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 Label fakeL; // Fake label for branch instructions. 522 Label* saveL = NULL; 523 uint save_bnum = 0; 524 bool is_branch = n->is_MachBranch(); 525 if (is_branch) { 526 MacroAssembler masm(&buf); 527 masm.bind(fakeL); 528 n->as_MachBranch()->save_label(&saveL, &save_bnum); 529 n->as_MachBranch()->label_set(&fakeL, 0); 530 } 531 n->emit(buf, this->regalloc()); 532 if (is_branch) // Restore label. 533 n->as_MachBranch()->label_set(saveL, save_bnum); 534 535 // End scratch_emit_size section. 536 set_in_scratch_emit_size(false); 537 538 return buf.insts_size(); 539 } 540 541 542 // ============================================================================ 543 //------------------------------Compile standard------------------------------- 544 debug_only( int Compile::_debug_idx = 100000; ) 545 546 // Compile a method. entry_bci is -1 for normal compilations and indicates 547 // the continuation bci for on stack replacement. 548 549 550 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 551 : Phase(Compiler), 552 _env(ci_env), 553 _log(ci_env->log()), 554 _compile_id(ci_env->compile_id()), 555 _save_argument_registers(false), 556 _stub_name(NULL), 557 _stub_function(NULL), 558 _stub_entry_point(NULL), 559 _method(target), 560 _entry_bci(osr_bci), 561 _initial_gvn(NULL), 562 _for_igvn(NULL), 563 _warm_calls(NULL), 564 _subsume_loads(subsume_loads), 565 _do_escape_analysis(do_escape_analysis), 566 _failure_reason(NULL), 567 _code_buffer("Compile::Fill_buffer"), 568 _orig_pc_slot(0), 569 _orig_pc_slot_offset_in_bytes(0), 570 _has_method_handle_invokes(false), 571 _mach_constant_base_node(NULL), 572 _node_bundling_limit(0), 573 _node_bundling_base(NULL), 574 _java_calls(0), 575 _inner_loops(0), 576 _scratch_const_size(-1), 577 _in_scratch_emit_size(false), 578 #ifndef PRODUCT 579 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 580 _printer(IdealGraphPrinter::printer()), 581 #endif 582 _congraph(NULL) { 583 C = this; 584 585 CompileWrapper cw(this); 586 #ifndef PRODUCT 587 if (TimeCompiler2) { 588 tty->print(" "); 589 target->holder()->name()->print(); 590 tty->print("."); 591 target->print_short_name(); 592 tty->print(" "); 593 } 594 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 595 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 596 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 597 if (!print_opto_assembly) { 598 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 599 if (print_assembly && !Disassembler::can_decode()) { 600 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 601 print_opto_assembly = true; 602 } 603 } 604 set_print_assembly(print_opto_assembly); 605 set_parsed_irreducible_loop(false); 606 #endif 607 608 if (ProfileTraps) { 609 // Make sure the method being compiled gets its own MDO, 610 // so we can at least track the decompile_count(). 611 method()->ensure_method_data(); 612 } 613 614 Init(::AliasLevel); 615 616 617 print_compile_messages(); 618 619 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 620 _ilt = InlineTree::build_inline_tree_root(); 621 else 622 _ilt = NULL; 623 624 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 625 assert(num_alias_types() >= AliasIdxRaw, ""); 626 627 #define MINIMUM_NODE_HASH 1023 628 // Node list that Iterative GVN will start with 629 Unique_Node_List for_igvn(comp_arena()); 630 set_for_igvn(&for_igvn); 631 632 // GVN that will be run immediately on new nodes 633 uint estimated_size = method()->code_size()*4+64; 634 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 635 PhaseGVN gvn(node_arena(), estimated_size); 636 set_initial_gvn(&gvn); 637 638 { // Scope for timing the parser 639 TracePhase t3("parse", &_t_parser, true); 640 641 // Put top into the hash table ASAP. 642 initial_gvn()->transform_no_reclaim(top()); 643 644 // Set up tf(), start(), and find a CallGenerator. 645 CallGenerator* cg = NULL; 646 if (is_osr_compilation()) { 647 const TypeTuple *domain = StartOSRNode::osr_domain(); 648 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 649 init_tf(TypeFunc::make(domain, range)); 650 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 651 initial_gvn()->set_type_bottom(s); 652 init_start(s); 653 cg = CallGenerator::for_osr(method(), entry_bci()); 654 } else { 655 // Normal case. 656 init_tf(TypeFunc::make(method())); 657 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 658 initial_gvn()->set_type_bottom(s); 659 init_start(s); 660 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 661 // With java.lang.ref.reference.get() we must go through the 662 // intrinsic when G1 is enabled - even when get() is the root 663 // method of the compile - so that, if necessary, the value in 664 // the referent field of the reference object gets recorded by 665 // the pre-barrier code. 666 // Specifically, if G1 is enabled, the value in the referent 667 // field is recorded by the G1 SATB pre barrier. This will 668 // result in the referent being marked live and the reference 669 // object removed from the list of discovered references during 670 // reference processing. 671 cg = find_intrinsic(method(), false); 672 } 673 if (cg == NULL) { 674 float past_uses = method()->interpreter_invocation_count(); 675 float expected_uses = past_uses; 676 cg = CallGenerator::for_inline(method(), expected_uses); 677 } 678 } 679 if (failing()) return; 680 if (cg == NULL) { 681 record_method_not_compilable_all_tiers("cannot parse method"); 682 return; 683 } 684 JVMState* jvms = build_start_state(start(), tf()); 685 if ((jvms = cg->generate(jvms)) == NULL) { 686 record_method_not_compilable("method parse failed"); 687 return; 688 } 689 GraphKit kit(jvms); 690 691 if (!kit.stopped()) { 692 // Accept return values, and transfer control we know not where. 693 // This is done by a special, unique ReturnNode bound to root. 694 return_values(kit.jvms()); 695 } 696 697 if (kit.has_exceptions()) { 698 // Any exceptions that escape from this call must be rethrown 699 // to whatever caller is dynamically above us on the stack. 700 // This is done by a special, unique RethrowNode bound to root. 701 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 702 } 703 704 if (!failing() && has_stringbuilder()) { 705 { 706 // remove useless nodes to make the usage analysis simpler 707 ResourceMark rm; 708 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 709 } 710 711 { 712 ResourceMark rm; 713 print_method("Before StringOpts", 3); 714 PhaseStringOpts pso(initial_gvn(), &for_igvn); 715 print_method("After StringOpts", 3); 716 } 717 718 // now inline anything that we skipped the first time around 719 while (_late_inlines.length() > 0) { 720 CallGenerator* cg = _late_inlines.pop(); 721 cg->do_late_inline(); 722 } 723 } 724 assert(_late_inlines.length() == 0, "should have been processed"); 725 726 print_method("Before RemoveUseless", 3); 727 728 // Remove clutter produced by parsing. 729 if (!failing()) { 730 ResourceMark rm; 731 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 732 } 733 } 734 735 // Note: Large methods are capped off in do_one_bytecode(). 736 if (failing()) return; 737 738 // After parsing, node notes are no longer automagic. 739 // They must be propagated by register_new_node_with_optimizer(), 740 // clone(), or the like. 741 set_default_node_notes(NULL); 742 743 for (;;) { 744 int successes = Inline_Warm(); 745 if (failing()) return; 746 if (successes == 0) break; 747 } 748 749 // Drain the list. 750 Finish_Warm(); 751 #ifndef PRODUCT 752 if (_printer) { 753 _printer->print_inlining(this); 754 } 755 #endif 756 757 if (failing()) return; 758 NOT_PRODUCT( verify_graph_edges(); ) 759 760 // Now optimize 761 Optimize(); 762 if (failing()) return; 763 NOT_PRODUCT( verify_graph_edges(); ) 764 765 #ifndef PRODUCT 766 if (PrintIdeal) { 767 ttyLocker ttyl; // keep the following output all in one block 768 // This output goes directly to the tty, not the compiler log. 769 // To enable tools to match it up with the compilation activity, 770 // be sure to tag this tty output with the compile ID. 771 if (xtty != NULL) { 772 xtty->head("ideal compile_id='%d'%s", compile_id(), 773 is_osr_compilation() ? " compile_kind='osr'" : 774 ""); 775 } 776 root()->dump(9999); 777 if (xtty != NULL) { 778 xtty->tail("ideal"); 779 } 780 } 781 #endif 782 783 // Now that we know the size of all the monitors we can add a fixed slot 784 // for the original deopt pc. 785 786 _orig_pc_slot = fixed_slots(); 787 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 788 set_fixed_slots(next_slot); 789 790 // Now generate code 791 Code_Gen(); 792 if (failing()) return; 793 794 // Check if we want to skip execution of all compiled code. 795 { 796 #ifndef PRODUCT 797 if (OptoNoExecute) { 798 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 799 return; 800 } 801 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 802 #endif 803 804 if (is_osr_compilation()) { 805 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 806 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 807 } else { 808 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 809 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 810 } 811 812 env()->register_method(_method, _entry_bci, 813 &_code_offsets, 814 _orig_pc_slot_offset_in_bytes, 815 code_buffer(), 816 frame_size_in_words(), _oop_map_set, 817 &_handler_table, &_inc_table, 818 compiler, 819 env()->comp_level(), 820 has_unsafe_access() 821 ); 822 } 823 } 824 825 //------------------------------Compile---------------------------------------- 826 // Compile a runtime stub 827 Compile::Compile( ciEnv* ci_env, 828 TypeFunc_generator generator, 829 address stub_function, 830 const char *stub_name, 831 int is_fancy_jump, 832 bool pass_tls, 833 bool save_arg_registers, 834 bool return_pc ) 835 : Phase(Compiler), 836 _env(ci_env), 837 _log(ci_env->log()), 838 _compile_id(-1), 839 _save_argument_registers(save_arg_registers), 840 _method(NULL), 841 _stub_name(stub_name), 842 _stub_function(stub_function), 843 _stub_entry_point(NULL), 844 _entry_bci(InvocationEntryBci), 845 _initial_gvn(NULL), 846 _for_igvn(NULL), 847 _warm_calls(NULL), 848 _orig_pc_slot(0), 849 _orig_pc_slot_offset_in_bytes(0), 850 _subsume_loads(true), 851 _do_escape_analysis(false), 852 _failure_reason(NULL), 853 _code_buffer("Compile::Fill_buffer"), 854 _has_method_handle_invokes(false), 855 _mach_constant_base_node(NULL), 856 _node_bundling_limit(0), 857 _node_bundling_base(NULL), 858 _java_calls(0), 859 _inner_loops(0), 860 #ifndef PRODUCT 861 _trace_opto_output(TraceOptoOutput), 862 _printer(NULL), 863 #endif 864 _congraph(NULL) { 865 C = this; 866 867 #ifndef PRODUCT 868 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 869 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 870 set_print_assembly(PrintFrameConverterAssembly); 871 set_parsed_irreducible_loop(false); 872 #endif 873 CompileWrapper cw(this); 874 Init(/*AliasLevel=*/ 0); 875 init_tf((*generator)()); 876 877 { 878 // The following is a dummy for the sake of GraphKit::gen_stub 879 Unique_Node_List for_igvn(comp_arena()); 880 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 881 PhaseGVN gvn(Thread::current()->resource_area(),255); 882 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 883 gvn.transform_no_reclaim(top()); 884 885 GraphKit kit; 886 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 887 } 888 889 NOT_PRODUCT( verify_graph_edges(); ) 890 Code_Gen(); 891 if (failing()) return; 892 893 894 // Entry point will be accessed using compile->stub_entry_point(); 895 if (code_buffer() == NULL) { 896 Matcher::soft_match_failure(); 897 } else { 898 if (PrintAssembly && (WizardMode || Verbose)) 899 tty->print_cr("### Stub::%s", stub_name); 900 901 if (!failing()) { 902 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 903 904 // Make the NMethod 905 // For now we mark the frame as never safe for profile stackwalking 906 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 907 code_buffer(), 908 CodeOffsets::frame_never_safe, 909 // _code_offsets.value(CodeOffsets::Frame_Complete), 910 frame_size_in_words(), 911 _oop_map_set, 912 save_arg_registers); 913 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 914 915 _stub_entry_point = rs->entry_point(); 916 } 917 } 918 } 919 920 #ifndef PRODUCT 921 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 922 if(PrintOpto && Verbose) { 923 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 924 } 925 } 926 #endif 927 928 void Compile::print_codes() { 929 } 930 931 //------------------------------Init------------------------------------------- 932 // Prepare for a single compilation 933 void Compile::Init(int aliaslevel) { 934 _unique = 0; 935 _regalloc = NULL; 936 937 _tf = NULL; // filled in later 938 _top = NULL; // cached later 939 _matcher = NULL; // filled in later 940 _cfg = NULL; // filled in later 941 942 set_24_bit_selection_and_mode(Use24BitFP, false); 943 944 _node_note_array = NULL; 945 _default_node_notes = NULL; 946 947 _immutable_memory = NULL; // filled in at first inquiry 948 949 // Globally visible Nodes 950 // First set TOP to NULL to give safe behavior during creation of RootNode 951 set_cached_top_node(NULL); 952 set_root(new (this, 3) RootNode()); 953 // Now that you have a Root to point to, create the real TOP 954 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 955 set_recent_alloc(NULL, NULL); 956 957 // Create Debug Information Recorder to record scopes, oopmaps, etc. 958 env()->set_oop_recorder(new OopRecorder(comp_arena())); 959 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 960 env()->set_dependencies(new Dependencies(env())); 961 962 _fixed_slots = 0; 963 set_has_split_ifs(false); 964 set_has_loops(has_method() && method()->has_loops()); // first approximation 965 set_has_stringbuilder(false); 966 _trap_can_recompile = false; // no traps emitted yet 967 _major_progress = true; // start out assuming good things will happen 968 set_has_unsafe_access(false); 969 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 970 set_decompile_count(0); 971 972 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 973 set_num_loop_opts(LoopOptsCount); 974 set_do_inlining(Inline); 975 set_max_inline_size(MaxInlineSize); 976 set_freq_inline_size(FreqInlineSize); 977 set_do_scheduling(OptoScheduling); 978 set_do_count_invocations(false); 979 set_do_method_data_update(false); 980 981 if (debug_info()->recording_non_safepoints()) { 982 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 983 (comp_arena(), 8, 0, NULL)); 984 set_default_node_notes(Node_Notes::make(this)); 985 } 986 987 // // -- Initialize types before each compile -- 988 // // Update cached type information 989 // if( _method && _method->constants() ) 990 // Type::update_loaded_types(_method, _method->constants()); 991 992 // Init alias_type map. 993 if (!_do_escape_analysis && aliaslevel == 3) 994 aliaslevel = 2; // No unique types without escape analysis 995 _AliasLevel = aliaslevel; 996 const int grow_ats = 16; 997 _max_alias_types = grow_ats; 998 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 999 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1000 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1001 { 1002 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1003 } 1004 // Initialize the first few types. 1005 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1006 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1007 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1008 _num_alias_types = AliasIdxRaw+1; 1009 // Zero out the alias type cache. 1010 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1011 // A NULL adr_type hits in the cache right away. Preload the right answer. 1012 probe_alias_cache(NULL)->_index = AliasIdxTop; 1013 1014 _intrinsics = NULL; 1015 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1016 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1017 register_library_intrinsics(); 1018 } 1019 1020 //---------------------------init_start---------------------------------------- 1021 // Install the StartNode on this compile object. 1022 void Compile::init_start(StartNode* s) { 1023 if (failing()) 1024 return; // already failing 1025 assert(s == start(), ""); 1026 } 1027 1028 StartNode* Compile::start() const { 1029 assert(!failing(), ""); 1030 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1031 Node* start = root()->fast_out(i); 1032 if( start->is_Start() ) 1033 return start->as_Start(); 1034 } 1035 ShouldNotReachHere(); 1036 return NULL; 1037 } 1038 1039 //-------------------------------immutable_memory------------------------------------- 1040 // Access immutable memory 1041 Node* Compile::immutable_memory() { 1042 if (_immutable_memory != NULL) { 1043 return _immutable_memory; 1044 } 1045 StartNode* s = start(); 1046 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1047 Node *p = s->fast_out(i); 1048 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1049 _immutable_memory = p; 1050 return _immutable_memory; 1051 } 1052 } 1053 ShouldNotReachHere(); 1054 return NULL; 1055 } 1056 1057 //----------------------set_cached_top_node------------------------------------ 1058 // Install the cached top node, and make sure Node::is_top works correctly. 1059 void Compile::set_cached_top_node(Node* tn) { 1060 if (tn != NULL) verify_top(tn); 1061 Node* old_top = _top; 1062 _top = tn; 1063 // Calling Node::setup_is_top allows the nodes the chance to adjust 1064 // their _out arrays. 1065 if (_top != NULL) _top->setup_is_top(); 1066 if (old_top != NULL) old_top->setup_is_top(); 1067 assert(_top == NULL || top()->is_top(), ""); 1068 } 1069 1070 #ifndef PRODUCT 1071 void Compile::verify_top(Node* tn) const { 1072 if (tn != NULL) { 1073 assert(tn->is_Con(), "top node must be a constant"); 1074 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1075 assert(tn->in(0) != NULL, "must have live top node"); 1076 } 1077 } 1078 #endif 1079 1080 1081 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1082 1083 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1084 guarantee(arr != NULL, ""); 1085 int num_blocks = arr->length(); 1086 if (grow_by < num_blocks) grow_by = num_blocks; 1087 int num_notes = grow_by * _node_notes_block_size; 1088 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1089 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1090 while (num_notes > 0) { 1091 arr->append(notes); 1092 notes += _node_notes_block_size; 1093 num_notes -= _node_notes_block_size; 1094 } 1095 assert(num_notes == 0, "exact multiple, please"); 1096 } 1097 1098 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1099 if (source == NULL || dest == NULL) return false; 1100 1101 if (dest->is_Con()) 1102 return false; // Do not push debug info onto constants. 1103 1104 #ifdef ASSERT 1105 // Leave a bread crumb trail pointing to the original node: 1106 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1107 dest->set_debug_orig(source); 1108 } 1109 #endif 1110 1111 if (node_note_array() == NULL) 1112 return false; // Not collecting any notes now. 1113 1114 // This is a copy onto a pre-existing node, which may already have notes. 1115 // If both nodes have notes, do not overwrite any pre-existing notes. 1116 Node_Notes* source_notes = node_notes_at(source->_idx); 1117 if (source_notes == NULL || source_notes->is_clear()) return false; 1118 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1119 if (dest_notes == NULL || dest_notes->is_clear()) { 1120 return set_node_notes_at(dest->_idx, source_notes); 1121 } 1122 1123 Node_Notes merged_notes = (*source_notes); 1124 // The order of operations here ensures that dest notes will win... 1125 merged_notes.update_from(dest_notes); 1126 return set_node_notes_at(dest->_idx, &merged_notes); 1127 } 1128 1129 1130 //--------------------------allow_range_check_smearing------------------------- 1131 // Gating condition for coalescing similar range checks. 1132 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1133 // single covering check that is at least as strong as any of them. 1134 // If the optimization succeeds, the simplified (strengthened) range check 1135 // will always succeed. If it fails, we will deopt, and then give up 1136 // on the optimization. 1137 bool Compile::allow_range_check_smearing() const { 1138 // If this method has already thrown a range-check, 1139 // assume it was because we already tried range smearing 1140 // and it failed. 1141 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1142 return !already_trapped; 1143 } 1144 1145 1146 //------------------------------flatten_alias_type----------------------------- 1147 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1148 int offset = tj->offset(); 1149 TypePtr::PTR ptr = tj->ptr(); 1150 1151 // Known instance (scalarizable allocation) alias only with itself. 1152 bool is_known_inst = tj->isa_oopptr() != NULL && 1153 tj->is_oopptr()->is_known_instance(); 1154 1155 // Process weird unsafe references. 1156 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1157 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1158 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1159 tj = TypeOopPtr::BOTTOM; 1160 ptr = tj->ptr(); 1161 offset = tj->offset(); 1162 } 1163 1164 // Array pointers need some flattening 1165 const TypeAryPtr *ta = tj->isa_aryptr(); 1166 if( ta && is_known_inst ) { 1167 if ( offset != Type::OffsetBot && 1168 offset > arrayOopDesc::length_offset_in_bytes() ) { 1169 offset = Type::OffsetBot; // Flatten constant access into array body only 1170 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1171 } 1172 } else if( ta && _AliasLevel >= 2 ) { 1173 // For arrays indexed by constant indices, we flatten the alias 1174 // space to include all of the array body. Only the header, klass 1175 // and array length can be accessed un-aliased. 1176 if( offset != Type::OffsetBot ) { 1177 if( ta->const_oop() ) { // methodDataOop or methodOop 1178 offset = Type::OffsetBot; // Flatten constant access into array body 1179 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1180 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1181 // range is OK as-is. 1182 tj = ta = TypeAryPtr::RANGE; 1183 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1184 tj = TypeInstPtr::KLASS; // all klass loads look alike 1185 ta = TypeAryPtr::RANGE; // generic ignored junk 1186 ptr = TypePtr::BotPTR; 1187 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1188 tj = TypeInstPtr::MARK; 1189 ta = TypeAryPtr::RANGE; // generic ignored junk 1190 ptr = TypePtr::BotPTR; 1191 } else { // Random constant offset into array body 1192 offset = Type::OffsetBot; // Flatten constant access into array body 1193 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1194 } 1195 } 1196 // Arrays of fixed size alias with arrays of unknown size. 1197 if (ta->size() != TypeInt::POS) { 1198 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1199 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1200 } 1201 // Arrays of known objects become arrays of unknown objects. 1202 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1203 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1204 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1205 } 1206 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1207 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1208 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1209 } 1210 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1211 // cannot be distinguished by bytecode alone. 1212 if (ta->elem() == TypeInt::BOOL) { 1213 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1214 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1215 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1216 } 1217 // During the 2nd round of IterGVN, NotNull castings are removed. 1218 // Make sure the Bottom and NotNull variants alias the same. 1219 // Also, make sure exact and non-exact variants alias the same. 1220 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1221 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1222 } 1223 } 1224 1225 // Oop pointers need some flattening 1226 const TypeInstPtr *to = tj->isa_instptr(); 1227 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1228 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1229 if( ptr == TypePtr::Constant ) { 1230 if (to->klass() != ciEnv::current()->Class_klass() || 1231 offset < k->size_helper() * wordSize) { 1232 // No constant oop pointers (such as Strings); they alias with 1233 // unknown strings. 1234 assert(!is_known_inst, "not scalarizable allocation"); 1235 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1236 } 1237 } else if( is_known_inst ) { 1238 tj = to; // Keep NotNull and klass_is_exact for instance type 1239 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1240 // During the 2nd round of IterGVN, NotNull castings are removed. 1241 // Make sure the Bottom and NotNull variants alias the same. 1242 // Also, make sure exact and non-exact variants alias the same. 1243 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1244 } 1245 // Canonicalize the holder of this field 1246 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1247 // First handle header references such as a LoadKlassNode, even if the 1248 // object's klass is unloaded at compile time (4965979). 1249 if (!is_known_inst) { // Do it only for non-instance types 1250 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1251 } 1252 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1253 // Static fields are in the space above the normal instance 1254 // fields in the java.lang.Class instance. 1255 if (to->klass() != ciEnv::current()->Class_klass()) { 1256 to = NULL; 1257 tj = TypeOopPtr::BOTTOM; 1258 offset = tj->offset(); 1259 } 1260 } else { 1261 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1262 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1263 if( is_known_inst ) { 1264 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1265 } else { 1266 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1267 } 1268 } 1269 } 1270 } 1271 1272 // Klass pointers to object array klasses need some flattening 1273 const TypeKlassPtr *tk = tj->isa_klassptr(); 1274 if( tk ) { 1275 // If we are referencing a field within a Klass, we need 1276 // to assume the worst case of an Object. Both exact and 1277 // inexact types must flatten to the same alias class. 1278 // Since the flattened result for a klass is defined to be 1279 // precisely java.lang.Object, use a constant ptr. 1280 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1281 1282 tj = tk = TypeKlassPtr::make(TypePtr::Constant, 1283 TypeKlassPtr::OBJECT->klass(), 1284 offset); 1285 } 1286 1287 ciKlass* klass = tk->klass(); 1288 if( klass->is_obj_array_klass() ) { 1289 ciKlass* k = TypeAryPtr::OOPS->klass(); 1290 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1291 k = TypeInstPtr::BOTTOM->klass(); 1292 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1293 } 1294 1295 // Check for precise loads from the primary supertype array and force them 1296 // to the supertype cache alias index. Check for generic array loads from 1297 // the primary supertype array and also force them to the supertype cache 1298 // alias index. Since the same load can reach both, we need to merge 1299 // these 2 disparate memories into the same alias class. Since the 1300 // primary supertype array is read-only, there's no chance of confusion 1301 // where we bypass an array load and an array store. 1302 uint off2 = offset - Klass::primary_supers_offset_in_bytes(); 1303 if( offset == Type::OffsetBot || 1304 off2 < Klass::primary_super_limit()*wordSize ) { 1305 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes(); 1306 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1307 } 1308 } 1309 1310 // Flatten all Raw pointers together. 1311 if (tj->base() == Type::RawPtr) 1312 tj = TypeRawPtr::BOTTOM; 1313 1314 if (tj->base() == Type::AnyPtr) 1315 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1316 1317 // Flatten all to bottom for now 1318 switch( _AliasLevel ) { 1319 case 0: 1320 tj = TypePtr::BOTTOM; 1321 break; 1322 case 1: // Flatten to: oop, static, field or array 1323 switch (tj->base()) { 1324 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1325 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1326 case Type::AryPtr: // do not distinguish arrays at all 1327 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1328 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1329 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1330 default: ShouldNotReachHere(); 1331 } 1332 break; 1333 case 2: // No collapsing at level 2; keep all splits 1334 case 3: // No collapsing at level 3; keep all splits 1335 break; 1336 default: 1337 Unimplemented(); 1338 } 1339 1340 offset = tj->offset(); 1341 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1342 1343 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1344 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1345 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1346 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1347 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1348 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1349 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1350 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1351 assert( tj->ptr() != TypePtr::TopPTR && 1352 tj->ptr() != TypePtr::AnyNull && 1353 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1354 // assert( tj->ptr() != TypePtr::Constant || 1355 // tj->base() == Type::RawPtr || 1356 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1357 1358 return tj; 1359 } 1360 1361 void Compile::AliasType::Init(int i, const TypePtr* at) { 1362 _index = i; 1363 _adr_type = at; 1364 _field = NULL; 1365 _is_rewritable = true; // default 1366 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1367 if (atoop != NULL && atoop->is_known_instance()) { 1368 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1369 _general_index = Compile::current()->get_alias_index(gt); 1370 } else { 1371 _general_index = 0; 1372 } 1373 } 1374 1375 //---------------------------------print_on------------------------------------ 1376 #ifndef PRODUCT 1377 void Compile::AliasType::print_on(outputStream* st) { 1378 if (index() < 10) 1379 st->print("@ <%d> ", index()); 1380 else st->print("@ <%d>", index()); 1381 st->print(is_rewritable() ? " " : " RO"); 1382 int offset = adr_type()->offset(); 1383 if (offset == Type::OffsetBot) 1384 st->print(" +any"); 1385 else st->print(" +%-3d", offset); 1386 st->print(" in "); 1387 adr_type()->dump_on(st); 1388 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1389 if (field() != NULL && tjp) { 1390 if (tjp->klass() != field()->holder() || 1391 tjp->offset() != field()->offset_in_bytes()) { 1392 st->print(" != "); 1393 field()->print(); 1394 st->print(" ***"); 1395 } 1396 } 1397 } 1398 1399 void print_alias_types() { 1400 Compile* C = Compile::current(); 1401 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1402 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1403 C->alias_type(idx)->print_on(tty); 1404 tty->cr(); 1405 } 1406 } 1407 #endif 1408 1409 1410 //----------------------------probe_alias_cache-------------------------------- 1411 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1412 intptr_t key = (intptr_t) adr_type; 1413 key ^= key >> logAliasCacheSize; 1414 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1415 } 1416 1417 1418 //-----------------------------grow_alias_types-------------------------------- 1419 void Compile::grow_alias_types() { 1420 const int old_ats = _max_alias_types; // how many before? 1421 const int new_ats = old_ats; // how many more? 1422 const int grow_ats = old_ats+new_ats; // how many now? 1423 _max_alias_types = grow_ats; 1424 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1425 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1426 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1427 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1428 } 1429 1430 1431 //--------------------------------find_alias_type------------------------------ 1432 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1433 if (_AliasLevel == 0) 1434 return alias_type(AliasIdxBot); 1435 1436 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1437 if (ace->_adr_type == adr_type) { 1438 return alias_type(ace->_index); 1439 } 1440 1441 // Handle special cases. 1442 if (adr_type == NULL) return alias_type(AliasIdxTop); 1443 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1444 1445 // Do it the slow way. 1446 const TypePtr* flat = flatten_alias_type(adr_type); 1447 1448 #ifdef ASSERT 1449 assert(flat == flatten_alias_type(flat), "idempotent"); 1450 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1451 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1452 const TypeOopPtr* foop = flat->is_oopptr(); 1453 // Scalarizable allocations have exact klass always. 1454 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1455 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1456 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1457 } 1458 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1459 #endif 1460 1461 int idx = AliasIdxTop; 1462 for (int i = 0; i < num_alias_types(); i++) { 1463 if (alias_type(i)->adr_type() == flat) { 1464 idx = i; 1465 break; 1466 } 1467 } 1468 1469 if (idx == AliasIdxTop) { 1470 if (no_create) return NULL; 1471 // Grow the array if necessary. 1472 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1473 // Add a new alias type. 1474 idx = _num_alias_types++; 1475 _alias_types[idx]->Init(idx, flat); 1476 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1477 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1478 if (flat->isa_instptr()) { 1479 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1480 && flat->is_instptr()->klass() == env()->Class_klass()) 1481 alias_type(idx)->set_rewritable(false); 1482 } 1483 if (flat->isa_klassptr()) { 1484 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) 1485 alias_type(idx)->set_rewritable(false); 1486 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1487 alias_type(idx)->set_rewritable(false); 1488 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) 1489 alias_type(idx)->set_rewritable(false); 1490 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) 1491 alias_type(idx)->set_rewritable(false); 1492 } 1493 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1494 // but the base pointer type is not distinctive enough to identify 1495 // references into JavaThread.) 1496 1497 // Check for final fields. 1498 const TypeInstPtr* tinst = flat->isa_instptr(); 1499 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1500 ciField* field; 1501 if (tinst->const_oop() != NULL && 1502 tinst->klass() == ciEnv::current()->Class_klass() && 1503 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1504 // static field 1505 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1506 field = k->get_field_by_offset(tinst->offset(), true); 1507 } else { 1508 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1509 field = k->get_field_by_offset(tinst->offset(), false); 1510 } 1511 assert(field == NULL || 1512 original_field == NULL || 1513 (field->holder() == original_field->holder() && 1514 field->offset() == original_field->offset() && 1515 field->is_static() == original_field->is_static()), "wrong field?"); 1516 // Set field() and is_rewritable() attributes. 1517 if (field != NULL) alias_type(idx)->set_field(field); 1518 } 1519 } 1520 1521 // Fill the cache for next time. 1522 ace->_adr_type = adr_type; 1523 ace->_index = idx; 1524 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1525 1526 // Might as well try to fill the cache for the flattened version, too. 1527 AliasCacheEntry* face = probe_alias_cache(flat); 1528 if (face->_adr_type == NULL) { 1529 face->_adr_type = flat; 1530 face->_index = idx; 1531 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1532 } 1533 1534 return alias_type(idx); 1535 } 1536 1537 1538 Compile::AliasType* Compile::alias_type(ciField* field) { 1539 const TypeOopPtr* t; 1540 if (field->is_static()) 1541 t = TypeInstPtr::make(field->holder()->java_mirror()); 1542 else 1543 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1544 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1545 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1546 return atp; 1547 } 1548 1549 1550 //------------------------------have_alias_type-------------------------------- 1551 bool Compile::have_alias_type(const TypePtr* adr_type) { 1552 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1553 if (ace->_adr_type == adr_type) { 1554 return true; 1555 } 1556 1557 // Handle special cases. 1558 if (adr_type == NULL) return true; 1559 if (adr_type == TypePtr::BOTTOM) return true; 1560 1561 return find_alias_type(adr_type, true, NULL) != NULL; 1562 } 1563 1564 //-----------------------------must_alias-------------------------------------- 1565 // True if all values of the given address type are in the given alias category. 1566 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1567 if (alias_idx == AliasIdxBot) return true; // the universal category 1568 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1569 if (alias_idx == AliasIdxTop) return false; // the empty category 1570 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1571 1572 // the only remaining possible overlap is identity 1573 int adr_idx = get_alias_index(adr_type); 1574 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1575 assert(adr_idx == alias_idx || 1576 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1577 && adr_type != TypeOopPtr::BOTTOM), 1578 "should not be testing for overlap with an unsafe pointer"); 1579 return adr_idx == alias_idx; 1580 } 1581 1582 //------------------------------can_alias-------------------------------------- 1583 // True if any values of the given address type are in the given alias category. 1584 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1585 if (alias_idx == AliasIdxTop) return false; // the empty category 1586 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1587 if (alias_idx == AliasIdxBot) return true; // the universal category 1588 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1589 1590 // the only remaining possible overlap is identity 1591 int adr_idx = get_alias_index(adr_type); 1592 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1593 return adr_idx == alias_idx; 1594 } 1595 1596 1597 1598 //---------------------------pop_warm_call------------------------------------- 1599 WarmCallInfo* Compile::pop_warm_call() { 1600 WarmCallInfo* wci = _warm_calls; 1601 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1602 return wci; 1603 } 1604 1605 //----------------------------Inline_Warm-------------------------------------- 1606 int Compile::Inline_Warm() { 1607 // If there is room, try to inline some more warm call sites. 1608 // %%% Do a graph index compaction pass when we think we're out of space? 1609 if (!InlineWarmCalls) return 0; 1610 1611 int calls_made_hot = 0; 1612 int room_to_grow = NodeCountInliningCutoff - unique(); 1613 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1614 int amount_grown = 0; 1615 WarmCallInfo* call; 1616 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1617 int est_size = (int)call->size(); 1618 if (est_size > (room_to_grow - amount_grown)) { 1619 // This one won't fit anyway. Get rid of it. 1620 call->make_cold(); 1621 continue; 1622 } 1623 call->make_hot(); 1624 calls_made_hot++; 1625 amount_grown += est_size; 1626 amount_to_grow -= est_size; 1627 } 1628 1629 if (calls_made_hot > 0) set_major_progress(); 1630 return calls_made_hot; 1631 } 1632 1633 1634 //----------------------------Finish_Warm-------------------------------------- 1635 void Compile::Finish_Warm() { 1636 if (!InlineWarmCalls) return; 1637 if (failing()) return; 1638 if (warm_calls() == NULL) return; 1639 1640 // Clean up loose ends, if we are out of space for inlining. 1641 WarmCallInfo* call; 1642 while ((call = pop_warm_call()) != NULL) { 1643 call->make_cold(); 1644 } 1645 } 1646 1647 //---------------------cleanup_loop_predicates----------------------- 1648 // Remove the opaque nodes that protect the predicates so that all unused 1649 // checks and uncommon_traps will be eliminated from the ideal graph 1650 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1651 if (predicate_count()==0) return; 1652 for (int i = predicate_count(); i > 0; i--) { 1653 Node * n = predicate_opaque1_node(i-1); 1654 assert(n->Opcode() == Op_Opaque1, "must be"); 1655 igvn.replace_node(n, n->in(1)); 1656 } 1657 assert(predicate_count()==0, "should be clean!"); 1658 } 1659 1660 //------------------------------Optimize--------------------------------------- 1661 // Given a graph, optimize it. 1662 void Compile::Optimize() { 1663 TracePhase t1("optimizer", &_t_optimizer, true); 1664 1665 #ifndef PRODUCT 1666 if (env()->break_at_compile()) { 1667 BREAKPOINT; 1668 } 1669 1670 #endif 1671 1672 ResourceMark rm; 1673 int loop_opts_cnt; 1674 1675 NOT_PRODUCT( verify_graph_edges(); ) 1676 1677 print_method("After Parsing"); 1678 1679 { 1680 // Iterative Global Value Numbering, including ideal transforms 1681 // Initialize IterGVN with types and values from parse-time GVN 1682 PhaseIterGVN igvn(initial_gvn()); 1683 { 1684 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1685 igvn.optimize(); 1686 } 1687 1688 print_method("Iter GVN 1", 2); 1689 1690 if (failing()) return; 1691 1692 // Perform escape analysis 1693 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1694 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true); 1695 ConnectionGraph::do_analysis(this, &igvn); 1696 1697 if (failing()) return; 1698 1699 igvn.optimize(); 1700 print_method("Iter GVN 3", 2); 1701 1702 if (failing()) return; 1703 1704 } 1705 1706 // Loop transforms on the ideal graph. Range Check Elimination, 1707 // peeling, unrolling, etc. 1708 1709 // Set loop opts counter 1710 loop_opts_cnt = num_loop_opts(); 1711 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1712 { 1713 TracePhase t2("idealLoop", &_t_idealLoop, true); 1714 PhaseIdealLoop ideal_loop( igvn, true ); 1715 loop_opts_cnt--; 1716 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1717 if (failing()) return; 1718 } 1719 // Loop opts pass if partial peeling occurred in previous pass 1720 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1721 TracePhase t3("idealLoop", &_t_idealLoop, true); 1722 PhaseIdealLoop ideal_loop( igvn, false ); 1723 loop_opts_cnt--; 1724 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1725 if (failing()) return; 1726 } 1727 // Loop opts pass for loop-unrolling before CCP 1728 if(major_progress() && (loop_opts_cnt > 0)) { 1729 TracePhase t4("idealLoop", &_t_idealLoop, true); 1730 PhaseIdealLoop ideal_loop( igvn, false ); 1731 loop_opts_cnt--; 1732 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1733 } 1734 if (!failing()) { 1735 // Verify that last round of loop opts produced a valid graph 1736 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1737 PhaseIdealLoop::verify(igvn); 1738 } 1739 } 1740 if (failing()) return; 1741 1742 // Conditional Constant Propagation; 1743 PhaseCCP ccp( &igvn ); 1744 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1745 { 1746 TracePhase t2("ccp", &_t_ccp, true); 1747 ccp.do_transform(); 1748 } 1749 print_method("PhaseCPP 1", 2); 1750 1751 assert( true, "Break here to ccp.dump_old2new_map()"); 1752 1753 // Iterative Global Value Numbering, including ideal transforms 1754 { 1755 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1756 igvn = ccp; 1757 igvn.optimize(); 1758 } 1759 1760 print_method("Iter GVN 2", 2); 1761 1762 if (failing()) return; 1763 1764 // Loop transforms on the ideal graph. Range Check Elimination, 1765 // peeling, unrolling, etc. 1766 if(loop_opts_cnt > 0) { 1767 debug_only( int cnt = 0; ); 1768 while(major_progress() && (loop_opts_cnt > 0)) { 1769 TracePhase t2("idealLoop", &_t_idealLoop, true); 1770 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1771 PhaseIdealLoop ideal_loop( igvn, true); 1772 loop_opts_cnt--; 1773 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1774 if (failing()) return; 1775 } 1776 } 1777 1778 { 1779 // Verify that all previous optimizations produced a valid graph 1780 // at least to this point, even if no loop optimizations were done. 1781 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1782 PhaseIdealLoop::verify(igvn); 1783 } 1784 1785 { 1786 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1787 PhaseMacroExpand mex(igvn); 1788 if (mex.expand_macro_nodes()) { 1789 assert(failing(), "must bail out w/ explicit message"); 1790 return; 1791 } 1792 } 1793 1794 } // (End scope of igvn; run destructor if necessary for asserts.) 1795 1796 // A method with only infinite loops has no edges entering loops from root 1797 { 1798 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1799 if (final_graph_reshaping()) { 1800 assert(failing(), "must bail out w/ explicit message"); 1801 return; 1802 } 1803 } 1804 1805 print_method("Optimize finished", 2); 1806 } 1807 1808 1809 //------------------------------Code_Gen--------------------------------------- 1810 // Given a graph, generate code for it 1811 void Compile::Code_Gen() { 1812 if (failing()) return; 1813 1814 // Perform instruction selection. You might think we could reclaim Matcher 1815 // memory PDQ, but actually the Matcher is used in generating spill code. 1816 // Internals of the Matcher (including some VectorSets) must remain live 1817 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1818 // set a bit in reclaimed memory. 1819 1820 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1821 // nodes. Mapping is only valid at the root of each matched subtree. 1822 NOT_PRODUCT( verify_graph_edges(); ) 1823 1824 Node_List proj_list; 1825 Matcher m(proj_list); 1826 _matcher = &m; 1827 { 1828 TracePhase t2("matcher", &_t_matcher, true); 1829 m.match(); 1830 } 1831 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1832 // nodes. Mapping is only valid at the root of each matched subtree. 1833 NOT_PRODUCT( verify_graph_edges(); ) 1834 1835 // If you have too many nodes, or if matching has failed, bail out 1836 check_node_count(0, "out of nodes matching instructions"); 1837 if (failing()) return; 1838 1839 // Build a proper-looking CFG 1840 PhaseCFG cfg(node_arena(), root(), m); 1841 _cfg = &cfg; 1842 { 1843 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1844 cfg.Dominators(); 1845 if (failing()) return; 1846 1847 NOT_PRODUCT( verify_graph_edges(); ) 1848 1849 cfg.Estimate_Block_Frequency(); 1850 cfg.GlobalCodeMotion(m,unique(),proj_list); 1851 1852 print_method("Global code motion", 2); 1853 1854 if (failing()) return; 1855 NOT_PRODUCT( verify_graph_edges(); ) 1856 1857 debug_only( cfg.verify(); ) 1858 } 1859 NOT_PRODUCT( verify_graph_edges(); ) 1860 1861 PhaseChaitin regalloc(unique(),cfg,m); 1862 _regalloc = ®alloc; 1863 { 1864 TracePhase t2("regalloc", &_t_registerAllocation, true); 1865 // Perform any platform dependent preallocation actions. This is used, 1866 // for example, to avoid taking an implicit null pointer exception 1867 // using the frame pointer on win95. 1868 _regalloc->pd_preallocate_hook(); 1869 1870 // Perform register allocation. After Chaitin, use-def chains are 1871 // no longer accurate (at spill code) and so must be ignored. 1872 // Node->LRG->reg mappings are still accurate. 1873 _regalloc->Register_Allocate(); 1874 1875 // Bail out if the allocator builds too many nodes 1876 if (failing()) return; 1877 } 1878 1879 // Prior to register allocation we kept empty basic blocks in case the 1880 // the allocator needed a place to spill. After register allocation we 1881 // are not adding any new instructions. If any basic block is empty, we 1882 // can now safely remove it. 1883 { 1884 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1885 cfg.remove_empty(); 1886 if (do_freq_based_layout()) { 1887 PhaseBlockLayout layout(cfg); 1888 } else { 1889 cfg.set_loop_alignment(); 1890 } 1891 cfg.fixup_flow(); 1892 } 1893 1894 // Perform any platform dependent postallocation verifications. 1895 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1896 1897 // Apply peephole optimizations 1898 if( OptoPeephole ) { 1899 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1900 PhasePeephole peep( _regalloc, cfg); 1901 peep.do_transform(); 1902 } 1903 1904 // Convert Nodes to instruction bits in a buffer 1905 { 1906 // %%%% workspace merge brought two timers together for one job 1907 TracePhase t2a("output", &_t_output, true); 1908 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1909 Output(); 1910 } 1911 1912 print_method("Final Code"); 1913 1914 // He's dead, Jim. 1915 _cfg = (PhaseCFG*)0xdeadbeef; 1916 _regalloc = (PhaseChaitin*)0xdeadbeef; 1917 } 1918 1919 1920 //------------------------------dump_asm--------------------------------------- 1921 // Dump formatted assembly 1922 #ifndef PRODUCT 1923 void Compile::dump_asm(int *pcs, uint pc_limit) { 1924 bool cut_short = false; 1925 tty->print_cr("#"); 1926 tty->print("# "); _tf->dump(); tty->cr(); 1927 tty->print_cr("#"); 1928 1929 // For all blocks 1930 int pc = 0x0; // Program counter 1931 char starts_bundle = ' '; 1932 _regalloc->dump_frame(); 1933 1934 Node *n = NULL; 1935 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1936 if (VMThread::should_terminate()) { cut_short = true; break; } 1937 Block *b = _cfg->_blocks[i]; 1938 if (b->is_connector() && !Verbose) continue; 1939 n = b->_nodes[0]; 1940 if (pcs && n->_idx < pc_limit) 1941 tty->print("%3.3x ", pcs[n->_idx]); 1942 else 1943 tty->print(" "); 1944 b->dump_head( &_cfg->_bbs ); 1945 if (b->is_connector()) { 1946 tty->print_cr(" # Empty connector block"); 1947 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1948 tty->print_cr(" # Block is sole successor of call"); 1949 } 1950 1951 // For all instructions 1952 Node *delay = NULL; 1953 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1954 if (VMThread::should_terminate()) { cut_short = true; break; } 1955 n = b->_nodes[j]; 1956 if (valid_bundle_info(n)) { 1957 Bundle *bundle = node_bundling(n); 1958 if (bundle->used_in_unconditional_delay()) { 1959 delay = n; 1960 continue; 1961 } 1962 if (bundle->starts_bundle()) 1963 starts_bundle = '+'; 1964 } 1965 1966 if (WizardMode) n->dump(); 1967 1968 if( !n->is_Region() && // Dont print in the Assembly 1969 !n->is_Phi() && // a few noisely useless nodes 1970 !n->is_Proj() && 1971 !n->is_MachTemp() && 1972 !n->is_SafePointScalarObject() && 1973 !n->is_Catch() && // Would be nice to print exception table targets 1974 !n->is_MergeMem() && // Not very interesting 1975 !n->is_top() && // Debug info table constants 1976 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1977 ) { 1978 if (pcs && n->_idx < pc_limit) 1979 tty->print("%3.3x", pcs[n->_idx]); 1980 else 1981 tty->print(" "); 1982 tty->print(" %c ", starts_bundle); 1983 starts_bundle = ' '; 1984 tty->print("\t"); 1985 n->format(_regalloc, tty); 1986 tty->cr(); 1987 } 1988 1989 // If we have an instruction with a delay slot, and have seen a delay, 1990 // then back up and print it 1991 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 1992 assert(delay != NULL, "no unconditional delay instruction"); 1993 if (WizardMode) delay->dump(); 1994 1995 if (node_bundling(delay)->starts_bundle()) 1996 starts_bundle = '+'; 1997 if (pcs && n->_idx < pc_limit) 1998 tty->print("%3.3x", pcs[n->_idx]); 1999 else 2000 tty->print(" "); 2001 tty->print(" %c ", starts_bundle); 2002 starts_bundle = ' '; 2003 tty->print("\t"); 2004 delay->format(_regalloc, tty); 2005 tty->print_cr(""); 2006 delay = NULL; 2007 } 2008 2009 // Dump the exception table as well 2010 if( n->is_Catch() && (Verbose || WizardMode) ) { 2011 // Print the exception table for this offset 2012 _handler_table.print_subtable_for(pc); 2013 } 2014 } 2015 2016 if (pcs && n->_idx < pc_limit) 2017 tty->print_cr("%3.3x", pcs[n->_idx]); 2018 else 2019 tty->print_cr(""); 2020 2021 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2022 2023 } // End of per-block dump 2024 tty->print_cr(""); 2025 2026 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2027 } 2028 #endif 2029 2030 //------------------------------Final_Reshape_Counts--------------------------- 2031 // This class defines counters to help identify when a method 2032 // may/must be executed using hardware with only 24-bit precision. 2033 struct Final_Reshape_Counts : public StackObj { 2034 int _call_count; // count non-inlined 'common' calls 2035 int _float_count; // count float ops requiring 24-bit precision 2036 int _double_count; // count double ops requiring more precision 2037 int _java_call_count; // count non-inlined 'java' calls 2038 int _inner_loop_count; // count loops which need alignment 2039 VectorSet _visited; // Visitation flags 2040 Node_List _tests; // Set of IfNodes & PCTableNodes 2041 2042 Final_Reshape_Counts() : 2043 _call_count(0), _float_count(0), _double_count(0), 2044 _java_call_count(0), _inner_loop_count(0), 2045 _visited( Thread::current()->resource_area() ) { } 2046 2047 void inc_call_count () { _call_count ++; } 2048 void inc_float_count () { _float_count ++; } 2049 void inc_double_count() { _double_count++; } 2050 void inc_java_call_count() { _java_call_count++; } 2051 void inc_inner_loop_count() { _inner_loop_count++; } 2052 2053 int get_call_count () const { return _call_count ; } 2054 int get_float_count () const { return _float_count ; } 2055 int get_double_count() const { return _double_count; } 2056 int get_java_call_count() const { return _java_call_count; } 2057 int get_inner_loop_count() const { return _inner_loop_count; } 2058 }; 2059 2060 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2061 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2062 // Make sure the offset goes inside the instance layout. 2063 return k->contains_field_offset(tp->offset()); 2064 // Note that OffsetBot and OffsetTop are very negative. 2065 } 2066 2067 // Eliminate trivially redundant StoreCMs and accumulate their 2068 // precedence edges. 2069 static void eliminate_redundant_card_marks(Node* n) { 2070 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2071 if (n->in(MemNode::Address)->outcnt() > 1) { 2072 // There are multiple users of the same address so it might be 2073 // possible to eliminate some of the StoreCMs 2074 Node* mem = n->in(MemNode::Memory); 2075 Node* adr = n->in(MemNode::Address); 2076 Node* val = n->in(MemNode::ValueIn); 2077 Node* prev = n; 2078 bool done = false; 2079 // Walk the chain of StoreCMs eliminating ones that match. As 2080 // long as it's a chain of single users then the optimization is 2081 // safe. Eliminating partially redundant StoreCMs would require 2082 // cloning copies down the other paths. 2083 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2084 if (adr == mem->in(MemNode::Address) && 2085 val == mem->in(MemNode::ValueIn)) { 2086 // redundant StoreCM 2087 if (mem->req() > MemNode::OopStore) { 2088 // Hasn't been processed by this code yet. 2089 n->add_prec(mem->in(MemNode::OopStore)); 2090 } else { 2091 // Already converted to precedence edge 2092 for (uint i = mem->req(); i < mem->len(); i++) { 2093 // Accumulate any precedence edges 2094 if (mem->in(i) != NULL) { 2095 n->add_prec(mem->in(i)); 2096 } 2097 } 2098 // Everything above this point has been processed. 2099 done = true; 2100 } 2101 // Eliminate the previous StoreCM 2102 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2103 assert(mem->outcnt() == 0, "should be dead"); 2104 mem->disconnect_inputs(NULL); 2105 } else { 2106 prev = mem; 2107 } 2108 mem = prev->in(MemNode::Memory); 2109 } 2110 } 2111 } 2112 2113 //------------------------------final_graph_reshaping_impl---------------------- 2114 // Implement items 1-5 from final_graph_reshaping below. 2115 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 2116 2117 if ( n->outcnt() == 0 ) return; // dead node 2118 uint nop = n->Opcode(); 2119 2120 // Check for 2-input instruction with "last use" on right input. 2121 // Swap to left input. Implements item (2). 2122 if( n->req() == 3 && // two-input instruction 2123 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2124 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2125 n->in(2)->outcnt() == 1 &&// right use IS a last use 2126 !n->in(2)->is_Con() ) { // right use is not a constant 2127 // Check for commutative opcode 2128 switch( nop ) { 2129 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2130 case Op_MaxI: case Op_MinI: 2131 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2132 case Op_AndL: case Op_XorL: case Op_OrL: 2133 case Op_AndI: case Op_XorI: case Op_OrI: { 2134 // Move "last use" input to left by swapping inputs 2135 n->swap_edges(1, 2); 2136 break; 2137 } 2138 default: 2139 break; 2140 } 2141 } 2142 2143 #ifdef ASSERT 2144 if( n->is_Mem() ) { 2145 Compile* C = Compile::current(); 2146 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2147 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2148 // oop will be recorded in oop map if load crosses safepoint 2149 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2150 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2151 "raw memory operations should have control edge"); 2152 } 2153 #endif 2154 // Count FPU ops and common calls, implements item (3) 2155 switch( nop ) { 2156 // Count all float operations that may use FPU 2157 case Op_AddF: 2158 case Op_SubF: 2159 case Op_MulF: 2160 case Op_DivF: 2161 case Op_NegF: 2162 case Op_ModF: 2163 case Op_ConvI2F: 2164 case Op_ConF: 2165 case Op_CmpF: 2166 case Op_CmpF3: 2167 // case Op_ConvL2F: // longs are split into 32-bit halves 2168 frc.inc_float_count(); 2169 break; 2170 2171 case Op_ConvF2D: 2172 case Op_ConvD2F: 2173 frc.inc_float_count(); 2174 frc.inc_double_count(); 2175 break; 2176 2177 // Count all double operations that may use FPU 2178 case Op_AddD: 2179 case Op_SubD: 2180 case Op_MulD: 2181 case Op_DivD: 2182 case Op_NegD: 2183 case Op_ModD: 2184 case Op_ConvI2D: 2185 case Op_ConvD2I: 2186 // case Op_ConvL2D: // handled by leaf call 2187 // case Op_ConvD2L: // handled by leaf call 2188 case Op_ConD: 2189 case Op_CmpD: 2190 case Op_CmpD3: 2191 frc.inc_double_count(); 2192 break; 2193 case Op_Opaque1: // Remove Opaque Nodes before matching 2194 case Op_Opaque2: // Remove Opaque Nodes before matching 2195 n->subsume_by(n->in(1)); 2196 break; 2197 case Op_CallStaticJava: 2198 case Op_CallJava: 2199 case Op_CallDynamicJava: 2200 frc.inc_java_call_count(); // Count java call site; 2201 case Op_CallRuntime: 2202 case Op_CallLeaf: 2203 case Op_CallLeafNoFP: { 2204 assert( n->is_Call(), "" ); 2205 CallNode *call = n->as_Call(); 2206 // Count call sites where the FP mode bit would have to be flipped. 2207 // Do not count uncommon runtime calls: 2208 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2209 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2210 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2211 frc.inc_call_count(); // Count the call site 2212 } else { // See if uncommon argument is shared 2213 Node *n = call->in(TypeFunc::Parms); 2214 int nop = n->Opcode(); 2215 // Clone shared simple arguments to uncommon calls, item (1). 2216 if( n->outcnt() > 1 && 2217 !n->is_Proj() && 2218 nop != Op_CreateEx && 2219 nop != Op_CheckCastPP && 2220 nop != Op_DecodeN && 2221 !n->is_Mem() ) { 2222 Node *x = n->clone(); 2223 call->set_req( TypeFunc::Parms, x ); 2224 } 2225 } 2226 break; 2227 } 2228 2229 case Op_StoreD: 2230 case Op_LoadD: 2231 case Op_LoadD_unaligned: 2232 frc.inc_double_count(); 2233 goto handle_mem; 2234 case Op_StoreF: 2235 case Op_LoadF: 2236 frc.inc_float_count(); 2237 goto handle_mem; 2238 2239 case Op_StoreCM: 2240 { 2241 // Convert OopStore dependence into precedence edge 2242 Node* prec = n->in(MemNode::OopStore); 2243 n->del_req(MemNode::OopStore); 2244 n->add_prec(prec); 2245 eliminate_redundant_card_marks(n); 2246 } 2247 2248 // fall through 2249 2250 case Op_StoreB: 2251 case Op_StoreC: 2252 case Op_StorePConditional: 2253 case Op_StoreI: 2254 case Op_StoreL: 2255 case Op_StoreIConditional: 2256 case Op_StoreLConditional: 2257 case Op_CompareAndSwapI: 2258 case Op_CompareAndSwapL: 2259 case Op_CompareAndSwapP: 2260 case Op_CompareAndSwapN: 2261 case Op_StoreP: 2262 case Op_StoreN: 2263 case Op_LoadB: 2264 case Op_LoadUB: 2265 case Op_LoadUS: 2266 case Op_LoadI: 2267 case Op_LoadUI2L: 2268 case Op_LoadKlass: 2269 case Op_LoadNKlass: 2270 case Op_LoadL: 2271 case Op_LoadL_unaligned: 2272 case Op_LoadPLocked: 2273 case Op_LoadLLocked: 2274 case Op_LoadP: 2275 case Op_LoadN: 2276 case Op_LoadRange: 2277 case Op_LoadS: { 2278 handle_mem: 2279 #ifdef ASSERT 2280 if( VerifyOptoOopOffsets ) { 2281 assert( n->is_Mem(), "" ); 2282 MemNode *mem = (MemNode*)n; 2283 // Check to see if address types have grounded out somehow. 2284 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2285 assert( !tp || oop_offset_is_sane(tp), "" ); 2286 } 2287 #endif 2288 break; 2289 } 2290 2291 case Op_AddP: { // Assert sane base pointers 2292 Node *addp = n->in(AddPNode::Address); 2293 assert( !addp->is_AddP() || 2294 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2295 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2296 "Base pointers must match" ); 2297 #ifdef _LP64 2298 if (UseCompressedOops && 2299 addp->Opcode() == Op_ConP && 2300 addp == n->in(AddPNode::Base) && 2301 n->in(AddPNode::Offset)->is_Con()) { 2302 // Use addressing with narrow klass to load with offset on x86. 2303 // On sparc loading 32-bits constant and decoding it have less 2304 // instructions (4) then load 64-bits constant (7). 2305 // Do this transformation here since IGVN will convert ConN back to ConP. 2306 const Type* t = addp->bottom_type(); 2307 if (t->isa_oopptr()) { 2308 Node* nn = NULL; 2309 2310 // Look for existing ConN node of the same exact type. 2311 Compile* C = Compile::current(); 2312 Node* r = C->root(); 2313 uint cnt = r->outcnt(); 2314 for (uint i = 0; i < cnt; i++) { 2315 Node* m = r->raw_out(i); 2316 if (m!= NULL && m->Opcode() == Op_ConN && 2317 m->bottom_type()->make_ptr() == t) { 2318 nn = m; 2319 break; 2320 } 2321 } 2322 if (nn != NULL) { 2323 // Decode a narrow oop to match address 2324 // [R12 + narrow_oop_reg<<3 + offset] 2325 nn = new (C, 2) DecodeNNode(nn, t); 2326 n->set_req(AddPNode::Base, nn); 2327 n->set_req(AddPNode::Address, nn); 2328 if (addp->outcnt() == 0) { 2329 addp->disconnect_inputs(NULL); 2330 } 2331 } 2332 } 2333 } 2334 #endif 2335 break; 2336 } 2337 2338 #ifdef _LP64 2339 case Op_CastPP: 2340 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2341 Compile* C = Compile::current(); 2342 Node* in1 = n->in(1); 2343 const Type* t = n->bottom_type(); 2344 Node* new_in1 = in1->clone(); 2345 new_in1->as_DecodeN()->set_type(t); 2346 2347 if (!Matcher::narrow_oop_use_complex_address()) { 2348 // 2349 // x86, ARM and friends can handle 2 adds in addressing mode 2350 // and Matcher can fold a DecodeN node into address by using 2351 // a narrow oop directly and do implicit NULL check in address: 2352 // 2353 // [R12 + narrow_oop_reg<<3 + offset] 2354 // NullCheck narrow_oop_reg 2355 // 2356 // On other platforms (Sparc) we have to keep new DecodeN node and 2357 // use it to do implicit NULL check in address: 2358 // 2359 // decode_not_null narrow_oop_reg, base_reg 2360 // [base_reg + offset] 2361 // NullCheck base_reg 2362 // 2363 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2364 // to keep the information to which NULL check the new DecodeN node 2365 // corresponds to use it as value in implicit_null_check(). 2366 // 2367 new_in1->set_req(0, n->in(0)); 2368 } 2369 2370 n->subsume_by(new_in1); 2371 if (in1->outcnt() == 0) { 2372 in1->disconnect_inputs(NULL); 2373 } 2374 } 2375 break; 2376 2377 case Op_CmpP: 2378 // Do this transformation here to preserve CmpPNode::sub() and 2379 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2380 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2381 Node* in1 = n->in(1); 2382 Node* in2 = n->in(2); 2383 if (!in1->is_DecodeN()) { 2384 in2 = in1; 2385 in1 = n->in(2); 2386 } 2387 assert(in1->is_DecodeN(), "sanity"); 2388 2389 Compile* C = Compile::current(); 2390 Node* new_in2 = NULL; 2391 if (in2->is_DecodeN()) { 2392 new_in2 = in2->in(1); 2393 } else if (in2->Opcode() == Op_ConP) { 2394 const Type* t = in2->bottom_type(); 2395 if (t == TypePtr::NULL_PTR) { 2396 // Don't convert CmpP null check into CmpN if compressed 2397 // oops implicit null check is not generated. 2398 // This will allow to generate normal oop implicit null check. 2399 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2400 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2401 // 2402 // This transformation together with CastPP transformation above 2403 // will generated code for implicit NULL checks for compressed oops. 2404 // 2405 // The original code after Optimize() 2406 // 2407 // LoadN memory, narrow_oop_reg 2408 // decode narrow_oop_reg, base_reg 2409 // CmpP base_reg, NULL 2410 // CastPP base_reg // NotNull 2411 // Load [base_reg + offset], val_reg 2412 // 2413 // after these transformations will be 2414 // 2415 // LoadN memory, narrow_oop_reg 2416 // CmpN narrow_oop_reg, NULL 2417 // decode_not_null narrow_oop_reg, base_reg 2418 // Load [base_reg + offset], val_reg 2419 // 2420 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2421 // since narrow oops can be used in debug info now (see the code in 2422 // final_graph_reshaping_walk()). 2423 // 2424 // At the end the code will be matched to 2425 // on x86: 2426 // 2427 // Load_narrow_oop memory, narrow_oop_reg 2428 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2429 // NullCheck narrow_oop_reg 2430 // 2431 // and on sparc: 2432 // 2433 // Load_narrow_oop memory, narrow_oop_reg 2434 // decode_not_null narrow_oop_reg, base_reg 2435 // Load [base_reg + offset], val_reg 2436 // NullCheck base_reg 2437 // 2438 } else if (t->isa_oopptr()) { 2439 new_in2 = ConNode::make(C, t->make_narrowoop()); 2440 } 2441 } 2442 if (new_in2 != NULL) { 2443 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2444 n->subsume_by( cmpN ); 2445 if (in1->outcnt() == 0) { 2446 in1->disconnect_inputs(NULL); 2447 } 2448 if (in2->outcnt() == 0) { 2449 in2->disconnect_inputs(NULL); 2450 } 2451 } 2452 } 2453 break; 2454 2455 case Op_DecodeN: 2456 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2457 // DecodeN could be pinned when it can't be fold into 2458 // an address expression, see the code for Op_CastPP above. 2459 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2460 break; 2461 2462 case Op_EncodeP: { 2463 Node* in1 = n->in(1); 2464 if (in1->is_DecodeN()) { 2465 n->subsume_by(in1->in(1)); 2466 } else if (in1->Opcode() == Op_ConP) { 2467 Compile* C = Compile::current(); 2468 const Type* t = in1->bottom_type(); 2469 if (t == TypePtr::NULL_PTR) { 2470 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2471 } else if (t->isa_oopptr()) { 2472 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2473 } 2474 } 2475 if (in1->outcnt() == 0) { 2476 in1->disconnect_inputs(NULL); 2477 } 2478 break; 2479 } 2480 2481 case Op_Proj: { 2482 if (OptimizeStringConcat) { 2483 ProjNode* p = n->as_Proj(); 2484 if (p->_is_io_use) { 2485 // Separate projections were used for the exception path which 2486 // are normally removed by a late inline. If it wasn't inlined 2487 // then they will hang around and should just be replaced with 2488 // the original one. 2489 Node* proj = NULL; 2490 // Replace with just one 2491 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2492 Node *use = i.get(); 2493 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2494 proj = use; 2495 break; 2496 } 2497 } 2498 assert(p != NULL, "must be found"); 2499 p->subsume_by(proj); 2500 } 2501 } 2502 break; 2503 } 2504 2505 case Op_Phi: 2506 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2507 // The EncodeP optimization may create Phi with the same edges 2508 // for all paths. It is not handled well by Register Allocator. 2509 Node* unique_in = n->in(1); 2510 assert(unique_in != NULL, ""); 2511 uint cnt = n->req(); 2512 for (uint i = 2; i < cnt; i++) { 2513 Node* m = n->in(i); 2514 assert(m != NULL, ""); 2515 if (unique_in != m) 2516 unique_in = NULL; 2517 } 2518 if (unique_in != NULL) { 2519 n->subsume_by(unique_in); 2520 } 2521 } 2522 break; 2523 2524 #endif 2525 2526 case Op_ModI: 2527 if (UseDivMod) { 2528 // Check if a%b and a/b both exist 2529 Node* d = n->find_similar(Op_DivI); 2530 if (d) { 2531 // Replace them with a fused divmod if supported 2532 Compile* C = Compile::current(); 2533 if (Matcher::has_match_rule(Op_DivModI)) { 2534 DivModINode* divmod = DivModINode::make(C, n); 2535 d->subsume_by(divmod->div_proj()); 2536 n->subsume_by(divmod->mod_proj()); 2537 } else { 2538 // replace a%b with a-((a/b)*b) 2539 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2540 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2541 n->subsume_by( sub ); 2542 } 2543 } 2544 } 2545 break; 2546 2547 case Op_ModL: 2548 if (UseDivMod) { 2549 // Check if a%b and a/b both exist 2550 Node* d = n->find_similar(Op_DivL); 2551 if (d) { 2552 // Replace them with a fused divmod if supported 2553 Compile* C = Compile::current(); 2554 if (Matcher::has_match_rule(Op_DivModL)) { 2555 DivModLNode* divmod = DivModLNode::make(C, n); 2556 d->subsume_by(divmod->div_proj()); 2557 n->subsume_by(divmod->mod_proj()); 2558 } else { 2559 // replace a%b with a-((a/b)*b) 2560 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2561 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2562 n->subsume_by( sub ); 2563 } 2564 } 2565 } 2566 break; 2567 2568 case Op_Load16B: 2569 case Op_Load8B: 2570 case Op_Load4B: 2571 case Op_Load8S: 2572 case Op_Load4S: 2573 case Op_Load2S: 2574 case Op_Load8C: 2575 case Op_Load4C: 2576 case Op_Load2C: 2577 case Op_Load4I: 2578 case Op_Load2I: 2579 case Op_Load2L: 2580 case Op_Load4F: 2581 case Op_Load2F: 2582 case Op_Load2D: 2583 case Op_Store16B: 2584 case Op_Store8B: 2585 case Op_Store4B: 2586 case Op_Store8C: 2587 case Op_Store4C: 2588 case Op_Store2C: 2589 case Op_Store4I: 2590 case Op_Store2I: 2591 case Op_Store2L: 2592 case Op_Store4F: 2593 case Op_Store2F: 2594 case Op_Store2D: 2595 break; 2596 2597 case Op_PackB: 2598 case Op_PackS: 2599 case Op_PackC: 2600 case Op_PackI: 2601 case Op_PackF: 2602 case Op_PackL: 2603 case Op_PackD: 2604 if (n->req()-1 > 2) { 2605 // Replace many operand PackNodes with a binary tree for matching 2606 PackNode* p = (PackNode*) n; 2607 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req()); 2608 n->subsume_by(btp); 2609 } 2610 break; 2611 case Op_Loop: 2612 case Op_CountedLoop: 2613 if (n->as_Loop()->is_inner_loop()) { 2614 frc.inc_inner_loop_count(); 2615 } 2616 break; 2617 case Op_LShiftI: 2618 case Op_RShiftI: 2619 case Op_URShiftI: 2620 case Op_LShiftL: 2621 case Op_RShiftL: 2622 case Op_URShiftL: 2623 if (Matcher::need_masked_shift_count) { 2624 // The cpu's shift instructions don't restrict the count to the 2625 // lower 5/6 bits. We need to do the masking ourselves. 2626 Node* in2 = n->in(2); 2627 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2628 const TypeInt* t = in2->find_int_type(); 2629 if (t != NULL && t->is_con()) { 2630 juint shift = t->get_con(); 2631 if (shift > mask) { // Unsigned cmp 2632 Compile* C = Compile::current(); 2633 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask))); 2634 } 2635 } else { 2636 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2637 Compile* C = Compile::current(); 2638 Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask))); 2639 n->set_req(2, shift); 2640 } 2641 } 2642 if (in2->outcnt() == 0) { // Remove dead node 2643 in2->disconnect_inputs(NULL); 2644 } 2645 } 2646 break; 2647 default: 2648 assert( !n->is_Call(), "" ); 2649 assert( !n->is_Mem(), "" ); 2650 break; 2651 } 2652 2653 // Collect CFG split points 2654 if (n->is_MultiBranch()) 2655 frc._tests.push(n); 2656 } 2657 2658 //------------------------------final_graph_reshaping_walk--------------------- 2659 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2660 // requires that the walk visits a node's inputs before visiting the node. 2661 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2662 ResourceArea *area = Thread::current()->resource_area(); 2663 Unique_Node_List sfpt(area); 2664 2665 frc._visited.set(root->_idx); // first, mark node as visited 2666 uint cnt = root->req(); 2667 Node *n = root; 2668 uint i = 0; 2669 while (true) { 2670 if (i < cnt) { 2671 // Place all non-visited non-null inputs onto stack 2672 Node* m = n->in(i); 2673 ++i; 2674 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2675 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2676 sfpt.push(m); 2677 cnt = m->req(); 2678 nstack.push(n, i); // put on stack parent and next input's index 2679 n = m; 2680 i = 0; 2681 } 2682 } else { 2683 // Now do post-visit work 2684 final_graph_reshaping_impl( n, frc ); 2685 if (nstack.is_empty()) 2686 break; // finished 2687 n = nstack.node(); // Get node from stack 2688 cnt = n->req(); 2689 i = nstack.index(); 2690 nstack.pop(); // Shift to the next node on stack 2691 } 2692 } 2693 2694 // Skip next transformation if compressed oops are not used. 2695 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks()) 2696 return; 2697 2698 // Go over safepoints nodes to skip DecodeN nodes for debug edges. 2699 // It could be done for an uncommon traps or any safepoints/calls 2700 // if the DecodeN node is referenced only in a debug info. 2701 while (sfpt.size() > 0) { 2702 n = sfpt.pop(); 2703 JVMState *jvms = n->as_SafePoint()->jvms(); 2704 assert(jvms != NULL, "sanity"); 2705 int start = jvms->debug_start(); 2706 int end = n->req(); 2707 bool is_uncommon = (n->is_CallStaticJava() && 2708 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2709 for (int j = start; j < end; j++) { 2710 Node* in = n->in(j); 2711 if (in->is_DecodeN()) { 2712 bool safe_to_skip = true; 2713 if (!is_uncommon ) { 2714 // Is it safe to skip? 2715 for (uint i = 0; i < in->outcnt(); i++) { 2716 Node* u = in->raw_out(i); 2717 if (!u->is_SafePoint() || 2718 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2719 safe_to_skip = false; 2720 } 2721 } 2722 } 2723 if (safe_to_skip) { 2724 n->set_req(j, in->in(1)); 2725 } 2726 if (in->outcnt() == 0) { 2727 in->disconnect_inputs(NULL); 2728 } 2729 } 2730 } 2731 } 2732 } 2733 2734 //------------------------------final_graph_reshaping-------------------------- 2735 // Final Graph Reshaping. 2736 // 2737 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2738 // and not commoned up and forced early. Must come after regular 2739 // optimizations to avoid GVN undoing the cloning. Clone constant 2740 // inputs to Loop Phis; these will be split by the allocator anyways. 2741 // Remove Opaque nodes. 2742 // (2) Move last-uses by commutative operations to the left input to encourage 2743 // Intel update-in-place two-address operations and better register usage 2744 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2745 // calls canonicalizing them back. 2746 // (3) Count the number of double-precision FP ops, single-precision FP ops 2747 // and call sites. On Intel, we can get correct rounding either by 2748 // forcing singles to memory (requires extra stores and loads after each 2749 // FP bytecode) or we can set a rounding mode bit (requires setting and 2750 // clearing the mode bit around call sites). The mode bit is only used 2751 // if the relative frequency of single FP ops to calls is low enough. 2752 // This is a key transform for SPEC mpeg_audio. 2753 // (4) Detect infinite loops; blobs of code reachable from above but not 2754 // below. Several of the Code_Gen algorithms fail on such code shapes, 2755 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2756 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2757 // Detection is by looking for IfNodes where only 1 projection is 2758 // reachable from below or CatchNodes missing some targets. 2759 // (5) Assert for insane oop offsets in debug mode. 2760 2761 bool Compile::final_graph_reshaping() { 2762 // an infinite loop may have been eliminated by the optimizer, 2763 // in which case the graph will be empty. 2764 if (root()->req() == 1) { 2765 record_method_not_compilable("trivial infinite loop"); 2766 return true; 2767 } 2768 2769 Final_Reshape_Counts frc; 2770 2771 // Visit everybody reachable! 2772 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2773 Node_Stack nstack(unique() >> 1); 2774 final_graph_reshaping_walk(nstack, root(), frc); 2775 2776 // Check for unreachable (from below) code (i.e., infinite loops). 2777 for( uint i = 0; i < frc._tests.size(); i++ ) { 2778 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2779 // Get number of CFG targets. 2780 // Note that PCTables include exception targets after calls. 2781 uint required_outcnt = n->required_outcnt(); 2782 if (n->outcnt() != required_outcnt) { 2783 // Check for a few special cases. Rethrow Nodes never take the 2784 // 'fall-thru' path, so expected kids is 1 less. 2785 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2786 if (n->in(0)->in(0)->is_Call()) { 2787 CallNode *call = n->in(0)->in(0)->as_Call(); 2788 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2789 required_outcnt--; // Rethrow always has 1 less kid 2790 } else if (call->req() > TypeFunc::Parms && 2791 call->is_CallDynamicJava()) { 2792 // Check for null receiver. In such case, the optimizer has 2793 // detected that the virtual call will always result in a null 2794 // pointer exception. The fall-through projection of this CatchNode 2795 // will not be populated. 2796 Node *arg0 = call->in(TypeFunc::Parms); 2797 if (arg0->is_Type() && 2798 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2799 required_outcnt--; 2800 } 2801 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2802 call->req() > TypeFunc::Parms+1 && 2803 call->is_CallStaticJava()) { 2804 // Check for negative array length. In such case, the optimizer has 2805 // detected that the allocation attempt will always result in an 2806 // exception. There is no fall-through projection of this CatchNode . 2807 Node *arg1 = call->in(TypeFunc::Parms+1); 2808 if (arg1->is_Type() && 2809 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2810 required_outcnt--; 2811 } 2812 } 2813 } 2814 } 2815 // Recheck with a better notion of 'required_outcnt' 2816 if (n->outcnt() != required_outcnt) { 2817 record_method_not_compilable("malformed control flow"); 2818 return true; // Not all targets reachable! 2819 } 2820 } 2821 // Check that I actually visited all kids. Unreached kids 2822 // must be infinite loops. 2823 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2824 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2825 record_method_not_compilable("infinite loop"); 2826 return true; // Found unvisited kid; must be unreach 2827 } 2828 } 2829 2830 // If original bytecodes contained a mixture of floats and doubles 2831 // check if the optimizer has made it homogenous, item (3). 2832 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2833 frc.get_float_count() > 32 && 2834 frc.get_double_count() == 0 && 2835 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2836 set_24_bit_selection_and_mode( false, true ); 2837 } 2838 2839 set_java_calls(frc.get_java_call_count()); 2840 set_inner_loops(frc.get_inner_loop_count()); 2841 2842 // No infinite loops, no reason to bail out. 2843 return false; 2844 } 2845 2846 //-----------------------------too_many_traps---------------------------------- 2847 // Report if there are too many traps at the current method and bci. 2848 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2849 bool Compile::too_many_traps(ciMethod* method, 2850 int bci, 2851 Deoptimization::DeoptReason reason) { 2852 ciMethodData* md = method->method_data(); 2853 if (md->is_empty()) { 2854 // Assume the trap has not occurred, or that it occurred only 2855 // because of a transient condition during start-up in the interpreter. 2856 return false; 2857 } 2858 if (md->has_trap_at(bci, reason) != 0) { 2859 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2860 // Also, if there are multiple reasons, or if there is no per-BCI record, 2861 // assume the worst. 2862 if (log()) 2863 log()->elem("observe trap='%s' count='%d'", 2864 Deoptimization::trap_reason_name(reason), 2865 md->trap_count(reason)); 2866 return true; 2867 } else { 2868 // Ignore method/bci and see if there have been too many globally. 2869 return too_many_traps(reason, md); 2870 } 2871 } 2872 2873 // Less-accurate variant which does not require a method and bci. 2874 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2875 ciMethodData* logmd) { 2876 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2877 // Too many traps globally. 2878 // Note that we use cumulative trap_count, not just md->trap_count. 2879 if (log()) { 2880 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2881 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2882 Deoptimization::trap_reason_name(reason), 2883 mcount, trap_count(reason)); 2884 } 2885 return true; 2886 } else { 2887 // The coast is clear. 2888 return false; 2889 } 2890 } 2891 2892 //--------------------------too_many_recompiles-------------------------------- 2893 // Report if there are too many recompiles at the current method and bci. 2894 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2895 // Is not eager to return true, since this will cause the compiler to use 2896 // Action_none for a trap point, to avoid too many recompilations. 2897 bool Compile::too_many_recompiles(ciMethod* method, 2898 int bci, 2899 Deoptimization::DeoptReason reason) { 2900 ciMethodData* md = method->method_data(); 2901 if (md->is_empty()) { 2902 // Assume the trap has not occurred, or that it occurred only 2903 // because of a transient condition during start-up in the interpreter. 2904 return false; 2905 } 2906 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2907 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2908 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2909 Deoptimization::DeoptReason per_bc_reason 2910 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2911 if ((per_bc_reason == Deoptimization::Reason_none 2912 || md->has_trap_at(bci, reason) != 0) 2913 // The trap frequency measure we care about is the recompile count: 2914 && md->trap_recompiled_at(bci) 2915 && md->overflow_recompile_count() >= bc_cutoff) { 2916 // Do not emit a trap here if it has already caused recompilations. 2917 // Also, if there are multiple reasons, or if there is no per-BCI record, 2918 // assume the worst. 2919 if (log()) 2920 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2921 Deoptimization::trap_reason_name(reason), 2922 md->trap_count(reason), 2923 md->overflow_recompile_count()); 2924 return true; 2925 } else if (trap_count(reason) != 0 2926 && decompile_count() >= m_cutoff) { 2927 // Too many recompiles globally, and we have seen this sort of trap. 2928 // Use cumulative decompile_count, not just md->decompile_count. 2929 if (log()) 2930 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2931 Deoptimization::trap_reason_name(reason), 2932 md->trap_count(reason), trap_count(reason), 2933 md->decompile_count(), decompile_count()); 2934 return true; 2935 } else { 2936 // The coast is clear. 2937 return false; 2938 } 2939 } 2940 2941 2942 #ifndef PRODUCT 2943 //------------------------------verify_graph_edges--------------------------- 2944 // Walk the Graph and verify that there is a one-to-one correspondence 2945 // between Use-Def edges and Def-Use edges in the graph. 2946 void Compile::verify_graph_edges(bool no_dead_code) { 2947 if (VerifyGraphEdges) { 2948 ResourceArea *area = Thread::current()->resource_area(); 2949 Unique_Node_List visited(area); 2950 // Call recursive graph walk to check edges 2951 _root->verify_edges(visited); 2952 if (no_dead_code) { 2953 // Now make sure that no visited node is used by an unvisited node. 2954 bool dead_nodes = 0; 2955 Unique_Node_List checked(area); 2956 while (visited.size() > 0) { 2957 Node* n = visited.pop(); 2958 checked.push(n); 2959 for (uint i = 0; i < n->outcnt(); i++) { 2960 Node* use = n->raw_out(i); 2961 if (checked.member(use)) continue; // already checked 2962 if (visited.member(use)) continue; // already in the graph 2963 if (use->is_Con()) continue; // a dead ConNode is OK 2964 // At this point, we have found a dead node which is DU-reachable. 2965 if (dead_nodes++ == 0) 2966 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2967 use->dump(2); 2968 tty->print_cr("---"); 2969 checked.push(use); // No repeats; pretend it is now checked. 2970 } 2971 } 2972 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2973 } 2974 } 2975 } 2976 #endif 2977 2978 // The Compile object keeps track of failure reasons separately from the ciEnv. 2979 // This is required because there is not quite a 1-1 relation between the 2980 // ciEnv and its compilation task and the Compile object. Note that one 2981 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 2982 // to backtrack and retry without subsuming loads. Other than this backtracking 2983 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 2984 // by the logic in C2Compiler. 2985 void Compile::record_failure(const char* reason) { 2986 if (log() != NULL) { 2987 log()->elem("failure reason='%s' phase='compile'", reason); 2988 } 2989 if (_failure_reason == NULL) { 2990 // Record the first failure reason. 2991 _failure_reason = reason; 2992 } 2993 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 2994 C->print_method(_failure_reason); 2995 } 2996 _root = NULL; // flush the graph, too 2997 } 2998 2999 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3000 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 3001 { 3002 if (dolog) { 3003 C = Compile::current(); 3004 _log = C->log(); 3005 } else { 3006 C = NULL; 3007 _log = NULL; 3008 } 3009 if (_log != NULL) { 3010 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 3011 _log->stamp(); 3012 _log->end_head(); 3013 } 3014 } 3015 3016 Compile::TracePhase::~TracePhase() { 3017 if (_log != NULL) { 3018 _log->done("phase nodes='%d'", C->unique()); 3019 } 3020 } 3021 3022 //============================================================================= 3023 // Two Constant's are equal when the type and the value are equal. 3024 bool Compile::Constant::operator==(const Constant& other) { 3025 if (type() != other.type() ) return false; 3026 if (can_be_reused() != other.can_be_reused()) return false; 3027 // For floating point values we compare the bit pattern. 3028 switch (type()) { 3029 case T_FLOAT: return (_value.i == other._value.i); 3030 case T_LONG: 3031 case T_DOUBLE: return (_value.j == other._value.j); 3032 case T_OBJECT: 3033 case T_ADDRESS: return (_value.l == other._value.l); 3034 case T_VOID: return (_value.l == other._value.l); // jump-table entries 3035 default: ShouldNotReachHere(); 3036 } 3037 return false; 3038 } 3039 3040 // Emit constants grouped in the following order: 3041 static BasicType type_order[] = { 3042 T_FLOAT, // 32-bit 3043 T_OBJECT, // 32 or 64-bit 3044 T_ADDRESS, // 32 or 64-bit 3045 T_DOUBLE, // 64-bit 3046 T_LONG, // 64-bit 3047 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons) 3048 T_ILLEGAL 3049 }; 3050 3051 static int type_to_size_in_bytes(BasicType t) { 3052 switch (t) { 3053 case T_LONG: return sizeof(jlong ); 3054 case T_FLOAT: return sizeof(jfloat ); 3055 case T_DOUBLE: return sizeof(jdouble); 3056 // We use T_VOID as marker for jump-table entries (labels) which 3057 // need an interal word relocation. 3058 case T_VOID: 3059 case T_ADDRESS: 3060 case T_OBJECT: return sizeof(jobject); 3061 } 3062 3063 ShouldNotReachHere(); 3064 return -1; 3065 } 3066 3067 void Compile::ConstantTable::calculate_offsets_and_size() { 3068 int size = 0; 3069 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 3070 BasicType type = type_order[t]; 3071 3072 for (int i = 0; i < _constants.length(); i++) { 3073 Constant con = _constants.at(i); 3074 if (con.type() != type) continue; // Skip other types. 3075 3076 // Align size for type. 3077 int typesize = type_to_size_in_bytes(con.type()); 3078 size = align_size_up(size, typesize); 3079 3080 // Set offset. 3081 con.set_offset(size); 3082 _constants.at_put(i, con); 3083 3084 // Add type size. 3085 size = size + typesize; 3086 } 3087 } 3088 3089 // Align size up to the next section start (which is insts; see 3090 // CodeBuffer::align_at_start). 3091 assert(_size == -1, "already set?"); 3092 _size = align_size_up(size, CodeEntryAlignment); 3093 3094 if (Matcher::constant_table_absolute_addressing) { 3095 set_table_base_offset(0); // No table base offset required 3096 } else { 3097 if (UseRDPCForConstantTableBase) { 3098 // table base offset is set in MachConstantBaseNode::emit 3099 } else { 3100 // When RDPC is not used, the table base is set into the middle of 3101 // the constant table. 3102 int half_size = _size / 2; 3103 assert(half_size * 2 == _size, "sanity"); 3104 set_table_base_offset(-half_size); 3105 } 3106 } 3107 } 3108 3109 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3110 MacroAssembler _masm(&cb); 3111 for (int t = 0; type_order[t] != T_ILLEGAL; t++) { 3112 BasicType type = type_order[t]; 3113 3114 for (int i = 0; i < _constants.length(); i++) { 3115 Constant con = _constants.at(i); 3116 if (con.type() != type) continue; // Skip other types. 3117 3118 address constant_addr; 3119 switch (con.type()) { 3120 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3121 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3122 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3123 case T_OBJECT: { 3124 jobject obj = con.get_jobject(); 3125 int oop_index = _masm.oop_recorder()->find_index(obj); 3126 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3127 break; 3128 } 3129 case T_ADDRESS: { 3130 address addr = (address) con.get_jobject(); 3131 constant_addr = _masm.address_constant(addr); 3132 break; 3133 } 3134 // We use T_VOID as marker for jump-table entries (labels) which 3135 // need an interal word relocation. 3136 case T_VOID: { 3137 // Write a dummy word. The real value is filled in later 3138 // in fill_jump_table_in_constant_table. 3139 address addr = (address) con.get_jobject(); 3140 constant_addr = _masm.address_constant(addr); 3141 break; 3142 } 3143 default: ShouldNotReachHere(); 3144 } 3145 assert(constant_addr != NULL, "consts section too small"); 3146 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3147 } 3148 } 3149 } 3150 3151 int Compile::ConstantTable::find_offset(Constant& con) const { 3152 int idx = _constants.find(con); 3153 assert(idx != -1, "constant must be in constant table"); 3154 int offset = _constants.at(idx).offset(); 3155 assert(offset != -1, "constant table not emitted yet?"); 3156 return offset; 3157 } 3158 3159 void Compile::ConstantTable::add(Constant& con) { 3160 if (con.can_be_reused()) { 3161 int idx = _constants.find(con); 3162 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3163 return; 3164 } 3165 } 3166 (void) _constants.append(con); 3167 } 3168 3169 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) { 3170 Constant con(type, value); 3171 add(con); 3172 return con; 3173 } 3174 3175 Compile::Constant Compile::ConstantTable::add(MachOper* oper) { 3176 jvalue value; 3177 BasicType type = oper->type()->basic_type(); 3178 switch (type) { 3179 case T_LONG: value.j = oper->constantL(); break; 3180 case T_FLOAT: value.f = oper->constantF(); break; 3181 case T_DOUBLE: value.d = oper->constantD(); break; 3182 case T_OBJECT: 3183 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3184 default: ShouldNotReachHere(); 3185 } 3186 return add(type, value); 3187 } 3188 3189 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) { 3190 jvalue value; 3191 // We can use the node pointer here to identify the right jump-table 3192 // as this method is called from Compile::Fill_buffer right before 3193 // the MachNodes are emitted and the jump-table is filled (means the 3194 // MachNode pointers do not change anymore). 3195 value.l = (jobject) n; 3196 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused. 3197 for (uint i = 0; i < n->outcnt(); i++) { 3198 add(con); 3199 } 3200 return con; 3201 } 3202 3203 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3204 // If called from Compile::scratch_emit_size do nothing. 3205 if (Compile::current()->in_scratch_emit_size()) return; 3206 3207 assert(labels.is_nonempty(), "must be"); 3208 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt())); 3209 3210 // Since MachConstantNode::constant_offset() also contains 3211 // table_base_offset() we need to subtract the table_base_offset() 3212 // to get the plain offset into the constant table. 3213 int offset = n->constant_offset() - table_base_offset(); 3214 3215 MacroAssembler _masm(&cb); 3216 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3217 3218 for (int i = 0; i < labels.length(); i++) { 3219 address* constant_addr = &jump_table_base[i]; 3220 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer"); 3221 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3222 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3223 } 3224 }