1 /* 2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/assembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "code/exceptionHandlerTable.hpp" 29 #include "code/nmethod.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "compiler/oopMap.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/block.hpp" 34 #include "opto/c2compiler.hpp" 35 #include "opto/callGenerator.hpp" 36 #include "opto/callnode.hpp" 37 #include "opto/cfgnode.hpp" 38 #include "opto/chaitin.hpp" 39 #include "opto/compile.hpp" 40 #include "opto/connode.hpp" 41 #include "opto/divnode.hpp" 42 #include "opto/escape.hpp" 43 #include "opto/idealGraphPrinter.hpp" 44 #include "opto/loopnode.hpp" 45 #include "opto/machnode.hpp" 46 #include "opto/macro.hpp" 47 #include "opto/matcher.hpp" 48 #include "opto/memnode.hpp" 49 #include "opto/mulnode.hpp" 50 #include "opto/node.hpp" 51 #include "opto/opcodes.hpp" 52 #include "opto/output.hpp" 53 #include "opto/parse.hpp" 54 #include "opto/phaseX.hpp" 55 #include "opto/rootnode.hpp" 56 #include "opto/runtime.hpp" 57 #include "opto/stringopts.hpp" 58 #include "opto/type.hpp" 59 #include "opto/vectornode.hpp" 60 #include "runtime/arguments.hpp" 61 #include "runtime/signature.hpp" 62 #include "runtime/stubRoutines.hpp" 63 #include "runtime/timer.hpp" 64 #include "utilities/copy.hpp" 65 #ifdef TARGET_ARCH_MODEL_x86_32 66 # include "adfiles/ad_x86_32.hpp" 67 #endif 68 #ifdef TARGET_ARCH_MODEL_x86_64 69 # include "adfiles/ad_x86_64.hpp" 70 #endif 71 #ifdef TARGET_ARCH_MODEL_sparc 72 # include "adfiles/ad_sparc.hpp" 73 #endif 74 #ifdef TARGET_ARCH_MODEL_zero 75 # include "adfiles/ad_zero.hpp" 76 #endif 77 #ifdef TARGET_ARCH_MODEL_arm 78 # include "adfiles/ad_arm.hpp" 79 #endif 80 #ifdef TARGET_ARCH_MODEL_ppc 81 # include "adfiles/ad_ppc.hpp" 82 #endif 83 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 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 // Remove useless macro and predicate opaq nodes 369 for (int i = C->macro_count()-1; i >= 0; i--) { 370 Node* n = C->macro_node(i); 371 if (!useful.member(n)) { 372 remove_macro_node(n); 373 } 374 } 375 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 376 } 377 378 //------------------------------frame_size_in_words----------------------------- 379 // frame_slots in units of words 380 int Compile::frame_size_in_words() const { 381 // shift is 0 in LP32 and 1 in LP64 382 const int shift = (LogBytesPerWord - LogBytesPerInt); 383 int words = _frame_slots >> shift; 384 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 385 return words; 386 } 387 388 // ============================================================================ 389 //------------------------------CompileWrapper--------------------------------- 390 class CompileWrapper : public StackObj { 391 Compile *const _compile; 392 public: 393 CompileWrapper(Compile* compile); 394 395 ~CompileWrapper(); 396 }; 397 398 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 399 // the Compile* pointer is stored in the current ciEnv: 400 ciEnv* env = compile->env(); 401 assert(env == ciEnv::current(), "must already be a ciEnv active"); 402 assert(env->compiler_data() == NULL, "compile already active?"); 403 env->set_compiler_data(compile); 404 assert(compile == Compile::current(), "sanity"); 405 406 compile->set_type_dict(NULL); 407 compile->set_type_hwm(NULL); 408 compile->set_type_last_size(0); 409 compile->set_last_tf(NULL, NULL); 410 compile->set_indexSet_arena(NULL); 411 compile->set_indexSet_free_block_list(NULL); 412 compile->init_type_arena(); 413 Type::Initialize(compile); 414 _compile->set_scratch_buffer_blob(NULL); 415 _compile->begin_method(); 416 } 417 CompileWrapper::~CompileWrapper() { 418 _compile->end_method(); 419 if (_compile->scratch_buffer_blob() != NULL) 420 BufferBlob::free(_compile->scratch_buffer_blob()); 421 _compile->env()->set_compiler_data(NULL); 422 } 423 424 425 //----------------------------print_compile_messages--------------------------- 426 void Compile::print_compile_messages() { 427 #ifndef PRODUCT 428 // Check if recompiling 429 if (_subsume_loads == false && PrintOpto) { 430 // Recompiling without allowing machine instructions to subsume loads 431 tty->print_cr("*********************************************************"); 432 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 433 tty->print_cr("*********************************************************"); 434 } 435 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 436 // Recompiling without escape analysis 437 tty->print_cr("*********************************************************"); 438 tty->print_cr("** Bailout: Recompile without escape analysis **"); 439 tty->print_cr("*********************************************************"); 440 } 441 if (env()->break_at_compile()) { 442 // Open the debugger when compiling this method. 443 tty->print("### Breaking when compiling: "); 444 method()->print_short_name(); 445 tty->cr(); 446 BREAKPOINT; 447 } 448 449 if( PrintOpto ) { 450 if (is_osr_compilation()) { 451 tty->print("[OSR]%3d", _compile_id); 452 } else { 453 tty->print("%3d", _compile_id); 454 } 455 } 456 #endif 457 } 458 459 460 //-----------------------init_scratch_buffer_blob------------------------------ 461 // Construct a temporary BufferBlob and cache it for this compile. 462 void Compile::init_scratch_buffer_blob(int const_size) { 463 // If there is already a scratch buffer blob allocated and the 464 // constant section is big enough, use it. Otherwise free the 465 // current and allocate a new one. 466 BufferBlob* blob = scratch_buffer_blob(); 467 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 468 // Use the current blob. 469 } else { 470 if (blob != NULL) { 471 BufferBlob::free(blob); 472 } 473 474 ResourceMark rm; 475 _scratch_const_size = const_size; 476 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 477 blob = BufferBlob::create("Compile::scratch_buffer", size); 478 // Record the buffer blob for next time. 479 set_scratch_buffer_blob(blob); 480 // Have we run out of code space? 481 if (scratch_buffer_blob() == NULL) { 482 // Let CompilerBroker disable further compilations. 483 record_failure("Not enough space for scratch buffer in CodeCache"); 484 return; 485 } 486 } 487 488 // Initialize the relocation buffers 489 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 490 set_scratch_locs_memory(locs_buf); 491 } 492 493 494 //-----------------------scratch_emit_size------------------------------------- 495 // Helper function that computes size by emitting code 496 uint Compile::scratch_emit_size(const Node* n) { 497 // Start scratch_emit_size section. 498 set_in_scratch_emit_size(true); 499 500 // Emit into a trash buffer and count bytes emitted. 501 // This is a pretty expensive way to compute a size, 502 // but it works well enough if seldom used. 503 // All common fixed-size instructions are given a size 504 // method by the AD file. 505 // Note that the scratch buffer blob and locs memory are 506 // allocated at the beginning of the compile task, and 507 // may be shared by several calls to scratch_emit_size. 508 // The allocation of the scratch buffer blob is particularly 509 // expensive, since it has to grab the code cache lock. 510 BufferBlob* blob = this->scratch_buffer_blob(); 511 assert(blob != NULL, "Initialize BufferBlob at start"); 512 assert(blob->size() > MAX_inst_size, "sanity"); 513 relocInfo* locs_buf = scratch_locs_memory(); 514 address blob_begin = blob->content_begin(); 515 address blob_end = (address)locs_buf; 516 assert(blob->content_contains(blob_end), "sanity"); 517 CodeBuffer buf(blob_begin, blob_end - blob_begin); 518 buf.initialize_consts_size(_scratch_const_size); 519 buf.initialize_stubs_size(MAX_stubs_size); 520 assert(locs_buf != NULL, "sanity"); 521 int lsize = MAX_locs_size / 3; 522 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 523 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 524 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 525 526 // Do the emission. 527 528 Label fakeL; // Fake label for branch instructions. 529 Label* saveL = NULL; 530 uint save_bnum = 0; 531 bool is_branch = n->is_MachBranch(); 532 if (is_branch) { 533 MacroAssembler masm(&buf); 534 masm.bind(fakeL); 535 n->as_MachBranch()->save_label(&saveL, &save_bnum); 536 n->as_MachBranch()->label_set(&fakeL, 0); 537 } 538 n->emit(buf, this->regalloc()); 539 if (is_branch) // Restore label. 540 n->as_MachBranch()->label_set(saveL, save_bnum); 541 542 // End scratch_emit_size section. 543 set_in_scratch_emit_size(false); 544 545 return buf.insts_size(); 546 } 547 548 549 // ============================================================================ 550 //------------------------------Compile standard------------------------------- 551 debug_only( int Compile::_debug_idx = 100000; ) 552 553 // Compile a method. entry_bci is -1 for normal compilations and indicates 554 // the continuation bci for on stack replacement. 555 556 557 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis ) 558 : Phase(Compiler), 559 _env(ci_env), 560 _log(ci_env->log()), 561 _compile_id(ci_env->compile_id()), 562 _save_argument_registers(false), 563 _stub_name(NULL), 564 _stub_function(NULL), 565 _stub_entry_point(NULL), 566 _method(target), 567 _entry_bci(osr_bci), 568 _initial_gvn(NULL), 569 _for_igvn(NULL), 570 _warm_calls(NULL), 571 _subsume_loads(subsume_loads), 572 _do_escape_analysis(do_escape_analysis), 573 _failure_reason(NULL), 574 _code_buffer("Compile::Fill_buffer"), 575 _orig_pc_slot(0), 576 _orig_pc_slot_offset_in_bytes(0), 577 _has_method_handle_invokes(false), 578 _mach_constant_base_node(NULL), 579 _node_bundling_limit(0), 580 _node_bundling_base(NULL), 581 _java_calls(0), 582 _inner_loops(0), 583 _scratch_const_size(-1), 584 _in_scratch_emit_size(false), 585 #ifndef PRODUCT 586 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 587 _printer(IdealGraphPrinter::printer()), 588 #endif 589 _congraph(NULL) { 590 C = this; 591 592 CompileWrapper cw(this); 593 #ifndef PRODUCT 594 if (TimeCompiler2) { 595 tty->print(" "); 596 target->holder()->name()->print(); 597 tty->print("."); 598 target->print_short_name(); 599 tty->print(" "); 600 } 601 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 602 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 603 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 604 if (!print_opto_assembly) { 605 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 606 if (print_assembly && !Disassembler::can_decode()) { 607 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 608 print_opto_assembly = true; 609 } 610 } 611 set_print_assembly(print_opto_assembly); 612 set_parsed_irreducible_loop(false); 613 #endif 614 615 if (ProfileTraps) { 616 // Make sure the method being compiled gets its own MDO, 617 // so we can at least track the decompile_count(). 618 method()->ensure_method_data(); 619 } 620 621 Init(::AliasLevel); 622 623 624 print_compile_messages(); 625 626 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 627 _ilt = InlineTree::build_inline_tree_root(); 628 else 629 _ilt = NULL; 630 631 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 632 assert(num_alias_types() >= AliasIdxRaw, ""); 633 634 #define MINIMUM_NODE_HASH 1023 635 // Node list that Iterative GVN will start with 636 Unique_Node_List for_igvn(comp_arena()); 637 set_for_igvn(&for_igvn); 638 639 // GVN that will be run immediately on new nodes 640 uint estimated_size = method()->code_size()*4+64; 641 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 642 PhaseGVN gvn(node_arena(), estimated_size); 643 set_initial_gvn(&gvn); 644 645 { // Scope for timing the parser 646 TracePhase t3("parse", &_t_parser, true); 647 648 // Put top into the hash table ASAP. 649 initial_gvn()->transform_no_reclaim(top()); 650 651 // Set up tf(), start(), and find a CallGenerator. 652 CallGenerator* cg = NULL; 653 if (is_osr_compilation()) { 654 const TypeTuple *domain = StartOSRNode::osr_domain(); 655 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 656 init_tf(TypeFunc::make(domain, range)); 657 StartNode* s = new (this, 2) StartOSRNode(root(), domain); 658 initial_gvn()->set_type_bottom(s); 659 init_start(s); 660 cg = CallGenerator::for_osr(method(), entry_bci()); 661 } else { 662 // Normal case. 663 init_tf(TypeFunc::make(method())); 664 StartNode* s = new (this, 2) StartNode(root(), tf()->domain()); 665 initial_gvn()->set_type_bottom(s); 666 init_start(s); 667 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 668 // With java.lang.ref.reference.get() we must go through the 669 // intrinsic when G1 is enabled - even when get() is the root 670 // method of the compile - so that, if necessary, the value in 671 // the referent field of the reference object gets recorded by 672 // the pre-barrier code. 673 // Specifically, if G1 is enabled, the value in the referent 674 // field is recorded by the G1 SATB pre barrier. This will 675 // result in the referent being marked live and the reference 676 // object removed from the list of discovered references during 677 // reference processing. 678 cg = find_intrinsic(method(), false); 679 } 680 if (cg == NULL) { 681 float past_uses = method()->interpreter_invocation_count(); 682 float expected_uses = past_uses; 683 cg = CallGenerator::for_inline(method(), expected_uses); 684 } 685 } 686 if (failing()) return; 687 if (cg == NULL) { 688 record_method_not_compilable_all_tiers("cannot parse method"); 689 return; 690 } 691 JVMState* jvms = build_start_state(start(), tf()); 692 if ((jvms = cg->generate(jvms)) == NULL) { 693 record_method_not_compilable("method parse failed"); 694 return; 695 } 696 GraphKit kit(jvms); 697 698 if (!kit.stopped()) { 699 // Accept return values, and transfer control we know not where. 700 // This is done by a special, unique ReturnNode bound to root. 701 return_values(kit.jvms()); 702 } 703 704 if (kit.has_exceptions()) { 705 // Any exceptions that escape from this call must be rethrown 706 // to whatever caller is dynamically above us on the stack. 707 // This is done by a special, unique RethrowNode bound to root. 708 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 709 } 710 711 if (!failing() && has_stringbuilder()) { 712 { 713 // remove useless nodes to make the usage analysis simpler 714 ResourceMark rm; 715 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 716 } 717 718 { 719 ResourceMark rm; 720 print_method("Before StringOpts", 3); 721 PhaseStringOpts pso(initial_gvn(), &for_igvn); 722 print_method("After StringOpts", 3); 723 } 724 725 // now inline anything that we skipped the first time around 726 while (_late_inlines.length() > 0) { 727 CallGenerator* cg = _late_inlines.pop(); 728 cg->do_late_inline(); 729 if (failing()) return; 730 } 731 } 732 assert(_late_inlines.length() == 0, "should have been processed"); 733 734 print_method("Before RemoveUseless", 3); 735 736 // Remove clutter produced by parsing. 737 if (!failing()) { 738 ResourceMark rm; 739 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 740 } 741 } 742 743 // Note: Large methods are capped off in do_one_bytecode(). 744 if (failing()) return; 745 746 // After parsing, node notes are no longer automagic. 747 // They must be propagated by register_new_node_with_optimizer(), 748 // clone(), or the like. 749 set_default_node_notes(NULL); 750 751 for (;;) { 752 int successes = Inline_Warm(); 753 if (failing()) return; 754 if (successes == 0) break; 755 } 756 757 // Drain the list. 758 Finish_Warm(); 759 #ifndef PRODUCT 760 if (_printer) { 761 _printer->print_inlining(this); 762 } 763 #endif 764 765 if (failing()) return; 766 NOT_PRODUCT( verify_graph_edges(); ) 767 768 // Now optimize 769 Optimize(); 770 if (failing()) return; 771 NOT_PRODUCT( verify_graph_edges(); ) 772 773 #ifndef PRODUCT 774 if (PrintIdeal) { 775 ttyLocker ttyl; // keep the following output all in one block 776 // This output goes directly to the tty, not the compiler log. 777 // To enable tools to match it up with the compilation activity, 778 // be sure to tag this tty output with the compile ID. 779 if (xtty != NULL) { 780 xtty->head("ideal compile_id='%d'%s", compile_id(), 781 is_osr_compilation() ? " compile_kind='osr'" : 782 ""); 783 } 784 root()->dump(9999); 785 if (xtty != NULL) { 786 xtty->tail("ideal"); 787 } 788 } 789 #endif 790 791 // Now that we know the size of all the monitors we can add a fixed slot 792 // for the original deopt pc. 793 794 _orig_pc_slot = fixed_slots(); 795 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 796 set_fixed_slots(next_slot); 797 798 // Now generate code 799 Code_Gen(); 800 if (failing()) return; 801 802 // Check if we want to skip execution of all compiled code. 803 { 804 #ifndef PRODUCT 805 if (OptoNoExecute) { 806 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 807 return; 808 } 809 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 810 #endif 811 812 if (is_osr_compilation()) { 813 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 814 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 815 } else { 816 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 817 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 818 } 819 820 env()->register_method(_method, _entry_bci, 821 &_code_offsets, 822 _orig_pc_slot_offset_in_bytes, 823 code_buffer(), 824 frame_size_in_words(), _oop_map_set, 825 &_handler_table, &_inc_table, 826 compiler, 827 env()->comp_level(), 828 has_unsafe_access() 829 ); 830 } 831 } 832 833 //------------------------------Compile---------------------------------------- 834 // Compile a runtime stub 835 Compile::Compile( ciEnv* ci_env, 836 TypeFunc_generator generator, 837 address stub_function, 838 const char *stub_name, 839 int is_fancy_jump, 840 bool pass_tls, 841 bool save_arg_registers, 842 bool return_pc ) 843 : Phase(Compiler), 844 _env(ci_env), 845 _log(ci_env->log()), 846 _compile_id(-1), 847 _save_argument_registers(save_arg_registers), 848 _method(NULL), 849 _stub_name(stub_name), 850 _stub_function(stub_function), 851 _stub_entry_point(NULL), 852 _entry_bci(InvocationEntryBci), 853 _initial_gvn(NULL), 854 _for_igvn(NULL), 855 _warm_calls(NULL), 856 _orig_pc_slot(0), 857 _orig_pc_slot_offset_in_bytes(0), 858 _subsume_loads(true), 859 _do_escape_analysis(false), 860 _failure_reason(NULL), 861 _code_buffer("Compile::Fill_buffer"), 862 _has_method_handle_invokes(false), 863 _mach_constant_base_node(NULL), 864 _node_bundling_limit(0), 865 _node_bundling_base(NULL), 866 _java_calls(0), 867 _inner_loops(0), 868 #ifndef PRODUCT 869 _trace_opto_output(TraceOptoOutput), 870 _printer(NULL), 871 #endif 872 _congraph(NULL) { 873 C = this; 874 875 #ifndef PRODUCT 876 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 877 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 878 set_print_assembly(PrintFrameConverterAssembly); 879 set_parsed_irreducible_loop(false); 880 #endif 881 CompileWrapper cw(this); 882 Init(/*AliasLevel=*/ 0); 883 init_tf((*generator)()); 884 885 { 886 // The following is a dummy for the sake of GraphKit::gen_stub 887 Unique_Node_List for_igvn(comp_arena()); 888 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 889 PhaseGVN gvn(Thread::current()->resource_area(),255); 890 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 891 gvn.transform_no_reclaim(top()); 892 893 GraphKit kit; 894 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 895 } 896 897 NOT_PRODUCT( verify_graph_edges(); ) 898 Code_Gen(); 899 if (failing()) return; 900 901 902 // Entry point will be accessed using compile->stub_entry_point(); 903 if (code_buffer() == NULL) { 904 Matcher::soft_match_failure(); 905 } else { 906 if (PrintAssembly && (WizardMode || Verbose)) 907 tty->print_cr("### Stub::%s", stub_name); 908 909 if (!failing()) { 910 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 911 912 // Make the NMethod 913 // For now we mark the frame as never safe for profile stackwalking 914 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 915 code_buffer(), 916 CodeOffsets::frame_never_safe, 917 // _code_offsets.value(CodeOffsets::Frame_Complete), 918 frame_size_in_words(), 919 _oop_map_set, 920 save_arg_registers); 921 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 922 923 _stub_entry_point = rs->entry_point(); 924 } 925 } 926 } 927 928 #ifndef PRODUCT 929 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) { 930 if(PrintOpto && Verbose) { 931 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr(); 932 } 933 } 934 #endif 935 936 void Compile::print_codes() { 937 } 938 939 //------------------------------Init------------------------------------------- 940 // Prepare for a single compilation 941 void Compile::Init(int aliaslevel) { 942 _unique = 0; 943 _regalloc = NULL; 944 945 _tf = NULL; // filled in later 946 _top = NULL; // cached later 947 _matcher = NULL; // filled in later 948 _cfg = NULL; // filled in later 949 950 set_24_bit_selection_and_mode(Use24BitFP, false); 951 952 _node_note_array = NULL; 953 _default_node_notes = NULL; 954 955 _immutable_memory = NULL; // filled in at first inquiry 956 957 // Globally visible Nodes 958 // First set TOP to NULL to give safe behavior during creation of RootNode 959 set_cached_top_node(NULL); 960 set_root(new (this, 3) RootNode()); 961 // Now that you have a Root to point to, create the real TOP 962 set_cached_top_node( new (this, 1) ConNode(Type::TOP) ); 963 set_recent_alloc(NULL, NULL); 964 965 // Create Debug Information Recorder to record scopes, oopmaps, etc. 966 env()->set_oop_recorder(new OopRecorder(comp_arena())); 967 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 968 env()->set_dependencies(new Dependencies(env())); 969 970 _fixed_slots = 0; 971 set_has_split_ifs(false); 972 set_has_loops(has_method() && method()->has_loops()); // first approximation 973 set_has_stringbuilder(false); 974 _trap_can_recompile = false; // no traps emitted yet 975 _major_progress = true; // start out assuming good things will happen 976 set_has_unsafe_access(false); 977 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 978 set_decompile_count(0); 979 980 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 981 set_num_loop_opts(LoopOptsCount); 982 set_do_inlining(Inline); 983 set_max_inline_size(MaxInlineSize); 984 set_freq_inline_size(FreqInlineSize); 985 set_do_scheduling(OptoScheduling); 986 set_do_count_invocations(false); 987 set_do_method_data_update(false); 988 989 if (debug_info()->recording_non_safepoints()) { 990 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 991 (comp_arena(), 8, 0, NULL)); 992 set_default_node_notes(Node_Notes::make(this)); 993 } 994 995 // // -- Initialize types before each compile -- 996 // // Update cached type information 997 // if( _method && _method->constants() ) 998 // Type::update_loaded_types(_method, _method->constants()); 999 1000 // Init alias_type map. 1001 if (!_do_escape_analysis && aliaslevel == 3) 1002 aliaslevel = 2; // No unique types without escape analysis 1003 _AliasLevel = aliaslevel; 1004 const int grow_ats = 16; 1005 _max_alias_types = grow_ats; 1006 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1007 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1008 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1009 { 1010 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1011 } 1012 // Initialize the first few types. 1013 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1014 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1015 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1016 _num_alias_types = AliasIdxRaw+1; 1017 // Zero out the alias type cache. 1018 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1019 // A NULL adr_type hits in the cache right away. Preload the right answer. 1020 probe_alias_cache(NULL)->_index = AliasIdxTop; 1021 1022 _intrinsics = NULL; 1023 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1024 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1025 register_library_intrinsics(); 1026 } 1027 1028 //---------------------------init_start---------------------------------------- 1029 // Install the StartNode on this compile object. 1030 void Compile::init_start(StartNode* s) { 1031 if (failing()) 1032 return; // already failing 1033 assert(s == start(), ""); 1034 } 1035 1036 StartNode* Compile::start() const { 1037 assert(!failing(), ""); 1038 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1039 Node* start = root()->fast_out(i); 1040 if( start->is_Start() ) 1041 return start->as_Start(); 1042 } 1043 ShouldNotReachHere(); 1044 return NULL; 1045 } 1046 1047 //-------------------------------immutable_memory------------------------------------- 1048 // Access immutable memory 1049 Node* Compile::immutable_memory() { 1050 if (_immutable_memory != NULL) { 1051 return _immutable_memory; 1052 } 1053 StartNode* s = start(); 1054 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1055 Node *p = s->fast_out(i); 1056 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1057 _immutable_memory = p; 1058 return _immutable_memory; 1059 } 1060 } 1061 ShouldNotReachHere(); 1062 return NULL; 1063 } 1064 1065 //----------------------set_cached_top_node------------------------------------ 1066 // Install the cached top node, and make sure Node::is_top works correctly. 1067 void Compile::set_cached_top_node(Node* tn) { 1068 if (tn != NULL) verify_top(tn); 1069 Node* old_top = _top; 1070 _top = tn; 1071 // Calling Node::setup_is_top allows the nodes the chance to adjust 1072 // their _out arrays. 1073 if (_top != NULL) _top->setup_is_top(); 1074 if (old_top != NULL) old_top->setup_is_top(); 1075 assert(_top == NULL || top()->is_top(), ""); 1076 } 1077 1078 #ifndef PRODUCT 1079 void Compile::verify_top(Node* tn) const { 1080 if (tn != NULL) { 1081 assert(tn->is_Con(), "top node must be a constant"); 1082 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1083 assert(tn->in(0) != NULL, "must have live top node"); 1084 } 1085 } 1086 #endif 1087 1088 1089 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1090 1091 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1092 guarantee(arr != NULL, ""); 1093 int num_blocks = arr->length(); 1094 if (grow_by < num_blocks) grow_by = num_blocks; 1095 int num_notes = grow_by * _node_notes_block_size; 1096 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1097 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1098 while (num_notes > 0) { 1099 arr->append(notes); 1100 notes += _node_notes_block_size; 1101 num_notes -= _node_notes_block_size; 1102 } 1103 assert(num_notes == 0, "exact multiple, please"); 1104 } 1105 1106 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1107 if (source == NULL || dest == NULL) return false; 1108 1109 if (dest->is_Con()) 1110 return false; // Do not push debug info onto constants. 1111 1112 #ifdef ASSERT 1113 // Leave a bread crumb trail pointing to the original node: 1114 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1115 dest->set_debug_orig(source); 1116 } 1117 #endif 1118 1119 if (node_note_array() == NULL) 1120 return false; // Not collecting any notes now. 1121 1122 // This is a copy onto a pre-existing node, which may already have notes. 1123 // If both nodes have notes, do not overwrite any pre-existing notes. 1124 Node_Notes* source_notes = node_notes_at(source->_idx); 1125 if (source_notes == NULL || source_notes->is_clear()) return false; 1126 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1127 if (dest_notes == NULL || dest_notes->is_clear()) { 1128 return set_node_notes_at(dest->_idx, source_notes); 1129 } 1130 1131 Node_Notes merged_notes = (*source_notes); 1132 // The order of operations here ensures that dest notes will win... 1133 merged_notes.update_from(dest_notes); 1134 return set_node_notes_at(dest->_idx, &merged_notes); 1135 } 1136 1137 1138 //--------------------------allow_range_check_smearing------------------------- 1139 // Gating condition for coalescing similar range checks. 1140 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1141 // single covering check that is at least as strong as any of them. 1142 // If the optimization succeeds, the simplified (strengthened) range check 1143 // will always succeed. If it fails, we will deopt, and then give up 1144 // on the optimization. 1145 bool Compile::allow_range_check_smearing() const { 1146 // If this method has already thrown a range-check, 1147 // assume it was because we already tried range smearing 1148 // and it failed. 1149 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1150 return !already_trapped; 1151 } 1152 1153 1154 //------------------------------flatten_alias_type----------------------------- 1155 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1156 int offset = tj->offset(); 1157 TypePtr::PTR ptr = tj->ptr(); 1158 1159 // Known instance (scalarizable allocation) alias only with itself. 1160 bool is_known_inst = tj->isa_oopptr() != NULL && 1161 tj->is_oopptr()->is_known_instance(); 1162 1163 // Process weird unsafe references. 1164 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1165 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1166 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1167 tj = TypeOopPtr::BOTTOM; 1168 ptr = tj->ptr(); 1169 offset = tj->offset(); 1170 } 1171 1172 // Array pointers need some flattening 1173 const TypeAryPtr *ta = tj->isa_aryptr(); 1174 if( ta && is_known_inst ) { 1175 if ( offset != Type::OffsetBot && 1176 offset > arrayOopDesc::length_offset_in_bytes() ) { 1177 offset = Type::OffsetBot; // Flatten constant access into array body only 1178 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1179 } 1180 } else if( ta && _AliasLevel >= 2 ) { 1181 // For arrays indexed by constant indices, we flatten the alias 1182 // space to include all of the array body. Only the header, klass 1183 // and array length can be accessed un-aliased. 1184 if( offset != Type::OffsetBot ) { 1185 if( ta->const_oop() ) { // methodDataOop or methodOop 1186 offset = Type::OffsetBot; // Flatten constant access into array body 1187 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1188 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1189 // range is OK as-is. 1190 tj = ta = TypeAryPtr::RANGE; 1191 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1192 tj = TypeInstPtr::KLASS; // all klass loads look alike 1193 ta = TypeAryPtr::RANGE; // generic ignored junk 1194 ptr = TypePtr::BotPTR; 1195 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1196 tj = TypeInstPtr::MARK; 1197 ta = TypeAryPtr::RANGE; // generic ignored junk 1198 ptr = TypePtr::BotPTR; 1199 } else { // Random constant offset into array body 1200 offset = Type::OffsetBot; // Flatten constant access into array body 1201 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1202 } 1203 } 1204 // Arrays of fixed size alias with arrays of unknown size. 1205 if (ta->size() != TypeInt::POS) { 1206 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1207 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1208 } 1209 // Arrays of known objects become arrays of unknown objects. 1210 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1211 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1212 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1213 } 1214 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1215 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1216 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1217 } 1218 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1219 // cannot be distinguished by bytecode alone. 1220 if (ta->elem() == TypeInt::BOOL) { 1221 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1222 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1223 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1224 } 1225 // During the 2nd round of IterGVN, NotNull castings are removed. 1226 // Make sure the Bottom and NotNull variants alias the same. 1227 // Also, make sure exact and non-exact variants alias the same. 1228 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1229 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1230 } 1231 } 1232 1233 // Oop pointers need some flattening 1234 const TypeInstPtr *to = tj->isa_instptr(); 1235 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1236 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1237 if( ptr == TypePtr::Constant ) { 1238 if (to->klass() != ciEnv::current()->Class_klass() || 1239 offset < k->size_helper() * wordSize) { 1240 // No constant oop pointers (such as Strings); they alias with 1241 // unknown strings. 1242 assert(!is_known_inst, "not scalarizable allocation"); 1243 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1244 } 1245 } else if( is_known_inst ) { 1246 tj = to; // Keep NotNull and klass_is_exact for instance type 1247 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1248 // During the 2nd round of IterGVN, NotNull castings are removed. 1249 // Make sure the Bottom and NotNull variants alias the same. 1250 // Also, make sure exact and non-exact variants alias the same. 1251 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1252 } 1253 // Canonicalize the holder of this field 1254 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1255 // First handle header references such as a LoadKlassNode, even if the 1256 // object's klass is unloaded at compile time (4965979). 1257 if (!is_known_inst) { // Do it only for non-instance types 1258 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1259 } 1260 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1261 // Static fields are in the space above the normal instance 1262 // fields in the java.lang.Class instance. 1263 if (to->klass() != ciEnv::current()->Class_klass()) { 1264 to = NULL; 1265 tj = TypeOopPtr::BOTTOM; 1266 offset = tj->offset(); 1267 } 1268 } else { 1269 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1270 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1271 if( is_known_inst ) { 1272 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1273 } else { 1274 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1275 } 1276 } 1277 } 1278 } 1279 1280 // Klass pointers to object array klasses need some flattening 1281 const TypeKlassPtr *tk = tj->isa_klassptr(); 1282 if( tk ) { 1283 // If we are referencing a field within a Klass, we need 1284 // to assume the worst case of an Object. Both exact and 1285 // inexact types must flatten to the same alias class so 1286 // use NotNull as the PTR. 1287 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1288 1289 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1290 TypeKlassPtr::OBJECT->klass(), 1291 offset); 1292 } 1293 1294 ciKlass* klass = tk->klass(); 1295 if( klass->is_obj_array_klass() ) { 1296 ciKlass* k = TypeAryPtr::OOPS->klass(); 1297 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1298 k = TypeInstPtr::BOTTOM->klass(); 1299 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1300 } 1301 1302 // Check for precise loads from the primary supertype array and force them 1303 // to the supertype cache alias index. Check for generic array loads from 1304 // the primary supertype array and also force them to the supertype cache 1305 // alias index. Since the same load can reach both, we need to merge 1306 // these 2 disparate memories into the same alias class. Since the 1307 // primary supertype array is read-only, there's no chance of confusion 1308 // where we bypass an array load and an array store. 1309 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1310 if (offset == Type::OffsetBot || 1311 (offset >= primary_supers_offset && 1312 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1313 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1314 offset = in_bytes(Klass::secondary_super_cache_offset()); 1315 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1316 } 1317 } 1318 1319 // Flatten all Raw pointers together. 1320 if (tj->base() == Type::RawPtr) 1321 tj = TypeRawPtr::BOTTOM; 1322 1323 if (tj->base() == Type::AnyPtr) 1324 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1325 1326 // Flatten all to bottom for now 1327 switch( _AliasLevel ) { 1328 case 0: 1329 tj = TypePtr::BOTTOM; 1330 break; 1331 case 1: // Flatten to: oop, static, field or array 1332 switch (tj->base()) { 1333 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1334 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1335 case Type::AryPtr: // do not distinguish arrays at all 1336 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1337 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1338 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1339 default: ShouldNotReachHere(); 1340 } 1341 break; 1342 case 2: // No collapsing at level 2; keep all splits 1343 case 3: // No collapsing at level 3; keep all splits 1344 break; 1345 default: 1346 Unimplemented(); 1347 } 1348 1349 offset = tj->offset(); 1350 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1351 1352 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1353 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1354 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1355 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1356 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1357 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1358 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1359 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1360 assert( tj->ptr() != TypePtr::TopPTR && 1361 tj->ptr() != TypePtr::AnyNull && 1362 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1363 // assert( tj->ptr() != TypePtr::Constant || 1364 // tj->base() == Type::RawPtr || 1365 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1366 1367 return tj; 1368 } 1369 1370 void Compile::AliasType::Init(int i, const TypePtr* at) { 1371 _index = i; 1372 _adr_type = at; 1373 _field = NULL; 1374 _is_rewritable = true; // default 1375 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1376 if (atoop != NULL && atoop->is_known_instance()) { 1377 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1378 _general_index = Compile::current()->get_alias_index(gt); 1379 } else { 1380 _general_index = 0; 1381 } 1382 } 1383 1384 //---------------------------------print_on------------------------------------ 1385 #ifndef PRODUCT 1386 void Compile::AliasType::print_on(outputStream* st) { 1387 if (index() < 10) 1388 st->print("@ <%d> ", index()); 1389 else st->print("@ <%d>", index()); 1390 st->print(is_rewritable() ? " " : " RO"); 1391 int offset = adr_type()->offset(); 1392 if (offset == Type::OffsetBot) 1393 st->print(" +any"); 1394 else st->print(" +%-3d", offset); 1395 st->print(" in "); 1396 adr_type()->dump_on(st); 1397 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1398 if (field() != NULL && tjp) { 1399 if (tjp->klass() != field()->holder() || 1400 tjp->offset() != field()->offset_in_bytes()) { 1401 st->print(" != "); 1402 field()->print(); 1403 st->print(" ***"); 1404 } 1405 } 1406 } 1407 1408 void print_alias_types() { 1409 Compile* C = Compile::current(); 1410 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1411 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1412 C->alias_type(idx)->print_on(tty); 1413 tty->cr(); 1414 } 1415 } 1416 #endif 1417 1418 1419 //----------------------------probe_alias_cache-------------------------------- 1420 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1421 intptr_t key = (intptr_t) adr_type; 1422 key ^= key >> logAliasCacheSize; 1423 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1424 } 1425 1426 1427 //-----------------------------grow_alias_types-------------------------------- 1428 void Compile::grow_alias_types() { 1429 const int old_ats = _max_alias_types; // how many before? 1430 const int new_ats = old_ats; // how many more? 1431 const int grow_ats = old_ats+new_ats; // how many now? 1432 _max_alias_types = grow_ats; 1433 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1434 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1435 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1436 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1437 } 1438 1439 1440 //--------------------------------find_alias_type------------------------------ 1441 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1442 if (_AliasLevel == 0) 1443 return alias_type(AliasIdxBot); 1444 1445 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1446 if (ace->_adr_type == adr_type) { 1447 return alias_type(ace->_index); 1448 } 1449 1450 // Handle special cases. 1451 if (adr_type == NULL) return alias_type(AliasIdxTop); 1452 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1453 1454 // Do it the slow way. 1455 const TypePtr* flat = flatten_alias_type(adr_type); 1456 1457 #ifdef ASSERT 1458 assert(flat == flatten_alias_type(flat), "idempotent"); 1459 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1460 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1461 const TypeOopPtr* foop = flat->is_oopptr(); 1462 // Scalarizable allocations have exact klass always. 1463 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1464 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1465 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1466 } 1467 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1468 #endif 1469 1470 int idx = AliasIdxTop; 1471 for (int i = 0; i < num_alias_types(); i++) { 1472 if (alias_type(i)->adr_type() == flat) { 1473 idx = i; 1474 break; 1475 } 1476 } 1477 1478 if (idx == AliasIdxTop) { 1479 if (no_create) return NULL; 1480 // Grow the array if necessary. 1481 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1482 // Add a new alias type. 1483 idx = _num_alias_types++; 1484 _alias_types[idx]->Init(idx, flat); 1485 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1486 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1487 if (flat->isa_instptr()) { 1488 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1489 && flat->is_instptr()->klass() == env()->Class_klass()) 1490 alias_type(idx)->set_rewritable(false); 1491 } 1492 if (flat->isa_klassptr()) { 1493 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1494 alias_type(idx)->set_rewritable(false); 1495 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1496 alias_type(idx)->set_rewritable(false); 1497 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1498 alias_type(idx)->set_rewritable(false); 1499 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1500 alias_type(idx)->set_rewritable(false); 1501 } 1502 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1503 // but the base pointer type is not distinctive enough to identify 1504 // references into JavaThread.) 1505 1506 // Check for final fields. 1507 const TypeInstPtr* tinst = flat->isa_instptr(); 1508 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1509 ciField* field; 1510 if (tinst->const_oop() != NULL && 1511 tinst->klass() == ciEnv::current()->Class_klass() && 1512 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1513 // static field 1514 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1515 field = k->get_field_by_offset(tinst->offset(), true); 1516 } else { 1517 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1518 field = k->get_field_by_offset(tinst->offset(), false); 1519 } 1520 assert(field == NULL || 1521 original_field == NULL || 1522 (field->holder() == original_field->holder() && 1523 field->offset() == original_field->offset() && 1524 field->is_static() == original_field->is_static()), "wrong field?"); 1525 // Set field() and is_rewritable() attributes. 1526 if (field != NULL) alias_type(idx)->set_field(field); 1527 } 1528 } 1529 1530 // Fill the cache for next time. 1531 ace->_adr_type = adr_type; 1532 ace->_index = idx; 1533 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1534 1535 // Might as well try to fill the cache for the flattened version, too. 1536 AliasCacheEntry* face = probe_alias_cache(flat); 1537 if (face->_adr_type == NULL) { 1538 face->_adr_type = flat; 1539 face->_index = idx; 1540 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1541 } 1542 1543 return alias_type(idx); 1544 } 1545 1546 1547 Compile::AliasType* Compile::alias_type(ciField* field) { 1548 const TypeOopPtr* t; 1549 if (field->is_static()) 1550 t = TypeInstPtr::make(field->holder()->java_mirror()); 1551 else 1552 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1553 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1554 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1555 return atp; 1556 } 1557 1558 1559 //------------------------------have_alias_type-------------------------------- 1560 bool Compile::have_alias_type(const TypePtr* adr_type) { 1561 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1562 if (ace->_adr_type == adr_type) { 1563 return true; 1564 } 1565 1566 // Handle special cases. 1567 if (adr_type == NULL) return true; 1568 if (adr_type == TypePtr::BOTTOM) return true; 1569 1570 return find_alias_type(adr_type, true, NULL) != NULL; 1571 } 1572 1573 //-----------------------------must_alias-------------------------------------- 1574 // True if all values of the given address type are in the given alias category. 1575 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1576 if (alias_idx == AliasIdxBot) return true; // the universal category 1577 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1578 if (alias_idx == AliasIdxTop) return false; // the empty category 1579 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1580 1581 // the only remaining possible overlap is identity 1582 int adr_idx = get_alias_index(adr_type); 1583 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1584 assert(adr_idx == alias_idx || 1585 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1586 && adr_type != TypeOopPtr::BOTTOM), 1587 "should not be testing for overlap with an unsafe pointer"); 1588 return adr_idx == alias_idx; 1589 } 1590 1591 //------------------------------can_alias-------------------------------------- 1592 // True if any values of the given address type are in the given alias category. 1593 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1594 if (alias_idx == AliasIdxTop) return false; // the empty category 1595 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1596 if (alias_idx == AliasIdxBot) return true; // the universal category 1597 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1598 1599 // the only remaining possible overlap is identity 1600 int adr_idx = get_alias_index(adr_type); 1601 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1602 return adr_idx == alias_idx; 1603 } 1604 1605 1606 1607 //---------------------------pop_warm_call------------------------------------- 1608 WarmCallInfo* Compile::pop_warm_call() { 1609 WarmCallInfo* wci = _warm_calls; 1610 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1611 return wci; 1612 } 1613 1614 //----------------------------Inline_Warm-------------------------------------- 1615 int Compile::Inline_Warm() { 1616 // If there is room, try to inline some more warm call sites. 1617 // %%% Do a graph index compaction pass when we think we're out of space? 1618 if (!InlineWarmCalls) return 0; 1619 1620 int calls_made_hot = 0; 1621 int room_to_grow = NodeCountInliningCutoff - unique(); 1622 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1623 int amount_grown = 0; 1624 WarmCallInfo* call; 1625 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1626 int est_size = (int)call->size(); 1627 if (est_size > (room_to_grow - amount_grown)) { 1628 // This one won't fit anyway. Get rid of it. 1629 call->make_cold(); 1630 continue; 1631 } 1632 call->make_hot(); 1633 calls_made_hot++; 1634 amount_grown += est_size; 1635 amount_to_grow -= est_size; 1636 } 1637 1638 if (calls_made_hot > 0) set_major_progress(); 1639 return calls_made_hot; 1640 } 1641 1642 1643 //----------------------------Finish_Warm-------------------------------------- 1644 void Compile::Finish_Warm() { 1645 if (!InlineWarmCalls) return; 1646 if (failing()) return; 1647 if (warm_calls() == NULL) return; 1648 1649 // Clean up loose ends, if we are out of space for inlining. 1650 WarmCallInfo* call; 1651 while ((call = pop_warm_call()) != NULL) { 1652 call->make_cold(); 1653 } 1654 } 1655 1656 //---------------------cleanup_loop_predicates----------------------- 1657 // Remove the opaque nodes that protect the predicates so that all unused 1658 // checks and uncommon_traps will be eliminated from the ideal graph 1659 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1660 if (predicate_count()==0) return; 1661 for (int i = predicate_count(); i > 0; i--) { 1662 Node * n = predicate_opaque1_node(i-1); 1663 assert(n->Opcode() == Op_Opaque1, "must be"); 1664 igvn.replace_node(n, n->in(1)); 1665 } 1666 assert(predicate_count()==0, "should be clean!"); 1667 } 1668 1669 //------------------------------Optimize--------------------------------------- 1670 // Given a graph, optimize it. 1671 void Compile::Optimize() { 1672 TracePhase t1("optimizer", &_t_optimizer, true); 1673 1674 #ifndef PRODUCT 1675 if (env()->break_at_compile()) { 1676 BREAKPOINT; 1677 } 1678 1679 #endif 1680 1681 ResourceMark rm; 1682 int loop_opts_cnt; 1683 1684 NOT_PRODUCT( verify_graph_edges(); ) 1685 1686 print_method("After Parsing"); 1687 1688 { 1689 // Iterative Global Value Numbering, including ideal transforms 1690 // Initialize IterGVN with types and values from parse-time GVN 1691 PhaseIterGVN igvn(initial_gvn()); 1692 { 1693 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1694 igvn.optimize(); 1695 } 1696 1697 print_method("Iter GVN 1", 2); 1698 1699 if (failing()) return; 1700 1701 // Perform escape analysis 1702 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1703 if (has_loops()) { 1704 // Cleanup graph (remove dead nodes). 1705 TracePhase t2("idealLoop", &_t_idealLoop, true); 1706 PhaseIdealLoop ideal_loop( igvn, false, true ); 1707 if (major_progress()) print_method("PhaseIdealLoop before EA", 2); 1708 if (failing()) return; 1709 } 1710 ConnectionGraph::do_analysis(this, &igvn); 1711 1712 if (failing()) return; 1713 1714 // Optimize out fields loads from scalar replaceable allocations. 1715 igvn.optimize(); 1716 print_method("Iter GVN after EA", 2); 1717 1718 if (failing()) return; 1719 1720 if (congraph() != NULL && macro_count() > 0) { 1721 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 1722 PhaseMacroExpand mexp(igvn); 1723 mexp.eliminate_macro_nodes(); 1724 igvn.set_delay_transform(false); 1725 1726 igvn.optimize(); 1727 print_method("Iter GVN after eliminating allocations and locks", 2); 1728 1729 if (failing()) return; 1730 } 1731 } 1732 1733 // Loop transforms on the ideal graph. Range Check Elimination, 1734 // peeling, unrolling, etc. 1735 1736 // Set loop opts counter 1737 loop_opts_cnt = num_loop_opts(); 1738 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1739 { 1740 TracePhase t2("idealLoop", &_t_idealLoop, true); 1741 PhaseIdealLoop ideal_loop( igvn, true ); 1742 loop_opts_cnt--; 1743 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1744 if (failing()) return; 1745 } 1746 // Loop opts pass if partial peeling occurred in previous pass 1747 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1748 TracePhase t3("idealLoop", &_t_idealLoop, true); 1749 PhaseIdealLoop ideal_loop( igvn, false ); 1750 loop_opts_cnt--; 1751 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1752 if (failing()) return; 1753 } 1754 // Loop opts pass for loop-unrolling before CCP 1755 if(major_progress() && (loop_opts_cnt > 0)) { 1756 TracePhase t4("idealLoop", &_t_idealLoop, true); 1757 PhaseIdealLoop ideal_loop( igvn, false ); 1758 loop_opts_cnt--; 1759 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1760 } 1761 if (!failing()) { 1762 // Verify that last round of loop opts produced a valid graph 1763 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1764 PhaseIdealLoop::verify(igvn); 1765 } 1766 } 1767 if (failing()) return; 1768 1769 // Conditional Constant Propagation; 1770 PhaseCCP ccp( &igvn ); 1771 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1772 { 1773 TracePhase t2("ccp", &_t_ccp, true); 1774 ccp.do_transform(); 1775 } 1776 print_method("PhaseCPP 1", 2); 1777 1778 assert( true, "Break here to ccp.dump_old2new_map()"); 1779 1780 // Iterative Global Value Numbering, including ideal transforms 1781 { 1782 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1783 igvn = ccp; 1784 igvn.optimize(); 1785 } 1786 1787 print_method("Iter GVN 2", 2); 1788 1789 if (failing()) return; 1790 1791 // Loop transforms on the ideal graph. Range Check Elimination, 1792 // peeling, unrolling, etc. 1793 if(loop_opts_cnt > 0) { 1794 debug_only( int cnt = 0; ); 1795 while(major_progress() && (loop_opts_cnt > 0)) { 1796 TracePhase t2("idealLoop", &_t_idealLoop, true); 1797 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1798 PhaseIdealLoop ideal_loop( igvn, true); 1799 loop_opts_cnt--; 1800 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1801 if (failing()) return; 1802 } 1803 } 1804 1805 { 1806 // Verify that all previous optimizations produced a valid graph 1807 // at least to this point, even if no loop optimizations were done. 1808 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1809 PhaseIdealLoop::verify(igvn); 1810 } 1811 1812 { 1813 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1814 PhaseMacroExpand mex(igvn); 1815 if (mex.expand_macro_nodes()) { 1816 assert(failing(), "must bail out w/ explicit message"); 1817 return; 1818 } 1819 } 1820 1821 } // (End scope of igvn; run destructor if necessary for asserts.) 1822 1823 // A method with only infinite loops has no edges entering loops from root 1824 { 1825 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1826 if (final_graph_reshaping()) { 1827 assert(failing(), "must bail out w/ explicit message"); 1828 return; 1829 } 1830 } 1831 1832 print_method("Optimize finished", 2); 1833 } 1834 1835 1836 //------------------------------Code_Gen--------------------------------------- 1837 // Given a graph, generate code for it 1838 void Compile::Code_Gen() { 1839 if (failing()) return; 1840 1841 // Perform instruction selection. You might think we could reclaim Matcher 1842 // memory PDQ, but actually the Matcher is used in generating spill code. 1843 // Internals of the Matcher (including some VectorSets) must remain live 1844 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1845 // set a bit in reclaimed memory. 1846 1847 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1848 // nodes. Mapping is only valid at the root of each matched subtree. 1849 NOT_PRODUCT( verify_graph_edges(); ) 1850 1851 Node_List proj_list; 1852 Matcher m(proj_list); 1853 _matcher = &m; 1854 { 1855 TracePhase t2("matcher", &_t_matcher, true); 1856 m.match(); 1857 } 1858 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1859 // nodes. Mapping is only valid at the root of each matched subtree. 1860 NOT_PRODUCT( verify_graph_edges(); ) 1861 1862 // If you have too many nodes, or if matching has failed, bail out 1863 check_node_count(0, "out of nodes matching instructions"); 1864 if (failing()) return; 1865 1866 // Build a proper-looking CFG 1867 PhaseCFG cfg(node_arena(), root(), m); 1868 _cfg = &cfg; 1869 { 1870 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1871 cfg.Dominators(); 1872 if (failing()) return; 1873 1874 NOT_PRODUCT( verify_graph_edges(); ) 1875 1876 cfg.Estimate_Block_Frequency(); 1877 cfg.GlobalCodeMotion(m,unique(),proj_list); 1878 if (failing()) return; 1879 1880 print_method("Global code motion", 2); 1881 1882 NOT_PRODUCT( verify_graph_edges(); ) 1883 1884 debug_only( cfg.verify(); ) 1885 } 1886 NOT_PRODUCT( verify_graph_edges(); ) 1887 1888 PhaseChaitin regalloc(unique(),cfg,m); 1889 _regalloc = ®alloc; 1890 { 1891 TracePhase t2("regalloc", &_t_registerAllocation, true); 1892 // Perform any platform dependent preallocation actions. This is used, 1893 // for example, to avoid taking an implicit null pointer exception 1894 // using the frame pointer on win95. 1895 _regalloc->pd_preallocate_hook(); 1896 1897 // Perform register allocation. After Chaitin, use-def chains are 1898 // no longer accurate (at spill code) and so must be ignored. 1899 // Node->LRG->reg mappings are still accurate. 1900 _regalloc->Register_Allocate(); 1901 1902 // Bail out if the allocator builds too many nodes 1903 if (failing()) return; 1904 } 1905 1906 // Prior to register allocation we kept empty basic blocks in case the 1907 // the allocator needed a place to spill. After register allocation we 1908 // are not adding any new instructions. If any basic block is empty, we 1909 // can now safely remove it. 1910 { 1911 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1912 cfg.remove_empty(); 1913 if (do_freq_based_layout()) { 1914 PhaseBlockLayout layout(cfg); 1915 } else { 1916 cfg.set_loop_alignment(); 1917 } 1918 cfg.fixup_flow(); 1919 } 1920 1921 // Perform any platform dependent postallocation verifications. 1922 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1923 1924 // Apply peephole optimizations 1925 if( OptoPeephole ) { 1926 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1927 PhasePeephole peep( _regalloc, cfg); 1928 peep.do_transform(); 1929 } 1930 1931 // Convert Nodes to instruction bits in a buffer 1932 { 1933 // %%%% workspace merge brought two timers together for one job 1934 TracePhase t2a("output", &_t_output, true); 1935 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1936 Output(); 1937 } 1938 1939 print_method("Final Code"); 1940 1941 // He's dead, Jim. 1942 _cfg = (PhaseCFG*)0xdeadbeef; 1943 _regalloc = (PhaseChaitin*)0xdeadbeef; 1944 } 1945 1946 1947 //------------------------------dump_asm--------------------------------------- 1948 // Dump formatted assembly 1949 #ifndef PRODUCT 1950 void Compile::dump_asm(int *pcs, uint pc_limit) { 1951 bool cut_short = false; 1952 tty->print_cr("#"); 1953 tty->print("# "); _tf->dump(); tty->cr(); 1954 tty->print_cr("#"); 1955 1956 // For all blocks 1957 int pc = 0x0; // Program counter 1958 char starts_bundle = ' '; 1959 _regalloc->dump_frame(); 1960 1961 Node *n = NULL; 1962 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1963 if (VMThread::should_terminate()) { cut_short = true; break; } 1964 Block *b = _cfg->_blocks[i]; 1965 if (b->is_connector() && !Verbose) continue; 1966 n = b->_nodes[0]; 1967 if (pcs && n->_idx < pc_limit) 1968 tty->print("%3.3x ", pcs[n->_idx]); 1969 else 1970 tty->print(" "); 1971 b->dump_head( &_cfg->_bbs ); 1972 if (b->is_connector()) { 1973 tty->print_cr(" # Empty connector block"); 1974 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1975 tty->print_cr(" # Block is sole successor of call"); 1976 } 1977 1978 // For all instructions 1979 Node *delay = NULL; 1980 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1981 if (VMThread::should_terminate()) { cut_short = true; break; } 1982 n = b->_nodes[j]; 1983 if (valid_bundle_info(n)) { 1984 Bundle *bundle = node_bundling(n); 1985 if (bundle->used_in_unconditional_delay()) { 1986 delay = n; 1987 continue; 1988 } 1989 if (bundle->starts_bundle()) 1990 starts_bundle = '+'; 1991 } 1992 1993 if (WizardMode) n->dump(); 1994 1995 if( !n->is_Region() && // Dont print in the Assembly 1996 !n->is_Phi() && // a few noisely useless nodes 1997 !n->is_Proj() && 1998 !n->is_MachTemp() && 1999 !n->is_SafePointScalarObject() && 2000 !n->is_Catch() && // Would be nice to print exception table targets 2001 !n->is_MergeMem() && // Not very interesting 2002 !n->is_top() && // Debug info table constants 2003 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2004 ) { 2005 if (pcs && n->_idx < pc_limit) 2006 tty->print("%3.3x", pcs[n->_idx]); 2007 else 2008 tty->print(" "); 2009 tty->print(" %c ", starts_bundle); 2010 starts_bundle = ' '; 2011 tty->print("\t"); 2012 n->format(_regalloc, tty); 2013 tty->cr(); 2014 } 2015 2016 // If we have an instruction with a delay slot, and have seen a delay, 2017 // then back up and print it 2018 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2019 assert(delay != NULL, "no unconditional delay instruction"); 2020 if (WizardMode) delay->dump(); 2021 2022 if (node_bundling(delay)->starts_bundle()) 2023 starts_bundle = '+'; 2024 if (pcs && n->_idx < pc_limit) 2025 tty->print("%3.3x", pcs[n->_idx]); 2026 else 2027 tty->print(" "); 2028 tty->print(" %c ", starts_bundle); 2029 starts_bundle = ' '; 2030 tty->print("\t"); 2031 delay->format(_regalloc, tty); 2032 tty->print_cr(""); 2033 delay = NULL; 2034 } 2035 2036 // Dump the exception table as well 2037 if( n->is_Catch() && (Verbose || WizardMode) ) { 2038 // Print the exception table for this offset 2039 _handler_table.print_subtable_for(pc); 2040 } 2041 } 2042 2043 if (pcs && n->_idx < pc_limit) 2044 tty->print_cr("%3.3x", pcs[n->_idx]); 2045 else 2046 tty->print_cr(""); 2047 2048 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2049 2050 } // End of per-block dump 2051 tty->print_cr(""); 2052 2053 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2054 } 2055 #endif 2056 2057 //------------------------------Final_Reshape_Counts--------------------------- 2058 // This class defines counters to help identify when a method 2059 // may/must be executed using hardware with only 24-bit precision. 2060 struct Final_Reshape_Counts : public StackObj { 2061 int _call_count; // count non-inlined 'common' calls 2062 int _float_count; // count float ops requiring 24-bit precision 2063 int _double_count; // count double ops requiring more precision 2064 int _java_call_count; // count non-inlined 'java' calls 2065 int _inner_loop_count; // count loops which need alignment 2066 VectorSet _visited; // Visitation flags 2067 Node_List _tests; // Set of IfNodes & PCTableNodes 2068 2069 Final_Reshape_Counts() : 2070 _call_count(0), _float_count(0), _double_count(0), 2071 _java_call_count(0), _inner_loop_count(0), 2072 _visited( Thread::current()->resource_area() ) { } 2073 2074 void inc_call_count () { _call_count ++; } 2075 void inc_float_count () { _float_count ++; } 2076 void inc_double_count() { _double_count++; } 2077 void inc_java_call_count() { _java_call_count++; } 2078 void inc_inner_loop_count() { _inner_loop_count++; } 2079 2080 int get_call_count () const { return _call_count ; } 2081 int get_float_count () const { return _float_count ; } 2082 int get_double_count() const { return _double_count; } 2083 int get_java_call_count() const { return _java_call_count; } 2084 int get_inner_loop_count() const { return _inner_loop_count; } 2085 }; 2086 2087 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2088 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2089 // Make sure the offset goes inside the instance layout. 2090 return k->contains_field_offset(tp->offset()); 2091 // Note that OffsetBot and OffsetTop are very negative. 2092 } 2093 2094 // Eliminate trivially redundant StoreCMs and accumulate their 2095 // precedence edges. 2096 static void eliminate_redundant_card_marks(Node* n) { 2097 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2098 if (n->in(MemNode::Address)->outcnt() > 1) { 2099 // There are multiple users of the same address so it might be 2100 // possible to eliminate some of the StoreCMs 2101 Node* mem = n->in(MemNode::Memory); 2102 Node* adr = n->in(MemNode::Address); 2103 Node* val = n->in(MemNode::ValueIn); 2104 Node* prev = n; 2105 bool done = false; 2106 // Walk the chain of StoreCMs eliminating ones that match. As 2107 // long as it's a chain of single users then the optimization is 2108 // safe. Eliminating partially redundant StoreCMs would require 2109 // cloning copies down the other paths. 2110 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2111 if (adr == mem->in(MemNode::Address) && 2112 val == mem->in(MemNode::ValueIn)) { 2113 // redundant StoreCM 2114 if (mem->req() > MemNode::OopStore) { 2115 // Hasn't been processed by this code yet. 2116 n->add_prec(mem->in(MemNode::OopStore)); 2117 } else { 2118 // Already converted to precedence edge 2119 for (uint i = mem->req(); i < mem->len(); i++) { 2120 // Accumulate any precedence edges 2121 if (mem->in(i) != NULL) { 2122 n->add_prec(mem->in(i)); 2123 } 2124 } 2125 // Everything above this point has been processed. 2126 done = true; 2127 } 2128 // Eliminate the previous StoreCM 2129 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2130 assert(mem->outcnt() == 0, "should be dead"); 2131 mem->disconnect_inputs(NULL); 2132 } else { 2133 prev = mem; 2134 } 2135 mem = prev->in(MemNode::Memory); 2136 } 2137 } 2138 } 2139 2140 //------------------------------final_graph_reshaping_impl---------------------- 2141 // Implement items 1-5 from final_graph_reshaping below. 2142 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 2143 2144 if ( n->outcnt() == 0 ) return; // dead node 2145 uint nop = n->Opcode(); 2146 2147 // Check for 2-input instruction with "last use" on right input. 2148 // Swap to left input. Implements item (2). 2149 if( n->req() == 3 && // two-input instruction 2150 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2151 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2152 n->in(2)->outcnt() == 1 &&// right use IS a last use 2153 !n->in(2)->is_Con() ) { // right use is not a constant 2154 // Check for commutative opcode 2155 switch( nop ) { 2156 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2157 case Op_MaxI: case Op_MinI: 2158 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2159 case Op_AndL: case Op_XorL: case Op_OrL: 2160 case Op_AndI: case Op_XorI: case Op_OrI: { 2161 // Move "last use" input to left by swapping inputs 2162 n->swap_edges(1, 2); 2163 break; 2164 } 2165 default: 2166 break; 2167 } 2168 } 2169 2170 #ifdef ASSERT 2171 if( n->is_Mem() ) { 2172 Compile* C = Compile::current(); 2173 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2174 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2175 // oop will be recorded in oop map if load crosses safepoint 2176 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2177 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2178 "raw memory operations should have control edge"); 2179 } 2180 #endif 2181 // Count FPU ops and common calls, implements item (3) 2182 switch( nop ) { 2183 // Count all float operations that may use FPU 2184 case Op_AddF: 2185 case Op_SubF: 2186 case Op_MulF: 2187 case Op_DivF: 2188 case Op_NegF: 2189 case Op_ModF: 2190 case Op_ConvI2F: 2191 case Op_ConF: 2192 case Op_CmpF: 2193 case Op_CmpF3: 2194 // case Op_ConvL2F: // longs are split into 32-bit halves 2195 frc.inc_float_count(); 2196 break; 2197 2198 case Op_ConvF2D: 2199 case Op_ConvD2F: 2200 frc.inc_float_count(); 2201 frc.inc_double_count(); 2202 break; 2203 2204 // Count all double operations that may use FPU 2205 case Op_AddD: 2206 case Op_SubD: 2207 case Op_MulD: 2208 case Op_DivD: 2209 case Op_NegD: 2210 case Op_ModD: 2211 case Op_ConvI2D: 2212 case Op_ConvD2I: 2213 // case Op_ConvL2D: // handled by leaf call 2214 // case Op_ConvD2L: // handled by leaf call 2215 case Op_ConD: 2216 case Op_CmpD: 2217 case Op_CmpD3: 2218 frc.inc_double_count(); 2219 break; 2220 case Op_Opaque1: // Remove Opaque Nodes before matching 2221 case Op_Opaque2: // Remove Opaque Nodes before matching 2222 n->subsume_by(n->in(1)); 2223 break; 2224 case Op_CallStaticJava: 2225 case Op_CallJava: 2226 case Op_CallDynamicJava: 2227 frc.inc_java_call_count(); // Count java call site; 2228 case Op_CallRuntime: 2229 case Op_CallLeaf: 2230 case Op_CallLeafNoFP: { 2231 assert( n->is_Call(), "" ); 2232 CallNode *call = n->as_Call(); 2233 // Count call sites where the FP mode bit would have to be flipped. 2234 // Do not count uncommon runtime calls: 2235 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2236 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2237 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2238 frc.inc_call_count(); // Count the call site 2239 } else { // See if uncommon argument is shared 2240 Node *n = call->in(TypeFunc::Parms); 2241 int nop = n->Opcode(); 2242 // Clone shared simple arguments to uncommon calls, item (1). 2243 if( n->outcnt() > 1 && 2244 !n->is_Proj() && 2245 nop != Op_CreateEx && 2246 nop != Op_CheckCastPP && 2247 nop != Op_DecodeN && 2248 !n->is_Mem() ) { 2249 Node *x = n->clone(); 2250 call->set_req( TypeFunc::Parms, x ); 2251 } 2252 } 2253 break; 2254 } 2255 2256 case Op_StoreD: 2257 case Op_LoadD: 2258 case Op_LoadD_unaligned: 2259 frc.inc_double_count(); 2260 goto handle_mem; 2261 case Op_StoreF: 2262 case Op_LoadF: 2263 frc.inc_float_count(); 2264 goto handle_mem; 2265 2266 case Op_StoreCM: 2267 { 2268 // Convert OopStore dependence into precedence edge 2269 Node* prec = n->in(MemNode::OopStore); 2270 n->del_req(MemNode::OopStore); 2271 n->add_prec(prec); 2272 eliminate_redundant_card_marks(n); 2273 } 2274 2275 // fall through 2276 2277 case Op_StoreB: 2278 case Op_StoreC: 2279 case Op_StorePConditional: 2280 case Op_StoreI: 2281 case Op_StoreL: 2282 case Op_StoreIConditional: 2283 case Op_StoreLConditional: 2284 case Op_CompareAndSwapI: 2285 case Op_CompareAndSwapL: 2286 case Op_CompareAndSwapP: 2287 case Op_CompareAndSwapN: 2288 case Op_StoreP: 2289 case Op_StoreN: 2290 case Op_LoadB: 2291 case Op_LoadUB: 2292 case Op_LoadUS: 2293 case Op_LoadI: 2294 case Op_LoadUI2L: 2295 case Op_LoadKlass: 2296 case Op_LoadNKlass: 2297 case Op_LoadL: 2298 case Op_LoadL_unaligned: 2299 case Op_LoadPLocked: 2300 case Op_LoadP: 2301 case Op_LoadN: 2302 case Op_LoadRange: 2303 case Op_LoadS: { 2304 handle_mem: 2305 #ifdef ASSERT 2306 if( VerifyOptoOopOffsets ) { 2307 assert( n->is_Mem(), "" ); 2308 MemNode *mem = (MemNode*)n; 2309 // Check to see if address types have grounded out somehow. 2310 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2311 assert( !tp || oop_offset_is_sane(tp), "" ); 2312 } 2313 #endif 2314 break; 2315 } 2316 2317 case Op_AddP: { // Assert sane base pointers 2318 Node *addp = n->in(AddPNode::Address); 2319 assert( !addp->is_AddP() || 2320 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2321 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2322 "Base pointers must match" ); 2323 #ifdef _LP64 2324 if (UseCompressedOops && 2325 addp->Opcode() == Op_ConP && 2326 addp == n->in(AddPNode::Base) && 2327 n->in(AddPNode::Offset)->is_Con()) { 2328 // Use addressing with narrow klass to load with offset on x86. 2329 // On sparc loading 32-bits constant and decoding it have less 2330 // instructions (4) then load 64-bits constant (7). 2331 // Do this transformation here since IGVN will convert ConN back to ConP. 2332 const Type* t = addp->bottom_type(); 2333 if (t->isa_oopptr()) { 2334 Node* nn = NULL; 2335 2336 // Look for existing ConN node of the same exact type. 2337 Compile* C = Compile::current(); 2338 Node* r = C->root(); 2339 uint cnt = r->outcnt(); 2340 for (uint i = 0; i < cnt; i++) { 2341 Node* m = r->raw_out(i); 2342 if (m!= NULL && m->Opcode() == Op_ConN && 2343 m->bottom_type()->make_ptr() == t) { 2344 nn = m; 2345 break; 2346 } 2347 } 2348 if (nn != NULL) { 2349 // Decode a narrow oop to match address 2350 // [R12 + narrow_oop_reg<<3 + offset] 2351 nn = new (C, 2) DecodeNNode(nn, t); 2352 n->set_req(AddPNode::Base, nn); 2353 n->set_req(AddPNode::Address, nn); 2354 if (addp->outcnt() == 0) { 2355 addp->disconnect_inputs(NULL); 2356 } 2357 } 2358 } 2359 } 2360 #endif 2361 break; 2362 } 2363 2364 #ifdef _LP64 2365 case Op_CastPP: 2366 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2367 Compile* C = Compile::current(); 2368 Node* in1 = n->in(1); 2369 const Type* t = n->bottom_type(); 2370 Node* new_in1 = in1->clone(); 2371 new_in1->as_DecodeN()->set_type(t); 2372 2373 if (!Matcher::narrow_oop_use_complex_address()) { 2374 // 2375 // x86, ARM and friends can handle 2 adds in addressing mode 2376 // and Matcher can fold a DecodeN node into address by using 2377 // a narrow oop directly and do implicit NULL check in address: 2378 // 2379 // [R12 + narrow_oop_reg<<3 + offset] 2380 // NullCheck narrow_oop_reg 2381 // 2382 // On other platforms (Sparc) we have to keep new DecodeN node and 2383 // use it to do implicit NULL check in address: 2384 // 2385 // decode_not_null narrow_oop_reg, base_reg 2386 // [base_reg + offset] 2387 // NullCheck base_reg 2388 // 2389 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2390 // to keep the information to which NULL check the new DecodeN node 2391 // corresponds to use it as value in implicit_null_check(). 2392 // 2393 new_in1->set_req(0, n->in(0)); 2394 } 2395 2396 n->subsume_by(new_in1); 2397 if (in1->outcnt() == 0) { 2398 in1->disconnect_inputs(NULL); 2399 } 2400 } 2401 break; 2402 2403 case Op_CmpP: 2404 // Do this transformation here to preserve CmpPNode::sub() and 2405 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2406 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) { 2407 Node* in1 = n->in(1); 2408 Node* in2 = n->in(2); 2409 if (!in1->is_DecodeN()) { 2410 in2 = in1; 2411 in1 = n->in(2); 2412 } 2413 assert(in1->is_DecodeN(), "sanity"); 2414 2415 Compile* C = Compile::current(); 2416 Node* new_in2 = NULL; 2417 if (in2->is_DecodeN()) { 2418 new_in2 = in2->in(1); 2419 } else if (in2->Opcode() == Op_ConP) { 2420 const Type* t = in2->bottom_type(); 2421 if (t == TypePtr::NULL_PTR) { 2422 // Don't convert CmpP null check into CmpN if compressed 2423 // oops implicit null check is not generated. 2424 // This will allow to generate normal oop implicit null check. 2425 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2426 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2427 // 2428 // This transformation together with CastPP transformation above 2429 // will generated code for implicit NULL checks for compressed oops. 2430 // 2431 // The original code after Optimize() 2432 // 2433 // LoadN memory, narrow_oop_reg 2434 // decode narrow_oop_reg, base_reg 2435 // CmpP base_reg, NULL 2436 // CastPP base_reg // NotNull 2437 // Load [base_reg + offset], val_reg 2438 // 2439 // after these transformations will be 2440 // 2441 // LoadN memory, narrow_oop_reg 2442 // CmpN narrow_oop_reg, NULL 2443 // decode_not_null narrow_oop_reg, base_reg 2444 // Load [base_reg + offset], val_reg 2445 // 2446 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2447 // since narrow oops can be used in debug info now (see the code in 2448 // final_graph_reshaping_walk()). 2449 // 2450 // At the end the code will be matched to 2451 // on x86: 2452 // 2453 // Load_narrow_oop memory, narrow_oop_reg 2454 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2455 // NullCheck narrow_oop_reg 2456 // 2457 // and on sparc: 2458 // 2459 // Load_narrow_oop memory, narrow_oop_reg 2460 // decode_not_null narrow_oop_reg, base_reg 2461 // Load [base_reg + offset], val_reg 2462 // NullCheck base_reg 2463 // 2464 } else if (t->isa_oopptr()) { 2465 new_in2 = ConNode::make(C, t->make_narrowoop()); 2466 } 2467 } 2468 if (new_in2 != NULL) { 2469 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2); 2470 n->subsume_by( cmpN ); 2471 if (in1->outcnt() == 0) { 2472 in1->disconnect_inputs(NULL); 2473 } 2474 if (in2->outcnt() == 0) { 2475 in2->disconnect_inputs(NULL); 2476 } 2477 } 2478 } 2479 break; 2480 2481 case Op_DecodeN: 2482 assert(!n->in(1)->is_EncodeP(), "should be optimized out"); 2483 // DecodeN could be pinned when it can't be fold into 2484 // an address expression, see the code for Op_CastPP above. 2485 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control"); 2486 break; 2487 2488 case Op_EncodeP: { 2489 Node* in1 = n->in(1); 2490 if (in1->is_DecodeN()) { 2491 n->subsume_by(in1->in(1)); 2492 } else if (in1->Opcode() == Op_ConP) { 2493 Compile* C = Compile::current(); 2494 const Type* t = in1->bottom_type(); 2495 if (t == TypePtr::NULL_PTR) { 2496 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2497 } else if (t->isa_oopptr()) { 2498 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2499 } 2500 } 2501 if (in1->outcnt() == 0) { 2502 in1->disconnect_inputs(NULL); 2503 } 2504 break; 2505 } 2506 2507 case Op_Proj: { 2508 if (OptimizeStringConcat) { 2509 ProjNode* p = n->as_Proj(); 2510 if (p->_is_io_use) { 2511 // Separate projections were used for the exception path which 2512 // are normally removed by a late inline. If it wasn't inlined 2513 // then they will hang around and should just be replaced with 2514 // the original one. 2515 Node* proj = NULL; 2516 // Replace with just one 2517 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2518 Node *use = i.get(); 2519 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2520 proj = use; 2521 break; 2522 } 2523 } 2524 assert(proj != NULL, "must be found"); 2525 p->subsume_by(proj); 2526 } 2527 } 2528 break; 2529 } 2530 2531 case Op_Phi: 2532 if (n->as_Phi()->bottom_type()->isa_narrowoop()) { 2533 // The EncodeP optimization may create Phi with the same edges 2534 // for all paths. It is not handled well by Register Allocator. 2535 Node* unique_in = n->in(1); 2536 assert(unique_in != NULL, ""); 2537 uint cnt = n->req(); 2538 for (uint i = 2; i < cnt; i++) { 2539 Node* m = n->in(i); 2540 assert(m != NULL, ""); 2541 if (unique_in != m) 2542 unique_in = NULL; 2543 } 2544 if (unique_in != NULL) { 2545 n->subsume_by(unique_in); 2546 } 2547 } 2548 break; 2549 2550 #endif 2551 2552 case Op_ModI: 2553 if (UseDivMod) { 2554 // Check if a%b and a/b both exist 2555 Node* d = n->find_similar(Op_DivI); 2556 if (d) { 2557 // Replace them with a fused divmod if supported 2558 Compile* C = Compile::current(); 2559 if (Matcher::has_match_rule(Op_DivModI)) { 2560 DivModINode* divmod = DivModINode::make(C, n); 2561 d->subsume_by(divmod->div_proj()); 2562 n->subsume_by(divmod->mod_proj()); 2563 } else { 2564 // replace a%b with a-((a/b)*b) 2565 Node* mult = new (C, 3) MulINode(d, d->in(2)); 2566 Node* sub = new (C, 3) SubINode(d->in(1), mult); 2567 n->subsume_by( sub ); 2568 } 2569 } 2570 } 2571 break; 2572 2573 case Op_ModL: 2574 if (UseDivMod) { 2575 // Check if a%b and a/b both exist 2576 Node* d = n->find_similar(Op_DivL); 2577 if (d) { 2578 // Replace them with a fused divmod if supported 2579 Compile* C = Compile::current(); 2580 if (Matcher::has_match_rule(Op_DivModL)) { 2581 DivModLNode* divmod = DivModLNode::make(C, n); 2582 d->subsume_by(divmod->div_proj()); 2583 n->subsume_by(divmod->mod_proj()); 2584 } else { 2585 // replace a%b with a-((a/b)*b) 2586 Node* mult = new (C, 3) MulLNode(d, d->in(2)); 2587 Node* sub = new (C, 3) SubLNode(d->in(1), mult); 2588 n->subsume_by( sub ); 2589 } 2590 } 2591 } 2592 break; 2593 2594 case Op_LoadVector: 2595 case Op_StoreVector: 2596 break; 2597 2598 case Op_PackB: 2599 case Op_PackS: 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 static int type_to_size_in_bytes(BasicType t) { 3041 switch (t) { 3042 case T_LONG: return sizeof(jlong ); 3043 case T_FLOAT: return sizeof(jfloat ); 3044 case T_DOUBLE: return sizeof(jdouble); 3045 // We use T_VOID as marker for jump-table entries (labels) which 3046 // need an internal word relocation. 3047 case T_VOID: 3048 case T_ADDRESS: 3049 case T_OBJECT: return sizeof(jobject); 3050 } 3051 3052 ShouldNotReachHere(); 3053 return -1; 3054 } 3055 3056 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3057 // sort descending 3058 if (a->freq() > b->freq()) return -1; 3059 if (a->freq() < b->freq()) return 1; 3060 return 0; 3061 } 3062 3063 void Compile::ConstantTable::calculate_offsets_and_size() { 3064 // First, sort the array by frequencies. 3065 _constants.sort(qsort_comparator); 3066 3067 #ifdef ASSERT 3068 // Make sure all jump-table entries were sorted to the end of the 3069 // array (they have a negative frequency). 3070 bool found_void = false; 3071 for (int i = 0; i < _constants.length(); i++) { 3072 Constant con = _constants.at(i); 3073 if (con.type() == T_VOID) 3074 found_void = true; // jump-tables 3075 else 3076 assert(!found_void, "wrong sorting"); 3077 } 3078 #endif 3079 3080 int offset = 0; 3081 for (int i = 0; i < _constants.length(); i++) { 3082 Constant* con = _constants.adr_at(i); 3083 3084 // Align offset for type. 3085 int typesize = type_to_size_in_bytes(con->type()); 3086 offset = align_size_up(offset, typesize); 3087 con->set_offset(offset); // set constant's offset 3088 3089 if (con->type() == T_VOID) { 3090 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3091 offset = offset + typesize * n->outcnt(); // expand jump-table 3092 } else { 3093 offset = offset + typesize; 3094 } 3095 } 3096 3097 // Align size up to the next section start (which is insts; see 3098 // CodeBuffer::align_at_start). 3099 assert(_size == -1, "already set?"); 3100 _size = align_size_up(offset, CodeEntryAlignment); 3101 } 3102 3103 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3104 MacroAssembler _masm(&cb); 3105 for (int i = 0; i < _constants.length(); i++) { 3106 Constant con = _constants.at(i); 3107 address constant_addr; 3108 switch (con.type()) { 3109 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3110 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3111 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3112 case T_OBJECT: { 3113 jobject obj = con.get_jobject(); 3114 int oop_index = _masm.oop_recorder()->find_index(obj); 3115 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3116 break; 3117 } 3118 case T_ADDRESS: { 3119 address addr = (address) con.get_jobject(); 3120 constant_addr = _masm.address_constant(addr); 3121 break; 3122 } 3123 // We use T_VOID as marker for jump-table entries (labels) which 3124 // need an internal word relocation. 3125 case T_VOID: { 3126 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3127 // Fill the jump-table with a dummy word. The real value is 3128 // filled in later in fill_jump_table. 3129 address dummy = (address) n; 3130 constant_addr = _masm.address_constant(dummy); 3131 // Expand jump-table 3132 for (uint i = 1; i < n->outcnt(); i++) { 3133 address temp_addr = _masm.address_constant(dummy + i); 3134 assert(temp_addr, "consts section too small"); 3135 } 3136 break; 3137 } 3138 default: ShouldNotReachHere(); 3139 } 3140 assert(constant_addr, "consts section too small"); 3141 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3142 } 3143 } 3144 3145 int Compile::ConstantTable::find_offset(Constant& con) const { 3146 int idx = _constants.find(con); 3147 assert(idx != -1, "constant must be in constant table"); 3148 int offset = _constants.at(idx).offset(); 3149 assert(offset != -1, "constant table not emitted yet?"); 3150 return offset; 3151 } 3152 3153 void Compile::ConstantTable::add(Constant& con) { 3154 if (con.can_be_reused()) { 3155 int idx = _constants.find(con); 3156 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3157 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3158 return; 3159 } 3160 } 3161 (void) _constants.append(con); 3162 } 3163 3164 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3165 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3166 Constant con(type, value, b->_freq); 3167 add(con); 3168 return con; 3169 } 3170 3171 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3172 jvalue value; 3173 BasicType type = oper->type()->basic_type(); 3174 switch (type) { 3175 case T_LONG: value.j = oper->constantL(); break; 3176 case T_FLOAT: value.f = oper->constantF(); break; 3177 case T_DOUBLE: value.d = oper->constantD(); break; 3178 case T_OBJECT: 3179 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3180 default: ShouldNotReachHere(); 3181 } 3182 return add(n, type, value); 3183 } 3184 3185 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3186 jvalue value; 3187 // We can use the node pointer here to identify the right jump-table 3188 // as this method is called from Compile::Fill_buffer right before 3189 // the MachNodes are emitted and the jump-table is filled (means the 3190 // MachNode pointers do not change anymore). 3191 value.l = (jobject) n; 3192 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3193 add(con); 3194 return con; 3195 } 3196 3197 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3198 // If called from Compile::scratch_emit_size do nothing. 3199 if (Compile::current()->in_scratch_emit_size()) return; 3200 3201 assert(labels.is_nonempty(), "must be"); 3202 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt())); 3203 3204 // Since MachConstantNode::constant_offset() also contains 3205 // table_base_offset() we need to subtract the table_base_offset() 3206 // to get the plain offset into the constant table. 3207 int offset = n->constant_offset() - table_base_offset(); 3208 3209 MacroAssembler _masm(&cb); 3210 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3211 3212 for (uint i = 0; i < n->outcnt(); i++) { 3213 address* constant_addr = &jump_table_base[i]; 3214 assert(*constant_addr == (((address) n) + i), err_msg("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); 3215 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3216 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3217 } 3218 }