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. 520 521 // Fake label for branch instructions. 522 Label fakeL; 523 if (n->is_Branch() && n->as_Mach()->ideal_Opcode() != Op_Jump) { 524 MacroAssembler masm(&buf); 525 masm.bind(fakeL); 526 n->as_Mach()->label_set(fakeL, 0); 527 } 528 n->emit(buf, this->regalloc()); 529 530 // End scratch_emit_size section. 531 set_in_scratch_emit_size(false); 532 533 return buf.insts_size(); 534 } 535 536 537 // ============================================================================ 538 //------------------------------Compile standard------------------------------- 539 debug_only( int Compile::_debug_idx = 100000; ) 540 541 // Compile a method. entry_bci is -1 for normal compilations and indicates 542 // the continuation bci for on stack replacement. 543 544 545 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 546 : Phase(Compiler), 547 _env(ci_env), 548 _log(ci_env->log()), 549 _compile_id(ci_env->compile_id()), 550 _save_argument_registers(false), 551 _stub_name(NULL), 552 _stub_function(NULL), 553 _stub_entry_point(NULL), 554 _method(target), 555 _entry_bci(osr_bci), 556 _initial_gvn(NULL), 557 _for_igvn(NULL), 558 _warm_calls(NULL), 559 _subsume_loads(subsume_loads), 560 _do_escape_analysis(do_escape_analysis), 561 _failure_reason(NULL), 562 _code_buffer("Compile::Fill_buffer"), 563 _orig_pc_slot(0), 564 _orig_pc_slot_offset_in_bytes(0), 565 _has_method_handle_invokes(false), 566 _mach_constant_base_node(NULL), 567 _node_bundling_limit(0), 568 _node_bundling_base(NULL), 569 _java_calls(0), 570 _inner_loops(0), 571 _scratch_const_size(-1), 572 _in_scratch_emit_size(false), 573 #ifndef PRODUCT 574 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 575 _printer(IdealGraphPrinter::printer()), 576 #endif 577 _congraph(NULL) { 578 C = this; 579 580 CompileWrapper cw(this); 581 #ifndef PRODUCT 582 if (TimeCompiler2) { 583 tty->print(" "); 584 target->holder()->name()->print(); 585 tty->print("."); 586 target->print_short_name(); 587 tty->print(" "); 588 } 589 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 590 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 591 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 592 if (!print_opto_assembly) { 593 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 594 if (print_assembly && !Disassembler::can_decode()) { 595 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 596 print_opto_assembly = true; 597 } 598 } 599 set_print_assembly(print_opto_assembly); 600 set_parsed_irreducible_loop(false); 601 #endif 602 603 if (ProfileTraps) { 604 // Make sure the method being compiled gets its own MDO, 605 // so we can at least track the decompile_count(). 606 method()->ensure_method_data(); 607 } 608 609 Init(::AliasLevel); 610 611 612 print_compile_messages(); 613 614 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 615 _ilt = InlineTree::build_inline_tree_root(); 616 else 617 _ilt = NULL; 618 619 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 620 assert(num_alias_types() >= AliasIdxRaw, ""); 621 622 #define MINIMUM_NODE_HASH 1023 623 // Node list that Iterative GVN will start with 624 Unique_Node_List for_igvn(comp_arena()); 625 set_for_igvn(&for_igvn); 626 627 // GVN that will be run immediately on new nodes 628 uint estimated_size = method()->code_size()*4+64; 629 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 630 PhaseGVN gvn(node_arena(), estimated_size); 631 set_initial_gvn(&gvn); 632 633 { // Scope for timing the parser 634 TracePhase t3("parse", &_t_parser, true); 635 636 // Put top into the hash table ASAP. 637 initial_gvn()->transform_no_reclaim(top()); 638 639 // Set up tf(), start(), and find a CallGenerator. 640 CallGenerator* cg = NULL; 641 if (is_osr_compilation()) { 642 const TypeTuple *domain = StartOSRNode::osr_domain(); 643 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 644 init_tf(TypeFunc::make(domain, range)); 645 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 646 initial_gvn()->set_type_bottom(s); 647 init_start(s); 648 cg = CallGenerator::for_osr(method(), entry_bci()); 649 } else { 650 // Normal case. 651 init_tf(TypeFunc::make(method())); 652 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 653 initial_gvn()->set_type_bottom(s); 654 init_start(s); 655 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 656 // With java.lang.ref.reference.get() we must go through the 657 // intrinsic when G1 is enabled - even when get() is the root 658 // method of the compile - so that, if necessary, the value in 659 // the referent field of the reference object gets recorded by 660 // the pre-barrier code. 661 // Specifically, if G1 is enabled, the value in the referent 662 // field is recorded by the G1 SATB pre barrier. This will 663 // result in the referent being marked live and the reference 664 // object removed from the list of discovered references during 665 // reference processing. 666 cg = find_intrinsic(method(), false); 667 } 668 if (cg == NULL) { 669 float past_uses = method()->interpreter_invocation_count(); 670 float expected_uses = past_uses; 671 cg = CallGenerator::for_inline(method(), expected_uses); 672 } 673 } 674 if (failing()) return; 675 if (cg == NULL) { 676 record_method_not_compilable_all_tiers("cannot parse method"); 677 return; 678 } 679 JVMState* jvms = build_start_state(start(), tf()); 680 if ((jvms = cg->generate(jvms)) == NULL) { 681 record_method_not_compilable("method parse failed"); 682 return; 683 } 684 GraphKit kit(jvms); 685 686 if (!kit.stopped()) { 687 // Accept return values, and transfer control we know not where. 688 // This is done by a special, unique ReturnNode bound to root. 689 return_values(kit.jvms()); 690 } 691 692 if (kit.has_exceptions()) { 693 // Any exceptions that escape from this call must be rethrown 694 // to whatever caller is dynamically above us on the stack. 695 // This is done by a special, unique RethrowNode bound to root. 696 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 697 } 698 699 if (!failing() && has_stringbuilder()) { 700 { 701 // remove useless nodes to make the usage analysis simpler 702 ResourceMark rm; 703 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 704 } 705 706 { 707 ResourceMark rm; 708 print_method("Before StringOpts", 3); 709 PhaseStringOpts pso(initial_gvn(), &for_igvn); 710 print_method("After StringOpts", 3); 711 } 712 713 // now inline anything that we skipped the first time around 714 while (_late_inlines.length() > 0) { 715 CallGenerator* cg = _late_inlines.pop(); 716 cg->do_late_inline(); 717 } 718 } 719 assert(_late_inlines.length() == 0, "should have been processed"); 720 721 print_method("Before RemoveUseless", 3); 722 723 // Remove clutter produced by parsing. 724 if (!failing()) { 725 ResourceMark rm; 726 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 727 } 728 } 729 730 // Note: Large methods are capped off in do_one_bytecode(). 731 if (failing()) return; 732 733 // After parsing, node notes are no longer automagic. 734 // They must be propagated by register_new_node_with_optimizer(), 735 // clone(), or the like. 736 set_default_node_notes(NULL); 737 738 for (;;) { 739 int successes = Inline_Warm(); 740 if (failing()) return; 741 if (successes == 0) break; 742 } 743 744 // Drain the list. 745 Finish_Warm(); 746 #ifndef PRODUCT 747 if (_printer) { 748 _printer->print_inlining(this); 749 } 750 #endif 751 752 if (failing()) return; 753 NOT_PRODUCT( verify_graph_edges(); ) 754 755 // Now optimize 756 Optimize(); 757 if (failing()) return; 758 NOT_PRODUCT( verify_graph_edges(); ) 759 760 #ifndef PRODUCT 761 if (PrintIdeal) { 762 ttyLocker ttyl; // keep the following output all in one block 763 // This output goes directly to the tty, not the compiler log. 764 // To enable tools to match it up with the compilation activity, 765 // be sure to tag this tty output with the compile ID. 766 if (xtty != NULL) { 767 xtty->head("ideal compile_id='%d'%s", compile_id(), 768 is_osr_compilation() ? " compile_kind='osr'" : 769 ""); 770 } 771 root()->dump(9999); 772 if (xtty != NULL) { 773 xtty->tail("ideal"); 774 } 775 } 776 #endif 777 778 // Now that we know the size of all the monitors we can add a fixed slot 779 // for the original deopt pc. 780 781 _orig_pc_slot = fixed_slots(); 782 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 783 set_fixed_slots(next_slot); 784 785 // Now generate code 786 Code_Gen(); 787 if (failing()) return; 788 789 // Check if we want to skip execution of all compiled code. 790 { 791 #ifndef PRODUCT 792 if (OptoNoExecute) { 793 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 794 return; 795 } 796 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 797 #endif 798 799 if (is_osr_compilation()) { 800 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 801 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 802 } else { 803 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 804 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 805 } 806 807 env()->register_method(_method, _entry_bci, 808 &_code_offsets, 809 _orig_pc_slot_offset_in_bytes, 810 code_buffer(), 811 frame_size_in_words(), _oop_map_set, 812 &_handler_table, &_inc_table, 813 compiler, 814 env()->comp_level(), 815 true, /*has_debug_info*/ 816 has_unsafe_access() 817 ); 818 } 819 } 820 821 //------------------------------Compile---------------------------------------- 822 // Compile a runtime stub 823 Compile::Compile( ciEnv* ci_env, 824 TypeFunc_generator generator, 825 address stub_function, 826 const char *stub_name, 827 int is_fancy_jump, 828 bool pass_tls, 829 bool save_arg_registers, 830 bool return_pc ) 831 : Phase(Compiler), 832 _env(ci_env), 833 _log(ci_env->log()), 834 _compile_id(-1), 835 _save_argument_registers(save_arg_registers), 836 _method(NULL), 837 _stub_name(stub_name), 838 _stub_function(stub_function), 839 _stub_entry_point(NULL), 840 _entry_bci(InvocationEntryBci), 841 _initial_gvn(NULL), 842 _for_igvn(NULL), 843 _warm_calls(NULL), 844 _orig_pc_slot(0), 845 _orig_pc_slot_offset_in_bytes(0), 846 _subsume_loads(true), 847 _do_escape_analysis(false), 848 _failure_reason(NULL), 849 _code_buffer("Compile::Fill_buffer"), 850 _has_method_handle_invokes(false), 851 _mach_constant_base_node(NULL), 852 _node_bundling_limit(0), 853 _node_bundling_base(NULL), 854 _java_calls(0), 855 _inner_loops(0), 856 #ifndef PRODUCT 857 _trace_opto_output(TraceOptoOutput), 858 _printer(NULL), 859 #endif 860 _congraph(NULL) { 861 C = this; 862 863 #ifndef PRODUCT 864 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 865 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 866 set_print_assembly(PrintFrameConverterAssembly); 867 set_parsed_irreducible_loop(false); 868 #endif 869 CompileWrapper cw(this); 870 Init(/*AliasLevel=*/ 0); 871 init_tf((*generator)()); 872 873 { 874 // The following is a dummy for the sake of GraphKit::gen_stub 875 Unique_Node_List for_igvn(comp_arena()); 876 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 877 PhaseGVN gvn(Thread::current()->resource_area(),255); 878 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 879 gvn.transform_no_reclaim(top()); 880 881 GraphKit kit; 882 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 883 } 884 885 NOT_PRODUCT( verify_graph_edges(); ) 886 Code_Gen(); 887 if (failing()) return; 888 889 890 // Entry point will be accessed using compile->stub_entry_point(); 891 if (code_buffer() == NULL) { 892 Matcher::soft_match_failure(); 893 } else { 894 if (PrintAssembly && (WizardMode || Verbose)) 895 tty->print_cr("### Stub::%s", stub_name); 896 897 if (!failing()) { 898 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 899 900 // Make the NMethod 901 // For now we mark the frame as never safe for profile stackwalking 902 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 903 code_buffer(), 904 CodeOffsets::frame_never_safe, 905 // _code_offsets.value(CodeOffsets::Frame_Complete), 906 frame_size_in_words(), 907 _oop_map_set, 908 save_arg_registers); 909 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 910 911 _stub_entry_point = rs->entry_point(); 912 } 913 } 914 } 915 916 #ifndef PRODUCT 917 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 918 if(PrintOpto && Verbose) { 919 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 920 } 921 } 922 #endif 923 924 void Compile::print_codes() { 925 } 926 927 //------------------------------Init------------------------------------------- 928 // Prepare for a single compilation 929 void Compile::Init(int aliaslevel) { 930 _unique = 0; 931 _regalloc = NULL; 932 933 _tf = NULL; // filled in later 934 _top = NULL; // cached later 935 _matcher = NULL; // filled in later 936 _cfg = NULL; // filled in later 937 938 set_24_bit_selection_and_mode(Use24BitFP, false); 939 940 _node_note_array = NULL; 941 _default_node_notes = NULL; 942 943 _immutable_memory = NULL; // filled in at first inquiry 944 945 // Globally visible Nodes 946 // First set TOP to NULL to give safe behavior during creation of RootNode 947 set_cached_top_node(NULL); 948 set_root(new (this, 3) RootNode()); 949 // Now that you have a Root to point to, create the real TOP 950 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 951 set_recent_alloc(NULL, NULL); 952 953 // Create Debug Information Recorder to record scopes, oopmaps, etc. 954 env()->set_oop_recorder(new OopRecorder(comp_arena())); 955 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 956 env()->set_dependencies(new Dependencies(env())); 957 958 _fixed_slots = 0; 959 set_has_split_ifs(false); 960 set_has_loops(has_method() && method()->has_loops()); // first approximation 961 set_has_stringbuilder(false); 962 _trap_can_recompile = false; // no traps emitted yet 963 _major_progress = true; // start out assuming good things will happen 964 set_has_unsafe_access(false); 965 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 966 set_decompile_count(0); 967 968 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 969 set_num_loop_opts(LoopOptsCount); 970 set_do_inlining(Inline); 971 set_max_inline_size(MaxInlineSize); 972 set_freq_inline_size(FreqInlineSize); 973 set_do_scheduling(OptoScheduling); 974 set_do_count_invocations(false); 975 set_do_method_data_update(false); 976 977 if (debug_info()->recording_non_safepoints()) { 978 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 979 (comp_arena(), 8, 0, NULL)); 980 set_default_node_notes(Node_Notes::make(this)); 981 } 982 983 // // -- Initialize types before each compile -- 984 // // Update cached type information 985 // if( _method && _method->constants() ) 986 // Type::update_loaded_types(_method, _method->constants()); 987 988 // Init alias_type map. 989 if (!_do_escape_analysis && aliaslevel == 3) 990 aliaslevel = 2; // No unique types without escape analysis 991 _AliasLevel = aliaslevel; 992 const int grow_ats = 16; 993 _max_alias_types = grow_ats; 994 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 995 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 996 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 997 { 998 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 999 } 1000 // Initialize the first few types. 1001 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1002 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1003 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1004 _num_alias_types = AliasIdxRaw+1; 1005 // Zero out the alias type cache. 1006 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1007 // A NULL adr_type hits in the cache right away. Preload the right answer. 1008 probe_alias_cache(NULL)->_index = AliasIdxTop; 1009 1010 _intrinsics = NULL; 1011 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1012 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1013 register_library_intrinsics(); 1014 } 1015 1016 //---------------------------init_start---------------------------------------- 1017 // Install the StartNode on this compile object. 1018 void Compile::init_start(StartNode* s) { 1019 if (failing()) 1020 return; // already failing 1021 assert(s == start(), ""); 1022 } 1023 1024 StartNode* Compile::start() const { 1025 assert(!failing(), ""); 1026 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1027 Node* start = root()->fast_out(i); 1028 if( start->is_Start() ) 1029 return start->as_Start(); 1030 } 1031 ShouldNotReachHere(); 1032 return NULL; 1033 } 1034 1035 //-------------------------------immutable_memory------------------------------------- 1036 // Access immutable memory 1037 Node* Compile::immutable_memory() { 1038 if (_immutable_memory != NULL) { 1039 return _immutable_memory; 1040 } 1041 StartNode* s = start(); 1042 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1043 Node *p = s->fast_out(i); 1044 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1045 _immutable_memory = p; 1046 return _immutable_memory; 1047 } 1048 } 1049 ShouldNotReachHere(); 1050 return NULL; 1051 } 1052 1053 //----------------------set_cached_top_node------------------------------------ 1054 // Install the cached top node, and make sure Node::is_top works correctly. 1055 void Compile::set_cached_top_node(Node* tn) { 1056 if (tn != NULL) verify_top(tn); 1057 Node* old_top = _top; 1058 _top = tn; 1059 // Calling Node::setup_is_top allows the nodes the chance to adjust 1060 // their _out arrays. 1061 if (_top != NULL) _top->setup_is_top(); 1062 if (old_top != NULL) old_top->setup_is_top(); 1063 assert(_top == NULL || top()->is_top(), ""); 1064 } 1065 1066 #ifndef PRODUCT 1067 void Compile::verify_top(Node* tn) const { 1068 if (tn != NULL) { 1069 assert(tn->is_Con(), "top node must be a constant"); 1070 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1071 assert(tn->in(0) != NULL, "must have live top node"); 1072 } 1073 } 1074 #endif 1075 1076 1077 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1078 1079 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1080 guarantee(arr != NULL, ""); 1081 int num_blocks = arr->length(); 1082 if (grow_by < num_blocks) grow_by = num_blocks; 1083 int num_notes = grow_by * _node_notes_block_size; 1084 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1085 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1086 while (num_notes > 0) { 1087 arr->append(notes); 1088 notes += _node_notes_block_size; 1089 num_notes -= _node_notes_block_size; 1090 } 1091 assert(num_notes == 0, "exact multiple, please"); 1092 } 1093 1094 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1095 if (source == NULL || dest == NULL) return false; 1096 1097 if (dest->is_Con()) 1098 return false; // Do not push debug info onto constants. 1099 1100 #ifdef ASSERT 1101 // Leave a bread crumb trail pointing to the original node: 1102 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1103 dest->set_debug_orig(source); 1104 } 1105 #endif 1106 1107 if (node_note_array() == NULL) 1108 return false; // Not collecting any notes now. 1109 1110 // This is a copy onto a pre-existing node, which may already have notes. 1111 // If both nodes have notes, do not overwrite any pre-existing notes. 1112 Node_Notes* source_notes = node_notes_at(source->_idx); 1113 if (source_notes == NULL || source_notes->is_clear()) return false; 1114 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1115 if (dest_notes == NULL || dest_notes->is_clear()) { 1116 return set_node_notes_at(dest->_idx, source_notes); 1117 } 1118 1119 Node_Notes merged_notes = (*source_notes); 1120 // The order of operations here ensures that dest notes will win... 1121 merged_notes.update_from(dest_notes); 1122 return set_node_notes_at(dest->_idx, &merged_notes); 1123 } 1124 1125 1126 //--------------------------allow_range_check_smearing------------------------- 1127 // Gating condition for coalescing similar range checks. 1128 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1129 // single covering check that is at least as strong as any of them. 1130 // If the optimization succeeds, the simplified (strengthened) range check 1131 // will always succeed. If it fails, we will deopt, and then give up 1132 // on the optimization. 1133 bool Compile::allow_range_check_smearing() const { 1134 // If this method has already thrown a range-check, 1135 // assume it was because we already tried range smearing 1136 // and it failed. 1137 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1138 return !already_trapped; 1139 } 1140 1141 1142 //------------------------------flatten_alias_type----------------------------- 1143 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1144 int offset = tj->offset(); 1145 TypePtr::PTR ptr = tj->ptr(); 1146 1147 // Known instance (scalarizable allocation) alias only with itself. 1148 bool is_known_inst = tj->isa_oopptr() != NULL && 1149 tj->is_oopptr()->is_known_instance(); 1150 1151 // Process weird unsafe references. 1152 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1153 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1154 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1155 tj = TypeOopPtr::BOTTOM; 1156 ptr = tj->ptr(); 1157 offset = tj->offset(); 1158 } 1159 1160 // Array pointers need some flattening 1161 const TypeAryPtr *ta = tj->isa_aryptr(); 1162 if( ta && is_known_inst ) { 1163 if ( offset != Type::OffsetBot && 1164 offset > arrayOopDesc::length_offset_in_bytes() ) { 1165 offset = Type::OffsetBot; // Flatten constant access into array body only 1166 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1167 } 1168 } else if( ta && _AliasLevel >= 2 ) { 1169 // For arrays indexed by constant indices, we flatten the alias 1170 // space to include all of the array body. Only the header, klass 1171 // and array length can be accessed un-aliased. 1172 if( offset != Type::OffsetBot ) { 1173 if( ta->const_oop() ) { // methodDataOop or methodOop 1174 offset = Type::OffsetBot; // Flatten constant access into array body 1175 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1176 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1177 // range is OK as-is. 1178 tj = ta = TypeAryPtr::RANGE; 1179 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1180 tj = TypeInstPtr::KLASS; // all klass loads look alike 1181 ta = TypeAryPtr::RANGE; // generic ignored junk 1182 ptr = TypePtr::BotPTR; 1183 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1184 tj = TypeInstPtr::MARK; 1185 ta = TypeAryPtr::RANGE; // generic ignored junk 1186 ptr = TypePtr::BotPTR; 1187 } else { // Random constant offset into array body 1188 offset = Type::OffsetBot; // Flatten constant access into array body 1189 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1190 } 1191 } 1192 // Arrays of fixed size alias with arrays of unknown size. 1193 if (ta->size() != TypeInt::POS) { 1194 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1195 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1196 } 1197 // Arrays of known objects become arrays of unknown objects. 1198 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1199 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1200 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1201 } 1202 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1203 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1204 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1205 } 1206 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1207 // cannot be distinguished by bytecode alone. 1208 if (ta->elem() == TypeInt::BOOL) { 1209 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1210 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1211 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1212 } 1213 // During the 2nd round of IterGVN, NotNull castings are removed. 1214 // Make sure the Bottom and NotNull variants alias the same. 1215 // Also, make sure exact and non-exact variants alias the same. 1216 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1217 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1218 } 1219 } 1220 1221 // Oop pointers need some flattening 1222 const TypeInstPtr *to = tj->isa_instptr(); 1223 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1224 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1225 if( ptr == TypePtr::Constant ) { 1226 if (to->klass() != ciEnv::current()->Class_klass() || 1227 offset < k->size_helper() * wordSize) { 1228 // No constant oop pointers (such as Strings); they alias with 1229 // unknown strings. 1230 assert(!is_known_inst, "not scalarizable allocation"); 1231 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1232 } 1233 } else if( is_known_inst ) { 1234 tj = to; // Keep NotNull and klass_is_exact for instance type 1235 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1236 // During the 2nd round of IterGVN, NotNull castings are removed. 1237 // Make sure the Bottom and NotNull variants alias the same. 1238 // Also, make sure exact and non-exact variants alias the same. 1239 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1240 } 1241 // Canonicalize the holder of this field 1242 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1243 // First handle header references such as a LoadKlassNode, even if the 1244 // object's klass is unloaded at compile time (4965979). 1245 if (!is_known_inst) { // Do it only for non-instance types 1246 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1247 } 1248 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1249 // Static fields are in the space above the normal instance 1250 // fields in the java.lang.Class instance. 1251 if (to->klass() != ciEnv::current()->Class_klass()) { 1252 to = NULL; 1253 tj = TypeOopPtr::BOTTOM; 1254 offset = tj->offset(); 1255 } 1256 } else { 1257 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1258 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1259 if( is_known_inst ) { 1260 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1261 } else { 1262 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1263 } 1264 } 1265 } 1266 } 1267 1268 // Klass pointers to object array klasses need some flattening 1269 const TypeKlassPtr *tk = tj->isa_klassptr(); 1270 if( tk ) { 1271 // If we are referencing a field within a Klass, we need 1272 // to assume the worst case of an Object. Both exact and 1273 // inexact types must flatten to the same alias class. 1274 // Since the flattened result for a klass is defined to be 1275 // precisely java.lang.Object, use a constant ptr. 1276 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1277 1278 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1279 TypeKlassPtr::OBJECT->klass(), 1280 offset); 1281 } 1282 1283 ciKlass* klass = tk->klass(); 1284 if( klass->is_obj_array_klass() ) { 1285 ciKlass* k = TypeAryPtr::OOPS->klass(); 1286 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1287 k = TypeInstPtr::BOTTOM->klass(); 1288 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1289 } 1290 1291 // Check for precise loads from the primary supertype array and force them 1292 // to the supertype cache alias index. Check for generic array loads from 1293 // the primary supertype array and also force them to the supertype cache 1294 // alias index. Since the same load can reach both, we need to merge 1295 // these 2 disparate memories into the same alias class. Since the 1296 // primary supertype array is read-only, there's no chance of confusion 1297 // where we bypass an array load and an array store. 1298 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1299 if( offset == Type::OffsetBot || 1300 off2 < Klass::primary_super_limit()*wordSize ) { 1301 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1302 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1303 } 1304 } 1305 1306 // Flatten all Raw pointers together. 1307 if (tj->base() == Type::RawPtr) 1308 tj = TypeRawPtr::BOTTOM; 1309 1310 if (tj->base() == Type::AnyPtr) 1311 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1312 1313 // Flatten all to bottom for now 1314 switch( _AliasLevel ) { 1315 case 0: 1316 tj = TypePtr::BOTTOM; 1317 break; 1318 case 1: // Flatten to: oop, static, field or array 1319 switch (tj->base()) { 1320 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1321 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1322 case Type::AryPtr: // do not distinguish arrays at all 1323 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1324 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1325 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1326 default: ShouldNotReachHere(); 1327 } 1328 break; 1329 case 2: // No collapsing at level 2; keep all splits 1330 case 3: // No collapsing at level 3; keep all splits 1331 break; 1332 default: 1333 Unimplemented(); 1334 } 1335 1336 offset = tj->offset(); 1337 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1338 1339 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1340 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1341 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1342 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1343 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1344 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1345 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1346 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1347 assert( tj->ptr() != TypePtr::TopPTR && 1348 tj->ptr() != TypePtr::AnyNull && 1349 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1350 // assert( tj->ptr() != TypePtr::Constant || 1351 // tj->base() == Type::RawPtr || 1352 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1353 1354 return tj; 1355 } 1356 1357 void Compile::AliasType::Init(int i, const TypePtr* at) { 1358 _index = i; 1359 _adr_type = at; 1360 _field = NULL; 1361 _is_rewritable = true; // default 1362 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1363 if (atoop != NULL && atoop->is_known_instance()) { 1364 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1365 _general_index = Compile::current()->get_alias_index(gt); 1366 } else { 1367 _general_index = 0; 1368 } 1369 } 1370 1371 //---------------------------------print_on------------------------------------ 1372 #ifndef PRODUCT 1373 void Compile::AliasType::print_on(outputStream* st) { 1374 if (index() < 10) 1375 st->print("@ <%d> ", index()); 1376 else st->print("@ <%d>", index()); 1377 st->print(is_rewritable() ? " " : " RO"); 1378 int offset = adr_type()->offset(); 1379 if (offset == Type::OffsetBot) 1380 st->print(" +any"); 1381 else st->print(" +%-3d", offset); 1382 st->print(" in "); 1383 adr_type()->dump_on(st); 1384 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1385 if (field() != NULL && tjp) { 1386 if (tjp->klass() != field()->holder() || 1387 tjp->offset() != field()->offset_in_bytes()) { 1388 st->print(" != "); 1389 field()->print(); 1390 st->print(" ***"); 1391 } 1392 } 1393 } 1394 1395 void print_alias_types() { 1396 Compile* C = Compile::current(); 1397 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1398 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1399 C->alias_type(idx)->print_on(tty); 1400 tty->cr(); 1401 } 1402 } 1403 #endif 1404 1405 1406 //----------------------------probe_alias_cache-------------------------------- 1407 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1408 intptr_t key = (intptr_t) adr_type; 1409 key ^= key >> logAliasCacheSize; 1410 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1411 } 1412 1413 1414 //-----------------------------grow_alias_types-------------------------------- 1415 void Compile::grow_alias_types() { 1416 const int old_ats = _max_alias_types; // how many before? 1417 const int new_ats = old_ats; // how many more? 1418 const int grow_ats = old_ats+new_ats; // how many now? 1419 _max_alias_types = grow_ats; 1420 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1421 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1422 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1423 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1424 } 1425 1426 1427 //--------------------------------find_alias_type------------------------------ 1428 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1429 if (_AliasLevel == 0) 1430 return alias_type(AliasIdxBot); 1431 1432 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1433 if (ace->_adr_type == adr_type) { 1434 return alias_type(ace->_index); 1435 } 1436 1437 // Handle special cases. 1438 if (adr_type == NULL) return alias_type(AliasIdxTop); 1439 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1440 1441 // Do it the slow way. 1442 const TypePtr* flat = flatten_alias_type(adr_type); 1443 1444 #ifdef ASSERT 1445 assert(flat == flatten_alias_type(flat), "idempotent"); 1446 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1447 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1448 const TypeOopPtr* foop = flat->is_oopptr(); 1449 // Scalarizable allocations have exact klass always. 1450 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1451 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1452 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1453 } 1454 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1455 #endif 1456 1457 int idx = AliasIdxTop; 1458 for (int i = 0; i < num_alias_types(); i++) { 1459 if (alias_type(i)->adr_type() == flat) { 1460 idx = i; 1461 break; 1462 } 1463 } 1464 1465 if (idx == AliasIdxTop) { 1466 if (no_create) return NULL; 1467 // Grow the array if necessary. 1468 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1469 // Add a new alias type. 1470 idx = _num_alias_types++; 1471 _alias_types[idx]->Init(idx, flat); 1472 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1473 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1474 if (flat->isa_instptr()) { 1475 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1476 && flat->is_instptr()->klass() == env()->Class_klass()) 1477 alias_type(idx)->set_rewritable(false); 1478 } 1479 if (flat->isa_klassptr()) { 1480 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1481 alias_type(idx)->set_rewritable(false); 1482 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1483 alias_type(idx)->set_rewritable(false); 1484 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1485 alias_type(idx)->set_rewritable(false); 1486 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1487 alias_type(idx)->set_rewritable(false); 1488 } 1489 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1490 // but the base pointer type is not distinctive enough to identify 1491 // references into JavaThread.) 1492 1493 // Check for final fields. 1494 const TypeInstPtr* tinst = flat->isa_instptr(); 1495 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1496 ciField* field; 1497 if (tinst->const_oop() != NULL && 1498 tinst->klass() == ciEnv::current()->Class_klass() && 1499 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1500 // static field 1501 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1502 field = k->get_field_by_offset(tinst->offset(), true); 1503 } else { 1504 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1505 field = k->get_field_by_offset(tinst->offset(), false); 1506 } 1507 assert(field == NULL || 1508 original_field == NULL || 1509 (field->holder() == original_field->holder() && 1510 field->offset() == original_field->offset() && 1511 field->is_static() == original_field->is_static()), "wrong field?"); 1512 // Set field() and is_rewritable() attributes. 1513 if (field != NULL) alias_type(idx)->set_field(field); 1514 } 1515 } 1516 1517 // Fill the cache for next time. 1518 ace->_adr_type = adr_type; 1519 ace->_index = idx; 1520 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1521 1522 // Might as well try to fill the cache for the flattened version, too. 1523 AliasCacheEntry* face = probe_alias_cache(flat); 1524 if (face->_adr_type == NULL) { 1525 face->_adr_type = flat; 1526 face->_index = idx; 1527 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1528 } 1529 1530 return alias_type(idx); 1531 } 1532 1533 1534 Compile::AliasType* Compile::alias_type(ciField* field) { 1535 const TypeOopPtr* t; 1536 if (field->is_static()) 1537 t = TypeInstPtr::make(field->holder()->java_mirror()); 1538 else 1539 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1540 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1541 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1542 return atp; 1543 } 1544 1545 1546 //------------------------------have_alias_type-------------------------------- 1547 bool Compile::have_alias_type(const TypePtr* adr_type) { 1548 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1549 if (ace->_adr_type == adr_type) { 1550 return true; 1551 } 1552 1553 // Handle special cases. 1554 if (adr_type == NULL) return true; 1555 if (adr_type == TypePtr::BOTTOM) return true; 1556 1557 return find_alias_type(adr_type, true, NULL) != NULL; 1558 } 1559 1560 //-----------------------------must_alias-------------------------------------- 1561 // True if all values of the given address type are in the given alias category. 1562 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1563 if (alias_idx == AliasIdxBot) return true; // the universal category 1564 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1565 if (alias_idx == AliasIdxTop) return false; // the empty category 1566 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1567 1568 // the only remaining possible overlap is identity 1569 int adr_idx = get_alias_index(adr_type); 1570 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1571 assert(adr_idx == alias_idx || 1572 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1573 && adr_type != TypeOopPtr::BOTTOM), 1574 "should not be testing for overlap with an unsafe pointer"); 1575 return adr_idx == alias_idx; 1576 } 1577 1578 //------------------------------can_alias-------------------------------------- 1579 // True if any values of the given address type are in the given alias category. 1580 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1581 if (alias_idx == AliasIdxTop) return false; // the empty category 1582 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1583 if (alias_idx == AliasIdxBot) return true; // the universal category 1584 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1585 1586 // the only remaining possible overlap is identity 1587 int adr_idx = get_alias_index(adr_type); 1588 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1589 return adr_idx == alias_idx; 1590 } 1591 1592 1593 1594 //---------------------------pop_warm_call------------------------------------- 1595 WarmCallInfo* Compile::pop_warm_call() { 1596 WarmCallInfo* wci = _warm_calls; 1597 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1598 return wci; 1599 } 1600 1601 //----------------------------Inline_Warm-------------------------------------- 1602 int Compile::Inline_Warm() { 1603 // If there is room, try to inline some more warm call sites. 1604 // %%% Do a graph index compaction pass when we think we're out of space? 1605 if (!InlineWarmCalls) return 0; 1606 1607 int calls_made_hot = 0; 1608 int room_to_grow = NodeCountInliningCutoff - unique(); 1609 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1610 int amount_grown = 0; 1611 WarmCallInfo* call; 1612 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1613 int est_size = (int)call->size(); 1614 if (est_size > (room_to_grow - amount_grown)) { 1615 // This one won't fit anyway. Get rid of it. 1616 call->make_cold(); 1617 continue; 1618 } 1619 call->make_hot(); 1620 calls_made_hot++; 1621 amount_grown += est_size; 1622 amount_to_grow -= est_size; 1623 } 1624 1625 if (calls_made_hot > 0) set_major_progress(); 1626 return calls_made_hot; 1627 } 1628 1629 1630 //----------------------------Finish_Warm-------------------------------------- 1631 void Compile::Finish_Warm() { 1632 if (!InlineWarmCalls) return; 1633 if (failing()) return; 1634 if (warm_calls() == NULL) return; 1635 1636 // Clean up loose ends, if we are out of space for inlining. 1637 WarmCallInfo* call; 1638 while ((call = pop_warm_call()) != NULL) { 1639 call->make_cold(); 1640 } 1641 } 1642 1643 //---------------------cleanup_loop_predicates----------------------- 1644 // Remove the opaque nodes that protect the predicates so that all unused 1645 // checks and uncommon_traps will be eliminated from the ideal graph 1646 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1647 if (predicate_count()==0) return; 1648 for (int i = predicate_count(); i > 0; i--) { 1649 Node * n = predicate_opaque1_node(i-1); 1650 assert(n->Opcode() == Op_Opaque1, "must be"); 1651 igvn.replace_node(n, n->in(1)); 1652 } 1653 assert(predicate_count()==0, "should be clean!"); 1654 } 1655 1656 //------------------------------Optimize--------------------------------------- 1657 // Given a graph, optimize it. 1658 void Compile::Optimize() { 1659 TracePhase t1("optimizer", &_t_optimizer, true); 1660 1661 #ifndef PRODUCT 1662 if (env()->break_at_compile()) { 1663 BREAKPOINT; 1664 } 1665 1666 #endif 1667 1668 ResourceMark rm; 1669 int loop_opts_cnt; 1670 1671 NOT_PRODUCT( verify_graph_edges(); ) 1672 1673 print_method("After Parsing"); 1674 1675 { 1676 // Iterative Global Value Numbering, including ideal transforms 1677 // Initialize IterGVN with types and values from parse-time GVN 1678 PhaseIterGVN igvn(initial_gvn()); 1679 { 1680 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1681 igvn.optimize(); 1682 } 1683 1684 print_method("Iter GVN 1", 2); 1685 1686 if (failing()) return; 1687 1688 // Perform escape analysis 1689 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1690 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 1691 ConnectionGraph::do_analysis(this, &igvn); 1692 1693 if (failing()) return; 1694 1695 igvn.optimize(); 1696 print_method("Iter GVN 3", 2); 1697 1698 if (failing()) return; 1699 1700 } 1701 1702 // Loop transforms on the ideal graph. Range Check Elimination, 1703 // peeling, unrolling, etc. 1704 1705 // Set loop opts counter 1706 loop_opts_cnt = num_loop_opts(); 1707 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1708 { 1709 TracePhase t2("idealLoop", &_t_idealLoop, true); 1710 PhaseIdealLoop ideal_loop( igvn, true ); 1711 loop_opts_cnt--; 1712 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1713 if (failing()) return; 1714 } 1715 // Loop opts pass if partial peeling occurred in previous pass 1716 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1717 TracePhase t3("idealLoop", &_t_idealLoop, true); 1718 PhaseIdealLoop ideal_loop( igvn, false ); 1719 loop_opts_cnt--; 1720 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1721 if (failing()) return; 1722 } 1723 // Loop opts pass for loop-unrolling before CCP 1724 if(major_progress() && (loop_opts_cnt > 0)) { 1725 TracePhase t4("idealLoop", &_t_idealLoop, true); 1726 PhaseIdealLoop ideal_loop( igvn, false ); 1727 loop_opts_cnt--; 1728 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1729 } 1730 if (!failing()) { 1731 // Verify that last round of loop opts produced a valid graph 1732 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1733 PhaseIdealLoop::verify(igvn); 1734 } 1735 } 1736 if (failing()) return; 1737 1738 // Conditional Constant Propagation; 1739 PhaseCCP ccp( &igvn ); 1740 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1741 { 1742 TracePhase t2("ccp", &_t_ccp, true); 1743 ccp.do_transform(); 1744 } 1745 print_method("PhaseCPP 1", 2); 1746 1747 assert( true, "Break here to ccp.dump_old2new_map()"); 1748 1749 // Iterative Global Value Numbering, including ideal transforms 1750 { 1751 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1752 igvn = ccp; 1753 igvn.optimize(); 1754 } 1755 1756 print_method("Iter GVN 2", 2); 1757 1758 if (failing()) return; 1759 1760 // Loop transforms on the ideal graph. Range Check Elimination, 1761 // peeling, unrolling, etc. 1762 if(loop_opts_cnt > 0) { 1763 debug_only( int cnt = 0; ); 1764 while(major_progress() && (loop_opts_cnt > 0)) { 1765 TracePhase t2("idealLoop", &_t_idealLoop, true); 1766 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1767 PhaseIdealLoop ideal_loop( igvn, true); 1768 loop_opts_cnt--; 1769 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1770 if (failing()) return; 1771 } 1772 } 1773 1774 { 1775 // Verify that all previous optimizations produced a valid graph 1776 // at least to this point, even if no loop optimizations were done. 1777 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1778 PhaseIdealLoop::verify(igvn); 1779 } 1780 1781 { 1782 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1783 PhaseMacroExpand mex(igvn); 1784 if (mex.expand_macro_nodes()) { 1785 assert(failing(), "must bail out w/ explicit message"); 1786 return; 1787 } 1788 } 1789 1790 } // (End scope of igvn; run destructor if necessary for asserts.) 1791 1792 // A method with only infinite loops has no edges entering loops from root 1793 { 1794 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1795 if (final_graph_reshaping()) { 1796 assert(failing(), "must bail out w/ explicit message"); 1797 return; 1798 } 1799 } 1800 1801 print_method("Optimize finished", 2); 1802 } 1803 1804 1805 //------------------------------Code_Gen--------------------------------------- 1806 // Given a graph, generate code for it 1807 void Compile::Code_Gen() { 1808 if (failing()) return; 1809 1810 // Perform instruction selection. You might think we could reclaim Matcher 1811 // memory PDQ, but actually the Matcher is used in generating spill code. 1812 // Internals of the Matcher (including some VectorSets) must remain live 1813 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1814 // set a bit in reclaimed memory. 1815 1816 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1817 // nodes. Mapping is only valid at the root of each matched subtree. 1818 NOT_PRODUCT( verify_graph_edges(); ) 1819 1820 Node_List proj_list; 1821 Matcher m(proj_list); 1822 _matcher = &m; 1823 { 1824 TracePhase t2("matcher", &_t_matcher, true); 1825 m.match(); 1826 } 1827 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1828 // nodes. Mapping is only valid at the root of each matched subtree. 1829 NOT_PRODUCT( verify_graph_edges(); ) 1830 1831 // If you have too many nodes, or if matching has failed, bail out 1832 check_node_count(0, "out of nodes matching instructions"); 1833 if (failing()) return; 1834 1835 // Build a proper-looking CFG 1836 PhaseCFG cfg(node_arena(), root(), m); 1837 _cfg = &cfg; 1838 { 1839 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1840 cfg.Dominators(); 1841 if (failing()) return; 1842 1843 NOT_PRODUCT( verify_graph_edges(); ) 1844 1845 cfg.Estimate_Block_Frequency(); 1846 cfg.GlobalCodeMotion(m,unique(),proj_list); 1847 1848 print_method("Global code motion", 2); 1849 1850 if (failing()) return; 1851 NOT_PRODUCT( verify_graph_edges(); ) 1852 1853 debug_only( cfg.verify(); ) 1854 } 1855 NOT_PRODUCT( verify_graph_edges(); ) 1856 1857 PhaseChaitin regalloc(unique(),cfg,m); 1858 _regalloc = ®alloc; 1859 { 1860 TracePhase t2("regalloc", &_t_registerAllocation, true); 1861 // Perform any platform dependent preallocation actions. This is used, 1862 // for example, to avoid taking an implicit null pointer exception 1863 // using the frame pointer on win95. 1864 _regalloc->pd_preallocate_hook(); 1865 1866 // Perform register allocation. After Chaitin, use-def chains are 1867 // no longer accurate (at spill code) and so must be ignored. 1868 // Node->LRG->reg mappings are still accurate. 1869 _regalloc->Register_Allocate(); 1870 1871 // Bail out if the allocator builds too many nodes 1872 if (failing()) return; 1873 } 1874 1875 // Prior to register allocation we kept empty basic blocks in case the 1876 // the allocator needed a place to spill. After register allocation we 1877 // are not adding any new instructions. If any basic block is empty, we 1878 // can now safely remove it. 1879 { 1880 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1881 cfg.remove_empty(); 1882 if (do_freq_based_layout()) { 1883 PhaseBlockLayout layout(cfg); 1884 } else { 1885 cfg.set_loop_alignment(); 1886 } 1887 cfg.fixup_flow(); 1888 } 1889 1890 // Perform any platform dependent postallocation verifications. 1891 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1892 1893 // Apply peephole optimizations 1894 if( OptoPeephole ) { 1895 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1896 PhasePeephole peep( _regalloc, cfg); 1897 peep.do_transform(); 1898 } 1899 1900 // Convert Nodes to instruction bits in a buffer 1901 { 1902 // %%%% workspace merge brought two timers together for one job 1903 TracePhase t2a("output", &_t_output, true); 1904 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1905 Output(); 1906 } 1907 1908 print_method("Final Code"); 1909 1910 // He's dead, Jim. 1911 _cfg = (PhaseCFG*)0xdeadbeef; 1912 _regalloc = (PhaseChaitin*)0xdeadbeef; 1913 } 1914 1915 1916 //------------------------------dump_asm--------------------------------------- 1917 // Dump formatted assembly 1918 #ifndef PRODUCT 1919 void Compile::dump_asm(int *pcs, uint pc_limit) { 1920 bool cut_short = false; 1921 tty->print_cr("#"); 1922 tty->print("# "); _tf->dump(); tty->cr(); 1923 tty->print_cr("#"); 1924 1925 // For all blocks 1926 int pc = 0x0; // Program counter 1927 char starts_bundle = ' '; 1928 _regalloc->dump_frame(); 1929 1930 Node *n = NULL; 1931 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1932 if (VMThread::should_terminate()) { cut_short = true; break; } 1933 Block *b = _cfg->_blocks[i]; 1934 if (b->is_connector() && !Verbose) continue; 1935 n = b->_nodes[0]; 1936 if (pcs && n->_idx < pc_limit) 1937 tty->print("%3.3x ", pcs[n->_idx]); 1938 else 1939 tty->print(" "); 1940 b->dump_head( &_cfg->_bbs ); 1941 if (b->is_connector()) { 1942 tty->print_cr(" # Empty connector block"); 1943 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1944 tty->print_cr(" # Block is sole successor of call"); 1945 } 1946 1947 // For all instructions 1948 Node *delay = NULL; 1949 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1950 if (VMThread::should_terminate()) { cut_short = true; break; } 1951 n = b->_nodes[j]; 1952 if (valid_bundle_info(n)) { 1953 Bundle *bundle = node_bundling(n); 1954 if (bundle->used_in_unconditional_delay()) { 1955 delay = n; 1956 continue; 1957 } 1958 if (bundle->starts_bundle()) 1959 starts_bundle = '+'; 1960 } 1961 1962 if (WizardMode) n->dump(); 1963 1964 if( !n->is_Region() && // Dont print in the Assembly 1965 !n->is_Phi() && // a few noisely useless nodes 1966 !n->is_Proj() && 1967 !n->is_MachTemp() && 1968 !n->is_SafePointScalarObject() && 1969 !n->is_Catch() && // Would be nice to print exception table targets 1970 !n->is_MergeMem() && // Not very interesting 1971 !n->is_top() && // Debug info table constants 1972 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1973 ) { 1974 if (pcs && n->_idx < pc_limit) 1975 tty->print("%3.3x", pcs[n->_idx]); 1976 else 1977 tty->print(" "); 1978 tty->print(" %c ", starts_bundle); 1979 starts_bundle = ' '; 1980 tty->print("\t"); 1981 n->format(_regalloc, tty); 1982 tty->cr(); 1983 } 1984 1985 // If we have an instruction with a delay slot, and have seen a delay, 1986 // then back up and print it 1987 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1988 assert(delay != NULL, "no unconditional delay instruction"); 1989 if (WizardMode) delay->dump(); 1990 1991 if (node_bundling(delay)->starts_bundle()) 1992 starts_bundle = '+'; 1993 if (pcs && n->_idx < pc_limit) 1994 tty->print("%3.3x", pcs[n->_idx]); 1995 else 1996 tty->print(" "); 1997 tty->print(" %c ", starts_bundle); 1998 starts_bundle = ' '; 1999 tty->print("\t"); 2000 delay->format(_regalloc, tty); 2001 tty->print_cr(""); 2002 delay = NULL; 2003 } 2004 2005 // Dump the exception table as well 2006 if( n->is_Catch() && (Verbose || WizardMode) ) { 2007 // Print the exception table for this offset 2008 _handler_table.print_subtable_for(pc); 2009 } 2010 } 2011 2012 if (pcs && n->_idx < pc_limit) 2013 tty->print_cr("%3.3x", pcs[n->_idx]); 2014 else 2015 tty->print_cr(""); 2016 2017 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2018 2019 } // End of per-block dump 2020 tty->print_cr(""); 2021 2022 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2023 } 2024 #endif 2025 2026 //------------------------------Final_Reshape_Counts--------------------------- 2027 // This class defines counters to help identify when a method 2028 // may/must be executed using hardware with only 24-bit precision. 2029 struct Final_Reshape_Counts : public StackObj { 2030 int _call_count; // count non-inlined 'common' calls 2031 int _float_count; // count float ops requiring 24-bit precision 2032 int _double_count; // count double ops requiring more precision 2033 int _java_call_count; // count non-inlined 'java' calls 2034 int _inner_loop_count; // count loops which need alignment 2035 VectorSet _visited; // Visitation flags 2036 Node_List _tests; // Set of IfNodes & PCTableNodes 2037 2038 Final_Reshape_Counts() : 2039 _call_count(0), _float_count(0), _double_count(0), 2040 _java_call_count(0), _inner_loop_count(0), 2041 _visited( Thread::current()->resource_area() ) { } 2042 2043 void inc_call_count () { _call_count ++; } 2044 void inc_float_count () { _float_count ++; } 2045 void inc_double_count() { _double_count++; } 2046 void inc_java_call_count() { _java_call_count++; } 2047 void inc_inner_loop_count() { _inner_loop_count++; } 2048 2049 int get_call_count () const { return _call_count ; } 2050 int get_float_count () const { return _float_count ; } 2051 int get_double_count() const { return _double_count; } 2052 int get_java_call_count() const { return _java_call_count; } 2053 int get_inner_loop_count() const { return _inner_loop_count; } 2054 }; 2055 2056 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2057 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2058 // Make sure the offset goes inside the instance layout. 2059 return k->contains_field_offset(tp->offset()); 2060 // Note that OffsetBot and OffsetTop are very negative. 2061 } 2062 2063 // Eliminate trivially redundant StoreCMs and accumulate their 2064 // precedence edges. 2065 static void eliminate_redundant_card_marks(Node* n) { 2066 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2067 if (n->in(MemNode::Address)->outcnt() > 1) { 2068 // There are multiple users of the same address so it might be 2069 // possible to eliminate some of the StoreCMs 2070 Node* mem = n->in(MemNode::Memory); 2071 Node* adr = n->in(MemNode::Address); 2072 Node* val = n->in(MemNode::ValueIn); 2073 Node* prev = n; 2074 bool done = false; 2075 // Walk the chain of StoreCMs eliminating ones that match. As 2076 // long as it's a chain of single users then the optimization is 2077 // safe. Eliminating partially redundant StoreCMs would require 2078 // cloning copies down the other paths. 2079 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2080 if (adr == mem->in(MemNode::Address) && 2081 val == mem->in(MemNode::ValueIn)) { 2082 // redundant StoreCM 2083 if (mem->req() > MemNode::OopStore) { 2084 // Hasn't been processed by this code yet. 2085 n->add_prec(mem->in(MemNode::OopStore)); 2086 } else { 2087 // Already converted to precedence edge 2088 for (uint i = mem->req(); i < mem->len(); i++) { 2089 // Accumulate any precedence edges 2090 if (mem->in(i) != NULL) { 2091 n->add_prec(mem->in(i)); 2092 } 2093 } 2094 // Everything above this point has been processed. 2095 done = true; 2096 } 2097 // Eliminate the previous StoreCM 2098 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2099 assert(mem->outcnt() == 0, "should be dead"); 2100 mem->disconnect_inputs(NULL); 2101 } else { 2102 prev = mem; 2103 } 2104 mem = prev->in(MemNode::Memory); 2105 } 2106 } 2107 } 2108 2109 //------------------------------final_graph_reshaping_impl---------------------- 2110 // Implement items 1-5 from final_graph_reshaping below. 2111 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 2112 2113 if ( n->outcnt() == 0 ) return; // dead node 2114 uint nop = n->Opcode(); 2115 2116 // Check for 2-input instruction with "last use" on right input. 2117 // Swap to left input. Implements item (2). 2118 if( n->req() == 3 && // two-input instruction 2119 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2120 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2121 n->in(2)->outcnt() == 1 &&// right use IS a last use 2122 !n->in(2)->is_Con() ) { // right use is not a constant 2123 // Check for commutative opcode 2124 switch( nop ) { 2125 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2126 case Op_MaxI: case Op_MinI: 2127 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2128 case Op_AndL: case Op_XorL: case Op_OrL: 2129 case Op_AndI: case Op_XorI: case Op_OrI: { 2130 // Move "last use" input to left by swapping inputs 2131 n->swap_edges(1, 2); 2132 break; 2133 } 2134 default: 2135 break; 2136 } 2137 } 2138 2139 #ifdef ASSERT 2140 if( n->is_Mem() ) { 2141 Compile* C = Compile::current(); 2142 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2143 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2144 // oop will be recorded in oop map if load crosses safepoint 2145 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2146 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2147 "raw memory operations should have control edge"); 2148 } 2149 #endif 2150 // Count FPU ops and common calls, implements item (3) 2151 switch( nop ) { 2152 // Count all float operations that may use FPU 2153 case Op_AddF: 2154 case Op_SubF: 2155 case Op_MulF: 2156 case Op_DivF: 2157 case Op_NegF: 2158 case Op_ModF: 2159 case Op_ConvI2F: 2160 case Op_ConF: 2161 case Op_CmpF: 2162 case Op_CmpF3: 2163 // case Op_ConvL2F: // longs are split into 32-bit halves 2164 frc.inc_float_count(); 2165 break; 2166 2167 case Op_ConvF2D: 2168 case Op_ConvD2F: 2169 frc.inc_float_count(); 2170 frc.inc_double_count(); 2171 break; 2172 2173 // Count all double operations that may use FPU 2174 case Op_AddD: 2175 case Op_SubD: 2176 case Op_MulD: 2177 case Op_DivD: 2178 case Op_NegD: 2179 case Op_ModD: 2180 case Op_ConvI2D: 2181 case Op_ConvD2I: 2182 // case Op_ConvL2D: // handled by leaf call 2183 // case Op_ConvD2L: // handled by leaf call 2184 case Op_ConD: 2185 case Op_CmpD: 2186 case Op_CmpD3: 2187 frc.inc_double_count(); 2188 break; 2189 case Op_Opaque1: // Remove Opaque Nodes before matching 2190 case Op_Opaque2: // Remove Opaque Nodes before matching 2191 n->subsume_by(n->in(1)); 2192 break; 2193 case Op_CallStaticJava: 2194 case Op_CallJava: 2195 case Op_CallDynamicJava: 2196 frc.inc_java_call_count(); // Count java call site; 2197 case Op_CallRuntime: 2198 case Op_CallLeaf: 2199 case Op_CallLeafNoFP: { 2200 assert( n->is_Call(), "" ); 2201 CallNode *call = n->as_Call(); 2202 // Count call sites where the FP mode bit would have to be flipped. 2203 // Do not count uncommon runtime calls: 2204 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2205 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2206 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2207 frc.inc_call_count(); // Count the call site 2208 } else { // See if uncommon argument is shared 2209 Node *n = call->in(TypeFunc::Parms); 2210 int nop = n->Opcode(); 2211 // Clone shared simple arguments to uncommon calls, item (1). 2212 if( n->outcnt() > 1 && 2213 !n->is_Proj() && 2214 nop != Op_CreateEx && 2215 nop != Op_CheckCastPP && 2216 nop != Op_DecodeN && 2217 !n->is_Mem() ) { 2218 Node *x = n->clone(); 2219 call->set_req( TypeFunc::Parms, x ); 2220 } 2221 } 2222 break; 2223 } 2224 2225 case Op_StoreD: 2226 case Op_LoadD: 2227 case Op_LoadD_unaligned: 2228 frc.inc_double_count(); 2229 goto handle_mem; 2230 case Op_StoreF: 2231 case Op_LoadF: 2232 frc.inc_float_count(); 2233 goto handle_mem; 2234 2235 case Op_StoreCM: 2236 { 2237 // Convert OopStore dependence into precedence edge 2238 Node* prec = n->in(MemNode::OopStore); 2239 n->del_req(MemNode::OopStore); 2240 n->add_prec(prec); 2241 eliminate_redundant_card_marks(n); 2242 } 2243 2244 // fall through 2245 2246 case Op_StoreB: 2247 case Op_StoreC: 2248 case Op_StorePConditional: 2249 case Op_StoreI: 2250 case Op_StoreL: 2251 case Op_StoreIConditional: 2252 case Op_StoreLConditional: 2253 case Op_CompareAndSwapI: 2254 case Op_CompareAndSwapL: 2255 case Op_CompareAndSwapP: 2256 case Op_CompareAndSwapN: 2257 case Op_StoreP: 2258 case Op_StoreN: 2259 case Op_LoadB: 2260 case Op_LoadUB: 2261 case Op_LoadUS: 2262 case Op_LoadI: 2263 case Op_LoadUI2L: 2264 case Op_LoadKlass: 2265 case Op_LoadNKlass: 2266 case Op_LoadL: 2267 case Op_LoadL_unaligned: 2268 case Op_LoadPLocked: 2269 case Op_LoadLLocked: 2270 case Op_LoadP: 2271 case Op_LoadN: 2272 case Op_LoadRange: 2273 case Op_LoadS: { 2274 handle_mem: 2275 #ifdef ASSERT 2276 if( VerifyOptoOopOffsets ) { 2277 assert( n->is_Mem(), "" ); 2278 MemNode *mem = (MemNode*)n; 2279 // Check to see if address types have grounded out somehow. 2280 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2281 assert( !tp || oop_offset_is_sane(tp), "" ); 2282 } 2283 #endif 2284 break; 2285 } 2286 2287 case Op_AddP: { // Assert sane base pointers 2288 Node *addp = n->in(AddPNode::Address); 2289 assert( !addp->is_AddP() || 2290 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2291 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2292 "Base pointers must match" ); 2293 #ifdef _LP64 2294 if (UseCompressedOops && 2295 addp->Opcode() == Op_ConP && 2296 addp == n->in(AddPNode::Base) && 2297 n->in(AddPNode::Offset)->is_Con()) { 2298 // Use addressing with narrow klass to load with offset on x86. 2299 // On sparc loading 32-bits constant and decoding it have less 2300 // instructions (4) then load 64-bits constant (7). 2301 // Do this transformation here since IGVN will convert ConN back to ConP. 2302 const Type* t = addp->bottom_type(); 2303 if (t->isa_oopptr()) { 2304 Node* nn = NULL; 2305 2306 // Look for existing ConN node of the same exact type. 2307 Compile* C = Compile::current(); 2308 Node* r = C->root(); 2309 uint cnt = r->outcnt(); 2310 for (uint i = 0; i < cnt; i++) { 2311 Node* m = r->raw_out(i); 2312 if (m!= NULL && m->Opcode() == Op_ConN && 2313 m->bottom_type()->make_ptr() == t) { 2314 nn = m; 2315 break; 2316 } 2317 } 2318 if (nn != NULL) { 2319 // Decode a narrow oop to match address 2320 // [R12 + narrow_oop_reg<<3 + offset] 2321 nn = new (C, 2) DecodeNNode(nn, t); 2322 n->set_req(AddPNode::Base, nn); 2323 n->set_req(AddPNode::Address, nn); 2324 if (addp->outcnt() == 0) { 2325 addp->disconnect_inputs(NULL); 2326 } 2327 } 2328 } 2329 } 2330 #endif 2331 break; 2332 } 2333 2334 #ifdef _LP64 2335 case Op_CastPP: 2336 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2337 Compile* C = Compile::current(); 2338 Node* in1 = n->in(1); 2339 const Type* t = n->bottom_type(); 2340 Node* new_in1 = in1->clone(); 2341 new_in1->as_DecodeN()->set_type(t); 2342 2343 if (!Matcher::narrow_oop_use_complex_address()) { 2344 // 2345 // x86, ARM and friends can handle 2 adds in addressing mode 2346 // and Matcher can fold a DecodeN node into address by using 2347 // a narrow oop directly and do implicit NULL check in address: 2348 // 2349 // [R12 + narrow_oop_reg<<3 + offset] 2350 // NullCheck narrow_oop_reg 2351 // 2352 // On other platforms (Sparc) we have to keep new DecodeN node and 2353 // use it to do implicit NULL check in address: 2354 // 2355 // decode_not_null narrow_oop_reg, base_reg 2356 // [base_reg + offset] 2357 // NullCheck base_reg 2358 // 2359 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2360 // to keep the information to which NULL check the new DecodeN node 2361 // corresponds to use it as value in implicit_null_check(). 2362 // 2363 new_in1->set_req(0, n->in(0)); 2364 } 2365 2366 n->subsume_by(new_in1); 2367 if (in1->outcnt() == 0) { 2368 in1->disconnect_inputs(NULL); 2369 } 2370 } 2371 break; 2372 2373 case Op_CmpP: 2374 // Do this transformation here to preserve CmpPNode::sub() and 2375 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2376 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2377 Node* in1 = n->in(1); 2378 Node* in2 = n->in(2); 2379 if (!in1->is_DecodeN()) { 2380 in2 = in1; 2381 in1 = n->in(2); 2382 } 2383 assert(in1->is_DecodeN(), "sanity"); 2384 2385 Compile* C = Compile::current(); 2386 Node* new_in2 = NULL; 2387 if (in2->is_DecodeN()) { 2388 new_in2 = in2->in(1); 2389 } else if (in2->Opcode() == Op_ConP) { 2390 const Type* t = in2->bottom_type(); 2391 if (t == TypePtr::NULL_PTR) { 2392 // Don't convert CmpP null check into CmpN if compressed 2393 // oops implicit null check is not generated. 2394 // This will allow to generate normal oop implicit null check. 2395 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2396 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2397 // 2398 // This transformation together with CastPP transformation above 2399 // will generated code for implicit NULL checks for compressed oops. 2400 // 2401 // The original code after Optimize() 2402 // 2403 // LoadN memory, narrow_oop_reg 2404 // decode narrow_oop_reg, base_reg 2405 // CmpP base_reg, NULL 2406 // CastPP base_reg // NotNull 2407 // Load [base_reg + offset], val_reg 2408 // 2409 // after these transformations will be 2410 // 2411 // LoadN memory, narrow_oop_reg 2412 // CmpN narrow_oop_reg, NULL 2413 // decode_not_null narrow_oop_reg, base_reg 2414 // Load [base_reg + offset], val_reg 2415 // 2416 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2417 // since narrow oops can be used in debug info now (see the code in 2418 // final_graph_reshaping_walk()). 2419 // 2420 // At the end the code will be matched to 2421 // on x86: 2422 // 2423 // Load_narrow_oop memory, narrow_oop_reg 2424 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2425 // NullCheck narrow_oop_reg 2426 // 2427 // and on sparc: 2428 // 2429 // Load_narrow_oop memory, narrow_oop_reg 2430 // decode_not_null narrow_oop_reg, base_reg 2431 // Load [base_reg + offset], val_reg 2432 // NullCheck base_reg 2433 // 2434 } else if (t->isa_oopptr()) { 2435 new_in2 = ConNode::make(C, t->make_narrowoop()); 2436 } 2437 } 2438 if (new_in2 != NULL) { 2439 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2440 n->subsume_by( cmpN ); 2441 if (in1->outcnt() == 0) { 2442 in1->disconnect_inputs(NULL); 2443 } 2444 if (in2->outcnt() == 0) { 2445 in2->disconnect_inputs(NULL); 2446 } 2447 } 2448 } 2449 break; 2450 2451 case Op_DecodeN: 2452 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2453 // DecodeN could be pinned when it can't be fold into 2454 // an address expression, see the code for Op_CastPP above. 2455 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2456 break; 2457 2458 case Op_EncodeP: { 2459 Node* in1 = n->in(1); 2460 if (in1->is_DecodeN()) { 2461 n->subsume_by(in1->in(1)); 2462 } else if (in1->Opcode() == Op_ConP) { 2463 Compile* C = Compile::current(); 2464 const Type* t = in1->bottom_type(); 2465 if (t == TypePtr::NULL_PTR) { 2466 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2467 } else if (t->isa_oopptr()) { 2468 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2469 } 2470 } 2471 if (in1->outcnt() == 0) { 2472 in1->disconnect_inputs(NULL); 2473 } 2474 break; 2475 } 2476 2477 case Op_Proj: { 2478 if (OptimizeStringConcat) { 2479 ProjNode* p = n->as_Proj(); 2480 if (p->_is_io_use) { 2481 // Separate projections were used for the exception path which 2482 // are normally removed by a late inline. If it wasn't inlined 2483 // then they will hang around and should just be replaced with 2484 // the original one. 2485 Node* proj = NULL; 2486 // Replace with just one 2487 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2488 Node *use = i.get(); 2489 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2490 proj = use; 2491 break; 2492 } 2493 } 2494 assert(p != NULL, "must be found"); 2495 p->subsume_by(proj); 2496 } 2497 } 2498 break; 2499 } 2500 2501 case Op_Phi: 2502 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2503 // The EncodeP optimization may create Phi with the same edges 2504 // for all paths. It is not handled well by Register Allocator. 2505 Node* unique_in = n->in(1); 2506 assert(unique_in != NULL, ""); 2507 uint cnt = n->req(); 2508 for (uint i = 2; i < cnt; i++) { 2509 Node* m = n->in(i); 2510 assert(m != NULL, ""); 2511 if (unique_in != m) 2512 unique_in = NULL; 2513 } 2514 if (unique_in != NULL) { 2515 n->subsume_by(unique_in); 2516 } 2517 } 2518 break; 2519 2520 #endif 2521 2522 case Op_ModI: 2523 if (UseDivMod) { 2524 // Check if a%b and a/b both exist 2525 Node* d = n->find_similar(Op_DivI); 2526 if (d) { 2527 // Replace them with a fused divmod if supported 2528 Compile* C = Compile::current(); 2529 if (Matcher::has_match_rule(Op_DivModI)) { 2530 DivModINode* divmod = DivModINode::make(C, n); 2531 d->subsume_by(divmod->div_proj()); 2532 n->subsume_by(divmod->mod_proj()); 2533 } else { 2534 // replace a%b with a-((a/b)*b) 2535 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2536 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2537 n->subsume_by( sub ); 2538 } 2539 } 2540 } 2541 break; 2542 2543 case Op_ModL: 2544 if (UseDivMod) { 2545 // Check if a%b and a/b both exist 2546 Node* d = n->find_similar(Op_DivL); 2547 if (d) { 2548 // Replace them with a fused divmod if supported 2549 Compile* C = Compile::current(); 2550 if (Matcher::has_match_rule(Op_DivModL)) { 2551 DivModLNode* divmod = DivModLNode::make(C, n); 2552 d->subsume_by(divmod->div_proj()); 2553 n->subsume_by(divmod->mod_proj()); 2554 } else { 2555 // replace a%b with a-((a/b)*b) 2556 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2557 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2558 n->subsume_by( sub ); 2559 } 2560 } 2561 } 2562 break; 2563 2564 case Op_Load16B: 2565 case Op_Load8B: 2566 case Op_Load4B: 2567 case Op_Load8S: 2568 case Op_Load4S: 2569 case Op_Load2S: 2570 case Op_Load8C: 2571 case Op_Load4C: 2572 case Op_Load2C: 2573 case Op_Load4I: 2574 case Op_Load2I: 2575 case Op_Load2L: 2576 case Op_Load4F: 2577 case Op_Load2F: 2578 case Op_Load2D: 2579 case Op_Store16B: 2580 case Op_Store8B: 2581 case Op_Store4B: 2582 case Op_Store8C: 2583 case Op_Store4C: 2584 case Op_Store2C: 2585 case Op_Store4I: 2586 case Op_Store2I: 2587 case Op_Store2L: 2588 case Op_Store4F: 2589 case Op_Store2F: 2590 case Op_Store2D: 2591 break; 2592 2593 case Op_PackB: 2594 case Op_PackS: 2595 case Op_PackC: 2596 case Op_PackI: 2597 case Op_PackF: 2598 case Op_PackL: 2599 case Op_PackD: 2600 if (n->req()-1 > 2) { 2601 // Replace many operand PackNodes with a binary tree for matching 2602 PackNode* p = (PackNode*) n; 2603 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2604 n->subsume_by(btp); 2605 } 2606 break; 2607 case Op_Loop: 2608 case Op_CountedLoop: 2609 if (n->as_Loop()->is_inner_loop()) { 2610 frc.inc_inner_loop_count(); 2611 } 2612 break; 2613 case Op_LShiftI: 2614 case Op_RShiftI: 2615 case Op_URShiftI: 2616 case Op_LShiftL: 2617 case Op_RShiftL: 2618 case Op_URShiftL: 2619 if (Matcher::need_masked_shift_count) { 2620 // The cpu's shift instructions don't restrict the count to the 2621 // lower 5/6 bits. We need to do the masking ourselves. 2622 Node* in2 = n->in(2); 2623 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2624 const TypeInt* t = in2->find_int_type(); 2625 if (t != NULL && t->is_con()) { 2626 juint shift = t->get_con(); 2627 if (shift > mask) { // Unsigned cmp 2628 Compile* C = Compile::current(); 2629 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask))); 2630 } 2631 } else { 2632 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2633 Compile* C = Compile::current(); 2634 Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask))); 2635 n->set_req(2, shift); 2636 } 2637 } 2638 if (in2->outcnt() == 0) { // Remove dead node 2639 in2->disconnect_inputs(NULL); 2640 } 2641 } 2642 break; 2643 default: 2644 assert( !n->is_Call(), "" ); 2645 assert( !n->is_Mem(), "" ); 2646 break; 2647 } 2648 2649 // Collect CFG split points 2650 if (n->is_MultiBranch()) 2651 frc._tests.push(n); 2652 } 2653 2654 //------------------------------final_graph_reshaping_walk--------------------- 2655 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2656 // requires that the walk visits a node's inputs before visiting the node. 2657 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2658 ResourceArea *area = Thread::current()->resource_area(); 2659 Unique_Node_List sfpt(area); 2660 2661 frc._visited.set(root->_idx); // first, mark node as visited 2662 uint cnt = root->req(); 2663 Node *n = root; 2664 uint i = 0; 2665 while (true) { 2666 if (i < cnt) { 2667 // Place all non-visited non-null inputs onto stack 2668 Node* m = n->in(i); 2669 ++i; 2670 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2671 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2672 sfpt.push(m); 2673 cnt = m->req(); 2674 nstack.push(n, i); // put on stack parent and next input's index 2675 n = m; 2676 i = 0; 2677 } 2678 } else { 2679 // Now do post-visit work 2680 final_graph_reshaping_impl( n, frc ); 2681 if (nstack.is_empty()) 2682 break; // finished 2683 n = nstack.node(); // Get node from stack 2684 cnt = n->req(); 2685 i = nstack.index(); 2686 nstack.pop(); // Shift to the next node on stack 2687 } 2688 } 2689 2690 // Skip next transformation if compressed oops are not used. 2691 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) 2692 return; 2693 2694 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2695 // It could be done for an uncommon traps or any safepoints/calls 2696 // if the DecodeN node is referenced only in a debug info. 2697 while (sfpt.size() > 0) { 2698 n = sfpt.pop(); 2699 JVMState *jvms = n->as_SafePoint()->jvms(); 2700 assert(jvms != NULL, "sanity"); 2701 int start = jvms->debug_start(); 2702 int end = n->req(); 2703 bool is_uncommon = (n->is_CallStaticJava() && 2704 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2705 for (int j = start; j < end; j++) { 2706 Node* in = n->in(j); 2707 if (in->is_DecodeN()) { 2708 bool safe_to_skip = true; 2709 if (!is_uncommon ) { 2710 // Is it safe to skip? 2711 for (uint i = 0; i < in->outcnt(); i++) { 2712 Node* u = in->raw_out(i); 2713 if (!u->is_SafePoint() || 2714 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2715 safe_to_skip = false; 2716 } 2717 } 2718 } 2719 if (safe_to_skip) { 2720 n->set_req(j, in->in(1)); 2721 } 2722 if (in->outcnt() == 0) { 2723 in->disconnect_inputs(NULL); 2724 } 2725 } 2726 } 2727 } 2728 } 2729 2730 //------------------------------final_graph_reshaping-------------------------- 2731 // Final Graph Reshaping. 2732 // 2733 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2734 // and not commoned up and forced early. Must come after regular 2735 // optimizations to avoid GVN undoing the cloning. Clone constant 2736 // inputs to Loop Phis; these will be split by the allocator anyways. 2737 // Remove Opaque nodes. 2738 // (2) Move last-uses by commutative operations to the left input to encourage 2739 // Intel update-in-place two-address operations and better register usage 2740 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2741 // calls canonicalizing them back. 2742 // (3) Count the number of double-precision FP ops, single-precision FP ops 2743 // and call sites. On Intel, we can get correct rounding either by 2744 // forcing singles to memory (requires extra stores and loads after each 2745 // FP bytecode) or we can set a rounding mode bit (requires setting and 2746 // clearing the mode bit around call sites). The mode bit is only used 2747 // if the relative frequency of single FP ops to calls is low enough. 2748 // This is a key transform for SPEC mpeg_audio. 2749 // (4) Detect infinite loops; blobs of code reachable from above but not 2750 // below. Several of the Code_Gen algorithms fail on such code shapes, 2751 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2752 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2753 // Detection is by looking for IfNodes where only 1 projection is 2754 // reachable from below or CatchNodes missing some targets. 2755 // (5) Assert for insane oop offsets in debug mode. 2756 2757 bool Compile::final_graph_reshaping() { 2758 // an infinite loop may have been eliminated by the optimizer, 2759 // in which case the graph will be empty. 2760 if (root()->req() == 1) { 2761 record_method_not_compilable("trivial infinite loop"); 2762 return true; 2763 } 2764 2765 Final_Reshape_Counts frc; 2766 2767 // Visit everybody reachable! 2768 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2769 Node_Stack nstack(unique() >> 1); 2770 final_graph_reshaping_walk(nstack, root(), frc); 2771 2772 // Check for unreachable (from below) code (i.e., infinite loops). 2773 for( uint i = 0; i < frc._tests.size(); i++ ) { 2774 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2775 // Get number of CFG targets. 2776 // Note that PCTables include exception targets after calls. 2777 uint required_outcnt = n->required_outcnt(); 2778 if (n->outcnt() != required_outcnt) { 2779 // Check for a few special cases. Rethrow Nodes never take the 2780 // 'fall-thru' path, so expected kids is 1 less. 2781 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2782 if (n->in(0)->in(0)->is_Call()) { 2783 CallNode *call = n->in(0)->in(0)->as_Call(); 2784 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2785 required_outcnt--; // Rethrow always has 1 less kid 2786 } else if (call->req() > TypeFunc::Parms && 2787 call->is_CallDynamicJava()) { 2788 // Check for null receiver. In such case, the optimizer has 2789 // detected that the virtual call will always result in a null 2790 // pointer exception. The fall-through projection of this CatchNode 2791 // will not be populated. 2792 Node *arg0 = call->in(TypeFunc::Parms); 2793 if (arg0->is_Type() && 2794 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2795 required_outcnt--; 2796 } 2797 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2798 call->req() > TypeFunc::Parms+1 && 2799 call->is_CallStaticJava()) { 2800 // Check for negative array length. In such case, the optimizer has 2801 // detected that the allocation attempt will always result in an 2802 // exception. There is no fall-through projection of this CatchNode . 2803 Node *arg1 = call->in(TypeFunc::Parms+1); 2804 if (arg1->is_Type() && 2805 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2806 required_outcnt--; 2807 } 2808 } 2809 } 2810 } 2811 // Recheck with a better notion of 'required_outcnt' 2812 if (n->outcnt() != required_outcnt) { 2813 record_method_not_compilable("malformed control flow"); 2814 return true; // Not all targets reachable! 2815 } 2816 } 2817 // Check that I actually visited all kids. Unreached kids 2818 // must be infinite loops. 2819 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2820 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2821 record_method_not_compilable("infinite loop"); 2822 return true; // Found unvisited kid; must be unreach 2823 } 2824 } 2825 2826 // If original bytecodes contained a mixture of floats and doubles 2827 // check if the optimizer has made it homogenous, item (3). 2828 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2829 frc.get_float_count() > 32 && 2830 frc.get_double_count() == 0 && 2831 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2832 set_24_bit_selection_and_mode( false, true ); 2833 } 2834 2835 set_java_calls(frc.get_java_call_count()); 2836 set_inner_loops(frc.get_inner_loop_count()); 2837 2838 // No infinite loops, no reason to bail out. 2839 return false; 2840 } 2841 2842 //-----------------------------too_many_traps---------------------------------- 2843 // Report if there are too many traps at the current method and bci. 2844 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2845 bool Compile::too_many_traps(ciMethod* method, 2846 int bci, 2847 Deoptimization::DeoptReason reason) { 2848 ciMethodData* md = method->method_data(); 2849 if (md->is_empty()) { 2850 // Assume the trap has not occurred, or that it occurred only 2851 // because of a transient condition during start-up in the interpreter. 2852 return false; 2853 } 2854 if (md->has_trap_at(bci, reason) != 0) { 2855 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2856 // Also, if there are multiple reasons, or if there is no per-BCI record, 2857 // assume the worst. 2858 if (log()) 2859 log()->elem("observe trap='%s' count='%d'", 2860 Deoptimization::trap_reason_name(reason), 2861 md->trap_count(reason)); 2862 return true; 2863 } else { 2864 // Ignore method/bci and see if there have been too many globally. 2865 return too_many_traps(reason, md); 2866 } 2867 } 2868 2869 // Less-accurate variant which does not require a method and bci. 2870 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2871 ciMethodData* logmd) { 2872 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2873 // Too many traps globally. 2874 // Note that we use cumulative trap_count, not just md->trap_count. 2875 if (log()) { 2876 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2877 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2878 Deoptimization::trap_reason_name(reason), 2879 mcount, trap_count(reason)); 2880 } 2881 return true; 2882 } else { 2883 // The coast is clear. 2884 return false; 2885 } 2886 } 2887 2888 //--------------------------too_many_recompiles-------------------------------- 2889 // Report if there are too many recompiles at the current method and bci. 2890 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2891 // Is not eager to return true, since this will cause the compiler to use 2892 // Action_none for a trap point, to avoid too many recompilations. 2893 bool Compile::too_many_recompiles(ciMethod* method, 2894 int bci, 2895 Deoptimization::DeoptReason reason) { 2896 ciMethodData* md = method->method_data(); 2897 if (md->is_empty()) { 2898 // Assume the trap has not occurred, or that it occurred only 2899 // because of a transient condition during start-up in the interpreter. 2900 return false; 2901 } 2902 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2903 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2904 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2905 Deoptimization::DeoptReason per_bc_reason 2906 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2907 if ((per_bc_reason == Deoptimization::Reason_none 2908 || md->has_trap_at(bci, reason) != 0) 2909 // The trap frequency measure we care about is the recompile count: 2910 && md->trap_recompiled_at(bci) 2911 && md->overflow_recompile_count() >= bc_cutoff) { 2912 // Do not emit a trap here if it has already caused recompilations. 2913 // Also, if there are multiple reasons, or if there is no per-BCI record, 2914 // assume the worst. 2915 if (log()) 2916 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2917 Deoptimization::trap_reason_name(reason), 2918 md->trap_count(reason), 2919 md->overflow_recompile_count()); 2920 return true; 2921 } else if (trap_count(reason) != 0 2922 && decompile_count() >= m_cutoff) { 2923 // Too many recompiles globally, and we have seen this sort of trap. 2924 // Use cumulative decompile_count, not just md->decompile_count. 2925 if (log()) 2926 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2927 Deoptimization::trap_reason_name(reason), 2928 md->trap_count(reason), trap_count(reason), 2929 md->decompile_count(), decompile_count()); 2930 return true; 2931 } else { 2932 // The coast is clear. 2933 return false; 2934 } 2935 } 2936 2937 2938 #ifndef PRODUCT 2939 //------------------------------verify_graph_edges--------------------------- 2940 // Walk the Graph and verify that there is a one-to-one correspondence 2941 // between Use-Def edges and Def-Use edges in the graph. 2942 void Compile::verify_graph_edges(bool no_dead_code) { 2943 if (VerifyGraphEdges) { 2944 ResourceArea *area = Thread::current()->resource_area(); 2945 Unique_Node_List visited(area); 2946 // Call recursive graph walk to check edges 2947 _root->verify_edges(visited); 2948 if (no_dead_code) { 2949 // Now make sure that no visited node is used by an unvisited node. 2950 bool dead_nodes = 0; 2951 Unique_Node_List checked(area); 2952 while (visited.size() > 0) { 2953 Node* n = visited.pop(); 2954 checked.push(n); 2955 for (uint i = 0; i < n->outcnt(); i++) { 2956 Node* use = n->raw_out(i); 2957 if (checked.member(use)) continue; // already checked 2958 if (visited.member(use)) continue; // already in the graph 2959 if (use->is_Con()) continue; // a dead ConNode is OK 2960 // At this point, we have found a dead node which is DU-reachable. 2961 if (dead_nodes++ == 0) 2962 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2963 use->dump(2); 2964 tty->print_cr("---"); 2965 checked.push(use); // No repeats; pretend it is now checked. 2966 } 2967 } 2968 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2969 } 2970 } 2971 } 2972 #endif 2973 2974 // The Compile object keeps track of failure reasons separately from the ciEnv. 2975 // This is required because there is not quite a 1-1 relation between the 2976 // ciEnv and its compilation task and the Compile object. Note that one 2977 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2978 // to backtrack and retry without subsuming loads. Other than this backtracking 2979 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2980 // by the logic in C2Compiler. 2981 void Compile::record_failure(const char* reason) { 2982 if (log() != NULL) { 2983 log()->elem("failure reason='%s' phase='compile'", reason); 2984 } 2985 if (_failure_reason == NULL) { 2986 // Record the first failure reason. 2987 _failure_reason = reason; 2988 } 2989 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2990 C->print_method(_failure_reason); 2991 } 2992 _root = NULL; // flush the graph, too 2993 } 2994 2995 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 2996 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 2997 { 2998 if (dolog) { 2999 C = Compile::current(); 3000 _log = C->log(); 3001 } else { 3002 C = NULL; 3003 _log = NULL; 3004 } 3005 if (_log != NULL) { 3006 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 3007 _log->stamp(); 3008 _log->end_head(); 3009 } 3010 } 3011 3012 Compile::TracePhase::~TracePhase() { 3013 if (_log != NULL) { 3014 _log->done("phase nodes='%d'", C->unique()); 3015 } 3016 } 3017 3018 //============================================================================= 3019 // Two Constant's are equal when the type and the value are equal. 3020 bool Compile::Constant::operator==(const Constant& other) { 3021 if (type() != other.type() ) return false; 3022 if (can_be_reused() != other.can_be_reused()) return false; 3023 // For floating point values we compare the bit pattern. 3024 switch (type()) { 3025 case T_FLOAT: return (_value.i == other._value.i); 3026 case T_LONG: 3027 case T_DOUBLE: return (_value.j == other._value.j); 3028 case T_OBJECT: 3029 case T_ADDRESS: return (_value.l == other._value.l); 3030 case T_VOID: return (_value.l == other._value.l); // jump-table entries 3031 default: ShouldNotReachHere(); 3032 } 3033 return false; 3034 } 3035 3036 // Emit constants grouped in the following order: 3037 static BasicType type_order[] = { 3038 T_FLOAT, // 32-bit 3039 T_OBJECT, // 32 or 64-bit 3040 T_ADDRESS, // 32 or 64-bit 3041 T_DOUBLE, // 64-bit 3042 T_LONG, // 64-bit 3043 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons) 3044 T_ILLEGAL 3045 }; 3046 3047 static int type_to_size_in_bytes(BasicType t) { 3048 switch (t) { 3049 case T_LONG: return sizeof(jlong ); 3050 case T_FLOAT: return sizeof(jfloat ); 3051 case T_DOUBLE: return sizeof(jdouble); 3052 // We use T_VOID as marker for jump-table entries (labels) which 3053 // need an interal word relocation. 3054 case T_VOID: 3055 case T_ADDRESS: 3056 case T_OBJECT: return sizeof(jobject); 3057 } 3058 3059 ShouldNotReachHere(); 3060 return -1; 3061 } 3062 3063 void Compile::ConstantTable::calculate_offsets_and_size() { 3064 int size = 0; 3065 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 3066 BasicType type = type_order[t]; 3067 3068 for (int i = 0; i < _constants.length(); i++) { 3069 Constant con = _constants.at(i); 3070 if (con.type() != type) continue; // Skip other types. 3071 3072 // Align size for type. 3073 int typesize = type_to_size_in_bytes(con.type()); 3074 size = align_size_up(size, typesize); 3075 3076 // Set offset. 3077 con.set_offset(size); 3078 _constants.at_put(i, con); 3079 3080 // Add type size. 3081 size = size + typesize; 3082 } 3083 } 3084 3085 // Align size up to the next section start (which is insts; see 3086 // CodeBuffer::align_at_start). 3087 assert(_size == -1, "already set?"); 3088 _size = align_size_up(size, CodeEntryAlignment); 3089 3090 if (Matcher::constant_table_absolute_addressing) { 3091 set_table_base_offset(0); // No table base offset required 3092 } else { 3093 if (UseRDPCForConstantTableBase) { 3094 // table base offset is set in MachConstantBaseNode::emit 3095 } else { 3096 // When RDPC is not used, the table base is set into the middle of 3097 // the constant table. 3098 int half_size = _size / 2; 3099 assert(half_size * 2 == _size, "sanity"); 3100 set_table_base_offset(-half_size); 3101 } 3102 } 3103 } 3104 3105 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3106 MacroAssembler _masm(&cb); 3107 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 3108 BasicType type = type_order[t]; 3109 3110 for (int i = 0; i < _constants.length(); i++) { 3111 Constant con = _constants.at(i); 3112 if (con.type() != type) continue; // Skip other types. 3113 3114 address constant_addr; 3115 switch (con.type()) { 3116 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3117 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3118 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3119 case T_OBJECT: { 3120 jobject obj = con.get_jobject(); 3121 int oop_index = _masm.oop_recorder()->find_index(obj); 3122 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3123 break; 3124 } 3125 case T_ADDRESS: { 3126 address addr = (address) con.get_jobject(); 3127 constant_addr = _masm.address_constant(addr); 3128 break; 3129 } 3130 // We use T_VOID as marker for jump-table entries (labels) which 3131 // need an interal word relocation. 3132 case T_VOID: { 3133 // Write a dummy word. The real value is filled in later 3134 // in fill_jump_table_in_constant_table. 3135 address addr = (address) con.get_jobject(); 3136 constant_addr = _masm.address_constant(addr); 3137 break; 3138 } 3139 default: ShouldNotReachHere(); 3140 } 3141 assert(constant_addr != NULL, "consts section too small"); 3142 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3143 } 3144 } 3145 } 3146 3147 int Compile::ConstantTable::find_offset(Constant& con) const { 3148 int idx = _constants.find(con); 3149 assert(idx != -1, "constant must be in constant table"); 3150 int offset = _constants.at(idx).offset(); 3151 assert(offset != -1, "constant table not emitted yet?"); 3152 return offset; 3153 } 3154 3155 void Compile::ConstantTable::add(Constant& con) { 3156 if (con.can_be_reused()) { 3157 int idx = _constants.find(con); 3158 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3159 return; 3160 } 3161 } 3162 (void) _constants.append(con); 3163 } 3164 3165 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) { 3166 Constant con(type, value); 3167 add(con); 3168 return con; 3169 } 3170 3171 Compile::Constant Compile::ConstantTable::add(MachOper* oper) { 3172 jvalue value; 3173 BasicType type = oper->type()->basic_type(); 3174 switch (type) { 3175 case T_LONG: value.j = oper->constantL(); break; 3176 case T_FLOAT: value.f = oper->constantF(); break; 3177 case T_DOUBLE: value.d = oper->constantD(); break; 3178 case T_OBJECT: 3179 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3180 default: ShouldNotReachHere(); 3181 } 3182 return add(type, value); 3183 } 3184 3185 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) { 3186 jvalue value; 3187 // We can use the node pointer here to identify the right jump-table 3188 // as this method is called from Compile::Fill_buffer right before 3189 // the MachNodes are emitted and the jump-table is filled (means the 3190 // MachNode pointers do not change anymore). 3191 value.l = (jobject) n; 3192 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused. 3193 for (uint i = 0; i < n->outcnt(); i++) { 3194 add(con); 3195 } 3196 return con; 3197 } 3198 3199 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3200 // If called from Compile::scratch_emit_size do nothing. 3201 if (Compile::current()->in_scratch_emit_size()) return; 3202 3203 assert(labels.is_nonempty(), "must be"); 3204 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt())); 3205 3206 // Since MachConstantNode::constant_offset() also contains 3207 // table_base_offset() we need to subtract the table_base_offset() 3208 // to get the plain offset into the constant table. 3209 int offset = n->constant_offset() - table_base_offset(); 3210 3211 MacroAssembler _masm(&cb); 3212 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3213 3214 for (int i = 0; i < labels.length(); i++) { 3215 address* constant_addr = &jump_table_base[i]; 3216 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer"); 3217 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3218 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3219 } 3220 }