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) 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) 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 SharedRuntime::is_wide_vector(max_vector_size()) 830 ); 831 832 if (log() != NULL) // Print code cache state into compiler log 833 log()->code_cache_state(); 834 } 835 } 836 837 //------------------------------Compile---------------------------------------- 838 // Compile a runtime stub 839 Compile::Compile( ciEnv* ci_env, 840 TypeFunc_generator generator, 841 address stub_function, 842 const char *stub_name, 843 int is_fancy_jump, 844 bool pass_tls, 845 bool save_arg_registers, 846 bool return_pc ) 847 : Phase(Compiler), 848 _env(ci_env), 849 _log(ci_env->log()), 850 _compile_id(-1), 851 _save_argument_registers(save_arg_registers), 852 _method(NULL), 853 _stub_name(stub_name), 854 _stub_function(stub_function), 855 _stub_entry_point(NULL), 856 _entry_bci(InvocationEntryBci), 857 _initial_gvn(NULL), 858 _for_igvn(NULL), 859 _warm_calls(NULL), 860 _orig_pc_slot(0), 861 _orig_pc_slot_offset_in_bytes(0), 862 _subsume_loads(true), 863 _do_escape_analysis(false), 864 _failure_reason(NULL), 865 _code_buffer("Compile::Fill_buffer"), 866 _has_method_handle_invokes(false), 867 _mach_constant_base_node(NULL), 868 _node_bundling_limit(0), 869 _node_bundling_base(NULL), 870 _java_calls(0), 871 _inner_loops(0), 872 #ifndef PRODUCT 873 _trace_opto_output(TraceOptoOutput), 874 _printer(NULL), 875 #endif 876 _congraph(NULL) { 877 C = this; 878 879 #ifndef PRODUCT 880 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 881 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 882 set_print_assembly(PrintFrameConverterAssembly); 883 set_parsed_irreducible_loop(false); 884 #endif 885 CompileWrapper cw(this); 886 Init(/*AliasLevel=*/ 0); 887 init_tf((*generator)()); 888 889 { 890 // The following is a dummy for the sake of GraphKit::gen_stub 891 Unique_Node_List for_igvn(comp_arena()); 892 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 893 PhaseGVN gvn(Thread::current()->resource_area(),255); 894 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 895 gvn.transform_no_reclaim(top()); 896 897 GraphKit kit; 898 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 899 } 900 901 NOT_PRODUCT( verify_graph_edges(); ) 902 Code_Gen(); 903 if (failing()) return; 904 905 906 // Entry point will be accessed using compile->stub_entry_point(); 907 if (code_buffer() == NULL) { 908 Matcher::soft_match_failure(); 909 } else { 910 if (PrintAssembly && (WizardMode || Verbose)) 911 tty->print_cr("### Stub::%s", stub_name); 912 913 if (!failing()) { 914 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 915 916 // Make the NMethod 917 // For now we mark the frame as never safe for profile stackwalking 918 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 919 code_buffer(), 920 CodeOffsets::frame_never_safe, 921 // _code_offsets.value(CodeOffsets::Frame_Complete), 922 frame_size_in_words(), 923 _oop_map_set, 924 save_arg_registers); 925 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 926 927 _stub_entry_point = rs->entry_point(); 928 } 929 } 930 } 931 932 //------------------------------Init------------------------------------------- 933 // Prepare for a single compilation 934 void Compile::Init(int aliaslevel) { 935 _unique = 0; 936 _regalloc = NULL; 937 938 _tf = NULL; // filled in later 939 _top = NULL; // cached later 940 _matcher = NULL; // filled in later 941 _cfg = NULL; // filled in later 942 943 set_24_bit_selection_and_mode(Use24BitFP, false); 944 945 _node_note_array = NULL; 946 _default_node_notes = NULL; 947 948 _immutable_memory = NULL; // filled in at first inquiry 949 950 // Globally visible Nodes 951 // First set TOP to NULL to give safe behavior during creation of RootNode 952 set_cached_top_node(NULL); 953 set_root(new (this) RootNode()); 954 // Now that you have a Root to point to, create the real TOP 955 set_cached_top_node( new (this) ConNode(Type::TOP) ); 956 set_recent_alloc(NULL, NULL); 957 958 // Create Debug Information Recorder to record scopes, oopmaps, etc. 959 env()->set_oop_recorder(new OopRecorder(env()->arena())); 960 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 961 env()->set_dependencies(new Dependencies(env())); 962 963 _fixed_slots = 0; 964 set_has_split_ifs(false); 965 set_has_loops(has_method() && method()->has_loops()); // first approximation 966 set_has_stringbuilder(false); 967 _trap_can_recompile = false; // no traps emitted yet 968 _major_progress = true; // start out assuming good things will happen 969 set_has_unsafe_access(false); 970 set_max_vector_size(0); 971 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 972 set_decompile_count(0); 973 974 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 975 set_num_loop_opts(LoopOptsCount); 976 set_do_inlining(Inline); 977 set_max_inline_size(MaxInlineSize); 978 set_freq_inline_size(FreqInlineSize); 979 set_do_scheduling(OptoScheduling); 980 set_do_count_invocations(false); 981 set_do_method_data_update(false); 982 983 if (debug_info()->recording_non_safepoints()) { 984 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 985 (comp_arena(), 8, 0, NULL)); 986 set_default_node_notes(Node_Notes::make(this)); 987 } 988 989 // // -- Initialize types before each compile -- 990 // // Update cached type information 991 // if( _method && _method->constants() ) 992 // Type::update_loaded_types(_method, _method->constants()); 993 994 // Init alias_type map. 995 if (!_do_escape_analysis && aliaslevel == 3) 996 aliaslevel = 2; // No unique types without escape analysis 997 _AliasLevel = aliaslevel; 998 const int grow_ats = 16; 999 _max_alias_types = grow_ats; 1000 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1001 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1002 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1003 { 1004 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1005 } 1006 // Initialize the first few types. 1007 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1008 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1009 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1010 _num_alias_types = AliasIdxRaw+1; 1011 // Zero out the alias type cache. 1012 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1013 // A NULL adr_type hits in the cache right away. Preload the right answer. 1014 probe_alias_cache(NULL)->_index = AliasIdxTop; 1015 1016 _intrinsics = NULL; 1017 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1018 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1019 register_library_intrinsics(); 1020 } 1021 1022 //---------------------------init_start---------------------------------------- 1023 // Install the StartNode on this compile object. 1024 void Compile::init_start(StartNode* s) { 1025 if (failing()) 1026 return; // already failing 1027 assert(s == start(), ""); 1028 } 1029 1030 StartNode* Compile::start() const { 1031 assert(!failing(), ""); 1032 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1033 Node* start = root()->fast_out(i); 1034 if( start->is_Start() ) 1035 return start->as_Start(); 1036 } 1037 ShouldNotReachHere(); 1038 return NULL; 1039 } 1040 1041 //-------------------------------immutable_memory------------------------------------- 1042 // Access immutable memory 1043 Node* Compile::immutable_memory() { 1044 if (_immutable_memory != NULL) { 1045 return _immutable_memory; 1046 } 1047 StartNode* s = start(); 1048 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1049 Node *p = s->fast_out(i); 1050 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1051 _immutable_memory = p; 1052 return _immutable_memory; 1053 } 1054 } 1055 ShouldNotReachHere(); 1056 return NULL; 1057 } 1058 1059 //----------------------set_cached_top_node------------------------------------ 1060 // Install the cached top node, and make sure Node::is_top works correctly. 1061 void Compile::set_cached_top_node(Node* tn) { 1062 if (tn != NULL) verify_top(tn); 1063 Node* old_top = _top; 1064 _top = tn; 1065 // Calling Node::setup_is_top allows the nodes the chance to adjust 1066 // their _out arrays. 1067 if (_top != NULL) _top->setup_is_top(); 1068 if (old_top != NULL) old_top->setup_is_top(); 1069 assert(_top == NULL || top()->is_top(), ""); 1070 } 1071 1072 #ifndef PRODUCT 1073 void Compile::verify_top(Node* tn) const { 1074 if (tn != NULL) { 1075 assert(tn->is_Con(), "top node must be a constant"); 1076 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1077 assert(tn->in(0) != NULL, "must have live top node"); 1078 } 1079 } 1080 #endif 1081 1082 1083 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1084 1085 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1086 guarantee(arr != NULL, ""); 1087 int num_blocks = arr->length(); 1088 if (grow_by < num_blocks) grow_by = num_blocks; 1089 int num_notes = grow_by * _node_notes_block_size; 1090 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1091 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1092 while (num_notes > 0) { 1093 arr->append(notes); 1094 notes += _node_notes_block_size; 1095 num_notes -= _node_notes_block_size; 1096 } 1097 assert(num_notes == 0, "exact multiple, please"); 1098 } 1099 1100 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1101 if (source == NULL || dest == NULL) return false; 1102 1103 if (dest->is_Con()) 1104 return false; // Do not push debug info onto constants. 1105 1106 #ifdef ASSERT 1107 // Leave a bread crumb trail pointing to the original node: 1108 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1109 dest->set_debug_orig(source); 1110 } 1111 #endif 1112 1113 if (node_note_array() == NULL) 1114 return false; // Not collecting any notes now. 1115 1116 // This is a copy onto a pre-existing node, which may already have notes. 1117 // If both nodes have notes, do not overwrite any pre-existing notes. 1118 Node_Notes* source_notes = node_notes_at(source->_idx); 1119 if (source_notes == NULL || source_notes->is_clear()) return false; 1120 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1121 if (dest_notes == NULL || dest_notes->is_clear()) { 1122 return set_node_notes_at(dest->_idx, source_notes); 1123 } 1124 1125 Node_Notes merged_notes = (*source_notes); 1126 // The order of operations here ensures that dest notes will win... 1127 merged_notes.update_from(dest_notes); 1128 return set_node_notes_at(dest->_idx, &merged_notes); 1129 } 1130 1131 1132 //--------------------------allow_range_check_smearing------------------------- 1133 // Gating condition for coalescing similar range checks. 1134 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1135 // single covering check that is at least as strong as any of them. 1136 // If the optimization succeeds, the simplified (strengthened) range check 1137 // will always succeed. If it fails, we will deopt, and then give up 1138 // on the optimization. 1139 bool Compile::allow_range_check_smearing() const { 1140 // If this method has already thrown a range-check, 1141 // assume it was because we already tried range smearing 1142 // and it failed. 1143 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1144 return !already_trapped; 1145 } 1146 1147 1148 //------------------------------flatten_alias_type----------------------------- 1149 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1150 int offset = tj->offset(); 1151 TypePtr::PTR ptr = tj->ptr(); 1152 1153 // Known instance (scalarizable allocation) alias only with itself. 1154 bool is_known_inst = tj->isa_oopptr() != NULL && 1155 tj->is_oopptr()->is_known_instance(); 1156 1157 // Process weird unsafe references. 1158 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1159 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1160 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1161 tj = TypeOopPtr::BOTTOM; 1162 ptr = tj->ptr(); 1163 offset = tj->offset(); 1164 } 1165 1166 // Array pointers need some flattening 1167 const TypeAryPtr *ta = tj->isa_aryptr(); 1168 if( ta && is_known_inst ) { 1169 if ( offset != Type::OffsetBot && 1170 offset > arrayOopDesc::length_offset_in_bytes() ) { 1171 offset = Type::OffsetBot; // Flatten constant access into array body only 1172 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1173 } 1174 } else if( ta && _AliasLevel >= 2 ) { 1175 // For arrays indexed by constant indices, we flatten the alias 1176 // space to include all of the array body. Only the header, klass 1177 // and array length can be accessed un-aliased. 1178 if( offset != Type::OffsetBot ) { 1179 if( ta->const_oop() ) { // MethodData* or Method* 1180 offset = Type::OffsetBot; // Flatten constant access into array body 1181 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1182 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1183 // range is OK as-is. 1184 tj = ta = TypeAryPtr::RANGE; 1185 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1186 tj = TypeInstPtr::KLASS; // all klass loads look alike 1187 ta = TypeAryPtr::RANGE; // generic ignored junk 1188 ptr = TypePtr::BotPTR; 1189 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1190 tj = TypeInstPtr::MARK; 1191 ta = TypeAryPtr::RANGE; // generic ignored junk 1192 ptr = TypePtr::BotPTR; 1193 } else { // Random constant offset into array body 1194 offset = Type::OffsetBot; // Flatten constant access into array body 1195 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1196 } 1197 } 1198 // Arrays of fixed size alias with arrays of unknown size. 1199 if (ta->size() != TypeInt::POS) { 1200 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1201 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1202 } 1203 // Arrays of known objects become arrays of unknown objects. 1204 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1205 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1206 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1207 } 1208 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1209 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1210 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1211 } 1212 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1213 // cannot be distinguished by bytecode alone. 1214 if (ta->elem() == TypeInt::BOOL) { 1215 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1216 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1217 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1218 } 1219 // During the 2nd round of IterGVN, NotNull castings are removed. 1220 // Make sure the Bottom and NotNull variants alias the same. 1221 // Also, make sure exact and non-exact variants alias the same. 1222 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { 1223 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1224 } 1225 } 1226 1227 // Oop pointers need some flattening 1228 const TypeInstPtr *to = tj->isa_instptr(); 1229 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1230 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1231 if( ptr == TypePtr::Constant ) { 1232 if (to->klass() != ciEnv::current()->Class_klass() || 1233 offset < k->size_helper() * wordSize) { 1234 // No constant oop pointers (such as Strings); they alias with 1235 // unknown strings. 1236 assert(!is_known_inst, "not scalarizable allocation"); 1237 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1238 } 1239 } else if( is_known_inst ) { 1240 tj = to; // Keep NotNull and klass_is_exact for instance type 1241 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1242 // During the 2nd round of IterGVN, NotNull castings are removed. 1243 // Make sure the Bottom and NotNull variants alias the same. 1244 // Also, make sure exact and non-exact variants alias the same. 1245 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1246 } 1247 // Canonicalize the holder of this field 1248 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1249 // First handle header references such as a LoadKlassNode, even if the 1250 // object's klass is unloaded at compile time (4965979). 1251 if (!is_known_inst) { // Do it only for non-instance types 1252 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1253 } 1254 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1255 // Static fields are in the space above the normal instance 1256 // fields in the java.lang.Class instance. 1257 if (to->klass() != ciEnv::current()->Class_klass()) { 1258 to = NULL; 1259 tj = TypeOopPtr::BOTTOM; 1260 offset = tj->offset(); 1261 } 1262 } else { 1263 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1264 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1265 if( is_known_inst ) { 1266 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1267 } else { 1268 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1269 } 1270 } 1271 } 1272 } 1273 1274 // Klass pointers to object array klasses need some flattening 1275 const TypeKlassPtr *tk = tj->isa_klassptr(); 1276 if( tk ) { 1277 // If we are referencing a field within a Klass, we need 1278 // to assume the worst case of an Object. Both exact and 1279 // inexact types must flatten to the same alias class so 1280 // use NotNull as the PTR. 1281 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1282 1283 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1284 TypeKlassPtr::OBJECT->klass(), 1285 offset); 1286 } 1287 1288 ciKlass* klass = tk->klass(); 1289 if( klass->is_obj_array_klass() ) { 1290 ciKlass* k = TypeAryPtr::OOPS->klass(); 1291 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1292 k = TypeInstPtr::BOTTOM->klass(); 1293 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1294 } 1295 1296 // Check for precise loads from the primary supertype array and force them 1297 // to the supertype cache alias index. Check for generic array loads from 1298 // the primary supertype array and also force them to the supertype cache 1299 // alias index. Since the same load can reach both, we need to merge 1300 // these 2 disparate memories into the same alias class. Since the 1301 // primary supertype array is read-only, there's no chance of confusion 1302 // where we bypass an array load and an array store. 1303 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1304 if (offset == Type::OffsetBot || 1305 (offset >= primary_supers_offset && 1306 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1307 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1308 offset = in_bytes(Klass::secondary_super_cache_offset()); 1309 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1310 } 1311 } 1312 1313 // Flatten all Raw pointers together. 1314 if (tj->base() == Type::RawPtr) 1315 tj = TypeRawPtr::BOTTOM; 1316 1317 if (tj->base() == Type::AnyPtr) 1318 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1319 1320 // Flatten all to bottom for now 1321 switch( _AliasLevel ) { 1322 case 0: 1323 tj = TypePtr::BOTTOM; 1324 break; 1325 case 1: // Flatten to: oop, static, field or array 1326 switch (tj->base()) { 1327 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1328 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1329 case Type::AryPtr: // do not distinguish arrays at all 1330 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1331 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1332 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1333 default: ShouldNotReachHere(); 1334 } 1335 break; 1336 case 2: // No collapsing at level 2; keep all splits 1337 case 3: // No collapsing at level 3; keep all splits 1338 break; 1339 default: 1340 Unimplemented(); 1341 } 1342 1343 offset = tj->offset(); 1344 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1345 1346 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1347 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1348 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1349 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1350 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1351 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1352 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1353 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1354 assert( tj->ptr() != TypePtr::TopPTR && 1355 tj->ptr() != TypePtr::AnyNull && 1356 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1357 // assert( tj->ptr() != TypePtr::Constant || 1358 // tj->base() == Type::RawPtr || 1359 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1360 1361 return tj; 1362 } 1363 1364 void Compile::AliasType::Init(int i, const TypePtr* at) { 1365 _index = i; 1366 _adr_type = at; 1367 _field = NULL; 1368 _is_rewritable = true; // default 1369 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1370 if (atoop != NULL && atoop->is_known_instance()) { 1371 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1372 _general_index = Compile::current()->get_alias_index(gt); 1373 } else { 1374 _general_index = 0; 1375 } 1376 } 1377 1378 //---------------------------------print_on------------------------------------ 1379 #ifndef PRODUCT 1380 void Compile::AliasType::print_on(outputStream* st) { 1381 if (index() < 10) 1382 st->print("@ <%d> ", index()); 1383 else st->print("@ <%d>", index()); 1384 st->print(is_rewritable() ? " " : " RO"); 1385 int offset = adr_type()->offset(); 1386 if (offset == Type::OffsetBot) 1387 st->print(" +any"); 1388 else st->print(" +%-3d", offset); 1389 st->print(" in "); 1390 adr_type()->dump_on(st); 1391 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1392 if (field() != NULL && tjp) { 1393 if (tjp->klass() != field()->holder() || 1394 tjp->offset() != field()->offset_in_bytes()) { 1395 st->print(" != "); 1396 field()->print(); 1397 st->print(" ***"); 1398 } 1399 } 1400 } 1401 1402 void print_alias_types() { 1403 Compile* C = Compile::current(); 1404 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1405 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1406 C->alias_type(idx)->print_on(tty); 1407 tty->cr(); 1408 } 1409 } 1410 #endif 1411 1412 1413 //----------------------------probe_alias_cache-------------------------------- 1414 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1415 intptr_t key = (intptr_t) adr_type; 1416 key ^= key >> logAliasCacheSize; 1417 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1418 } 1419 1420 1421 //-----------------------------grow_alias_types-------------------------------- 1422 void Compile::grow_alias_types() { 1423 const int old_ats = _max_alias_types; // how many before? 1424 const int new_ats = old_ats; // how many more? 1425 const int grow_ats = old_ats+new_ats; // how many now? 1426 _max_alias_types = grow_ats; 1427 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1428 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1429 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1430 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1431 } 1432 1433 1434 //--------------------------------find_alias_type------------------------------ 1435 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1436 if (_AliasLevel == 0) 1437 return alias_type(AliasIdxBot); 1438 1439 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1440 if (ace->_adr_type == adr_type) { 1441 return alias_type(ace->_index); 1442 } 1443 1444 // Handle special cases. 1445 if (adr_type == NULL) return alias_type(AliasIdxTop); 1446 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1447 1448 // Do it the slow way. 1449 const TypePtr* flat = flatten_alias_type(adr_type); 1450 1451 #ifdef ASSERT 1452 assert(flat == flatten_alias_type(flat), "idempotent"); 1453 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1454 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1455 const TypeOopPtr* foop = flat->is_oopptr(); 1456 // Scalarizable allocations have exact klass always. 1457 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1458 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1459 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1460 } 1461 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1462 #endif 1463 1464 int idx = AliasIdxTop; 1465 for (int i = 0; i < num_alias_types(); i++) { 1466 if (alias_type(i)->adr_type() == flat) { 1467 idx = i; 1468 break; 1469 } 1470 } 1471 1472 if (idx == AliasIdxTop) { 1473 if (no_create) return NULL; 1474 // Grow the array if necessary. 1475 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1476 // Add a new alias type. 1477 idx = _num_alias_types++; 1478 _alias_types[idx]->Init(idx, flat); 1479 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1480 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1481 if (flat->isa_instptr()) { 1482 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1483 && flat->is_instptr()->klass() == env()->Class_klass()) 1484 alias_type(idx)->set_rewritable(false); 1485 } 1486 if (flat->isa_klassptr()) { 1487 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1488 alias_type(idx)->set_rewritable(false); 1489 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1490 alias_type(idx)->set_rewritable(false); 1491 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1492 alias_type(idx)->set_rewritable(false); 1493 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1494 alias_type(idx)->set_rewritable(false); 1495 } 1496 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1497 // but the base pointer type is not distinctive enough to identify 1498 // references into JavaThread.) 1499 1500 // Check for final fields. 1501 const TypeInstPtr* tinst = flat->isa_instptr(); 1502 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1503 ciField* field; 1504 if (tinst->const_oop() != NULL && 1505 tinst->klass() == ciEnv::current()->Class_klass() && 1506 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1507 // static field 1508 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1509 field = k->get_field_by_offset(tinst->offset(), true); 1510 } else { 1511 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1512 field = k->get_field_by_offset(tinst->offset(), false); 1513 } 1514 assert(field == NULL || 1515 original_field == NULL || 1516 (field->holder() == original_field->holder() && 1517 field->offset() == original_field->offset() && 1518 field->is_static() == original_field->is_static()), "wrong field?"); 1519 // Set field() and is_rewritable() attributes. 1520 if (field != NULL) alias_type(idx)->set_field(field); 1521 } 1522 } 1523 1524 // Fill the cache for next time. 1525 ace->_adr_type = adr_type; 1526 ace->_index = idx; 1527 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1528 1529 // Might as well try to fill the cache for the flattened version, too. 1530 AliasCacheEntry* face = probe_alias_cache(flat); 1531 if (face->_adr_type == NULL) { 1532 face->_adr_type = flat; 1533 face->_index = idx; 1534 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1535 } 1536 1537 return alias_type(idx); 1538 } 1539 1540 1541 Compile::AliasType* Compile::alias_type(ciField* field) { 1542 const TypeOopPtr* t; 1543 if (field->is_static()) 1544 t = TypeInstPtr::make(field->holder()->java_mirror()); 1545 else 1546 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1547 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1548 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); 1549 return atp; 1550 } 1551 1552 1553 //------------------------------have_alias_type-------------------------------- 1554 bool Compile::have_alias_type(const TypePtr* adr_type) { 1555 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1556 if (ace->_adr_type == adr_type) { 1557 return true; 1558 } 1559 1560 // Handle special cases. 1561 if (adr_type == NULL) return true; 1562 if (adr_type == TypePtr::BOTTOM) return true; 1563 1564 return find_alias_type(adr_type, true, NULL) != NULL; 1565 } 1566 1567 //-----------------------------must_alias-------------------------------------- 1568 // True if all values of the given address type are in the given alias category. 1569 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1570 if (alias_idx == AliasIdxBot) return true; // the universal category 1571 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1572 if (alias_idx == AliasIdxTop) return false; // the empty category 1573 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1574 1575 // the only remaining possible overlap is identity 1576 int adr_idx = get_alias_index(adr_type); 1577 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1578 assert(adr_idx == alias_idx || 1579 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1580 && adr_type != TypeOopPtr::BOTTOM), 1581 "should not be testing for overlap with an unsafe pointer"); 1582 return adr_idx == alias_idx; 1583 } 1584 1585 //------------------------------can_alias-------------------------------------- 1586 // True if any values of the given address type are in the given alias category. 1587 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1588 if (alias_idx == AliasIdxTop) return false; // the empty category 1589 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1590 if (alias_idx == AliasIdxBot) return true; // the universal category 1591 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1592 1593 // the only remaining possible overlap is identity 1594 int adr_idx = get_alias_index(adr_type); 1595 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1596 return adr_idx == alias_idx; 1597 } 1598 1599 1600 1601 //---------------------------pop_warm_call------------------------------------- 1602 WarmCallInfo* Compile::pop_warm_call() { 1603 WarmCallInfo* wci = _warm_calls; 1604 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1605 return wci; 1606 } 1607 1608 //----------------------------Inline_Warm-------------------------------------- 1609 int Compile::Inline_Warm() { 1610 // If there is room, try to inline some more warm call sites. 1611 // %%% Do a graph index compaction pass when we think we're out of space? 1612 if (!InlineWarmCalls) return 0; 1613 1614 int calls_made_hot = 0; 1615 int room_to_grow = NodeCountInliningCutoff - unique(); 1616 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1617 int amount_grown = 0; 1618 WarmCallInfo* call; 1619 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1620 int est_size = (int)call->size(); 1621 if (est_size > (room_to_grow - amount_grown)) { 1622 // This one won't fit anyway. Get rid of it. 1623 call->make_cold(); 1624 continue; 1625 } 1626 call->make_hot(); 1627 calls_made_hot++; 1628 amount_grown += est_size; 1629 amount_to_grow -= est_size; 1630 } 1631 1632 if (calls_made_hot > 0) set_major_progress(); 1633 return calls_made_hot; 1634 } 1635 1636 1637 //----------------------------Finish_Warm-------------------------------------- 1638 void Compile::Finish_Warm() { 1639 if (!InlineWarmCalls) return; 1640 if (failing()) return; 1641 if (warm_calls() == NULL) return; 1642 1643 // Clean up loose ends, if we are out of space for inlining. 1644 WarmCallInfo* call; 1645 while ((call = pop_warm_call()) != NULL) { 1646 call->make_cold(); 1647 } 1648 } 1649 1650 //---------------------cleanup_loop_predicates----------------------- 1651 // Remove the opaque nodes that protect the predicates so that all unused 1652 // checks and uncommon_traps will be eliminated from the ideal graph 1653 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1654 if (predicate_count()==0) return; 1655 for (int i = predicate_count(); i > 0; i--) { 1656 Node * n = predicate_opaque1_node(i-1); 1657 assert(n->Opcode() == Op_Opaque1, "must be"); 1658 igvn.replace_node(n, n->in(1)); 1659 } 1660 assert(predicate_count()==0, "should be clean!"); 1661 } 1662 1663 //------------------------------Optimize--------------------------------------- 1664 // Given a graph, optimize it. 1665 void Compile::Optimize() { 1666 TracePhase t1("optimizer", &_t_optimizer, true); 1667 1668 #ifndef PRODUCT 1669 if (env()->break_at_compile()) { 1670 BREAKPOINT; 1671 } 1672 1673 #endif 1674 1675 ResourceMark rm; 1676 int loop_opts_cnt; 1677 1678 NOT_PRODUCT( verify_graph_edges(); ) 1679 1680 print_method("After Parsing"); 1681 1682 { 1683 // Iterative Global Value Numbering, including ideal transforms 1684 // Initialize IterGVN with types and values from parse-time GVN 1685 PhaseIterGVN igvn(initial_gvn()); 1686 { 1687 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1688 igvn.optimize(); 1689 } 1690 1691 print_method("Iter GVN 1", 2); 1692 1693 if (failing()) return; 1694 1695 // Perform escape analysis 1696 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 1697 if (has_loops()) { 1698 // Cleanup graph (remove dead nodes). 1699 TracePhase t2("idealLoop", &_t_idealLoop, true); 1700 PhaseIdealLoop ideal_loop( igvn, false, true ); 1701 if (major_progress()) print_method("PhaseIdealLoop before EA", 2); 1702 if (failing()) return; 1703 } 1704 ConnectionGraph::do_analysis(this, &igvn); 1705 1706 if (failing()) return; 1707 1708 // Optimize out fields loads from scalar replaceable allocations. 1709 igvn.optimize(); 1710 print_method("Iter GVN after EA", 2); 1711 1712 if (failing()) return; 1713 1714 if (congraph() != NULL && macro_count() > 0) { 1715 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 1716 PhaseMacroExpand mexp(igvn); 1717 mexp.eliminate_macro_nodes(); 1718 igvn.set_delay_transform(false); 1719 1720 igvn.optimize(); 1721 print_method("Iter GVN after eliminating allocations and locks", 2); 1722 1723 if (failing()) return; 1724 } 1725 } 1726 1727 // Loop transforms on the ideal graph. Range Check Elimination, 1728 // peeling, unrolling, etc. 1729 1730 // Set loop opts counter 1731 loop_opts_cnt = num_loop_opts(); 1732 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 1733 { 1734 TracePhase t2("idealLoop", &_t_idealLoop, true); 1735 PhaseIdealLoop ideal_loop( igvn, true ); 1736 loop_opts_cnt--; 1737 if (major_progress()) print_method("PhaseIdealLoop 1", 2); 1738 if (failing()) return; 1739 } 1740 // Loop opts pass if partial peeling occurred in previous pass 1741 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 1742 TracePhase t3("idealLoop", &_t_idealLoop, true); 1743 PhaseIdealLoop ideal_loop( igvn, false ); 1744 loop_opts_cnt--; 1745 if (major_progress()) print_method("PhaseIdealLoop 2", 2); 1746 if (failing()) return; 1747 } 1748 // Loop opts pass for loop-unrolling before CCP 1749 if(major_progress() && (loop_opts_cnt > 0)) { 1750 TracePhase t4("idealLoop", &_t_idealLoop, true); 1751 PhaseIdealLoop ideal_loop( igvn, false ); 1752 loop_opts_cnt--; 1753 if (major_progress()) print_method("PhaseIdealLoop 3", 2); 1754 } 1755 if (!failing()) { 1756 // Verify that last round of loop opts produced a valid graph 1757 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1758 PhaseIdealLoop::verify(igvn); 1759 } 1760 } 1761 if (failing()) return; 1762 1763 // Conditional Constant Propagation; 1764 PhaseCCP ccp( &igvn ); 1765 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 1766 { 1767 TracePhase t2("ccp", &_t_ccp, true); 1768 ccp.do_transform(); 1769 } 1770 print_method("PhaseCPP 1", 2); 1771 1772 assert( true, "Break here to ccp.dump_old2new_map()"); 1773 1774 // Iterative Global Value Numbering, including ideal transforms 1775 { 1776 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 1777 igvn = ccp; 1778 igvn.optimize(); 1779 } 1780 1781 print_method("Iter GVN 2", 2); 1782 1783 if (failing()) return; 1784 1785 // Loop transforms on the ideal graph. Range Check Elimination, 1786 // peeling, unrolling, etc. 1787 if(loop_opts_cnt > 0) { 1788 debug_only( int cnt = 0; ); 1789 while(major_progress() && (loop_opts_cnt > 0)) { 1790 TracePhase t2("idealLoop", &_t_idealLoop, true); 1791 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 1792 PhaseIdealLoop ideal_loop( igvn, true); 1793 loop_opts_cnt--; 1794 if (major_progress()) print_method("PhaseIdealLoop iterations", 2); 1795 if (failing()) return; 1796 } 1797 } 1798 1799 { 1800 // Verify that all previous optimizations produced a valid graph 1801 // at least to this point, even if no loop optimizations were done. 1802 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 1803 PhaseIdealLoop::verify(igvn); 1804 } 1805 1806 { 1807 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 1808 PhaseMacroExpand mex(igvn); 1809 if (mex.expand_macro_nodes()) { 1810 assert(failing(), "must bail out w/ explicit message"); 1811 return; 1812 } 1813 } 1814 1815 } // (End scope of igvn; run destructor if necessary for asserts.) 1816 1817 // A method with only infinite loops has no edges entering loops from root 1818 { 1819 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 1820 if (final_graph_reshaping()) { 1821 assert(failing(), "must bail out w/ explicit message"); 1822 return; 1823 } 1824 } 1825 1826 print_method("Optimize finished", 2); 1827 } 1828 1829 1830 //------------------------------Code_Gen--------------------------------------- 1831 // Given a graph, generate code for it 1832 void Compile::Code_Gen() { 1833 if (failing()) return; 1834 1835 // Perform instruction selection. You might think we could reclaim Matcher 1836 // memory PDQ, but actually the Matcher is used in generating spill code. 1837 // Internals of the Matcher (including some VectorSets) must remain live 1838 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 1839 // set a bit in reclaimed memory. 1840 1841 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1842 // nodes. Mapping is only valid at the root of each matched subtree. 1843 NOT_PRODUCT( verify_graph_edges(); ) 1844 1845 Node_List proj_list; 1846 Matcher m(proj_list); 1847 _matcher = &m; 1848 { 1849 TracePhase t2("matcher", &_t_matcher, true); 1850 m.match(); 1851 } 1852 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 1853 // nodes. Mapping is only valid at the root of each matched subtree. 1854 NOT_PRODUCT( verify_graph_edges(); ) 1855 1856 // If you have too many nodes, or if matching has failed, bail out 1857 check_node_count(0, "out of nodes matching instructions"); 1858 if (failing()) return; 1859 1860 // Build a proper-looking CFG 1861 PhaseCFG cfg(node_arena(), root(), m); 1862 _cfg = &cfg; 1863 { 1864 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 1865 cfg.Dominators(); 1866 if (failing()) return; 1867 1868 NOT_PRODUCT( verify_graph_edges(); ) 1869 1870 cfg.Estimate_Block_Frequency(); 1871 cfg.GlobalCodeMotion(m,unique(),proj_list); 1872 if (failing()) return; 1873 1874 print_method("Global code motion", 2); 1875 1876 NOT_PRODUCT( verify_graph_edges(); ) 1877 1878 debug_only( cfg.verify(); ) 1879 } 1880 NOT_PRODUCT( verify_graph_edges(); ) 1881 1882 PhaseChaitin regalloc(unique(),cfg,m); 1883 _regalloc = ®alloc; 1884 { 1885 TracePhase t2("regalloc", &_t_registerAllocation, true); 1886 // Perform any platform dependent preallocation actions. This is used, 1887 // for example, to avoid taking an implicit null pointer exception 1888 // using the frame pointer on win95. 1889 _regalloc->pd_preallocate_hook(); 1890 1891 // Perform register allocation. After Chaitin, use-def chains are 1892 // no longer accurate (at spill code) and so must be ignored. 1893 // Node->LRG->reg mappings are still accurate. 1894 _regalloc->Register_Allocate(); 1895 1896 // Bail out if the allocator builds too many nodes 1897 if (failing()) return; 1898 } 1899 1900 // Prior to register allocation we kept empty basic blocks in case the 1901 // the allocator needed a place to spill. After register allocation we 1902 // are not adding any new instructions. If any basic block is empty, we 1903 // can now safely remove it. 1904 { 1905 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 1906 cfg.remove_empty(); 1907 if (do_freq_based_layout()) { 1908 PhaseBlockLayout layout(cfg); 1909 } else { 1910 cfg.set_loop_alignment(); 1911 } 1912 cfg.fixup_flow(); 1913 } 1914 1915 // Perform any platform dependent postallocation verifications. 1916 debug_only( _regalloc->pd_postallocate_verify_hook(); ) 1917 1918 // Apply peephole optimizations 1919 if( OptoPeephole ) { 1920 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 1921 PhasePeephole peep( _regalloc, cfg); 1922 peep.do_transform(); 1923 } 1924 1925 // Convert Nodes to instruction bits in a buffer 1926 { 1927 // %%%% workspace merge brought two timers together for one job 1928 TracePhase t2a("output", &_t_output, true); 1929 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 1930 Output(); 1931 } 1932 1933 print_method("Final Code"); 1934 1935 // He's dead, Jim. 1936 _cfg = (PhaseCFG*)0xdeadbeef; 1937 _regalloc = (PhaseChaitin*)0xdeadbeef; 1938 } 1939 1940 1941 //------------------------------dump_asm--------------------------------------- 1942 // Dump formatted assembly 1943 #ifndef PRODUCT 1944 void Compile::dump_asm(int *pcs, uint pc_limit) { 1945 bool cut_short = false; 1946 tty->print_cr("#"); 1947 tty->print("# "); _tf->dump(); tty->cr(); 1948 tty->print_cr("#"); 1949 1950 // For all blocks 1951 int pc = 0x0; // Program counter 1952 char starts_bundle = ' '; 1953 _regalloc->dump_frame(); 1954 1955 Node *n = NULL; 1956 for( uint i=0; i<_cfg->_num_blocks; i++ ) { 1957 if (VMThread::should_terminate()) { cut_short = true; break; } 1958 Block *b = _cfg->_blocks[i]; 1959 if (b->is_connector() && !Verbose) continue; 1960 n = b->_nodes[0]; 1961 if (pcs && n->_idx < pc_limit) 1962 tty->print("%3.3x ", pcs[n->_idx]); 1963 else 1964 tty->print(" "); 1965 b->dump_head( &_cfg->_bbs ); 1966 if (b->is_connector()) { 1967 tty->print_cr(" # Empty connector block"); 1968 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 1969 tty->print_cr(" # Block is sole successor of call"); 1970 } 1971 1972 // For all instructions 1973 Node *delay = NULL; 1974 for( uint j = 0; j<b->_nodes.size(); j++ ) { 1975 if (VMThread::should_terminate()) { cut_short = true; break; } 1976 n = b->_nodes[j]; 1977 if (valid_bundle_info(n)) { 1978 Bundle *bundle = node_bundling(n); 1979 if (bundle->used_in_unconditional_delay()) { 1980 delay = n; 1981 continue; 1982 } 1983 if (bundle->starts_bundle()) 1984 starts_bundle = '+'; 1985 } 1986 1987 if (WizardMode) n->dump(); 1988 1989 if( !n->is_Region() && // Dont print in the Assembly 1990 !n->is_Phi() && // a few noisely useless nodes 1991 !n->is_Proj() && 1992 !n->is_MachTemp() && 1993 !n->is_SafePointScalarObject() && 1994 !n->is_Catch() && // Would be nice to print exception table targets 1995 !n->is_MergeMem() && // Not very interesting 1996 !n->is_top() && // Debug info table constants 1997 !(n->is_Con() && !n->is_Mach())// Debug info table constants 1998 ) { 1999 if (pcs && n->_idx < pc_limit) 2000 tty->print("%3.3x", pcs[n->_idx]); 2001 else 2002 tty->print(" "); 2003 tty->print(" %c ", starts_bundle); 2004 starts_bundle = ' '; 2005 tty->print("\t"); 2006 n->format(_regalloc, tty); 2007 tty->cr(); 2008 } 2009 2010 // If we have an instruction with a delay slot, and have seen a delay, 2011 // then back up and print it 2012 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2013 assert(delay != NULL, "no unconditional delay instruction"); 2014 if (WizardMode) delay->dump(); 2015 2016 if (node_bundling(delay)->starts_bundle()) 2017 starts_bundle = '+'; 2018 if (pcs && n->_idx < pc_limit) 2019 tty->print("%3.3x", pcs[n->_idx]); 2020 else 2021 tty->print(" "); 2022 tty->print(" %c ", starts_bundle); 2023 starts_bundle = ' '; 2024 tty->print("\t"); 2025 delay->format(_regalloc, tty); 2026 tty->print_cr(""); 2027 delay = NULL; 2028 } 2029 2030 // Dump the exception table as well 2031 if( n->is_Catch() && (Verbose || WizardMode) ) { 2032 // Print the exception table for this offset 2033 _handler_table.print_subtable_for(pc); 2034 } 2035 } 2036 2037 if (pcs && n->_idx < pc_limit) 2038 tty->print_cr("%3.3x", pcs[n->_idx]); 2039 else 2040 tty->print_cr(""); 2041 2042 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2043 2044 } // End of per-block dump 2045 tty->print_cr(""); 2046 2047 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2048 } 2049 #endif 2050 2051 //------------------------------Final_Reshape_Counts--------------------------- 2052 // This class defines counters to help identify when a method 2053 // may/must be executed using hardware with only 24-bit precision. 2054 struct Final_Reshape_Counts : public StackObj { 2055 int _call_count; // count non-inlined 'common' calls 2056 int _float_count; // count float ops requiring 24-bit precision 2057 int _double_count; // count double ops requiring more precision 2058 int _java_call_count; // count non-inlined 'java' calls 2059 int _inner_loop_count; // count loops which need alignment 2060 VectorSet _visited; // Visitation flags 2061 Node_List _tests; // Set of IfNodes & PCTableNodes 2062 2063 Final_Reshape_Counts() : 2064 _call_count(0), _float_count(0), _double_count(0), 2065 _java_call_count(0), _inner_loop_count(0), 2066 _visited( Thread::current()->resource_area() ) { } 2067 2068 void inc_call_count () { _call_count ++; } 2069 void inc_float_count () { _float_count ++; } 2070 void inc_double_count() { _double_count++; } 2071 void inc_java_call_count() { _java_call_count++; } 2072 void inc_inner_loop_count() { _inner_loop_count++; } 2073 2074 int get_call_count () const { return _call_count ; } 2075 int get_float_count () const { return _float_count ; } 2076 int get_double_count() const { return _double_count; } 2077 int get_java_call_count() const { return _java_call_count; } 2078 int get_inner_loop_count() const { return _inner_loop_count; } 2079 }; 2080 2081 #ifdef ASSERT 2082 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2083 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2084 // Make sure the offset goes inside the instance layout. 2085 return k->contains_field_offset(tp->offset()); 2086 // Note that OffsetBot and OffsetTop are very negative. 2087 } 2088 #endif 2089 2090 // Eliminate trivially redundant StoreCMs and accumulate their 2091 // precedence edges. 2092 static void eliminate_redundant_card_marks(Node* n) { 2093 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2094 if (n->in(MemNode::Address)->outcnt() > 1) { 2095 // There are multiple users of the same address so it might be 2096 // possible to eliminate some of the StoreCMs 2097 Node* mem = n->in(MemNode::Memory); 2098 Node* adr = n->in(MemNode::Address); 2099 Node* val = n->in(MemNode::ValueIn); 2100 Node* prev = n; 2101 bool done = false; 2102 // Walk the chain of StoreCMs eliminating ones that match. As 2103 // long as it's a chain of single users then the optimization is 2104 // safe. Eliminating partially redundant StoreCMs would require 2105 // cloning copies down the other paths. 2106 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2107 if (adr == mem->in(MemNode::Address) && 2108 val == mem->in(MemNode::ValueIn)) { 2109 // redundant StoreCM 2110 if (mem->req() > MemNode::OopStore) { 2111 // Hasn't been processed by this code yet. 2112 n->add_prec(mem->in(MemNode::OopStore)); 2113 } else { 2114 // Already converted to precedence edge 2115 for (uint i = mem->req(); i < mem->len(); i++) { 2116 // Accumulate any precedence edges 2117 if (mem->in(i) != NULL) { 2118 n->add_prec(mem->in(i)); 2119 } 2120 } 2121 // Everything above this point has been processed. 2122 done = true; 2123 } 2124 // Eliminate the previous StoreCM 2125 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2126 assert(mem->outcnt() == 0, "should be dead"); 2127 mem->disconnect_inputs(NULL); 2128 } else { 2129 prev = mem; 2130 } 2131 mem = prev->in(MemNode::Memory); 2132 } 2133 } 2134 } 2135 2136 //------------------------------final_graph_reshaping_impl---------------------- 2137 // Implement items 1-5 from final_graph_reshaping below. 2138 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) { 2139 2140 if ( n->outcnt() == 0 ) return; // dead node 2141 uint nop = n->Opcode(); 2142 2143 // Check for 2-input instruction with "last use" on right input. 2144 // Swap to left input. Implements item (2). 2145 if( n->req() == 3 && // two-input instruction 2146 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2147 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2148 n->in(2)->outcnt() == 1 &&// right use IS a last use 2149 !n->in(2)->is_Con() ) { // right use is not a constant 2150 // Check for commutative opcode 2151 switch( nop ) { 2152 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2153 case Op_MaxI: case Op_MinI: 2154 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2155 case Op_AndL: case Op_XorL: case Op_OrL: 2156 case Op_AndI: case Op_XorI: case Op_OrI: { 2157 // Move "last use" input to left by swapping inputs 2158 n->swap_edges(1, 2); 2159 break; 2160 } 2161 default: 2162 break; 2163 } 2164 } 2165 2166 #ifdef ASSERT 2167 if( n->is_Mem() ) { 2168 Compile* C = Compile::current(); 2169 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type()); 2170 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2171 // oop will be recorded in oop map if load crosses safepoint 2172 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2173 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2174 "raw memory operations should have control edge"); 2175 } 2176 #endif 2177 // Count FPU ops and common calls, implements item (3) 2178 switch( nop ) { 2179 // Count all float operations that may use FPU 2180 case Op_AddF: 2181 case Op_SubF: 2182 case Op_MulF: 2183 case Op_DivF: 2184 case Op_NegF: 2185 case Op_ModF: 2186 case Op_ConvI2F: 2187 case Op_ConF: 2188 case Op_CmpF: 2189 case Op_CmpF3: 2190 // case Op_ConvL2F: // longs are split into 32-bit halves 2191 frc.inc_float_count(); 2192 break; 2193 2194 case Op_ConvF2D: 2195 case Op_ConvD2F: 2196 frc.inc_float_count(); 2197 frc.inc_double_count(); 2198 break; 2199 2200 // Count all double operations that may use FPU 2201 case Op_AddD: 2202 case Op_SubD: 2203 case Op_MulD: 2204 case Op_DivD: 2205 case Op_NegD: 2206 case Op_ModD: 2207 case Op_ConvI2D: 2208 case Op_ConvD2I: 2209 // case Op_ConvL2D: // handled by leaf call 2210 // case Op_ConvD2L: // handled by leaf call 2211 case Op_ConD: 2212 case Op_CmpD: 2213 case Op_CmpD3: 2214 frc.inc_double_count(); 2215 break; 2216 case Op_Opaque1: // Remove Opaque Nodes before matching 2217 case Op_Opaque2: // Remove Opaque Nodes before matching 2218 n->subsume_by(n->in(1)); 2219 break; 2220 case Op_CallStaticJava: 2221 case Op_CallJava: 2222 case Op_CallDynamicJava: 2223 frc.inc_java_call_count(); // Count java call site; 2224 case Op_CallRuntime: 2225 case Op_CallLeaf: 2226 case Op_CallLeafNoFP: { 2227 assert( n->is_Call(), "" ); 2228 CallNode *call = n->as_Call(); 2229 // Count call sites where the FP mode bit would have to be flipped. 2230 // Do not count uncommon runtime calls: 2231 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2232 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2233 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2234 frc.inc_call_count(); // Count the call site 2235 } else { // See if uncommon argument is shared 2236 Node *n = call->in(TypeFunc::Parms); 2237 int nop = n->Opcode(); 2238 // Clone shared simple arguments to uncommon calls, item (1). 2239 if( n->outcnt() > 1 && 2240 !n->is_Proj() && 2241 nop != Op_CreateEx && 2242 nop != Op_CheckCastPP && 2243 nop != Op_DecodeN && 2244 nop != Op_DecodeNKlass && 2245 !n->is_Mem() ) { 2246 Node *x = n->clone(); 2247 call->set_req( TypeFunc::Parms, x ); 2248 } 2249 } 2250 break; 2251 } 2252 2253 case Op_StoreD: 2254 case Op_LoadD: 2255 case Op_LoadD_unaligned: 2256 frc.inc_double_count(); 2257 goto handle_mem; 2258 case Op_StoreF: 2259 case Op_LoadF: 2260 frc.inc_float_count(); 2261 goto handle_mem; 2262 2263 case Op_StoreCM: 2264 { 2265 // Convert OopStore dependence into precedence edge 2266 Node* prec = n->in(MemNode::OopStore); 2267 n->del_req(MemNode::OopStore); 2268 n->add_prec(prec); 2269 eliminate_redundant_card_marks(n); 2270 } 2271 2272 // fall through 2273 2274 case Op_StoreB: 2275 case Op_StoreC: 2276 case Op_StorePConditional: 2277 case Op_StoreI: 2278 case Op_StoreL: 2279 case Op_StoreIConditional: 2280 case Op_StoreLConditional: 2281 case Op_CompareAndSwapI: 2282 case Op_CompareAndSwapL: 2283 case Op_CompareAndSwapP: 2284 case Op_CompareAndSwapN: 2285 case Op_GetAndAddI: 2286 case Op_GetAndAddL: 2287 case Op_GetAndSetI: 2288 case Op_GetAndSetL: 2289 case Op_GetAndSetP: 2290 case Op_GetAndSetN: 2291 case Op_StoreP: 2292 case Op_StoreN: 2293 case Op_StoreNKlass: 2294 case Op_LoadB: 2295 case Op_LoadUB: 2296 case Op_LoadUS: 2297 case Op_LoadI: 2298 case Op_LoadKlass: 2299 case Op_LoadNKlass: 2300 case Op_LoadL: 2301 case Op_LoadL_unaligned: 2302 case Op_LoadPLocked: 2303 case Op_LoadP: 2304 case Op_LoadN: 2305 case Op_LoadRange: 2306 case Op_LoadS: { 2307 handle_mem: 2308 #ifdef ASSERT 2309 if( VerifyOptoOopOffsets ) { 2310 assert( n->is_Mem(), "" ); 2311 MemNode *mem = (MemNode*)n; 2312 // Check to see if address types have grounded out somehow. 2313 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2314 assert( !tp || oop_offset_is_sane(tp), "" ); 2315 } 2316 #endif 2317 break; 2318 } 2319 2320 case Op_AddP: { // Assert sane base pointers 2321 Node *addp = n->in(AddPNode::Address); 2322 assert( !addp->is_AddP() || 2323 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2324 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2325 "Base pointers must match" ); 2326 #ifdef _LP64 2327 if ((UseCompressedOops || UseCompressedKlassPointers) && 2328 addp->Opcode() == Op_ConP && 2329 addp == n->in(AddPNode::Base) && 2330 n->in(AddPNode::Offset)->is_Con()) { 2331 // Use addressing with narrow klass to load with offset on x86. 2332 // On sparc loading 32-bits constant and decoding it have less 2333 // instructions (4) then load 64-bits constant (7). 2334 // Do this transformation here since IGVN will convert ConN back to ConP. 2335 const Type* t = addp->bottom_type(); 2336 if (t->isa_oopptr() || t->isa_klassptr()) { 2337 Node* nn = NULL; 2338 2339 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2340 2341 // Look for existing ConN node of the same exact type. 2342 Compile* C = Compile::current(); 2343 Node* r = C->root(); 2344 uint cnt = r->outcnt(); 2345 for (uint i = 0; i < cnt; i++) { 2346 Node* m = r->raw_out(i); 2347 if (m!= NULL && m->Opcode() == op && 2348 m->bottom_type()->make_ptr() == t) { 2349 nn = m; 2350 break; 2351 } 2352 } 2353 if (nn != NULL) { 2354 // Decode a narrow oop to match address 2355 // [R12 + narrow_oop_reg<<3 + offset] 2356 if (t->isa_oopptr()) { 2357 nn = new (C) DecodeNNode(nn, t); 2358 } else { 2359 nn = new (C) DecodeNKlassNode(nn, t); 2360 } 2361 n->set_req(AddPNode::Base, nn); 2362 n->set_req(AddPNode::Address, nn); 2363 if (addp->outcnt() == 0) { 2364 addp->disconnect_inputs(NULL); 2365 } 2366 } 2367 } 2368 } 2369 #endif 2370 break; 2371 } 2372 2373 #ifdef _LP64 2374 case Op_CastPP: 2375 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2376 Compile* C = Compile::current(); 2377 Node* in1 = n->in(1); 2378 const Type* t = n->bottom_type(); 2379 Node* new_in1 = in1->clone(); 2380 new_in1->as_DecodeN()->set_type(t); 2381 2382 if (!Matcher::narrow_oop_use_complex_address()) { 2383 // 2384 // x86, ARM and friends can handle 2 adds in addressing mode 2385 // and Matcher can fold a DecodeN node into address by using 2386 // a narrow oop directly and do implicit NULL check in address: 2387 // 2388 // [R12 + narrow_oop_reg<<3 + offset] 2389 // NullCheck narrow_oop_reg 2390 // 2391 // On other platforms (Sparc) we have to keep new DecodeN node and 2392 // use it to do implicit NULL check in address: 2393 // 2394 // decode_not_null narrow_oop_reg, base_reg 2395 // [base_reg + offset] 2396 // NullCheck base_reg 2397 // 2398 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2399 // to keep the information to which NULL check the new DecodeN node 2400 // corresponds to use it as value in implicit_null_check(). 2401 // 2402 new_in1->set_req(0, n->in(0)); 2403 } 2404 2405 n->subsume_by(new_in1); 2406 if (in1->outcnt() == 0) { 2407 in1->disconnect_inputs(NULL); 2408 } 2409 } 2410 break; 2411 2412 case Op_CmpP: 2413 // Do this transformation here to preserve CmpPNode::sub() and 2414 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2415 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2416 Node* in1 = n->in(1); 2417 Node* in2 = n->in(2); 2418 if (!in1->is_DecodeNarrowPtr()) { 2419 in2 = in1; 2420 in1 = n->in(2); 2421 } 2422 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2423 2424 Compile* C = Compile::current(); 2425 Node* new_in2 = NULL; 2426 if (in2->is_DecodeNarrowPtr()) { 2427 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2428 new_in2 = in2->in(1); 2429 } else if (in2->Opcode() == Op_ConP) { 2430 const Type* t = in2->bottom_type(); 2431 if (t == TypePtr::NULL_PTR) { 2432 assert(in1->is_DecodeN(), "compare klass to null?"); 2433 // Don't convert CmpP null check into CmpN if compressed 2434 // oops implicit null check is not generated. 2435 // This will allow to generate normal oop implicit null check. 2436 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2437 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR); 2438 // 2439 // This transformation together with CastPP transformation above 2440 // will generated code for implicit NULL checks for compressed oops. 2441 // 2442 // The original code after Optimize() 2443 // 2444 // LoadN memory, narrow_oop_reg 2445 // decode narrow_oop_reg, base_reg 2446 // CmpP base_reg, NULL 2447 // CastPP base_reg // NotNull 2448 // Load [base_reg + offset], val_reg 2449 // 2450 // after these transformations will be 2451 // 2452 // LoadN memory, narrow_oop_reg 2453 // CmpN narrow_oop_reg, NULL 2454 // decode_not_null narrow_oop_reg, base_reg 2455 // Load [base_reg + offset], val_reg 2456 // 2457 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2458 // since narrow oops can be used in debug info now (see the code in 2459 // final_graph_reshaping_walk()). 2460 // 2461 // At the end the code will be matched to 2462 // on x86: 2463 // 2464 // Load_narrow_oop memory, narrow_oop_reg 2465 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2466 // NullCheck narrow_oop_reg 2467 // 2468 // and on sparc: 2469 // 2470 // Load_narrow_oop memory, narrow_oop_reg 2471 // decode_not_null narrow_oop_reg, base_reg 2472 // Load [base_reg + offset], val_reg 2473 // NullCheck base_reg 2474 // 2475 } else if (t->isa_oopptr()) { 2476 new_in2 = ConNode::make(C, t->make_narrowoop()); 2477 } else if (t->isa_klassptr()) { 2478 new_in2 = ConNode::make(C, t->make_narrowklass()); 2479 } 2480 } 2481 if (new_in2 != NULL) { 2482 Node* cmpN = new (C) CmpNNode(in1->in(1), new_in2); 2483 n->subsume_by( cmpN ); 2484 if (in1->outcnt() == 0) { 2485 in1->disconnect_inputs(NULL); 2486 } 2487 if (in2->outcnt() == 0) { 2488 in2->disconnect_inputs(NULL); 2489 } 2490 } 2491 } 2492 break; 2493 2494 case Op_DecodeN: 2495 case Op_DecodeNKlass: 2496 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2497 // DecodeN could be pinned when it can't be fold into 2498 // an address expression, see the code for Op_CastPP above. 2499 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2500 break; 2501 2502 case Op_EncodeP: 2503 case Op_EncodePKlass: { 2504 Node* in1 = n->in(1); 2505 if (in1->is_DecodeNarrowPtr()) { 2506 n->subsume_by(in1->in(1)); 2507 } else if (in1->Opcode() == Op_ConP) { 2508 Compile* C = Compile::current(); 2509 const Type* t = in1->bottom_type(); 2510 if (t == TypePtr::NULL_PTR) { 2511 assert(t->isa_oopptr(), "null klass?"); 2512 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR)); 2513 } else if (t->isa_oopptr()) { 2514 n->subsume_by(ConNode::make(C, t->make_narrowoop())); 2515 } else if (t->isa_klassptr()) { 2516 n->subsume_by(ConNode::make(C, t->make_narrowklass())); 2517 } 2518 } 2519 if (in1->outcnt() == 0) { 2520 in1->disconnect_inputs(NULL); 2521 } 2522 break; 2523 } 2524 2525 case Op_Proj: { 2526 if (OptimizeStringConcat) { 2527 ProjNode* p = n->as_Proj(); 2528 if (p->_is_io_use) { 2529 // Separate projections were used for the exception path which 2530 // are normally removed by a late inline. If it wasn't inlined 2531 // then they will hang around and should just be replaced with 2532 // the original one. 2533 Node* proj = NULL; 2534 // Replace with just one 2535 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2536 Node *use = i.get(); 2537 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2538 proj = use; 2539 break; 2540 } 2541 } 2542 assert(proj != NULL, "must be found"); 2543 p->subsume_by(proj); 2544 } 2545 } 2546 break; 2547 } 2548 2549 case Op_Phi: 2550 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2551 // The EncodeP optimization may create Phi with the same edges 2552 // for all paths. It is not handled well by Register Allocator. 2553 Node* unique_in = n->in(1); 2554 assert(unique_in != NULL, ""); 2555 uint cnt = n->req(); 2556 for (uint i = 2; i < cnt; i++) { 2557 Node* m = n->in(i); 2558 assert(m != NULL, ""); 2559 if (unique_in != m) 2560 unique_in = NULL; 2561 } 2562 if (unique_in != NULL) { 2563 n->subsume_by(unique_in); 2564 } 2565 } 2566 break; 2567 2568 #endif 2569 2570 case Op_ModI: 2571 if (UseDivMod) { 2572 // Check if a%b and a/b both exist 2573 Node* d = n->find_similar(Op_DivI); 2574 if (d) { 2575 // Replace them with a fused divmod if supported 2576 Compile* C = Compile::current(); 2577 if (Matcher::has_match_rule(Op_DivModI)) { 2578 DivModINode* divmod = DivModINode::make(C, n); 2579 d->subsume_by(divmod->div_proj()); 2580 n->subsume_by(divmod->mod_proj()); 2581 } else { 2582 // replace a%b with a-((a/b)*b) 2583 Node* mult = new (C) MulINode(d, d->in(2)); 2584 Node* sub = new (C) SubINode(d->in(1), mult); 2585 n->subsume_by( sub ); 2586 } 2587 } 2588 } 2589 break; 2590 2591 case Op_ModL: 2592 if (UseDivMod) { 2593 // Check if a%b and a/b both exist 2594 Node* d = n->find_similar(Op_DivL); 2595 if (d) { 2596 // Replace them with a fused divmod if supported 2597 Compile* C = Compile::current(); 2598 if (Matcher::has_match_rule(Op_DivModL)) { 2599 DivModLNode* divmod = DivModLNode::make(C, n); 2600 d->subsume_by(divmod->div_proj()); 2601 n->subsume_by(divmod->mod_proj()); 2602 } else { 2603 // replace a%b with a-((a/b)*b) 2604 Node* mult = new (C) MulLNode(d, d->in(2)); 2605 Node* sub = new (C) SubLNode(d->in(1), mult); 2606 n->subsume_by( sub ); 2607 } 2608 } 2609 } 2610 break; 2611 2612 case Op_LoadVector: 2613 case Op_StoreVector: 2614 break; 2615 2616 case Op_PackB: 2617 case Op_PackS: 2618 case Op_PackI: 2619 case Op_PackF: 2620 case Op_PackL: 2621 case Op_PackD: 2622 if (n->req()-1 > 2) { 2623 // Replace many operand PackNodes with a binary tree for matching 2624 PackNode* p = (PackNode*) n; 2625 Node* btp = p->binary_tree_pack(Compile::current(), 1, n->req()); 2626 n->subsume_by(btp); 2627 } 2628 break; 2629 case Op_Loop: 2630 case Op_CountedLoop: 2631 if (n->as_Loop()->is_inner_loop()) { 2632 frc.inc_inner_loop_count(); 2633 } 2634 break; 2635 case Op_LShiftI: 2636 case Op_RShiftI: 2637 case Op_URShiftI: 2638 case Op_LShiftL: 2639 case Op_RShiftL: 2640 case Op_URShiftL: 2641 if (Matcher::need_masked_shift_count) { 2642 // The cpu's shift instructions don't restrict the count to the 2643 // lower 5/6 bits. We need to do the masking ourselves. 2644 Node* in2 = n->in(2); 2645 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2646 const TypeInt* t = in2->find_int_type(); 2647 if (t != NULL && t->is_con()) { 2648 juint shift = t->get_con(); 2649 if (shift > mask) { // Unsigned cmp 2650 Compile* C = Compile::current(); 2651 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask))); 2652 } 2653 } else { 2654 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2655 Compile* C = Compile::current(); 2656 Node* shift = new (C) AndINode(in2, ConNode::make(C, TypeInt::make(mask))); 2657 n->set_req(2, shift); 2658 } 2659 } 2660 if (in2->outcnt() == 0) { // Remove dead node 2661 in2->disconnect_inputs(NULL); 2662 } 2663 } 2664 break; 2665 default: 2666 assert( !n->is_Call(), "" ); 2667 assert( !n->is_Mem(), "" ); 2668 break; 2669 } 2670 2671 // Collect CFG split points 2672 if (n->is_MultiBranch()) 2673 frc._tests.push(n); 2674 } 2675 2676 //------------------------------final_graph_reshaping_walk--------------------- 2677 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 2678 // requires that the walk visits a node's inputs before visiting the node. 2679 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 2680 ResourceArea *area = Thread::current()->resource_area(); 2681 Unique_Node_List sfpt(area); 2682 2683 frc._visited.set(root->_idx); // first, mark node as visited 2684 uint cnt = root->req(); 2685 Node *n = root; 2686 uint i = 0; 2687 while (true) { 2688 if (i < cnt) { 2689 // Place all non-visited non-null inputs onto stack 2690 Node* m = n->in(i); 2691 ++i; 2692 if (m != NULL && !frc._visited.test_set(m->_idx)) { 2693 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 2694 sfpt.push(m); 2695 cnt = m->req(); 2696 nstack.push(n, i); // put on stack parent and next input's index 2697 n = m; 2698 i = 0; 2699 } 2700 } else { 2701 // Now do post-visit work 2702 final_graph_reshaping_impl( n, frc ); 2703 if (nstack.is_empty()) 2704 break; // finished 2705 n = nstack.node(); // Get node from stack 2706 cnt = n->req(); 2707 i = nstack.index(); 2708 nstack.pop(); // Shift to the next node on stack 2709 } 2710 } 2711 2712 // Skip next transformation if compressed oops are not used. 2713 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 2714 (!UseCompressedOops && !UseCompressedKlassPointers)) 2715 return; 2716 2717 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 2718 // It could be done for an uncommon traps or any safepoints/calls 2719 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 2720 while (sfpt.size() > 0) { 2721 n = sfpt.pop(); 2722 JVMState *jvms = n->as_SafePoint()->jvms(); 2723 assert(jvms != NULL, "sanity"); 2724 int start = jvms->debug_start(); 2725 int end = n->req(); 2726 bool is_uncommon = (n->is_CallStaticJava() && 2727 n->as_CallStaticJava()->uncommon_trap_request() != 0); 2728 for (int j = start; j < end; j++) { 2729 Node* in = n->in(j); 2730 if (in->is_DecodeNarrowPtr()) { 2731 bool safe_to_skip = true; 2732 if (!is_uncommon ) { 2733 // Is it safe to skip? 2734 for (uint i = 0; i < in->outcnt(); i++) { 2735 Node* u = in->raw_out(i); 2736 if (!u->is_SafePoint() || 2737 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 2738 safe_to_skip = false; 2739 } 2740 } 2741 } 2742 if (safe_to_skip) { 2743 n->set_req(j, in->in(1)); 2744 } 2745 if (in->outcnt() == 0) { 2746 in->disconnect_inputs(NULL); 2747 } 2748 } 2749 } 2750 } 2751 } 2752 2753 //------------------------------final_graph_reshaping-------------------------- 2754 // Final Graph Reshaping. 2755 // 2756 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 2757 // and not commoned up and forced early. Must come after regular 2758 // optimizations to avoid GVN undoing the cloning. Clone constant 2759 // inputs to Loop Phis; these will be split by the allocator anyways. 2760 // Remove Opaque nodes. 2761 // (2) Move last-uses by commutative operations to the left input to encourage 2762 // Intel update-in-place two-address operations and better register usage 2763 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 2764 // calls canonicalizing them back. 2765 // (3) Count the number of double-precision FP ops, single-precision FP ops 2766 // and call sites. On Intel, we can get correct rounding either by 2767 // forcing singles to memory (requires extra stores and loads after each 2768 // FP bytecode) or we can set a rounding mode bit (requires setting and 2769 // clearing the mode bit around call sites). The mode bit is only used 2770 // if the relative frequency of single FP ops to calls is low enough. 2771 // This is a key transform for SPEC mpeg_audio. 2772 // (4) Detect infinite loops; blobs of code reachable from above but not 2773 // below. Several of the Code_Gen algorithms fail on such code shapes, 2774 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 2775 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 2776 // Detection is by looking for IfNodes where only 1 projection is 2777 // reachable from below or CatchNodes missing some targets. 2778 // (5) Assert for insane oop offsets in debug mode. 2779 2780 bool Compile::final_graph_reshaping() { 2781 // an infinite loop may have been eliminated by the optimizer, 2782 // in which case the graph will be empty. 2783 if (root()->req() == 1) { 2784 record_method_not_compilable("trivial infinite loop"); 2785 return true; 2786 } 2787 2788 Final_Reshape_Counts frc; 2789 2790 // Visit everybody reachable! 2791 // Allocate stack of size C->unique()/2 to avoid frequent realloc 2792 Node_Stack nstack(unique() >> 1); 2793 final_graph_reshaping_walk(nstack, root(), frc); 2794 2795 // Check for unreachable (from below) code (i.e., infinite loops). 2796 for( uint i = 0; i < frc._tests.size(); i++ ) { 2797 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 2798 // Get number of CFG targets. 2799 // Note that PCTables include exception targets after calls. 2800 uint required_outcnt = n->required_outcnt(); 2801 if (n->outcnt() != required_outcnt) { 2802 // Check for a few special cases. Rethrow Nodes never take the 2803 // 'fall-thru' path, so expected kids is 1 less. 2804 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 2805 if (n->in(0)->in(0)->is_Call()) { 2806 CallNode *call = n->in(0)->in(0)->as_Call(); 2807 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 2808 required_outcnt--; // Rethrow always has 1 less kid 2809 } else if (call->req() > TypeFunc::Parms && 2810 call->is_CallDynamicJava()) { 2811 // Check for null receiver. In such case, the optimizer has 2812 // detected that the virtual call will always result in a null 2813 // pointer exception. The fall-through projection of this CatchNode 2814 // will not be populated. 2815 Node *arg0 = call->in(TypeFunc::Parms); 2816 if (arg0->is_Type() && 2817 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 2818 required_outcnt--; 2819 } 2820 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 2821 call->req() > TypeFunc::Parms+1 && 2822 call->is_CallStaticJava()) { 2823 // Check for negative array length. In such case, the optimizer has 2824 // detected that the allocation attempt will always result in an 2825 // exception. There is no fall-through projection of this CatchNode . 2826 Node *arg1 = call->in(TypeFunc::Parms+1); 2827 if (arg1->is_Type() && 2828 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 2829 required_outcnt--; 2830 } 2831 } 2832 } 2833 } 2834 // Recheck with a better notion of 'required_outcnt' 2835 if (n->outcnt() != required_outcnt) { 2836 record_method_not_compilable("malformed control flow"); 2837 return true; // Not all targets reachable! 2838 } 2839 } 2840 // Check that I actually visited all kids. Unreached kids 2841 // must be infinite loops. 2842 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 2843 if (!frc._visited.test(n->fast_out(j)->_idx)) { 2844 record_method_not_compilable("infinite loop"); 2845 return true; // Found unvisited kid; must be unreach 2846 } 2847 } 2848 2849 // If original bytecodes contained a mixture of floats and doubles 2850 // check if the optimizer has made it homogenous, item (3). 2851 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 2852 frc.get_float_count() > 32 && 2853 frc.get_double_count() == 0 && 2854 (10 * frc.get_call_count() < frc.get_float_count()) ) { 2855 set_24_bit_selection_and_mode( false, true ); 2856 } 2857 2858 set_java_calls(frc.get_java_call_count()); 2859 set_inner_loops(frc.get_inner_loop_count()); 2860 2861 // No infinite loops, no reason to bail out. 2862 return false; 2863 } 2864 2865 //-----------------------------too_many_traps---------------------------------- 2866 // Report if there are too many traps at the current method and bci. 2867 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 2868 bool Compile::too_many_traps(ciMethod* method, 2869 int bci, 2870 Deoptimization::DeoptReason reason) { 2871 ciMethodData* md = method->method_data(); 2872 if (md->is_empty()) { 2873 // Assume the trap has not occurred, or that it occurred only 2874 // because of a transient condition during start-up in the interpreter. 2875 return false; 2876 } 2877 if (md->has_trap_at(bci, reason) != 0) { 2878 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 2879 // Also, if there are multiple reasons, or if there is no per-BCI record, 2880 // assume the worst. 2881 if (log()) 2882 log()->elem("observe trap='%s' count='%d'", 2883 Deoptimization::trap_reason_name(reason), 2884 md->trap_count(reason)); 2885 return true; 2886 } else { 2887 // Ignore method/bci and see if there have been too many globally. 2888 return too_many_traps(reason, md); 2889 } 2890 } 2891 2892 // Less-accurate variant which does not require a method and bci. 2893 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 2894 ciMethodData* logmd) { 2895 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 2896 // Too many traps globally. 2897 // Note that we use cumulative trap_count, not just md->trap_count. 2898 if (log()) { 2899 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 2900 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 2901 Deoptimization::trap_reason_name(reason), 2902 mcount, trap_count(reason)); 2903 } 2904 return true; 2905 } else { 2906 // The coast is clear. 2907 return false; 2908 } 2909 } 2910 2911 //--------------------------too_many_recompiles-------------------------------- 2912 // Report if there are too many recompiles at the current method and bci. 2913 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 2914 // Is not eager to return true, since this will cause the compiler to use 2915 // Action_none for a trap point, to avoid too many recompilations. 2916 bool Compile::too_many_recompiles(ciMethod* method, 2917 int bci, 2918 Deoptimization::DeoptReason reason) { 2919 ciMethodData* md = method->method_data(); 2920 if (md->is_empty()) { 2921 // Assume the trap has not occurred, or that it occurred only 2922 // because of a transient condition during start-up in the interpreter. 2923 return false; 2924 } 2925 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 2926 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 2927 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 2928 Deoptimization::DeoptReason per_bc_reason 2929 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 2930 if ((per_bc_reason == Deoptimization::Reason_none 2931 || md->has_trap_at(bci, reason) != 0) 2932 // The trap frequency measure we care about is the recompile count: 2933 && md->trap_recompiled_at(bci) 2934 && md->overflow_recompile_count() >= bc_cutoff) { 2935 // Do not emit a trap here if it has already caused recompilations. 2936 // Also, if there are multiple reasons, or if there is no per-BCI record, 2937 // assume the worst. 2938 if (log()) 2939 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 2940 Deoptimization::trap_reason_name(reason), 2941 md->trap_count(reason), 2942 md->overflow_recompile_count()); 2943 return true; 2944 } else if (trap_count(reason) != 0 2945 && decompile_count() >= m_cutoff) { 2946 // Too many recompiles globally, and we have seen this sort of trap. 2947 // Use cumulative decompile_count, not just md->decompile_count. 2948 if (log()) 2949 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 2950 Deoptimization::trap_reason_name(reason), 2951 md->trap_count(reason), trap_count(reason), 2952 md->decompile_count(), decompile_count()); 2953 return true; 2954 } else { 2955 // The coast is clear. 2956 return false; 2957 } 2958 } 2959 2960 2961 #ifndef PRODUCT 2962 //------------------------------verify_graph_edges--------------------------- 2963 // Walk the Graph and verify that there is a one-to-one correspondence 2964 // between Use-Def edges and Def-Use edges in the graph. 2965 void Compile::verify_graph_edges(bool no_dead_code) { 2966 if (VerifyGraphEdges) { 2967 ResourceArea *area = Thread::current()->resource_area(); 2968 Unique_Node_List visited(area); 2969 // Call recursive graph walk to check edges 2970 _root->verify_edges(visited); 2971 if (no_dead_code) { 2972 // Now make sure that no visited node is used by an unvisited node. 2973 bool dead_nodes = 0; 2974 Unique_Node_List checked(area); 2975 while (visited.size() > 0) { 2976 Node* n = visited.pop(); 2977 checked.push(n); 2978 for (uint i = 0; i < n->outcnt(); i++) { 2979 Node* use = n->raw_out(i); 2980 if (checked.member(use)) continue; // already checked 2981 if (visited.member(use)) continue; // already in the graph 2982 if (use->is_Con()) continue; // a dead ConNode is OK 2983 // At this point, we have found a dead node which is DU-reachable. 2984 if (dead_nodes++ == 0) 2985 tty->print_cr("*** Dead nodes reachable via DU edges:"); 2986 use->dump(2); 2987 tty->print_cr("---"); 2988 checked.push(use); // No repeats; pretend it is now checked. 2989 } 2990 } 2991 assert(dead_nodes == 0, "using nodes must be reachable from root"); 2992 } 2993 } 2994 } 2995 #endif 2996 2997 // The Compile object keeps track of failure reasons separately from the ciEnv. 2998 // This is required because there is not quite a 1-1 relation between the 2999 // ciEnv and its compilation task and the Compile object. Note that one 3000 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3001 // to backtrack and retry without subsuming loads. Other than this backtracking 3002 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3003 // by the logic in C2Compiler. 3004 void Compile::record_failure(const char* reason) { 3005 if (log() != NULL) { 3006 log()->elem("failure reason='%s' phase='compile'", reason); 3007 } 3008 if (_failure_reason == NULL) { 3009 // Record the first failure reason. 3010 _failure_reason = reason; 3011 } 3012 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3013 C->print_method(_failure_reason); 3014 } 3015 _root = NULL; // flush the graph, too 3016 } 3017 3018 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3019 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false) 3020 { 3021 if (dolog) { 3022 C = Compile::current(); 3023 _log = C->log(); 3024 } else { 3025 C = NULL; 3026 _log = NULL; 3027 } 3028 if (_log != NULL) { 3029 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique()); 3030 _log->stamp(); 3031 _log->end_head(); 3032 } 3033 } 3034 3035 Compile::TracePhase::~TracePhase() { 3036 if (_log != NULL) { 3037 _log->done("phase nodes='%d'", C->unique()); 3038 } 3039 } 3040 3041 //============================================================================= 3042 // Two Constant's are equal when the type and the value are equal. 3043 bool Compile::Constant::operator==(const Constant& other) { 3044 if (type() != other.type() ) return false; 3045 if (can_be_reused() != other.can_be_reused()) return false; 3046 // For floating point values we compare the bit pattern. 3047 switch (type()) { 3048 case T_FLOAT: return (_v._value.i == other._v._value.i); 3049 case T_LONG: 3050 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3051 case T_OBJECT: 3052 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3053 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3054 case T_METADATA: return (_v._metadata == other._v._metadata); 3055 default: ShouldNotReachHere(); 3056 } 3057 return false; 3058 } 3059 3060 static int type_to_size_in_bytes(BasicType t) { 3061 switch (t) { 3062 case T_LONG: return sizeof(jlong ); 3063 case T_FLOAT: return sizeof(jfloat ); 3064 case T_DOUBLE: return sizeof(jdouble); 3065 case T_METADATA: return sizeof(Metadata*); 3066 // We use T_VOID as marker for jump-table entries (labels) which 3067 // need an internal word relocation. 3068 case T_VOID: 3069 case T_ADDRESS: 3070 case T_OBJECT: return sizeof(jobject); 3071 } 3072 3073 ShouldNotReachHere(); 3074 return -1; 3075 } 3076 3077 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3078 // sort descending 3079 if (a->freq() > b->freq()) return -1; 3080 if (a->freq() < b->freq()) return 1; 3081 return 0; 3082 } 3083 3084 void Compile::ConstantTable::calculate_offsets_and_size() { 3085 // First, sort the array by frequencies. 3086 _constants.sort(qsort_comparator); 3087 3088 #ifdef ASSERT 3089 // Make sure all jump-table entries were sorted to the end of the 3090 // array (they have a negative frequency). 3091 bool found_void = false; 3092 for (int i = 0; i < _constants.length(); i++) { 3093 Constant con = _constants.at(i); 3094 if (con.type() == T_VOID) 3095 found_void = true; // jump-tables 3096 else 3097 assert(!found_void, "wrong sorting"); 3098 } 3099 #endif 3100 3101 int offset = 0; 3102 for (int i = 0; i < _constants.length(); i++) { 3103 Constant* con = _constants.adr_at(i); 3104 3105 // Align offset for type. 3106 int typesize = type_to_size_in_bytes(con->type()); 3107 offset = align_size_up(offset, typesize); 3108 con->set_offset(offset); // set constant's offset 3109 3110 if (con->type() == T_VOID) { 3111 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3112 offset = offset + typesize * n->outcnt(); // expand jump-table 3113 } else { 3114 offset = offset + typesize; 3115 } 3116 } 3117 3118 // Align size up to the next section start (which is insts; see 3119 // CodeBuffer::align_at_start). 3120 assert(_size == -1, "already set?"); 3121 _size = align_size_up(offset, CodeEntryAlignment); 3122 } 3123 3124 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3125 MacroAssembler _masm(&cb); 3126 for (int i = 0; i < _constants.length(); i++) { 3127 Constant con = _constants.at(i); 3128 address constant_addr; 3129 switch (con.type()) { 3130 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3131 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3132 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3133 case T_OBJECT: { 3134 jobject obj = con.get_jobject(); 3135 int oop_index = _masm.oop_recorder()->find_index(obj); 3136 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3137 break; 3138 } 3139 case T_ADDRESS: { 3140 address addr = (address) con.get_jobject(); 3141 constant_addr = _masm.address_constant(addr); 3142 break; 3143 } 3144 // We use T_VOID as marker for jump-table entries (labels) which 3145 // need an internal word relocation. 3146 case T_VOID: { 3147 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3148 // Fill the jump-table with a dummy word. The real value is 3149 // filled in later in fill_jump_table. 3150 address dummy = (address) n; 3151 constant_addr = _masm.address_constant(dummy); 3152 // Expand jump-table 3153 for (uint i = 1; i < n->outcnt(); i++) { 3154 address temp_addr = _masm.address_constant(dummy + i); 3155 assert(temp_addr, "consts section too small"); 3156 } 3157 break; 3158 } 3159 case T_METADATA: { 3160 Metadata* obj = con.get_metadata(); 3161 int metadata_index = _masm.oop_recorder()->find_index(obj); 3162 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3163 break; 3164 } 3165 default: ShouldNotReachHere(); 3166 } 3167 assert(constant_addr, "consts section too small"); 3168 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3169 } 3170 } 3171 3172 int Compile::ConstantTable::find_offset(Constant& con) const { 3173 int idx = _constants.find(con); 3174 assert(idx != -1, "constant must be in constant table"); 3175 int offset = _constants.at(idx).offset(); 3176 assert(offset != -1, "constant table not emitted yet?"); 3177 return offset; 3178 } 3179 3180 void Compile::ConstantTable::add(Constant& con) { 3181 if (con.can_be_reused()) { 3182 int idx = _constants.find(con); 3183 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3184 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3185 return; 3186 } 3187 } 3188 (void) _constants.append(con); 3189 } 3190 3191 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3192 Block* b = Compile::current()->cfg()->_bbs[n->_idx]; 3193 Constant con(type, value, b->_freq); 3194 add(con); 3195 return con; 3196 } 3197 3198 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3199 Constant con(metadata); 3200 add(con); 3201 return con; 3202 } 3203 3204 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3205 jvalue value; 3206 BasicType type = oper->type()->basic_type(); 3207 switch (type) { 3208 case T_LONG: value.j = oper->constantL(); break; 3209 case T_FLOAT: value.f = oper->constantF(); break; 3210 case T_DOUBLE: value.d = oper->constantD(); break; 3211 case T_OBJECT: 3212 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3213 case T_METADATA: return add((Metadata*)oper->constant()); break; 3214 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3215 } 3216 return add(n, type, value); 3217 } 3218 3219 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3220 jvalue value; 3221 // We can use the node pointer here to identify the right jump-table 3222 // as this method is called from Compile::Fill_buffer right before 3223 // the MachNodes are emitted and the jump-table is filled (means the 3224 // MachNode pointers do not change anymore). 3225 value.l = (jobject) n; 3226 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3227 add(con); 3228 return con; 3229 } 3230 3231 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3232 // If called from Compile::scratch_emit_size do nothing. 3233 if (Compile::current()->in_scratch_emit_size()) return; 3234 3235 assert(labels.is_nonempty(), "must be"); 3236 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3237 3238 // Since MachConstantNode::constant_offset() also contains 3239 // table_base_offset() we need to subtract the table_base_offset() 3240 // to get the plain offset into the constant table. 3241 int offset = n->constant_offset() - table_base_offset(); 3242 3243 MacroAssembler _masm(&cb); 3244 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3245 3246 for (uint i = 0; i < n->outcnt(); i++) { 3247 address* constant_addr = &jump_table_base[i]; 3248 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); 3249 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3250 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3251 } 3252 }