take the 2662 // 'fall-thru' path, so expected kids is 1 less. 2663 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2664 if (n->in(0)->in(0)->is_Call()) { 2665 CallNode *call = n->in(0)->in(0)->as_Call(); 2666 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2667 required_outcnt--; // Rethrow always has 1 less kid 2668 } else if (call->req() > TypeFunc::Parms && 2669 call->is_CallDynamicJava()) { 2670 // Check for null receiver. In such case, the optimizer has 2671 // detected that the virtual call will always result in a null 2672 // pointer exception. The fall-through projection of this CatchNode 2673 // will not be populated. 2674 Node *arg0 = call->in(TypeFunc::Parms); 2675 if (arg0->is_Type() && 2676 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2677 required_outcnt--; 2678 } 2679 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2680 call->req() > TypeFunc::Parms+1 && 2681 call->is_CallStaticJava()) { 2682 // Check for negative array length. In such case, the optimizer has 2683 // detected that the allocation attempt will always result in an 2684 // exception. There is no fall-through projection of this CatchNode . 2685 Node *arg1 = call->in(TypeFunc::Parms+1); 2686 if (arg1->is_Type() && 2687 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2688 required_outcnt--; 2689 } 2690 } 2691 } 2692 } 2693 // Recheck with a better notion of 'required_outcnt' 2694 if (n->outcnt() != required_outcnt) { 2695 record_method_not_compilable("malformed control flow"); 2696 return true; // Not all targets reachable! 2697 } 2698 } 2699 // Check that I actually visited all kids. Unreached kids 2700 // must be infinite loops. 2701 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2702 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2703 record_method_not_compilable("infinite loop"); 2704 return true; // Found unvisited kid; must be unreach 2705 } 2706 } 2707 2708 // If original bytecodes contained a mixture of floats and doubles 2709 // check if the optimizer has made it homogenous, item (3). 2710 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2711 frc.get_float_count() > 32 && 2712 frc.get_double_count() == 0 && 2713 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2714 set_24_bit_selection_and_mode( false, true ); 2715 } 2716 2717 set_java_calls(frc.get_java_call_count()); 2718 set_inner_loops(frc.get_inner_loop_count()); 2719 2720 // No infinite loops, no reason to bail out. 2721 return false; 2722} 2723 2724//-----------------------------too_many_traps---------------------------------- 2725// Report if there are too many traps at the current method and bci. 2726// Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2727bool Compile::too_many_traps(ciMethod* method, 2728 int bci, 2729 Deoptimization::DeoptReason reason) { 2730 ciMethodData* md = method->method_data(); 2731 if (md->is_empty()) { 2732 // Assume the trap has not occurred, or that it occurred only 2733 // because of a transient condition during start-up in the interpreter. 2734 return false; 2735 } 2736 if (md->has_trap_at(bci, reason) != 0) { 2737 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2738 // Also, if there are multiple reasons, or if there is no per-BCI record, 2739 // assume the worst. 2740 if (log()) 2741 log()->elem("observe trap='%s' count='%d'", 2742 Deoptimization::trap_reason_name(reason), 2743 md->trap_count(reason)); 2744 return true; 2745 } else { 2746 // Ignore method/bci and see if there have been too many globally. 2747 return too_many_traps(reason, md); 2748 } 2749} 2750 2751// Less-accurate variant which does not require a method and bci. 2752bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2753 ciMethodData* logmd) { 2754 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2755 // Too many traps globally. 2756 // Note that we use cumulative trap_count, not just md->trap_count. 2757 if (log()) { 2758 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2759 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2760 Deoptimization::trap_reason_name(reason), 2761 mcount, trap_count(reason)); 2762 } 2763 return true; 2764 } else { 2765 // The coast is clear. 2766 return false; 2767 } 2768} 2769 2770//--------------------------too_many_recompiles-------------------------------- 2771// Report if there are too many recompiles at the current method and bci. 2772// Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2773// Is not eager to return true, since this will cause the compiler to use 2774// Action_none for a trap point, to avoid too many recompilations. 2775bool Compile::too_many_recompiles(ciMethod* method, 2776 int bci, 2777 Deoptimization::DeoptReason reason) { 2778 ciMethodData* md = method->method_data(); 2779 if (md->is_empty()) { 2780 // Assume the trap has not occurred, or that it occurred only 2781 // because of a transient condition during start-up in the interpreter. 2782 return false; 2783 } 2784 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2785 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2786 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2787 Deoptimization::DeoptReason per_bc_reason 2788 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2789 if ((per_bc_reason == Deoptimization::Reason_none 2790 || md->has_trap_at(bci, reason) != 0) 2791 // The trap frequency measure we care about is the recompile count: 2792 && md->trap_recompiled_at(bci) 2793 && md->overflow_recompile_count() >= bc_cutoff) { 2794 // Do not emit a trap here if it has already caused recompilations. 2795 // Also, if there are multiple reasons, or if there is no per-BCI record, 2796 // assume the worst. 2797 if (log()) 2798 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2799 Deoptimization::trap_reason_name(reason), 2800 md->trap_count(reason), 2801 md->overflow_recompile_count()); 2802 return true; 2803 } else if (trap_count(reason) != 0 2804 && decompile_count() >= m_cutoff) { 2805 // Too many recompiles globally, and we have seen this sort of trap. 2806 // Use cumulative decompile_count, not just md->decompile_count. 2807 if (log()) 2808 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2809 Deoptimization::trap_reason_name(reason), 2810 md->trap_count(reason), trap_count(reason), 2811 md->decompile_count(), decompile_count()); 2812 return true; 2813 } else { 2814 // The coast is clear. 2815 return false; 2816 } 2817} 2818 2819 2820#ifndef PRODUCT 2821//------------------------------verify_graph_edges--------------------------- 2822// Walk the Graph and verify that there is a one-to-one correspondence 2823// between Use-Def edges and Def-Use edges in the graph. 2824void Compile::verify_graph_edges(bool no_dead_code) { 2825 if (VerifyGraphEdges) { 2826 ResourceArea *area = Thread::current()->resource_area(); 2827 Unique_Node_List visited(area); 2828 // Call recursive graph walk to check edges 2829 _root->verify_edges(visited); 2830 if (no_dead_code) { 2831 // Now make sure that no visited node is used by an unvisited node. 2832 bool dead_nodes = 0; 2833 Unique_Node_List checked(area); 2834 while (visited.size() > 0) { 2835 Node* n = visited.pop(); 2836 checked.push(n); 2837 for (uint i = 0; i < n->outcnt(); i++) { 2838 Node* use = n->raw_out(i); 2839 if (checked.member(use)) continue; // already checked 2840 if (visited.member(use)) continue; // already in the graph 2841 if (use->is_Con()) continue; // a dead ConNode is OK 2842 // At this point, we have found a dead node which is DU-reachable. 2843 if (dead_nodes++ == 0) 2844 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2845 use->dump(2); 2846 tty->print_cr("---"); 2847 checked.push(use); // No repeats; pretend it is now checked. 2848 } 2849 } 2850 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2851 } 2852 } 2853} 2854#endif 2855 2856// The Compile object keeps track of failure reasons separately from the ciEnv. 2857// This is required because there is not quite a 1-1 relation between the 2858// ciEnv and its compilation task and the Compile object. Note that one 2859// ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2860// to backtrack and retry without subsuming loads. Other than this backtracking 2861// behavior, the Compile's failure reason is quietly copied up to the ciEnv 2862// by the logic in C2Compiler. 2863void Compile::record_failure(const char* reason) { 2864 if (log() != NULL) { 2865 log()->elem("failure reason='%s' phase='compile'", reason); 2866 } 2867 if (_failure_reason == NULL) { 2868 // Record the first failure reason. 2869 _failure_reason = reason; 2870 } 2871 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2872 C->print_method(_failure_reason); 2873 } 2874 _root = NULL; // flush the graph, too 2875} 2876 2877Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2878 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2879{ 2880 if (dolog) { 2881 C = Compile::current(); 2882 _log = C->log(); 2883 } else { 2884 C = NULL; 2885 _log = NULL; 2886 } 2887 if (_log != NULL) { 2888 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2889 _log->stamp(); 2890 _log->end_head(); 2891 } 2892} 2893 2894Compile::TracePhase::~TracePhase() { 2895 if (_log != NULL) { 2896 _log->done("phase nodes='%d'", C->unique()); 2897 } 2898} 2899 2900//============================================================================= 2901// Two Constant's are equal when the type and the value are equal. 2902bool Compile::Constant::operator==(const Constant& other) { 2903 if (type() != other.type() ) return false; 2904 if (can_be_reused() != other.can_be_reused()) return false; 2905 // For floating point values we compare the bit pattern. 2906 switch (type()) { 2907 case T_FLOAT: return (_value.i == other._value.i); 2908 case T_LONG: 2909 case T_DOUBLE: return (_value.j == other._value.j); 2910 case T_OBJECT: 2911 case T_ADDRESS: return (_value.l == other._value.l); 2912 case T_VOID: return (_value.l == other._value.l); // jump-table entries 2913 default: ShouldNotReachHere(); 2914 } 2915 return false; 2916} 2917 2918// Emit constants grouped in the following order: 2919static BasicType type_order[] = { 2920 T_FLOAT, // 32-bit 2921 T_OBJECT, // 32 or 64-bit 2922 T_ADDRESS, // 32 or 64-bit 2923 T_DOUBLE, // 64-bit 2924 T_LONG, // 64-bit 2925 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons) 2926 T_ILLEGAL 2927}; 2928 2929static int type_to_size_in_bytes(BasicType t) { 2930 switch (t) { 2931 case T_LONG: return sizeof(jlong ); 2932 case T_FLOAT: return sizeof(jfloat ); 2933 case T_DOUBLE: return sizeof(jdouble); 2934 // We use T_VOID as marker for jump-table entries (labels) which 2935 // need an interal word relocation. 2936 case T_VOID: 2937 case T_ADDRESS: 2938 case T_OBJECT: return sizeof(jobject); 2939 } 2940 2941 ShouldNotReachHere(); 2942 return -1; 2943} 2944 2945void Compile::ConstantTable::calculate_offsets_and_size() { 2946 int size = 0; 2947 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 2948 BasicType type = type_order[t]; 2949 2950 for (int i = 0; i < _constants.length(); i++) { 2951 Constant con = _constants.at(i); 2952 if (con.type() != type) continue; // Skip other types. 2953 2954 // Align size for type. 2955 int typesize = type_to_size_in_bytes(con.type()); 2956 size = align_size_up(size, typesize); 2957 2958 // Set offset. 2959 con.set_offset(size); 2960 _constants.at_put(i, con); 2961 2962 // Add type size. 2963 size = size + typesize; 2964 } 2965 } 2966 2967 // Align size up to the next section start (which is insts; see 2968 // CodeBuffer::align_at_start). 2969 assert(_size == -1, "already set?"); 2970 _size = align_size_up(size, CodeEntryAlignment); 2971 2972 if (Matcher::constant_table_absolute_addressing) { 2973 set_table_base_offset(0); // No table base offset required 2974 } else { 2975 if (UseRDPCForConstantTableBase) { 2976 // table base offset is set in MachConstantBaseNode::emit 2977 } else { 2978 // When RDPC is not used, the table base is set into the middle of 2979 // the constant table. 2980 int half_size = _size / 2; 2981 assert(half_size * 2 == _size, "sanity"); 2982 set_table_base_offset(-half_size); 2983 } 2984 } 2985} 2986 2987void Compile::ConstantTable::emit(CodeBuffer& cb) { 2988 MacroAssembler _masm(&cb); 2989 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 2990 BasicType type = type_order[t]; 2991 2992 for (int i = 0; i < _constants.length(); i++) { 2993 Constant con = _constants.at(i); 2994 if (con.type() != type) continue; // Skip other types. 2995 2996 address constant_addr; 2997 switch (con.type()) { 2998 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 2999 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3000 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3001 case T_OBJECT: { 3002 jobject obj = con.get_jobject(); 3003 int oop_index = _masm.oop_recorder()->find_index(obj); 3004 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3005 break; 3006 } 3007 case T_ADDRESS: { 3008 address addr = (address) con.get_jobject(); 3009 constant_addr = _masm.address_constant(addr); 3010 break; 3011 } 3012 // We use T_VOID as marker for jump-table entries (labels) which 3013 // need an interal word relocation. 3014 case T_VOID: { 3015 // Write a dummy word. The real value is filled in later 3016 // in fill_jump_table_in_constant_table. 3017 address addr = (address) con.get_jobject(); 3018 constant_addr = _masm.address_constant(addr); 3019 break; 3020 } 3021 default: ShouldNotReachHere(); 3022 } 3023 assert(constant_addr != NULL, "consts section too small"); 3024 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3025 } 3026 } 3027} 3028 3029int Compile::ConstantTable::find_offset(Constant& con) const { 3030 int idx = _constants.find(con); 3031 assert(idx != -1, "constant must be in constant table"); 3032 int offset = _constants.at(idx).offset(); 3033 assert(offset != -1, "constant table not emitted yet?"); 3034 return offset; 3035} 3036 3037void Compile::ConstantTable::add(Constant& con) { 3038 if (con.can_be_reused()) { 3039 int idx = _constants.find(con); 3040 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3041 return; 3042 } 3043 } 3044 (void) _constants.append(con); 3045} 3046 3047Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) { 3048 Constant con(type, value); 3049 add(con); 3050 return con; 3051} 3052 3053Compile::Constant Compile::ConstantTable::add(MachOper* oper) { 3054 jvalue value; 3055 BasicType type = oper->type()->basic_type(); 3056 switch (type) { 3057 case T_LONG: value.j = oper->constantL(); break; 3058 case T_FLOAT: value.f = oper->constantF(); break; 3059 case T_DOUBLE: value.d = oper->constantD(); break; 3060 case T_OBJECT: 3061 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3062 default: ShouldNotReachHere(); 3063 } 3064 return add(type, value); 3065} 3066 3067Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) { 3068 jvalue value; 3069 // We can use the node pointer here to identify the right jump-table 3070 // as this method is called from Compile::Fill_buffer right before 3071 // the MachNodes are emitted and the jump-table is filled (means the 3072 // MachNode pointers do not change anymore). 3073 value.l = (jobject) n; 3074 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused. 3075 for (uint i = 0; i < n->outcnt(); i++) { 3076 add(con); 3077 } 3078 return con; 3079} 3080 3081void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3082 // If called from Compile::scratch_emit_size do nothing. 3083 if (Compile::current()->in_scratch_emit_size()) return; 3084 3085 assert(labels.is_nonempty(), "must be"); 3086 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt())); 3087 3088 // Since MachConstantNode::constant_offset() also contains 3089 // table_base_offset() we need to subtract the table_base_offset() 3090 // to get the plain offset into the constant table. 3091 int offset = n->constant_offset() - table_base_offset(); 3092 3093 MacroAssembler _masm(&cb); 3094 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3095 3096 for (int i = 0; i < labels.length(); i++) { 3097 address* constant_addr = &jump_table_base[i]; 3098 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer"); 3099 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3100 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3101 } 3102} 3103