1 /* 2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compilationFailureInfo.hpp" 33 #include "compiler/compilationMemoryStatistic.hpp" 34 #include "compiler/compileBroker.hpp" 35 #include "compiler/compileLog.hpp" 36 #include "compiler/compiler_globals.hpp" 37 #include "compiler/disassembler.hpp" 38 #include "compiler/oopMap.hpp" 39 #include "gc/shared/barrierSet.hpp" 40 #include "gc/shared/c2/barrierSetC2.hpp" 41 #include "jfr/jfrEvents.hpp" 42 #include "jvm_io.h" 43 #include "memory/allocation.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "opto/addnode.hpp" 46 #include "opto/block.hpp" 47 #include "opto/c2compiler.hpp" 48 #include "opto/callGenerator.hpp" 49 #include "opto/callnode.hpp" 50 #include "opto/castnode.hpp" 51 #include "opto/cfgnode.hpp" 52 #include "opto/chaitin.hpp" 53 #include "opto/compile.hpp" 54 #include "opto/connode.hpp" 55 #include "opto/convertnode.hpp" 56 #include "opto/divnode.hpp" 57 #include "opto/escape.hpp" 58 #include "opto/idealGraphPrinter.hpp" 59 #include "opto/locknode.hpp" 60 #include "opto/loopnode.hpp" 61 #include "opto/machnode.hpp" 62 #include "opto/macro.hpp" 63 #include "opto/matcher.hpp" 64 #include "opto/mathexactnode.hpp" 65 #include "opto/memnode.hpp" 66 #include "opto/mulnode.hpp" 67 #include "opto/narrowptrnode.hpp" 68 #include "opto/node.hpp" 69 #include "opto/opcodes.hpp" 70 #include "opto/output.hpp" 71 #include "opto/parse.hpp" 72 #include "opto/phaseX.hpp" 73 #include "opto/rootnode.hpp" 74 #include "opto/runtime.hpp" 75 #include "opto/stringopts.hpp" 76 #include "opto/type.hpp" 77 #include "opto/vector.hpp" 78 #include "opto/vectornode.hpp" 79 #include "runtime/globals_extension.hpp" 80 #include "runtime/sharedRuntime.hpp" 81 #include "runtime/signature.hpp" 82 #include "runtime/stubRoutines.hpp" 83 #include "runtime/timer.hpp" 84 #include "utilities/align.hpp" 85 #include "utilities/copy.hpp" 86 #include "utilities/macros.hpp" 87 #include "utilities/resourceHash.hpp" 88 89 // -------------------- Compile::mach_constant_base_node ----------------------- 90 // Constant table base node singleton. 91 MachConstantBaseNode* Compile::mach_constant_base_node() { 92 if (_mach_constant_base_node == nullptr) { 93 _mach_constant_base_node = new MachConstantBaseNode(); 94 _mach_constant_base_node->add_req(C->root()); 95 } 96 return _mach_constant_base_node; 97 } 98 99 100 /// Support for intrinsics. 101 102 // Return the index at which m must be inserted (or already exists). 103 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 104 class IntrinsicDescPair { 105 private: 106 ciMethod* _m; 107 bool _is_virtual; 108 public: 109 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 110 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 111 ciMethod* m= elt->method(); 112 ciMethod* key_m = key->_m; 113 if (key_m < m) return -1; 114 else if (key_m > m) return 1; 115 else { 116 bool is_virtual = elt->is_virtual(); 117 bool key_virtual = key->_is_virtual; 118 if (key_virtual < is_virtual) return -1; 119 else if (key_virtual > is_virtual) return 1; 120 else return 0; 121 } 122 } 123 }; 124 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 125 #ifdef ASSERT 126 for (int i = 1; i < _intrinsics.length(); i++) { 127 CallGenerator* cg1 = _intrinsics.at(i-1); 128 CallGenerator* cg2 = _intrinsics.at(i); 129 assert(cg1->method() != cg2->method() 130 ? cg1->method() < cg2->method() 131 : cg1->is_virtual() < cg2->is_virtual(), 132 "compiler intrinsics list must stay sorted"); 133 } 134 #endif 135 IntrinsicDescPair pair(m, is_virtual); 136 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 137 } 138 139 void Compile::register_intrinsic(CallGenerator* cg) { 140 bool found = false; 141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 142 assert(!found, "registering twice"); 143 _intrinsics.insert_before(index, cg); 144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 145 } 146 147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 148 assert(m->is_loaded(), "don't try this on unloaded methods"); 149 if (_intrinsics.length() > 0) { 150 bool found = false; 151 int index = intrinsic_insertion_index(m, is_virtual, found); 152 if (found) { 153 return _intrinsics.at(index); 154 } 155 } 156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 157 if (m->intrinsic_id() != vmIntrinsics::_none && 158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 160 if (cg != nullptr) { 161 // Save it for next time: 162 register_intrinsic(cg); 163 return cg; 164 } else { 165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 166 } 167 } 168 return nullptr; 169 } 170 171 // Compile::make_vm_intrinsic is defined in library_call.cpp. 172 173 #ifndef PRODUCT 174 // statistics gathering... 175 176 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 177 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 178 179 inline int as_int(vmIntrinsics::ID id) { 180 return vmIntrinsics::as_int(id); 181 } 182 183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 185 int oflags = _intrinsic_hist_flags[as_int(id)]; 186 assert(flags != 0, "what happened?"); 187 if (is_virtual) { 188 flags |= _intrinsic_virtual; 189 } 190 bool changed = (flags != oflags); 191 if ((flags & _intrinsic_worked) != 0) { 192 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 193 if (count == 1) { 194 changed = true; // first time 195 } 196 // increment the overall count also: 197 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 198 } 199 if (changed) { 200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 201 // Something changed about the intrinsic's virtuality. 202 if ((flags & _intrinsic_virtual) != 0) { 203 // This is the first use of this intrinsic as a virtual call. 204 if (oflags != 0) { 205 // We already saw it as a non-virtual, so note both cases. 206 flags |= _intrinsic_both; 207 } 208 } else if ((oflags & _intrinsic_both) == 0) { 209 // This is the first use of this intrinsic as a non-virtual 210 flags |= _intrinsic_both; 211 } 212 } 213 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 214 } 215 // update the overall flags also: 216 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 217 return changed; 218 } 219 220 static char* format_flags(int flags, char* buf) { 221 buf[0] = 0; 222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 227 if (buf[0] == 0) strcat(buf, ","); 228 assert(buf[0] == ',', "must be"); 229 return &buf[1]; 230 } 231 232 void Compile::print_intrinsic_statistics() { 233 char flagsbuf[100]; 234 ttyLocker ttyl; 235 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 236 tty->print_cr("Compiler intrinsic usage:"); 237 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 238 if (total == 0) total = 1; // avoid div0 in case of no successes 239 #define PRINT_STAT_LINE(name, c, f) \ 240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 241 for (auto id : EnumRange<vmIntrinsicID>{}) { 242 int flags = _intrinsic_hist_flags[as_int(id)]; 243 juint count = _intrinsic_hist_count[as_int(id)]; 244 if ((flags | count) != 0) { 245 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 246 } 247 } 248 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 249 if (xtty != nullptr) xtty->tail("statistics"); 250 } 251 252 void Compile::print_statistics() { 253 { ttyLocker ttyl; 254 if (xtty != nullptr) xtty->head("statistics type='opto'"); 255 Parse::print_statistics(); 256 PhaseStringOpts::print_statistics(); 257 PhaseCCP::print_statistics(); 258 PhaseRegAlloc::print_statistics(); 259 PhaseOutput::print_statistics(); 260 PhasePeephole::print_statistics(); 261 PhaseIdealLoop::print_statistics(); 262 ConnectionGraph::print_statistics(); 263 PhaseMacroExpand::print_statistics(); 264 if (xtty != nullptr) xtty->tail("statistics"); 265 } 266 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 267 // put this under its own <statistics> element. 268 print_intrinsic_statistics(); 269 } 270 } 271 #endif //PRODUCT 272 273 void Compile::gvn_replace_by(Node* n, Node* nn) { 274 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 275 Node* use = n->last_out(i); 276 bool is_in_table = initial_gvn()->hash_delete(use); 277 uint uses_found = 0; 278 for (uint j = 0; j < use->len(); j++) { 279 if (use->in(j) == n) { 280 if (j < use->req()) 281 use->set_req(j, nn); 282 else 283 use->set_prec(j, nn); 284 uses_found++; 285 } 286 } 287 if (is_in_table) { 288 // reinsert into table 289 initial_gvn()->hash_find_insert(use); 290 } 291 record_for_igvn(use); 292 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist); 293 i -= uses_found; // we deleted 1 or more copies of this edge 294 } 295 } 296 297 298 // Identify all nodes that are reachable from below, useful. 299 // Use breadth-first pass that records state in a Unique_Node_List, 300 // recursive traversal is slower. 301 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 302 int estimated_worklist_size = live_nodes(); 303 useful.map( estimated_worklist_size, nullptr ); // preallocate space 304 305 // Initialize worklist 306 if (root() != nullptr) { useful.push(root()); } 307 // If 'top' is cached, declare it useful to preserve cached node 308 if (cached_top_node()) { useful.push(cached_top_node()); } 309 310 // Push all useful nodes onto the list, breadthfirst 311 for( uint next = 0; next < useful.size(); ++next ) { 312 assert( next < unique(), "Unique useful nodes < total nodes"); 313 Node *n = useful.at(next); 314 uint max = n->len(); 315 for( uint i = 0; i < max; ++i ) { 316 Node *m = n->in(i); 317 if (not_a_node(m)) continue; 318 useful.push(m); 319 } 320 } 321 } 322 323 // Update dead_node_list with any missing dead nodes using useful 324 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 325 void Compile::update_dead_node_list(Unique_Node_List &useful) { 326 uint max_idx = unique(); 327 VectorSet& useful_node_set = useful.member_set(); 328 329 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 330 // If node with index node_idx is not in useful set, 331 // mark it as dead in dead node list. 332 if (!useful_node_set.test(node_idx)) { 333 record_dead_node(node_idx); 334 } 335 } 336 } 337 338 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 339 int shift = 0; 340 for (int i = 0; i < inlines->length(); i++) { 341 CallGenerator* cg = inlines->at(i); 342 if (useful.member(cg->call_node())) { 343 if (shift > 0) { 344 inlines->at_put(i - shift, cg); 345 } 346 } else { 347 shift++; // skip over the dead element 348 } 349 } 350 if (shift > 0) { 351 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 352 } 353 } 354 355 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 356 assert(dead != nullptr && dead->is_Call(), "sanity"); 357 int found = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 if (inlines->at(i)->call_node() == dead) { 360 inlines->remove_at(i); 361 found++; 362 NOT_DEBUG( break; ) // elements are unique, so exit early 363 } 364 } 365 assert(found <= 1, "not unique"); 366 } 367 368 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)> 369 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) { 370 for (int i = node_list.length() - 1; i >= 0; i--) { 371 N* node = node_list.at(i); 372 if (!useful.member(node)) { 373 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 374 } 375 } 376 } 377 378 void Compile::remove_useless_node(Node* dead) { 379 remove_modified_node(dead); 380 381 // Constant node that has no out-edges and has only one in-edge from 382 // root is usually dead. However, sometimes reshaping walk makes 383 // it reachable by adding use edges. So, we will NOT count Con nodes 384 // as dead to be conservative about the dead node count at any 385 // given time. 386 if (!dead->is_Con()) { 387 record_dead_node(dead->_idx); 388 } 389 if (dead->is_macro()) { 390 remove_macro_node(dead); 391 } 392 if (dead->is_expensive()) { 393 remove_expensive_node(dead); 394 } 395 if (dead->Opcode() == Op_Opaque4) { 396 remove_template_assertion_predicate_opaq(dead); 397 } 398 if (dead->is_ParsePredicate()) { 399 remove_parse_predicate(dead->as_ParsePredicate()); 400 } 401 if (dead->for_post_loop_opts_igvn()) { 402 remove_from_post_loop_opts_igvn(dead); 403 } 404 if (dead->is_Call()) { 405 remove_useless_late_inlines( &_late_inlines, dead); 406 remove_useless_late_inlines( &_string_late_inlines, dead); 407 remove_useless_late_inlines( &_boxing_late_inlines, dead); 408 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 409 410 if (dead->is_CallStaticJava()) { 411 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 412 } 413 } 414 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 415 bs->unregister_potential_barrier_node(dead); 416 } 417 418 // Disconnect all useless nodes by disconnecting those at the boundary. 419 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) { 420 uint next = 0; 421 while (next < useful.size()) { 422 Node *n = useful.at(next++); 423 if (n->is_SafePoint()) { 424 // We're done with a parsing phase. Replaced nodes are not valid 425 // beyond that point. 426 n->as_SafePoint()->delete_replaced_nodes(); 427 } 428 // Use raw traversal of out edges since this code removes out edges 429 int max = n->outcnt(); 430 for (int j = 0; j < max; ++j) { 431 Node* child = n->raw_out(j); 432 if (!useful.member(child)) { 433 assert(!child->is_top() || child != top(), 434 "If top is cached in Compile object it is in useful list"); 435 // Only need to remove this out-edge to the useless node 436 n->raw_del_out(j); 437 --j; 438 --max; 439 } 440 } 441 if (n->outcnt() == 1 && n->has_special_unique_user()) { 442 assert(useful.member(n->unique_out()), "do not push a useless node"); 443 worklist.push(n->unique_out()); 444 } 445 } 446 447 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 448 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes 449 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes 450 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 451 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 452 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 453 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 454 #ifdef ASSERT 455 if (_modified_nodes != nullptr) { 456 _modified_nodes->remove_useless_nodes(useful.member_set()); 457 } 458 #endif 459 460 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 461 bs->eliminate_useless_gc_barriers(useful, this); 462 // clean up the late inline lists 463 remove_useless_late_inlines( &_late_inlines, useful); 464 remove_useless_late_inlines( &_string_late_inlines, useful); 465 remove_useless_late_inlines( &_boxing_late_inlines, useful); 466 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 467 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 468 } 469 470 // ============================================================================ 471 //------------------------------CompileWrapper--------------------------------- 472 class CompileWrapper : public StackObj { 473 Compile *const _compile; 474 public: 475 CompileWrapper(Compile* compile); 476 477 ~CompileWrapper(); 478 }; 479 480 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 481 // the Compile* pointer is stored in the current ciEnv: 482 ciEnv* env = compile->env(); 483 assert(env == ciEnv::current(), "must already be a ciEnv active"); 484 assert(env->compiler_data() == nullptr, "compile already active?"); 485 env->set_compiler_data(compile); 486 assert(compile == Compile::current(), "sanity"); 487 488 compile->set_type_dict(nullptr); 489 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 490 compile->clone_map().set_clone_idx(0); 491 compile->set_type_last_size(0); 492 compile->set_last_tf(nullptr, nullptr); 493 compile->set_indexSet_arena(nullptr); 494 compile->set_indexSet_free_block_list(nullptr); 495 compile->init_type_arena(); 496 Type::Initialize(compile); 497 _compile->begin_method(); 498 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 499 } 500 CompileWrapper::~CompileWrapper() { 501 // simulate crash during compilation 502 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 503 504 _compile->end_method(); 505 _compile->env()->set_compiler_data(nullptr); 506 } 507 508 509 //----------------------------print_compile_messages--------------------------- 510 void Compile::print_compile_messages() { 511 #ifndef PRODUCT 512 // Check if recompiling 513 if (!subsume_loads() && PrintOpto) { 514 // Recompiling without allowing machine instructions to subsume loads 515 tty->print_cr("*********************************************************"); 516 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 517 tty->print_cr("*********************************************************"); 518 } 519 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 520 // Recompiling without escape analysis 521 tty->print_cr("*********************************************************"); 522 tty->print_cr("** Bailout: Recompile without escape analysis **"); 523 tty->print_cr("*********************************************************"); 524 } 525 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 526 // Recompiling without iterative escape analysis 527 tty->print_cr("*********************************************************"); 528 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 529 tty->print_cr("*********************************************************"); 530 } 531 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) { 532 // Recompiling without reducing allocation merges 533 tty->print_cr("*********************************************************"); 534 tty->print_cr("** Bailout: Recompile without reduce allocation merges **"); 535 tty->print_cr("*********************************************************"); 536 } 537 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 538 // Recompiling without boxing elimination 539 tty->print_cr("*********************************************************"); 540 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 541 tty->print_cr("*********************************************************"); 542 } 543 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 544 // Recompiling without locks coarsening 545 tty->print_cr("*********************************************************"); 546 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 547 tty->print_cr("*********************************************************"); 548 } 549 if (env()->break_at_compile()) { 550 // Open the debugger when compiling this method. 551 tty->print("### Breaking when compiling: "); 552 method()->print_short_name(); 553 tty->cr(); 554 BREAKPOINT; 555 } 556 557 if( PrintOpto ) { 558 if (is_osr_compilation()) { 559 tty->print("[OSR]%3d", _compile_id); 560 } else { 561 tty->print("%3d", _compile_id); 562 } 563 } 564 #endif 565 } 566 567 #ifndef PRODUCT 568 void Compile::print_ideal_ir(const char* phase_name) { 569 // keep the following output all in one block 570 // This output goes directly to the tty, not the compiler log. 571 // To enable tools to match it up with the compilation activity, 572 // be sure to tag this tty output with the compile ID. 573 574 // Node dumping can cause a safepoint, which can break the tty lock. 575 // Buffer all node dumps, so that all safepoints happen before we lock. 576 ResourceMark rm; 577 stringStream ss; 578 579 if (_output == nullptr) { 580 ss.print_cr("AFTER: %s", phase_name); 581 // Print out all nodes in ascending order of index. 582 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 583 } else { 584 // Dump the node blockwise if we have a scheduling 585 _output->print_scheduling(&ss); 586 } 587 588 // Check that the lock is not broken by a safepoint. 589 NoSafepointVerifier nsv; 590 ttyLocker ttyl; 591 if (xtty != nullptr) { 592 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 593 compile_id(), 594 is_osr_compilation() ? " compile_kind='osr'" : "", 595 phase_name); 596 } 597 598 tty->print("%s", ss.as_string()); 599 600 if (xtty != nullptr) { 601 xtty->tail("ideal"); 602 } 603 } 604 #endif 605 606 // ============================================================================ 607 //------------------------------Compile standard------------------------------- 608 609 // Compile a method. entry_bci is -1 for normal compilations and indicates 610 // the continuation bci for on stack replacement. 611 612 613 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 614 Options options, DirectiveSet* directive) 615 : Phase(Compiler), 616 _compile_id(ci_env->compile_id()), 617 _options(options), 618 _method(target), 619 _entry_bci(osr_bci), 620 _ilt(nullptr), 621 _stub_function(nullptr), 622 _stub_name(nullptr), 623 _stub_entry_point(nullptr), 624 _max_node_limit(MaxNodeLimit), 625 _post_loop_opts_phase(false), 626 _inlining_progress(false), 627 _inlining_incrementally(false), 628 _do_cleanup(false), 629 _has_reserved_stack_access(target->has_reserved_stack_access()), 630 #ifndef PRODUCT 631 _igv_idx(0), 632 _trace_opto_output(directive->TraceOptoOutputOption), 633 #endif 634 _has_method_handle_invokes(false), 635 _clinit_barrier_on_entry(false), 636 _stress_seed(0), 637 _comp_arena(mtCompiler), 638 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 639 _env(ci_env), 640 _directive(directive), 641 _log(ci_env->log()), 642 _first_failure_details(nullptr), 643 _intrinsics (comp_arena(), 0, 0, nullptr), 644 _macro_nodes (comp_arena(), 8, 0, nullptr), 645 _parse_predicates (comp_arena(), 8, 0, nullptr), 646 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr), 647 _expensive_nodes (comp_arena(), 8, 0, nullptr), 648 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 649 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 650 _coarsened_locks (comp_arena(), 8, 0, nullptr), 651 _congraph(nullptr), 652 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 653 _unique(0), 654 _dead_node_count(0), 655 _dead_node_list(comp_arena()), 656 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 657 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 658 _node_arena(&_node_arena_one), 659 _mach_constant_base_node(nullptr), 660 _Compile_types(mtCompiler), 661 _initial_gvn(nullptr), 662 _igvn_worklist(nullptr), 663 _types(nullptr), 664 _node_hash(nullptr), 665 _late_inlines(comp_arena(), 2, 0, nullptr), 666 _string_late_inlines(comp_arena(), 2, 0, nullptr), 667 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 668 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 669 _late_inlines_pos(0), 670 _number_of_mh_late_inlines(0), 671 _oom(false), 672 _print_inlining_stream(new (mtCompiler) stringStream()), 673 _print_inlining_list(nullptr), 674 _print_inlining_idx(0), 675 _print_inlining_output(nullptr), 676 _replay_inline_data(nullptr), 677 _java_calls(0), 678 _inner_loops(0), 679 _interpreter_frame_size(0), 680 _output(nullptr) 681 #ifndef PRODUCT 682 , _in_dump_cnt(0) 683 #endif 684 { 685 C = this; 686 CompileWrapper cw(this); 687 688 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 689 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 690 691 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 692 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 693 // We can always print a disassembly, either abstract (hex dump) or 694 // with the help of a suitable hsdis library. Thus, we should not 695 // couple print_assembly and print_opto_assembly controls. 696 // But: always print opto and regular assembly on compile command 'print'. 697 bool print_assembly = directive->PrintAssemblyOption; 698 set_print_assembly(print_opto_assembly || print_assembly); 699 #else 700 set_print_assembly(false); // must initialize. 701 #endif 702 703 #ifndef PRODUCT 704 set_parsed_irreducible_loop(false); 705 #endif 706 707 if (directive->ReplayInlineOption) { 708 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 709 } 710 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 711 set_print_intrinsics(directive->PrintIntrinsicsOption); 712 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 713 714 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 715 // Make sure the method being compiled gets its own MDO, 716 // so we can at least track the decompile_count(). 717 // Need MDO to record RTM code generation state. 718 method()->ensure_method_data(); 719 } 720 721 Init(/*do_aliasing=*/ true); 722 723 print_compile_messages(); 724 725 _ilt = InlineTree::build_inline_tree_root(); 726 727 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 728 assert(num_alias_types() >= AliasIdxRaw, ""); 729 730 #define MINIMUM_NODE_HASH 1023 731 732 // GVN that will be run immediately on new nodes 733 uint estimated_size = method()->code_size()*4+64; 734 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 735 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 736 _types = new (comp_arena()) Type_Array(comp_arena()); 737 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 738 PhaseGVN gvn; 739 set_initial_gvn(&gvn); 740 741 print_inlining_init(); 742 { // Scope for timing the parser 743 TracePhase tp("parse", &timers[_t_parser]); 744 745 // Put top into the hash table ASAP. 746 initial_gvn()->transform(top()); 747 748 // Set up tf(), start(), and find a CallGenerator. 749 CallGenerator* cg = nullptr; 750 if (is_osr_compilation()) { 751 const TypeTuple *domain = StartOSRNode::osr_domain(); 752 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 753 init_tf(TypeFunc::make(domain, range)); 754 StartNode* s = new StartOSRNode(root(), domain); 755 initial_gvn()->set_type_bottom(s); 756 init_start(s); 757 cg = CallGenerator::for_osr(method(), entry_bci()); 758 } else { 759 // Normal case. 760 init_tf(TypeFunc::make(method())); 761 StartNode* s = new StartNode(root(), tf()->domain()); 762 initial_gvn()->set_type_bottom(s); 763 init_start(s); 764 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 765 // With java.lang.ref.reference.get() we must go through the 766 // intrinsic - even when get() is the root 767 // method of the compile - so that, if necessary, the value in 768 // the referent field of the reference object gets recorded by 769 // the pre-barrier code. 770 cg = find_intrinsic(method(), false); 771 } 772 if (cg == nullptr) { 773 float past_uses = method()->interpreter_invocation_count(); 774 float expected_uses = past_uses; 775 cg = CallGenerator::for_inline(method(), expected_uses); 776 } 777 } 778 if (failing()) return; 779 if (cg == nullptr) { 780 const char* reason = InlineTree::check_can_parse(method()); 781 assert(reason != nullptr, "expect reason for parse failure"); 782 stringStream ss; 783 ss.print("cannot parse method: %s", reason); 784 record_method_not_compilable(ss.as_string()); 785 return; 786 } 787 788 gvn.set_type(root(), root()->bottom_type()); 789 790 JVMState* jvms = build_start_state(start(), tf()); 791 if ((jvms = cg->generate(jvms)) == nullptr) { 792 assert(failure_reason() != nullptr, "expect reason for parse failure"); 793 stringStream ss; 794 ss.print("method parse failed: %s", failure_reason()); 795 record_method_not_compilable(ss.as_string()); 796 return; 797 } 798 GraphKit kit(jvms); 799 800 if (!kit.stopped()) { 801 // Accept return values, and transfer control we know not where. 802 // This is done by a special, unique ReturnNode bound to root. 803 return_values(kit.jvms()); 804 } 805 806 if (kit.has_exceptions()) { 807 // Any exceptions that escape from this call must be rethrown 808 // to whatever caller is dynamically above us on the stack. 809 // This is done by a special, unique RethrowNode bound to root. 810 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 811 } 812 813 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 814 815 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 816 inline_string_calls(true); 817 } 818 819 if (failing()) return; 820 821 // Remove clutter produced by parsing. 822 if (!failing()) { 823 ResourceMark rm; 824 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 825 } 826 } 827 828 // Note: Large methods are capped off in do_one_bytecode(). 829 if (failing()) return; 830 831 // After parsing, node notes are no longer automagic. 832 // They must be propagated by register_new_node_with_optimizer(), 833 // clone(), or the like. 834 set_default_node_notes(nullptr); 835 836 #ifndef PRODUCT 837 if (should_print_igv(1)) { 838 _igv_printer->print_inlining(); 839 } 840 #endif 841 842 if (failing()) return; 843 NOT_PRODUCT( verify_graph_edges(); ) 844 845 // If any phase is randomized for stress testing, seed random number 846 // generation and log the seed for repeatability. 847 if (StressLCM || StressGCM || StressIGVN || StressCCP || 848 StressIncrementalInlining || StressMacroExpansion) { 849 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 850 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 851 FLAG_SET_ERGO(StressSeed, _stress_seed); 852 } else { 853 _stress_seed = StressSeed; 854 } 855 if (_log != nullptr) { 856 _log->elem("stress_test seed='%u'", _stress_seed); 857 } 858 } 859 860 // Now optimize 861 Optimize(); 862 if (failing()) return; 863 NOT_PRODUCT( verify_graph_edges(); ) 864 865 #ifndef PRODUCT 866 if (should_print_ideal()) { 867 print_ideal_ir("print_ideal"); 868 } 869 #endif 870 871 #ifdef ASSERT 872 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 873 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 874 #endif 875 876 // Dump compilation data to replay it. 877 if (directive->DumpReplayOption) { 878 env()->dump_replay_data(_compile_id); 879 } 880 if (directive->DumpInlineOption && (ilt() != nullptr)) { 881 env()->dump_inline_data(_compile_id); 882 } 883 884 // Now that we know the size of all the monitors we can add a fixed slot 885 // for the original deopt pc. 886 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 887 set_fixed_slots(next_slot); 888 889 // Compute when to use implicit null checks. Used by matching trap based 890 // nodes and NullCheck optimization. 891 set_allowed_deopt_reasons(); 892 893 // Now generate code 894 Code_Gen(); 895 } 896 897 //------------------------------Compile---------------------------------------- 898 // Compile a runtime stub 899 Compile::Compile( ciEnv* ci_env, 900 TypeFunc_generator generator, 901 address stub_function, 902 const char *stub_name, 903 int is_fancy_jump, 904 bool pass_tls, 905 bool return_pc, 906 DirectiveSet* directive) 907 : Phase(Compiler), 908 _compile_id(0), 909 _options(Options::for_runtime_stub()), 910 _method(nullptr), 911 _entry_bci(InvocationEntryBci), 912 _stub_function(stub_function), 913 _stub_name(stub_name), 914 _stub_entry_point(nullptr), 915 _max_node_limit(MaxNodeLimit), 916 _post_loop_opts_phase(false), 917 _inlining_progress(false), 918 _inlining_incrementally(false), 919 _has_reserved_stack_access(false), 920 #ifndef PRODUCT 921 _igv_idx(0), 922 _trace_opto_output(directive->TraceOptoOutputOption), 923 #endif 924 _has_method_handle_invokes(false), 925 _clinit_barrier_on_entry(false), 926 _stress_seed(0), 927 _comp_arena(mtCompiler), 928 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 929 _env(ci_env), 930 _directive(directive), 931 _log(ci_env->log()), 932 _first_failure_details(nullptr), 933 _congraph(nullptr), 934 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 935 _unique(0), 936 _dead_node_count(0), 937 _dead_node_list(comp_arena()), 938 _node_arena_one(mtCompiler), 939 _node_arena_two(mtCompiler), 940 _node_arena(&_node_arena_one), 941 _mach_constant_base_node(nullptr), 942 _Compile_types(mtCompiler), 943 _initial_gvn(nullptr), 944 _igvn_worklist(nullptr), 945 _types(nullptr), 946 _node_hash(nullptr), 947 _number_of_mh_late_inlines(0), 948 _oom(false), 949 _print_inlining_stream(new (mtCompiler) stringStream()), 950 _print_inlining_list(nullptr), 951 _print_inlining_idx(0), 952 _print_inlining_output(nullptr), 953 _replay_inline_data(nullptr), 954 _java_calls(0), 955 _inner_loops(0), 956 _interpreter_frame_size(0), 957 _output(nullptr), 958 #ifndef PRODUCT 959 _in_dump_cnt(0), 960 #endif 961 _allowed_reasons(0) { 962 C = this; 963 964 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 965 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 966 967 #ifndef PRODUCT 968 set_print_assembly(PrintFrameConverterAssembly); 969 set_parsed_irreducible_loop(false); 970 #else 971 set_print_assembly(false); // Must initialize. 972 #endif 973 set_has_irreducible_loop(false); // no loops 974 975 CompileWrapper cw(this); 976 Init(/*do_aliasing=*/ false); 977 init_tf((*generator)()); 978 979 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 980 _types = new (comp_arena()) Type_Array(comp_arena()); 981 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 982 { 983 PhaseGVN gvn; 984 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 985 gvn.transform(top()); 986 987 GraphKit kit; 988 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 989 } 990 991 NOT_PRODUCT( verify_graph_edges(); ) 992 993 Code_Gen(); 994 } 995 996 Compile::~Compile() { 997 delete _print_inlining_stream; 998 delete _first_failure_details; 999 }; 1000 1001 //------------------------------Init------------------------------------------- 1002 // Prepare for a single compilation 1003 void Compile::Init(bool aliasing) { 1004 _do_aliasing = aliasing; 1005 _unique = 0; 1006 _regalloc = nullptr; 1007 1008 _tf = nullptr; // filled in later 1009 _top = nullptr; // cached later 1010 _matcher = nullptr; // filled in later 1011 _cfg = nullptr; // filled in later 1012 1013 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 1014 1015 _node_note_array = nullptr; 1016 _default_node_notes = nullptr; 1017 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1018 1019 _immutable_memory = nullptr; // filled in at first inquiry 1020 1021 #ifdef ASSERT 1022 _phase_optimize_finished = false; 1023 _exception_backedge = false; 1024 _type_verify = nullptr; 1025 #endif 1026 1027 // Globally visible Nodes 1028 // First set TOP to null to give safe behavior during creation of RootNode 1029 set_cached_top_node(nullptr); 1030 set_root(new RootNode()); 1031 // Now that you have a Root to point to, create the real TOP 1032 set_cached_top_node( new ConNode(Type::TOP) ); 1033 set_recent_alloc(nullptr, nullptr); 1034 1035 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1036 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1037 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1038 env()->set_dependencies(new Dependencies(env())); 1039 1040 _fixed_slots = 0; 1041 set_has_split_ifs(false); 1042 set_has_loops(false); // first approximation 1043 set_has_stringbuilder(false); 1044 set_has_boxed_value(false); 1045 _trap_can_recompile = false; // no traps emitted yet 1046 _major_progress = true; // start out assuming good things will happen 1047 set_has_unsafe_access(false); 1048 set_max_vector_size(0); 1049 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1050 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1051 set_decompile_count(0); 1052 1053 #ifndef PRODUCT 1054 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter)); 1055 #endif 1056 1057 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1058 _loop_opts_cnt = LoopOptsCount; 1059 set_do_inlining(Inline); 1060 set_max_inline_size(MaxInlineSize); 1061 set_freq_inline_size(FreqInlineSize); 1062 set_do_scheduling(OptoScheduling); 1063 1064 set_do_vector_loop(false); 1065 set_has_monitors(false); 1066 1067 if (AllowVectorizeOnDemand) { 1068 if (has_method() && _directive->VectorizeOption) { 1069 set_do_vector_loop(true); 1070 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());}) 1071 } else if (has_method() && method()->name() != 0 && 1072 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1073 set_do_vector_loop(true); 1074 } 1075 } 1076 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1077 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());}) 1078 1079 set_rtm_state(NoRTM); // No RTM lock eliding by default 1080 _max_node_limit = _directive->MaxNodeLimitOption; 1081 1082 #if INCLUDE_RTM_OPT 1083 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1084 int rtm_state = method()->method_data()->rtm_state(); 1085 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1086 // Don't generate RTM lock eliding code. 1087 set_rtm_state(NoRTM); 1088 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1089 // Generate RTM lock eliding code without abort ratio calculation code. 1090 set_rtm_state(UseRTM); 1091 } else if (UseRTMDeopt) { 1092 // Generate RTM lock eliding code and include abort ratio calculation 1093 // code if UseRTMDeopt is on. 1094 set_rtm_state(ProfileRTM); 1095 } 1096 } 1097 #endif 1098 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1099 set_clinit_barrier_on_entry(true); 1100 } 1101 if (debug_info()->recording_non_safepoints()) { 1102 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1103 (comp_arena(), 8, 0, nullptr)); 1104 set_default_node_notes(Node_Notes::make(this)); 1105 } 1106 1107 const int grow_ats = 16; 1108 _max_alias_types = grow_ats; 1109 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1110 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1111 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1112 { 1113 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1114 } 1115 // Initialize the first few types. 1116 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1117 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1118 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1119 _num_alias_types = AliasIdxRaw+1; 1120 // Zero out the alias type cache. 1121 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1122 // A null adr_type hits in the cache right away. Preload the right answer. 1123 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1124 } 1125 1126 //---------------------------init_start---------------------------------------- 1127 // Install the StartNode on this compile object. 1128 void Compile::init_start(StartNode* s) { 1129 if (failing()) 1130 return; // already failing 1131 assert(s == start(), ""); 1132 } 1133 1134 /** 1135 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1136 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1137 * the ideal graph. 1138 */ 1139 StartNode* Compile::start() const { 1140 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1141 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1142 Node* start = root()->fast_out(i); 1143 if (start->is_Start()) { 1144 return start->as_Start(); 1145 } 1146 } 1147 fatal("Did not find Start node!"); 1148 return nullptr; 1149 } 1150 1151 //-------------------------------immutable_memory------------------------------------- 1152 // Access immutable memory 1153 Node* Compile::immutable_memory() { 1154 if (_immutable_memory != nullptr) { 1155 return _immutable_memory; 1156 } 1157 StartNode* s = start(); 1158 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1159 Node *p = s->fast_out(i); 1160 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1161 _immutable_memory = p; 1162 return _immutable_memory; 1163 } 1164 } 1165 ShouldNotReachHere(); 1166 return nullptr; 1167 } 1168 1169 //----------------------set_cached_top_node------------------------------------ 1170 // Install the cached top node, and make sure Node::is_top works correctly. 1171 void Compile::set_cached_top_node(Node* tn) { 1172 if (tn != nullptr) verify_top(tn); 1173 Node* old_top = _top; 1174 _top = tn; 1175 // Calling Node::setup_is_top allows the nodes the chance to adjust 1176 // their _out arrays. 1177 if (_top != nullptr) _top->setup_is_top(); 1178 if (old_top != nullptr) old_top->setup_is_top(); 1179 assert(_top == nullptr || top()->is_top(), ""); 1180 } 1181 1182 #ifdef ASSERT 1183 uint Compile::count_live_nodes_by_graph_walk() { 1184 Unique_Node_List useful(comp_arena()); 1185 // Get useful node list by walking the graph. 1186 identify_useful_nodes(useful); 1187 return useful.size(); 1188 } 1189 1190 void Compile::print_missing_nodes() { 1191 1192 // Return if CompileLog is null and PrintIdealNodeCount is false. 1193 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1194 return; 1195 } 1196 1197 // This is an expensive function. It is executed only when the user 1198 // specifies VerifyIdealNodeCount option or otherwise knows the 1199 // additional work that needs to be done to identify reachable nodes 1200 // by walking the flow graph and find the missing ones using 1201 // _dead_node_list. 1202 1203 Unique_Node_List useful(comp_arena()); 1204 // Get useful node list by walking the graph. 1205 identify_useful_nodes(useful); 1206 1207 uint l_nodes = C->live_nodes(); 1208 uint l_nodes_by_walk = useful.size(); 1209 1210 if (l_nodes != l_nodes_by_walk) { 1211 if (_log != nullptr) { 1212 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1213 _log->stamp(); 1214 _log->end_head(); 1215 } 1216 VectorSet& useful_member_set = useful.member_set(); 1217 int last_idx = l_nodes_by_walk; 1218 for (int i = 0; i < last_idx; i++) { 1219 if (useful_member_set.test(i)) { 1220 if (_dead_node_list.test(i)) { 1221 if (_log != nullptr) { 1222 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1223 } 1224 if (PrintIdealNodeCount) { 1225 // Print the log message to tty 1226 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1227 useful.at(i)->dump(); 1228 } 1229 } 1230 } 1231 else if (! _dead_node_list.test(i)) { 1232 if (_log != nullptr) { 1233 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1234 } 1235 if (PrintIdealNodeCount) { 1236 // Print the log message to tty 1237 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1238 } 1239 } 1240 } 1241 if (_log != nullptr) { 1242 _log->tail("mismatched_nodes"); 1243 } 1244 } 1245 } 1246 void Compile::record_modified_node(Node* n) { 1247 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1248 _modified_nodes->push(n); 1249 } 1250 } 1251 1252 void Compile::remove_modified_node(Node* n) { 1253 if (_modified_nodes != nullptr) { 1254 _modified_nodes->remove(n); 1255 } 1256 } 1257 #endif 1258 1259 #ifndef PRODUCT 1260 void Compile::verify_top(Node* tn) const { 1261 if (tn != nullptr) { 1262 assert(tn->is_Con(), "top node must be a constant"); 1263 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1264 assert(tn->in(0) != nullptr, "must have live top node"); 1265 } 1266 } 1267 #endif 1268 1269 1270 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1271 1272 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1273 guarantee(arr != nullptr, ""); 1274 int num_blocks = arr->length(); 1275 if (grow_by < num_blocks) grow_by = num_blocks; 1276 int num_notes = grow_by * _node_notes_block_size; 1277 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1278 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1279 while (num_notes > 0) { 1280 arr->append(notes); 1281 notes += _node_notes_block_size; 1282 num_notes -= _node_notes_block_size; 1283 } 1284 assert(num_notes == 0, "exact multiple, please"); 1285 } 1286 1287 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1288 if (source == nullptr || dest == nullptr) return false; 1289 1290 if (dest->is_Con()) 1291 return false; // Do not push debug info onto constants. 1292 1293 #ifdef ASSERT 1294 // Leave a bread crumb trail pointing to the original node: 1295 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1296 dest->set_debug_orig(source); 1297 } 1298 #endif 1299 1300 if (node_note_array() == nullptr) 1301 return false; // Not collecting any notes now. 1302 1303 // This is a copy onto a pre-existing node, which may already have notes. 1304 // If both nodes have notes, do not overwrite any pre-existing notes. 1305 Node_Notes* source_notes = node_notes_at(source->_idx); 1306 if (source_notes == nullptr || source_notes->is_clear()) return false; 1307 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1308 if (dest_notes == nullptr || dest_notes->is_clear()) { 1309 return set_node_notes_at(dest->_idx, source_notes); 1310 } 1311 1312 Node_Notes merged_notes = (*source_notes); 1313 // The order of operations here ensures that dest notes will win... 1314 merged_notes.update_from(dest_notes); 1315 return set_node_notes_at(dest->_idx, &merged_notes); 1316 } 1317 1318 1319 //--------------------------allow_range_check_smearing------------------------- 1320 // Gating condition for coalescing similar range checks. 1321 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1322 // single covering check that is at least as strong as any of them. 1323 // If the optimization succeeds, the simplified (strengthened) range check 1324 // will always succeed. If it fails, we will deopt, and then give up 1325 // on the optimization. 1326 bool Compile::allow_range_check_smearing() const { 1327 // If this method has already thrown a range-check, 1328 // assume it was because we already tried range smearing 1329 // and it failed. 1330 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1331 return !already_trapped; 1332 } 1333 1334 1335 //------------------------------flatten_alias_type----------------------------- 1336 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1337 assert(do_aliasing(), "Aliasing should be enabled"); 1338 int offset = tj->offset(); 1339 TypePtr::PTR ptr = tj->ptr(); 1340 1341 // Known instance (scalarizable allocation) alias only with itself. 1342 bool is_known_inst = tj->isa_oopptr() != nullptr && 1343 tj->is_oopptr()->is_known_instance(); 1344 1345 // Process weird unsafe references. 1346 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1347 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1348 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1349 tj = TypeOopPtr::BOTTOM; 1350 ptr = tj->ptr(); 1351 offset = tj->offset(); 1352 } 1353 1354 // Array pointers need some flattening 1355 const TypeAryPtr* ta = tj->isa_aryptr(); 1356 if (ta && ta->is_stable()) { 1357 // Erase stability property for alias analysis. 1358 tj = ta = ta->cast_to_stable(false); 1359 } 1360 if( ta && is_known_inst ) { 1361 if ( offset != Type::OffsetBot && 1362 offset > arrayOopDesc::length_offset_in_bytes() ) { 1363 offset = Type::OffsetBot; // Flatten constant access into array body only 1364 tj = ta = ta-> 1365 remove_speculative()-> 1366 cast_to_ptr_type(ptr)-> 1367 with_offset(offset); 1368 } 1369 } else if (ta) { 1370 // For arrays indexed by constant indices, we flatten the alias 1371 // space to include all of the array body. Only the header, klass 1372 // and array length can be accessed un-aliased. 1373 if( offset != Type::OffsetBot ) { 1374 if( ta->const_oop() ) { // MethodData* or Method* 1375 offset = Type::OffsetBot; // Flatten constant access into array body 1376 tj = ta = ta-> 1377 remove_speculative()-> 1378 cast_to_ptr_type(ptr)-> 1379 cast_to_exactness(false)-> 1380 with_offset(offset); 1381 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1382 // range is OK as-is. 1383 tj = ta = TypeAryPtr::RANGE; 1384 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1385 tj = TypeInstPtr::KLASS; // all klass loads look alike 1386 ta = TypeAryPtr::RANGE; // generic ignored junk 1387 ptr = TypePtr::BotPTR; 1388 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1389 tj = TypeInstPtr::MARK; 1390 ta = TypeAryPtr::RANGE; // generic ignored junk 1391 ptr = TypePtr::BotPTR; 1392 } else { // Random constant offset into array body 1393 offset = Type::OffsetBot; // Flatten constant access into array body 1394 tj = ta = ta-> 1395 remove_speculative()-> 1396 cast_to_ptr_type(ptr)-> 1397 cast_to_exactness(false)-> 1398 with_offset(offset); 1399 } 1400 } 1401 // Arrays of fixed size alias with arrays of unknown size. 1402 if (ta->size() != TypeInt::POS) { 1403 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1404 tj = ta = ta-> 1405 remove_speculative()-> 1406 cast_to_ptr_type(ptr)-> 1407 with_ary(tary)-> 1408 cast_to_exactness(false); 1409 } 1410 // Arrays of known objects become arrays of unknown objects. 1411 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1412 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1413 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1414 } 1415 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1416 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1418 } 1419 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1420 // cannot be distinguished by bytecode alone. 1421 if (ta->elem() == TypeInt::BOOL) { 1422 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1423 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1425 } 1426 // During the 2nd round of IterGVN, NotNull castings are removed. 1427 // Make sure the Bottom and NotNull variants alias the same. 1428 // Also, make sure exact and non-exact variants alias the same. 1429 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1430 tj = ta = ta-> 1431 remove_speculative()-> 1432 cast_to_ptr_type(TypePtr::BotPTR)-> 1433 cast_to_exactness(false)-> 1434 with_offset(offset); 1435 } 1436 } 1437 1438 // Oop pointers need some flattening 1439 const TypeInstPtr *to = tj->isa_instptr(); 1440 if (to && to != TypeOopPtr::BOTTOM) { 1441 ciInstanceKlass* ik = to->instance_klass(); 1442 if( ptr == TypePtr::Constant ) { 1443 if (ik != ciEnv::current()->Class_klass() || 1444 offset < ik->layout_helper_size_in_bytes()) { 1445 // No constant oop pointers (such as Strings); they alias with 1446 // unknown strings. 1447 assert(!is_known_inst, "not scalarizable allocation"); 1448 tj = to = to-> 1449 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1450 remove_speculative()-> 1451 cast_to_ptr_type(TypePtr::BotPTR)-> 1452 cast_to_exactness(false); 1453 } 1454 } else if( is_known_inst ) { 1455 tj = to; // Keep NotNull and klass_is_exact for instance type 1456 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1457 // During the 2nd round of IterGVN, NotNull castings are removed. 1458 // Make sure the Bottom and NotNull variants alias the same. 1459 // Also, make sure exact and non-exact variants alias the same. 1460 tj = to = to-> 1461 remove_speculative()-> 1462 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1463 cast_to_ptr_type(TypePtr::BotPTR)-> 1464 cast_to_exactness(false); 1465 } 1466 if (to->speculative() != nullptr) { 1467 tj = to = to->remove_speculative(); 1468 } 1469 // Canonicalize the holder of this field 1470 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1471 // First handle header references such as a LoadKlassNode, even if the 1472 // object's klass is unloaded at compile time (4965979). 1473 if (!is_known_inst) { // Do it only for non-instance types 1474 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1475 } 1476 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1477 // Static fields are in the space above the normal instance 1478 // fields in the java.lang.Class instance. 1479 if (ik != ciEnv::current()->Class_klass()) { 1480 to = nullptr; 1481 tj = TypeOopPtr::BOTTOM; 1482 offset = tj->offset(); 1483 } 1484 } else { 1485 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1486 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1487 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1488 if( is_known_inst ) { 1489 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id()); 1490 } else { 1491 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset); 1492 } 1493 } 1494 } 1495 } 1496 1497 // Klass pointers to object array klasses need some flattening 1498 const TypeKlassPtr *tk = tj->isa_klassptr(); 1499 if( tk ) { 1500 // If we are referencing a field within a Klass, we need 1501 // to assume the worst case of an Object. Both exact and 1502 // inexact types must flatten to the same alias class so 1503 // use NotNull as the PTR. 1504 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1505 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1506 env()->Object_klass(), 1507 offset); 1508 } 1509 1510 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1511 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1512 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1513 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1514 } else { 1515 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1516 } 1517 } 1518 1519 // Check for precise loads from the primary supertype array and force them 1520 // to the supertype cache alias index. Check for generic array loads from 1521 // the primary supertype array and also force them to the supertype cache 1522 // alias index. Since the same load can reach both, we need to merge 1523 // these 2 disparate memories into the same alias class. Since the 1524 // primary supertype array is read-only, there's no chance of confusion 1525 // where we bypass an array load and an array store. 1526 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1527 if (offset == Type::OffsetBot || 1528 (offset >= primary_supers_offset && 1529 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1530 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1531 offset = in_bytes(Klass::secondary_super_cache_offset()); 1532 tj = tk = tk->with_offset(offset); 1533 } 1534 } 1535 1536 // Flatten all Raw pointers together. 1537 if (tj->base() == Type::RawPtr) 1538 tj = TypeRawPtr::BOTTOM; 1539 1540 if (tj->base() == Type::AnyPtr) 1541 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1542 1543 offset = tj->offset(); 1544 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1545 1546 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1547 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1548 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1549 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1550 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1551 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1552 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1553 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1554 assert( tj->ptr() != TypePtr::TopPTR && 1555 tj->ptr() != TypePtr::AnyNull && 1556 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1557 // assert( tj->ptr() != TypePtr::Constant || 1558 // tj->base() == Type::RawPtr || 1559 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1560 1561 return tj; 1562 } 1563 1564 void Compile::AliasType::Init(int i, const TypePtr* at) { 1565 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1566 _index = i; 1567 _adr_type = at; 1568 _field = nullptr; 1569 _element = nullptr; 1570 _is_rewritable = true; // default 1571 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1572 if (atoop != nullptr && atoop->is_known_instance()) { 1573 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1574 _general_index = Compile::current()->get_alias_index(gt); 1575 } else { 1576 _general_index = 0; 1577 } 1578 } 1579 1580 BasicType Compile::AliasType::basic_type() const { 1581 if (element() != nullptr) { 1582 const Type* element = adr_type()->is_aryptr()->elem(); 1583 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1584 } if (field() != nullptr) { 1585 return field()->layout_type(); 1586 } else { 1587 return T_ILLEGAL; // unknown 1588 } 1589 } 1590 1591 //---------------------------------print_on------------------------------------ 1592 #ifndef PRODUCT 1593 void Compile::AliasType::print_on(outputStream* st) { 1594 if (index() < 10) 1595 st->print("@ <%d> ", index()); 1596 else st->print("@ <%d>", index()); 1597 st->print(is_rewritable() ? " " : " RO"); 1598 int offset = adr_type()->offset(); 1599 if (offset == Type::OffsetBot) 1600 st->print(" +any"); 1601 else st->print(" +%-3d", offset); 1602 st->print(" in "); 1603 adr_type()->dump_on(st); 1604 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1605 if (field() != nullptr && tjp) { 1606 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1607 tjp->offset() != field()->offset_in_bytes()) { 1608 st->print(" != "); 1609 field()->print(); 1610 st->print(" ***"); 1611 } 1612 } 1613 } 1614 1615 void print_alias_types() { 1616 Compile* C = Compile::current(); 1617 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1618 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1619 C->alias_type(idx)->print_on(tty); 1620 tty->cr(); 1621 } 1622 } 1623 #endif 1624 1625 1626 //----------------------------probe_alias_cache-------------------------------- 1627 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1628 intptr_t key = (intptr_t) adr_type; 1629 key ^= key >> logAliasCacheSize; 1630 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1631 } 1632 1633 1634 //-----------------------------grow_alias_types-------------------------------- 1635 void Compile::grow_alias_types() { 1636 const int old_ats = _max_alias_types; // how many before? 1637 const int new_ats = old_ats; // how many more? 1638 const int grow_ats = old_ats+new_ats; // how many now? 1639 _max_alias_types = grow_ats; 1640 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1641 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1642 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1643 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1644 } 1645 1646 1647 //--------------------------------find_alias_type------------------------------ 1648 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1649 if (!do_aliasing()) { 1650 return alias_type(AliasIdxBot); 1651 } 1652 1653 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1654 if (ace->_adr_type == adr_type) { 1655 return alias_type(ace->_index); 1656 } 1657 1658 // Handle special cases. 1659 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1660 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1661 1662 // Do it the slow way. 1663 const TypePtr* flat = flatten_alias_type(adr_type); 1664 1665 #ifdef ASSERT 1666 { 1667 ResourceMark rm; 1668 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1669 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1670 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1671 Type::str(adr_type)); 1672 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1673 const TypeOopPtr* foop = flat->is_oopptr(); 1674 // Scalarizable allocations have exact klass always. 1675 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1676 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1677 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1678 Type::str(foop), Type::str(xoop)); 1679 } 1680 } 1681 #endif 1682 1683 int idx = AliasIdxTop; 1684 for (int i = 0; i < num_alias_types(); i++) { 1685 if (alias_type(i)->adr_type() == flat) { 1686 idx = i; 1687 break; 1688 } 1689 } 1690 1691 if (idx == AliasIdxTop) { 1692 if (no_create) return nullptr; 1693 // Grow the array if necessary. 1694 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1695 // Add a new alias type. 1696 idx = _num_alias_types++; 1697 _alias_types[idx]->Init(idx, flat); 1698 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1699 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1700 if (flat->isa_instptr()) { 1701 if (flat->offset() == java_lang_Class::klass_offset() 1702 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1703 alias_type(idx)->set_rewritable(false); 1704 } 1705 if (flat->isa_aryptr()) { 1706 #ifdef ASSERT 1707 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1708 // (T_BYTE has the weakest alignment and size restrictions...) 1709 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1710 #endif 1711 if (flat->offset() == TypePtr::OffsetBot) { 1712 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1713 } 1714 } 1715 if (flat->isa_klassptr()) { 1716 if (UseCompactObjectHeaders) { 1717 if (flat->offset() == in_bytes(Klass::prototype_header_offset())) 1718 alias_type(idx)->set_rewritable(false); 1719 } 1720 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1721 alias_type(idx)->set_rewritable(false); 1722 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1723 alias_type(idx)->set_rewritable(false); 1724 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1725 alias_type(idx)->set_rewritable(false); 1726 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1727 alias_type(idx)->set_rewritable(false); 1728 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1729 alias_type(idx)->set_rewritable(false); 1730 } 1731 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1732 // but the base pointer type is not distinctive enough to identify 1733 // references into JavaThread.) 1734 1735 // Check for final fields. 1736 const TypeInstPtr* tinst = flat->isa_instptr(); 1737 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1738 ciField* field; 1739 if (tinst->const_oop() != nullptr && 1740 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1741 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1742 // static field 1743 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1744 field = k->get_field_by_offset(tinst->offset(), true); 1745 } else { 1746 ciInstanceKlass *k = tinst->instance_klass(); 1747 field = k->get_field_by_offset(tinst->offset(), false); 1748 } 1749 assert(field == nullptr || 1750 original_field == nullptr || 1751 (field->holder() == original_field->holder() && 1752 field->offset_in_bytes() == original_field->offset_in_bytes() && 1753 field->is_static() == original_field->is_static()), "wrong field?"); 1754 // Set field() and is_rewritable() attributes. 1755 if (field != nullptr) alias_type(idx)->set_field(field); 1756 } 1757 } 1758 1759 // Fill the cache for next time. 1760 ace->_adr_type = adr_type; 1761 ace->_index = idx; 1762 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1763 1764 // Might as well try to fill the cache for the flattened version, too. 1765 AliasCacheEntry* face = probe_alias_cache(flat); 1766 if (face->_adr_type == nullptr) { 1767 face->_adr_type = flat; 1768 face->_index = idx; 1769 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1770 } 1771 1772 return alias_type(idx); 1773 } 1774 1775 1776 Compile::AliasType* Compile::alias_type(ciField* field) { 1777 const TypeOopPtr* t; 1778 if (field->is_static()) 1779 t = TypeInstPtr::make(field->holder()->java_mirror()); 1780 else 1781 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1782 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1783 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1784 return atp; 1785 } 1786 1787 1788 //------------------------------have_alias_type-------------------------------- 1789 bool Compile::have_alias_type(const TypePtr* adr_type) { 1790 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1791 if (ace->_adr_type == adr_type) { 1792 return true; 1793 } 1794 1795 // Handle special cases. 1796 if (adr_type == nullptr) return true; 1797 if (adr_type == TypePtr::BOTTOM) return true; 1798 1799 return find_alias_type(adr_type, true, nullptr) != nullptr; 1800 } 1801 1802 //-----------------------------must_alias-------------------------------------- 1803 // True if all values of the given address type are in the given alias category. 1804 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1805 if (alias_idx == AliasIdxBot) return true; // the universal category 1806 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1807 if (alias_idx == AliasIdxTop) return false; // the empty category 1808 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1809 1810 // the only remaining possible overlap is identity 1811 int adr_idx = get_alias_index(adr_type); 1812 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1813 assert(adr_idx == alias_idx || 1814 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1815 && adr_type != TypeOopPtr::BOTTOM), 1816 "should not be testing for overlap with an unsafe pointer"); 1817 return adr_idx == alias_idx; 1818 } 1819 1820 //------------------------------can_alias-------------------------------------- 1821 // True if any values of the given address type are in the given alias category. 1822 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1823 if (alias_idx == AliasIdxTop) return false; // the empty category 1824 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1825 // Known instance doesn't alias with bottom memory 1826 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1827 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1828 1829 // the only remaining possible overlap is identity 1830 int adr_idx = get_alias_index(adr_type); 1831 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1832 return adr_idx == alias_idx; 1833 } 1834 1835 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1836 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1837 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1838 if (parse_predicate_count() == 0) { 1839 return; 1840 } 1841 for (int i = 0; i < parse_predicate_count(); i++) { 1842 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1843 parse_predicate->mark_useless(); 1844 igvn._worklist.push(parse_predicate); 1845 } 1846 _parse_predicates.clear(); 1847 } 1848 1849 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1850 if (!n->for_post_loop_opts_igvn()) { 1851 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1852 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1853 _for_post_loop_igvn.append(n); 1854 } 1855 } 1856 1857 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1858 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1859 _for_post_loop_igvn.remove(n); 1860 } 1861 1862 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1863 // Verify that all previous optimizations produced a valid graph 1864 // at least to this point, even if no loop optimizations were done. 1865 PhaseIdealLoop::verify(igvn); 1866 1867 C->set_post_loop_opts_phase(); // no more loop opts allowed 1868 1869 assert(!C->major_progress(), "not cleared"); 1870 1871 if (_for_post_loop_igvn.length() > 0) { 1872 while (_for_post_loop_igvn.length() > 0) { 1873 Node* n = _for_post_loop_igvn.pop(); 1874 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1875 igvn._worklist.push(n); 1876 } 1877 igvn.optimize(); 1878 if (failing()) return; 1879 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1880 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1881 1882 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1883 if (C->major_progress()) { 1884 C->clear_major_progress(); // ensure that major progress is now clear 1885 } 1886 } 1887 } 1888 1889 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1890 if (OptimizeUnstableIf) { 1891 _unstable_if_traps.append(trap); 1892 } 1893 } 1894 1895 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1896 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1897 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1898 Node* n = trap->uncommon_trap(); 1899 if (!useful.member(n)) { 1900 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1901 } 1902 } 1903 } 1904 1905 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1906 // or fold-compares case. Return true if succeed or not found. 1907 // 1908 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1909 // false and ask fold-compares to yield. 1910 // 1911 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1912 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1913 // when deoptimization does happen. 1914 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1915 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1916 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1917 if (trap->uncommon_trap() == unc) { 1918 if (yield && trap->modified()) { 1919 return false; 1920 } 1921 _unstable_if_traps.delete_at(i); 1922 break; 1923 } 1924 } 1925 return true; 1926 } 1927 1928 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1929 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1930 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1931 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1932 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1933 CallStaticJavaNode* unc = trap->uncommon_trap(); 1934 int next_bci = trap->next_bci(); 1935 bool modified = trap->modified(); 1936 1937 if (next_bci != -1 && !modified) { 1938 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1939 JVMState* jvms = unc->jvms(); 1940 ciMethod* method = jvms->method(); 1941 ciBytecodeStream iter(method); 1942 1943 iter.force_bci(jvms->bci()); 1944 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1945 Bytecodes::Code c = iter.cur_bc(); 1946 Node* lhs = nullptr; 1947 Node* rhs = nullptr; 1948 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1949 lhs = unc->peek_operand(0); 1950 rhs = unc->peek_operand(1); 1951 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 1952 lhs = unc->peek_operand(0); 1953 } 1954 1955 ResourceMark rm; 1956 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 1957 assert(live_locals.is_valid(), "broken liveness info"); 1958 int len = (int)live_locals.size(); 1959 1960 for (int i = 0; i < len; i++) { 1961 Node* local = unc->local(jvms, i); 1962 // kill local using the liveness of next_bci. 1963 // give up when the local looks like an operand to secure reexecution. 1964 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 1965 uint idx = jvms->locoff() + i; 1966 #ifdef ASSERT 1967 if (PrintOpto && Verbose) { 1968 tty->print("[unstable_if] kill local#%d: ", idx); 1969 local->dump(); 1970 tty->cr(); 1971 } 1972 #endif 1973 igvn.replace_input_of(unc, idx, top()); 1974 modified = true; 1975 } 1976 } 1977 } 1978 1979 // keep the mondified trap for late query 1980 if (modified) { 1981 trap->set_modified(); 1982 } else { 1983 _unstable_if_traps.delete_at(i); 1984 } 1985 } 1986 igvn.optimize(); 1987 } 1988 1989 // StringOpts and late inlining of string methods 1990 void Compile::inline_string_calls(bool parse_time) { 1991 { 1992 // remove useless nodes to make the usage analysis simpler 1993 ResourceMark rm; 1994 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 1995 } 1996 1997 { 1998 ResourceMark rm; 1999 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2000 PhaseStringOpts pso(initial_gvn()); 2001 print_method(PHASE_AFTER_STRINGOPTS, 3); 2002 } 2003 2004 // now inline anything that we skipped the first time around 2005 if (!parse_time) { 2006 _late_inlines_pos = _late_inlines.length(); 2007 } 2008 2009 while (_string_late_inlines.length() > 0) { 2010 CallGenerator* cg = _string_late_inlines.pop(); 2011 cg->do_late_inline(); 2012 if (failing()) return; 2013 } 2014 _string_late_inlines.trunc_to(0); 2015 } 2016 2017 // Late inlining of boxing methods 2018 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2019 if (_boxing_late_inlines.length() > 0) { 2020 assert(has_boxed_value(), "inconsistent"); 2021 2022 set_inlining_incrementally(true); 2023 2024 igvn_worklist()->ensure_empty(); // should be done with igvn 2025 2026 _late_inlines_pos = _late_inlines.length(); 2027 2028 while (_boxing_late_inlines.length() > 0) { 2029 CallGenerator* cg = _boxing_late_inlines.pop(); 2030 cg->do_late_inline(); 2031 if (failing()) return; 2032 } 2033 _boxing_late_inlines.trunc_to(0); 2034 2035 inline_incrementally_cleanup(igvn); 2036 2037 set_inlining_incrementally(false); 2038 } 2039 } 2040 2041 bool Compile::inline_incrementally_one() { 2042 assert(IncrementalInline, "incremental inlining should be on"); 2043 2044 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2045 2046 set_inlining_progress(false); 2047 set_do_cleanup(false); 2048 2049 for (int i = 0; i < _late_inlines.length(); i++) { 2050 _late_inlines_pos = i+1; 2051 CallGenerator* cg = _late_inlines.at(i); 2052 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2053 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2054 cg->do_late_inline(); 2055 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2056 if (failing()) { 2057 return false; 2058 } else if (inlining_progress()) { 2059 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2060 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2061 break; // process one call site at a time 2062 } 2063 } else { 2064 // Ignore late inline direct calls when inlining is not allowed. 2065 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2066 } 2067 } 2068 // Remove processed elements. 2069 _late_inlines.remove_till(_late_inlines_pos); 2070 _late_inlines_pos = 0; 2071 2072 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2073 2074 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2075 2076 set_inlining_progress(false); 2077 set_do_cleanup(false); 2078 2079 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2080 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2081 } 2082 2083 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2084 { 2085 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2086 ResourceMark rm; 2087 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2088 } 2089 { 2090 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2091 igvn.reset_from_gvn(initial_gvn()); 2092 igvn.optimize(); 2093 if (failing()) return; 2094 } 2095 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2096 } 2097 2098 // Perform incremental inlining until bound on number of live nodes is reached 2099 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2100 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2101 2102 set_inlining_incrementally(true); 2103 uint low_live_nodes = 0; 2104 2105 while (_late_inlines.length() > 0) { 2106 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2107 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2108 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2109 // PhaseIdealLoop is expensive so we only try it once we are 2110 // out of live nodes and we only try it again if the previous 2111 // helped got the number of nodes down significantly 2112 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2113 if (failing()) return; 2114 low_live_nodes = live_nodes(); 2115 _major_progress = true; 2116 } 2117 2118 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2119 bool do_print_inlining = print_inlining() || print_intrinsics(); 2120 if (do_print_inlining || log() != nullptr) { 2121 // Print inlining message for candidates that we couldn't inline for lack of space. 2122 for (int i = 0; i < _late_inlines.length(); i++) { 2123 CallGenerator* cg = _late_inlines.at(i); 2124 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2125 if (do_print_inlining) { 2126 cg->print_inlining_late(InliningResult::FAILURE, msg); 2127 } 2128 log_late_inline_failure(cg, msg); 2129 } 2130 } 2131 break; // finish 2132 } 2133 } 2134 2135 igvn_worklist()->ensure_empty(); // should be done with igvn 2136 2137 while (inline_incrementally_one()) { 2138 assert(!failing(), "inconsistent"); 2139 } 2140 if (failing()) return; 2141 2142 inline_incrementally_cleanup(igvn); 2143 2144 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2145 2146 if (failing()) return; 2147 2148 if (_late_inlines.length() == 0) { 2149 break; // no more progress 2150 } 2151 } 2152 2153 igvn_worklist()->ensure_empty(); // should be done with igvn 2154 2155 if (_string_late_inlines.length() > 0) { 2156 assert(has_stringbuilder(), "inconsistent"); 2157 2158 inline_string_calls(false); 2159 2160 if (failing()) return; 2161 2162 inline_incrementally_cleanup(igvn); 2163 } 2164 2165 set_inlining_incrementally(false); 2166 } 2167 2168 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2169 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2170 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2171 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2172 // as if "inlining_incrementally() == true" were set. 2173 assert(inlining_incrementally() == false, "not allowed"); 2174 assert(_modified_nodes == nullptr, "not allowed"); 2175 assert(_late_inlines.length() > 0, "sanity"); 2176 2177 while (_late_inlines.length() > 0) { 2178 igvn_worklist()->ensure_empty(); // should be done with igvn 2179 2180 while (inline_incrementally_one()) { 2181 assert(!failing(), "inconsistent"); 2182 } 2183 if (failing()) return; 2184 2185 inline_incrementally_cleanup(igvn); 2186 } 2187 } 2188 2189 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2190 if (_loop_opts_cnt > 0) { 2191 while (major_progress() && (_loop_opts_cnt > 0)) { 2192 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2193 PhaseIdealLoop::optimize(igvn, mode); 2194 _loop_opts_cnt--; 2195 if (failing()) return false; 2196 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2197 } 2198 } 2199 return true; 2200 } 2201 2202 // Remove edges from "root" to each SafePoint at a backward branch. 2203 // They were inserted during parsing (see add_safepoint()) to make 2204 // infinite loops without calls or exceptions visible to root, i.e., 2205 // useful. 2206 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2207 Node *r = root(); 2208 if (r != nullptr) { 2209 for (uint i = r->req(); i < r->len(); ++i) { 2210 Node *n = r->in(i); 2211 if (n != nullptr && n->is_SafePoint()) { 2212 r->rm_prec(i); 2213 if (n->outcnt() == 0) { 2214 igvn.remove_dead_node(n); 2215 } 2216 --i; 2217 } 2218 } 2219 // Parsing may have added top inputs to the root node (Path 2220 // leading to the Halt node proven dead). Make sure we get a 2221 // chance to clean them up. 2222 igvn._worklist.push(r); 2223 igvn.optimize(); 2224 } 2225 } 2226 2227 //------------------------------Optimize--------------------------------------- 2228 // Given a graph, optimize it. 2229 void Compile::Optimize() { 2230 TracePhase tp("optimizer", &timers[_t_optimizer]); 2231 2232 #ifndef PRODUCT 2233 if (env()->break_at_compile()) { 2234 BREAKPOINT; 2235 } 2236 2237 #endif 2238 2239 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2240 #ifdef ASSERT 2241 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2242 #endif 2243 2244 ResourceMark rm; 2245 2246 print_inlining_reinit(); 2247 2248 NOT_PRODUCT( verify_graph_edges(); ) 2249 2250 print_method(PHASE_AFTER_PARSING, 1); 2251 2252 { 2253 // Iterative Global Value Numbering, including ideal transforms 2254 // Initialize IterGVN with types and values from parse-time GVN 2255 PhaseIterGVN igvn(initial_gvn()); 2256 #ifdef ASSERT 2257 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2258 #endif 2259 { 2260 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2261 igvn.optimize(); 2262 } 2263 2264 if (failing()) return; 2265 2266 print_method(PHASE_ITER_GVN1, 2); 2267 2268 process_for_unstable_if_traps(igvn); 2269 2270 if (failing()) return; 2271 2272 inline_incrementally(igvn); 2273 2274 print_method(PHASE_INCREMENTAL_INLINE, 2); 2275 2276 if (failing()) return; 2277 2278 if (eliminate_boxing()) { 2279 // Inline valueOf() methods now. 2280 inline_boxing_calls(igvn); 2281 2282 if (failing()) return; 2283 2284 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2285 inline_incrementally(igvn); 2286 } 2287 2288 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2289 2290 if (failing()) return; 2291 } 2292 2293 // Remove the speculative part of types and clean up the graph from 2294 // the extra CastPP nodes whose only purpose is to carry them. Do 2295 // that early so that optimizations are not disrupted by the extra 2296 // CastPP nodes. 2297 remove_speculative_types(igvn); 2298 2299 if (failing()) return; 2300 2301 // No more new expensive nodes will be added to the list from here 2302 // so keep only the actual candidates for optimizations. 2303 cleanup_expensive_nodes(igvn); 2304 2305 if (failing()) return; 2306 2307 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2308 if (EnableVectorSupport && has_vbox_nodes()) { 2309 TracePhase tp("", &timers[_t_vector]); 2310 PhaseVector pv(igvn); 2311 pv.optimize_vector_boxes(); 2312 if (failing()) return; 2313 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2314 } 2315 assert(!has_vbox_nodes(), "sanity"); 2316 2317 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2318 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2319 igvn_worklist()->ensure_empty(); // should be done with igvn 2320 { 2321 ResourceMark rm; 2322 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2323 } 2324 igvn.reset_from_gvn(initial_gvn()); 2325 igvn.optimize(); 2326 if (failing()) return; 2327 } 2328 2329 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2330 // safepoints 2331 remove_root_to_sfpts_edges(igvn); 2332 2333 if (failing()) return; 2334 2335 // Perform escape analysis 2336 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2337 if (has_loops()) { 2338 // Cleanup graph (remove dead nodes). 2339 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2340 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2341 if (failing()) return; 2342 } 2343 bool progress; 2344 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2345 do { 2346 ConnectionGraph::do_analysis(this, &igvn); 2347 2348 if (failing()) return; 2349 2350 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2351 2352 // Optimize out fields loads from scalar replaceable allocations. 2353 igvn.optimize(); 2354 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2355 2356 if (failing()) return; 2357 2358 if (congraph() != nullptr && macro_count() > 0) { 2359 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2360 PhaseMacroExpand mexp(igvn); 2361 mexp.eliminate_macro_nodes(); 2362 if (failing()) return; 2363 2364 igvn.set_delay_transform(false); 2365 igvn.optimize(); 2366 if (failing()) return; 2367 2368 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2369 } 2370 2371 ConnectionGraph::verify_ram_nodes(this, root()); 2372 if (failing()) return; 2373 2374 progress = do_iterative_escape_analysis() && 2375 (macro_count() < mcount) && 2376 ConnectionGraph::has_candidates(this); 2377 // Try again if candidates exist and made progress 2378 // by removing some allocations and/or locks. 2379 } while (progress); 2380 } 2381 2382 // Loop transforms on the ideal graph. Range Check Elimination, 2383 // peeling, unrolling, etc. 2384 2385 // Set loop opts counter 2386 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2387 { 2388 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2389 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2390 _loop_opts_cnt--; 2391 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2392 if (failing()) return; 2393 } 2394 // Loop opts pass if partial peeling occurred in previous pass 2395 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2396 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2397 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2398 _loop_opts_cnt--; 2399 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2400 if (failing()) return; 2401 } 2402 // Loop opts pass for loop-unrolling before CCP 2403 if(major_progress() && (_loop_opts_cnt > 0)) { 2404 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2405 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2406 _loop_opts_cnt--; 2407 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2408 } 2409 if (!failing()) { 2410 // Verify that last round of loop opts produced a valid graph 2411 PhaseIdealLoop::verify(igvn); 2412 } 2413 } 2414 if (failing()) return; 2415 2416 // Conditional Constant Propagation; 2417 print_method(PHASE_BEFORE_CCP1, 2); 2418 PhaseCCP ccp( &igvn ); 2419 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2420 { 2421 TracePhase tp("ccp", &timers[_t_ccp]); 2422 ccp.do_transform(); 2423 } 2424 print_method(PHASE_CCP1, 2); 2425 2426 assert( true, "Break here to ccp.dump_old2new_map()"); 2427 2428 // Iterative Global Value Numbering, including ideal transforms 2429 { 2430 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2431 igvn.reset_from_igvn(&ccp); 2432 igvn.optimize(); 2433 } 2434 print_method(PHASE_ITER_GVN2, 2); 2435 2436 if (failing()) return; 2437 2438 // Loop transforms on the ideal graph. Range Check Elimination, 2439 // peeling, unrolling, etc. 2440 if (!optimize_loops(igvn, LoopOptsDefault)) { 2441 return; 2442 } 2443 2444 if (failing()) return; 2445 2446 C->clear_major_progress(); // ensure that major progress is now clear 2447 2448 process_for_post_loop_opts_igvn(igvn); 2449 2450 if (failing()) return; 2451 2452 #ifdef ASSERT 2453 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2454 #endif 2455 2456 { 2457 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2458 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3); 2459 PhaseMacroExpand mex(igvn); 2460 if (mex.expand_macro_nodes()) { 2461 assert(failing(), "must bail out w/ explicit message"); 2462 return; 2463 } 2464 print_method(PHASE_AFTER_MACRO_EXPANSION, 2); 2465 } 2466 2467 { 2468 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2469 if (bs->expand_barriers(this, igvn)) { 2470 assert(failing(), "must bail out w/ explicit message"); 2471 return; 2472 } 2473 print_method(PHASE_BARRIER_EXPANSION, 2); 2474 } 2475 2476 if (C->max_vector_size() > 0) { 2477 C->optimize_logic_cones(igvn); 2478 igvn.optimize(); 2479 if (failing()) return; 2480 } 2481 2482 DEBUG_ONLY( _modified_nodes = nullptr; ) 2483 2484 assert(igvn._worklist.size() == 0, "not empty"); 2485 2486 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2487 2488 if (_late_inlines.length() > 0) { 2489 // More opportunities to optimize virtual and MH calls. 2490 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2491 process_late_inline_calls_no_inline(igvn); 2492 if (failing()) return; 2493 } 2494 } // (End scope of igvn; run destructor if necessary for asserts.) 2495 2496 check_no_dead_use(); 2497 2498 process_print_inlining(); 2499 2500 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2501 // to remove hashes to unlock nodes for modifications. 2502 C->node_hash()->clear(); 2503 2504 // A method with only infinite loops has no edges entering loops from root 2505 { 2506 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2507 if (final_graph_reshaping()) { 2508 assert(failing(), "must bail out w/ explicit message"); 2509 return; 2510 } 2511 } 2512 2513 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2514 DEBUG_ONLY(set_phase_optimize_finished();) 2515 } 2516 2517 #ifdef ASSERT 2518 void Compile::check_no_dead_use() const { 2519 ResourceMark rm; 2520 Unique_Node_List wq; 2521 wq.push(root()); 2522 for (uint i = 0; i < wq.size(); ++i) { 2523 Node* n = wq.at(i); 2524 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2525 Node* u = n->fast_out(j); 2526 if (u->outcnt() == 0 && !u->is_Con()) { 2527 u->dump(); 2528 fatal("no reachable node should have no use"); 2529 } 2530 wq.push(u); 2531 } 2532 } 2533 } 2534 #endif 2535 2536 void Compile::inline_vector_reboxing_calls() { 2537 if (C->_vector_reboxing_late_inlines.length() > 0) { 2538 _late_inlines_pos = C->_late_inlines.length(); 2539 while (_vector_reboxing_late_inlines.length() > 0) { 2540 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2541 cg->do_late_inline(); 2542 if (failing()) return; 2543 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2544 } 2545 _vector_reboxing_late_inlines.trunc_to(0); 2546 } 2547 } 2548 2549 bool Compile::has_vbox_nodes() { 2550 if (C->_vector_reboxing_late_inlines.length() > 0) { 2551 return true; 2552 } 2553 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2554 Node * n = C->macro_node(macro_idx); 2555 assert(n->is_macro(), "only macro nodes expected here"); 2556 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2557 return true; 2558 } 2559 } 2560 return false; 2561 } 2562 2563 //---------------------------- Bitwise operation packing optimization --------------------------- 2564 2565 static bool is_vector_unary_bitwise_op(Node* n) { 2566 return n->Opcode() == Op_XorV && 2567 VectorNode::is_vector_bitwise_not_pattern(n); 2568 } 2569 2570 static bool is_vector_binary_bitwise_op(Node* n) { 2571 switch (n->Opcode()) { 2572 case Op_AndV: 2573 case Op_OrV: 2574 return true; 2575 2576 case Op_XorV: 2577 return !is_vector_unary_bitwise_op(n); 2578 2579 default: 2580 return false; 2581 } 2582 } 2583 2584 static bool is_vector_ternary_bitwise_op(Node* n) { 2585 return n->Opcode() == Op_MacroLogicV; 2586 } 2587 2588 static bool is_vector_bitwise_op(Node* n) { 2589 return is_vector_unary_bitwise_op(n) || 2590 is_vector_binary_bitwise_op(n) || 2591 is_vector_ternary_bitwise_op(n); 2592 } 2593 2594 static bool is_vector_bitwise_cone_root(Node* n) { 2595 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2596 return false; 2597 } 2598 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2599 if (is_vector_bitwise_op(n->fast_out(i))) { 2600 return false; 2601 } 2602 } 2603 return true; 2604 } 2605 2606 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2607 uint cnt = 0; 2608 if (is_vector_bitwise_op(n)) { 2609 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2610 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2611 for (uint i = 1; i < inp_cnt; i++) { 2612 Node* in = n->in(i); 2613 bool skip = VectorNode::is_all_ones_vector(in); 2614 if (!skip && !inputs.member(in)) { 2615 inputs.push(in); 2616 cnt++; 2617 } 2618 } 2619 assert(cnt <= 1, "not unary"); 2620 } else { 2621 uint last_req = inp_cnt; 2622 if (is_vector_ternary_bitwise_op(n)) { 2623 last_req = inp_cnt - 1; // skip last input 2624 } 2625 for (uint i = 1; i < last_req; i++) { 2626 Node* def = n->in(i); 2627 if (!inputs.member(def)) { 2628 inputs.push(def); 2629 cnt++; 2630 } 2631 } 2632 } 2633 } else { // not a bitwise operations 2634 if (!inputs.member(n)) { 2635 inputs.push(n); 2636 cnt++; 2637 } 2638 } 2639 return cnt; 2640 } 2641 2642 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2643 Unique_Node_List useful_nodes; 2644 C->identify_useful_nodes(useful_nodes); 2645 2646 for (uint i = 0; i < useful_nodes.size(); i++) { 2647 Node* n = useful_nodes.at(i); 2648 if (is_vector_bitwise_cone_root(n)) { 2649 list.push(n); 2650 } 2651 } 2652 } 2653 2654 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2655 const TypeVect* vt, 2656 Unique_Node_List& partition, 2657 Unique_Node_List& inputs) { 2658 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2659 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2660 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2661 2662 Node* in1 = inputs.at(0); 2663 Node* in2 = inputs.at(1); 2664 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2665 2666 uint func = compute_truth_table(partition, inputs); 2667 2668 Node* pn = partition.at(partition.size() - 1); 2669 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2670 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2671 } 2672 2673 static uint extract_bit(uint func, uint pos) { 2674 return (func & (1 << pos)) >> pos; 2675 } 2676 2677 // 2678 // A macro logic node represents a truth table. It has 4 inputs, 2679 // First three inputs corresponds to 3 columns of a truth table 2680 // and fourth input captures the logic function. 2681 // 2682 // eg. fn = (in1 AND in2) OR in3; 2683 // 2684 // MacroNode(in1,in2,in3,fn) 2685 // 2686 // ----------------- 2687 // in1 in2 in3 fn 2688 // ----------------- 2689 // 0 0 0 0 2690 // 0 0 1 1 2691 // 0 1 0 0 2692 // 0 1 1 1 2693 // 1 0 0 0 2694 // 1 0 1 1 2695 // 1 1 0 1 2696 // 1 1 1 1 2697 // 2698 2699 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2700 int res = 0; 2701 for (int i = 0; i < 8; i++) { 2702 int bit1 = extract_bit(in1, i); 2703 int bit2 = extract_bit(in2, i); 2704 int bit3 = extract_bit(in3, i); 2705 2706 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2707 int func_bit = extract_bit(func, func_bit_pos); 2708 2709 res |= func_bit << i; 2710 } 2711 return res; 2712 } 2713 2714 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2715 assert(n != nullptr, ""); 2716 assert(eval_map.contains(n), "absent"); 2717 return *(eval_map.get(n)); 2718 } 2719 2720 static void eval_operands(Node* n, 2721 uint& func1, uint& func2, uint& func3, 2722 ResourceHashtable<Node*,uint>& eval_map) { 2723 assert(is_vector_bitwise_op(n), ""); 2724 2725 if (is_vector_unary_bitwise_op(n)) { 2726 Node* opnd = n->in(1); 2727 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2728 opnd = n->in(2); 2729 } 2730 func1 = eval_operand(opnd, eval_map); 2731 } else if (is_vector_binary_bitwise_op(n)) { 2732 func1 = eval_operand(n->in(1), eval_map); 2733 func2 = eval_operand(n->in(2), eval_map); 2734 } else { 2735 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2736 func1 = eval_operand(n->in(1), eval_map); 2737 func2 = eval_operand(n->in(2), eval_map); 2738 func3 = eval_operand(n->in(3), eval_map); 2739 } 2740 } 2741 2742 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2743 assert(inputs.size() <= 3, "sanity"); 2744 ResourceMark rm; 2745 uint res = 0; 2746 ResourceHashtable<Node*,uint> eval_map; 2747 2748 // Populate precomputed functions for inputs. 2749 // Each input corresponds to one column of 3 input truth-table. 2750 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2751 0xCC, // (_, b, _) -> b 2752 0xF0 }; // (a, _, _) -> a 2753 for (uint i = 0; i < inputs.size(); i++) { 2754 eval_map.put(inputs.at(i), input_funcs[2-i]); 2755 } 2756 2757 for (uint i = 0; i < partition.size(); i++) { 2758 Node* n = partition.at(i); 2759 2760 uint func1 = 0, func2 = 0, func3 = 0; 2761 eval_operands(n, func1, func2, func3, eval_map); 2762 2763 switch (n->Opcode()) { 2764 case Op_OrV: 2765 assert(func3 == 0, "not binary"); 2766 res = func1 | func2; 2767 break; 2768 case Op_AndV: 2769 assert(func3 == 0, "not binary"); 2770 res = func1 & func2; 2771 break; 2772 case Op_XorV: 2773 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2774 assert(func2 == 0 && func3 == 0, "not unary"); 2775 res = (~func1) & 0xFF; 2776 } else { 2777 assert(func3 == 0, "not binary"); 2778 res = func1 ^ func2; 2779 } 2780 break; 2781 case Op_MacroLogicV: 2782 // Ordering of inputs may change during evaluation of sub-tree 2783 // containing MacroLogic node as a child node, thus a re-evaluation 2784 // makes sure that function is evaluated in context of current 2785 // inputs. 2786 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2787 break; 2788 2789 default: assert(false, "not supported: %s", n->Name()); 2790 } 2791 assert(res <= 0xFF, "invalid"); 2792 eval_map.put(n, res); 2793 } 2794 return res; 2795 } 2796 2797 // Criteria under which nodes gets packed into a macro logic node:- 2798 // 1) Parent and both child nodes are all unmasked or masked with 2799 // same predicates. 2800 // 2) Masked parent can be packed with left child if it is predicated 2801 // and both have same predicates. 2802 // 3) Masked parent can be packed with right child if its un-predicated 2803 // or has matching predication condition. 2804 // 4) An unmasked parent can be packed with an unmasked child. 2805 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2806 assert(partition.size() == 0, "not empty"); 2807 assert(inputs.size() == 0, "not empty"); 2808 if (is_vector_ternary_bitwise_op(n)) { 2809 return false; 2810 } 2811 2812 bool is_unary_op = is_vector_unary_bitwise_op(n); 2813 if (is_unary_op) { 2814 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2815 return false; // too few inputs 2816 } 2817 2818 bool pack_left_child = true; 2819 bool pack_right_child = true; 2820 2821 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2822 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2823 2824 int left_child_input_cnt = 0; 2825 int right_child_input_cnt = 0; 2826 2827 bool parent_is_predicated = n->is_predicated_vector(); 2828 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2829 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2830 2831 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2832 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2833 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2834 2835 do { 2836 if (pack_left_child && left_child_LOP && 2837 ((!parent_is_predicated && !left_child_predicated) || 2838 ((parent_is_predicated && left_child_predicated && 2839 parent_pred == left_child_pred)))) { 2840 partition.push(n->in(1)); 2841 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2842 } else { 2843 inputs.push(n->in(1)); 2844 left_child_input_cnt = 1; 2845 } 2846 2847 if (pack_right_child && right_child_LOP && 2848 (!right_child_predicated || 2849 (right_child_predicated && parent_is_predicated && 2850 parent_pred == right_child_pred))) { 2851 partition.push(n->in(2)); 2852 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2853 } else { 2854 inputs.push(n->in(2)); 2855 right_child_input_cnt = 1; 2856 } 2857 2858 if (inputs.size() > 3) { 2859 assert(partition.size() > 0, ""); 2860 inputs.clear(); 2861 partition.clear(); 2862 if (left_child_input_cnt > right_child_input_cnt) { 2863 pack_left_child = false; 2864 } else { 2865 pack_right_child = false; 2866 } 2867 } else { 2868 break; 2869 } 2870 } while(true); 2871 2872 if(partition.size()) { 2873 partition.push(n); 2874 } 2875 2876 return (partition.size() == 2 || partition.size() == 3) && 2877 (inputs.size() == 2 || inputs.size() == 3); 2878 } 2879 2880 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2881 assert(is_vector_bitwise_op(n), "not a root"); 2882 2883 visited.set(n->_idx); 2884 2885 // 1) Do a DFS walk over the logic cone. 2886 for (uint i = 1; i < n->req(); i++) { 2887 Node* in = n->in(i); 2888 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2889 process_logic_cone_root(igvn, in, visited); 2890 } 2891 } 2892 2893 // 2) Bottom up traversal: Merge node[s] with 2894 // the parent to form macro logic node. 2895 Unique_Node_List partition; 2896 Unique_Node_List inputs; 2897 if (compute_logic_cone(n, partition, inputs)) { 2898 const TypeVect* vt = n->bottom_type()->is_vect(); 2899 Node* pn = partition.at(partition.size() - 1); 2900 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2901 if (mask == nullptr || 2902 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2903 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2904 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 2905 igvn.replace_node(n, macro_logic); 2906 } 2907 } 2908 } 2909 2910 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2911 ResourceMark rm; 2912 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2913 Unique_Node_List list; 2914 collect_logic_cone_roots(list); 2915 2916 while (list.size() > 0) { 2917 Node* n = list.pop(); 2918 const TypeVect* vt = n->bottom_type()->is_vect(); 2919 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2920 if (supported) { 2921 VectorSet visited(comp_arena()); 2922 process_logic_cone_root(igvn, n, visited); 2923 } 2924 } 2925 } 2926 } 2927 2928 //------------------------------Code_Gen--------------------------------------- 2929 // Given a graph, generate code for it 2930 void Compile::Code_Gen() { 2931 if (failing()) { 2932 return; 2933 } 2934 2935 // Perform instruction selection. You might think we could reclaim Matcher 2936 // memory PDQ, but actually the Matcher is used in generating spill code. 2937 // Internals of the Matcher (including some VectorSets) must remain live 2938 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2939 // set a bit in reclaimed memory. 2940 2941 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2942 // nodes. Mapping is only valid at the root of each matched subtree. 2943 NOT_PRODUCT( verify_graph_edges(); ) 2944 2945 Matcher matcher; 2946 _matcher = &matcher; 2947 { 2948 TracePhase tp("matcher", &timers[_t_matcher]); 2949 matcher.match(); 2950 if (failing()) { 2951 return; 2952 } 2953 } 2954 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2955 // nodes. Mapping is only valid at the root of each matched subtree. 2956 NOT_PRODUCT( verify_graph_edges(); ) 2957 2958 // If you have too many nodes, or if matching has failed, bail out 2959 check_node_count(0, "out of nodes matching instructions"); 2960 if (failing()) { 2961 return; 2962 } 2963 2964 print_method(PHASE_MATCHING, 2); 2965 2966 // Build a proper-looking CFG 2967 PhaseCFG cfg(node_arena(), root(), matcher); 2968 _cfg = &cfg; 2969 { 2970 TracePhase tp("scheduler", &timers[_t_scheduler]); 2971 bool success = cfg.do_global_code_motion(); 2972 if (!success) { 2973 return; 2974 } 2975 2976 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2977 NOT_PRODUCT( verify_graph_edges(); ) 2978 cfg.verify(); 2979 } 2980 2981 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2982 _regalloc = ®alloc; 2983 { 2984 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2985 // Perform register allocation. After Chaitin, use-def chains are 2986 // no longer accurate (at spill code) and so must be ignored. 2987 // Node->LRG->reg mappings are still accurate. 2988 _regalloc->Register_Allocate(); 2989 2990 // Bail out if the allocator builds too many nodes 2991 if (failing()) { 2992 return; 2993 } 2994 2995 print_method(PHASE_REGISTER_ALLOCATION, 2); 2996 } 2997 2998 // Prior to register allocation we kept empty basic blocks in case the 2999 // the allocator needed a place to spill. After register allocation we 3000 // are not adding any new instructions. If any basic block is empty, we 3001 // can now safely remove it. 3002 { 3003 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 3004 cfg.remove_empty_blocks(); 3005 if (do_freq_based_layout()) { 3006 PhaseBlockLayout layout(cfg); 3007 } else { 3008 cfg.set_loop_alignment(); 3009 } 3010 cfg.fixup_flow(); 3011 cfg.remove_unreachable_blocks(); 3012 cfg.verify_dominator_tree(); 3013 print_method(PHASE_BLOCK_ORDERING, 3); 3014 } 3015 3016 // Apply peephole optimizations 3017 if( OptoPeephole ) { 3018 TracePhase tp("peephole", &timers[_t_peephole]); 3019 PhasePeephole peep( _regalloc, cfg); 3020 peep.do_transform(); 3021 print_method(PHASE_PEEPHOLE, 3); 3022 } 3023 3024 // Do late expand if CPU requires this. 3025 if (Matcher::require_postalloc_expand) { 3026 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 3027 cfg.postalloc_expand(_regalloc); 3028 print_method(PHASE_POSTALLOC_EXPAND, 3); 3029 } 3030 3031 // Convert Nodes to instruction bits in a buffer 3032 { 3033 TracePhase tp("output", &timers[_t_output]); 3034 PhaseOutput output; 3035 output.Output(); 3036 if (failing()) return; 3037 output.install(); 3038 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3039 } 3040 3041 // He's dead, Jim. 3042 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3043 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3044 } 3045 3046 //------------------------------Final_Reshape_Counts--------------------------- 3047 // This class defines counters to help identify when a method 3048 // may/must be executed using hardware with only 24-bit precision. 3049 struct Final_Reshape_Counts : public StackObj { 3050 int _call_count; // count non-inlined 'common' calls 3051 int _float_count; // count float ops requiring 24-bit precision 3052 int _double_count; // count double ops requiring more precision 3053 int _java_call_count; // count non-inlined 'java' calls 3054 int _inner_loop_count; // count loops which need alignment 3055 VectorSet _visited; // Visitation flags 3056 Node_List _tests; // Set of IfNodes & PCTableNodes 3057 3058 Final_Reshape_Counts() : 3059 _call_count(0), _float_count(0), _double_count(0), 3060 _java_call_count(0), _inner_loop_count(0) { } 3061 3062 void inc_call_count () { _call_count ++; } 3063 void inc_float_count () { _float_count ++; } 3064 void inc_double_count() { _double_count++; } 3065 void inc_java_call_count() { _java_call_count++; } 3066 void inc_inner_loop_count() { _inner_loop_count++; } 3067 3068 int get_call_count () const { return _call_count ; } 3069 int get_float_count () const { return _float_count ; } 3070 int get_double_count() const { return _double_count; } 3071 int get_java_call_count() const { return _java_call_count; } 3072 int get_inner_loop_count() const { return _inner_loop_count; } 3073 }; 3074 3075 // Eliminate trivially redundant StoreCMs and accumulate their 3076 // precedence edges. 3077 void Compile::eliminate_redundant_card_marks(Node* n) { 3078 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3079 if (n->in(MemNode::Address)->outcnt() > 1) { 3080 // There are multiple users of the same address so it might be 3081 // possible to eliminate some of the StoreCMs 3082 Node* mem = n->in(MemNode::Memory); 3083 Node* adr = n->in(MemNode::Address); 3084 Node* val = n->in(MemNode::ValueIn); 3085 Node* prev = n; 3086 bool done = false; 3087 // Walk the chain of StoreCMs eliminating ones that match. As 3088 // long as it's a chain of single users then the optimization is 3089 // safe. Eliminating partially redundant StoreCMs would require 3090 // cloning copies down the other paths. 3091 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3092 if (adr == mem->in(MemNode::Address) && 3093 val == mem->in(MemNode::ValueIn)) { 3094 // redundant StoreCM 3095 if (mem->req() > MemNode::OopStore) { 3096 // Hasn't been processed by this code yet. 3097 n->add_prec(mem->in(MemNode::OopStore)); 3098 } else { 3099 // Already converted to precedence edge 3100 for (uint i = mem->req(); i < mem->len(); i++) { 3101 // Accumulate any precedence edges 3102 if (mem->in(i) != nullptr) { 3103 n->add_prec(mem->in(i)); 3104 } 3105 } 3106 // Everything above this point has been processed. 3107 done = true; 3108 } 3109 // Eliminate the previous StoreCM 3110 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3111 assert(mem->outcnt() == 0, "should be dead"); 3112 mem->disconnect_inputs(this); 3113 } else { 3114 prev = mem; 3115 } 3116 mem = prev->in(MemNode::Memory); 3117 } 3118 } 3119 } 3120 3121 //------------------------------final_graph_reshaping_impl---------------------- 3122 // Implement items 1-5 from final_graph_reshaping below. 3123 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3124 3125 if ( n->outcnt() == 0 ) return; // dead node 3126 uint nop = n->Opcode(); 3127 3128 // Check for 2-input instruction with "last use" on right input. 3129 // Swap to left input. Implements item (2). 3130 if( n->req() == 3 && // two-input instruction 3131 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3132 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3133 n->in(2)->outcnt() == 1 &&// right use IS a last use 3134 !n->in(2)->is_Con() ) { // right use is not a constant 3135 // Check for commutative opcode 3136 switch( nop ) { 3137 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3138 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3139 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3140 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3141 case Op_AndL: case Op_XorL: case Op_OrL: 3142 case Op_AndI: case Op_XorI: case Op_OrI: { 3143 // Move "last use" input to left by swapping inputs 3144 n->swap_edges(1, 2); 3145 break; 3146 } 3147 default: 3148 break; 3149 } 3150 } 3151 3152 #ifdef ASSERT 3153 if( n->is_Mem() ) { 3154 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3155 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3156 // oop will be recorded in oop map if load crosses safepoint 3157 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3158 LoadNode::is_immutable_value(n->in(MemNode::Address)))), 3159 "raw memory operations should have control edge"); 3160 } 3161 if (n->is_MemBar()) { 3162 MemBarNode* mb = n->as_MemBar(); 3163 if (mb->trailing_store() || mb->trailing_load_store()) { 3164 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3165 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3166 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3167 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3168 } else if (mb->leading()) { 3169 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3170 } 3171 } 3172 #endif 3173 // Count FPU ops and common calls, implements item (3) 3174 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3175 if (!gc_handled) { 3176 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3177 } 3178 3179 // Collect CFG split points 3180 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3181 frc._tests.push(n); 3182 } 3183 } 3184 3185 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3186 switch( nop ) { 3187 // Count all float operations that may use FPU 3188 case Op_AddF: 3189 case Op_SubF: 3190 case Op_MulF: 3191 case Op_DivF: 3192 case Op_NegF: 3193 case Op_ModF: 3194 case Op_ConvI2F: 3195 case Op_ConF: 3196 case Op_CmpF: 3197 case Op_CmpF3: 3198 case Op_StoreF: 3199 case Op_LoadF: 3200 // case Op_ConvL2F: // longs are split into 32-bit halves 3201 frc.inc_float_count(); 3202 break; 3203 3204 case Op_ConvF2D: 3205 case Op_ConvD2F: 3206 frc.inc_float_count(); 3207 frc.inc_double_count(); 3208 break; 3209 3210 // Count all double operations that may use FPU 3211 case Op_AddD: 3212 case Op_SubD: 3213 case Op_MulD: 3214 case Op_DivD: 3215 case Op_NegD: 3216 case Op_ModD: 3217 case Op_ConvI2D: 3218 case Op_ConvD2I: 3219 // case Op_ConvL2D: // handled by leaf call 3220 // case Op_ConvD2L: // handled by leaf call 3221 case Op_ConD: 3222 case Op_CmpD: 3223 case Op_CmpD3: 3224 case Op_StoreD: 3225 case Op_LoadD: 3226 case Op_LoadD_unaligned: 3227 frc.inc_double_count(); 3228 break; 3229 case Op_Opaque1: // Remove Opaque Nodes before matching 3230 case Op_Opaque3: 3231 n->subsume_by(n->in(1), this); 3232 break; 3233 case Op_CallStaticJava: 3234 case Op_CallJava: 3235 case Op_CallDynamicJava: 3236 frc.inc_java_call_count(); // Count java call site; 3237 case Op_CallRuntime: 3238 case Op_CallLeaf: 3239 case Op_CallLeafVector: 3240 case Op_CallLeafNoFP: { 3241 assert (n->is_Call(), ""); 3242 CallNode *call = n->as_Call(); 3243 // Count call sites where the FP mode bit would have to be flipped. 3244 // Do not count uncommon runtime calls: 3245 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3246 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3247 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3248 frc.inc_call_count(); // Count the call site 3249 } else { // See if uncommon argument is shared 3250 Node *n = call->in(TypeFunc::Parms); 3251 int nop = n->Opcode(); 3252 // Clone shared simple arguments to uncommon calls, item (1). 3253 if (n->outcnt() > 1 && 3254 !n->is_Proj() && 3255 nop != Op_CreateEx && 3256 nop != Op_CheckCastPP && 3257 nop != Op_DecodeN && 3258 nop != Op_DecodeNKlass && 3259 !n->is_Mem() && 3260 !n->is_Phi()) { 3261 Node *x = n->clone(); 3262 call->set_req(TypeFunc::Parms, x); 3263 } 3264 } 3265 break; 3266 } 3267 3268 case Op_StoreCM: 3269 { 3270 // Convert OopStore dependence into precedence edge 3271 Node* prec = n->in(MemNode::OopStore); 3272 n->del_req(MemNode::OopStore); 3273 n->add_prec(prec); 3274 eliminate_redundant_card_marks(n); 3275 } 3276 3277 // fall through 3278 3279 case Op_StoreB: 3280 case Op_StoreC: 3281 case Op_StoreI: 3282 case Op_StoreL: 3283 case Op_CompareAndSwapB: 3284 case Op_CompareAndSwapS: 3285 case Op_CompareAndSwapI: 3286 case Op_CompareAndSwapL: 3287 case Op_CompareAndSwapP: 3288 case Op_CompareAndSwapN: 3289 case Op_WeakCompareAndSwapB: 3290 case Op_WeakCompareAndSwapS: 3291 case Op_WeakCompareAndSwapI: 3292 case Op_WeakCompareAndSwapL: 3293 case Op_WeakCompareAndSwapP: 3294 case Op_WeakCompareAndSwapN: 3295 case Op_CompareAndExchangeB: 3296 case Op_CompareAndExchangeS: 3297 case Op_CompareAndExchangeI: 3298 case Op_CompareAndExchangeL: 3299 case Op_CompareAndExchangeP: 3300 case Op_CompareAndExchangeN: 3301 case Op_GetAndAddS: 3302 case Op_GetAndAddB: 3303 case Op_GetAndAddI: 3304 case Op_GetAndAddL: 3305 case Op_GetAndSetS: 3306 case Op_GetAndSetB: 3307 case Op_GetAndSetI: 3308 case Op_GetAndSetL: 3309 case Op_GetAndSetP: 3310 case Op_GetAndSetN: 3311 case Op_StoreP: 3312 case Op_StoreN: 3313 case Op_StoreNKlass: 3314 case Op_LoadB: 3315 case Op_LoadUB: 3316 case Op_LoadUS: 3317 case Op_LoadI: 3318 case Op_LoadKlass: 3319 case Op_LoadNKlass: 3320 case Op_LoadL: 3321 case Op_LoadL_unaligned: 3322 case Op_LoadP: 3323 case Op_LoadN: 3324 case Op_LoadRange: 3325 case Op_LoadS: 3326 break; 3327 3328 case Op_AddP: { // Assert sane base pointers 3329 Node *addp = n->in(AddPNode::Address); 3330 assert( !addp->is_AddP() || 3331 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3332 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3333 "Base pointers must match (addp %u)", addp->_idx ); 3334 #ifdef _LP64 3335 if ((UseCompressedOops || UseCompressedClassPointers) && 3336 addp->Opcode() == Op_ConP && 3337 addp == n->in(AddPNode::Base) && 3338 n->in(AddPNode::Offset)->is_Con()) { 3339 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3340 // on the platform and on the compressed oops mode. 3341 // Use addressing with narrow klass to load with offset on x86. 3342 // Some platforms can use the constant pool to load ConP. 3343 // Do this transformation here since IGVN will convert ConN back to ConP. 3344 const Type* t = addp->bottom_type(); 3345 bool is_oop = t->isa_oopptr() != nullptr; 3346 bool is_klass = t->isa_klassptr() != nullptr; 3347 3348 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3349 (is_klass && Matcher::const_klass_prefer_decode())) { 3350 Node* nn = nullptr; 3351 3352 int op = is_oop ? Op_ConN : Op_ConNKlass; 3353 3354 // Look for existing ConN node of the same exact type. 3355 Node* r = root(); 3356 uint cnt = r->outcnt(); 3357 for (uint i = 0; i < cnt; i++) { 3358 Node* m = r->raw_out(i); 3359 if (m!= nullptr && m->Opcode() == op && 3360 m->bottom_type()->make_ptr() == t) { 3361 nn = m; 3362 break; 3363 } 3364 } 3365 if (nn != nullptr) { 3366 // Decode a narrow oop to match address 3367 // [R12 + narrow_oop_reg<<3 + offset] 3368 if (is_oop) { 3369 nn = new DecodeNNode(nn, t); 3370 } else { 3371 nn = new DecodeNKlassNode(nn, t); 3372 } 3373 // Check for succeeding AddP which uses the same Base. 3374 // Otherwise we will run into the assertion above when visiting that guy. 3375 for (uint i = 0; i < n->outcnt(); ++i) { 3376 Node *out_i = n->raw_out(i); 3377 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3378 out_i->set_req(AddPNode::Base, nn); 3379 #ifdef ASSERT 3380 for (uint j = 0; j < out_i->outcnt(); ++j) { 3381 Node *out_j = out_i->raw_out(j); 3382 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3383 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3384 } 3385 #endif 3386 } 3387 } 3388 n->set_req(AddPNode::Base, nn); 3389 n->set_req(AddPNode::Address, nn); 3390 if (addp->outcnt() == 0) { 3391 addp->disconnect_inputs(this); 3392 } 3393 } 3394 } 3395 } 3396 #endif 3397 break; 3398 } 3399 3400 case Op_CastPP: { 3401 // Remove CastPP nodes to gain more freedom during scheduling but 3402 // keep the dependency they encode as control or precedence edges 3403 // (if control is set already) on memory operations. Some CastPP 3404 // nodes don't have a control (don't carry a dependency): skip 3405 // those. 3406 if (n->in(0) != nullptr) { 3407 ResourceMark rm; 3408 Unique_Node_List wq; 3409 wq.push(n); 3410 for (uint next = 0; next < wq.size(); ++next) { 3411 Node *m = wq.at(next); 3412 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3413 Node* use = m->fast_out(i); 3414 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3415 use->ensure_control_or_add_prec(n->in(0)); 3416 } else { 3417 switch(use->Opcode()) { 3418 case Op_AddP: 3419 case Op_DecodeN: 3420 case Op_DecodeNKlass: 3421 case Op_CheckCastPP: 3422 case Op_CastPP: 3423 wq.push(use); 3424 break; 3425 } 3426 } 3427 } 3428 } 3429 } 3430 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3431 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3432 Node* in1 = n->in(1); 3433 const Type* t = n->bottom_type(); 3434 Node* new_in1 = in1->clone(); 3435 new_in1->as_DecodeN()->set_type(t); 3436 3437 if (!Matcher::narrow_oop_use_complex_address()) { 3438 // 3439 // x86, ARM and friends can handle 2 adds in addressing mode 3440 // and Matcher can fold a DecodeN node into address by using 3441 // a narrow oop directly and do implicit null check in address: 3442 // 3443 // [R12 + narrow_oop_reg<<3 + offset] 3444 // NullCheck narrow_oop_reg 3445 // 3446 // On other platforms (Sparc) we have to keep new DecodeN node and 3447 // use it to do implicit null check in address: 3448 // 3449 // decode_not_null narrow_oop_reg, base_reg 3450 // [base_reg + offset] 3451 // NullCheck base_reg 3452 // 3453 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3454 // to keep the information to which null check the new DecodeN node 3455 // corresponds to use it as value in implicit_null_check(). 3456 // 3457 new_in1->set_req(0, n->in(0)); 3458 } 3459 3460 n->subsume_by(new_in1, this); 3461 if (in1->outcnt() == 0) { 3462 in1->disconnect_inputs(this); 3463 } 3464 } else { 3465 n->subsume_by(n->in(1), this); 3466 if (n->outcnt() == 0) { 3467 n->disconnect_inputs(this); 3468 } 3469 } 3470 break; 3471 } 3472 #ifdef _LP64 3473 case Op_CmpP: 3474 // Do this transformation here to preserve CmpPNode::sub() and 3475 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3476 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3477 Node* in1 = n->in(1); 3478 Node* in2 = n->in(2); 3479 if (!in1->is_DecodeNarrowPtr()) { 3480 in2 = in1; 3481 in1 = n->in(2); 3482 } 3483 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3484 3485 Node* new_in2 = nullptr; 3486 if (in2->is_DecodeNarrowPtr()) { 3487 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3488 new_in2 = in2->in(1); 3489 } else if (in2->Opcode() == Op_ConP) { 3490 const Type* t = in2->bottom_type(); 3491 if (t == TypePtr::NULL_PTR) { 3492 assert(in1->is_DecodeN(), "compare klass to null?"); 3493 // Don't convert CmpP null check into CmpN if compressed 3494 // oops implicit null check is not generated. 3495 // This will allow to generate normal oop implicit null check. 3496 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3497 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3498 // 3499 // This transformation together with CastPP transformation above 3500 // will generated code for implicit null checks for compressed oops. 3501 // 3502 // The original code after Optimize() 3503 // 3504 // LoadN memory, narrow_oop_reg 3505 // decode narrow_oop_reg, base_reg 3506 // CmpP base_reg, nullptr 3507 // CastPP base_reg // NotNull 3508 // Load [base_reg + offset], val_reg 3509 // 3510 // after these transformations will be 3511 // 3512 // LoadN memory, narrow_oop_reg 3513 // CmpN narrow_oop_reg, nullptr 3514 // decode_not_null narrow_oop_reg, base_reg 3515 // Load [base_reg + offset], val_reg 3516 // 3517 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3518 // since narrow oops can be used in debug info now (see the code in 3519 // final_graph_reshaping_walk()). 3520 // 3521 // At the end the code will be matched to 3522 // on x86: 3523 // 3524 // Load_narrow_oop memory, narrow_oop_reg 3525 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3526 // NullCheck narrow_oop_reg 3527 // 3528 // and on sparc: 3529 // 3530 // Load_narrow_oop memory, narrow_oop_reg 3531 // decode_not_null narrow_oop_reg, base_reg 3532 // Load [base_reg + offset], val_reg 3533 // NullCheck base_reg 3534 // 3535 } else if (t->isa_oopptr()) { 3536 new_in2 = ConNode::make(t->make_narrowoop()); 3537 } else if (t->isa_klassptr()) { 3538 new_in2 = ConNode::make(t->make_narrowklass()); 3539 } 3540 } 3541 if (new_in2 != nullptr) { 3542 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3543 n->subsume_by(cmpN, this); 3544 if (in1->outcnt() == 0) { 3545 in1->disconnect_inputs(this); 3546 } 3547 if (in2->outcnt() == 0) { 3548 in2->disconnect_inputs(this); 3549 } 3550 } 3551 } 3552 break; 3553 3554 case Op_DecodeN: 3555 case Op_DecodeNKlass: 3556 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3557 // DecodeN could be pinned when it can't be fold into 3558 // an address expression, see the code for Op_CastPP above. 3559 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3560 break; 3561 3562 case Op_EncodeP: 3563 case Op_EncodePKlass: { 3564 Node* in1 = n->in(1); 3565 if (in1->is_DecodeNarrowPtr()) { 3566 n->subsume_by(in1->in(1), this); 3567 } else if (in1->Opcode() == Op_ConP) { 3568 const Type* t = in1->bottom_type(); 3569 if (t == TypePtr::NULL_PTR) { 3570 assert(t->isa_oopptr(), "null klass?"); 3571 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3572 } else if (t->isa_oopptr()) { 3573 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3574 } else if (t->isa_klassptr()) { 3575 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3576 } 3577 } 3578 if (in1->outcnt() == 0) { 3579 in1->disconnect_inputs(this); 3580 } 3581 break; 3582 } 3583 3584 case Op_Proj: { 3585 if (OptimizeStringConcat || IncrementalInline) { 3586 ProjNode* proj = n->as_Proj(); 3587 if (proj->_is_io_use) { 3588 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3589 // Separate projections were used for the exception path which 3590 // are normally removed by a late inline. If it wasn't inlined 3591 // then they will hang around and should just be replaced with 3592 // the original one. Merge them. 3593 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3594 if (non_io_proj != nullptr) { 3595 proj->subsume_by(non_io_proj , this); 3596 } 3597 } 3598 } 3599 break; 3600 } 3601 3602 case Op_Phi: 3603 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3604 // The EncodeP optimization may create Phi with the same edges 3605 // for all paths. It is not handled well by Register Allocator. 3606 Node* unique_in = n->in(1); 3607 assert(unique_in != nullptr, ""); 3608 uint cnt = n->req(); 3609 for (uint i = 2; i < cnt; i++) { 3610 Node* m = n->in(i); 3611 assert(m != nullptr, ""); 3612 if (unique_in != m) 3613 unique_in = nullptr; 3614 } 3615 if (unique_in != nullptr) { 3616 n->subsume_by(unique_in, this); 3617 } 3618 } 3619 break; 3620 3621 #endif 3622 3623 #ifdef ASSERT 3624 case Op_CastII: 3625 // Verify that all range check dependent CastII nodes were removed. 3626 if (n->isa_CastII()->has_range_check()) { 3627 n->dump(3); 3628 assert(false, "Range check dependent CastII node was not removed"); 3629 } 3630 break; 3631 #endif 3632 3633 case Op_ModI: 3634 if (UseDivMod) { 3635 // Check if a%b and a/b both exist 3636 Node* d = n->find_similar(Op_DivI); 3637 if (d) { 3638 // Replace them with a fused divmod if supported 3639 if (Matcher::has_match_rule(Op_DivModI)) { 3640 DivModINode* divmod = DivModINode::make(n); 3641 d->subsume_by(divmod->div_proj(), this); 3642 n->subsume_by(divmod->mod_proj(), this); 3643 } else { 3644 // replace a%b with a-((a/b)*b) 3645 Node* mult = new MulINode(d, d->in(2)); 3646 Node* sub = new SubINode(d->in(1), mult); 3647 n->subsume_by(sub, this); 3648 } 3649 } 3650 } 3651 break; 3652 3653 case Op_ModL: 3654 if (UseDivMod) { 3655 // Check if a%b and a/b both exist 3656 Node* d = n->find_similar(Op_DivL); 3657 if (d) { 3658 // Replace them with a fused divmod if supported 3659 if (Matcher::has_match_rule(Op_DivModL)) { 3660 DivModLNode* divmod = DivModLNode::make(n); 3661 d->subsume_by(divmod->div_proj(), this); 3662 n->subsume_by(divmod->mod_proj(), this); 3663 } else { 3664 // replace a%b with a-((a/b)*b) 3665 Node* mult = new MulLNode(d, d->in(2)); 3666 Node* sub = new SubLNode(d->in(1), mult); 3667 n->subsume_by(sub, this); 3668 } 3669 } 3670 } 3671 break; 3672 3673 case Op_UModI: 3674 if (UseDivMod) { 3675 // Check if a%b and a/b both exist 3676 Node* d = n->find_similar(Op_UDivI); 3677 if (d) { 3678 // Replace them with a fused unsigned divmod if supported 3679 if (Matcher::has_match_rule(Op_UDivModI)) { 3680 UDivModINode* divmod = UDivModINode::make(n); 3681 d->subsume_by(divmod->div_proj(), this); 3682 n->subsume_by(divmod->mod_proj(), this); 3683 } else { 3684 // replace a%b with a-((a/b)*b) 3685 Node* mult = new MulINode(d, d->in(2)); 3686 Node* sub = new SubINode(d->in(1), mult); 3687 n->subsume_by(sub, this); 3688 } 3689 } 3690 } 3691 break; 3692 3693 case Op_UModL: 3694 if (UseDivMod) { 3695 // Check if a%b and a/b both exist 3696 Node* d = n->find_similar(Op_UDivL); 3697 if (d) { 3698 // Replace them with a fused unsigned divmod if supported 3699 if (Matcher::has_match_rule(Op_UDivModL)) { 3700 UDivModLNode* divmod = UDivModLNode::make(n); 3701 d->subsume_by(divmod->div_proj(), this); 3702 n->subsume_by(divmod->mod_proj(), this); 3703 } else { 3704 // replace a%b with a-((a/b)*b) 3705 Node* mult = new MulLNode(d, d->in(2)); 3706 Node* sub = new SubLNode(d->in(1), mult); 3707 n->subsume_by(sub, this); 3708 } 3709 } 3710 } 3711 break; 3712 3713 case Op_LoadVector: 3714 case Op_StoreVector: 3715 #ifdef ASSERT 3716 // Add VerifyVectorAlignment node between adr and load / store. 3717 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) { 3718 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() : 3719 n->as_StoreVector()->must_verify_alignment(); 3720 if (must_verify_alignment) { 3721 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() : 3722 n->as_StoreVector()->memory_size(); 3723 // The memory access should be aligned to the vector width in bytes. 3724 // However, the underlying array is possibly less well aligned, but at least 3725 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in 3726 // a loop we can expect at least the following alignment: 3727 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes); 3728 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range"); 3729 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2"); 3730 // Create mask from alignment. e.g. 0b1000 -> 0b0111 3731 jlong mask = guaranteed_alignment - 1; 3732 Node* mask_con = ConLNode::make(mask); 3733 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con); 3734 n->set_req(MemNode::Address, va); 3735 } 3736 } 3737 #endif 3738 break; 3739 3740 case Op_LoadVectorGather: 3741 case Op_StoreVectorScatter: 3742 case Op_LoadVectorGatherMasked: 3743 case Op_StoreVectorScatterMasked: 3744 case Op_VectorCmpMasked: 3745 case Op_VectorMaskGen: 3746 case Op_LoadVectorMasked: 3747 case Op_StoreVectorMasked: 3748 break; 3749 3750 case Op_AddReductionVI: 3751 case Op_AddReductionVL: 3752 case Op_AddReductionVF: 3753 case Op_AddReductionVD: 3754 case Op_MulReductionVI: 3755 case Op_MulReductionVL: 3756 case Op_MulReductionVF: 3757 case Op_MulReductionVD: 3758 case Op_MinReductionV: 3759 case Op_MaxReductionV: 3760 case Op_AndReductionV: 3761 case Op_OrReductionV: 3762 case Op_XorReductionV: 3763 break; 3764 3765 case Op_PackB: 3766 case Op_PackS: 3767 case Op_PackI: 3768 case Op_PackF: 3769 case Op_PackL: 3770 case Op_PackD: 3771 if (n->req()-1 > 2) { 3772 // Replace many operand PackNodes with a binary tree for matching 3773 PackNode* p = (PackNode*) n; 3774 Node* btp = p->binary_tree_pack(1, n->req()); 3775 n->subsume_by(btp, this); 3776 } 3777 break; 3778 case Op_Loop: 3779 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3780 case Op_CountedLoop: 3781 case Op_LongCountedLoop: 3782 case Op_OuterStripMinedLoop: 3783 if (n->as_Loop()->is_inner_loop()) { 3784 frc.inc_inner_loop_count(); 3785 } 3786 n->as_Loop()->verify_strip_mined(0); 3787 break; 3788 case Op_LShiftI: 3789 case Op_RShiftI: 3790 case Op_URShiftI: 3791 case Op_LShiftL: 3792 case Op_RShiftL: 3793 case Op_URShiftL: 3794 if (Matcher::need_masked_shift_count) { 3795 // The cpu's shift instructions don't restrict the count to the 3796 // lower 5/6 bits. We need to do the masking ourselves. 3797 Node* in2 = n->in(2); 3798 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3799 const TypeInt* t = in2->find_int_type(); 3800 if (t != nullptr && t->is_con()) { 3801 juint shift = t->get_con(); 3802 if (shift > mask) { // Unsigned cmp 3803 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3804 } 3805 } else { 3806 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3807 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3808 n->set_req(2, shift); 3809 } 3810 } 3811 if (in2->outcnt() == 0) { // Remove dead node 3812 in2->disconnect_inputs(this); 3813 } 3814 } 3815 break; 3816 case Op_MemBarStoreStore: 3817 case Op_MemBarRelease: 3818 // Break the link with AllocateNode: it is no longer useful and 3819 // confuses register allocation. 3820 if (n->req() > MemBarNode::Precedent) { 3821 n->set_req(MemBarNode::Precedent, top()); 3822 } 3823 break; 3824 case Op_MemBarAcquire: { 3825 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3826 // At parse time, the trailing MemBarAcquire for a volatile load 3827 // is created with an edge to the load. After optimizations, 3828 // that input may be a chain of Phis. If those phis have no 3829 // other use, then the MemBarAcquire keeps them alive and 3830 // register allocation can be confused. 3831 dead_nodes.push(n->in(MemBarNode::Precedent)); 3832 n->set_req(MemBarNode::Precedent, top()); 3833 } 3834 break; 3835 } 3836 case Op_Blackhole: 3837 break; 3838 case Op_RangeCheck: { 3839 RangeCheckNode* rc = n->as_RangeCheck(); 3840 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3841 n->subsume_by(iff, this); 3842 frc._tests.push(iff); 3843 break; 3844 } 3845 case Op_ConvI2L: { 3846 if (!Matcher::convi2l_type_required) { 3847 // Code generation on some platforms doesn't need accurate 3848 // ConvI2L types. Widening the type can help remove redundant 3849 // address computations. 3850 n->as_Type()->set_type(TypeLong::INT); 3851 ResourceMark rm; 3852 Unique_Node_List wq; 3853 wq.push(n); 3854 for (uint next = 0; next < wq.size(); next++) { 3855 Node *m = wq.at(next); 3856 3857 for(;;) { 3858 // Loop over all nodes with identical inputs edges as m 3859 Node* k = m->find_similar(m->Opcode()); 3860 if (k == nullptr) { 3861 break; 3862 } 3863 // Push their uses so we get a chance to remove node made 3864 // redundant 3865 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3866 Node* u = k->fast_out(i); 3867 if (u->Opcode() == Op_LShiftL || 3868 u->Opcode() == Op_AddL || 3869 u->Opcode() == Op_SubL || 3870 u->Opcode() == Op_AddP) { 3871 wq.push(u); 3872 } 3873 } 3874 // Replace all nodes with identical edges as m with m 3875 k->subsume_by(m, this); 3876 } 3877 } 3878 } 3879 break; 3880 } 3881 case Op_CmpUL: { 3882 if (!Matcher::has_match_rule(Op_CmpUL)) { 3883 // No support for unsigned long comparisons 3884 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3885 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3886 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3887 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3888 Node* andl = new AndLNode(orl, remove_sign_mask); 3889 Node* cmp = new CmpLNode(andl, n->in(2)); 3890 n->subsume_by(cmp, this); 3891 } 3892 break; 3893 } 3894 default: 3895 assert(!n->is_Call(), ""); 3896 assert(!n->is_Mem(), ""); 3897 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3898 break; 3899 } 3900 } 3901 3902 //------------------------------final_graph_reshaping_walk--------------------- 3903 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3904 // requires that the walk visits a node's inputs before visiting the node. 3905 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3906 Unique_Node_List sfpt; 3907 3908 frc._visited.set(root->_idx); // first, mark node as visited 3909 uint cnt = root->req(); 3910 Node *n = root; 3911 uint i = 0; 3912 while (true) { 3913 if (i < cnt) { 3914 // Place all non-visited non-null inputs onto stack 3915 Node* m = n->in(i); 3916 ++i; 3917 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3918 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3919 // compute worst case interpreter size in case of a deoptimization 3920 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3921 3922 sfpt.push(m); 3923 } 3924 cnt = m->req(); 3925 nstack.push(n, i); // put on stack parent and next input's index 3926 n = m; 3927 i = 0; 3928 } 3929 } else { 3930 // Now do post-visit work 3931 final_graph_reshaping_impl(n, frc, dead_nodes); 3932 if (nstack.is_empty()) 3933 break; // finished 3934 n = nstack.node(); // Get node from stack 3935 cnt = n->req(); 3936 i = nstack.index(); 3937 nstack.pop(); // Shift to the next node on stack 3938 } 3939 } 3940 3941 // Skip next transformation if compressed oops are not used. 3942 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3943 (!UseCompressedOops && !UseCompressedClassPointers)) 3944 return; 3945 3946 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3947 // It could be done for an uncommon traps or any safepoints/calls 3948 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3949 while (sfpt.size() > 0) { 3950 n = sfpt.pop(); 3951 JVMState *jvms = n->as_SafePoint()->jvms(); 3952 assert(jvms != nullptr, "sanity"); 3953 int start = jvms->debug_start(); 3954 int end = n->req(); 3955 bool is_uncommon = (n->is_CallStaticJava() && 3956 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3957 for (int j = start; j < end; j++) { 3958 Node* in = n->in(j); 3959 if (in->is_DecodeNarrowPtr()) { 3960 bool safe_to_skip = true; 3961 if (!is_uncommon ) { 3962 // Is it safe to skip? 3963 for (uint i = 0; i < in->outcnt(); i++) { 3964 Node* u = in->raw_out(i); 3965 if (!u->is_SafePoint() || 3966 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3967 safe_to_skip = false; 3968 } 3969 } 3970 } 3971 if (safe_to_skip) { 3972 n->set_req(j, in->in(1)); 3973 } 3974 if (in->outcnt() == 0) { 3975 in->disconnect_inputs(this); 3976 } 3977 } 3978 } 3979 } 3980 } 3981 3982 //------------------------------final_graph_reshaping-------------------------- 3983 // Final Graph Reshaping. 3984 // 3985 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3986 // and not commoned up and forced early. Must come after regular 3987 // optimizations to avoid GVN undoing the cloning. Clone constant 3988 // inputs to Loop Phis; these will be split by the allocator anyways. 3989 // Remove Opaque nodes. 3990 // (2) Move last-uses by commutative operations to the left input to encourage 3991 // Intel update-in-place two-address operations and better register usage 3992 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3993 // calls canonicalizing them back. 3994 // (3) Count the number of double-precision FP ops, single-precision FP ops 3995 // and call sites. On Intel, we can get correct rounding either by 3996 // forcing singles to memory (requires extra stores and loads after each 3997 // FP bytecode) or we can set a rounding mode bit (requires setting and 3998 // clearing the mode bit around call sites). The mode bit is only used 3999 // if the relative frequency of single FP ops to calls is low enough. 4000 // This is a key transform for SPEC mpeg_audio. 4001 // (4) Detect infinite loops; blobs of code reachable from above but not 4002 // below. Several of the Code_Gen algorithms fail on such code shapes, 4003 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4004 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4005 // Detection is by looking for IfNodes where only 1 projection is 4006 // reachable from below or CatchNodes missing some targets. 4007 // (5) Assert for insane oop offsets in debug mode. 4008 4009 bool Compile::final_graph_reshaping() { 4010 // an infinite loop may have been eliminated by the optimizer, 4011 // in which case the graph will be empty. 4012 if (root()->req() == 1) { 4013 // Do not compile method that is only a trivial infinite loop, 4014 // since the content of the loop may have been eliminated. 4015 record_method_not_compilable("trivial infinite loop"); 4016 return true; 4017 } 4018 4019 // Expensive nodes have their control input set to prevent the GVN 4020 // from freely commoning them. There's no GVN beyond this point so 4021 // no need to keep the control input. We want the expensive nodes to 4022 // be freely moved to the least frequent code path by gcm. 4023 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4024 for (int i = 0; i < expensive_count(); i++) { 4025 _expensive_nodes.at(i)->set_req(0, nullptr); 4026 } 4027 4028 Final_Reshape_Counts frc; 4029 4030 // Visit everybody reachable! 4031 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4032 Node_Stack nstack(live_nodes() >> 1); 4033 Unique_Node_List dead_nodes; 4034 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 4035 4036 // Check for unreachable (from below) code (i.e., infinite loops). 4037 for( uint i = 0; i < frc._tests.size(); i++ ) { 4038 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4039 // Get number of CFG targets. 4040 // Note that PCTables include exception targets after calls. 4041 uint required_outcnt = n->required_outcnt(); 4042 if (n->outcnt() != required_outcnt) { 4043 // Check for a few special cases. Rethrow Nodes never take the 4044 // 'fall-thru' path, so expected kids is 1 less. 4045 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4046 if (n->in(0)->in(0)->is_Call()) { 4047 CallNode* call = n->in(0)->in(0)->as_Call(); 4048 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4049 required_outcnt--; // Rethrow always has 1 less kid 4050 } else if (call->req() > TypeFunc::Parms && 4051 call->is_CallDynamicJava()) { 4052 // Check for null receiver. In such case, the optimizer has 4053 // detected that the virtual call will always result in a null 4054 // pointer exception. The fall-through projection of this CatchNode 4055 // will not be populated. 4056 Node* arg0 = call->in(TypeFunc::Parms); 4057 if (arg0->is_Type() && 4058 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4059 required_outcnt--; 4060 } 4061 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4062 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4063 // Check for illegal array length. In such case, the optimizer has 4064 // detected that the allocation attempt will always result in an 4065 // exception. There is no fall-through projection of this CatchNode . 4066 assert(call->is_CallStaticJava(), "static call expected"); 4067 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4068 uint valid_length_test_input = call->req() - 1; 4069 Node* valid_length_test = call->in(valid_length_test_input); 4070 call->del_req(valid_length_test_input); 4071 if (valid_length_test->find_int_con(1) == 0) { 4072 required_outcnt--; 4073 } 4074 dead_nodes.push(valid_length_test); 4075 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4076 continue; 4077 } 4078 } 4079 } 4080 4081 // Recheck with a better notion of 'required_outcnt' 4082 if (n->outcnt() != required_outcnt) { 4083 record_method_not_compilable("malformed control flow"); 4084 return true; // Not all targets reachable! 4085 } 4086 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4087 CallNode* call = n->in(0)->in(0)->as_Call(); 4088 if (call->entry_point() == OptoRuntime::new_array_Java() || 4089 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4090 assert(call->is_CallStaticJava(), "static call expected"); 4091 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4092 uint valid_length_test_input = call->req() - 1; 4093 dead_nodes.push(call->in(valid_length_test_input)); 4094 call->del_req(valid_length_test_input); // valid length test useless now 4095 } 4096 } 4097 // Check that I actually visited all kids. Unreached kids 4098 // must be infinite loops. 4099 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4100 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4101 record_method_not_compilable("infinite loop"); 4102 return true; // Found unvisited kid; must be unreach 4103 } 4104 4105 // Here so verification code in final_graph_reshaping_walk() 4106 // always see an OuterStripMinedLoopEnd 4107 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4108 IfNode* init_iff = n->as_If(); 4109 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4110 n->subsume_by(iff, this); 4111 } 4112 } 4113 4114 while (dead_nodes.size() > 0) { 4115 Node* m = dead_nodes.pop(); 4116 if (m->outcnt() == 0 && m != top()) { 4117 for (uint j = 0; j < m->req(); j++) { 4118 Node* in = m->in(j); 4119 if (in != nullptr) { 4120 dead_nodes.push(in); 4121 } 4122 } 4123 m->disconnect_inputs(this); 4124 } 4125 } 4126 4127 #ifdef IA32 4128 // If original bytecodes contained a mixture of floats and doubles 4129 // check if the optimizer has made it homogeneous, item (3). 4130 if (UseSSE == 0 && 4131 frc.get_float_count() > 32 && 4132 frc.get_double_count() == 0 && 4133 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4134 set_24_bit_selection_and_mode(false, true); 4135 } 4136 #endif // IA32 4137 4138 set_java_calls(frc.get_java_call_count()); 4139 set_inner_loops(frc.get_inner_loop_count()); 4140 4141 // No infinite loops, no reason to bail out. 4142 return false; 4143 } 4144 4145 //-----------------------------too_many_traps---------------------------------- 4146 // Report if there are too many traps at the current method and bci. 4147 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4148 bool Compile::too_many_traps(ciMethod* method, 4149 int bci, 4150 Deoptimization::DeoptReason reason) { 4151 ciMethodData* md = method->method_data(); 4152 if (md->is_empty()) { 4153 // Assume the trap has not occurred, or that it occurred only 4154 // because of a transient condition during start-up in the interpreter. 4155 return false; 4156 } 4157 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4158 if (md->has_trap_at(bci, m, reason) != 0) { 4159 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4160 // Also, if there are multiple reasons, or if there is no per-BCI record, 4161 // assume the worst. 4162 if (log()) 4163 log()->elem("observe trap='%s' count='%d'", 4164 Deoptimization::trap_reason_name(reason), 4165 md->trap_count(reason)); 4166 return true; 4167 } else { 4168 // Ignore method/bci and see if there have been too many globally. 4169 return too_many_traps(reason, md); 4170 } 4171 } 4172 4173 // Less-accurate variant which does not require a method and bci. 4174 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4175 ciMethodData* logmd) { 4176 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4177 // Too many traps globally. 4178 // Note that we use cumulative trap_count, not just md->trap_count. 4179 if (log()) { 4180 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4181 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4182 Deoptimization::trap_reason_name(reason), 4183 mcount, trap_count(reason)); 4184 } 4185 return true; 4186 } else { 4187 // The coast is clear. 4188 return false; 4189 } 4190 } 4191 4192 //--------------------------too_many_recompiles-------------------------------- 4193 // Report if there are too many recompiles at the current method and bci. 4194 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4195 // Is not eager to return true, since this will cause the compiler to use 4196 // Action_none for a trap point, to avoid too many recompilations. 4197 bool Compile::too_many_recompiles(ciMethod* method, 4198 int bci, 4199 Deoptimization::DeoptReason reason) { 4200 ciMethodData* md = method->method_data(); 4201 if (md->is_empty()) { 4202 // Assume the trap has not occurred, or that it occurred only 4203 // because of a transient condition during start-up in the interpreter. 4204 return false; 4205 } 4206 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4207 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4208 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4209 Deoptimization::DeoptReason per_bc_reason 4210 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4211 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4212 if ((per_bc_reason == Deoptimization::Reason_none 4213 || md->has_trap_at(bci, m, reason) != 0) 4214 // The trap frequency measure we care about is the recompile count: 4215 && md->trap_recompiled_at(bci, m) 4216 && md->overflow_recompile_count() >= bc_cutoff) { 4217 // Do not emit a trap here if it has already caused recompilations. 4218 // Also, if there are multiple reasons, or if there is no per-BCI record, 4219 // assume the worst. 4220 if (log()) 4221 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4222 Deoptimization::trap_reason_name(reason), 4223 md->trap_count(reason), 4224 md->overflow_recompile_count()); 4225 return true; 4226 } else if (trap_count(reason) != 0 4227 && decompile_count() >= m_cutoff) { 4228 // Too many recompiles globally, and we have seen this sort of trap. 4229 // Use cumulative decompile_count, not just md->decompile_count. 4230 if (log()) 4231 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4232 Deoptimization::trap_reason_name(reason), 4233 md->trap_count(reason), trap_count(reason), 4234 md->decompile_count(), decompile_count()); 4235 return true; 4236 } else { 4237 // The coast is clear. 4238 return false; 4239 } 4240 } 4241 4242 // Compute when not to trap. Used by matching trap based nodes and 4243 // NullCheck optimization. 4244 void Compile::set_allowed_deopt_reasons() { 4245 _allowed_reasons = 0; 4246 if (is_method_compilation()) { 4247 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4248 assert(rs < BitsPerInt, "recode bit map"); 4249 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4250 _allowed_reasons |= nth_bit(rs); 4251 } 4252 } 4253 } 4254 } 4255 4256 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4257 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4258 } 4259 4260 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4261 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4262 } 4263 4264 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4265 if (holder->is_initialized()) { 4266 return false; 4267 } 4268 if (holder->is_being_initialized()) { 4269 if (accessing_method->holder() == holder) { 4270 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4271 // <init>, or a static method. In all those cases, there was an initialization 4272 // barrier on the holder klass passed. 4273 if (accessing_method->is_static_initializer() || 4274 accessing_method->is_object_initializer() || 4275 accessing_method->is_static()) { 4276 return false; 4277 } 4278 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4279 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4280 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4281 // child class can become fully initialized while its parent class is still being initialized. 4282 if (accessing_method->is_static_initializer()) { 4283 return false; 4284 } 4285 } 4286 ciMethod* root = method(); // the root method of compilation 4287 if (root != accessing_method) { 4288 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4289 } 4290 } 4291 return true; 4292 } 4293 4294 #ifndef PRODUCT 4295 //------------------------------verify_bidirectional_edges--------------------- 4296 // For each input edge to a node (ie - for each Use-Def edge), verify that 4297 // there is a corresponding Def-Use edge. 4298 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4299 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4300 uint stack_size = live_nodes() >> 4; 4301 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4302 nstack.push(_root); 4303 4304 while (nstack.size() > 0) { 4305 Node* n = nstack.pop(); 4306 if (visited.member(n)) { 4307 continue; 4308 } 4309 visited.push(n); 4310 4311 // Walk over all input edges, checking for correspondence 4312 uint length = n->len(); 4313 for (uint i = 0; i < length; i++) { 4314 Node* in = n->in(i); 4315 if (in != nullptr && !visited.member(in)) { 4316 nstack.push(in); // Put it on stack 4317 } 4318 if (in != nullptr && !in->is_top()) { 4319 // Count instances of `next` 4320 int cnt = 0; 4321 for (uint idx = 0; idx < in->_outcnt; idx++) { 4322 if (in->_out[idx] == n) { 4323 cnt++; 4324 } 4325 } 4326 assert(cnt > 0, "Failed to find Def-Use edge."); 4327 // Check for duplicate edges 4328 // walk the input array downcounting the input edges to n 4329 for (uint j = 0; j < length; j++) { 4330 if (n->in(j) == in) { 4331 cnt--; 4332 } 4333 } 4334 assert(cnt == 0, "Mismatched edge count."); 4335 } else if (in == nullptr) { 4336 assert(i == 0 || i >= n->req() || 4337 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4338 (n->is_Unlock() && i == (n->req() - 1)) || 4339 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4340 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4341 } else { 4342 assert(in->is_top(), "sanity"); 4343 // Nothing to check. 4344 } 4345 } 4346 } 4347 } 4348 4349 //------------------------------verify_graph_edges--------------------------- 4350 // Walk the Graph and verify that there is a one-to-one correspondence 4351 // between Use-Def edges and Def-Use edges in the graph. 4352 void Compile::verify_graph_edges(bool no_dead_code) { 4353 if (VerifyGraphEdges) { 4354 Unique_Node_List visited; 4355 4356 // Call graph walk to check edges 4357 verify_bidirectional_edges(visited); 4358 if (no_dead_code) { 4359 // Now make sure that no visited node is used by an unvisited node. 4360 bool dead_nodes = false; 4361 Unique_Node_List checked; 4362 while (visited.size() > 0) { 4363 Node* n = visited.pop(); 4364 checked.push(n); 4365 for (uint i = 0; i < n->outcnt(); i++) { 4366 Node* use = n->raw_out(i); 4367 if (checked.member(use)) continue; // already checked 4368 if (visited.member(use)) continue; // already in the graph 4369 if (use->is_Con()) continue; // a dead ConNode is OK 4370 // At this point, we have found a dead node which is DU-reachable. 4371 if (!dead_nodes) { 4372 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4373 dead_nodes = true; 4374 } 4375 use->dump(2); 4376 tty->print_cr("---"); 4377 checked.push(use); // No repeats; pretend it is now checked. 4378 } 4379 } 4380 assert(!dead_nodes, "using nodes must be reachable from root"); 4381 } 4382 } 4383 } 4384 #endif 4385 4386 // The Compile object keeps track of failure reasons separately from the ciEnv. 4387 // This is required because there is not quite a 1-1 relation between the 4388 // ciEnv and its compilation task and the Compile object. Note that one 4389 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4390 // to backtrack and retry without subsuming loads. Other than this backtracking 4391 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4392 // by the logic in C2Compiler. 4393 void Compile::record_failure(const char* reason) { 4394 if (log() != nullptr) { 4395 log()->elem("failure reason='%s' phase='compile'", reason); 4396 } 4397 if (_failure_reason.get() == nullptr) { 4398 // Record the first failure reason. 4399 _failure_reason.set(reason); 4400 if (CaptureBailoutInformation) { 4401 _first_failure_details = new CompilationFailureInfo(reason); 4402 } 4403 } 4404 4405 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4406 C->print_method(PHASE_FAILURE, 1); 4407 } 4408 _root = nullptr; // flush the graph, too 4409 } 4410 4411 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4412 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4413 _compile(Compile::current()), 4414 _log(nullptr), 4415 _phase_name(name), 4416 _dolog(CITimeVerbose) 4417 { 4418 assert(_compile != nullptr, "sanity check"); 4419 if (_dolog) { 4420 _log = _compile->log(); 4421 } 4422 if (_log != nullptr) { 4423 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4424 _log->stamp(); 4425 _log->end_head(); 4426 } 4427 } 4428 4429 Compile::TracePhase::~TracePhase() { 4430 if (_compile->failing()) return; 4431 #ifdef ASSERT 4432 if (PrintIdealNodeCount) { 4433 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4434 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4435 } 4436 4437 if (VerifyIdealNodeCount) { 4438 _compile->print_missing_nodes(); 4439 } 4440 #endif 4441 4442 if (_log != nullptr) { 4443 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4444 } 4445 } 4446 4447 //----------------------------static_subtype_check----------------------------- 4448 // Shortcut important common cases when superklass is exact: 4449 // (0) superklass is java.lang.Object (can occur in reflective code) 4450 // (1) subklass is already limited to a subtype of superklass => always ok 4451 // (2) subklass does not overlap with superklass => always fail 4452 // (3) superklass has NO subtypes and we can check with a simple compare. 4453 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4454 if (skip) { 4455 return SSC_full_test; // Let caller generate the general case. 4456 } 4457 4458 if (subk->is_java_subtype_of(superk)) { 4459 return SSC_always_true; // (0) and (1) this test cannot fail 4460 } 4461 4462 if (!subk->maybe_java_subtype_of(superk)) { 4463 return SSC_always_false; // (2) true path dead; no dynamic test needed 4464 } 4465 4466 const Type* superelem = superk; 4467 if (superk->isa_aryklassptr()) { 4468 int ignored; 4469 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4470 } 4471 4472 if (superelem->isa_instklassptr()) { 4473 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4474 if (!ik->has_subklass()) { 4475 if (!ik->is_final()) { 4476 // Add a dependency if there is a chance of a later subclass. 4477 dependencies()->assert_leaf_type(ik); 4478 } 4479 if (!superk->maybe_java_subtype_of(subk)) { 4480 return SSC_always_false; 4481 } 4482 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4483 } 4484 } else { 4485 // A primitive array type has no subtypes. 4486 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4487 } 4488 4489 return SSC_full_test; 4490 } 4491 4492 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4493 #ifdef _LP64 4494 // The scaled index operand to AddP must be a clean 64-bit value. 4495 // Java allows a 32-bit int to be incremented to a negative 4496 // value, which appears in a 64-bit register as a large 4497 // positive number. Using that large positive number as an 4498 // operand in pointer arithmetic has bad consequences. 4499 // On the other hand, 32-bit overflow is rare, and the possibility 4500 // can often be excluded, if we annotate the ConvI2L node with 4501 // a type assertion that its value is known to be a small positive 4502 // number. (The prior range check has ensured this.) 4503 // This assertion is used by ConvI2LNode::Ideal. 4504 int index_max = max_jint - 1; // array size is max_jint, index is one less 4505 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4506 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4507 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4508 #endif 4509 return idx; 4510 } 4511 4512 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4513 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4514 if (ctrl != nullptr) { 4515 // Express control dependency by a CastII node with a narrow type. 4516 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4517 // node from floating above the range check during loop optimizations. Otherwise, the 4518 // ConvI2L node may be eliminated independently of the range check, causing the data path 4519 // to become TOP while the control path is still there (although it's unreachable). 4520 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4521 value = phase->transform(value); 4522 } 4523 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4524 return phase->transform(new ConvI2LNode(value, ltype)); 4525 } 4526 4527 // The message about the current inlining is accumulated in 4528 // _print_inlining_stream and transferred into the _print_inlining_list 4529 // once we know whether inlining succeeds or not. For regular 4530 // inlining, messages are appended to the buffer pointed by 4531 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4532 // a new buffer is added after _print_inlining_idx in the list. This 4533 // way we can update the inlining message for late inlining call site 4534 // when the inlining is attempted again. 4535 void Compile::print_inlining_init() { 4536 if (print_inlining() || print_intrinsics()) { 4537 // print_inlining_init is actually called several times. 4538 print_inlining_reset(); 4539 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4540 } 4541 } 4542 4543 void Compile::print_inlining_reinit() { 4544 if (print_inlining() || print_intrinsics()) { 4545 print_inlining_reset(); 4546 } 4547 } 4548 4549 void Compile::print_inlining_reset() { 4550 _print_inlining_stream->reset(); 4551 } 4552 4553 void Compile::print_inlining_commit() { 4554 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4555 // Transfer the message from _print_inlining_stream to the current 4556 // _print_inlining_list buffer and clear _print_inlining_stream. 4557 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4558 print_inlining_reset(); 4559 } 4560 4561 void Compile::print_inlining_push() { 4562 // Add new buffer to the _print_inlining_list at current position 4563 _print_inlining_idx++; 4564 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4565 } 4566 4567 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4568 return _print_inlining_list->at(_print_inlining_idx); 4569 } 4570 4571 void Compile::print_inlining_update(CallGenerator* cg) { 4572 if (print_inlining() || print_intrinsics()) { 4573 if (cg->is_late_inline()) { 4574 if (print_inlining_current()->cg() != cg && 4575 (print_inlining_current()->cg() != nullptr || 4576 print_inlining_current()->ss()->size() != 0)) { 4577 print_inlining_push(); 4578 } 4579 print_inlining_commit(); 4580 print_inlining_current()->set_cg(cg); 4581 } else { 4582 if (print_inlining_current()->cg() != nullptr) { 4583 print_inlining_push(); 4584 } 4585 print_inlining_commit(); 4586 } 4587 } 4588 } 4589 4590 void Compile::print_inlining_move_to(CallGenerator* cg) { 4591 // We resume inlining at a late inlining call site. Locate the 4592 // corresponding inlining buffer so that we can update it. 4593 if (print_inlining() || print_intrinsics()) { 4594 for (int i = 0; i < _print_inlining_list->length(); i++) { 4595 if (_print_inlining_list->at(i)->cg() == cg) { 4596 _print_inlining_idx = i; 4597 return; 4598 } 4599 } 4600 ShouldNotReachHere(); 4601 } 4602 } 4603 4604 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4605 if (print_inlining() || print_intrinsics()) { 4606 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4607 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4608 // replace message with new message 4609 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4610 print_inlining_commit(); 4611 print_inlining_current()->set_cg(cg); 4612 } 4613 } 4614 4615 void Compile::print_inlining_assert_ready() { 4616 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 4617 } 4618 4619 void Compile::process_print_inlining() { 4620 assert(_late_inlines.length() == 0, "not drained yet"); 4621 if (print_inlining() || print_intrinsics()) { 4622 ResourceMark rm; 4623 stringStream ss; 4624 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 4625 for (int i = 0; i < _print_inlining_list->length(); i++) { 4626 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4627 ss.print("%s", pib->ss()->freeze()); 4628 delete pib; 4629 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 4630 } 4631 // Reset _print_inlining_list, it only contains destructed objects. 4632 // It is on the arena, so it will be freed when the arena is reset. 4633 _print_inlining_list = nullptr; 4634 // _print_inlining_stream won't be used anymore, either. 4635 print_inlining_reset(); 4636 size_t end = ss.size(); 4637 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4638 strncpy(_print_inlining_output, ss.freeze(), end+1); 4639 _print_inlining_output[end] = 0; 4640 } 4641 } 4642 4643 void Compile::dump_print_inlining() { 4644 if (_print_inlining_output != nullptr) { 4645 tty->print_raw(_print_inlining_output); 4646 } 4647 } 4648 4649 void Compile::log_late_inline(CallGenerator* cg) { 4650 if (log() != nullptr) { 4651 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4652 cg->unique_id()); 4653 JVMState* p = cg->call_node()->jvms(); 4654 while (p != nullptr) { 4655 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4656 p = p->caller(); 4657 } 4658 log()->tail("late_inline"); 4659 } 4660 } 4661 4662 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4663 log_late_inline(cg); 4664 if (log() != nullptr) { 4665 log()->inline_fail(msg); 4666 } 4667 } 4668 4669 void Compile::log_inline_id(CallGenerator* cg) { 4670 if (log() != nullptr) { 4671 // The LogCompilation tool needs a unique way to identify late 4672 // inline call sites. This id must be unique for this call site in 4673 // this compilation. Try to have it unique across compilations as 4674 // well because it can be convenient when grepping through the log 4675 // file. 4676 // Distinguish OSR compilations from others in case CICountOSR is 4677 // on. 4678 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4679 cg->set_unique_id(id); 4680 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4681 } 4682 } 4683 4684 void Compile::log_inline_failure(const char* msg) { 4685 if (C->log() != nullptr) { 4686 C->log()->inline_fail(msg); 4687 } 4688 } 4689 4690 4691 // Dump inlining replay data to the stream. 4692 // Don't change thread state and acquire any locks. 4693 void Compile::dump_inline_data(outputStream* out) { 4694 InlineTree* inl_tree = ilt(); 4695 if (inl_tree != nullptr) { 4696 out->print(" inline %d", inl_tree->count()); 4697 inl_tree->dump_replay_data(out); 4698 } 4699 } 4700 4701 void Compile::dump_inline_data_reduced(outputStream* out) { 4702 assert(ReplayReduce, ""); 4703 4704 InlineTree* inl_tree = ilt(); 4705 if (inl_tree == nullptr) { 4706 return; 4707 } 4708 // Enable iterative replay file reduction 4709 // Output "compile" lines for depth 1 subtrees, 4710 // simulating that those trees were compiled 4711 // instead of inlined. 4712 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4713 InlineTree* sub = inl_tree->subtrees().at(i); 4714 if (sub->inline_level() != 1) { 4715 continue; 4716 } 4717 4718 ciMethod* method = sub->method(); 4719 int entry_bci = -1; 4720 int comp_level = env()->task()->comp_level(); 4721 out->print("compile "); 4722 method->dump_name_as_ascii(out); 4723 out->print(" %d %d", entry_bci, comp_level); 4724 out->print(" inline %d", sub->count()); 4725 sub->dump_replay_data(out, -1); 4726 out->cr(); 4727 } 4728 } 4729 4730 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4731 if (n1->Opcode() < n2->Opcode()) return -1; 4732 else if (n1->Opcode() > n2->Opcode()) return 1; 4733 4734 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4735 for (uint i = 1; i < n1->req(); i++) { 4736 if (n1->in(i) < n2->in(i)) return -1; 4737 else if (n1->in(i) > n2->in(i)) return 1; 4738 } 4739 4740 return 0; 4741 } 4742 4743 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4744 Node* n1 = *n1p; 4745 Node* n2 = *n2p; 4746 4747 return cmp_expensive_nodes(n1, n2); 4748 } 4749 4750 void Compile::sort_expensive_nodes() { 4751 if (!expensive_nodes_sorted()) { 4752 _expensive_nodes.sort(cmp_expensive_nodes); 4753 } 4754 } 4755 4756 bool Compile::expensive_nodes_sorted() const { 4757 for (int i = 1; i < _expensive_nodes.length(); i++) { 4758 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4759 return false; 4760 } 4761 } 4762 return true; 4763 } 4764 4765 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4766 if (_expensive_nodes.length() == 0) { 4767 return false; 4768 } 4769 4770 assert(OptimizeExpensiveOps, "optimization off?"); 4771 4772 // Take this opportunity to remove dead nodes from the list 4773 int j = 0; 4774 for (int i = 0; i < _expensive_nodes.length(); i++) { 4775 Node* n = _expensive_nodes.at(i); 4776 if (!n->is_unreachable(igvn)) { 4777 assert(n->is_expensive(), "should be expensive"); 4778 _expensive_nodes.at_put(j, n); 4779 j++; 4780 } 4781 } 4782 _expensive_nodes.trunc_to(j); 4783 4784 // Then sort the list so that similar nodes are next to each other 4785 // and check for at least two nodes of identical kind with same data 4786 // inputs. 4787 sort_expensive_nodes(); 4788 4789 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4790 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4791 return true; 4792 } 4793 } 4794 4795 return false; 4796 } 4797 4798 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4799 if (_expensive_nodes.length() == 0) { 4800 return; 4801 } 4802 4803 assert(OptimizeExpensiveOps, "optimization off?"); 4804 4805 // Sort to bring similar nodes next to each other and clear the 4806 // control input of nodes for which there's only a single copy. 4807 sort_expensive_nodes(); 4808 4809 int j = 0; 4810 int identical = 0; 4811 int i = 0; 4812 bool modified = false; 4813 for (; i < _expensive_nodes.length()-1; i++) { 4814 assert(j <= i, "can't write beyond current index"); 4815 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4816 identical++; 4817 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4818 continue; 4819 } 4820 if (identical > 0) { 4821 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4822 identical = 0; 4823 } else { 4824 Node* n = _expensive_nodes.at(i); 4825 igvn.replace_input_of(n, 0, nullptr); 4826 igvn.hash_insert(n); 4827 modified = true; 4828 } 4829 } 4830 if (identical > 0) { 4831 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4832 } else if (_expensive_nodes.length() >= 1) { 4833 Node* n = _expensive_nodes.at(i); 4834 igvn.replace_input_of(n, 0, nullptr); 4835 igvn.hash_insert(n); 4836 modified = true; 4837 } 4838 _expensive_nodes.trunc_to(j); 4839 if (modified) { 4840 igvn.optimize(); 4841 } 4842 } 4843 4844 void Compile::add_expensive_node(Node * n) { 4845 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4846 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4847 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4848 if (OptimizeExpensiveOps) { 4849 _expensive_nodes.append(n); 4850 } else { 4851 // Clear control input and let IGVN optimize expensive nodes if 4852 // OptimizeExpensiveOps is off. 4853 n->set_req(0, nullptr); 4854 } 4855 } 4856 4857 /** 4858 * Track coarsened Lock and Unlock nodes. 4859 */ 4860 4861 class Lock_List : public Node_List { 4862 uint _origin_cnt; 4863 public: 4864 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4865 uint origin_cnt() const { return _origin_cnt; } 4866 }; 4867 4868 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4869 int length = locks.length(); 4870 if (length > 0) { 4871 // Have to keep this list until locks elimination during Macro nodes elimination. 4872 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4873 AbstractLockNode* alock = locks.at(0); 4874 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4875 for (int i = 0; i < length; i++) { 4876 AbstractLockNode* lock = locks.at(i); 4877 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4878 locks_list->push(lock); 4879 BoxLockNode* this_box = lock->box_node()->as_BoxLock(); 4880 if (this_box != box) { 4881 // Locking regions (BoxLock) could be Unbalanced here: 4882 // - its coarsened locks were eliminated in earlier 4883 // macro nodes elimination followed by loop unroll 4884 // - it is OSR locking region (no Lock node) 4885 // Preserve Unbalanced status in such cases. 4886 if (!this_box->is_unbalanced()) { 4887 this_box->set_coarsened(); 4888 } 4889 if (!box->is_unbalanced()) { 4890 box->set_coarsened(); 4891 } 4892 } 4893 } 4894 _coarsened_locks.append(locks_list); 4895 } 4896 } 4897 4898 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4899 int count = coarsened_count(); 4900 for (int i = 0; i < count; i++) { 4901 Node_List* locks_list = _coarsened_locks.at(i); 4902 for (uint j = 0; j < locks_list->size(); j++) { 4903 Node* lock = locks_list->at(j); 4904 assert(lock->is_AbstractLock(), "sanity"); 4905 if (!useful.member(lock)) { 4906 locks_list->yank(lock); 4907 } 4908 } 4909 } 4910 } 4911 4912 void Compile::remove_coarsened_lock(Node* n) { 4913 if (n->is_AbstractLock()) { 4914 int count = coarsened_count(); 4915 for (int i = 0; i < count; i++) { 4916 Node_List* locks_list = _coarsened_locks.at(i); 4917 locks_list->yank(n); 4918 } 4919 } 4920 } 4921 4922 bool Compile::coarsened_locks_consistent() { 4923 int count = coarsened_count(); 4924 for (int i = 0; i < count; i++) { 4925 bool unbalanced = false; 4926 bool modified = false; // track locks kind modifications 4927 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4928 uint size = locks_list->size(); 4929 if (size == 0) { 4930 unbalanced = false; // All locks were eliminated - good 4931 } else if (size != locks_list->origin_cnt()) { 4932 unbalanced = true; // Some locks were removed from list 4933 } else { 4934 for (uint j = 0; j < size; j++) { 4935 Node* lock = locks_list->at(j); 4936 // All nodes in group should have the same state (modified or not) 4937 if (!lock->as_AbstractLock()->is_coarsened()) { 4938 if (j == 0) { 4939 // first on list was modified, the rest should be too for consistency 4940 modified = true; 4941 } else if (!modified) { 4942 // this lock was modified but previous locks on the list were not 4943 unbalanced = true; 4944 break; 4945 } 4946 } else if (modified) { 4947 // previous locks on list were modified but not this lock 4948 unbalanced = true; 4949 break; 4950 } 4951 } 4952 } 4953 if (unbalanced) { 4954 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4955 #ifdef ASSERT 4956 if (PrintEliminateLocks) { 4957 tty->print_cr("=== unbalanced coarsened locks ==="); 4958 for (uint l = 0; l < size; l++) { 4959 locks_list->at(l)->dump(); 4960 } 4961 } 4962 #endif 4963 record_failure(C2Compiler::retry_no_locks_coarsening()); 4964 return false; 4965 } 4966 } 4967 return true; 4968 } 4969 4970 // Mark locking regions (identified by BoxLockNode) as unbalanced if 4971 // locks coarsening optimization removed Lock/Unlock nodes from them. 4972 // Such regions become unbalanced because coarsening only removes part 4973 // of Lock/Unlock nodes in region. As result we can't execute other 4974 // locks elimination optimizations which assume all code paths have 4975 // corresponding pair of Lock/Unlock nodes - they are balanced. 4976 void Compile::mark_unbalanced_boxes() const { 4977 int count = coarsened_count(); 4978 for (int i = 0; i < count; i++) { 4979 Node_List* locks_list = _coarsened_locks.at(i); 4980 uint size = locks_list->size(); 4981 if (size > 0) { 4982 AbstractLockNode* alock = locks_list->at(0)->as_AbstractLock(); 4983 BoxLockNode* box = alock->box_node()->as_BoxLock(); 4984 if (alock->is_coarsened()) { 4985 // coarsened_locks_consistent(), which is called before this method, verifies 4986 // that the rest of Lock/Unlock nodes on locks_list are also coarsened. 4987 assert(!box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4988 for (uint j = 1; j < size; j++) { 4989 assert(locks_list->at(j)->as_AbstractLock()->is_coarsened(), "only coarsened locks are expected here"); 4990 BoxLockNode* this_box = locks_list->at(j)->as_AbstractLock()->box_node()->as_BoxLock(); 4991 if (box != this_box) { 4992 assert(!this_box->is_eliminated(), "regions with coarsened locks should not be marked as eliminated"); 4993 box->set_unbalanced(); 4994 this_box->set_unbalanced(); 4995 } 4996 } 4997 } 4998 } 4999 } 5000 } 5001 5002 /** 5003 * Remove the speculative part of types and clean up the graph 5004 */ 5005 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 5006 if (UseTypeSpeculation) { 5007 Unique_Node_List worklist; 5008 worklist.push(root()); 5009 int modified = 0; 5010 // Go over all type nodes that carry a speculative type, drop the 5011 // speculative part of the type and enqueue the node for an igvn 5012 // which may optimize it out. 5013 for (uint next = 0; next < worklist.size(); ++next) { 5014 Node *n = worklist.at(next); 5015 if (n->is_Type()) { 5016 TypeNode* tn = n->as_Type(); 5017 const Type* t = tn->type(); 5018 const Type* t_no_spec = t->remove_speculative(); 5019 if (t_no_spec != t) { 5020 bool in_hash = igvn.hash_delete(n); 5021 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table"); 5022 tn->set_type(t_no_spec); 5023 igvn.hash_insert(n); 5024 igvn._worklist.push(n); // give it a chance to go away 5025 modified++; 5026 } 5027 } 5028 // Iterate over outs - endless loops is unreachable from below 5029 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 5030 Node *m = n->fast_out(i); 5031 if (not_a_node(m)) { 5032 continue; 5033 } 5034 worklist.push(m); 5035 } 5036 } 5037 // Drop the speculative part of all types in the igvn's type table 5038 igvn.remove_speculative_types(); 5039 if (modified > 0) { 5040 igvn.optimize(); 5041 if (failing()) return; 5042 } 5043 #ifdef ASSERT 5044 // Verify that after the IGVN is over no speculative type has resurfaced 5045 worklist.clear(); 5046 worklist.push(root()); 5047 for (uint next = 0; next < worklist.size(); ++next) { 5048 Node *n = worklist.at(next); 5049 const Type* t = igvn.type_or_null(n); 5050 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 5051 if (n->is_Type()) { 5052 t = n->as_Type()->type(); 5053 assert(t == t->remove_speculative(), "no more speculative types"); 5054 } 5055 // Iterate over outs - endless loops is unreachable from below 5056 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 5057 Node *m = n->fast_out(i); 5058 if (not_a_node(m)) { 5059 continue; 5060 } 5061 worklist.push(m); 5062 } 5063 } 5064 igvn.check_no_speculative_types(); 5065 #endif 5066 } 5067 } 5068 5069 // Auxiliary methods to support randomized stressing/fuzzing. 5070 5071 int Compile::random() { 5072 _stress_seed = os::next_random(_stress_seed); 5073 return static_cast<int>(_stress_seed); 5074 } 5075 5076 // This method can be called the arbitrary number of times, with current count 5077 // as the argument. The logic allows selecting a single candidate from the 5078 // running list of candidates as follows: 5079 // int count = 0; 5080 // Cand* selected = null; 5081 // while(cand = cand->next()) { 5082 // if (randomized_select(++count)) { 5083 // selected = cand; 5084 // } 5085 // } 5086 // 5087 // Including count equalizes the chances any candidate is "selected". 5088 // This is useful when we don't have the complete list of candidates to choose 5089 // from uniformly. In this case, we need to adjust the randomicity of the 5090 // selection, or else we will end up biasing the selection towards the latter 5091 // candidates. 5092 // 5093 // Quick back-envelope calculation shows that for the list of n candidates 5094 // the equal probability for the candidate to persist as "best" can be 5095 // achieved by replacing it with "next" k-th candidate with the probability 5096 // of 1/k. It can be easily shown that by the end of the run, the 5097 // probability for any candidate is converged to 1/n, thus giving the 5098 // uniform distribution among all the candidates. 5099 // 5100 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5101 #define RANDOMIZED_DOMAIN_POW 29 5102 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5103 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5104 bool Compile::randomized_select(int count) { 5105 assert(count > 0, "only positive"); 5106 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5107 } 5108 5109 CloneMap& Compile::clone_map() { return _clone_map; } 5110 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5111 5112 void NodeCloneInfo::dump_on(outputStream* st) const { 5113 st->print(" {%d:%d} ", idx(), gen()); 5114 } 5115 5116 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5117 uint64_t val = value(old->_idx); 5118 NodeCloneInfo cio(val); 5119 assert(val != 0, "old node should be in the map"); 5120 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5121 insert(nnn->_idx, cin.get()); 5122 #ifndef PRODUCT 5123 if (is_debug()) { 5124 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5125 } 5126 #endif 5127 } 5128 5129 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5130 NodeCloneInfo cio(value(old->_idx)); 5131 if (cio.get() == 0) { 5132 cio.set(old->_idx, 0); 5133 insert(old->_idx, cio.get()); 5134 #ifndef PRODUCT 5135 if (is_debug()) { 5136 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5137 } 5138 #endif 5139 } 5140 clone(old, nnn, gen); 5141 } 5142 5143 int CloneMap::max_gen() const { 5144 int g = 0; 5145 DictI di(_dict); 5146 for(; di.test(); ++di) { 5147 int t = gen(di._key); 5148 if (g < t) { 5149 g = t; 5150 #ifndef PRODUCT 5151 if (is_debug()) { 5152 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5153 } 5154 #endif 5155 } 5156 } 5157 return g; 5158 } 5159 5160 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5161 uint64_t val = value(key); 5162 if (val != 0) { 5163 NodeCloneInfo ni(val); 5164 ni.dump_on(st); 5165 } 5166 } 5167 5168 void Compile::shuffle_macro_nodes() { 5169 if (_macro_nodes.length() < 2) { 5170 return; 5171 } 5172 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) { 5173 uint j = C->random() % (i + 1); 5174 swap(_macro_nodes.at(i), _macro_nodes.at(j)); 5175 } 5176 } 5177 5178 // Move Allocate nodes to the start of the list 5179 void Compile::sort_macro_nodes() { 5180 int count = macro_count(); 5181 int allocates = 0; 5182 for (int i = 0; i < count; i++) { 5183 Node* n = macro_node(i); 5184 if (n->is_Allocate()) { 5185 if (i != allocates) { 5186 Node* tmp = macro_node(allocates); 5187 _macro_nodes.at_put(allocates, n); 5188 _macro_nodes.at_put(i, tmp); 5189 } 5190 allocates++; 5191 } 5192 } 5193 } 5194 5195 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5196 if (failing()) { return; } 5197 EventCompilerPhase event(UNTIMED); 5198 if (event.should_commit()) { 5199 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5200 } 5201 #ifndef PRODUCT 5202 ResourceMark rm; 5203 stringStream ss; 5204 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5205 int iter = ++_igv_phase_iter[cpt]; 5206 if (iter > 1) { 5207 ss.print(" %d", iter); 5208 } 5209 if (n != nullptr) { 5210 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5211 } 5212 5213 const char* name = ss.as_string(); 5214 if (should_print_igv(level)) { 5215 _igv_printer->print_method(name, level); 5216 } 5217 if (should_print_phase(cpt)) { 5218 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5219 } 5220 #endif 5221 C->_latest_stage_start_counter.stamp(); 5222 } 5223 5224 // Only used from CompileWrapper 5225 void Compile::begin_method() { 5226 #ifndef PRODUCT 5227 if (_method != nullptr && should_print_igv(1)) { 5228 _igv_printer->begin_method(); 5229 } 5230 #endif 5231 C->_latest_stage_start_counter.stamp(); 5232 } 5233 5234 // Only used from CompileWrapper 5235 void Compile::end_method() { 5236 EventCompilerPhase event(UNTIMED); 5237 if (event.should_commit()) { 5238 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5239 } 5240 5241 #ifndef PRODUCT 5242 if (_method != nullptr && should_print_igv(1)) { 5243 _igv_printer->end_method(); 5244 } 5245 #endif 5246 } 5247 5248 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5249 #ifndef PRODUCT 5250 if (_directive->should_print_phase(cpt)) { 5251 return true; 5252 } 5253 #endif 5254 return false; 5255 } 5256 5257 bool Compile::should_print_igv(const int level) { 5258 #ifndef PRODUCT 5259 if (PrintIdealGraphLevel < 0) { // disabled by the user 5260 return false; 5261 } 5262 5263 bool need = directive()->IGVPrintLevelOption >= level; 5264 if (need && !_igv_printer) { 5265 _igv_printer = IdealGraphPrinter::printer(); 5266 _igv_printer->set_compile(this); 5267 } 5268 return need; 5269 #else 5270 return false; 5271 #endif 5272 } 5273 5274 #ifndef PRODUCT 5275 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5276 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5277 5278 // Called from debugger. Prints method to the default file with the default phase name. 5279 // This works regardless of any Ideal Graph Visualizer flags set or not. 5280 void igv_print() { 5281 Compile::current()->igv_print_method_to_file(); 5282 } 5283 5284 // Same as igv_print() above but with a specified phase name. 5285 void igv_print(const char* phase_name) { 5286 Compile::current()->igv_print_method_to_file(phase_name); 5287 } 5288 5289 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5290 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5291 // This works regardless of any Ideal Graph Visualizer flags set or not. 5292 void igv_print(bool network) { 5293 if (network) { 5294 Compile::current()->igv_print_method_to_network(); 5295 } else { 5296 Compile::current()->igv_print_method_to_file(); 5297 } 5298 } 5299 5300 // Same as igv_print(bool network) above but with a specified phase name. 5301 void igv_print(bool network, const char* phase_name) { 5302 if (network) { 5303 Compile::current()->igv_print_method_to_network(phase_name); 5304 } else { 5305 Compile::current()->igv_print_method_to_file(phase_name); 5306 } 5307 } 5308 5309 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5310 void igv_print_default() { 5311 Compile::current()->print_method(PHASE_DEBUG, 0); 5312 } 5313 5314 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5315 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5316 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5317 void igv_append() { 5318 Compile::current()->igv_print_method_to_file("Debug", true); 5319 } 5320 5321 // Same as igv_append() above but with a specified phase name. 5322 void igv_append(const char* phase_name) { 5323 Compile::current()->igv_print_method_to_file(phase_name, true); 5324 } 5325 5326 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5327 const char* file_name = "custom_debug.xml"; 5328 if (_debug_file_printer == nullptr) { 5329 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5330 } else { 5331 _debug_file_printer->update_compiled_method(C->method()); 5332 } 5333 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5334 _debug_file_printer->print(phase_name, (Node*)C->root()); 5335 } 5336 5337 void Compile::igv_print_method_to_network(const char* phase_name) { 5338 if (_debug_network_printer == nullptr) { 5339 _debug_network_printer = new IdealGraphPrinter(C); 5340 } else { 5341 _debug_network_printer->update_compiled_method(C->method()); 5342 } 5343 tty->print_cr("Method printed over network stream to IGV"); 5344 _debug_network_printer->print(phase_name, (Node*)C->root()); 5345 } 5346 #endif 5347 5348 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5349 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5350 return value; 5351 } 5352 Node* result = nullptr; 5353 if (bt == T_BYTE) { 5354 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5355 result = new RShiftINode(result, phase->intcon(24)); 5356 } else if (bt == T_BOOLEAN) { 5357 result = new AndINode(value, phase->intcon(0xFF)); 5358 } else if (bt == T_CHAR) { 5359 result = new AndINode(value,phase->intcon(0xFFFF)); 5360 } else { 5361 assert(bt == T_SHORT, "unexpected narrow type"); 5362 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5363 result = new RShiftINode(result, phase->intcon(16)); 5364 } 5365 if (transform_res) { 5366 result = phase->transform(result); 5367 } 5368 return result; 5369 } 5370 5371 void Compile::record_method_not_compilable_oom() { 5372 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5373 }