1 /* 2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compilationFailureInfo.hpp" 33 #include "compiler/compilationMemoryStatistic.hpp" 34 #include "compiler/compileBroker.hpp" 35 #include "compiler/compileLog.hpp" 36 #include "compiler/compiler_globals.hpp" 37 #include "compiler/disassembler.hpp" 38 #include "compiler/oopMap.hpp" 39 #include "gc/shared/barrierSet.hpp" 40 #include "gc/shared/c2/barrierSetC2.hpp" 41 #include "jfr/jfrEvents.hpp" 42 #include "jvm_io.h" 43 #include "memory/allocation.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "opto/addnode.hpp" 46 #include "opto/block.hpp" 47 #include "opto/c2compiler.hpp" 48 #include "opto/callGenerator.hpp" 49 #include "opto/callnode.hpp" 50 #include "opto/castnode.hpp" 51 #include "opto/cfgnode.hpp" 52 #include "opto/chaitin.hpp" 53 #include "opto/compile.hpp" 54 #include "opto/connode.hpp" 55 #include "opto/convertnode.hpp" 56 #include "opto/divnode.hpp" 57 #include "opto/escape.hpp" 58 #include "opto/idealGraphPrinter.hpp" 59 #include "opto/inlinetypenode.hpp" 60 #include "opto/loopnode.hpp" 61 #include "opto/machnode.hpp" 62 #include "opto/macro.hpp" 63 #include "opto/matcher.hpp" 64 #include "opto/mathexactnode.hpp" 65 #include "opto/memnode.hpp" 66 #include "opto/mulnode.hpp" 67 #include "opto/narrowptrnode.hpp" 68 #include "opto/node.hpp" 69 #include "opto/opcodes.hpp" 70 #include "opto/output.hpp" 71 #include "opto/parse.hpp" 72 #include "opto/phaseX.hpp" 73 #include "opto/rootnode.hpp" 74 #include "opto/runtime.hpp" 75 #include "opto/stringopts.hpp" 76 #include "opto/type.hpp" 77 #include "opto/vector.hpp" 78 #include "opto/vectornode.hpp" 79 #include "runtime/globals_extension.hpp" 80 #include "runtime/sharedRuntime.hpp" 81 #include "runtime/signature.hpp" 82 #include "runtime/stubRoutines.hpp" 83 #include "runtime/timer.hpp" 84 #include "utilities/align.hpp" 85 #include "utilities/copy.hpp" 86 #include "utilities/macros.hpp" 87 #include "utilities/resourceHash.hpp" 88 89 // -------------------- Compile::mach_constant_base_node ----------------------- 90 // Constant table base node singleton. 91 MachConstantBaseNode* Compile::mach_constant_base_node() { 92 if (_mach_constant_base_node == nullptr) { 93 _mach_constant_base_node = new MachConstantBaseNode(); 94 _mach_constant_base_node->add_req(C->root()); 95 } 96 return _mach_constant_base_node; 97 } 98 99 100 /// Support for intrinsics. 101 102 // Return the index at which m must be inserted (or already exists). 103 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 104 class IntrinsicDescPair { 105 private: 106 ciMethod* _m; 107 bool _is_virtual; 108 public: 109 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 110 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 111 ciMethod* m= elt->method(); 112 ciMethod* key_m = key->_m; 113 if (key_m < m) return -1; 114 else if (key_m > m) return 1; 115 else { 116 bool is_virtual = elt->is_virtual(); 117 bool key_virtual = key->_is_virtual; 118 if (key_virtual < is_virtual) return -1; 119 else if (key_virtual > is_virtual) return 1; 120 else return 0; 121 } 122 } 123 }; 124 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 125 #ifdef ASSERT 126 for (int i = 1; i < _intrinsics.length(); i++) { 127 CallGenerator* cg1 = _intrinsics.at(i-1); 128 CallGenerator* cg2 = _intrinsics.at(i); 129 assert(cg1->method() != cg2->method() 130 ? cg1->method() < cg2->method() 131 : cg1->is_virtual() < cg2->is_virtual(), 132 "compiler intrinsics list must stay sorted"); 133 } 134 #endif 135 IntrinsicDescPair pair(m, is_virtual); 136 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 137 } 138 139 void Compile::register_intrinsic(CallGenerator* cg) { 140 bool found = false; 141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 142 assert(!found, "registering twice"); 143 _intrinsics.insert_before(index, cg); 144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 145 } 146 147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 148 assert(m->is_loaded(), "don't try this on unloaded methods"); 149 if (_intrinsics.length() > 0) { 150 bool found = false; 151 int index = intrinsic_insertion_index(m, is_virtual, found); 152 if (found) { 153 return _intrinsics.at(index); 154 } 155 } 156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 157 if (m->intrinsic_id() != vmIntrinsics::_none && 158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 160 if (cg != nullptr) { 161 // Save it for next time: 162 register_intrinsic(cg); 163 return cg; 164 } else { 165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 166 } 167 } 168 return nullptr; 169 } 170 171 // Compile::make_vm_intrinsic is defined in library_call.cpp. 172 173 #ifndef PRODUCT 174 // statistics gathering... 175 176 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 177 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 178 179 inline int as_int(vmIntrinsics::ID id) { 180 return vmIntrinsics::as_int(id); 181 } 182 183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 185 int oflags = _intrinsic_hist_flags[as_int(id)]; 186 assert(flags != 0, "what happened?"); 187 if (is_virtual) { 188 flags |= _intrinsic_virtual; 189 } 190 bool changed = (flags != oflags); 191 if ((flags & _intrinsic_worked) != 0) { 192 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 193 if (count == 1) { 194 changed = true; // first time 195 } 196 // increment the overall count also: 197 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 198 } 199 if (changed) { 200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 201 // Something changed about the intrinsic's virtuality. 202 if ((flags & _intrinsic_virtual) != 0) { 203 // This is the first use of this intrinsic as a virtual call. 204 if (oflags != 0) { 205 // We already saw it as a non-virtual, so note both cases. 206 flags |= _intrinsic_both; 207 } 208 } else if ((oflags & _intrinsic_both) == 0) { 209 // This is the first use of this intrinsic as a non-virtual 210 flags |= _intrinsic_both; 211 } 212 } 213 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 214 } 215 // update the overall flags also: 216 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 217 return changed; 218 } 219 220 static char* format_flags(int flags, char* buf) { 221 buf[0] = 0; 222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 227 if (buf[0] == 0) strcat(buf, ","); 228 assert(buf[0] == ',', "must be"); 229 return &buf[1]; 230 } 231 232 void Compile::print_intrinsic_statistics() { 233 char flagsbuf[100]; 234 ttyLocker ttyl; 235 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 236 tty->print_cr("Compiler intrinsic usage:"); 237 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 238 if (total == 0) total = 1; // avoid div0 in case of no successes 239 #define PRINT_STAT_LINE(name, c, f) \ 240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 241 for (auto id : EnumRange<vmIntrinsicID>{}) { 242 int flags = _intrinsic_hist_flags[as_int(id)]; 243 juint count = _intrinsic_hist_count[as_int(id)]; 244 if ((flags | count) != 0) { 245 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 246 } 247 } 248 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 249 if (xtty != nullptr) xtty->tail("statistics"); 250 } 251 252 void Compile::print_statistics() { 253 { ttyLocker ttyl; 254 if (xtty != nullptr) xtty->head("statistics type='opto'"); 255 Parse::print_statistics(); 256 PhaseStringOpts::print_statistics(); 257 PhaseCCP::print_statistics(); 258 PhaseRegAlloc::print_statistics(); 259 PhaseOutput::print_statistics(); 260 PhasePeephole::print_statistics(); 261 PhaseIdealLoop::print_statistics(); 262 ConnectionGraph::print_statistics(); 263 PhaseMacroExpand::print_statistics(); 264 if (xtty != nullptr) xtty->tail("statistics"); 265 } 266 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 267 // put this under its own <statistics> element. 268 print_intrinsic_statistics(); 269 } 270 } 271 #endif //PRODUCT 272 273 void Compile::gvn_replace_by(Node* n, Node* nn) { 274 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 275 Node* use = n->last_out(i); 276 bool is_in_table = initial_gvn()->hash_delete(use); 277 uint uses_found = 0; 278 for (uint j = 0; j < use->len(); j++) { 279 if (use->in(j) == n) { 280 if (j < use->req()) 281 use->set_req(j, nn); 282 else 283 use->set_prec(j, nn); 284 uses_found++; 285 } 286 } 287 if (is_in_table) { 288 // reinsert into table 289 initial_gvn()->hash_find_insert(use); 290 } 291 record_for_igvn(use); 292 PhaseIterGVN::add_users_of_use_to_worklist(nn, use, *_igvn_worklist); 293 i -= uses_found; // we deleted 1 or more copies of this edge 294 } 295 } 296 297 298 // Identify all nodes that are reachable from below, useful. 299 // Use breadth-first pass that records state in a Unique_Node_List, 300 // recursive traversal is slower. 301 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 302 int estimated_worklist_size = live_nodes(); 303 useful.map( estimated_worklist_size, nullptr ); // preallocate space 304 305 // Initialize worklist 306 if (root() != nullptr) { useful.push(root()); } 307 // If 'top' is cached, declare it useful to preserve cached node 308 if (cached_top_node()) { useful.push(cached_top_node()); } 309 310 // Push all useful nodes onto the list, breadthfirst 311 for( uint next = 0; next < useful.size(); ++next ) { 312 assert( next < unique(), "Unique useful nodes < total nodes"); 313 Node *n = useful.at(next); 314 uint max = n->len(); 315 for( uint i = 0; i < max; ++i ) { 316 Node *m = n->in(i); 317 if (not_a_node(m)) continue; 318 useful.push(m); 319 } 320 } 321 } 322 323 // Update dead_node_list with any missing dead nodes using useful 324 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 325 void Compile::update_dead_node_list(Unique_Node_List &useful) { 326 uint max_idx = unique(); 327 VectorSet& useful_node_set = useful.member_set(); 328 329 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 330 // If node with index node_idx is not in useful set, 331 // mark it as dead in dead node list. 332 if (!useful_node_set.test(node_idx)) { 333 record_dead_node(node_idx); 334 } 335 } 336 } 337 338 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 339 int shift = 0; 340 for (int i = 0; i < inlines->length(); i++) { 341 CallGenerator* cg = inlines->at(i); 342 if (useful.member(cg->call_node())) { 343 if (shift > 0) { 344 inlines->at_put(i - shift, cg); 345 } 346 } else { 347 shift++; // skip over the dead element 348 } 349 } 350 if (shift > 0) { 351 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 352 } 353 } 354 355 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 356 assert(dead != nullptr && dead->is_Call(), "sanity"); 357 int found = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 if (inlines->at(i)->call_node() == dead) { 360 inlines->remove_at(i); 361 found++; 362 NOT_DEBUG( break; ) // elements are unique, so exit early 363 } 364 } 365 assert(found <= 1, "not unique"); 366 } 367 368 template<typename N, ENABLE_IF_SDEFN(std::is_base_of<Node, N>::value)> 369 void Compile::remove_useless_nodes(GrowableArray<N*>& node_list, Unique_Node_List& useful) { 370 for (int i = node_list.length() - 1; i >= 0; i--) { 371 N* node = node_list.at(i); 372 if (!useful.member(node)) { 373 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 374 } 375 } 376 } 377 378 void Compile::remove_useless_node(Node* dead) { 379 remove_modified_node(dead); 380 381 // Constant node that has no out-edges and has only one in-edge from 382 // root is usually dead. However, sometimes reshaping walk makes 383 // it reachable by adding use edges. So, we will NOT count Con nodes 384 // as dead to be conservative about the dead node count at any 385 // given time. 386 if (!dead->is_Con()) { 387 record_dead_node(dead->_idx); 388 } 389 if (dead->is_macro()) { 390 remove_macro_node(dead); 391 } 392 if (dead->is_expensive()) { 393 remove_expensive_node(dead); 394 } 395 if (dead->Opcode() == Op_Opaque4) { 396 remove_template_assertion_predicate_opaq(dead); 397 } 398 if (dead->is_ParsePredicate()) { 399 remove_parse_predicate(dead->as_ParsePredicate()); 400 } 401 if (dead->for_post_loop_opts_igvn()) { 402 remove_from_post_loop_opts_igvn(dead); 403 } 404 if (dead->is_InlineType()) { 405 remove_inline_type(dead); 406 } 407 if (dead->is_Call()) { 408 remove_useless_late_inlines( &_late_inlines, dead); 409 remove_useless_late_inlines( &_string_late_inlines, dead); 410 remove_useless_late_inlines( &_boxing_late_inlines, dead); 411 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 412 413 if (dead->is_CallStaticJava()) { 414 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 415 } 416 } 417 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 418 bs->unregister_potential_barrier_node(dead); 419 } 420 421 // Disconnect all useless nodes by disconnecting those at the boundary. 422 void Compile::disconnect_useless_nodes(Unique_Node_List& useful, Unique_Node_List& worklist) { 423 uint next = 0; 424 while (next < useful.size()) { 425 Node *n = useful.at(next++); 426 if (n->is_SafePoint()) { 427 // We're done with a parsing phase. Replaced nodes are not valid 428 // beyond that point. 429 n->as_SafePoint()->delete_replaced_nodes(); 430 } 431 // Use raw traversal of out edges since this code removes out edges 432 int max = n->outcnt(); 433 for (int j = 0; j < max; ++j) { 434 Node* child = n->raw_out(j); 435 if (!useful.member(child)) { 436 assert(!child->is_top() || child != top(), 437 "If top is cached in Compile object it is in useful list"); 438 // Only need to remove this out-edge to the useless node 439 n->raw_del_out(j); 440 --j; 441 --max; 442 } 443 } 444 if (n->outcnt() == 1 && n->has_special_unique_user()) { 445 assert(useful.member(n->unique_out()), "do not push a useless node"); 446 worklist.push(n->unique_out()); 447 } 448 if (n->outcnt() == 0) { 449 worklist.push(n); 450 } 451 } 452 453 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 454 remove_useless_nodes(_parse_predicates, useful); // remove useless Parse Predicate nodes 455 remove_useless_nodes(_template_assertion_predicate_opaqs, useful); // remove useless Assertion Predicate opaque nodes 456 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 457 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 458 remove_useless_nodes(_inline_type_nodes, useful); // remove useless inline type nodes 459 #ifdef ASSERT 460 if (_modified_nodes != nullptr) { 461 _modified_nodes->remove_useless_nodes(useful.member_set()); 462 } 463 #endif 464 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 465 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 466 #ifdef ASSERT 467 if (_modified_nodes != nullptr) { 468 _modified_nodes->remove_useless_nodes(useful.member_set()); 469 } 470 #endif 471 472 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 473 bs->eliminate_useless_gc_barriers(useful, this); 474 // clean up the late inline lists 475 remove_useless_late_inlines( &_late_inlines, useful); 476 remove_useless_late_inlines( &_string_late_inlines, useful); 477 remove_useless_late_inlines( &_boxing_late_inlines, useful); 478 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 479 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 480 } 481 482 // ============================================================================ 483 //------------------------------CompileWrapper--------------------------------- 484 class CompileWrapper : public StackObj { 485 Compile *const _compile; 486 public: 487 CompileWrapper(Compile* compile); 488 489 ~CompileWrapper(); 490 }; 491 492 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 493 // the Compile* pointer is stored in the current ciEnv: 494 ciEnv* env = compile->env(); 495 assert(env == ciEnv::current(), "must already be a ciEnv active"); 496 assert(env->compiler_data() == nullptr, "compile already active?"); 497 env->set_compiler_data(compile); 498 assert(compile == Compile::current(), "sanity"); 499 500 compile->set_type_dict(nullptr); 501 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 502 compile->clone_map().set_clone_idx(0); 503 compile->set_type_last_size(0); 504 compile->set_last_tf(nullptr, nullptr); 505 compile->set_indexSet_arena(nullptr); 506 compile->set_indexSet_free_block_list(nullptr); 507 compile->init_type_arena(); 508 Type::Initialize(compile); 509 _compile->begin_method(); 510 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 511 } 512 CompileWrapper::~CompileWrapper() { 513 // simulate crash during compilation 514 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 515 516 _compile->end_method(); 517 _compile->env()->set_compiler_data(nullptr); 518 } 519 520 521 //----------------------------print_compile_messages--------------------------- 522 void Compile::print_compile_messages() { 523 #ifndef PRODUCT 524 // Check if recompiling 525 if (!subsume_loads() && PrintOpto) { 526 // Recompiling without allowing machine instructions to subsume loads 527 tty->print_cr("*********************************************************"); 528 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 529 tty->print_cr("*********************************************************"); 530 } 531 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 532 // Recompiling without escape analysis 533 tty->print_cr("*********************************************************"); 534 tty->print_cr("** Bailout: Recompile without escape analysis **"); 535 tty->print_cr("*********************************************************"); 536 } 537 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 538 // Recompiling without iterative escape analysis 539 tty->print_cr("*********************************************************"); 540 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 541 tty->print_cr("*********************************************************"); 542 } 543 if (do_reduce_allocation_merges() != ReduceAllocationMerges && PrintOpto) { 544 // Recompiling without reducing allocation merges 545 tty->print_cr("*********************************************************"); 546 tty->print_cr("** Bailout: Recompile without reduce allocation merges **"); 547 tty->print_cr("*********************************************************"); 548 } 549 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 550 // Recompiling without boxing elimination 551 tty->print_cr("*********************************************************"); 552 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 553 tty->print_cr("*********************************************************"); 554 } 555 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 556 // Recompiling without locks coarsening 557 tty->print_cr("*********************************************************"); 558 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 559 tty->print_cr("*********************************************************"); 560 } 561 if (env()->break_at_compile()) { 562 // Open the debugger when compiling this method. 563 tty->print("### Breaking when compiling: "); 564 method()->print_short_name(); 565 tty->cr(); 566 BREAKPOINT; 567 } 568 569 if( PrintOpto ) { 570 if (is_osr_compilation()) { 571 tty->print("[OSR]%3d", _compile_id); 572 } else { 573 tty->print("%3d", _compile_id); 574 } 575 } 576 #endif 577 } 578 579 #ifndef PRODUCT 580 void Compile::print_ideal_ir(const char* phase_name) { 581 // keep the following output all in one block 582 // This output goes directly to the tty, not the compiler log. 583 // To enable tools to match it up with the compilation activity, 584 // be sure to tag this tty output with the compile ID. 585 586 // Node dumping can cause a safepoint, which can break the tty lock. 587 // Buffer all node dumps, so that all safepoints happen before we lock. 588 ResourceMark rm; 589 stringStream ss; 590 591 if (_output == nullptr) { 592 ss.print_cr("AFTER: %s", phase_name); 593 // Print out all nodes in ascending order of index. 594 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$", &ss); 595 } else { 596 // Dump the node blockwise if we have a scheduling 597 _output->print_scheduling(&ss); 598 } 599 600 // Check that the lock is not broken by a safepoint. 601 NoSafepointVerifier nsv; 602 ttyLocker ttyl; 603 if (xtty != nullptr) { 604 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 605 compile_id(), 606 is_osr_compilation() ? " compile_kind='osr'" : "", 607 phase_name); 608 } 609 610 tty->print("%s", ss.as_string()); 611 612 if (xtty != nullptr) { 613 xtty->tail("ideal"); 614 } 615 } 616 #endif 617 618 // ============================================================================ 619 //------------------------------Compile standard------------------------------- 620 621 // Compile a method. entry_bci is -1 for normal compilations and indicates 622 // the continuation bci for on stack replacement. 623 624 625 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 626 Options options, DirectiveSet* directive) 627 : Phase(Compiler), 628 _compile_id(ci_env->compile_id()), 629 _options(options), 630 _method(target), 631 _entry_bci(osr_bci), 632 _ilt(nullptr), 633 _stub_function(nullptr), 634 _stub_name(nullptr), 635 _stub_entry_point(nullptr), 636 _max_node_limit(MaxNodeLimit), 637 _post_loop_opts_phase(false), 638 _inlining_progress(false), 639 _inlining_incrementally(false), 640 _do_cleanup(false), 641 _has_reserved_stack_access(target->has_reserved_stack_access()), 642 _has_circular_inline_type(false), 643 #ifndef PRODUCT 644 _igv_idx(0), 645 _trace_opto_output(directive->TraceOptoOutputOption), 646 #endif 647 _has_method_handle_invokes(false), 648 _clinit_barrier_on_entry(false), 649 _stress_seed(0), 650 _comp_arena(mtCompiler), 651 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 652 _env(ci_env), 653 _directive(directive), 654 _log(ci_env->log()), 655 _failure_reason(nullptr), 656 _first_failure_details(nullptr), 657 _intrinsics (comp_arena(), 0, 0, nullptr), 658 _macro_nodes (comp_arena(), 8, 0, nullptr), 659 _parse_predicates (comp_arena(), 8, 0, nullptr), 660 _template_assertion_predicate_opaqs (comp_arena(), 8, 0, nullptr), 661 _expensive_nodes (comp_arena(), 8, 0, nullptr), 662 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 663 _inline_type_nodes (comp_arena(), 8, 0, nullptr), 664 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 665 _coarsened_locks (comp_arena(), 8, 0, nullptr), 666 _congraph(nullptr), 667 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 668 _unique(0), 669 _dead_node_count(0), 670 _dead_node_list(comp_arena()), 671 _node_arena_one(mtCompiler, Arena::Tag::tag_node), 672 _node_arena_two(mtCompiler, Arena::Tag::tag_node), 673 _node_arena(&_node_arena_one), 674 _mach_constant_base_node(nullptr), 675 _Compile_types(mtCompiler), 676 _initial_gvn(nullptr), 677 _igvn_worklist(nullptr), 678 _types(nullptr), 679 _node_hash(nullptr), 680 _late_inlines(comp_arena(), 2, 0, nullptr), 681 _string_late_inlines(comp_arena(), 2, 0, nullptr), 682 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 683 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 684 _late_inlines_pos(0), 685 _number_of_mh_late_inlines(0), 686 _oom(false), 687 _print_inlining_stream(new (mtCompiler) stringStream()), 688 _print_inlining_list(nullptr), 689 _print_inlining_idx(0), 690 _print_inlining_output(nullptr), 691 _replay_inline_data(nullptr), 692 _java_calls(0), 693 _inner_loops(0), 694 _interpreter_frame_size(0), 695 _output(nullptr) 696 #ifndef PRODUCT 697 , _in_dump_cnt(0) 698 #endif 699 { 700 C = this; 701 CompileWrapper cw(this); 702 703 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 704 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 705 706 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 707 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 708 // We can always print a disassembly, either abstract (hex dump) or 709 // with the help of a suitable hsdis library. Thus, we should not 710 // couple print_assembly and print_opto_assembly controls. 711 // But: always print opto and regular assembly on compile command 'print'. 712 bool print_assembly = directive->PrintAssemblyOption; 713 set_print_assembly(print_opto_assembly || print_assembly); 714 #else 715 set_print_assembly(false); // must initialize. 716 #endif 717 718 #ifndef PRODUCT 719 set_parsed_irreducible_loop(false); 720 #endif 721 722 if (directive->ReplayInlineOption) { 723 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 724 } 725 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 726 set_print_intrinsics(directive->PrintIntrinsicsOption); 727 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 728 729 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 730 // Make sure the method being compiled gets its own MDO, 731 // so we can at least track the decompile_count(). 732 // Need MDO to record RTM code generation state. 733 method()->ensure_method_data(); 734 } 735 736 Init(/*do_aliasing=*/ true); 737 738 print_compile_messages(); 739 740 _ilt = InlineTree::build_inline_tree_root(); 741 742 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 743 assert(num_alias_types() >= AliasIdxRaw, ""); 744 745 #define MINIMUM_NODE_HASH 1023 746 747 // GVN that will be run immediately on new nodes 748 uint estimated_size = method()->code_size()*4+64; 749 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 750 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 751 _types = new (comp_arena()) Type_Array(comp_arena()); 752 _node_hash = new (comp_arena()) NodeHash(comp_arena(), estimated_size); 753 PhaseGVN gvn; 754 set_initial_gvn(&gvn); 755 756 print_inlining_init(); 757 { // Scope for timing the parser 758 TracePhase tp("parse", &timers[_t_parser]); 759 760 // Put top into the hash table ASAP. 761 initial_gvn()->transform(top()); 762 763 // Set up tf(), start(), and find a CallGenerator. 764 CallGenerator* cg = nullptr; 765 if (is_osr_compilation()) { 766 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true)); 767 StartNode* s = new StartOSRNode(root(), tf()->domain_sig()); 768 initial_gvn()->set_type_bottom(s); 769 init_start(s); 770 cg = CallGenerator::for_osr(method(), entry_bci()); 771 } else { 772 // Normal case. 773 init_tf(TypeFunc::make(method())); 774 StartNode* s = new StartNode(root(), tf()->domain_cc()); 775 initial_gvn()->set_type_bottom(s); 776 init_start(s); 777 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 778 // With java.lang.ref.reference.get() we must go through the 779 // intrinsic - even when get() is the root 780 // method of the compile - so that, if necessary, the value in 781 // the referent field of the reference object gets recorded by 782 // the pre-barrier code. 783 cg = find_intrinsic(method(), false); 784 } 785 if (cg == nullptr) { 786 float past_uses = method()->interpreter_invocation_count(); 787 float expected_uses = past_uses; 788 cg = CallGenerator::for_inline(method(), expected_uses); 789 } 790 } 791 if (failing()) return; 792 if (cg == nullptr) { 793 const char* reason = InlineTree::check_can_parse(method()); 794 assert(reason != nullptr, "expect reason for parse failure"); 795 stringStream ss; 796 ss.print("cannot parse method: %s", reason); 797 record_method_not_compilable(ss.as_string()); 798 return; 799 } 800 801 gvn.set_type(root(), root()->bottom_type()); 802 803 JVMState* jvms = build_start_state(start(), tf()); 804 if ((jvms = cg->generate(jvms)) == nullptr) { 805 assert(failure_reason() != nullptr, "expect reason for parse failure"); 806 stringStream ss; 807 ss.print("method parse failed: %s", failure_reason()); 808 record_method_not_compilable(ss.as_string()); 809 return; 810 } 811 GraphKit kit(jvms); 812 813 if (!kit.stopped()) { 814 // Accept return values, and transfer control we know not where. 815 // This is done by a special, unique ReturnNode bound to root. 816 return_values(kit.jvms()); 817 } 818 819 if (kit.has_exceptions()) { 820 // Any exceptions that escape from this call must be rethrown 821 // to whatever caller is dynamically above us on the stack. 822 // This is done by a special, unique RethrowNode bound to root. 823 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 824 } 825 826 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 827 828 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 829 inline_string_calls(true); 830 } 831 832 if (failing()) return; 833 834 // Remove clutter produced by parsing. 835 if (!failing()) { 836 ResourceMark rm; 837 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 838 } 839 } 840 841 // Note: Large methods are capped off in do_one_bytecode(). 842 if (failing()) return; 843 844 // After parsing, node notes are no longer automagic. 845 // They must be propagated by register_new_node_with_optimizer(), 846 // clone(), or the like. 847 set_default_node_notes(nullptr); 848 849 #ifndef PRODUCT 850 if (should_print_igv(1)) { 851 _igv_printer->print_inlining(); 852 } 853 #endif 854 855 if (failing()) return; 856 NOT_PRODUCT( verify_graph_edges(); ) 857 858 // If any phase is randomized for stress testing, seed random number 859 // generation and log the seed for repeatability. 860 if (StressLCM || StressGCM || StressIGVN || StressCCP || 861 StressIncrementalInlining || StressMacroExpansion) { 862 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && directive->RepeatCompilationOption)) { 863 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 864 FLAG_SET_ERGO(StressSeed, _stress_seed); 865 } else { 866 _stress_seed = StressSeed; 867 } 868 if (_log != nullptr) { 869 _log->elem("stress_test seed='%u'", _stress_seed); 870 } 871 } 872 873 // Now optimize 874 Optimize(); 875 if (failing()) return; 876 NOT_PRODUCT( verify_graph_edges(); ) 877 878 #ifndef PRODUCT 879 if (should_print_ideal()) { 880 print_ideal_ir("print_ideal"); 881 } 882 #endif 883 884 #ifdef ASSERT 885 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 886 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 887 #endif 888 889 // Dump compilation data to replay it. 890 if (directive->DumpReplayOption) { 891 env()->dump_replay_data(_compile_id); 892 } 893 if (directive->DumpInlineOption && (ilt() != nullptr)) { 894 env()->dump_inline_data(_compile_id); 895 } 896 897 // Now that we know the size of all the monitors we can add a fixed slot 898 // for the original deopt pc. 899 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 900 if (needs_stack_repair()) { 901 // One extra slot for the special stack increment value 902 next_slot += 2; 903 } 904 // TODO 8284443 Only reserve extra slot if needed 905 if (InlineTypeReturnedAsFields) { 906 // One extra slot to hold the IsInit information for a nullable 907 // inline type return if we run out of registers. 908 next_slot += 2; 909 } 910 set_fixed_slots(next_slot); 911 912 // Compute when to use implicit null checks. Used by matching trap based 913 // nodes and NullCheck optimization. 914 set_allowed_deopt_reasons(); 915 916 // Now generate code 917 Code_Gen(); 918 } 919 920 //------------------------------Compile---------------------------------------- 921 // Compile a runtime stub 922 Compile::Compile( ciEnv* ci_env, 923 TypeFunc_generator generator, 924 address stub_function, 925 const char *stub_name, 926 int is_fancy_jump, 927 bool pass_tls, 928 bool return_pc, 929 DirectiveSet* directive) 930 : Phase(Compiler), 931 _compile_id(0), 932 _options(Options::for_runtime_stub()), 933 _method(nullptr), 934 _entry_bci(InvocationEntryBci), 935 _stub_function(stub_function), 936 _stub_name(stub_name), 937 _stub_entry_point(nullptr), 938 _max_node_limit(MaxNodeLimit), 939 _post_loop_opts_phase(false), 940 _inlining_progress(false), 941 _inlining_incrementally(false), 942 _has_reserved_stack_access(false), 943 _has_circular_inline_type(false), 944 #ifndef PRODUCT 945 _igv_idx(0), 946 _trace_opto_output(directive->TraceOptoOutputOption), 947 #endif 948 _has_method_handle_invokes(false), 949 _clinit_barrier_on_entry(false), 950 _stress_seed(0), 951 _comp_arena(mtCompiler), 952 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 953 _env(ci_env), 954 _directive(directive), 955 _log(ci_env->log()), 956 _failure_reason(nullptr), 957 _first_failure_details(nullptr), 958 _congraph(nullptr), 959 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 960 _unique(0), 961 _dead_node_count(0), 962 _dead_node_list(comp_arena()), 963 _node_arena_one(mtCompiler), 964 _node_arena_two(mtCompiler), 965 _node_arena(&_node_arena_one), 966 _mach_constant_base_node(nullptr), 967 _Compile_types(mtCompiler), 968 _initial_gvn(nullptr), 969 _igvn_worklist(nullptr), 970 _types(nullptr), 971 _node_hash(nullptr), 972 _number_of_mh_late_inlines(0), 973 _oom(false), 974 _print_inlining_stream(new (mtCompiler) stringStream()), 975 _print_inlining_list(nullptr), 976 _print_inlining_idx(0), 977 _print_inlining_output(nullptr), 978 _replay_inline_data(nullptr), 979 _java_calls(0), 980 _inner_loops(0), 981 _interpreter_frame_size(0), 982 _output(nullptr), 983 #ifndef PRODUCT 984 _in_dump_cnt(0), 985 #endif 986 _allowed_reasons(0) { 987 C = this; 988 989 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 990 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 991 992 #ifndef PRODUCT 993 set_print_assembly(PrintFrameConverterAssembly); 994 set_parsed_irreducible_loop(false); 995 #else 996 set_print_assembly(false); // Must initialize. 997 #endif 998 set_has_irreducible_loop(false); // no loops 999 1000 CompileWrapper cw(this); 1001 Init(/*do_aliasing=*/ false); 1002 init_tf((*generator)()); 1003 1004 _igvn_worklist = new (comp_arena()) Unique_Node_List(comp_arena()); 1005 _types = new (comp_arena()) Type_Array(comp_arena()); 1006 _node_hash = new (comp_arena()) NodeHash(comp_arena(), 255); 1007 { 1008 PhaseGVN gvn; 1009 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1010 gvn.transform(top()); 1011 1012 GraphKit kit; 1013 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1014 } 1015 1016 NOT_PRODUCT( verify_graph_edges(); ) 1017 1018 Code_Gen(); 1019 } 1020 1021 Compile::~Compile() { 1022 delete _print_inlining_stream; 1023 delete _first_failure_details; 1024 }; 1025 1026 //------------------------------Init------------------------------------------- 1027 // Prepare for a single compilation 1028 void Compile::Init(bool aliasing) { 1029 _do_aliasing = aliasing; 1030 _unique = 0; 1031 _regalloc = nullptr; 1032 1033 _tf = nullptr; // filled in later 1034 _top = nullptr; // cached later 1035 _matcher = nullptr; // filled in later 1036 _cfg = nullptr; // filled in later 1037 1038 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 1039 1040 _node_note_array = nullptr; 1041 _default_node_notes = nullptr; 1042 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 1043 1044 _immutable_memory = nullptr; // filled in at first inquiry 1045 1046 #ifdef ASSERT 1047 _phase_optimize_finished = false; 1048 _exception_backedge = false; 1049 _type_verify = nullptr; 1050 #endif 1051 1052 // Globally visible Nodes 1053 // First set TOP to null to give safe behavior during creation of RootNode 1054 set_cached_top_node(nullptr); 1055 set_root(new RootNode()); 1056 // Now that you have a Root to point to, create the real TOP 1057 set_cached_top_node( new ConNode(Type::TOP) ); 1058 set_recent_alloc(nullptr, nullptr); 1059 1060 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1061 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1062 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1063 env()->set_dependencies(new Dependencies(env())); 1064 1065 _fixed_slots = 0; 1066 set_has_split_ifs(false); 1067 set_has_loops(false); // first approximation 1068 set_has_stringbuilder(false); 1069 set_has_boxed_value(false); 1070 _trap_can_recompile = false; // no traps emitted yet 1071 _major_progress = true; // start out assuming good things will happen 1072 set_has_unsafe_access(false); 1073 set_max_vector_size(0); 1074 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1075 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1076 set_decompile_count(0); 1077 1078 #ifndef PRODUCT 1079 Copy::zero_to_bytes(_igv_phase_iter, sizeof(_igv_phase_iter)); 1080 #endif 1081 1082 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1083 _loop_opts_cnt = LoopOptsCount; 1084 _has_flat_accesses = false; 1085 _flat_accesses_share_alias = true; 1086 _scalarize_in_safepoints = false; 1087 1088 set_do_inlining(Inline); 1089 set_max_inline_size(MaxInlineSize); 1090 set_freq_inline_size(FreqInlineSize); 1091 set_do_scheduling(OptoScheduling); 1092 1093 set_do_vector_loop(false); 1094 set_has_monitors(false); 1095 1096 if (AllowVectorizeOnDemand) { 1097 if (has_method() && _directive->VectorizeOption) { 1098 set_do_vector_loop(true); 1099 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());}) 1100 } else if (has_method() && method()->name() != 0 && 1101 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1102 set_do_vector_loop(true); 1103 } 1104 } 1105 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1106 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());}) 1107 1108 set_rtm_state(NoRTM); // No RTM lock eliding by default 1109 _max_node_limit = _directive->MaxNodeLimitOption; 1110 1111 #if INCLUDE_RTM_OPT 1112 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1113 int rtm_state = method()->method_data()->rtm_state(); 1114 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1115 // Don't generate RTM lock eliding code. 1116 set_rtm_state(NoRTM); 1117 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1118 // Generate RTM lock eliding code without abort ratio calculation code. 1119 set_rtm_state(UseRTM); 1120 } else if (UseRTMDeopt) { 1121 // Generate RTM lock eliding code and include abort ratio calculation 1122 // code if UseRTMDeopt is on. 1123 set_rtm_state(ProfileRTM); 1124 } 1125 } 1126 #endif 1127 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1128 set_clinit_barrier_on_entry(true); 1129 } 1130 if (debug_info()->recording_non_safepoints()) { 1131 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1132 (comp_arena(), 8, 0, nullptr)); 1133 set_default_node_notes(Node_Notes::make(this)); 1134 } 1135 1136 const int grow_ats = 16; 1137 _max_alias_types = grow_ats; 1138 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1139 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1140 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1141 { 1142 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1143 } 1144 // Initialize the first few types. 1145 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1146 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1147 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1148 _num_alias_types = AliasIdxRaw+1; 1149 // Zero out the alias type cache. 1150 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1151 // A null adr_type hits in the cache right away. Preload the right answer. 1152 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1153 } 1154 1155 //---------------------------init_start---------------------------------------- 1156 // Install the StartNode on this compile object. 1157 void Compile::init_start(StartNode* s) { 1158 if (failing()) 1159 return; // already failing 1160 assert(s == start(), ""); 1161 } 1162 1163 /** 1164 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1165 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1166 * the ideal graph. 1167 */ 1168 StartNode* Compile::start() const { 1169 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1170 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1171 Node* start = root()->fast_out(i); 1172 if (start->is_Start()) { 1173 return start->as_Start(); 1174 } 1175 } 1176 fatal("Did not find Start node!"); 1177 return nullptr; 1178 } 1179 1180 //-------------------------------immutable_memory------------------------------------- 1181 // Access immutable memory 1182 Node* Compile::immutable_memory() { 1183 if (_immutable_memory != nullptr) { 1184 return _immutable_memory; 1185 } 1186 StartNode* s = start(); 1187 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1188 Node *p = s->fast_out(i); 1189 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1190 _immutable_memory = p; 1191 return _immutable_memory; 1192 } 1193 } 1194 ShouldNotReachHere(); 1195 return nullptr; 1196 } 1197 1198 //----------------------set_cached_top_node------------------------------------ 1199 // Install the cached top node, and make sure Node::is_top works correctly. 1200 void Compile::set_cached_top_node(Node* tn) { 1201 if (tn != nullptr) verify_top(tn); 1202 Node* old_top = _top; 1203 _top = tn; 1204 // Calling Node::setup_is_top allows the nodes the chance to adjust 1205 // their _out arrays. 1206 if (_top != nullptr) _top->setup_is_top(); 1207 if (old_top != nullptr) old_top->setup_is_top(); 1208 assert(_top == nullptr || top()->is_top(), ""); 1209 } 1210 1211 #ifdef ASSERT 1212 uint Compile::count_live_nodes_by_graph_walk() { 1213 Unique_Node_List useful(comp_arena()); 1214 // Get useful node list by walking the graph. 1215 identify_useful_nodes(useful); 1216 return useful.size(); 1217 } 1218 1219 void Compile::print_missing_nodes() { 1220 1221 // Return if CompileLog is null and PrintIdealNodeCount is false. 1222 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1223 return; 1224 } 1225 1226 // This is an expensive function. It is executed only when the user 1227 // specifies VerifyIdealNodeCount option or otherwise knows the 1228 // additional work that needs to be done to identify reachable nodes 1229 // by walking the flow graph and find the missing ones using 1230 // _dead_node_list. 1231 1232 Unique_Node_List useful(comp_arena()); 1233 // Get useful node list by walking the graph. 1234 identify_useful_nodes(useful); 1235 1236 uint l_nodes = C->live_nodes(); 1237 uint l_nodes_by_walk = useful.size(); 1238 1239 if (l_nodes != l_nodes_by_walk) { 1240 if (_log != nullptr) { 1241 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1242 _log->stamp(); 1243 _log->end_head(); 1244 } 1245 VectorSet& useful_member_set = useful.member_set(); 1246 int last_idx = l_nodes_by_walk; 1247 for (int i = 0; i < last_idx; i++) { 1248 if (useful_member_set.test(i)) { 1249 if (_dead_node_list.test(i)) { 1250 if (_log != nullptr) { 1251 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1252 } 1253 if (PrintIdealNodeCount) { 1254 // Print the log message to tty 1255 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1256 useful.at(i)->dump(); 1257 } 1258 } 1259 } 1260 else if (! _dead_node_list.test(i)) { 1261 if (_log != nullptr) { 1262 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1263 } 1264 if (PrintIdealNodeCount) { 1265 // Print the log message to tty 1266 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1267 } 1268 } 1269 } 1270 if (_log != nullptr) { 1271 _log->tail("mismatched_nodes"); 1272 } 1273 } 1274 } 1275 void Compile::record_modified_node(Node* n) { 1276 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1277 _modified_nodes->push(n); 1278 } 1279 } 1280 1281 void Compile::remove_modified_node(Node* n) { 1282 if (_modified_nodes != nullptr) { 1283 _modified_nodes->remove(n); 1284 } 1285 } 1286 #endif 1287 1288 #ifndef PRODUCT 1289 void Compile::verify_top(Node* tn) const { 1290 if (tn != nullptr) { 1291 assert(tn->is_Con(), "top node must be a constant"); 1292 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1293 assert(tn->in(0) != nullptr, "must have live top node"); 1294 } 1295 } 1296 #endif 1297 1298 1299 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1300 1301 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1302 guarantee(arr != nullptr, ""); 1303 int num_blocks = arr->length(); 1304 if (grow_by < num_blocks) grow_by = num_blocks; 1305 int num_notes = grow_by * _node_notes_block_size; 1306 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1307 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1308 while (num_notes > 0) { 1309 arr->append(notes); 1310 notes += _node_notes_block_size; 1311 num_notes -= _node_notes_block_size; 1312 } 1313 assert(num_notes == 0, "exact multiple, please"); 1314 } 1315 1316 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1317 if (source == nullptr || dest == nullptr) return false; 1318 1319 if (dest->is_Con()) 1320 return false; // Do not push debug info onto constants. 1321 1322 #ifdef ASSERT 1323 // Leave a bread crumb trail pointing to the original node: 1324 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1325 dest->set_debug_orig(source); 1326 } 1327 #endif 1328 1329 if (node_note_array() == nullptr) 1330 return false; // Not collecting any notes now. 1331 1332 // This is a copy onto a pre-existing node, which may already have notes. 1333 // If both nodes have notes, do not overwrite any pre-existing notes. 1334 Node_Notes* source_notes = node_notes_at(source->_idx); 1335 if (source_notes == nullptr || source_notes->is_clear()) return false; 1336 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1337 if (dest_notes == nullptr || dest_notes->is_clear()) { 1338 return set_node_notes_at(dest->_idx, source_notes); 1339 } 1340 1341 Node_Notes merged_notes = (*source_notes); 1342 // The order of operations here ensures that dest notes will win... 1343 merged_notes.update_from(dest_notes); 1344 return set_node_notes_at(dest->_idx, &merged_notes); 1345 } 1346 1347 1348 //--------------------------allow_range_check_smearing------------------------- 1349 // Gating condition for coalescing similar range checks. 1350 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1351 // single covering check that is at least as strong as any of them. 1352 // If the optimization succeeds, the simplified (strengthened) range check 1353 // will always succeed. If it fails, we will deopt, and then give up 1354 // on the optimization. 1355 bool Compile::allow_range_check_smearing() const { 1356 // If this method has already thrown a range-check, 1357 // assume it was because we already tried range smearing 1358 // and it failed. 1359 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1360 return !already_trapped; 1361 } 1362 1363 1364 //------------------------------flatten_alias_type----------------------------- 1365 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1366 assert(do_aliasing(), "Aliasing should be enabled"); 1367 int offset = tj->offset(); 1368 TypePtr::PTR ptr = tj->ptr(); 1369 1370 // Known instance (scalarizable allocation) alias only with itself. 1371 bool is_known_inst = tj->isa_oopptr() != nullptr && 1372 tj->is_oopptr()->is_known_instance(); 1373 1374 // Process weird unsafe references. 1375 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1376 bool default_value_load = EnableValhalla && tj->is_instptr()->instance_klass() == ciEnv::current()->Class_klass(); 1377 assert(InlineUnsafeOps || StressReflectiveCode || default_value_load, "indeterminate pointers come only from unsafe ops"); 1378 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1379 tj = TypeOopPtr::BOTTOM; 1380 ptr = tj->ptr(); 1381 offset = tj->offset(); 1382 } 1383 1384 // Array pointers need some flattening 1385 const TypeAryPtr* ta = tj->isa_aryptr(); 1386 if (ta && ta->is_stable()) { 1387 // Erase stability property for alias analysis. 1388 tj = ta = ta->cast_to_stable(false); 1389 } 1390 if (ta && ta->is_not_flat()) { 1391 // Erase not flat property for alias analysis. 1392 tj = ta = ta->cast_to_not_flat(false); 1393 } 1394 if (ta && ta->is_not_null_free()) { 1395 // Erase not null free property for alias analysis. 1396 tj = ta = ta->cast_to_not_null_free(false); 1397 } 1398 1399 if( ta && is_known_inst ) { 1400 if ( offset != Type::OffsetBot && 1401 offset > arrayOopDesc::length_offset_in_bytes() ) { 1402 offset = Type::OffsetBot; // Flatten constant access into array body only 1403 tj = ta = ta-> 1404 remove_speculative()-> 1405 cast_to_ptr_type(ptr)-> 1406 with_offset(offset); 1407 } 1408 } else if (ta) { 1409 // For arrays indexed by constant indices, we flatten the alias 1410 // space to include all of the array body. Only the header, klass 1411 // and array length can be accessed un-aliased. 1412 // For flat inline type array, each field has its own slice so 1413 // we must include the field offset. 1414 if( offset != Type::OffsetBot ) { 1415 if( ta->const_oop() ) { // MethodData* or Method* 1416 offset = Type::OffsetBot; // Flatten constant access into array body 1417 tj = ta = ta-> 1418 remove_speculative()-> 1419 cast_to_ptr_type(ptr)-> 1420 cast_to_exactness(false)-> 1421 with_offset(offset); 1422 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1423 // range is OK as-is. 1424 tj = ta = TypeAryPtr::RANGE; 1425 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1426 tj = TypeInstPtr::KLASS; // all klass loads look alike 1427 ta = TypeAryPtr::RANGE; // generic ignored junk 1428 ptr = TypePtr::BotPTR; 1429 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1430 tj = TypeInstPtr::MARK; 1431 ta = TypeAryPtr::RANGE; // generic ignored junk 1432 ptr = TypePtr::BotPTR; 1433 } else { // Random constant offset into array body 1434 offset = Type::OffsetBot; // Flatten constant access into array body 1435 tj = ta = ta-> 1436 remove_speculative()-> 1437 cast_to_ptr_type(ptr)-> 1438 cast_to_exactness(false)-> 1439 with_offset(offset); 1440 } 1441 } 1442 // Arrays of fixed size alias with arrays of unknown size. 1443 if (ta->size() != TypeInt::POS) { 1444 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1445 tj = ta = ta-> 1446 remove_speculative()-> 1447 cast_to_ptr_type(ptr)-> 1448 with_ary(tary)-> 1449 cast_to_exactness(false); 1450 } 1451 // Arrays of known objects become arrays of unknown objects. 1452 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1453 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1454 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset()); 1455 } 1456 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1457 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1458 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), ta->field_offset()); 1459 } 1460 // Initially all flattened array accesses share a single slice 1461 if (ta->is_flat() && ta->elem() != TypeInstPtr::BOTTOM && _flat_accesses_share_alias) { 1462 const TypeAry* tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size(), /* stable= */ false, /* flat= */ true); 1463 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,Type::Offset(offset), Type::Offset(Type::OffsetBot)); 1464 } 1465 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1466 // cannot be distinguished by bytecode alone. 1467 if (ta->elem() == TypeInt::BOOL) { 1468 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1469 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1470 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1471 } 1472 // During the 2nd round of IterGVN, NotNull castings are removed. 1473 // Make sure the Bottom and NotNull variants alias the same. 1474 // Also, make sure exact and non-exact variants alias the same. 1475 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1476 tj = ta = ta-> 1477 remove_speculative()-> 1478 cast_to_ptr_type(TypePtr::BotPTR)-> 1479 cast_to_exactness(false)-> 1480 with_offset(offset); 1481 } 1482 } 1483 1484 // Oop pointers need some flattening 1485 const TypeInstPtr *to = tj->isa_instptr(); 1486 if (to && to != TypeOopPtr::BOTTOM) { 1487 ciInstanceKlass* ik = to->instance_klass(); 1488 if( ptr == TypePtr::Constant ) { 1489 if (ik != ciEnv::current()->Class_klass() || 1490 offset < ik->layout_helper_size_in_bytes()) { 1491 // No constant oop pointers (such as Strings); they alias with 1492 // unknown strings. 1493 assert(!is_known_inst, "not scalarizable allocation"); 1494 tj = to = to-> 1495 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1496 remove_speculative()-> 1497 cast_to_ptr_type(TypePtr::BotPTR)-> 1498 cast_to_exactness(false); 1499 } 1500 } else if( is_known_inst ) { 1501 tj = to; // Keep NotNull and klass_is_exact for instance type 1502 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1503 // During the 2nd round of IterGVN, NotNull castings are removed. 1504 // Make sure the Bottom and NotNull variants alias the same. 1505 // Also, make sure exact and non-exact variants alias the same. 1506 tj = to = to-> 1507 remove_speculative()-> 1508 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1509 cast_to_ptr_type(TypePtr::BotPTR)-> 1510 cast_to_exactness(false); 1511 } 1512 if (to->speculative() != nullptr) { 1513 tj = to = to->remove_speculative(); 1514 } 1515 // Canonicalize the holder of this field 1516 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1517 // First handle header references such as a LoadKlassNode, even if the 1518 // object's klass is unloaded at compile time (4965979). 1519 if (!is_known_inst) { // Do it only for non-instance types 1520 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, Type::Offset(offset)); 1521 } 1522 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1523 // Static fields are in the space above the normal instance 1524 // fields in the java.lang.Class instance. 1525 if (ik != ciEnv::current()->Class_klass()) { 1526 to = nullptr; 1527 tj = TypeOopPtr::BOTTOM; 1528 offset = tj->offset(); 1529 } 1530 } else { 1531 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1532 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1533 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1534 if( is_known_inst ) { 1535 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, Type::Offset(offset), to->instance_id()); 1536 } else { 1537 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, Type::Offset(offset)); 1538 } 1539 } 1540 } 1541 } 1542 1543 // Klass pointers to object array klasses need some flattening 1544 const TypeKlassPtr *tk = tj->isa_klassptr(); 1545 if( tk ) { 1546 // If we are referencing a field within a Klass, we need 1547 // to assume the worst case of an Object. Both exact and 1548 // inexact types must flatten to the same alias class so 1549 // use NotNull as the PTR. 1550 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1551 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1552 env()->Object_klass(), 1553 Type::Offset(offset)); 1554 } 1555 1556 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1557 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1558 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1559 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), Type::Offset(offset)); 1560 } else { 1561 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()); 1562 } 1563 } 1564 // Check for precise loads from the primary supertype array and force them 1565 // to the supertype cache alias index. Check for generic array loads from 1566 // the primary supertype array and also force them to the supertype cache 1567 // alias index. Since the same load can reach both, we need to merge 1568 // these 2 disparate memories into the same alias class. Since the 1569 // primary supertype array is read-only, there's no chance of confusion 1570 // where we bypass an array load and an array store. 1571 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1572 if (offset == Type::OffsetBot || 1573 (offset >= primary_supers_offset && 1574 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1575 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1576 offset = in_bytes(Klass::secondary_super_cache_offset()); 1577 tj = tk = tk->with_offset(offset); 1578 } 1579 } 1580 1581 // Flatten all Raw pointers together. 1582 if (tj->base() == Type::RawPtr) 1583 tj = TypeRawPtr::BOTTOM; 1584 1585 if (tj->base() == Type::AnyPtr) 1586 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1587 1588 offset = tj->offset(); 1589 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1590 1591 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1592 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1593 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1594 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1595 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1596 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1597 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1598 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1599 assert( tj->ptr() != TypePtr::TopPTR && 1600 tj->ptr() != TypePtr::AnyNull && 1601 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1602 // assert( tj->ptr() != TypePtr::Constant || 1603 // tj->base() == Type::RawPtr || 1604 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1605 1606 return tj; 1607 } 1608 1609 void Compile::AliasType::Init(int i, const TypePtr* at) { 1610 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1611 _index = i; 1612 _adr_type = at; 1613 _field = nullptr; 1614 _element = nullptr; 1615 _is_rewritable = true; // default 1616 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1617 if (atoop != nullptr && atoop->is_known_instance()) { 1618 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1619 _general_index = Compile::current()->get_alias_index(gt); 1620 } else { 1621 _general_index = 0; 1622 } 1623 } 1624 1625 BasicType Compile::AliasType::basic_type() const { 1626 if (element() != nullptr) { 1627 const Type* element = adr_type()->is_aryptr()->elem(); 1628 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1629 } if (field() != nullptr) { 1630 return field()->layout_type(); 1631 } else { 1632 return T_ILLEGAL; // unknown 1633 } 1634 } 1635 1636 //---------------------------------print_on------------------------------------ 1637 #ifndef PRODUCT 1638 void Compile::AliasType::print_on(outputStream* st) { 1639 if (index() < 10) 1640 st->print("@ <%d> ", index()); 1641 else st->print("@ <%d>", index()); 1642 st->print(is_rewritable() ? " " : " RO"); 1643 int offset = adr_type()->offset(); 1644 if (offset == Type::OffsetBot) 1645 st->print(" +any"); 1646 else st->print(" +%-3d", offset); 1647 st->print(" in "); 1648 adr_type()->dump_on(st); 1649 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1650 if (field() != nullptr && tjp) { 1651 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1652 tjp->offset() != field()->offset_in_bytes()) { 1653 st->print(" != "); 1654 field()->print(); 1655 st->print(" ***"); 1656 } 1657 } 1658 } 1659 1660 void print_alias_types() { 1661 Compile* C = Compile::current(); 1662 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1663 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1664 C->alias_type(idx)->print_on(tty); 1665 tty->cr(); 1666 } 1667 } 1668 #endif 1669 1670 1671 //----------------------------probe_alias_cache-------------------------------- 1672 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1673 intptr_t key = (intptr_t) adr_type; 1674 key ^= key >> logAliasCacheSize; 1675 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1676 } 1677 1678 1679 //-----------------------------grow_alias_types-------------------------------- 1680 void Compile::grow_alias_types() { 1681 const int old_ats = _max_alias_types; // how many before? 1682 const int new_ats = old_ats; // how many more? 1683 const int grow_ats = old_ats+new_ats; // how many now? 1684 _max_alias_types = grow_ats; 1685 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1686 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1687 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1688 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1689 } 1690 1691 1692 //--------------------------------find_alias_type------------------------------ 1693 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) { 1694 if (!do_aliasing()) { 1695 return alias_type(AliasIdxBot); 1696 } 1697 1698 AliasCacheEntry* ace = nullptr; 1699 if (!uncached) { 1700 ace = probe_alias_cache(adr_type); 1701 if (ace->_adr_type == adr_type) { 1702 return alias_type(ace->_index); 1703 } 1704 } 1705 1706 // Handle special cases. 1707 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1708 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1709 1710 // Do it the slow way. 1711 const TypePtr* flat = flatten_alias_type(adr_type); 1712 1713 #ifdef ASSERT 1714 { 1715 ResourceMark rm; 1716 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1717 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1718 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1719 Type::str(adr_type)); 1720 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1721 const TypeOopPtr* foop = flat->is_oopptr(); 1722 // Scalarizable allocations have exact klass always. 1723 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1724 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1725 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1726 Type::str(foop), Type::str(xoop)); 1727 } 1728 } 1729 #endif 1730 1731 int idx = AliasIdxTop; 1732 for (int i = 0; i < num_alias_types(); i++) { 1733 if (alias_type(i)->adr_type() == flat) { 1734 idx = i; 1735 break; 1736 } 1737 } 1738 1739 if (idx == AliasIdxTop) { 1740 if (no_create) return nullptr; 1741 // Grow the array if necessary. 1742 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1743 // Add a new alias type. 1744 idx = _num_alias_types++; 1745 _alias_types[idx]->Init(idx, flat); 1746 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1747 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1748 if (flat->isa_instptr()) { 1749 if (flat->offset() == java_lang_Class::klass_offset() 1750 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1751 alias_type(idx)->set_rewritable(false); 1752 } 1753 ciField* field = nullptr; 1754 if (flat->isa_aryptr()) { 1755 #ifdef ASSERT 1756 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1757 // (T_BYTE has the weakest alignment and size restrictions...) 1758 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1759 #endif 1760 const Type* elemtype = flat->is_aryptr()->elem(); 1761 if (flat->offset() == TypePtr::OffsetBot) { 1762 alias_type(idx)->set_element(elemtype); 1763 } 1764 int field_offset = flat->is_aryptr()->field_offset().get(); 1765 if (flat->is_flat() && 1766 field_offset != Type::OffsetBot) { 1767 ciInlineKlass* vk = elemtype->inline_klass(); 1768 field_offset += vk->first_field_offset(); 1769 field = vk->get_field_by_offset(field_offset, false); 1770 } 1771 } 1772 if (flat->isa_klassptr()) { 1773 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1774 alias_type(idx)->set_rewritable(false); 1775 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1776 alias_type(idx)->set_rewritable(false); 1777 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1778 alias_type(idx)->set_rewritable(false); 1779 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1780 alias_type(idx)->set_rewritable(false); 1781 if (flat->offset() == in_bytes(Klass::layout_helper_offset())) 1782 alias_type(idx)->set_rewritable(false); 1783 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1784 alias_type(idx)->set_rewritable(false); 1785 } 1786 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1787 // but the base pointer type is not distinctive enough to identify 1788 // references into JavaThread.) 1789 1790 // Check for final fields. 1791 const TypeInstPtr* tinst = flat->isa_instptr(); 1792 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1793 if (tinst->const_oop() != nullptr && 1794 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1795 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1796 // static field 1797 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1798 field = k->get_field_by_offset(tinst->offset(), true); 1799 } else if (tinst->is_inlinetypeptr()) { 1800 // Inline type field 1801 ciInlineKlass* vk = tinst->inline_klass(); 1802 field = vk->get_field_by_offset(tinst->offset(), false); 1803 } else { 1804 ciInstanceKlass *k = tinst->instance_klass(); 1805 field = k->get_field_by_offset(tinst->offset(), false); 1806 } 1807 } 1808 assert(field == nullptr || 1809 original_field == nullptr || 1810 (field->holder() == original_field->holder() && 1811 field->offset_in_bytes() == original_field->offset_in_bytes() && 1812 field->is_static() == original_field->is_static()), "wrong field?"); 1813 // Set field() and is_rewritable() attributes. 1814 if (field != nullptr) { 1815 alias_type(idx)->set_field(field); 1816 if (flat->isa_aryptr()) { 1817 // Fields of flat arrays are rewritable although they are declared final 1818 assert(flat->is_flat(), "must be a flat array"); 1819 alias_type(idx)->set_rewritable(true); 1820 } 1821 } 1822 } 1823 1824 // Fill the cache for next time. 1825 if (!uncached) { 1826 ace->_adr_type = adr_type; 1827 ace->_index = idx; 1828 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1829 1830 // Might as well try to fill the cache for the flattened version, too. 1831 AliasCacheEntry* face = probe_alias_cache(flat); 1832 if (face->_adr_type == nullptr) { 1833 face->_adr_type = flat; 1834 face->_index = idx; 1835 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1836 } 1837 } 1838 1839 return alias_type(idx); 1840 } 1841 1842 1843 Compile::AliasType* Compile::alias_type(ciField* field) { 1844 const TypeOopPtr* t; 1845 if (field->is_static()) 1846 t = TypeInstPtr::make(field->holder()->java_mirror()); 1847 else 1848 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1849 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1850 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1851 return atp; 1852 } 1853 1854 1855 //------------------------------have_alias_type-------------------------------- 1856 bool Compile::have_alias_type(const TypePtr* adr_type) { 1857 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1858 if (ace->_adr_type == adr_type) { 1859 return true; 1860 } 1861 1862 // Handle special cases. 1863 if (adr_type == nullptr) return true; 1864 if (adr_type == TypePtr::BOTTOM) return true; 1865 1866 return find_alias_type(adr_type, true, nullptr) != nullptr; 1867 } 1868 1869 //-----------------------------must_alias-------------------------------------- 1870 // True if all values of the given address type are in the given alias category. 1871 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1872 if (alias_idx == AliasIdxBot) return true; // the universal category 1873 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1874 if (alias_idx == AliasIdxTop) return false; // the empty category 1875 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1876 1877 // the only remaining possible overlap is identity 1878 int adr_idx = get_alias_index(adr_type); 1879 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1880 assert(adr_idx == alias_idx || 1881 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1882 && adr_type != TypeOopPtr::BOTTOM), 1883 "should not be testing for overlap with an unsafe pointer"); 1884 return adr_idx == alias_idx; 1885 } 1886 1887 //------------------------------can_alias-------------------------------------- 1888 // True if any values of the given address type are in the given alias category. 1889 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1890 if (alias_idx == AliasIdxTop) return false; // the empty category 1891 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1892 // Known instance doesn't alias with bottom memory 1893 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1894 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1895 1896 // the only remaining possible overlap is identity 1897 int adr_idx = get_alias_index(adr_type); 1898 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1899 return adr_idx == alias_idx; 1900 } 1901 1902 // Mark all ParsePredicateNodes as useless. They will later be removed from the graph in IGVN together with their 1903 // uncommon traps if no Runtime Predicates were created from the Parse Predicates. 1904 void Compile::mark_parse_predicate_nodes_useless(PhaseIterGVN& igvn) { 1905 if (parse_predicate_count() == 0) { 1906 return; 1907 } 1908 for (int i = 0; i < parse_predicate_count(); i++) { 1909 ParsePredicateNode* parse_predicate = _parse_predicates.at(i); 1910 parse_predicate->mark_useless(); 1911 igvn._worklist.push(parse_predicate); 1912 } 1913 _parse_predicates.clear(); 1914 } 1915 1916 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1917 if (!n->for_post_loop_opts_igvn()) { 1918 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1919 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1920 _for_post_loop_igvn.append(n); 1921 } 1922 } 1923 1924 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1925 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1926 _for_post_loop_igvn.remove(n); 1927 } 1928 1929 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1930 // Verify that all previous optimizations produced a valid graph 1931 // at least to this point, even if no loop optimizations were done. 1932 PhaseIdealLoop::verify(igvn); 1933 1934 C->set_post_loop_opts_phase(); // no more loop opts allowed 1935 1936 assert(!C->major_progress(), "not cleared"); 1937 1938 if (_for_post_loop_igvn.length() > 0) { 1939 while (_for_post_loop_igvn.length() > 0) { 1940 Node* n = _for_post_loop_igvn.pop(); 1941 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1942 igvn._worklist.push(n); 1943 } 1944 igvn.optimize(); 1945 if (failing()) return; 1946 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1947 assert(C->parse_predicate_count() == 0, "all parse predicates should have been removed now"); 1948 1949 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1950 if (C->major_progress()) { 1951 C->clear_major_progress(); // ensure that major progress is now clear 1952 } 1953 } 1954 } 1955 1956 void Compile::add_inline_type(Node* n) { 1957 assert(n->is_InlineType(), "unexpected node"); 1958 _inline_type_nodes.push(n); 1959 } 1960 1961 void Compile::remove_inline_type(Node* n) { 1962 assert(n->is_InlineType(), "unexpected node"); 1963 if (_inline_type_nodes.contains(n)) { 1964 _inline_type_nodes.remove(n); 1965 } 1966 } 1967 1968 // Does the return value keep otherwise useless inline type allocations alive? 1969 static bool return_val_keeps_allocations_alive(Node* ret_val) { 1970 ResourceMark rm; 1971 Unique_Node_List wq; 1972 wq.push(ret_val); 1973 bool some_allocations = false; 1974 for (uint i = 0; i < wq.size(); i++) { 1975 Node* n = wq.at(i); 1976 if (n->outcnt() > 1) { 1977 // Some other use for the allocation 1978 return false; 1979 } else if (n->is_InlineType()) { 1980 wq.push(n->in(1)); 1981 } else if (n->is_Phi()) { 1982 for (uint j = 1; j < n->req(); j++) { 1983 wq.push(n->in(j)); 1984 } 1985 } else if (n->is_CheckCastPP() && 1986 n->in(1)->is_Proj() && 1987 n->in(1)->in(0)->is_Allocate()) { 1988 some_allocations = true; 1989 } else if (n->is_CheckCastPP()) { 1990 wq.push(n->in(1)); 1991 } 1992 } 1993 return some_allocations; 1994 } 1995 1996 void Compile::process_inline_types(PhaseIterGVN &igvn, bool remove) { 1997 // Make sure that the return value does not keep an otherwise unused allocation alive 1998 if (tf()->returns_inline_type_as_fields()) { 1999 Node* ret = nullptr; 2000 for (uint i = 1; i < root()->req(); i++) { 2001 Node* in = root()->in(i); 2002 if (in->Opcode() == Op_Return) { 2003 assert(ret == nullptr, "only one return"); 2004 ret = in; 2005 } 2006 } 2007 if (ret != nullptr) { 2008 Node* ret_val = ret->in(TypeFunc::Parms); 2009 if (igvn.type(ret_val)->isa_oopptr() && 2010 return_val_keeps_allocations_alive(ret_val)) { 2011 igvn.replace_input_of(ret, TypeFunc::Parms, InlineTypeNode::tagged_klass(igvn.type(ret_val)->inline_klass(), igvn)); 2012 assert(ret_val->outcnt() == 0, "should be dead now"); 2013 igvn.remove_dead_node(ret_val); 2014 } 2015 } 2016 } 2017 if (_inline_type_nodes.length() == 0) { 2018 return; 2019 } 2020 // Scalarize inline types in safepoint debug info. 2021 // Delay this until all inlining is over to avoid getting inconsistent debug info. 2022 set_scalarize_in_safepoints(true); 2023 for (int i = _inline_type_nodes.length()-1; i >= 0; i--) { 2024 _inline_type_nodes.at(i)->as_InlineType()->make_scalar_in_safepoints(&igvn); 2025 } 2026 if (remove) { 2027 // Remove inline type nodes by replacing them with their oop input 2028 while (_inline_type_nodes.length() > 0) { 2029 InlineTypeNode* vt = _inline_type_nodes.pop()->as_InlineType(); 2030 if (vt->outcnt() == 0) { 2031 igvn.remove_dead_node(vt); 2032 continue; 2033 } 2034 for (DUIterator i = vt->outs(); vt->has_out(i); i++) { 2035 DEBUG_ONLY(bool must_be_buffered = false); 2036 Node* u = vt->out(i); 2037 // Check if any users are blackholes. If so, rewrite them to use either the 2038 // allocated buffer, or individual components, instead of the inline type node 2039 // that goes away. 2040 if (u->is_Blackhole()) { 2041 BlackholeNode* bh = u->as_Blackhole(); 2042 2043 // Unlink the old input 2044 int idx = bh->find_edge(vt); 2045 assert(idx != -1, "The edge should be there"); 2046 bh->del_req(idx); 2047 --i; 2048 2049 if (vt->is_allocated(&igvn)) { 2050 // Already has the allocated instance, blackhole that 2051 bh->add_req(vt->get_oop()); 2052 } else { 2053 // Not allocated yet, blackhole the components 2054 for (uint c = 0; c < vt->field_count(); c++) { 2055 bh->add_req(vt->field_value(c)); 2056 } 2057 } 2058 2059 // Node modified, record for IGVN 2060 igvn.record_for_igvn(bh); 2061 } 2062 #ifdef ASSERT 2063 // Verify that inline type is buffered when replacing by oop 2064 else if (u->is_InlineType()) { 2065 // InlineType uses don't need buffering because they are about to be replaced as well 2066 } else if (u->is_Phi()) { 2067 // TODO 8302217 Remove this once InlineTypeNodes are reliably pushed through 2068 } else { 2069 must_be_buffered = true; 2070 } 2071 if (must_be_buffered && !vt->is_allocated(&igvn)) { 2072 vt->dump(0); 2073 u->dump(0); 2074 assert(false, "Should have been buffered"); 2075 } 2076 #endif 2077 } 2078 igvn.replace_node(vt, vt->get_oop()); 2079 } 2080 } 2081 igvn.optimize(); 2082 } 2083 2084 void Compile::adjust_flat_array_access_aliases(PhaseIterGVN& igvn) { 2085 if (!_has_flat_accesses) { 2086 return; 2087 } 2088 // Initially, all flat array accesses share the same slice to 2089 // keep dependencies with Object[] array accesses (that could be 2090 // to a flat array) correct. We're done with parsing so we 2091 // now know all flat array accesses in this compile 2092 // unit. Let's move flat array accesses to their own slice, 2093 // one per element field. This should help memory access 2094 // optimizations. 2095 ResourceMark rm; 2096 Unique_Node_List wq; 2097 wq.push(root()); 2098 2099 Node_List mergememnodes; 2100 Node_List memnodes; 2101 2102 // Alias index currently shared by all flat memory accesses 2103 int index = get_alias_index(TypeAryPtr::INLINES); 2104 2105 // Find MergeMem nodes and flat array accesses 2106 for (uint i = 0; i < wq.size(); i++) { 2107 Node* n = wq.at(i); 2108 if (n->is_Mem()) { 2109 const TypePtr* adr_type = nullptr; 2110 if (n->Opcode() == Op_StoreCM) { 2111 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type())); 2112 } else { 2113 adr_type = get_adr_type(get_alias_index(n->adr_type())); 2114 } 2115 if (adr_type == TypeAryPtr::INLINES) { 2116 memnodes.push(n); 2117 } 2118 } else if (n->is_MergeMem()) { 2119 MergeMemNode* mm = n->as_MergeMem(); 2120 if (mm->memory_at(index) != mm->base_memory()) { 2121 mergememnodes.push(n); 2122 } 2123 } 2124 for (uint j = 0; j < n->req(); j++) { 2125 Node* m = n->in(j); 2126 if (m != nullptr) { 2127 wq.push(m); 2128 } 2129 } 2130 } 2131 2132 if (memnodes.size() > 0) { 2133 _flat_accesses_share_alias = false; 2134 2135 // We are going to change the slice for the flat array 2136 // accesses so we need to clear the cache entries that refer to 2137 // them. 2138 for (uint i = 0; i < AliasCacheSize; i++) { 2139 AliasCacheEntry* ace = &_alias_cache[i]; 2140 if (ace->_adr_type != nullptr && 2141 ace->_adr_type->is_flat()) { 2142 ace->_adr_type = nullptr; 2143 ace->_index = (i != 0) ? 0 : AliasIdxTop; // Make sure the nullptr adr_type resolves to AliasIdxTop 2144 } 2145 } 2146 2147 // Find what aliases we are going to add 2148 int start_alias = num_alias_types()-1; 2149 int stop_alias = 0; 2150 2151 for (uint i = 0; i < memnodes.size(); i++) { 2152 Node* m = memnodes.at(i); 2153 const TypePtr* adr_type = nullptr; 2154 if (m->Opcode() == Op_StoreCM) { 2155 adr_type = m->in(MemNode::OopStore)->adr_type(); 2156 if (adr_type != TypeAryPtr::INLINES) { 2157 // store was optimized out and we lost track of the adr_type 2158 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address), 2159 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore), 2160 get_alias_index(adr_type)); 2161 igvn.register_new_node_with_optimizer(clone); 2162 igvn.replace_node(m, clone); 2163 } 2164 } else { 2165 adr_type = m->adr_type(); 2166 #ifdef ASSERT 2167 m->as_Mem()->set_adr_type(adr_type); 2168 #endif 2169 } 2170 int idx = get_alias_index(adr_type); 2171 start_alias = MIN2(start_alias, idx); 2172 stop_alias = MAX2(stop_alias, idx); 2173 } 2174 2175 assert(stop_alias >= start_alias, "should have expanded aliases"); 2176 2177 Node_Stack stack(0); 2178 #ifdef ASSERT 2179 VectorSet seen(Thread::current()->resource_area()); 2180 #endif 2181 // Now let's fix the memory graph so each flat array access 2182 // is moved to the right slice. Start from the MergeMem nodes. 2183 uint last = unique(); 2184 for (uint i = 0; i < mergememnodes.size(); i++) { 2185 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem(); 2186 Node* n = current->memory_at(index); 2187 MergeMemNode* mm = nullptr; 2188 do { 2189 // Follow memory edges through memory accesses, phis and 2190 // narrow membars and push nodes on the stack. Once we hit 2191 // bottom memory, we pop element off the stack one at a 2192 // time, in reverse order, and move them to the right slice 2193 // by changing their memory edges. 2194 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::INLINES) { 2195 assert(!seen.test_set(n->_idx), ""); 2196 // Uses (a load for instance) will need to be moved to the 2197 // right slice as well and will get a new memory state 2198 // that we don't know yet. The use could also be the 2199 // backedge of a loop. We put a place holder node between 2200 // the memory node and its uses. We replace that place 2201 // holder with the correct memory state once we know it, 2202 // i.e. when nodes are popped off the stack. Using the 2203 // place holder make the logic work in the presence of 2204 // loops. 2205 if (n->outcnt() > 1) { 2206 Node* place_holder = nullptr; 2207 assert(!n->has_out_with(Op_Node), ""); 2208 for (DUIterator k = n->outs(); n->has_out(k); k++) { 2209 Node* u = n->out(k); 2210 if (u != current && u->_idx < last) { 2211 bool success = false; 2212 for (uint l = 0; l < u->req(); l++) { 2213 if (!stack.is_empty() && u == stack.node() && l == stack.index()) { 2214 continue; 2215 } 2216 Node* in = u->in(l); 2217 if (in == n) { 2218 if (place_holder == nullptr) { 2219 place_holder = new Node(1); 2220 place_holder->init_req(0, n); 2221 } 2222 igvn.replace_input_of(u, l, place_holder); 2223 success = true; 2224 } 2225 } 2226 if (success) { 2227 --k; 2228 } 2229 } 2230 } 2231 } 2232 if (n->is_Phi()) { 2233 stack.push(n, 1); 2234 n = n->in(1); 2235 } else if (n->is_Mem()) { 2236 stack.push(n, n->req()); 2237 n = n->in(MemNode::Memory); 2238 } else { 2239 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, ""); 2240 stack.push(n, n->req()); 2241 n = n->in(0)->in(TypeFunc::Memory); 2242 } 2243 } else { 2244 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), ""); 2245 // Build a new MergeMem node to carry the new memory state 2246 // as we build it. IGVN should fold extraneous MergeMem 2247 // nodes. 2248 mm = MergeMemNode::make(n); 2249 igvn.register_new_node_with_optimizer(mm); 2250 while (stack.size() > 0) { 2251 Node* m = stack.node(); 2252 uint idx = stack.index(); 2253 if (m->is_Mem()) { 2254 // Move memory node to its new slice 2255 const TypePtr* adr_type = m->adr_type(); 2256 int alias = get_alias_index(adr_type); 2257 Node* prev = mm->memory_at(alias); 2258 igvn.replace_input_of(m, MemNode::Memory, prev); 2259 mm->set_memory_at(alias, m); 2260 } else if (m->is_Phi()) { 2261 // We need as many new phis as there are new aliases 2262 igvn.replace_input_of(m, idx, mm); 2263 if (idx == m->req()-1) { 2264 Node* r = m->in(0); 2265 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2266 const TypePtr* adr_type = get_adr_type(j); 2267 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) { 2268 continue; 2269 } 2270 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j)); 2271 igvn.register_new_node_with_optimizer(phi); 2272 for (uint k = 1; k < m->req(); k++) { 2273 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j)); 2274 } 2275 mm->set_memory_at(j, phi); 2276 } 2277 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2278 igvn.register_new_node_with_optimizer(base_phi); 2279 for (uint k = 1; k < m->req(); k++) { 2280 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory()); 2281 } 2282 mm->set_base_memory(base_phi); 2283 } 2284 } else { 2285 // This is a MemBarCPUOrder node from 2286 // Parse::array_load()/Parse::array_store(), in the 2287 // branch that handles flat arrays hidden under 2288 // an Object[] array. We also need one new membar per 2289 // new alias to keep the unknown access that the 2290 // membars protect properly ordered with accesses to 2291 // known flat array. 2292 assert(m->is_Proj(), "projection expected"); 2293 Node* ctrl = m->in(0)->in(TypeFunc::Control); 2294 igvn.replace_input_of(m->in(0), TypeFunc::Control, top()); 2295 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2296 const TypePtr* adr_type = get_adr_type(j); 2297 if (!adr_type->isa_aryptr() || !adr_type->is_flat() || j == (uint)index) { 2298 continue; 2299 } 2300 MemBarNode* mb = new MemBarCPUOrderNode(this, j, nullptr); 2301 igvn.register_new_node_with_optimizer(mb); 2302 Node* mem = mm->memory_at(j); 2303 mb->init_req(TypeFunc::Control, ctrl); 2304 mb->init_req(TypeFunc::Memory, mem); 2305 ctrl = new ProjNode(mb, TypeFunc::Control); 2306 igvn.register_new_node_with_optimizer(ctrl); 2307 mem = new ProjNode(mb, TypeFunc::Memory); 2308 igvn.register_new_node_with_optimizer(mem); 2309 mm->set_memory_at(j, mem); 2310 } 2311 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl); 2312 } 2313 if (idx < m->req()-1) { 2314 idx += 1; 2315 stack.set_index(idx); 2316 n = m->in(idx); 2317 break; 2318 } 2319 // Take care of place holder nodes 2320 if (m->has_out_with(Op_Node)) { 2321 Node* place_holder = m->find_out_with(Op_Node); 2322 if (place_holder != nullptr) { 2323 Node* mm_clone = mm->clone(); 2324 igvn.register_new_node_with_optimizer(mm_clone); 2325 Node* hook = new Node(1); 2326 hook->init_req(0, mm); 2327 igvn.replace_node(place_holder, mm_clone); 2328 hook->destruct(&igvn); 2329 } 2330 assert(!m->has_out_with(Op_Node), "place holder should be gone now"); 2331 } 2332 stack.pop(); 2333 } 2334 } 2335 } while(stack.size() > 0); 2336 // Fix the memory state at the MergeMem we started from 2337 igvn.rehash_node_delayed(current); 2338 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2339 const TypePtr* adr_type = get_adr_type(j); 2340 if (!adr_type->isa_aryptr() || !adr_type->is_flat()) { 2341 continue; 2342 } 2343 current->set_memory_at(j, mm); 2344 } 2345 current->set_memory_at(index, current->base_memory()); 2346 } 2347 igvn.optimize(); 2348 } 2349 print_method(PHASE_SPLIT_INLINES_ARRAY, 2); 2350 #ifdef ASSERT 2351 if (!_flat_accesses_share_alias) { 2352 wq.clear(); 2353 wq.push(root()); 2354 for (uint i = 0; i < wq.size(); i++) { 2355 Node* n = wq.at(i); 2356 assert(n->adr_type() != TypeAryPtr::INLINES, "should have been removed from the graph"); 2357 for (uint j = 0; j < n->req(); j++) { 2358 Node* m = n->in(j); 2359 if (m != nullptr) { 2360 wq.push(m); 2361 } 2362 } 2363 } 2364 } 2365 #endif 2366 } 2367 2368 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 2369 if (OptimizeUnstableIf) { 2370 _unstable_if_traps.append(trap); 2371 } 2372 } 2373 2374 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 2375 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 2376 UnstableIfTrap* trap = _unstable_if_traps.at(i); 2377 Node* n = trap->uncommon_trap(); 2378 if (!useful.member(n)) { 2379 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 2380 } 2381 } 2382 } 2383 2384 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 2385 // or fold-compares case. Return true if succeed or not found. 2386 // 2387 // In rare cases, the found trap has been processed. It is too late to delete it. Return 2388 // false and ask fold-compares to yield. 2389 // 2390 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 2391 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 2392 // when deoptimization does happen. 2393 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 2394 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 2395 UnstableIfTrap* trap = _unstable_if_traps.at(i); 2396 if (trap->uncommon_trap() == unc) { 2397 if (yield && trap->modified()) { 2398 return false; 2399 } 2400 _unstable_if_traps.delete_at(i); 2401 break; 2402 } 2403 } 2404 return true; 2405 } 2406 2407 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 2408 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 2409 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 2410 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 2411 UnstableIfTrap* trap = _unstable_if_traps.at(i); 2412 CallStaticJavaNode* unc = trap->uncommon_trap(); 2413 int next_bci = trap->next_bci(); 2414 bool modified = trap->modified(); 2415 2416 if (next_bci != -1 && !modified) { 2417 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 2418 JVMState* jvms = unc->jvms(); 2419 ciMethod* method = jvms->method(); 2420 ciBytecodeStream iter(method); 2421 2422 iter.force_bci(jvms->bci()); 2423 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 2424 Bytecodes::Code c = iter.cur_bc(); 2425 Node* lhs = nullptr; 2426 Node* rhs = nullptr; 2427 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 2428 lhs = unc->peek_operand(0); 2429 rhs = unc->peek_operand(1); 2430 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 2431 lhs = unc->peek_operand(0); 2432 } 2433 2434 ResourceMark rm; 2435 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 2436 assert(live_locals.is_valid(), "broken liveness info"); 2437 int len = (int)live_locals.size(); 2438 2439 for (int i = 0; i < len; i++) { 2440 Node* local = unc->local(jvms, i); 2441 // kill local using the liveness of next_bci. 2442 // give up when the local looks like an operand to secure reexecution. 2443 if (!live_locals.at(i) && !local->is_top() && local != lhs && local != rhs) { 2444 uint idx = jvms->locoff() + i; 2445 #ifdef ASSERT 2446 if (PrintOpto && Verbose) { 2447 tty->print("[unstable_if] kill local#%d: ", idx); 2448 local->dump(); 2449 tty->cr(); 2450 } 2451 #endif 2452 igvn.replace_input_of(unc, idx, top()); 2453 modified = true; 2454 } 2455 } 2456 } 2457 2458 // keep the modified trap for late query 2459 if (modified) { 2460 trap->set_modified(); 2461 } else { 2462 _unstable_if_traps.delete_at(i); 2463 } 2464 } 2465 igvn.optimize(); 2466 } 2467 2468 // StringOpts and late inlining of string methods 2469 void Compile::inline_string_calls(bool parse_time) { 2470 { 2471 // remove useless nodes to make the usage analysis simpler 2472 ResourceMark rm; 2473 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2474 } 2475 2476 { 2477 ResourceMark rm; 2478 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2479 PhaseStringOpts pso(initial_gvn()); 2480 print_method(PHASE_AFTER_STRINGOPTS, 3); 2481 } 2482 2483 // now inline anything that we skipped the first time around 2484 if (!parse_time) { 2485 _late_inlines_pos = _late_inlines.length(); 2486 } 2487 2488 while (_string_late_inlines.length() > 0) { 2489 CallGenerator* cg = _string_late_inlines.pop(); 2490 cg->do_late_inline(); 2491 if (failing()) return; 2492 } 2493 _string_late_inlines.trunc_to(0); 2494 } 2495 2496 // Late inlining of boxing methods 2497 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2498 if (_boxing_late_inlines.length() > 0) { 2499 assert(has_boxed_value(), "inconsistent"); 2500 2501 set_inlining_incrementally(true); 2502 2503 igvn_worklist()->ensure_empty(); // should be done with igvn 2504 2505 _late_inlines_pos = _late_inlines.length(); 2506 2507 while (_boxing_late_inlines.length() > 0) { 2508 CallGenerator* cg = _boxing_late_inlines.pop(); 2509 cg->do_late_inline(); 2510 if (failing()) return; 2511 } 2512 _boxing_late_inlines.trunc_to(0); 2513 2514 inline_incrementally_cleanup(igvn); 2515 2516 set_inlining_incrementally(false); 2517 } 2518 } 2519 2520 bool Compile::inline_incrementally_one() { 2521 assert(IncrementalInline, "incremental inlining should be on"); 2522 2523 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2524 2525 set_inlining_progress(false); 2526 set_do_cleanup(false); 2527 2528 for (int i = 0; i < _late_inlines.length(); i++) { 2529 _late_inlines_pos = i+1; 2530 CallGenerator* cg = _late_inlines.at(i); 2531 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2532 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2533 cg->do_late_inline(); 2534 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2535 if (failing()) { 2536 return false; 2537 } else if (inlining_progress()) { 2538 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2539 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2540 break; // process one call site at a time 2541 } 2542 } else { 2543 // Ignore late inline direct calls when inlining is not allowed. 2544 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2545 } 2546 } 2547 // Remove processed elements. 2548 _late_inlines.remove_till(_late_inlines_pos); 2549 _late_inlines_pos = 0; 2550 2551 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2552 2553 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2554 2555 set_inlining_progress(false); 2556 set_do_cleanup(false); 2557 2558 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2559 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2560 } 2561 2562 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2563 { 2564 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2565 ResourceMark rm; 2566 PhaseRemoveUseless pru(initial_gvn(), *igvn_worklist()); 2567 } 2568 { 2569 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2570 igvn.reset_from_gvn(initial_gvn()); 2571 igvn.optimize(); 2572 if (failing()) return; 2573 } 2574 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2575 } 2576 2577 // Perform incremental inlining until bound on number of live nodes is reached 2578 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2579 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2580 2581 set_inlining_incrementally(true); 2582 uint low_live_nodes = 0; 2583 2584 while (_late_inlines.length() > 0) { 2585 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2586 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2587 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2588 // PhaseIdealLoop is expensive so we only try it once we are 2589 // out of live nodes and we only try it again if the previous 2590 // helped got the number of nodes down significantly 2591 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2592 if (failing()) return; 2593 low_live_nodes = live_nodes(); 2594 _major_progress = true; 2595 } 2596 2597 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2598 bool do_print_inlining = print_inlining() || print_intrinsics(); 2599 if (do_print_inlining || log() != nullptr) { 2600 // Print inlining message for candidates that we couldn't inline for lack of space. 2601 for (int i = 0; i < _late_inlines.length(); i++) { 2602 CallGenerator* cg = _late_inlines.at(i); 2603 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2604 if (do_print_inlining) { 2605 cg->print_inlining_late(InliningResult::FAILURE, msg); 2606 } 2607 log_late_inline_failure(cg, msg); 2608 } 2609 } 2610 break; // finish 2611 } 2612 } 2613 2614 igvn_worklist()->ensure_empty(); // should be done with igvn 2615 2616 while (inline_incrementally_one()) { 2617 assert(!failing(), "inconsistent"); 2618 } 2619 if (failing()) return; 2620 2621 inline_incrementally_cleanup(igvn); 2622 2623 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2624 2625 if (failing()) return; 2626 2627 if (_late_inlines.length() == 0) { 2628 break; // no more progress 2629 } 2630 } 2631 2632 igvn_worklist()->ensure_empty(); // should be done with igvn 2633 2634 if (_string_late_inlines.length() > 0) { 2635 assert(has_stringbuilder(), "inconsistent"); 2636 2637 inline_string_calls(false); 2638 2639 if (failing()) return; 2640 2641 inline_incrementally_cleanup(igvn); 2642 } 2643 2644 set_inlining_incrementally(false); 2645 } 2646 2647 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2648 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2649 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2650 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2651 // as if "inlining_incrementally() == true" were set. 2652 assert(inlining_incrementally() == false, "not allowed"); 2653 #ifdef ASSERT 2654 Unique_Node_List* modified_nodes = _modified_nodes; 2655 _modified_nodes = nullptr; 2656 #endif 2657 assert(_late_inlines.length() > 0, "sanity"); 2658 2659 while (_late_inlines.length() > 0) { 2660 igvn_worklist()->ensure_empty(); // should be done with igvn 2661 2662 while (inline_incrementally_one()) { 2663 assert(!failing(), "inconsistent"); 2664 } 2665 if (failing()) return; 2666 2667 inline_incrementally_cleanup(igvn); 2668 } 2669 DEBUG_ONLY( _modified_nodes = modified_nodes; ) 2670 } 2671 2672 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2673 if (_loop_opts_cnt > 0) { 2674 while (major_progress() && (_loop_opts_cnt > 0)) { 2675 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2676 PhaseIdealLoop::optimize(igvn, mode); 2677 _loop_opts_cnt--; 2678 if (failing()) return false; 2679 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2680 } 2681 } 2682 return true; 2683 } 2684 2685 // Remove edges from "root" to each SafePoint at a backward branch. 2686 // They were inserted during parsing (see add_safepoint()) to make 2687 // infinite loops without calls or exceptions visible to root, i.e., 2688 // useful. 2689 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2690 Node *r = root(); 2691 if (r != nullptr) { 2692 for (uint i = r->req(); i < r->len(); ++i) { 2693 Node *n = r->in(i); 2694 if (n != nullptr && n->is_SafePoint()) { 2695 r->rm_prec(i); 2696 if (n->outcnt() == 0) { 2697 igvn.remove_dead_node(n); 2698 } 2699 --i; 2700 } 2701 } 2702 // Parsing may have added top inputs to the root node (Path 2703 // leading to the Halt node proven dead). Make sure we get a 2704 // chance to clean them up. 2705 igvn._worklist.push(r); 2706 igvn.optimize(); 2707 } 2708 } 2709 2710 //------------------------------Optimize--------------------------------------- 2711 // Given a graph, optimize it. 2712 void Compile::Optimize() { 2713 TracePhase tp("optimizer", &timers[_t_optimizer]); 2714 2715 #ifndef PRODUCT 2716 if (env()->break_at_compile()) { 2717 BREAKPOINT; 2718 } 2719 2720 #endif 2721 2722 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2723 #ifdef ASSERT 2724 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2725 #endif 2726 2727 ResourceMark rm; 2728 2729 print_inlining_reinit(); 2730 2731 NOT_PRODUCT( verify_graph_edges(); ) 2732 2733 print_method(PHASE_AFTER_PARSING, 1); 2734 2735 { 2736 // Iterative Global Value Numbering, including ideal transforms 2737 // Initialize IterGVN with types and values from parse-time GVN 2738 PhaseIterGVN igvn(initial_gvn()); 2739 #ifdef ASSERT 2740 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2741 #endif 2742 { 2743 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2744 igvn.optimize(); 2745 } 2746 2747 if (failing()) return; 2748 2749 print_method(PHASE_ITER_GVN1, 2); 2750 2751 process_for_unstable_if_traps(igvn); 2752 2753 if (failing()) return; 2754 2755 inline_incrementally(igvn); 2756 2757 print_method(PHASE_INCREMENTAL_INLINE, 2); 2758 2759 if (failing()) return; 2760 2761 if (eliminate_boxing()) { 2762 // Inline valueOf() methods now. 2763 inline_boxing_calls(igvn); 2764 2765 if (failing()) return; 2766 2767 if (AlwaysIncrementalInline || StressIncrementalInlining) { 2768 inline_incrementally(igvn); 2769 } 2770 2771 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2772 2773 if (failing()) return; 2774 } 2775 2776 // Remove the speculative part of types and clean up the graph from 2777 // the extra CastPP nodes whose only purpose is to carry them. Do 2778 // that early so that optimizations are not disrupted by the extra 2779 // CastPP nodes. 2780 remove_speculative_types(igvn); 2781 2782 if (failing()) return; 2783 2784 // No more new expensive nodes will be added to the list from here 2785 // so keep only the actual candidates for optimizations. 2786 cleanup_expensive_nodes(igvn); 2787 2788 if (failing()) return; 2789 2790 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2791 if (EnableVectorSupport && has_vbox_nodes()) { 2792 TracePhase tp("", &timers[_t_vector]); 2793 PhaseVector pv(igvn); 2794 pv.optimize_vector_boxes(); 2795 if (failing()) return; 2796 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2797 } 2798 assert(!has_vbox_nodes(), "sanity"); 2799 2800 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2801 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2802 igvn_worklist()->ensure_empty(); // should be done with igvn 2803 { 2804 ResourceMark rm; 2805 PhaseRenumberLive prl(initial_gvn(), *igvn_worklist()); 2806 } 2807 igvn.reset_from_gvn(initial_gvn()); 2808 igvn.optimize(); 2809 if (failing()) return; 2810 } 2811 2812 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2813 // safepoints 2814 remove_root_to_sfpts_edges(igvn); 2815 2816 // Process inline type nodes now that all inlining is over 2817 process_inline_types(igvn); 2818 2819 adjust_flat_array_access_aliases(igvn); 2820 2821 if (failing()) return; 2822 2823 // Perform escape analysis 2824 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2825 if (has_loops()) { 2826 // Cleanup graph (remove dead nodes). 2827 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2828 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2829 if (failing()) return; 2830 } 2831 bool progress; 2832 print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2833 do { 2834 ConnectionGraph::do_analysis(this, &igvn); 2835 2836 if (failing()) return; 2837 2838 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2839 2840 // Optimize out fields loads from scalar replaceable allocations. 2841 igvn.optimize(); 2842 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2843 2844 if (failing()) return; 2845 2846 if (congraph() != nullptr && macro_count() > 0) { 2847 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2848 PhaseMacroExpand mexp(igvn); 2849 mexp.eliminate_macro_nodes(); 2850 if (failing()) return; 2851 2852 igvn.set_delay_transform(false); 2853 igvn.optimize(); 2854 if (failing()) return; 2855 2856 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2857 } 2858 2859 ConnectionGraph::verify_ram_nodes(this, root()); 2860 if (failing()) return; 2861 2862 progress = do_iterative_escape_analysis() && 2863 (macro_count() < mcount) && 2864 ConnectionGraph::has_candidates(this); 2865 // Try again if candidates exist and made progress 2866 // by removing some allocations and/or locks. 2867 } while (progress); 2868 } 2869 2870 // Loop transforms on the ideal graph. Range Check Elimination, 2871 // peeling, unrolling, etc. 2872 2873 // Set loop opts counter 2874 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2875 { 2876 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2877 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2878 _loop_opts_cnt--; 2879 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2880 if (failing()) return; 2881 } 2882 // Loop opts pass if partial peeling occurred in previous pass 2883 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2884 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2885 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2886 _loop_opts_cnt--; 2887 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2888 if (failing()) return; 2889 } 2890 // Loop opts pass for loop-unrolling before CCP 2891 if(major_progress() && (_loop_opts_cnt > 0)) { 2892 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2893 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2894 _loop_opts_cnt--; 2895 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2896 } 2897 if (!failing()) { 2898 // Verify that last round of loop opts produced a valid graph 2899 PhaseIdealLoop::verify(igvn); 2900 } 2901 } 2902 if (failing()) return; 2903 2904 // Conditional Constant Propagation; 2905 print_method(PHASE_BEFORE_CCP1, 2); 2906 PhaseCCP ccp( &igvn ); 2907 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2908 { 2909 TracePhase tp("ccp", &timers[_t_ccp]); 2910 ccp.do_transform(); 2911 } 2912 print_method(PHASE_CCP1, 2); 2913 2914 assert( true, "Break here to ccp.dump_old2new_map()"); 2915 2916 // Iterative Global Value Numbering, including ideal transforms 2917 { 2918 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2919 igvn.reset_from_igvn(&ccp); 2920 igvn.optimize(); 2921 } 2922 print_method(PHASE_ITER_GVN2, 2); 2923 2924 if (failing()) return; 2925 2926 // Loop transforms on the ideal graph. Range Check Elimination, 2927 // peeling, unrolling, etc. 2928 if (!optimize_loops(igvn, LoopOptsDefault)) { 2929 return; 2930 } 2931 2932 if (failing()) return; 2933 2934 C->clear_major_progress(); // ensure that major progress is now clear 2935 2936 process_for_post_loop_opts_igvn(igvn); 2937 2938 if (failing()) return; 2939 2940 #ifdef ASSERT 2941 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2942 #endif 2943 2944 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2945 2946 if (_late_inlines.length() > 0) { 2947 // More opportunities to optimize virtual and MH calls. 2948 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2949 process_late_inline_calls_no_inline(igvn); 2950 } 2951 2952 { 2953 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2954 print_method(PHASE_BEFORE_MACRO_EXPANSION, 3); 2955 PhaseMacroExpand mex(igvn); 2956 if (mex.expand_macro_nodes()) { 2957 assert(failing(), "must bail out w/ explicit message"); 2958 return; 2959 } 2960 print_method(PHASE_AFTER_MACRO_EXPANSION, 2); 2961 } 2962 2963 // Process inline type nodes again and remove them. From here 2964 // on we don't need to keep track of field values anymore. 2965 process_inline_types(igvn, /* remove= */ true); 2966 2967 { 2968 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2969 if (bs->expand_barriers(this, igvn)) { 2970 assert(failing(), "must bail out w/ explicit message"); 2971 return; 2972 } 2973 print_method(PHASE_BARRIER_EXPANSION, 2); 2974 } 2975 2976 if (C->max_vector_size() > 0) { 2977 C->optimize_logic_cones(igvn); 2978 igvn.optimize(); 2979 if (failing()) return; 2980 } 2981 2982 DEBUG_ONLY( _modified_nodes = nullptr; ) 2983 DEBUG_ONLY( _late_inlines.clear(); ) 2984 2985 assert(igvn._worklist.size() == 0, "not empty"); 2986 } // (End scope of igvn; run destructor if necessary for asserts.) 2987 2988 check_no_dead_use(); 2989 2990 process_print_inlining(); 2991 2992 // We will never use the NodeHash table any more. Clear it so that final_graph_reshaping does not have 2993 // to remove hashes to unlock nodes for modifications. 2994 C->node_hash()->clear(); 2995 2996 // A method with only infinite loops has no edges entering loops from root 2997 { 2998 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2999 if (final_graph_reshaping()) { 3000 assert(failing(), "must bail out w/ explicit message"); 3001 return; 3002 } 3003 } 3004 3005 print_method(PHASE_OPTIMIZE_FINISHED, 2); 3006 DEBUG_ONLY(set_phase_optimize_finished();) 3007 } 3008 3009 #ifdef ASSERT 3010 void Compile::check_no_dead_use() const { 3011 ResourceMark rm; 3012 Unique_Node_List wq; 3013 wq.push(root()); 3014 for (uint i = 0; i < wq.size(); ++i) { 3015 Node* n = wq.at(i); 3016 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 3017 Node* u = n->fast_out(j); 3018 if (u->outcnt() == 0 && !u->is_Con()) { 3019 u->dump(); 3020 fatal("no reachable node should have no use"); 3021 } 3022 wq.push(u); 3023 } 3024 } 3025 } 3026 #endif 3027 3028 void Compile::inline_vector_reboxing_calls() { 3029 if (C->_vector_reboxing_late_inlines.length() > 0) { 3030 _late_inlines_pos = C->_late_inlines.length(); 3031 while (_vector_reboxing_late_inlines.length() > 0) { 3032 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 3033 cg->do_late_inline(); 3034 if (failing()) return; 3035 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 3036 } 3037 _vector_reboxing_late_inlines.trunc_to(0); 3038 } 3039 } 3040 3041 bool Compile::has_vbox_nodes() { 3042 if (C->_vector_reboxing_late_inlines.length() > 0) { 3043 return true; 3044 } 3045 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 3046 Node * n = C->macro_node(macro_idx); 3047 assert(n->is_macro(), "only macro nodes expected here"); 3048 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 3049 return true; 3050 } 3051 } 3052 return false; 3053 } 3054 3055 //---------------------------- Bitwise operation packing optimization --------------------------- 3056 3057 static bool is_vector_unary_bitwise_op(Node* n) { 3058 return n->Opcode() == Op_XorV && 3059 VectorNode::is_vector_bitwise_not_pattern(n); 3060 } 3061 3062 static bool is_vector_binary_bitwise_op(Node* n) { 3063 switch (n->Opcode()) { 3064 case Op_AndV: 3065 case Op_OrV: 3066 return true; 3067 3068 case Op_XorV: 3069 return !is_vector_unary_bitwise_op(n); 3070 3071 default: 3072 return false; 3073 } 3074 } 3075 3076 static bool is_vector_ternary_bitwise_op(Node* n) { 3077 return n->Opcode() == Op_MacroLogicV; 3078 } 3079 3080 static bool is_vector_bitwise_op(Node* n) { 3081 return is_vector_unary_bitwise_op(n) || 3082 is_vector_binary_bitwise_op(n) || 3083 is_vector_ternary_bitwise_op(n); 3084 } 3085 3086 static bool is_vector_bitwise_cone_root(Node* n) { 3087 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 3088 return false; 3089 } 3090 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3091 if (is_vector_bitwise_op(n->fast_out(i))) { 3092 return false; 3093 } 3094 } 3095 return true; 3096 } 3097 3098 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 3099 uint cnt = 0; 3100 if (is_vector_bitwise_op(n)) { 3101 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 3102 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 3103 for (uint i = 1; i < inp_cnt; i++) { 3104 Node* in = n->in(i); 3105 bool skip = VectorNode::is_all_ones_vector(in); 3106 if (!skip && !inputs.member(in)) { 3107 inputs.push(in); 3108 cnt++; 3109 } 3110 } 3111 assert(cnt <= 1, "not unary"); 3112 } else { 3113 uint last_req = inp_cnt; 3114 if (is_vector_ternary_bitwise_op(n)) { 3115 last_req = inp_cnt - 1; // skip last input 3116 } 3117 for (uint i = 1; i < last_req; i++) { 3118 Node* def = n->in(i); 3119 if (!inputs.member(def)) { 3120 inputs.push(def); 3121 cnt++; 3122 } 3123 } 3124 } 3125 } else { // not a bitwise operations 3126 if (!inputs.member(n)) { 3127 inputs.push(n); 3128 cnt++; 3129 } 3130 } 3131 return cnt; 3132 } 3133 3134 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 3135 Unique_Node_List useful_nodes; 3136 C->identify_useful_nodes(useful_nodes); 3137 3138 for (uint i = 0; i < useful_nodes.size(); i++) { 3139 Node* n = useful_nodes.at(i); 3140 if (is_vector_bitwise_cone_root(n)) { 3141 list.push(n); 3142 } 3143 } 3144 } 3145 3146 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 3147 const TypeVect* vt, 3148 Unique_Node_List& partition, 3149 Unique_Node_List& inputs) { 3150 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 3151 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 3152 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 3153 3154 Node* in1 = inputs.at(0); 3155 Node* in2 = inputs.at(1); 3156 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 3157 3158 uint func = compute_truth_table(partition, inputs); 3159 3160 Node* pn = partition.at(partition.size() - 1); 3161 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 3162 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 3163 } 3164 3165 static uint extract_bit(uint func, uint pos) { 3166 return (func & (1 << pos)) >> pos; 3167 } 3168 3169 // 3170 // A macro logic node represents a truth table. It has 4 inputs, 3171 // First three inputs corresponds to 3 columns of a truth table 3172 // and fourth input captures the logic function. 3173 // 3174 // eg. fn = (in1 AND in2) OR in3; 3175 // 3176 // MacroNode(in1,in2,in3,fn) 3177 // 3178 // ----------------- 3179 // in1 in2 in3 fn 3180 // ----------------- 3181 // 0 0 0 0 3182 // 0 0 1 1 3183 // 0 1 0 0 3184 // 0 1 1 1 3185 // 1 0 0 0 3186 // 1 0 1 1 3187 // 1 1 0 1 3188 // 1 1 1 1 3189 // 3190 3191 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 3192 int res = 0; 3193 for (int i = 0; i < 8; i++) { 3194 int bit1 = extract_bit(in1, i); 3195 int bit2 = extract_bit(in2, i); 3196 int bit3 = extract_bit(in3, i); 3197 3198 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 3199 int func_bit = extract_bit(func, func_bit_pos); 3200 3201 res |= func_bit << i; 3202 } 3203 return res; 3204 } 3205 3206 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 3207 assert(n != nullptr, ""); 3208 assert(eval_map.contains(n), "absent"); 3209 return *(eval_map.get(n)); 3210 } 3211 3212 static void eval_operands(Node* n, 3213 uint& func1, uint& func2, uint& func3, 3214 ResourceHashtable<Node*,uint>& eval_map) { 3215 assert(is_vector_bitwise_op(n), ""); 3216 3217 if (is_vector_unary_bitwise_op(n)) { 3218 Node* opnd = n->in(1); 3219 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 3220 opnd = n->in(2); 3221 } 3222 func1 = eval_operand(opnd, eval_map); 3223 } else if (is_vector_binary_bitwise_op(n)) { 3224 func1 = eval_operand(n->in(1), eval_map); 3225 func2 = eval_operand(n->in(2), eval_map); 3226 } else { 3227 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 3228 func1 = eval_operand(n->in(1), eval_map); 3229 func2 = eval_operand(n->in(2), eval_map); 3230 func3 = eval_operand(n->in(3), eval_map); 3231 } 3232 } 3233 3234 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 3235 assert(inputs.size() <= 3, "sanity"); 3236 ResourceMark rm; 3237 uint res = 0; 3238 ResourceHashtable<Node*,uint> eval_map; 3239 3240 // Populate precomputed functions for inputs. 3241 // Each input corresponds to one column of 3 input truth-table. 3242 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 3243 0xCC, // (_, b, _) -> b 3244 0xF0 }; // (a, _, _) -> a 3245 for (uint i = 0; i < inputs.size(); i++) { 3246 eval_map.put(inputs.at(i), input_funcs[2-i]); 3247 } 3248 3249 for (uint i = 0; i < partition.size(); i++) { 3250 Node* n = partition.at(i); 3251 3252 uint func1 = 0, func2 = 0, func3 = 0; 3253 eval_operands(n, func1, func2, func3, eval_map); 3254 3255 switch (n->Opcode()) { 3256 case Op_OrV: 3257 assert(func3 == 0, "not binary"); 3258 res = func1 | func2; 3259 break; 3260 case Op_AndV: 3261 assert(func3 == 0, "not binary"); 3262 res = func1 & func2; 3263 break; 3264 case Op_XorV: 3265 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 3266 assert(func2 == 0 && func3 == 0, "not unary"); 3267 res = (~func1) & 0xFF; 3268 } else { 3269 assert(func3 == 0, "not binary"); 3270 res = func1 ^ func2; 3271 } 3272 break; 3273 case Op_MacroLogicV: 3274 // Ordering of inputs may change during evaluation of sub-tree 3275 // containing MacroLogic node as a child node, thus a re-evaluation 3276 // makes sure that function is evaluated in context of current 3277 // inputs. 3278 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 3279 break; 3280 3281 default: assert(false, "not supported: %s", n->Name()); 3282 } 3283 assert(res <= 0xFF, "invalid"); 3284 eval_map.put(n, res); 3285 } 3286 return res; 3287 } 3288 3289 // Criteria under which nodes gets packed into a macro logic node:- 3290 // 1) Parent and both child nodes are all unmasked or masked with 3291 // same predicates. 3292 // 2) Masked parent can be packed with left child if it is predicated 3293 // and both have same predicates. 3294 // 3) Masked parent can be packed with right child if its un-predicated 3295 // or has matching predication condition. 3296 // 4) An unmasked parent can be packed with an unmasked child. 3297 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 3298 assert(partition.size() == 0, "not empty"); 3299 assert(inputs.size() == 0, "not empty"); 3300 if (is_vector_ternary_bitwise_op(n)) { 3301 return false; 3302 } 3303 3304 bool is_unary_op = is_vector_unary_bitwise_op(n); 3305 if (is_unary_op) { 3306 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 3307 return false; // too few inputs 3308 } 3309 3310 bool pack_left_child = true; 3311 bool pack_right_child = true; 3312 3313 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 3314 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 3315 3316 int left_child_input_cnt = 0; 3317 int right_child_input_cnt = 0; 3318 3319 bool parent_is_predicated = n->is_predicated_vector(); 3320 bool left_child_predicated = n->in(1)->is_predicated_vector(); 3321 bool right_child_predicated = n->in(2)->is_predicated_vector(); 3322 3323 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 3324 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 3325 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 3326 3327 do { 3328 if (pack_left_child && left_child_LOP && 3329 ((!parent_is_predicated && !left_child_predicated) || 3330 ((parent_is_predicated && left_child_predicated && 3331 parent_pred == left_child_pred)))) { 3332 partition.push(n->in(1)); 3333 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 3334 } else { 3335 inputs.push(n->in(1)); 3336 left_child_input_cnt = 1; 3337 } 3338 3339 if (pack_right_child && right_child_LOP && 3340 (!right_child_predicated || 3341 (right_child_predicated && parent_is_predicated && 3342 parent_pred == right_child_pred))) { 3343 partition.push(n->in(2)); 3344 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 3345 } else { 3346 inputs.push(n->in(2)); 3347 right_child_input_cnt = 1; 3348 } 3349 3350 if (inputs.size() > 3) { 3351 assert(partition.size() > 0, ""); 3352 inputs.clear(); 3353 partition.clear(); 3354 if (left_child_input_cnt > right_child_input_cnt) { 3355 pack_left_child = false; 3356 } else { 3357 pack_right_child = false; 3358 } 3359 } else { 3360 break; 3361 } 3362 } while(true); 3363 3364 if(partition.size()) { 3365 partition.push(n); 3366 } 3367 3368 return (partition.size() == 2 || partition.size() == 3) && 3369 (inputs.size() == 2 || inputs.size() == 3); 3370 } 3371 3372 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 3373 assert(is_vector_bitwise_op(n), "not a root"); 3374 3375 visited.set(n->_idx); 3376 3377 // 1) Do a DFS walk over the logic cone. 3378 for (uint i = 1; i < n->req(); i++) { 3379 Node* in = n->in(i); 3380 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 3381 process_logic_cone_root(igvn, in, visited); 3382 } 3383 } 3384 3385 // 2) Bottom up traversal: Merge node[s] with 3386 // the parent to form macro logic node. 3387 Unique_Node_List partition; 3388 Unique_Node_List inputs; 3389 if (compute_logic_cone(n, partition, inputs)) { 3390 const TypeVect* vt = n->bottom_type()->is_vect(); 3391 Node* pn = partition.at(partition.size() - 1); 3392 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 3393 if (mask == nullptr || 3394 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 3395 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 3396 VectorNode::trace_new_vector(macro_logic, "MacroLogic"); 3397 igvn.replace_node(n, macro_logic); 3398 } 3399 } 3400 } 3401 3402 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 3403 ResourceMark rm; 3404 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 3405 Unique_Node_List list; 3406 collect_logic_cone_roots(list); 3407 3408 while (list.size() > 0) { 3409 Node* n = list.pop(); 3410 const TypeVect* vt = n->bottom_type()->is_vect(); 3411 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 3412 if (supported) { 3413 VectorSet visited(comp_arena()); 3414 process_logic_cone_root(igvn, n, visited); 3415 } 3416 } 3417 } 3418 } 3419 3420 //------------------------------Code_Gen--------------------------------------- 3421 // Given a graph, generate code for it 3422 void Compile::Code_Gen() { 3423 if (failing()) { 3424 return; 3425 } 3426 3427 // Perform instruction selection. You might think we could reclaim Matcher 3428 // memory PDQ, but actually the Matcher is used in generating spill code. 3429 // Internals of the Matcher (including some VectorSets) must remain live 3430 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 3431 // set a bit in reclaimed memory. 3432 3433 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3434 // nodes. Mapping is only valid at the root of each matched subtree. 3435 NOT_PRODUCT( verify_graph_edges(); ) 3436 3437 Matcher matcher; 3438 _matcher = &matcher; 3439 { 3440 TracePhase tp("matcher", &timers[_t_matcher]); 3441 matcher.match(); 3442 if (failing()) { 3443 return; 3444 } 3445 } 3446 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 3447 // nodes. Mapping is only valid at the root of each matched subtree. 3448 NOT_PRODUCT( verify_graph_edges(); ) 3449 3450 // If you have too many nodes, or if matching has failed, bail out 3451 check_node_count(0, "out of nodes matching instructions"); 3452 if (failing()) { 3453 return; 3454 } 3455 3456 print_method(PHASE_MATCHING, 2); 3457 3458 // Build a proper-looking CFG 3459 PhaseCFG cfg(node_arena(), root(), matcher); 3460 _cfg = &cfg; 3461 { 3462 TracePhase tp("scheduler", &timers[_t_scheduler]); 3463 bool success = cfg.do_global_code_motion(); 3464 if (!success) { 3465 return; 3466 } 3467 3468 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 3469 NOT_PRODUCT( verify_graph_edges(); ) 3470 cfg.verify(); 3471 } 3472 3473 PhaseChaitin regalloc(unique(), cfg, matcher, false); 3474 _regalloc = ®alloc; 3475 { 3476 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 3477 // Perform register allocation. After Chaitin, use-def chains are 3478 // no longer accurate (at spill code) and so must be ignored. 3479 // Node->LRG->reg mappings are still accurate. 3480 _regalloc->Register_Allocate(); 3481 3482 // Bail out if the allocator builds too many nodes 3483 if (failing()) { 3484 return; 3485 } 3486 3487 print_method(PHASE_REGISTER_ALLOCATION, 2); 3488 } 3489 3490 // Prior to register allocation we kept empty basic blocks in case the 3491 // the allocator needed a place to spill. After register allocation we 3492 // are not adding any new instructions. If any basic block is empty, we 3493 // can now safely remove it. 3494 { 3495 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 3496 cfg.remove_empty_blocks(); 3497 if (do_freq_based_layout()) { 3498 PhaseBlockLayout layout(cfg); 3499 } else { 3500 cfg.set_loop_alignment(); 3501 } 3502 cfg.fixup_flow(); 3503 cfg.remove_unreachable_blocks(); 3504 cfg.verify_dominator_tree(); 3505 print_method(PHASE_BLOCK_ORDERING, 3); 3506 } 3507 3508 // Apply peephole optimizations 3509 if( OptoPeephole ) { 3510 TracePhase tp("peephole", &timers[_t_peephole]); 3511 PhasePeephole peep( _regalloc, cfg); 3512 peep.do_transform(); 3513 print_method(PHASE_PEEPHOLE, 3); 3514 } 3515 3516 // Do late expand if CPU requires this. 3517 if (Matcher::require_postalloc_expand) { 3518 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 3519 cfg.postalloc_expand(_regalloc); 3520 print_method(PHASE_POSTALLOC_EXPAND, 3); 3521 } 3522 3523 // Convert Nodes to instruction bits in a buffer 3524 { 3525 TracePhase tp("output", &timers[_t_output]); 3526 PhaseOutput output; 3527 output.Output(); 3528 if (failing()) return; 3529 output.install(); 3530 print_method(PHASE_FINAL_CODE, 1); // Compile::_output is not null here 3531 } 3532 3533 // He's dead, Jim. 3534 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 3535 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 3536 } 3537 3538 //------------------------------Final_Reshape_Counts--------------------------- 3539 // This class defines counters to help identify when a method 3540 // may/must be executed using hardware with only 24-bit precision. 3541 struct Final_Reshape_Counts : public StackObj { 3542 int _call_count; // count non-inlined 'common' calls 3543 int _float_count; // count float ops requiring 24-bit precision 3544 int _double_count; // count double ops requiring more precision 3545 int _java_call_count; // count non-inlined 'java' calls 3546 int _inner_loop_count; // count loops which need alignment 3547 VectorSet _visited; // Visitation flags 3548 Node_List _tests; // Set of IfNodes & PCTableNodes 3549 3550 Final_Reshape_Counts() : 3551 _call_count(0), _float_count(0), _double_count(0), 3552 _java_call_count(0), _inner_loop_count(0) { } 3553 3554 void inc_call_count () { _call_count ++; } 3555 void inc_float_count () { _float_count ++; } 3556 void inc_double_count() { _double_count++; } 3557 void inc_java_call_count() { _java_call_count++; } 3558 void inc_inner_loop_count() { _inner_loop_count++; } 3559 3560 int get_call_count () const { return _call_count ; } 3561 int get_float_count () const { return _float_count ; } 3562 int get_double_count() const { return _double_count; } 3563 int get_java_call_count() const { return _java_call_count; } 3564 int get_inner_loop_count() const { return _inner_loop_count; } 3565 }; 3566 3567 // Eliminate trivially redundant StoreCMs and accumulate their 3568 // precedence edges. 3569 void Compile::eliminate_redundant_card_marks(Node* n) { 3570 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3571 if (n->in(MemNode::Address)->outcnt() > 1) { 3572 // There are multiple users of the same address so it might be 3573 // possible to eliminate some of the StoreCMs 3574 Node* mem = n->in(MemNode::Memory); 3575 Node* adr = n->in(MemNode::Address); 3576 Node* val = n->in(MemNode::ValueIn); 3577 Node* prev = n; 3578 bool done = false; 3579 // Walk the chain of StoreCMs eliminating ones that match. As 3580 // long as it's a chain of single users then the optimization is 3581 // safe. Eliminating partially redundant StoreCMs would require 3582 // cloning copies down the other paths. 3583 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3584 if (adr == mem->in(MemNode::Address) && 3585 val == mem->in(MemNode::ValueIn)) { 3586 // redundant StoreCM 3587 if (mem->req() > MemNode::OopStore) { 3588 // Hasn't been processed by this code yet. 3589 n->add_prec(mem->in(MemNode::OopStore)); 3590 } else { 3591 // Already converted to precedence edge 3592 for (uint i = mem->req(); i < mem->len(); i++) { 3593 // Accumulate any precedence edges 3594 if (mem->in(i) != nullptr) { 3595 n->add_prec(mem->in(i)); 3596 } 3597 } 3598 // Everything above this point has been processed. 3599 done = true; 3600 } 3601 // Eliminate the previous StoreCM 3602 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3603 assert(mem->outcnt() == 0, "should be dead"); 3604 mem->disconnect_inputs(this); 3605 } else { 3606 prev = mem; 3607 } 3608 mem = prev->in(MemNode::Memory); 3609 } 3610 } 3611 } 3612 3613 3614 //------------------------------final_graph_reshaping_impl---------------------- 3615 // Implement items 1-5 from final_graph_reshaping below. 3616 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3617 3618 if ( n->outcnt() == 0 ) return; // dead node 3619 uint nop = n->Opcode(); 3620 3621 // Check for 2-input instruction with "last use" on right input. 3622 // Swap to left input. Implements item (2). 3623 if( n->req() == 3 && // two-input instruction 3624 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3625 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3626 n->in(2)->outcnt() == 1 &&// right use IS a last use 3627 !n->in(2)->is_Con() ) { // right use is not a constant 3628 // Check for commutative opcode 3629 switch( nop ) { 3630 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3631 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3632 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3633 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3634 case Op_AndL: case Op_XorL: case Op_OrL: 3635 case Op_AndI: case Op_XorI: case Op_OrI: { 3636 // Move "last use" input to left by swapping inputs 3637 n->swap_edges(1, 2); 3638 break; 3639 } 3640 default: 3641 break; 3642 } 3643 } 3644 3645 #ifdef ASSERT 3646 if( n->is_Mem() ) { 3647 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3648 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3649 // oop will be recorded in oop map if load crosses safepoint 3650 (n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3651 LoadNode::is_immutable_value(n->in(MemNode::Address)))), 3652 "raw memory operations should have control edge"); 3653 } 3654 if (n->is_MemBar()) { 3655 MemBarNode* mb = n->as_MemBar(); 3656 if (mb->trailing_store() || mb->trailing_load_store()) { 3657 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3658 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3659 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3660 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3661 } else if (mb->leading()) { 3662 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3663 } 3664 } 3665 #endif 3666 // Count FPU ops and common calls, implements item (3) 3667 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3668 if (!gc_handled) { 3669 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3670 } 3671 3672 // Collect CFG split points 3673 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3674 frc._tests.push(n); 3675 } 3676 } 3677 3678 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3679 switch( nop ) { 3680 // Count all float operations that may use FPU 3681 case Op_AddF: 3682 case Op_SubF: 3683 case Op_MulF: 3684 case Op_DivF: 3685 case Op_NegF: 3686 case Op_ModF: 3687 case Op_ConvI2F: 3688 case Op_ConF: 3689 case Op_CmpF: 3690 case Op_CmpF3: 3691 case Op_StoreF: 3692 case Op_LoadF: 3693 // case Op_ConvL2F: // longs are split into 32-bit halves 3694 frc.inc_float_count(); 3695 break; 3696 3697 case Op_ConvF2D: 3698 case Op_ConvD2F: 3699 frc.inc_float_count(); 3700 frc.inc_double_count(); 3701 break; 3702 3703 // Count all double operations that may use FPU 3704 case Op_AddD: 3705 case Op_SubD: 3706 case Op_MulD: 3707 case Op_DivD: 3708 case Op_NegD: 3709 case Op_ModD: 3710 case Op_ConvI2D: 3711 case Op_ConvD2I: 3712 // case Op_ConvL2D: // handled by leaf call 3713 // case Op_ConvD2L: // handled by leaf call 3714 case Op_ConD: 3715 case Op_CmpD: 3716 case Op_CmpD3: 3717 case Op_StoreD: 3718 case Op_LoadD: 3719 case Op_LoadD_unaligned: 3720 frc.inc_double_count(); 3721 break; 3722 case Op_Opaque1: // Remove Opaque Nodes before matching 3723 case Op_Opaque3: 3724 n->subsume_by(n->in(1), this); 3725 break; 3726 case Op_CallStaticJava: 3727 case Op_CallJava: 3728 case Op_CallDynamicJava: 3729 frc.inc_java_call_count(); // Count java call site; 3730 case Op_CallRuntime: 3731 case Op_CallLeaf: 3732 case Op_CallLeafVector: 3733 case Op_CallLeafNoFP: { 3734 assert (n->is_Call(), ""); 3735 CallNode *call = n->as_Call(); 3736 // Count call sites where the FP mode bit would have to be flipped. 3737 // Do not count uncommon runtime calls: 3738 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3739 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3740 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3741 frc.inc_call_count(); // Count the call site 3742 } else { // See if uncommon argument is shared 3743 Node *n = call->in(TypeFunc::Parms); 3744 int nop = n->Opcode(); 3745 // Clone shared simple arguments to uncommon calls, item (1). 3746 if (n->outcnt() > 1 && 3747 !n->is_Proj() && 3748 nop != Op_CreateEx && 3749 nop != Op_CheckCastPP && 3750 nop != Op_DecodeN && 3751 nop != Op_DecodeNKlass && 3752 !n->is_Mem() && 3753 !n->is_Phi()) { 3754 Node *x = n->clone(); 3755 call->set_req(TypeFunc::Parms, x); 3756 } 3757 } 3758 break; 3759 } 3760 3761 case Op_StoreCM: 3762 { 3763 // Convert OopStore dependence into precedence edge 3764 Node* prec = n->in(MemNode::OopStore); 3765 n->del_req(MemNode::OopStore); 3766 if (prec->is_MergeMem()) { 3767 MergeMemNode* mm = prec->as_MergeMem(); 3768 Node* base = mm->base_memory(); 3769 for (int i = AliasIdxRaw + 1; i < num_alias_types(); i++) { 3770 const TypePtr* adr_type = get_adr_type(i); 3771 if (adr_type->is_flat()) { 3772 Node* m = mm->memory_at(i); 3773 n->add_prec(m); 3774 } 3775 } 3776 if (mm->outcnt() == 0) { 3777 mm->disconnect_inputs(this); 3778 } 3779 } else { 3780 n->add_prec(prec); 3781 } 3782 eliminate_redundant_card_marks(n); 3783 } 3784 3785 // fall through 3786 3787 case Op_StoreB: 3788 case Op_StoreC: 3789 case Op_StoreI: 3790 case Op_StoreL: 3791 case Op_CompareAndSwapB: 3792 case Op_CompareAndSwapS: 3793 case Op_CompareAndSwapI: 3794 case Op_CompareAndSwapL: 3795 case Op_CompareAndSwapP: 3796 case Op_CompareAndSwapN: 3797 case Op_WeakCompareAndSwapB: 3798 case Op_WeakCompareAndSwapS: 3799 case Op_WeakCompareAndSwapI: 3800 case Op_WeakCompareAndSwapL: 3801 case Op_WeakCompareAndSwapP: 3802 case Op_WeakCompareAndSwapN: 3803 case Op_CompareAndExchangeB: 3804 case Op_CompareAndExchangeS: 3805 case Op_CompareAndExchangeI: 3806 case Op_CompareAndExchangeL: 3807 case Op_CompareAndExchangeP: 3808 case Op_CompareAndExchangeN: 3809 case Op_GetAndAddS: 3810 case Op_GetAndAddB: 3811 case Op_GetAndAddI: 3812 case Op_GetAndAddL: 3813 case Op_GetAndSetS: 3814 case Op_GetAndSetB: 3815 case Op_GetAndSetI: 3816 case Op_GetAndSetL: 3817 case Op_GetAndSetP: 3818 case Op_GetAndSetN: 3819 case Op_StoreP: 3820 case Op_StoreN: 3821 case Op_StoreNKlass: 3822 case Op_LoadB: 3823 case Op_LoadUB: 3824 case Op_LoadUS: 3825 case Op_LoadI: 3826 case Op_LoadKlass: 3827 case Op_LoadNKlass: 3828 case Op_LoadL: 3829 case Op_LoadL_unaligned: 3830 case Op_LoadP: 3831 case Op_LoadN: 3832 case Op_LoadRange: 3833 case Op_LoadS: 3834 break; 3835 3836 case Op_AddP: { // Assert sane base pointers 3837 Node *addp = n->in(AddPNode::Address); 3838 assert( !addp->is_AddP() || 3839 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3840 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3841 "Base pointers must match (addp %u)", addp->_idx ); 3842 #ifdef _LP64 3843 if ((UseCompressedOops || UseCompressedClassPointers) && 3844 addp->Opcode() == Op_ConP && 3845 addp == n->in(AddPNode::Base) && 3846 n->in(AddPNode::Offset)->is_Con()) { 3847 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3848 // on the platform and on the compressed oops mode. 3849 // Use addressing with narrow klass to load with offset on x86. 3850 // Some platforms can use the constant pool to load ConP. 3851 // Do this transformation here since IGVN will convert ConN back to ConP. 3852 const Type* t = addp->bottom_type(); 3853 bool is_oop = t->isa_oopptr() != nullptr; 3854 bool is_klass = t->isa_klassptr() != nullptr; 3855 3856 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3857 (is_klass && Matcher::const_klass_prefer_decode())) { 3858 Node* nn = nullptr; 3859 3860 int op = is_oop ? Op_ConN : Op_ConNKlass; 3861 3862 // Look for existing ConN node of the same exact type. 3863 Node* r = root(); 3864 uint cnt = r->outcnt(); 3865 for (uint i = 0; i < cnt; i++) { 3866 Node* m = r->raw_out(i); 3867 if (m!= nullptr && m->Opcode() == op && 3868 m->bottom_type()->make_ptr() == t) { 3869 nn = m; 3870 break; 3871 } 3872 } 3873 if (nn != nullptr) { 3874 // Decode a narrow oop to match address 3875 // [R12 + narrow_oop_reg<<3 + offset] 3876 if (is_oop) { 3877 nn = new DecodeNNode(nn, t); 3878 } else { 3879 nn = new DecodeNKlassNode(nn, t); 3880 } 3881 // Check for succeeding AddP which uses the same Base. 3882 // Otherwise we will run into the assertion above when visiting that guy. 3883 for (uint i = 0; i < n->outcnt(); ++i) { 3884 Node *out_i = n->raw_out(i); 3885 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3886 out_i->set_req(AddPNode::Base, nn); 3887 #ifdef ASSERT 3888 for (uint j = 0; j < out_i->outcnt(); ++j) { 3889 Node *out_j = out_i->raw_out(j); 3890 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3891 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3892 } 3893 #endif 3894 } 3895 } 3896 n->set_req(AddPNode::Base, nn); 3897 n->set_req(AddPNode::Address, nn); 3898 if (addp->outcnt() == 0) { 3899 addp->disconnect_inputs(this); 3900 } 3901 } 3902 } 3903 } 3904 #endif 3905 break; 3906 } 3907 3908 case Op_CastPP: { 3909 // Remove CastPP nodes to gain more freedom during scheduling but 3910 // keep the dependency they encode as control or precedence edges 3911 // (if control is set already) on memory operations. Some CastPP 3912 // nodes don't have a control (don't carry a dependency): skip 3913 // those. 3914 if (n->in(0) != nullptr) { 3915 ResourceMark rm; 3916 Unique_Node_List wq; 3917 wq.push(n); 3918 for (uint next = 0; next < wq.size(); ++next) { 3919 Node *m = wq.at(next); 3920 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3921 Node* use = m->fast_out(i); 3922 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3923 use->ensure_control_or_add_prec(n->in(0)); 3924 } else { 3925 switch(use->Opcode()) { 3926 case Op_AddP: 3927 case Op_DecodeN: 3928 case Op_DecodeNKlass: 3929 case Op_CheckCastPP: 3930 case Op_CastPP: 3931 wq.push(use); 3932 break; 3933 } 3934 } 3935 } 3936 } 3937 } 3938 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3939 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3940 Node* in1 = n->in(1); 3941 const Type* t = n->bottom_type(); 3942 Node* new_in1 = in1->clone(); 3943 new_in1->as_DecodeN()->set_type(t); 3944 3945 if (!Matcher::narrow_oop_use_complex_address()) { 3946 // 3947 // x86, ARM and friends can handle 2 adds in addressing mode 3948 // and Matcher can fold a DecodeN node into address by using 3949 // a narrow oop directly and do implicit null check in address: 3950 // 3951 // [R12 + narrow_oop_reg<<3 + offset] 3952 // NullCheck narrow_oop_reg 3953 // 3954 // On other platforms (Sparc) we have to keep new DecodeN node and 3955 // use it to do implicit null check in address: 3956 // 3957 // decode_not_null narrow_oop_reg, base_reg 3958 // [base_reg + offset] 3959 // NullCheck base_reg 3960 // 3961 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3962 // to keep the information to which null check the new DecodeN node 3963 // corresponds to use it as value in implicit_null_check(). 3964 // 3965 new_in1->set_req(0, n->in(0)); 3966 } 3967 3968 n->subsume_by(new_in1, this); 3969 if (in1->outcnt() == 0) { 3970 in1->disconnect_inputs(this); 3971 } 3972 } else { 3973 n->subsume_by(n->in(1), this); 3974 if (n->outcnt() == 0) { 3975 n->disconnect_inputs(this); 3976 } 3977 } 3978 break; 3979 } 3980 #ifdef _LP64 3981 case Op_CmpP: 3982 // Do this transformation here to preserve CmpPNode::sub() and 3983 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3984 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3985 Node* in1 = n->in(1); 3986 Node* in2 = n->in(2); 3987 if (!in1->is_DecodeNarrowPtr()) { 3988 in2 = in1; 3989 in1 = n->in(2); 3990 } 3991 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3992 3993 Node* new_in2 = nullptr; 3994 if (in2->is_DecodeNarrowPtr()) { 3995 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3996 new_in2 = in2->in(1); 3997 } else if (in2->Opcode() == Op_ConP) { 3998 const Type* t = in2->bottom_type(); 3999 if (t == TypePtr::NULL_PTR) { 4000 assert(in1->is_DecodeN(), "compare klass to null?"); 4001 // Don't convert CmpP null check into CmpN if compressed 4002 // oops implicit null check is not generated. 4003 // This will allow to generate normal oop implicit null check. 4004 if (Matcher::gen_narrow_oop_implicit_null_checks()) 4005 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 4006 // 4007 // This transformation together with CastPP transformation above 4008 // will generated code for implicit null checks for compressed oops. 4009 // 4010 // The original code after Optimize() 4011 // 4012 // LoadN memory, narrow_oop_reg 4013 // decode narrow_oop_reg, base_reg 4014 // CmpP base_reg, nullptr 4015 // CastPP base_reg // NotNull 4016 // Load [base_reg + offset], val_reg 4017 // 4018 // after these transformations will be 4019 // 4020 // LoadN memory, narrow_oop_reg 4021 // CmpN narrow_oop_reg, nullptr 4022 // decode_not_null narrow_oop_reg, base_reg 4023 // Load [base_reg + offset], val_reg 4024 // 4025 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 4026 // since narrow oops can be used in debug info now (see the code in 4027 // final_graph_reshaping_walk()). 4028 // 4029 // At the end the code will be matched to 4030 // on x86: 4031 // 4032 // Load_narrow_oop memory, narrow_oop_reg 4033 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 4034 // NullCheck narrow_oop_reg 4035 // 4036 // and on sparc: 4037 // 4038 // Load_narrow_oop memory, narrow_oop_reg 4039 // decode_not_null narrow_oop_reg, base_reg 4040 // Load [base_reg + offset], val_reg 4041 // NullCheck base_reg 4042 // 4043 } else if (t->isa_oopptr()) { 4044 new_in2 = ConNode::make(t->make_narrowoop()); 4045 } else if (t->isa_klassptr()) { 4046 new_in2 = ConNode::make(t->make_narrowklass()); 4047 } 4048 } 4049 if (new_in2 != nullptr) { 4050 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 4051 n->subsume_by(cmpN, this); 4052 if (in1->outcnt() == 0) { 4053 in1->disconnect_inputs(this); 4054 } 4055 if (in2->outcnt() == 0) { 4056 in2->disconnect_inputs(this); 4057 } 4058 } 4059 } 4060 break; 4061 4062 case Op_DecodeN: 4063 case Op_DecodeNKlass: 4064 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 4065 // DecodeN could be pinned when it can't be fold into 4066 // an address expression, see the code for Op_CastPP above. 4067 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 4068 break; 4069 4070 case Op_EncodeP: 4071 case Op_EncodePKlass: { 4072 Node* in1 = n->in(1); 4073 if (in1->is_DecodeNarrowPtr()) { 4074 n->subsume_by(in1->in(1), this); 4075 } else if (in1->Opcode() == Op_ConP) { 4076 const Type* t = in1->bottom_type(); 4077 if (t == TypePtr::NULL_PTR) { 4078 assert(t->isa_oopptr(), "null klass?"); 4079 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 4080 } else if (t->isa_oopptr()) { 4081 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 4082 } else if (t->isa_klassptr()) { 4083 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 4084 } 4085 } 4086 if (in1->outcnt() == 0) { 4087 in1->disconnect_inputs(this); 4088 } 4089 break; 4090 } 4091 4092 case Op_Proj: { 4093 if (OptimizeStringConcat || IncrementalInline) { 4094 ProjNode* proj = n->as_Proj(); 4095 if (proj->_is_io_use) { 4096 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 4097 // Separate projections were used for the exception path which 4098 // are normally removed by a late inline. If it wasn't inlined 4099 // then they will hang around and should just be replaced with 4100 // the original one. Merge them. 4101 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 4102 if (non_io_proj != nullptr) { 4103 proj->subsume_by(non_io_proj , this); 4104 } 4105 } 4106 } 4107 break; 4108 } 4109 4110 case Op_Phi: 4111 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 4112 // The EncodeP optimization may create Phi with the same edges 4113 // for all paths. It is not handled well by Register Allocator. 4114 Node* unique_in = n->in(1); 4115 assert(unique_in != nullptr, ""); 4116 uint cnt = n->req(); 4117 for (uint i = 2; i < cnt; i++) { 4118 Node* m = n->in(i); 4119 assert(m != nullptr, ""); 4120 if (unique_in != m) 4121 unique_in = nullptr; 4122 } 4123 if (unique_in != nullptr) { 4124 n->subsume_by(unique_in, this); 4125 } 4126 } 4127 break; 4128 4129 #endif 4130 4131 #ifdef ASSERT 4132 case Op_CastII: 4133 // Verify that all range check dependent CastII nodes were removed. 4134 if (n->isa_CastII()->has_range_check()) { 4135 n->dump(3); 4136 assert(false, "Range check dependent CastII node was not removed"); 4137 } 4138 break; 4139 #endif 4140 4141 case Op_ModI: 4142 if (UseDivMod) { 4143 // Check if a%b and a/b both exist 4144 Node* d = n->find_similar(Op_DivI); 4145 if (d) { 4146 // Replace them with a fused divmod if supported 4147 if (Matcher::has_match_rule(Op_DivModI)) { 4148 DivModINode* divmod = DivModINode::make(n); 4149 d->subsume_by(divmod->div_proj(), this); 4150 n->subsume_by(divmod->mod_proj(), this); 4151 } else { 4152 // replace a%b with a-((a/b)*b) 4153 Node* mult = new MulINode(d, d->in(2)); 4154 Node* sub = new SubINode(d->in(1), mult); 4155 n->subsume_by(sub, this); 4156 } 4157 } 4158 } 4159 break; 4160 4161 case Op_ModL: 4162 if (UseDivMod) { 4163 // Check if a%b and a/b both exist 4164 Node* d = n->find_similar(Op_DivL); 4165 if (d) { 4166 // Replace them with a fused divmod if supported 4167 if (Matcher::has_match_rule(Op_DivModL)) { 4168 DivModLNode* divmod = DivModLNode::make(n); 4169 d->subsume_by(divmod->div_proj(), this); 4170 n->subsume_by(divmod->mod_proj(), this); 4171 } else { 4172 // replace a%b with a-((a/b)*b) 4173 Node* mult = new MulLNode(d, d->in(2)); 4174 Node* sub = new SubLNode(d->in(1), mult); 4175 n->subsume_by(sub, this); 4176 } 4177 } 4178 } 4179 break; 4180 4181 case Op_UModI: 4182 if (UseDivMod) { 4183 // Check if a%b and a/b both exist 4184 Node* d = n->find_similar(Op_UDivI); 4185 if (d) { 4186 // Replace them with a fused unsigned divmod if supported 4187 if (Matcher::has_match_rule(Op_UDivModI)) { 4188 UDivModINode* divmod = UDivModINode::make(n); 4189 d->subsume_by(divmod->div_proj(), this); 4190 n->subsume_by(divmod->mod_proj(), this); 4191 } else { 4192 // replace a%b with a-((a/b)*b) 4193 Node* mult = new MulINode(d, d->in(2)); 4194 Node* sub = new SubINode(d->in(1), mult); 4195 n->subsume_by(sub, this); 4196 } 4197 } 4198 } 4199 break; 4200 4201 case Op_UModL: 4202 if (UseDivMod) { 4203 // Check if a%b and a/b both exist 4204 Node* d = n->find_similar(Op_UDivL); 4205 if (d) { 4206 // Replace them with a fused unsigned divmod if supported 4207 if (Matcher::has_match_rule(Op_UDivModL)) { 4208 UDivModLNode* divmod = UDivModLNode::make(n); 4209 d->subsume_by(divmod->div_proj(), this); 4210 n->subsume_by(divmod->mod_proj(), this); 4211 } else { 4212 // replace a%b with a-((a/b)*b) 4213 Node* mult = new MulLNode(d, d->in(2)); 4214 Node* sub = new SubLNode(d->in(1), mult); 4215 n->subsume_by(sub, this); 4216 } 4217 } 4218 } 4219 break; 4220 4221 case Op_LoadVector: 4222 case Op_StoreVector: 4223 #ifdef ASSERT 4224 // Add VerifyVectorAlignment node between adr and load / store. 4225 if (VerifyAlignVector && Matcher::has_match_rule(Op_VerifyVectorAlignment)) { 4226 bool must_verify_alignment = n->is_LoadVector() ? n->as_LoadVector()->must_verify_alignment() : 4227 n->as_StoreVector()->must_verify_alignment(); 4228 if (must_verify_alignment) { 4229 jlong vector_width = n->is_LoadVector() ? n->as_LoadVector()->memory_size() : 4230 n->as_StoreVector()->memory_size(); 4231 // The memory access should be aligned to the vector width in bytes. 4232 // However, the underlying array is possibly less well aligned, but at least 4233 // to ObjectAlignmentInBytes. Hence, even if multiple arrays are accessed in 4234 // a loop we can expect at least the following alignment: 4235 jlong guaranteed_alignment = MIN2(vector_width, (jlong)ObjectAlignmentInBytes); 4236 assert(2 <= guaranteed_alignment && guaranteed_alignment <= 64, "alignment must be in range"); 4237 assert(is_power_of_2(guaranteed_alignment), "alignment must be power of 2"); 4238 // Create mask from alignment. e.g. 0b1000 -> 0b0111 4239 jlong mask = guaranteed_alignment - 1; 4240 Node* mask_con = ConLNode::make(mask); 4241 VerifyVectorAlignmentNode* va = new VerifyVectorAlignmentNode(n->in(MemNode::Address), mask_con); 4242 n->set_req(MemNode::Address, va); 4243 } 4244 } 4245 #endif 4246 break; 4247 4248 case Op_LoadVectorGather: 4249 case Op_StoreVectorScatter: 4250 case Op_LoadVectorGatherMasked: 4251 case Op_StoreVectorScatterMasked: 4252 case Op_VectorCmpMasked: 4253 case Op_VectorMaskGen: 4254 case Op_LoadVectorMasked: 4255 case Op_StoreVectorMasked: 4256 break; 4257 4258 case Op_AddReductionVI: 4259 case Op_AddReductionVL: 4260 case Op_AddReductionVF: 4261 case Op_AddReductionVD: 4262 case Op_MulReductionVI: 4263 case Op_MulReductionVL: 4264 case Op_MulReductionVF: 4265 case Op_MulReductionVD: 4266 case Op_MinReductionV: 4267 case Op_MaxReductionV: 4268 case Op_AndReductionV: 4269 case Op_OrReductionV: 4270 case Op_XorReductionV: 4271 break; 4272 4273 case Op_PackB: 4274 case Op_PackS: 4275 case Op_PackI: 4276 case Op_PackF: 4277 case Op_PackL: 4278 case Op_PackD: 4279 if (n->req()-1 > 2) { 4280 // Replace many operand PackNodes with a binary tree for matching 4281 PackNode* p = (PackNode*) n; 4282 Node* btp = p->binary_tree_pack(1, n->req()); 4283 n->subsume_by(btp, this); 4284 } 4285 break; 4286 case Op_Loop: 4287 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 4288 case Op_CountedLoop: 4289 case Op_LongCountedLoop: 4290 case Op_OuterStripMinedLoop: 4291 if (n->as_Loop()->is_inner_loop()) { 4292 frc.inc_inner_loop_count(); 4293 } 4294 n->as_Loop()->verify_strip_mined(0); 4295 break; 4296 case Op_LShiftI: 4297 case Op_RShiftI: 4298 case Op_URShiftI: 4299 case Op_LShiftL: 4300 case Op_RShiftL: 4301 case Op_URShiftL: 4302 if (Matcher::need_masked_shift_count) { 4303 // The cpu's shift instructions don't restrict the count to the 4304 // lower 5/6 bits. We need to do the masking ourselves. 4305 Node* in2 = n->in(2); 4306 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 4307 const TypeInt* t = in2->find_int_type(); 4308 if (t != nullptr && t->is_con()) { 4309 juint shift = t->get_con(); 4310 if (shift > mask) { // Unsigned cmp 4311 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 4312 } 4313 } else { 4314 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 4315 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 4316 n->set_req(2, shift); 4317 } 4318 } 4319 if (in2->outcnt() == 0) { // Remove dead node 4320 in2->disconnect_inputs(this); 4321 } 4322 } 4323 break; 4324 case Op_MemBarStoreStore: 4325 case Op_MemBarRelease: 4326 // Break the link with AllocateNode: it is no longer useful and 4327 // confuses register allocation. 4328 if (n->req() > MemBarNode::Precedent) { 4329 n->set_req(MemBarNode::Precedent, top()); 4330 } 4331 break; 4332 case Op_MemBarAcquire: { 4333 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 4334 // At parse time, the trailing MemBarAcquire for a volatile load 4335 // is created with an edge to the load. After optimizations, 4336 // that input may be a chain of Phis. If those phis have no 4337 // other use, then the MemBarAcquire keeps them alive and 4338 // register allocation can be confused. 4339 dead_nodes.push(n->in(MemBarNode::Precedent)); 4340 n->set_req(MemBarNode::Precedent, top()); 4341 } 4342 break; 4343 } 4344 case Op_Blackhole: 4345 break; 4346 case Op_RangeCheck: { 4347 RangeCheckNode* rc = n->as_RangeCheck(); 4348 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 4349 n->subsume_by(iff, this); 4350 frc._tests.push(iff); 4351 break; 4352 } 4353 case Op_ConvI2L: { 4354 if (!Matcher::convi2l_type_required) { 4355 // Code generation on some platforms doesn't need accurate 4356 // ConvI2L types. Widening the type can help remove redundant 4357 // address computations. 4358 n->as_Type()->set_type(TypeLong::INT); 4359 ResourceMark rm; 4360 Unique_Node_List wq; 4361 wq.push(n); 4362 for (uint next = 0; next < wq.size(); next++) { 4363 Node *m = wq.at(next); 4364 4365 for(;;) { 4366 // Loop over all nodes with identical inputs edges as m 4367 Node* k = m->find_similar(m->Opcode()); 4368 if (k == nullptr) { 4369 break; 4370 } 4371 // Push their uses so we get a chance to remove node made 4372 // redundant 4373 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 4374 Node* u = k->fast_out(i); 4375 if (u->Opcode() == Op_LShiftL || 4376 u->Opcode() == Op_AddL || 4377 u->Opcode() == Op_SubL || 4378 u->Opcode() == Op_AddP) { 4379 wq.push(u); 4380 } 4381 } 4382 // Replace all nodes with identical edges as m with m 4383 k->subsume_by(m, this); 4384 } 4385 } 4386 } 4387 break; 4388 } 4389 case Op_CmpUL: { 4390 if (!Matcher::has_match_rule(Op_CmpUL)) { 4391 // No support for unsigned long comparisons 4392 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 4393 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 4394 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 4395 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 4396 Node* andl = new AndLNode(orl, remove_sign_mask); 4397 Node* cmp = new CmpLNode(andl, n->in(2)); 4398 n->subsume_by(cmp, this); 4399 } 4400 break; 4401 } 4402 #ifdef ASSERT 4403 case Op_InlineType: { 4404 n->dump(-1); 4405 assert(false, "inline type node was not removed"); 4406 break; 4407 } 4408 #endif 4409 default: 4410 assert(!n->is_Call(), ""); 4411 assert(!n->is_Mem(), ""); 4412 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 4413 break; 4414 } 4415 } 4416 4417 //------------------------------final_graph_reshaping_walk--------------------- 4418 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 4419 // requires that the walk visits a node's inputs before visiting the node. 4420 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 4421 Unique_Node_List sfpt; 4422 4423 frc._visited.set(root->_idx); // first, mark node as visited 4424 uint cnt = root->req(); 4425 Node *n = root; 4426 uint i = 0; 4427 while (true) { 4428 if (i < cnt) { 4429 // Place all non-visited non-null inputs onto stack 4430 Node* m = n->in(i); 4431 ++i; 4432 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 4433 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 4434 // compute worst case interpreter size in case of a deoptimization 4435 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 4436 4437 sfpt.push(m); 4438 } 4439 cnt = m->req(); 4440 nstack.push(n, i); // put on stack parent and next input's index 4441 n = m; 4442 i = 0; 4443 } 4444 } else { 4445 // Now do post-visit work 4446 final_graph_reshaping_impl(n, frc, dead_nodes); 4447 if (nstack.is_empty()) 4448 break; // finished 4449 n = nstack.node(); // Get node from stack 4450 cnt = n->req(); 4451 i = nstack.index(); 4452 nstack.pop(); // Shift to the next node on stack 4453 } 4454 } 4455 4456 // Skip next transformation if compressed oops are not used. 4457 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 4458 (!UseCompressedOops && !UseCompressedClassPointers)) 4459 return; 4460 4461 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 4462 // It could be done for an uncommon traps or any safepoints/calls 4463 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 4464 while (sfpt.size() > 0) { 4465 n = sfpt.pop(); 4466 JVMState *jvms = n->as_SafePoint()->jvms(); 4467 assert(jvms != nullptr, "sanity"); 4468 int start = jvms->debug_start(); 4469 int end = n->req(); 4470 bool is_uncommon = (n->is_CallStaticJava() && 4471 n->as_CallStaticJava()->uncommon_trap_request() != 0); 4472 for (int j = start; j < end; j++) { 4473 Node* in = n->in(j); 4474 if (in->is_DecodeNarrowPtr()) { 4475 bool safe_to_skip = true; 4476 if (!is_uncommon ) { 4477 // Is it safe to skip? 4478 for (uint i = 0; i < in->outcnt(); i++) { 4479 Node* u = in->raw_out(i); 4480 if (!u->is_SafePoint() || 4481 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 4482 safe_to_skip = false; 4483 } 4484 } 4485 } 4486 if (safe_to_skip) { 4487 n->set_req(j, in->in(1)); 4488 } 4489 if (in->outcnt() == 0) { 4490 in->disconnect_inputs(this); 4491 } 4492 } 4493 } 4494 } 4495 } 4496 4497 //------------------------------final_graph_reshaping-------------------------- 4498 // Final Graph Reshaping. 4499 // 4500 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 4501 // and not commoned up and forced early. Must come after regular 4502 // optimizations to avoid GVN undoing the cloning. Clone constant 4503 // inputs to Loop Phis; these will be split by the allocator anyways. 4504 // Remove Opaque nodes. 4505 // (2) Move last-uses by commutative operations to the left input to encourage 4506 // Intel update-in-place two-address operations and better register usage 4507 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 4508 // calls canonicalizing them back. 4509 // (3) Count the number of double-precision FP ops, single-precision FP ops 4510 // and call sites. On Intel, we can get correct rounding either by 4511 // forcing singles to memory (requires extra stores and loads after each 4512 // FP bytecode) or we can set a rounding mode bit (requires setting and 4513 // clearing the mode bit around call sites). The mode bit is only used 4514 // if the relative frequency of single FP ops to calls is low enough. 4515 // This is a key transform for SPEC mpeg_audio. 4516 // (4) Detect infinite loops; blobs of code reachable from above but not 4517 // below. Several of the Code_Gen algorithms fail on such code shapes, 4518 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4519 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4520 // Detection is by looking for IfNodes where only 1 projection is 4521 // reachable from below or CatchNodes missing some targets. 4522 // (5) Assert for insane oop offsets in debug mode. 4523 4524 bool Compile::final_graph_reshaping() { 4525 // an infinite loop may have been eliminated by the optimizer, 4526 // in which case the graph will be empty. 4527 if (root()->req() == 1) { 4528 // Do not compile method that is only a trivial infinite loop, 4529 // since the content of the loop may have been eliminated. 4530 record_method_not_compilable("trivial infinite loop"); 4531 return true; 4532 } 4533 4534 // Expensive nodes have their control input set to prevent the GVN 4535 // from freely commoning them. There's no GVN beyond this point so 4536 // no need to keep the control input. We want the expensive nodes to 4537 // be freely moved to the least frequent code path by gcm. 4538 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4539 for (int i = 0; i < expensive_count(); i++) { 4540 _expensive_nodes.at(i)->set_req(0, nullptr); 4541 } 4542 4543 Final_Reshape_Counts frc; 4544 4545 // Visit everybody reachable! 4546 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4547 Node_Stack nstack(live_nodes() >> 1); 4548 Unique_Node_List dead_nodes; 4549 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 4550 4551 // Check for unreachable (from below) code (i.e., infinite loops). 4552 for( uint i = 0; i < frc._tests.size(); i++ ) { 4553 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4554 // Get number of CFG targets. 4555 // Note that PCTables include exception targets after calls. 4556 uint required_outcnt = n->required_outcnt(); 4557 if (n->outcnt() != required_outcnt) { 4558 // Check for a few special cases. Rethrow Nodes never take the 4559 // 'fall-thru' path, so expected kids is 1 less. 4560 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4561 if (n->in(0)->in(0)->is_Call()) { 4562 CallNode* call = n->in(0)->in(0)->as_Call(); 4563 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4564 required_outcnt--; // Rethrow always has 1 less kid 4565 } else if (call->req() > TypeFunc::Parms && 4566 call->is_CallDynamicJava()) { 4567 // Check for null receiver. In such case, the optimizer has 4568 // detected that the virtual call will always result in a null 4569 // pointer exception. The fall-through projection of this CatchNode 4570 // will not be populated. 4571 Node* arg0 = call->in(TypeFunc::Parms); 4572 if (arg0->is_Type() && 4573 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4574 required_outcnt--; 4575 } 4576 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 4577 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4578 // Check for illegal array length. In such case, the optimizer has 4579 // detected that the allocation attempt will always result in an 4580 // exception. There is no fall-through projection of this CatchNode . 4581 assert(call->is_CallStaticJava(), "static call expected"); 4582 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4583 uint valid_length_test_input = call->req() - 1; 4584 Node* valid_length_test = call->in(valid_length_test_input); 4585 call->del_req(valid_length_test_input); 4586 if (valid_length_test->find_int_con(1) == 0) { 4587 required_outcnt--; 4588 } 4589 dead_nodes.push(valid_length_test); 4590 assert(n->outcnt() == required_outcnt, "malformed control flow"); 4591 continue; 4592 } 4593 } 4594 } 4595 4596 // Recheck with a better notion of 'required_outcnt' 4597 if (n->outcnt() != required_outcnt) { 4598 record_method_not_compilable("malformed control flow"); 4599 return true; // Not all targets reachable! 4600 } 4601 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4602 CallNode* call = n->in(0)->in(0)->as_Call(); 4603 if (call->entry_point() == OptoRuntime::new_array_Java() || 4604 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4605 assert(call->is_CallStaticJava(), "static call expected"); 4606 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4607 uint valid_length_test_input = call->req() - 1; 4608 dead_nodes.push(call->in(valid_length_test_input)); 4609 call->del_req(valid_length_test_input); // valid length test useless now 4610 } 4611 } 4612 // Check that I actually visited all kids. Unreached kids 4613 // must be infinite loops. 4614 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4615 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4616 record_method_not_compilable("infinite loop"); 4617 return true; // Found unvisited kid; must be unreach 4618 } 4619 4620 // Here so verification code in final_graph_reshaping_walk() 4621 // always see an OuterStripMinedLoopEnd 4622 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4623 IfNode* init_iff = n->as_If(); 4624 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4625 n->subsume_by(iff, this); 4626 } 4627 } 4628 4629 while (dead_nodes.size() > 0) { 4630 Node* m = dead_nodes.pop(); 4631 if (m->outcnt() == 0 && m != top()) { 4632 for (uint j = 0; j < m->req(); j++) { 4633 Node* in = m->in(j); 4634 if (in != nullptr) { 4635 dead_nodes.push(in); 4636 } 4637 } 4638 m->disconnect_inputs(this); 4639 } 4640 } 4641 4642 #ifdef IA32 4643 // If original bytecodes contained a mixture of floats and doubles 4644 // check if the optimizer has made it homogeneous, item (3). 4645 if (UseSSE == 0 && 4646 frc.get_float_count() > 32 && 4647 frc.get_double_count() == 0 && 4648 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4649 set_24_bit_selection_and_mode(false, true); 4650 } 4651 #endif // IA32 4652 4653 set_java_calls(frc.get_java_call_count()); 4654 set_inner_loops(frc.get_inner_loop_count()); 4655 4656 // No infinite loops, no reason to bail out. 4657 return false; 4658 } 4659 4660 //-----------------------------too_many_traps---------------------------------- 4661 // Report if there are too many traps at the current method and bci. 4662 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4663 bool Compile::too_many_traps(ciMethod* method, 4664 int bci, 4665 Deoptimization::DeoptReason reason) { 4666 ciMethodData* md = method->method_data(); 4667 if (md->is_empty()) { 4668 // Assume the trap has not occurred, or that it occurred only 4669 // because of a transient condition during start-up in the interpreter. 4670 return false; 4671 } 4672 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4673 if (md->has_trap_at(bci, m, reason) != 0) { 4674 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4675 // Also, if there are multiple reasons, or if there is no per-BCI record, 4676 // assume the worst. 4677 if (log()) 4678 log()->elem("observe trap='%s' count='%d'", 4679 Deoptimization::trap_reason_name(reason), 4680 md->trap_count(reason)); 4681 return true; 4682 } else { 4683 // Ignore method/bci and see if there have been too many globally. 4684 return too_many_traps(reason, md); 4685 } 4686 } 4687 4688 // Less-accurate variant which does not require a method and bci. 4689 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4690 ciMethodData* logmd) { 4691 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4692 // Too many traps globally. 4693 // Note that we use cumulative trap_count, not just md->trap_count. 4694 if (log()) { 4695 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4696 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4697 Deoptimization::trap_reason_name(reason), 4698 mcount, trap_count(reason)); 4699 } 4700 return true; 4701 } else { 4702 // The coast is clear. 4703 return false; 4704 } 4705 } 4706 4707 //--------------------------too_many_recompiles-------------------------------- 4708 // Report if there are too many recompiles at the current method and bci. 4709 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4710 // Is not eager to return true, since this will cause the compiler to use 4711 // Action_none for a trap point, to avoid too many recompilations. 4712 bool Compile::too_many_recompiles(ciMethod* method, 4713 int bci, 4714 Deoptimization::DeoptReason reason) { 4715 ciMethodData* md = method->method_data(); 4716 if (md->is_empty()) { 4717 // Assume the trap has not occurred, or that it occurred only 4718 // because of a transient condition during start-up in the interpreter. 4719 return false; 4720 } 4721 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4722 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4723 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4724 Deoptimization::DeoptReason per_bc_reason 4725 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4726 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4727 if ((per_bc_reason == Deoptimization::Reason_none 4728 || md->has_trap_at(bci, m, reason) != 0) 4729 // The trap frequency measure we care about is the recompile count: 4730 && md->trap_recompiled_at(bci, m) 4731 && md->overflow_recompile_count() >= bc_cutoff) { 4732 // Do not emit a trap here if it has already caused recompilations. 4733 // Also, if there are multiple reasons, or if there is no per-BCI record, 4734 // assume the worst. 4735 if (log()) 4736 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4737 Deoptimization::trap_reason_name(reason), 4738 md->trap_count(reason), 4739 md->overflow_recompile_count()); 4740 return true; 4741 } else if (trap_count(reason) != 0 4742 && decompile_count() >= m_cutoff) { 4743 // Too many recompiles globally, and we have seen this sort of trap. 4744 // Use cumulative decompile_count, not just md->decompile_count. 4745 if (log()) 4746 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4747 Deoptimization::trap_reason_name(reason), 4748 md->trap_count(reason), trap_count(reason), 4749 md->decompile_count(), decompile_count()); 4750 return true; 4751 } else { 4752 // The coast is clear. 4753 return false; 4754 } 4755 } 4756 4757 // Compute when not to trap. Used by matching trap based nodes and 4758 // NullCheck optimization. 4759 void Compile::set_allowed_deopt_reasons() { 4760 _allowed_reasons = 0; 4761 if (is_method_compilation()) { 4762 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4763 assert(rs < BitsPerInt, "recode bit map"); 4764 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4765 _allowed_reasons |= nth_bit(rs); 4766 } 4767 } 4768 } 4769 } 4770 4771 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4772 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4773 } 4774 4775 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4776 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4777 } 4778 4779 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4780 if (holder->is_initialized()) { 4781 return false; 4782 } 4783 if (holder->is_being_initialized()) { 4784 if (accessing_method->holder() == holder) { 4785 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4786 // <init>, or a static method. In all those cases, there was an initialization 4787 // barrier on the holder klass passed. 4788 if (accessing_method->is_class_initializer() || 4789 accessing_method->is_object_constructor() || 4790 accessing_method->is_static()) { 4791 return false; 4792 } 4793 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4794 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4795 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4796 // child class can become fully initialized while its parent class is still being initialized. 4797 if (accessing_method->is_class_initializer()) { 4798 return false; 4799 } 4800 } 4801 ciMethod* root = method(); // the root method of compilation 4802 if (root != accessing_method) { 4803 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4804 } 4805 } 4806 return true; 4807 } 4808 4809 #ifndef PRODUCT 4810 //------------------------------verify_bidirectional_edges--------------------- 4811 // For each input edge to a node (ie - for each Use-Def edge), verify that 4812 // there is a corresponding Def-Use edge. 4813 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4814 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4815 uint stack_size = live_nodes() >> 4; 4816 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4817 nstack.push(_root); 4818 4819 while (nstack.size() > 0) { 4820 Node* n = nstack.pop(); 4821 if (visited.member(n)) { 4822 continue; 4823 } 4824 visited.push(n); 4825 4826 // Walk over all input edges, checking for correspondence 4827 uint length = n->len(); 4828 for (uint i = 0; i < length; i++) { 4829 Node* in = n->in(i); 4830 if (in != nullptr && !visited.member(in)) { 4831 nstack.push(in); // Put it on stack 4832 } 4833 if (in != nullptr && !in->is_top()) { 4834 // Count instances of `next` 4835 int cnt = 0; 4836 for (uint idx = 0; idx < in->_outcnt; idx++) { 4837 if (in->_out[idx] == n) { 4838 cnt++; 4839 } 4840 } 4841 assert(cnt > 0, "Failed to find Def-Use edge."); 4842 // Check for duplicate edges 4843 // walk the input array downcounting the input edges to n 4844 for (uint j = 0; j < length; j++) { 4845 if (n->in(j) == in) { 4846 cnt--; 4847 } 4848 } 4849 assert(cnt == 0, "Mismatched edge count."); 4850 } else if (in == nullptr) { 4851 assert(i == 0 || i >= n->req() || 4852 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4853 (n->is_Allocate() && i >= AllocateNode::InlineType) || 4854 (n->is_Unlock() && i == (n->req() - 1)) || 4855 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4856 "only region, phi, arraycopy, allocate, unlock or membar nodes have null data edges"); 4857 } else { 4858 assert(in->is_top(), "sanity"); 4859 // Nothing to check. 4860 } 4861 } 4862 } 4863 } 4864 4865 //------------------------------verify_graph_edges--------------------------- 4866 // Walk the Graph and verify that there is a one-to-one correspondence 4867 // between Use-Def edges and Def-Use edges in the graph. 4868 void Compile::verify_graph_edges(bool no_dead_code) { 4869 if (VerifyGraphEdges) { 4870 Unique_Node_List visited; 4871 4872 // Call graph walk to check edges 4873 verify_bidirectional_edges(visited); 4874 if (no_dead_code) { 4875 // Now make sure that no visited node is used by an unvisited node. 4876 bool dead_nodes = false; 4877 Unique_Node_List checked; 4878 while (visited.size() > 0) { 4879 Node* n = visited.pop(); 4880 checked.push(n); 4881 for (uint i = 0; i < n->outcnt(); i++) { 4882 Node* use = n->raw_out(i); 4883 if (checked.member(use)) continue; // already checked 4884 if (visited.member(use)) continue; // already in the graph 4885 if (use->is_Con()) continue; // a dead ConNode is OK 4886 // At this point, we have found a dead node which is DU-reachable. 4887 if (!dead_nodes) { 4888 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4889 dead_nodes = true; 4890 } 4891 use->dump(2); 4892 tty->print_cr("---"); 4893 checked.push(use); // No repeats; pretend it is now checked. 4894 } 4895 } 4896 assert(!dead_nodes, "using nodes must be reachable from root"); 4897 } 4898 } 4899 } 4900 #endif 4901 4902 // The Compile object keeps track of failure reasons separately from the ciEnv. 4903 // This is required because there is not quite a 1-1 relation between the 4904 // ciEnv and its compilation task and the Compile object. Note that one 4905 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4906 // to backtrack and retry without subsuming loads. Other than this backtracking 4907 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4908 // by the logic in C2Compiler. 4909 void Compile::record_failure(const char* reason) { 4910 if (log() != nullptr) { 4911 log()->elem("failure reason='%s' phase='compile'", reason); 4912 } 4913 if (_failure_reason == nullptr) { 4914 // Record the first failure reason. 4915 _failure_reason = reason; 4916 if (CaptureBailoutInformation) { 4917 _first_failure_details = new CompilationFailureInfo(reason); 4918 } 4919 } 4920 4921 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4922 C->print_method(PHASE_FAILURE, 1); 4923 } 4924 _root = nullptr; // flush the graph, too 4925 } 4926 4927 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4928 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4929 _compile(Compile::current()), 4930 _log(nullptr), 4931 _phase_name(name), 4932 _dolog(CITimeVerbose) 4933 { 4934 assert(_compile != nullptr, "sanity check"); 4935 if (_dolog) { 4936 _log = _compile->log(); 4937 } 4938 if (_log != nullptr) { 4939 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4940 _log->stamp(); 4941 _log->end_head(); 4942 } 4943 } 4944 4945 Compile::TracePhase::~TracePhase() { 4946 if (_compile->failing()) return; 4947 #ifdef ASSERT 4948 if (PrintIdealNodeCount) { 4949 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4950 _phase_name, _compile->unique(), _compile->live_nodes(), _compile->count_live_nodes_by_graph_walk()); 4951 } 4952 4953 if (VerifyIdealNodeCount) { 4954 _compile->print_missing_nodes(); 4955 } 4956 #endif 4957 4958 if (_log != nullptr) { 4959 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, _compile->unique(), _compile->live_nodes()); 4960 } 4961 } 4962 4963 //----------------------------static_subtype_check----------------------------- 4964 // Shortcut important common cases when superklass is exact: 4965 // (0) superklass is java.lang.Object (can occur in reflective code) 4966 // (1) subklass is already limited to a subtype of superklass => always ok 4967 // (2) subklass does not overlap with superklass => always fail 4968 // (3) superklass has NO subtypes and we can check with a simple compare. 4969 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4970 if (skip) { 4971 return SSC_full_test; // Let caller generate the general case. 4972 } 4973 4974 if (subk->is_java_subtype_of(superk)) { 4975 return SSC_always_true; // (0) and (1) this test cannot fail 4976 } 4977 4978 if (!subk->maybe_java_subtype_of(superk)) { 4979 return SSC_always_false; // (2) true path dead; no dynamic test needed 4980 } 4981 4982 const Type* superelem = superk; 4983 if (superk->isa_aryklassptr()) { 4984 int ignored; 4985 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4986 4987 // TODO 8325106 Fix comment 4988 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays. 4989 // [QMyValue is a subtype of [LMyValue but the klass for [QMyValue is not equal to 4990 // the klass for [LMyValue. Perform a full test. 4991 if (!superk->is_aryklassptr()->is_null_free() && superk->is_aryklassptr()->elem()->isa_instklassptr() && 4992 superk->is_aryklassptr()->elem()->is_instklassptr()->instance_klass()->is_inlinetype()) { 4993 return SSC_full_test; 4994 } 4995 } 4996 4997 if (superelem->isa_instklassptr()) { 4998 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4999 if (!ik->has_subklass()) { 5000 if (!ik->is_final()) { 5001 // Add a dependency if there is a chance of a later subclass. 5002 dependencies()->assert_leaf_type(ik); 5003 } 5004 if (!superk->maybe_java_subtype_of(subk)) { 5005 return SSC_always_false; 5006 } 5007 return SSC_easy_test; // (3) caller can do a simple ptr comparison 5008 } 5009 } else { 5010 // A primitive array type has no subtypes. 5011 return SSC_easy_test; // (3) caller can do a simple ptr comparison 5012 } 5013 5014 return SSC_full_test; 5015 } 5016 5017 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 5018 #ifdef _LP64 5019 // The scaled index operand to AddP must be a clean 64-bit value. 5020 // Java allows a 32-bit int to be incremented to a negative 5021 // value, which appears in a 64-bit register as a large 5022 // positive number. Using that large positive number as an 5023 // operand in pointer arithmetic has bad consequences. 5024 // On the other hand, 32-bit overflow is rare, and the possibility 5025 // can often be excluded, if we annotate the ConvI2L node with 5026 // a type assertion that its value is known to be a small positive 5027 // number. (The prior range check has ensured this.) 5028 // This assertion is used by ConvI2LNode::Ideal. 5029 int index_max = max_jint - 1; // array size is max_jint, index is one less 5030 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 5031 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 5032 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 5033 #endif 5034 return idx; 5035 } 5036 5037 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 5038 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 5039 if (ctrl != nullptr) { 5040 // Express control dependency by a CastII node with a narrow type. 5041 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 5042 // node from floating above the range check during loop optimizations. Otherwise, the 5043 // ConvI2L node may be eliminated independently of the range check, causing the data path 5044 // to become TOP while the control path is still there (although it's unreachable). 5045 value = new CastIINode(ctrl, value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 5046 value = phase->transform(value); 5047 } 5048 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 5049 return phase->transform(new ConvI2LNode(value, ltype)); 5050 } 5051 5052 // The message about the current inlining is accumulated in 5053 // _print_inlining_stream and transferred into the _print_inlining_list 5054 // once we know whether inlining succeeds or not. For regular 5055 // inlining, messages are appended to the buffer pointed by 5056 // _print_inlining_idx in the _print_inlining_list. For late inlining, 5057 // a new buffer is added after _print_inlining_idx in the list. This 5058 // way we can update the inlining message for late inlining call site 5059 // when the inlining is attempted again. 5060 void Compile::print_inlining_init() { 5061 if (print_inlining() || print_intrinsics()) { 5062 // print_inlining_init is actually called several times. 5063 print_inlining_reset(); 5064 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 5065 } 5066 } 5067 5068 void Compile::print_inlining_reinit() { 5069 if (print_inlining() || print_intrinsics()) { 5070 print_inlining_reset(); 5071 } 5072 } 5073 5074 void Compile::print_inlining_reset() { 5075 _print_inlining_stream->reset(); 5076 } 5077 5078 void Compile::print_inlining_commit() { 5079 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 5080 // Transfer the message from _print_inlining_stream to the current 5081 // _print_inlining_list buffer and clear _print_inlining_stream. 5082 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 5083 print_inlining_reset(); 5084 } 5085 5086 void Compile::print_inlining_push() { 5087 // Add new buffer to the _print_inlining_list at current position 5088 _print_inlining_idx++; 5089 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 5090 } 5091 5092 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 5093 return _print_inlining_list->at(_print_inlining_idx); 5094 } 5095 5096 void Compile::print_inlining_update(CallGenerator* cg) { 5097 if (print_inlining() || print_intrinsics()) { 5098 if (cg->is_late_inline()) { 5099 if (print_inlining_current()->cg() != cg && 5100 (print_inlining_current()->cg() != nullptr || 5101 print_inlining_current()->ss()->size() != 0)) { 5102 print_inlining_push(); 5103 } 5104 print_inlining_commit(); 5105 print_inlining_current()->set_cg(cg); 5106 } else { 5107 if (print_inlining_current()->cg() != nullptr) { 5108 print_inlining_push(); 5109 } 5110 print_inlining_commit(); 5111 } 5112 } 5113 } 5114 5115 void Compile::print_inlining_move_to(CallGenerator* cg) { 5116 // We resume inlining at a late inlining call site. Locate the 5117 // corresponding inlining buffer so that we can update it. 5118 if (print_inlining() || print_intrinsics()) { 5119 for (int i = 0; i < _print_inlining_list->length(); i++) { 5120 if (_print_inlining_list->at(i)->cg() == cg) { 5121 _print_inlining_idx = i; 5122 return; 5123 } 5124 } 5125 ShouldNotReachHere(); 5126 } 5127 } 5128 5129 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 5130 if (print_inlining() || print_intrinsics()) { 5131 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 5132 assert(print_inlining_current()->cg() == cg, "wrong entry"); 5133 // replace message with new message 5134 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 5135 print_inlining_commit(); 5136 print_inlining_current()->set_cg(cg); 5137 } 5138 } 5139 5140 void Compile::print_inlining_assert_ready() { 5141 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 5142 } 5143 5144 void Compile::process_print_inlining() { 5145 assert(_late_inlines.length() == 0, "not drained yet"); 5146 if (print_inlining() || print_intrinsics()) { 5147 ResourceMark rm; 5148 stringStream ss; 5149 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 5150 for (int i = 0; i < _print_inlining_list->length(); i++) { 5151 PrintInliningBuffer* pib = _print_inlining_list->at(i); 5152 ss.print("%s", pib->ss()->freeze()); 5153 delete pib; 5154 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 5155 } 5156 // Reset _print_inlining_list, it only contains destructed objects. 5157 // It is on the arena, so it will be freed when the arena is reset. 5158 _print_inlining_list = nullptr; 5159 // _print_inlining_stream won't be used anymore, either. 5160 print_inlining_reset(); 5161 size_t end = ss.size(); 5162 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 5163 strncpy(_print_inlining_output, ss.freeze(), end+1); 5164 _print_inlining_output[end] = 0; 5165 } 5166 } 5167 5168 void Compile::dump_print_inlining() { 5169 if (_print_inlining_output != nullptr) { 5170 tty->print_raw(_print_inlining_output); 5171 } 5172 } 5173 5174 void Compile::log_late_inline(CallGenerator* cg) { 5175 if (log() != nullptr) { 5176 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 5177 cg->unique_id()); 5178 JVMState* p = cg->call_node()->jvms(); 5179 while (p != nullptr) { 5180 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 5181 p = p->caller(); 5182 } 5183 log()->tail("late_inline"); 5184 } 5185 } 5186 5187 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 5188 log_late_inline(cg); 5189 if (log() != nullptr) { 5190 log()->inline_fail(msg); 5191 } 5192 } 5193 5194 void Compile::log_inline_id(CallGenerator* cg) { 5195 if (log() != nullptr) { 5196 // The LogCompilation tool needs a unique way to identify late 5197 // inline call sites. This id must be unique for this call site in 5198 // this compilation. Try to have it unique across compilations as 5199 // well because it can be convenient when grepping through the log 5200 // file. 5201 // Distinguish OSR compilations from others in case CICountOSR is 5202 // on. 5203 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 5204 cg->set_unique_id(id); 5205 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 5206 } 5207 } 5208 5209 void Compile::log_inline_failure(const char* msg) { 5210 if (C->log() != nullptr) { 5211 C->log()->inline_fail(msg); 5212 } 5213 } 5214 5215 5216 // Dump inlining replay data to the stream. 5217 // Don't change thread state and acquire any locks. 5218 void Compile::dump_inline_data(outputStream* out) { 5219 InlineTree* inl_tree = ilt(); 5220 if (inl_tree != nullptr) { 5221 out->print(" inline %d", inl_tree->count()); 5222 inl_tree->dump_replay_data(out); 5223 } 5224 } 5225 5226 void Compile::dump_inline_data_reduced(outputStream* out) { 5227 assert(ReplayReduce, ""); 5228 5229 InlineTree* inl_tree = ilt(); 5230 if (inl_tree == nullptr) { 5231 return; 5232 } 5233 // Enable iterative replay file reduction 5234 // Output "compile" lines for depth 1 subtrees, 5235 // simulating that those trees were compiled 5236 // instead of inlined. 5237 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 5238 InlineTree* sub = inl_tree->subtrees().at(i); 5239 if (sub->inline_level() != 1) { 5240 continue; 5241 } 5242 5243 ciMethod* method = sub->method(); 5244 int entry_bci = -1; 5245 int comp_level = env()->task()->comp_level(); 5246 out->print("compile "); 5247 method->dump_name_as_ascii(out); 5248 out->print(" %d %d", entry_bci, comp_level); 5249 out->print(" inline %d", sub->count()); 5250 sub->dump_replay_data(out, -1); 5251 out->cr(); 5252 } 5253 } 5254 5255 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 5256 if (n1->Opcode() < n2->Opcode()) return -1; 5257 else if (n1->Opcode() > n2->Opcode()) return 1; 5258 5259 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 5260 for (uint i = 1; i < n1->req(); i++) { 5261 if (n1->in(i) < n2->in(i)) return -1; 5262 else if (n1->in(i) > n2->in(i)) return 1; 5263 } 5264 5265 return 0; 5266 } 5267 5268 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 5269 Node* n1 = *n1p; 5270 Node* n2 = *n2p; 5271 5272 return cmp_expensive_nodes(n1, n2); 5273 } 5274 5275 void Compile::sort_expensive_nodes() { 5276 if (!expensive_nodes_sorted()) { 5277 _expensive_nodes.sort(cmp_expensive_nodes); 5278 } 5279 } 5280 5281 bool Compile::expensive_nodes_sorted() const { 5282 for (int i = 1; i < _expensive_nodes.length(); i++) { 5283 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 5284 return false; 5285 } 5286 } 5287 return true; 5288 } 5289 5290 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 5291 if (_expensive_nodes.length() == 0) { 5292 return false; 5293 } 5294 5295 assert(OptimizeExpensiveOps, "optimization off?"); 5296 5297 // Take this opportunity to remove dead nodes from the list 5298 int j = 0; 5299 for (int i = 0; i < _expensive_nodes.length(); i++) { 5300 Node* n = _expensive_nodes.at(i); 5301 if (!n->is_unreachable(igvn)) { 5302 assert(n->is_expensive(), "should be expensive"); 5303 _expensive_nodes.at_put(j, n); 5304 j++; 5305 } 5306 } 5307 _expensive_nodes.trunc_to(j); 5308 5309 // Then sort the list so that similar nodes are next to each other 5310 // and check for at least two nodes of identical kind with same data 5311 // inputs. 5312 sort_expensive_nodes(); 5313 5314 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 5315 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 5316 return true; 5317 } 5318 } 5319 5320 return false; 5321 } 5322 5323 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 5324 if (_expensive_nodes.length() == 0) { 5325 return; 5326 } 5327 5328 assert(OptimizeExpensiveOps, "optimization off?"); 5329 5330 // Sort to bring similar nodes next to each other and clear the 5331 // control input of nodes for which there's only a single copy. 5332 sort_expensive_nodes(); 5333 5334 int j = 0; 5335 int identical = 0; 5336 int i = 0; 5337 bool modified = false; 5338 for (; i < _expensive_nodes.length()-1; i++) { 5339 assert(j <= i, "can't write beyond current index"); 5340 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 5341 identical++; 5342 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 5343 continue; 5344 } 5345 if (identical > 0) { 5346 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 5347 identical = 0; 5348 } else { 5349 Node* n = _expensive_nodes.at(i); 5350 igvn.replace_input_of(n, 0, nullptr); 5351 igvn.hash_insert(n); 5352 modified = true; 5353 } 5354 } 5355 if (identical > 0) { 5356 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 5357 } else if (_expensive_nodes.length() >= 1) { 5358 Node* n = _expensive_nodes.at(i); 5359 igvn.replace_input_of(n, 0, nullptr); 5360 igvn.hash_insert(n); 5361 modified = true; 5362 } 5363 _expensive_nodes.trunc_to(j); 5364 if (modified) { 5365 igvn.optimize(); 5366 } 5367 } 5368 5369 void Compile::add_expensive_node(Node * n) { 5370 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 5371 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 5372 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 5373 if (OptimizeExpensiveOps) { 5374 _expensive_nodes.append(n); 5375 } else { 5376 // Clear control input and let IGVN optimize expensive nodes if 5377 // OptimizeExpensiveOps is off. 5378 n->set_req(0, nullptr); 5379 } 5380 } 5381 5382 /** 5383 * Track coarsened Lock and Unlock nodes. 5384 */ 5385 5386 class Lock_List : public Node_List { 5387 uint _origin_cnt; 5388 public: 5389 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 5390 uint origin_cnt() const { return _origin_cnt; } 5391 }; 5392 5393 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 5394 int length = locks.length(); 5395 if (length > 0) { 5396 // Have to keep this list until locks elimination during Macro nodes elimination. 5397 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 5398 for (int i = 0; i < length; i++) { 5399 AbstractLockNode* lock = locks.at(i); 5400 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 5401 locks_list->push(lock); 5402 } 5403 _coarsened_locks.append(locks_list); 5404 } 5405 } 5406 5407 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 5408 int count = coarsened_count(); 5409 for (int i = 0; i < count; i++) { 5410 Node_List* locks_list = _coarsened_locks.at(i); 5411 for (uint j = 0; j < locks_list->size(); j++) { 5412 Node* lock = locks_list->at(j); 5413 assert(lock->is_AbstractLock(), "sanity"); 5414 if (!useful.member(lock)) { 5415 locks_list->yank(lock); 5416 } 5417 } 5418 } 5419 } 5420 5421 void Compile::remove_coarsened_lock(Node* n) { 5422 if (n->is_AbstractLock()) { 5423 int count = coarsened_count(); 5424 for (int i = 0; i < count; i++) { 5425 Node_List* locks_list = _coarsened_locks.at(i); 5426 locks_list->yank(n); 5427 } 5428 } 5429 } 5430 5431 bool Compile::coarsened_locks_consistent() { 5432 int count = coarsened_count(); 5433 for (int i = 0; i < count; i++) { 5434 bool unbalanced = false; 5435 bool modified = false; // track locks kind modifications 5436 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 5437 uint size = locks_list->size(); 5438 if (size == 0) { 5439 unbalanced = false; // All locks were eliminated - good 5440 } else if (size != locks_list->origin_cnt()) { 5441 unbalanced = true; // Some locks were removed from list 5442 } else { 5443 for (uint j = 0; j < size; j++) { 5444 Node* lock = locks_list->at(j); 5445 // All nodes in group should have the same state (modified or not) 5446 if (!lock->as_AbstractLock()->is_coarsened()) { 5447 if (j == 0) { 5448 // first on list was modified, the rest should be too for consistency 5449 modified = true; 5450 } else if (!modified) { 5451 // this lock was modified but previous locks on the list were not 5452 unbalanced = true; 5453 break; 5454 } 5455 } else if (modified) { 5456 // previous locks on list were modified but not this lock 5457 unbalanced = true; 5458 break; 5459 } 5460 } 5461 } 5462 if (unbalanced) { 5463 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 5464 #ifdef ASSERT 5465 if (PrintEliminateLocks) { 5466 tty->print_cr("=== unbalanced coarsened locks ==="); 5467 for (uint l = 0; l < size; l++) { 5468 locks_list->at(l)->dump(); 5469 } 5470 } 5471 #endif 5472 record_failure(C2Compiler::retry_no_locks_coarsening()); 5473 return false; 5474 } 5475 } 5476 return true; 5477 } 5478 5479 /** 5480 * Remove the speculative part of types and clean up the graph 5481 */ 5482 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 5483 if (UseTypeSpeculation) { 5484 Unique_Node_List worklist; 5485 worklist.push(root()); 5486 int modified = 0; 5487 // Go over all type nodes that carry a speculative type, drop the 5488 // speculative part of the type and enqueue the node for an igvn 5489 // which may optimize it out. 5490 for (uint next = 0; next < worklist.size(); ++next) { 5491 Node *n = worklist.at(next); 5492 if (n->is_Type()) { 5493 TypeNode* tn = n->as_Type(); 5494 const Type* t = tn->type(); 5495 const Type* t_no_spec = t->remove_speculative(); 5496 if (t_no_spec != t) { 5497 bool in_hash = igvn.hash_delete(n); 5498 #ifdef ASSERT 5499 if (!in_hash) { 5500 tty->print_cr("current graph:"); 5501 n->dump_bfs(MaxNodeLimit, nullptr, "S$"); 5502 tty->cr(); 5503 tty->print_cr("erroneous node:"); 5504 n->dump(); 5505 assert(false, "node should be in igvn hash table"); 5506 } 5507 #endif 5508 tn->set_type(t_no_spec); 5509 igvn.hash_insert(n); 5510 igvn._worklist.push(n); // give it a chance to go away 5511 modified++; 5512 } 5513 } 5514 // Iterate over outs - endless loops is unreachable from below 5515 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 5516 Node *m = n->fast_out(i); 5517 if (not_a_node(m)) { 5518 continue; 5519 } 5520 worklist.push(m); 5521 } 5522 } 5523 // Drop the speculative part of all types in the igvn's type table 5524 igvn.remove_speculative_types(); 5525 if (modified > 0) { 5526 igvn.optimize(); 5527 if (failing()) return; 5528 } 5529 #ifdef ASSERT 5530 // Verify that after the IGVN is over no speculative type has resurfaced 5531 worklist.clear(); 5532 worklist.push(root()); 5533 for (uint next = 0; next < worklist.size(); ++next) { 5534 Node *n = worklist.at(next); 5535 const Type* t = igvn.type_or_null(n); 5536 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 5537 if (n->is_Type()) { 5538 t = n->as_Type()->type(); 5539 assert(t == t->remove_speculative(), "no more speculative types"); 5540 } 5541 // Iterate over outs - endless loops is unreachable from below 5542 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 5543 Node *m = n->fast_out(i); 5544 if (not_a_node(m)) { 5545 continue; 5546 } 5547 worklist.push(m); 5548 } 5549 } 5550 igvn.check_no_speculative_types(); 5551 #endif 5552 } 5553 } 5554 5555 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 5556 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 5557 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 5558 if (!EnableValhalla || ta == nullptr || tb == nullptr || 5559 ta->is_zero_type() || tb->is_zero_type() || 5560 !ta->can_be_inline_type() || !tb->can_be_inline_type()) { 5561 // Use old acmp if one operand is null or not an inline type 5562 return new CmpPNode(a, b); 5563 } else if (ta->is_inlinetypeptr() || tb->is_inlinetypeptr()) { 5564 // We know that one operand is an inline type. Therefore, 5565 // new acmp will only return true if both operands are nullptr. 5566 // Check if both operands are null by or'ing the oops. 5567 a = phase->transform(new CastP2XNode(nullptr, a)); 5568 b = phase->transform(new CastP2XNode(nullptr, b)); 5569 a = phase->transform(new OrXNode(a, b)); 5570 return new CmpXNode(a, phase->MakeConX(0)); 5571 } 5572 // Use new acmp 5573 return nullptr; 5574 } 5575 5576 // Auxiliary methods to support randomized stressing/fuzzing. 5577 5578 int Compile::random() { 5579 _stress_seed = os::next_random(_stress_seed); 5580 return static_cast<int>(_stress_seed); 5581 } 5582 5583 // This method can be called the arbitrary number of times, with current count 5584 // as the argument. The logic allows selecting a single candidate from the 5585 // running list of candidates as follows: 5586 // int count = 0; 5587 // Cand* selected = null; 5588 // while(cand = cand->next()) { 5589 // if (randomized_select(++count)) { 5590 // selected = cand; 5591 // } 5592 // } 5593 // 5594 // Including count equalizes the chances any candidate is "selected". 5595 // This is useful when we don't have the complete list of candidates to choose 5596 // from uniformly. In this case, we need to adjust the randomicity of the 5597 // selection, or else we will end up biasing the selection towards the latter 5598 // candidates. 5599 // 5600 // Quick back-envelope calculation shows that for the list of n candidates 5601 // the equal probability for the candidate to persist as "best" can be 5602 // achieved by replacing it with "next" k-th candidate with the probability 5603 // of 1/k. It can be easily shown that by the end of the run, the 5604 // probability for any candidate is converged to 1/n, thus giving the 5605 // uniform distribution among all the candidates. 5606 // 5607 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5608 #define RANDOMIZED_DOMAIN_POW 29 5609 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5610 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5611 bool Compile::randomized_select(int count) { 5612 assert(count > 0, "only positive"); 5613 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5614 } 5615 5616 CloneMap& Compile::clone_map() { return _clone_map; } 5617 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5618 5619 void NodeCloneInfo::dump_on(outputStream* st) const { 5620 st->print(" {%d:%d} ", idx(), gen()); 5621 } 5622 5623 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5624 uint64_t val = value(old->_idx); 5625 NodeCloneInfo cio(val); 5626 assert(val != 0, "old node should be in the map"); 5627 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5628 insert(nnn->_idx, cin.get()); 5629 #ifndef PRODUCT 5630 if (is_debug()) { 5631 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5632 } 5633 #endif 5634 } 5635 5636 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5637 NodeCloneInfo cio(value(old->_idx)); 5638 if (cio.get() == 0) { 5639 cio.set(old->_idx, 0); 5640 insert(old->_idx, cio.get()); 5641 #ifndef PRODUCT 5642 if (is_debug()) { 5643 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5644 } 5645 #endif 5646 } 5647 clone(old, nnn, gen); 5648 } 5649 5650 int CloneMap::max_gen() const { 5651 int g = 0; 5652 DictI di(_dict); 5653 for(; di.test(); ++di) { 5654 int t = gen(di._key); 5655 if (g < t) { 5656 g = t; 5657 #ifndef PRODUCT 5658 if (is_debug()) { 5659 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5660 } 5661 #endif 5662 } 5663 } 5664 return g; 5665 } 5666 5667 void CloneMap::dump(node_idx_t key, outputStream* st) const { 5668 uint64_t val = value(key); 5669 if (val != 0) { 5670 NodeCloneInfo ni(val); 5671 ni.dump_on(st); 5672 } 5673 } 5674 5675 void Compile::shuffle_macro_nodes() { 5676 if (_macro_nodes.length() < 2) { 5677 return; 5678 } 5679 for (uint i = _macro_nodes.length() - 1; i >= 1; i--) { 5680 uint j = C->random() % (i + 1); 5681 swap(_macro_nodes.at(i), _macro_nodes.at(j)); 5682 } 5683 } 5684 5685 // Move Allocate nodes to the start of the list 5686 void Compile::sort_macro_nodes() { 5687 int count = macro_count(); 5688 int allocates = 0; 5689 for (int i = 0; i < count; i++) { 5690 Node* n = macro_node(i); 5691 if (n->is_Allocate()) { 5692 if (i != allocates) { 5693 Node* tmp = macro_node(allocates); 5694 _macro_nodes.at_put(allocates, n); 5695 _macro_nodes.at_put(i, tmp); 5696 } 5697 allocates++; 5698 } 5699 } 5700 } 5701 5702 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5703 if (failing()) { return; } 5704 EventCompilerPhase event(UNTIMED); 5705 if (event.should_commit()) { 5706 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5707 } 5708 #ifndef PRODUCT 5709 ResourceMark rm; 5710 stringStream ss; 5711 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5712 int iter = ++_igv_phase_iter[cpt]; 5713 if (iter > 1) { 5714 ss.print(" %d", iter); 5715 } 5716 if (n != nullptr) { 5717 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5718 } 5719 5720 const char* name = ss.as_string(); 5721 if (should_print_igv(level)) { 5722 _igv_printer->print_method(name, level); 5723 } 5724 if (should_print_phase(cpt)) { 5725 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5726 } 5727 #endif 5728 C->_latest_stage_start_counter.stamp(); 5729 } 5730 5731 // Only used from CompileWrapper 5732 void Compile::begin_method() { 5733 #ifndef PRODUCT 5734 if (_method != nullptr && should_print_igv(1)) { 5735 _igv_printer->begin_method(); 5736 } 5737 #endif 5738 C->_latest_stage_start_counter.stamp(); 5739 } 5740 5741 // Only used from CompileWrapper 5742 void Compile::end_method() { 5743 EventCompilerPhase event(UNTIMED); 5744 if (event.should_commit()) { 5745 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5746 } 5747 5748 #ifndef PRODUCT 5749 if (_method != nullptr && should_print_igv(1)) { 5750 _igv_printer->end_method(); 5751 } 5752 #endif 5753 } 5754 5755 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5756 #ifndef PRODUCT 5757 if (_directive->should_print_phase(cpt)) { 5758 return true; 5759 } 5760 #endif 5761 return false; 5762 } 5763 5764 bool Compile::should_print_igv(const int level) { 5765 #ifndef PRODUCT 5766 if (PrintIdealGraphLevel < 0) { // disabled by the user 5767 return false; 5768 } 5769 5770 bool need = directive()->IGVPrintLevelOption >= level; 5771 if (need && !_igv_printer) { 5772 _igv_printer = IdealGraphPrinter::printer(); 5773 _igv_printer->set_compile(this); 5774 } 5775 return need; 5776 #else 5777 return false; 5778 #endif 5779 } 5780 5781 #ifndef PRODUCT 5782 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5783 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5784 5785 // Called from debugger. Prints method to the default file with the default phase name. 5786 // This works regardless of any Ideal Graph Visualizer flags set or not. 5787 void igv_print() { 5788 Compile::current()->igv_print_method_to_file(); 5789 } 5790 5791 // Same as igv_print() above but with a specified phase name. 5792 void igv_print(const char* phase_name) { 5793 Compile::current()->igv_print_method_to_file(phase_name); 5794 } 5795 5796 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5797 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5798 // This works regardless of any Ideal Graph Visualizer flags set or not. 5799 void igv_print(bool network) { 5800 if (network) { 5801 Compile::current()->igv_print_method_to_network(); 5802 } else { 5803 Compile::current()->igv_print_method_to_file(); 5804 } 5805 } 5806 5807 // Same as igv_print(bool network) above but with a specified phase name. 5808 void igv_print(bool network, const char* phase_name) { 5809 if (network) { 5810 Compile::current()->igv_print_method_to_network(phase_name); 5811 } else { 5812 Compile::current()->igv_print_method_to_file(phase_name); 5813 } 5814 } 5815 5816 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5817 void igv_print_default() { 5818 Compile::current()->print_method(PHASE_DEBUG, 0); 5819 } 5820 5821 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5822 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5823 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5824 void igv_append() { 5825 Compile::current()->igv_print_method_to_file("Debug", true); 5826 } 5827 5828 // Same as igv_append() above but with a specified phase name. 5829 void igv_append(const char* phase_name) { 5830 Compile::current()->igv_print_method_to_file(phase_name, true); 5831 } 5832 5833 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5834 const char* file_name = "custom_debug.xml"; 5835 if (_debug_file_printer == nullptr) { 5836 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5837 } else { 5838 _debug_file_printer->update_compiled_method(C->method()); 5839 } 5840 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5841 _debug_file_printer->print(phase_name, (Node*)C->root()); 5842 } 5843 5844 void Compile::igv_print_method_to_network(const char* phase_name) { 5845 if (_debug_network_printer == nullptr) { 5846 _debug_network_printer = new IdealGraphPrinter(C); 5847 } else { 5848 _debug_network_printer->update_compiled_method(C->method()); 5849 } 5850 tty->print_cr("Method printed over network stream to IGV"); 5851 _debug_network_printer->print(phase_name, (Node*)C->root()); 5852 } 5853 #endif 5854 5855 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5856 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5857 return value; 5858 } 5859 Node* result = nullptr; 5860 if (bt == T_BYTE) { 5861 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5862 result = new RShiftINode(result, phase->intcon(24)); 5863 } else if (bt == T_BOOLEAN) { 5864 result = new AndINode(value, phase->intcon(0xFF)); 5865 } else if (bt == T_CHAR) { 5866 result = new AndINode(value,phase->intcon(0xFFFF)); 5867 } else { 5868 assert(bt == T_SHORT, "unexpected narrow type"); 5869 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5870 result = new RShiftINode(result, phase->intcon(16)); 5871 } 5872 if (transform_res) { 5873 result = phase->transform(result); 5874 } 5875 return result; 5876 } 5877 5878 void Compile::record_method_not_compilable_oom() { 5879 record_method_not_compilable(CompilationMemoryStatistic::failure_reason_memlimit()); 5880 }