compile.cpp revision 3602:da91efe96a93
1/* 2 * Copyright (c) 1997, 2012, 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#ifdef TARGET_ARCH_MODEL_arm 78# include "adfiles/ad_arm.hpp" 79#endif 80#ifdef TARGET_ARCH_MODEL_ppc 81# include "adfiles/ad_ppc.hpp" 82#endif 83 84 85// -------------------- Compile::mach_constant_base_node ----------------------- 86// Constant table base node singleton. 87MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == NULL) { 89 _mach_constant_base_node = new (C) MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93} 94 95 96/// Support for intrinsics. 97 98// Return the index at which m must be inserted (or already exists). 99// The sort order is by the address of the ciMethod, with is_virtual as minor key. 100int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 101#ifdef ASSERT 102 for (int i = 1; i < _intrinsics->length(); i++) { 103 CallGenerator* cg1 = _intrinsics->at(i-1); 104 CallGenerator* cg2 = _intrinsics->at(i); 105 assert(cg1->method() != cg2->method() 106 ? cg1->method() < cg2->method() 107 : cg1->is_virtual() < cg2->is_virtual(), 108 "compiler intrinsics list must stay sorted"); 109 } 110#endif 111 // Binary search sorted list, in decreasing intervals [lo, hi]. 112 int lo = 0, hi = _intrinsics->length()-1; 113 while (lo <= hi) { 114 int mid = (uint)(hi + lo) / 2; 115 ciMethod* mid_m = _intrinsics->at(mid)->method(); 116 if (m < mid_m) { 117 hi = mid-1; 118 } else if (m > mid_m) { 119 lo = mid+1; 120 } else { 121 // look at minor sort key 122 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 123 if (is_virtual < mid_virt) { 124 hi = mid-1; 125 } else if (is_virtual > mid_virt) { 126 lo = mid+1; 127 } else { 128 return mid; // exact match 129 } 130 } 131 } 132 return lo; // inexact match 133} 134 135void Compile::register_intrinsic(CallGenerator* cg) { 136 if (_intrinsics == NULL) { 137 _intrinsics = new GrowableArray<CallGenerator*>(60); 138 } 139 // This code is stolen from ciObjectFactory::insert. 140 // Really, GrowableArray should have methods for 141 // insert_at, remove_at, and binary_search. 142 int len = _intrinsics->length(); 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 144 if (index == len) { 145 _intrinsics->append(cg); 146 } else { 147#ifdef ASSERT 148 CallGenerator* oldcg = _intrinsics->at(index); 149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 150#endif 151 _intrinsics->append(_intrinsics->at(len-1)); 152 int pos; 153 for (pos = len-2; pos >= index; pos--) { 154 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 155 } 156 _intrinsics->at_put(index, cg); 157 } 158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 159} 160 161CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 162 assert(m->is_loaded(), "don't try this on unloaded methods"); 163 if (_intrinsics != NULL) { 164 int index = intrinsic_insertion_index(m, is_virtual); 165 if (index < _intrinsics->length() 166 && _intrinsics->at(index)->method() == m 167 && _intrinsics->at(index)->is_virtual() == is_virtual) { 168 return _intrinsics->at(index); 169 } 170 } 171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 172 if (m->intrinsic_id() != vmIntrinsics::_none && 173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 175 if (cg != NULL) { 176 // Save it for next time: 177 register_intrinsic(cg); 178 return cg; 179 } else { 180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 181 } 182 } 183 return NULL; 184} 185 186// Compile:: register_library_intrinsics and make_vm_intrinsic are defined 187// in library_call.cpp. 188 189 190#ifndef PRODUCT 191// statistics gathering... 192 193juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 194jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 195 196bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 198 int oflags = _intrinsic_hist_flags[id]; 199 assert(flags != 0, "what happened?"); 200 if (is_virtual) { 201 flags |= _intrinsic_virtual; 202 } 203 bool changed = (flags != oflags); 204 if ((flags & _intrinsic_worked) != 0) { 205 juint count = (_intrinsic_hist_count[id] += 1); 206 if (count == 1) { 207 changed = true; // first time 208 } 209 // increment the overall count also: 210 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 211 } 212 if (changed) { 213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 214 // Something changed about the intrinsic's virtuality. 215 if ((flags & _intrinsic_virtual) != 0) { 216 // This is the first use of this intrinsic as a virtual call. 217 if (oflags != 0) { 218 // We already saw it as a non-virtual, so note both cases. 219 flags |= _intrinsic_both; 220 } 221 } else if ((oflags & _intrinsic_both) == 0) { 222 // This is the first use of this intrinsic as a non-virtual 223 flags |= _intrinsic_both; 224 } 225 } 226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 227 } 228 // update the overall flags also: 229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 230 return changed; 231} 232 233static char* format_flags(int flags, char* buf) { 234 buf[0] = 0; 235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 240 if (buf[0] == 0) strcat(buf, ","); 241 assert(buf[0] == ',', "must be"); 242 return &buf[1]; 243} 244 245void Compile::print_intrinsic_statistics() { 246 char flagsbuf[100]; 247 ttyLocker ttyl; 248 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 249 tty->print_cr("Compiler intrinsic usage:"); 250 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 251 if (total == 0) total = 1; // avoid div0 in case of no successes 252 #define PRINT_STAT_LINE(name, c, f) \ 253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 255 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 256 int flags = _intrinsic_hist_flags[id]; 257 juint count = _intrinsic_hist_count[id]; 258 if ((flags | count) != 0) { 259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 260 } 261 } 262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 263 if (xtty != NULL) xtty->tail("statistics"); 264} 265 266void Compile::print_statistics() { 267 { ttyLocker ttyl; 268 if (xtty != NULL) xtty->head("statistics type='opto'"); 269 Parse::print_statistics(); 270 PhaseCCP::print_statistics(); 271 PhaseRegAlloc::print_statistics(); 272 Scheduling::print_statistics(); 273 PhasePeephole::print_statistics(); 274 PhaseIdealLoop::print_statistics(); 275 if (xtty != NULL) xtty->tail("statistics"); 276 } 277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 278 // put this under its own <statistics> element. 279 print_intrinsic_statistics(); 280 } 281} 282#endif //PRODUCT 283 284// Support for bundling info 285Bundle* Compile::node_bundling(const Node *n) { 286 assert(valid_bundle_info(n), "oob"); 287 return &_node_bundling_base[n->_idx]; 288} 289 290bool Compile::valid_bundle_info(const Node *n) { 291 return (_node_bundling_limit > n->_idx); 292} 293 294 295void Compile::gvn_replace_by(Node* n, Node* nn) { 296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 297 Node* use = n->last_out(i); 298 bool is_in_table = initial_gvn()->hash_delete(use); 299 uint uses_found = 0; 300 for (uint j = 0; j < use->len(); j++) { 301 if (use->in(j) == n) { 302 if (j < use->req()) 303 use->set_req(j, nn); 304 else 305 use->set_prec(j, nn); 306 uses_found++; 307 } 308 } 309 if (is_in_table) { 310 // reinsert into table 311 initial_gvn()->hash_find_insert(use); 312 } 313 record_for_igvn(use); 314 i -= uses_found; // we deleted 1 or more copies of this edge 315 } 316} 317 318 319 320 321// Identify all nodes that are reachable from below, useful. 322// Use breadth-first pass that records state in a Unique_Node_List, 323// recursive traversal is slower. 324void Compile::identify_useful_nodes(Unique_Node_List &useful) { 325 int estimated_worklist_size = unique(); 326 useful.map( estimated_worklist_size, NULL ); // preallocate space 327 328 // Initialize worklist 329 if (root() != NULL) { useful.push(root()); } 330 // If 'top' is cached, declare it useful to preserve cached node 331 if( cached_top_node() ) { useful.push(cached_top_node()); } 332 333 // Push all useful nodes onto the list, breadthfirst 334 for( uint next = 0; next < useful.size(); ++next ) { 335 assert( next < unique(), "Unique useful nodes < total nodes"); 336 Node *n = useful.at(next); 337 uint max = n->len(); 338 for( uint i = 0; i < max; ++i ) { 339 Node *m = n->in(i); 340 if( m == NULL ) continue; 341 useful.push(m); 342 } 343 } 344} 345 346// Disconnect all useless nodes by disconnecting those at the boundary. 347void Compile::remove_useless_nodes(Unique_Node_List &useful) { 348 uint next = 0; 349 while (next < useful.size()) { 350 Node *n = useful.at(next++); 351 // Use raw traversal of out edges since this code removes out edges 352 int max = n->outcnt(); 353 for (int j = 0; j < max; ++j) { 354 Node* child = n->raw_out(j); 355 if (! useful.member(child)) { 356 assert(!child->is_top() || child != top(), 357 "If top is cached in Compile object it is in useful list"); 358 // Only need to remove this out-edge to the useless node 359 n->raw_del_out(j); 360 --j; 361 --max; 362 } 363 } 364 if (n->outcnt() == 1 && n->has_special_unique_user()) { 365 record_for_igvn(n->unique_out()); 366 } 367 } 368 // Remove useless macro and predicate opaq nodes 369 for (int i = C->macro_count()-1; i >= 0; i--) { 370 Node* n = C->macro_node(i); 371 if (!useful.member(n)) { 372 remove_macro_node(n); 373 } 374 } 375 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 376} 377 378//------------------------------frame_size_in_words----------------------------- 379// frame_slots in units of words 380int Compile::frame_size_in_words() const { 381 // shift is 0 in LP32 and 1 in LP64 382 const int shift = (LogBytesPerWord - LogBytesPerInt); 383 int words = _frame_slots >> shift; 384 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 385 return words; 386} 387 388// ============================================================================ 389//------------------------------CompileWrapper--------------------------------- 390class CompileWrapper : public StackObj { 391 Compile *const _compile; 392 public: 393 CompileWrapper(Compile* compile); 394 395 ~CompileWrapper(); 396}; 397 398CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 399 // the Compile* pointer is stored in the current ciEnv: 400 ciEnv* env = compile->env(); 401 assert(env == ciEnv::current(), "must already be a ciEnv active"); 402 assert(env->compiler_data() == NULL, "compile already active?"); 403 env->set_compiler_data(compile); 404 assert(compile == Compile::current(), "sanity"); 405 406 compile->set_type_dict(NULL); 407 compile->set_type_hwm(NULL); 408 compile->set_type_last_size(0); 409 compile->set_last_tf(NULL, NULL); 410 compile->set_indexSet_arena(NULL); 411 compile->set_indexSet_free_block_list(NULL); 412 compile->init_type_arena(); 413 Type::Initialize(compile); 414 _compile->set_scratch_buffer_blob(NULL); 415 _compile->begin_method(); 416} 417CompileWrapper::~CompileWrapper() { 418 _compile->end_method(); 419 if (_compile->scratch_buffer_blob() != NULL) 420 BufferBlob::free(_compile->scratch_buffer_blob()); 421 _compile->env()->set_compiler_data(NULL); 422} 423 424 425//----------------------------print_compile_messages--------------------------- 426void Compile::print_compile_messages() { 427#ifndef PRODUCT 428 // Check if recompiling 429 if (_subsume_loads == false && PrintOpto) { 430 // Recompiling without allowing machine instructions to subsume loads 431 tty->print_cr("*********************************************************"); 432 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 433 tty->print_cr("*********************************************************"); 434 } 435 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 436 // Recompiling without escape analysis 437 tty->print_cr("*********************************************************"); 438 tty->print_cr("** Bailout: Recompile without escape analysis **"); 439 tty->print_cr("*********************************************************"); 440 } 441 if (env()->break_at_compile()) { 442 // Open the debugger when compiling this method. 443 tty->print("### Breaking when compiling: "); 444 method()->print_short_name(); 445 tty->cr(); 446 BREAKPOINT; 447 } 448 449 if( PrintOpto ) { 450 if (is_osr_compilation()) { 451 tty->print("[OSR]%3d", _compile_id); 452 } else { 453 tty->print("%3d", _compile_id); 454 } 455 } 456#endif 457} 458 459 460//-----------------------init_scratch_buffer_blob------------------------------ 461// Construct a temporary BufferBlob and cache it for this compile. 462void Compile::init_scratch_buffer_blob(int const_size) { 463 // If there is already a scratch buffer blob allocated and the 464 // constant section is big enough, use it. Otherwise free the 465 // current and allocate a new one. 466 BufferBlob* blob = scratch_buffer_blob(); 467 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 468 // Use the current blob. 469 } else { 470 if (blob != NULL) { 471 BufferBlob::free(blob); 472 } 473 474 ResourceMark rm; 475 _scratch_const_size = const_size; 476 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 477 blob = BufferBlob::create("Compile::scratch_buffer", size); 478 // Record the buffer blob for next time. 479 set_scratch_buffer_blob(blob); 480 // Have we run out of code space? 481 if (scratch_buffer_blob() == NULL) { 482 // Let CompilerBroker disable further compilations. 483 record_failure("Not enough space for scratch buffer in CodeCache"); 484 return; 485 } 486 } 487 488 // Initialize the relocation buffers 489 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 490 set_scratch_locs_memory(locs_buf); 491} 492 493 494//-----------------------scratch_emit_size------------------------------------- 495// Helper function that computes size by emitting code 496uint Compile::scratch_emit_size(const Node* n) { 497 // Start scratch_emit_size section. 498 set_in_scratch_emit_size(true); 499 500 // Emit into a trash buffer and count bytes emitted. 501 // This is a pretty expensive way to compute a size, 502 // but it works well enough if seldom used. 503 // All common fixed-size instructions are given a size 504 // method by the AD file. 505 // Note that the scratch buffer blob and locs memory are 506 // allocated at the beginning of the compile task, and 507 // may be shared by several calls to scratch_emit_size. 508 // The allocation of the scratch buffer blob is particularly 509 // expensive, since it has to grab the code cache lock. 510 BufferBlob* blob = this->scratch_buffer_blob(); 511 assert(blob != NULL, "Initialize BufferBlob at start"); 512 assert(blob->size() > MAX_inst_size, "sanity"); 513 relocInfo* locs_buf = scratch_locs_memory(); 514 address blob_begin = blob->content_begin(); 515 address blob_end = (address)locs_buf; 516 assert(blob->content_contains(blob_end), "sanity"); 517 CodeBuffer buf(blob_begin, blob_end - blob_begin); 518 buf.initialize_consts_size(_scratch_const_size); 519 buf.initialize_stubs_size(MAX_stubs_size); 520 assert(locs_buf != NULL, "sanity"); 521 int lsize = MAX_locs_size / 3; 522 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 523 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 524 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 525 526 // Do the emission. 527 528 Label fakeL; // Fake label for branch instructions. 529 Label* saveL = NULL; 530 uint save_bnum = 0; 531 bool is_branch = n->is_MachBranch(); 532 if (is_branch) { 533 MacroAssembler masm(&buf); 534 masm.bind(fakeL); 535 n->as_MachBranch()->save_label(&saveL, &save_bnum); 536 n->as_MachBranch()->label_set(&fakeL, 0); 537 } 538 n->emit(buf, this->regalloc()); 539 if (is_branch) // Restore label. 540 n->as_MachBranch()->label_set(saveL, save_bnum); 541 542 // End scratch_emit_size section. 543 set_in_scratch_emit_size(false); 544 545 return buf.insts_size(); 546} 547 548 549// ============================================================================ 550//------------------------------Compile standard------------------------------- 551debug_only( int Compile::_debug_idx = 100000; ) 552 553// Compile a method. entry_bci is -1 for normal compilations and indicates 554// the continuation bci for on stack replacement. 555 556 557Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 558 : Phase(Compiler), 559 _env(ci_env), 560 _log(ci_env->log()), 561 _compile_id(ci_env->compile_id()), 562 _save_argument_registers(false), 563 _stub_name(NULL), 564 _stub_function(NULL), 565 _stub_entry_point(NULL), 566 _method(target), 567 _entry_bci(osr_bci), 568 _initial_gvn(NULL), 569 _for_igvn(NULL), 570 _warm_calls(NULL), 571 _subsume_loads(subsume_loads), 572 _do_escape_analysis(do_escape_analysis), 573 _failure_reason(NULL), 574 _code_buffer("Compile::Fill_buffer"), 575 _orig_pc_slot(0), 576 _orig_pc_slot_offset_in_bytes(0), 577 _has_method_handle_invokes(false), 578 _mach_constant_base_node(NULL), 579 _node_bundling_limit(0), 580 _node_bundling_base(NULL), 581 _java_calls(0), 582 _inner_loops(0), 583 _scratch_const_size(-1), 584 _in_scratch_emit_size(false), 585#ifndef PRODUCT 586 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 587 _printer(IdealGraphPrinter::printer()), 588#endif 589 _congraph(NULL) { 590 C = this; 591 592 CompileWrapper cw(this); 593#ifndef PRODUCT 594 if (TimeCompiler2) { 595 tty->print(" "); 596 target->holder()->name()->print(); 597 tty->print("."); 598 target->print_short_name(); 599 tty->print(" "); 600 } 601 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 602 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 603 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 604 if (!print_opto_assembly) { 605 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 606 if (print_assembly && !Disassembler::can_decode()) { 607 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 608 print_opto_assembly = true; 609 } 610 } 611 set_print_assembly(print_opto_assembly); 612 set_parsed_irreducible_loop(false); 613#endif 614 615 if (ProfileTraps) { 616 // Make sure the method being compiled gets its own MDO, 617 // so we can at least track the decompile_count(). 618 method()->ensure_method_data(); 619 } 620 621 Init(::AliasLevel); 622 623 624 print_compile_messages(); 625 626 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 627 _ilt = InlineTree::build_inline_tree_root(); 628 else 629 _ilt = NULL; 630 631 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 632 assert(num_alias_types() >= AliasIdxRaw, ""); 633 634#define MINIMUM_NODE_HASH 1023 635 // Node list that Iterative GVN will start with 636 Unique_Node_List for_igvn(comp_arena()); 637 set_for_igvn(&for_igvn); 638 639 // GVN that will be run immediately on new nodes 640 uint estimated_size = method()->code_size()*4+64; 641 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 642 PhaseGVN gvn(node_arena(), estimated_size); 643 set_initial_gvn(&gvn); 644 645 { // Scope for timing the parser 646 TracePhase t3("parse", &_t_parser, true); 647 648 // Put top into the hash table ASAP. 649 initial_gvn()->transform_no_reclaim(top()); 650 651 // Set up tf(), start(), and find a CallGenerator. 652 CallGenerator* cg = NULL; 653 if (is_osr_compilation()) { 654 const TypeTuple *domain = StartOSRNode::osr_domain(); 655 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 656 init_tf(TypeFunc::make(domain, range)); 657 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 658 initial_gvn()->set_type_bottom(s); 659 init_start(s); 660 cg = CallGenerator::for_osr(method(), entry_bci()); 661 } else { 662 // Normal case. 663 init_tf(TypeFunc::make(method())); 664 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 665 initial_gvn()->set_type_bottom(s); 666 init_start(s); 667 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 668 // With java.lang.ref.reference.get() we must go through the 669 // intrinsic when G1 is enabled - even when get() is the root 670 // method of the compile - so that, if necessary, the value in 671 // the referent field of the reference object gets recorded by 672 // the pre-barrier code. 673 // Specifically, if G1 is enabled, the value in the referent 674 // field is recorded by the G1 SATB pre barrier. This will 675 // result in the referent being marked live and the reference 676 // object removed from the list of discovered references during 677 // reference processing. 678 cg = find_intrinsic(method(), false); 679 } 680 if (cg == NULL) { 681 float past_uses = method()->interpreter_invocation_count(); 682 float expected_uses = past_uses; 683 cg = CallGenerator::for_inline(method(), expected_uses); 684 } 685 } 686 if (failing()) return; 687 if (cg == NULL) { 688 record_method_not_compilable_all_tiers("cannot parse method"); 689 return; 690 } 691 JVMState* jvms = build_start_state(start(), tf()); 692 if ((jvms = cg->generate(jvms)) == NULL) { 693 record_method_not_compilable("method parse failed"); 694 return; 695 } 696 GraphKit kit(jvms); 697 698 if (!kit.stopped()) { 699 // Accept return values, and transfer control we know not where. 700 // This is done by a special, unique ReturnNode bound to root. 701 return_values(kit.jvms()); 702 } 703 704 if (kit.has_exceptions()) { 705 // Any exceptions that escape from this call must be rethrown 706 // to whatever caller is dynamically above us on the stack. 707 // This is done by a special, unique RethrowNode bound to root. 708 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 709 } 710 711 if (!failing() && has_stringbuilder()) { 712 { 713 // remove useless nodes to make the usage analysis simpler 714 ResourceMark rm; 715 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 716 } 717 718 { 719 ResourceMark rm; 720 print_method("Before StringOpts", 3); 721 PhaseStringOpts pso(initial_gvn(), &for_igvn); 722 print_method("After StringOpts", 3); 723 } 724 725 // now inline anything that we skipped the first time around 726 while (_late_inlines.length() > 0) { 727 CallGenerator* cg = _late_inlines.pop(); 728 cg->do_late_inline(); 729 if (failing()) return; 730 } 731 } 732 assert(_late_inlines.length() == 0, "should have been processed"); 733 734 print_method("Before RemoveUseless", 3); 735 736 // Remove clutter produced by parsing. 737 if (!failing()) { 738 ResourceMark rm; 739 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 740 } 741 } 742 743 // Note: Large methods are capped off in do_one_bytecode(). 744 if (failing()) return; 745 746 // After parsing, node notes are no longer automagic. 747 // They must be propagated by register_new_node_with_optimizer(), 748 // clone(), or the like. 749 set_default_node_notes(NULL); 750 751 for (;;) { 752 int successes = Inline_Warm(); 753 if (failing()) return; 754 if (successes == 0) break; 755 } 756 757 // Drain the list. 758 Finish_Warm(); 759#ifndef PRODUCT 760 if (_printer) { 761 _printer->print_inlining(this); 762 } 763#endif 764 765 if (failing()) return; 766 NOT_PRODUCT( verify_graph_edges(); ) 767 768 // Now optimize 769 Optimize(); 770 if (failing()) return; 771 NOT_PRODUCT( verify_graph_edges(); ) 772 773#ifndef PRODUCT 774 if (PrintIdeal) { 775 ttyLocker ttyl; // keep the following output all in one block 776 // This output goes directly to the tty, not the compiler log. 777 // To enable tools to match it up with the compilation activity, 778 // be sure to tag this tty output with the compile ID. 779 if (xtty != NULL) { 780 xtty->head("ideal compile_id='%d'%s", compile_id(), 781 is_osr_compilation() ? " compile_kind='osr'" : 782 ""); 783 } 784 root()->dump(9999); 785 if (xtty != NULL) { 786 xtty->tail("ideal"); 787 } 788 } 789#endif 790 791 // Now that we know the size of all the monitors we can add a fixed slot 792 // for the original deopt pc. 793 794 _orig_pc_slot = fixed_slots(); 795 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 796 set_fixed_slots(next_slot); 797 798 // Now generate code 799 Code_Gen(); 800 if (failing()) return; 801 802 // Check if we want to skip execution of all compiled code. 803 { 804#ifndef PRODUCT 805 if (OptoNoExecute) { 806 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 807 return; 808 } 809 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 810#endif 811 812 if (is_osr_compilation()) { 813 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 814 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 815 } else { 816 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 817 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 818 } 819 820 env()->register_method(_method, _entry_bci, 821 &_code_offsets, 822 _orig_pc_slot_offset_in_bytes, 823 code_buffer(), 824 frame_size_in_words(), _oop_map_set, 825 &_handler_table, &_inc_table, 826 compiler, 827 env()->comp_level(), 828 has_unsafe_access() 829 ); 830 } 831} 832 833//------------------------------Compile---------------------------------------- 834// Compile a runtime stub 835Compile::Compile( ciEnv* ci_env, 836 TypeFunc_generator generator, 837 address stub_function, 838 const char *stub_name, 839 int is_fancy_jump, 840 bool pass_tls, 841 bool save_arg_registers, 842 bool return_pc ) 843 : Phase(Compiler), 844 _env(ci_env), 845 _log(ci_env->log()), 846 _compile_id(-1), 847 _save_argument_registers(save_arg_registers), 848 _method(NULL), 849 _stub_name(stub_name), 850 _stub_function(stub_function), 851 _stub_entry_point(NULL), 852 _entry_bci(InvocationEntryBci), 853 _initial_gvn(NULL), 854 _for_igvn(NULL), 855 _warm_calls(NULL), 856 _orig_pc_slot(0), 857 _orig_pc_slot_offset_in_bytes(0), 858 _subsume_loads(true), 859 _do_escape_analysis(false), 860 _failure_reason(NULL), 861 _code_buffer("Compile::Fill_buffer"), 862 _has_method_handle_invokes(false), 863 _mach_constant_base_node(NULL), 864 _node_bundling_limit(0), 865 _node_bundling_base(NULL), 866 _java_calls(0), 867 _inner_loops(0), 868#ifndef PRODUCT 869 _trace_opto_output(TraceOptoOutput), 870 _printer(NULL), 871#endif 872 _congraph(NULL) { 873 C = this; 874 875#ifndef PRODUCT 876 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 877 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 878 set_print_assembly(PrintFrameConverterAssembly); 879 set_parsed_irreducible_loop(false); 880#endif 881 CompileWrapper cw(this); 882 Init(/*AliasLevel=*/ 0); 883 init_tf((*generator)()); 884 885 { 886 // The following is a dummy for the sake of GraphKit::gen_stub 887 Unique_Node_List for_igvn(comp_arena()); 888 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 889 PhaseGVN gvn(Thread::current()->resource_area(),255); 890 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 891 gvn.transform_no_reclaim(top()); 892 893 GraphKit kit; 894 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 895 } 896 897 NOT_PRODUCT( verify_graph_edges(); ) 898 Code_Gen(); 899 if (failing()) return; 900 901 902 // Entry point will be accessed using compile->stub_entry_point(); 903 if (code_buffer() == NULL) { 904 Matcher::soft_match_failure(); 905 } else { 906 if (PrintAssembly && (WizardMode || Verbose)) 907 tty->print_cr("### Stub::%s", stub_name); 908 909 if (!failing()) { 910 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 911 912 // Make the NMethod 913 // For now we mark the frame as never safe for profile stackwalking 914 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 915 code_buffer(), 916 CodeOffsets::frame_never_safe, 917 // _code_offsets.value(CodeOffsets::Frame_Complete), 918 frame_size_in_words(), 919 _oop_map_set, 920 save_arg_registers); 921 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 922 923 _stub_entry_point = rs->entry_point(); 924 } 925 } 926} 927 928//------------------------------Init------------------------------------------- 929// Prepare for a single compilation 930void Compile::Init(int aliaslevel) { 931 _unique = 0; 932 _regalloc = NULL; 933 934 _tf = NULL; // filled in later 935 _top = NULL; // cached later 936 _matcher = NULL; // filled in later 937 _cfg = NULL; // filled in later 938 939 set_24_bit_selection_and_mode(Use24BitFP, false); 940 941 _node_note_array = NULL; 942 _default_node_notes = NULL; 943 944 _immutable_memory = NULL; // filled in at first inquiry 945 946 // Globally visible Nodes 947 // First set TOP to NULL to give safe behavior during creation of RootNode 948 set_cached_top_node(NULL); 949 set_root(new (this, 3) RootNode()); 950 // Now that you have a Root to point to, create the real TOP 951 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 952 set_recent_alloc(NULL, NULL); 953 954 // Create Debug Information Recorder to record scopes, oopmaps, etc. 955 env()->set_oop_recorder(new OopRecorder(env()->arena())); 956 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 957 env()->set_dependencies(new Dependencies(env())); 958 959 _fixed_slots = 0; 960 set_has_split_ifs(false); 961 set_has_loops(has_method() && method()->has_loops()); // first approximation 962 set_has_stringbuilder(false); 963 _trap_can_recompile = false; // no traps emitted yet 964 _major_progress = true; // start out assuming good things will happen 965 set_has_unsafe_access(false); 966 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 967 set_decompile_count(0); 968 969 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 970 set_num_loop_opts(LoopOptsCount); 971 set_do_inlining(Inline); 972 set_max_inline_size(MaxInlineSize); 973 set_freq_inline_size(FreqInlineSize); 974 set_do_scheduling(OptoScheduling); 975 set_do_count_invocations(false); 976 set_do_method_data_update(false); 977 978 if (debug_info()->recording_non_safepoints()) { 979 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 980 (comp_arena(), 8, 0, NULL)); 981 set_default_node_notes(Node_Notes::make(this)); 982 } 983 984 // // -- Initialize types before each compile -- 985 // // Update cached type information 986 // if( _method && _method->constants() ) 987 // Type::update_loaded_types(_method, _method->constants()); 988 989 // Init alias_type map. 990 if (!_do_escape_analysis && aliaslevel == 3) 991 aliaslevel = 2; // No unique types without escape analysis 992 _AliasLevel = aliaslevel; 993 const int grow_ats = 16; 994 _max_alias_types = grow_ats; 995 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 996 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 997 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 998 { 999 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1000 } 1001 // Initialize the first few types. 1002 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1003 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1004 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1005 _num_alias_types = AliasIdxRaw+1; 1006 // Zero out the alias type cache. 1007 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1008 // A NULL adr_type hits in the cache right away. Preload the right answer. 1009 probe_alias_cache(NULL)->_index = AliasIdxTop; 1010 1011 _intrinsics = NULL; 1012 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1013 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1014 register_library_intrinsics(); 1015} 1016 1017//---------------------------init_start---------------------------------------- 1018// Install the StartNode on this compile object. 1019void Compile::init_start(StartNode* s) { 1020 if (failing()) 1021 return; // already failing 1022 assert(s == start(), ""); 1023} 1024 1025StartNode* Compile::start() const { 1026 assert(!failing(), ""); 1027 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1028 Node* start = root()->fast_out(i); 1029 if( start->is_Start() ) 1030 return start->as_Start(); 1031 } 1032 ShouldNotReachHere(); 1033 return NULL; 1034} 1035 1036//-------------------------------immutable_memory------------------------------------- 1037// Access immutable memory 1038Node* Compile::immutable_memory() { 1039 if (_immutable_memory != NULL) { 1040 return _immutable_memory; 1041 } 1042 StartNode* s = start(); 1043 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1044 Node *p = s->fast_out(i); 1045 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1046 _immutable_memory = p; 1047 return _immutable_memory; 1048 } 1049 } 1050 ShouldNotReachHere(); 1051 return NULL; 1052} 1053 1054//----------------------set_cached_top_node------------------------------------ 1055// Install the cached top node, and make sure Node::is_top works correctly. 1056void Compile::set_cached_top_node(Node* tn) { 1057 if (tn != NULL) verify_top(tn); 1058 Node* old_top = _top; 1059 _top = tn; 1060 // Calling Node::setup_is_top allows the nodes the chance to adjust 1061 // their _out arrays. 1062 if (_top != NULL) _top->setup_is_top(); 1063 if (old_top != NULL) old_top->setup_is_top(); 1064 assert(_top == NULL || top()->is_top(), ""); 1065} 1066 1067#ifndef PRODUCT 1068void Compile::verify_top(Node* tn) const { 1069 if (tn != NULL) { 1070 assert(tn->is_Con(), "top node must be a constant"); 1071 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1072 assert(tn->in(0) != NULL, "must have live top node"); 1073 } 1074} 1075#endif 1076 1077 1078///-------------------Managing Per-Node Debug & Profile Info------------------- 1079 1080void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1081 guarantee(arr != NULL, ""); 1082 int num_blocks = arr->length(); 1083 if (grow_by < num_blocks) grow_by = num_blocks; 1084 int num_notes = grow_by * _node_notes_block_size; 1085 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1086 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1087 while (num_notes > 0) { 1088 arr->append(notes); 1089 notes += _node_notes_block_size; 1090 num_notes -= _node_notes_block_size; 1091 } 1092 assert(num_notes == 0, "exact multiple, please"); 1093} 1094 1095bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1096 if (source == NULL || dest == NULL) return false; 1097 1098 if (dest->is_Con()) 1099 return false; // Do not push debug info onto constants. 1100 1101#ifdef ASSERT 1102 // Leave a bread crumb trail pointing to the original node: 1103 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1104 dest->set_debug_orig(source); 1105 } 1106#endif 1107 1108 if (node_note_array() == NULL) 1109 return false; // Not collecting any notes now. 1110 1111 // This is a copy onto a pre-existing node, which may already have notes. 1112 // If both nodes have notes, do not overwrite any pre-existing notes. 1113 Node_Notes* source_notes = node_notes_at(source->_idx); 1114 if (source_notes == NULL || source_notes->is_clear()) return false; 1115 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1116 if (dest_notes == NULL || dest_notes->is_clear()) { 1117 return set_node_notes_at(dest->_idx, source_notes); 1118 } 1119 1120 Node_Notes merged_notes = (*source_notes); 1121 // The order of operations here ensures that dest notes will win... 1122 merged_notes.update_from(dest_notes); 1123 return set_node_notes_at(dest->_idx, &merged_notes); 1124} 1125 1126 1127//--------------------------allow_range_check_smearing------------------------- 1128// Gating condition for coalescing similar range checks. 1129// Sometimes we try 'speculatively' replacing a series of a range checks by a 1130// single covering check that is at least as strong as any of them. 1131// If the optimization succeeds, the simplified (strengthened) range check 1132// will always succeed. If it fails, we will deopt, and then give up 1133// on the optimization. 1134bool Compile::allow_range_check_smearing() const { 1135 // If this method has already thrown a range-check, 1136 // assume it was because we already tried range smearing 1137 // and it failed. 1138 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1139 return !already_trapped; 1140} 1141 1142 1143//------------------------------flatten_alias_type----------------------------- 1144const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1145 int offset = tj->offset(); 1146 TypePtr::PTR ptr = tj->ptr(); 1147 1148 // Known instance (scalarizable allocation) alias only with itself. 1149 bool is_known_inst = tj->isa_oopptr() != NULL && 1150 tj->is_oopptr()->is_known_instance(); 1151 1152 // Process weird unsafe references. 1153 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1154 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1155 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1156 tj = TypeOopPtr::BOTTOM; 1157 ptr = tj->ptr(); 1158 offset = tj->offset(); 1159 } 1160 1161 // Array pointers need some flattening 1162 const TypeAryPtr *ta = tj->isa_aryptr(); 1163 if( ta && is_known_inst ) { 1164 if ( offset != Type::OffsetBot && 1165 offset > arrayOopDesc::length_offset_in_bytes() ) { 1166 offset = Type::OffsetBot; // Flatten constant access into array body only 1167 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1168 } 1169 } else if( ta && _AliasLevel >= 2 ) { 1170 // For arrays indexed by constant indices, we flatten the alias 1171 // space to include all of the array body. Only the header, klass 1172 // and array length can be accessed un-aliased. 1173 if( offset != Type::OffsetBot ) { 1174 if( ta->const_oop() ) { // MethodData* or Method* 1175 offset = Type::OffsetBot; // Flatten constant access into array body 1176 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1177 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1178 // range is OK as-is. 1179 tj = ta = TypeAryPtr::RANGE; 1180 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1181 tj = TypeInstPtr::KLASS; // all klass loads look alike 1182 ta = TypeAryPtr::RANGE; // generic ignored junk 1183 ptr = TypePtr::BotPTR; 1184 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1185 tj = TypeInstPtr::MARK; 1186 ta = TypeAryPtr::RANGE; // generic ignored junk 1187 ptr = TypePtr::BotPTR; 1188 } else { // Random constant offset into array body 1189 offset = Type::OffsetBot; // Flatten constant access into array body 1190 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1191 } 1192 } 1193 // Arrays of fixed size alias with arrays of unknown size. 1194 if (ta->size() != TypeInt::POS) { 1195 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1196 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1197 } 1198 // Arrays of known objects become arrays of unknown objects. 1199 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1200 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1201 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1202 } 1203 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1204 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1205 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1206 } 1207 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1208 // cannot be distinguished by bytecode alone. 1209 if (ta->elem() == TypeInt::BOOL) { 1210 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1211 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1212 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1213 } 1214 // During the 2nd round of IterGVN, NotNull castings are removed. 1215 // Make sure the Bottom and NotNull variants alias the same. 1216 // Also, make sure exact and non-exact variants alias the same. 1217 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1218 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1219 } 1220 } 1221 1222 // Oop pointers need some flattening 1223 const TypeInstPtr *to = tj->isa_instptr(); 1224 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1225 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1226 if( ptr == TypePtr::Constant ) { 1227 if (to->klass() != ciEnv::current()->Class_klass() || 1228 offset < k->size_helper() * wordSize) { 1229 // No constant oop pointers (such as Strings); they alias with 1230 // unknown strings. 1231 assert(!is_known_inst, "not scalarizable allocation"); 1232 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1233 } 1234 } else if( is_known_inst ) { 1235 tj = to; // Keep NotNull and klass_is_exact for instance type 1236 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1237 // During the 2nd round of IterGVN, NotNull castings are removed. 1238 // Make sure the Bottom and NotNull variants alias the same. 1239 // Also, make sure exact and non-exact variants alias the same. 1240 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1241 } 1242 // Canonicalize the holder of this field 1243 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1244 // First handle header references such as a LoadKlassNode, even if the 1245 // object's klass is unloaded at compile time (4965979). 1246 if (!is_known_inst) { // Do it only for non-instance types 1247 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1248 } 1249 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1250 // Static fields are in the space above the normal instance 1251 // fields in the java.lang.Class instance. 1252 if (to->klass() != ciEnv::current()->Class_klass()) { 1253 to = NULL; 1254 tj = TypeOopPtr::BOTTOM; 1255 offset = tj->offset(); 1256 } 1257 } else { 1258 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1259 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1260 if( is_known_inst ) { 1261 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1262 } else { 1263 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1264 } 1265 } 1266 } 1267 } 1268 1269 // Klass pointers to object array klasses need some flattening 1270 const TypeKlassPtr *tk = tj->isa_klassptr(); 1271 if( tk ) { 1272 // If we are referencing a field within a Klass, we need 1273 // to assume the worst case of an Object. Both exact and 1274 // inexact types must flatten to the same alias class so 1275 // use NotNull as the PTR. 1276 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1277 1278 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1279 TypeKlassPtr::OBJECT->klass(), 1280 offset); 1281 } 1282 1283 ciKlass* klass = tk->klass(); 1284 if( klass->is_obj_array_klass() ) { 1285 ciKlass* k = TypeAryPtr::OOPS->klass(); 1286 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1287 k = TypeInstPtr::BOTTOM->klass(); 1288 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1289 } 1290 1291 // Check for precise loads from the primary supertype array and force them 1292 // to the supertype cache alias index. Check for generic array loads from 1293 // the primary supertype array and also force them to the supertype cache 1294 // alias index. Since the same load can reach both, we need to merge 1295 // these 2 disparate memories into the same alias class. Since the 1296 // primary supertype array is read-only, there's no chance of confusion 1297 // where we bypass an array load and an array store. 1298 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1299 if (offset == Type::OffsetBot || 1300 (offset >= primary_supers_offset && 1301 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1302 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1303 offset = in_bytes(Klass::secondary_super_cache_offset()); 1304 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1305 } 1306 } 1307 1308 // Flatten all Raw pointers together. 1309 if (tj->base() == Type::RawPtr) 1310 tj = TypeRawPtr::BOTTOM; 1311 1312 if (tj->base() == Type::AnyPtr) 1313 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1314 1315 // Flatten all to bottom for now 1316 switch( _AliasLevel ) { 1317 case 0: 1318 tj = TypePtr::BOTTOM; 1319 break; 1320 case 1: // Flatten to: oop, static, field or array 1321 switch (tj->base()) { 1322 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1323 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1324 case Type::AryPtr: // do not distinguish arrays at all 1325 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1326 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1327 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1328 default: ShouldNotReachHere(); 1329 } 1330 break; 1331 case 2: // No collapsing at level 2; keep all splits 1332 case 3: // No collapsing at level 3; keep all splits 1333 break; 1334 default: 1335 Unimplemented(); 1336 } 1337 1338 offset = tj->offset(); 1339 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1340 1341 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1342 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1343 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1344 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1345 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1346 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1347 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1348 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1349 assert( tj->ptr() != TypePtr::TopPTR && 1350 tj->ptr() != TypePtr::AnyNull && 1351 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1352// assert( tj->ptr() != TypePtr::Constant || 1353// tj->base() == Type::RawPtr || 1354// tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1355 1356 return tj; 1357} 1358 1359void Compile::AliasType::Init(int i, const TypePtr* at) { 1360 _index = i; 1361 _adr_type = at; 1362 _field = NULL; 1363 _is_rewritable = true; // default 1364 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1365 if (atoop != NULL && atoop->is_known_instance()) { 1366 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1367 _general_index = Compile::current()->get_alias_index(gt); 1368 } else { 1369 _general_index = 0; 1370 } 1371} 1372 1373//---------------------------------print_on------------------------------------ 1374#ifndef PRODUCT 1375void Compile::AliasType::print_on(outputStream* st) { 1376 if (index() < 10) 1377 st->print("@ <%d> ", index()); 1378 else st->print("@ <%d>", index()); 1379 st->print(is_rewritable() ? " " : " RO"); 1380 int offset = adr_type()->offset(); 1381 if (offset == Type::OffsetBot) 1382 st->print(" +any"); 1383 else st->print(" +%-3d", offset); 1384 st->print(" in "); 1385 adr_type()->dump_on(st); 1386 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1387 if (field() != NULL && tjp) { 1388 if (tjp->klass() != field()->holder() || 1389 tjp->offset() != field()->offset_in_bytes()) { 1390 st->print(" != "); 1391 field()->print(); 1392 st->print(" ***"); 1393 } 1394 } 1395} 1396 1397void print_alias_types() { 1398 Compile* C = Compile::current(); 1399 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1400 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1401 C->alias_type(idx)->print_on(tty); 1402 tty->cr(); 1403 } 1404} 1405#endif 1406 1407 1408//----------------------------probe_alias_cache-------------------------------- 1409Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1410 intptr_t key = (intptr_t) adr_type; 1411 key ^= key >> logAliasCacheSize; 1412 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1413} 1414 1415 1416//-----------------------------grow_alias_types-------------------------------- 1417void Compile::grow_alias_types() { 1418 const int old_ats = _max_alias_types; // how many before? 1419 const int new_ats = old_ats; // how many more? 1420 const int grow_ats = old_ats+new_ats; // how many now? 1421 _max_alias_types = grow_ats; 1422 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1423 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1424 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1425 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1426} 1427 1428 1429//--------------------------------find_alias_type------------------------------ 1430Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1431 if (_AliasLevel == 0) 1432 return alias_type(AliasIdxBot); 1433 1434 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1435 if (ace->_adr_type == adr_type) { 1436 return alias_type(ace->_index); 1437 } 1438 1439 // Handle special cases. 1440 if (adr_type == NULL) return alias_type(AliasIdxTop); 1441 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1442 1443 // Do it the slow way. 1444 const TypePtr* flat = flatten_alias_type(adr_type); 1445 1446#ifdef ASSERT 1447 assert(flat == flatten_alias_type(flat), "idempotent"); 1448 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1449 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1450 const TypeOopPtr* foop = flat->is_oopptr(); 1451 // Scalarizable allocations have exact klass always. 1452 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1453 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1454 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1455 } 1456 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1457#endif 1458 1459 int idx = AliasIdxTop; 1460 for (int i = 0; i < num_alias_types(); i++) { 1461 if (alias_type(i)->adr_type() == flat) { 1462 idx = i; 1463 break; 1464 } 1465 } 1466 1467 if (idx == AliasIdxTop) { 1468 if (no_create) return NULL; 1469 // Grow the array if necessary. 1470 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1471 // Add a new alias type. 1472 idx = _num_alias_types++; 1473 _alias_types[idx]->Init(idx, flat); 1474 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1475 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1476 if (flat->isa_instptr()) { 1477 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1478 && flat->is_instptr()->klass() == env()->Class_klass()) 1479 alias_type(idx)->set_rewritable(false); 1480 } 1481 if (flat->isa_klassptr()) { 1482 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1483 alias_type(idx)->set_rewritable(false); 1484 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1485 alias_type(idx)->set_rewritable(false); 1486 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1487 alias_type(idx)->set_rewritable(false); 1488 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1489 alias_type(idx)->set_rewritable(false); 1490 } 1491 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1492 // but the base pointer type is not distinctive enough to identify 1493 // references into JavaThread.) 1494 1495 // Check for final fields. 1496 const TypeInstPtr* tinst = flat->isa_instptr(); 1497 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1498 ciField* field; 1499 if (tinst->const_oop() != NULL && 1500 tinst->klass() == ciEnv::current()->Class_klass() && 1501 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1502 // static field 1503 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1504 field = k->get_field_by_offset(tinst->offset(), true); 1505 } else { 1506 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1507 field = k->get_field_by_offset(tinst->offset(), false); 1508 } 1509 assert(field == NULL || 1510 original_field == NULL || 1511 (field->holder() == original_field->holder() && 1512 field->offset() == original_field->offset() && 1513 field->is_static() == original_field->is_static()), "wrong field?"); 1514 // Set field() and is_rewritable() attributes. 1515 if (field != NULL) alias_type(idx)->set_field(field); 1516 } 1517 } 1518 1519 // Fill the cache for next time. 1520 ace->_adr_type = adr_type; 1521 ace->_index = idx; 1522 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1523 1524 // Might as well try to fill the cache for the flattened version, too. 1525 AliasCacheEntry* face = probe_alias_cache(flat); 1526 if (face->_adr_type == NULL) { 1527 face->_adr_type = flat; 1528 face->_index = idx; 1529 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1530 } 1531 1532 return alias_type(idx); 1533} 1534 1535 1536Compile::AliasType* Compile::alias_type(ciField* field) { 1537 const TypeOopPtr* t; 1538 if (field->is_static()) 1539 t = TypeInstPtr::make(field->holder()->java_mirror()); 1540 else 1541 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1542 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1543 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1544 return atp; 1545} 1546 1547 1548//------------------------------have_alias_type-------------------------------- 1549bool Compile::have_alias_type(const TypePtr* adr_type) { 1550 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1551 if (ace->_adr_type == adr_type) { 1552 return true; 1553 } 1554 1555 // Handle special cases. 1556 if (adr_type == NULL) return true; 1557 if (adr_type == TypePtr::BOTTOM) return true; 1558 1559 return find_alias_type(adr_type, true, NULL) != NULL; 1560} 1561 1562//-----------------------------must_alias-------------------------------------- 1563// True if all values of the given address type are in the given alias category. 1564bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1565 if (alias_idx == AliasIdxBot) return true; // the universal category 1566 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1567 if (alias_idx == AliasIdxTop) return false; // the empty category 1568 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1569 1570 // the only remaining possible overlap is identity 1571 int adr_idx = get_alias_index(adr_type); 1572 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1573 assert(adr_idx == alias_idx || 1574 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1575 && adr_type != TypeOopPtr::BOTTOM), 1576 "should not be testing for overlap with an unsafe pointer"); 1577 return adr_idx == alias_idx; 1578} 1579 1580//------------------------------can_alias-------------------------------------- 1581// True if any values of the given address type are in the given alias category. 1582bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1583 if (alias_idx == AliasIdxTop) return false; // the empty category 1584 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1585 if (alias_idx == AliasIdxBot) return true; // the universal category 1586 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1587 1588 // the only remaining possible overlap is identity 1589 int adr_idx = get_alias_index(adr_type); 1590 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1591 return adr_idx == alias_idx; 1592} 1593 1594 1595 1596//---------------------------pop_warm_call------------------------------------- 1597WarmCallInfo* Compile::pop_warm_call() { 1598 WarmCallInfo* wci = _warm_calls; 1599 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1600 return wci; 1601} 1602 1603//----------------------------Inline_Warm-------------------------------------- 1604int Compile::Inline_Warm() { 1605 // If there is room, try to inline some more warm call sites. 1606 // %%% Do a graph index compaction pass when we think we're out of space? 1607 if (!InlineWarmCalls) return 0; 1608 1609 int calls_made_hot = 0; 1610 int room_to_grow = NodeCountInliningCutoff - unique(); 1611 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1612 int amount_grown = 0; 1613 WarmCallInfo* call; 1614 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1615 int est_size = (int)call->size(); 1616 if (est_size > (room_to_grow - amount_grown)) { 1617 // This one won't fit anyway. Get rid of it. 1618 call->make_cold(); 1619 continue; 1620 } 1621 call->make_hot(); 1622 calls_made_hot++; 1623 amount_grown += est_size; 1624 amount_to_grow -= est_size; 1625 } 1626 1627 if (calls_made_hot > 0) set_major_progress(); 1628 return calls_made_hot; 1629} 1630 1631 1632//----------------------------Finish_Warm-------------------------------------- 1633void Compile::Finish_Warm() { 1634 if (!InlineWarmCalls) return; 1635 if (failing()) return; 1636 if (warm_calls() == NULL) return; 1637 1638 // Clean up loose ends, if we are out of space for inlining. 1639 WarmCallInfo* call; 1640 while ((call = pop_warm_call()) != NULL) { 1641 call->make_cold(); 1642 } 1643} 1644 1645//---------------------cleanup_loop_predicates----------------------- 1646// Remove the opaque nodes that protect the predicates so that all unused 1647// checks and uncommon_traps will be eliminated from the ideal graph 1648void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1649 if (predicate_count()==0) return; 1650 for (int i = predicate_count(); i > 0; i--) { 1651 Node * n = predicate_opaque1_node(i-1); 1652 assert(n->Opcode() == Op_Opaque1, "must be"); 1653 igvn.replace_node(n, n->in(1)); 1654 } 1655 assert(predicate_count()==0, "should be clean!"); 1656} 1657 1658//------------------------------Optimize--------------------------------------- 1659// Given a graph, optimize it. 1660void Compile::Optimize() { 1661 TracePhase t1("optimizer", &_t_optimizer, true); 1662 1663#ifndef PRODUCT 1664 if (env()->break_at_compile()) { 1665 BREAKPOINT; 1666 } 1667 1668#endif 1669 1670 ResourceMark rm; 1671 int loop_opts_cnt; 1672 1673 NOT_PRODUCT( verify_graph_edges(); ) 1674 1675 print_method("After Parsing"); 1676 1677 { 1678 // Iterative Global Value Numbering, including ideal transforms 1679 // Initialize IterGVN with types and values from parse-time GVN 1680 PhaseIterGVN igvn(initial_gvn()); 1681 { 1682 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1683 igvn.optimize(); 1684 } 1685 1686 print_method("Iter GVN 1", 2); 1687 1688 if (failing()) return; 1689 1690 // Perform escape analysis 1691 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1692 if (has_loops()) { 1693 // Cleanup graph (remove dead nodes). 1694 TracePhase t2("idealLoop", &_t_idealLoop, true); 1695 PhaseIdealLoop ideal_loop( igvn, false, true ); 1696 if (major_progress()) print_method("PhaseIdealLoop before EA", 2); 1697 if (failing()) return; 1698 } 1699 ConnectionGraph::do_analysis(this, &igvn); 1700 1701 if (failing()) return; 1702 1703 // Optimize out fields loads from scalar replaceable allocations. 1704 igvn.optimize(); 1705 print_method("Iter GVN after EA", 2); 1706 1707 if (failing()) return; 1708 1709 if (congraph() != NULL && macro_count() > 0) { 1710 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 1711 PhaseMacroExpand mexp(igvn); 1712 mexp.eliminate_macro_nodes(); 1713 igvn.set_delay_transform(false); 1714 1715 igvn.optimize(); 1716 print_method("Iter GVN after eliminating allocations and locks", 2); 1717 1718 if (failing()) return; 1719 } 1720 } 1721 1722 // Loop transforms on the ideal graph. Range Check Elimination, 1723 // peeling, unrolling, etc. 1724 1725 // Set loop opts counter 1726 loop_opts_cnt = num_loop_opts(); 1727 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1728 { 1729 TracePhase t2("idealLoop", &_t_idealLoop, true); 1730 PhaseIdealLoop ideal_loop( igvn, true ); 1731 loop_opts_cnt--; 1732 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1733 if (failing()) return; 1734 } 1735 // Loop opts pass if partial peeling occurred in previous pass 1736 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1737 TracePhase t3("idealLoop", &_t_idealLoop, true); 1738 PhaseIdealLoop ideal_loop( igvn, false ); 1739 loop_opts_cnt--; 1740 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1741 if (failing()) return; 1742 } 1743 // Loop opts pass for loop-unrolling before CCP 1744 if(major_progress() && (loop_opts_cnt > 0)) { 1745 TracePhase t4("idealLoop", &_t_idealLoop, true); 1746 PhaseIdealLoop ideal_loop( igvn, false ); 1747 loop_opts_cnt--; 1748 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1749 } 1750 if (!failing()) { 1751 // Verify that last round of loop opts produced a valid graph 1752 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1753 PhaseIdealLoop::verify(igvn); 1754 } 1755 } 1756 if (failing()) return; 1757 1758 // Conditional Constant Propagation; 1759 PhaseCCP ccp( &igvn ); 1760 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1761 { 1762 TracePhase t2("ccp", &_t_ccp, true); 1763 ccp.do_transform(); 1764 } 1765 print_method("PhaseCPP 1", 2); 1766 1767 assert( true, "Break here to ccp.dump_old2new_map()"); 1768 1769 // Iterative Global Value Numbering, including ideal transforms 1770 { 1771 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1772 igvn = ccp; 1773 igvn.optimize(); 1774 } 1775 1776 print_method("Iter GVN 2", 2); 1777 1778 if (failing()) return; 1779 1780 // Loop transforms on the ideal graph. Range Check Elimination, 1781 // peeling, unrolling, etc. 1782 if(loop_opts_cnt > 0) { 1783 debug_only( int cnt = 0; ); 1784 while(major_progress() && (loop_opts_cnt > 0)) { 1785 TracePhase t2("idealLoop", &_t_idealLoop, true); 1786 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1787 PhaseIdealLoop ideal_loop( igvn, true); 1788 loop_opts_cnt--; 1789 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1790 if (failing()) return; 1791 } 1792 } 1793 1794 { 1795 // Verify that all previous optimizations produced a valid graph 1796 // at least to this point, even if no loop optimizations were done. 1797 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1798 PhaseIdealLoop::verify(igvn); 1799 } 1800 1801 { 1802 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1803 PhaseMacroExpand mex(igvn); 1804 if (mex.expand_macro_nodes()) { 1805 assert(failing(), "must bail out w/ explicit message"); 1806 return; 1807 } 1808 } 1809 1810 } // (End scope of igvn; run destructor if necessary for asserts.) 1811 1812 // A method with only infinite loops has no edges entering loops from root 1813 { 1814 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1815 if (final_graph_reshaping()) { 1816 assert(failing(), "must bail out w/ explicit message"); 1817 return; 1818 } 1819 } 1820 1821 print_method("Optimize finished", 2); 1822} 1823 1824 1825//------------------------------Code_Gen--------------------------------------- 1826// Given a graph, generate code for it 1827void Compile::Code_Gen() { 1828 if (failing()) return; 1829 1830 // Perform instruction selection. You might think we could reclaim Matcher 1831 // memory PDQ, but actually the Matcher is used in generating spill code. 1832 // Internals of the Matcher (including some VectorSets) must remain live 1833 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1834 // set a bit in reclaimed memory. 1835 1836 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1837 // nodes. Mapping is only valid at the root of each matched subtree. 1838 NOT_PRODUCT( verify_graph_edges(); ) 1839 1840 Node_List proj_list; 1841 Matcher m(proj_list); 1842 _matcher = &m; 1843 { 1844 TracePhase t2("matcher", &_t_matcher, true); 1845 m.match(); 1846 } 1847 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1848 // nodes. Mapping is only valid at the root of each matched subtree. 1849 NOT_PRODUCT( verify_graph_edges(); ) 1850 1851 // If you have too many nodes, or if matching has failed, bail out 1852 check_node_count(0, "out of nodes matching instructions"); 1853 if (failing()) return; 1854 1855 // Build a proper-looking CFG 1856 PhaseCFG cfg(node_arena(), root(), m); 1857 _cfg = &cfg; 1858 { 1859 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1860 cfg.Dominators(); 1861 if (failing()) return; 1862 1863 NOT_PRODUCT( verify_graph_edges(); ) 1864 1865 cfg.Estimate_Block_Frequency(); 1866 cfg.GlobalCodeMotion(m,unique(),proj_list); 1867 if (failing()) return; 1868 1869 print_method("Global code motion", 2); 1870 1871 NOT_PRODUCT( verify_graph_edges(); ) 1872 1873 debug_only( cfg.verify(); ) 1874 } 1875 NOT_PRODUCT( verify_graph_edges(); ) 1876 1877 PhaseChaitin regalloc(unique(),cfg,m); 1878 _regalloc = ®alloc; 1879 { 1880 TracePhase t2("regalloc", &_t_registerAllocation, true); 1881 // Perform any platform dependent preallocation actions. This is used, 1882 // for example, to avoid taking an implicit null pointer exception 1883 // using the frame pointer on win95. 1884 _regalloc->pd_preallocate_hook(); 1885 1886 // Perform register allocation. After Chaitin, use-def chains are 1887 // no longer accurate (at spill code) and so must be ignored. 1888 // Node->LRG->reg mappings are still accurate. 1889 _regalloc->Register_Allocate(); 1890 1891 // Bail out if the allocator builds too many nodes 1892 if (failing()) return; 1893 } 1894 1895 // Prior to register allocation we kept empty basic blocks in case the 1896 // the allocator needed a place to spill. After register allocation we 1897 // are not adding any new instructions. If any basic block is empty, we 1898 // can now safely remove it. 1899 { 1900 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1901 cfg.remove_empty(); 1902 if (do_freq_based_layout()) { 1903 PhaseBlockLayout layout(cfg); 1904 } else { 1905 cfg.set_loop_alignment(); 1906 } 1907 cfg.fixup_flow(); 1908 } 1909 1910 // Perform any platform dependent postallocation verifications. 1911 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1912 1913 // Apply peephole optimizations 1914 if( OptoPeephole ) { 1915 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1916 PhasePeephole peep( _regalloc, cfg); 1917 peep.do_transform(); 1918 } 1919 1920 // Convert Nodes to instruction bits in a buffer 1921 { 1922 // %%%% workspace merge brought two timers together for one job 1923 TracePhase t2a("output", &_t_output, true); 1924 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1925 Output(); 1926 } 1927 1928 print_method("Final Code"); 1929 1930 // He's dead, Jim. 1931 _cfg = (PhaseCFG*)0xdeadbeef; 1932 _regalloc = (PhaseChaitin*)0xdeadbeef; 1933} 1934 1935 1936//------------------------------dump_asm--------------------------------------- 1937// Dump formatted assembly 1938#ifndef PRODUCT 1939void Compile::dump_asm(int *pcs, uint pc_limit) { 1940 bool cut_short = false; 1941 tty->print_cr("#"); 1942 tty->print("# "); _tf->dump(); tty->cr(); 1943 tty->print_cr("#"); 1944 1945 // For all blocks 1946 int pc = 0x0; // Program counter 1947 char starts_bundle = ' '; 1948 _regalloc->dump_frame(); 1949 1950 Node *n = NULL; 1951 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1952 if (VMThread::should_terminate()) { cut_short = true; break; } 1953 Block *b = _cfg->_blocks[i]; 1954 if (b->is_connector() && !Verbose) continue; 1955 n = b->_nodes[0]; 1956 if (pcs && n->_idx < pc_limit) 1957 tty->print("%3.3x ", pcs[n->_idx]); 1958 else 1959 tty->print(" "); 1960 b->dump_head( &_cfg->_bbs ); 1961 if (b->is_connector()) { 1962 tty->print_cr(" # Empty connector block"); 1963 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1964 tty->print_cr(" # Block is sole successor of call"); 1965 } 1966 1967 // For all instructions 1968 Node *delay = NULL; 1969 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1970 if (VMThread::should_terminate()) { cut_short = true; break; } 1971 n = b->_nodes[j]; 1972 if (valid_bundle_info(n)) { 1973 Bundle *bundle = node_bundling(n); 1974 if (bundle->used_in_unconditional_delay()) { 1975 delay = n; 1976 continue; 1977 } 1978 if (bundle->starts_bundle()) 1979 starts_bundle = '+'; 1980 } 1981 1982 if (WizardMode) n->dump(); 1983 1984 if( !n->is_Region() && // Dont print in the Assembly 1985 !n->is_Phi() && // a few noisely useless nodes 1986 !n->is_Proj() && 1987 !n->is_MachTemp() && 1988 !n->is_SafePointScalarObject() && 1989 !n->is_Catch() && // Would be nice to print exception table targets 1990 !n->is_MergeMem() && // Not very interesting 1991 !n->is_top() && // Debug info table constants 1992 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1993 ) { 1994 if (pcs && n->_idx < pc_limit) 1995 tty->print("%3.3x", pcs[n->_idx]); 1996 else 1997 tty->print(" "); 1998 tty->print(" %c ", starts_bundle); 1999 starts_bundle = ' '; 2000 tty->print("\t"); 2001 n->format(_regalloc, tty); 2002 tty->cr(); 2003 } 2004 2005 // If we have an instruction with a delay slot, and have seen a delay, 2006 // then back up and print it 2007 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2008 assert(delay != NULL, "no unconditional delay instruction"); 2009 if (WizardMode) delay->dump(); 2010 2011 if (node_bundling(delay)->starts_bundle()) 2012 starts_bundle = '+'; 2013 if (pcs && n->_idx < pc_limit) 2014 tty->print("%3.3x", pcs[n->_idx]); 2015 else 2016 tty->print(" "); 2017 tty->print(" %c ", starts_bundle); 2018 starts_bundle = ' '; 2019 tty->print("\t"); 2020 delay->format(_regalloc, tty); 2021 tty->print_cr(""); 2022 delay = NULL; 2023 } 2024 2025 // Dump the exception table as well 2026 if( n->is_Catch() && (Verbose || WizardMode) ) { 2027 // Print the exception table for this offset 2028 _handler_table.print_subtable_for(pc); 2029 } 2030 } 2031 2032 if (pcs && n->_idx < pc_limit) 2033 tty->print_cr("%3.3x", pcs[n->_idx]); 2034 else 2035 tty->print_cr(""); 2036 2037 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2038 2039 } // End of per-block dump 2040 tty->print_cr(""); 2041 2042 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2043} 2044#endif 2045 2046//------------------------------Final_Reshape_Counts--------------------------- 2047// This class defines counters to help identify when a method 2048// may/must be executed using hardware with only 24-bit precision. 2049struct Final_Reshape_Counts : public StackObj { 2050 int _call_count; // count non-inlined 'common' calls 2051 int _float_count; // count float ops requiring 24-bit precision 2052 int _double_count; // count double ops requiring more precision 2053 int _java_call_count; // count non-inlined 'java' calls 2054 int _inner_loop_count; // count loops which need alignment 2055 VectorSet _visited; // Visitation flags 2056 Node_List _tests; // Set of IfNodes & PCTableNodes 2057 2058 Final_Reshape_Counts() : 2059 _call_count(0), _float_count(0), _double_count(0), 2060 _java_call_count(0), _inner_loop_count(0), 2061 _visited( Thread::current()->resource_area() ) { } 2062 2063 void inc_call_count () { _call_count ++; } 2064 void inc_float_count () { _float_count ++; } 2065 void inc_double_count() { _double_count++; } 2066 void inc_java_call_count() { _java_call_count++; } 2067 void inc_inner_loop_count() { _inner_loop_count++; } 2068 2069 int get_call_count () const { return _call_count ; } 2070 int get_float_count () const { return _float_count ; } 2071 int get_double_count() const { return _double_count; } 2072 int get_java_call_count() const { return _java_call_count; } 2073 int get_inner_loop_count() const { return _inner_loop_count; } 2074}; 2075 2076static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2077 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2078 // Make sure the offset goes inside the instance layout. 2079 return k->contains_field_offset(tp->offset()); 2080 // Note that OffsetBot and OffsetTop are very negative. 2081} 2082 2083// Eliminate trivially redundant StoreCMs and accumulate their 2084// precedence edges. 2085static void eliminate_redundant_card_marks(Node* n) { 2086 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2087 if (n->in(MemNode::Address)->outcnt() > 1) { 2088 // There are multiple users of the same address so it might be 2089 // possible to eliminate some of the StoreCMs 2090 Node* mem = n->in(MemNode::Memory); 2091 Node* adr = n->in(MemNode::Address); 2092 Node* val = n->in(MemNode::ValueIn); 2093 Node* prev = n; 2094 bool done = false; 2095 // Walk the chain of StoreCMs eliminating ones that match. As 2096 // long as it's a chain of single users then the optimization is 2097 // safe. Eliminating partially redundant StoreCMs would require 2098 // cloning copies down the other paths. 2099 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2100 if (adr == mem->in(MemNode::Address) && 2101 val == mem->in(MemNode::ValueIn)) { 2102 // redundant StoreCM 2103 if (mem->req() > MemNode::OopStore) { 2104 // Hasn't been processed by this code yet. 2105 n->add_prec(mem->in(MemNode::OopStore)); 2106 } else { 2107 // Already converted to precedence edge 2108 for (uint i = mem->req(); i < mem->len(); i++) { 2109 // Accumulate any precedence edges 2110 if (mem->in(i) != NULL) { 2111 n->add_prec(mem->in(i)); 2112 } 2113 } 2114 // Everything above this point has been processed. 2115 done = true; 2116 } 2117 // Eliminate the previous StoreCM 2118 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2119 assert(mem->outcnt() == 0, "should be dead"); 2120 mem->disconnect_inputs(NULL); 2121 } else { 2122 prev = mem; 2123 } 2124 mem = prev->in(MemNode::Memory); 2125 } 2126 } 2127} 2128 2129//------------------------------final_graph_reshaping_impl---------------------- 2130// Implement items 1-5 from final_graph_reshaping below. 2131static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 2132 2133 if ( n->outcnt() == 0 ) return; // dead node 2134 uint nop = n->Opcode(); 2135 2136 // Check for 2-input instruction with "last use" on right input. 2137 // Swap to left input. Implements item (2). 2138 if( n->req() == 3 && // two-input instruction 2139 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2140 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2141 n->in(2)->outcnt() == 1 &&// right use IS a last use 2142 !n->in(2)->is_Con() ) { // right use is not a constant 2143 // Check for commutative opcode 2144 switch( nop ) { 2145 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2146 case Op_MaxI: case Op_MinI: 2147 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2148 case Op_AndL: case Op_XorL: case Op_OrL: 2149 case Op_AndI: case Op_XorI: case Op_OrI: { 2150 // Move "last use" input to left by swapping inputs 2151 n->swap_edges(1, 2); 2152 break; 2153 } 2154 default: 2155 break; 2156 } 2157 } 2158 2159#ifdef ASSERT 2160 if( n->is_Mem() ) { 2161 Compile* C = Compile::current(); 2162 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2163 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2164 // oop will be recorded in oop map if load crosses safepoint 2165 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2166 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2167 "raw memory operations should have control edge"); 2168 } 2169#endif 2170 // Count FPU ops and common calls, implements item (3) 2171 switch( nop ) { 2172 // Count all float operations that may use FPU 2173 case Op_AddF: 2174 case Op_SubF: 2175 case Op_MulF: 2176 case Op_DivF: 2177 case Op_NegF: 2178 case Op_ModF: 2179 case Op_ConvI2F: 2180 case Op_ConF: 2181 case Op_CmpF: 2182 case Op_CmpF3: 2183 // case Op_ConvL2F: // longs are split into 32-bit halves 2184 frc.inc_float_count(); 2185 break; 2186 2187 case Op_ConvF2D: 2188 case Op_ConvD2F: 2189 frc.inc_float_count(); 2190 frc.inc_double_count(); 2191 break; 2192 2193 // Count all double operations that may use FPU 2194 case Op_AddD: 2195 case Op_SubD: 2196 case Op_MulD: 2197 case Op_DivD: 2198 case Op_NegD: 2199 case Op_ModD: 2200 case Op_ConvI2D: 2201 case Op_ConvD2I: 2202 // case Op_ConvL2D: // handled by leaf call 2203 // case Op_ConvD2L: // handled by leaf call 2204 case Op_ConD: 2205 case Op_CmpD: 2206 case Op_CmpD3: 2207 frc.inc_double_count(); 2208 break; 2209 case Op_Opaque1: // Remove Opaque Nodes before matching 2210 case Op_Opaque2: // Remove Opaque Nodes before matching 2211 n->subsume_by(n->in(1)); 2212 break; 2213 case Op_CallStaticJava: 2214 case Op_CallJava: 2215 case Op_CallDynamicJava: 2216 frc.inc_java_call_count(); // Count java call site; 2217 case Op_CallRuntime: 2218 case Op_CallLeaf: 2219 case Op_CallLeafNoFP: { 2220 assert( n->is_Call(), "" ); 2221 CallNode *call = n->as_Call(); 2222 // Count call sites where the FP mode bit would have to be flipped. 2223 // Do not count uncommon runtime calls: 2224 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2225 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2226 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2227 frc.inc_call_count(); // Count the call site 2228 } else { // See if uncommon argument is shared 2229 Node *n = call->in(TypeFunc::Parms); 2230 int nop = n->Opcode(); 2231 // Clone shared simple arguments to uncommon calls, item (1). 2232 if( n->outcnt() > 1 && 2233 !n->is_Proj() && 2234 nop != Op_CreateEx && 2235 nop != Op_CheckCastPP && 2236 nop != Op_DecodeN && 2237 !n->is_Mem() ) { 2238 Node *x = n->clone(); 2239 call->set_req( TypeFunc::Parms, x ); 2240 } 2241 } 2242 break; 2243 } 2244 2245 case Op_StoreD: 2246 case Op_LoadD: 2247 case Op_LoadD_unaligned: 2248 frc.inc_double_count(); 2249 goto handle_mem; 2250 case Op_StoreF: 2251 case Op_LoadF: 2252 frc.inc_float_count(); 2253 goto handle_mem; 2254 2255 case Op_StoreCM: 2256 { 2257 // Convert OopStore dependence into precedence edge 2258 Node* prec = n->in(MemNode::OopStore); 2259 n->del_req(MemNode::OopStore); 2260 n->add_prec(prec); 2261 eliminate_redundant_card_marks(n); 2262 } 2263 2264 // fall through 2265 2266 case Op_StoreB: 2267 case Op_StoreC: 2268 case Op_StorePConditional: 2269 case Op_StoreI: 2270 case Op_StoreL: 2271 case Op_StoreIConditional: 2272 case Op_StoreLConditional: 2273 case Op_CompareAndSwapI: 2274 case Op_CompareAndSwapL: 2275 case Op_CompareAndSwapP: 2276 case Op_CompareAndSwapN: 2277 case Op_StoreP: 2278 case Op_StoreN: 2279 case Op_LoadB: 2280 case Op_LoadUB: 2281 case Op_LoadUS: 2282 case Op_LoadI: 2283 case Op_LoadUI2L: 2284 case Op_LoadKlass: 2285 case Op_LoadNKlass: 2286 case Op_LoadL: 2287 case Op_LoadL_unaligned: 2288 case Op_LoadPLocked: 2289 case Op_LoadP: 2290 case Op_LoadN: 2291 case Op_LoadRange: 2292 case Op_LoadS: { 2293 handle_mem: 2294#ifdef ASSERT 2295 if( VerifyOptoOopOffsets ) { 2296 assert( n->is_Mem(), "" ); 2297 MemNode *mem = (MemNode*)n; 2298 // Check to see if address types have grounded out somehow. 2299 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2300 assert( !tp || oop_offset_is_sane(tp), "" ); 2301 } 2302#endif 2303 break; 2304 } 2305 2306 case Op_AddP: { // Assert sane base pointers 2307 Node *addp = n->in(AddPNode::Address); 2308 assert( !addp->is_AddP() || 2309 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2310 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2311 "Base pointers must match" ); 2312#ifdef _LP64 2313 if (UseCompressedOops && 2314 addp->Opcode() == Op_ConP && 2315 addp == n->in(AddPNode::Base) && 2316 n->in(AddPNode::Offset)->is_Con()) { 2317 // Use addressing with narrow klass to load with offset on x86. 2318 // On sparc loading 32-bits constant and decoding it have less 2319 // instructions (4) then load 64-bits constant (7). 2320 // Do this transformation here since IGVN will convert ConN back to ConP. 2321 const Type* t = addp->bottom_type(); 2322 if (t->isa_oopptr()) { 2323 Node* nn = NULL; 2324 2325 // Look for existing ConN node of the same exact type. 2326 Compile* C = Compile::current(); 2327 Node* r = C->root(); 2328 uint cnt = r->outcnt(); 2329 for (uint i = 0; i < cnt; i++) { 2330 Node* m = r->raw_out(i); 2331 if (m!= NULL && m->Opcode() == Op_ConN && 2332 m->bottom_type()->make_ptr() == t) { 2333 nn = m; 2334 break; 2335 } 2336 } 2337 if (nn != NULL) { 2338 // Decode a narrow oop to match address 2339 // [R12 + narrow_oop_reg<<3 + offset] 2340 nn = new (C, 2) DecodeNNode(nn, t); 2341 n->set_req(AddPNode::Base, nn); 2342 n->set_req(AddPNode::Address, nn); 2343 if (addp->outcnt() == 0) { 2344 addp->disconnect_inputs(NULL); 2345 } 2346 } 2347 } 2348 } 2349#endif 2350 break; 2351 } 2352 2353#ifdef _LP64 2354 case Op_CastPP: 2355 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2356 Compile* C = Compile::current(); 2357 Node* in1 = n->in(1); 2358 const Type* t = n->bottom_type(); 2359 Node* new_in1 = in1->clone(); 2360 new_in1->as_DecodeN()->set_type(t); 2361 2362 if (!Matcher::narrow_oop_use_complex_address()) { 2363 // 2364 // x86, ARM and friends can handle 2 adds in addressing mode 2365 // and Matcher can fold a DecodeN node into address by using 2366 // a narrow oop directly and do implicit NULL check in address: 2367 // 2368 // [R12 + narrow_oop_reg<<3 + offset] 2369 // NullCheck narrow_oop_reg 2370 // 2371 // On other platforms (Sparc) we have to keep new DecodeN node and 2372 // use it to do implicit NULL check in address: 2373 // 2374 // decode_not_null narrow_oop_reg, base_reg 2375 // [base_reg + offset] 2376 // NullCheck base_reg 2377 // 2378 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2379 // to keep the information to which NULL check the new DecodeN node 2380 // corresponds to use it as value in implicit_null_check(). 2381 // 2382 new_in1->set_req(0, n->in(0)); 2383 } 2384 2385 n->subsume_by(new_in1); 2386 if (in1->outcnt() == 0) { 2387 in1->disconnect_inputs(NULL); 2388 } 2389 } 2390 break; 2391 2392 case Op_CmpP: 2393 // Do this transformation here to preserve CmpPNode::sub() and 2394 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2395 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2396 Node* in1 = n->in(1); 2397 Node* in2 = n->in(2); 2398 if (!in1->is_DecodeN()) { 2399 in2 = in1; 2400 in1 = n->in(2); 2401 } 2402 assert(in1->is_DecodeN(), "sanity"); 2403 2404 Compile* C = Compile::current(); 2405 Node* new_in2 = NULL; 2406 if (in2->is_DecodeN()) { 2407 new_in2 = in2->in(1); 2408 } else if (in2->Opcode() == Op_ConP) { 2409 const Type* t = in2->bottom_type(); 2410 if (t == TypePtr::NULL_PTR) { 2411 // Don't convert CmpP null check into CmpN if compressed 2412 // oops implicit null check is not generated. 2413 // This will allow to generate normal oop implicit null check. 2414 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2415 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2416 // 2417 // This transformation together with CastPP transformation above 2418 // will generated code for implicit NULL checks for compressed oops. 2419 // 2420 // The original code after Optimize() 2421 // 2422 // LoadN memory, narrow_oop_reg 2423 // decode narrow_oop_reg, base_reg 2424 // CmpP base_reg, NULL 2425 // CastPP base_reg // NotNull 2426 // Load [base_reg + offset], val_reg 2427 // 2428 // after these transformations will be 2429 // 2430 // LoadN memory, narrow_oop_reg 2431 // CmpN narrow_oop_reg, NULL 2432 // decode_not_null narrow_oop_reg, base_reg 2433 // Load [base_reg + offset], val_reg 2434 // 2435 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2436 // since narrow oops can be used in debug info now (see the code in 2437 // final_graph_reshaping_walk()). 2438 // 2439 // At the end the code will be matched to 2440 // on x86: 2441 // 2442 // Load_narrow_oop memory, narrow_oop_reg 2443 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2444 // NullCheck narrow_oop_reg 2445 // 2446 // and on sparc: 2447 // 2448 // Load_narrow_oop memory, narrow_oop_reg 2449 // decode_not_null narrow_oop_reg, base_reg 2450 // Load [base_reg + offset], val_reg 2451 // NullCheck base_reg 2452 // 2453 } else if (t->isa_oopptr()) { 2454 new_in2 = ConNode::make(C, t->make_narrowoop()); 2455 } 2456 } 2457 if (new_in2 != NULL) { 2458 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2459 n->subsume_by( cmpN ); 2460 if (in1->outcnt() == 0) { 2461 in1->disconnect_inputs(NULL); 2462 } 2463 if (in2->outcnt() == 0) { 2464 in2->disconnect_inputs(NULL); 2465 } 2466 } 2467 } 2468 break; 2469 2470 case Op_DecodeN: 2471 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2472 // DecodeN could be pinned when it can't be fold into 2473 // an address expression, see the code for Op_CastPP above. 2474 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2475 break; 2476 2477 case Op_EncodeP: { 2478 Node* in1 = n->in(1); 2479 if (in1->is_DecodeN()) { 2480 n->subsume_by(in1->in(1)); 2481 } else if (in1->Opcode() == Op_ConP) { 2482 Compile* C = Compile::current(); 2483 const Type* t = in1->bottom_type(); 2484 if (t == TypePtr::NULL_PTR) { 2485 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2486 } else if (t->isa_oopptr()) { 2487 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2488 } 2489 } 2490 if (in1->outcnt() == 0) { 2491 in1->disconnect_inputs(NULL); 2492 } 2493 break; 2494 } 2495 2496 case Op_Proj: { 2497 if (OptimizeStringConcat) { 2498 ProjNode* p = n->as_Proj(); 2499 if (p->_is_io_use) { 2500 // Separate projections were used for the exception path which 2501 // are normally removed by a late inline. If it wasn't inlined 2502 // then they will hang around and should just be replaced with 2503 // the original one. 2504 Node* proj = NULL; 2505 // Replace with just one 2506 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2507 Node *use = i.get(); 2508 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2509 proj = use; 2510 break; 2511 } 2512 } 2513 assert(proj != NULL, "must be found"); 2514 p->subsume_by(proj); 2515 } 2516 } 2517 break; 2518 } 2519 2520 case Op_Phi: 2521 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2522 // The EncodeP optimization may create Phi with the same edges 2523 // for all paths. It is not handled well by Register Allocator. 2524 Node* unique_in = n->in(1); 2525 assert(unique_in != NULL, ""); 2526 uint cnt = n->req(); 2527 for (uint i = 2; i < cnt; i++) { 2528 Node* m = n->in(i); 2529 assert(m != NULL, ""); 2530 if (unique_in != m) 2531 unique_in = NULL; 2532 } 2533 if (unique_in != NULL) { 2534 n->subsume_by(unique_in); 2535 } 2536 } 2537 break; 2538 2539#endif 2540 2541 case Op_ModI: 2542 if (UseDivMod) { 2543 // Check if a%b and a/b both exist 2544 Node* d = n->find_similar(Op_DivI); 2545 if (d) { 2546 // Replace them with a fused divmod if supported 2547 Compile* C = Compile::current(); 2548 if (Matcher::has_match_rule(Op_DivModI)) { 2549 DivModINode* divmod = DivModINode::make(C, n); 2550 d->subsume_by(divmod->div_proj()); 2551 n->subsume_by(divmod->mod_proj()); 2552 } else { 2553 // replace a%b with a-((a/b)*b) 2554 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2555 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2556 n->subsume_by( sub ); 2557 } 2558 } 2559 } 2560 break; 2561 2562 case Op_ModL: 2563 if (UseDivMod) { 2564 // Check if a%b and a/b both exist 2565 Node* d = n->find_similar(Op_DivL); 2566 if (d) { 2567 // Replace them with a fused divmod if supported 2568 Compile* C = Compile::current(); 2569 if (Matcher::has_match_rule(Op_DivModL)) { 2570 DivModLNode* divmod = DivModLNode::make(C, n); 2571 d->subsume_by(divmod->div_proj()); 2572 n->subsume_by(divmod->mod_proj()); 2573 } else { 2574 // replace a%b with a-((a/b)*b) 2575 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2576 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2577 n->subsume_by( sub ); 2578 } 2579 } 2580 } 2581 break; 2582 2583 case Op_LoadVector: 2584 case Op_StoreVector: 2585 break; 2586 2587 case Op_PackB: 2588 case Op_PackS: 2589 case Op_PackI: 2590 case Op_PackF: 2591 case Op_PackL: 2592 case Op_PackD: 2593 if (n->req()-1 > 2) { 2594 // Replace many operand PackNodes with a binary tree for matching 2595 PackNode* p = (PackNode*) n; 2596 Node* btp = p->binary_tree_pack(Compile::current(), 1, n->req()); 2597 n->subsume_by(btp); 2598 } 2599 break; 2600 case Op_Loop: 2601 case Op_CountedLoop: 2602 if (n->as_Loop()->is_inner_loop()) { 2603 frc.inc_inner_loop_count(); 2604 } 2605 break; 2606 case Op_LShiftI: 2607 case Op_RShiftI: 2608 case Op_URShiftI: 2609 case Op_LShiftL: 2610 case Op_RShiftL: 2611 case Op_URShiftL: 2612 if (Matcher::need_masked_shift_count) { 2613 // The cpu's shift instructions don't restrict the count to the 2614 // lower 5/6 bits. We need to do the masking ourselves. 2615 Node* in2 = n->in(2); 2616 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2617 const TypeInt* t = in2->find_int_type(); 2618 if (t != NULL && t->is_con()) { 2619 juint shift = t->get_con(); 2620 if (shift > mask) { // Unsigned cmp 2621 Compile* C = Compile::current(); 2622 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask))); 2623 } 2624 } else { 2625 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2626 Compile* C = Compile::current(); 2627 Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask))); 2628 n->set_req(2, shift); 2629 } 2630 } 2631 if (in2->outcnt() == 0) { // Remove dead node 2632 in2->disconnect_inputs(NULL); 2633 } 2634 } 2635 break; 2636 default: 2637 assert( !n->is_Call(), "" ); 2638 assert( !n->is_Mem(), "" ); 2639 break; 2640 } 2641 2642 // Collect CFG split points 2643 if (n->is_MultiBranch()) 2644 frc._tests.push(n); 2645} 2646 2647//------------------------------final_graph_reshaping_walk--------------------- 2648// Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2649// requires that the walk visits a node's inputs before visiting the node. 2650static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2651 ResourceArea *area = Thread::current()->resource_area(); 2652 Unique_Node_List sfpt(area); 2653 2654 frc._visited.set(root->_idx); // first, mark node as visited 2655 uint cnt = root->req(); 2656 Node *n = root; 2657 uint i = 0; 2658 while (true) { 2659 if (i < cnt) { 2660 // Place all non-visited non-null inputs onto stack 2661 Node* m = n->in(i); 2662 ++i; 2663 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2664 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2665 sfpt.push(m); 2666 cnt = m->req(); 2667 nstack.push(n, i); // put on stack parent and next input's index 2668 n = m; 2669 i = 0; 2670 } 2671 } else { 2672 // Now do post-visit work 2673 final_graph_reshaping_impl( n, frc ); 2674 if (nstack.is_empty()) 2675 break; // finished 2676 n = nstack.node(); // Get node from stack 2677 cnt = n->req(); 2678 i = nstack.index(); 2679 nstack.pop(); // Shift to the next node on stack 2680 } 2681 } 2682 2683 // Skip next transformation if compressed oops are not used. 2684 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) 2685 return; 2686 2687 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2688 // It could be done for an uncommon traps or any safepoints/calls 2689 // if the DecodeN node is referenced only in a debug info. 2690 while (sfpt.size() > 0) { 2691 n = sfpt.pop(); 2692 JVMState *jvms = n->as_SafePoint()->jvms(); 2693 assert(jvms != NULL, "sanity"); 2694 int start = jvms->debug_start(); 2695 int end = n->req(); 2696 bool is_uncommon = (n->is_CallStaticJava() && 2697 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2698 for (int j = start; j < end; j++) { 2699 Node* in = n->in(j); 2700 if (in->is_DecodeN()) { 2701 bool safe_to_skip = true; 2702 if (!is_uncommon ) { 2703 // Is it safe to skip? 2704 for (uint i = 0; i < in->outcnt(); i++) { 2705 Node* u = in->raw_out(i); 2706 if (!u->is_SafePoint() || 2707 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2708 safe_to_skip = false; 2709 } 2710 } 2711 } 2712 if (safe_to_skip) { 2713 n->set_req(j, in->in(1)); 2714 } 2715 if (in->outcnt() == 0) { 2716 in->disconnect_inputs(NULL); 2717 } 2718 } 2719 } 2720 } 2721} 2722 2723//------------------------------final_graph_reshaping-------------------------- 2724// Final Graph Reshaping. 2725// 2726// (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2727// and not commoned up and forced early. Must come after regular 2728// optimizations to avoid GVN undoing the cloning. Clone constant 2729// inputs to Loop Phis; these will be split by the allocator anyways. 2730// Remove Opaque nodes. 2731// (2) Move last-uses by commutative operations to the left input to encourage 2732// Intel update-in-place two-address operations and better register usage 2733// on RISCs. Must come after regular optimizations to avoid GVN Ideal 2734// calls canonicalizing them back. 2735// (3) Count the number of double-precision FP ops, single-precision FP ops 2736// and call sites. On Intel, we can get correct rounding either by 2737// forcing singles to memory (requires extra stores and loads after each 2738// FP bytecode) or we can set a rounding mode bit (requires setting and 2739// clearing the mode bit around call sites). The mode bit is only used 2740// if the relative frequency of single FP ops to calls is low enough. 2741// This is a key transform for SPEC mpeg_audio. 2742// (4) Detect infinite loops; blobs of code reachable from above but not 2743// below. Several of the Code_Gen algorithms fail on such code shapes, 2744// so we simply bail out. Happens a lot in ZKM.jar, but also happens 2745// from time to time in other codes (such as -Xcomp finalizer loops, etc). 2746// Detection is by looking for IfNodes where only 1 projection is 2747// reachable from below or CatchNodes missing some targets. 2748// (5) Assert for insane oop offsets in debug mode. 2749 2750bool Compile::final_graph_reshaping() { 2751 // an infinite loop may have been eliminated by the optimizer, 2752 // in which case the graph will be empty. 2753 if (root()->req() == 1) { 2754 record_method_not_compilable("trivial infinite loop"); 2755 return true; 2756 } 2757 2758 Final_Reshape_Counts frc; 2759 2760 // Visit everybody reachable! 2761 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2762 Node_Stack nstack(unique() >> 1); 2763 final_graph_reshaping_walk(nstack, root(), frc); 2764 2765 // Check for unreachable (from below) code (i.e., infinite loops). 2766 for( uint i = 0; i < frc._tests.size(); i++ ) { 2767 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2768 // Get number of CFG targets. 2769 // Note that PCTables include exception targets after calls. 2770 uint required_outcnt = n->required_outcnt(); 2771 if (n->outcnt() != required_outcnt) { 2772 // Check for a few special cases. Rethrow Nodes never take the 2773 // 'fall-thru' path, so expected kids is 1 less. 2774 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2775 if (n->in(0)->in(0)->is_Call()) { 2776 CallNode *call = n->in(0)->in(0)->as_Call(); 2777 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2778 required_outcnt--; // Rethrow always has 1 less kid 2779 } else if (call->req() > TypeFunc::Parms && 2780 call->is_CallDynamicJava()) { 2781 // Check for null receiver. In such case, the optimizer has 2782 // detected that the virtual call will always result in a null 2783 // pointer exception. The fall-through projection of this CatchNode 2784 // will not be populated. 2785 Node *arg0 = call->in(TypeFunc::Parms); 2786 if (arg0->is_Type() && 2787 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2788 required_outcnt--; 2789 } 2790 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2791 call->req() > TypeFunc::Parms+1 && 2792 call->is_CallStaticJava()) { 2793 // Check for negative array length. In such case, the optimizer has 2794 // detected that the allocation attempt will always result in an 2795 // exception. There is no fall-through projection of this CatchNode . 2796 Node *arg1 = call->in(TypeFunc::Parms+1); 2797 if (arg1->is_Type() && 2798 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2799 required_outcnt--; 2800 } 2801 } 2802 } 2803 } 2804 // Recheck with a better notion of 'required_outcnt' 2805 if (n->outcnt() != required_outcnt) { 2806 record_method_not_compilable("malformed control flow"); 2807 return true; // Not all targets reachable! 2808 } 2809 } 2810 // Check that I actually visited all kids. Unreached kids 2811 // must be infinite loops. 2812 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2813 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2814 record_method_not_compilable("infinite loop"); 2815 return true; // Found unvisited kid; must be unreach 2816 } 2817 } 2818 2819 // If original bytecodes contained a mixture of floats and doubles 2820 // check if the optimizer has made it homogenous, item (3). 2821 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2822 frc.get_float_count() > 32 && 2823 frc.get_double_count() == 0 && 2824 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2825 set_24_bit_selection_and_mode( false, true ); 2826 } 2827 2828 set_java_calls(frc.get_java_call_count()); 2829 set_inner_loops(frc.get_inner_loop_count()); 2830 2831 // No infinite loops, no reason to bail out. 2832 return false; 2833} 2834 2835//-----------------------------too_many_traps---------------------------------- 2836// Report if there are too many traps at the current method and bci. 2837// Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2838bool Compile::too_many_traps(ciMethod* method, 2839 int bci, 2840 Deoptimization::DeoptReason reason) { 2841 ciMethodData* md = method->method_data(); 2842 if (md->is_empty()) { 2843 // Assume the trap has not occurred, or that it occurred only 2844 // because of a transient condition during start-up in the interpreter. 2845 return false; 2846 } 2847 if (md->has_trap_at(bci, reason) != 0) { 2848 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2849 // Also, if there are multiple reasons, or if there is no per-BCI record, 2850 // assume the worst. 2851 if (log()) 2852 log()->elem("observe trap='%s' count='%d'", 2853 Deoptimization::trap_reason_name(reason), 2854 md->trap_count(reason)); 2855 return true; 2856 } else { 2857 // Ignore method/bci and see if there have been too many globally. 2858 return too_many_traps(reason, md); 2859 } 2860} 2861 2862// Less-accurate variant which does not require a method and bci. 2863bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2864 ciMethodData* logmd) { 2865 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2866 // Too many traps globally. 2867 // Note that we use cumulative trap_count, not just md->trap_count. 2868 if (log()) { 2869 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2870 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2871 Deoptimization::trap_reason_name(reason), 2872 mcount, trap_count(reason)); 2873 } 2874 return true; 2875 } else { 2876 // The coast is clear. 2877 return false; 2878 } 2879} 2880 2881//--------------------------too_many_recompiles-------------------------------- 2882// Report if there are too many recompiles at the current method and bci. 2883// Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2884// Is not eager to return true, since this will cause the compiler to use 2885// Action_none for a trap point, to avoid too many recompilations. 2886bool Compile::too_many_recompiles(ciMethod* method, 2887 int bci, 2888 Deoptimization::DeoptReason reason) { 2889 ciMethodData* md = method->method_data(); 2890 if (md->is_empty()) { 2891 // Assume the trap has not occurred, or that it occurred only 2892 // because of a transient condition during start-up in the interpreter. 2893 return false; 2894 } 2895 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2896 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2897 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2898 Deoptimization::DeoptReason per_bc_reason 2899 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2900 if ((per_bc_reason == Deoptimization::Reason_none 2901 || md->has_trap_at(bci, reason) != 0) 2902 // The trap frequency measure we care about is the recompile count: 2903 && md->trap_recompiled_at(bci) 2904 && md->overflow_recompile_count() >= bc_cutoff) { 2905 // Do not emit a trap here if it has already caused recompilations. 2906 // Also, if there are multiple reasons, or if there is no per-BCI record, 2907 // assume the worst. 2908 if (log()) 2909 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2910 Deoptimization::trap_reason_name(reason), 2911 md->trap_count(reason), 2912 md->overflow_recompile_count()); 2913 return true; 2914 } else if (trap_count(reason) != 0 2915 && decompile_count() >= m_cutoff) { 2916 // Too many recompiles globally, and we have seen this sort of trap. 2917 // Use cumulative decompile_count, not just md->decompile_count. 2918 if (log()) 2919 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2920 Deoptimization::trap_reason_name(reason), 2921 md->trap_count(reason), trap_count(reason), 2922 md->decompile_count(), decompile_count()); 2923 return true; 2924 } else { 2925 // The coast is clear. 2926 return false; 2927 } 2928} 2929 2930 2931#ifndef PRODUCT 2932//------------------------------verify_graph_edges--------------------------- 2933// Walk the Graph and verify that there is a one-to-one correspondence 2934// between Use-Def edges and Def-Use edges in the graph. 2935void Compile::verify_graph_edges(bool no_dead_code) { 2936 if (VerifyGraphEdges) { 2937 ResourceArea *area = Thread::current()->resource_area(); 2938 Unique_Node_List visited(area); 2939 // Call recursive graph walk to check edges 2940 _root->verify_edges(visited); 2941 if (no_dead_code) { 2942 // Now make sure that no visited node is used by an unvisited node. 2943 bool dead_nodes = 0; 2944 Unique_Node_List checked(area); 2945 while (visited.size() > 0) { 2946 Node* n = visited.pop(); 2947 checked.push(n); 2948 for (uint i = 0; i < n->outcnt(); i++) { 2949 Node* use = n->raw_out(i); 2950 if (checked.member(use)) continue; // already checked 2951 if (visited.member(use)) continue; // already in the graph 2952 if (use->is_Con()) continue; // a dead ConNode is OK 2953 // At this point, we have found a dead node which is DU-reachable. 2954 if (dead_nodes++ == 0) 2955 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2956 use->dump(2); 2957 tty->print_cr("---"); 2958 checked.push(use); // No repeats; pretend it is now checked. 2959 } 2960 } 2961 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2962 } 2963 } 2964} 2965#endif 2966 2967// The Compile object keeps track of failure reasons separately from the ciEnv. 2968// This is required because there is not quite a 1-1 relation between the 2969// ciEnv and its compilation task and the Compile object. Note that one 2970// ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2971// to backtrack and retry without subsuming loads. Other than this backtracking 2972// behavior, the Compile's failure reason is quietly copied up to the ciEnv 2973// by the logic in C2Compiler. 2974void Compile::record_failure(const char* reason) { 2975 if (log() != NULL) { 2976 log()->elem("failure reason='%s' phase='compile'", reason); 2977 } 2978 if (_failure_reason == NULL) { 2979 // Record the first failure reason. 2980 _failure_reason = reason; 2981 } 2982 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2983 C->print_method(_failure_reason); 2984 } 2985 _root = NULL; // flush the graph, too 2986} 2987 2988Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2989 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2990{ 2991 if (dolog) { 2992 C = Compile::current(); 2993 _log = C->log(); 2994 } else { 2995 C = NULL; 2996 _log = NULL; 2997 } 2998 if (_log != NULL) { 2999 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 3000 _log->stamp(); 3001 _log->end_head(); 3002 } 3003} 3004 3005Compile::TracePhase::~TracePhase() { 3006 if (_log != NULL) { 3007 _log->done("phase nodes='%d'", C->unique()); 3008 } 3009} 3010 3011//============================================================================= 3012// Two Constant's are equal when the type and the value are equal. 3013bool Compile::Constant::operator==(const Constant& other) { 3014 if (type() != other.type() ) return false; 3015 if (can_be_reused() != other.can_be_reused()) return false; 3016 // For floating point values we compare the bit pattern. 3017 switch (type()) { 3018 case T_FLOAT: return (_v._value.i == other._v._value.i); 3019 case T_LONG: 3020 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3021 case T_OBJECT: 3022 case T_METADATA: return (_v._metadata == other._v._metadata); 3023 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3024 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3025 default: ShouldNotReachHere(); 3026 } 3027 return false; 3028} 3029 3030static int type_to_size_in_bytes(BasicType t) { 3031 switch (t) { 3032 case T_LONG: return sizeof(jlong ); 3033 case T_FLOAT: return sizeof(jfloat ); 3034 case T_DOUBLE: return sizeof(jdouble); 3035 case T_METADATA: return sizeof(Metadata*); 3036 // We use T_VOID as marker for jump-table entries (labels) which 3037 // need an internal word relocation. 3038 case T_VOID: 3039 case T_ADDRESS: 3040 case T_OBJECT: return sizeof(jobject); 3041 } 3042 3043 ShouldNotReachHere(); 3044 return -1; 3045} 3046 3047int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3048 // sort descending 3049 if (a->freq() > b->freq()) return -1; 3050 if (a->freq() < b->freq()) return 1; 3051 return 0; 3052} 3053 3054void Compile::ConstantTable::calculate_offsets_and_size() { 3055 // First, sort the array by frequencies. 3056 _constants.sort(qsort_comparator); 3057 3058#ifdef ASSERT 3059 // Make sure all jump-table entries were sorted to the end of the 3060 // array (they have a negative frequency). 3061 bool found_void = false; 3062 for (int i = 0; i < _constants.length(); i++) { 3063 Constant con = _constants.at(i); 3064 if (con.type() == T_VOID) 3065 found_void = true; // jump-tables 3066 else 3067 assert(!found_void, "wrong sorting"); 3068 } 3069#endif 3070 3071 int offset = 0; 3072 for (int i = 0; i < _constants.length(); i++) { 3073 Constant* con = _constants.adr_at(i); 3074 3075 // Align offset for type. 3076 int typesize = type_to_size_in_bytes(con->type()); 3077 offset = align_size_up(offset, typesize); 3078 con->set_offset(offset); // set constant's offset 3079 3080 if (con->type() == T_VOID) { 3081 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3082 offset = offset + typesize * n->outcnt(); // expand jump-table 3083 } else { 3084 offset = offset + typesize; 3085 } 3086 } 3087 3088 // Align size up to the next section start (which is insts; see 3089 // CodeBuffer::align_at_start). 3090 assert(_size == -1, "already set?"); 3091 _size = align_size_up(offset, CodeEntryAlignment); 3092} 3093 3094void Compile::ConstantTable::emit(CodeBuffer& cb) { 3095 MacroAssembler _masm(&cb); 3096 for (int i = 0; i < _constants.length(); i++) { 3097 Constant con = _constants.at(i); 3098 address constant_addr; 3099 switch (con.type()) { 3100 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3101 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3102 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3103 case T_OBJECT: { 3104 jobject obj = con.get_jobject(); 3105 int oop_index = _masm.oop_recorder()->find_index(obj); 3106 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3107 break; 3108 } 3109 case T_ADDRESS: { 3110 address addr = (address) con.get_jobject(); 3111 constant_addr = _masm.address_constant(addr); 3112 break; 3113 } 3114 // We use T_VOID as marker for jump-table entries (labels) which 3115 // need an internal word relocation. 3116 case T_VOID: { 3117 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3118 // Fill the jump-table with a dummy word. The real value is 3119 // filled in later in fill_jump_table. 3120 address dummy = (address) n; 3121 constant_addr = _masm.address_constant(dummy); 3122 // Expand jump-table 3123 for (uint i = 1; i < n->outcnt(); i++) { 3124 address temp_addr = _masm.address_constant(dummy + i); 3125 assert(temp_addr, "consts section too small"); 3126 } 3127 break; 3128 } 3129 case T_METADATA: { 3130 Metadata* obj = con.get_metadata(); 3131 int metadata_index = _masm.oop_recorder()->find_index(obj); 3132 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3133 break; 3134 } 3135 default: ShouldNotReachHere(); 3136 } 3137 assert(constant_addr, "consts section too small"); 3138 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3139 } 3140} 3141 3142int Compile::ConstantTable::find_offset(Constant& con) const { 3143 int idx = _constants.find(con); 3144 assert(idx != -1, "constant must be in constant table"); 3145 int offset = _constants.at(idx).offset(); 3146 assert(offset != -1, "constant table not emitted yet?"); 3147 return offset; 3148} 3149 3150void Compile::ConstantTable::add(Constant& con) { 3151 if (con.can_be_reused()) { 3152 int idx = _constants.find(con); 3153 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3154 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3155 return; 3156 } 3157 } 3158 (void) _constants.append(con); 3159} 3160 3161Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3162 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3163 Constant con(type, value, b->_freq); 3164 add(con); 3165 return con; 3166} 3167 3168Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3169 Constant con(metadata); 3170 add(con); 3171 return con; 3172} 3173 3174Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3175 jvalue value; 3176 BasicType type = oper->type()->basic_type(); 3177 switch (type) { 3178 case T_LONG: value.j = oper->constantL(); break; 3179 case T_FLOAT: value.f = oper->constantF(); break; 3180 case T_DOUBLE: value.d = oper->constantD(); break; 3181 case T_OBJECT: 3182 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3183 case T_METADATA: return add((Metadata*)oper->constant()); break; 3184 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3185 } 3186 return add(n, type, value); 3187} 3188 3189Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3190 jvalue value; 3191 // We can use the node pointer here to identify the right jump-table 3192 // as this method is called from Compile::Fill_buffer right before 3193 // the MachNodes are emitted and the jump-table is filled (means the 3194 // MachNode pointers do not change anymore). 3195 value.l = (jobject) n; 3196 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3197 add(con); 3198 return con; 3199} 3200 3201void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3202 // If called from Compile::scratch_emit_size do nothing. 3203 if (Compile::current()->in_scratch_emit_size()) return; 3204 3205 assert(labels.is_nonempty(), "must be"); 3206 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3207 3208 // Since MachConstantNode::constant_offset() also contains 3209 // table_base_offset() we need to subtract the table_base_offset() 3210 // to get the plain offset into the constant table. 3211 int offset = n->constant_offset() - table_base_offset(); 3212 3213 MacroAssembler _masm(&cb); 3214 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3215 3216 for (uint i = 0; i < n->outcnt(); i++) { 3217 address* constant_addr = &jump_table_base[i]; 3218 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); 3219 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3220 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3221 } 3222} 3223