1 /* 2 * Copyright (c) 1997, 2011, 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. 87 MachConstantBaseNode* 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. 100 int 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 } 134 135 void 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 } 160 161 CallGenerator* 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... 192 193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 195 196 bool 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 } 232 233 static 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 } 244 245 void 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 } 265 266 void 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 285 Bundle* Compile::node_bundling(const Node *n) { 286 assert(valid_bundle_info(n), "oob"); 287 return &_node_bundling_base[n->_idx]; 288 } 289 290 bool Compile::valid_bundle_info(const Node *n) { 291 return (_node_bundling_limit > n->_idx); 292 } 293 294 295 void 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. 324 void 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. 347 void 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 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 369 } 370 371 //------------------------------frame_size_in_words----------------------------- 372 // frame_slots in units of words 373 int Compile::frame_size_in_words() const { 374 // shift is 0 in LP32 and 1 in LP64 375 const int shift = (LogBytesPerWord - LogBytesPerInt); 376 int words = _frame_slots >> shift; 377 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 378 return words; 379 } 380 381 // ============================================================================ 382 //------------------------------CompileWrapper--------------------------------- 383 class CompileWrapper : public StackObj { 384 Compile *const _compile; 385 public: 386 CompileWrapper(Compile* compile); 387 388 ~CompileWrapper(); 389 }; 390 391 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 392 // the Compile* pointer is stored in the current ciEnv: 393 ciEnv* env = compile->env(); 394 assert(env == ciEnv::current(), "must already be a ciEnv active"); 395 assert(env->compiler_data() == NULL, "compile already active?"); 396 env->set_compiler_data(compile); 397 assert(compile == Compile::current(), "sanity"); 398 399 compile->set_type_dict(NULL); 400 compile->set_type_hwm(NULL); 401 compile->set_type_last_size(0); 402 compile->set_last_tf(NULL, NULL); 403 compile->set_indexSet_arena(NULL); 404 compile->set_indexSet_free_block_list(NULL); 405 compile->init_type_arena(); 406 Type::Initialize(compile); 407 _compile->set_scratch_buffer_blob(NULL); 408 _compile->begin_method(); 409 } 410 CompileWrapper::~CompileWrapper() { 411 _compile->end_method(); 412 if (_compile->scratch_buffer_blob() != NULL) 413 BufferBlob::free(_compile->scratch_buffer_blob()); 414 _compile->env()->set_compiler_data(NULL); 415 } 416 417 418 //----------------------------print_compile_messages--------------------------- 419 void Compile::print_compile_messages() { 420 #ifndef PRODUCT 421 // Check if recompiling 422 if (_subsume_loads == false && PrintOpto) { 423 // Recompiling without allowing machine instructions to subsume loads 424 tty->print_cr("*********************************************************"); 425 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 426 tty->print_cr("*********************************************************"); 427 } 428 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 429 // Recompiling without escape analysis 430 tty->print_cr("*********************************************************"); 431 tty->print_cr("** Bailout: Recompile without escape analysis **"); 432 tty->print_cr("*********************************************************"); 433 } 434 if (env()->break_at_compile()) { 435 // Open the debugger when compiling this method. 436 tty->print("### Breaking when compiling: "); 437 method()->print_short_name(); 438 tty->cr(); 439 BREAKPOINT; 440 } 441 442 if( PrintOpto ) { 443 if (is_osr_compilation()) { 444 tty->print("[OSR]%3d", _compile_id); 445 } else { 446 tty->print("%3d", _compile_id); 447 } 448 } 449 #endif 450 } 451 452 453 //-----------------------init_scratch_buffer_blob------------------------------ 454 // Construct a temporary BufferBlob and cache it for this compile. 455 void Compile::init_scratch_buffer_blob(int const_size) { 456 // If there is already a scratch buffer blob allocated and the 457 // constant section is big enough, use it. Otherwise free the 458 // current and allocate a new one. 459 BufferBlob* blob = scratch_buffer_blob(); 460 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 461 // Use the current blob. 462 } else { 463 if (blob != NULL) { 464 BufferBlob::free(blob); 465 } 466 467 ResourceMark rm; 468 _scratch_const_size = const_size; 469 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 470 blob = BufferBlob::create("Compile::scratch_buffer", size); 471 // Record the buffer blob for next time. 472 set_scratch_buffer_blob(blob); 473 // Have we run out of code space? 474 if (scratch_buffer_blob() == NULL) { 475 // Let CompilerBroker disable further compilations. 476 record_failure("Not enough space for scratch buffer in CodeCache"); 477 return; 478 } 479 } 480 481 // Initialize the relocation buffers 482 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 483 set_scratch_locs_memory(locs_buf); 484 } 485 486 487 //-----------------------scratch_emit_size------------------------------------- 488 // Helper function that computes size by emitting code 489 uint Compile::scratch_emit_size(const Node* n) { 490 // Start scratch_emit_size section. 491 set_in_scratch_emit_size(true); 492 493 // Emit into a trash buffer and count bytes emitted. 494 // This is a pretty expensive way to compute a size, 495 // but it works well enough if seldom used. 496 // All common fixed-size instructions are given a size 497 // method by the AD file. 498 // Note that the scratch buffer blob and locs memory are 499 // allocated at the beginning of the compile task, and 500 // may be shared by several calls to scratch_emit_size. 501 // The allocation of the scratch buffer blob is particularly 502 // expensive, since it has to grab the code cache lock. 503 BufferBlob* blob = this->scratch_buffer_blob(); 504 assert(blob != NULL, "Initialize BufferBlob at start"); 505 assert(blob->size() > MAX_inst_size, "sanity"); 506 relocInfo* locs_buf = scratch_locs_memory(); 507 address blob_begin = blob->content_begin(); 508 address blob_end = (address)locs_buf; 509 assert(blob->content_contains(blob_end), "sanity"); 510 CodeBuffer buf(blob_begin, blob_end - blob_begin); 511 buf.initialize_consts_size(_scratch_const_size); 512 buf.initialize_stubs_size(MAX_stubs_size); 513 assert(locs_buf != NULL, "sanity"); 514 int lsize = MAX_locs_size / 3; 515 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 516 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 517 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 518 519 // Do the emission. 520 n->emit(buf, this->regalloc()); 521 522 // End scratch_emit_size section. 523 set_in_scratch_emit_size(false); 524 525 return buf.insts_size(); 526 } 527 528 529 // ============================================================================ 530 //------------------------------Compile standard------------------------------- 531 debug_only( int Compile::_debug_idx = 100000; ) 532 533 // Compile a method. entry_bci is -1 for normal compilations and indicates 534 // the continuation bci for on stack replacement. 535 536 537 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 538 : Phase(Compiler), 539 _env(ci_env), 540 _log(ci_env->log()), 541 _compile_id(ci_env->compile_id()), 542 _save_argument_registers(false), 543 _stub_name(NULL), 544 _stub_function(NULL), 545 _stub_entry_point(NULL), 546 _method(target), 547 _entry_bci(osr_bci), 548 _initial_gvn(NULL), 549 _for_igvn(NULL), 550 _warm_calls(NULL), 551 _subsume_loads(subsume_loads), 552 _do_escape_analysis(do_escape_analysis), 553 _failure_reason(NULL), 554 _code_buffer("Compile::Fill_buffer"), 555 _orig_pc_slot(0), 556 _orig_pc_slot_offset_in_bytes(0), 557 _has_method_handle_invokes(false), 558 _mach_constant_base_node(NULL), 559 _node_bundling_limit(0), 560 _node_bundling_base(NULL), 561 _java_calls(0), 562 _inner_loops(0), 563 _scratch_const_size(-1), 564 _in_scratch_emit_size(false), 565 #ifndef PRODUCT 566 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 567 _printer(IdealGraphPrinter::printer()), 568 #endif 569 _congraph(NULL) { 570 C = this; 571 572 CompileWrapper cw(this); 573 #ifndef PRODUCT 574 if (TimeCompiler2) { 575 tty->print(" "); 576 target->holder()->name()->print(); 577 tty->print("."); 578 target->print_short_name(); 579 tty->print(" "); 580 } 581 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 582 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 583 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 584 if (!print_opto_assembly) { 585 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 586 if (print_assembly && !Disassembler::can_decode()) { 587 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 588 print_opto_assembly = true; 589 } 590 } 591 set_print_assembly(print_opto_assembly); 592 set_parsed_irreducible_loop(false); 593 #endif 594 595 if (ProfileTraps) { 596 // Make sure the method being compiled gets its own MDO, 597 // so we can at least track the decompile_count(). 598 method()->ensure_method_data(); 599 } 600 601 Init(::AliasLevel); 602 603 604 print_compile_messages(); 605 606 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 607 _ilt = InlineTree::build_inline_tree_root(); 608 else 609 _ilt = NULL; 610 611 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 612 assert(num_alias_types() >= AliasIdxRaw, ""); 613 614 #define MINIMUM_NODE_HASH 1023 615 // Node list that Iterative GVN will start with 616 Unique_Node_List for_igvn(comp_arena()); 617 set_for_igvn(&for_igvn); 618 619 // GVN that will be run immediately on new nodes 620 uint estimated_size = method()->code_size()*4+64; 621 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 622 PhaseGVN gvn(node_arena(), estimated_size); 623 set_initial_gvn(&gvn); 624 625 { // Scope for timing the parser 626 TracePhase t3("parse", &_t_parser, true); 627 628 // Put top into the hash table ASAP. 629 initial_gvn()->transform_no_reclaim(top()); 630 631 // Set up tf(), start(), and find a CallGenerator. 632 CallGenerator* cg; 633 if (is_osr_compilation()) { 634 const TypeTuple *domain = StartOSRNode::osr_domain(); 635 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 636 init_tf(TypeFunc::make(domain, range)); 637 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 638 initial_gvn()->set_type_bottom(s); 639 init_start(s); 640 cg = CallGenerator::for_osr(method(), entry_bci()); 641 } else { 642 // Normal case. 643 init_tf(TypeFunc::make(method())); 644 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 645 initial_gvn()->set_type_bottom(s); 646 init_start(s); 647 float past_uses = method()->interpreter_invocation_count(); 648 float expected_uses = past_uses; 649 cg = CallGenerator::for_inline(method(), expected_uses); 650 } 651 if (failing()) return; 652 if (cg == NULL) { 653 record_method_not_compilable_all_tiers("cannot parse method"); 654 return; 655 } 656 JVMState* jvms = build_start_state(start(), tf()); 657 if ((jvms = cg->generate(jvms)) == NULL) { 658 record_method_not_compilable("method parse failed"); 659 return; 660 } 661 GraphKit kit(jvms); 662 663 if (!kit.stopped()) { 664 // Accept return values, and transfer control we know not where. 665 // This is done by a special, unique ReturnNode bound to root. 666 return_values(kit.jvms()); 667 } 668 669 if (kit.has_exceptions()) { 670 // Any exceptions that escape from this call must be rethrown 671 // to whatever caller is dynamically above us on the stack. 672 // This is done by a special, unique RethrowNode bound to root. 673 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 674 } 675 676 if (!failing() && has_stringbuilder()) { 677 { 678 // remove useless nodes to make the usage analysis simpler 679 ResourceMark rm; 680 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 681 } 682 683 { 684 ResourceMark rm; 685 print_method("Before StringOpts", 3); 686 PhaseStringOpts pso(initial_gvn(), &for_igvn); 687 print_method("After StringOpts", 3); 688 } 689 690 // now inline anything that we skipped the first time around 691 while (_late_inlines.length() > 0) { 692 CallGenerator* cg = _late_inlines.pop(); 693 cg->do_late_inline(); 694 } 695 } 696 assert(_late_inlines.length() == 0, "should have been processed"); 697 698 print_method("Before RemoveUseless", 3); 699 700 // Remove clutter produced by parsing. 701 if (!failing()) { 702 ResourceMark rm; 703 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 704 } 705 } 706 707 // Note: Large methods are capped off in do_one_bytecode(). 708 if (failing()) return; 709 710 // After parsing, node notes are no longer automagic. 711 // They must be propagated by register_new_node_with_optimizer(), 712 // clone(), or the like. 713 set_default_node_notes(NULL); 714 715 for (;;) { 716 int successes = Inline_Warm(); 717 if (failing()) return; 718 if (successes == 0) break; 719 } 720 721 // Drain the list. 722 Finish_Warm(); 723 #ifndef PRODUCT 724 if (_printer) { 725 _printer->print_inlining(this); 726 } 727 #endif 728 729 if (failing()) return; 730 NOT_PRODUCT( verify_graph_edges(); ) 731 732 // Now optimize 733 Optimize(); 734 if (failing()) return; 735 NOT_PRODUCT( verify_graph_edges(); ) 736 737 #ifndef PRODUCT 738 if (PrintIdeal) { 739 ttyLocker ttyl; // keep the following output all in one block 740 // This output goes directly to the tty, not the compiler log. 741 // To enable tools to match it up with the compilation activity, 742 // be sure to tag this tty output with the compile ID. 743 if (xtty != NULL) { 744 xtty->head("ideal compile_id='%d'%s", compile_id(), 745 is_osr_compilation() ? " compile_kind='osr'" : 746 ""); 747 } 748 root()->dump(9999); 749 if (xtty != NULL) { 750 xtty->tail("ideal"); 751 } 752 } 753 #endif 754 755 // Now that we know the size of all the monitors we can add a fixed slot 756 // for the original deopt pc. 757 758 _orig_pc_slot = fixed_slots(); 759 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 760 set_fixed_slots(next_slot); 761 762 // Now generate code 763 Code_Gen(); 764 if (failing()) return; 765 766 // Check if we want to skip execution of all compiled code. 767 { 768 #ifndef PRODUCT 769 if (OptoNoExecute) { 770 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 771 return; 772 } 773 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 774 #endif 775 776 if (is_osr_compilation()) { 777 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 778 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 779 } else { 780 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 781 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 782 } 783 784 env()->register_method(_method, _entry_bci, 785 &_code_offsets, 786 _orig_pc_slot_offset_in_bytes, 787 code_buffer(), 788 frame_size_in_words(), _oop_map_set, 789 &_handler_table, &_inc_table, 790 compiler, 791 env()->comp_level(), 792 true, /*has_debug_info*/ 793 has_unsafe_access() 794 ); 795 } 796 } 797 798 //------------------------------Compile---------------------------------------- 799 // Compile a runtime stub 800 Compile::Compile( ciEnv* ci_env, 801 TypeFunc_generator generator, 802 address stub_function, 803 const char *stub_name, 804 int is_fancy_jump, 805 bool pass_tls, 806 bool save_arg_registers, 807 bool return_pc ) 808 : Phase(Compiler), 809 _env(ci_env), 810 _log(ci_env->log()), 811 _compile_id(-1), 812 _save_argument_registers(save_arg_registers), 813 _method(NULL), 814 _stub_name(stub_name), 815 _stub_function(stub_function), 816 _stub_entry_point(NULL), 817 _entry_bci(InvocationEntryBci), 818 _initial_gvn(NULL), 819 _for_igvn(NULL), 820 _warm_calls(NULL), 821 _orig_pc_slot(0), 822 _orig_pc_slot_offset_in_bytes(0), 823 _subsume_loads(true), 824 _do_escape_analysis(false), 825 _failure_reason(NULL), 826 _code_buffer("Compile::Fill_buffer"), 827 _has_method_handle_invokes(false), 828 _mach_constant_base_node(NULL), 829 _node_bundling_limit(0), 830 _node_bundling_base(NULL), 831 _java_calls(0), 832 _inner_loops(0), 833 #ifndef PRODUCT 834 _trace_opto_output(TraceOptoOutput), 835 _printer(NULL), 836 #endif 837 _congraph(NULL) { 838 C = this; 839 840 #ifndef PRODUCT 841 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 842 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 843 set_print_assembly(PrintFrameConverterAssembly); 844 set_parsed_irreducible_loop(false); 845 #endif 846 CompileWrapper cw(this); 847 Init(/*AliasLevel=*/ 0); 848 init_tf((*generator)()); 849 850 { 851 // The following is a dummy for the sake of GraphKit::gen_stub 852 Unique_Node_List for_igvn(comp_arena()); 853 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 854 PhaseGVN gvn(Thread::current()->resource_area(),255); 855 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 856 gvn.transform_no_reclaim(top()); 857 858 GraphKit kit; 859 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 860 } 861 862 NOT_PRODUCT( verify_graph_edges(); ) 863 Code_Gen(); 864 if (failing()) return; 865 866 867 // Entry point will be accessed using compile->stub_entry_point(); 868 if (code_buffer() == NULL) { 869 Matcher::soft_match_failure(); 870 } else { 871 if (PrintAssembly && (WizardMode || Verbose)) 872 tty->print_cr("### Stub::%s", stub_name); 873 874 if (!failing()) { 875 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 876 877 // Make the NMethod 878 // For now we mark the frame as never safe for profile stackwalking 879 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 880 code_buffer(), 881 CodeOffsets::frame_never_safe, 882 // _code_offsets.value(CodeOffsets::Frame_Complete), 883 frame_size_in_words(), 884 _oop_map_set, 885 save_arg_registers); 886 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 887 888 _stub_entry_point = rs->entry_point(); 889 } 890 } 891 } 892 893 #ifndef PRODUCT 894 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 895 if(PrintOpto && Verbose) { 896 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 897 } 898 } 899 #endif 900 901 void Compile::print_codes() { 902 } 903 904 //------------------------------Init------------------------------------------- 905 // Prepare for a single compilation 906 void Compile::Init(int aliaslevel) { 907 _unique = 0; 908 _regalloc = NULL; 909 910 _tf = NULL; // filled in later 911 _top = NULL; // cached later 912 _matcher = NULL; // filled in later 913 _cfg = NULL; // filled in later 914 915 set_24_bit_selection_and_mode(Use24BitFP, false); 916 917 _node_note_array = NULL; 918 _default_node_notes = NULL; 919 920 _immutable_memory = NULL; // filled in at first inquiry 921 922 // Globally visible Nodes 923 // First set TOP to NULL to give safe behavior during creation of RootNode 924 set_cached_top_node(NULL); 925 set_root(new (this, 3) RootNode()); 926 // Now that you have a Root to point to, create the real TOP 927 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 928 set_recent_alloc(NULL, NULL); 929 930 // Create Debug Information Recorder to record scopes, oopmaps, etc. 931 env()->set_oop_recorder(new OopRecorder(comp_arena())); 932 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 933 env()->set_dependencies(new Dependencies(env())); 934 935 _fixed_slots = 0; 936 set_has_split_ifs(false); 937 set_has_loops(has_method() && method()->has_loops()); // first approximation 938 set_has_stringbuilder(false); 939 _trap_can_recompile = false; // no traps emitted yet 940 _major_progress = true; // start out assuming good things will happen 941 set_has_unsafe_access(false); 942 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 943 set_decompile_count(0); 944 945 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 946 set_num_loop_opts(LoopOptsCount); 947 set_do_inlining(Inline); 948 set_max_inline_size(MaxInlineSize); 949 set_freq_inline_size(FreqInlineSize); 950 set_do_scheduling(OptoScheduling); 951 set_do_count_invocations(false); 952 set_do_method_data_update(false); 953 954 if (debug_info()->recording_non_safepoints()) { 955 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 956 (comp_arena(), 8, 0, NULL)); 957 set_default_node_notes(Node_Notes::make(this)); 958 } 959 960 // // -- Initialize types before each compile -- 961 // // Update cached type information 962 // if( _method && _method->constants() ) 963 // Type::update_loaded_types(_method, _method->constants()); 964 965 // Init alias_type map. 966 if (!_do_escape_analysis && aliaslevel == 3) 967 aliaslevel = 2; // No unique types without escape analysis 968 _AliasLevel = aliaslevel; 969 const int grow_ats = 16; 970 _max_alias_types = grow_ats; 971 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 972 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 973 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 974 { 975 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 976 } 977 // Initialize the first few types. 978 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 979 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 980 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 981 _num_alias_types = AliasIdxRaw+1; 982 // Zero out the alias type cache. 983 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 984 // A NULL adr_type hits in the cache right away. Preload the right answer. 985 probe_alias_cache(NULL)->_index = AliasIdxTop; 986 987 _intrinsics = NULL; 988 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 989 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 990 register_library_intrinsics(); 991 } 992 993 //---------------------------init_start---------------------------------------- 994 // Install the StartNode on this compile object. 995 void Compile::init_start(StartNode* s) { 996 if (failing()) 997 return; // already failing 998 assert(s == start(), ""); 999 } 1000 1001 StartNode* Compile::start() const { 1002 assert(!failing(), ""); 1003 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1004 Node* start = root()->fast_out(i); 1005 if( start->is_Start() ) 1006 return start->as_Start(); 1007 } 1008 ShouldNotReachHere(); 1009 return NULL; 1010 } 1011 1012 //-------------------------------immutable_memory------------------------------------- 1013 // Access immutable memory 1014 Node* Compile::immutable_memory() { 1015 if (_immutable_memory != NULL) { 1016 return _immutable_memory; 1017 } 1018 StartNode* s = start(); 1019 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1020 Node *p = s->fast_out(i); 1021 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1022 _immutable_memory = p; 1023 return _immutable_memory; 1024 } 1025 } 1026 ShouldNotReachHere(); 1027 return NULL; 1028 } 1029 1030 //----------------------set_cached_top_node------------------------------------ 1031 // Install the cached top node, and make sure Node::is_top works correctly. 1032 void Compile::set_cached_top_node(Node* tn) { 1033 if (tn != NULL) verify_top(tn); 1034 Node* old_top = _top; 1035 _top = tn; 1036 // Calling Node::setup_is_top allows the nodes the chance to adjust 1037 // their _out arrays. 1038 if (_top != NULL) _top->setup_is_top(); 1039 if (old_top != NULL) old_top->setup_is_top(); 1040 assert(_top == NULL || top()->is_top(), ""); 1041 } 1042 1043 #ifndef PRODUCT 1044 void Compile::verify_top(Node* tn) const { 1045 if (tn != NULL) { 1046 assert(tn->is_Con(), "top node must be a constant"); 1047 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1048 assert(tn->in(0) != NULL, "must have live top node"); 1049 } 1050 } 1051 #endif 1052 1053 1054 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1055 1056 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1057 guarantee(arr != NULL, ""); 1058 int num_blocks = arr->length(); 1059 if (grow_by < num_blocks) grow_by = num_blocks; 1060 int num_notes = grow_by * _node_notes_block_size; 1061 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1062 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1063 while (num_notes > 0) { 1064 arr->append(notes); 1065 notes += _node_notes_block_size; 1066 num_notes -= _node_notes_block_size; 1067 } 1068 assert(num_notes == 0, "exact multiple, please"); 1069 } 1070 1071 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1072 if (source == NULL || dest == NULL) return false; 1073 1074 if (dest->is_Con()) 1075 return false; // Do not push debug info onto constants. 1076 1077 #ifdef ASSERT 1078 // Leave a bread crumb trail pointing to the original node: 1079 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1080 dest->set_debug_orig(source); 1081 } 1082 #endif 1083 1084 if (node_note_array() == NULL) 1085 return false; // Not collecting any notes now. 1086 1087 // This is a copy onto a pre-existing node, which may already have notes. 1088 // If both nodes have notes, do not overwrite any pre-existing notes. 1089 Node_Notes* source_notes = node_notes_at(source->_idx); 1090 if (source_notes == NULL || source_notes->is_clear()) return false; 1091 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1092 if (dest_notes == NULL || dest_notes->is_clear()) { 1093 return set_node_notes_at(dest->_idx, source_notes); 1094 } 1095 1096 Node_Notes merged_notes = (*source_notes); 1097 // The order of operations here ensures that dest notes will win... 1098 merged_notes.update_from(dest_notes); 1099 return set_node_notes_at(dest->_idx, &merged_notes); 1100 } 1101 1102 1103 //--------------------------allow_range_check_smearing------------------------- 1104 // Gating condition for coalescing similar range checks. 1105 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1106 // single covering check that is at least as strong as any of them. 1107 // If the optimization succeeds, the simplified (strengthened) range check 1108 // will always succeed. If it fails, we will deopt, and then give up 1109 // on the optimization. 1110 bool Compile::allow_range_check_smearing() const { 1111 // If this method has already thrown a range-check, 1112 // assume it was because we already tried range smearing 1113 // and it failed. 1114 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1115 return !already_trapped; 1116 } 1117 1118 1119 //------------------------------flatten_alias_type----------------------------- 1120 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1121 int offset = tj->offset(); 1122 TypePtr::PTR ptr = tj->ptr(); 1123 1124 // Known instance (scalarizable allocation) alias only with itself. 1125 bool is_known_inst = tj->isa_oopptr() != NULL && 1126 tj->is_oopptr()->is_known_instance(); 1127 1128 // Process weird unsafe references. 1129 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1130 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1131 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1132 tj = TypeOopPtr::BOTTOM; 1133 ptr = tj->ptr(); 1134 offset = tj->offset(); 1135 } 1136 1137 // Array pointers need some flattening 1138 const TypeAryPtr *ta = tj->isa_aryptr(); 1139 if( ta && is_known_inst ) { 1140 if ( offset != Type::OffsetBot && 1141 offset > arrayOopDesc::length_offset_in_bytes() ) { 1142 offset = Type::OffsetBot; // Flatten constant access into array body only 1143 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1144 } 1145 } else if( ta && _AliasLevel >= 2 ) { 1146 // For arrays indexed by constant indices, we flatten the alias 1147 // space to include all of the array body. Only the header, klass 1148 // and array length can be accessed un-aliased. 1149 if( offset != Type::OffsetBot ) { 1150 if( ta->const_oop() ) { // methodDataOop or methodOop 1151 offset = Type::OffsetBot; // Flatten constant access into array body 1152 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1153 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1154 // range is OK as-is. 1155 tj = ta = TypeAryPtr::RANGE; 1156 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1157 tj = TypeInstPtr::KLASS; // all klass loads look alike 1158 ta = TypeAryPtr::RANGE; // generic ignored junk 1159 ptr = TypePtr::BotPTR; 1160 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1161 tj = TypeInstPtr::MARK; 1162 ta = TypeAryPtr::RANGE; // generic ignored junk 1163 ptr = TypePtr::BotPTR; 1164 } else { // Random constant offset into array body 1165 offset = Type::OffsetBot; // Flatten constant access into array body 1166 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1167 } 1168 } 1169 // Arrays of fixed size alias with arrays of unknown size. 1170 if (ta->size() != TypeInt::POS) { 1171 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1172 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1173 } 1174 // Arrays of known objects become arrays of unknown objects. 1175 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1176 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1177 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1178 } 1179 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1180 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1181 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1182 } 1183 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1184 // cannot be distinguished by bytecode alone. 1185 if (ta->elem() == TypeInt::BOOL) { 1186 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1187 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1188 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1189 } 1190 // During the 2nd round of IterGVN, NotNull castings are removed. 1191 // Make sure the Bottom and NotNull variants alias the same. 1192 // Also, make sure exact and non-exact variants alias the same. 1193 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1194 if (ta->const_oop()) { 1195 tj = ta = TypeAryPtr::make(TypePtr::Constant,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1196 } else { 1197 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1198 } 1199 } 1200 } 1201 1202 // Oop pointers need some flattening 1203 const TypeInstPtr *to = tj->isa_instptr(); 1204 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1205 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1206 if( ptr == TypePtr::Constant ) { 1207 if (to->klass() != ciEnv::current()->Class_klass() || 1208 offset < k->size_helper() * wordSize) { 1209 // No constant oop pointers (such as Strings); they alias with 1210 // unknown strings. 1211 assert(!is_known_inst, "not scalarizable allocation"); 1212 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1213 } 1214 } else if( is_known_inst ) { 1215 tj = to; // Keep NotNull and klass_is_exact for instance type 1216 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1217 // During the 2nd round of IterGVN, NotNull castings are removed. 1218 // Make sure the Bottom and NotNull variants alias the same. 1219 // Also, make sure exact and non-exact variants alias the same. 1220 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1221 } 1222 // Canonicalize the holder of this field 1223 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1224 // First handle header references such as a LoadKlassNode, even if the 1225 // object's klass is unloaded at compile time (4965979). 1226 if (!is_known_inst) { // Do it only for non-instance types 1227 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1228 } 1229 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1230 // Static fields are in the space above the normal instance 1231 // fields in the java.lang.Class instance. 1232 if (to->klass() != ciEnv::current()->Class_klass()) { 1233 to = NULL; 1234 tj = TypeOopPtr::BOTTOM; 1235 offset = tj->offset(); 1236 } 1237 } else { 1238 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1239 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1240 if( is_known_inst ) { 1241 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1242 } else { 1243 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1244 } 1245 } 1246 } 1247 } 1248 1249 // Klass pointers to object array klasses need some flattening 1250 const TypeKlassPtr *tk = tj->isa_klassptr(); 1251 if( tk ) { 1252 // If we are referencing a field within a Klass, we need 1253 // to assume the worst case of an Object. Both exact and 1254 // inexact types must flatten to the same alias class. 1255 // Since the flattened result for a klass is defined to be 1256 // precisely java.lang.Object, use a constant ptr. 1257 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1258 1259 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1260 TypeKlassPtr::OBJECT->klass(), 1261 offset); 1262 } 1263 1264 ciKlass* klass = tk->klass(); 1265 if( klass->is_obj_array_klass() ) { 1266 ciKlass* k = TypeAryPtr::OOPS->klass(); 1267 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1268 k = TypeInstPtr::BOTTOM->klass(); 1269 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1270 } 1271 1272 // Check for precise loads from the primary supertype array and force them 1273 // to the supertype cache alias index. Check for generic array loads from 1274 // the primary supertype array and also force them to the supertype cache 1275 // alias index. Since the same load can reach both, we need to merge 1276 // these 2 disparate memories into the same alias class. Since the 1277 // primary supertype array is read-only, there's no chance of confusion 1278 // where we bypass an array load and an array store. 1279 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1280 if( offset == Type::OffsetBot || 1281 off2 < Klass::primary_super_limit()*wordSize ) { 1282 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1283 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1284 } 1285 } 1286 1287 // Flatten all Raw pointers together. 1288 if (tj->base() == Type::RawPtr) 1289 tj = TypeRawPtr::BOTTOM; 1290 1291 if (tj->base() == Type::AnyPtr) 1292 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1293 1294 // Flatten all to bottom for now 1295 switch( _AliasLevel ) { 1296 case 0: 1297 tj = TypePtr::BOTTOM; 1298 break; 1299 case 1: // Flatten to: oop, static, field or array 1300 switch (tj->base()) { 1301 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1302 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1303 case Type::AryPtr: // do not distinguish arrays at all 1304 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1305 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1306 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1307 default: ShouldNotReachHere(); 1308 } 1309 break; 1310 case 2: // No collapsing at level 2; keep all splits 1311 case 3: // No collapsing at level 3; keep all splits 1312 break; 1313 default: 1314 Unimplemented(); 1315 } 1316 1317 offset = tj->offset(); 1318 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1319 1320 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1321 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1322 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1323 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1324 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1325 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1326 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1327 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1328 assert( tj->ptr() != TypePtr::TopPTR && 1329 tj->ptr() != TypePtr::AnyNull && 1330 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1331 // assert( tj->ptr() != TypePtr::Constant || 1332 // tj->base() == Type::RawPtr || 1333 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1334 1335 return tj; 1336 } 1337 1338 void Compile::AliasType::Init(int i, const TypePtr* at) { 1339 _index = i; 1340 _adr_type = at; 1341 _field = NULL; 1342 _is_rewritable = true; // default 1343 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1344 if (atoop != NULL && atoop->is_known_instance()) { 1345 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1346 _general_index = Compile::current()->get_alias_index(gt); 1347 } else { 1348 _general_index = 0; 1349 } 1350 } 1351 1352 //---------------------------------print_on------------------------------------ 1353 #ifndef PRODUCT 1354 void Compile::AliasType::print_on(outputStream* st) { 1355 if (index() < 10) 1356 st->print("@ <%d> ", index()); 1357 else st->print("@ <%d>", index()); 1358 st->print(is_rewritable() ? " " : " RO"); 1359 int offset = adr_type()->offset(); 1360 if (offset == Type::OffsetBot) 1361 st->print(" +any"); 1362 else st->print(" +%-3d", offset); 1363 st->print(" in "); 1364 adr_type()->dump_on(st); 1365 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1366 if (field() != NULL && tjp) { 1367 if (tjp->klass() != field()->holder() || 1368 tjp->offset() != field()->offset_in_bytes()) { 1369 st->print(" != "); 1370 field()->print(); 1371 st->print(" ***"); 1372 } 1373 } 1374 } 1375 1376 void print_alias_types() { 1377 Compile* C = Compile::current(); 1378 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1379 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1380 C->alias_type(idx)->print_on(tty); 1381 tty->cr(); 1382 } 1383 } 1384 #endif 1385 1386 1387 //----------------------------probe_alias_cache-------------------------------- 1388 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1389 intptr_t key = (intptr_t) adr_type; 1390 key ^= key >> logAliasCacheSize; 1391 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1392 } 1393 1394 1395 //-----------------------------grow_alias_types-------------------------------- 1396 void Compile::grow_alias_types() { 1397 const int old_ats = _max_alias_types; // how many before? 1398 const int new_ats = old_ats; // how many more? 1399 const int grow_ats = old_ats+new_ats; // how many now? 1400 _max_alias_types = grow_ats; 1401 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1402 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1403 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1404 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1405 } 1406 1407 1408 //--------------------------------find_alias_type------------------------------ 1409 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1410 if (_AliasLevel == 0) 1411 return alias_type(AliasIdxBot); 1412 1413 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1414 if (ace->_adr_type == adr_type) { 1415 return alias_type(ace->_index); 1416 } 1417 1418 // Handle special cases. 1419 if (adr_type == NULL) return alias_type(AliasIdxTop); 1420 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1421 1422 // Do it the slow way. 1423 const TypePtr* flat = flatten_alias_type(adr_type); 1424 1425 #ifdef ASSERT 1426 assert(flat == flatten_alias_type(flat), "idempotent"); 1427 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1428 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1429 const TypeOopPtr* foop = flat->is_oopptr(); 1430 // Scalarizable allocations have exact klass always. 1431 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1432 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1433 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1434 } 1435 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1436 #endif 1437 1438 int idx = AliasIdxTop; 1439 for (int i = 0; i < num_alias_types(); i++) { 1440 if (alias_type(i)->adr_type() == flat) { 1441 idx = i; 1442 break; 1443 } 1444 } 1445 1446 if (idx == AliasIdxTop) { 1447 if (no_create) return NULL; 1448 // Grow the array if necessary. 1449 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1450 // Add a new alias type. 1451 idx = _num_alias_types++; 1452 _alias_types[idx]->Init(idx, flat); 1453 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1454 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1455 if (flat->isa_instptr()) { 1456 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1457 && flat->is_instptr()->klass() == env()->Class_klass()) 1458 alias_type(idx)->set_rewritable(false); 1459 } 1460 if (flat->isa_klassptr()) { 1461 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1462 alias_type(idx)->set_rewritable(false); 1463 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1464 alias_type(idx)->set_rewritable(false); 1465 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1466 alias_type(idx)->set_rewritable(false); 1467 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1468 alias_type(idx)->set_rewritable(false); 1469 } 1470 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1471 // but the base pointer type is not distinctive enough to identify 1472 // references into JavaThread.) 1473 1474 // Check for final fields. 1475 const TypeInstPtr* tinst = flat->isa_instptr(); 1476 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1477 ciField* field; 1478 if (tinst->const_oop() != NULL && 1479 tinst->klass() == ciEnv::current()->Class_klass() && 1480 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1481 // static field 1482 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1483 field = k->get_field_by_offset(tinst->offset(), true); 1484 } else { 1485 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1486 field = k->get_field_by_offset(tinst->offset(), false); 1487 } 1488 assert(field == NULL || 1489 original_field == NULL || 1490 (field->holder() == original_field->holder() && 1491 field->offset() == original_field->offset() && 1492 field->is_static() == original_field->is_static()), "wrong field?"); 1493 // Set field() and is_rewritable() attributes. 1494 if (field != NULL) alias_type(idx)->set_field(field); 1495 } 1496 } 1497 1498 // Fill the cache for next time. 1499 ace->_adr_type = adr_type; 1500 ace->_index = idx; 1501 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1502 1503 // Might as well try to fill the cache for the flattened version, too. 1504 AliasCacheEntry* face = probe_alias_cache(flat); 1505 if (face->_adr_type == NULL) { 1506 face->_adr_type = flat; 1507 face->_index = idx; 1508 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1509 } 1510 1511 return alias_type(idx); 1512 } 1513 1514 1515 Compile::AliasType* Compile::alias_type(ciField* field) { 1516 const TypeOopPtr* t; 1517 if (field->is_static()) 1518 t = TypeInstPtr::make(field->holder()->java_mirror()); 1519 else 1520 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1521 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1522 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1523 return atp; 1524 } 1525 1526 1527 //------------------------------have_alias_type-------------------------------- 1528 bool Compile::have_alias_type(const TypePtr* adr_type) { 1529 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1530 if (ace->_adr_type == adr_type) { 1531 return true; 1532 } 1533 1534 // Handle special cases. 1535 if (adr_type == NULL) return true; 1536 if (adr_type == TypePtr::BOTTOM) return true; 1537 1538 return find_alias_type(adr_type, true, NULL) != NULL; 1539 } 1540 1541 //-----------------------------must_alias-------------------------------------- 1542 // True if all values of the given address type are in the given alias category. 1543 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1544 if (alias_idx == AliasIdxBot) return true; // the universal category 1545 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1546 if (alias_idx == AliasIdxTop) return false; // the empty category 1547 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1548 1549 // the only remaining possible overlap is identity 1550 int adr_idx = get_alias_index(adr_type); 1551 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1552 assert(adr_idx == alias_idx || 1553 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1554 && adr_type != TypeOopPtr::BOTTOM), 1555 "should not be testing for overlap with an unsafe pointer"); 1556 return adr_idx == alias_idx; 1557 } 1558 1559 //------------------------------can_alias-------------------------------------- 1560 // True if any values of the given address type are in the given alias category. 1561 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1562 if (alias_idx == AliasIdxTop) return false; // the empty category 1563 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1564 if (alias_idx == AliasIdxBot) return true; // the universal category 1565 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1566 1567 // the only remaining possible overlap is identity 1568 int adr_idx = get_alias_index(adr_type); 1569 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1570 return adr_idx == alias_idx; 1571 } 1572 1573 1574 1575 //---------------------------pop_warm_call------------------------------------- 1576 WarmCallInfo* Compile::pop_warm_call() { 1577 WarmCallInfo* wci = _warm_calls; 1578 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1579 return wci; 1580 } 1581 1582 //----------------------------Inline_Warm-------------------------------------- 1583 int Compile::Inline_Warm() { 1584 // If there is room, try to inline some more warm call sites. 1585 // %%% Do a graph index compaction pass when we think we're out of space? 1586 if (!InlineWarmCalls) return 0; 1587 1588 int calls_made_hot = 0; 1589 int room_to_grow = NodeCountInliningCutoff - unique(); 1590 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1591 int amount_grown = 0; 1592 WarmCallInfo* call; 1593 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1594 int est_size = (int)call->size(); 1595 if (est_size > (room_to_grow - amount_grown)) { 1596 // This one won't fit anyway. Get rid of it. 1597 call->make_cold(); 1598 continue; 1599 } 1600 call->make_hot(); 1601 calls_made_hot++; 1602 amount_grown += est_size; 1603 amount_to_grow -= est_size; 1604 } 1605 1606 if (calls_made_hot > 0) set_major_progress(); 1607 return calls_made_hot; 1608 } 1609 1610 1611 //----------------------------Finish_Warm-------------------------------------- 1612 void Compile::Finish_Warm() { 1613 if (!InlineWarmCalls) return; 1614 if (failing()) return; 1615 if (warm_calls() == NULL) return; 1616 1617 // Clean up loose ends, if we are out of space for inlining. 1618 WarmCallInfo* call; 1619 while ((call = pop_warm_call()) != NULL) { 1620 call->make_cold(); 1621 } 1622 } 1623 1624 //---------------------cleanup_loop_predicates----------------------- 1625 // Remove the opaque nodes that protect the predicates so that all unused 1626 // checks and uncommon_traps will be eliminated from the ideal graph 1627 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1628 if (predicate_count()==0) return; 1629 for (int i = predicate_count(); i > 0; i--) { 1630 Node * n = predicate_opaque1_node(i-1); 1631 assert(n->Opcode() == Op_Opaque1, "must be"); 1632 igvn.replace_node(n, n->in(1)); 1633 } 1634 assert(predicate_count()==0, "should be clean!"); 1635 igvn.optimize(); 1636 } 1637 1638 //------------------------------Optimize--------------------------------------- 1639 // Given a graph, optimize it. 1640 void Compile::Optimize() { 1641 TracePhase t1("optimizer", &_t_optimizer, true); 1642 1643 #ifndef PRODUCT 1644 if (env()->break_at_compile()) { 1645 BREAKPOINT; 1646 } 1647 1648 #endif 1649 1650 ResourceMark rm; 1651 int loop_opts_cnt; 1652 1653 NOT_PRODUCT( verify_graph_edges(); ) 1654 1655 print_method("After Parsing"); 1656 1657 { 1658 // Iterative Global Value Numbering, including ideal transforms 1659 // Initialize IterGVN with types and values from parse-time GVN 1660 PhaseIterGVN igvn(initial_gvn()); 1661 { 1662 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1663 igvn.optimize(); 1664 } 1665 1666 print_method("Iter GVN 1", 2); 1667 1668 if (failing()) return; 1669 1670 // Perform escape analysis 1671 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1672 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 1673 ConnectionGraph::do_analysis(this, &igvn); 1674 1675 if (failing()) return; 1676 1677 igvn.optimize(); 1678 print_method("Iter GVN 3", 2); 1679 1680 if (failing()) return; 1681 1682 } 1683 1684 // Loop transforms on the ideal graph. Range Check Elimination, 1685 // peeling, unrolling, etc. 1686 1687 // Set loop opts counter 1688 loop_opts_cnt = num_loop_opts(); 1689 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1690 { 1691 TracePhase t2("idealLoop", &_t_idealLoop, true); 1692 PhaseIdealLoop ideal_loop( igvn, true, UseLoopPredicate); 1693 loop_opts_cnt--; 1694 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1695 if (failing()) return; 1696 } 1697 // Loop opts pass if partial peeling occurred in previous pass 1698 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1699 TracePhase t3("idealLoop", &_t_idealLoop, true); 1700 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1701 loop_opts_cnt--; 1702 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1703 if (failing()) return; 1704 } 1705 // Loop opts pass for loop-unrolling before CCP 1706 if(major_progress() && (loop_opts_cnt > 0)) { 1707 TracePhase t4("idealLoop", &_t_idealLoop, true); 1708 PhaseIdealLoop ideal_loop( igvn, false, UseLoopPredicate); 1709 loop_opts_cnt--; 1710 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1711 } 1712 if (!failing()) { 1713 // Verify that last round of loop opts produced a valid graph 1714 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1715 PhaseIdealLoop::verify(igvn); 1716 } 1717 } 1718 if (failing()) return; 1719 1720 // Conditional Constant Propagation; 1721 PhaseCCP ccp( &igvn ); 1722 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1723 { 1724 TracePhase t2("ccp", &_t_ccp, true); 1725 ccp.do_transform(); 1726 } 1727 print_method("PhaseCPP 1", 2); 1728 1729 assert( true, "Break here to ccp.dump_old2new_map()"); 1730 1731 // Iterative Global Value Numbering, including ideal transforms 1732 { 1733 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1734 igvn = ccp; 1735 igvn.optimize(); 1736 } 1737 1738 print_method("Iter GVN 2", 2); 1739 1740 if (failing()) return; 1741 1742 // Loop transforms on the ideal graph. Range Check Elimination, 1743 // peeling, unrolling, etc. 1744 if(loop_opts_cnt > 0) { 1745 debug_only( int cnt = 0; ); 1746 bool loop_predication = UseLoopPredicate; 1747 while(major_progress() && (loop_opts_cnt > 0)) { 1748 TracePhase t2("idealLoop", &_t_idealLoop, true); 1749 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1750 PhaseIdealLoop ideal_loop( igvn, true, loop_predication); 1751 loop_opts_cnt--; 1752 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1753 if (failing()) return; 1754 // Perform loop predication optimization during first iteration after CCP. 1755 // After that switch it off and cleanup unused loop predicates. 1756 if (loop_predication) { 1757 loop_predication = false; 1758 cleanup_loop_predicates(igvn); 1759 if (failing()) return; 1760 } 1761 } 1762 } 1763 1764 { 1765 // Verify that all previous optimizations produced a valid graph 1766 // at least to this point, even if no loop optimizations were done. 1767 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1768 PhaseIdealLoop::verify(igvn); 1769 } 1770 1771 { 1772 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1773 PhaseMacroExpand mex(igvn); 1774 if (mex.expand_macro_nodes()) { 1775 assert(failing(), "must bail out w/ explicit message"); 1776 return; 1777 } 1778 } 1779 1780 } // (End scope of igvn; run destructor if necessary for asserts.) 1781 1782 // A method with only infinite loops has no edges entering loops from root 1783 { 1784 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1785 if (final_graph_reshaping()) { 1786 assert(failing(), "must bail out w/ explicit message"); 1787 return; 1788 } 1789 } 1790 1791 print_method("Optimize finished", 2); 1792 } 1793 1794 1795 //------------------------------Code_Gen--------------------------------------- 1796 // Given a graph, generate code for it 1797 void Compile::Code_Gen() { 1798 if (failing()) return; 1799 1800 // Perform instruction selection. You might think we could reclaim Matcher 1801 // memory PDQ, but actually the Matcher is used in generating spill code. 1802 // Internals of the Matcher (including some VectorSets) must remain live 1803 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1804 // set a bit in reclaimed memory. 1805 1806 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1807 // nodes. Mapping is only valid at the root of each matched subtree. 1808 NOT_PRODUCT( verify_graph_edges(); ) 1809 1810 Node_List proj_list; 1811 Matcher m(proj_list); 1812 _matcher = &m; 1813 { 1814 TracePhase t2("matcher", &_t_matcher, true); 1815 m.match(); 1816 } 1817 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1818 // nodes. Mapping is only valid at the root of each matched subtree. 1819 NOT_PRODUCT( verify_graph_edges(); ) 1820 1821 // If you have too many nodes, or if matching has failed, bail out 1822 check_node_count(0, "out of nodes matching instructions"); 1823 if (failing()) return; 1824 1825 // Build a proper-looking CFG 1826 PhaseCFG cfg(node_arena(), root(), m); 1827 _cfg = &cfg; 1828 { 1829 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1830 cfg.Dominators(); 1831 if (failing()) return; 1832 1833 NOT_PRODUCT( verify_graph_edges(); ) 1834 1835 cfg.Estimate_Block_Frequency(); 1836 cfg.GlobalCodeMotion(m,unique(),proj_list); 1837 1838 print_method("Global code motion", 2); 1839 1840 if (failing()) return; 1841 NOT_PRODUCT( verify_graph_edges(); ) 1842 1843 debug_only( cfg.verify(); ) 1844 } 1845 NOT_PRODUCT( verify_graph_edges(); ) 1846 1847 PhaseChaitin regalloc(unique(),cfg,m); 1848 _regalloc = ®alloc; 1849 { 1850 TracePhase t2("regalloc", &_t_registerAllocation, true); 1851 // Perform any platform dependent preallocation actions. This is used, 1852 // for example, to avoid taking an implicit null pointer exception 1853 // using the frame pointer on win95. 1854 _regalloc->pd_preallocate_hook(); 1855 1856 // Perform register allocation. After Chaitin, use-def chains are 1857 // no longer accurate (at spill code) and so must be ignored. 1858 // Node->LRG->reg mappings are still accurate. 1859 _regalloc->Register_Allocate(); 1860 1861 // Bail out if the allocator builds too many nodes 1862 if (failing()) return; 1863 } 1864 1865 // Prior to register allocation we kept empty basic blocks in case the 1866 // the allocator needed a place to spill. After register allocation we 1867 // are not adding any new instructions. If any basic block is empty, we 1868 // can now safely remove it. 1869 { 1870 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1871 cfg.remove_empty(); 1872 if (do_freq_based_layout()) { 1873 PhaseBlockLayout layout(cfg); 1874 } else { 1875 cfg.set_loop_alignment(); 1876 } 1877 cfg.fixup_flow(); 1878 } 1879 1880 // Perform any platform dependent postallocation verifications. 1881 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1882 1883 // Apply peephole optimizations 1884 if( OptoPeephole ) { 1885 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1886 PhasePeephole peep( _regalloc, cfg); 1887 peep.do_transform(); 1888 } 1889 1890 // Convert Nodes to instruction bits in a buffer 1891 { 1892 // %%%% workspace merge brought two timers together for one job 1893 TracePhase t2a("output", &_t_output, true); 1894 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1895 Output(); 1896 } 1897 1898 print_method("Final Code"); 1899 1900 // He's dead, Jim. 1901 _cfg = (PhaseCFG*)0xdeadbeef; 1902 _regalloc = (PhaseChaitin*)0xdeadbeef; 1903 } 1904 1905 1906 //------------------------------dump_asm--------------------------------------- 1907 // Dump formatted assembly 1908 #ifndef PRODUCT 1909 void Compile::dump_asm(int *pcs, uint pc_limit) { 1910 bool cut_short = false; 1911 tty->print_cr("#"); 1912 tty->print("# "); _tf->dump(); tty->cr(); 1913 tty->print_cr("#"); 1914 1915 // For all blocks 1916 int pc = 0x0; // Program counter 1917 char starts_bundle = ' '; 1918 _regalloc->dump_frame(); 1919 1920 Node *n = NULL; 1921 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1922 if (VMThread::should_terminate()) { cut_short = true; break; } 1923 Block *b = _cfg->_blocks[i]; 1924 if (b->is_connector() && !Verbose) continue; 1925 n = b->_nodes[0]; 1926 if (pcs && n->_idx < pc_limit) 1927 tty->print("%3.3x ", pcs[n->_idx]); 1928 else 1929 tty->print(" "); 1930 b->dump_head( &_cfg->_bbs ); 1931 if (b->is_connector()) { 1932 tty->print_cr(" # Empty connector block"); 1933 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1934 tty->print_cr(" # Block is sole successor of call"); 1935 } 1936 1937 // For all instructions 1938 Node *delay = NULL; 1939 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1940 if (VMThread::should_terminate()) { cut_short = true; break; } 1941 n = b->_nodes[j]; 1942 if (valid_bundle_info(n)) { 1943 Bundle *bundle = node_bundling(n); 1944 if (bundle->used_in_unconditional_delay()) { 1945 delay = n; 1946 continue; 1947 } 1948 if (bundle->starts_bundle()) 1949 starts_bundle = '+'; 1950 } 1951 1952 if (WizardMode) n->dump(); 1953 1954 if( !n->is_Region() && // Dont print in the Assembly 1955 !n->is_Phi() && // a few noisely useless nodes 1956 !n->is_Proj() && 1957 !n->is_MachTemp() && 1958 !n->is_SafePointScalarObject() && 1959 !n->is_Catch() && // Would be nice to print exception table targets 1960 !n->is_MergeMem() && // Not very interesting 1961 !n->is_top() && // Debug info table constants 1962 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1963 ) { 1964 if (pcs && n->_idx < pc_limit) 1965 tty->print("%3.3x", pcs[n->_idx]); 1966 else 1967 tty->print(" "); 1968 tty->print(" %c ", starts_bundle); 1969 starts_bundle = ' '; 1970 tty->print("\t"); 1971 n->format(_regalloc, tty); 1972 tty->cr(); 1973 } 1974 1975 // If we have an instruction with a delay slot, and have seen a delay, 1976 // then back up and print it 1977 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1978 assert(delay != NULL, "no unconditional delay instruction"); 1979 if (WizardMode) delay->dump(); 1980 1981 if (node_bundling(delay)->starts_bundle()) 1982 starts_bundle = '+'; 1983 if (pcs && n->_idx < pc_limit) 1984 tty->print("%3.3x", pcs[n->_idx]); 1985 else 1986 tty->print(" "); 1987 tty->print(" %c ", starts_bundle); 1988 starts_bundle = ' '; 1989 tty->print("\t"); 1990 delay->format(_regalloc, tty); 1991 tty->print_cr(""); 1992 delay = NULL; 1993 } 1994 1995 // Dump the exception table as well 1996 if( n->is_Catch() && (Verbose || WizardMode) ) { 1997 // Print the exception table for this offset 1998 _handler_table.print_subtable_for(pc); 1999 } 2000 } 2001 2002 if (pcs && n->_idx < pc_limit) 2003 tty->print_cr("%3.3x", pcs[n->_idx]); 2004 else 2005 tty->print_cr(""); 2006 2007 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2008 2009 } // End of per-block dump 2010 tty->print_cr(""); 2011 2012 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2013 } 2014 #endif 2015 2016 //------------------------------Final_Reshape_Counts--------------------------- 2017 // This class defines counters to help identify when a method 2018 // may/must be executed using hardware with only 24-bit precision. 2019 struct Final_Reshape_Counts : public StackObj { 2020 int _call_count; // count non-inlined 'common' calls 2021 int _float_count; // count float ops requiring 24-bit precision 2022 int _double_count; // count double ops requiring more precision 2023 int _java_call_count; // count non-inlined 'java' calls 2024 int _inner_loop_count; // count loops which need alignment 2025 VectorSet _visited; // Visitation flags 2026 Node_List _tests; // Set of IfNodes & PCTableNodes 2027 2028 Final_Reshape_Counts() : 2029 _call_count(0), _float_count(0), _double_count(0), 2030 _java_call_count(0), _inner_loop_count(0), 2031 _visited( Thread::current()->resource_area() ) { } 2032 2033 void inc_call_count () { _call_count ++; } 2034 void inc_float_count () { _float_count ++; } 2035 void inc_double_count() { _double_count++; } 2036 void inc_java_call_count() { _java_call_count++; } 2037 void inc_inner_loop_count() { _inner_loop_count++; } 2038 2039 int get_call_count () const { return _call_count ; } 2040 int get_float_count () const { return _float_count ; } 2041 int get_double_count() const { return _double_count; } 2042 int get_java_call_count() const { return _java_call_count; } 2043 int get_inner_loop_count() const { return _inner_loop_count; } 2044 }; 2045 2046 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2047 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2048 // Make sure the offset goes inside the instance layout. 2049 return k->contains_field_offset(tp->offset()); 2050 // Note that OffsetBot and OffsetTop are very negative. 2051 } 2052 2053 //------------------------------final_graph_reshaping_impl---------------------- 2054 // Implement items 1-5 from final_graph_reshaping below. 2055 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 2056 2057 if ( n->outcnt() == 0 ) return; // dead node 2058 uint nop = n->Opcode(); 2059 2060 // Check for 2-input instruction with "last use" on right input. 2061 // Swap to left input. Implements item (2). 2062 if( n->req() == 3 && // two-input instruction 2063 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2064 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2065 n->in(2)->outcnt() == 1 &&// right use IS a last use 2066 !n->in(2)->is_Con() ) { // right use is not a constant 2067 // Check for commutative opcode 2068 switch( nop ) { 2069 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2070 case Op_MaxI: case Op_MinI: 2071 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2072 case Op_AndL: case Op_XorL: case Op_OrL: 2073 case Op_AndI: case Op_XorI: case Op_OrI: { 2074 // Move "last use" input to left by swapping inputs 2075 n->swap_edges(1, 2); 2076 break; 2077 } 2078 default: 2079 break; 2080 } 2081 } 2082 2083 #ifdef ASSERT 2084 if( n->is_Mem() ) { 2085 Compile* C = Compile::current(); 2086 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2087 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2088 // oop will be recorded in oop map if load crosses safepoint 2089 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2090 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2091 "raw memory operations should have control edge"); 2092 } 2093 #endif 2094 // Count FPU ops and common calls, implements item (3) 2095 switch( nop ) { 2096 // Count all float operations that may use FPU 2097 case Op_AddF: 2098 case Op_SubF: 2099 case Op_MulF: 2100 case Op_DivF: 2101 case Op_NegF: 2102 case Op_ModF: 2103 case Op_ConvI2F: 2104 case Op_ConF: 2105 case Op_CmpF: 2106 case Op_CmpF3: 2107 // case Op_ConvL2F: // longs are split into 32-bit halves 2108 frc.inc_float_count(); 2109 break; 2110 2111 case Op_ConvF2D: 2112 case Op_ConvD2F: 2113 frc.inc_float_count(); 2114 frc.inc_double_count(); 2115 break; 2116 2117 // Count all double operations that may use FPU 2118 case Op_AddD: 2119 case Op_SubD: 2120 case Op_MulD: 2121 case Op_DivD: 2122 case Op_NegD: 2123 case Op_ModD: 2124 case Op_ConvI2D: 2125 case Op_ConvD2I: 2126 // case Op_ConvL2D: // handled by leaf call 2127 // case Op_ConvD2L: // handled by leaf call 2128 case Op_ConD: 2129 case Op_CmpD: 2130 case Op_CmpD3: 2131 frc.inc_double_count(); 2132 break; 2133 case Op_Opaque1: // Remove Opaque Nodes before matching 2134 case Op_Opaque2: // Remove Opaque Nodes before matching 2135 n->subsume_by(n->in(1)); 2136 break; 2137 case Op_CallStaticJava: 2138 case Op_CallJava: 2139 case Op_CallDynamicJava: 2140 frc.inc_java_call_count(); // Count java call site; 2141 case Op_CallRuntime: 2142 case Op_CallLeaf: 2143 case Op_CallLeafNoFP: { 2144 assert( n->is_Call(), "" ); 2145 CallNode *call = n->as_Call(); 2146 // Count call sites where the FP mode bit would have to be flipped. 2147 // Do not count uncommon runtime calls: 2148 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2149 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2150 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2151 frc.inc_call_count(); // Count the call site 2152 } else { // See if uncommon argument is shared 2153 Node *n = call->in(TypeFunc::Parms); 2154 int nop = n->Opcode(); 2155 // Clone shared simple arguments to uncommon calls, item (1). 2156 if( n->outcnt() > 1 && 2157 !n->is_Proj() && 2158 nop != Op_CreateEx && 2159 nop != Op_CheckCastPP && 2160 nop != Op_DecodeN && 2161 !n->is_Mem() ) { 2162 Node *x = n->clone(); 2163 call->set_req( TypeFunc::Parms, x ); 2164 } 2165 } 2166 break; 2167 } 2168 2169 case Op_StoreD: 2170 case Op_LoadD: 2171 case Op_LoadD_unaligned: 2172 frc.inc_double_count(); 2173 goto handle_mem; 2174 case Op_StoreF: 2175 case Op_LoadF: 2176 frc.inc_float_count(); 2177 goto handle_mem; 2178 2179 case Op_StoreB: 2180 case Op_StoreC: 2181 case Op_StoreCM: 2182 case Op_StorePConditional: 2183 case Op_StoreI: 2184 case Op_StoreL: 2185 case Op_StoreIConditional: 2186 case Op_StoreLConditional: 2187 case Op_CompareAndSwapI: 2188 case Op_CompareAndSwapL: 2189 case Op_CompareAndSwapP: 2190 case Op_CompareAndSwapN: 2191 case Op_StoreP: 2192 case Op_StoreN: 2193 case Op_LoadB: 2194 case Op_LoadUB: 2195 case Op_LoadUS: 2196 case Op_LoadI: 2197 case Op_LoadUI2L: 2198 case Op_LoadKlass: 2199 case Op_LoadNKlass: 2200 case Op_LoadL: 2201 case Op_LoadL_unaligned: 2202 case Op_LoadPLocked: 2203 case Op_LoadLLocked: 2204 case Op_LoadP: 2205 case Op_LoadN: 2206 case Op_LoadRange: 2207 case Op_LoadS: { 2208 handle_mem: 2209 #ifdef ASSERT 2210 if( VerifyOptoOopOffsets ) { 2211 assert( n->is_Mem(), "" ); 2212 MemNode *mem = (MemNode*)n; 2213 // Check to see if address types have grounded out somehow. 2214 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2215 assert( !tp || oop_offset_is_sane(tp), "" ); 2216 } 2217 #endif 2218 break; 2219 } 2220 2221 case Op_AddP: { // Assert sane base pointers 2222 Node *addp = n->in(AddPNode::Address); 2223 assert( !addp->is_AddP() || 2224 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2225 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2226 "Base pointers must match" ); 2227 #ifdef _LP64 2228 if (UseCompressedOops && 2229 addp->Opcode() == Op_ConP && 2230 addp == n->in(AddPNode::Base) && 2231 n->in(AddPNode::Offset)->is_Con()) { 2232 // Use addressing with narrow klass to load with offset on x86. 2233 // On sparc loading 32-bits constant and decoding it have less 2234 // instructions (4) then load 64-bits constant (7). 2235 // Do this transformation here since IGVN will convert ConN back to ConP. 2236 const Type* t = addp->bottom_type(); 2237 if (t->isa_oopptr()) { 2238 Node* nn = NULL; 2239 2240 // Look for existing ConN node of the same exact type. 2241 Compile* C = Compile::current(); 2242 Node* r = C->root(); 2243 uint cnt = r->outcnt(); 2244 for (uint i = 0; i < cnt; i++) { 2245 Node* m = r->raw_out(i); 2246 if (m!= NULL && m->Opcode() == Op_ConN && 2247 m->bottom_type()->make_ptr() == t) { 2248 nn = m; 2249 break; 2250 } 2251 } 2252 if (nn != NULL) { 2253 // Decode a narrow oop to match address 2254 // [R12 + narrow_oop_reg<<3 + offset] 2255 nn = new (C, 2) DecodeNNode(nn, t); 2256 n->set_req(AddPNode::Base, nn); 2257 n->set_req(AddPNode::Address, nn); 2258 if (addp->outcnt() == 0) { 2259 addp->disconnect_inputs(NULL); 2260 } 2261 } 2262 } 2263 } 2264 #endif 2265 break; 2266 } 2267 2268 #ifdef _LP64 2269 case Op_CastPP: 2270 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2271 Compile* C = Compile::current(); 2272 Node* in1 = n->in(1); 2273 const Type* t = n->bottom_type(); 2274 Node* new_in1 = in1->clone(); 2275 new_in1->as_DecodeN()->set_type(t); 2276 2277 if (!Matcher::narrow_oop_use_complex_address()) { 2278 // 2279 // x86, ARM and friends can handle 2 adds in addressing mode 2280 // and Matcher can fold a DecodeN node into address by using 2281 // a narrow oop directly and do implicit NULL check in address: 2282 // 2283 // [R12 + narrow_oop_reg<<3 + offset] 2284 // NullCheck narrow_oop_reg 2285 // 2286 // On other platforms (Sparc) we have to keep new DecodeN node and 2287 // use it to do implicit NULL check in address: 2288 // 2289 // decode_not_null narrow_oop_reg, base_reg 2290 // [base_reg + offset] 2291 // NullCheck base_reg 2292 // 2293 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2294 // to keep the information to which NULL check the new DecodeN node 2295 // corresponds to use it as value in implicit_null_check(). 2296 // 2297 new_in1->set_req(0, n->in(0)); 2298 } 2299 2300 n->subsume_by(new_in1); 2301 if (in1->outcnt() == 0) { 2302 in1->disconnect_inputs(NULL); 2303 } 2304 } 2305 break; 2306 2307 case Op_CmpP: 2308 // Do this transformation here to preserve CmpPNode::sub() and 2309 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2310 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2311 Node* in1 = n->in(1); 2312 Node* in2 = n->in(2); 2313 if (!in1->is_DecodeN()) { 2314 in2 = in1; 2315 in1 = n->in(2); 2316 } 2317 assert(in1->is_DecodeN(), "sanity"); 2318 2319 Compile* C = Compile::current(); 2320 Node* new_in2 = NULL; 2321 if (in2->is_DecodeN()) { 2322 new_in2 = in2->in(1); 2323 } else if (in2->Opcode() == Op_ConP) { 2324 const Type* t = in2->bottom_type(); 2325 if (t == TypePtr::NULL_PTR) { 2326 // Don't convert CmpP null check into CmpN if compressed 2327 // oops implicit null check is not generated. 2328 // This will allow to generate normal oop implicit null check. 2329 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2330 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2331 // 2332 // This transformation together with CastPP transformation above 2333 // will generated code for implicit NULL checks for compressed oops. 2334 // 2335 // The original code after Optimize() 2336 // 2337 // LoadN memory, narrow_oop_reg 2338 // decode narrow_oop_reg, base_reg 2339 // CmpP base_reg, NULL 2340 // CastPP base_reg // NotNull 2341 // Load [base_reg + offset], val_reg 2342 // 2343 // after these transformations will be 2344 // 2345 // LoadN memory, narrow_oop_reg 2346 // CmpN narrow_oop_reg, NULL 2347 // decode_not_null narrow_oop_reg, base_reg 2348 // Load [base_reg + offset], val_reg 2349 // 2350 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2351 // since narrow oops can be used in debug info now (see the code in 2352 // final_graph_reshaping_walk()). 2353 // 2354 // At the end the code will be matched to 2355 // on x86: 2356 // 2357 // Load_narrow_oop memory, narrow_oop_reg 2358 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2359 // NullCheck narrow_oop_reg 2360 // 2361 // and on sparc: 2362 // 2363 // Load_narrow_oop memory, narrow_oop_reg 2364 // decode_not_null narrow_oop_reg, base_reg 2365 // Load [base_reg + offset], val_reg 2366 // NullCheck base_reg 2367 // 2368 } else if (t->isa_oopptr()) { 2369 new_in2 = ConNode::make(C, t->make_narrowoop()); 2370 } 2371 } 2372 if (new_in2 != NULL) { 2373 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2374 n->subsume_by( cmpN ); 2375 if (in1->outcnt() == 0) { 2376 in1->disconnect_inputs(NULL); 2377 } 2378 if (in2->outcnt() == 0) { 2379 in2->disconnect_inputs(NULL); 2380 } 2381 } 2382 } 2383 break; 2384 2385 case Op_DecodeN: 2386 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2387 // DecodeN could be pinned when it can't be fold into 2388 // an address expression, see the code for Op_CastPP above. 2389 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2390 break; 2391 2392 case Op_EncodeP: { 2393 Node* in1 = n->in(1); 2394 if (in1->is_DecodeN()) { 2395 n->subsume_by(in1->in(1)); 2396 } else if (in1->Opcode() == Op_ConP) { 2397 Compile* C = Compile::current(); 2398 const Type* t = in1->bottom_type(); 2399 if (t == TypePtr::NULL_PTR) { 2400 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2401 } else if (t->isa_oopptr()) { 2402 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2403 } 2404 } 2405 if (in1->outcnt() == 0) { 2406 in1->disconnect_inputs(NULL); 2407 } 2408 break; 2409 } 2410 2411 case Op_Proj: { 2412 if (OptimizeStringConcat) { 2413 ProjNode* p = n->as_Proj(); 2414 if (p->_is_io_use) { 2415 // Separate projections were used for the exception path which 2416 // are normally removed by a late inline. If it wasn't inlined 2417 // then they will hang around and should just be replaced with 2418 // the original one. 2419 Node* proj = NULL; 2420 // Replace with just one 2421 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2422 Node *use = i.get(); 2423 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2424 proj = use; 2425 break; 2426 } 2427 } 2428 assert(p != NULL, "must be found"); 2429 p->subsume_by(proj); 2430 } 2431 } 2432 break; 2433 } 2434 2435 case Op_Phi: 2436 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2437 // The EncodeP optimization may create Phi with the same edges 2438 // for all paths. It is not handled well by Register Allocator. 2439 Node* unique_in = n->in(1); 2440 assert(unique_in != NULL, ""); 2441 uint cnt = n->req(); 2442 for (uint i = 2; i < cnt; i++) { 2443 Node* m = n->in(i); 2444 assert(m != NULL, ""); 2445 if (unique_in != m) 2446 unique_in = NULL; 2447 } 2448 if (unique_in != NULL) { 2449 n->subsume_by(unique_in); 2450 } 2451 } 2452 break; 2453 2454 #endif 2455 2456 case Op_ModI: 2457 if (UseDivMod) { 2458 // Check if a%b and a/b both exist 2459 Node* d = n->find_similar(Op_DivI); 2460 if (d) { 2461 // Replace them with a fused divmod if supported 2462 Compile* C = Compile::current(); 2463 if (Matcher::has_match_rule(Op_DivModI)) { 2464 DivModINode* divmod = DivModINode::make(C, n); 2465 d->subsume_by(divmod->div_proj()); 2466 n->subsume_by(divmod->mod_proj()); 2467 } else { 2468 // replace a%b with a-((a/b)*b) 2469 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2470 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2471 n->subsume_by( sub ); 2472 } 2473 } 2474 } 2475 break; 2476 2477 case Op_ModL: 2478 if (UseDivMod) { 2479 // Check if a%b and a/b both exist 2480 Node* d = n->find_similar(Op_DivL); 2481 if (d) { 2482 // Replace them with a fused divmod if supported 2483 Compile* C = Compile::current(); 2484 if (Matcher::has_match_rule(Op_DivModL)) { 2485 DivModLNode* divmod = DivModLNode::make(C, n); 2486 d->subsume_by(divmod->div_proj()); 2487 n->subsume_by(divmod->mod_proj()); 2488 } else { 2489 // replace a%b with a-((a/b)*b) 2490 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2491 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2492 n->subsume_by( sub ); 2493 } 2494 } 2495 } 2496 break; 2497 2498 case Op_Load16B: 2499 case Op_Load8B: 2500 case Op_Load4B: 2501 case Op_Load8S: 2502 case Op_Load4S: 2503 case Op_Load2S: 2504 case Op_Load8C: 2505 case Op_Load4C: 2506 case Op_Load2C: 2507 case Op_Load4I: 2508 case Op_Load2I: 2509 case Op_Load2L: 2510 case Op_Load4F: 2511 case Op_Load2F: 2512 case Op_Load2D: 2513 case Op_Store16B: 2514 case Op_Store8B: 2515 case Op_Store4B: 2516 case Op_Store8C: 2517 case Op_Store4C: 2518 case Op_Store2C: 2519 case Op_Store4I: 2520 case Op_Store2I: 2521 case Op_Store2L: 2522 case Op_Store4F: 2523 case Op_Store2F: 2524 case Op_Store2D: 2525 break; 2526 2527 case Op_PackB: 2528 case Op_PackS: 2529 case Op_PackC: 2530 case Op_PackI: 2531 case Op_PackF: 2532 case Op_PackL: 2533 case Op_PackD: 2534 if (n->req()-1 > 2) { 2535 // Replace many operand PackNodes with a binary tree for matching 2536 PackNode* p = (PackNode*) n; 2537 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2538 n->subsume_by(btp); 2539 } 2540 break; 2541 case Op_Loop: 2542 case Op_CountedLoop: 2543 if (n->as_Loop()->is_inner_loop()) { 2544 frc.inc_inner_loop_count(); 2545 } 2546 break; 2547 default: 2548 assert( !n->is_Call(), "" ); 2549 assert( !n->is_Mem(), "" ); 2550 break; 2551 } 2552 2553 // Collect CFG split points 2554 if (n->is_MultiBranch()) 2555 frc._tests.push(n); 2556 } 2557 2558 //------------------------------final_graph_reshaping_walk--------------------- 2559 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2560 // requires that the walk visits a node's inputs before visiting the node. 2561 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2562 ResourceArea *area = Thread::current()->resource_area(); 2563 Unique_Node_List sfpt(area); 2564 2565 frc._visited.set(root->_idx); // first, mark node as visited 2566 uint cnt = root->req(); 2567 Node *n = root; 2568 uint i = 0; 2569 while (true) { 2570 if (i < cnt) { 2571 // Place all non-visited non-null inputs onto stack 2572 Node* m = n->in(i); 2573 ++i; 2574 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2575 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2576 sfpt.push(m); 2577 cnt = m->req(); 2578 nstack.push(n, i); // put on stack parent and next input's index 2579 n = m; 2580 i = 0; 2581 } 2582 } else { 2583 // Now do post-visit work 2584 final_graph_reshaping_impl( n, frc ); 2585 if (nstack.is_empty()) 2586 break; // finished 2587 n = nstack.node(); // Get node from stack 2588 cnt = n->req(); 2589 i = nstack.index(); 2590 nstack.pop(); // Shift to the next node on stack 2591 } 2592 } 2593 2594 // Skip next transformation if compressed oops are not used. 2595 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) 2596 return; 2597 2598 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2599 // It could be done for an uncommon traps or any safepoints/calls 2600 // if the DecodeN node is referenced only in a debug info. 2601 while (sfpt.size() > 0) { 2602 n = sfpt.pop(); 2603 JVMState *jvms = n->as_SafePoint()->jvms(); 2604 assert(jvms != NULL, "sanity"); 2605 int start = jvms->debug_start(); 2606 int end = n->req(); 2607 bool is_uncommon = (n->is_CallStaticJava() && 2608 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2609 for (int j = start; j < end; j++) { 2610 Node* in = n->in(j); 2611 if (in->is_DecodeN()) { 2612 bool safe_to_skip = true; 2613 if (!is_uncommon ) { 2614 // Is it safe to skip? 2615 for (uint i = 0; i < in->outcnt(); i++) { 2616 Node* u = in->raw_out(i); 2617 if (!u->is_SafePoint() || 2618 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2619 safe_to_skip = false; 2620 } 2621 } 2622 } 2623 if (safe_to_skip) { 2624 n->set_req(j, in->in(1)); 2625 } 2626 if (in->outcnt() == 0) { 2627 in->disconnect_inputs(NULL); 2628 } 2629 } 2630 } 2631 } 2632 } 2633 2634 //------------------------------final_graph_reshaping-------------------------- 2635 // Final Graph Reshaping. 2636 // 2637 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2638 // and not commoned up and forced early. Must come after regular 2639 // optimizations to avoid GVN undoing the cloning. Clone constant 2640 // inputs to Loop Phis; these will be split by the allocator anyways. 2641 // Remove Opaque nodes. 2642 // (2) Move last-uses by commutative operations to the left input to encourage 2643 // Intel update-in-place two-address operations and better register usage 2644 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2645 // calls canonicalizing them back. 2646 // (3) Count the number of double-precision FP ops, single-precision FP ops 2647 // and call sites. On Intel, we can get correct rounding either by 2648 // forcing singles to memory (requires extra stores and loads after each 2649 // FP bytecode) or we can set a rounding mode bit (requires setting and 2650 // clearing the mode bit around call sites). The mode bit is only used 2651 // if the relative frequency of single FP ops to calls is low enough. 2652 // This is a key transform for SPEC mpeg_audio. 2653 // (4) Detect infinite loops; blobs of code reachable from above but not 2654 // below. Several of the Code_Gen algorithms fail on such code shapes, 2655 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2656 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2657 // Detection is by looking for IfNodes where only 1 projection is 2658 // reachable from below or CatchNodes missing some targets. 2659 // (5) Assert for insane oop offsets in debug mode. 2660 2661 bool Compile::final_graph_reshaping() { 2662 // an infinite loop may have been eliminated by the optimizer, 2663 // in which case the graph will be empty. 2664 if (root()->req() == 1) { 2665 record_method_not_compilable("trivial infinite loop"); 2666 return true; 2667 } 2668 2669 Final_Reshape_Counts frc; 2670 2671 // Visit everybody reachable! 2672 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2673 Node_Stack nstack(unique() >> 1); 2674 final_graph_reshaping_walk(nstack, root(), frc); 2675 2676 // Check for unreachable (from below) code (i.e., infinite loops). 2677 for( uint i = 0; i < frc._tests.size(); i++ ) { 2678 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2679 // Get number of CFG targets. 2680 // Note that PCTables include exception targets after calls. 2681 uint required_outcnt = n->required_outcnt(); 2682 if (n->outcnt() != required_outcnt) { 2683 // Check for a few special cases. Rethrow Nodes never take the 2684 // 'fall-thru' path, so expected kids is 1 less. 2685 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2686 if (n->in(0)->in(0)->is_Call()) { 2687 CallNode *call = n->in(0)->in(0)->as_Call(); 2688 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2689 required_outcnt--; // Rethrow always has 1 less kid 2690 } else if (call->req() > TypeFunc::Parms && 2691 call->is_CallDynamicJava()) { 2692 // Check for null receiver. In such case, the optimizer has 2693 // detected that the virtual call will always result in a null 2694 // pointer exception. The fall-through projection of this CatchNode 2695 // will not be populated. 2696 Node *arg0 = call->in(TypeFunc::Parms); 2697 if (arg0->is_Type() && 2698 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2699 required_outcnt--; 2700 } 2701 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2702 call->req() > TypeFunc::Parms+1 && 2703 call->is_CallStaticJava()) { 2704 // Check for negative array length. In such case, the optimizer has 2705 // detected that the allocation attempt will always result in an 2706 // exception. There is no fall-through projection of this CatchNode . 2707 Node *arg1 = call->in(TypeFunc::Parms+1); 2708 if (arg1->is_Type() && 2709 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2710 required_outcnt--; 2711 } 2712 } 2713 } 2714 } 2715 // Recheck with a better notion of 'required_outcnt' 2716 if (n->outcnt() != required_outcnt) { 2717 record_method_not_compilable("malformed control flow"); 2718 return true; // Not all targets reachable! 2719 } 2720 } 2721 // Check that I actually visited all kids. Unreached kids 2722 // must be infinite loops. 2723 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2724 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2725 record_method_not_compilable("infinite loop"); 2726 return true; // Found unvisited kid; must be unreach 2727 } 2728 } 2729 2730 // If original bytecodes contained a mixture of floats and doubles 2731 // check if the optimizer has made it homogenous, item (3). 2732 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2733 frc.get_float_count() > 32 && 2734 frc.get_double_count() == 0 && 2735 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2736 set_24_bit_selection_and_mode( false, true ); 2737 } 2738 2739 set_java_calls(frc.get_java_call_count()); 2740 set_inner_loops(frc.get_inner_loop_count()); 2741 2742 // No infinite loops, no reason to bail out. 2743 return false; 2744 } 2745 2746 //-----------------------------too_many_traps---------------------------------- 2747 // Report if there are too many traps at the current method and bci. 2748 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2749 bool Compile::too_many_traps(ciMethod* method, 2750 int bci, 2751 Deoptimization::DeoptReason reason) { 2752 ciMethodData* md = method->method_data(); 2753 if (md->is_empty()) { 2754 // Assume the trap has not occurred, or that it occurred only 2755 // because of a transient condition during start-up in the interpreter. 2756 return false; 2757 } 2758 if (md->has_trap_at(bci, reason) != 0) { 2759 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2760 // Also, if there are multiple reasons, or if there is no per-BCI record, 2761 // assume the worst. 2762 if (log()) 2763 log()->elem("observe trap='%s' count='%d'", 2764 Deoptimization::trap_reason_name(reason), 2765 md->trap_count(reason)); 2766 return true; 2767 } else { 2768 // Ignore method/bci and see if there have been too many globally. 2769 return too_many_traps(reason, md); 2770 } 2771 } 2772 2773 // Less-accurate variant which does not require a method and bci. 2774 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2775 ciMethodData* logmd) { 2776 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2777 // Too many traps globally. 2778 // Note that we use cumulative trap_count, not just md->trap_count. 2779 if (log()) { 2780 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2781 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2782 Deoptimization::trap_reason_name(reason), 2783 mcount, trap_count(reason)); 2784 } 2785 return true; 2786 } else { 2787 // The coast is clear. 2788 return false; 2789 } 2790 } 2791 2792 //--------------------------too_many_recompiles-------------------------------- 2793 // Report if there are too many recompiles at the current method and bci. 2794 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2795 // Is not eager to return true, since this will cause the compiler to use 2796 // Action_none for a trap point, to avoid too many recompilations. 2797 bool Compile::too_many_recompiles(ciMethod* method, 2798 int bci, 2799 Deoptimization::DeoptReason reason) { 2800 ciMethodData* md = method->method_data(); 2801 if (md->is_empty()) { 2802 // Assume the trap has not occurred, or that it occurred only 2803 // because of a transient condition during start-up in the interpreter. 2804 return false; 2805 } 2806 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2807 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2808 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2809 Deoptimization::DeoptReason per_bc_reason 2810 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2811 if ((per_bc_reason == Deoptimization::Reason_none 2812 || md->has_trap_at(bci, reason) != 0) 2813 // The trap frequency measure we care about is the recompile count: 2814 && md->trap_recompiled_at(bci) 2815 && md->overflow_recompile_count() >= bc_cutoff) { 2816 // Do not emit a trap here if it has already caused recompilations. 2817 // Also, if there are multiple reasons, or if there is no per-BCI record, 2818 // assume the worst. 2819 if (log()) 2820 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2821 Deoptimization::trap_reason_name(reason), 2822 md->trap_count(reason), 2823 md->overflow_recompile_count()); 2824 return true; 2825 } else if (trap_count(reason) != 0 2826 && decompile_count() >= m_cutoff) { 2827 // Too many recompiles globally, and we have seen this sort of trap. 2828 // Use cumulative decompile_count, not just md->decompile_count. 2829 if (log()) 2830 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2831 Deoptimization::trap_reason_name(reason), 2832 md->trap_count(reason), trap_count(reason), 2833 md->decompile_count(), decompile_count()); 2834 return true; 2835 } else { 2836 // The coast is clear. 2837 return false; 2838 } 2839 } 2840 2841 2842 #ifndef PRODUCT 2843 //------------------------------verify_graph_edges--------------------------- 2844 // Walk the Graph and verify that there is a one-to-one correspondence 2845 // between Use-Def edges and Def-Use edges in the graph. 2846 void Compile::verify_graph_edges(bool no_dead_code) { 2847 if (VerifyGraphEdges) { 2848 ResourceArea *area = Thread::current()->resource_area(); 2849 Unique_Node_List visited(area); 2850 // Call recursive graph walk to check edges 2851 _root->verify_edges(visited); 2852 if (no_dead_code) { 2853 // Now make sure that no visited node is used by an unvisited node. 2854 bool dead_nodes = 0; 2855 Unique_Node_List checked(area); 2856 while (visited.size() > 0) { 2857 Node* n = visited.pop(); 2858 checked.push(n); 2859 for (uint i = 0; i < n->outcnt(); i++) { 2860 Node* use = n->raw_out(i); 2861 if (checked.member(use)) continue; // already checked 2862 if (visited.member(use)) continue; // already in the graph 2863 if (use->is_Con()) continue; // a dead ConNode is OK 2864 // At this point, we have found a dead node which is DU-reachable. 2865 if (dead_nodes++ == 0) 2866 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2867 use->dump(2); 2868 tty->print_cr("---"); 2869 checked.push(use); // No repeats; pretend it is now checked. 2870 } 2871 } 2872 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2873 } 2874 } 2875 } 2876 #endif 2877 2878 // The Compile object keeps track of failure reasons separately from the ciEnv. 2879 // This is required because there is not quite a 1-1 relation between the 2880 // ciEnv and its compilation task and the Compile object. Note that one 2881 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2882 // to backtrack and retry without subsuming loads. Other than this backtracking 2883 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2884 // by the logic in C2Compiler. 2885 void Compile::record_failure(const char* reason) { 2886 if (log() != NULL) { 2887 log()->elem("failure reason='%s' phase='compile'", reason); 2888 } 2889 if (_failure_reason == NULL) { 2890 // Record the first failure reason. 2891 _failure_reason = reason; 2892 } 2893 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2894 C->print_method(_failure_reason); 2895 } 2896 _root = NULL; // flush the graph, too 2897 } 2898 2899 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2900 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2901 { 2902 if (dolog) { 2903 C = Compile::current(); 2904 _log = C->log(); 2905 } else { 2906 C = NULL; 2907 _log = NULL; 2908 } 2909 if (_log != NULL) { 2910 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 2911 _log->stamp(); 2912 _log->end_head(); 2913 } 2914 } 2915 2916 Compile::TracePhase::~TracePhase() { 2917 if (_log != NULL) { 2918 _log->done("phase nodes='%d'", C->unique()); 2919 } 2920 } 2921 2922 //============================================================================= 2923 // Two Constant's are equal when the type and the value are equal. 2924 bool Compile::Constant::operator==(const Constant& other) { 2925 if (type() != other.type() ) return false; 2926 if (can_be_reused() != other.can_be_reused()) return false; 2927 // For floating point values we compare the bit pattern. 2928 switch (type()) { 2929 case T_FLOAT: return (_value.i == other._value.i); 2930 case T_LONG: 2931 case T_DOUBLE: return (_value.j == other._value.j); 2932 case T_OBJECT: 2933 case T_ADDRESS: return (_value.l == other._value.l); 2934 case T_VOID: return (_value.l == other._value.l); // jump-table entries 2935 default: ShouldNotReachHere(); 2936 } 2937 return false; 2938 } 2939 2940 // Emit constants grouped in the following order: 2941 static BasicType type_order[] = { 2942 T_FLOAT, // 32-bit 2943 T_OBJECT, // 32 or 64-bit 2944 T_ADDRESS, // 32 or 64-bit 2945 T_DOUBLE, // 64-bit 2946 T_LONG, // 64-bit 2947 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons) 2948 T_ILLEGAL 2949 }; 2950 2951 static int type_to_size_in_bytes(BasicType t) { 2952 switch (t) { 2953 case T_LONG: return sizeof(jlong ); 2954 case T_FLOAT: return sizeof(jfloat ); 2955 case T_DOUBLE: return sizeof(jdouble); 2956 // We use T_VOID as marker for jump-table entries (labels) which 2957 // need an interal word relocation. 2958 case T_VOID: 2959 case T_ADDRESS: 2960 case T_OBJECT: return sizeof(jobject); 2961 } 2962 2963 ShouldNotReachHere(); 2964 return -1; 2965 } 2966 2967 void Compile::ConstantTable::calculate_offsets_and_size() { 2968 int size = 0; 2969 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 2970 BasicType type = type_order[t]; 2971 2972 for (int i = 0; i < _constants.length(); i++) { 2973 Constant con = _constants.at(i); 2974 if (con.type() != type) continue; // Skip other types. 2975 2976 // Align size for type. 2977 int typesize = type_to_size_in_bytes(con.type()); 2978 size = align_size_up(size, typesize); 2979 2980 // Set offset. 2981 con.set_offset(size); 2982 _constants.at_put(i, con); 2983 2984 // Add type size. 2985 size = size + typesize; 2986 } 2987 } 2988 2989 // Align size up to the next section start (which is insts; see 2990 // CodeBuffer::align_at_start). 2991 assert(_size == -1, "already set?"); 2992 _size = align_size_up(size, CodeEntryAlignment); 2993 2994 if (Matcher::constant_table_absolute_addressing) { 2995 set_table_base_offset(0); // No table base offset required 2996 } else { 2997 if (UseRDPCForConstantTableBase) { 2998 // table base offset is set in MachConstantBaseNode::emit 2999 } else { 3000 // When RDPC is not used, the table base is set into the middle of 3001 // the constant table. 3002 int half_size = _size / 2; 3003 assert(half_size * 2 == _size, "sanity"); 3004 set_table_base_offset(-half_size); 3005 } 3006 } 3007 } 3008 3009 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3010 MacroAssembler _masm(&cb); 3011 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 3012 BasicType type = type_order[t]; 3013 3014 for (int i = 0; i < _constants.length(); i++) { 3015 Constant con = _constants.at(i); 3016 if (con.type() != type) continue; // Skip other types. 3017 3018 address constant_addr; 3019 switch (con.type()) { 3020 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3021 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3022 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3023 case T_OBJECT: { 3024 jobject obj = con.get_jobject(); 3025 int oop_index = _masm.oop_recorder()->find_index(obj); 3026 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3027 break; 3028 } 3029 case T_ADDRESS: { 3030 address addr = (address) con.get_jobject(); 3031 constant_addr = _masm.address_constant(addr); 3032 break; 3033 } 3034 // We use T_VOID as marker for jump-table entries (labels) which 3035 // need an interal word relocation. 3036 case T_VOID: { 3037 // Write a dummy word. The real value is filled in later 3038 // in fill_jump_table_in_constant_table. 3039 address addr = (address) con.get_jobject(); 3040 constant_addr = _masm.address_constant(addr); 3041 break; 3042 } 3043 default: ShouldNotReachHere(); 3044 } 3045 assert(constant_addr != NULL, "consts section too small"); 3046 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3047 } 3048 } 3049 } 3050 3051 int Compile::ConstantTable::find_offset(Constant& con) const { 3052 int idx = _constants.find(con); 3053 assert(idx != -1, "constant must be in constant table"); 3054 int offset = _constants.at(idx).offset(); 3055 assert(offset != -1, "constant table not emitted yet?"); 3056 return offset; 3057 } 3058 3059 void Compile::ConstantTable::add(Constant& con) { 3060 if (con.can_be_reused()) { 3061 int idx = _constants.find(con); 3062 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3063 return; 3064 } 3065 } 3066 (void) _constants.append(con); 3067 } 3068 3069 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) { 3070 Constant con(type, value); 3071 add(con); 3072 return con; 3073 } 3074 3075 Compile::Constant Compile::ConstantTable::add(MachOper* oper) { 3076 jvalue value; 3077 BasicType type = oper->type()->basic_type(); 3078 switch (type) { 3079 case T_LONG: value.j = oper->constantL(); break; 3080 case T_FLOAT: value.f = oper->constantF(); break; 3081 case T_DOUBLE: value.d = oper->constantD(); break; 3082 case T_OBJECT: 3083 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3084 default: ShouldNotReachHere(); 3085 } 3086 return add(type, value); 3087 } 3088 3089 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) { 3090 jvalue value; 3091 // We can use the node pointer here to identify the right jump-table 3092 // as this method is called from Compile::Fill_buffer right before 3093 // the MachNodes are emitted and the jump-table is filled (means the 3094 // MachNode pointers do not change anymore). 3095 value.l = (jobject) n; 3096 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused. 3097 for (uint i = 0; i < n->outcnt(); i++) { 3098 add(con); 3099 } 3100 return con; 3101 } 3102 3103 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3104 // If called from Compile::scratch_emit_size do nothing. 3105 if (Compile::current()->in_scratch_emit_size()) return; 3106 3107 assert(labels.is_nonempty(), "must be"); 3108 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt())); 3109 3110 // Since MachConstantNode::constant_offset() also contains 3111 // table_base_offset() we need to subtract the table_base_offset() 3112 // to get the plain offset into the constant table. 3113 int offset = n->constant_offset() - table_base_offset(); 3114 3115 MacroAssembler _masm(&cb); 3116 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3117 3118 for (int i = 0; i < labels.length(); i++) { 3119 address* constant_addr = &jump_table_base[i]; 3120 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer"); 3121 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3122 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3123 } 3124 }