1 /* 2 * Copyright (c) 1997, 2023, 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/compilationMemoryStatistic.hpp" 33 #include "compiler/compileBroker.hpp" 34 #include "compiler/compileLog.hpp" 35 #include "compiler/disassembler.hpp" 36 #include "compiler/oopMap.hpp" 37 #include "gc/shared/barrierSet.hpp" 38 #include "gc/shared/c2/barrierSetC2.hpp" 39 #include "jfr/jfrEvents.hpp" 40 #include "jvm_io.h" 41 #include "memory/allocation.hpp" 42 #include "memory/resourceArea.hpp" 43 #include "opto/addnode.hpp" 44 #include "opto/block.hpp" 45 #include "opto/c2compiler.hpp" 46 #include "opto/callGenerator.hpp" 47 #include "opto/callnode.hpp" 48 #include "opto/castnode.hpp" 49 #include "opto/cfgnode.hpp" 50 #include "opto/chaitin.hpp" 51 #include "opto/compile.hpp" 52 #include "opto/connode.hpp" 53 #include "opto/convertnode.hpp" 54 #include "opto/divnode.hpp" 55 #include "opto/escape.hpp" 56 #include "opto/idealGraphPrinter.hpp" 57 #include "opto/inlinetypenode.hpp" 58 #include "opto/loopnode.hpp" 59 #include "opto/machnode.hpp" 60 #include "opto/macro.hpp" 61 #include "opto/matcher.hpp" 62 #include "opto/mathexactnode.hpp" 63 #include "opto/memnode.hpp" 64 #include "opto/mulnode.hpp" 65 #include "opto/narrowptrnode.hpp" 66 #include "opto/node.hpp" 67 #include "opto/opcodes.hpp" 68 #include "opto/output.hpp" 69 #include "opto/parse.hpp" 70 #include "opto/phaseX.hpp" 71 #include "opto/rootnode.hpp" 72 #include "opto/runtime.hpp" 73 #include "opto/stringopts.hpp" 74 #include "opto/type.hpp" 75 #include "opto/vector.hpp" 76 #include "opto/vectornode.hpp" 77 #include "runtime/globals_extension.hpp" 78 #include "runtime/sharedRuntime.hpp" 79 #include "runtime/signature.hpp" 80 #include "runtime/stubRoutines.hpp" 81 #include "runtime/timer.hpp" 82 #include "utilities/align.hpp" 83 #include "utilities/copy.hpp" 84 #include "utilities/macros.hpp" 85 #include "utilities/resourceHash.hpp" 86 87 // -------------------- Compile::mach_constant_base_node ----------------------- 88 // Constant table base node singleton. 89 MachConstantBaseNode* Compile::mach_constant_base_node() { 90 if (_mach_constant_base_node == nullptr) { 91 _mach_constant_base_node = new MachConstantBaseNode(); 92 _mach_constant_base_node->add_req(C->root()); 93 } 94 return _mach_constant_base_node; 95 } 96 97 98 /// Support for intrinsics. 99 100 // Return the index at which m must be inserted (or already exists). 101 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 102 class IntrinsicDescPair { 103 private: 104 ciMethod* _m; 105 bool _is_virtual; 106 public: 107 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 108 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 109 ciMethod* m= elt->method(); 110 ciMethod* key_m = key->_m; 111 if (key_m < m) return -1; 112 else if (key_m > m) return 1; 113 else { 114 bool is_virtual = elt->is_virtual(); 115 bool key_virtual = key->_is_virtual; 116 if (key_virtual < is_virtual) return -1; 117 else if (key_virtual > is_virtual) return 1; 118 else return 0; 119 } 120 } 121 }; 122 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 123 #ifdef ASSERT 124 for (int i = 1; i < _intrinsics.length(); i++) { 125 CallGenerator* cg1 = _intrinsics.at(i-1); 126 CallGenerator* cg2 = _intrinsics.at(i); 127 assert(cg1->method() != cg2->method() 128 ? cg1->method() < cg2->method() 129 : cg1->is_virtual() < cg2->is_virtual(), 130 "compiler intrinsics list must stay sorted"); 131 } 132 #endif 133 IntrinsicDescPair pair(m, is_virtual); 134 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 135 } 136 137 void Compile::register_intrinsic(CallGenerator* cg) { 138 bool found = false; 139 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 140 assert(!found, "registering twice"); 141 _intrinsics.insert_before(index, cg); 142 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 143 } 144 145 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 146 assert(m->is_loaded(), "don't try this on unloaded methods"); 147 if (_intrinsics.length() > 0) { 148 bool found = false; 149 int index = intrinsic_insertion_index(m, is_virtual, found); 150 if (found) { 151 return _intrinsics.at(index); 152 } 153 } 154 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 155 if (m->intrinsic_id() != vmIntrinsics::_none && 156 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 157 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 158 if (cg != nullptr) { 159 // Save it for next time: 160 register_intrinsic(cg); 161 return cg; 162 } else { 163 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 164 } 165 } 166 return nullptr; 167 } 168 169 // Compile::make_vm_intrinsic is defined in library_call.cpp. 170 171 #ifndef PRODUCT 172 // statistics gathering... 173 174 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 175 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 176 177 inline int as_int(vmIntrinsics::ID id) { 178 return vmIntrinsics::as_int(id); 179 } 180 181 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 182 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 183 int oflags = _intrinsic_hist_flags[as_int(id)]; 184 assert(flags != 0, "what happened?"); 185 if (is_virtual) { 186 flags |= _intrinsic_virtual; 187 } 188 bool changed = (flags != oflags); 189 if ((flags & _intrinsic_worked) != 0) { 190 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 191 if (count == 1) { 192 changed = true; // first time 193 } 194 // increment the overall count also: 195 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 196 } 197 if (changed) { 198 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 199 // Something changed about the intrinsic's virtuality. 200 if ((flags & _intrinsic_virtual) != 0) { 201 // This is the first use of this intrinsic as a virtual call. 202 if (oflags != 0) { 203 // We already saw it as a non-virtual, so note both cases. 204 flags |= _intrinsic_both; 205 } 206 } else if ((oflags & _intrinsic_both) == 0) { 207 // This is the first use of this intrinsic as a non-virtual 208 flags |= _intrinsic_both; 209 } 210 } 211 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 212 } 213 // update the overall flags also: 214 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 215 return changed; 216 } 217 218 static char* format_flags(int flags, char* buf) { 219 buf[0] = 0; 220 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 221 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 222 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 223 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 224 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 225 if (buf[0] == 0) strcat(buf, ","); 226 assert(buf[0] == ',', "must be"); 227 return &buf[1]; 228 } 229 230 void Compile::print_intrinsic_statistics() { 231 char flagsbuf[100]; 232 ttyLocker ttyl; 233 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 234 tty->print_cr("Compiler intrinsic usage:"); 235 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 236 if (total == 0) total = 1; // avoid div0 in case of no successes 237 #define PRINT_STAT_LINE(name, c, f) \ 238 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 239 for (auto id : EnumRange<vmIntrinsicID>{}) { 240 int flags = _intrinsic_hist_flags[as_int(id)]; 241 juint count = _intrinsic_hist_count[as_int(id)]; 242 if ((flags | count) != 0) { 243 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 244 } 245 } 246 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 247 if (xtty != nullptr) xtty->tail("statistics"); 248 } 249 250 void Compile::print_statistics() { 251 { ttyLocker ttyl; 252 if (xtty != nullptr) xtty->head("statistics type='opto'"); 253 Parse::print_statistics(); 254 PhaseStringOpts::print_statistics(); 255 PhaseCCP::print_statistics(); 256 PhaseRegAlloc::print_statistics(); 257 PhaseOutput::print_statistics(); 258 PhasePeephole::print_statistics(); 259 PhaseIdealLoop::print_statistics(); 260 ConnectionGraph::print_statistics(); 261 PhaseMacroExpand::print_statistics(); 262 if (xtty != nullptr) xtty->tail("statistics"); 263 } 264 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 265 // put this under its own <statistics> element. 266 print_intrinsic_statistics(); 267 } 268 } 269 #endif //PRODUCT 270 271 void Compile::gvn_replace_by(Node* n, Node* nn) { 272 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 273 Node* use = n->last_out(i); 274 bool is_in_table = initial_gvn()->hash_delete(use); 275 uint uses_found = 0; 276 for (uint j = 0; j < use->len(); j++) { 277 if (use->in(j) == n) { 278 if (j < use->req()) 279 use->set_req(j, nn); 280 else 281 use->set_prec(j, nn); 282 uses_found++; 283 } 284 } 285 if (is_in_table) { 286 // reinsert into table 287 initial_gvn()->hash_find_insert(use); 288 } 289 record_for_igvn(use); 290 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist); 291 i -= uses_found; // we deleted 1 or more copies of this edge 292 } 293 } 294 295 296 // Identify all nodes that are reachable from below, useful. 297 // Use breadth-first pass that records state in a Unique_Node_List, 298 // recursive traversal is slower. 299 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 300 int estimated_worklist_size = live_nodes(); 301 useful.map( estimated_worklist_size, nullptr ); // preallocate space 302 303 // Initialize worklist 304 if (root() != nullptr) { useful.push(root()); } 305 // If 'top' is cached, declare it useful to preserve cached node 306 if (cached_top_node()) { useful.push(cached_top_node()); } 307 308 // Push all useful nodes onto the list, breadthfirst 309 for( uint next = 0; next < useful.size(); ++next ) { 310 assert( next < unique(), "Unique useful nodes < total nodes"); 311 Node *n = useful.at(next); 312 uint max = n->len(); 313 for( uint i = 0; i < max; ++i ) { 314 Node *m = n->in(i); 315 if (not_a_node(m)) continue; 316 useful.push(m); 317 } 318 } 319 } 320 321 // Update dead_node_list with any missing dead nodes using useful 322 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 323 void Compile::update_dead_node_list(Unique_Node_List &useful) { 324 uint max_idx = unique(); 325 VectorSet& useful_node_set = useful.member_set(); 326 327 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 328 // If node with index node_idx is not in useful set, 329 // mark it as dead in dead node list. 330 if (!useful_node_set.test(node_idx)) { 331 record_dead_node(node_idx); 332 } 333 } 334 } 335 336 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 337 int shift = 0; 338 for (int i = 0; i < inlines->length(); i++) { 339 CallGenerator* cg = inlines->at(i); 340 if (useful.member(cg->call_node())) { 341 if (shift > 0) { 342 inlines->at_put(i - shift, cg); 343 } 344 } else { 345 shift++; // skip over the dead element 346 } 347 } 348 if (shift > 0) { 349 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 350 } 351 } 352 353 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 354 assert(dead != nullptr && dead->is_Call(), "sanity"); 355 int found = 0; 356 for (int i = 0; i < inlines->length(); i++) { 357 if (inlines->at(i)->call_node() == dead) { 358 inlines->remove_at(i); 359 found++; 360 NOT_DEBUG( break; ) // elements are unique, so exit early 361 } 362 } 363 assert(found <= 1, "not unique"); 364 } 365 366 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)> 367 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) { 368 for (int i = node_list.length() - 1; i >= 0; i--) { 369 N* node = node_list.at(i); 370 if (!useful.member(node)) { 371 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 372 } 373 } 374 } 375 376 void Compile::remove_useless_node(Node* dead) { 377 remove_modified_node(dead); 378 379 // Constant node that has no out-edges and has only one in-edge from 380 // root is usually dead. However, sometimes reshaping walk makes 381 // it reachable by adding use edges. So, we will NOT count Con nodes 382 // as dead to be conservative about the dead node count at any 383 // given time. 384 if (!dead->is_Con()) { 385 record_dead_node(dead->_idx); 386 } 387 if (dead->is_macro()) { 388 remove_macro_node(dead); 389 } 390 if (dead->is_expensive()) { 391 remove_expensive_node(dead); 392 } 393 if (dead->Opcode() == Op_Opaque4) { 394 remove_template_assertion_predicate_opaq(dead); 395 } 396 if (dead->is_ParsePredicate()) { 397 remove_parse_predicate(dead->as_ParsePredicate()); 398 } 399 if (dead->for_post_loop_opts_igvn()) { 400 remove_from_post_loop_opts_igvn(dead); 401 } 402 if (dead->is_InlineType()) { 403 remove_inline_type(dead); 404 } 405 if (dead->is_Call()) { 406 remove_useless_late_inlines( &_late_inlines, dead); 407 remove_useless_late_inlines( &_string_late_inlines, dead); 408 remove_useless_late_inlines( &_boxing_late_inlines, dead); 409 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 410 411 if (dead->is_CallStaticJava()) { 412 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 413 } 414 } 415 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 416 bs->unregister_potential_barrier_node(dead); 417 } 418 419 // Disconnect all useless nodes by disconnecting those at the boundary. 420 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) { 421 uint next = 0; 422 while (next < useful.size()) { 423 Node *n = useful.at(next++); 424 if (n->is_SafePoint()) { 425 // We're done with a parsing phase. Replaced nodes are not valid 426 // beyond that point. 427 n->as_SafePoint()->delete_replaced_nodes(); 428 } 429 // Use raw traversal of out edges since this code removes out edges 430 int max = n->outcnt(); 431 for (int j = 0; j < max; ++j) { 432 Node* child = n->raw_out(j); 433 if (!useful.member(child)) { 434 assert(!child->is_top() || child != top(), 435 "If top is cached in Compile object it is in useful list"); 436 // Only need to remove this out-edge to the useless node 437 n->raw_del_out(j); 438 --j; 439 --max; 440 } 441 } 442 if (n->outcnt() == 1 && n->has_special_unique_user()) { 443 assert(useful.member(n->unique_out()), "do not push a useless node"); 444 worklist.push(n->unique_out()); 445 } 446 if (n->outcnt() == 0) { 447 worklist.push(n); 448 } 449 } 450 451 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 452 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes 453 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes 454 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 455 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 456 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes 457 #ifdef ASSERT 458 if (_modified_nodes != nullptr) { 459 _modified_nodes->remove_useless_nodes(useful.member_set()); 460 } 461 #endif 462 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 463 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 464 #ifdef ASSERT 465 if (_modified_nodes != nullptr) { 466 _modified_nodes->remove_useless_nodes(useful.member_set()); 467 } 468 #endif 469 470 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 471 bs->eliminate_useless_gc_barriers(useful, this); 472 // clean up the late inline lists 473 remove_useless_late_inlines( &_late_inlines, useful); 474 remove_useless_late_inlines( &_string_late_inlines, useful); 475 remove_useless_late_inlines( &_boxing_late_inlines, useful); 476 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 477 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 478 } 479 480 // ============================================================================ 481 //------------------------------CompileWrapper--------------------------------- 482 class CompileWrapper : public StackObj { 483 Compile *const _compile; 484 public: 485 CompileWrapper(Compile* compile); 486 487 ~CompileWrapper(); 488 }; 489 490 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 491 // the Compile* pointer is stored in the current ciEnv: 492 ciEnv* env = compile->env(); 493 assert(env == ciEnv::current(), "must already be a ciEnv active"); 494 assert(env->compiler_data() == nullptr, "compile already active?"); 495 env->set_compiler_data(compile); 496 assert(compile == Compile::current(), "sanity"); 497 498 compile->set_type_dict(nullptr); 499 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 500 compile->clone_map().set_clone_idx(0); 501 compile->set_type_last_size(0); 502 compile->set_last_tf(nullptr, nullptr); 503 compile->set_indexSet_arena(nullptr); 504 compile->set_indexSet_free_block_list(nullptr); 505 compile->init_type_arena(); 506 Type::Initialize(compile); 507 _compile->begin_method(); 508 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 509 } 510 CompileWrapper::~CompileWrapper() { 511 // simulate crash during compilation 512 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 513 514 _compile->end_method(); 515 _compile->env()->set_compiler_data(nullptr); 516 } 517 518 519 //----------------------------print_compile_messages--------------------------- 520 void Compile::print_compile_messages() { 521 #ifndef PRODUCT 522 // Check if recompiling 523 if (!subsume_loads() && PrintOpto) { 524 // Recompiling without allowing machine instructions to subsume loads 525 tty->print_cr("*********************************************************"); 526 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 527 tty->print_cr("*********************************************************"); 528 } 529 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 530 // Recompiling without escape analysis 531 tty->print_cr("*********************************************************"); 532 tty->print_cr("** Bailout: Recompile without escape analysis **"); 533 tty->print_cr("*********************************************************"); 534 } 535 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 536 // Recompiling without iterative escape analysis 537 tty->print_cr("*********************************************************"); 538 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 539 tty->print_cr("*********************************************************"); 540 } 541 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) { 542 // Recompiling without reducing allocation merges 543 tty->print_cr("*********************************************************"); 544 tty->print_cr("** Bailout: Recompile without reduce allocation merges **"); 545 tty->print_cr("*********************************************************"); 546 } 547 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 548 // Recompiling without boxing elimination 549 tty->print_cr("*********************************************************"); 550 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 551 tty->print_cr("*********************************************************"); 552 } 553 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 554 // Recompiling without locks coarsening 555 tty->print_cr("*********************************************************"); 556 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 557 tty->print_cr("*********************************************************"); 558 } 559 if (env()->break_at_compile()) { 560 // Open the debugger when compiling this method. 561 tty->print("### Breaking when compiling: "); 562 method()->print_short_name(); 563 tty->cr(); 564 BREAKPOINT; 565 } 566 567 if( PrintOpto ) { 568 if (is_osr_compilation()) { 569 tty->print("[OSR]%3d", _compile_id); 570 } else { 571 tty->print("%3d", _compile_id); 572 } 573 } 574 #endif 575 } 576 577 #ifndef PRODUCT 578 void Compile::print_ideal_ir(const char* phase_name) { 579 // keep the following output all in one block 580 // This output goes directly to the tty, not the compiler log. 581 // To enable tools to match it up with the compilation activity, 582 // be sure to tag this tty output with the compile ID. 583 584 // Node dumping can cause a safepoint, which can break the tty lock. 585 // Buffer all node dumps, so that all safepoints happen before we lock. 586 ResourceMark rm; 587 stringStream ss; 588 589 if (_output == nullptr) { 590 ss.print_cr("AFTER: %s", phase_name); 591 // Print out all nodes in ascending order of index. 592 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 593 } else { 594 // Dump the node blockwise if we have a scheduling 595 _output->print_scheduling(&ss); 596 } 597 598 // Check that the lock is not broken by a safepoint. 599 NoSafepointVerifier nsv; 600 ttyLocker ttyl; 601 if (xtty != nullptr) { 602 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 603 compile_id(), 604 is_osr_compilation() ? " compile_kind='osr'" : "", 605 phase_name); 606 xtty->print("%s", ss.as_string()); // print to tty would use xml escape encoding 607 xtty->tail("ideal"); 608 } else { 609 tty->print("%s", ss.as_string()); 610 } 611 } 612 #endif 613 614 // ============================================================================ 615 //------------------------------Compile standard------------------------------- 616 617 // Compile a method. entry_bci is -1 for normal compilations and indicates 618 // the continuation bci for on stack replacement. 619 620 621 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 622 Options options, DirectiveSet* directive) 623 : Phase(Compiler), 624 _compile_id(ci_env->compile_id()), 625 _options(options), 626 _method(target), 627 _entry_bci(osr_bci), 628 _ilt(nullptr), 629 _stub_function(nullptr), 630 _stub_name(nullptr), 631 _stub_entry_point(nullptr), 632 _max_node_limit(MaxNodeLimit), 633 _post_loop_opts_phase(false), 634 _inlining_progress(false), 635 _inlining_incrementally(false), 636 _do_cleanup(false), 637 _has_reserved_stack_access(target->has_reserved_stack_access()), 638 _has_circular_inline_type(false), 639 #ifndef PRODUCT 640 _igv_idx(0), 641 _trace_opto_output(directive->TraceOptoOutputOption), 642 #endif 643 _has_method_handle_invokes(false), 644 _clinit_barrier_on_entry(false), 645 _stress_seed(0), 646 _comp_arena(mtCompiler), 647 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 648 _env(ci_env), 649 _directive(directive), 650 _log(ci_env->log()), 651 _failure_reason(nullptr), 652 _intrinsics (comp_arena(), 0, 0, nullptr), 653 _macro_nodes (comp_arena(), 8, 0, nullptr), 654 _parse_predicates (comp_arena(), 8, 0, nullptr), 655 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr), 656 _expensive_nodes (comp_arena(), 8, 0, nullptr), 657 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 658 _inline_type_nodes (comp_arena(), 8, 0, nullptr), 659 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 660 _coarsened_locks (comp_arena(), 8, 0, nullptr), 661 _congraph(nullptr), 662 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 663 _unique(0), 664 _dead_node_count(0), 665 _dead_node_list(comp_arena()), 666 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 667 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 668 _node_arena(&_node_arena_one), 669 _mach_constant_base_node(nullptr), 670 _Compile_types(mtCompiler), 671 _initial_gvn(nullptr), 672 _igvn_worklist(nullptr), 673 _types(nullptr), 674 _node_hash(nullptr), 675 _late_inlines(comp_arena(), 2, 0, nullptr), 676 _string_late_inlines(comp_arena(), 2, 0, nullptr), 677 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 678 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 679 _late_inlines_pos(0), 680 _number_of_mh_late_inlines(0), 681 _oom(false), 682 _print_inlining_stream(new (mtCompiler) stringStream()), 683 _print_inlining_list(nullptr), 684 _print_inlining_idx(0), 685 _print_inlining_output(nullptr), 686 _replay_inline_data(nullptr), 687 _java_calls(0), 688 _inner_loops(0), 689 _interpreter_frame_size(0), 690 _output(nullptr) 691 #ifndef PRODUCT 692 , _in_dump_cnt(0) 693 #endif 694 { 695 C = this; 696 CompileWrapper cw(this); 697 698 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 699 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 700 701 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 702 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 703 // We can always print a disassembly, either abstract (hex dump) or 704 // with the help of a suitable hsdis library. Thus, we should not 705 // couple print_assembly and print_opto_assembly controls. 706 // But: always print opto and regular assembly on compile command 'print'. 707 bool print_assembly = directive->PrintAssemblyOption; 708 set_print_assembly(print_opto_assembly || print_assembly); 709 #else 710 set_print_assembly(false); // must initialize. 711 #endif 712 713 #ifndef PRODUCT 714 set_parsed_irreducible_loop(false); 715 #endif 716 717 if (directive->ReplayInlineOption) { 718 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 719 } 720 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 721 set_print_intrinsics(directive->PrintIntrinsicsOption); 722 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 723 724 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 725 // Make sure the method being compiled gets its own MDO, 726 // so we can at least track the decompile_count(). 727 // Need MDO to record RTM code generation state. 728 method()->ensure_method_data(); 729 } 730 731 Init(/*do_aliasing=*/ true); 732 733 print_compile_messages(); 734 735 _ilt = InlineTree::build_inline_tree_root(); 736 737 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 738 assert(num_alias_types() >= AliasIdxRaw, ""); 739 740 #define MINIMUM_NODE_HASH 1023 741 742 // GVN that will be run immediately on new nodes 743 uint estimated_size = method()->code_size()*4+64; 744 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 745 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 746 _types = new (comp_arena()) Type_Array(comp_arena()); 747 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 748 PhaseGVN gvn; 749 set_initial_gvn(&gvn); 750 751 print_inlining_init(); 752 { // Scope for timing the parser 753 TracePhase tp("parse", &timers[_t_parser]); 754 755 // Put top into the hash table ASAP. 756 initial_gvn()->transform_no_reclaim(top()); 757 758 // Set up tf(), start(), and find a CallGenerator. 759 CallGenerator* cg = nullptr; 760 if (is_osr_compilation()) { 761 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true)); 762 StartNode* s = new StartOSRNode(root(), tf()->domain_sig()); 763 initial_gvn()->set_type_bottom(s); 764 init_start(s); 765 cg = CallGenerator::for_osr(method(), entry_bci()); 766 } else { 767 // Normal case. 768 init_tf(TypeFunc::make(method())); 769 StartNode* s = new StartNode(root(), tf()->domain_cc()); 770 initial_gvn()->set_type_bottom(s); 771 init_start(s); 772 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 773 // With java.lang.ref.reference.get() we must go through the 774 // intrinsic - even when get() is the root 775 // method of the compile - so that, if necessary, the value in 776 // the referent field of the reference object gets recorded by 777 // the pre-barrier code. 778 cg = find_intrinsic(method(), false); 779 } 780 if (cg == nullptr) { 781 float past_uses = method()->interpreter_invocation_count(); 782 float expected_uses = past_uses; 783 cg = CallGenerator::for_inline(method(), expected_uses); 784 } 785 } 786 if (failing()) return; 787 if (cg == nullptr) { 788 const char* reason = InlineTree::check_can_parse(method()); 789 assert(reason != nullptr, "expect reason for parse failure"); 790 stringStream ss; 791 ss.print("cannot parse method: %s", reason); 792 record_method_not_compilable(ss.as_string()); 793 return; 794 } 795 796 gvn.set_type(root(), root()->bottom_type()); 797 798 JVMState* jvms = build_start_state(start(), tf()); 799 if ((jvms = cg->generate(jvms)) == nullptr) { 800 assert(failure_reason() != nullptr, "expect reason for parse failure"); 801 stringStream ss; 802 ss.print("method parse failed: %s", failure_reason()); 803 record_method_not_compilable(ss.as_string()); 804 return; 805 } 806 GraphKit kit(jvms); 807 808 if (!kit.stopped()) { 809 // Accept return values, and transfer control we know not where. 810 // This is done by a special, unique ReturnNode bound to root. 811 return_values(kit.jvms()); 812 } 813 814 if (kit.has_exceptions()) { 815 // Any exceptions that escape from this call must be rethrown 816 // to whatever caller is dynamically above us on the stack. 817 // This is done by a special, unique RethrowNode bound to root. 818 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 819 } 820 821 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 822 823 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 824 inline_string_calls(true); 825 } 826 827 if (failing()) return; 828 829 // Remove clutter produced by parsing. 830 if (!failing()) { 831 ResourceMark rm; 832 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 833 } 834 } 835 836 // Note: Large methods are capped off in do_one_bytecode(). 837 if (failing()) return; 838 839 // After parsing, node notes are no longer automagic. 840 // They must be propagated by register_new_node_with_optimizer(), 841 // clone(), or the like. 842 set_default_node_notes(nullptr); 843 844 #ifndef PRODUCT 845 if (should_print_igv(1)) { 846 _igv_printer->print_inlining(); 847 } 848 #endif 849 850 if (failing()) return; 851 NOT_PRODUCT( verify_graph_edges(); ) 852 853 // If any phase is randomized for stress testing, seed random number 854 // generation and log the seed for repeatability. 855 if (StressLCM || StressGCM || StressIGVN || StressCCP || StressIncrementalInlining) { 856 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 857 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 858 FLAG_SET_ERGO(StressSeed, _stress_seed); 859 } else { 860 _stress_seed = StressSeed; 861 } 862 if (_log != nullptr) { 863 _log->elem("stress_test seed='%u'", _stress_seed); 864 } 865 } 866 867 // Now optimize 868 Optimize(); 869 if (failing()) return; 870 NOT_PRODUCT( verify_graph_edges(); ) 871 872 #ifndef PRODUCT 873 if (should_print_ideal()) { 874 print_ideal_ir("print_ideal"); 875 } 876 #endif 877 878 #ifdef ASSERT 879 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 880 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 881 #endif 882 883 // Dump compilation data to replay it. 884 if (directive->DumpReplayOption) { 885 env()->dump_replay_data(_compile_id); 886 } 887 if (directive->DumpInlineOption && (ilt() != nullptr)) { 888 env()->dump_inline_data(_compile_id); 889 } 890 891 // Now that we know the size of all the monitors we can add a fixed slot 892 // for the original deopt pc. 893 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 894 if (needs_stack_repair()) { 895 // One extra slot for the special stack increment value 896 next_slot += 2; 897 } 898 // TODO 8284443 Only reserve extra slot if needed 899 if (InlineTypeReturnedAsFields) { 900 // One extra slot to hold the IsInit information for a nullable 901 // inline type return if we run out of registers. 902 next_slot += 2; 903 } 904 set_fixed_slots(next_slot); 905 906 // Compute when to use implicit null checks. Used by matching trap based 907 // nodes and NullCheck optimization. 908 set_allowed_deopt_reasons(); 909 910 // Now generate code 911 Code_Gen(); 912 } 913 914 //------------------------------Compile---------------------------------------- 915 // Compile a runtime stub 916 Compile::Compile( ciEnv* ci_env, 917 TypeFunc_generator generator, 918 address stub_function, 919 const char *stub_name, 920 int is_fancy_jump, 921 bool pass_tls, 922 bool return_pc, 923 DirectiveSet* directive) 924 : Phase(Compiler), 925 _compile_id(0), 926 _options(Options::for_runtime_stub()), 927 _method(nullptr), 928 _entry_bci(InvocationEntryBci), 929 _stub_function(stub_function), 930 _stub_name(stub_name), 931 _stub_entry_point(nullptr), 932 _max_node_limit(MaxNodeLimit), 933 _post_loop_opts_phase(false), 934 _inlining_progress(false), 935 _inlining_incrementally(false), 936 _has_reserved_stack_access(false), 937 _has_circular_inline_type(false), 938 #ifndef PRODUCT 939 _igv_idx(0), 940 _trace_opto_output(directive->TraceOptoOutputOption), 941 #endif 942 _has_method_handle_invokes(false), 943 _clinit_barrier_on_entry(false), 944 _stress_seed(0), 945 _comp_arena(mtCompiler), 946 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 947 _env(ci_env), 948 _directive(directive), 949 _log(ci_env->log()), 950 _failure_reason(nullptr), 951 _congraph(nullptr), 952 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 953 _unique(0), 954 _dead_node_count(0), 955 _dead_node_list(comp_arena()), 956 _node_arena_one(mtCompiler), 957 _node_arena_two(mtCompiler), 958 _node_arena(&_node_arena_one), 959 _mach_constant_base_node(nullptr), 960 _Compile_types(mtCompiler), 961 _initial_gvn(nullptr), 962 _igvn_worklist(nullptr), 963 _types(nullptr), 964 _node_hash(nullptr), 965 _number_of_mh_late_inlines(0), 966 _oom(false), 967 _print_inlining_stream(new (mtCompiler) stringStream()), 968 _print_inlining_list(nullptr), 969 _print_inlining_idx(0), 970 _print_inlining_output(nullptr), 971 _replay_inline_data(nullptr), 972 _java_calls(0), 973 _inner_loops(0), 974 _interpreter_frame_size(0), 975 _output(nullptr), 976 #ifndef PRODUCT 977 _in_dump_cnt(0), 978 #endif 979 _allowed_reasons(0) { 980 C = this; 981 982 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 983 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 984 985 #ifndef PRODUCT 986 set_print_assembly(PrintFrameConverterAssembly); 987 set_parsed_irreducible_loop(false); 988 #else 989 set_print_assembly(false); // Must initialize. 990 #endif 991 set_has_irreducible_loop(false); // no loops 992 993 CompileWrapper cw(this); 994 Init(/*do_aliasing=*/ false); 995 init_tf((*generator)()); 996 997 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 998 _types = new (comp_arena()) Type_Array(comp_arena()); 999 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 1000 { 1001 PhaseGVN gvn; 1002 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1003 gvn.transform_no_reclaim(top()); 1004 1005 GraphKit kit; 1006 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1007 } 1008 1009 NOT_PRODUCT( verify_graph_edges(); ) 1010 1011 Code_Gen(); 1012 } 1013 1014 //------------------------------Init------------------------------------------- 1015 // Prepare for a single compilation 1016 void Compile::Init(bool aliasing) { 1017 _do_aliasing = aliasing; 1018 _unique = 0; 1019 _regalloc = nullptr; 1020 1021 _tf = nullptr; // filled in later 1022 _top = nullptr; // cached later 1023 _matcher = nullptr; // filled in later 1024 _cfg = nullptr; // filled in later 1025 1026 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 1027 1028 _node_note_array = nullptr; 1029 _default_node_notes = nullptr; 1030 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1031 1032 _immutable_memory = nullptr; // filled in at first inquiry 1033 1034 #ifdef ASSERT 1035 _phase_optimize_finished = false; 1036 _exception_backedge = false; 1037 _type_verify = nullptr; 1038 #endif 1039 1040 // Globally visible Nodes 1041 // First set TOP to null to give safe behavior during creation of RootNode 1042 set_cached_top_node(nullptr); 1043 set_root(new RootNode()); 1044 // Now that you have a Root to point to, create the real TOP 1045 set_cached_top_node( new ConNode(Type::TOP) ); 1046 set_recent_alloc(nullptr, nullptr); 1047 1048 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1049 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1050 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1051 env()->set_dependencies(new Dependencies(env())); 1052 1053 _fixed_slots = 0; 1054 set_has_split_ifs(false); 1055 set_has_loops(false); // first approximation 1056 set_has_stringbuilder(false); 1057 set_has_boxed_value(false); 1058 _trap_can_recompile = false; // no traps emitted yet 1059 _major_progress = true; // start out assuming good things will happen 1060 set_has_unsafe_access(false); 1061 set_max_vector_size(0); 1062 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1063 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1064 set_decompile_count(0); 1065 1066 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1067 _loop_opts_cnt = LoopOptsCount; 1068 _has_flat_accesses = false; 1069 _flat_accesses_share_alias = true; 1070 _scalarize_in_safepoints = false; 1071 1072 set_do_inlining(Inline); 1073 set_max_inline_size(MaxInlineSize); 1074 set_freq_inline_size(FreqInlineSize); 1075 set_do_scheduling(OptoScheduling); 1076 1077 set_do_vector_loop(false); 1078 set_has_monitors(false); 1079 1080 if (AllowVectorizeOnDemand) { 1081 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1082 set_do_vector_loop(true); 1083 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());}) 1084 } else if (has_method() && method()->name() != 0 && 1085 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1086 set_do_vector_loop(true); 1087 } 1088 } 1089 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1090 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());}) 1091 1092 set_rtm_state(NoRTM); // No RTM lock eliding by default 1093 _max_node_limit = _directive->MaxNodeLimitOption; 1094 1095 #if INCLUDE_RTM_OPT 1096 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1097 int rtm_state = method()->method_data()->rtm_state(); 1098 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1099 // Don't generate RTM lock eliding code. 1100 set_rtm_state(NoRTM); 1101 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1102 // Generate RTM lock eliding code without abort ratio calculation code. 1103 set_rtm_state(UseRTM); 1104 } else if (UseRTMDeopt) { 1105 // Generate RTM lock eliding code and include abort ratio calculation 1106 // code if UseRTMDeopt is on. 1107 set_rtm_state(ProfileRTM); 1108 } 1109 } 1110 #endif 1111 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1112 set_clinit_barrier_on_entry(true); 1113 } 1114 if (debug_info()->recording_non_safepoints()) { 1115 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1116 (comp_arena(), 8, 0, nullptr)); 1117 set_default_node_notes(Node_Notes::make(this)); 1118 } 1119 1120 const int grow_ats = 16; 1121 _max_alias_types = grow_ats; 1122 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1123 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1124 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1125 { 1126 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1127 } 1128 // Initialize the first few types. 1129 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1130 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1131 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1132 _num_alias_types = AliasIdxRaw+1; 1133 // Zero out the alias type cache. 1134 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1135 // A null adr_type hits in the cache right away. Preload the right answer. 1136 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1137 } 1138 1139 //---------------------------init_start---------------------------------------- 1140 // Install the StartNode on this compile object. 1141 void Compile::init_start(StartNode* s) { 1142 if (failing()) 1143 return; // already failing 1144 assert(s == start(), ""); 1145 } 1146 1147 /** 1148 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1149 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1150 * the ideal graph. 1151 */ 1152 StartNode* Compile::start() const { 1153 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1154 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1155 Node* start = root()->fast_out(i); 1156 if (start->is_Start()) { 1157 return start->as_Start(); 1158 } 1159 } 1160 fatal("Did not find Start node!"); 1161 return nullptr; 1162 } 1163 1164 //-------------------------------immutable_memory------------------------------------- 1165 // Access immutable memory 1166 Node* Compile::immutable_memory() { 1167 if (_immutable_memory != nullptr) { 1168 return _immutable_memory; 1169 } 1170 StartNode* s = start(); 1171 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1172 Node *p = s->fast_out(i); 1173 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1174 _immutable_memory = p; 1175 return _immutable_memory; 1176 } 1177 } 1178 ShouldNotReachHere(); 1179 return nullptr; 1180 } 1181 1182 //----------------------set_cached_top_node------------------------------------ 1183 // Install the cached top node, and make sure Node::is_top works correctly. 1184 void Compile::set_cached_top_node(Node* tn) { 1185 if (tn != nullptr) verify_top(tn); 1186 Node* old_top = _top; 1187 _top = tn; 1188 // Calling Node::setup_is_top allows the nodes the chance to adjust 1189 // their _out arrays. 1190 if (_top != nullptr) _top->setup_is_top(); 1191 if (old_top != nullptr) old_top->setup_is_top(); 1192 assert(_top == nullptr || top()->is_top(), ""); 1193 } 1194 1195 #ifdef ASSERT 1196 uint Compile::count_live_nodes_by_graph_walk() { 1197 Unique_Node_List useful(comp_arena()); 1198 // Get useful node list by walking the graph. 1199 identify_useful_nodes(useful); 1200 return useful.size(); 1201 } 1202 1203 void Compile::print_missing_nodes() { 1204 1205 // Return if CompileLog is null and PrintIdealNodeCount is false. 1206 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1207 return; 1208 } 1209 1210 // This is an expensive function. It is executed only when the user 1211 // specifies VerifyIdealNodeCount option or otherwise knows the 1212 // additional work that needs to be done to identify reachable nodes 1213 // by walking the flow graph and find the missing ones using 1214 // _dead_node_list. 1215 1216 Unique_Node_List useful(comp_arena()); 1217 // Get useful node list by walking the graph. 1218 identify_useful_nodes(useful); 1219 1220 uint l_nodes = C->live_nodes(); 1221 uint l_nodes_by_walk = useful.size(); 1222 1223 if (l_nodes != l_nodes_by_walk) { 1224 if (_log != nullptr) { 1225 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1226 _log->stamp(); 1227 _log->end_head(); 1228 } 1229 VectorSet& useful_member_set = useful.member_set(); 1230 int last_idx = l_nodes_by_walk; 1231 for (int i = 0; i < last_idx; i++) { 1232 if (useful_member_set.test(i)) { 1233 if (_dead_node_list.test(i)) { 1234 if (_log != nullptr) { 1235 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1236 } 1237 if (PrintIdealNodeCount) { 1238 // Print the log message to tty 1239 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1240 useful.at(i)->dump(); 1241 } 1242 } 1243 } 1244 else if (! _dead_node_list.test(i)) { 1245 if (_log != nullptr) { 1246 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1247 } 1248 if (PrintIdealNodeCount) { 1249 // Print the log message to tty 1250 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1251 } 1252 } 1253 } 1254 if (_log != nullptr) { 1255 _log->tail("mismatched_nodes"); 1256 } 1257 } 1258 } 1259 void Compile::record_modified_node(Node* n) { 1260 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1261 _modified_nodes->push(n); 1262 } 1263 } 1264 1265 void Compile::remove_modified_node(Node* n) { 1266 if (_modified_nodes != nullptr) { 1267 _modified_nodes->remove(n); 1268 } 1269 } 1270 #endif 1271 1272 #ifndef PRODUCT 1273 void Compile::verify_top(Node* tn) const { 1274 if (tn != nullptr) { 1275 assert(tn->is_Con(), "top node must be a constant"); 1276 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1277 assert(tn->in(0) != nullptr, "must have live top node"); 1278 } 1279 } 1280 #endif 1281 1282 1283 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1284 1285 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1286 guarantee(arr != nullptr, ""); 1287 int num_blocks = arr->length(); 1288 if (grow_by < num_blocks) grow_by = num_blocks; 1289 int num_notes = grow_by * _node_notes_block_size; 1290 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1291 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1292 while (num_notes > 0) { 1293 arr->append(notes); 1294 notes += _node_notes_block_size; 1295 num_notes -= _node_notes_block_size; 1296 } 1297 assert(num_notes == 0, "exact multiple, please"); 1298 } 1299 1300 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1301 if (source == nullptr || dest == nullptr) return false; 1302 1303 if (dest->is_Con()) 1304 return false; // Do not push debug info onto constants. 1305 1306 #ifdef ASSERT 1307 // Leave a bread crumb trail pointing to the original node: 1308 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1309 dest->set_debug_orig(source); 1310 } 1311 #endif 1312 1313 if (node_note_array() == nullptr) 1314 return false; // Not collecting any notes now. 1315 1316 // This is a copy onto a pre-existing node, which may already have notes. 1317 // If both nodes have notes, do not overwrite any pre-existing notes. 1318 Node_Notes* source_notes = node_notes_at(source->_idx); 1319 if (source_notes == nullptr || source_notes->is_clear()) return false; 1320 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1321 if (dest_notes == nullptr || dest_notes->is_clear()) { 1322 return set_node_notes_at(dest->_idx, source_notes); 1323 } 1324 1325 Node_Notes merged_notes = (*source_notes); 1326 // The order of operations here ensures that dest notes will win... 1327 merged_notes.update_from(dest_notes); 1328 return set_node_notes_at(dest->_idx, &merged_notes); 1329 } 1330 1331 1332 //--------------------------allow_range_check_smearing------------------------- 1333 // Gating condition for coalescing similar range checks. 1334 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1335 // single covering check that is at least as strong as any of them. 1336 // If the optimization succeeds, the simplified (strengthened) range check 1337 // will always succeed. If it fails, we will deopt, and then give up 1338 // on the optimization. 1339 bool Compile::allow_range_check_smearing() const { 1340 // If this method has already thrown a range-check, 1341 // assume it was because we already tried range smearing 1342 // and it failed. 1343 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1344 return !already_trapped; 1345 } 1346 1347 1348 //------------------------------flatten_alias_type----------------------------- 1349 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1350 assert(do_aliasing(), "Aliasing should be enabled"); 1351 int offset = tj->offset(); 1352 TypePtr::PTR ptr = tj->ptr(); 1353 1354 // Known instance (scalarizable allocation) alias only with itself. 1355 bool is_known_inst = tj->isa_oopptr() != nullptr && 1356 tj->is_oopptr()->is_known_instance(); 1357 1358 // Process weird unsafe references. 1359 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1360 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass(); 1361 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops"); 1362 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1363 tj = TypeOopPtr::BOTTOM; 1364 ptr = tj->ptr(); 1365 offset = tj->offset(); 1366 } 1367 1368 // Array pointers need some flattening 1369 const TypeAryPtr* ta = tj->isa_aryptr(); 1370 if (ta && ta->is_stable()) { 1371 // Erase stability property for alias analysis. 1372 tj = ta = ta->cast_to_stable(false); 1373 } 1374 if (ta && ta->is_not_flat()) { 1375 // Erase not flat property for alias analysis. 1376 tj = ta = ta->cast_to_not_flat(false); 1377 } 1378 if (ta && ta->is_not_null_free()) { 1379 // Erase not null free property for alias analysis. 1380 tj = ta = ta->cast_to_not_null_free(false); 1381 } 1382 1383 if( ta && is_known_inst ) { 1384 if ( offset != Type::OffsetBot && 1385 offset > arrayOopDesc::length_offset_in_bytes() ) { 1386 offset = Type::OffsetBot; // Flatten constant access into array body only 1387 tj = ta = ta-> 1388 remove_speculative()-> 1389 cast_to_ptr_type(ptr)-> 1390 with_offset(offset); 1391 } 1392 } else if (ta) { 1393 // For arrays indexed by constant indices, we flatten the alias 1394 // space to include all of the array body. Only the header, klass 1395 // and array length can be accessed un-aliased. 1396 // For flat inline type array, each field has its own slice so 1397 // we must include the field offset. 1398 if( offset != Type::OffsetBot ) { 1399 if( ta->const_oop() ) { // MethodData* or Method* 1400 offset = Type::OffsetBot; // Flatten constant access into array body 1401 tj = ta = ta-> 1402 remove_speculative()-> 1403 cast_to_ptr_type(ptr)-> 1404 cast_to_exactness(false)-> 1405 with_offset(offset); 1406 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1407 // range is OK as-is. 1408 tj = ta = TypeAryPtr::RANGE; 1409 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1410 tj = TypeInstPtr::KLASS; // all klass loads look alike 1411 ta = TypeAryPtr::RANGE; // generic ignored junk 1412 ptr = TypePtr::BotPTR; 1413 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1414 tj = TypeInstPtr::MARK; 1415 ta = TypeAryPtr::RANGE; // generic ignored junk 1416 ptr = TypePtr::BotPTR; 1417 } else { // Random constant offset into array body 1418 offset = Type::OffsetBot; // Flatten constant access into array body 1419 tj = ta = ta-> 1420 remove_speculative()-> 1421 cast_to_ptr_type(ptr)-> 1422 cast_to_exactness(false)-> 1423 with_offset(offset); 1424 } 1425 } 1426 // Arrays of fixed size alias with arrays of unknown size. 1427 if (ta->size() != TypeInt::POS) { 1428 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1429 tj = ta = ta-> 1430 remove_speculative()-> 1431 cast_to_ptr_type(ptr)-> 1432 with_ary(tary)-> 1433 cast_to_exactness(false); 1434 } 1435 // Arrays of known objects become arrays of unknown objects. 1436 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1437 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1438 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset()); 1439 } 1440 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1441 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1442 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset()); 1443 } 1444 // Initially all flattened array accesses share a single slice 1445 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) { 1446 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true); 1447 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot)); 1448 } 1449 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1450 // cannot be distinguished by bytecode alone. 1451 if (ta->elem() == TypeInt::BOOL) { 1452 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1453 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1454 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1455 } 1456 // During the 2nd round of IterGVN, NotNull castings are removed. 1457 // Make sure the Bottom and NotNull variants alias the same. 1458 // Also, make sure exact and non-exact variants alias the same. 1459 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1460 tj = ta = ta-> 1461 remove_speculative()-> 1462 cast_to_ptr_type(TypePtr::BotPTR)-> 1463 cast_to_exactness(false)-> 1464 with_offset(offset); 1465 } 1466 } 1467 1468 // Oop pointers need some flattening 1469 const TypeInstPtr *to = tj->isa_instptr(); 1470 if (to && to != TypeOopPtr::BOTTOM) { 1471 ciInstanceKlass* ik = to->instance_klass(); 1472 if( ptr == TypePtr::Constant ) { 1473 if (ik != ciEnv::current()->Class_klass() || 1474 offset < ik->layout_helper_size_in_bytes()) { 1475 // No constant oop pointers (such as Strings); they alias with 1476 // unknown strings. 1477 assert(!is_known_inst, "not scalarizable allocation"); 1478 tj = to = to-> 1479 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1480 remove_speculative()-> 1481 cast_to_ptr_type(TypePtr::BotPTR)-> 1482 cast_to_exactness(false); 1483 } 1484 } else if( is_known_inst ) { 1485 tj = to; // Keep NotNull and klass_is_exact for instance type 1486 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1487 // During the 2nd round of IterGVN, NotNull castings are removed. 1488 // Make sure the Bottom and NotNull variants alias the same. 1489 // Also, make sure exact and non-exact variants alias the same. 1490 tj = to = to-> 1491 remove_speculative()-> 1492 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1493 cast_to_ptr_type(TypePtr::BotPTR)-> 1494 cast_to_exactness(false); 1495 } 1496 if (to->speculative() != nullptr) { 1497 tj = to = to->remove_speculative(); 1498 } 1499 // Canonicalize the holder of this field 1500 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1501 // First handle header references such as a LoadKlassNode, even if the 1502 // object's klass is unloaded at compile time (4965979). 1503 if (!is_known_inst) { // Do it only for non-instance types 1504 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset)); 1505 } 1506 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1507 // Static fields are in the space above the normal instance 1508 // fields in the java.lang.Class instance. 1509 if (ik != ciEnv::current()->Class_klass()) { 1510 to = nullptr; 1511 tj = TypeOopPtr::BOTTOM; 1512 offset = tj->offset(); 1513 } 1514 } else { 1515 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1516 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1517 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1518 if( is_known_inst ) { 1519 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, Type::Offset(offset), to->instance_id()); 1520 } else { 1521 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, Type::Offset(offset)); 1522 } 1523 } 1524 } 1525 } 1526 1527 // Klass pointers to object array klasses need some flattening 1528 const TypeKlassPtr *tk = tj->isa_klassptr(); 1529 if( tk ) { 1530 // If we are referencing a field within a Klass, we need 1531 // to assume the worst case of an Object. Both exact and 1532 // inexact types must flatten to the same alias class so 1533 // use NotNull as the PTR. 1534 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1535 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1536 env()->Object_klass(), 1537 Type::Offset(offset)); 1538 } 1539 1540 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1541 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1542 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1543 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset)); 1544 } else { 1545 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, Type::Offset(offset), tk->is_not_flat(), tk->is_not_null_free(), tk->is_null_free()); 1546 } 1547 } 1548 // Check for precise loads from the primary supertype array and force them 1549 // to the supertype cache alias index. Check for generic array loads from 1550 // the primary supertype array and also force them to the supertype cache 1551 // alias index. Since the same load can reach both, we need to merge 1552 // these 2 disparate memories into the same alias class. Since the 1553 // primary supertype array is read-only, there's no chance of confusion 1554 // where we bypass an array load and an array store. 1555 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1556 if (offset == Type::OffsetBot || 1557 (offset >= primary_supers_offset && 1558 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1559 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1560 offset = in_bytes(Klass::secondary_super_cache_offset()); 1561 tj = tk = tk->with_offset(offset); 1562 } 1563 } 1564 1565 // Flatten all Raw pointers together. 1566 if (tj->base() == Type::RawPtr) 1567 tj = TypeRawPtr::BOTTOM; 1568 1569 if (tj->base() == Type::AnyPtr) 1570 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1571 1572 offset = tj->offset(); 1573 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1574 1575 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1576 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1577 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1578 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1579 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1580 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1581 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1582 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1583 assert( tj->ptr() != TypePtr::TopPTR && 1584 tj->ptr() != TypePtr::AnyNull && 1585 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1586 // assert( tj->ptr() != TypePtr::Constant || 1587 // tj->base() == Type::RawPtr || 1588 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1589 1590 return tj; 1591 } 1592 1593 void Compile::AliasType::Init(int i, const TypePtr* at) { 1594 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1595 _index = i; 1596 _adr_type = at; 1597 _field = nullptr; 1598 _element = nullptr; 1599 _is_rewritable = true; // default 1600 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1601 if (atoop != nullptr && atoop->is_known_instance()) { 1602 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1603 _general_index = Compile::current()->get_alias_index(gt); 1604 } else { 1605 _general_index = 0; 1606 } 1607 } 1608 1609 BasicType Compile::AliasType::basic_type() const { 1610 if (element() != nullptr) { 1611 const Type* element = adr_type()->is_aryptr()->elem(); 1612 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1613 } if (field() != nullptr) { 1614 return field()->layout_type(); 1615 } else { 1616 return T_ILLEGAL; // unknown 1617 } 1618 } 1619 1620 //---------------------------------print_on------------------------------------ 1621 #ifndef PRODUCT 1622 void Compile::AliasType::print_on(outputStream* st) { 1623 if (index() < 10) 1624 st->print("@ <%d> ", index()); 1625 else st->print("@ <%d>", index()); 1626 st->print(is_rewritable() ? " " : " RO"); 1627 int offset = adr_type()->offset(); 1628 if (offset == Type::OffsetBot) 1629 st->print(" +any"); 1630 else st->print(" +%-3d", offset); 1631 st->print(" in "); 1632 adr_type()->dump_on(st); 1633 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1634 if (field() != nullptr && tjp) { 1635 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1636 tjp->offset() != field()->offset_in_bytes()) { 1637 st->print(" != "); 1638 field()->print(); 1639 st->print(" ***"); 1640 } 1641 } 1642 } 1643 1644 void print_alias_types() { 1645 Compile* C = Compile::current(); 1646 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1647 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1648 C->alias_type(idx)->print_on(tty); 1649 tty->cr(); 1650 } 1651 } 1652 #endif 1653 1654 1655 //----------------------------probe_alias_cache-------------------------------- 1656 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1657 intptr_t key = (intptr_t) adr_type; 1658 key ^= key >> logAliasCacheSize; 1659 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1660 } 1661 1662 1663 //-----------------------------grow_alias_types-------------------------------- 1664 void Compile::grow_alias_types() { 1665 const int old_ats = _max_alias_types; // how many before? 1666 const int new_ats = old_ats; // how many more? 1667 const int grow_ats = old_ats+new_ats; // how many now? 1668 _max_alias_types = grow_ats; 1669 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1670 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1671 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1672 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1673 } 1674 1675 1676 //--------------------------------find_alias_type------------------------------ 1677 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) { 1678 if (!do_aliasing()) { 1679 return alias_type(AliasIdxBot); 1680 } 1681 1682 AliasCacheEntry* ace = nullptr; 1683 if (!uncached) { 1684 ace = probe_alias_cache(adr_type); 1685 if (ace->_adr_type == adr_type) { 1686 return alias_type(ace->_index); 1687 } 1688 } 1689 1690 // Handle special cases. 1691 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1692 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1693 1694 // Do it the slow way. 1695 const TypePtr* flat = flatten_alias_type(adr_type); 1696 1697 #ifdef ASSERT 1698 { 1699 ResourceMark rm; 1700 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1701 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1702 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1703 Type::str(adr_type)); 1704 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1705 const TypeOopPtr* foop = flat->is_oopptr(); 1706 // Scalarizable allocations have exact klass always. 1707 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1708 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1709 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1710 Type::str(foop), Type::str(xoop)); 1711 } 1712 } 1713 #endif 1714 1715 int idx = AliasIdxTop; 1716 for (int i = 0; i < num_alias_types(); i++) { 1717 if (alias_type(i)->adr_type() == flat) { 1718 idx = i; 1719 break; 1720 } 1721 } 1722 1723 if (idx == AliasIdxTop) { 1724 if (no_create) return nullptr; 1725 // Grow the array if necessary. 1726 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1727 // Add a new alias type. 1728 idx = _num_alias_types++; 1729 _alias_types[idx]->Init(idx, flat); 1730 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1731 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1732 if (flat->isa_instptr()) { 1733 if (flat->offset() == java_lang_Class::klass_offset() 1734 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1735 alias_type(idx)->set_rewritable(false); 1736 } 1737 ciField* field = nullptr; 1738 if (flat->isa_aryptr()) { 1739 #ifdef ASSERT 1740 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1741 // (T_BYTE has the weakest alignment and size restrictions...) 1742 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1743 #endif 1744 const Type* elemtype = flat->is_aryptr()->elem(); 1745 if (flat->offset() == TypePtr::OffsetBot) { 1746 alias_type(idx)->set_element(elemtype); 1747 } 1748 int field_offset = flat->is_aryptr()->field_offset().get(); 1749 if (flat->is_flat() && 1750 field_offset != Type::OffsetBot) { 1751 ciInlineKlass* vk = elemtype->inline_klass(); 1752 field_offset += vk->first_field_offset(); 1753 field = vk->get_field_by_offset(field_offset, false); 1754 } 1755 } 1756 if (flat->isa_klassptr()) { 1757 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1758 alias_type(idx)->set_rewritable(false); 1759 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1760 alias_type(idx)->set_rewritable(false); 1761 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1762 alias_type(idx)->set_rewritable(false); 1763 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1764 alias_type(idx)->set_rewritable(false); 1765 if (flat->offset() == in_bytes(Klass::layout_helper_offset())) 1766 alias_type(idx)->set_rewritable(false); 1767 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1768 alias_type(idx)->set_rewritable(false); 1769 } 1770 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1771 // but the base pointer type is not distinctive enough to identify 1772 // references into JavaThread.) 1773 1774 // Check for final fields. 1775 const TypeInstPtr* tinst = flat->isa_instptr(); 1776 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1777 if (tinst->const_oop() != nullptr && 1778 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1779 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1780 // static field 1781 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1782 field = k->get_field_by_offset(tinst->offset(), true); 1783 } else if (tinst->is_inlinetypeptr()) { 1784 // Inline type field 1785 ciInlineKlass* vk = tinst->inline_klass(); 1786 field = vk->get_field_by_offset(tinst->offset(), false); 1787 } else { 1788 ciInstanceKlass *k = tinst->instance_klass(); 1789 field = k->get_field_by_offset(tinst->offset(), false); 1790 } 1791 } 1792 assert(field == nullptr || 1793 original_field == nullptr || 1794 (field->holder() == original_field->holder() && 1795 field->offset_in_bytes() == original_field->offset_in_bytes() && 1796 field->is_static() == original_field->is_static()), "wrong field?"); 1797 // Set field() and is_rewritable() attributes. 1798 if (field != nullptr) { 1799 alias_type(idx)->set_field(field); 1800 if (flat->isa_aryptr()) { 1801 // Fields of flat arrays are rewritable although they are declared final 1802 assert(flat->is_flat(), "must be a flat array"); 1803 alias_type(idx)->set_rewritable(true); 1804 } 1805 } 1806 } 1807 1808 // Fill the cache for next time. 1809 if (!uncached) { 1810 ace->_adr_type = adr_type; 1811 ace->_index = idx; 1812 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1813 1814 // Might as well try to fill the cache for the flattened version, too. 1815 AliasCacheEntry* face = probe_alias_cache(flat); 1816 if (face->_adr_type == nullptr) { 1817 face->_adr_type = flat; 1818 face->_index = idx; 1819 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1820 } 1821 } 1822 1823 return alias_type(idx); 1824 } 1825 1826 1827 Compile::AliasType* Compile::alias_type(ciField* field) { 1828 const TypeOopPtr* t; 1829 if (field->is_static()) 1830 t = TypeInstPtr::make(field->holder()->java_mirror()); 1831 else 1832 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1833 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1834 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1835 return atp; 1836 } 1837 1838 1839 //------------------------------have_alias_type-------------------------------- 1840 bool Compile::have_alias_type(const TypePtr* adr_type) { 1841 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1842 if (ace->_adr_type == adr_type) { 1843 return true; 1844 } 1845 1846 // Handle special cases. 1847 if (adr_type == nullptr) return true; 1848 if (adr_type == TypePtr::BOTTOM) return true; 1849 1850 return find_alias_type(adr_type, true, nullptr) != nullptr; 1851 } 1852 1853 //-----------------------------must_alias-------------------------------------- 1854 // True if all values of the given address type are in the given alias category. 1855 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1856 if (alias_idx == AliasIdxBot) return true; // the universal category 1857 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1858 if (alias_idx == AliasIdxTop) return false; // the empty category 1859 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1860 1861 // the only remaining possible overlap is identity 1862 int adr_idx = get_alias_index(adr_type); 1863 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1864 assert(adr_idx == alias_idx || 1865 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1866 && adr_type != TypeOopPtr::BOTTOM), 1867 "should not be testing for overlap with an unsafe pointer"); 1868 return adr_idx == alias_idx; 1869 } 1870 1871 //------------------------------can_alias-------------------------------------- 1872 // True if any values of the given address type are in the given alias category. 1873 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1874 if (alias_idx == AliasIdxTop) return false; // the empty category 1875 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1876 // Known instance doesn't alias with bottom memory 1877 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1878 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1879 1880 // the only remaining possible overlap is identity 1881 int adr_idx = get_alias_index(adr_type); 1882 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1883 return adr_idx == alias_idx; 1884 } 1885 1886 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1887 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1888 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1889 if (parse_predicate_count() == 0) { 1890 return; 1891 } 1892 for (int i = 0; i < parse_predicate_count(); i++) { 1893 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1894 parse_predicate->mark_useless(); 1895 igvn._worklist.push(parse_predicate); 1896 } 1897 _parse_predicates.clear(); 1898 } 1899 1900 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1901 if (!n->for_post_loop_opts_igvn()) { 1902 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1903 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1904 _for_post_loop_igvn.append(n); 1905 } 1906 } 1907 1908 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1909 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1910 _for_post_loop_igvn.remove(n); 1911 } 1912 1913 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1914 // Verify that all previous optimizations produced a valid graph 1915 // at least to this point, even if no loop optimizations were done. 1916 PhaseIdealLoop::verify(igvn); 1917 1918 C->set_post_loop_opts_phase(); // no more loop opts allowed 1919 1920 assert(!C->major_progress(), "not cleared"); 1921 1922 if (_for_post_loop_igvn.length() > 0) { 1923 while (_for_post_loop_igvn.length() > 0) { 1924 Node* n = _for_post_loop_igvn.pop(); 1925 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1926 igvn._worklist.push(n); 1927 } 1928 igvn.optimize(); 1929 if (failing()) return; 1930 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1931 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1932 1933 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1934 if (C->major_progress()) { 1935 C->clear_major_progress(); // ensure that major progress is now clear 1936 } 1937 } 1938 } 1939 1940 void Compile::add_inline_type(Node* n) { 1941 assert(n->is_InlineType(), "unexpected node"); 1942 _inline_type_nodes.push(n); 1943 } 1944 1945 void Compile::remove_inline_type(Node* n) { 1946 assert(n->is_InlineType(), "unexpected node"); 1947 if (_inline_type_nodes.contains(n)) { 1948 _inline_type_nodes.remove(n); 1949 } 1950 } 1951 1952 // Does the return value keep otherwise useless inline type allocations alive? 1953 static bool return_val_keeps_allocations_alive(Node* ret_val) { 1954 ResourceMark rm; 1955 Unique_Node_List wq; 1956 wq.push(ret_val); 1957 bool some_allocations = false; 1958 for (uint i = 0; i < wq.size(); i++) { 1959 Node* n = wq.at(i); 1960 if (n->outcnt() > 1) { 1961 // Some other use for the allocation 1962 return false; 1963 } else if (n->is_InlineType()) { 1964 wq.push(n->in(1)); 1965 } else if (n->is_Phi()) { 1966 for (uint j = 1; j < n->req(); j++) { 1967 wq.push(n->in(j)); 1968 } 1969 } else if (n->is_CheckCastPP() && 1970 n->in(1)->is_Proj() && 1971 n->in(1)->in(0)->is_Allocate()) { 1972 some_allocations = true; 1973 } else if (n->is_CheckCastPP()) { 1974 wq.push(n->in(1)); 1975 } 1976 } 1977 return some_allocations; 1978 } 1979 1980 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) { 1981 // Make sure that the return value does not keep an otherwise unused allocation alive 1982 if (tf()->returns_inline_type_as_fields()) { 1983 Node* ret = nullptr; 1984 for (uint i = 1; i < root()->req(); i++) { 1985 Node* in = root()->in(i); 1986 if (in->Opcode() == Op_Return) { 1987 assert(ret == nullptr, "only one return"); 1988 ret = in; 1989 } 1990 } 1991 if (ret != nullptr) { 1992 Node* ret_val = ret->in(TypeFunc::Parms); 1993 if (igvn.type(ret_val)->isa_oopptr() && 1994 return_val_keeps_allocations_alive(ret_val)) { 1995 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn)); 1996 assert(ret_val->outcnt() == 0, "should be dead now"); 1997 igvn.remove_dead_node(ret_val); 1998 } 1999 } 2000 } 2001 if (_inline_type_nodes.length() == 0) { 2002 return; 2003 } 2004 // Scalarize inline types in safepoint debug info. 2005 // Delay this until all inlining is over to avoid getting inconsistent debug info. 2006 set_scalarize_in_safepoints(true); 2007 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) { 2008 _inline_type_nodes.at(i)->as_InlineType()->make_scalar_in_safepoints(&igvn); 2009 } 2010 if (remove) { 2011 // Remove inline type nodes by replacing them with their oop input 2012 while (_inline_type_nodes.length() > 0) { 2013 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType(); 2014 if (vt->outcnt() == 0) { 2015 igvn.remove_dead_node(vt); 2016 continue; 2017 } 2018 for (DUIterator i = vt->outs(); vt->has_out(i); i++) { 2019 DEBUG_ONLY(bool must_be_buffered = false); 2020 Node* u = vt->out(i); 2021 // Check if any users are blackholes. If so, rewrite them to use either the 2022 // allocated buffer, or individual components, instead of the inline type node 2023 // that goes away. 2024 if (u->is_Blackhole()) { 2025 BlackholeNode* bh = u->as_Blackhole(); 2026 2027 // Unlink the old input 2028 int idx = bh->find_edge(vt); 2029 assert(idx != -1, "The edge should be there"); 2030 bh->del_req(idx); 2031 --i; 2032 2033 if (vt->is_allocated(&igvn)) { 2034 // Already has the allocated instance, blackhole that 2035 bh->add_req(vt->get_oop()); 2036 } else { 2037 // Not allocated yet, blackhole the components 2038 for (uint c = 0; c < vt->field_count(); c++) { 2039 bh->add_req(vt->field_value(c)); 2040 } 2041 } 2042 2043 // Node modified, record for IGVN 2044 igvn.record_for_igvn(bh); 2045 } 2046 #ifdef ASSERT 2047 // Verify that inline type is buffered when replacing by oop 2048 else if (u->is_InlineType()) { 2049 // InlineType uses don't need buffering because they are about to be replaced as well 2050 } else if (u->is_Phi()) { 2051 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through 2052 } else { 2053 must_be_buffered = true; 2054 } 2055 if (must_be_buffered && !vt->is_allocated(&igvn)) { 2056 vt->dump(0); 2057 u->dump(0); 2058 assert(false, "Should have been buffered"); 2059 } 2060 #endif 2061 } 2062 igvn.replace_node(vt, vt->get_oop()); 2063 } 2064 } 2065 igvn.optimize(); 2066 } 2067 2068 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) { 2069 if (!_has_flat_accesses) { 2070 return; 2071 } 2072 // Initially, all flat array accesses share the same slice to 2073 // keep dependencies with Object[] array accesses (that could be 2074 // to a flat array) correct. We're done with parsing so we 2075 // now know all flat array accesses in this compile 2076 // unit. Let's move flat array accesses to their own slice, 2077 // one per element field. This should help memory access 2078 // optimizations. 2079 ResourceMark rm; 2080 Unique_Node_List wq; 2081 wq.push(root()); 2082 2083 Node_List mergememnodes; 2084 Node_List memnodes; 2085 2086 // Alias index currently shared by all flat memory accesses 2087 int index = get_alias_index(TypeAryPtr::INLINES); 2088 2089 // Find MergeMem nodes and flat array accesses 2090 for (uint i = 0; i < wq.size(); i++) { 2091 Node* n = wq.at(i); 2092 if (n->is_Mem()) { 2093 const TypePtr* adr_type = nullptr; 2094 if (n->Opcode() == Op_StoreCM) { 2095 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type())); 2096 } else { 2097 adr_type = get_adr_type(get_alias_index(n->adr_type())); 2098 } 2099 if (adr_type == TypeAryPtr::INLINES) { 2100 memnodes.push(n); 2101 } 2102 } else if (n->is_MergeMem()) { 2103 MergeMemNode* mm = n->as_MergeMem(); 2104 if (mm->memory_at(index) != mm->base_memory()) { 2105 mergememnodes.push(n); 2106 } 2107 } 2108 for (uint j = 0; j < n->req(); j++) { 2109 Node* m = n->in(j); 2110 if (m != nullptr) { 2111 wq.push(m); 2112 } 2113 } 2114 } 2115 2116 if (memnodes.size() > 0) { 2117 _flat_accesses_share_alias = false; 2118 2119 // We are going to change the slice for the flat array 2120 // accesses so we need to clear the cache entries that refer to 2121 // them. 2122 for (uint i = 0; i < AliasCacheSize; i++) { 2123 AliasCacheEntry* ace = &_alias_cache[i]; 2124 if (ace->_adr_type != nullptr && 2125 ace->_adr_type->is_flat()) { 2126 ace->_adr_type = nullptr; 2127 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop 2128 } 2129 } 2130 2131 // Find what aliases we are going to add 2132 int start_alias = num_alias_types()-1; 2133 int stop_alias = 0; 2134 2135 for (uint i = 0; i < memnodes.size(); i++) { 2136 Node* m = memnodes.at(i); 2137 const TypePtr* adr_type = nullptr; 2138 if (m->Opcode() == Op_StoreCM) { 2139 adr_type = m->in(MemNode::OopStore)->adr_type(); 2140 if (adr_type != TypeAryPtr::INLINES) { 2141 // store was optimized out and we lost track of the adr_type 2142 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address), 2143 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore), 2144 get_alias_index(adr_type)); 2145 igvn.register_new_node_with_optimizer(clone); 2146 igvn.replace_node(m, clone); 2147 } 2148 } else { 2149 adr_type = m->adr_type(); 2150 #ifdef ASSERT 2151 m->as_Mem()->set_adr_type(adr_type); 2152 #endif 2153 } 2154 int idx = get_alias_index(adr_type); 2155 start_alias = MIN2(start_alias, idx); 2156 stop_alias = MAX2(stop_alias, idx); 2157 } 2158 2159 assert(stop_alias >= start_alias, "should have expanded aliases"); 2160 2161 Node_Stack stack(0); 2162 #ifdef ASSERT 2163 VectorSet seen(Thread::current()->resource_area()); 2164 #endif 2165 // Now let's fix the memory graph so each flat array access 2166 // is moved to the right slice. Start from the MergeMem nodes. 2167 uint last = unique(); 2168 for (uint i = 0; i < mergememnodes.size(); i++) { 2169 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem(); 2170 Node* n = current->memory_at(index); 2171 MergeMemNode* mm = nullptr; 2172 do { 2173 // Follow memory edges through memory accesses, phis and 2174 // narrow membars and push nodes on the stack. Once we hit 2175 // bottom memory, we pop element off the stack one at a 2176 // time, in reverse order, and move them to the right slice 2177 // by changing their memory edges. 2178 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) { 2179 assert(!seen.test_set(n->_idx), ""); 2180 // Uses (a load for instance) will need to be moved to the 2181 // right slice as well and will get a new memory state 2182 // that we don't know yet. The use could also be the 2183 // backedge of a loop. We put a place holder node between 2184 // the memory node and its uses. We replace that place 2185 // holder with the correct memory state once we know it, 2186 // i.e. when nodes are popped off the stack. Using the 2187 // place holder make the logic work in the presence of 2188 // loops. 2189 if (n->outcnt() > 1) { 2190 Node* place_holder = nullptr; 2191 assert(!n->has_out_with(Op_Node), ""); 2192 for (DUIterator k = n->outs(); n->has_out(k); k++) { 2193 Node* u = n->out(k); 2194 if (u != current && u->_idx < last) { 2195 bool success = false; 2196 for (uint l = 0; l < u->req(); l++) { 2197 if (!stack.is_empty() && u == stack.node() && l == stack.index()) { 2198 continue; 2199 } 2200 Node* in = u->in(l); 2201 if (in == n) { 2202 if (place_holder == nullptr) { 2203 place_holder = new Node(1); 2204 place_holder->init_req(0, n); 2205 } 2206 igvn.replace_input_of(u, l, place_holder); 2207 success = true; 2208 } 2209 } 2210 if (success) { 2211 --k; 2212 } 2213 } 2214 } 2215 } 2216 if (n->is_Phi()) { 2217 stack.push(n, 1); 2218 n = n->in(1); 2219 } else if (n->is_Mem()) { 2220 stack.push(n, n->req()); 2221 n = n->in(MemNode::Memory); 2222 } else { 2223 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, ""); 2224 stack.push(n, n->req()); 2225 n = n->in(0)->in(TypeFunc::Memory); 2226 } 2227 } else { 2228 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), ""); 2229 // Build a new MergeMem node to carry the new memory state 2230 // as we build it. IGVN should fold extraneous MergeMem 2231 // nodes. 2232 mm = MergeMemNode::make(n); 2233 igvn.register_new_node_with_optimizer(mm); 2234 while (stack.size() > 0) { 2235 Node* m = stack.node(); 2236 uint idx = stack.index(); 2237 if (m->is_Mem()) { 2238 // Move memory node to its new slice 2239 const TypePtr* adr_type = m->adr_type(); 2240 int alias = get_alias_index(adr_type); 2241 Node* prev = mm->memory_at(alias); 2242 igvn.replace_input_of(m, MemNode::Memory, prev); 2243 mm->set_memory_at(alias, m); 2244 } else if (m->is_Phi()) { 2245 // We need as many new phis as there are new aliases 2246 igvn.replace_input_of(m, idx, mm); 2247 if (idx == m->req()-1) { 2248 Node* r = m->in(0); 2249 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2250 const TypePtr* adr_type = get_adr_type(j); 2251 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) { 2252 continue; 2253 } 2254 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j)); 2255 igvn.register_new_node_with_optimizer(phi); 2256 for (uint k = 1; k < m->req(); k++) { 2257 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j)); 2258 } 2259 mm->set_memory_at(j, phi); 2260 } 2261 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2262 igvn.register_new_node_with_optimizer(base_phi); 2263 for (uint k = 1; k < m->req(); k++) { 2264 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory()); 2265 } 2266 mm->set_base_memory(base_phi); 2267 } 2268 } else { 2269 // This is a MemBarCPUOrder node from 2270 // Parse::array_load()/Parse::array_store(), in the 2271 // branch that handles flat arrays hidden under 2272 // an Object[] array. We also need one new membar per 2273 // new alias to keep the unknown access that the 2274 // membars protect properly ordered with accesses to 2275 // known flat array. 2276 assert(m->is_Proj(), "projection expected"); 2277 Node* ctrl = m->in(0)->in(TypeFunc::Control); 2278 igvn.replace_input_of(m->in(0), TypeFunc::Control, top()); 2279 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2280 const TypePtr* adr_type = get_adr_type(j); 2281 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) { 2282 continue; 2283 } 2284 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr); 2285 igvn.register_new_node_with_optimizer(mb); 2286 Node* mem = mm->memory_at(j); 2287 mb->init_req(TypeFunc::Control, ctrl); 2288 mb->init_req(TypeFunc::Memory, mem); 2289 ctrl = new ProjNode(mb, TypeFunc::Control); 2290 igvn.register_new_node_with_optimizer(ctrl); 2291 mem = new ProjNode(mb, TypeFunc::Memory); 2292 igvn.register_new_node_with_optimizer(mem); 2293 mm->set_memory_at(j, mem); 2294 } 2295 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl); 2296 } 2297 if (idx < m->req()-1) { 2298 idx += 1; 2299 stack.set_index(idx); 2300 n = m->in(idx); 2301 break; 2302 } 2303 // Take care of place holder nodes 2304 if (m->has_out_with(Op_Node)) { 2305 Node* place_holder = m->find_out_with(Op_Node); 2306 if (place_holder != nullptr) { 2307 Node* mm_clone = mm->clone(); 2308 igvn.register_new_node_with_optimizer(mm_clone); 2309 Node* hook = new Node(1); 2310 hook->init_req(0, mm); 2311 igvn.replace_node(place_holder, mm_clone); 2312 hook->destruct(&igvn); 2313 } 2314 assert(!m->has_out_with(Op_Node), "place holder should be gone now"); 2315 } 2316 stack.pop(); 2317 } 2318 } 2319 } while(stack.size() > 0); 2320 // Fix the memory state at the MergeMem we started from 2321 igvn.rehash_node_delayed(current); 2322 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2323 const TypePtr* adr_type = get_adr_type(j); 2324 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) { 2325 continue; 2326 } 2327 current->set_memory_at(j, mm); 2328 } 2329 current->set_memory_at(index, current->base_memory()); 2330 } 2331 igvn.optimize(); 2332 } 2333 print_method(PHASE_SPLIT_INLINES_ARRAY, 2); 2334 #ifdef ASSERT 2335 if (!_flat_accesses_share_alias) { 2336 wq.clear(); 2337 wq.push(root()); 2338 for (uint i = 0; i < wq.size(); i++) { 2339 Node* n = wq.at(i); 2340 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph"); 2341 for (uint j = 0; j < n->req(); j++) { 2342 Node* m = n->in(j); 2343 if (m != nullptr) { 2344 wq.push(m); 2345 } 2346 } 2347 } 2348 } 2349 #endif 2350 } 2351 2352 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 2353 if (OptimizeUnstableIf) { 2354 _unstable_if_traps.append(trap); 2355 } 2356 } 2357 2358 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 2359 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 2360 UnstableIfTrap* trap = _unstable_if_traps.at(i); 2361 Node* n = trap->uncommon_trap(); 2362 if (!useful.member(n)) { 2363 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 2364 } 2365 } 2366 } 2367 2368 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 2369 // or fold-compares case. Return true if succeed or not found. 2370 // 2371 // In rare cases, the found trap has been processed. It is too late to delete it. Return 2372 // false and ask fold-compares to yield. 2373 // 2374 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 2375 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 2376 // when deoptimization does happen. 2377 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 2378 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 2379 UnstableIfTrap* trap = _unstable_if_traps.at(i); 2380 if (trap->uncommon_trap() == unc) { 2381 if (yield && trap->modified()) { 2382 return false; 2383 } 2384 _unstable_if_traps.delete_at(i); 2385 break; 2386 } 2387 } 2388 return true; 2389 } 2390 2391 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 2392 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 2393 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 2394 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 2395 UnstableIfTrap* trap = _unstable_if_traps.at(i); 2396 CallStaticJavaNode* unc = trap->uncommon_trap(); 2397 int next_bci = trap->next_bci(); 2398 bool modified = trap->modified(); 2399 2400 if (next_bci != -1 && !modified) { 2401 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 2402 JVMState* jvms = unc->jvms(); 2403 ciMethod* method = jvms->method(); 2404 ciBytecodeStream iter(method); 2405 2406 iter.force_bci(jvms->bci()); 2407 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 2408 Bytecodes::Code c = iter.cur_bc(); 2409 Node* lhs = nullptr; 2410 Node* rhs = nullptr; 2411 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 2412 lhs = unc->peek_operand(0); 2413 rhs = unc->peek_operand(1); 2414 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 2415 lhs = unc->peek_operand(0); 2416 } 2417 2418 ResourceMark rm; 2419 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 2420 assert(live_locals.is_valid(), "broken liveness info"); 2421 int len = (int)live_locals.size(); 2422 2423 for (int i = 0; i < len; i++) { 2424 Node* local = unc->local(jvms, i); 2425 // kill local using the liveness of next_bci. 2426 // give up when the local looks like an operand to secure reexecution. 2427 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) { 2428 uint idx = jvms->locoff() + i; 2429 #ifdef ASSERT 2430 if (PrintOpto && Verbose) { 2431 tty->print("[unstable_if] kill local#%d: ", idx); 2432 local->dump(); 2433 tty->cr(); 2434 } 2435 #endif 2436 igvn.replace_input_of(unc, idx, top()); 2437 modified = true; 2438 } 2439 } 2440 } 2441 2442 // keep the modified trap for late query 2443 if (modified) { 2444 trap->set_modified(); 2445 } else { 2446 _unstable_if_traps.delete_at(i); 2447 } 2448 } 2449 igvn.optimize(); 2450 } 2451 2452 // StringOpts and late inlining of string methods 2453 void Compile::inline_string_calls(bool parse_time) { 2454 { 2455 // remove useless nodes to make the usage analysis simpler 2456 ResourceMark rm; 2457 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2458 } 2459 2460 { 2461 ResourceMark rm; 2462 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2463 PhaseStringOpts pso(initial_gvn()); 2464 print_method(PHASE_AFTER_STRINGOPTS, 3); 2465 } 2466 2467 // now inline anything that we skipped the first time around 2468 if (!parse_time) { 2469 _late_inlines_pos = _late_inlines.length(); 2470 } 2471 2472 while (_string_late_inlines.length() > 0) { 2473 CallGenerator* cg = _string_late_inlines.pop(); 2474 cg->do_late_inline(); 2475 if (failing()) return; 2476 } 2477 _string_late_inlines.trunc_to(0); 2478 } 2479 2480 // Late inlining of boxing methods 2481 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2482 if (_boxing_late_inlines.length() > 0) { 2483 assert(has_boxed_value(), "inconsistent"); 2484 2485 set_inlining_incrementally(true); 2486 2487 igvn_worklist()->ensure_empty(); // should be done with igvn 2488 2489 _late_inlines_pos = _late_inlines.length(); 2490 2491 while (_boxing_late_inlines.length() > 0) { 2492 CallGenerator* cg = _boxing_late_inlines.pop(); 2493 cg->do_late_inline(); 2494 if (failing()) return; 2495 } 2496 _boxing_late_inlines.trunc_to(0); 2497 2498 inline_incrementally_cleanup(igvn); 2499 2500 set_inlining_incrementally(false); 2501 } 2502 } 2503 2504 bool Compile::inline_incrementally_one() { 2505 assert(IncrementalInline, "incremental inlining should be on"); 2506 2507 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2508 2509 set_inlining_progress(false); 2510 set_do_cleanup(false); 2511 2512 for (int i = 0; i < _late_inlines.length(); i++) { 2513 _late_inlines_pos = i+1; 2514 CallGenerator* cg = _late_inlines.at(i); 2515 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2516 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2517 cg->do_late_inline(); 2518 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2519 if (failing()) { 2520 return false; 2521 } else if (inlining_progress()) { 2522 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2523 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2524 break; // process one call site at a time 2525 } 2526 } else { 2527 // Ignore late inline direct calls when inlining is not allowed. 2528 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2529 } 2530 } 2531 // Remove processed elements. 2532 _late_inlines.remove_till(_late_inlines_pos); 2533 _late_inlines_pos = 0; 2534 2535 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2536 2537 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2538 2539 set_inlining_progress(false); 2540 set_do_cleanup(false); 2541 2542 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2543 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2544 } 2545 2546 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2547 { 2548 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2549 ResourceMark rm; 2550 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2551 } 2552 { 2553 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2554 igvn.reset_from_gvn(initial_gvn()); 2555 igvn.optimize(); 2556 if (failing()) return; 2557 } 2558 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2559 } 2560 2561 // Perform incremental inlining until bound on number of live nodes is reached 2562 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2563 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2564 2565 set_inlining_incrementally(true); 2566 uint low_live_nodes = 0; 2567 2568 while (_late_inlines.length() > 0) { 2569 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2570 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2571 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2572 // PhaseIdealLoop is expensive so we only try it once we are 2573 // out of live nodes and we only try it again if the previous 2574 // helped got the number of nodes down significantly 2575 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2576 if (failing()) return; 2577 low_live_nodes = live_nodes(); 2578 _major_progress = true; 2579 } 2580 2581 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2582 bool do_print_inlining = print_inlining() || print_intrinsics(); 2583 if (do_print_inlining || log() != nullptr) { 2584 // Print inlining message for candidates that we couldn't inline for lack of space. 2585 for (int i = 0; i < _late_inlines.length(); i++) { 2586 CallGenerator* cg = _late_inlines.at(i); 2587 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2588 if (do_print_inlining) { 2589 cg->print_inlining_late(InliningResult::FAILURE, msg); 2590 } 2591 log_late_inline_failure(cg, msg); 2592 } 2593 } 2594 break; // finish 2595 } 2596 } 2597 2598 igvn_worklist()->ensure_empty(); // should be done with igvn 2599 2600 while (inline_incrementally_one()) { 2601 assert(!failing(), "inconsistent"); 2602 } 2603 if (failing()) return; 2604 2605 inline_incrementally_cleanup(igvn); 2606 2607 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2608 2609 if (failing()) return; 2610 2611 if (_late_inlines.length() == 0) { 2612 break; // no more progress 2613 } 2614 } 2615 2616 igvn_worklist()->ensure_empty(); // should be done with igvn 2617 2618 if (_string_late_inlines.length() > 0) { 2619 assert(has_stringbuilder(), "inconsistent"); 2620 2621 inline_string_calls(false); 2622 2623 if (failing()) return; 2624 2625 inline_incrementally_cleanup(igvn); 2626 } 2627 2628 set_inlining_incrementally(false); 2629 } 2630 2631 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2632 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2633 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2634 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2635 // as if "inlining_incrementally() == true" were set. 2636 assert(inlining_incrementally() == false, "not allowed"); 2637 #ifdef ASSERT 2638 Unique_Node_List* modified_nodes = _modified_nodes; 2639 _modified_nodes = nullptr; 2640 #endif 2641 assert(_late_inlines.length() > 0, "sanity"); 2642 2643 while (_late_inlines.length() > 0) { 2644 igvn_worklist()->ensure_empty(); // should be done with igvn 2645 2646 while (inline_incrementally_one()) { 2647 assert(!failing(), "inconsistent"); 2648 } 2649 if (failing()) return; 2650 2651 inline_incrementally_cleanup(igvn); 2652 } 2653 DEBUG_ONLY( _modified_nodes = modified_nodes; ) 2654 } 2655 2656 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2657 if (_loop_opts_cnt > 0) { 2658 while (major_progress() && (_loop_opts_cnt > 0)) { 2659 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2660 PhaseIdealLoop::optimize(igvn, mode); 2661 _loop_opts_cnt--; 2662 if (failing()) return false; 2663 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2664 } 2665 } 2666 return true; 2667 } 2668 2669 // Remove edges from "root" to each SafePoint at a backward branch. 2670 // They were inserted during parsing (see add_safepoint()) to make 2671 // infinite loops without calls or exceptions visible to root, i.e., 2672 // useful. 2673 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2674 Node *r = root(); 2675 if (r != nullptr) { 2676 for (uint i = r->req(); i < r->len(); ++i) { 2677 Node *n = r->in(i); 2678 if (n != nullptr && n->is_SafePoint()) { 2679 r->rm_prec(i); 2680 if (n->outcnt() == 0) { 2681 igvn.remove_dead_node(n); 2682 } 2683 --i; 2684 } 2685 } 2686 // Parsing may have added top inputs to the root node (Path 2687 // leading to the Halt node proven dead). Make sure we get a 2688 // chance to clean them up. 2689 igvn._worklist.push(r); 2690 igvn.optimize(); 2691 } 2692 } 2693 2694 //------------------------------Optimize--------------------------------------- 2695 // Given a graph, optimize it. 2696 void Compile::Optimize() { 2697 TracePhase tp("optimizer", &timers[_t_optimizer]); 2698 2699 #ifndef PRODUCT 2700 if (env()->break_at_compile()) { 2701 BREAKPOINT; 2702 } 2703 2704 #endif 2705 2706 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2707 #ifdef ASSERT 2708 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2709 #endif 2710 2711 ResourceMark rm; 2712 2713 print_inlining_reinit(); 2714 2715 NOT_PRODUCT( verify_graph_edges(); ) 2716 2717 print_method(PHASE_AFTER_PARSING, 1); 2718 2719 { 2720 // Iterative Global Value Numbering, including ideal transforms 2721 // Initialize IterGVN with types and values from parse-time GVN 2722 PhaseIterGVN igvn(initial_gvn()); 2723 #ifdef ASSERT 2724 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2725 #endif 2726 { 2727 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2728 igvn.optimize(); 2729 } 2730 2731 if (failing()) return; 2732 2733 print_method(PHASE_ITER_GVN1, 2); 2734 2735 process_for_unstable_if_traps(igvn); 2736 2737 if (failing()) return; 2738 2739 inline_incrementally(igvn); 2740 2741 print_method(PHASE_INCREMENTAL_INLINE, 2); 2742 2743 if (failing()) return; 2744 2745 if (eliminate_boxing()) { 2746 // Inline valueOf() methods now. 2747 inline_boxing_calls(igvn); 2748 2749 if (failing()) return; 2750 2751 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2752 inline_incrementally(igvn); 2753 } 2754 2755 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2756 2757 if (failing()) return; 2758 } 2759 2760 // Remove the speculative part of types and clean up the graph from 2761 // the extra CastPP nodes whose only purpose is to carry them. Do 2762 // that early so that optimizations are not disrupted by the extra 2763 // CastPP nodes. 2764 remove_speculative_types(igvn); 2765 2766 if (failing()) return; 2767 2768 // No more new expensive nodes will be added to the list from here 2769 // so keep only the actual candidates for optimizations. 2770 cleanup_expensive_nodes(igvn); 2771 2772 if (failing()) return; 2773 2774 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2775 if (EnableVectorSupport && has_vbox_nodes()) { 2776 TracePhase tp("", &timers[_t_vector]); 2777 PhaseVector pv(igvn); 2778 pv.optimize_vector_boxes(); 2779 if (failing()) return; 2780 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2781 } 2782 assert(!has_vbox_nodes(), "sanity"); 2783 2784 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2785 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2786 igvn_worklist()->ensure_empty(); // should be done with igvn 2787 { 2788 ResourceMark rm; 2789 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2790 } 2791 igvn.reset_from_gvn(initial_gvn()); 2792 igvn.optimize(); 2793 if (failing()) return; 2794 } 2795 2796 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2797 // safepoints 2798 remove_root_to_sfpts_edges(igvn); 2799 2800 // Process inline type nodes now that all inlining is over 2801 process_inline_types(igvn); 2802 2803 adjust_flat_array_access_aliases(igvn); 2804 2805 if (failing()) return; 2806 2807 // Perform escape analysis 2808 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2809 if (has_loops()) { 2810 // Cleanup graph (remove dead nodes). 2811 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2812 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2813 if (failing()) return; 2814 } 2815 bool progress; 2816 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2817 do { 2818 ConnectionGraph::do_analysis(this, &igvn); 2819 2820 if (failing()) return; 2821 2822 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2823 2824 // Optimize out fields loads from scalar replaceable allocations. 2825 igvn.optimize(); 2826 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2827 2828 if (failing()) return; 2829 2830 if (congraph() != nullptr && macro_count() > 0) { 2831 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2832 PhaseMacroExpand mexp(igvn); 2833 mexp.eliminate_macro_nodes(); 2834 if (failing()) return; 2835 2836 igvn.set_delay_transform(false); 2837 igvn.optimize(); 2838 if (failing()) return; 2839 2840 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2841 } 2842 2843 ConnectionGraph::verify_ram_nodes(this, root()); 2844 if (failing()) return; 2845 2846 progress = do_iterative_escape_analysis() && 2847 (macro_count() < mcount) && 2848 ConnectionGraph::has_candidates(this); 2849 // Try again if candidates exist and made progress 2850 // by removing some allocations and/or locks. 2851 } while (progress); 2852 } 2853 2854 // Loop transforms on the ideal graph. Range Check Elimination, 2855 // peeling, unrolling, etc. 2856 2857 // Set loop opts counter 2858 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2859 { 2860 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2861 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2862 _loop_opts_cnt--; 2863 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2864 if (failing()) return; 2865 } 2866 // Loop opts pass if partial peeling occurred in previous pass 2867 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2868 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2869 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2870 _loop_opts_cnt--; 2871 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2872 if (failing()) return; 2873 } 2874 // Loop opts pass for loop-unrolling before CCP 2875 if(major_progress() && (_loop_opts_cnt > 0)) { 2876 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2877 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2878 _loop_opts_cnt--; 2879 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2880 } 2881 if (!failing()) { 2882 // Verify that last round of loop opts produced a valid graph 2883 PhaseIdealLoop::verify(igvn); 2884 } 2885 } 2886 if (failing()) return; 2887 2888 // Conditional Constant Propagation; 2889 PhaseCCP ccp( &igvn ); 2890 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2891 { 2892 TracePhase tp("ccp", &timers[_t_ccp]); 2893 ccp.do_transform(); 2894 } 2895 print_method(PHASE_CCP1, 2); 2896 2897 assert( true, "Break here to ccp.dump_old2new_map()"); 2898 2899 // Iterative Global Value Numbering, including ideal transforms 2900 { 2901 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2902 igvn.reset_from_igvn(&ccp); 2903 igvn.optimize(); 2904 } 2905 print_method(PHASE_ITER_GVN2, 2); 2906 2907 if (failing()) return; 2908 2909 // Loop transforms on the ideal graph. Range Check Elimination, 2910 // peeling, unrolling, etc. 2911 if (!optimize_loops(igvn, LoopOptsDefault)) { 2912 return; 2913 } 2914 2915 if (failing()) return; 2916 2917 C->clear_major_progress(); // ensure that major progress is now clear 2918 2919 process_for_post_loop_opts_igvn(igvn); 2920 2921 if (failing()) return; 2922 2923 #ifdef ASSERT 2924 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2925 #endif 2926 2927 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2928 2929 if (_late_inlines.length() > 0) { 2930 // More opportunities to optimize virtual and MH calls. 2931 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2932 process_late_inline_calls_no_inline(igvn); 2933 } 2934 2935 { 2936 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2937 PhaseMacroExpand mex(igvn); 2938 if (mex.expand_macro_nodes()) { 2939 assert(failing(), "must bail out w/ explicit message"); 2940 return; 2941 } 2942 print_method(PHASE_MACRO_EXPANSION, 2); 2943 } 2944 2945 // Process inline type nodes again and remove them. From here 2946 // on we don't need to keep track of field values anymore. 2947 process_inline_types(igvn, /* remove= */ true); 2948 2949 { 2950 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2951 if (bs->expand_barriers(this, igvn)) { 2952 assert(failing(), "must bail out w/ explicit message"); 2953 return; 2954 } 2955 print_method(PHASE_BARRIER_EXPANSION, 2); 2956 } 2957 2958 if (C->max_vector_size() > 0) { 2959 C->optimize_logic_cones(igvn); 2960 igvn.optimize(); 2961 if (failing()) return; 2962 } 2963 2964 DEBUG_ONLY( _modified_nodes = nullptr; ) 2965 DEBUG_ONLY( _late_inlines.clear(); ) 2966 2967 assert(igvn._worklist.size() == 0, "not empty"); 2968 } // (End scope of igvn; run destructor if necessary for asserts.) 2969 2970 check_no_dead_use(); 2971 2972 process_print_inlining(); 2973 2974 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2975 // to remove hashes to unlock nodes for modifications. 2976 C->node_hash()->clear(); 2977 2978 // A method with only infinite loops has no edges entering loops from root 2979 { 2980 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2981 if (final_graph_reshaping()) { 2982 assert(failing(), "must bail out w/ explicit message"); 2983 return; 2984 } 2985 } 2986 2987 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2988 DEBUG_ONLY(set_phase_optimize_finished();) 2989 } 2990 2991 #ifdef ASSERT 2992 void Compile::check_no_dead_use() const { 2993 ResourceMark rm; 2994 Unique_Node_List wq; 2995 wq.push(root()); 2996 for (uint i = 0; i < wq.size(); ++i) { 2997 Node* n = wq.at(i); 2998 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2999 Node* u = n->fast_out(j); 3000 if (u->outcnt() == 0 && !u->is_Con()) { 3001 u->dump(); 3002 fatal("no reachable node should have no use"); 3003 } 3004 wq.push(u); 3005 } 3006 } 3007 } 3008 #endif 3009 3010 void Compile::inline_vector_reboxing_calls() { 3011 if (C->_vector_reboxing_late_inlines.length() > 0) { 3012 _late_inlines_pos = C->_late_inlines.length(); 3013 while (_vector_reboxing_late_inlines.length() > 0) { 3014 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 3015 cg->do_late_inline(); 3016 if (failing()) return; 3017 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 3018 } 3019 _vector_reboxing_late_inlines.trunc_to(0); 3020 } 3021 } 3022 3023 bool Compile::has_vbox_nodes() { 3024 if (C->_vector_reboxing_late_inlines.length() > 0) { 3025 return true; 3026 } 3027 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 3028 Node * n = C->macro_node(macro_idx); 3029 assert(n->is_macro(), "only macro nodes expected here"); 3030 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 3031 return true; 3032 } 3033 } 3034 return false; 3035 } 3036 3037 //---------------------------- Bitwise operation packing optimization --------------------------- 3038 3039 static bool is_vector_unary_bitwise_op(Node* n) { 3040 return n->Opcode() == Op_XorV && 3041 VectorNode::is_vector_bitwise_not_pattern(n); 3042 } 3043 3044 static bool is_vector_binary_bitwise_op(Node* n) { 3045 switch (n->Opcode()) { 3046 case Op_AndV: 3047 case Op_OrV: 3048 return true; 3049 3050 case Op_XorV: 3051 return !is_vector_unary_bitwise_op(n); 3052 3053 default: 3054 return false; 3055 } 3056 } 3057 3058 static bool is_vector_ternary_bitwise_op(Node* n) { 3059 return n->Opcode() == Op_MacroLogicV; 3060 } 3061 3062 static bool is_vector_bitwise_op(Node* n) { 3063 return is_vector_unary_bitwise_op(n) || 3064 is_vector_binary_bitwise_op(n) || 3065 is_vector_ternary_bitwise_op(n); 3066 } 3067 3068 static bool is_vector_bitwise_cone_root(Node* n) { 3069 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 3070 return false; 3071 } 3072 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3073 if (is_vector_bitwise_op(n->fast_out(i))) { 3074 return false; 3075 } 3076 } 3077 return true; 3078 } 3079 3080 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 3081 uint cnt = 0; 3082 if (is_vector_bitwise_op(n)) { 3083 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 3084 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 3085 for (uint i = 1; i < inp_cnt; i++) { 3086 Node* in = n->in(i); 3087 bool skip = VectorNode::is_all_ones_vector(in); 3088 if (!skip && !inputs.member(in)) { 3089 inputs.push(in); 3090 cnt++; 3091 } 3092 } 3093 assert(cnt <= 1, "not unary"); 3094 } else { 3095 uint last_req = inp_cnt; 3096 if (is_vector_ternary_bitwise_op(n)) { 3097 last_req = inp_cnt - 1; // skip last input 3098 } 3099 for (uint i = 1; i < last_req; i++) { 3100 Node* def = n->in(i); 3101 if (!inputs.member(def)) { 3102 inputs.push(def); 3103 cnt++; 3104 } 3105 } 3106 } 3107 } else { // not a bitwise operations 3108 if (!inputs.member(n)) { 3109 inputs.push(n); 3110 cnt++; 3111 } 3112 } 3113 return cnt; 3114 } 3115 3116 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 3117 Unique_Node_List useful_nodes; 3118 C->identify_useful_nodes(useful_nodes); 3119 3120 for (uint i = 0; i < useful_nodes.size(); i++) { 3121 Node* n = useful_nodes.at(i); 3122 if (is_vector_bitwise_cone_root(n)) { 3123 list.push(n); 3124 } 3125 } 3126 } 3127 3128 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 3129 const TypeVect* vt, 3130 Unique_Node_List& partition, 3131 Unique_Node_List& inputs) { 3132 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 3133 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 3134 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 3135 3136 Node* in1 = inputs.at(0); 3137 Node* in2 = inputs.at(1); 3138 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 3139 3140 uint func = compute_truth_table(partition, inputs); 3141 3142 Node* pn = partition.at(partition.size() - 1); 3143 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 3144 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 3145 } 3146 3147 static uint extract_bit(uint func, uint pos) { 3148 return (func & (1 << pos)) >> pos; 3149 } 3150 3151 // 3152 // A macro logic node represents a truth table. It has 4 inputs, 3153 // First three inputs corresponds to 3 columns of a truth table 3154 // and fourth input captures the logic function. 3155 // 3156 // eg. fn = (in1 AND in2) OR in3; 3157 // 3158 // MacroNode(in1,in2,in3,fn) 3159 // 3160 // ----------------- 3161 // in1 in2 in3 fn 3162 // ----------------- 3163 // 0 0 0 0 3164 // 0 0 1 1 3165 // 0 1 0 0 3166 // 0 1 1 1 3167 // 1 0 0 0 3168 // 1 0 1 1 3169 // 1 1 0 1 3170 // 1 1 1 1 3171 // 3172 3173 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 3174 int res = 0; 3175 for (int i = 0; i < 8; i++) { 3176 int bit1 = extract_bit(in1, i); 3177 int bit2 = extract_bit(in2, i); 3178 int bit3 = extract_bit(in3, i); 3179 3180 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 3181 int func_bit = extract_bit(func, func_bit_pos); 3182 3183 res |= func_bit << i; 3184 } 3185 return res; 3186 } 3187 3188 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 3189 assert(n != nullptr, ""); 3190 assert(eval_map.contains(n), "absent"); 3191 return *(eval_map.get(n)); 3192 } 3193 3194 static void eval_operands(Node* n, 3195 uint& func1, uint& func2, uint& func3, 3196 ResourceHashtable<Node*,uint>& eval_map) { 3197 assert(is_vector_bitwise_op(n), ""); 3198 3199 if (is_vector_unary_bitwise_op(n)) { 3200 Node* opnd = n->in(1); 3201 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 3202 opnd = n->in(2); 3203 } 3204 func1 = eval_operand(opnd, eval_map); 3205 } else if (is_vector_binary_bitwise_op(n)) { 3206 func1 = eval_operand(n->in(1), eval_map); 3207 func2 = eval_operand(n->in(2), eval_map); 3208 } else { 3209 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 3210 func1 = eval_operand(n->in(1), eval_map); 3211 func2 = eval_operand(n->in(2), eval_map); 3212 func3 = eval_operand(n->in(3), eval_map); 3213 } 3214 } 3215 3216 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 3217 assert(inputs.size() <= 3, "sanity"); 3218 ResourceMark rm; 3219 uint res = 0; 3220 ResourceHashtable<Node*,uint> eval_map; 3221 3222 // Populate precomputed functions for inputs. 3223 // Each input corresponds to one column of 3 input truth-table. 3224 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 3225 0xCC, // (_, b, _) -> b 3226 0xF0 }; // (a, _, _) -> a 3227 for (uint i = 0; i < inputs.size(); i++) { 3228 eval_map.put(inputs.at(i), input_funcs[2-i]); 3229 } 3230 3231 for (uint i = 0; i < partition.size(); i++) { 3232 Node* n = partition.at(i); 3233 3234 uint func1 = 0, func2 = 0, func3 = 0; 3235 eval_operands(n, func1, func2, func3, eval_map); 3236 3237 switch (n->Opcode()) { 3238 case Op_OrV: 3239 assert(func3 == 0, "not binary"); 3240 res = func1 | func2; 3241 break; 3242 case Op_AndV: 3243 assert(func3 == 0, "not binary"); 3244 res = func1 & func2; 3245 break; 3246 case Op_XorV: 3247 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 3248 assert(func2 == 0 && func3 == 0, "not unary"); 3249 res = (~func1) & 0xFF; 3250 } else { 3251 assert(func3 == 0, "not binary"); 3252 res = func1 ^ func2; 3253 } 3254 break; 3255 case Op_MacroLogicV: 3256 // Ordering of inputs may change during evaluation of sub-tree 3257 // containing MacroLogic node as a child node, thus a re-evaluation 3258 // makes sure that function is evaluated in context of current 3259 // inputs. 3260 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 3261 break; 3262 3263 default: assert(false, "not supported: %s", n->Name()); 3264 } 3265 assert(res <= 0xFF, "invalid"); 3266 eval_map.put(n, res); 3267 } 3268 return res; 3269 } 3270 3271 // Criteria under which nodes gets packed into a macro logic node:- 3272 // 1) Parent and both child nodes are all unmasked or masked with 3273 // same predicates. 3274 // 2) Masked parent can be packed with left child if it is predicated 3275 // and both have same predicates. 3276 // 3) Masked parent can be packed with right child if its un-predicated 3277 // or has matching predication condition. 3278 // 4) An unmasked parent can be packed with an unmasked child. 3279 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 3280 assert(partition.size() == 0, "not empty"); 3281 assert(inputs.size() == 0, "not empty"); 3282 if (is_vector_ternary_bitwise_op(n)) { 3283 return false; 3284 } 3285 3286 bool is_unary_op = is_vector_unary_bitwise_op(n); 3287 if (is_unary_op) { 3288 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 3289 return false; // too few inputs 3290 } 3291 3292 bool pack_left_child = true; 3293 bool pack_right_child = true; 3294 3295 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 3296 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 3297 3298 int left_child_input_cnt = 0; 3299 int right_child_input_cnt = 0; 3300 3301 bool parent_is_predicated = n->is_predicated_vector(); 3302 bool left_child_predicated = n->in(1)->is_predicated_vector(); 3303 bool right_child_predicated = n->in(2)->is_predicated_vector(); 3304 3305 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 3306 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 3307 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 3308 3309 do { 3310 if (pack_left_child && left_child_LOP && 3311 ((!parent_is_predicated && !left_child_predicated) || 3312 ((parent_is_predicated && left_child_predicated && 3313 parent_pred == left_child_pred)))) { 3314 partition.push(n->in(1)); 3315 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 3316 } else { 3317 inputs.push(n->in(1)); 3318 left_child_input_cnt = 1; 3319 } 3320 3321 if (pack_right_child && right_child_LOP && 3322 (!right_child_predicated || 3323 (right_child_predicated && parent_is_predicated && 3324 parent_pred == right_child_pred))) { 3325 partition.push(n->in(2)); 3326 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 3327 } else { 3328 inputs.push(n->in(2)); 3329 right_child_input_cnt = 1; 3330 } 3331 3332 if (inputs.size() > 3) { 3333 assert(partition.size() > 0, ""); 3334 inputs.clear(); 3335 partition.clear(); 3336 if (left_child_input_cnt > right_child_input_cnt) { 3337 pack_left_child = false; 3338 } else { 3339 pack_right_child = false; 3340 } 3341 } else { 3342 break; 3343 } 3344 } while(true); 3345 3346 if(partition.size()) { 3347 partition.push(n); 3348 } 3349 3350 return (partition.size() == 2 || partition.size() == 3) && 3351 (inputs.size() == 2 || inputs.size() == 3); 3352 } 3353 3354 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 3355 assert(is_vector_bitwise_op(n), "not a root"); 3356 3357 visited.set(n->_idx); 3358 3359 // 1) Do a DFS walk over the logic cone. 3360 for (uint i = 1; i < n->req(); i++) { 3361 Node* in = n->in(i); 3362 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 3363 process_logic_cone_root(igvn, in, visited); 3364 } 3365 } 3366 3367 // 2) Bottom up traversal: Merge node[s] with 3368 // the parent to form macro logic node. 3369 Unique_Node_List partition; 3370 Unique_Node_List inputs; 3371 if (compute_logic_cone(n, partition, inputs)) { 3372 const TypeVect* vt = n->bottom_type()->is_vect(); 3373 Node* pn = partition.at(partition.size() - 1); 3374 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 3375 if (mask == nullptr || 3376 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 3377 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 3378 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 3379 igvn.replace_node(n, macro_logic); 3380 } 3381 } 3382 } 3383 3384 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 3385 ResourceMark rm; 3386 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 3387 Unique_Node_List list; 3388 collect_logic_cone_roots(list); 3389 3390 while (list.size() > 0) { 3391 Node* n = list.pop(); 3392 const TypeVect* vt = n->bottom_type()->is_vect(); 3393 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 3394 if (supported) { 3395 VectorSet visited(comp_arena()); 3396 process_logic_cone_root(igvn, n, visited); 3397 } 3398 } 3399 } 3400 } 3401 3402 //------------------------------Code_Gen--------------------------------------- 3403 // Given a graph, generate code for it 3404 void Compile::Code_Gen() { 3405 if (failing()) { 3406 return; 3407 } 3408 3409 // Perform instruction selection. You might think we could reclaim Matcher 3410 // memory PDQ, but actually the Matcher is used in generating spill code. 3411 // Internals of the Matcher (including some VectorSets) must remain live 3412 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 3413 // set a bit in reclaimed memory. 3414 3415 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3416 // nodes. Mapping is only valid at the root of each matched subtree. 3417 NOT_PRODUCT( verify_graph_edges(); ) 3418 3419 Matcher matcher; 3420 _matcher = &matcher; 3421 { 3422 TracePhase tp("matcher", &timers[_t_matcher]); 3423 matcher.match(); 3424 if (failing()) { 3425 return; 3426 } 3427 } 3428 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3429 // nodes. Mapping is only valid at the root of each matched subtree. 3430 NOT_PRODUCT( verify_graph_edges(); ) 3431 3432 // If you have too many nodes, or if matching has failed, bail out 3433 check_node_count(0, "out of nodes matching instructions"); 3434 if (failing()) { 3435 return; 3436 } 3437 3438 print_method(PHASE_MATCHING, 2); 3439 3440 // Build a proper-looking CFG 3441 PhaseCFG cfg(node_arena(), root(), matcher); 3442 _cfg = &cfg; 3443 { 3444 TracePhase tp("scheduler", &timers[_t_scheduler]); 3445 bool success = cfg.do_global_code_motion(); 3446 if (!success) { 3447 return; 3448 } 3449 3450 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 3451 NOT_PRODUCT( verify_graph_edges(); ) 3452 cfg.verify(); 3453 } 3454 3455 PhaseChaitin regalloc(unique(), cfg, matcher, false); 3456 _regalloc = ®alloc; 3457 { 3458 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 3459 // Perform register allocation. After Chaitin, use-def chains are 3460 // no longer accurate (at spill code) and so must be ignored. 3461 // Node->LRG->reg mappings are still accurate. 3462 _regalloc->Register_Allocate(); 3463 3464 // Bail out if the allocator builds too many nodes 3465 if (failing()) { 3466 return; 3467 } 3468 } 3469 3470 // Prior to register allocation we kept empty basic blocks in case the 3471 // the allocator needed a place to spill. After register allocation we 3472 // are not adding any new instructions. If any basic block is empty, we 3473 // can now safely remove it. 3474 { 3475 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 3476 cfg.remove_empty_blocks(); 3477 if (do_freq_based_layout()) { 3478 PhaseBlockLayout layout(cfg); 3479 } else { 3480 cfg.set_loop_alignment(); 3481 } 3482 cfg.fixup_flow(); 3483 cfg.remove_unreachable_blocks(); 3484 cfg.verify_dominator_tree(); 3485 } 3486 3487 // Apply peephole optimizations 3488 if( OptoPeephole ) { 3489 TracePhase tp("peephole", &timers[_t_peephole]); 3490 PhasePeephole peep( _regalloc, cfg); 3491 peep.do_transform(); 3492 } 3493 3494 // Do late expand if CPU requires this. 3495 if (Matcher::require_postalloc_expand) { 3496 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 3497 cfg.postalloc_expand(_regalloc); 3498 } 3499 3500 // Convert Nodes to instruction bits in a buffer 3501 { 3502 TracePhase tp("output", &timers[_t_output]); 3503 PhaseOutput output; 3504 output.Output(); 3505 if (failing()) return; 3506 output.install(); 3507 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3508 } 3509 3510 // He's dead, Jim. 3511 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3512 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3513 } 3514 3515 //------------------------------Final_Reshape_Counts--------------------------- 3516 // This class defines counters to help identify when a method 3517 // may/must be executed using hardware with only 24-bit precision. 3518 struct Final_Reshape_Counts : public StackObj { 3519 int _call_count; // count non-inlined 'common' calls 3520 int _float_count; // count float ops requiring 24-bit precision 3521 int _double_count; // count double ops requiring more precision 3522 int _java_call_count; // count non-inlined 'java' calls 3523 int _inner_loop_count; // count loops which need alignment 3524 VectorSet _visited; // Visitation flags 3525 Node_List _tests; // Set of IfNodes & PCTableNodes 3526 3527 Final_Reshape_Counts() : 3528 _call_count(0), _float_count(0), _double_count(0), 3529 _java_call_count(0), _inner_loop_count(0) { } 3530 3531 void inc_call_count () { _call_count ++; } 3532 void inc_float_count () { _float_count ++; } 3533 void inc_double_count() { _double_count++; } 3534 void inc_java_call_count() { _java_call_count++; } 3535 void inc_inner_loop_count() { _inner_loop_count++; } 3536 3537 int get_call_count () const { return _call_count ; } 3538 int get_float_count () const { return _float_count ; } 3539 int get_double_count() const { return _double_count; } 3540 int get_java_call_count() const { return _java_call_count; } 3541 int get_inner_loop_count() const { return _inner_loop_count; } 3542 }; 3543 3544 // Eliminate trivially redundant StoreCMs and accumulate their 3545 // precedence edges. 3546 void Compile::eliminate_redundant_card_marks(Node* n) { 3547 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3548 if (n->in(MemNode::Address)->outcnt() > 1) { 3549 // There are multiple users of the same address so it might be 3550 // possible to eliminate some of the StoreCMs 3551 Node* mem = n->in(MemNode::Memory); 3552 Node* adr = n->in(MemNode::Address); 3553 Node* val = n->in(MemNode::ValueIn); 3554 Node* prev = n; 3555 bool done = false; 3556 // Walk the chain of StoreCMs eliminating ones that match. As 3557 // long as it's a chain of single users then the optimization is 3558 // safe. Eliminating partially redundant StoreCMs would require 3559 // cloning copies down the other paths. 3560 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3561 if (adr == mem->in(MemNode::Address) && 3562 val == mem->in(MemNode::ValueIn)) { 3563 // redundant StoreCM 3564 if (mem->req() > MemNode::OopStore) { 3565 // Hasn't been processed by this code yet. 3566 n->add_prec(mem->in(MemNode::OopStore)); 3567 } else { 3568 // Already converted to precedence edge 3569 for (uint i = mem->req(); i < mem->len(); i++) { 3570 // Accumulate any precedence edges 3571 if (mem->in(i) != nullptr) { 3572 n->add_prec(mem->in(i)); 3573 } 3574 } 3575 // Everything above this point has been processed. 3576 done = true; 3577 } 3578 // Eliminate the previous StoreCM 3579 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3580 assert(mem->outcnt() == 0, "should be dead"); 3581 mem->disconnect_inputs(this); 3582 } else { 3583 prev = mem; 3584 } 3585 mem = prev->in(MemNode::Memory); 3586 } 3587 } 3588 } 3589 3590 3591 //------------------------------final_graph_reshaping_impl---------------------- 3592 // Implement items 1-5 from final_graph_reshaping below. 3593 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3594 3595 if ( n->outcnt() == 0 ) return; // dead node 3596 uint nop = n->Opcode(); 3597 3598 // Check for 2-input instruction with "last use" on right input. 3599 // Swap to left input. Implements item (2). 3600 if( n->req() == 3 && // two-input instruction 3601 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3602 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3603 n->in(2)->outcnt() == 1 &&// right use IS a last use 3604 !n->in(2)->is_Con() ) { // right use is not a constant 3605 // Check for commutative opcode 3606 switch( nop ) { 3607 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3608 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3609 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3610 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3611 case Op_AndL: case Op_XorL: case Op_OrL: 3612 case Op_AndI: case Op_XorI: case Op_OrI: { 3613 // Move "last use" input to left by swapping inputs 3614 n->swap_edges(1, 2); 3615 break; 3616 } 3617 default: 3618 break; 3619 } 3620 } 3621 3622 #ifdef ASSERT 3623 if( n->is_Mem() ) { 3624 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3625 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3626 // oop will be recorded in oop map if load crosses safepoint 3627 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3628 LoadNode::is_immutable_value(n->in(MemNode::Address))), 3629 "raw memory operations should have control edge"); 3630 } 3631 if (n->is_MemBar()) { 3632 MemBarNode* mb = n->as_MemBar(); 3633 if (mb->trailing_store() || mb->trailing_load_store()) { 3634 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3635 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3636 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3637 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3638 } else if (mb->leading()) { 3639 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3640 } 3641 } 3642 #endif 3643 // Count FPU ops and common calls, implements item (3) 3644 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3645 if (!gc_handled) { 3646 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3647 } 3648 3649 // Collect CFG split points 3650 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3651 frc._tests.push(n); 3652 } 3653 } 3654 3655 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3656 switch( nop ) { 3657 // Count all float operations that may use FPU 3658 case Op_AddF: 3659 case Op_SubF: 3660 case Op_MulF: 3661 case Op_DivF: 3662 case Op_NegF: 3663 case Op_ModF: 3664 case Op_ConvI2F: 3665 case Op_ConF: 3666 case Op_CmpF: 3667 case Op_CmpF3: 3668 case Op_StoreF: 3669 case Op_LoadF: 3670 // case Op_ConvL2F: // longs are split into 32-bit halves 3671 frc.inc_float_count(); 3672 break; 3673 3674 case Op_ConvF2D: 3675 case Op_ConvD2F: 3676 frc.inc_float_count(); 3677 frc.inc_double_count(); 3678 break; 3679 3680 // Count all double operations that may use FPU 3681 case Op_AddD: 3682 case Op_SubD: 3683 case Op_MulD: 3684 case Op_DivD: 3685 case Op_NegD: 3686 case Op_ModD: 3687 case Op_ConvI2D: 3688 case Op_ConvD2I: 3689 // case Op_ConvL2D: // handled by leaf call 3690 // case Op_ConvD2L: // handled by leaf call 3691 case Op_ConD: 3692 case Op_CmpD: 3693 case Op_CmpD3: 3694 case Op_StoreD: 3695 case Op_LoadD: 3696 case Op_LoadD_unaligned: 3697 frc.inc_double_count(); 3698 break; 3699 case Op_Opaque1: // Remove Opaque Nodes before matching 3700 case Op_Opaque3: 3701 n->subsume_by(n->in(1), this); 3702 break; 3703 case Op_CallStaticJava: 3704 case Op_CallJava: 3705 case Op_CallDynamicJava: 3706 frc.inc_java_call_count(); // Count java call site; 3707 case Op_CallRuntime: 3708 case Op_CallLeaf: 3709 case Op_CallLeafVector: 3710 case Op_CallLeafNoFP: { 3711 assert (n->is_Call(), ""); 3712 CallNode *call = n->as_Call(); 3713 // Count call sites where the FP mode bit would have to be flipped. 3714 // Do not count uncommon runtime calls: 3715 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3716 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3717 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3718 frc.inc_call_count(); // Count the call site 3719 } else { // See if uncommon argument is shared 3720 Node *n = call->in(TypeFunc::Parms); 3721 int nop = n->Opcode(); 3722 // Clone shared simple arguments to uncommon calls, item (1). 3723 if (n->outcnt() > 1 && 3724 !n->is_Proj() && 3725 nop != Op_CreateEx && 3726 nop != Op_CheckCastPP && 3727 nop != Op_DecodeN && 3728 nop != Op_DecodeNKlass && 3729 !n->is_Mem() && 3730 !n->is_Phi()) { 3731 Node *x = n->clone(); 3732 call->set_req(TypeFunc::Parms, x); 3733 } 3734 } 3735 break; 3736 } 3737 3738 case Op_StoreCM: 3739 { 3740 // Convert OopStore dependence into precedence edge 3741 Node* prec = n->in(MemNode::OopStore); 3742 n->del_req(MemNode::OopStore); 3743 if (prec->is_MergeMem()) { 3744 MergeMemNode* mm = prec->as_MergeMem(); 3745 Node* base = mm->base_memory(); 3746 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) { 3747 const TypePtr* adr_type = get_adr_type(i); 3748 if (adr_type->is_flat()) { 3749 Node* m = mm->memory_at(i); 3750 n->add_prec(m); 3751 } 3752 } 3753 if (mm->outcnt() == 0) { 3754 mm->disconnect_inputs(this); 3755 } 3756 } else { 3757 n->add_prec(prec); 3758 } 3759 eliminate_redundant_card_marks(n); 3760 } 3761 3762 // fall through 3763 3764 case Op_StoreB: 3765 case Op_StoreC: 3766 case Op_StoreI: 3767 case Op_StoreL: 3768 case Op_CompareAndSwapB: 3769 case Op_CompareAndSwapS: 3770 case Op_CompareAndSwapI: 3771 case Op_CompareAndSwapL: 3772 case Op_CompareAndSwapP: 3773 case Op_CompareAndSwapN: 3774 case Op_WeakCompareAndSwapB: 3775 case Op_WeakCompareAndSwapS: 3776 case Op_WeakCompareAndSwapI: 3777 case Op_WeakCompareAndSwapL: 3778 case Op_WeakCompareAndSwapP: 3779 case Op_WeakCompareAndSwapN: 3780 case Op_CompareAndExchangeB: 3781 case Op_CompareAndExchangeS: 3782 case Op_CompareAndExchangeI: 3783 case Op_CompareAndExchangeL: 3784 case Op_CompareAndExchangeP: 3785 case Op_CompareAndExchangeN: 3786 case Op_GetAndAddS: 3787 case Op_GetAndAddB: 3788 case Op_GetAndAddI: 3789 case Op_GetAndAddL: 3790 case Op_GetAndSetS: 3791 case Op_GetAndSetB: 3792 case Op_GetAndSetI: 3793 case Op_GetAndSetL: 3794 case Op_GetAndSetP: 3795 case Op_GetAndSetN: 3796 case Op_StoreP: 3797 case Op_StoreN: 3798 case Op_StoreNKlass: 3799 case Op_LoadB: 3800 case Op_LoadUB: 3801 case Op_LoadUS: 3802 case Op_LoadI: 3803 case Op_LoadKlass: 3804 case Op_LoadNKlass: 3805 case Op_LoadL: 3806 case Op_LoadL_unaligned: 3807 case Op_LoadP: 3808 case Op_LoadN: 3809 case Op_LoadRange: 3810 case Op_LoadS: 3811 break; 3812 3813 case Op_AddP: { // Assert sane base pointers 3814 Node *addp = n->in(AddPNode::Address); 3815 assert( !addp->is_AddP() || 3816 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3817 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3818 "Base pointers must match (addp %u)", addp->_idx ); 3819 #ifdef _LP64 3820 if ((UseCompressedOops || UseCompressedClassPointers) && 3821 addp->Opcode() == Op_ConP && 3822 addp == n->in(AddPNode::Base) && 3823 n->in(AddPNode::Offset)->is_Con()) { 3824 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3825 // on the platform and on the compressed oops mode. 3826 // Use addressing with narrow klass to load with offset on x86. 3827 // Some platforms can use the constant pool to load ConP. 3828 // Do this transformation here since IGVN will convert ConN back to ConP. 3829 const Type* t = addp->bottom_type(); 3830 bool is_oop = t->isa_oopptr() != nullptr; 3831 bool is_klass = t->isa_klassptr() != nullptr; 3832 3833 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3834 (is_klass && Matcher::const_klass_prefer_decode())) { 3835 Node* nn = nullptr; 3836 3837 int op = is_oop ? Op_ConN : Op_ConNKlass; 3838 3839 // Look for existing ConN node of the same exact type. 3840 Node* r = root(); 3841 uint cnt = r->outcnt(); 3842 for (uint i = 0; i < cnt; i++) { 3843 Node* m = r->raw_out(i); 3844 if (m!= nullptr && m->Opcode() == op && 3845 m->bottom_type()->make_ptr() == t) { 3846 nn = m; 3847 break; 3848 } 3849 } 3850 if (nn != nullptr) { 3851 // Decode a narrow oop to match address 3852 // [R12 + narrow_oop_reg<<3 + offset] 3853 if (is_oop) { 3854 nn = new DecodeNNode(nn, t); 3855 } else { 3856 nn = new DecodeNKlassNode(nn, t); 3857 } 3858 // Check for succeeding AddP which uses the same Base. 3859 // Otherwise we will run into the assertion above when visiting that guy. 3860 for (uint i = 0; i < n->outcnt(); ++i) { 3861 Node *out_i = n->raw_out(i); 3862 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3863 out_i->set_req(AddPNode::Base, nn); 3864 #ifdef ASSERT 3865 for (uint j = 0; j < out_i->outcnt(); ++j) { 3866 Node *out_j = out_i->raw_out(j); 3867 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3868 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3869 } 3870 #endif 3871 } 3872 } 3873 n->set_req(AddPNode::Base, nn); 3874 n->set_req(AddPNode::Address, nn); 3875 if (addp->outcnt() == 0) { 3876 addp->disconnect_inputs(this); 3877 } 3878 } 3879 } 3880 } 3881 #endif 3882 break; 3883 } 3884 3885 case Op_CastPP: { 3886 // Remove CastPP nodes to gain more freedom during scheduling but 3887 // keep the dependency they encode as control or precedence edges 3888 // (if control is set already) on memory operations. Some CastPP 3889 // nodes don't have a control (don't carry a dependency): skip 3890 // those. 3891 if (n->in(0) != nullptr) { 3892 ResourceMark rm; 3893 Unique_Node_List wq; 3894 wq.push(n); 3895 for (uint next = 0; next < wq.size(); ++next) { 3896 Node *m = wq.at(next); 3897 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3898 Node* use = m->fast_out(i); 3899 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3900 use->ensure_control_or_add_prec(n->in(0)); 3901 } else { 3902 switch(use->Opcode()) { 3903 case Op_AddP: 3904 case Op_DecodeN: 3905 case Op_DecodeNKlass: 3906 case Op_CheckCastPP: 3907 case Op_CastPP: 3908 wq.push(use); 3909 break; 3910 } 3911 } 3912 } 3913 } 3914 } 3915 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3916 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3917 Node* in1 = n->in(1); 3918 const Type* t = n->bottom_type(); 3919 Node* new_in1 = in1->clone(); 3920 new_in1->as_DecodeN()->set_type(t); 3921 3922 if (!Matcher::narrow_oop_use_complex_address()) { 3923 // 3924 // x86, ARM and friends can handle 2 adds in addressing mode 3925 // and Matcher can fold a DecodeN node into address by using 3926 // a narrow oop directly and do implicit null check in address: 3927 // 3928 // [R12 + narrow_oop_reg<<3 + offset] 3929 // NullCheck narrow_oop_reg 3930 // 3931 // On other platforms (Sparc) we have to keep new DecodeN node and 3932 // use it to do implicit null check in address: 3933 // 3934 // decode_not_null narrow_oop_reg, base_reg 3935 // [base_reg + offset] 3936 // NullCheck base_reg 3937 // 3938 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3939 // to keep the information to which null check the new DecodeN node 3940 // corresponds to use it as value in implicit_null_check(). 3941 // 3942 new_in1->set_req(0, n->in(0)); 3943 } 3944 3945 n->subsume_by(new_in1, this); 3946 if (in1->outcnt() == 0) { 3947 in1->disconnect_inputs(this); 3948 } 3949 } else { 3950 n->subsume_by(n->in(1), this); 3951 if (n->outcnt() == 0) { 3952 n->disconnect_inputs(this); 3953 } 3954 } 3955 break; 3956 } 3957 #ifdef _LP64 3958 case Op_CmpP: 3959 // Do this transformation here to preserve CmpPNode::sub() and 3960 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3961 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3962 Node* in1 = n->in(1); 3963 Node* in2 = n->in(2); 3964 if (!in1->is_DecodeNarrowPtr()) { 3965 in2 = in1; 3966 in1 = n->in(2); 3967 } 3968 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3969 3970 Node* new_in2 = nullptr; 3971 if (in2->is_DecodeNarrowPtr()) { 3972 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3973 new_in2 = in2->in(1); 3974 } else if (in2->Opcode() == Op_ConP) { 3975 const Type* t = in2->bottom_type(); 3976 if (t == TypePtr::NULL_PTR) { 3977 assert(in1->is_DecodeN(), "compare klass to null?"); 3978 // Don't convert CmpP null check into CmpN if compressed 3979 // oops implicit null check is not generated. 3980 // This will allow to generate normal oop implicit null check. 3981 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3982 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3983 // 3984 // This transformation together with CastPP transformation above 3985 // will generated code for implicit null checks for compressed oops. 3986 // 3987 // The original code after Optimize() 3988 // 3989 // LoadN memory, narrow_oop_reg 3990 // decode narrow_oop_reg, base_reg 3991 // CmpP base_reg, nullptr 3992 // CastPP base_reg // NotNull 3993 // Load [base_reg + offset], val_reg 3994 // 3995 // after these transformations will be 3996 // 3997 // LoadN memory, narrow_oop_reg 3998 // CmpN narrow_oop_reg, nullptr 3999 // decode_not_null narrow_oop_reg, base_reg 4000 // Load [base_reg + offset], val_reg 4001 // 4002 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 4003 // since narrow oops can be used in debug info now (see the code in 4004 // final_graph_reshaping_walk()). 4005 // 4006 // At the end the code will be matched to 4007 // on x86: 4008 // 4009 // Load_narrow_oop memory, narrow_oop_reg 4010 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 4011 // NullCheck narrow_oop_reg 4012 // 4013 // and on sparc: 4014 // 4015 // Load_narrow_oop memory, narrow_oop_reg 4016 // decode_not_null narrow_oop_reg, base_reg 4017 // Load [base_reg + offset], val_reg 4018 // NullCheck base_reg 4019 // 4020 } else if (t->isa_oopptr()) { 4021 new_in2 = ConNode::make(t->make_narrowoop()); 4022 } else if (t->isa_klassptr()) { 4023 new_in2 = ConNode::make(t->make_narrowklass()); 4024 } 4025 } 4026 if (new_in2 != nullptr) { 4027 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 4028 n->subsume_by(cmpN, this); 4029 if (in1->outcnt() == 0) { 4030 in1->disconnect_inputs(this); 4031 } 4032 if (in2->outcnt() == 0) { 4033 in2->disconnect_inputs(this); 4034 } 4035 } 4036 } 4037 break; 4038 4039 case Op_DecodeN: 4040 case Op_DecodeNKlass: 4041 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 4042 // DecodeN could be pinned when it can't be fold into 4043 // an address expression, see the code for Op_CastPP above. 4044 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 4045 break; 4046 4047 case Op_EncodeP: 4048 case Op_EncodePKlass: { 4049 Node* in1 = n->in(1); 4050 if (in1->is_DecodeNarrowPtr()) { 4051 n->subsume_by(in1->in(1), this); 4052 } else if (in1->Opcode() == Op_ConP) { 4053 const Type* t = in1->bottom_type(); 4054 if (t == TypePtr::NULL_PTR) { 4055 assert(t->isa_oopptr(), "null klass?"); 4056 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 4057 } else if (t->isa_oopptr()) { 4058 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 4059 } else if (t->isa_klassptr()) { 4060 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 4061 } 4062 } 4063 if (in1->outcnt() == 0) { 4064 in1->disconnect_inputs(this); 4065 } 4066 break; 4067 } 4068 4069 case Op_Proj: { 4070 if (OptimizeStringConcat || IncrementalInline) { 4071 ProjNode* proj = n->as_Proj(); 4072 if (proj->_is_io_use) { 4073 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 4074 // Separate projections were used for the exception path which 4075 // are normally removed by a late inline. If it wasn't inlined 4076 // then they will hang around and should just be replaced with 4077 // the original one. Merge them. 4078 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 4079 if (non_io_proj != nullptr) { 4080 proj->subsume_by(non_io_proj , this); 4081 } 4082 } 4083 } 4084 break; 4085 } 4086 4087 case Op_Phi: 4088 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 4089 // The EncodeP optimization may create Phi with the same edges 4090 // for all paths. It is not handled well by Register Allocator. 4091 Node* unique_in = n->in(1); 4092 assert(unique_in != nullptr, ""); 4093 uint cnt = n->req(); 4094 for (uint i = 2; i < cnt; i++) { 4095 Node* m = n->in(i); 4096 assert(m != nullptr, ""); 4097 if (unique_in != m) 4098 unique_in = nullptr; 4099 } 4100 if (unique_in != nullptr) { 4101 n->subsume_by(unique_in, this); 4102 } 4103 } 4104 break; 4105 4106 #endif 4107 4108 #ifdef ASSERT 4109 case Op_CastII: 4110 // Verify that all range check dependent CastII nodes were removed. 4111 if (n->isa_CastII()->has_range_check()) { 4112 n->dump(3); 4113 assert(false, "Range check dependent CastII node was not removed"); 4114 } 4115 break; 4116 #endif 4117 4118 case Op_ModI: 4119 if (UseDivMod) { 4120 // Check if a%b and a/b both exist 4121 Node* d = n->find_similar(Op_DivI); 4122 if (d) { 4123 // Replace them with a fused divmod if supported 4124 if (Matcher::has_match_rule(Op_DivModI)) { 4125 DivModINode* divmod = DivModINode::make(n); 4126 d->subsume_by(divmod->div_proj(), this); 4127 n->subsume_by(divmod->mod_proj(), this); 4128 } else { 4129 // replace a%b with a-((a/b)*b) 4130 Node* mult = new MulINode(d, d->in(2)); 4131 Node* sub = new SubINode(d->in(1), mult); 4132 n->subsume_by(sub, this); 4133 } 4134 } 4135 } 4136 break; 4137 4138 case Op_ModL: 4139 if (UseDivMod) { 4140 // Check if a%b and a/b both exist 4141 Node* d = n->find_similar(Op_DivL); 4142 if (d) { 4143 // Replace them with a fused divmod if supported 4144 if (Matcher::has_match_rule(Op_DivModL)) { 4145 DivModLNode* divmod = DivModLNode::make(n); 4146 d->subsume_by(divmod->div_proj(), this); 4147 n->subsume_by(divmod->mod_proj(), this); 4148 } else { 4149 // replace a%b with a-((a/b)*b) 4150 Node* mult = new MulLNode(d, d->in(2)); 4151 Node* sub = new SubLNode(d->in(1), mult); 4152 n->subsume_by(sub, this); 4153 } 4154 } 4155 } 4156 break; 4157 4158 case Op_UModI: 4159 if (UseDivMod) { 4160 // Check if a%b and a/b both exist 4161 Node* d = n->find_similar(Op_UDivI); 4162 if (d) { 4163 // Replace them with a fused unsigned divmod if supported 4164 if (Matcher::has_match_rule(Op_UDivModI)) { 4165 UDivModINode* divmod = UDivModINode::make(n); 4166 d->subsume_by(divmod->div_proj(), this); 4167 n->subsume_by(divmod->mod_proj(), this); 4168 } else { 4169 // replace a%b with a-((a/b)*b) 4170 Node* mult = new MulINode(d, d->in(2)); 4171 Node* sub = new SubINode(d->in(1), mult); 4172 n->subsume_by(sub, this); 4173 } 4174 } 4175 } 4176 break; 4177 4178 case Op_UModL: 4179 if (UseDivMod) { 4180 // Check if a%b and a/b both exist 4181 Node* d = n->find_similar(Op_UDivL); 4182 if (d) { 4183 // Replace them with a fused unsigned divmod if supported 4184 if (Matcher::has_match_rule(Op_UDivModL)) { 4185 UDivModLNode* divmod = UDivModLNode::make(n); 4186 d->subsume_by(divmod->div_proj(), this); 4187 n->subsume_by(divmod->mod_proj(), this); 4188 } else { 4189 // replace a%b with a-((a/b)*b) 4190 Node* mult = new MulLNode(d, d->in(2)); 4191 Node* sub = new SubLNode(d->in(1), mult); 4192 n->subsume_by(sub, this); 4193 } 4194 } 4195 } 4196 break; 4197 4198 case Op_LoadVector: 4199 case Op_StoreVector: 4200 case Op_LoadVectorGather: 4201 case Op_StoreVectorScatter: 4202 case Op_LoadVectorGatherMasked: 4203 case Op_StoreVectorScatterMasked: 4204 case Op_VectorCmpMasked: 4205 case Op_VectorMaskGen: 4206 case Op_LoadVectorMasked: 4207 case Op_StoreVectorMasked: 4208 break; 4209 4210 case Op_AddReductionVI: 4211 case Op_AddReductionVL: 4212 case Op_AddReductionVF: 4213 case Op_AddReductionVD: 4214 case Op_MulReductionVI: 4215 case Op_MulReductionVL: 4216 case Op_MulReductionVF: 4217 case Op_MulReductionVD: 4218 case Op_MinReductionV: 4219 case Op_MaxReductionV: 4220 case Op_AndReductionV: 4221 case Op_OrReductionV: 4222 case Op_XorReductionV: 4223 break; 4224 4225 case Op_PackB: 4226 case Op_PackS: 4227 case Op_PackI: 4228 case Op_PackF: 4229 case Op_PackL: 4230 case Op_PackD: 4231 if (n->req()-1 > 2) { 4232 // Replace many operand PackNodes with a binary tree for matching 4233 PackNode* p = (PackNode*) n; 4234 Node* btp = p->binary_tree_pack(1, n->req()); 4235 n->subsume_by(btp, this); 4236 } 4237 break; 4238 case Op_Loop: 4239 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 4240 case Op_CountedLoop: 4241 case Op_LongCountedLoop: 4242 case Op_OuterStripMinedLoop: 4243 if (n->as_Loop()->is_inner_loop()) { 4244 frc.inc_inner_loop_count(); 4245 } 4246 n->as_Loop()->verify_strip_mined(0); 4247 break; 4248 case Op_LShiftI: 4249 case Op_RShiftI: 4250 case Op_URShiftI: 4251 case Op_LShiftL: 4252 case Op_RShiftL: 4253 case Op_URShiftL: 4254 if (Matcher::need_masked_shift_count) { 4255 // The cpu's shift instructions don't restrict the count to the 4256 // lower 5/6 bits. We need to do the masking ourselves. 4257 Node* in2 = n->in(2); 4258 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 4259 const TypeInt* t = in2->find_int_type(); 4260 if (t != nullptr && t->is_con()) { 4261 juint shift = t->get_con(); 4262 if (shift > mask) { // Unsigned cmp 4263 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 4264 } 4265 } else { 4266 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 4267 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 4268 n->set_req(2, shift); 4269 } 4270 } 4271 if (in2->outcnt() == 0) { // Remove dead node 4272 in2->disconnect_inputs(this); 4273 } 4274 } 4275 break; 4276 case Op_MemBarStoreStore: 4277 case Op_MemBarRelease: 4278 // Break the link with AllocateNode: it is no longer useful and 4279 // confuses register allocation. 4280 if (n->req() > MemBarNode::Precedent) { 4281 n->set_req(MemBarNode::Precedent, top()); 4282 } 4283 break; 4284 case Op_MemBarAcquire: { 4285 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 4286 // At parse time, the trailing MemBarAcquire for a volatile load 4287 // is created with an edge to the load. After optimizations, 4288 // that input may be a chain of Phis. If those phis have no 4289 // other use, then the MemBarAcquire keeps them alive and 4290 // register allocation can be confused. 4291 dead_nodes.push(n->in(MemBarNode::Precedent)); 4292 n->set_req(MemBarNode::Precedent, top()); 4293 } 4294 break; 4295 } 4296 case Op_Blackhole: 4297 break; 4298 case Op_RangeCheck: { 4299 RangeCheckNode* rc = n->as_RangeCheck(); 4300 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 4301 n->subsume_by(iff, this); 4302 frc._tests.push(iff); 4303 break; 4304 } 4305 case Op_ConvI2L: { 4306 if (!Matcher::convi2l_type_required) { 4307 // Code generation on some platforms doesn't need accurate 4308 // ConvI2L types. Widening the type can help remove redundant 4309 // address computations. 4310 n->as_Type()->set_type(TypeLong::INT); 4311 ResourceMark rm; 4312 Unique_Node_List wq; 4313 wq.push(n); 4314 for (uint next = 0; next < wq.size(); next++) { 4315 Node *m = wq.at(next); 4316 4317 for(;;) { 4318 // Loop over all nodes with identical inputs edges as m 4319 Node* k = m->find_similar(m->Opcode()); 4320 if (k == nullptr) { 4321 break; 4322 } 4323 // Push their uses so we get a chance to remove node made 4324 // redundant 4325 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 4326 Node* u = k->fast_out(i); 4327 if (u->Opcode() == Op_LShiftL || 4328 u->Opcode() == Op_AddL || 4329 u->Opcode() == Op_SubL || 4330 u->Opcode() == Op_AddP) { 4331 wq.push(u); 4332 } 4333 } 4334 // Replace all nodes with identical edges as m with m 4335 k->subsume_by(m, this); 4336 } 4337 } 4338 } 4339 break; 4340 } 4341 case Op_CmpUL: { 4342 if (!Matcher::has_match_rule(Op_CmpUL)) { 4343 // No support for unsigned long comparisons 4344 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 4345 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 4346 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 4347 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 4348 Node* andl = new AndLNode(orl, remove_sign_mask); 4349 Node* cmp = new CmpLNode(andl, n->in(2)); 4350 n->subsume_by(cmp, this); 4351 } 4352 break; 4353 } 4354 #ifdef ASSERT 4355 case Op_InlineType: { 4356 n->dump(-1); 4357 assert(false, "inline type node was not removed"); 4358 break; 4359 } 4360 #endif 4361 default: 4362 assert(!n->is_Call(), ""); 4363 assert(!n->is_Mem(), ""); 4364 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 4365 break; 4366 } 4367 } 4368 4369 //------------------------------final_graph_reshaping_walk--------------------- 4370 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 4371 // requires that the walk visits a node's inputs before visiting the node. 4372 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 4373 Unique_Node_List sfpt; 4374 4375 frc._visited.set(root->_idx); // first, mark node as visited 4376 uint cnt = root->req(); 4377 Node *n = root; 4378 uint i = 0; 4379 while (true) { 4380 if (i < cnt) { 4381 // Place all non-visited non-null inputs onto stack 4382 Node* m = n->in(i); 4383 ++i; 4384 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 4385 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 4386 // compute worst case interpreter size in case of a deoptimization 4387 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 4388 4389 sfpt.push(m); 4390 } 4391 cnt = m->req(); 4392 nstack.push(n, i); // put on stack parent and next input's index 4393 n = m; 4394 i = 0; 4395 } 4396 } else { 4397 // Now do post-visit work 4398 final_graph_reshaping_impl(n, frc, dead_nodes); 4399 if (nstack.is_empty()) 4400 break; // finished 4401 n = nstack.node(); // Get node from stack 4402 cnt = n->req(); 4403 i = nstack.index(); 4404 nstack.pop(); // Shift to the next node on stack 4405 } 4406 } 4407 4408 // Skip next transformation if compressed oops are not used. 4409 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 4410 (!UseCompressedOops && !UseCompressedClassPointers)) 4411 return; 4412 4413 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 4414 // It could be done for an uncommon traps or any safepoints/calls 4415 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 4416 while (sfpt.size() > 0) { 4417 n = sfpt.pop(); 4418 JVMState *jvms = n->as_SafePoint()->jvms(); 4419 assert(jvms != nullptr, "sanity"); 4420 int start = jvms->debug_start(); 4421 int end = n->req(); 4422 bool is_uncommon = (n->is_CallStaticJava() && 4423 n->as_CallStaticJava()->uncommon_trap_request() != 0); 4424 for (int j = start; j < end; j++) { 4425 Node* in = n->in(j); 4426 if (in->is_DecodeNarrowPtr()) { 4427 bool safe_to_skip = true; 4428 if (!is_uncommon ) { 4429 // Is it safe to skip? 4430 for (uint i = 0; i < in->outcnt(); i++) { 4431 Node* u = in->raw_out(i); 4432 if (!u->is_SafePoint() || 4433 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 4434 safe_to_skip = false; 4435 } 4436 } 4437 } 4438 if (safe_to_skip) { 4439 n->set_req(j, in->in(1)); 4440 } 4441 if (in->outcnt() == 0) { 4442 in->disconnect_inputs(this); 4443 } 4444 } 4445 } 4446 } 4447 } 4448 4449 //------------------------------final_graph_reshaping-------------------------- 4450 // Final Graph Reshaping. 4451 // 4452 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 4453 // and not commoned up and forced early. Must come after regular 4454 // optimizations to avoid GVN undoing the cloning. Clone constant 4455 // inputs to Loop Phis; these will be split by the allocator anyways. 4456 // Remove Opaque nodes. 4457 // (2) Move last-uses by commutative operations to the left input to encourage 4458 // Intel update-in-place two-address operations and better register usage 4459 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 4460 // calls canonicalizing them back. 4461 // (3) Count the number of double-precision FP ops, single-precision FP ops 4462 // and call sites. On Intel, we can get correct rounding either by 4463 // forcing singles to memory (requires extra stores and loads after each 4464 // FP bytecode) or we can set a rounding mode bit (requires setting and 4465 // clearing the mode bit around call sites). The mode bit is only used 4466 // if the relative frequency of single FP ops to calls is low enough. 4467 // This is a key transform for SPEC mpeg_audio. 4468 // (4) Detect infinite loops; blobs of code reachable from above but not 4469 // below. Several of the Code_Gen algorithms fail on such code shapes, 4470 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4471 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4472 // Detection is by looking for IfNodes where only 1 projection is 4473 // reachable from below or CatchNodes missing some targets. 4474 // (5) Assert for insane oop offsets in debug mode. 4475 4476 bool Compile::final_graph_reshaping() { 4477 // an infinite loop may have been eliminated by the optimizer, 4478 // in which case the graph will be empty. 4479 if (root()->req() == 1) { 4480 // Do not compile method that is only a trivial infinite loop, 4481 // since the content of the loop may have been eliminated. 4482 record_method_not_compilable("trivial infinite loop"); 4483 return true; 4484 } 4485 4486 // Expensive nodes have their control input set to prevent the GVN 4487 // from freely commoning them. There's no GVN beyond this point so 4488 // no need to keep the control input. We want the expensive nodes to 4489 // be freely moved to the least frequent code path by gcm. 4490 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4491 for (int i = 0; i < expensive_count(); i++) { 4492 _expensive_nodes.at(i)->set_req(0, nullptr); 4493 } 4494 4495 Final_Reshape_Counts frc; 4496 4497 // Visit everybody reachable! 4498 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4499 Node_Stack nstack(live_nodes() >> 1); 4500 Unique_Node_List dead_nodes; 4501 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 4502 4503 // Check for unreachable (from below) code (i.e., infinite loops). 4504 for( uint i = 0; i < frc._tests.size(); i++ ) { 4505 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4506 // Get number of CFG targets. 4507 // Note that PCTables include exception targets after calls. 4508 uint required_outcnt = n->required_outcnt(); 4509 if (n->outcnt() != required_outcnt) { 4510 // Check for a few special cases. Rethrow Nodes never take the 4511 // 'fall-thru' path, so expected kids is 1 less. 4512 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4513 if (n->in(0)->in(0)->is_Call()) { 4514 CallNode* call = n->in(0)->in(0)->as_Call(); 4515 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4516 required_outcnt--; // Rethrow always has 1 less kid 4517 } else if (call->req() > TypeFunc::Parms && 4518 call->is_CallDynamicJava()) { 4519 // Check for null receiver. In such case, the optimizer has 4520 // detected that the virtual call will always result in a null 4521 // pointer exception. The fall-through projection of this CatchNode 4522 // will not be populated. 4523 Node* arg0 = call->in(TypeFunc::Parms); 4524 if (arg0->is_Type() && 4525 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4526 required_outcnt--; 4527 } 4528 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4529 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4530 // Check for illegal array length. In such case, the optimizer has 4531 // detected that the allocation attempt will always result in an 4532 // exception. There is no fall-through projection of this CatchNode . 4533 assert(call->is_CallStaticJava(), "static call expected"); 4534 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4535 uint valid_length_test_input = call->req() - 1; 4536 Node* valid_length_test = call->in(valid_length_test_input); 4537 call->del_req(valid_length_test_input); 4538 if (valid_length_test->find_int_con(1) == 0) { 4539 required_outcnt--; 4540 } 4541 dead_nodes.push(valid_length_test); 4542 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4543 continue; 4544 } 4545 } 4546 } 4547 4548 // Recheck with a better notion of 'required_outcnt' 4549 if (n->outcnt() != required_outcnt) { 4550 record_method_not_compilable("malformed control flow"); 4551 return true; // Not all targets reachable! 4552 } 4553 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4554 CallNode* call = n->in(0)->in(0)->as_Call(); 4555 if (call->entry_point() == OptoRuntime::new_array_Java() || 4556 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4557 assert(call->is_CallStaticJava(), "static call expected"); 4558 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4559 uint valid_length_test_input = call->req() - 1; 4560 dead_nodes.push(call->in(valid_length_test_input)); 4561 call->del_req(valid_length_test_input); // valid length test useless now 4562 } 4563 } 4564 // Check that I actually visited all kids. Unreached kids 4565 // must be infinite loops. 4566 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4567 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4568 record_method_not_compilable("infinite loop"); 4569 return true; // Found unvisited kid; must be unreach 4570 } 4571 4572 // Here so verification code in final_graph_reshaping_walk() 4573 // always see an OuterStripMinedLoopEnd 4574 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4575 IfNode* init_iff = n->as_If(); 4576 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4577 n->subsume_by(iff, this); 4578 } 4579 } 4580 4581 while (dead_nodes.size() > 0) { 4582 Node* m = dead_nodes.pop(); 4583 if (m->outcnt() == 0 && m != top()) { 4584 for (uint j = 0; j < m->req(); j++) { 4585 Node* in = m->in(j); 4586 if (in != nullptr) { 4587 dead_nodes.push(in); 4588 } 4589 } 4590 m->disconnect_inputs(this); 4591 } 4592 } 4593 4594 #ifdef IA32 4595 // If original bytecodes contained a mixture of floats and doubles 4596 // check if the optimizer has made it homogeneous, item (3). 4597 if (UseSSE == 0 && 4598 frc.get_float_count() > 32 && 4599 frc.get_double_count() == 0 && 4600 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4601 set_24_bit_selection_and_mode(false, true); 4602 } 4603 #endif // IA32 4604 4605 set_java_calls(frc.get_java_call_count()); 4606 set_inner_loops(frc.get_inner_loop_count()); 4607 4608 // No infinite loops, no reason to bail out. 4609 return false; 4610 } 4611 4612 //-----------------------------too_many_traps---------------------------------- 4613 // Report if there are too many traps at the current method and bci. 4614 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4615 bool Compile::too_many_traps(ciMethod* method, 4616 int bci, 4617 Deoptimization::DeoptReason reason) { 4618 ciMethodData* md = method->method_data(); 4619 if (md->is_empty()) { 4620 // Assume the trap has not occurred, or that it occurred only 4621 // because of a transient condition during start-up in the interpreter. 4622 return false; 4623 } 4624 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4625 if (md->has_trap_at(bci, m, reason) != 0) { 4626 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4627 // Also, if there are multiple reasons, or if there is no per-BCI record, 4628 // assume the worst. 4629 if (log()) 4630 log()->elem("observe trap='%s' count='%d'", 4631 Deoptimization::trap_reason_name(reason), 4632 md->trap_count(reason)); 4633 return true; 4634 } else { 4635 // Ignore method/bci and see if there have been too many globally. 4636 return too_many_traps(reason, md); 4637 } 4638 } 4639 4640 // Less-accurate variant which does not require a method and bci. 4641 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4642 ciMethodData* logmd) { 4643 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4644 // Too many traps globally. 4645 // Note that we use cumulative trap_count, not just md->trap_count. 4646 if (log()) { 4647 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4648 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4649 Deoptimization::trap_reason_name(reason), 4650 mcount, trap_count(reason)); 4651 } 4652 return true; 4653 } else { 4654 // The coast is clear. 4655 return false; 4656 } 4657 } 4658 4659 //--------------------------too_many_recompiles-------------------------------- 4660 // Report if there are too many recompiles at the current method and bci. 4661 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4662 // Is not eager to return true, since this will cause the compiler to use 4663 // Action_none for a trap point, to avoid too many recompilations. 4664 bool Compile::too_many_recompiles(ciMethod* method, 4665 int bci, 4666 Deoptimization::DeoptReason reason) { 4667 ciMethodData* md = method->method_data(); 4668 if (md->is_empty()) { 4669 // Assume the trap has not occurred, or that it occurred only 4670 // because of a transient condition during start-up in the interpreter. 4671 return false; 4672 } 4673 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4674 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4675 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4676 Deoptimization::DeoptReason per_bc_reason 4677 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4678 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4679 if ((per_bc_reason == Deoptimization::Reason_none 4680 || md->has_trap_at(bci, m, reason) != 0) 4681 // The trap frequency measure we care about is the recompile count: 4682 && md->trap_recompiled_at(bci, m) 4683 && md->overflow_recompile_count() >= bc_cutoff) { 4684 // Do not emit a trap here if it has already caused recompilations. 4685 // Also, if there are multiple reasons, or if there is no per-BCI record, 4686 // assume the worst. 4687 if (log()) 4688 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4689 Deoptimization::trap_reason_name(reason), 4690 md->trap_count(reason), 4691 md->overflow_recompile_count()); 4692 return true; 4693 } else if (trap_count(reason) != 0 4694 && decompile_count() >= m_cutoff) { 4695 // Too many recompiles globally, and we have seen this sort of trap. 4696 // Use cumulative decompile_count, not just md->decompile_count. 4697 if (log()) 4698 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4699 Deoptimization::trap_reason_name(reason), 4700 md->trap_count(reason), trap_count(reason), 4701 md->decompile_count(), decompile_count()); 4702 return true; 4703 } else { 4704 // The coast is clear. 4705 return false; 4706 } 4707 } 4708 4709 // Compute when not to trap. Used by matching trap based nodes and 4710 // NullCheck optimization. 4711 void Compile::set_allowed_deopt_reasons() { 4712 _allowed_reasons = 0; 4713 if (is_method_compilation()) { 4714 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4715 assert(rs < BitsPerInt, "recode bit map"); 4716 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4717 _allowed_reasons |= nth_bit(rs); 4718 } 4719 } 4720 } 4721 } 4722 4723 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4724 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4725 } 4726 4727 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4728 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4729 } 4730 4731 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4732 if (holder->is_initialized()) { 4733 return false; 4734 } 4735 if (holder->is_being_initialized()) { 4736 if (accessing_method->holder() == holder) { 4737 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4738 // <init>, or a static method. In all those cases, there was an initialization 4739 // barrier on the holder klass passed. 4740 if (accessing_method->is_class_initializer() || 4741 accessing_method->is_object_constructor() || 4742 accessing_method->is_static()) { 4743 return false; 4744 } 4745 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4746 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4747 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4748 // child class can become fully initialized while its parent class is still being initialized. 4749 if (accessing_method->is_class_initializer()) { 4750 return false; 4751 } 4752 } 4753 ciMethod* root = method(); // the root method of compilation 4754 if (root != accessing_method) { 4755 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4756 } 4757 } 4758 return true; 4759 } 4760 4761 #ifndef PRODUCT 4762 //------------------------------verify_bidirectional_edges--------------------- 4763 // For each input edge to a node (ie - for each Use-Def edge), verify that 4764 // there is a corresponding Def-Use edge. 4765 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4766 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4767 uint stack_size = live_nodes() >> 4; 4768 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4769 nstack.push(_root); 4770 4771 while (nstack.size() > 0) { 4772 Node* n = nstack.pop(); 4773 if (visited.member(n)) { 4774 continue; 4775 } 4776 visited.push(n); 4777 4778 // Walk over all input edges, checking for correspondence 4779 uint length = n->len(); 4780 for (uint i = 0; i < length; i++) { 4781 Node* in = n->in(i); 4782 if (in != nullptr && !visited.member(in)) { 4783 nstack.push(in); // Put it on stack 4784 } 4785 if (in != nullptr && !in->is_top()) { 4786 // Count instances of `next` 4787 int cnt = 0; 4788 for (uint idx = 0; idx < in->_outcnt; idx++) { 4789 if (in->_out[idx] == n) { 4790 cnt++; 4791 } 4792 } 4793 assert(cnt > 0, "Failed to find Def-Use edge."); 4794 // Check for duplicate edges 4795 // walk the input array downcounting the input edges to n 4796 for (uint j = 0; j < length; j++) { 4797 if (n->in(j) == in) { 4798 cnt--; 4799 } 4800 } 4801 assert(cnt == 0, "Mismatched edge count."); 4802 } else if (in == nullptr) { 4803 assert(i == 0 || i >= n->req() || 4804 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4805 (n->is_Allocate() && i >= AllocateNode::InlineType) || 4806 (n->is_Unlock() && i == (n->req() - 1)) || 4807 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4808 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges"); 4809 } else { 4810 assert(in->is_top(), "sanity"); 4811 // Nothing to check. 4812 } 4813 } 4814 } 4815 } 4816 4817 //------------------------------verify_graph_edges--------------------------- 4818 // Walk the Graph and verify that there is a one-to-one correspondence 4819 // between Use-Def edges and Def-Use edges in the graph. 4820 void Compile::verify_graph_edges(bool no_dead_code) { 4821 if (VerifyGraphEdges) { 4822 Unique_Node_List visited; 4823 4824 // Call graph walk to check edges 4825 verify_bidirectional_edges(visited); 4826 if (no_dead_code) { 4827 // Now make sure that no visited node is used by an unvisited node. 4828 bool dead_nodes = false; 4829 Unique_Node_List checked; 4830 while (visited.size() > 0) { 4831 Node* n = visited.pop(); 4832 checked.push(n); 4833 for (uint i = 0; i < n->outcnt(); i++) { 4834 Node* use = n->raw_out(i); 4835 if (checked.member(use)) continue; // already checked 4836 if (visited.member(use)) continue; // already in the graph 4837 if (use->is_Con()) continue; // a dead ConNode is OK 4838 // At this point, we have found a dead node which is DU-reachable. 4839 if (!dead_nodes) { 4840 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4841 dead_nodes = true; 4842 } 4843 use->dump(2); 4844 tty->print_cr("---"); 4845 checked.push(use); // No repeats; pretend it is now checked. 4846 } 4847 } 4848 assert(!dead_nodes, "using nodes must be reachable from root"); 4849 } 4850 } 4851 } 4852 #endif 4853 4854 // The Compile object keeps track of failure reasons separately from the ciEnv. 4855 // This is required because there is not quite a 1-1 relation between the 4856 // ciEnv and its compilation task and the Compile object. Note that one 4857 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4858 // to backtrack and retry without subsuming loads. Other than this backtracking 4859 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4860 // by the logic in C2Compiler. 4861 void Compile::record_failure(const char* reason) { 4862 if (log() != nullptr) { 4863 log()->elem("failure reason='%s' phase='compile'", reason); 4864 } 4865 if (_failure_reason == nullptr) { 4866 // Record the first failure reason. 4867 _failure_reason = reason; 4868 } 4869 4870 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4871 C->print_method(PHASE_FAILURE, 1); 4872 } 4873 _root = nullptr; // flush the graph, too 4874 } 4875 4876 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4877 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4878 _compile(Compile::current()), 4879 _log(nullptr), 4880 _phase_name(name), 4881 _dolog(CITimeVerbose) 4882 { 4883 assert(_compile != nullptr, "sanity check"); 4884 if (_dolog) { 4885 _log = _compile->log(); 4886 } 4887 if (_log != nullptr) { 4888 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4889 _log->stamp(); 4890 _log->end_head(); 4891 } 4892 } 4893 4894 Compile::TracePhase::~TracePhase() { 4895 if (_compile->failing()) return; 4896 #ifdef ASSERT 4897 if (PrintIdealNodeCount) { 4898 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4899 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4900 } 4901 4902 if (VerifyIdealNodeCount) { 4903 _compile->print_missing_nodes(); 4904 } 4905 #endif 4906 4907 if (_log != nullptr) { 4908 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4909 } 4910 } 4911 4912 //----------------------------static_subtype_check----------------------------- 4913 // Shortcut important common cases when superklass is exact: 4914 // (0) superklass is java.lang.Object (can occur in reflective code) 4915 // (1) subklass is already limited to a subtype of superklass => always ok 4916 // (2) subklass does not overlap with superklass => always fail 4917 // (3) superklass has NO subtypes and we can check with a simple compare. 4918 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4919 if (skip) { 4920 return SSC_full_test; // Let caller generate the general case. 4921 } 4922 4923 if (subk->is_java_subtype_of(superk)) { 4924 return SSC_always_true; // (0) and (1) this test cannot fail 4925 } 4926 4927 if (!subk->maybe_java_subtype_of(superk)) { 4928 return SSC_always_false; // (2) true path dead; no dynamic test needed 4929 } 4930 4931 const Type* superelem = superk; 4932 if (superk->isa_aryklassptr()) { 4933 int ignored; 4934 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4935 4936 // TODO 8325106 Fix comment 4937 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays. 4938 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to 4939 // the klass for [LMyValue. Perform a full test. 4940 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() && 4941 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) { 4942 return SSC_full_test; 4943 } 4944 } 4945 4946 if (superelem->isa_instklassptr()) { 4947 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4948 if (!ik->has_subklass()) { 4949 if (!ik->is_final()) { 4950 // Add a dependency if there is a chance of a later subclass. 4951 dependencies()->assert_leaf_type(ik); 4952 } 4953 if (!superk->maybe_java_subtype_of(subk)) { 4954 return SSC_always_false; 4955 } 4956 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4957 } 4958 } else { 4959 // A primitive array type has no subtypes. 4960 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4961 } 4962 4963 return SSC_full_test; 4964 } 4965 4966 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4967 #ifdef _LP64 4968 // The scaled index operand to AddP must be a clean 64-bit value. 4969 // Java allows a 32-bit int to be incremented to a negative 4970 // value, which appears in a 64-bit register as a large 4971 // positive number. Using that large positive number as an 4972 // operand in pointer arithmetic has bad consequences. 4973 // On the other hand, 32-bit overflow is rare, and the possibility 4974 // can often be excluded, if we annotate the ConvI2L node with 4975 // a type assertion that its value is known to be a small positive 4976 // number. (The prior range check has ensured this.) 4977 // This assertion is used by ConvI2LNode::Ideal. 4978 int index_max = max_jint - 1; // array size is max_jint, index is one less 4979 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4980 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4981 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4982 #endif 4983 return idx; 4984 } 4985 4986 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4987 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4988 if (ctrl != nullptr) { 4989 // Express control dependency by a CastII node with a narrow type. 4990 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4991 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4992 // node from floating above the range check during loop optimizations. Otherwise, the 4993 // ConvI2L node may be eliminated independently of the range check, causing the data path 4994 // to become TOP while the control path is still there (although it's unreachable). 4995 value->set_req(0, ctrl); 4996 value = phase->transform(value); 4997 } 4998 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4999 return phase->transform(new ConvI2LNode(value, ltype)); 5000 } 5001 5002 // The message about the current inlining is accumulated in 5003 // _print_inlining_stream and transferred into the _print_inlining_list 5004 // once we know whether inlining succeeds or not. For regular 5005 // inlining, messages are appended to the buffer pointed by 5006 // _print_inlining_idx in the _print_inlining_list. For late inlining, 5007 // a new buffer is added after _print_inlining_idx in the list. This 5008 // way we can update the inlining message for late inlining call site 5009 // when the inlining is attempted again. 5010 void Compile::print_inlining_init() { 5011 if (print_inlining() || print_intrinsics()) { 5012 // print_inlining_init is actually called several times. 5013 print_inlining_reset(); 5014 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 5015 } 5016 } 5017 5018 void Compile::print_inlining_reinit() { 5019 if (print_inlining() || print_intrinsics()) { 5020 print_inlining_reset(); 5021 } 5022 } 5023 5024 void Compile::print_inlining_reset() { 5025 _print_inlining_stream->reset(); 5026 } 5027 5028 void Compile::print_inlining_commit() { 5029 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 5030 // Transfer the message from _print_inlining_stream to the current 5031 // _print_inlining_list buffer and clear _print_inlining_stream. 5032 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 5033 print_inlining_reset(); 5034 } 5035 5036 void Compile::print_inlining_push() { 5037 // Add new buffer to the _print_inlining_list at current position 5038 _print_inlining_idx++; 5039 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 5040 } 5041 5042 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 5043 return _print_inlining_list->at(_print_inlining_idx); 5044 } 5045 5046 void Compile::print_inlining_update(CallGenerator* cg) { 5047 if (print_inlining() || print_intrinsics()) { 5048 if (cg->is_late_inline()) { 5049 if (print_inlining_current()->cg() != cg && 5050 (print_inlining_current()->cg() != nullptr || 5051 print_inlining_current()->ss()->size() != 0)) { 5052 print_inlining_push(); 5053 } 5054 print_inlining_commit(); 5055 print_inlining_current()->set_cg(cg); 5056 } else { 5057 if (print_inlining_current()->cg() != nullptr) { 5058 print_inlining_push(); 5059 } 5060 print_inlining_commit(); 5061 } 5062 } 5063 } 5064 5065 void Compile::print_inlining_move_to(CallGenerator* cg) { 5066 // We resume inlining at a late inlining call site. Locate the 5067 // corresponding inlining buffer so that we can update it. 5068 if (print_inlining() || print_intrinsics()) { 5069 for (int i = 0; i < _print_inlining_list->length(); i++) { 5070 if (_print_inlining_list->at(i)->cg() == cg) { 5071 _print_inlining_idx = i; 5072 return; 5073 } 5074 } 5075 ShouldNotReachHere(); 5076 } 5077 } 5078 5079 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 5080 if (print_inlining() || print_intrinsics()) { 5081 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 5082 assert(print_inlining_current()->cg() == cg, "wrong entry"); 5083 // replace message with new message 5084 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 5085 print_inlining_commit(); 5086 print_inlining_current()->set_cg(cg); 5087 } 5088 } 5089 5090 void Compile::print_inlining_assert_ready() { 5091 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 5092 } 5093 5094 void Compile::process_print_inlining() { 5095 assert(_late_inlines.length() == 0, "not drained yet"); 5096 if (print_inlining() || print_intrinsics()) { 5097 ResourceMark rm; 5098 stringStream ss; 5099 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 5100 for (int i = 0; i < _print_inlining_list->length(); i++) { 5101 PrintInliningBuffer* pib = _print_inlining_list->at(i); 5102 ss.print("%s", pib->ss()->freeze()); 5103 delete pib; 5104 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 5105 } 5106 // Reset _print_inlining_list, it only contains destructed objects. 5107 // It is on the arena, so it will be freed when the arena is reset. 5108 _print_inlining_list = nullptr; 5109 // _print_inlining_stream won't be used anymore, either. 5110 print_inlining_reset(); 5111 size_t end = ss.size(); 5112 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 5113 strncpy(_print_inlining_output, ss.freeze(), end+1); 5114 _print_inlining_output[end] = 0; 5115 } 5116 } 5117 5118 void Compile::dump_print_inlining() { 5119 if (_print_inlining_output != nullptr) { 5120 tty->print_raw(_print_inlining_output); 5121 } 5122 } 5123 5124 void Compile::log_late_inline(CallGenerator* cg) { 5125 if (log() != nullptr) { 5126 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 5127 cg->unique_id()); 5128 JVMState* p = cg->call_node()->jvms(); 5129 while (p != nullptr) { 5130 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 5131 p = p->caller(); 5132 } 5133 log()->tail("late_inline"); 5134 } 5135 } 5136 5137 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 5138 log_late_inline(cg); 5139 if (log() != nullptr) { 5140 log()->inline_fail(msg); 5141 } 5142 } 5143 5144 void Compile::log_inline_id(CallGenerator* cg) { 5145 if (log() != nullptr) { 5146 // The LogCompilation tool needs a unique way to identify late 5147 // inline call sites. This id must be unique for this call site in 5148 // this compilation. Try to have it unique across compilations as 5149 // well because it can be convenient when grepping through the log 5150 // file. 5151 // Distinguish OSR compilations from others in case CICountOSR is 5152 // on. 5153 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 5154 cg->set_unique_id(id); 5155 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 5156 } 5157 } 5158 5159 void Compile::log_inline_failure(const char* msg) { 5160 if (C->log() != nullptr) { 5161 C->log()->inline_fail(msg); 5162 } 5163 } 5164 5165 5166 // Dump inlining replay data to the stream. 5167 // Don't change thread state and acquire any locks. 5168 void Compile::dump_inline_data(outputStream* out) { 5169 InlineTree* inl_tree = ilt(); 5170 if (inl_tree != nullptr) { 5171 out->print(" inline %d", inl_tree->count()); 5172 inl_tree->dump_replay_data(out); 5173 } 5174 } 5175 5176 void Compile::dump_inline_data_reduced(outputStream* out) { 5177 assert(ReplayReduce, ""); 5178 5179 InlineTree* inl_tree = ilt(); 5180 if (inl_tree == nullptr) { 5181 return; 5182 } 5183 // Enable iterative replay file reduction 5184 // Output "compile" lines for depth 1 subtrees, 5185 // simulating that those trees were compiled 5186 // instead of inlined. 5187 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 5188 InlineTree* sub = inl_tree->subtrees().at(i); 5189 if (sub->inline_level() != 1) { 5190 continue; 5191 } 5192 5193 ciMethod* method = sub->method(); 5194 int entry_bci = -1; 5195 int comp_level = env()->task()->comp_level(); 5196 out->print("compile "); 5197 method->dump_name_as_ascii(out); 5198 out->print(" %d %d", entry_bci, comp_level); 5199 out->print(" inline %d", sub->count()); 5200 sub->dump_replay_data(out, -1); 5201 out->cr(); 5202 } 5203 } 5204 5205 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 5206 if (n1->Opcode() < n2->Opcode()) return -1; 5207 else if (n1->Opcode() > n2->Opcode()) return 1; 5208 5209 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 5210 for (uint i = 1; i < n1->req(); i++) { 5211 if (n1->in(i) < n2->in(i)) return -1; 5212 else if (n1->in(i) > n2->in(i)) return 1; 5213 } 5214 5215 return 0; 5216 } 5217 5218 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 5219 Node* n1 = *n1p; 5220 Node* n2 = *n2p; 5221 5222 return cmp_expensive_nodes(n1, n2); 5223 } 5224 5225 void Compile::sort_expensive_nodes() { 5226 if (!expensive_nodes_sorted()) { 5227 _expensive_nodes.sort(cmp_expensive_nodes); 5228 } 5229 } 5230 5231 bool Compile::expensive_nodes_sorted() const { 5232 for (int i = 1; i < _expensive_nodes.length(); i++) { 5233 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 5234 return false; 5235 } 5236 } 5237 return true; 5238 } 5239 5240 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 5241 if (_expensive_nodes.length() == 0) { 5242 return false; 5243 } 5244 5245 assert(OptimizeExpensiveOps, "optimization off?"); 5246 5247 // Take this opportunity to remove dead nodes from the list 5248 int j = 0; 5249 for (int i = 0; i < _expensive_nodes.length(); i++) { 5250 Node* n = _expensive_nodes.at(i); 5251 if (!n->is_unreachable(igvn)) { 5252 assert(n->is_expensive(), "should be expensive"); 5253 _expensive_nodes.at_put(j, n); 5254 j++; 5255 } 5256 } 5257 _expensive_nodes.trunc_to(j); 5258 5259 // Then sort the list so that similar nodes are next to each other 5260 // and check for at least two nodes of identical kind with same data 5261 // inputs. 5262 sort_expensive_nodes(); 5263 5264 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 5265 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 5266 return true; 5267 } 5268 } 5269 5270 return false; 5271 } 5272 5273 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 5274 if (_expensive_nodes.length() == 0) { 5275 return; 5276 } 5277 5278 assert(OptimizeExpensiveOps, "optimization off?"); 5279 5280 // Sort to bring similar nodes next to each other and clear the 5281 // control input of nodes for which there's only a single copy. 5282 sort_expensive_nodes(); 5283 5284 int j = 0; 5285 int identical = 0; 5286 int i = 0; 5287 bool modified = false; 5288 for (; i < _expensive_nodes.length()-1; i++) { 5289 assert(j <= i, "can't write beyond current index"); 5290 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 5291 identical++; 5292 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 5293 continue; 5294 } 5295 if (identical > 0) { 5296 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 5297 identical = 0; 5298 } else { 5299 Node* n = _expensive_nodes.at(i); 5300 igvn.replace_input_of(n, 0, nullptr); 5301 igvn.hash_insert(n); 5302 modified = true; 5303 } 5304 } 5305 if (identical > 0) { 5306 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 5307 } else if (_expensive_nodes.length() >= 1) { 5308 Node* n = _expensive_nodes.at(i); 5309 igvn.replace_input_of(n, 0, nullptr); 5310 igvn.hash_insert(n); 5311 modified = true; 5312 } 5313 _expensive_nodes.trunc_to(j); 5314 if (modified) { 5315 igvn.optimize(); 5316 } 5317 } 5318 5319 void Compile::add_expensive_node(Node * n) { 5320 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 5321 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 5322 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 5323 if (OptimizeExpensiveOps) { 5324 _expensive_nodes.append(n); 5325 } else { 5326 // Clear control input and let IGVN optimize expensive nodes if 5327 // OptimizeExpensiveOps is off. 5328 n->set_req(0, nullptr); 5329 } 5330 } 5331 5332 /** 5333 * Track coarsened Lock and Unlock nodes. 5334 */ 5335 5336 class Lock_List : public Node_List { 5337 uint _origin_cnt; 5338 public: 5339 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 5340 uint origin_cnt() const { return _origin_cnt; } 5341 }; 5342 5343 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 5344 int length = locks.length(); 5345 if (length > 0) { 5346 // Have to keep this list until locks elimination during Macro nodes elimination. 5347 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 5348 for (int i = 0; i < length; i++) { 5349 AbstractLockNode* lock = locks.at(i); 5350 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 5351 locks_list->push(lock); 5352 } 5353 _coarsened_locks.append(locks_list); 5354 } 5355 } 5356 5357 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 5358 int count = coarsened_count(); 5359 for (int i = 0; i < count; i++) { 5360 Node_List* locks_list = _coarsened_locks.at(i); 5361 for (uint j = 0; j < locks_list->size(); j++) { 5362 Node* lock = locks_list->at(j); 5363 assert(lock->is_AbstractLock(), "sanity"); 5364 if (!useful.member(lock)) { 5365 locks_list->yank(lock); 5366 } 5367 } 5368 } 5369 } 5370 5371 void Compile::remove_coarsened_lock(Node* n) { 5372 if (n->is_AbstractLock()) { 5373 int count = coarsened_count(); 5374 for (int i = 0; i < count; i++) { 5375 Node_List* locks_list = _coarsened_locks.at(i); 5376 locks_list->yank(n); 5377 } 5378 } 5379 } 5380 5381 bool Compile::coarsened_locks_consistent() { 5382 int count = coarsened_count(); 5383 for (int i = 0; i < count; i++) { 5384 bool unbalanced = false; 5385 bool modified = false; // track locks kind modifications 5386 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 5387 uint size = locks_list->size(); 5388 if (size == 0) { 5389 unbalanced = false; // All locks were eliminated - good 5390 } else if (size != locks_list->origin_cnt()) { 5391 unbalanced = true; // Some locks were removed from list 5392 } else { 5393 for (uint j = 0; j < size; j++) { 5394 Node* lock = locks_list->at(j); 5395 // All nodes in group should have the same state (modified or not) 5396 if (!lock->as_AbstractLock()->is_coarsened()) { 5397 if (j == 0) { 5398 // first on list was modified, the rest should be too for consistency 5399 modified = true; 5400 } else if (!modified) { 5401 // this lock was modified but previous locks on the list were not 5402 unbalanced = true; 5403 break; 5404 } 5405 } else if (modified) { 5406 // previous locks on list were modified but not this lock 5407 unbalanced = true; 5408 break; 5409 } 5410 } 5411 } 5412 if (unbalanced) { 5413 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 5414 #ifdef ASSERT 5415 if (PrintEliminateLocks) { 5416 tty->print_cr("=== unbalanced coarsened locks ==="); 5417 for (uint l = 0; l < size; l++) { 5418 locks_list->at(l)->dump(); 5419 } 5420 } 5421 #endif 5422 record_failure(C2Compiler::retry_no_locks_coarsening()); 5423 return false; 5424 } 5425 } 5426 return true; 5427 } 5428 5429 /** 5430 * Remove the speculative part of types and clean up the graph 5431 */ 5432 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 5433 if (UseTypeSpeculation) { 5434 Unique_Node_List worklist; 5435 worklist.push(root()); 5436 int modified = 0; 5437 // Go over all type nodes that carry a speculative type, drop the 5438 // speculative part of the type and enqueue the node for an igvn 5439 // which may optimize it out. 5440 for (uint next = 0; next < worklist.size(); ++next) { 5441 Node *n = worklist.at(next); 5442 if (n->is_Type()) { 5443 TypeNode* tn = n->as_Type(); 5444 const Type* t = tn->type(); 5445 const Type* t_no_spec = t->remove_speculative(); 5446 if (t_no_spec != t) { 5447 bool in_hash = igvn.hash_delete(n); 5448 #ifdef ASSERT 5449 if (!in_hash) { 5450 tty->print_cr("current graph:"); 5451 n->dump_bfs(MaxNodeLimit, nullptr, "S$"); 5452 tty->cr(); 5453 tty->print_cr("erroneous node:"); 5454 n->dump(); 5455 assert(false, "node should be in igvn hash table"); 5456 } 5457 #endif 5458 tn->set_type(t_no_spec); 5459 igvn.hash_insert(n); 5460 igvn._worklist.push(n); // give it a chance to go away 5461 modified++; 5462 } 5463 } 5464 // Iterate over outs - endless loops is unreachable from below 5465 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 5466 Node *m = n->fast_out(i); 5467 if (not_a_node(m)) { 5468 continue; 5469 } 5470 worklist.push(m); 5471 } 5472 } 5473 // Drop the speculative part of all types in the igvn's type table 5474 igvn.remove_speculative_types(); 5475 if (modified > 0) { 5476 igvn.optimize(); 5477 if (failing()) return; 5478 } 5479 #ifdef ASSERT 5480 // Verify that after the IGVN is over no speculative type has resurfaced 5481 worklist.clear(); 5482 worklist.push(root()); 5483 for (uint next = 0; next < worklist.size(); ++next) { 5484 Node *n = worklist.at(next); 5485 const Type* t = igvn.type_or_null(n); 5486 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 5487 if (n->is_Type()) { 5488 t = n->as_Type()->type(); 5489 assert(t == t->remove_speculative(), "no more speculative types"); 5490 } 5491 // Iterate over outs - endless loops is unreachable from below 5492 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 5493 Node *m = n->fast_out(i); 5494 if (not_a_node(m)) { 5495 continue; 5496 } 5497 worklist.push(m); 5498 } 5499 } 5500 igvn.check_no_speculative_types(); 5501 #endif 5502 } 5503 } 5504 5505 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 5506 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 5507 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 5508 if (!EnableValhalla || ta == nullptr || tb == nullptr || 5509 ta->is_zero_type() || tb->is_zero_type() || 5510 !ta->can_be_inline_type() || !tb->can_be_inline_type()) { 5511 // Use old acmp if one operand is null or not an inline type 5512 return new CmpPNode(a, b); 5513 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) { 5514 // We know that one operand is an inline type. Therefore, 5515 // new acmp will only return true if both operands are nullptr. 5516 // Check if both operands are null by or'ing the oops. 5517 a = phase->transform(new CastP2XNode(nullptr, a)); 5518 b = phase->transform(new CastP2XNode(nullptr, b)); 5519 a = phase->transform(new OrXNode(a, b)); 5520 return new CmpXNode(a, phase->MakeConX(0)); 5521 } 5522 // Use new acmp 5523 return nullptr; 5524 } 5525 5526 // Auxiliary methods to support randomized stressing/fuzzing. 5527 5528 int Compile::random() { 5529 _stress_seed = os::next_random(_stress_seed); 5530 return static_cast<int>(_stress_seed); 5531 } 5532 5533 // This method can be called the arbitrary number of times, with current count 5534 // as the argument. The logic allows selecting a single candidate from the 5535 // running list of candidates as follows: 5536 // int count = 0; 5537 // Cand* selected = null; 5538 // while(cand = cand->next()) { 5539 // if (randomized_select(++count)) { 5540 // selected = cand; 5541 // } 5542 // } 5543 // 5544 // Including count equalizes the chances any candidate is "selected". 5545 // This is useful when we don't have the complete list of candidates to choose 5546 // from uniformly. In this case, we need to adjust the randomicity of the 5547 // selection, or else we will end up biasing the selection towards the latter 5548 // candidates. 5549 // 5550 // Quick back-envelope calculation shows that for the list of n candidates 5551 // the equal probability for the candidate to persist as "best" can be 5552 // achieved by replacing it with "next" k-th candidate with the probability 5553 // of 1/k. It can be easily shown that by the end of the run, the 5554 // probability for any candidate is converged to 1/n, thus giving the 5555 // uniform distribution among all the candidates. 5556 // 5557 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5558 #define RANDOMIZED_DOMAIN_POW 29 5559 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5560 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5561 bool Compile::randomized_select(int count) { 5562 assert(count > 0, "only positive"); 5563 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5564 } 5565 5566 CloneMap& Compile::clone_map() { return _clone_map; } 5567 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5568 5569 void NodeCloneInfo::dump_on(outputStream* st) const { 5570 st->print(" {%d:%d} ", idx(), gen()); 5571 } 5572 5573 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5574 uint64_t val = value(old->_idx); 5575 NodeCloneInfo cio(val); 5576 assert(val != 0, "old node should be in the map"); 5577 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5578 insert(nnn->_idx, cin.get()); 5579 #ifndef PRODUCT 5580 if (is_debug()) { 5581 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5582 } 5583 #endif 5584 } 5585 5586 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5587 NodeCloneInfo cio(value(old->_idx)); 5588 if (cio.get() == 0) { 5589 cio.set(old->_idx, 0); 5590 insert(old->_idx, cio.get()); 5591 #ifndef PRODUCT 5592 if (is_debug()) { 5593 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5594 } 5595 #endif 5596 } 5597 clone(old, nnn, gen); 5598 } 5599 5600 int CloneMap::max_gen() const { 5601 int g = 0; 5602 DictI di(_dict); 5603 for(; di.test(); ++di) { 5604 int t = gen(di._key); 5605 if (g < t) { 5606 g = t; 5607 #ifndef PRODUCT 5608 if (is_debug()) { 5609 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5610 } 5611 #endif 5612 } 5613 } 5614 return g; 5615 } 5616 5617 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5618 uint64_t val = value(key); 5619 if (val != 0) { 5620 NodeCloneInfo ni(val); 5621 ni.dump_on(st); 5622 } 5623 } 5624 5625 // Move Allocate nodes to the start of the list 5626 void Compile::sort_macro_nodes() { 5627 int count = macro_count(); 5628 int allocates = 0; 5629 for (int i = 0; i < count; i++) { 5630 Node* n = macro_node(i); 5631 if (n->is_Allocate()) { 5632 if (i != allocates) { 5633 Node* tmp = macro_node(allocates); 5634 _macro_nodes.at_put(allocates, n); 5635 _macro_nodes.at_put(i, tmp); 5636 } 5637 allocates++; 5638 } 5639 } 5640 } 5641 5642 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5643 if (failing()) { return; } 5644 EventCompilerPhase event; 5645 if (event.should_commit()) { 5646 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5647 } 5648 #ifndef PRODUCT 5649 ResourceMark rm; 5650 stringStream ss; 5651 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5652 if (n != nullptr) { 5653 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5654 } 5655 5656 const char* name = ss.as_string(); 5657 if (should_print_igv(level)) { 5658 _igv_printer->print_method(name, level); 5659 } 5660 if (should_print_phase(cpt)) { 5661 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5662 } 5663 #endif 5664 C->_latest_stage_start_counter.stamp(); 5665 } 5666 5667 // Only used from CompileWrapper 5668 void Compile::begin_method() { 5669 #ifndef PRODUCT 5670 if (_method != nullptr && should_print_igv(1)) { 5671 _igv_printer->begin_method(); 5672 } 5673 #endif 5674 C->_latest_stage_start_counter.stamp(); 5675 } 5676 5677 // Only used from CompileWrapper 5678 void Compile::end_method() { 5679 EventCompilerPhase event; 5680 if (event.should_commit()) { 5681 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5682 } 5683 5684 #ifndef PRODUCT 5685 if (_method != nullptr && should_print_igv(1)) { 5686 _igv_printer->end_method(); 5687 } 5688 #endif 5689 } 5690 5691 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5692 #ifndef PRODUCT 5693 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) { 5694 return true; 5695 } 5696 #endif 5697 return false; 5698 } 5699 5700 bool Compile::should_print_igv(const int level) { 5701 #ifndef PRODUCT 5702 if (PrintIdealGraphLevel < 0) { // disabled by the user 5703 return false; 5704 } 5705 5706 bool need = directive()->IGVPrintLevelOption >= level; 5707 if (need && !_igv_printer) { 5708 _igv_printer = IdealGraphPrinter::printer(); 5709 _igv_printer->set_compile(this); 5710 } 5711 return need; 5712 #else 5713 return false; 5714 #endif 5715 } 5716 5717 #ifndef PRODUCT 5718 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5719 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5720 5721 // Called from debugger. Prints method to the default file with the default phase name. 5722 // This works regardless of any Ideal Graph Visualizer flags set or not. 5723 void igv_print() { 5724 Compile::current()->igv_print_method_to_file(); 5725 } 5726 5727 // Same as igv_print() above but with a specified phase name. 5728 void igv_print(const char* phase_name) { 5729 Compile::current()->igv_print_method_to_file(phase_name); 5730 } 5731 5732 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5733 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5734 // This works regardless of any Ideal Graph Visualizer flags set or not. 5735 void igv_print(bool network) { 5736 if (network) { 5737 Compile::current()->igv_print_method_to_network(); 5738 } else { 5739 Compile::current()->igv_print_method_to_file(); 5740 } 5741 } 5742 5743 // Same as igv_print(bool network) above but with a specified phase name. 5744 void igv_print(bool network, const char* phase_name) { 5745 if (network) { 5746 Compile::current()->igv_print_method_to_network(phase_name); 5747 } else { 5748 Compile::current()->igv_print_method_to_file(phase_name); 5749 } 5750 } 5751 5752 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5753 void igv_print_default() { 5754 Compile::current()->print_method(PHASE_DEBUG, 0); 5755 } 5756 5757 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5758 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5759 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5760 void igv_append() { 5761 Compile::current()->igv_print_method_to_file("Debug", true); 5762 } 5763 5764 // Same as igv_append() above but with a specified phase name. 5765 void igv_append(const char* phase_name) { 5766 Compile::current()->igv_print_method_to_file(phase_name, true); 5767 } 5768 5769 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5770 const char* file_name = "custom_debug.xml"; 5771 if (_debug_file_printer == nullptr) { 5772 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5773 } else { 5774 _debug_file_printer->update_compiled_method(C->method()); 5775 } 5776 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5777 _debug_file_printer->print(phase_name, (Node*)C->root()); 5778 } 5779 5780 void Compile::igv_print_method_to_network(const char* phase_name) { 5781 if (_debug_network_printer == nullptr) { 5782 _debug_network_printer = new IdealGraphPrinter(C); 5783 } else { 5784 _debug_network_printer->update_compiled_method(C->method()); 5785 } 5786 tty->print_cr("Method printed over network stream to IGV"); 5787 _debug_network_printer->print(phase_name, (Node*)C->root()); 5788 } 5789 #endif 5790 5791 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5792 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5793 return value; 5794 } 5795 Node* result = nullptr; 5796 if (bt == T_BYTE) { 5797 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5798 result = new RShiftINode(result, phase->intcon(24)); 5799 } else if (bt == T_BOOLEAN) { 5800 result = new AndINode(value, phase->intcon(0xFF)); 5801 } else if (bt == T_CHAR) { 5802 result = new AndINode(value,phase->intcon(0xFFFF)); 5803 } else { 5804 assert(bt == T_SHORT, "unexpected narrow type"); 5805 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5806 result = new RShiftINode(result, phase->intcon(16)); 5807 } 5808 if (transform_res) { 5809 result = phase->transform(result); 5810 } 5811 return result; 5812 } 5813 5814 void Compile::record_method_not_compilable_oom() { 5815 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5816 }