1 /* 2 * Copyright (c) 1997, 2021, 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 "jvm_io.h" 27 #include "asm/macroAssembler.hpp" 28 #include "asm/macroAssembler.inline.hpp" 29 #include "ci/ciReplay.hpp" 30 #include "classfile/javaClasses.hpp" 31 #include "code/exceptionHandlerTable.hpp" 32 #include "code/nmethod.hpp" 33 #include "compiler/compileBroker.hpp" 34 #include "compiler/compileLog.hpp" 35 #include "compiler/disassembler.hpp" 36 #include "compiler/oopMap.hpp" 37 #include "gc/shared/barrierSet.hpp" 38 #include "gc/shared/c2/barrierSetC2.hpp" 39 #include "jfr/jfrEvents.hpp" 40 #include "memory/resourceArea.hpp" 41 #include "opto/addnode.hpp" 42 #include "opto/block.hpp" 43 #include "opto/c2compiler.hpp" 44 #include "opto/callGenerator.hpp" 45 #include "opto/callnode.hpp" 46 #include "opto/castnode.hpp" 47 #include "opto/cfgnode.hpp" 48 #include "opto/chaitin.hpp" 49 #include "opto/compile.hpp" 50 #include "opto/connode.hpp" 51 #include "opto/convertnode.hpp" 52 #include "opto/divnode.hpp" 53 #include "opto/escape.hpp" 54 #include "opto/idealGraphPrinter.hpp" 55 #include "opto/loopnode.hpp" 56 #include "opto/machnode.hpp" 57 #include "opto/macro.hpp" 58 #include "opto/matcher.hpp" 59 #include "opto/mathexactnode.hpp" 60 #include "opto/memnode.hpp" 61 #include "opto/mulnode.hpp" 62 #include "opto/narrowptrnode.hpp" 63 #include "opto/node.hpp" 64 #include "opto/opcodes.hpp" 65 #include "opto/output.hpp" 66 #include "opto/parse.hpp" 67 #include "opto/phaseX.hpp" 68 #include "opto/rootnode.hpp" 69 #include "opto/runtime.hpp" 70 #include "opto/stringopts.hpp" 71 #include "opto/type.hpp" 72 #include "opto/vector.hpp" 73 #include "opto/vectornode.hpp" 74 #include "runtime/globals_extension.hpp" 75 #include "runtime/sharedRuntime.hpp" 76 #include "runtime/signature.hpp" 77 #include "runtime/stubRoutines.hpp" 78 #include "runtime/timer.hpp" 79 #include "utilities/align.hpp" 80 #include "utilities/copy.hpp" 81 #include "utilities/macros.hpp" 82 #include "utilities/resourceHash.hpp" 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 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 class IntrinsicDescPair { 101 private: 102 ciMethod* _m; 103 bool _is_virtual; 104 public: 105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 107 ciMethod* m= elt->method(); 108 ciMethod* key_m = key->_m; 109 if (key_m < m) return -1; 110 else if (key_m > m) return 1; 111 else { 112 bool is_virtual = elt->is_virtual(); 113 bool key_virtual = key->_is_virtual; 114 if (key_virtual < is_virtual) return -1; 115 else if (key_virtual > is_virtual) return 1; 116 else return 0; 117 } 118 } 119 }; 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 121 #ifdef ASSERT 122 for (int i = 1; i < _intrinsics.length(); i++) { 123 CallGenerator* cg1 = _intrinsics.at(i-1); 124 CallGenerator* cg2 = _intrinsics.at(i); 125 assert(cg1->method() != cg2->method() 126 ? cg1->method() < cg2->method() 127 : cg1->is_virtual() < cg2->is_virtual(), 128 "compiler intrinsics list must stay sorted"); 129 } 130 #endif 131 IntrinsicDescPair pair(m, is_virtual); 132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 bool found = false; 137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 138 assert(!found, "registering twice"); 139 _intrinsics.insert_before(index, cg); 140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 141 } 142 143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 144 assert(m->is_loaded(), "don't try this on unloaded methods"); 145 if (_intrinsics.length() > 0) { 146 bool found = false; 147 int index = intrinsic_insertion_index(m, is_virtual, found); 148 if (found) { 149 return _intrinsics.at(index); 150 } 151 } 152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 153 if (m->intrinsic_id() != vmIntrinsics::_none && 154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 156 if (cg != NULL) { 157 // Save it for next time: 158 register_intrinsic(cg); 159 return cg; 160 } else { 161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 162 } 163 } 164 return NULL; 165 } 166 167 // Compile::make_vm_intrinsic is defined in library_call.cpp. 168 169 #ifndef PRODUCT 170 // statistics gathering... 171 172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 174 175 inline int as_int(vmIntrinsics::ID id) { 176 return vmIntrinsics::as_int(id); 177 } 178 179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 181 int oflags = _intrinsic_hist_flags[as_int(id)]; 182 assert(flags != 0, "what happened?"); 183 if (is_virtual) { 184 flags |= _intrinsic_virtual; 185 } 186 bool changed = (flags != oflags); 187 if ((flags & _intrinsic_worked) != 0) { 188 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 189 if (count == 1) { 190 changed = true; // first time 191 } 192 // increment the overall count also: 193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 194 } 195 if (changed) { 196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 197 // Something changed about the intrinsic's virtuality. 198 if ((flags & _intrinsic_virtual) != 0) { 199 // This is the first use of this intrinsic as a virtual call. 200 if (oflags != 0) { 201 // We already saw it as a non-virtual, so note both cases. 202 flags |= _intrinsic_both; 203 } 204 } else if ((oflags & _intrinsic_both) == 0) { 205 // This is the first use of this intrinsic as a non-virtual 206 flags |= _intrinsic_both; 207 } 208 } 209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 210 } 211 // update the overall flags also: 212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 213 return changed; 214 } 215 216 static char* format_flags(int flags, char* buf) { 217 buf[0] = 0; 218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 223 if (buf[0] == 0) strcat(buf, ","); 224 assert(buf[0] == ',', "must be"); 225 return &buf[1]; 226 } 227 228 void Compile::print_intrinsic_statistics() { 229 char flagsbuf[100]; 230 ttyLocker ttyl; 231 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 232 tty->print_cr("Compiler intrinsic usage:"); 233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 234 if (total == 0) total = 1; // avoid div0 in case of no successes 235 #define PRINT_STAT_LINE(name, c, f) \ 236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 237 for (auto id : EnumRange<vmIntrinsicID>{}) { 238 int flags = _intrinsic_hist_flags[as_int(id)]; 239 juint count = _intrinsic_hist_count[as_int(id)]; 240 if ((flags | count) != 0) { 241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 242 } 243 } 244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 245 if (xtty != NULL) xtty->tail("statistics"); 246 } 247 248 void Compile::print_statistics() { 249 { ttyLocker ttyl; 250 if (xtty != NULL) xtty->head("statistics type='opto'"); 251 Parse::print_statistics(); 252 PhaseCCP::print_statistics(); 253 PhaseRegAlloc::print_statistics(); 254 PhaseOutput::print_statistics(); 255 PhasePeephole::print_statistics(); 256 PhaseIdealLoop::print_statistics(); 257 if (xtty != NULL) xtty->tail("statistics"); 258 } 259 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 260 // put this under its own <statistics> element. 261 print_intrinsic_statistics(); 262 } 263 } 264 #endif //PRODUCT 265 266 void Compile::gvn_replace_by(Node* n, Node* nn) { 267 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 268 Node* use = n->last_out(i); 269 bool is_in_table = initial_gvn()->hash_delete(use); 270 uint uses_found = 0; 271 for (uint j = 0; j < use->len(); j++) { 272 if (use->in(j) == n) { 273 if (j < use->req()) 274 use->set_req(j, nn); 275 else 276 use->set_prec(j, nn); 277 uses_found++; 278 } 279 } 280 if (is_in_table) { 281 // reinsert into table 282 initial_gvn()->hash_find_insert(use); 283 } 284 record_for_igvn(use); 285 i -= uses_found; // we deleted 1 or more copies of this edge 286 } 287 } 288 289 290 // Identify all nodes that are reachable from below, useful. 291 // Use breadth-first pass that records state in a Unique_Node_List, 292 // recursive traversal is slower. 293 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 294 int estimated_worklist_size = live_nodes(); 295 useful.map( estimated_worklist_size, NULL ); // preallocate space 296 297 // Initialize worklist 298 if (root() != NULL) { useful.push(root()); } 299 // If 'top' is cached, declare it useful to preserve cached node 300 if( cached_top_node() ) { useful.push(cached_top_node()); } 301 302 // Push all useful nodes onto the list, breadthfirst 303 for( uint next = 0; next < useful.size(); ++next ) { 304 assert( next < unique(), "Unique useful nodes < total nodes"); 305 Node *n = useful.at(next); 306 uint max = n->len(); 307 for( uint i = 0; i < max; ++i ) { 308 Node *m = n->in(i); 309 if (not_a_node(m)) continue; 310 useful.push(m); 311 } 312 } 313 } 314 315 // Update dead_node_list with any missing dead nodes using useful 316 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 317 void Compile::update_dead_node_list(Unique_Node_List &useful) { 318 uint max_idx = unique(); 319 VectorSet& useful_node_set = useful.member_set(); 320 321 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 322 // If node with index node_idx is not in useful set, 323 // mark it as dead in dead node list. 324 if (!useful_node_set.test(node_idx)) { 325 record_dead_node(node_idx); 326 } 327 } 328 } 329 330 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 331 int shift = 0; 332 for (int i = 0; i < inlines->length(); i++) { 333 CallGenerator* cg = inlines->at(i); 334 if (useful.member(cg->call_node())) { 335 if (shift > 0) { 336 inlines->at_put(i - shift, cg); 337 } 338 } else { 339 shift++; // skip over the dead element 340 } 341 } 342 if (shift > 0) { 343 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 344 } 345 } 346 347 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 348 assert(dead != NULL && dead->is_Call(), "sanity"); 349 int found = 0; 350 for (int i = 0; i < inlines->length(); i++) { 351 if (inlines->at(i)->call_node() == dead) { 352 inlines->remove_at(i); 353 found++; 354 NOT_DEBUG( break; ) // elements are unique, so exit early 355 } 356 } 357 assert(found <= 1, "not unique"); 358 } 359 360 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) { 361 for (int i = node_list.length() - 1; i >= 0; i--) { 362 Node* n = node_list.at(i); 363 if (!useful.member(n)) { 364 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 365 } 366 } 367 } 368 369 void Compile::remove_useless_node(Node* dead) { 370 remove_modified_node(dead); 371 372 // Constant node that has no out-edges and has only one in-edge from 373 // root is usually dead. However, sometimes reshaping walk makes 374 // it reachable by adding use edges. So, we will NOT count Con nodes 375 // as dead to be conservative about the dead node count at any 376 // given time. 377 if (!dead->is_Con()) { 378 record_dead_node(dead->_idx); 379 } 380 if (dead->is_macro()) { 381 remove_macro_node(dead); 382 } 383 if (dead->is_expensive()) { 384 remove_expensive_node(dead); 385 } 386 if (dead->Opcode() == Op_Opaque4) { 387 remove_skeleton_predicate_opaq(dead); 388 } 389 if (dead->for_post_loop_opts_igvn()) { 390 remove_from_post_loop_opts_igvn(dead); 391 } 392 if (dead->is_Call()) { 393 remove_useless_late_inlines( &_late_inlines, dead); 394 remove_useless_late_inlines( &_string_late_inlines, dead); 395 remove_useless_late_inlines( &_boxing_late_inlines, dead); 396 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 397 } 398 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 399 bs->unregister_potential_barrier_node(dead); 400 } 401 402 // Disconnect all useless nodes by disconnecting those at the boundary. 403 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) { 404 uint next = 0; 405 while (next < useful.size()) { 406 Node *n = useful.at(next++); 407 if (n->is_SafePoint()) { 408 // We're done with a parsing phase. Replaced nodes are not valid 409 // beyond that point. 410 n->as_SafePoint()->delete_replaced_nodes(); 411 } 412 // Use raw traversal of out edges since this code removes out edges 413 int max = n->outcnt(); 414 for (int j = 0; j < max; ++j) { 415 Node* child = n->raw_out(j); 416 if (!useful.member(child)) { 417 assert(!child->is_top() || child != top(), 418 "If top is cached in Compile object it is in useful list"); 419 // Only need to remove this out-edge to the useless node 420 n->raw_del_out(j); 421 --j; 422 --max; 423 } 424 } 425 if (n->outcnt() == 1 && n->has_special_unique_user()) { 426 worklist->push(n->unique_out()); 427 } 428 } 429 430 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 431 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes 432 remove_useless_nodes(_skeleton_predicate_opaqs, useful); 433 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 434 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 435 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 436 #ifdef ASSERT 437 if (_modified_nodes != NULL) { 438 _modified_nodes->remove_useless_nodes(useful.member_set()); 439 } 440 #endif 441 442 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 443 bs->eliminate_useless_gc_barriers(useful, this); 444 // clean up the late inline lists 445 remove_useless_late_inlines( &_late_inlines, useful); 446 remove_useless_late_inlines( &_string_late_inlines, useful); 447 remove_useless_late_inlines( &_boxing_late_inlines, useful); 448 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 449 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 450 } 451 452 // ============================================================================ 453 //------------------------------CompileWrapper--------------------------------- 454 class CompileWrapper : public StackObj { 455 Compile *const _compile; 456 public: 457 CompileWrapper(Compile* compile); 458 459 ~CompileWrapper(); 460 }; 461 462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 463 // the Compile* pointer is stored in the current ciEnv: 464 ciEnv* env = compile->env(); 465 assert(env == ciEnv::current(), "must already be a ciEnv active"); 466 assert(env->compiler_data() == NULL, "compile already active?"); 467 env->set_compiler_data(compile); 468 assert(compile == Compile::current(), "sanity"); 469 470 compile->set_type_dict(NULL); 471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 472 compile->clone_map().set_clone_idx(0); 473 compile->set_type_last_size(0); 474 compile->set_last_tf(NULL, NULL); 475 compile->set_indexSet_arena(NULL); 476 compile->set_indexSet_free_block_list(NULL); 477 compile->init_type_arena(); 478 Type::Initialize(compile); 479 _compile->begin_method(); 480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 481 } 482 CompileWrapper::~CompileWrapper() { 483 _compile->end_method(); 484 _compile->env()->set_compiler_data(NULL); 485 } 486 487 488 //----------------------------print_compile_messages--------------------------- 489 void Compile::print_compile_messages() { 490 #ifndef PRODUCT 491 // Check if recompiling 492 if (_subsume_loads == false && PrintOpto) { 493 // Recompiling without allowing machine instructions to subsume loads 494 tty->print_cr("*********************************************************"); 495 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 496 tty->print_cr("*********************************************************"); 497 } 498 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 499 // Recompiling without escape analysis 500 tty->print_cr("*********************************************************"); 501 tty->print_cr("** Bailout: Recompile without escape analysis **"); 502 tty->print_cr("*********************************************************"); 503 } 504 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 505 // Recompiling without boxing elimination 506 tty->print_cr("*********************************************************"); 507 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 508 tty->print_cr("*********************************************************"); 509 } 510 if ((_do_locks_coarsening != EliminateLocks) && PrintOpto) { 511 // Recompiling without locks coarsening 512 tty->print_cr("*********************************************************"); 513 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 514 tty->print_cr("*********************************************************"); 515 } 516 if (env()->break_at_compile()) { 517 // Open the debugger when compiling this method. 518 tty->print("### Breaking when compiling: "); 519 method()->print_short_name(); 520 tty->cr(); 521 BREAKPOINT; 522 } 523 524 if( PrintOpto ) { 525 if (is_osr_compilation()) { 526 tty->print("[OSR]%3d", _compile_id); 527 } else { 528 tty->print("%3d", _compile_id); 529 } 530 } 531 #endif 532 } 533 534 // ============================================================================ 535 //------------------------------Compile standard------------------------------- 536 debug_only( int Compile::_debug_idx = 100000; ) 537 538 // Compile a method. entry_bci is -1 for normal compilations and indicates 539 // the continuation bci for on stack replacement. 540 541 542 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 543 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, 544 bool do_locks_coarsening, bool install_code, DirectiveSet* directive) 545 : Phase(Compiler), 546 _compile_id(ci_env->compile_id()), 547 _subsume_loads(subsume_loads), 548 _do_escape_analysis(do_escape_analysis), 549 _install_code(install_code), 550 _eliminate_boxing(eliminate_boxing), 551 _do_locks_coarsening(do_locks_coarsening), 552 _method(target), 553 _entry_bci(osr_bci), 554 _stub_function(NULL), 555 _stub_name(NULL), 556 _stub_entry_point(NULL), 557 _max_node_limit(MaxNodeLimit), 558 _post_loop_opts_phase(false), 559 _inlining_progress(false), 560 _inlining_incrementally(false), 561 _do_cleanup(false), 562 _has_reserved_stack_access(target->has_reserved_stack_access()), 563 #ifndef PRODUCT 564 _igv_idx(0), 565 _trace_opto_output(directive->TraceOptoOutputOption), 566 _print_ideal(directive->PrintIdealOption), 567 #endif 568 _has_method_handle_invokes(false), 569 _clinit_barrier_on_entry(false), 570 _stress_seed(0), 571 _comp_arena(mtCompiler), 572 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 573 _env(ci_env), 574 _directive(directive), 575 _log(ci_env->log()), 576 _failure_reason(NULL), 577 _intrinsics (comp_arena(), 0, 0, NULL), 578 _macro_nodes (comp_arena(), 8, 0, NULL), 579 _predicate_opaqs (comp_arena(), 8, 0, NULL), 580 _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL), 581 _expensive_nodes (comp_arena(), 8, 0, NULL), 582 _for_post_loop_igvn(comp_arena(), 8, 0, NULL), 583 _coarsened_locks (comp_arena(), 8, 0, NULL), 584 _congraph(NULL), 585 NOT_PRODUCT(_printer(NULL) COMMA) 586 _dead_node_list(comp_arena()), 587 _dead_node_count(0), 588 _node_arena(mtCompiler), 589 _old_arena(mtCompiler), 590 _mach_constant_base_node(NULL), 591 _Compile_types(mtCompiler), 592 _initial_gvn(NULL), 593 _for_igvn(NULL), 594 _late_inlines(comp_arena(), 2, 0, NULL), 595 _string_late_inlines(comp_arena(), 2, 0, NULL), 596 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 597 _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL), 598 _late_inlines_pos(0), 599 _number_of_mh_late_inlines(0), 600 _native_invokers(comp_arena(), 1, 0, NULL), 601 _print_inlining_stream(NULL), 602 _print_inlining_list(NULL), 603 _print_inlining_idx(0), 604 _print_inlining_output(NULL), 605 _replay_inline_data(NULL), 606 _java_calls(0), 607 _inner_loops(0), 608 _interpreter_frame_size(0) 609 #ifndef PRODUCT 610 , _in_dump_cnt(0) 611 #endif 612 { 613 C = this; 614 CompileWrapper cw(this); 615 616 if (CITimeVerbose) { 617 tty->print(" "); 618 target->holder()->name()->print(); 619 tty->print("."); 620 target->print_short_name(); 621 tty->print(" "); 622 } 623 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 624 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 625 626 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 627 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 628 // We can always print a disassembly, either abstract (hex dump) or 629 // with the help of a suitable hsdis library. Thus, we should not 630 // couple print_assembly and print_opto_assembly controls. 631 // But: always print opto and regular assembly on compile command 'print'. 632 bool print_assembly = directive->PrintAssemblyOption; 633 set_print_assembly(print_opto_assembly || print_assembly); 634 #else 635 set_print_assembly(false); // must initialize. 636 #endif 637 638 #ifndef PRODUCT 639 set_parsed_irreducible_loop(false); 640 641 if (directive->ReplayInlineOption) { 642 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 643 } 644 #endif 645 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 646 set_print_intrinsics(directive->PrintIntrinsicsOption); 647 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 648 649 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 650 // Make sure the method being compiled gets its own MDO, 651 // so we can at least track the decompile_count(). 652 // Need MDO to record RTM code generation state. 653 method()->ensure_method_data(); 654 } 655 656 Init(::AliasLevel); 657 658 659 print_compile_messages(); 660 661 _ilt = InlineTree::build_inline_tree_root(); 662 663 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 664 assert(num_alias_types() >= AliasIdxRaw, ""); 665 666 #define MINIMUM_NODE_HASH 1023 667 // Node list that Iterative GVN will start with 668 Unique_Node_List for_igvn(comp_arena()); 669 set_for_igvn(&for_igvn); 670 671 // GVN that will be run immediately on new nodes 672 uint estimated_size = method()->code_size()*4+64; 673 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 674 PhaseGVN gvn(node_arena(), estimated_size); 675 set_initial_gvn(&gvn); 676 677 print_inlining_init(); 678 { // Scope for timing the parser 679 TracePhase tp("parse", &timers[_t_parser]); 680 681 // Put top into the hash table ASAP. 682 initial_gvn()->transform_no_reclaim(top()); 683 684 // Set up tf(), start(), and find a CallGenerator. 685 CallGenerator* cg = NULL; 686 if (is_osr_compilation()) { 687 const TypeTuple *domain = StartOSRNode::osr_domain(); 688 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 689 init_tf(TypeFunc::make(domain, range)); 690 StartNode* s = new StartOSRNode(root(), domain); 691 initial_gvn()->set_type_bottom(s); 692 init_start(s); 693 cg = CallGenerator::for_osr(method(), entry_bci()); 694 } else { 695 // Normal case. 696 init_tf(TypeFunc::make(method())); 697 StartNode* s = new StartNode(root(), tf()->domain()); 698 initial_gvn()->set_type_bottom(s); 699 init_start(s); 700 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 701 // With java.lang.ref.reference.get() we must go through the 702 // intrinsic - even when get() is the root 703 // method of the compile - so that, if necessary, the value in 704 // the referent field of the reference object gets recorded by 705 // the pre-barrier code. 706 cg = find_intrinsic(method(), false); 707 } 708 if (cg == NULL) { 709 float past_uses = method()->interpreter_invocation_count(); 710 float expected_uses = past_uses; 711 cg = CallGenerator::for_inline(method(), expected_uses); 712 } 713 } 714 if (failing()) return; 715 if (cg == NULL) { 716 record_method_not_compilable("cannot parse method"); 717 return; 718 } 719 JVMState* jvms = build_start_state(start(), tf()); 720 if ((jvms = cg->generate(jvms)) == NULL) { 721 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 722 record_method_not_compilable("method parse failed"); 723 } 724 return; 725 } 726 GraphKit kit(jvms); 727 728 if (!kit.stopped()) { 729 // Accept return values, and transfer control we know not where. 730 // This is done by a special, unique ReturnNode bound to root. 731 return_values(kit.jvms()); 732 } 733 734 if (kit.has_exceptions()) { 735 // Any exceptions that escape from this call must be rethrown 736 // to whatever caller is dynamically above us on the stack. 737 // This is done by a special, unique RethrowNode bound to root. 738 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 739 } 740 741 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 742 743 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 744 inline_string_calls(true); 745 } 746 747 if (failing()) return; 748 749 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 750 751 // Remove clutter produced by parsing. 752 if (!failing()) { 753 ResourceMark rm; 754 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 755 } 756 } 757 758 // Note: Large methods are capped off in do_one_bytecode(). 759 if (failing()) return; 760 761 // After parsing, node notes are no longer automagic. 762 // They must be propagated by register_new_node_with_optimizer(), 763 // clone(), or the like. 764 set_default_node_notes(NULL); 765 766 #ifndef PRODUCT 767 if (should_print(1)) { 768 _printer->print_inlining(); 769 } 770 #endif 771 772 if (failing()) return; 773 NOT_PRODUCT( verify_graph_edges(); ) 774 775 // If any phase is randomized for stress testing, seed random number 776 // generation and log the seed for repeatability. 777 if (StressLCM || StressGCM || StressIGVN || StressCCP) { 778 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) { 779 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 780 FLAG_SET_ERGO(StressSeed, _stress_seed); 781 } else { 782 _stress_seed = StressSeed; 783 } 784 if (_log != NULL) { 785 _log->elem("stress_test seed='%u'", _stress_seed); 786 } 787 } 788 789 // Now optimize 790 Optimize(); 791 if (failing()) return; 792 NOT_PRODUCT( verify_graph_edges(); ) 793 794 #ifndef PRODUCT 795 if (print_ideal()) { 796 ttyLocker ttyl; // keep the following output all in one block 797 // This output goes directly to the tty, not the compiler log. 798 // To enable tools to match it up with the compilation activity, 799 // be sure to tag this tty output with the compile ID. 800 if (xtty != NULL) { 801 xtty->head("ideal compile_id='%d'%s", compile_id(), 802 is_osr_compilation() ? " compile_kind='osr'" : 803 ""); 804 } 805 root()->dump(9999); 806 if (xtty != NULL) { 807 xtty->tail("ideal"); 808 } 809 } 810 #endif 811 812 #ifdef ASSERT 813 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 814 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 815 #endif 816 817 // Dump compilation data to replay it. 818 if (directive->DumpReplayOption) { 819 env()->dump_replay_data(_compile_id); 820 } 821 if (directive->DumpInlineOption && (ilt() != NULL)) { 822 env()->dump_inline_data(_compile_id); 823 } 824 825 // Now that we know the size of all the monitors we can add a fixed slot 826 // for the original deopt pc. 827 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 828 set_fixed_slots(next_slot); 829 830 // Compute when to use implicit null checks. Used by matching trap based 831 // nodes and NullCheck optimization. 832 set_allowed_deopt_reasons(); 833 834 // Now generate code 835 Code_Gen(); 836 } 837 838 //------------------------------Compile---------------------------------------- 839 // Compile a runtime stub 840 Compile::Compile( ciEnv* ci_env, 841 TypeFunc_generator generator, 842 address stub_function, 843 const char *stub_name, 844 int is_fancy_jump, 845 bool pass_tls, 846 bool return_pc, 847 DirectiveSet* directive) 848 : Phase(Compiler), 849 _compile_id(0), 850 _subsume_loads(true), 851 _do_escape_analysis(false), 852 _install_code(true), 853 _eliminate_boxing(false), 854 _do_locks_coarsening(false), 855 _method(NULL), 856 _entry_bci(InvocationEntryBci), 857 _stub_function(stub_function), 858 _stub_name(stub_name), 859 _stub_entry_point(NULL), 860 _max_node_limit(MaxNodeLimit), 861 _post_loop_opts_phase(false), 862 _inlining_progress(false), 863 _inlining_incrementally(false), 864 _has_reserved_stack_access(false), 865 #ifndef PRODUCT 866 _igv_idx(0), 867 _trace_opto_output(directive->TraceOptoOutputOption), 868 _print_ideal(directive->PrintIdealOption), 869 #endif 870 _has_method_handle_invokes(false), 871 _clinit_barrier_on_entry(false), 872 _stress_seed(0), 873 _comp_arena(mtCompiler), 874 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 875 _env(ci_env), 876 _directive(directive), 877 _log(ci_env->log()), 878 _failure_reason(NULL), 879 _congraph(NULL), 880 NOT_PRODUCT(_printer(NULL) COMMA) 881 _dead_node_list(comp_arena()), 882 _dead_node_count(0), 883 _node_arena(mtCompiler), 884 _old_arena(mtCompiler), 885 _mach_constant_base_node(NULL), 886 _Compile_types(mtCompiler), 887 _initial_gvn(NULL), 888 _for_igvn(NULL), 889 _number_of_mh_late_inlines(0), 890 _native_invokers(), 891 _print_inlining_stream(NULL), 892 _print_inlining_list(NULL), 893 _print_inlining_idx(0), 894 _print_inlining_output(NULL), 895 _replay_inline_data(NULL), 896 _java_calls(0), 897 _inner_loops(0), 898 _interpreter_frame_size(0), 899 #ifndef PRODUCT 900 _in_dump_cnt(0), 901 #endif 902 _allowed_reasons(0) { 903 C = this; 904 905 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 906 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 907 908 #ifndef PRODUCT 909 set_print_assembly(PrintFrameConverterAssembly); 910 set_parsed_irreducible_loop(false); 911 #else 912 set_print_assembly(false); // Must initialize. 913 #endif 914 set_has_irreducible_loop(false); // no loops 915 916 CompileWrapper cw(this); 917 Init(/*AliasLevel=*/ 0); 918 init_tf((*generator)()); 919 920 { 921 // The following is a dummy for the sake of GraphKit::gen_stub 922 Unique_Node_List for_igvn(comp_arena()); 923 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 924 PhaseGVN gvn(Thread::current()->resource_area(),255); 925 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 926 gvn.transform_no_reclaim(top()); 927 928 GraphKit kit; 929 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 930 } 931 932 NOT_PRODUCT( verify_graph_edges(); ) 933 934 Code_Gen(); 935 } 936 937 //------------------------------Init------------------------------------------- 938 // Prepare for a single compilation 939 void Compile::Init(int aliaslevel) { 940 _unique = 0; 941 _regalloc = NULL; 942 943 _tf = NULL; // filled in later 944 _top = NULL; // cached later 945 _matcher = NULL; // filled in later 946 _cfg = NULL; // filled in later 947 948 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 949 950 _node_note_array = NULL; 951 _default_node_notes = NULL; 952 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 953 954 _immutable_memory = NULL; // filled in at first inquiry 955 956 // Globally visible Nodes 957 // First set TOP to NULL to give safe behavior during creation of RootNode 958 set_cached_top_node(NULL); 959 set_root(new RootNode()); 960 // Now that you have a Root to point to, create the real TOP 961 set_cached_top_node( new ConNode(Type::TOP) ); 962 set_recent_alloc(NULL, NULL); 963 964 // Create Debug Information Recorder to record scopes, oopmaps, etc. 965 env()->set_oop_recorder(new OopRecorder(env()->arena())); 966 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 967 env()->set_dependencies(new Dependencies(env())); 968 969 _fixed_slots = 0; 970 set_has_split_ifs(false); 971 set_has_loops(false); // first approximation 972 set_has_stringbuilder(false); 973 set_has_boxed_value(false); 974 _trap_can_recompile = false; // no traps emitted yet 975 _major_progress = true; // start out assuming good things will happen 976 set_has_unsafe_access(false); 977 set_max_vector_size(0); 978 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 979 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 980 set_decompile_count(0); 981 982 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 983 _loop_opts_cnt = LoopOptsCount; 984 set_do_inlining(Inline); 985 set_max_inline_size(MaxInlineSize); 986 set_freq_inline_size(FreqInlineSize); 987 set_do_scheduling(OptoScheduling); 988 989 set_do_vector_loop(false); 990 reset_max_monitors(); 991 992 if (AllowVectorizeOnDemand) { 993 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 994 set_do_vector_loop(true); 995 NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii());}) 996 } else if (has_method() && method()->name() != 0 && 997 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 998 set_do_vector_loop(true); 999 } 1000 } 1001 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1002 NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n", method()->name()->as_quoted_ascii());}) 1003 1004 set_age_code(has_method() && method()->profile_aging()); 1005 set_rtm_state(NoRTM); // No RTM lock eliding by default 1006 _max_node_limit = _directive->MaxNodeLimitOption; 1007 1008 #if INCLUDE_RTM_OPT 1009 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1010 int rtm_state = method()->method_data()->rtm_state(); 1011 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1012 // Don't generate RTM lock eliding code. 1013 set_rtm_state(NoRTM); 1014 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1015 // Generate RTM lock eliding code without abort ratio calculation code. 1016 set_rtm_state(UseRTM); 1017 } else if (UseRTMDeopt) { 1018 // Generate RTM lock eliding code and include abort ratio calculation 1019 // code if UseRTMDeopt is on. 1020 set_rtm_state(ProfileRTM); 1021 } 1022 } 1023 #endif 1024 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1025 set_clinit_barrier_on_entry(true); 1026 } 1027 if (debug_info()->recording_non_safepoints()) { 1028 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1029 (comp_arena(), 8, 0, NULL)); 1030 set_default_node_notes(Node_Notes::make(this)); 1031 } 1032 1033 // // -- Initialize types before each compile -- 1034 // // Update cached type information 1035 // if( _method && _method->constants() ) 1036 // Type::update_loaded_types(_method, _method->constants()); 1037 1038 // Init alias_type map. 1039 if (!_do_escape_analysis && aliaslevel == 3) 1040 aliaslevel = 2; // No unique types without escape analysis 1041 _AliasLevel = aliaslevel; 1042 const int grow_ats = 16; 1043 _max_alias_types = grow_ats; 1044 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1045 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1046 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1047 { 1048 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1049 } 1050 // Initialize the first few types. 1051 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1052 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1053 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1054 _num_alias_types = AliasIdxRaw+1; 1055 // Zero out the alias type cache. 1056 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1057 // A NULL adr_type hits in the cache right away. Preload the right answer. 1058 probe_alias_cache(NULL)->_index = AliasIdxTop; 1059 1060 #ifdef ASSERT 1061 _type_verify_symmetry = true; 1062 _phase_optimize_finished = false; 1063 _exception_backedge = false; 1064 #endif 1065 } 1066 1067 //---------------------------init_start---------------------------------------- 1068 // Install the StartNode on this compile object. 1069 void Compile::init_start(StartNode* s) { 1070 if (failing()) 1071 return; // already failing 1072 assert(s == start(), ""); 1073 } 1074 1075 /** 1076 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1077 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1078 * the ideal graph. 1079 */ 1080 StartNode* Compile::start() const { 1081 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1082 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1083 Node* start = root()->fast_out(i); 1084 if (start->is_Start()) { 1085 return start->as_Start(); 1086 } 1087 } 1088 fatal("Did not find Start node!"); 1089 return NULL; 1090 } 1091 1092 //-------------------------------immutable_memory------------------------------------- 1093 // Access immutable memory 1094 Node* Compile::immutable_memory() { 1095 if (_immutable_memory != NULL) { 1096 return _immutable_memory; 1097 } 1098 StartNode* s = start(); 1099 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1100 Node *p = s->fast_out(i); 1101 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1102 _immutable_memory = p; 1103 return _immutable_memory; 1104 } 1105 } 1106 ShouldNotReachHere(); 1107 return NULL; 1108 } 1109 1110 //----------------------set_cached_top_node------------------------------------ 1111 // Install the cached top node, and make sure Node::is_top works correctly. 1112 void Compile::set_cached_top_node(Node* tn) { 1113 if (tn != NULL) verify_top(tn); 1114 Node* old_top = _top; 1115 _top = tn; 1116 // Calling Node::setup_is_top allows the nodes the chance to adjust 1117 // their _out arrays. 1118 if (_top != NULL) _top->setup_is_top(); 1119 if (old_top != NULL) old_top->setup_is_top(); 1120 assert(_top == NULL || top()->is_top(), ""); 1121 } 1122 1123 #ifdef ASSERT 1124 uint Compile::count_live_nodes_by_graph_walk() { 1125 Unique_Node_List useful(comp_arena()); 1126 // Get useful node list by walking the graph. 1127 identify_useful_nodes(useful); 1128 return useful.size(); 1129 } 1130 1131 void Compile::print_missing_nodes() { 1132 1133 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1134 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1135 return; 1136 } 1137 1138 // This is an expensive function. It is executed only when the user 1139 // specifies VerifyIdealNodeCount option or otherwise knows the 1140 // additional work that needs to be done to identify reachable nodes 1141 // by walking the flow graph and find the missing ones using 1142 // _dead_node_list. 1143 1144 Unique_Node_List useful(comp_arena()); 1145 // Get useful node list by walking the graph. 1146 identify_useful_nodes(useful); 1147 1148 uint l_nodes = C->live_nodes(); 1149 uint l_nodes_by_walk = useful.size(); 1150 1151 if (l_nodes != l_nodes_by_walk) { 1152 if (_log != NULL) { 1153 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1154 _log->stamp(); 1155 _log->end_head(); 1156 } 1157 VectorSet& useful_member_set = useful.member_set(); 1158 int last_idx = l_nodes_by_walk; 1159 for (int i = 0; i < last_idx; i++) { 1160 if (useful_member_set.test(i)) { 1161 if (_dead_node_list.test(i)) { 1162 if (_log != NULL) { 1163 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1164 } 1165 if (PrintIdealNodeCount) { 1166 // Print the log message to tty 1167 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1168 useful.at(i)->dump(); 1169 } 1170 } 1171 } 1172 else if (! _dead_node_list.test(i)) { 1173 if (_log != NULL) { 1174 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1175 } 1176 if (PrintIdealNodeCount) { 1177 // Print the log message to tty 1178 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1179 } 1180 } 1181 } 1182 if (_log != NULL) { 1183 _log->tail("mismatched_nodes"); 1184 } 1185 } 1186 } 1187 void Compile::record_modified_node(Node* n) { 1188 if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) { 1189 _modified_nodes->push(n); 1190 } 1191 } 1192 1193 void Compile::remove_modified_node(Node* n) { 1194 if (_modified_nodes != NULL) { 1195 _modified_nodes->remove(n); 1196 } 1197 } 1198 #endif 1199 1200 #ifndef PRODUCT 1201 void Compile::verify_top(Node* tn) const { 1202 if (tn != NULL) { 1203 assert(tn->is_Con(), "top node must be a constant"); 1204 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1205 assert(tn->in(0) != NULL, "must have live top node"); 1206 } 1207 } 1208 #endif 1209 1210 1211 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1212 1213 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1214 guarantee(arr != NULL, ""); 1215 int num_blocks = arr->length(); 1216 if (grow_by < num_blocks) grow_by = num_blocks; 1217 int num_notes = grow_by * _node_notes_block_size; 1218 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1219 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1220 while (num_notes > 0) { 1221 arr->append(notes); 1222 notes += _node_notes_block_size; 1223 num_notes -= _node_notes_block_size; 1224 } 1225 assert(num_notes == 0, "exact multiple, please"); 1226 } 1227 1228 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1229 if (source == NULL || dest == NULL) return false; 1230 1231 if (dest->is_Con()) 1232 return false; // Do not push debug info onto constants. 1233 1234 #ifdef ASSERT 1235 // Leave a bread crumb trail pointing to the original node: 1236 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1237 dest->set_debug_orig(source); 1238 } 1239 #endif 1240 1241 if (node_note_array() == NULL) 1242 return false; // Not collecting any notes now. 1243 1244 // This is a copy onto a pre-existing node, which may already have notes. 1245 // If both nodes have notes, do not overwrite any pre-existing notes. 1246 Node_Notes* source_notes = node_notes_at(source->_idx); 1247 if (source_notes == NULL || source_notes->is_clear()) return false; 1248 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1249 if (dest_notes == NULL || dest_notes->is_clear()) { 1250 return set_node_notes_at(dest->_idx, source_notes); 1251 } 1252 1253 Node_Notes merged_notes = (*source_notes); 1254 // The order of operations here ensures that dest notes will win... 1255 merged_notes.update_from(dest_notes); 1256 return set_node_notes_at(dest->_idx, &merged_notes); 1257 } 1258 1259 1260 //--------------------------allow_range_check_smearing------------------------- 1261 // Gating condition for coalescing similar range checks. 1262 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1263 // single covering check that is at least as strong as any of them. 1264 // If the optimization succeeds, the simplified (strengthened) range check 1265 // will always succeed. If it fails, we will deopt, and then give up 1266 // on the optimization. 1267 bool Compile::allow_range_check_smearing() const { 1268 // If this method has already thrown a range-check, 1269 // assume it was because we already tried range smearing 1270 // and it failed. 1271 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1272 return !already_trapped; 1273 } 1274 1275 1276 //------------------------------flatten_alias_type----------------------------- 1277 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1278 int offset = tj->offset(); 1279 TypePtr::PTR ptr = tj->ptr(); 1280 1281 // Known instance (scalarizable allocation) alias only with itself. 1282 bool is_known_inst = tj->isa_oopptr() != NULL && 1283 tj->is_oopptr()->is_known_instance(); 1284 1285 // Process weird unsafe references. 1286 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1287 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1288 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1289 tj = TypeOopPtr::BOTTOM; 1290 ptr = tj->ptr(); 1291 offset = tj->offset(); 1292 } 1293 1294 // Array pointers need some flattening 1295 const TypeAryPtr *ta = tj->isa_aryptr(); 1296 if (ta && ta->is_stable()) { 1297 // Erase stability property for alias analysis. 1298 tj = ta = ta->cast_to_stable(false); 1299 } 1300 if( ta && is_known_inst ) { 1301 if ( offset != Type::OffsetBot && 1302 offset > arrayOopDesc::length_offset_in_bytes() ) { 1303 offset = Type::OffsetBot; // Flatten constant access into array body only 1304 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1305 } 1306 } else if( ta && _AliasLevel >= 2 ) { 1307 // For arrays indexed by constant indices, we flatten the alias 1308 // space to include all of the array body. Only the header, klass 1309 // and array length can be accessed un-aliased. 1310 if( offset != Type::OffsetBot ) { 1311 if( ta->const_oop() ) { // MethodData* or Method* 1312 offset = Type::OffsetBot; // Flatten constant access into array body 1313 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1314 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1315 // range is OK as-is. 1316 tj = ta = TypeAryPtr::RANGE; 1317 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1318 tj = TypeInstPtr::KLASS; // all klass loads look alike 1319 ta = TypeAryPtr::RANGE; // generic ignored junk 1320 ptr = TypePtr::BotPTR; 1321 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1322 tj = TypeInstPtr::MARK; 1323 ta = TypeAryPtr::RANGE; // generic ignored junk 1324 ptr = TypePtr::BotPTR; 1325 } else { // Random constant offset into array body 1326 offset = Type::OffsetBot; // Flatten constant access into array body 1327 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1328 } 1329 } 1330 // Arrays of fixed size alias with arrays of unknown size. 1331 if (ta->size() != TypeInt::POS) { 1332 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1333 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1334 } 1335 // Arrays of known objects become arrays of unknown objects. 1336 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1337 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1338 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1339 } 1340 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1341 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1342 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1343 } 1344 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1345 // cannot be distinguished by bytecode alone. 1346 if (ta->elem() == TypeInt::BOOL) { 1347 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1348 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1349 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1350 } 1351 // During the 2nd round of IterGVN, NotNull castings are removed. 1352 // Make sure the Bottom and NotNull variants alias the same. 1353 // Also, make sure exact and non-exact variants alias the same. 1354 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1355 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1356 } 1357 } 1358 1359 // Oop pointers need some flattening 1360 const TypeInstPtr *to = tj->isa_instptr(); 1361 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1362 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1363 if( ptr == TypePtr::Constant ) { 1364 if (to->klass() != ciEnv::current()->Class_klass() || 1365 offset < k->layout_helper_size_in_bytes()) { 1366 // No constant oop pointers (such as Strings); they alias with 1367 // unknown strings. 1368 assert(!is_known_inst, "not scalarizable allocation"); 1369 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1370 } 1371 } else if( is_known_inst ) { 1372 tj = to; // Keep NotNull and klass_is_exact for instance type 1373 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1374 // During the 2nd round of IterGVN, NotNull castings are removed. 1375 // Make sure the Bottom and NotNull variants alias the same. 1376 // Also, make sure exact and non-exact variants alias the same. 1377 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1378 } 1379 if (to->speculative() != NULL) { 1380 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1381 } 1382 // Canonicalize the holder of this field 1383 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1384 // First handle header references such as a LoadKlassNode, even if the 1385 // object's klass is unloaded at compile time (4965979). 1386 if (!is_known_inst) { // Do it only for non-instance types 1387 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1388 } 1389 } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) { 1390 // Static fields are in the space above the normal instance 1391 // fields in the java.lang.Class instance. 1392 if (to->klass() != ciEnv::current()->Class_klass()) { 1393 to = NULL; 1394 tj = TypeOopPtr::BOTTOM; 1395 offset = tj->offset(); 1396 } 1397 } else { 1398 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1399 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1400 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1401 if( is_known_inst ) { 1402 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1403 } else { 1404 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1405 } 1406 } 1407 } 1408 } 1409 1410 // Klass pointers to object array klasses need some flattening 1411 const TypeKlassPtr *tk = tj->isa_klassptr(); 1412 if( tk ) { 1413 // If we are referencing a field within a Klass, we need 1414 // to assume the worst case of an Object. Both exact and 1415 // inexact types must flatten to the same alias class so 1416 // use NotNull as the PTR. 1417 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1418 1419 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1420 TypeKlassPtr::OBJECT->klass(), 1421 offset); 1422 } 1423 1424 ciKlass* klass = tk->klass(); 1425 if( klass->is_obj_array_klass() ) { 1426 ciKlass* k = TypeAryPtr::OOPS->klass(); 1427 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1428 k = TypeInstPtr::BOTTOM->klass(); 1429 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1430 } 1431 1432 // Check for precise loads from the primary supertype array and force them 1433 // to the supertype cache alias index. Check for generic array loads from 1434 // the primary supertype array and also force them to the supertype cache 1435 // alias index. Since the same load can reach both, we need to merge 1436 // these 2 disparate memories into the same alias class. Since the 1437 // primary supertype array is read-only, there's no chance of confusion 1438 // where we bypass an array load and an array store. 1439 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1440 if (offset == Type::OffsetBot || 1441 (offset >= primary_supers_offset && 1442 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1443 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1444 offset = in_bytes(Klass::secondary_super_cache_offset()); 1445 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1446 } 1447 } 1448 1449 // Flatten all Raw pointers together. 1450 if (tj->base() == Type::RawPtr) 1451 tj = TypeRawPtr::BOTTOM; 1452 1453 if (tj->base() == Type::AnyPtr) 1454 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1455 1456 // Flatten all to bottom for now 1457 switch( _AliasLevel ) { 1458 case 0: 1459 tj = TypePtr::BOTTOM; 1460 break; 1461 case 1: // Flatten to: oop, static, field or array 1462 switch (tj->base()) { 1463 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1464 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1465 case Type::AryPtr: // do not distinguish arrays at all 1466 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1467 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1468 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1469 default: ShouldNotReachHere(); 1470 } 1471 break; 1472 case 2: // No collapsing at level 2; keep all splits 1473 case 3: // No collapsing at level 3; keep all splits 1474 break; 1475 default: 1476 Unimplemented(); 1477 } 1478 1479 offset = tj->offset(); 1480 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1481 1482 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1483 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1484 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1485 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1486 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1487 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1488 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1489 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1490 assert( tj->ptr() != TypePtr::TopPTR && 1491 tj->ptr() != TypePtr::AnyNull && 1492 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1493 // assert( tj->ptr() != TypePtr::Constant || 1494 // tj->base() == Type::RawPtr || 1495 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1496 1497 return tj; 1498 } 1499 1500 void Compile::AliasType::Init(int i, const TypePtr* at) { 1501 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1502 _index = i; 1503 _adr_type = at; 1504 _field = NULL; 1505 _element = NULL; 1506 _is_rewritable = true; // default 1507 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1508 if (atoop != NULL && atoop->is_known_instance()) { 1509 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1510 _general_index = Compile::current()->get_alias_index(gt); 1511 } else { 1512 _general_index = 0; 1513 } 1514 } 1515 1516 BasicType Compile::AliasType::basic_type() const { 1517 if (element() != NULL) { 1518 const Type* element = adr_type()->is_aryptr()->elem(); 1519 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1520 } if (field() != NULL) { 1521 return field()->layout_type(); 1522 } else { 1523 return T_ILLEGAL; // unknown 1524 } 1525 } 1526 1527 //---------------------------------print_on------------------------------------ 1528 #ifndef PRODUCT 1529 void Compile::AliasType::print_on(outputStream* st) { 1530 if (index() < 10) 1531 st->print("@ <%d> ", index()); 1532 else st->print("@ <%d>", index()); 1533 st->print(is_rewritable() ? " " : " RO"); 1534 int offset = adr_type()->offset(); 1535 if (offset == Type::OffsetBot) 1536 st->print(" +any"); 1537 else st->print(" +%-3d", offset); 1538 st->print(" in "); 1539 adr_type()->dump_on(st); 1540 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1541 if (field() != NULL && tjp) { 1542 if (tjp->klass() != field()->holder() || 1543 tjp->offset() != field()->offset_in_bytes()) { 1544 st->print(" != "); 1545 field()->print(); 1546 st->print(" ***"); 1547 } 1548 } 1549 } 1550 1551 void print_alias_types() { 1552 Compile* C = Compile::current(); 1553 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1554 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1555 C->alias_type(idx)->print_on(tty); 1556 tty->cr(); 1557 } 1558 } 1559 #endif 1560 1561 1562 //----------------------------probe_alias_cache-------------------------------- 1563 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1564 intptr_t key = (intptr_t) adr_type; 1565 key ^= key >> logAliasCacheSize; 1566 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1567 } 1568 1569 1570 //-----------------------------grow_alias_types-------------------------------- 1571 void Compile::grow_alias_types() { 1572 const int old_ats = _max_alias_types; // how many before? 1573 const int new_ats = old_ats; // how many more? 1574 const int grow_ats = old_ats+new_ats; // how many now? 1575 _max_alias_types = grow_ats; 1576 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1577 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1578 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1579 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1580 } 1581 1582 1583 //--------------------------------find_alias_type------------------------------ 1584 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1585 if (_AliasLevel == 0) 1586 return alias_type(AliasIdxBot); 1587 1588 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1589 if (ace->_adr_type == adr_type) { 1590 return alias_type(ace->_index); 1591 } 1592 1593 // Handle special cases. 1594 if (adr_type == NULL) return alias_type(AliasIdxTop); 1595 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1596 1597 // Do it the slow way. 1598 const TypePtr* flat = flatten_alias_type(adr_type); 1599 1600 #ifdef ASSERT 1601 { 1602 ResourceMark rm; 1603 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1604 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1605 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1606 Type::str(adr_type)); 1607 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1608 const TypeOopPtr* foop = flat->is_oopptr(); 1609 // Scalarizable allocations have exact klass always. 1610 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1611 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1612 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1613 Type::str(foop), Type::str(xoop)); 1614 } 1615 } 1616 #endif 1617 1618 int idx = AliasIdxTop; 1619 for (int i = 0; i < num_alias_types(); i++) { 1620 if (alias_type(i)->adr_type() == flat) { 1621 idx = i; 1622 break; 1623 } 1624 } 1625 1626 if (idx == AliasIdxTop) { 1627 if (no_create) return NULL; 1628 // Grow the array if necessary. 1629 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1630 // Add a new alias type. 1631 idx = _num_alias_types++; 1632 _alias_types[idx]->Init(idx, flat); 1633 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1634 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1635 if (flat->isa_instptr()) { 1636 if (flat->offset() == java_lang_Class::klass_offset() 1637 && flat->is_instptr()->klass() == env()->Class_klass()) 1638 alias_type(idx)->set_rewritable(false); 1639 } 1640 if (flat->isa_aryptr()) { 1641 #ifdef ASSERT 1642 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1643 // (T_BYTE has the weakest alignment and size restrictions...) 1644 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1645 #endif 1646 if (flat->offset() == TypePtr::OffsetBot) { 1647 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1648 } 1649 } 1650 if (flat->isa_klassptr()) { 1651 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1652 alias_type(idx)->set_rewritable(false); 1653 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1654 alias_type(idx)->set_rewritable(false); 1655 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1656 alias_type(idx)->set_rewritable(false); 1657 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1658 alias_type(idx)->set_rewritable(false); 1659 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1660 alias_type(idx)->set_rewritable(false); 1661 } 1662 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1663 // but the base pointer type is not distinctive enough to identify 1664 // references into JavaThread.) 1665 1666 // Check for final fields. 1667 const TypeInstPtr* tinst = flat->isa_instptr(); 1668 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1669 ciField* field; 1670 if (tinst->const_oop() != NULL && 1671 tinst->klass() == ciEnv::current()->Class_klass() && 1672 tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) { 1673 // static field 1674 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1675 field = k->get_field_by_offset(tinst->offset(), true); 1676 } else { 1677 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1678 field = k->get_field_by_offset(tinst->offset(), false); 1679 } 1680 assert(field == NULL || 1681 original_field == NULL || 1682 (field->holder() == original_field->holder() && 1683 field->offset() == original_field->offset() && 1684 field->is_static() == original_field->is_static()), "wrong field?"); 1685 // Set field() and is_rewritable() attributes. 1686 if (field != NULL) alias_type(idx)->set_field(field); 1687 } 1688 } 1689 1690 // Fill the cache for next time. 1691 ace->_adr_type = adr_type; 1692 ace->_index = idx; 1693 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1694 1695 // Might as well try to fill the cache for the flattened version, too. 1696 AliasCacheEntry* face = probe_alias_cache(flat); 1697 if (face->_adr_type == NULL) { 1698 face->_adr_type = flat; 1699 face->_index = idx; 1700 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1701 } 1702 1703 return alias_type(idx); 1704 } 1705 1706 1707 Compile::AliasType* Compile::alias_type(ciField* field) { 1708 const TypeOopPtr* t; 1709 if (field->is_static()) 1710 t = TypeInstPtr::make(field->holder()->java_mirror()); 1711 else 1712 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1713 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1714 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1715 return atp; 1716 } 1717 1718 1719 //------------------------------have_alias_type-------------------------------- 1720 bool Compile::have_alias_type(const TypePtr* adr_type) { 1721 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1722 if (ace->_adr_type == adr_type) { 1723 return true; 1724 } 1725 1726 // Handle special cases. 1727 if (adr_type == NULL) return true; 1728 if (adr_type == TypePtr::BOTTOM) return true; 1729 1730 return find_alias_type(adr_type, true, NULL) != NULL; 1731 } 1732 1733 //-----------------------------must_alias-------------------------------------- 1734 // True if all values of the given address type are in the given alias category. 1735 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1736 if (alias_idx == AliasIdxBot) return true; // the universal category 1737 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1738 if (alias_idx == AliasIdxTop) return false; // the empty category 1739 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1740 1741 // the only remaining possible overlap is identity 1742 int adr_idx = get_alias_index(adr_type); 1743 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1744 assert(adr_idx == alias_idx || 1745 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1746 && adr_type != TypeOopPtr::BOTTOM), 1747 "should not be testing for overlap with an unsafe pointer"); 1748 return adr_idx == alias_idx; 1749 } 1750 1751 //------------------------------can_alias-------------------------------------- 1752 // True if any values of the given address type are in the given alias category. 1753 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1754 if (alias_idx == AliasIdxTop) return false; // the empty category 1755 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1756 // Known instance doesn't alias with bottom memory 1757 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1758 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1759 1760 // the only remaining possible overlap is identity 1761 int adr_idx = get_alias_index(adr_type); 1762 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1763 return adr_idx == alias_idx; 1764 } 1765 1766 //---------------------cleanup_loop_predicates----------------------- 1767 // Remove the opaque nodes that protect the predicates so that all unused 1768 // checks and uncommon_traps will be eliminated from the ideal graph 1769 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1770 if (predicate_count()==0) return; 1771 for (int i = predicate_count(); i > 0; i--) { 1772 Node * n = predicate_opaque1_node(i-1); 1773 assert(n->Opcode() == Op_Opaque1, "must be"); 1774 igvn.replace_node(n, n->in(1)); 1775 } 1776 assert(predicate_count()==0, "should be clean!"); 1777 } 1778 1779 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1780 if (!n->for_post_loop_opts_igvn()) { 1781 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1782 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1783 _for_post_loop_igvn.append(n); 1784 } 1785 } 1786 1787 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1788 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1789 _for_post_loop_igvn.remove(n); 1790 } 1791 1792 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1793 // Verify that all previous optimizations produced a valid graph 1794 // at least to this point, even if no loop optimizations were done. 1795 PhaseIdealLoop::verify(igvn); 1796 1797 C->set_post_loop_opts_phase(); // no more loop opts allowed 1798 1799 assert(!C->major_progress(), "not cleared"); 1800 1801 if (_for_post_loop_igvn.length() > 0) { 1802 while (_for_post_loop_igvn.length() > 0) { 1803 Node* n = _for_post_loop_igvn.pop(); 1804 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1805 igvn._worklist.push(n); 1806 } 1807 igvn.optimize(); 1808 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1809 1810 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1811 if (C->major_progress()) { 1812 C->clear_major_progress(); // ensure that major progress is now clear 1813 } 1814 } 1815 } 1816 1817 // StringOpts and late inlining of string methods 1818 void Compile::inline_string_calls(bool parse_time) { 1819 { 1820 // remove useless nodes to make the usage analysis simpler 1821 ResourceMark rm; 1822 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1823 } 1824 1825 { 1826 ResourceMark rm; 1827 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1828 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1829 print_method(PHASE_AFTER_STRINGOPTS, 3); 1830 } 1831 1832 // now inline anything that we skipped the first time around 1833 if (!parse_time) { 1834 _late_inlines_pos = _late_inlines.length(); 1835 } 1836 1837 while (_string_late_inlines.length() > 0) { 1838 CallGenerator* cg = _string_late_inlines.pop(); 1839 cg->do_late_inline(); 1840 if (failing()) return; 1841 } 1842 _string_late_inlines.trunc_to(0); 1843 } 1844 1845 // Late inlining of boxing methods 1846 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1847 if (_boxing_late_inlines.length() > 0) { 1848 assert(has_boxed_value(), "inconsistent"); 1849 1850 PhaseGVN* gvn = initial_gvn(); 1851 set_inlining_incrementally(true); 1852 1853 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1854 for_igvn()->clear(); 1855 gvn->replace_with(&igvn); 1856 1857 _late_inlines_pos = _late_inlines.length(); 1858 1859 while (_boxing_late_inlines.length() > 0) { 1860 CallGenerator* cg = _boxing_late_inlines.pop(); 1861 cg->do_late_inline(); 1862 if (failing()) return; 1863 } 1864 _boxing_late_inlines.trunc_to(0); 1865 1866 inline_incrementally_cleanup(igvn); 1867 1868 set_inlining_incrementally(false); 1869 } 1870 } 1871 1872 bool Compile::inline_incrementally_one() { 1873 assert(IncrementalInline, "incremental inlining should be on"); 1874 1875 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1876 1877 set_inlining_progress(false); 1878 set_do_cleanup(false); 1879 1880 for (int i = 0; i < _late_inlines.length(); i++) { 1881 _late_inlines_pos = i+1; 1882 CallGenerator* cg = _late_inlines.at(i); 1883 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 1884 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 1885 cg->do_late_inline(); 1886 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 1887 if (failing()) { 1888 return false; 1889 } else if (inlining_progress()) { 1890 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 1891 print_method(PHASE_INCREMENTAL_INLINE_STEP, cg->call_node(), 3); 1892 break; // process one call site at a time 1893 } 1894 } else { 1895 // Ignore late inline direct calls when inlining is not allowed. 1896 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 1897 } 1898 } 1899 // Remove processed elements. 1900 _late_inlines.remove_till(_late_inlines_pos); 1901 _late_inlines_pos = 0; 1902 1903 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 1904 1905 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 1906 1907 set_inlining_progress(false); 1908 set_do_cleanup(false); 1909 1910 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 1911 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 1912 } 1913 1914 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 1915 { 1916 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1917 ResourceMark rm; 1918 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1919 } 1920 { 1921 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1922 igvn = PhaseIterGVN(initial_gvn()); 1923 igvn.optimize(); 1924 } 1925 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 1926 } 1927 1928 // Perform incremental inlining until bound on number of live nodes is reached 1929 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1930 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1931 1932 set_inlining_incrementally(true); 1933 uint low_live_nodes = 0; 1934 1935 while (_late_inlines.length() > 0) { 1936 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1937 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1938 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 1939 // PhaseIdealLoop is expensive so we only try it once we are 1940 // out of live nodes and we only try it again if the previous 1941 // helped got the number of nodes down significantly 1942 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 1943 if (failing()) return; 1944 low_live_nodes = live_nodes(); 1945 _major_progress = true; 1946 } 1947 1948 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1949 bool do_print_inlining = print_inlining() || print_intrinsics(); 1950 if (do_print_inlining || log() != NULL) { 1951 // Print inlining message for candidates that we couldn't inline for lack of space. 1952 for (int i = 0; i < _late_inlines.length(); i++) { 1953 CallGenerator* cg = _late_inlines.at(i); 1954 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 1955 if (do_print_inlining) { 1956 cg->print_inlining_late(msg); 1957 } 1958 log_late_inline_failure(cg, msg); 1959 } 1960 } 1961 break; // finish 1962 } 1963 } 1964 1965 for_igvn()->clear(); 1966 initial_gvn()->replace_with(&igvn); 1967 1968 while (inline_incrementally_one()) { 1969 assert(!failing(), "inconsistent"); 1970 } 1971 if (failing()) return; 1972 1973 inline_incrementally_cleanup(igvn); 1974 1975 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 1976 1977 if (failing()) return; 1978 1979 if (_late_inlines.length() == 0) { 1980 break; // no more progress 1981 } 1982 } 1983 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1984 1985 if (_string_late_inlines.length() > 0) { 1986 assert(has_stringbuilder(), "inconsistent"); 1987 for_igvn()->clear(); 1988 initial_gvn()->replace_with(&igvn); 1989 1990 inline_string_calls(false); 1991 1992 if (failing()) return; 1993 1994 inline_incrementally_cleanup(igvn); 1995 } 1996 1997 set_inlining_incrementally(false); 1998 } 1999 2000 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2001 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2002 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2003 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL" 2004 // as if "inlining_incrementally() == true" were set. 2005 assert(inlining_incrementally() == false, "not allowed"); 2006 assert(_modified_nodes == NULL, "not allowed"); 2007 assert(_late_inlines.length() > 0, "sanity"); 2008 2009 while (_late_inlines.length() > 0) { 2010 for_igvn()->clear(); 2011 initial_gvn()->replace_with(&igvn); 2012 2013 while (inline_incrementally_one()) { 2014 assert(!failing(), "inconsistent"); 2015 } 2016 if (failing()) return; 2017 2018 inline_incrementally_cleanup(igvn); 2019 } 2020 } 2021 2022 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2023 if (_loop_opts_cnt > 0) { 2024 while (major_progress() && (_loop_opts_cnt > 0)) { 2025 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2026 PhaseIdealLoop::optimize(igvn, mode); 2027 _loop_opts_cnt--; 2028 if (failing()) return false; 2029 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2030 } 2031 } 2032 return true; 2033 } 2034 2035 // Remove edges from "root" to each SafePoint at a backward branch. 2036 // They were inserted during parsing (see add_safepoint()) to make 2037 // infinite loops without calls or exceptions visible to root, i.e., 2038 // useful. 2039 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2040 Node *r = root(); 2041 if (r != NULL) { 2042 for (uint i = r->req(); i < r->len(); ++i) { 2043 Node *n = r->in(i); 2044 if (n != NULL && n->is_SafePoint()) { 2045 r->rm_prec(i); 2046 if (n->outcnt() == 0) { 2047 igvn.remove_dead_node(n); 2048 } 2049 --i; 2050 } 2051 } 2052 // Parsing may have added top inputs to the root node (Path 2053 // leading to the Halt node proven dead). Make sure we get a 2054 // chance to clean them up. 2055 igvn._worklist.push(r); 2056 igvn.optimize(); 2057 } 2058 } 2059 2060 //------------------------------Optimize--------------------------------------- 2061 // Given a graph, optimize it. 2062 void Compile::Optimize() { 2063 TracePhase tp("optimizer", &timers[_t_optimizer]); 2064 2065 #ifndef PRODUCT 2066 if (env()->break_at_compile()) { 2067 BREAKPOINT; 2068 } 2069 2070 #endif 2071 2072 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2073 #ifdef ASSERT 2074 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2075 #endif 2076 2077 ResourceMark rm; 2078 2079 print_inlining_reinit(); 2080 2081 NOT_PRODUCT( verify_graph_edges(); ) 2082 2083 print_method(PHASE_AFTER_PARSING); 2084 2085 { 2086 // Iterative Global Value Numbering, including ideal transforms 2087 // Initialize IterGVN with types and values from parse-time GVN 2088 PhaseIterGVN igvn(initial_gvn()); 2089 #ifdef ASSERT 2090 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2091 #endif 2092 { 2093 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2094 igvn.optimize(); 2095 } 2096 2097 if (failing()) return; 2098 2099 print_method(PHASE_ITER_GVN1, 2); 2100 2101 inline_incrementally(igvn); 2102 2103 print_method(PHASE_INCREMENTAL_INLINE, 2); 2104 2105 if (failing()) return; 2106 2107 if (eliminate_boxing()) { 2108 // Inline valueOf() methods now. 2109 inline_boxing_calls(igvn); 2110 2111 if (AlwaysIncrementalInline) { 2112 inline_incrementally(igvn); 2113 } 2114 2115 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2116 2117 if (failing()) return; 2118 } 2119 2120 // Remove the speculative part of types and clean up the graph from 2121 // the extra CastPP nodes whose only purpose is to carry them. Do 2122 // that early so that optimizations are not disrupted by the extra 2123 // CastPP nodes. 2124 remove_speculative_types(igvn); 2125 2126 // No more new expensive nodes will be added to the list from here 2127 // so keep only the actual candidates for optimizations. 2128 cleanup_expensive_nodes(igvn); 2129 2130 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2131 if (EnableVectorSupport && has_vbox_nodes()) { 2132 TracePhase tp("", &timers[_t_vector]); 2133 PhaseVector pv(igvn); 2134 pv.optimize_vector_boxes(); 2135 2136 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2137 } 2138 assert(!has_vbox_nodes(), "sanity"); 2139 2140 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2141 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2142 initial_gvn()->replace_with(&igvn); 2143 for_igvn()->clear(); 2144 Unique_Node_List new_worklist(C->comp_arena()); 2145 { 2146 ResourceMark rm; 2147 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2148 } 2149 Unique_Node_List* save_for_igvn = for_igvn(); 2150 set_for_igvn(&new_worklist); 2151 igvn = PhaseIterGVN(initial_gvn()); 2152 igvn.optimize(); 2153 set_for_igvn(save_for_igvn); 2154 } 2155 2156 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2157 // safepoints 2158 remove_root_to_sfpts_edges(igvn); 2159 2160 // Perform escape analysis 2161 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2162 if (has_loops()) { 2163 // Cleanup graph (remove dead nodes). 2164 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2165 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2166 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2167 if (failing()) return; 2168 } 2169 ConnectionGraph::do_analysis(this, &igvn); 2170 2171 if (failing()) return; 2172 2173 // Optimize out fields loads from scalar replaceable allocations. 2174 igvn.optimize(); 2175 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2176 2177 if (failing()) return; 2178 2179 if (congraph() != NULL && macro_count() > 0) { 2180 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2181 PhaseMacroExpand mexp(igvn); 2182 mexp.eliminate_macro_nodes(); 2183 igvn.set_delay_transform(false); 2184 2185 igvn.optimize(); 2186 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2187 2188 if (failing()) return; 2189 } 2190 } 2191 2192 // Loop transforms on the ideal graph. Range Check Elimination, 2193 // peeling, unrolling, etc. 2194 2195 // Set loop opts counter 2196 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2197 { 2198 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2199 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2200 _loop_opts_cnt--; 2201 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2202 if (failing()) return; 2203 } 2204 // Loop opts pass if partial peeling occurred in previous pass 2205 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2206 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2207 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2208 _loop_opts_cnt--; 2209 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2210 if (failing()) return; 2211 } 2212 // Loop opts pass for loop-unrolling before CCP 2213 if(major_progress() && (_loop_opts_cnt > 0)) { 2214 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2215 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2216 _loop_opts_cnt--; 2217 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2218 } 2219 if (!failing()) { 2220 // Verify that last round of loop opts produced a valid graph 2221 PhaseIdealLoop::verify(igvn); 2222 } 2223 } 2224 if (failing()) return; 2225 2226 // Conditional Constant Propagation; 2227 PhaseCCP ccp( &igvn ); 2228 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2229 { 2230 TracePhase tp("ccp", &timers[_t_ccp]); 2231 ccp.do_transform(); 2232 } 2233 print_method(PHASE_CCP1, 2); 2234 2235 assert( true, "Break here to ccp.dump_old2new_map()"); 2236 2237 // Iterative Global Value Numbering, including ideal transforms 2238 { 2239 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2240 igvn = ccp; 2241 igvn.optimize(); 2242 } 2243 print_method(PHASE_ITER_GVN2, 2); 2244 2245 if (failing()) return; 2246 2247 // Loop transforms on the ideal graph. Range Check Elimination, 2248 // peeling, unrolling, etc. 2249 if (!optimize_loops(igvn, LoopOptsDefault)) { 2250 return; 2251 } 2252 2253 if (failing()) return; 2254 2255 C->clear_major_progress(); // ensure that major progress is now clear 2256 2257 process_for_post_loop_opts_igvn(igvn); 2258 2259 #ifdef ASSERT 2260 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2261 #endif 2262 2263 { 2264 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2265 PhaseMacroExpand mex(igvn); 2266 if (mex.expand_macro_nodes()) { 2267 assert(failing(), "must bail out w/ explicit message"); 2268 return; 2269 } 2270 print_method(PHASE_MACRO_EXPANSION, 2); 2271 } 2272 2273 { 2274 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2275 if (bs->expand_barriers(this, igvn)) { 2276 assert(failing(), "must bail out w/ explicit message"); 2277 return; 2278 } 2279 print_method(PHASE_BARRIER_EXPANSION, 2); 2280 } 2281 2282 if (C->max_vector_size() > 0) { 2283 C->optimize_logic_cones(igvn); 2284 igvn.optimize(); 2285 } 2286 2287 DEBUG_ONLY( _modified_nodes = NULL; ) 2288 2289 assert(igvn._worklist.size() == 0, "not empty"); 2290 2291 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2292 2293 if (_late_inlines.length() > 0) { 2294 // More opportunities to optimize virtual and MH calls. 2295 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2296 process_late_inline_calls_no_inline(igvn); 2297 } 2298 } // (End scope of igvn; run destructor if necessary for asserts.) 2299 2300 check_no_dead_use(); 2301 2302 process_print_inlining(); 2303 2304 // A method with only infinite loops has no edges entering loops from root 2305 { 2306 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2307 if (final_graph_reshaping()) { 2308 assert(failing(), "must bail out w/ explicit message"); 2309 return; 2310 } 2311 } 2312 2313 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2314 DEBUG_ONLY(set_phase_optimize_finished();) 2315 } 2316 2317 #ifdef ASSERT 2318 void Compile::check_no_dead_use() const { 2319 ResourceMark rm; 2320 Unique_Node_List wq; 2321 wq.push(root()); 2322 for (uint i = 0; i < wq.size(); ++i) { 2323 Node* n = wq.at(i); 2324 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2325 Node* u = n->fast_out(j); 2326 if (u->outcnt() == 0 && !u->is_Con()) { 2327 u->dump(); 2328 fatal("no reachable node should have no use"); 2329 } 2330 wq.push(u); 2331 } 2332 } 2333 } 2334 #endif 2335 2336 void Compile::inline_vector_reboxing_calls() { 2337 if (C->_vector_reboxing_late_inlines.length() > 0) { 2338 _late_inlines_pos = C->_late_inlines.length(); 2339 while (_vector_reboxing_late_inlines.length() > 0) { 2340 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2341 cg->do_late_inline(); 2342 if (failing()) return; 2343 print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node()); 2344 } 2345 _vector_reboxing_late_inlines.trunc_to(0); 2346 } 2347 } 2348 2349 bool Compile::has_vbox_nodes() { 2350 if (C->_vector_reboxing_late_inlines.length() > 0) { 2351 return true; 2352 } 2353 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2354 Node * n = C->macro_node(macro_idx); 2355 assert(n->is_macro(), "only macro nodes expected here"); 2356 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2357 return true; 2358 } 2359 } 2360 return false; 2361 } 2362 2363 //---------------------------- Bitwise operation packing optimization --------------------------- 2364 2365 static bool is_vector_unary_bitwise_op(Node* n) { 2366 return n->Opcode() == Op_XorV && 2367 VectorNode::is_vector_bitwise_not_pattern(n); 2368 } 2369 2370 static bool is_vector_binary_bitwise_op(Node* n) { 2371 switch (n->Opcode()) { 2372 case Op_AndV: 2373 case Op_OrV: 2374 return true; 2375 2376 case Op_XorV: 2377 return !is_vector_unary_bitwise_op(n); 2378 2379 default: 2380 return false; 2381 } 2382 } 2383 2384 static bool is_vector_ternary_bitwise_op(Node* n) { 2385 return n->Opcode() == Op_MacroLogicV; 2386 } 2387 2388 static bool is_vector_bitwise_op(Node* n) { 2389 return is_vector_unary_bitwise_op(n) || 2390 is_vector_binary_bitwise_op(n) || 2391 is_vector_ternary_bitwise_op(n); 2392 } 2393 2394 static bool is_vector_bitwise_cone_root(Node* n) { 2395 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2396 return false; 2397 } 2398 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2399 if (is_vector_bitwise_op(n->fast_out(i))) { 2400 return false; 2401 } 2402 } 2403 return true; 2404 } 2405 2406 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2407 uint cnt = 0; 2408 if (is_vector_bitwise_op(n)) { 2409 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2410 for (uint i = 1; i < n->req(); i++) { 2411 Node* in = n->in(i); 2412 bool skip = VectorNode::is_all_ones_vector(in); 2413 if (!skip && !inputs.member(in)) { 2414 inputs.push(in); 2415 cnt++; 2416 } 2417 } 2418 assert(cnt <= 1, "not unary"); 2419 } else { 2420 uint last_req = n->req(); 2421 if (is_vector_ternary_bitwise_op(n)) { 2422 last_req = n->req() - 1; // skip last input 2423 } 2424 for (uint i = 1; i < last_req; i++) { 2425 Node* def = n->in(i); 2426 if (!inputs.member(def)) { 2427 inputs.push(def); 2428 cnt++; 2429 } 2430 } 2431 } 2432 partition.push(n); 2433 } else { // not a bitwise operations 2434 if (!inputs.member(n)) { 2435 inputs.push(n); 2436 cnt++; 2437 } 2438 } 2439 return cnt; 2440 } 2441 2442 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2443 Unique_Node_List useful_nodes; 2444 C->identify_useful_nodes(useful_nodes); 2445 2446 for (uint i = 0; i < useful_nodes.size(); i++) { 2447 Node* n = useful_nodes.at(i); 2448 if (is_vector_bitwise_cone_root(n)) { 2449 list.push(n); 2450 } 2451 } 2452 } 2453 2454 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2455 const TypeVect* vt, 2456 Unique_Node_List& partition, 2457 Unique_Node_List& inputs) { 2458 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2459 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2460 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2461 2462 Node* in1 = inputs.at(0); 2463 Node* in2 = inputs.at(1); 2464 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2465 2466 uint func = compute_truth_table(partition, inputs); 2467 return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt)); 2468 } 2469 2470 static uint extract_bit(uint func, uint pos) { 2471 return (func & (1 << pos)) >> pos; 2472 } 2473 2474 // 2475 // A macro logic node represents a truth table. It has 4 inputs, 2476 // First three inputs corresponds to 3 columns of a truth table 2477 // and fourth input captures the logic function. 2478 // 2479 // eg. fn = (in1 AND in2) OR in3; 2480 // 2481 // MacroNode(in1,in2,in3,fn) 2482 // 2483 // ----------------- 2484 // in1 in2 in3 fn 2485 // ----------------- 2486 // 0 0 0 0 2487 // 0 0 1 1 2488 // 0 1 0 0 2489 // 0 1 1 1 2490 // 1 0 0 0 2491 // 1 0 1 1 2492 // 1 1 0 1 2493 // 1 1 1 1 2494 // 2495 2496 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2497 int res = 0; 2498 for (int i = 0; i < 8; i++) { 2499 int bit1 = extract_bit(in1, i); 2500 int bit2 = extract_bit(in2, i); 2501 int bit3 = extract_bit(in3, i); 2502 2503 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2504 int func_bit = extract_bit(func, func_bit_pos); 2505 2506 res |= func_bit << i; 2507 } 2508 return res; 2509 } 2510 2511 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2512 assert(n != NULL, ""); 2513 assert(eval_map.contains(n), "absent"); 2514 return *(eval_map.get(n)); 2515 } 2516 2517 static void eval_operands(Node* n, 2518 uint& func1, uint& func2, uint& func3, 2519 ResourceHashtable<Node*,uint>& eval_map) { 2520 assert(is_vector_bitwise_op(n), ""); 2521 2522 if (is_vector_unary_bitwise_op(n)) { 2523 Node* opnd = n->in(1); 2524 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2525 opnd = n->in(2); 2526 } 2527 func1 = eval_operand(opnd, eval_map); 2528 } else if (is_vector_binary_bitwise_op(n)) { 2529 func1 = eval_operand(n->in(1), eval_map); 2530 func2 = eval_operand(n->in(2), eval_map); 2531 } else { 2532 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2533 func1 = eval_operand(n->in(1), eval_map); 2534 func2 = eval_operand(n->in(2), eval_map); 2535 func3 = eval_operand(n->in(3), eval_map); 2536 } 2537 } 2538 2539 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2540 assert(inputs.size() <= 3, "sanity"); 2541 ResourceMark rm; 2542 uint res = 0; 2543 ResourceHashtable<Node*,uint> eval_map; 2544 2545 // Populate precomputed functions for inputs. 2546 // Each input corresponds to one column of 3 input truth-table. 2547 uint input_funcs[] = { 0xAA, // (_, _, a) -> a 2548 0xCC, // (_, b, _) -> b 2549 0xF0 }; // (c, _, _) -> c 2550 for (uint i = 0; i < inputs.size(); i++) { 2551 eval_map.put(inputs.at(i), input_funcs[i]); 2552 } 2553 2554 for (uint i = 0; i < partition.size(); i++) { 2555 Node* n = partition.at(i); 2556 2557 uint func1 = 0, func2 = 0, func3 = 0; 2558 eval_operands(n, func1, func2, func3, eval_map); 2559 2560 switch (n->Opcode()) { 2561 case Op_OrV: 2562 assert(func3 == 0, "not binary"); 2563 res = func1 | func2; 2564 break; 2565 case Op_AndV: 2566 assert(func3 == 0, "not binary"); 2567 res = func1 & func2; 2568 break; 2569 case Op_XorV: 2570 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2571 assert(func2 == 0 && func3 == 0, "not unary"); 2572 res = (~func1) & 0xFF; 2573 } else { 2574 assert(func3 == 0, "not binary"); 2575 res = func1 ^ func2; 2576 } 2577 break; 2578 case Op_MacroLogicV: 2579 // Ordering of inputs may change during evaluation of sub-tree 2580 // containing MacroLogic node as a child node, thus a re-evaluation 2581 // makes sure that function is evaluated in context of current 2582 // inputs. 2583 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2584 break; 2585 2586 default: assert(false, "not supported: %s", n->Name()); 2587 } 2588 assert(res <= 0xFF, "invalid"); 2589 eval_map.put(n, res); 2590 } 2591 return res; 2592 } 2593 2594 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2595 assert(partition.size() == 0, "not empty"); 2596 assert(inputs.size() == 0, "not empty"); 2597 if (is_vector_ternary_bitwise_op(n)) { 2598 return false; 2599 } 2600 2601 bool is_unary_op = is_vector_unary_bitwise_op(n); 2602 if (is_unary_op) { 2603 assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary"); 2604 return false; // too few inputs 2605 } 2606 2607 assert(is_vector_binary_bitwise_op(n), "not binary"); 2608 Node* in1 = n->in(1); 2609 Node* in2 = n->in(2); 2610 2611 int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs); 2612 int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs); 2613 partition.push(n); 2614 2615 // Too many inputs? 2616 if (inputs.size() > 3) { 2617 partition.clear(); 2618 inputs.clear(); 2619 { // Recompute in2 inputs 2620 Unique_Node_List not_used; 2621 in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used); 2622 } 2623 // Pick the node with minimum number of inputs. 2624 if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) { 2625 return false; // still too many inputs 2626 } 2627 // Recompute partition & inputs. 2628 Node* child = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2); 2629 collect_unique_inputs(child, partition, inputs); 2630 2631 Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1); 2632 inputs.push(other_input); 2633 2634 partition.push(n); 2635 } 2636 2637 return (partition.size() == 2 || partition.size() == 3) && 2638 (inputs.size() == 2 || inputs.size() == 3); 2639 } 2640 2641 2642 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2643 assert(is_vector_bitwise_op(n), "not a root"); 2644 2645 visited.set(n->_idx); 2646 2647 // 1) Do a DFS walk over the logic cone. 2648 for (uint i = 1; i < n->req(); i++) { 2649 Node* in = n->in(i); 2650 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2651 process_logic_cone_root(igvn, in, visited); 2652 } 2653 } 2654 2655 // 2) Bottom up traversal: Merge node[s] with 2656 // the parent to form macro logic node. 2657 Unique_Node_List partition; 2658 Unique_Node_List inputs; 2659 if (compute_logic_cone(n, partition, inputs)) { 2660 const TypeVect* vt = n->bottom_type()->is_vect(); 2661 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2662 igvn.replace_node(n, macro_logic); 2663 } 2664 } 2665 2666 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2667 ResourceMark rm; 2668 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2669 Unique_Node_List list; 2670 collect_logic_cone_roots(list); 2671 2672 while (list.size() > 0) { 2673 Node* n = list.pop(); 2674 const TypeVect* vt = n->bottom_type()->is_vect(); 2675 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2676 if (supported) { 2677 VectorSet visited(comp_arena()); 2678 process_logic_cone_root(igvn, n, visited); 2679 } 2680 } 2681 } 2682 } 2683 2684 //------------------------------Code_Gen--------------------------------------- 2685 // Given a graph, generate code for it 2686 void Compile::Code_Gen() { 2687 if (failing()) { 2688 return; 2689 } 2690 2691 // Perform instruction selection. You might think we could reclaim Matcher 2692 // memory PDQ, but actually the Matcher is used in generating spill code. 2693 // Internals of the Matcher (including some VectorSets) must remain live 2694 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2695 // set a bit in reclaimed memory. 2696 2697 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2698 // nodes. Mapping is only valid at the root of each matched subtree. 2699 NOT_PRODUCT( verify_graph_edges(); ) 2700 2701 Matcher matcher; 2702 _matcher = &matcher; 2703 { 2704 TracePhase tp("matcher", &timers[_t_matcher]); 2705 matcher.match(); 2706 if (failing()) { 2707 return; 2708 } 2709 } 2710 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2711 // nodes. Mapping is only valid at the root of each matched subtree. 2712 NOT_PRODUCT( verify_graph_edges(); ) 2713 2714 // If you have too many nodes, or if matching has failed, bail out 2715 check_node_count(0, "out of nodes matching instructions"); 2716 if (failing()) { 2717 return; 2718 } 2719 2720 print_method(PHASE_MATCHING, 2); 2721 2722 // Build a proper-looking CFG 2723 PhaseCFG cfg(node_arena(), root(), matcher); 2724 _cfg = &cfg; 2725 { 2726 TracePhase tp("scheduler", &timers[_t_scheduler]); 2727 bool success = cfg.do_global_code_motion(); 2728 if (!success) { 2729 return; 2730 } 2731 2732 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2733 NOT_PRODUCT( verify_graph_edges(); ) 2734 cfg.verify(); 2735 } 2736 2737 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2738 _regalloc = ®alloc; 2739 { 2740 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2741 // Perform register allocation. After Chaitin, use-def chains are 2742 // no longer accurate (at spill code) and so must be ignored. 2743 // Node->LRG->reg mappings are still accurate. 2744 _regalloc->Register_Allocate(); 2745 2746 // Bail out if the allocator builds too many nodes 2747 if (failing()) { 2748 return; 2749 } 2750 } 2751 2752 // Prior to register allocation we kept empty basic blocks in case the 2753 // the allocator needed a place to spill. After register allocation we 2754 // are not adding any new instructions. If any basic block is empty, we 2755 // can now safely remove it. 2756 { 2757 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2758 cfg.remove_empty_blocks(); 2759 if (do_freq_based_layout()) { 2760 PhaseBlockLayout layout(cfg); 2761 } else { 2762 cfg.set_loop_alignment(); 2763 } 2764 cfg.fixup_flow(); 2765 cfg.remove_unreachable_blocks(); 2766 cfg.verify_dominator_tree(); 2767 } 2768 2769 // Apply peephole optimizations 2770 if( OptoPeephole ) { 2771 TracePhase tp("peephole", &timers[_t_peephole]); 2772 PhasePeephole peep( _regalloc, cfg); 2773 peep.do_transform(); 2774 } 2775 2776 // Do late expand if CPU requires this. 2777 if (Matcher::require_postalloc_expand) { 2778 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2779 cfg.postalloc_expand(_regalloc); 2780 } 2781 2782 // Convert Nodes to instruction bits in a buffer 2783 { 2784 TracePhase tp("output", &timers[_t_output]); 2785 PhaseOutput output; 2786 output.Output(); 2787 if (failing()) return; 2788 output.install(); 2789 } 2790 2791 print_method(PHASE_FINAL_CODE); 2792 2793 // He's dead, Jim. 2794 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2795 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2796 } 2797 2798 //------------------------------Final_Reshape_Counts--------------------------- 2799 // This class defines counters to help identify when a method 2800 // may/must be executed using hardware with only 24-bit precision. 2801 struct Final_Reshape_Counts : public StackObj { 2802 int _call_count; // count non-inlined 'common' calls 2803 int _float_count; // count float ops requiring 24-bit precision 2804 int _double_count; // count double ops requiring more precision 2805 int _java_call_count; // count non-inlined 'java' calls 2806 int _inner_loop_count; // count loops which need alignment 2807 VectorSet _visited; // Visitation flags 2808 Node_List _tests; // Set of IfNodes & PCTableNodes 2809 2810 Final_Reshape_Counts() : 2811 _call_count(0), _float_count(0), _double_count(0), 2812 _java_call_count(0), _inner_loop_count(0) { } 2813 2814 void inc_call_count () { _call_count ++; } 2815 void inc_float_count () { _float_count ++; } 2816 void inc_double_count() { _double_count++; } 2817 void inc_java_call_count() { _java_call_count++; } 2818 void inc_inner_loop_count() { _inner_loop_count++; } 2819 2820 int get_call_count () const { return _call_count ; } 2821 int get_float_count () const { return _float_count ; } 2822 int get_double_count() const { return _double_count; } 2823 int get_java_call_count() const { return _java_call_count; } 2824 int get_inner_loop_count() const { return _inner_loop_count; } 2825 }; 2826 2827 // Eliminate trivially redundant StoreCMs and accumulate their 2828 // precedence edges. 2829 void Compile::eliminate_redundant_card_marks(Node* n) { 2830 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2831 if (n->in(MemNode::Address)->outcnt() > 1) { 2832 // There are multiple users of the same address so it might be 2833 // possible to eliminate some of the StoreCMs 2834 Node* mem = n->in(MemNode::Memory); 2835 Node* adr = n->in(MemNode::Address); 2836 Node* val = n->in(MemNode::ValueIn); 2837 Node* prev = n; 2838 bool done = false; 2839 // Walk the chain of StoreCMs eliminating ones that match. As 2840 // long as it's a chain of single users then the optimization is 2841 // safe. Eliminating partially redundant StoreCMs would require 2842 // cloning copies down the other paths. 2843 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2844 if (adr == mem->in(MemNode::Address) && 2845 val == mem->in(MemNode::ValueIn)) { 2846 // redundant StoreCM 2847 if (mem->req() > MemNode::OopStore) { 2848 // Hasn't been processed by this code yet. 2849 n->add_prec(mem->in(MemNode::OopStore)); 2850 } else { 2851 // Already converted to precedence edge 2852 for (uint i = mem->req(); i < mem->len(); i++) { 2853 // Accumulate any precedence edges 2854 if (mem->in(i) != NULL) { 2855 n->add_prec(mem->in(i)); 2856 } 2857 } 2858 // Everything above this point has been processed. 2859 done = true; 2860 } 2861 // Eliminate the previous StoreCM 2862 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2863 assert(mem->outcnt() == 0, "should be dead"); 2864 mem->disconnect_inputs(this); 2865 } else { 2866 prev = mem; 2867 } 2868 mem = prev->in(MemNode::Memory); 2869 } 2870 } 2871 } 2872 2873 //------------------------------final_graph_reshaping_impl---------------------- 2874 // Implement items 1-5 from final_graph_reshaping below. 2875 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2876 2877 if ( n->outcnt() == 0 ) return; // dead node 2878 uint nop = n->Opcode(); 2879 2880 // Check for 2-input instruction with "last use" on right input. 2881 // Swap to left input. Implements item (2). 2882 if( n->req() == 3 && // two-input instruction 2883 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2884 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2885 n->in(2)->outcnt() == 1 &&// right use IS a last use 2886 !n->in(2)->is_Con() ) { // right use is not a constant 2887 // Check for commutative opcode 2888 switch( nop ) { 2889 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2890 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 2891 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 2892 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2893 case Op_AndL: case Op_XorL: case Op_OrL: 2894 case Op_AndI: case Op_XorI: case Op_OrI: { 2895 // Move "last use" input to left by swapping inputs 2896 n->swap_edges(1, 2); 2897 break; 2898 } 2899 default: 2900 break; 2901 } 2902 } 2903 2904 #ifdef ASSERT 2905 if( n->is_Mem() ) { 2906 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2907 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2908 // oop will be recorded in oop map if load crosses safepoint 2909 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2910 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2911 "raw memory operations should have control edge"); 2912 } 2913 if (n->is_MemBar()) { 2914 MemBarNode* mb = n->as_MemBar(); 2915 if (mb->trailing_store() || mb->trailing_load_store()) { 2916 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2917 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 2918 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2919 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2920 } else if (mb->leading()) { 2921 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2922 } 2923 } 2924 #endif 2925 // Count FPU ops and common calls, implements item (3) 2926 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2927 if (!gc_handled) { 2928 final_graph_reshaping_main_switch(n, frc, nop); 2929 } 2930 2931 // Collect CFG split points 2932 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2933 frc._tests.push(n); 2934 } 2935 } 2936 2937 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2938 switch( nop ) { 2939 // Count all float operations that may use FPU 2940 case Op_AddF: 2941 case Op_SubF: 2942 case Op_MulF: 2943 case Op_DivF: 2944 case Op_NegF: 2945 case Op_ModF: 2946 case Op_ConvI2F: 2947 case Op_ConF: 2948 case Op_CmpF: 2949 case Op_CmpF3: 2950 case Op_StoreF: 2951 case Op_LoadF: 2952 // case Op_ConvL2F: // longs are split into 32-bit halves 2953 frc.inc_float_count(); 2954 break; 2955 2956 case Op_ConvF2D: 2957 case Op_ConvD2F: 2958 frc.inc_float_count(); 2959 frc.inc_double_count(); 2960 break; 2961 2962 // Count all double operations that may use FPU 2963 case Op_AddD: 2964 case Op_SubD: 2965 case Op_MulD: 2966 case Op_DivD: 2967 case Op_NegD: 2968 case Op_ModD: 2969 case Op_ConvI2D: 2970 case Op_ConvD2I: 2971 // case Op_ConvL2D: // handled by leaf call 2972 // case Op_ConvD2L: // handled by leaf call 2973 case Op_ConD: 2974 case Op_CmpD: 2975 case Op_CmpD3: 2976 case Op_StoreD: 2977 case Op_LoadD: 2978 case Op_LoadD_unaligned: 2979 frc.inc_double_count(); 2980 break; 2981 case Op_Opaque1: // Remove Opaque Nodes before matching 2982 case Op_Opaque2: // Remove Opaque Nodes before matching 2983 case Op_Opaque3: 2984 n->subsume_by(n->in(1), this); 2985 break; 2986 case Op_CallStaticJava: 2987 case Op_CallJava: 2988 case Op_CallDynamicJava: 2989 frc.inc_java_call_count(); // Count java call site; 2990 case Op_CallRuntime: 2991 case Op_CallLeaf: 2992 case Op_CallLeafVector: 2993 case Op_CallNative: 2994 case Op_CallLeafNoFP: { 2995 assert (n->is_Call(), ""); 2996 CallNode *call = n->as_Call(); 2997 // Count call sites where the FP mode bit would have to be flipped. 2998 // Do not count uncommon runtime calls: 2999 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3000 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3001 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3002 frc.inc_call_count(); // Count the call site 3003 } else { // See if uncommon argument is shared 3004 Node *n = call->in(TypeFunc::Parms); 3005 int nop = n->Opcode(); 3006 // Clone shared simple arguments to uncommon calls, item (1). 3007 if (n->outcnt() > 1 && 3008 !n->is_Proj() && 3009 nop != Op_CreateEx && 3010 nop != Op_CheckCastPP && 3011 nop != Op_DecodeN && 3012 nop != Op_DecodeNKlass && 3013 !n->is_Mem() && 3014 !n->is_Phi()) { 3015 Node *x = n->clone(); 3016 call->set_req(TypeFunc::Parms, x); 3017 } 3018 } 3019 break; 3020 } 3021 3022 case Op_StoreCM: 3023 { 3024 // Convert OopStore dependence into precedence edge 3025 Node* prec = n->in(MemNode::OopStore); 3026 n->del_req(MemNode::OopStore); 3027 n->add_prec(prec); 3028 eliminate_redundant_card_marks(n); 3029 } 3030 3031 // fall through 3032 3033 case Op_StoreB: 3034 case Op_StoreC: 3035 case Op_StorePConditional: 3036 case Op_StoreI: 3037 case Op_StoreL: 3038 case Op_StoreIConditional: 3039 case Op_StoreLConditional: 3040 case Op_CompareAndSwapB: 3041 case Op_CompareAndSwapS: 3042 case Op_CompareAndSwapI: 3043 case Op_CompareAndSwapL: 3044 case Op_CompareAndSwapP: 3045 case Op_CompareAndSwapN: 3046 case Op_WeakCompareAndSwapB: 3047 case Op_WeakCompareAndSwapS: 3048 case Op_WeakCompareAndSwapI: 3049 case Op_WeakCompareAndSwapL: 3050 case Op_WeakCompareAndSwapP: 3051 case Op_WeakCompareAndSwapN: 3052 case Op_CompareAndExchangeB: 3053 case Op_CompareAndExchangeS: 3054 case Op_CompareAndExchangeI: 3055 case Op_CompareAndExchangeL: 3056 case Op_CompareAndExchangeP: 3057 case Op_CompareAndExchangeN: 3058 case Op_GetAndAddS: 3059 case Op_GetAndAddB: 3060 case Op_GetAndAddI: 3061 case Op_GetAndAddL: 3062 case Op_GetAndSetS: 3063 case Op_GetAndSetB: 3064 case Op_GetAndSetI: 3065 case Op_GetAndSetL: 3066 case Op_GetAndSetP: 3067 case Op_GetAndSetN: 3068 case Op_StoreP: 3069 case Op_StoreN: 3070 case Op_StoreNKlass: 3071 case Op_LoadB: 3072 case Op_LoadUB: 3073 case Op_LoadUS: 3074 case Op_LoadI: 3075 case Op_LoadKlass: 3076 case Op_LoadNKlass: 3077 case Op_LoadL: 3078 case Op_LoadL_unaligned: 3079 case Op_LoadPLocked: 3080 case Op_LoadP: 3081 case Op_LoadN: 3082 case Op_LoadRange: 3083 case Op_LoadS: 3084 break; 3085 3086 case Op_AddP: { // Assert sane base pointers 3087 Node *addp = n->in(AddPNode::Address); 3088 assert( !addp->is_AddP() || 3089 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3090 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3091 "Base pointers must match (addp %u)", addp->_idx ); 3092 #ifdef _LP64 3093 if ((UseCompressedOops || UseCompressedClassPointers) && 3094 addp->Opcode() == Op_ConP && 3095 addp == n->in(AddPNode::Base) && 3096 n->in(AddPNode::Offset)->is_Con()) { 3097 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3098 // on the platform and on the compressed oops mode. 3099 // Use addressing with narrow klass to load with offset on x86. 3100 // Some platforms can use the constant pool to load ConP. 3101 // Do this transformation here since IGVN will convert ConN back to ConP. 3102 const Type* t = addp->bottom_type(); 3103 bool is_oop = t->isa_oopptr() != NULL; 3104 bool is_klass = t->isa_klassptr() != NULL; 3105 3106 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3107 (is_klass && Matcher::const_klass_prefer_decode())) { 3108 Node* nn = NULL; 3109 3110 int op = is_oop ? Op_ConN : Op_ConNKlass; 3111 3112 // Look for existing ConN node of the same exact type. 3113 Node* r = root(); 3114 uint cnt = r->outcnt(); 3115 for (uint i = 0; i < cnt; i++) { 3116 Node* m = r->raw_out(i); 3117 if (m!= NULL && m->Opcode() == op && 3118 m->bottom_type()->make_ptr() == t) { 3119 nn = m; 3120 break; 3121 } 3122 } 3123 if (nn != NULL) { 3124 // Decode a narrow oop to match address 3125 // [R12 + narrow_oop_reg<<3 + offset] 3126 if (is_oop) { 3127 nn = new DecodeNNode(nn, t); 3128 } else { 3129 nn = new DecodeNKlassNode(nn, t); 3130 } 3131 // Check for succeeding AddP which uses the same Base. 3132 // Otherwise we will run into the assertion above when visiting that guy. 3133 for (uint i = 0; i < n->outcnt(); ++i) { 3134 Node *out_i = n->raw_out(i); 3135 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3136 out_i->set_req(AddPNode::Base, nn); 3137 #ifdef ASSERT 3138 for (uint j = 0; j < out_i->outcnt(); ++j) { 3139 Node *out_j = out_i->raw_out(j); 3140 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3141 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3142 } 3143 #endif 3144 } 3145 } 3146 n->set_req(AddPNode::Base, nn); 3147 n->set_req(AddPNode::Address, nn); 3148 if (addp->outcnt() == 0) { 3149 addp->disconnect_inputs(this); 3150 } 3151 } 3152 } 3153 } 3154 #endif 3155 break; 3156 } 3157 3158 case Op_CastPP: { 3159 // Remove CastPP nodes to gain more freedom during scheduling but 3160 // keep the dependency they encode as control or precedence edges 3161 // (if control is set already) on memory operations. Some CastPP 3162 // nodes don't have a control (don't carry a dependency): skip 3163 // those. 3164 if (n->in(0) != NULL) { 3165 ResourceMark rm; 3166 Unique_Node_List wq; 3167 wq.push(n); 3168 for (uint next = 0; next < wq.size(); ++next) { 3169 Node *m = wq.at(next); 3170 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3171 Node* use = m->fast_out(i); 3172 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3173 use->ensure_control_or_add_prec(n->in(0)); 3174 } else { 3175 switch(use->Opcode()) { 3176 case Op_AddP: 3177 case Op_DecodeN: 3178 case Op_DecodeNKlass: 3179 case Op_CheckCastPP: 3180 case Op_CastPP: 3181 wq.push(use); 3182 break; 3183 } 3184 } 3185 } 3186 } 3187 } 3188 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3189 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3190 Node* in1 = n->in(1); 3191 const Type* t = n->bottom_type(); 3192 Node* new_in1 = in1->clone(); 3193 new_in1->as_DecodeN()->set_type(t); 3194 3195 if (!Matcher::narrow_oop_use_complex_address()) { 3196 // 3197 // x86, ARM and friends can handle 2 adds in addressing mode 3198 // and Matcher can fold a DecodeN node into address by using 3199 // a narrow oop directly and do implicit NULL check in address: 3200 // 3201 // [R12 + narrow_oop_reg<<3 + offset] 3202 // NullCheck narrow_oop_reg 3203 // 3204 // On other platforms (Sparc) we have to keep new DecodeN node and 3205 // use it to do implicit NULL check in address: 3206 // 3207 // decode_not_null narrow_oop_reg, base_reg 3208 // [base_reg + offset] 3209 // NullCheck base_reg 3210 // 3211 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3212 // to keep the information to which NULL check the new DecodeN node 3213 // corresponds to use it as value in implicit_null_check(). 3214 // 3215 new_in1->set_req(0, n->in(0)); 3216 } 3217 3218 n->subsume_by(new_in1, this); 3219 if (in1->outcnt() == 0) { 3220 in1->disconnect_inputs(this); 3221 } 3222 } else { 3223 n->subsume_by(n->in(1), this); 3224 if (n->outcnt() == 0) { 3225 n->disconnect_inputs(this); 3226 } 3227 } 3228 break; 3229 } 3230 #ifdef _LP64 3231 case Op_CmpP: 3232 // Do this transformation here to preserve CmpPNode::sub() and 3233 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3234 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3235 Node* in1 = n->in(1); 3236 Node* in2 = n->in(2); 3237 if (!in1->is_DecodeNarrowPtr()) { 3238 in2 = in1; 3239 in1 = n->in(2); 3240 } 3241 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3242 3243 Node* new_in2 = NULL; 3244 if (in2->is_DecodeNarrowPtr()) { 3245 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3246 new_in2 = in2->in(1); 3247 } else if (in2->Opcode() == Op_ConP) { 3248 const Type* t = in2->bottom_type(); 3249 if (t == TypePtr::NULL_PTR) { 3250 assert(in1->is_DecodeN(), "compare klass to null?"); 3251 // Don't convert CmpP null check into CmpN if compressed 3252 // oops implicit null check is not generated. 3253 // This will allow to generate normal oop implicit null check. 3254 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3255 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3256 // 3257 // This transformation together with CastPP transformation above 3258 // will generated code for implicit NULL checks for compressed oops. 3259 // 3260 // The original code after Optimize() 3261 // 3262 // LoadN memory, narrow_oop_reg 3263 // decode narrow_oop_reg, base_reg 3264 // CmpP base_reg, NULL 3265 // CastPP base_reg // NotNull 3266 // Load [base_reg + offset], val_reg 3267 // 3268 // after these transformations will be 3269 // 3270 // LoadN memory, narrow_oop_reg 3271 // CmpN narrow_oop_reg, NULL 3272 // decode_not_null narrow_oop_reg, base_reg 3273 // Load [base_reg + offset], val_reg 3274 // 3275 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3276 // since narrow oops can be used in debug info now (see the code in 3277 // final_graph_reshaping_walk()). 3278 // 3279 // At the end the code will be matched to 3280 // on x86: 3281 // 3282 // Load_narrow_oop memory, narrow_oop_reg 3283 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3284 // NullCheck narrow_oop_reg 3285 // 3286 // and on sparc: 3287 // 3288 // Load_narrow_oop memory, narrow_oop_reg 3289 // decode_not_null narrow_oop_reg, base_reg 3290 // Load [base_reg + offset], val_reg 3291 // NullCheck base_reg 3292 // 3293 } else if (t->isa_oopptr()) { 3294 new_in2 = ConNode::make(t->make_narrowoop()); 3295 } else if (t->isa_klassptr()) { 3296 new_in2 = ConNode::make(t->make_narrowklass()); 3297 } 3298 } 3299 if (new_in2 != NULL) { 3300 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3301 n->subsume_by(cmpN, this); 3302 if (in1->outcnt() == 0) { 3303 in1->disconnect_inputs(this); 3304 } 3305 if (in2->outcnt() == 0) { 3306 in2->disconnect_inputs(this); 3307 } 3308 } 3309 } 3310 break; 3311 3312 case Op_DecodeN: 3313 case Op_DecodeNKlass: 3314 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3315 // DecodeN could be pinned when it can't be fold into 3316 // an address expression, see the code for Op_CastPP above. 3317 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3318 break; 3319 3320 case Op_EncodeP: 3321 case Op_EncodePKlass: { 3322 Node* in1 = n->in(1); 3323 if (in1->is_DecodeNarrowPtr()) { 3324 n->subsume_by(in1->in(1), this); 3325 } else if (in1->Opcode() == Op_ConP) { 3326 const Type* t = in1->bottom_type(); 3327 if (t == TypePtr::NULL_PTR) { 3328 assert(t->isa_oopptr(), "null klass?"); 3329 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3330 } else if (t->isa_oopptr()) { 3331 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3332 } else if (t->isa_klassptr()) { 3333 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3334 } 3335 } 3336 if (in1->outcnt() == 0) { 3337 in1->disconnect_inputs(this); 3338 } 3339 break; 3340 } 3341 3342 case Op_Proj: { 3343 if (OptimizeStringConcat || IncrementalInline) { 3344 ProjNode* proj = n->as_Proj(); 3345 if (proj->_is_io_use) { 3346 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3347 // Separate projections were used for the exception path which 3348 // are normally removed by a late inline. If it wasn't inlined 3349 // then they will hang around and should just be replaced with 3350 // the original one. Merge them. 3351 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3352 if (non_io_proj != NULL) { 3353 proj->subsume_by(non_io_proj , this); 3354 } 3355 } 3356 } 3357 break; 3358 } 3359 3360 case Op_Phi: 3361 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3362 // The EncodeP optimization may create Phi with the same edges 3363 // for all paths. It is not handled well by Register Allocator. 3364 Node* unique_in = n->in(1); 3365 assert(unique_in != NULL, ""); 3366 uint cnt = n->req(); 3367 for (uint i = 2; i < cnt; i++) { 3368 Node* m = n->in(i); 3369 assert(m != NULL, ""); 3370 if (unique_in != m) 3371 unique_in = NULL; 3372 } 3373 if (unique_in != NULL) { 3374 n->subsume_by(unique_in, this); 3375 } 3376 } 3377 break; 3378 3379 #endif 3380 3381 #ifdef ASSERT 3382 case Op_CastII: 3383 // Verify that all range check dependent CastII nodes were removed. 3384 if (n->isa_CastII()->has_range_check()) { 3385 n->dump(3); 3386 assert(false, "Range check dependent CastII node was not removed"); 3387 } 3388 break; 3389 #endif 3390 3391 case Op_ModI: 3392 if (UseDivMod) { 3393 // Check if a%b and a/b both exist 3394 Node* d = n->find_similar(Op_DivI); 3395 if (d) { 3396 // Replace them with a fused divmod if supported 3397 if (Matcher::has_match_rule(Op_DivModI)) { 3398 DivModINode* divmod = DivModINode::make(n); 3399 d->subsume_by(divmod->div_proj(), this); 3400 n->subsume_by(divmod->mod_proj(), this); 3401 } else { 3402 // replace a%b with a-((a/b)*b) 3403 Node* mult = new MulINode(d, d->in(2)); 3404 Node* sub = new SubINode(d->in(1), mult); 3405 n->subsume_by(sub, this); 3406 } 3407 } 3408 } 3409 break; 3410 3411 case Op_ModL: 3412 if (UseDivMod) { 3413 // Check if a%b and a/b both exist 3414 Node* d = n->find_similar(Op_DivL); 3415 if (d) { 3416 // Replace them with a fused divmod if supported 3417 if (Matcher::has_match_rule(Op_DivModL)) { 3418 DivModLNode* divmod = DivModLNode::make(n); 3419 d->subsume_by(divmod->div_proj(), this); 3420 n->subsume_by(divmod->mod_proj(), this); 3421 } else { 3422 // replace a%b with a-((a/b)*b) 3423 Node* mult = new MulLNode(d, d->in(2)); 3424 Node* sub = new SubLNode(d->in(1), mult); 3425 n->subsume_by(sub, this); 3426 } 3427 } 3428 } 3429 break; 3430 3431 case Op_LoadVector: 3432 case Op_StoreVector: 3433 case Op_LoadVectorGather: 3434 case Op_StoreVectorScatter: 3435 case Op_VectorCmpMasked: 3436 case Op_VectorMaskGen: 3437 case Op_LoadVectorMasked: 3438 case Op_StoreVectorMasked: 3439 break; 3440 3441 case Op_AddReductionVI: 3442 case Op_AddReductionVL: 3443 case Op_AddReductionVF: 3444 case Op_AddReductionVD: 3445 case Op_MulReductionVI: 3446 case Op_MulReductionVL: 3447 case Op_MulReductionVF: 3448 case Op_MulReductionVD: 3449 case Op_MinReductionV: 3450 case Op_MaxReductionV: 3451 case Op_AndReductionV: 3452 case Op_OrReductionV: 3453 case Op_XorReductionV: 3454 break; 3455 3456 case Op_PackB: 3457 case Op_PackS: 3458 case Op_PackI: 3459 case Op_PackF: 3460 case Op_PackL: 3461 case Op_PackD: 3462 if (n->req()-1 > 2) { 3463 // Replace many operand PackNodes with a binary tree for matching 3464 PackNode* p = (PackNode*) n; 3465 Node* btp = p->binary_tree_pack(1, n->req()); 3466 n->subsume_by(btp, this); 3467 } 3468 break; 3469 case Op_Loop: 3470 assert(!n->as_Loop()->is_transformed_long_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3471 case Op_CountedLoop: 3472 case Op_LongCountedLoop: 3473 case Op_OuterStripMinedLoop: 3474 if (n->as_Loop()->is_inner_loop()) { 3475 frc.inc_inner_loop_count(); 3476 } 3477 n->as_Loop()->verify_strip_mined(0); 3478 break; 3479 case Op_LShiftI: 3480 case Op_RShiftI: 3481 case Op_URShiftI: 3482 case Op_LShiftL: 3483 case Op_RShiftL: 3484 case Op_URShiftL: 3485 if (Matcher::need_masked_shift_count) { 3486 // The cpu's shift instructions don't restrict the count to the 3487 // lower 5/6 bits. We need to do the masking ourselves. 3488 Node* in2 = n->in(2); 3489 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3490 const TypeInt* t = in2->find_int_type(); 3491 if (t != NULL && t->is_con()) { 3492 juint shift = t->get_con(); 3493 if (shift > mask) { // Unsigned cmp 3494 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3495 } 3496 } else { 3497 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3498 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3499 n->set_req(2, shift); 3500 } 3501 } 3502 if (in2->outcnt() == 0) { // Remove dead node 3503 in2->disconnect_inputs(this); 3504 } 3505 } 3506 break; 3507 case Op_MemBarStoreStore: 3508 case Op_MemBarRelease: 3509 // Break the link with AllocateNode: it is no longer useful and 3510 // confuses register allocation. 3511 if (n->req() > MemBarNode::Precedent) { 3512 n->set_req(MemBarNode::Precedent, top()); 3513 } 3514 break; 3515 case Op_MemBarAcquire: { 3516 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3517 // At parse time, the trailing MemBarAcquire for a volatile load 3518 // is created with an edge to the load. After optimizations, 3519 // that input may be a chain of Phis. If those phis have no 3520 // other use, then the MemBarAcquire keeps them alive and 3521 // register allocation can be confused. 3522 ResourceMark rm; 3523 Unique_Node_List wq; 3524 wq.push(n->in(MemBarNode::Precedent)); 3525 n->set_req(MemBarNode::Precedent, top()); 3526 while (wq.size() > 0) { 3527 Node* m = wq.pop(); 3528 if (m->outcnt() == 0 && m != top()) { 3529 for (uint j = 0; j < m->req(); j++) { 3530 Node* in = m->in(j); 3531 if (in != NULL) { 3532 wq.push(in); 3533 } 3534 } 3535 m->disconnect_inputs(this); 3536 } 3537 } 3538 } 3539 break; 3540 } 3541 case Op_Blackhole: 3542 break; 3543 case Op_RangeCheck: { 3544 RangeCheckNode* rc = n->as_RangeCheck(); 3545 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3546 n->subsume_by(iff, this); 3547 frc._tests.push(iff); 3548 break; 3549 } 3550 case Op_ConvI2L: { 3551 if (!Matcher::convi2l_type_required) { 3552 // Code generation on some platforms doesn't need accurate 3553 // ConvI2L types. Widening the type can help remove redundant 3554 // address computations. 3555 n->as_Type()->set_type(TypeLong::INT); 3556 ResourceMark rm; 3557 Unique_Node_List wq; 3558 wq.push(n); 3559 for (uint next = 0; next < wq.size(); next++) { 3560 Node *m = wq.at(next); 3561 3562 for(;;) { 3563 // Loop over all nodes with identical inputs edges as m 3564 Node* k = m->find_similar(m->Opcode()); 3565 if (k == NULL) { 3566 break; 3567 } 3568 // Push their uses so we get a chance to remove node made 3569 // redundant 3570 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3571 Node* u = k->fast_out(i); 3572 if (u->Opcode() == Op_LShiftL || 3573 u->Opcode() == Op_AddL || 3574 u->Opcode() == Op_SubL || 3575 u->Opcode() == Op_AddP) { 3576 wq.push(u); 3577 } 3578 } 3579 // Replace all nodes with identical edges as m with m 3580 k->subsume_by(m, this); 3581 } 3582 } 3583 } 3584 break; 3585 } 3586 case Op_CmpUL: { 3587 if (!Matcher::has_match_rule(Op_CmpUL)) { 3588 // No support for unsigned long comparisons 3589 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3590 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3591 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3592 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3593 Node* andl = new AndLNode(orl, remove_sign_mask); 3594 Node* cmp = new CmpLNode(andl, n->in(2)); 3595 n->subsume_by(cmp, this); 3596 } 3597 break; 3598 } 3599 default: 3600 assert(!n->is_Call(), ""); 3601 assert(!n->is_Mem(), ""); 3602 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3603 break; 3604 } 3605 } 3606 3607 //------------------------------final_graph_reshaping_walk--------------------- 3608 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3609 // requires that the walk visits a node's inputs before visiting the node. 3610 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3611 Unique_Node_List sfpt; 3612 3613 frc._visited.set(root->_idx); // first, mark node as visited 3614 uint cnt = root->req(); 3615 Node *n = root; 3616 uint i = 0; 3617 while (true) { 3618 if (i < cnt) { 3619 // Place all non-visited non-null inputs onto stack 3620 Node* m = n->in(i); 3621 ++i; 3622 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3623 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3624 // compute worst case interpreter size in case of a deoptimization 3625 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3626 3627 sfpt.push(m); 3628 } 3629 cnt = m->req(); 3630 nstack.push(n, i); // put on stack parent and next input's index 3631 n = m; 3632 i = 0; 3633 } 3634 } else { 3635 // Now do post-visit work 3636 final_graph_reshaping_impl( n, frc ); 3637 if (nstack.is_empty()) 3638 break; // finished 3639 n = nstack.node(); // Get node from stack 3640 cnt = n->req(); 3641 i = nstack.index(); 3642 nstack.pop(); // Shift to the next node on stack 3643 } 3644 } 3645 3646 // Skip next transformation if compressed oops are not used. 3647 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3648 (!UseCompressedOops && !UseCompressedClassPointers)) 3649 return; 3650 3651 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3652 // It could be done for an uncommon traps or any safepoints/calls 3653 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3654 while (sfpt.size() > 0) { 3655 n = sfpt.pop(); 3656 JVMState *jvms = n->as_SafePoint()->jvms(); 3657 assert(jvms != NULL, "sanity"); 3658 int start = jvms->debug_start(); 3659 int end = n->req(); 3660 bool is_uncommon = (n->is_CallStaticJava() && 3661 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3662 for (int j = start; j < end; j++) { 3663 Node* in = n->in(j); 3664 if (in->is_DecodeNarrowPtr()) { 3665 bool safe_to_skip = true; 3666 if (!is_uncommon ) { 3667 // Is it safe to skip? 3668 for (uint i = 0; i < in->outcnt(); i++) { 3669 Node* u = in->raw_out(i); 3670 if (!u->is_SafePoint() || 3671 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3672 safe_to_skip = false; 3673 } 3674 } 3675 } 3676 if (safe_to_skip) { 3677 n->set_req(j, in->in(1)); 3678 } 3679 if (in->outcnt() == 0) { 3680 in->disconnect_inputs(this); 3681 } 3682 } 3683 } 3684 } 3685 } 3686 3687 //------------------------------final_graph_reshaping-------------------------- 3688 // Final Graph Reshaping. 3689 // 3690 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3691 // and not commoned up and forced early. Must come after regular 3692 // optimizations to avoid GVN undoing the cloning. Clone constant 3693 // inputs to Loop Phis; these will be split by the allocator anyways. 3694 // Remove Opaque nodes. 3695 // (2) Move last-uses by commutative operations to the left input to encourage 3696 // Intel update-in-place two-address operations and better register usage 3697 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3698 // calls canonicalizing them back. 3699 // (3) Count the number of double-precision FP ops, single-precision FP ops 3700 // and call sites. On Intel, we can get correct rounding either by 3701 // forcing singles to memory (requires extra stores and loads after each 3702 // FP bytecode) or we can set a rounding mode bit (requires setting and 3703 // clearing the mode bit around call sites). The mode bit is only used 3704 // if the relative frequency of single FP ops to calls is low enough. 3705 // This is a key transform for SPEC mpeg_audio. 3706 // (4) Detect infinite loops; blobs of code reachable from above but not 3707 // below. Several of the Code_Gen algorithms fail on such code shapes, 3708 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3709 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3710 // Detection is by looking for IfNodes where only 1 projection is 3711 // reachable from below or CatchNodes missing some targets. 3712 // (5) Assert for insane oop offsets in debug mode. 3713 3714 bool Compile::final_graph_reshaping() { 3715 // an infinite loop may have been eliminated by the optimizer, 3716 // in which case the graph will be empty. 3717 if (root()->req() == 1) { 3718 record_method_not_compilable("trivial infinite loop"); 3719 return true; 3720 } 3721 3722 // Expensive nodes have their control input set to prevent the GVN 3723 // from freely commoning them. There's no GVN beyond this point so 3724 // no need to keep the control input. We want the expensive nodes to 3725 // be freely moved to the least frequent code path by gcm. 3726 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3727 for (int i = 0; i < expensive_count(); i++) { 3728 _expensive_nodes.at(i)->set_req(0, NULL); 3729 } 3730 3731 Final_Reshape_Counts frc; 3732 3733 // Visit everybody reachable! 3734 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3735 Node_Stack nstack(live_nodes() >> 1); 3736 final_graph_reshaping_walk(nstack, root(), frc); 3737 3738 // Check for unreachable (from below) code (i.e., infinite loops). 3739 for( uint i = 0; i < frc._tests.size(); i++ ) { 3740 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3741 // Get number of CFG targets. 3742 // Note that PCTables include exception targets after calls. 3743 uint required_outcnt = n->required_outcnt(); 3744 if (n->outcnt() != required_outcnt) { 3745 // Check for a few special cases. Rethrow Nodes never take the 3746 // 'fall-thru' path, so expected kids is 1 less. 3747 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3748 if (n->in(0)->in(0)->is_Call()) { 3749 CallNode *call = n->in(0)->in(0)->as_Call(); 3750 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3751 required_outcnt--; // Rethrow always has 1 less kid 3752 } else if (call->req() > TypeFunc::Parms && 3753 call->is_CallDynamicJava()) { 3754 // Check for null receiver. In such case, the optimizer has 3755 // detected that the virtual call will always result in a null 3756 // pointer exception. The fall-through projection of this CatchNode 3757 // will not be populated. 3758 Node *arg0 = call->in(TypeFunc::Parms); 3759 if (arg0->is_Type() && 3760 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3761 required_outcnt--; 3762 } 3763 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3764 call->req() > TypeFunc::Parms+1 && 3765 call->is_CallStaticJava()) { 3766 // Check for negative array length. In such case, the optimizer has 3767 // detected that the allocation attempt will always result in an 3768 // exception. There is no fall-through projection of this CatchNode . 3769 Node *arg1 = call->in(TypeFunc::Parms+1); 3770 if (arg1->is_Type() && 3771 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3772 required_outcnt--; 3773 } 3774 } 3775 } 3776 } 3777 // Recheck with a better notion of 'required_outcnt' 3778 if (n->outcnt() != required_outcnt) { 3779 record_method_not_compilable("malformed control flow"); 3780 return true; // Not all targets reachable! 3781 } 3782 } 3783 // Check that I actually visited all kids. Unreached kids 3784 // must be infinite loops. 3785 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3786 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3787 record_method_not_compilable("infinite loop"); 3788 return true; // Found unvisited kid; must be unreach 3789 } 3790 3791 // Here so verification code in final_graph_reshaping_walk() 3792 // always see an OuterStripMinedLoopEnd 3793 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 3794 IfNode* init_iff = n->as_If(); 3795 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3796 n->subsume_by(iff, this); 3797 } 3798 } 3799 3800 #ifdef IA32 3801 // If original bytecodes contained a mixture of floats and doubles 3802 // check if the optimizer has made it homogenous, item (3). 3803 if (UseSSE == 0 && 3804 frc.get_float_count() > 32 && 3805 frc.get_double_count() == 0 && 3806 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3807 set_24_bit_selection_and_mode(false, true); 3808 } 3809 #endif // IA32 3810 3811 set_java_calls(frc.get_java_call_count()); 3812 set_inner_loops(frc.get_inner_loop_count()); 3813 3814 // No infinite loops, no reason to bail out. 3815 return false; 3816 } 3817 3818 //-----------------------------too_many_traps---------------------------------- 3819 // Report if there are too many traps at the current method and bci. 3820 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3821 bool Compile::too_many_traps(ciMethod* method, 3822 int bci, 3823 Deoptimization::DeoptReason reason) { 3824 ciMethodData* md = method->method_data(); 3825 if (md->is_empty()) { 3826 // Assume the trap has not occurred, or that it occurred only 3827 // because of a transient condition during start-up in the interpreter. 3828 return false; 3829 } 3830 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3831 if (md->has_trap_at(bci, m, reason) != 0) { 3832 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3833 // Also, if there are multiple reasons, or if there is no per-BCI record, 3834 // assume the worst. 3835 if (log()) 3836 log()->elem("observe trap='%s' count='%d'", 3837 Deoptimization::trap_reason_name(reason), 3838 md->trap_count(reason)); 3839 return true; 3840 } else { 3841 // Ignore method/bci and see if there have been too many globally. 3842 return too_many_traps(reason, md); 3843 } 3844 } 3845 3846 // Less-accurate variant which does not require a method and bci. 3847 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3848 ciMethodData* logmd) { 3849 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3850 // Too many traps globally. 3851 // Note that we use cumulative trap_count, not just md->trap_count. 3852 if (log()) { 3853 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3854 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3855 Deoptimization::trap_reason_name(reason), 3856 mcount, trap_count(reason)); 3857 } 3858 return true; 3859 } else { 3860 // The coast is clear. 3861 return false; 3862 } 3863 } 3864 3865 //--------------------------too_many_recompiles-------------------------------- 3866 // Report if there are too many recompiles at the current method and bci. 3867 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3868 // Is not eager to return true, since this will cause the compiler to use 3869 // Action_none for a trap point, to avoid too many recompilations. 3870 bool Compile::too_many_recompiles(ciMethod* method, 3871 int bci, 3872 Deoptimization::DeoptReason reason) { 3873 ciMethodData* md = method->method_data(); 3874 if (md->is_empty()) { 3875 // Assume the trap has not occurred, or that it occurred only 3876 // because of a transient condition during start-up in the interpreter. 3877 return false; 3878 } 3879 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3880 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3881 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3882 Deoptimization::DeoptReason per_bc_reason 3883 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3884 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3885 if ((per_bc_reason == Deoptimization::Reason_none 3886 || md->has_trap_at(bci, m, reason) != 0) 3887 // The trap frequency measure we care about is the recompile count: 3888 && md->trap_recompiled_at(bci, m) 3889 && md->overflow_recompile_count() >= bc_cutoff) { 3890 // Do not emit a trap here if it has already caused recompilations. 3891 // Also, if there are multiple reasons, or if there is no per-BCI record, 3892 // assume the worst. 3893 if (log()) 3894 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3895 Deoptimization::trap_reason_name(reason), 3896 md->trap_count(reason), 3897 md->overflow_recompile_count()); 3898 return true; 3899 } else if (trap_count(reason) != 0 3900 && decompile_count() >= m_cutoff) { 3901 // Too many recompiles globally, and we have seen this sort of trap. 3902 // Use cumulative decompile_count, not just md->decompile_count. 3903 if (log()) 3904 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3905 Deoptimization::trap_reason_name(reason), 3906 md->trap_count(reason), trap_count(reason), 3907 md->decompile_count(), decompile_count()); 3908 return true; 3909 } else { 3910 // The coast is clear. 3911 return false; 3912 } 3913 } 3914 3915 // Compute when not to trap. Used by matching trap based nodes and 3916 // NullCheck optimization. 3917 void Compile::set_allowed_deopt_reasons() { 3918 _allowed_reasons = 0; 3919 if (is_method_compilation()) { 3920 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3921 assert(rs < BitsPerInt, "recode bit map"); 3922 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3923 _allowed_reasons |= nth_bit(rs); 3924 } 3925 } 3926 } 3927 } 3928 3929 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 3930 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 3931 } 3932 3933 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 3934 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 3935 } 3936 3937 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 3938 if (holder->is_initialized()) { 3939 return false; 3940 } 3941 if (holder->is_being_initialized()) { 3942 if (accessing_method->holder() == holder) { 3943 // Access inside a class. The barrier can be elided when access happens in <clinit>, 3944 // <init>, or a static method. In all those cases, there was an initialization 3945 // barrier on the holder klass passed. 3946 if (accessing_method->is_static_initializer() || 3947 accessing_method->is_object_initializer() || 3948 accessing_method->is_static()) { 3949 return false; 3950 } 3951 } else if (accessing_method->holder()->is_subclass_of(holder)) { 3952 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 3953 // In case of <init> or a static method, the barrier is on the subclass is not enough: 3954 // child class can become fully initialized while its parent class is still being initialized. 3955 if (accessing_method->is_static_initializer()) { 3956 return false; 3957 } 3958 } 3959 ciMethod* root = method(); // the root method of compilation 3960 if (root != accessing_method) { 3961 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 3962 } 3963 } 3964 return true; 3965 } 3966 3967 #ifndef PRODUCT 3968 //------------------------------verify_graph_edges--------------------------- 3969 // Walk the Graph and verify that there is a one-to-one correspondence 3970 // between Use-Def edges and Def-Use edges in the graph. 3971 void Compile::verify_graph_edges(bool no_dead_code) { 3972 if (VerifyGraphEdges) { 3973 Unique_Node_List visited; 3974 // Call recursive graph walk to check edges 3975 _root->verify_edges(visited); 3976 if (no_dead_code) { 3977 // Now make sure that no visited node is used by an unvisited node. 3978 bool dead_nodes = false; 3979 Unique_Node_List checked; 3980 while (visited.size() > 0) { 3981 Node* n = visited.pop(); 3982 checked.push(n); 3983 for (uint i = 0; i < n->outcnt(); i++) { 3984 Node* use = n->raw_out(i); 3985 if (checked.member(use)) continue; // already checked 3986 if (visited.member(use)) continue; // already in the graph 3987 if (use->is_Con()) continue; // a dead ConNode is OK 3988 // At this point, we have found a dead node which is DU-reachable. 3989 if (!dead_nodes) { 3990 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3991 dead_nodes = true; 3992 } 3993 use->dump(2); 3994 tty->print_cr("---"); 3995 checked.push(use); // No repeats; pretend it is now checked. 3996 } 3997 } 3998 assert(!dead_nodes, "using nodes must be reachable from root"); 3999 } 4000 } 4001 } 4002 #endif 4003 4004 // The Compile object keeps track of failure reasons separately from the ciEnv. 4005 // This is required because there is not quite a 1-1 relation between the 4006 // ciEnv and its compilation task and the Compile object. Note that one 4007 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4008 // to backtrack and retry without subsuming loads. Other than this backtracking 4009 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4010 // by the logic in C2Compiler. 4011 void Compile::record_failure(const char* reason) { 4012 if (log() != NULL) { 4013 log()->elem("failure reason='%s' phase='compile'", reason); 4014 } 4015 if (_failure_reason == NULL) { 4016 // Record the first failure reason. 4017 _failure_reason = reason; 4018 } 4019 4020 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4021 C->print_method(PHASE_FAILURE); 4022 } 4023 _root = NULL; // flush the graph, too 4024 } 4025 4026 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4027 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4028 _phase_name(name), _dolog(CITimeVerbose) 4029 { 4030 if (_dolog) { 4031 C = Compile::current(); 4032 _log = C->log(); 4033 } else { 4034 C = NULL; 4035 _log = NULL; 4036 } 4037 if (_log != NULL) { 4038 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4039 _log->stamp(); 4040 _log->end_head(); 4041 } 4042 } 4043 4044 Compile::TracePhase::~TracePhase() { 4045 4046 C = Compile::current(); 4047 if (_dolog) { 4048 _log = C->log(); 4049 } else { 4050 _log = NULL; 4051 } 4052 4053 #ifdef ASSERT 4054 if (PrintIdealNodeCount) { 4055 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4056 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4057 } 4058 4059 if (VerifyIdealNodeCount) { 4060 Compile::current()->print_missing_nodes(); 4061 } 4062 #endif 4063 4064 if (_log != NULL) { 4065 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4066 } 4067 } 4068 4069 //----------------------------static_subtype_check----------------------------- 4070 // Shortcut important common cases when superklass is exact: 4071 // (0) superklass is java.lang.Object (can occur in reflective code) 4072 // (1) subklass is already limited to a subtype of superklass => always ok 4073 // (2) subklass does not overlap with superklass => always fail 4074 // (3) superklass has NO subtypes and we can check with a simple compare. 4075 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4076 if (StressReflectiveCode) { 4077 return SSC_full_test; // Let caller generate the general case. 4078 } 4079 4080 if (superk == env()->Object_klass()) { 4081 return SSC_always_true; // (0) this test cannot fail 4082 } 4083 4084 ciType* superelem = superk; 4085 ciType* subelem = subk; 4086 if (superelem->is_array_klass()) { 4087 superelem = superelem->as_array_klass()->base_element_type(); 4088 } 4089 if (subelem->is_array_klass()) { 4090 subelem = subelem->as_array_klass()->base_element_type(); 4091 } 4092 4093 if (!subk->is_interface()) { // cannot trust static interface types yet 4094 if (subk->is_subtype_of(superk)) { 4095 return SSC_always_true; // (1) false path dead; no dynamic test needed 4096 } 4097 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4098 !(subelem->is_klass() && subelem->as_klass()->is_interface()) && 4099 !superk->is_subtype_of(subk)) { 4100 return SSC_always_false; // (2) true path dead; no dynamic test needed 4101 } 4102 } 4103 4104 // If casting to an instance klass, it must have no subtypes 4105 if (superk->is_interface()) { 4106 // Cannot trust interfaces yet. 4107 // %%% S.B. superk->nof_implementors() == 1 4108 } else if (superelem->is_instance_klass()) { 4109 ciInstanceKlass* ik = superelem->as_instance_klass(); 4110 if (!ik->has_subklass() && !ik->is_interface()) { 4111 if (!ik->is_final()) { 4112 // Add a dependency if there is a chance of a later subclass. 4113 dependencies()->assert_leaf_type(ik); 4114 } 4115 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4116 } 4117 } else { 4118 // A primitive array type has no subtypes. 4119 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4120 } 4121 4122 return SSC_full_test; 4123 } 4124 4125 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4126 #ifdef _LP64 4127 // The scaled index operand to AddP must be a clean 64-bit value. 4128 // Java allows a 32-bit int to be incremented to a negative 4129 // value, which appears in a 64-bit register as a large 4130 // positive number. Using that large positive number as an 4131 // operand in pointer arithmetic has bad consequences. 4132 // On the other hand, 32-bit overflow is rare, and the possibility 4133 // can often be excluded, if we annotate the ConvI2L node with 4134 // a type assertion that its value is known to be a small positive 4135 // number. (The prior range check has ensured this.) 4136 // This assertion is used by ConvI2LNode::Ideal. 4137 int index_max = max_jint - 1; // array size is max_jint, index is one less 4138 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4139 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4140 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4141 #endif 4142 return idx; 4143 } 4144 4145 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4146 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4147 if (ctrl != NULL) { 4148 // Express control dependency by a CastII node with a narrow type. 4149 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4150 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4151 // node from floating above the range check during loop optimizations. Otherwise, the 4152 // ConvI2L node may be eliminated independently of the range check, causing the data path 4153 // to become TOP while the control path is still there (although it's unreachable). 4154 value->set_req(0, ctrl); 4155 value = phase->transform(value); 4156 } 4157 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4158 return phase->transform(new ConvI2LNode(value, ltype)); 4159 } 4160 4161 void Compile::print_inlining_stream_free() { 4162 if (_print_inlining_stream != NULL) { 4163 _print_inlining_stream->~stringStream(); 4164 _print_inlining_stream = NULL; 4165 } 4166 } 4167 4168 // The message about the current inlining is accumulated in 4169 // _print_inlining_stream and transfered into the _print_inlining_list 4170 // once we know whether inlining succeeds or not. For regular 4171 // inlining, messages are appended to the buffer pointed by 4172 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4173 // a new buffer is added after _print_inlining_idx in the list. This 4174 // way we can update the inlining message for late inlining call site 4175 // when the inlining is attempted again. 4176 void Compile::print_inlining_init() { 4177 if (print_inlining() || print_intrinsics()) { 4178 // print_inlining_init is actually called several times. 4179 print_inlining_stream_free(); 4180 _print_inlining_stream = new stringStream(); 4181 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4182 } 4183 } 4184 4185 void Compile::print_inlining_reinit() { 4186 if (print_inlining() || print_intrinsics()) { 4187 print_inlining_stream_free(); 4188 // Re allocate buffer when we change ResourceMark 4189 _print_inlining_stream = new stringStream(); 4190 } 4191 } 4192 4193 void Compile::print_inlining_reset() { 4194 _print_inlining_stream->reset(); 4195 } 4196 4197 void Compile::print_inlining_commit() { 4198 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4199 // Transfer the message from _print_inlining_stream to the current 4200 // _print_inlining_list buffer and clear _print_inlining_stream. 4201 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4202 print_inlining_reset(); 4203 } 4204 4205 void Compile::print_inlining_push() { 4206 // Add new buffer to the _print_inlining_list at current position 4207 _print_inlining_idx++; 4208 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4209 } 4210 4211 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4212 return _print_inlining_list->at(_print_inlining_idx); 4213 } 4214 4215 void Compile::print_inlining_update(CallGenerator* cg) { 4216 if (print_inlining() || print_intrinsics()) { 4217 if (cg->is_late_inline()) { 4218 if (print_inlining_current()->cg() != cg && 4219 (print_inlining_current()->cg() != NULL || 4220 print_inlining_current()->ss()->size() != 0)) { 4221 print_inlining_push(); 4222 } 4223 print_inlining_commit(); 4224 print_inlining_current()->set_cg(cg); 4225 } else { 4226 if (print_inlining_current()->cg() != NULL) { 4227 print_inlining_push(); 4228 } 4229 print_inlining_commit(); 4230 } 4231 } 4232 } 4233 4234 void Compile::print_inlining_move_to(CallGenerator* cg) { 4235 // We resume inlining at a late inlining call site. Locate the 4236 // corresponding inlining buffer so that we can update it. 4237 if (print_inlining() || print_intrinsics()) { 4238 for (int i = 0; i < _print_inlining_list->length(); i++) { 4239 if (_print_inlining_list->at(i)->cg() == cg) { 4240 _print_inlining_idx = i; 4241 return; 4242 } 4243 } 4244 ShouldNotReachHere(); 4245 } 4246 } 4247 4248 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4249 if (print_inlining() || print_intrinsics()) { 4250 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4251 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4252 // replace message with new message 4253 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4254 print_inlining_commit(); 4255 print_inlining_current()->set_cg(cg); 4256 } 4257 } 4258 4259 void Compile::print_inlining_assert_ready() { 4260 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4261 } 4262 4263 void Compile::process_print_inlining() { 4264 assert(_late_inlines.length() == 0, "not drained yet"); 4265 if (print_inlining() || print_intrinsics()) { 4266 ResourceMark rm; 4267 stringStream ss; 4268 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4269 for (int i = 0; i < _print_inlining_list->length(); i++) { 4270 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4271 ss.print("%s", pib->ss()->as_string()); 4272 delete pib; 4273 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL)); 4274 } 4275 // Reset _print_inlining_list, it only contains destructed objects. 4276 // It is on the arena, so it will be freed when the arena is reset. 4277 _print_inlining_list = NULL; 4278 // _print_inlining_stream won't be used anymore, either. 4279 print_inlining_stream_free(); 4280 size_t end = ss.size(); 4281 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4282 strncpy(_print_inlining_output, ss.base(), end+1); 4283 _print_inlining_output[end] = 0; 4284 } 4285 } 4286 4287 void Compile::dump_print_inlining() { 4288 if (_print_inlining_output != NULL) { 4289 tty->print_raw(_print_inlining_output); 4290 } 4291 } 4292 4293 void Compile::log_late_inline(CallGenerator* cg) { 4294 if (log() != NULL) { 4295 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4296 cg->unique_id()); 4297 JVMState* p = cg->call_node()->jvms(); 4298 while (p != NULL) { 4299 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4300 p = p->caller(); 4301 } 4302 log()->tail("late_inline"); 4303 } 4304 } 4305 4306 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4307 log_late_inline(cg); 4308 if (log() != NULL) { 4309 log()->inline_fail(msg); 4310 } 4311 } 4312 4313 void Compile::log_inline_id(CallGenerator* cg) { 4314 if (log() != NULL) { 4315 // The LogCompilation tool needs a unique way to identify late 4316 // inline call sites. This id must be unique for this call site in 4317 // this compilation. Try to have it unique across compilations as 4318 // well because it can be convenient when grepping through the log 4319 // file. 4320 // Distinguish OSR compilations from others in case CICountOSR is 4321 // on. 4322 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4323 cg->set_unique_id(id); 4324 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4325 } 4326 } 4327 4328 void Compile::log_inline_failure(const char* msg) { 4329 if (C->log() != NULL) { 4330 C->log()->inline_fail(msg); 4331 } 4332 } 4333 4334 4335 // Dump inlining replay data to the stream. 4336 // Don't change thread state and acquire any locks. 4337 void Compile::dump_inline_data(outputStream* out) { 4338 InlineTree* inl_tree = ilt(); 4339 if (inl_tree != NULL) { 4340 out->print(" inline %d", inl_tree->count()); 4341 inl_tree->dump_replay_data(out); 4342 } 4343 } 4344 4345 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4346 if (n1->Opcode() < n2->Opcode()) return -1; 4347 else if (n1->Opcode() > n2->Opcode()) return 1; 4348 4349 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4350 for (uint i = 1; i < n1->req(); i++) { 4351 if (n1->in(i) < n2->in(i)) return -1; 4352 else if (n1->in(i) > n2->in(i)) return 1; 4353 } 4354 4355 return 0; 4356 } 4357 4358 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4359 Node* n1 = *n1p; 4360 Node* n2 = *n2p; 4361 4362 return cmp_expensive_nodes(n1, n2); 4363 } 4364 4365 void Compile::sort_expensive_nodes() { 4366 if (!expensive_nodes_sorted()) { 4367 _expensive_nodes.sort(cmp_expensive_nodes); 4368 } 4369 } 4370 4371 bool Compile::expensive_nodes_sorted() const { 4372 for (int i = 1; i < _expensive_nodes.length(); i++) { 4373 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4374 return false; 4375 } 4376 } 4377 return true; 4378 } 4379 4380 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4381 if (_expensive_nodes.length() == 0) { 4382 return false; 4383 } 4384 4385 assert(OptimizeExpensiveOps, "optimization off?"); 4386 4387 // Take this opportunity to remove dead nodes from the list 4388 int j = 0; 4389 for (int i = 0; i < _expensive_nodes.length(); i++) { 4390 Node* n = _expensive_nodes.at(i); 4391 if (!n->is_unreachable(igvn)) { 4392 assert(n->is_expensive(), "should be expensive"); 4393 _expensive_nodes.at_put(j, n); 4394 j++; 4395 } 4396 } 4397 _expensive_nodes.trunc_to(j); 4398 4399 // Then sort the list so that similar nodes are next to each other 4400 // and check for at least two nodes of identical kind with same data 4401 // inputs. 4402 sort_expensive_nodes(); 4403 4404 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4405 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4406 return true; 4407 } 4408 } 4409 4410 return false; 4411 } 4412 4413 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4414 if (_expensive_nodes.length() == 0) { 4415 return; 4416 } 4417 4418 assert(OptimizeExpensiveOps, "optimization off?"); 4419 4420 // Sort to bring similar nodes next to each other and clear the 4421 // control input of nodes for which there's only a single copy. 4422 sort_expensive_nodes(); 4423 4424 int j = 0; 4425 int identical = 0; 4426 int i = 0; 4427 bool modified = false; 4428 for (; i < _expensive_nodes.length()-1; i++) { 4429 assert(j <= i, "can't write beyond current index"); 4430 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4431 identical++; 4432 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4433 continue; 4434 } 4435 if (identical > 0) { 4436 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4437 identical = 0; 4438 } else { 4439 Node* n = _expensive_nodes.at(i); 4440 igvn.replace_input_of(n, 0, NULL); 4441 igvn.hash_insert(n); 4442 modified = true; 4443 } 4444 } 4445 if (identical > 0) { 4446 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4447 } else if (_expensive_nodes.length() >= 1) { 4448 Node* n = _expensive_nodes.at(i); 4449 igvn.replace_input_of(n, 0, NULL); 4450 igvn.hash_insert(n); 4451 modified = true; 4452 } 4453 _expensive_nodes.trunc_to(j); 4454 if (modified) { 4455 igvn.optimize(); 4456 } 4457 } 4458 4459 void Compile::add_expensive_node(Node * n) { 4460 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4461 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4462 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4463 if (OptimizeExpensiveOps) { 4464 _expensive_nodes.append(n); 4465 } else { 4466 // Clear control input and let IGVN optimize expensive nodes if 4467 // OptimizeExpensiveOps is off. 4468 n->set_req(0, NULL); 4469 } 4470 } 4471 4472 /** 4473 * Track coarsened Lock and Unlock nodes. 4474 */ 4475 4476 class Lock_List : public Node_List { 4477 uint _origin_cnt; 4478 public: 4479 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4480 uint origin_cnt() const { return _origin_cnt; } 4481 }; 4482 4483 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4484 int length = locks.length(); 4485 if (length > 0) { 4486 // Have to keep this list until locks elimination during Macro nodes elimination. 4487 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4488 for (int i = 0; i < length; i++) { 4489 AbstractLockNode* lock = locks.at(i); 4490 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4491 locks_list->push(lock); 4492 } 4493 _coarsened_locks.append(locks_list); 4494 } 4495 } 4496 4497 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4498 int count = coarsened_count(); 4499 for (int i = 0; i < count; i++) { 4500 Node_List* locks_list = _coarsened_locks.at(i); 4501 for (uint j = 0; j < locks_list->size(); j++) { 4502 Node* lock = locks_list->at(j); 4503 assert(lock->is_AbstractLock(), "sanity"); 4504 if (!useful.member(lock)) { 4505 locks_list->yank(lock); 4506 } 4507 } 4508 } 4509 } 4510 4511 void Compile::remove_coarsened_lock(Node* n) { 4512 if (n->is_AbstractLock()) { 4513 int count = coarsened_count(); 4514 for (int i = 0; i < count; i++) { 4515 Node_List* locks_list = _coarsened_locks.at(i); 4516 locks_list->yank(n); 4517 } 4518 } 4519 } 4520 4521 bool Compile::coarsened_locks_consistent() { 4522 int count = coarsened_count(); 4523 for (int i = 0; i < count; i++) { 4524 bool unbalanced = false; 4525 bool modified = false; // track locks kind modifications 4526 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4527 uint size = locks_list->size(); 4528 if (size == 0) { 4529 unbalanced = false; // All locks were eliminated - good 4530 } else if (size != locks_list->origin_cnt()) { 4531 unbalanced = true; // Some locks were removed from list 4532 } else { 4533 for (uint j = 0; j < size; j++) { 4534 Node* lock = locks_list->at(j); 4535 // All nodes in group should have the same state (modified or not) 4536 if (!lock->as_AbstractLock()->is_coarsened()) { 4537 if (j == 0) { 4538 // first on list was modified, the rest should be too for consistency 4539 modified = true; 4540 } else if (!modified) { 4541 // this lock was modified but previous locks on the list were not 4542 unbalanced = true; 4543 break; 4544 } 4545 } else if (modified) { 4546 // previous locks on list were modified but not this lock 4547 unbalanced = true; 4548 break; 4549 } 4550 } 4551 } 4552 if (unbalanced) { 4553 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4554 #ifdef ASSERT 4555 if (PrintEliminateLocks) { 4556 tty->print_cr("=== unbalanced coarsened locks ==="); 4557 for (uint l = 0; l < size; l++) { 4558 locks_list->at(l)->dump(); 4559 } 4560 } 4561 #endif 4562 record_failure(C2Compiler::retry_no_locks_coarsening()); 4563 return false; 4564 } 4565 } 4566 return true; 4567 } 4568 4569 /** 4570 * Remove the speculative part of types and clean up the graph 4571 */ 4572 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4573 if (UseTypeSpeculation) { 4574 Unique_Node_List worklist; 4575 worklist.push(root()); 4576 int modified = 0; 4577 // Go over all type nodes that carry a speculative type, drop the 4578 // speculative part of the type and enqueue the node for an igvn 4579 // which may optimize it out. 4580 for (uint next = 0; next < worklist.size(); ++next) { 4581 Node *n = worklist.at(next); 4582 if (n->is_Type()) { 4583 TypeNode* tn = n->as_Type(); 4584 const Type* t = tn->type(); 4585 const Type* t_no_spec = t->remove_speculative(); 4586 if (t_no_spec != t) { 4587 bool in_hash = igvn.hash_delete(n); 4588 assert(in_hash, "node should be in igvn hash table"); 4589 tn->set_type(t_no_spec); 4590 igvn.hash_insert(n); 4591 igvn._worklist.push(n); // give it a chance to go away 4592 modified++; 4593 } 4594 } 4595 // Iterate over outs - endless loops is unreachable from below 4596 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4597 Node *m = n->fast_out(i); 4598 if (not_a_node(m)) { 4599 continue; 4600 } 4601 worklist.push(m); 4602 } 4603 } 4604 // Drop the speculative part of all types in the igvn's type table 4605 igvn.remove_speculative_types(); 4606 if (modified > 0) { 4607 igvn.optimize(); 4608 } 4609 #ifdef ASSERT 4610 // Verify that after the IGVN is over no speculative type has resurfaced 4611 worklist.clear(); 4612 worklist.push(root()); 4613 for (uint next = 0; next < worklist.size(); ++next) { 4614 Node *n = worklist.at(next); 4615 const Type* t = igvn.type_or_null(n); 4616 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4617 if (n->is_Type()) { 4618 t = n->as_Type()->type(); 4619 assert(t == t->remove_speculative(), "no more speculative types"); 4620 } 4621 // Iterate over outs - endless loops is unreachable from below 4622 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4623 Node *m = n->fast_out(i); 4624 if (not_a_node(m)) { 4625 continue; 4626 } 4627 worklist.push(m); 4628 } 4629 } 4630 igvn.check_no_speculative_types(); 4631 #endif 4632 } 4633 } 4634 4635 // Auxiliary methods to support randomized stressing/fuzzing. 4636 4637 int Compile::random() { 4638 _stress_seed = os::next_random(_stress_seed); 4639 return static_cast<int>(_stress_seed); 4640 } 4641 4642 // This method can be called the arbitrary number of times, with current count 4643 // as the argument. The logic allows selecting a single candidate from the 4644 // running list of candidates as follows: 4645 // int count = 0; 4646 // Cand* selected = null; 4647 // while(cand = cand->next()) { 4648 // if (randomized_select(++count)) { 4649 // selected = cand; 4650 // } 4651 // } 4652 // 4653 // Including count equalizes the chances any candidate is "selected". 4654 // This is useful when we don't have the complete list of candidates to choose 4655 // from uniformly. In this case, we need to adjust the randomicity of the 4656 // selection, or else we will end up biasing the selection towards the latter 4657 // candidates. 4658 // 4659 // Quick back-envelope calculation shows that for the list of n candidates 4660 // the equal probability for the candidate to persist as "best" can be 4661 // achieved by replacing it with "next" k-th candidate with the probability 4662 // of 1/k. It can be easily shown that by the end of the run, the 4663 // probability for any candidate is converged to 1/n, thus giving the 4664 // uniform distribution among all the candidates. 4665 // 4666 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4667 #define RANDOMIZED_DOMAIN_POW 29 4668 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4669 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4670 bool Compile::randomized_select(int count) { 4671 assert(count > 0, "only positive"); 4672 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4673 } 4674 4675 CloneMap& Compile::clone_map() { return _clone_map; } 4676 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4677 4678 void NodeCloneInfo::dump() const { 4679 tty->print(" {%d:%d} ", idx(), gen()); 4680 } 4681 4682 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4683 uint64_t val = value(old->_idx); 4684 NodeCloneInfo cio(val); 4685 assert(val != 0, "old node should be in the map"); 4686 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4687 insert(nnn->_idx, cin.get()); 4688 #ifndef PRODUCT 4689 if (is_debug()) { 4690 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4691 } 4692 #endif 4693 } 4694 4695 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4696 NodeCloneInfo cio(value(old->_idx)); 4697 if (cio.get() == 0) { 4698 cio.set(old->_idx, 0); 4699 insert(old->_idx, cio.get()); 4700 #ifndef PRODUCT 4701 if (is_debug()) { 4702 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4703 } 4704 #endif 4705 } 4706 clone(old, nnn, gen); 4707 } 4708 4709 int CloneMap::max_gen() const { 4710 int g = 0; 4711 DictI di(_dict); 4712 for(; di.test(); ++di) { 4713 int t = gen(di._key); 4714 if (g < t) { 4715 g = t; 4716 #ifndef PRODUCT 4717 if (is_debug()) { 4718 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4719 } 4720 #endif 4721 } 4722 } 4723 return g; 4724 } 4725 4726 void CloneMap::dump(node_idx_t key) const { 4727 uint64_t val = value(key); 4728 if (val != 0) { 4729 NodeCloneInfo ni(val); 4730 ni.dump(); 4731 } 4732 } 4733 4734 // Move Allocate nodes to the start of the list 4735 void Compile::sort_macro_nodes() { 4736 int count = macro_count(); 4737 int allocates = 0; 4738 for (int i = 0; i < count; i++) { 4739 Node* n = macro_node(i); 4740 if (n->is_Allocate()) { 4741 if (i != allocates) { 4742 Node* tmp = macro_node(allocates); 4743 _macro_nodes.at_put(allocates, n); 4744 _macro_nodes.at_put(i, tmp); 4745 } 4746 allocates++; 4747 } 4748 } 4749 } 4750 4751 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level) { 4752 EventCompilerPhase event; 4753 if (event.should_commit()) { 4754 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 4755 } 4756 #ifndef PRODUCT 4757 if (should_print(level)) { 4758 _printer->print_method(name, level); 4759 } 4760 #endif 4761 C->_latest_stage_start_counter.stamp(); 4762 } 4763 4764 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) { 4765 char output[1024]; 4766 #ifndef PRODUCT 4767 if (idx != 0) { 4768 jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx); 4769 } else { 4770 jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt)); 4771 } 4772 #endif 4773 print_method(cpt, output, level); 4774 } 4775 4776 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) { 4777 ResourceMark rm; 4778 stringStream ss; 4779 ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt)); 4780 if (n != NULL) { 4781 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 4782 } else { 4783 ss.print_raw(": NULL"); 4784 } 4785 C->print_method(cpt, ss.as_string(), level); 4786 } 4787 4788 void Compile::end_method(int level) { 4789 EventCompilerPhase event; 4790 if (event.should_commit()) { 4791 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level); 4792 } 4793 4794 #ifndef PRODUCT 4795 if (_method != NULL && should_print(level)) { 4796 _printer->end_method(); 4797 } 4798 #endif 4799 } 4800 4801 4802 #ifndef PRODUCT 4803 IdealGraphPrinter* Compile::_debug_file_printer = NULL; 4804 IdealGraphPrinter* Compile::_debug_network_printer = NULL; 4805 4806 // Called from debugger. Prints method to the default file with the default phase name. 4807 // This works regardless of any Ideal Graph Visualizer flags set or not. 4808 void igv_print() { 4809 Compile::current()->igv_print_method_to_file(); 4810 } 4811 4812 // Same as igv_print() above but with a specified phase name. 4813 void igv_print(const char* phase_name) { 4814 Compile::current()->igv_print_method_to_file(phase_name); 4815 } 4816 4817 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 4818 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 4819 // This works regardless of any Ideal Graph Visualizer flags set or not. 4820 void igv_print(bool network) { 4821 if (network) { 4822 Compile::current()->igv_print_method_to_network(); 4823 } else { 4824 Compile::current()->igv_print_method_to_file(); 4825 } 4826 } 4827 4828 // Same as igv_print(bool network) above but with a specified phase name. 4829 void igv_print(bool network, const char* phase_name) { 4830 if (network) { 4831 Compile::current()->igv_print_method_to_network(phase_name); 4832 } else { 4833 Compile::current()->igv_print_method_to_file(phase_name); 4834 } 4835 } 4836 4837 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 4838 void igv_print_default() { 4839 Compile::current()->print_method(PHASE_DEBUG, 0); 4840 } 4841 4842 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 4843 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 4844 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 4845 void igv_append() { 4846 Compile::current()->igv_print_method_to_file("Debug", true); 4847 } 4848 4849 // Same as igv_append() above but with a specified phase name. 4850 void igv_append(const char* phase_name) { 4851 Compile::current()->igv_print_method_to_file(phase_name, true); 4852 } 4853 4854 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 4855 const char* file_name = "custom_debug.xml"; 4856 if (_debug_file_printer == NULL) { 4857 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 4858 } else { 4859 _debug_file_printer->update_compiled_method(C->method()); 4860 } 4861 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 4862 _debug_file_printer->print(phase_name, (Node*)C->root()); 4863 } 4864 4865 void Compile::igv_print_method_to_network(const char* phase_name) { 4866 if (_debug_network_printer == NULL) { 4867 _debug_network_printer = new IdealGraphPrinter(C); 4868 } else { 4869 _debug_network_printer->update_compiled_method(C->method()); 4870 } 4871 tty->print_cr("Method printed over network stream to IGV"); 4872 _debug_network_printer->print(phase_name, (Node*)C->root()); 4873 } 4874 #endif 4875 4876 void Compile::add_native_invoker(RuntimeStub* stub) { 4877 _native_invokers.append(stub); 4878 } 4879 4880 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 4881 if (type != NULL && phase->type(value)->higher_equal(type)) { 4882 return value; 4883 } 4884 Node* result = NULL; 4885 if (bt == T_BYTE) { 4886 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 4887 result = new RShiftINode(result, phase->intcon(24)); 4888 } else if (bt == T_BOOLEAN) { 4889 result = new AndINode(value, phase->intcon(0xFF)); 4890 } else if (bt == T_CHAR) { 4891 result = new AndINode(value,phase->intcon(0xFFFF)); 4892 } else { 4893 assert(bt == T_SHORT, "unexpected narrow type"); 4894 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 4895 result = new RShiftINode(result, phase->intcon(16)); 4896 } 4897 if (transform_res) { 4898 result = phase->transform(result); 4899 } 4900 return result; 4901 } 4902