1 /* 2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/javaClasses.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "jfr/jfrEvents.hpp" 39 #include "jvm_io.h" 40 #include "memory/allocation.hpp" 41 #include "memory/resourceArea.hpp" 42 #include "opto/addnode.hpp" 43 #include "opto/block.hpp" 44 #include "opto/c2compiler.hpp" 45 #include "opto/callGenerator.hpp" 46 #include "opto/callnode.hpp" 47 #include "opto/castnode.hpp" 48 #include "opto/cfgnode.hpp" 49 #include "opto/chaitin.hpp" 50 #include "opto/compile.hpp" 51 #include "opto/connode.hpp" 52 #include "opto/convertnode.hpp" 53 #include "opto/divnode.hpp" 54 #include "opto/escape.hpp" 55 #include "opto/idealGraphPrinter.hpp" 56 #include "opto/loopnode.hpp" 57 #include "opto/machnode.hpp" 58 #include "opto/macro.hpp" 59 #include "opto/matcher.hpp" 60 #include "opto/mathexactnode.hpp" 61 #include "opto/memnode.hpp" 62 #include "opto/mulnode.hpp" 63 #include "opto/narrowptrnode.hpp" 64 #include "opto/node.hpp" 65 #include "opto/opcodes.hpp" 66 #include "opto/output.hpp" 67 #include "opto/parse.hpp" 68 #include "opto/phaseX.hpp" 69 #include "opto/rootnode.hpp" 70 #include "opto/runtime.hpp" 71 #include "opto/stringopts.hpp" 72 #include "opto/type.hpp" 73 #include "opto/vector.hpp" 74 #include "opto/vectornode.hpp" 75 #include "runtime/globals_extension.hpp" 76 #include "runtime/sharedRuntime.hpp" 77 #include "runtime/signature.hpp" 78 #include "runtime/stubRoutines.hpp" 79 #include "runtime/timer.hpp" 80 #include "utilities/align.hpp" 81 #include "utilities/copy.hpp" 82 #include "utilities/macros.hpp" 83 #include "utilities/resourceHash.hpp" 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == nullptr) { 89 _mach_constant_base_node = new MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 class IntrinsicDescPair { 101 private: 102 ciMethod* _m; 103 bool _is_virtual; 104 public: 105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 107 ciMethod* m= elt->method(); 108 ciMethod* key_m = key->_m; 109 if (key_m < m) return -1; 110 else if (key_m > m) return 1; 111 else { 112 bool is_virtual = elt->is_virtual(); 113 bool key_virtual = key->_is_virtual; 114 if (key_virtual < is_virtual) return -1; 115 else if (key_virtual > is_virtual) return 1; 116 else return 0; 117 } 118 } 119 }; 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 121 #ifdef ASSERT 122 for (int i = 1; i < _intrinsics.length(); i++) { 123 CallGenerator* cg1 = _intrinsics.at(i-1); 124 CallGenerator* cg2 = _intrinsics.at(i); 125 assert(cg1->method() != cg2->method() 126 ? cg1->method() < cg2->method() 127 : cg1->is_virtual() < cg2->is_virtual(), 128 "compiler intrinsics list must stay sorted"); 129 } 130 #endif 131 IntrinsicDescPair pair(m, is_virtual); 132 return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 bool found = false; 137 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 138 assert(!found, "registering twice"); 139 _intrinsics.insert_before(index, cg); 140 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 141 } 142 143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 144 assert(m->is_loaded(), "don't try this on unloaded methods"); 145 if (_intrinsics.length() > 0) { 146 bool found = false; 147 int index = intrinsic_insertion_index(m, is_virtual, found); 148 if (found) { 149 return _intrinsics.at(index); 150 } 151 } 152 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 153 if (m->intrinsic_id() != vmIntrinsics::_none && 154 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 155 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 156 if (cg != nullptr) { 157 // Save it for next time: 158 register_intrinsic(cg); 159 return cg; 160 } else { 161 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 162 } 163 } 164 return nullptr; 165 } 166 167 // Compile::make_vm_intrinsic is defined in library_call.cpp. 168 169 #ifndef PRODUCT 170 // statistics gathering... 171 172 juint Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0}; 173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0}; 174 175 inline int as_int(vmIntrinsics::ID id) { 176 return vmIntrinsics::as_int(id); 177 } 178 179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 180 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 181 int oflags = _intrinsic_hist_flags[as_int(id)]; 182 assert(flags != 0, "what happened?"); 183 if (is_virtual) { 184 flags |= _intrinsic_virtual; 185 } 186 bool changed = (flags != oflags); 187 if ((flags & _intrinsic_worked) != 0) { 188 juint count = (_intrinsic_hist_count[as_int(id)] += 1); 189 if (count == 1) { 190 changed = true; // first time 191 } 192 // increment the overall count also: 193 _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1; 194 } 195 if (changed) { 196 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 197 // Something changed about the intrinsic's virtuality. 198 if ((flags & _intrinsic_virtual) != 0) { 199 // This is the first use of this intrinsic as a virtual call. 200 if (oflags != 0) { 201 // We already saw it as a non-virtual, so note both cases. 202 flags |= _intrinsic_both; 203 } 204 } else if ((oflags & _intrinsic_both) == 0) { 205 // This is the first use of this intrinsic as a non-virtual 206 flags |= _intrinsic_both; 207 } 208 } 209 _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags); 210 } 211 // update the overall flags also: 212 _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags; 213 return changed; 214 } 215 216 static char* format_flags(int flags, char* buf) { 217 buf[0] = 0; 218 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 219 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 220 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 221 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 222 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 223 if (buf[0] == 0) strcat(buf, ","); 224 assert(buf[0] == ',', "must be"); 225 return &buf[1]; 226 } 227 228 void Compile::print_intrinsic_statistics() { 229 char flagsbuf[100]; 230 ttyLocker ttyl; 231 if (xtty != nullptr) xtty->head("statistics type='intrinsic'"); 232 tty->print_cr("Compiler intrinsic usage:"); 233 juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)]; 234 if (total == 0) total = 1; // avoid div0 in case of no successes 235 #define PRINT_STAT_LINE(name, c, f) \ 236 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 237 for (auto id : EnumRange<vmIntrinsicID>{}) { 238 int flags = _intrinsic_hist_flags[as_int(id)]; 239 juint count = _intrinsic_hist_count[as_int(id)]; 240 if ((flags | count) != 0) { 241 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 242 } 243 } 244 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf)); 245 if (xtty != nullptr) xtty->tail("statistics"); 246 } 247 248 void Compile::print_statistics() { 249 { ttyLocker ttyl; 250 if (xtty != nullptr) xtty->head("statistics type='opto'"); 251 Parse::print_statistics(); 252 PhaseStringOpts::print_statistics(); 253 PhaseCCP::print_statistics(); 254 PhaseRegAlloc::print_statistics(); 255 PhaseOutput::print_statistics(); 256 PhasePeephole::print_statistics(); 257 PhaseIdealLoop::print_statistics(); 258 ConnectionGraph::print_statistics(); 259 PhaseMacroExpand::print_statistics(); 260 if (xtty != nullptr) xtty->tail("statistics"); 261 } 262 if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) { 263 // put this under its own <statistics> element. 264 print_intrinsic_statistics(); 265 } 266 } 267 #endif //PRODUCT 268 269 void Compile::gvn_replace_by(Node* n, Node* nn) { 270 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 271 Node* use = n->last_out(i); 272 bool is_in_table = initial_gvn()->hash_delete(use); 273 uint uses_found = 0; 274 for (uint j = 0; j < use->len(); j++) { 275 if (use->in(j) == n) { 276 if (j < use->req()) 277 use->set_req(j, nn); 278 else 279 use->set_prec(j, nn); 280 uses_found++; 281 } 282 } 283 if (is_in_table) { 284 // reinsert into table 285 initial_gvn()->hash_find_insert(use); 286 } 287 record_for_igvn(use); 288 i -= uses_found; // we deleted 1 or more copies of this edge 289 } 290 } 291 292 293 // Identify all nodes that are reachable from below, useful. 294 // Use breadth-first pass that records state in a Unique_Node_List, 295 // recursive traversal is slower. 296 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 297 int estimated_worklist_size = live_nodes(); 298 useful.map( estimated_worklist_size, nullptr ); // preallocate space 299 300 // Initialize worklist 301 if (root() != nullptr) { useful.push(root()); } 302 // If 'top' is cached, declare it useful to preserve cached node 303 if (cached_top_node()) { useful.push(cached_top_node()); } 304 305 // Push all useful nodes onto the list, breadthfirst 306 for( uint next = 0; next < useful.size(); ++next ) { 307 assert( next < unique(), "Unique useful nodes < total nodes"); 308 Node *n = useful.at(next); 309 uint max = n->len(); 310 for( uint i = 0; i < max; ++i ) { 311 Node *m = n->in(i); 312 if (not_a_node(m)) continue; 313 useful.push(m); 314 } 315 } 316 } 317 318 // Update dead_node_list with any missing dead nodes using useful 319 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 320 void Compile::update_dead_node_list(Unique_Node_List &useful) { 321 uint max_idx = unique(); 322 VectorSet& useful_node_set = useful.member_set(); 323 324 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 325 // If node with index node_idx is not in useful set, 326 // mark it as dead in dead node list. 327 if (!useful_node_set.test(node_idx)) { 328 record_dead_node(node_idx); 329 } 330 } 331 } 332 333 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 334 int shift = 0; 335 for (int i = 0; i < inlines->length(); i++) { 336 CallGenerator* cg = inlines->at(i); 337 if (useful.member(cg->call_node())) { 338 if (shift > 0) { 339 inlines->at_put(i - shift, cg); 340 } 341 } else { 342 shift++; // skip over the dead element 343 } 344 } 345 if (shift > 0) { 346 inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array 347 } 348 } 349 350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) { 351 assert(dead != nullptr && dead->is_Call(), "sanity"); 352 int found = 0; 353 for (int i = 0; i < inlines->length(); i++) { 354 if (inlines->at(i)->call_node() == dead) { 355 inlines->remove_at(i); 356 found++; 357 NOT_DEBUG( break; ) // elements are unique, so exit early 358 } 359 } 360 assert(found <= 1, "not unique"); 361 } 362 363 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) { 364 for (int i = node_list.length() - 1; i >= 0; i--) { 365 Node* n = node_list.at(i); 366 if (!useful.member(n)) { 367 node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 368 } 369 } 370 } 371 372 void Compile::remove_useless_node(Node* dead) { 373 remove_modified_node(dead); 374 375 // Constant node that has no out-edges and has only one in-edge from 376 // root is usually dead. However, sometimes reshaping walk makes 377 // it reachable by adding use edges. So, we will NOT count Con nodes 378 // as dead to be conservative about the dead node count at any 379 // given time. 380 if (!dead->is_Con()) { 381 record_dead_node(dead->_idx); 382 } 383 if (dead->is_macro()) { 384 remove_macro_node(dead); 385 } 386 if (dead->is_expensive()) { 387 remove_expensive_node(dead); 388 } 389 if (dead->Opcode() == Op_Opaque4) { 390 remove_skeleton_predicate_opaq(dead); 391 } 392 if (dead->for_post_loop_opts_igvn()) { 393 remove_from_post_loop_opts_igvn(dead); 394 } 395 if (dead->is_Call()) { 396 remove_useless_late_inlines( &_late_inlines, dead); 397 remove_useless_late_inlines( &_string_late_inlines, dead); 398 remove_useless_late_inlines( &_boxing_late_inlines, dead); 399 remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead); 400 401 if (dead->is_CallStaticJava()) { 402 remove_unstable_if_trap(dead->as_CallStaticJava(), false); 403 } 404 } 405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 406 bs->unregister_potential_barrier_node(dead); 407 } 408 409 // Disconnect all useless nodes by disconnecting those at the boundary. 410 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) { 411 uint next = 0; 412 while (next < useful.size()) { 413 Node *n = useful.at(next++); 414 if (n->is_SafePoint()) { 415 // We're done with a parsing phase. Replaced nodes are not valid 416 // beyond that point. 417 n->as_SafePoint()->delete_replaced_nodes(); 418 } 419 // Use raw traversal of out edges since this code removes out edges 420 int max = n->outcnt(); 421 for (int j = 0; j < max; ++j) { 422 Node* child = n->raw_out(j); 423 if (!useful.member(child)) { 424 assert(!child->is_top() || child != top(), 425 "If top is cached in Compile object it is in useful list"); 426 // Only need to remove this out-edge to the useless node 427 n->raw_del_out(j); 428 --j; 429 --max; 430 } 431 } 432 if (n->outcnt() == 1 && n->has_special_unique_user()) { 433 worklist->push(n->unique_out()); 434 } 435 } 436 437 remove_useless_nodes(_macro_nodes, useful); // remove useless macro nodes 438 remove_useless_nodes(_predicate_opaqs, useful); // remove useless predicate opaque nodes 439 remove_useless_nodes(_skeleton_predicate_opaqs, useful); 440 remove_useless_nodes(_expensive_nodes, useful); // remove useless expensive nodes 441 remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass 442 remove_useless_unstable_if_traps(useful); // remove useless unstable_if traps 443 remove_useless_coarsened_locks(useful); // remove useless coarsened locks nodes 444 #ifdef ASSERT 445 if (_modified_nodes != nullptr) { 446 _modified_nodes->remove_useless_nodes(useful.member_set()); 447 } 448 #endif 449 450 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 451 bs->eliminate_useless_gc_barriers(useful, this); 452 // clean up the late inline lists 453 remove_useless_late_inlines( &_late_inlines, useful); 454 remove_useless_late_inlines( &_string_late_inlines, useful); 455 remove_useless_late_inlines( &_boxing_late_inlines, useful); 456 remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful); 457 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 458 } 459 460 // ============================================================================ 461 //------------------------------CompileWrapper--------------------------------- 462 class CompileWrapper : public StackObj { 463 Compile *const _compile; 464 public: 465 CompileWrapper(Compile* compile); 466 467 ~CompileWrapper(); 468 }; 469 470 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 471 // the Compile* pointer is stored in the current ciEnv: 472 ciEnv* env = compile->env(); 473 assert(env == ciEnv::current(), "must already be a ciEnv active"); 474 assert(env->compiler_data() == nullptr, "compile already active?"); 475 env->set_compiler_data(compile); 476 assert(compile == Compile::current(), "sanity"); 477 478 compile->set_type_dict(nullptr); 479 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 480 compile->clone_map().set_clone_idx(0); 481 compile->set_type_last_size(0); 482 compile->set_last_tf(nullptr, nullptr); 483 compile->set_indexSet_arena(nullptr); 484 compile->set_indexSet_free_block_list(nullptr); 485 compile->init_type_arena(); 486 Type::Initialize(compile); 487 _compile->begin_method(); 488 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 489 } 490 CompileWrapper::~CompileWrapper() { 491 // simulate crash during compilation 492 assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned"); 493 494 _compile->end_method(); 495 _compile->env()->set_compiler_data(nullptr); 496 } 497 498 499 //----------------------------print_compile_messages--------------------------- 500 void Compile::print_compile_messages() { 501 #ifndef PRODUCT 502 // Check if recompiling 503 if (!subsume_loads() && PrintOpto) { 504 // Recompiling without allowing machine instructions to subsume loads 505 tty->print_cr("*********************************************************"); 506 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 507 tty->print_cr("*********************************************************"); 508 } 509 if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) { 510 // Recompiling without escape analysis 511 tty->print_cr("*********************************************************"); 512 tty->print_cr("** Bailout: Recompile without escape analysis **"); 513 tty->print_cr("*********************************************************"); 514 } 515 if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) { 516 // Recompiling without iterative escape analysis 517 tty->print_cr("*********************************************************"); 518 tty->print_cr("** Bailout: Recompile without iterative escape analysis**"); 519 tty->print_cr("*********************************************************"); 520 } 521 if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) { 522 // Recompiling without boxing elimination 523 tty->print_cr("*********************************************************"); 524 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 525 tty->print_cr("*********************************************************"); 526 } 527 if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) { 528 // Recompiling without locks coarsening 529 tty->print_cr("*********************************************************"); 530 tty->print_cr("** Bailout: Recompile without locks coarsening **"); 531 tty->print_cr("*********************************************************"); 532 } 533 if (env()->break_at_compile()) { 534 // Open the debugger when compiling this method. 535 tty->print("### Breaking when compiling: "); 536 method()->print_short_name(); 537 tty->cr(); 538 BREAKPOINT; 539 } 540 541 if( PrintOpto ) { 542 if (is_osr_compilation()) { 543 tty->print("[OSR]%3d", _compile_id); 544 } else { 545 tty->print("%3d", _compile_id); 546 } 547 } 548 #endif 549 } 550 551 #ifndef PRODUCT 552 void Compile::print_ideal_ir(const char* phase_name) { 553 ttyLocker ttyl; 554 // keep the following output all in one block 555 // This output goes directly to the tty, not the compiler log. 556 // To enable tools to match it up with the compilation activity, 557 // be sure to tag this tty output with the compile ID. 558 if (xtty != nullptr) { 559 xtty->head("ideal compile_id='%d'%s compile_phase='%s'", 560 compile_id(), 561 is_osr_compilation() ? " compile_kind='osr'" : "", 562 phase_name); 563 } 564 if (_output == nullptr) { 565 tty->print_cr("AFTER: %s", phase_name); 566 // Print out all nodes in ascending order of index. 567 root()->dump_bfs(MaxNodeLimit, nullptr, "+S$"); 568 } else { 569 // Dump the node blockwise if we have a scheduling 570 _output->print_scheduling(); 571 } 572 573 if (xtty != nullptr) { 574 xtty->tail("ideal"); 575 } 576 } 577 #endif 578 579 // ============================================================================ 580 //------------------------------Compile standard------------------------------- 581 debug_only( int Compile::_debug_idx = 100000; ) 582 583 // Compile a method. entry_bci is -1 for normal compilations and indicates 584 // the continuation bci for on stack replacement. 585 586 587 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci, 588 Options options, DirectiveSet* directive) 589 : Phase(Compiler), 590 _compile_id(ci_env->compile_id()), 591 _options(options), 592 _method(target), 593 _entry_bci(osr_bci), 594 _ilt(nullptr), 595 _stub_function(nullptr), 596 _stub_name(nullptr), 597 _stub_entry_point(nullptr), 598 _max_node_limit(MaxNodeLimit), 599 _post_loop_opts_phase(false), 600 _inlining_progress(false), 601 _inlining_incrementally(false), 602 _do_cleanup(false), 603 _has_reserved_stack_access(target->has_reserved_stack_access()), 604 #ifndef PRODUCT 605 _igv_idx(0), 606 _trace_opto_output(directive->TraceOptoOutputOption), 607 #endif 608 _has_method_handle_invokes(false), 609 _clinit_barrier_on_entry(false), 610 _stress_seed(0), 611 _comp_arena(mtCompiler), 612 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 613 _env(ci_env), 614 _directive(directive), 615 _log(ci_env->log()), 616 _failure_reason(nullptr), 617 _intrinsics (comp_arena(), 0, 0, nullptr), 618 _macro_nodes (comp_arena(), 8, 0, nullptr), 619 _predicate_opaqs (comp_arena(), 8, 0, nullptr), 620 _skeleton_predicate_opaqs (comp_arena(), 8, 0, nullptr), 621 _expensive_nodes (comp_arena(), 8, 0, nullptr), 622 _for_post_loop_igvn(comp_arena(), 8, 0, nullptr), 623 _unstable_if_traps (comp_arena(), 8, 0, nullptr), 624 _coarsened_locks (comp_arena(), 8, 0, nullptr), 625 _congraph(nullptr), 626 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 627 _dead_node_list(comp_arena()), 628 _dead_node_count(0), 629 _node_arena(mtCompiler), 630 _old_arena(mtCompiler), 631 _mach_constant_base_node(nullptr), 632 _Compile_types(mtCompiler), 633 _initial_gvn(nullptr), 634 _for_igvn(nullptr), 635 _late_inlines(comp_arena(), 2, 0, nullptr), 636 _string_late_inlines(comp_arena(), 2, 0, nullptr), 637 _boxing_late_inlines(comp_arena(), 2, 0, nullptr), 638 _vector_reboxing_late_inlines(comp_arena(), 2, 0, nullptr), 639 _late_inlines_pos(0), 640 _number_of_mh_late_inlines(0), 641 _print_inlining_stream(new (mtCompiler) stringStream()), 642 _print_inlining_list(nullptr), 643 _print_inlining_idx(0), 644 _print_inlining_output(nullptr), 645 _replay_inline_data(nullptr), 646 _java_calls(0), 647 _inner_loops(0), 648 _interpreter_frame_size(0), 649 _output(nullptr) 650 #ifndef PRODUCT 651 , _in_dump_cnt(0) 652 #endif 653 { 654 C = this; 655 CompileWrapper cw(this); 656 657 if (CITimeVerbose) { 658 tty->print(" "); 659 target->holder()->name()->print(); 660 tty->print("."); 661 target->print_short_name(); 662 tty->print(" "); 663 } 664 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 665 TraceTime t2(nullptr, &_t_methodCompilation, CITime, false); 666 667 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 668 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 669 // We can always print a disassembly, either abstract (hex dump) or 670 // with the help of a suitable hsdis library. Thus, we should not 671 // couple print_assembly and print_opto_assembly controls. 672 // But: always print opto and regular assembly on compile command 'print'. 673 bool print_assembly = directive->PrintAssemblyOption; 674 set_print_assembly(print_opto_assembly || print_assembly); 675 #else 676 set_print_assembly(false); // must initialize. 677 #endif 678 679 #ifndef PRODUCT 680 set_parsed_irreducible_loop(false); 681 682 if (directive->ReplayInlineOption) { 683 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 684 } 685 #endif 686 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 687 set_print_intrinsics(directive->PrintIntrinsicsOption); 688 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 689 690 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 691 // Make sure the method being compiled gets its own MDO, 692 // so we can at least track the decompile_count(). 693 // Need MDO to record RTM code generation state. 694 method()->ensure_method_data(); 695 } 696 697 Init(/*do_aliasing=*/ true); 698 699 print_compile_messages(); 700 701 _ilt = InlineTree::build_inline_tree_root(); 702 703 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 704 assert(num_alias_types() >= AliasIdxRaw, ""); 705 706 #define MINIMUM_NODE_HASH 1023 707 // Node list that Iterative GVN will start with 708 Unique_Node_List for_igvn(comp_arena()); 709 set_for_igvn(&for_igvn); 710 711 // GVN that will be run immediately on new nodes 712 uint estimated_size = method()->code_size()*4+64; 713 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 714 PhaseGVN gvn(node_arena(), estimated_size); 715 set_initial_gvn(&gvn); 716 717 print_inlining_init(); 718 { // Scope for timing the parser 719 TracePhase tp("parse", &timers[_t_parser]); 720 721 // Put top into the hash table ASAP. 722 initial_gvn()->transform_no_reclaim(top()); 723 724 // Set up tf(), start(), and find a CallGenerator. 725 CallGenerator* cg = nullptr; 726 if (is_osr_compilation()) { 727 const TypeTuple *domain = StartOSRNode::osr_domain(); 728 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 729 init_tf(TypeFunc::make(domain, range)); 730 StartNode* s = new StartOSRNode(root(), domain); 731 initial_gvn()->set_type_bottom(s); 732 init_start(s); 733 cg = CallGenerator::for_osr(method(), entry_bci()); 734 } else { 735 // Normal case. 736 init_tf(TypeFunc::make(method())); 737 StartNode* s = new StartNode(root(), tf()->domain()); 738 initial_gvn()->set_type_bottom(s); 739 init_start(s); 740 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 741 // With java.lang.ref.reference.get() we must go through the 742 // intrinsic - even when get() is the root 743 // method of the compile - so that, if necessary, the value in 744 // the referent field of the reference object gets recorded by 745 // the pre-barrier code. 746 cg = find_intrinsic(method(), false); 747 } 748 if (cg == nullptr) { 749 float past_uses = method()->interpreter_invocation_count(); 750 float expected_uses = past_uses; 751 cg = CallGenerator::for_inline(method(), expected_uses); 752 } 753 } 754 if (failing()) return; 755 if (cg == nullptr) { 756 const char* reason = InlineTree::check_can_parse(method()); 757 assert(reason != nullptr, "expect reason for parse failure"); 758 stringStream ss; 759 ss.print("cannot parse method: %s", reason); 760 record_method_not_compilable(ss.as_string()); 761 return; 762 } 763 764 gvn.set_type(root(), root()->bottom_type()); 765 766 JVMState* jvms = build_start_state(start(), tf()); 767 if ((jvms = cg->generate(jvms)) == nullptr) { 768 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 769 assert(failure_reason() != nullptr, "expect reason for parse failure"); 770 stringStream ss; 771 ss.print("method parse failed: %s", failure_reason()); 772 record_method_not_compilable(ss.as_string()); 773 } 774 return; 775 } 776 GraphKit kit(jvms); 777 778 if (!kit.stopped()) { 779 // Accept return values, and transfer control we know not where. 780 // This is done by a special, unique ReturnNode bound to root. 781 return_values(kit.jvms()); 782 } 783 784 if (kit.has_exceptions()) { 785 // Any exceptions that escape from this call must be rethrown 786 // to whatever caller is dynamically above us on the stack. 787 // This is done by a special, unique RethrowNode bound to root. 788 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 789 } 790 791 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 792 793 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 794 inline_string_calls(true); 795 } 796 797 if (failing()) return; 798 799 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 800 801 // Remove clutter produced by parsing. 802 if (!failing()) { 803 ResourceMark rm; 804 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 805 } 806 } 807 808 // Note: Large methods are capped off in do_one_bytecode(). 809 if (failing()) return; 810 811 // After parsing, node notes are no longer automagic. 812 // They must be propagated by register_new_node_with_optimizer(), 813 // clone(), or the like. 814 set_default_node_notes(nullptr); 815 816 #ifndef PRODUCT 817 if (should_print_igv(1)) { 818 _igv_printer->print_inlining(); 819 } 820 #endif 821 822 if (failing()) return; 823 NOT_PRODUCT( verify_graph_edges(); ) 824 825 // If any phase is randomized for stress testing, seed random number 826 // generation and log the seed for repeatability. 827 if (StressLCM || StressGCM || StressIGVN || StressCCP) { 828 if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) { 829 _stress_seed = static_cast<uint>(Ticks::now().nanoseconds()); 830 FLAG_SET_ERGO(StressSeed, _stress_seed); 831 } else { 832 _stress_seed = StressSeed; 833 } 834 if (_log != nullptr) { 835 _log->elem("stress_test seed='%u'", _stress_seed); 836 } 837 } 838 839 // Now optimize 840 Optimize(); 841 if (failing()) return; 842 NOT_PRODUCT( verify_graph_edges(); ) 843 844 #ifndef PRODUCT 845 if (should_print_ideal()) { 846 print_ideal_ir("print_ideal"); 847 } 848 #endif 849 850 #ifdef ASSERT 851 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 852 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 853 #endif 854 855 // Dump compilation data to replay it. 856 if (directive->DumpReplayOption) { 857 env()->dump_replay_data(_compile_id); 858 } 859 if (directive->DumpInlineOption && (ilt() != nullptr)) { 860 env()->dump_inline_data(_compile_id); 861 } 862 863 // Now that we know the size of all the monitors we can add a fixed slot 864 // for the original deopt pc. 865 int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size); 866 set_fixed_slots(next_slot); 867 868 // Compute when to use implicit null checks. Used by matching trap based 869 // nodes and NullCheck optimization. 870 set_allowed_deopt_reasons(); 871 872 // Now generate code 873 Code_Gen(); 874 } 875 876 //------------------------------Compile---------------------------------------- 877 // Compile a runtime stub 878 Compile::Compile( ciEnv* ci_env, 879 TypeFunc_generator generator, 880 address stub_function, 881 const char *stub_name, 882 int is_fancy_jump, 883 bool pass_tls, 884 bool return_pc, 885 DirectiveSet* directive) 886 : Phase(Compiler), 887 _compile_id(0), 888 _options(Options::for_runtime_stub()), 889 _method(nullptr), 890 _entry_bci(InvocationEntryBci), 891 _stub_function(stub_function), 892 _stub_name(stub_name), 893 _stub_entry_point(nullptr), 894 _max_node_limit(MaxNodeLimit), 895 _post_loop_opts_phase(false), 896 _inlining_progress(false), 897 _inlining_incrementally(false), 898 _has_reserved_stack_access(false), 899 #ifndef PRODUCT 900 _igv_idx(0), 901 _trace_opto_output(directive->TraceOptoOutputOption), 902 #endif 903 _has_method_handle_invokes(false), 904 _clinit_barrier_on_entry(false), 905 _stress_seed(0), 906 _comp_arena(mtCompiler), 907 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 908 _env(ci_env), 909 _directive(directive), 910 _log(ci_env->log()), 911 _failure_reason(nullptr), 912 _congraph(nullptr), 913 NOT_PRODUCT(_igv_printer(nullptr) COMMA) 914 _dead_node_list(comp_arena()), 915 _dead_node_count(0), 916 _node_arena(mtCompiler), 917 _old_arena(mtCompiler), 918 _mach_constant_base_node(nullptr), 919 _Compile_types(mtCompiler), 920 _initial_gvn(nullptr), 921 _for_igvn(nullptr), 922 _number_of_mh_late_inlines(0), 923 _print_inlining_stream(new (mtCompiler) stringStream()), 924 _print_inlining_list(nullptr), 925 _print_inlining_idx(0), 926 _print_inlining_output(nullptr), 927 _replay_inline_data(nullptr), 928 _java_calls(0), 929 _inner_loops(0), 930 _interpreter_frame_size(0), 931 _output(nullptr), 932 #ifndef PRODUCT 933 _in_dump_cnt(0), 934 #endif 935 _allowed_reasons(0) { 936 C = this; 937 938 TraceTime t1(nullptr, &_t_totalCompilation, CITime, false); 939 TraceTime t2(nullptr, &_t_stubCompilation, CITime, false); 940 941 #ifndef PRODUCT 942 set_print_assembly(PrintFrameConverterAssembly); 943 set_parsed_irreducible_loop(false); 944 #else 945 set_print_assembly(false); // Must initialize. 946 #endif 947 set_has_irreducible_loop(false); // no loops 948 949 CompileWrapper cw(this); 950 Init(/*do_aliasing=*/ false); 951 init_tf((*generator)()); 952 953 { 954 // The following is a dummy for the sake of GraphKit::gen_stub 955 Unique_Node_List for_igvn(comp_arena()); 956 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 957 PhaseGVN gvn(Thread::current()->resource_area(),255); 958 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 959 gvn.transform_no_reclaim(top()); 960 961 GraphKit kit; 962 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 963 } 964 965 NOT_PRODUCT( verify_graph_edges(); ) 966 967 Code_Gen(); 968 } 969 970 //------------------------------Init------------------------------------------- 971 // Prepare for a single compilation 972 void Compile::Init(bool aliasing) { 973 _do_aliasing = aliasing; 974 _unique = 0; 975 _regalloc = nullptr; 976 977 _tf = nullptr; // filled in later 978 _top = nullptr; // cached later 979 _matcher = nullptr; // filled in later 980 _cfg = nullptr; // filled in later 981 982 IA32_ONLY( set_24_bit_selection_and_mode(true, false); ) 983 984 _node_note_array = nullptr; 985 _default_node_notes = nullptr; 986 DEBUG_ONLY( _modified_nodes = nullptr; ) // Used in Optimize() 987 988 _immutable_memory = nullptr; // filled in at first inquiry 989 990 #ifdef ASSERT 991 _phase_optimize_finished = false; 992 _exception_backedge = false; 993 _type_verify = nullptr; 994 #endif 995 996 // Globally visible Nodes 997 // First set TOP to null to give safe behavior during creation of RootNode 998 set_cached_top_node(nullptr); 999 set_root(new RootNode()); 1000 // Now that you have a Root to point to, create the real TOP 1001 set_cached_top_node( new ConNode(Type::TOP) ); 1002 set_recent_alloc(nullptr, nullptr); 1003 1004 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1005 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1006 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1007 env()->set_dependencies(new Dependencies(env())); 1008 1009 _fixed_slots = 0; 1010 set_has_split_ifs(false); 1011 set_has_loops(false); // first approximation 1012 set_has_stringbuilder(false); 1013 set_has_boxed_value(false); 1014 _trap_can_recompile = false; // no traps emitted yet 1015 _major_progress = true; // start out assuming good things will happen 1016 set_has_unsafe_access(false); 1017 set_max_vector_size(0); 1018 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1019 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1020 set_decompile_count(0); 1021 1022 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1023 _loop_opts_cnt = LoopOptsCount; 1024 set_do_inlining(Inline); 1025 set_max_inline_size(MaxInlineSize); 1026 set_freq_inline_size(FreqInlineSize); 1027 set_do_scheduling(OptoScheduling); 1028 1029 set_do_vector_loop(false); 1030 set_has_monitors(false); 1031 1032 if (AllowVectorizeOnDemand) { 1033 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1034 set_do_vector_loop(true); 1035 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());}) 1036 } else if (has_method() && method()->name() != 0 && 1037 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1038 set_do_vector_loop(true); 1039 } 1040 } 1041 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1042 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());}) 1043 1044 set_rtm_state(NoRTM); // No RTM lock eliding by default 1045 _max_node_limit = _directive->MaxNodeLimitOption; 1046 1047 #if INCLUDE_RTM_OPT 1048 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != nullptr)) { 1049 int rtm_state = method()->method_data()->rtm_state(); 1050 if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) { 1051 // Don't generate RTM lock eliding code. 1052 set_rtm_state(NoRTM); 1053 } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1054 // Generate RTM lock eliding code without abort ratio calculation code. 1055 set_rtm_state(UseRTM); 1056 } else if (UseRTMDeopt) { 1057 // Generate RTM lock eliding code and include abort ratio calculation 1058 // code if UseRTMDeopt is on. 1059 set_rtm_state(ProfileRTM); 1060 } 1061 } 1062 #endif 1063 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1064 set_clinit_barrier_on_entry(true); 1065 } 1066 if (debug_info()->recording_non_safepoints()) { 1067 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1068 (comp_arena(), 8, 0, nullptr)); 1069 set_default_node_notes(Node_Notes::make(this)); 1070 } 1071 1072 const int grow_ats = 16; 1073 _max_alias_types = grow_ats; 1074 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1075 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1076 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1077 { 1078 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1079 } 1080 // Initialize the first few types. 1081 _alias_types[AliasIdxTop]->Init(AliasIdxTop, nullptr); 1082 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1083 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1084 _num_alias_types = AliasIdxRaw+1; 1085 // Zero out the alias type cache. 1086 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1087 // A null adr_type hits in the cache right away. Preload the right answer. 1088 probe_alias_cache(nullptr)->_index = AliasIdxTop; 1089 } 1090 1091 //---------------------------init_start---------------------------------------- 1092 // Install the StartNode on this compile object. 1093 void Compile::init_start(StartNode* s) { 1094 if (failing()) 1095 return; // already failing 1096 assert(s == start(), ""); 1097 } 1098 1099 /** 1100 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1101 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1102 * the ideal graph. 1103 */ 1104 StartNode* Compile::start() const { 1105 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1106 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1107 Node* start = root()->fast_out(i); 1108 if (start->is_Start()) { 1109 return start->as_Start(); 1110 } 1111 } 1112 fatal("Did not find Start node!"); 1113 return nullptr; 1114 } 1115 1116 //-------------------------------immutable_memory------------------------------------- 1117 // Access immutable memory 1118 Node* Compile::immutable_memory() { 1119 if (_immutable_memory != nullptr) { 1120 return _immutable_memory; 1121 } 1122 StartNode* s = start(); 1123 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1124 Node *p = s->fast_out(i); 1125 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1126 _immutable_memory = p; 1127 return _immutable_memory; 1128 } 1129 } 1130 ShouldNotReachHere(); 1131 return nullptr; 1132 } 1133 1134 //----------------------set_cached_top_node------------------------------------ 1135 // Install the cached top node, and make sure Node::is_top works correctly. 1136 void Compile::set_cached_top_node(Node* tn) { 1137 if (tn != nullptr) verify_top(tn); 1138 Node* old_top = _top; 1139 _top = tn; 1140 // Calling Node::setup_is_top allows the nodes the chance to adjust 1141 // their _out arrays. 1142 if (_top != nullptr) _top->setup_is_top(); 1143 if (old_top != nullptr) old_top->setup_is_top(); 1144 assert(_top == nullptr || top()->is_top(), ""); 1145 } 1146 1147 #ifdef ASSERT 1148 uint Compile::count_live_nodes_by_graph_walk() { 1149 Unique_Node_List useful(comp_arena()); 1150 // Get useful node list by walking the graph. 1151 identify_useful_nodes(useful); 1152 return useful.size(); 1153 } 1154 1155 void Compile::print_missing_nodes() { 1156 1157 // Return if CompileLog is null and PrintIdealNodeCount is false. 1158 if ((_log == nullptr) && (! PrintIdealNodeCount)) { 1159 return; 1160 } 1161 1162 // This is an expensive function. It is executed only when the user 1163 // specifies VerifyIdealNodeCount option or otherwise knows the 1164 // additional work that needs to be done to identify reachable nodes 1165 // by walking the flow graph and find the missing ones using 1166 // _dead_node_list. 1167 1168 Unique_Node_List useful(comp_arena()); 1169 // Get useful node list by walking the graph. 1170 identify_useful_nodes(useful); 1171 1172 uint l_nodes = C->live_nodes(); 1173 uint l_nodes_by_walk = useful.size(); 1174 1175 if (l_nodes != l_nodes_by_walk) { 1176 if (_log != nullptr) { 1177 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1178 _log->stamp(); 1179 _log->end_head(); 1180 } 1181 VectorSet& useful_member_set = useful.member_set(); 1182 int last_idx = l_nodes_by_walk; 1183 for (int i = 0; i < last_idx; i++) { 1184 if (useful_member_set.test(i)) { 1185 if (_dead_node_list.test(i)) { 1186 if (_log != nullptr) { 1187 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1188 } 1189 if (PrintIdealNodeCount) { 1190 // Print the log message to tty 1191 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1192 useful.at(i)->dump(); 1193 } 1194 } 1195 } 1196 else if (! _dead_node_list.test(i)) { 1197 if (_log != nullptr) { 1198 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1199 } 1200 if (PrintIdealNodeCount) { 1201 // Print the log message to tty 1202 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1203 } 1204 } 1205 } 1206 if (_log != nullptr) { 1207 _log->tail("mismatched_nodes"); 1208 } 1209 } 1210 } 1211 void Compile::record_modified_node(Node* n) { 1212 if (_modified_nodes != nullptr && !_inlining_incrementally && !n->is_Con()) { 1213 _modified_nodes->push(n); 1214 } 1215 } 1216 1217 void Compile::remove_modified_node(Node* n) { 1218 if (_modified_nodes != nullptr) { 1219 _modified_nodes->remove(n); 1220 } 1221 } 1222 #endif 1223 1224 #ifndef PRODUCT 1225 void Compile::verify_top(Node* tn) const { 1226 if (tn != nullptr) { 1227 assert(tn->is_Con(), "top node must be a constant"); 1228 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1229 assert(tn->in(0) != nullptr, "must have live top node"); 1230 } 1231 } 1232 #endif 1233 1234 1235 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1236 1237 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1238 guarantee(arr != nullptr, ""); 1239 int num_blocks = arr->length(); 1240 if (grow_by < num_blocks) grow_by = num_blocks; 1241 int num_notes = grow_by * _node_notes_block_size; 1242 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1243 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1244 while (num_notes > 0) { 1245 arr->append(notes); 1246 notes += _node_notes_block_size; 1247 num_notes -= _node_notes_block_size; 1248 } 1249 assert(num_notes == 0, "exact multiple, please"); 1250 } 1251 1252 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1253 if (source == nullptr || dest == nullptr) return false; 1254 1255 if (dest->is_Con()) 1256 return false; // Do not push debug info onto constants. 1257 1258 #ifdef ASSERT 1259 // Leave a bread crumb trail pointing to the original node: 1260 if (dest != nullptr && dest != source && dest->debug_orig() == nullptr) { 1261 dest->set_debug_orig(source); 1262 } 1263 #endif 1264 1265 if (node_note_array() == nullptr) 1266 return false; // Not collecting any notes now. 1267 1268 // This is a copy onto a pre-existing node, which may already have notes. 1269 // If both nodes have notes, do not overwrite any pre-existing notes. 1270 Node_Notes* source_notes = node_notes_at(source->_idx); 1271 if (source_notes == nullptr || source_notes->is_clear()) return false; 1272 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1273 if (dest_notes == nullptr || dest_notes->is_clear()) { 1274 return set_node_notes_at(dest->_idx, source_notes); 1275 } 1276 1277 Node_Notes merged_notes = (*source_notes); 1278 // The order of operations here ensures that dest notes will win... 1279 merged_notes.update_from(dest_notes); 1280 return set_node_notes_at(dest->_idx, &merged_notes); 1281 } 1282 1283 1284 //--------------------------allow_range_check_smearing------------------------- 1285 // Gating condition for coalescing similar range checks. 1286 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1287 // single covering check that is at least as strong as any of them. 1288 // If the optimization succeeds, the simplified (strengthened) range check 1289 // will always succeed. If it fails, we will deopt, and then give up 1290 // on the optimization. 1291 bool Compile::allow_range_check_smearing() const { 1292 // If this method has already thrown a range-check, 1293 // assume it was because we already tried range smearing 1294 // and it failed. 1295 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1296 return !already_trapped; 1297 } 1298 1299 1300 //------------------------------flatten_alias_type----------------------------- 1301 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1302 assert(do_aliasing(), "Aliasing should be enabled"); 1303 int offset = tj->offset(); 1304 TypePtr::PTR ptr = tj->ptr(); 1305 1306 // Known instance (scalarizable allocation) alias only with itself. 1307 bool is_known_inst = tj->isa_oopptr() != nullptr && 1308 tj->is_oopptr()->is_known_instance(); 1309 1310 // Process weird unsafe references. 1311 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1312 assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops"); 1313 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1314 tj = TypeOopPtr::BOTTOM; 1315 ptr = tj->ptr(); 1316 offset = tj->offset(); 1317 } 1318 1319 // Array pointers need some flattening 1320 const TypeAryPtr* ta = tj->isa_aryptr(); 1321 if (ta && ta->is_stable()) { 1322 // Erase stability property for alias analysis. 1323 tj = ta = ta->cast_to_stable(false); 1324 } 1325 if( ta && is_known_inst ) { 1326 if ( offset != Type::OffsetBot && 1327 offset > arrayOopDesc::length_offset_in_bytes() ) { 1328 offset = Type::OffsetBot; // Flatten constant access into array body only 1329 tj = ta = ta-> 1330 remove_speculative()-> 1331 cast_to_ptr_type(ptr)-> 1332 with_offset(offset); 1333 } 1334 } else if (ta) { 1335 // For arrays indexed by constant indices, we flatten the alias 1336 // space to include all of the array body. Only the header, klass 1337 // and array length can be accessed un-aliased. 1338 if( offset != Type::OffsetBot ) { 1339 if( ta->const_oop() ) { // MethodData* or Method* 1340 offset = Type::OffsetBot; // Flatten constant access into array body 1341 tj = ta = ta-> 1342 remove_speculative()-> 1343 cast_to_ptr_type(ptr)-> 1344 cast_to_exactness(false)-> 1345 with_offset(offset); 1346 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1347 // range is OK as-is. 1348 tj = ta = TypeAryPtr::RANGE; 1349 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1350 tj = TypeInstPtr::KLASS; // all klass loads look alike 1351 ta = TypeAryPtr::RANGE; // generic ignored junk 1352 ptr = TypePtr::BotPTR; 1353 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1354 tj = TypeInstPtr::MARK; 1355 ta = TypeAryPtr::RANGE; // generic ignored junk 1356 ptr = TypePtr::BotPTR; 1357 } else { // Random constant offset into array body 1358 offset = Type::OffsetBot; // Flatten constant access into array body 1359 tj = ta = ta-> 1360 remove_speculative()-> 1361 cast_to_ptr_type(ptr)-> 1362 cast_to_exactness(false)-> 1363 with_offset(offset); 1364 } 1365 } 1366 // Arrays of fixed size alias with arrays of unknown size. 1367 if (ta->size() != TypeInt::POS) { 1368 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1369 tj = ta = ta-> 1370 remove_speculative()-> 1371 cast_to_ptr_type(ptr)-> 1372 with_ary(tary)-> 1373 cast_to_exactness(false); 1374 } 1375 // Arrays of known objects become arrays of unknown objects. 1376 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1377 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1378 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1379 } 1380 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1381 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1382 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,nullptr,false,offset); 1383 } 1384 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1385 // cannot be distinguished by bytecode alone. 1386 if (ta->elem() == TypeInt::BOOL) { 1387 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1388 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1389 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1390 } 1391 // During the 2nd round of IterGVN, NotNull castings are removed. 1392 // Make sure the Bottom and NotNull variants alias the same. 1393 // Also, make sure exact and non-exact variants alias the same. 1394 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != nullptr) { 1395 tj = ta = ta-> 1396 remove_speculative()-> 1397 cast_to_ptr_type(TypePtr::BotPTR)-> 1398 cast_to_exactness(false)-> 1399 with_offset(offset); 1400 } 1401 } 1402 1403 // Oop pointers need some flattening 1404 const TypeInstPtr *to = tj->isa_instptr(); 1405 if (to && to != TypeOopPtr::BOTTOM) { 1406 ciInstanceKlass* ik = to->instance_klass(); 1407 if( ptr == TypePtr::Constant ) { 1408 if (ik != ciEnv::current()->Class_klass() || 1409 offset < ik->layout_helper_size_in_bytes()) { 1410 // No constant oop pointers (such as Strings); they alias with 1411 // unknown strings. 1412 assert(!is_known_inst, "not scalarizable allocation"); 1413 tj = to = to-> 1414 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1415 remove_speculative()-> 1416 cast_to_ptr_type(TypePtr::BotPTR)-> 1417 cast_to_exactness(false); 1418 } 1419 } else if( is_known_inst ) { 1420 tj = to; // Keep NotNull and klass_is_exact for instance type 1421 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1422 // During the 2nd round of IterGVN, NotNull castings are removed. 1423 // Make sure the Bottom and NotNull variants alias the same. 1424 // Also, make sure exact and non-exact variants alias the same. 1425 tj = to = to-> 1426 remove_speculative()-> 1427 cast_to_instance_id(TypeOopPtr::InstanceBot)-> 1428 cast_to_ptr_type(TypePtr::BotPTR)-> 1429 cast_to_exactness(false); 1430 } 1431 if (to->speculative() != nullptr) { 1432 tj = to = to->remove_speculative(); 1433 } 1434 // Canonicalize the holder of this field 1435 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1436 // First handle header references such as a LoadKlassNode, even if the 1437 // object's klass is unloaded at compile time (4965979). 1438 if (!is_known_inst) { // Do it only for non-instance types 1439 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, nullptr, offset); 1440 } 1441 } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) { 1442 // Static fields are in the space above the normal instance 1443 // fields in the java.lang.Class instance. 1444 if (ik != ciEnv::current()->Class_klass()) { 1445 to = nullptr; 1446 tj = TypeOopPtr::BOTTOM; 1447 offset = tj->offset(); 1448 } 1449 } else { 1450 ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset); 1451 assert(offset < canonical_holder->layout_helper_size_in_bytes(), ""); 1452 if (!ik->equals(canonical_holder) || tj->offset() != offset) { 1453 if( is_known_inst ) { 1454 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, nullptr, offset, to->instance_id()); 1455 } else { 1456 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, nullptr, offset); 1457 } 1458 } 1459 } 1460 } 1461 1462 // Klass pointers to object array klasses need some flattening 1463 const TypeKlassPtr *tk = tj->isa_klassptr(); 1464 if( tk ) { 1465 // If we are referencing a field within a Klass, we need 1466 // to assume the worst case of an Object. Both exact and 1467 // inexact types must flatten to the same alias class so 1468 // use NotNull as the PTR. 1469 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1470 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, 1471 env()->Object_klass(), 1472 offset); 1473 } 1474 1475 if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) { 1476 ciKlass* k = ciObjArrayKlass::make(env()->Object_klass()); 1477 if (!k || !k->is_loaded()) { // Only fails for some -Xcomp runs 1478 tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset); 1479 } else { 1480 tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset); 1481 } 1482 } 1483 1484 // Check for precise loads from the primary supertype array and force them 1485 // to the supertype cache alias index. Check for generic array loads from 1486 // the primary supertype array and also force them to the supertype cache 1487 // alias index. Since the same load can reach both, we need to merge 1488 // these 2 disparate memories into the same alias class. Since the 1489 // primary supertype array is read-only, there's no chance of confusion 1490 // where we bypass an array load and an array store. 1491 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1492 if (offset == Type::OffsetBot || 1493 (offset >= primary_supers_offset && 1494 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1495 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1496 offset = in_bytes(Klass::secondary_super_cache_offset()); 1497 tj = tk = tk->with_offset(offset); 1498 } 1499 } 1500 1501 // Flatten all Raw pointers together. 1502 if (tj->base() == Type::RawPtr) 1503 tj = TypeRawPtr::BOTTOM; 1504 1505 if (tj->base() == Type::AnyPtr) 1506 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1507 1508 offset = tj->offset(); 1509 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1510 1511 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1512 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1513 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1514 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1515 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1516 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1517 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr), 1518 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1519 assert( tj->ptr() != TypePtr::TopPTR && 1520 tj->ptr() != TypePtr::AnyNull && 1521 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1522 // assert( tj->ptr() != TypePtr::Constant || 1523 // tj->base() == Type::RawPtr || 1524 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1525 1526 return tj; 1527 } 1528 1529 void Compile::AliasType::Init(int i, const TypePtr* at) { 1530 assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index"); 1531 _index = i; 1532 _adr_type = at; 1533 _field = nullptr; 1534 _element = nullptr; 1535 _is_rewritable = true; // default 1536 const TypeOopPtr *atoop = (at != nullptr) ? at->isa_oopptr() : nullptr; 1537 if (atoop != nullptr && atoop->is_known_instance()) { 1538 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1539 _general_index = Compile::current()->get_alias_index(gt); 1540 } else { 1541 _general_index = 0; 1542 } 1543 } 1544 1545 BasicType Compile::AliasType::basic_type() const { 1546 if (element() != nullptr) { 1547 const Type* element = adr_type()->is_aryptr()->elem(); 1548 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1549 } if (field() != nullptr) { 1550 return field()->layout_type(); 1551 } else { 1552 return T_ILLEGAL; // unknown 1553 } 1554 } 1555 1556 //---------------------------------print_on------------------------------------ 1557 #ifndef PRODUCT 1558 void Compile::AliasType::print_on(outputStream* st) { 1559 if (index() < 10) 1560 st->print("@ <%d> ", index()); 1561 else st->print("@ <%d>", index()); 1562 st->print(is_rewritable() ? " " : " RO"); 1563 int offset = adr_type()->offset(); 1564 if (offset == Type::OffsetBot) 1565 st->print(" +any"); 1566 else st->print(" +%-3d", offset); 1567 st->print(" in "); 1568 adr_type()->dump_on(st); 1569 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1570 if (field() != nullptr && tjp) { 1571 if (tjp->is_instptr()->instance_klass() != field()->holder() || 1572 tjp->offset() != field()->offset_in_bytes()) { 1573 st->print(" != "); 1574 field()->print(); 1575 st->print(" ***"); 1576 } 1577 } 1578 } 1579 1580 void print_alias_types() { 1581 Compile* C = Compile::current(); 1582 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1583 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1584 C->alias_type(idx)->print_on(tty); 1585 tty->cr(); 1586 } 1587 } 1588 #endif 1589 1590 1591 //----------------------------probe_alias_cache-------------------------------- 1592 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1593 intptr_t key = (intptr_t) adr_type; 1594 key ^= key >> logAliasCacheSize; 1595 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1596 } 1597 1598 1599 //-----------------------------grow_alias_types-------------------------------- 1600 void Compile::grow_alias_types() { 1601 const int old_ats = _max_alias_types; // how many before? 1602 const int new_ats = old_ats; // how many more? 1603 const int grow_ats = old_ats+new_ats; // how many now? 1604 _max_alias_types = grow_ats; 1605 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1606 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1607 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1608 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1609 } 1610 1611 1612 //--------------------------------find_alias_type------------------------------ 1613 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1614 if (!do_aliasing()) { 1615 return alias_type(AliasIdxBot); 1616 } 1617 1618 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1619 if (ace->_adr_type == adr_type) { 1620 return alias_type(ace->_index); 1621 } 1622 1623 // Handle special cases. 1624 if (adr_type == nullptr) return alias_type(AliasIdxTop); 1625 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1626 1627 // Do it the slow way. 1628 const TypePtr* flat = flatten_alias_type(adr_type); 1629 1630 #ifdef ASSERT 1631 { 1632 ResourceMark rm; 1633 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1634 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1635 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1636 Type::str(adr_type)); 1637 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1638 const TypeOopPtr* foop = flat->is_oopptr(); 1639 // Scalarizable allocations have exact klass always. 1640 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1641 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1642 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1643 Type::str(foop), Type::str(xoop)); 1644 } 1645 } 1646 #endif 1647 1648 int idx = AliasIdxTop; 1649 for (int i = 0; i < num_alias_types(); i++) { 1650 if (alias_type(i)->adr_type() == flat) { 1651 idx = i; 1652 break; 1653 } 1654 } 1655 1656 if (idx == AliasIdxTop) { 1657 if (no_create) return nullptr; 1658 // Grow the array if necessary. 1659 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1660 // Add a new alias type. 1661 idx = _num_alias_types++; 1662 _alias_types[idx]->Init(idx, flat); 1663 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1664 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1665 if (flat->isa_instptr()) { 1666 if (flat->offset() == java_lang_Class::klass_offset() 1667 && flat->is_instptr()->instance_klass() == env()->Class_klass()) 1668 alias_type(idx)->set_rewritable(false); 1669 } 1670 if (flat->isa_aryptr()) { 1671 #ifdef ASSERT 1672 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1673 // (T_BYTE has the weakest alignment and size restrictions...) 1674 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1675 #endif 1676 if (flat->offset() == TypePtr::OffsetBot) { 1677 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1678 } 1679 } 1680 if (flat->isa_klassptr()) { 1681 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1682 alias_type(idx)->set_rewritable(false); 1683 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1684 alias_type(idx)->set_rewritable(false); 1685 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1686 alias_type(idx)->set_rewritable(false); 1687 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1688 alias_type(idx)->set_rewritable(false); 1689 if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset())) 1690 alias_type(idx)->set_rewritable(false); 1691 } 1692 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1693 // but the base pointer type is not distinctive enough to identify 1694 // references into JavaThread.) 1695 1696 // Check for final fields. 1697 const TypeInstPtr* tinst = flat->isa_instptr(); 1698 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1699 ciField* field; 1700 if (tinst->const_oop() != nullptr && 1701 tinst->instance_klass() == ciEnv::current()->Class_klass() && 1702 tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) { 1703 // static field 1704 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1705 field = k->get_field_by_offset(tinst->offset(), true); 1706 } else { 1707 ciInstanceKlass *k = tinst->instance_klass(); 1708 field = k->get_field_by_offset(tinst->offset(), false); 1709 } 1710 assert(field == nullptr || 1711 original_field == nullptr || 1712 (field->holder() == original_field->holder() && 1713 field->offset_in_bytes() == original_field->offset_in_bytes() && 1714 field->is_static() == original_field->is_static()), "wrong field?"); 1715 // Set field() and is_rewritable() attributes. 1716 if (field != nullptr) alias_type(idx)->set_field(field); 1717 } 1718 } 1719 1720 // Fill the cache for next time. 1721 ace->_adr_type = adr_type; 1722 ace->_index = idx; 1723 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1724 1725 // Might as well try to fill the cache for the flattened version, too. 1726 AliasCacheEntry* face = probe_alias_cache(flat); 1727 if (face->_adr_type == nullptr) { 1728 face->_adr_type = flat; 1729 face->_index = idx; 1730 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1731 } 1732 1733 return alias_type(idx); 1734 } 1735 1736 1737 Compile::AliasType* Compile::alias_type(ciField* field) { 1738 const TypeOopPtr* t; 1739 if (field->is_static()) 1740 t = TypeInstPtr::make(field->holder()->java_mirror()); 1741 else 1742 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1743 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1744 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1745 return atp; 1746 } 1747 1748 1749 //------------------------------have_alias_type-------------------------------- 1750 bool Compile::have_alias_type(const TypePtr* adr_type) { 1751 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1752 if (ace->_adr_type == adr_type) { 1753 return true; 1754 } 1755 1756 // Handle special cases. 1757 if (adr_type == nullptr) return true; 1758 if (adr_type == TypePtr::BOTTOM) return true; 1759 1760 return find_alias_type(adr_type, true, nullptr) != nullptr; 1761 } 1762 1763 //-----------------------------must_alias-------------------------------------- 1764 // True if all values of the given address type are in the given alias category. 1765 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1766 if (alias_idx == AliasIdxBot) return true; // the universal category 1767 if (adr_type == nullptr) return true; // null serves as TypePtr::TOP 1768 if (alias_idx == AliasIdxTop) return false; // the empty category 1769 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1770 1771 // the only remaining possible overlap is identity 1772 int adr_idx = get_alias_index(adr_type); 1773 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1774 assert(adr_idx == alias_idx || 1775 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1776 && adr_type != TypeOopPtr::BOTTOM), 1777 "should not be testing for overlap with an unsafe pointer"); 1778 return adr_idx == alias_idx; 1779 } 1780 1781 //------------------------------can_alias-------------------------------------- 1782 // True if any values of the given address type are in the given alias category. 1783 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1784 if (alias_idx == AliasIdxTop) return false; // the empty category 1785 if (adr_type == nullptr) return false; // null serves as TypePtr::TOP 1786 // Known instance doesn't alias with bottom memory 1787 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1788 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1789 1790 // the only remaining possible overlap is identity 1791 int adr_idx = get_alias_index(adr_type); 1792 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1793 return adr_idx == alias_idx; 1794 } 1795 1796 //---------------------cleanup_loop_predicates----------------------- 1797 // Remove the opaque nodes that protect the predicates so that all unused 1798 // checks and uncommon_traps will be eliminated from the ideal graph 1799 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1800 if (predicate_count()==0) return; 1801 for (int i = predicate_count(); i > 0; i--) { 1802 Node * n = predicate_opaque1_node(i-1); 1803 assert(n->Opcode() == Op_Opaque1, "must be"); 1804 igvn.replace_node(n, n->in(1)); 1805 } 1806 assert(predicate_count()==0, "should be clean!"); 1807 } 1808 1809 void Compile::record_for_post_loop_opts_igvn(Node* n) { 1810 if (!n->for_post_loop_opts_igvn()) { 1811 assert(!_for_post_loop_igvn.contains(n), "duplicate"); 1812 n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1813 _for_post_loop_igvn.append(n); 1814 } 1815 } 1816 1817 void Compile::remove_from_post_loop_opts_igvn(Node* n) { 1818 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1819 _for_post_loop_igvn.remove(n); 1820 } 1821 1822 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) { 1823 // Verify that all previous optimizations produced a valid graph 1824 // at least to this point, even if no loop optimizations were done. 1825 PhaseIdealLoop::verify(igvn); 1826 1827 C->set_post_loop_opts_phase(); // no more loop opts allowed 1828 1829 assert(!C->major_progress(), "not cleared"); 1830 1831 if (_for_post_loop_igvn.length() > 0) { 1832 while (_for_post_loop_igvn.length() > 0) { 1833 Node* n = _for_post_loop_igvn.pop(); 1834 n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn); 1835 igvn._worklist.push(n); 1836 } 1837 igvn.optimize(); 1838 assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed"); 1839 1840 // Sometimes IGVN sets major progress (e.g., when processing loop nodes). 1841 if (C->major_progress()) { 1842 C->clear_major_progress(); // ensure that major progress is now clear 1843 } 1844 } 1845 } 1846 1847 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) { 1848 if (OptimizeUnstableIf) { 1849 _unstable_if_traps.append(trap); 1850 } 1851 } 1852 1853 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) { 1854 for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) { 1855 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1856 Node* n = trap->uncommon_trap(); 1857 if (!useful.member(n)) { 1858 _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed) 1859 } 1860 } 1861 } 1862 1863 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead 1864 // or fold-compares case. Return true if succeed or not found. 1865 // 1866 // In rare cases, the found trap has been processed. It is too late to delete it. Return 1867 // false and ask fold-compares to yield. 1868 // 1869 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused 1870 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path 1871 // when deoptimization does happen. 1872 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) { 1873 for (int i = 0; i < _unstable_if_traps.length(); ++i) { 1874 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1875 if (trap->uncommon_trap() == unc) { 1876 if (yield && trap->modified()) { 1877 return false; 1878 } 1879 _unstable_if_traps.delete_at(i); 1880 break; 1881 } 1882 } 1883 return true; 1884 } 1885 1886 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path. 1887 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering. 1888 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) { 1889 for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) { 1890 UnstableIfTrap* trap = _unstable_if_traps.at(i); 1891 CallStaticJavaNode* unc = trap->uncommon_trap(); 1892 int next_bci = trap->next_bci(); 1893 bool modified = trap->modified(); 1894 1895 if (next_bci != -1 && !modified) { 1896 assert(!_dead_node_list.test(unc->_idx), "changing a dead node!"); 1897 JVMState* jvms = unc->jvms(); 1898 ciMethod* method = jvms->method(); 1899 ciBytecodeStream iter(method); 1900 1901 iter.force_bci(jvms->bci()); 1902 assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if"); 1903 Bytecodes::Code c = iter.cur_bc(); 1904 Node* lhs = nullptr; 1905 Node* rhs = nullptr; 1906 if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) { 1907 lhs = unc->peek_operand(0); 1908 rhs = unc->peek_operand(1); 1909 } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) { 1910 lhs = unc->peek_operand(0); 1911 } 1912 1913 ResourceMark rm; 1914 const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci); 1915 assert(live_locals.is_valid(), "broken liveness info"); 1916 int len = (int)live_locals.size(); 1917 1918 for (int i = 0; i < len; i++) { 1919 Node* local = unc->local(jvms, i); 1920 // kill local using the liveness of next_bci. 1921 // give up when the local looks like an operand to secure reexecution. 1922 if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) { 1923 uint idx = jvms->locoff() + i; 1924 #ifdef ASSERT 1925 if (Verbose) { 1926 tty->print("[unstable_if] kill local#%d: ", idx); 1927 local->dump(); 1928 tty->cr(); 1929 } 1930 #endif 1931 igvn.replace_input_of(unc, idx, top()); 1932 modified = true; 1933 } 1934 } 1935 } 1936 1937 // keep the mondified trap for late query 1938 if (modified) { 1939 trap->set_modified(); 1940 } else { 1941 _unstable_if_traps.delete_at(i); 1942 } 1943 } 1944 igvn.optimize(); 1945 } 1946 1947 // StringOpts and late inlining of string methods 1948 void Compile::inline_string_calls(bool parse_time) { 1949 { 1950 // remove useless nodes to make the usage analysis simpler 1951 ResourceMark rm; 1952 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1953 } 1954 1955 { 1956 ResourceMark rm; 1957 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1958 PhaseStringOpts pso(initial_gvn()); 1959 print_method(PHASE_AFTER_STRINGOPTS, 3); 1960 } 1961 1962 // now inline anything that we skipped the first time around 1963 if (!parse_time) { 1964 _late_inlines_pos = _late_inlines.length(); 1965 } 1966 1967 while (_string_late_inlines.length() > 0) { 1968 CallGenerator* cg = _string_late_inlines.pop(); 1969 cg->do_late_inline(); 1970 if (failing()) return; 1971 } 1972 _string_late_inlines.trunc_to(0); 1973 } 1974 1975 // Late inlining of boxing methods 1976 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1977 if (_boxing_late_inlines.length() > 0) { 1978 assert(has_boxed_value(), "inconsistent"); 1979 1980 PhaseGVN* gvn = initial_gvn(); 1981 set_inlining_incrementally(true); 1982 1983 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1984 for_igvn()->clear(); 1985 gvn->replace_with(&igvn); 1986 1987 _late_inlines_pos = _late_inlines.length(); 1988 1989 while (_boxing_late_inlines.length() > 0) { 1990 CallGenerator* cg = _boxing_late_inlines.pop(); 1991 cg->do_late_inline(); 1992 if (failing()) return; 1993 } 1994 _boxing_late_inlines.trunc_to(0); 1995 1996 inline_incrementally_cleanup(igvn); 1997 1998 set_inlining_incrementally(false); 1999 } 2000 } 2001 2002 bool Compile::inline_incrementally_one() { 2003 assert(IncrementalInline, "incremental inlining should be on"); 2004 2005 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2006 2007 set_inlining_progress(false); 2008 set_do_cleanup(false); 2009 2010 for (int i = 0; i < _late_inlines.length(); i++) { 2011 _late_inlines_pos = i+1; 2012 CallGenerator* cg = _late_inlines.at(i); 2013 bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline(); 2014 if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call 2015 cg->do_late_inline(); 2016 assert(_late_inlines.at(i) == cg, "no insertions before current position allowed"); 2017 if (failing()) { 2018 return false; 2019 } else if (inlining_progress()) { 2020 _late_inlines_pos = i+1; // restore the position in case new elements were inserted 2021 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node()); 2022 break; // process one call site at a time 2023 } 2024 } else { 2025 // Ignore late inline direct calls when inlining is not allowed. 2026 // They are left in the late inline list when node budget is exhausted until the list is fully drained. 2027 } 2028 } 2029 // Remove processed elements. 2030 _late_inlines.remove_till(_late_inlines_pos); 2031 _late_inlines_pos = 0; 2032 2033 assert(inlining_progress() || _late_inlines.length() == 0, "no progress"); 2034 2035 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2036 2037 set_inlining_progress(false); 2038 set_do_cleanup(false); 2039 2040 bool force_cleanup = directive()->IncrementalInlineForceCleanupOption; 2041 return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup; 2042 } 2043 2044 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2045 { 2046 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2047 ResourceMark rm; 2048 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2049 } 2050 { 2051 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2052 igvn = PhaseIterGVN(initial_gvn()); 2053 igvn.optimize(); 2054 } 2055 print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3); 2056 } 2057 2058 // Perform incremental inlining until bound on number of live nodes is reached 2059 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2060 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2061 2062 set_inlining_incrementally(true); 2063 uint low_live_nodes = 0; 2064 2065 while (_late_inlines.length() > 0) { 2066 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2067 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2068 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2069 // PhaseIdealLoop is expensive so we only try it once we are 2070 // out of live nodes and we only try it again if the previous 2071 // helped got the number of nodes down significantly 2072 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2073 if (failing()) return; 2074 low_live_nodes = live_nodes(); 2075 _major_progress = true; 2076 } 2077 2078 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2079 bool do_print_inlining = print_inlining() || print_intrinsics(); 2080 if (do_print_inlining || log() != nullptr) { 2081 // Print inlining message for candidates that we couldn't inline for lack of space. 2082 for (int i = 0; i < _late_inlines.length(); i++) { 2083 CallGenerator* cg = _late_inlines.at(i); 2084 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 2085 if (do_print_inlining) { 2086 cg->print_inlining_late(msg); 2087 } 2088 log_late_inline_failure(cg, msg); 2089 } 2090 } 2091 break; // finish 2092 } 2093 } 2094 2095 for_igvn()->clear(); 2096 initial_gvn()->replace_with(&igvn); 2097 2098 while (inline_incrementally_one()) { 2099 assert(!failing(), "inconsistent"); 2100 } 2101 if (failing()) return; 2102 2103 inline_incrementally_cleanup(igvn); 2104 2105 print_method(PHASE_INCREMENTAL_INLINE_STEP, 3); 2106 2107 if (failing()) return; 2108 2109 if (_late_inlines.length() == 0) { 2110 break; // no more progress 2111 } 2112 } 2113 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2114 2115 if (_string_late_inlines.length() > 0) { 2116 assert(has_stringbuilder(), "inconsistent"); 2117 for_igvn()->clear(); 2118 initial_gvn()->replace_with(&igvn); 2119 2120 inline_string_calls(false); 2121 2122 if (failing()) return; 2123 2124 inline_incrementally_cleanup(igvn); 2125 } 2126 2127 set_inlining_incrementally(false); 2128 } 2129 2130 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) { 2131 // "inlining_incrementally() == false" is used to signal that no inlining is allowed 2132 // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details). 2133 // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == nullptr" 2134 // as if "inlining_incrementally() == true" were set. 2135 assert(inlining_incrementally() == false, "not allowed"); 2136 assert(_modified_nodes == nullptr, "not allowed"); 2137 assert(_late_inlines.length() > 0, "sanity"); 2138 2139 while (_late_inlines.length() > 0) { 2140 for_igvn()->clear(); 2141 initial_gvn()->replace_with(&igvn); 2142 2143 while (inline_incrementally_one()) { 2144 assert(!failing(), "inconsistent"); 2145 } 2146 if (failing()) return; 2147 2148 inline_incrementally_cleanup(igvn); 2149 } 2150 } 2151 2152 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2153 if (_loop_opts_cnt > 0) { 2154 while (major_progress() && (_loop_opts_cnt > 0)) { 2155 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2156 PhaseIdealLoop::optimize(igvn, mode); 2157 _loop_opts_cnt--; 2158 if (failing()) return false; 2159 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2160 } 2161 } 2162 return true; 2163 } 2164 2165 // Remove edges from "root" to each SafePoint at a backward branch. 2166 // They were inserted during parsing (see add_safepoint()) to make 2167 // infinite loops without calls or exceptions visible to root, i.e., 2168 // useful. 2169 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2170 Node *r = root(); 2171 if (r != nullptr) { 2172 for (uint i = r->req(); i < r->len(); ++i) { 2173 Node *n = r->in(i); 2174 if (n != nullptr && n->is_SafePoint()) { 2175 r->rm_prec(i); 2176 if (n->outcnt() == 0) { 2177 igvn.remove_dead_node(n); 2178 } 2179 --i; 2180 } 2181 } 2182 // Parsing may have added top inputs to the root node (Path 2183 // leading to the Halt node proven dead). Make sure we get a 2184 // chance to clean them up. 2185 igvn._worklist.push(r); 2186 igvn.optimize(); 2187 } 2188 } 2189 2190 //------------------------------Optimize--------------------------------------- 2191 // Given a graph, optimize it. 2192 void Compile::Optimize() { 2193 TracePhase tp("optimizer", &timers[_t_optimizer]); 2194 2195 #ifndef PRODUCT 2196 if (env()->break_at_compile()) { 2197 BREAKPOINT; 2198 } 2199 2200 #endif 2201 2202 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2203 #ifdef ASSERT 2204 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2205 #endif 2206 2207 ResourceMark rm; 2208 2209 print_inlining_reinit(); 2210 2211 NOT_PRODUCT( verify_graph_edges(); ) 2212 2213 print_method(PHASE_AFTER_PARSING, 1); 2214 2215 { 2216 // Iterative Global Value Numbering, including ideal transforms 2217 // Initialize IterGVN with types and values from parse-time GVN 2218 PhaseIterGVN igvn(initial_gvn()); 2219 #ifdef ASSERT 2220 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2221 #endif 2222 { 2223 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2224 igvn.optimize(); 2225 } 2226 2227 if (failing()) return; 2228 2229 print_method(PHASE_ITER_GVN1, 2); 2230 2231 process_for_unstable_if_traps(igvn); 2232 2233 inline_incrementally(igvn); 2234 2235 print_method(PHASE_INCREMENTAL_INLINE, 2); 2236 2237 if (failing()) return; 2238 2239 if (eliminate_boxing()) { 2240 // Inline valueOf() methods now. 2241 inline_boxing_calls(igvn); 2242 2243 if (AlwaysIncrementalInline) { 2244 inline_incrementally(igvn); 2245 } 2246 2247 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2248 2249 if (failing()) return; 2250 } 2251 2252 // Remove the speculative part of types and clean up the graph from 2253 // the extra CastPP nodes whose only purpose is to carry them. Do 2254 // that early so that optimizations are not disrupted by the extra 2255 // CastPP nodes. 2256 remove_speculative_types(igvn); 2257 2258 // No more new expensive nodes will be added to the list from here 2259 // so keep only the actual candidates for optimizations. 2260 cleanup_expensive_nodes(igvn); 2261 2262 assert(EnableVectorSupport || !has_vbox_nodes(), "sanity"); 2263 if (EnableVectorSupport && has_vbox_nodes()) { 2264 TracePhase tp("", &timers[_t_vector]); 2265 PhaseVector pv(igvn); 2266 pv.optimize_vector_boxes(); 2267 2268 print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2); 2269 } 2270 assert(!has_vbox_nodes(), "sanity"); 2271 2272 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2273 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2274 initial_gvn()->replace_with(&igvn); 2275 Unique_Node_List* old_worklist = for_igvn(); 2276 old_worklist->clear(); 2277 Unique_Node_List new_worklist(C->comp_arena()); 2278 { 2279 ResourceMark rm; 2280 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2281 } 2282 Unique_Node_List* save_for_igvn = for_igvn(); 2283 set_for_igvn(&new_worklist); 2284 igvn = PhaseIterGVN(initial_gvn()); 2285 igvn.optimize(); 2286 set_for_igvn(old_worklist); // new_worklist is dead beyond this point 2287 } 2288 2289 // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop 2290 // safepoints 2291 remove_root_to_sfpts_edges(igvn); 2292 2293 // Perform escape analysis 2294 if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) { 2295 if (has_loops()) { 2296 // Cleanup graph (remove dead nodes). 2297 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2298 PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll); 2299 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2300 if (failing()) return; 2301 } 2302 bool progress; 2303 do { 2304 ConnectionGraph::do_analysis(this, &igvn); 2305 2306 if (failing()) return; 2307 2308 int mcount = macro_count(); // Record number of allocations and locks before IGVN 2309 2310 // Optimize out fields loads from scalar replaceable allocations. 2311 igvn.optimize(); 2312 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2313 2314 if (failing()) return; 2315 2316 if (congraph() != nullptr && macro_count() > 0) { 2317 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2318 PhaseMacroExpand mexp(igvn); 2319 mexp.eliminate_macro_nodes(); 2320 igvn.set_delay_transform(false); 2321 2322 igvn.optimize(); 2323 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2324 2325 if (failing()) return; 2326 } 2327 progress = do_iterative_escape_analysis() && 2328 (macro_count() < mcount) && 2329 ConnectionGraph::has_candidates(this); 2330 // Try again if candidates exist and made progress 2331 // by removing some allocations and/or locks. 2332 } while (progress); 2333 } 2334 2335 // Loop transforms on the ideal graph. Range Check Elimination, 2336 // peeling, unrolling, etc. 2337 2338 // Set loop opts counter 2339 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2340 { 2341 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2342 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2343 _loop_opts_cnt--; 2344 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2345 if (failing()) return; 2346 } 2347 // Loop opts pass if partial peeling occurred in previous pass 2348 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2349 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2350 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2351 _loop_opts_cnt--; 2352 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2353 if (failing()) return; 2354 } 2355 // Loop opts pass for loop-unrolling before CCP 2356 if(major_progress() && (_loop_opts_cnt > 0)) { 2357 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2358 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2359 _loop_opts_cnt--; 2360 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2361 } 2362 if (!failing()) { 2363 // Verify that last round of loop opts produced a valid graph 2364 PhaseIdealLoop::verify(igvn); 2365 } 2366 } 2367 if (failing()) return; 2368 2369 // Conditional Constant Propagation; 2370 PhaseCCP ccp( &igvn ); 2371 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2372 { 2373 TracePhase tp("ccp", &timers[_t_ccp]); 2374 ccp.do_transform(); 2375 } 2376 print_method(PHASE_CCP1, 2); 2377 2378 assert( true, "Break here to ccp.dump_old2new_map()"); 2379 2380 // Iterative Global Value Numbering, including ideal transforms 2381 { 2382 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2383 igvn = ccp; 2384 igvn.optimize(); 2385 } 2386 print_method(PHASE_ITER_GVN2, 2); 2387 2388 if (failing()) return; 2389 2390 // Loop transforms on the ideal graph. Range Check Elimination, 2391 // peeling, unrolling, etc. 2392 if (!optimize_loops(igvn, LoopOptsDefault)) { 2393 return; 2394 } 2395 2396 if (failing()) return; 2397 2398 C->clear_major_progress(); // ensure that major progress is now clear 2399 2400 process_for_post_loop_opts_igvn(igvn); 2401 2402 #ifdef ASSERT 2403 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2404 #endif 2405 2406 { 2407 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2408 PhaseMacroExpand mex(igvn); 2409 if (mex.expand_macro_nodes()) { 2410 assert(failing(), "must bail out w/ explicit message"); 2411 return; 2412 } 2413 print_method(PHASE_MACRO_EXPANSION, 2); 2414 } 2415 2416 { 2417 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2418 if (bs->expand_barriers(this, igvn)) { 2419 assert(failing(), "must bail out w/ explicit message"); 2420 return; 2421 } 2422 print_method(PHASE_BARRIER_EXPANSION, 2); 2423 } 2424 2425 if (C->max_vector_size() > 0) { 2426 C->optimize_logic_cones(igvn); 2427 igvn.optimize(); 2428 } 2429 2430 DEBUG_ONLY( _modified_nodes = nullptr; ) 2431 2432 assert(igvn._worklist.size() == 0, "not empty"); 2433 2434 assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty"); 2435 2436 if (_late_inlines.length() > 0) { 2437 // More opportunities to optimize virtual and MH calls. 2438 // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option. 2439 process_late_inline_calls_no_inline(igvn); 2440 } 2441 } // (End scope of igvn; run destructor if necessary for asserts.) 2442 2443 check_no_dead_use(); 2444 2445 process_print_inlining(); 2446 2447 // A method with only infinite loops has no edges entering loops from root 2448 { 2449 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2450 if (final_graph_reshaping()) { 2451 assert(failing(), "must bail out w/ explicit message"); 2452 return; 2453 } 2454 } 2455 2456 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2457 DEBUG_ONLY(set_phase_optimize_finished();) 2458 } 2459 2460 #ifdef ASSERT 2461 void Compile::check_no_dead_use() const { 2462 ResourceMark rm; 2463 Unique_Node_List wq; 2464 wq.push(root()); 2465 for (uint i = 0; i < wq.size(); ++i) { 2466 Node* n = wq.at(i); 2467 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { 2468 Node* u = n->fast_out(j); 2469 if (u->outcnt() == 0 && !u->is_Con()) { 2470 u->dump(); 2471 fatal("no reachable node should have no use"); 2472 } 2473 wq.push(u); 2474 } 2475 } 2476 } 2477 #endif 2478 2479 void Compile::inline_vector_reboxing_calls() { 2480 if (C->_vector_reboxing_late_inlines.length() > 0) { 2481 _late_inlines_pos = C->_late_inlines.length(); 2482 while (_vector_reboxing_late_inlines.length() > 0) { 2483 CallGenerator* cg = _vector_reboxing_late_inlines.pop(); 2484 cg->do_late_inline(); 2485 if (failing()) return; 2486 print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node()); 2487 } 2488 _vector_reboxing_late_inlines.trunc_to(0); 2489 } 2490 } 2491 2492 bool Compile::has_vbox_nodes() { 2493 if (C->_vector_reboxing_late_inlines.length() > 0) { 2494 return true; 2495 } 2496 for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) { 2497 Node * n = C->macro_node(macro_idx); 2498 assert(n->is_macro(), "only macro nodes expected here"); 2499 if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) { 2500 return true; 2501 } 2502 } 2503 return false; 2504 } 2505 2506 //---------------------------- Bitwise operation packing optimization --------------------------- 2507 2508 static bool is_vector_unary_bitwise_op(Node* n) { 2509 return n->Opcode() == Op_XorV && 2510 VectorNode::is_vector_bitwise_not_pattern(n); 2511 } 2512 2513 static bool is_vector_binary_bitwise_op(Node* n) { 2514 switch (n->Opcode()) { 2515 case Op_AndV: 2516 case Op_OrV: 2517 return true; 2518 2519 case Op_XorV: 2520 return !is_vector_unary_bitwise_op(n); 2521 2522 default: 2523 return false; 2524 } 2525 } 2526 2527 static bool is_vector_ternary_bitwise_op(Node* n) { 2528 return n->Opcode() == Op_MacroLogicV; 2529 } 2530 2531 static bool is_vector_bitwise_op(Node* n) { 2532 return is_vector_unary_bitwise_op(n) || 2533 is_vector_binary_bitwise_op(n) || 2534 is_vector_ternary_bitwise_op(n); 2535 } 2536 2537 static bool is_vector_bitwise_cone_root(Node* n) { 2538 if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) { 2539 return false; 2540 } 2541 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 2542 if (is_vector_bitwise_op(n->fast_out(i))) { 2543 return false; 2544 } 2545 } 2546 return true; 2547 } 2548 2549 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) { 2550 uint cnt = 0; 2551 if (is_vector_bitwise_op(n)) { 2552 uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req(); 2553 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2554 for (uint i = 1; i < inp_cnt; i++) { 2555 Node* in = n->in(i); 2556 bool skip = VectorNode::is_all_ones_vector(in); 2557 if (!skip && !inputs.member(in)) { 2558 inputs.push(in); 2559 cnt++; 2560 } 2561 } 2562 assert(cnt <= 1, "not unary"); 2563 } else { 2564 uint last_req = inp_cnt; 2565 if (is_vector_ternary_bitwise_op(n)) { 2566 last_req = inp_cnt - 1; // skip last input 2567 } 2568 for (uint i = 1; i < last_req; i++) { 2569 Node* def = n->in(i); 2570 if (!inputs.member(def)) { 2571 inputs.push(def); 2572 cnt++; 2573 } 2574 } 2575 } 2576 } else { // not a bitwise operations 2577 if (!inputs.member(n)) { 2578 inputs.push(n); 2579 cnt++; 2580 } 2581 } 2582 return cnt; 2583 } 2584 2585 void Compile::collect_logic_cone_roots(Unique_Node_List& list) { 2586 Unique_Node_List useful_nodes; 2587 C->identify_useful_nodes(useful_nodes); 2588 2589 for (uint i = 0; i < useful_nodes.size(); i++) { 2590 Node* n = useful_nodes.at(i); 2591 if (is_vector_bitwise_cone_root(n)) { 2592 list.push(n); 2593 } 2594 } 2595 } 2596 2597 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn, 2598 const TypeVect* vt, 2599 Unique_Node_List& partition, 2600 Unique_Node_List& inputs) { 2601 assert(partition.size() == 2 || partition.size() == 3, "not supported"); 2602 assert(inputs.size() == 2 || inputs.size() == 3, "not supported"); 2603 assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported"); 2604 2605 Node* in1 = inputs.at(0); 2606 Node* in2 = inputs.at(1); 2607 Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2); 2608 2609 uint func = compute_truth_table(partition, inputs); 2610 2611 Node* pn = partition.at(partition.size() - 1); 2612 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2613 return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt)); 2614 } 2615 2616 static uint extract_bit(uint func, uint pos) { 2617 return (func & (1 << pos)) >> pos; 2618 } 2619 2620 // 2621 // A macro logic node represents a truth table. It has 4 inputs, 2622 // First three inputs corresponds to 3 columns of a truth table 2623 // and fourth input captures the logic function. 2624 // 2625 // eg. fn = (in1 AND in2) OR in3; 2626 // 2627 // MacroNode(in1,in2,in3,fn) 2628 // 2629 // ----------------- 2630 // in1 in2 in3 fn 2631 // ----------------- 2632 // 0 0 0 0 2633 // 0 0 1 1 2634 // 0 1 0 0 2635 // 0 1 1 1 2636 // 1 0 0 0 2637 // 1 0 1 1 2638 // 1 1 0 1 2639 // 1 1 1 1 2640 // 2641 2642 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) { 2643 int res = 0; 2644 for (int i = 0; i < 8; i++) { 2645 int bit1 = extract_bit(in1, i); 2646 int bit2 = extract_bit(in2, i); 2647 int bit3 = extract_bit(in3, i); 2648 2649 int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3); 2650 int func_bit = extract_bit(func, func_bit_pos); 2651 2652 res |= func_bit << i; 2653 } 2654 return res; 2655 } 2656 2657 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) { 2658 assert(n != nullptr, ""); 2659 assert(eval_map.contains(n), "absent"); 2660 return *(eval_map.get(n)); 2661 } 2662 2663 static void eval_operands(Node* n, 2664 uint& func1, uint& func2, uint& func3, 2665 ResourceHashtable<Node*,uint>& eval_map) { 2666 assert(is_vector_bitwise_op(n), ""); 2667 2668 if (is_vector_unary_bitwise_op(n)) { 2669 Node* opnd = n->in(1); 2670 if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) { 2671 opnd = n->in(2); 2672 } 2673 func1 = eval_operand(opnd, eval_map); 2674 } else if (is_vector_binary_bitwise_op(n)) { 2675 func1 = eval_operand(n->in(1), eval_map); 2676 func2 = eval_operand(n->in(2), eval_map); 2677 } else { 2678 assert(is_vector_ternary_bitwise_op(n), "unknown operation"); 2679 func1 = eval_operand(n->in(1), eval_map); 2680 func2 = eval_operand(n->in(2), eval_map); 2681 func3 = eval_operand(n->in(3), eval_map); 2682 } 2683 } 2684 2685 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) { 2686 assert(inputs.size() <= 3, "sanity"); 2687 ResourceMark rm; 2688 uint res = 0; 2689 ResourceHashtable<Node*,uint> eval_map; 2690 2691 // Populate precomputed functions for inputs. 2692 // Each input corresponds to one column of 3 input truth-table. 2693 uint input_funcs[] = { 0xAA, // (_, _, c) -> c 2694 0xCC, // (_, b, _) -> b 2695 0xF0 }; // (a, _, _) -> a 2696 for (uint i = 0; i < inputs.size(); i++) { 2697 eval_map.put(inputs.at(i), input_funcs[2-i]); 2698 } 2699 2700 for (uint i = 0; i < partition.size(); i++) { 2701 Node* n = partition.at(i); 2702 2703 uint func1 = 0, func2 = 0, func3 = 0; 2704 eval_operands(n, func1, func2, func3, eval_map); 2705 2706 switch (n->Opcode()) { 2707 case Op_OrV: 2708 assert(func3 == 0, "not binary"); 2709 res = func1 | func2; 2710 break; 2711 case Op_AndV: 2712 assert(func3 == 0, "not binary"); 2713 res = func1 & func2; 2714 break; 2715 case Op_XorV: 2716 if (VectorNode::is_vector_bitwise_not_pattern(n)) { 2717 assert(func2 == 0 && func3 == 0, "not unary"); 2718 res = (~func1) & 0xFF; 2719 } else { 2720 assert(func3 == 0, "not binary"); 2721 res = func1 ^ func2; 2722 } 2723 break; 2724 case Op_MacroLogicV: 2725 // Ordering of inputs may change during evaluation of sub-tree 2726 // containing MacroLogic node as a child node, thus a re-evaluation 2727 // makes sure that function is evaluated in context of current 2728 // inputs. 2729 res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3); 2730 break; 2731 2732 default: assert(false, "not supported: %s", n->Name()); 2733 } 2734 assert(res <= 0xFF, "invalid"); 2735 eval_map.put(n, res); 2736 } 2737 return res; 2738 } 2739 2740 // Criteria under which nodes gets packed into a macro logic node:- 2741 // 1) Parent and both child nodes are all unmasked or masked with 2742 // same predicates. 2743 // 2) Masked parent can be packed with left child if it is predicated 2744 // and both have same predicates. 2745 // 3) Masked parent can be packed with right child if its un-predicated 2746 // or has matching predication condition. 2747 // 4) An unmasked parent can be packed with an unmasked child. 2748 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) { 2749 assert(partition.size() == 0, "not empty"); 2750 assert(inputs.size() == 0, "not empty"); 2751 if (is_vector_ternary_bitwise_op(n)) { 2752 return false; 2753 } 2754 2755 bool is_unary_op = is_vector_unary_bitwise_op(n); 2756 if (is_unary_op) { 2757 assert(collect_unique_inputs(n, inputs) == 1, "not unary"); 2758 return false; // too few inputs 2759 } 2760 2761 bool pack_left_child = true; 2762 bool pack_right_child = true; 2763 2764 bool left_child_LOP = is_vector_bitwise_op(n->in(1)); 2765 bool right_child_LOP = is_vector_bitwise_op(n->in(2)); 2766 2767 int left_child_input_cnt = 0; 2768 int right_child_input_cnt = 0; 2769 2770 bool parent_is_predicated = n->is_predicated_vector(); 2771 bool left_child_predicated = n->in(1)->is_predicated_vector(); 2772 bool right_child_predicated = n->in(2)->is_predicated_vector(); 2773 2774 Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : nullptr; 2775 Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2776 Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : nullptr; 2777 2778 do { 2779 if (pack_left_child && left_child_LOP && 2780 ((!parent_is_predicated && !left_child_predicated) || 2781 ((parent_is_predicated && left_child_predicated && 2782 parent_pred == left_child_pred)))) { 2783 partition.push(n->in(1)); 2784 left_child_input_cnt = collect_unique_inputs(n->in(1), inputs); 2785 } else { 2786 inputs.push(n->in(1)); 2787 left_child_input_cnt = 1; 2788 } 2789 2790 if (pack_right_child && right_child_LOP && 2791 (!right_child_predicated || 2792 (right_child_predicated && parent_is_predicated && 2793 parent_pred == right_child_pred))) { 2794 partition.push(n->in(2)); 2795 right_child_input_cnt = collect_unique_inputs(n->in(2), inputs); 2796 } else { 2797 inputs.push(n->in(2)); 2798 right_child_input_cnt = 1; 2799 } 2800 2801 if (inputs.size() > 3) { 2802 assert(partition.size() > 0, ""); 2803 inputs.clear(); 2804 partition.clear(); 2805 if (left_child_input_cnt > right_child_input_cnt) { 2806 pack_left_child = false; 2807 } else { 2808 pack_right_child = false; 2809 } 2810 } else { 2811 break; 2812 } 2813 } while(true); 2814 2815 if(partition.size()) { 2816 partition.push(n); 2817 } 2818 2819 return (partition.size() == 2 || partition.size() == 3) && 2820 (inputs.size() == 2 || inputs.size() == 3); 2821 } 2822 2823 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) { 2824 assert(is_vector_bitwise_op(n), "not a root"); 2825 2826 visited.set(n->_idx); 2827 2828 // 1) Do a DFS walk over the logic cone. 2829 for (uint i = 1; i < n->req(); i++) { 2830 Node* in = n->in(i); 2831 if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) { 2832 process_logic_cone_root(igvn, in, visited); 2833 } 2834 } 2835 2836 // 2) Bottom up traversal: Merge node[s] with 2837 // the parent to form macro logic node. 2838 Unique_Node_List partition; 2839 Unique_Node_List inputs; 2840 if (compute_logic_cone(n, partition, inputs)) { 2841 const TypeVect* vt = n->bottom_type()->is_vect(); 2842 Node* pn = partition.at(partition.size() - 1); 2843 Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : nullptr; 2844 if (mask == nullptr || 2845 Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) { 2846 Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs); 2847 #ifdef ASSERT 2848 if (TraceNewVectors) { 2849 tty->print("new Vector node: "); 2850 macro_logic->dump(); 2851 } 2852 #endif 2853 igvn.replace_node(n, macro_logic); 2854 } 2855 } 2856 } 2857 2858 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) { 2859 ResourceMark rm; 2860 if (Matcher::match_rule_supported(Op_MacroLogicV)) { 2861 Unique_Node_List list; 2862 collect_logic_cone_roots(list); 2863 2864 while (list.size() > 0) { 2865 Node* n = list.pop(); 2866 const TypeVect* vt = n->bottom_type()->is_vect(); 2867 bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()); 2868 if (supported) { 2869 VectorSet visited(comp_arena()); 2870 process_logic_cone_root(igvn, n, visited); 2871 } 2872 } 2873 } 2874 } 2875 2876 //------------------------------Code_Gen--------------------------------------- 2877 // Given a graph, generate code for it 2878 void Compile::Code_Gen() { 2879 if (failing()) { 2880 return; 2881 } 2882 2883 // Perform instruction selection. You might think we could reclaim Matcher 2884 // memory PDQ, but actually the Matcher is used in generating spill code. 2885 // Internals of the Matcher (including some VectorSets) must remain live 2886 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2887 // set a bit in reclaimed memory. 2888 2889 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2890 // nodes. Mapping is only valid at the root of each matched subtree. 2891 NOT_PRODUCT( verify_graph_edges(); ) 2892 2893 Matcher matcher; 2894 _matcher = &matcher; 2895 { 2896 TracePhase tp("matcher", &timers[_t_matcher]); 2897 matcher.match(); 2898 if (failing()) { 2899 return; 2900 } 2901 } 2902 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2903 // nodes. Mapping is only valid at the root of each matched subtree. 2904 NOT_PRODUCT( verify_graph_edges(); ) 2905 2906 // If you have too many nodes, or if matching has failed, bail out 2907 check_node_count(0, "out of nodes matching instructions"); 2908 if (failing()) { 2909 return; 2910 } 2911 2912 print_method(PHASE_MATCHING, 2); 2913 2914 // Build a proper-looking CFG 2915 PhaseCFG cfg(node_arena(), root(), matcher); 2916 _cfg = &cfg; 2917 { 2918 TracePhase tp("scheduler", &timers[_t_scheduler]); 2919 bool success = cfg.do_global_code_motion(); 2920 if (!success) { 2921 return; 2922 } 2923 2924 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2925 NOT_PRODUCT( verify_graph_edges(); ) 2926 cfg.verify(); 2927 } 2928 2929 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2930 _regalloc = ®alloc; 2931 { 2932 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2933 // Perform register allocation. After Chaitin, use-def chains are 2934 // no longer accurate (at spill code) and so must be ignored. 2935 // Node->LRG->reg mappings are still accurate. 2936 _regalloc->Register_Allocate(); 2937 2938 // Bail out if the allocator builds too many nodes 2939 if (failing()) { 2940 return; 2941 } 2942 } 2943 2944 // Prior to register allocation we kept empty basic blocks in case the 2945 // the allocator needed a place to spill. After register allocation we 2946 // are not adding any new instructions. If any basic block is empty, we 2947 // can now safely remove it. 2948 { 2949 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2950 cfg.remove_empty_blocks(); 2951 if (do_freq_based_layout()) { 2952 PhaseBlockLayout layout(cfg); 2953 } else { 2954 cfg.set_loop_alignment(); 2955 } 2956 cfg.fixup_flow(); 2957 cfg.remove_unreachable_blocks(); 2958 cfg.verify_dominator_tree(); 2959 } 2960 2961 // Apply peephole optimizations 2962 if( OptoPeephole ) { 2963 TracePhase tp("peephole", &timers[_t_peephole]); 2964 PhasePeephole peep( _regalloc, cfg); 2965 peep.do_transform(); 2966 } 2967 2968 // Do late expand if CPU requires this. 2969 if (Matcher::require_postalloc_expand) { 2970 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2971 cfg.postalloc_expand(_regalloc); 2972 } 2973 2974 // Convert Nodes to instruction bits in a buffer 2975 { 2976 TracePhase tp("output", &timers[_t_output]); 2977 PhaseOutput output; 2978 output.Output(); 2979 if (failing()) return; 2980 output.install(); 2981 } 2982 2983 print_method(PHASE_FINAL_CODE, 1); 2984 2985 // He's dead, Jim. 2986 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2987 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2988 } 2989 2990 //------------------------------Final_Reshape_Counts--------------------------- 2991 // This class defines counters to help identify when a method 2992 // may/must be executed using hardware with only 24-bit precision. 2993 struct Final_Reshape_Counts : public StackObj { 2994 int _call_count; // count non-inlined 'common' calls 2995 int _float_count; // count float ops requiring 24-bit precision 2996 int _double_count; // count double ops requiring more precision 2997 int _java_call_count; // count non-inlined 'java' calls 2998 int _inner_loop_count; // count loops which need alignment 2999 VectorSet _visited; // Visitation flags 3000 Node_List _tests; // Set of IfNodes & PCTableNodes 3001 3002 Final_Reshape_Counts() : 3003 _call_count(0), _float_count(0), _double_count(0), 3004 _java_call_count(0), _inner_loop_count(0) { } 3005 3006 void inc_call_count () { _call_count ++; } 3007 void inc_float_count () { _float_count ++; } 3008 void inc_double_count() { _double_count++; } 3009 void inc_java_call_count() { _java_call_count++; } 3010 void inc_inner_loop_count() { _inner_loop_count++; } 3011 3012 int get_call_count () const { return _call_count ; } 3013 int get_float_count () const { return _float_count ; } 3014 int get_double_count() const { return _double_count; } 3015 int get_java_call_count() const { return _java_call_count; } 3016 int get_inner_loop_count() const { return _inner_loop_count; } 3017 }; 3018 3019 // Eliminate trivially redundant StoreCMs and accumulate their 3020 // precedence edges. 3021 void Compile::eliminate_redundant_card_marks(Node* n) { 3022 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3023 if (n->in(MemNode::Address)->outcnt() > 1) { 3024 // There are multiple users of the same address so it might be 3025 // possible to eliminate some of the StoreCMs 3026 Node* mem = n->in(MemNode::Memory); 3027 Node* adr = n->in(MemNode::Address); 3028 Node* val = n->in(MemNode::ValueIn); 3029 Node* prev = n; 3030 bool done = false; 3031 // Walk the chain of StoreCMs eliminating ones that match. As 3032 // long as it's a chain of single users then the optimization is 3033 // safe. Eliminating partially redundant StoreCMs would require 3034 // cloning copies down the other paths. 3035 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3036 if (adr == mem->in(MemNode::Address) && 3037 val == mem->in(MemNode::ValueIn)) { 3038 // redundant StoreCM 3039 if (mem->req() > MemNode::OopStore) { 3040 // Hasn't been processed by this code yet. 3041 n->add_prec(mem->in(MemNode::OopStore)); 3042 } else { 3043 // Already converted to precedence edge 3044 for (uint i = mem->req(); i < mem->len(); i++) { 3045 // Accumulate any precedence edges 3046 if (mem->in(i) != nullptr) { 3047 n->add_prec(mem->in(i)); 3048 } 3049 } 3050 // Everything above this point has been processed. 3051 done = true; 3052 } 3053 // Eliminate the previous StoreCM 3054 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3055 assert(mem->outcnt() == 0, "should be dead"); 3056 mem->disconnect_inputs(this); 3057 } else { 3058 prev = mem; 3059 } 3060 mem = prev->in(MemNode::Memory); 3061 } 3062 } 3063 } 3064 3065 //------------------------------final_graph_reshaping_impl---------------------- 3066 // Implement items 1-5 from final_graph_reshaping below. 3067 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3068 3069 if ( n->outcnt() == 0 ) return; // dead node 3070 uint nop = n->Opcode(); 3071 3072 // Check for 2-input instruction with "last use" on right input. 3073 // Swap to left input. Implements item (2). 3074 if( n->req() == 3 && // two-input instruction 3075 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3076 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3077 n->in(2)->outcnt() == 1 &&// right use IS a last use 3078 !n->in(2)->is_Con() ) { // right use is not a constant 3079 // Check for commutative opcode 3080 switch( nop ) { 3081 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3082 case Op_MaxI: case Op_MaxL: case Op_MaxF: case Op_MaxD: 3083 case Op_MinI: case Op_MinL: case Op_MinF: case Op_MinD: 3084 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3085 case Op_AndL: case Op_XorL: case Op_OrL: 3086 case Op_AndI: case Op_XorI: case Op_OrI: { 3087 // Move "last use" input to left by swapping inputs 3088 n->swap_edges(1, 2); 3089 break; 3090 } 3091 default: 3092 break; 3093 } 3094 } 3095 3096 #ifdef ASSERT 3097 if( n->is_Mem() ) { 3098 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3099 assert( n->in(0) != nullptr || alias_idx != Compile::AliasIdxRaw || 3100 // oop will be recorded in oop map if load crosses safepoint 3101 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3102 LoadNode::is_immutable_value(n->in(MemNode::Address))), 3103 "raw memory operations should have control edge"); 3104 } 3105 if (n->is_MemBar()) { 3106 MemBarNode* mb = n->as_MemBar(); 3107 if (mb->trailing_store() || mb->trailing_load_store()) { 3108 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3109 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3110 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3111 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3112 } else if (mb->leading()) { 3113 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3114 } 3115 } 3116 #endif 3117 // Count FPU ops and common calls, implements item (3) 3118 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes); 3119 if (!gc_handled) { 3120 final_graph_reshaping_main_switch(n, frc, nop, dead_nodes); 3121 } 3122 3123 // Collect CFG split points 3124 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3125 frc._tests.push(n); 3126 } 3127 } 3128 3129 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) { 3130 switch( nop ) { 3131 // Count all float operations that may use FPU 3132 case Op_AddF: 3133 case Op_SubF: 3134 case Op_MulF: 3135 case Op_DivF: 3136 case Op_NegF: 3137 case Op_ModF: 3138 case Op_ConvI2F: 3139 case Op_ConF: 3140 case Op_CmpF: 3141 case Op_CmpF3: 3142 case Op_StoreF: 3143 case Op_LoadF: 3144 // case Op_ConvL2F: // longs are split into 32-bit halves 3145 frc.inc_float_count(); 3146 break; 3147 3148 case Op_ConvF2D: 3149 case Op_ConvD2F: 3150 frc.inc_float_count(); 3151 frc.inc_double_count(); 3152 break; 3153 3154 // Count all double operations that may use FPU 3155 case Op_AddD: 3156 case Op_SubD: 3157 case Op_MulD: 3158 case Op_DivD: 3159 case Op_NegD: 3160 case Op_ModD: 3161 case Op_ConvI2D: 3162 case Op_ConvD2I: 3163 // case Op_ConvL2D: // handled by leaf call 3164 // case Op_ConvD2L: // handled by leaf call 3165 case Op_ConD: 3166 case Op_CmpD: 3167 case Op_CmpD3: 3168 case Op_StoreD: 3169 case Op_LoadD: 3170 case Op_LoadD_unaligned: 3171 frc.inc_double_count(); 3172 break; 3173 case Op_Opaque1: // Remove Opaque Nodes before matching 3174 case Op_Opaque3: 3175 n->subsume_by(n->in(1), this); 3176 break; 3177 case Op_CallStaticJava: 3178 case Op_CallJava: 3179 case Op_CallDynamicJava: 3180 frc.inc_java_call_count(); // Count java call site; 3181 case Op_CallRuntime: 3182 case Op_CallLeaf: 3183 case Op_CallLeafVector: 3184 case Op_CallLeafNoFP: { 3185 assert (n->is_Call(), ""); 3186 CallNode *call = n->as_Call(); 3187 // Count call sites where the FP mode bit would have to be flipped. 3188 // Do not count uncommon runtime calls: 3189 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3190 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3191 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3192 frc.inc_call_count(); // Count the call site 3193 } else { // See if uncommon argument is shared 3194 Node *n = call->in(TypeFunc::Parms); 3195 int nop = n->Opcode(); 3196 // Clone shared simple arguments to uncommon calls, item (1). 3197 if (n->outcnt() > 1 && 3198 !n->is_Proj() && 3199 nop != Op_CreateEx && 3200 nop != Op_CheckCastPP && 3201 nop != Op_DecodeN && 3202 nop != Op_DecodeNKlass && 3203 !n->is_Mem() && 3204 !n->is_Phi()) { 3205 Node *x = n->clone(); 3206 call->set_req(TypeFunc::Parms, x); 3207 } 3208 } 3209 break; 3210 } 3211 3212 case Op_StoreCM: 3213 { 3214 // Convert OopStore dependence into precedence edge 3215 Node* prec = n->in(MemNode::OopStore); 3216 n->del_req(MemNode::OopStore); 3217 n->add_prec(prec); 3218 eliminate_redundant_card_marks(n); 3219 } 3220 3221 // fall through 3222 3223 case Op_StoreB: 3224 case Op_StoreC: 3225 case Op_StoreI: 3226 case Op_StoreL: 3227 case Op_CompareAndSwapB: 3228 case Op_CompareAndSwapS: 3229 case Op_CompareAndSwapI: 3230 case Op_CompareAndSwapL: 3231 case Op_CompareAndSwapP: 3232 case Op_CompareAndSwapN: 3233 case Op_WeakCompareAndSwapB: 3234 case Op_WeakCompareAndSwapS: 3235 case Op_WeakCompareAndSwapI: 3236 case Op_WeakCompareAndSwapL: 3237 case Op_WeakCompareAndSwapP: 3238 case Op_WeakCompareAndSwapN: 3239 case Op_CompareAndExchangeB: 3240 case Op_CompareAndExchangeS: 3241 case Op_CompareAndExchangeI: 3242 case Op_CompareAndExchangeL: 3243 case Op_CompareAndExchangeP: 3244 case Op_CompareAndExchangeN: 3245 case Op_GetAndAddS: 3246 case Op_GetAndAddB: 3247 case Op_GetAndAddI: 3248 case Op_GetAndAddL: 3249 case Op_GetAndSetS: 3250 case Op_GetAndSetB: 3251 case Op_GetAndSetI: 3252 case Op_GetAndSetL: 3253 case Op_GetAndSetP: 3254 case Op_GetAndSetN: 3255 case Op_StoreP: 3256 case Op_StoreN: 3257 case Op_StoreNKlass: 3258 case Op_LoadB: 3259 case Op_LoadUB: 3260 case Op_LoadUS: 3261 case Op_LoadI: 3262 case Op_LoadKlass: 3263 case Op_LoadNKlass: 3264 case Op_LoadL: 3265 case Op_LoadL_unaligned: 3266 case Op_LoadP: 3267 case Op_LoadN: 3268 case Op_LoadRange: 3269 case Op_LoadS: 3270 break; 3271 3272 case Op_AddP: { // Assert sane base pointers 3273 Node *addp = n->in(AddPNode::Address); 3274 assert( !addp->is_AddP() || 3275 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3276 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3277 "Base pointers must match (addp %u)", addp->_idx ); 3278 #ifdef _LP64 3279 if ((UseCompressedOops || UseCompressedClassPointers) && 3280 addp->Opcode() == Op_ConP && 3281 addp == n->in(AddPNode::Base) && 3282 n->in(AddPNode::Offset)->is_Con()) { 3283 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3284 // on the platform and on the compressed oops mode. 3285 // Use addressing with narrow klass to load with offset on x86. 3286 // Some platforms can use the constant pool to load ConP. 3287 // Do this transformation here since IGVN will convert ConN back to ConP. 3288 const Type* t = addp->bottom_type(); 3289 bool is_oop = t->isa_oopptr() != nullptr; 3290 bool is_klass = t->isa_klassptr() != nullptr; 3291 3292 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3293 (is_klass && Matcher::const_klass_prefer_decode())) { 3294 Node* nn = nullptr; 3295 3296 int op = is_oop ? Op_ConN : Op_ConNKlass; 3297 3298 // Look for existing ConN node of the same exact type. 3299 Node* r = root(); 3300 uint cnt = r->outcnt(); 3301 for (uint i = 0; i < cnt; i++) { 3302 Node* m = r->raw_out(i); 3303 if (m!= nullptr && m->Opcode() == op && 3304 m->bottom_type()->make_ptr() == t) { 3305 nn = m; 3306 break; 3307 } 3308 } 3309 if (nn != nullptr) { 3310 // Decode a narrow oop to match address 3311 // [R12 + narrow_oop_reg<<3 + offset] 3312 if (is_oop) { 3313 nn = new DecodeNNode(nn, t); 3314 } else { 3315 nn = new DecodeNKlassNode(nn, t); 3316 } 3317 // Check for succeeding AddP which uses the same Base. 3318 // Otherwise we will run into the assertion above when visiting that guy. 3319 for (uint i = 0; i < n->outcnt(); ++i) { 3320 Node *out_i = n->raw_out(i); 3321 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3322 out_i->set_req(AddPNode::Base, nn); 3323 #ifdef ASSERT 3324 for (uint j = 0; j < out_i->outcnt(); ++j) { 3325 Node *out_j = out_i->raw_out(j); 3326 assert(out_j == nullptr || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3327 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3328 } 3329 #endif 3330 } 3331 } 3332 n->set_req(AddPNode::Base, nn); 3333 n->set_req(AddPNode::Address, nn); 3334 if (addp->outcnt() == 0) { 3335 addp->disconnect_inputs(this); 3336 } 3337 } 3338 } 3339 } 3340 #endif 3341 break; 3342 } 3343 3344 case Op_CastPP: { 3345 // Remove CastPP nodes to gain more freedom during scheduling but 3346 // keep the dependency they encode as control or precedence edges 3347 // (if control is set already) on memory operations. Some CastPP 3348 // nodes don't have a control (don't carry a dependency): skip 3349 // those. 3350 if (n->in(0) != nullptr) { 3351 ResourceMark rm; 3352 Unique_Node_List wq; 3353 wq.push(n); 3354 for (uint next = 0; next < wq.size(); ++next) { 3355 Node *m = wq.at(next); 3356 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3357 Node* use = m->fast_out(i); 3358 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3359 use->ensure_control_or_add_prec(n->in(0)); 3360 } else { 3361 switch(use->Opcode()) { 3362 case Op_AddP: 3363 case Op_DecodeN: 3364 case Op_DecodeNKlass: 3365 case Op_CheckCastPP: 3366 case Op_CastPP: 3367 wq.push(use); 3368 break; 3369 } 3370 } 3371 } 3372 } 3373 } 3374 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3375 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3376 Node* in1 = n->in(1); 3377 const Type* t = n->bottom_type(); 3378 Node* new_in1 = in1->clone(); 3379 new_in1->as_DecodeN()->set_type(t); 3380 3381 if (!Matcher::narrow_oop_use_complex_address()) { 3382 // 3383 // x86, ARM and friends can handle 2 adds in addressing mode 3384 // and Matcher can fold a DecodeN node into address by using 3385 // a narrow oop directly and do implicit null check in address: 3386 // 3387 // [R12 + narrow_oop_reg<<3 + offset] 3388 // NullCheck narrow_oop_reg 3389 // 3390 // On other platforms (Sparc) we have to keep new DecodeN node and 3391 // use it to do implicit null check in address: 3392 // 3393 // decode_not_null narrow_oop_reg, base_reg 3394 // [base_reg + offset] 3395 // NullCheck base_reg 3396 // 3397 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3398 // to keep the information to which null check the new DecodeN node 3399 // corresponds to use it as value in implicit_null_check(). 3400 // 3401 new_in1->set_req(0, n->in(0)); 3402 } 3403 3404 n->subsume_by(new_in1, this); 3405 if (in1->outcnt() == 0) { 3406 in1->disconnect_inputs(this); 3407 } 3408 } else { 3409 n->subsume_by(n->in(1), this); 3410 if (n->outcnt() == 0) { 3411 n->disconnect_inputs(this); 3412 } 3413 } 3414 break; 3415 } 3416 #ifdef _LP64 3417 case Op_CmpP: 3418 // Do this transformation here to preserve CmpPNode::sub() and 3419 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3420 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3421 Node* in1 = n->in(1); 3422 Node* in2 = n->in(2); 3423 if (!in1->is_DecodeNarrowPtr()) { 3424 in2 = in1; 3425 in1 = n->in(2); 3426 } 3427 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3428 3429 Node* new_in2 = nullptr; 3430 if (in2->is_DecodeNarrowPtr()) { 3431 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3432 new_in2 = in2->in(1); 3433 } else if (in2->Opcode() == Op_ConP) { 3434 const Type* t = in2->bottom_type(); 3435 if (t == TypePtr::NULL_PTR) { 3436 assert(in1->is_DecodeN(), "compare klass to null?"); 3437 // Don't convert CmpP null check into CmpN if compressed 3438 // oops implicit null check is not generated. 3439 // This will allow to generate normal oop implicit null check. 3440 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3441 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3442 // 3443 // This transformation together with CastPP transformation above 3444 // will generated code for implicit null checks for compressed oops. 3445 // 3446 // The original code after Optimize() 3447 // 3448 // LoadN memory, narrow_oop_reg 3449 // decode narrow_oop_reg, base_reg 3450 // CmpP base_reg, nullptr 3451 // CastPP base_reg // NotNull 3452 // Load [base_reg + offset], val_reg 3453 // 3454 // after these transformations will be 3455 // 3456 // LoadN memory, narrow_oop_reg 3457 // CmpN narrow_oop_reg, nullptr 3458 // decode_not_null narrow_oop_reg, base_reg 3459 // Load [base_reg + offset], val_reg 3460 // 3461 // and the uncommon path (== nullptr) will use narrow_oop_reg directly 3462 // since narrow oops can be used in debug info now (see the code in 3463 // final_graph_reshaping_walk()). 3464 // 3465 // At the end the code will be matched to 3466 // on x86: 3467 // 3468 // Load_narrow_oop memory, narrow_oop_reg 3469 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3470 // NullCheck narrow_oop_reg 3471 // 3472 // and on sparc: 3473 // 3474 // Load_narrow_oop memory, narrow_oop_reg 3475 // decode_not_null narrow_oop_reg, base_reg 3476 // Load [base_reg + offset], val_reg 3477 // NullCheck base_reg 3478 // 3479 } else if (t->isa_oopptr()) { 3480 new_in2 = ConNode::make(t->make_narrowoop()); 3481 } else if (t->isa_klassptr()) { 3482 new_in2 = ConNode::make(t->make_narrowklass()); 3483 } 3484 } 3485 if (new_in2 != nullptr) { 3486 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3487 n->subsume_by(cmpN, this); 3488 if (in1->outcnt() == 0) { 3489 in1->disconnect_inputs(this); 3490 } 3491 if (in2->outcnt() == 0) { 3492 in2->disconnect_inputs(this); 3493 } 3494 } 3495 } 3496 break; 3497 3498 case Op_DecodeN: 3499 case Op_DecodeNKlass: 3500 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3501 // DecodeN could be pinned when it can't be fold into 3502 // an address expression, see the code for Op_CastPP above. 3503 assert(n->in(0) == nullptr || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3504 break; 3505 3506 case Op_EncodeP: 3507 case Op_EncodePKlass: { 3508 Node* in1 = n->in(1); 3509 if (in1->is_DecodeNarrowPtr()) { 3510 n->subsume_by(in1->in(1), this); 3511 } else if (in1->Opcode() == Op_ConP) { 3512 const Type* t = in1->bottom_type(); 3513 if (t == TypePtr::NULL_PTR) { 3514 assert(t->isa_oopptr(), "null klass?"); 3515 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3516 } else if (t->isa_oopptr()) { 3517 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3518 } else if (t->isa_klassptr()) { 3519 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3520 } 3521 } 3522 if (in1->outcnt() == 0) { 3523 in1->disconnect_inputs(this); 3524 } 3525 break; 3526 } 3527 3528 case Op_Proj: { 3529 if (OptimizeStringConcat || IncrementalInline) { 3530 ProjNode* proj = n->as_Proj(); 3531 if (proj->_is_io_use) { 3532 assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, ""); 3533 // Separate projections were used for the exception path which 3534 // are normally removed by a late inline. If it wasn't inlined 3535 // then they will hang around and should just be replaced with 3536 // the original one. Merge them. 3537 Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/); 3538 if (non_io_proj != nullptr) { 3539 proj->subsume_by(non_io_proj , this); 3540 } 3541 } 3542 } 3543 break; 3544 } 3545 3546 case Op_Phi: 3547 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3548 // The EncodeP optimization may create Phi with the same edges 3549 // for all paths. It is not handled well by Register Allocator. 3550 Node* unique_in = n->in(1); 3551 assert(unique_in != nullptr, ""); 3552 uint cnt = n->req(); 3553 for (uint i = 2; i < cnt; i++) { 3554 Node* m = n->in(i); 3555 assert(m != nullptr, ""); 3556 if (unique_in != m) 3557 unique_in = nullptr; 3558 } 3559 if (unique_in != nullptr) { 3560 n->subsume_by(unique_in, this); 3561 } 3562 } 3563 break; 3564 3565 #endif 3566 3567 #ifdef ASSERT 3568 case Op_CastII: 3569 // Verify that all range check dependent CastII nodes were removed. 3570 if (n->isa_CastII()->has_range_check()) { 3571 n->dump(3); 3572 assert(false, "Range check dependent CastII node was not removed"); 3573 } 3574 break; 3575 #endif 3576 3577 case Op_ModI: 3578 if (UseDivMod) { 3579 // Check if a%b and a/b both exist 3580 Node* d = n->find_similar(Op_DivI); 3581 if (d) { 3582 // Replace them with a fused divmod if supported 3583 if (Matcher::has_match_rule(Op_DivModI)) { 3584 DivModINode* divmod = DivModINode::make(n); 3585 d->subsume_by(divmod->div_proj(), this); 3586 n->subsume_by(divmod->mod_proj(), this); 3587 } else { 3588 // replace a%b with a-((a/b)*b) 3589 Node* mult = new MulINode(d, d->in(2)); 3590 Node* sub = new SubINode(d->in(1), mult); 3591 n->subsume_by(sub, this); 3592 } 3593 } 3594 } 3595 break; 3596 3597 case Op_ModL: 3598 if (UseDivMod) { 3599 // Check if a%b and a/b both exist 3600 Node* d = n->find_similar(Op_DivL); 3601 if (d) { 3602 // Replace them with a fused divmod if supported 3603 if (Matcher::has_match_rule(Op_DivModL)) { 3604 DivModLNode* divmod = DivModLNode::make(n); 3605 d->subsume_by(divmod->div_proj(), this); 3606 n->subsume_by(divmod->mod_proj(), this); 3607 } else { 3608 // replace a%b with a-((a/b)*b) 3609 Node* mult = new MulLNode(d, d->in(2)); 3610 Node* sub = new SubLNode(d->in(1), mult); 3611 n->subsume_by(sub, this); 3612 } 3613 } 3614 } 3615 break; 3616 3617 case Op_UModI: 3618 if (UseDivMod) { 3619 // Check if a%b and a/b both exist 3620 Node* d = n->find_similar(Op_UDivI); 3621 if (d) { 3622 // Replace them with a fused unsigned divmod if supported 3623 if (Matcher::has_match_rule(Op_UDivModI)) { 3624 UDivModINode* divmod = UDivModINode::make(n); 3625 d->subsume_by(divmod->div_proj(), this); 3626 n->subsume_by(divmod->mod_proj(), this); 3627 } else { 3628 // replace a%b with a-((a/b)*b) 3629 Node* mult = new MulINode(d, d->in(2)); 3630 Node* sub = new SubINode(d->in(1), mult); 3631 n->subsume_by(sub, this); 3632 } 3633 } 3634 } 3635 break; 3636 3637 case Op_UModL: 3638 if (UseDivMod) { 3639 // Check if a%b and a/b both exist 3640 Node* d = n->find_similar(Op_UDivL); 3641 if (d) { 3642 // Replace them with a fused unsigned divmod if supported 3643 if (Matcher::has_match_rule(Op_UDivModL)) { 3644 UDivModLNode* divmod = UDivModLNode::make(n); 3645 d->subsume_by(divmod->div_proj(), this); 3646 n->subsume_by(divmod->mod_proj(), this); 3647 } else { 3648 // replace a%b with a-((a/b)*b) 3649 Node* mult = new MulLNode(d, d->in(2)); 3650 Node* sub = new SubLNode(d->in(1), mult); 3651 n->subsume_by(sub, this); 3652 } 3653 } 3654 } 3655 break; 3656 3657 case Op_LoadVector: 3658 case Op_StoreVector: 3659 case Op_LoadVectorGather: 3660 case Op_StoreVectorScatter: 3661 case Op_LoadVectorGatherMasked: 3662 case Op_StoreVectorScatterMasked: 3663 case Op_VectorCmpMasked: 3664 case Op_VectorMaskGen: 3665 case Op_LoadVectorMasked: 3666 case Op_StoreVectorMasked: 3667 break; 3668 3669 case Op_AddReductionVI: 3670 case Op_AddReductionVL: 3671 case Op_AddReductionVF: 3672 case Op_AddReductionVD: 3673 case Op_MulReductionVI: 3674 case Op_MulReductionVL: 3675 case Op_MulReductionVF: 3676 case Op_MulReductionVD: 3677 case Op_MinReductionV: 3678 case Op_MaxReductionV: 3679 case Op_AndReductionV: 3680 case Op_OrReductionV: 3681 case Op_XorReductionV: 3682 break; 3683 3684 case Op_PackB: 3685 case Op_PackS: 3686 case Op_PackI: 3687 case Op_PackF: 3688 case Op_PackL: 3689 case Op_PackD: 3690 if (n->req()-1 > 2) { 3691 // Replace many operand PackNodes with a binary tree for matching 3692 PackNode* p = (PackNode*) n; 3693 Node* btp = p->binary_tree_pack(1, n->req()); 3694 n->subsume_by(btp, this); 3695 } 3696 break; 3697 case Op_Loop: 3698 assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop"); 3699 case Op_CountedLoop: 3700 case Op_LongCountedLoop: 3701 case Op_OuterStripMinedLoop: 3702 if (n->as_Loop()->is_inner_loop()) { 3703 frc.inc_inner_loop_count(); 3704 } 3705 n->as_Loop()->verify_strip_mined(0); 3706 break; 3707 case Op_LShiftI: 3708 case Op_RShiftI: 3709 case Op_URShiftI: 3710 case Op_LShiftL: 3711 case Op_RShiftL: 3712 case Op_URShiftL: 3713 if (Matcher::need_masked_shift_count) { 3714 // The cpu's shift instructions don't restrict the count to the 3715 // lower 5/6 bits. We need to do the masking ourselves. 3716 Node* in2 = n->in(2); 3717 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3718 const TypeInt* t = in2->find_int_type(); 3719 if (t != nullptr && t->is_con()) { 3720 juint shift = t->get_con(); 3721 if (shift > mask) { // Unsigned cmp 3722 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3723 } 3724 } else { 3725 if (t == nullptr || t->_lo < 0 || t->_hi > (int)mask) { 3726 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3727 n->set_req(2, shift); 3728 } 3729 } 3730 if (in2->outcnt() == 0) { // Remove dead node 3731 in2->disconnect_inputs(this); 3732 } 3733 } 3734 break; 3735 case Op_MemBarStoreStore: 3736 case Op_MemBarRelease: 3737 // Break the link with AllocateNode: it is no longer useful and 3738 // confuses register allocation. 3739 if (n->req() > MemBarNode::Precedent) { 3740 n->set_req(MemBarNode::Precedent, top()); 3741 } 3742 break; 3743 case Op_MemBarAcquire: { 3744 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3745 // At parse time, the trailing MemBarAcquire for a volatile load 3746 // is created with an edge to the load. After optimizations, 3747 // that input may be a chain of Phis. If those phis have no 3748 // other use, then the MemBarAcquire keeps them alive and 3749 // register allocation can be confused. 3750 dead_nodes.push(n->in(MemBarNode::Precedent)); 3751 n->set_req(MemBarNode::Precedent, top()); 3752 } 3753 break; 3754 } 3755 case Op_Blackhole: 3756 break; 3757 case Op_RangeCheck: { 3758 RangeCheckNode* rc = n->as_RangeCheck(); 3759 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3760 n->subsume_by(iff, this); 3761 frc._tests.push(iff); 3762 break; 3763 } 3764 case Op_ConvI2L: { 3765 if (!Matcher::convi2l_type_required) { 3766 // Code generation on some platforms doesn't need accurate 3767 // ConvI2L types. Widening the type can help remove redundant 3768 // address computations. 3769 n->as_Type()->set_type(TypeLong::INT); 3770 ResourceMark rm; 3771 Unique_Node_List wq; 3772 wq.push(n); 3773 for (uint next = 0; next < wq.size(); next++) { 3774 Node *m = wq.at(next); 3775 3776 for(;;) { 3777 // Loop over all nodes with identical inputs edges as m 3778 Node* k = m->find_similar(m->Opcode()); 3779 if (k == nullptr) { 3780 break; 3781 } 3782 // Push their uses so we get a chance to remove node made 3783 // redundant 3784 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3785 Node* u = k->fast_out(i); 3786 if (u->Opcode() == Op_LShiftL || 3787 u->Opcode() == Op_AddL || 3788 u->Opcode() == Op_SubL || 3789 u->Opcode() == Op_AddP) { 3790 wq.push(u); 3791 } 3792 } 3793 // Replace all nodes with identical edges as m with m 3794 k->subsume_by(m, this); 3795 } 3796 } 3797 } 3798 break; 3799 } 3800 case Op_CmpUL: { 3801 if (!Matcher::has_match_rule(Op_CmpUL)) { 3802 // No support for unsigned long comparisons 3803 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3804 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3805 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3806 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3807 Node* andl = new AndLNode(orl, remove_sign_mask); 3808 Node* cmp = new CmpLNode(andl, n->in(2)); 3809 n->subsume_by(cmp, this); 3810 } 3811 break; 3812 } 3813 default: 3814 assert(!n->is_Call(), ""); 3815 assert(!n->is_Mem(), ""); 3816 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3817 break; 3818 } 3819 } 3820 3821 //------------------------------final_graph_reshaping_walk--------------------- 3822 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3823 // requires that the walk visits a node's inputs before visiting the node. 3824 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) { 3825 Unique_Node_List sfpt; 3826 3827 frc._visited.set(root->_idx); // first, mark node as visited 3828 uint cnt = root->req(); 3829 Node *n = root; 3830 uint i = 0; 3831 while (true) { 3832 if (i < cnt) { 3833 // Place all non-visited non-null inputs onto stack 3834 Node* m = n->in(i); 3835 ++i; 3836 if (m != nullptr && !frc._visited.test_set(m->_idx)) { 3837 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != nullptr) { 3838 // compute worst case interpreter size in case of a deoptimization 3839 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3840 3841 sfpt.push(m); 3842 } 3843 cnt = m->req(); 3844 nstack.push(n, i); // put on stack parent and next input's index 3845 n = m; 3846 i = 0; 3847 } 3848 } else { 3849 // Now do post-visit work 3850 final_graph_reshaping_impl(n, frc, dead_nodes); 3851 if (nstack.is_empty()) 3852 break; // finished 3853 n = nstack.node(); // Get node from stack 3854 cnt = n->req(); 3855 i = nstack.index(); 3856 nstack.pop(); // Shift to the next node on stack 3857 } 3858 } 3859 3860 // Skip next transformation if compressed oops are not used. 3861 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3862 (!UseCompressedOops && !UseCompressedClassPointers)) 3863 return; 3864 3865 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3866 // It could be done for an uncommon traps or any safepoints/calls 3867 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3868 while (sfpt.size() > 0) { 3869 n = sfpt.pop(); 3870 JVMState *jvms = n->as_SafePoint()->jvms(); 3871 assert(jvms != nullptr, "sanity"); 3872 int start = jvms->debug_start(); 3873 int end = n->req(); 3874 bool is_uncommon = (n->is_CallStaticJava() && 3875 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3876 for (int j = start; j < end; j++) { 3877 Node* in = n->in(j); 3878 if (in->is_DecodeNarrowPtr()) { 3879 bool safe_to_skip = true; 3880 if (!is_uncommon ) { 3881 // Is it safe to skip? 3882 for (uint i = 0; i < in->outcnt(); i++) { 3883 Node* u = in->raw_out(i); 3884 if (!u->is_SafePoint() || 3885 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3886 safe_to_skip = false; 3887 } 3888 } 3889 } 3890 if (safe_to_skip) { 3891 n->set_req(j, in->in(1)); 3892 } 3893 if (in->outcnt() == 0) { 3894 in->disconnect_inputs(this); 3895 } 3896 } 3897 } 3898 } 3899 } 3900 3901 //------------------------------final_graph_reshaping-------------------------- 3902 // Final Graph Reshaping. 3903 // 3904 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3905 // and not commoned up and forced early. Must come after regular 3906 // optimizations to avoid GVN undoing the cloning. Clone constant 3907 // inputs to Loop Phis; these will be split by the allocator anyways. 3908 // Remove Opaque nodes. 3909 // (2) Move last-uses by commutative operations to the left input to encourage 3910 // Intel update-in-place two-address operations and better register usage 3911 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3912 // calls canonicalizing them back. 3913 // (3) Count the number of double-precision FP ops, single-precision FP ops 3914 // and call sites. On Intel, we can get correct rounding either by 3915 // forcing singles to memory (requires extra stores and loads after each 3916 // FP bytecode) or we can set a rounding mode bit (requires setting and 3917 // clearing the mode bit around call sites). The mode bit is only used 3918 // if the relative frequency of single FP ops to calls is low enough. 3919 // This is a key transform for SPEC mpeg_audio. 3920 // (4) Detect infinite loops; blobs of code reachable from above but not 3921 // below. Several of the Code_Gen algorithms fail on such code shapes, 3922 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3923 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3924 // Detection is by looking for IfNodes where only 1 projection is 3925 // reachable from below or CatchNodes missing some targets. 3926 // (5) Assert for insane oop offsets in debug mode. 3927 3928 bool Compile::final_graph_reshaping() { 3929 // an infinite loop may have been eliminated by the optimizer, 3930 // in which case the graph will be empty. 3931 if (root()->req() == 1) { 3932 // Do not compile method that is only a trivial infinite loop, 3933 // since the content of the loop may have been eliminated. 3934 record_method_not_compilable("trivial infinite loop"); 3935 return true; 3936 } 3937 3938 // Expensive nodes have their control input set to prevent the GVN 3939 // from freely commoning them. There's no GVN beyond this point so 3940 // no need to keep the control input. We want the expensive nodes to 3941 // be freely moved to the least frequent code path by gcm. 3942 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3943 for (int i = 0; i < expensive_count(); i++) { 3944 _expensive_nodes.at(i)->set_req(0, nullptr); 3945 } 3946 3947 Final_Reshape_Counts frc; 3948 3949 // Visit everybody reachable! 3950 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3951 Node_Stack nstack(live_nodes() >> 1); 3952 Unique_Node_List dead_nodes; 3953 final_graph_reshaping_walk(nstack, root(), frc, dead_nodes); 3954 3955 // Check for unreachable (from below) code (i.e., infinite loops). 3956 for( uint i = 0; i < frc._tests.size(); i++ ) { 3957 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3958 // Get number of CFG targets. 3959 // Note that PCTables include exception targets after calls. 3960 uint required_outcnt = n->required_outcnt(); 3961 if (n->outcnt() != required_outcnt) { 3962 // Check for a few special cases. Rethrow Nodes never take the 3963 // 'fall-thru' path, so expected kids is 1 less. 3964 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3965 if (n->in(0)->in(0)->is_Call()) { 3966 CallNode* call = n->in(0)->in(0)->as_Call(); 3967 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3968 required_outcnt--; // Rethrow always has 1 less kid 3969 } else if (call->req() > TypeFunc::Parms && 3970 call->is_CallDynamicJava()) { 3971 // Check for null receiver. In such case, the optimizer has 3972 // detected that the virtual call will always result in a null 3973 // pointer exception. The fall-through projection of this CatchNode 3974 // will not be populated. 3975 Node* arg0 = call->in(TypeFunc::Parms); 3976 if (arg0->is_Type() && 3977 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3978 required_outcnt--; 3979 } 3980 } else if (call->entry_point() == OptoRuntime::new_array_Java() || 3981 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 3982 // Check for illegal array length. In such case, the optimizer has 3983 // detected that the allocation attempt will always result in an 3984 // exception. There is no fall-through projection of this CatchNode . 3985 assert(call->is_CallStaticJava(), "static call expected"); 3986 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 3987 uint valid_length_test_input = call->req() - 1; 3988 Node* valid_length_test = call->in(valid_length_test_input); 3989 call->del_req(valid_length_test_input); 3990 if (valid_length_test->find_int_con(1) == 0) { 3991 required_outcnt--; 3992 } 3993 dead_nodes.push(valid_length_test); 3994 assert(n->outcnt() == required_outcnt, "malformed control flow"); 3995 continue; 3996 } 3997 } 3998 } 3999 4000 // Recheck with a better notion of 'required_outcnt' 4001 if (n->outcnt() != required_outcnt) { 4002 DEBUG_ONLY( n->dump_bfs(1, 0, "-"); ); 4003 assert(false, "malformed control flow"); 4004 record_method_not_compilable("malformed control flow"); 4005 return true; // Not all targets reachable! 4006 } 4007 } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) { 4008 CallNode* call = n->in(0)->in(0)->as_Call(); 4009 if (call->entry_point() == OptoRuntime::new_array_Java() || 4010 call->entry_point() == OptoRuntime::new_array_nozero_Java()) { 4011 assert(call->is_CallStaticJava(), "static call expected"); 4012 assert(call->req() == call->jvms()->endoff() + 1, "missing extra input"); 4013 uint valid_length_test_input = call->req() - 1; 4014 dead_nodes.push(call->in(valid_length_test_input)); 4015 call->del_req(valid_length_test_input); // valid length test useless now 4016 } 4017 } 4018 // Check that I actually visited all kids. Unreached kids 4019 // must be infinite loops. 4020 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4021 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4022 record_method_not_compilable("infinite loop"); 4023 return true; // Found unvisited kid; must be unreach 4024 } 4025 4026 // Here so verification code in final_graph_reshaping_walk() 4027 // always see an OuterStripMinedLoopEnd 4028 if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) { 4029 IfNode* init_iff = n->as_If(); 4030 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4031 n->subsume_by(iff, this); 4032 } 4033 } 4034 4035 while (dead_nodes.size() > 0) { 4036 Node* m = dead_nodes.pop(); 4037 if (m->outcnt() == 0 && m != top()) { 4038 for (uint j = 0; j < m->req(); j++) { 4039 Node* in = m->in(j); 4040 if (in != nullptr) { 4041 dead_nodes.push(in); 4042 } 4043 } 4044 m->disconnect_inputs(this); 4045 } 4046 } 4047 4048 #ifdef IA32 4049 // If original bytecodes contained a mixture of floats and doubles 4050 // check if the optimizer has made it homogeneous, item (3). 4051 if (UseSSE == 0 && 4052 frc.get_float_count() > 32 && 4053 frc.get_double_count() == 0 && 4054 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4055 set_24_bit_selection_and_mode(false, true); 4056 } 4057 #endif // IA32 4058 4059 set_java_calls(frc.get_java_call_count()); 4060 set_inner_loops(frc.get_inner_loop_count()); 4061 4062 // No infinite loops, no reason to bail out. 4063 return false; 4064 } 4065 4066 //-----------------------------too_many_traps---------------------------------- 4067 // Report if there are too many traps at the current method and bci. 4068 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4069 bool Compile::too_many_traps(ciMethod* method, 4070 int bci, 4071 Deoptimization::DeoptReason reason) { 4072 ciMethodData* md = method->method_data(); 4073 if (md->is_empty()) { 4074 // Assume the trap has not occurred, or that it occurred only 4075 // because of a transient condition during start-up in the interpreter. 4076 return false; 4077 } 4078 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4079 if (md->has_trap_at(bci, m, reason) != 0) { 4080 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4081 // Also, if there are multiple reasons, or if there is no per-BCI record, 4082 // assume the worst. 4083 if (log()) 4084 log()->elem("observe trap='%s' count='%d'", 4085 Deoptimization::trap_reason_name(reason), 4086 md->trap_count(reason)); 4087 return true; 4088 } else { 4089 // Ignore method/bci and see if there have been too many globally. 4090 return too_many_traps(reason, md); 4091 } 4092 } 4093 4094 // Less-accurate variant which does not require a method and bci. 4095 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4096 ciMethodData* logmd) { 4097 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4098 // Too many traps globally. 4099 // Note that we use cumulative trap_count, not just md->trap_count. 4100 if (log()) { 4101 int mcount = (logmd == nullptr)? -1: (int)logmd->trap_count(reason); 4102 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4103 Deoptimization::trap_reason_name(reason), 4104 mcount, trap_count(reason)); 4105 } 4106 return true; 4107 } else { 4108 // The coast is clear. 4109 return false; 4110 } 4111 } 4112 4113 //--------------------------too_many_recompiles-------------------------------- 4114 // Report if there are too many recompiles at the current method and bci. 4115 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4116 // Is not eager to return true, since this will cause the compiler to use 4117 // Action_none for a trap point, to avoid too many recompilations. 4118 bool Compile::too_many_recompiles(ciMethod* method, 4119 int bci, 4120 Deoptimization::DeoptReason reason) { 4121 ciMethodData* md = method->method_data(); 4122 if (md->is_empty()) { 4123 // Assume the trap has not occurred, or that it occurred only 4124 // because of a transient condition during start-up in the interpreter. 4125 return false; 4126 } 4127 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4128 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4129 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4130 Deoptimization::DeoptReason per_bc_reason 4131 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4132 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : nullptr; 4133 if ((per_bc_reason == Deoptimization::Reason_none 4134 || md->has_trap_at(bci, m, reason) != 0) 4135 // The trap frequency measure we care about is the recompile count: 4136 && md->trap_recompiled_at(bci, m) 4137 && md->overflow_recompile_count() >= bc_cutoff) { 4138 // Do not emit a trap here if it has already caused recompilations. 4139 // Also, if there are multiple reasons, or if there is no per-BCI record, 4140 // assume the worst. 4141 if (log()) 4142 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4143 Deoptimization::trap_reason_name(reason), 4144 md->trap_count(reason), 4145 md->overflow_recompile_count()); 4146 return true; 4147 } else if (trap_count(reason) != 0 4148 && decompile_count() >= m_cutoff) { 4149 // Too many recompiles globally, and we have seen this sort of trap. 4150 // Use cumulative decompile_count, not just md->decompile_count. 4151 if (log()) 4152 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4153 Deoptimization::trap_reason_name(reason), 4154 md->trap_count(reason), trap_count(reason), 4155 md->decompile_count(), decompile_count()); 4156 return true; 4157 } else { 4158 // The coast is clear. 4159 return false; 4160 } 4161 } 4162 4163 // Compute when not to trap. Used by matching trap based nodes and 4164 // NullCheck optimization. 4165 void Compile::set_allowed_deopt_reasons() { 4166 _allowed_reasons = 0; 4167 if (is_method_compilation()) { 4168 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4169 assert(rs < BitsPerInt, "recode bit map"); 4170 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4171 _allowed_reasons |= nth_bit(rs); 4172 } 4173 } 4174 } 4175 } 4176 4177 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4178 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4179 } 4180 4181 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4182 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4183 } 4184 4185 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4186 if (holder->is_initialized()) { 4187 return false; 4188 } 4189 if (holder->is_being_initialized()) { 4190 if (accessing_method->holder() == holder) { 4191 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4192 // <init>, or a static method. In all those cases, there was an initialization 4193 // barrier on the holder klass passed. 4194 if (accessing_method->is_static_initializer() || 4195 accessing_method->is_object_initializer() || 4196 accessing_method->is_static()) { 4197 return false; 4198 } 4199 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4200 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4201 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4202 // child class can become fully initialized while its parent class is still being initialized. 4203 if (accessing_method->is_static_initializer()) { 4204 return false; 4205 } 4206 } 4207 ciMethod* root = method(); // the root method of compilation 4208 if (root != accessing_method) { 4209 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4210 } 4211 } 4212 return true; 4213 } 4214 4215 #ifndef PRODUCT 4216 //------------------------------verify_bidirectional_edges--------------------- 4217 // For each input edge to a node (ie - for each Use-Def edge), verify that 4218 // there is a corresponding Def-Use edge. 4219 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) { 4220 // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc 4221 uint stack_size = live_nodes() >> 4; 4222 Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize)); 4223 nstack.push(_root); 4224 4225 while (nstack.size() > 0) { 4226 Node* n = nstack.pop(); 4227 if (visited.member(n)) { 4228 continue; 4229 } 4230 visited.push(n); 4231 4232 // Walk over all input edges, checking for correspondence 4233 uint length = n->len(); 4234 for (uint i = 0; i < length; i++) { 4235 Node* in = n->in(i); 4236 if (in != nullptr && !visited.member(in)) { 4237 nstack.push(in); // Put it on stack 4238 } 4239 if (in != nullptr && !in->is_top()) { 4240 // Count instances of `next` 4241 int cnt = 0; 4242 for (uint idx = 0; idx < in->_outcnt; idx++) { 4243 if (in->_out[idx] == n) { 4244 cnt++; 4245 } 4246 } 4247 assert(cnt > 0, "Failed to find Def-Use edge."); 4248 // Check for duplicate edges 4249 // walk the input array downcounting the input edges to n 4250 for (uint j = 0; j < length; j++) { 4251 if (n->in(j) == in) { 4252 cnt--; 4253 } 4254 } 4255 assert(cnt == 0, "Mismatched edge count."); 4256 } else if (in == nullptr) { 4257 assert(i == 0 || i >= n->req() || 4258 n->is_Region() || n->is_Phi() || n->is_ArrayCopy() || 4259 (n->is_Unlock() && i == (n->req() - 1)) || 4260 (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion 4261 "only region, phi, arraycopy, unlock or membar nodes have null data edges"); 4262 } else { 4263 assert(in->is_top(), "sanity"); 4264 // Nothing to check. 4265 } 4266 } 4267 } 4268 } 4269 4270 //------------------------------verify_graph_edges--------------------------- 4271 // Walk the Graph and verify that there is a one-to-one correspondence 4272 // between Use-Def edges and Def-Use edges in the graph. 4273 void Compile::verify_graph_edges(bool no_dead_code) { 4274 if (VerifyGraphEdges) { 4275 Unique_Node_List visited; 4276 4277 // Call graph walk to check edges 4278 verify_bidirectional_edges(visited); 4279 if (no_dead_code) { 4280 // Now make sure that no visited node is used by an unvisited node. 4281 bool dead_nodes = false; 4282 Unique_Node_List checked; 4283 while (visited.size() > 0) { 4284 Node* n = visited.pop(); 4285 checked.push(n); 4286 for (uint i = 0; i < n->outcnt(); i++) { 4287 Node* use = n->raw_out(i); 4288 if (checked.member(use)) continue; // already checked 4289 if (visited.member(use)) continue; // already in the graph 4290 if (use->is_Con()) continue; // a dead ConNode is OK 4291 // At this point, we have found a dead node which is DU-reachable. 4292 if (!dead_nodes) { 4293 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4294 dead_nodes = true; 4295 } 4296 use->dump(2); 4297 tty->print_cr("---"); 4298 checked.push(use); // No repeats; pretend it is now checked. 4299 } 4300 } 4301 assert(!dead_nodes, "using nodes must be reachable from root"); 4302 } 4303 } 4304 } 4305 #endif 4306 4307 // The Compile object keeps track of failure reasons separately from the ciEnv. 4308 // This is required because there is not quite a 1-1 relation between the 4309 // ciEnv and its compilation task and the Compile object. Note that one 4310 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4311 // to backtrack and retry without subsuming loads. Other than this backtracking 4312 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4313 // by the logic in C2Compiler. 4314 void Compile::record_failure(const char* reason) { 4315 if (log() != nullptr) { 4316 log()->elem("failure reason='%s' phase='compile'", reason); 4317 } 4318 if (_failure_reason == nullptr) { 4319 // Record the first failure reason. 4320 _failure_reason = reason; 4321 } 4322 4323 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4324 C->print_method(PHASE_FAILURE, 1); 4325 } 4326 _root = nullptr; // flush the graph, too 4327 } 4328 4329 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4330 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4331 _phase_name(name), _dolog(CITimeVerbose) 4332 { 4333 if (_dolog) { 4334 C = Compile::current(); 4335 _log = C->log(); 4336 } else { 4337 C = nullptr; 4338 _log = nullptr; 4339 } 4340 if (_log != nullptr) { 4341 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4342 _log->stamp(); 4343 _log->end_head(); 4344 } 4345 } 4346 4347 Compile::TracePhase::~TracePhase() { 4348 4349 C = Compile::current(); 4350 if (_dolog) { 4351 _log = C->log(); 4352 } else { 4353 _log = nullptr; 4354 } 4355 4356 #ifdef ASSERT 4357 if (PrintIdealNodeCount) { 4358 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4359 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4360 } 4361 4362 if (VerifyIdealNodeCount) { 4363 Compile::current()->print_missing_nodes(); 4364 } 4365 #endif 4366 4367 if (_log != nullptr) { 4368 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4369 } 4370 } 4371 4372 //----------------------------static_subtype_check----------------------------- 4373 // Shortcut important common cases when superklass is exact: 4374 // (0) superklass is java.lang.Object (can occur in reflective code) 4375 // (1) subklass is already limited to a subtype of superklass => always ok 4376 // (2) subklass does not overlap with superklass => always fail 4377 // (3) superklass has NO subtypes and we can check with a simple compare. 4378 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) { 4379 if (skip) { 4380 return SSC_full_test; // Let caller generate the general case. 4381 } 4382 4383 if (subk->is_java_subtype_of(superk)) { 4384 return SSC_always_true; // (0) and (1) this test cannot fail 4385 } 4386 4387 if (!subk->maybe_java_subtype_of(superk)) { 4388 return SSC_always_false; // (2) true path dead; no dynamic test needed 4389 } 4390 4391 const Type* superelem = superk; 4392 if (superk->isa_aryklassptr()) { 4393 int ignored; 4394 superelem = superk->is_aryklassptr()->base_element_type(ignored); 4395 } 4396 4397 if (superelem->isa_instklassptr()) { 4398 ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass(); 4399 if (!ik->has_subklass()) { 4400 if (!ik->is_final()) { 4401 // Add a dependency if there is a chance of a later subclass. 4402 dependencies()->assert_leaf_type(ik); 4403 } 4404 if (!superk->maybe_java_subtype_of(subk)) { 4405 return SSC_always_false; 4406 } 4407 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4408 } 4409 } else { 4410 // A primitive array type has no subtypes. 4411 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4412 } 4413 4414 return SSC_full_test; 4415 } 4416 4417 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4418 #ifdef _LP64 4419 // The scaled index operand to AddP must be a clean 64-bit value. 4420 // Java allows a 32-bit int to be incremented to a negative 4421 // value, which appears in a 64-bit register as a large 4422 // positive number. Using that large positive number as an 4423 // operand in pointer arithmetic has bad consequences. 4424 // On the other hand, 32-bit overflow is rare, and the possibility 4425 // can often be excluded, if we annotate the ConvI2L node with 4426 // a type assertion that its value is known to be a small positive 4427 // number. (The prior range check has ensured this.) 4428 // This assertion is used by ConvI2LNode::Ideal. 4429 int index_max = max_jint - 1; // array size is max_jint, index is one less 4430 if (sizetype != nullptr) index_max = sizetype->_hi - 1; 4431 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4432 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4433 #endif 4434 return idx; 4435 } 4436 4437 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4438 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4439 if (ctrl != nullptr) { 4440 // Express control dependency by a CastII node with a narrow type. 4441 value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */); 4442 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4443 // node from floating above the range check during loop optimizations. Otherwise, the 4444 // ConvI2L node may be eliminated independently of the range check, causing the data path 4445 // to become TOP while the control path is still there (although it's unreachable). 4446 value->set_req(0, ctrl); 4447 value = phase->transform(value); 4448 } 4449 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4450 return phase->transform(new ConvI2LNode(value, ltype)); 4451 } 4452 4453 // The message about the current inlining is accumulated in 4454 // _print_inlining_stream and transferred into the _print_inlining_list 4455 // once we know whether inlining succeeds or not. For regular 4456 // inlining, messages are appended to the buffer pointed by 4457 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4458 // a new buffer is added after _print_inlining_idx in the list. This 4459 // way we can update the inlining message for late inlining call site 4460 // when the inlining is attempted again. 4461 void Compile::print_inlining_init() { 4462 if (print_inlining() || print_intrinsics()) { 4463 // print_inlining_init is actually called several times. 4464 print_inlining_reset(); 4465 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4466 } 4467 } 4468 4469 void Compile::print_inlining_reinit() { 4470 if (print_inlining() || print_intrinsics()) { 4471 print_inlining_reset(); 4472 } 4473 } 4474 4475 void Compile::print_inlining_reset() { 4476 _print_inlining_stream->reset(); 4477 } 4478 4479 void Compile::print_inlining_commit() { 4480 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4481 // Transfer the message from _print_inlining_stream to the current 4482 // _print_inlining_list buffer and clear _print_inlining_stream. 4483 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4484 print_inlining_reset(); 4485 } 4486 4487 void Compile::print_inlining_push() { 4488 // Add new buffer to the _print_inlining_list at current position 4489 _print_inlining_idx++; 4490 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4491 } 4492 4493 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4494 return _print_inlining_list->at(_print_inlining_idx); 4495 } 4496 4497 void Compile::print_inlining_update(CallGenerator* cg) { 4498 if (print_inlining() || print_intrinsics()) { 4499 if (cg->is_late_inline()) { 4500 if (print_inlining_current()->cg() != cg && 4501 (print_inlining_current()->cg() != nullptr || 4502 print_inlining_current()->ss()->size() != 0)) { 4503 print_inlining_push(); 4504 } 4505 print_inlining_commit(); 4506 print_inlining_current()->set_cg(cg); 4507 } else { 4508 if (print_inlining_current()->cg() != nullptr) { 4509 print_inlining_push(); 4510 } 4511 print_inlining_commit(); 4512 } 4513 } 4514 } 4515 4516 void Compile::print_inlining_move_to(CallGenerator* cg) { 4517 // We resume inlining at a late inlining call site. Locate the 4518 // corresponding inlining buffer so that we can update it. 4519 if (print_inlining() || print_intrinsics()) { 4520 for (int i = 0; i < _print_inlining_list->length(); i++) { 4521 if (_print_inlining_list->at(i)->cg() == cg) { 4522 _print_inlining_idx = i; 4523 return; 4524 } 4525 } 4526 ShouldNotReachHere(); 4527 } 4528 } 4529 4530 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4531 if (print_inlining() || print_intrinsics()) { 4532 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4533 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4534 // replace message with new message 4535 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4536 print_inlining_commit(); 4537 print_inlining_current()->set_cg(cg); 4538 } 4539 } 4540 4541 void Compile::print_inlining_assert_ready() { 4542 assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data"); 4543 } 4544 4545 void Compile::process_print_inlining() { 4546 assert(_late_inlines.length() == 0, "not drained yet"); 4547 if (print_inlining() || print_intrinsics()) { 4548 ResourceMark rm; 4549 stringStream ss; 4550 assert(_print_inlining_list != nullptr, "process_print_inlining should be called only once."); 4551 for (int i = 0; i < _print_inlining_list->length(); i++) { 4552 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4553 ss.print("%s", pib->ss()->freeze()); 4554 delete pib; 4555 DEBUG_ONLY(_print_inlining_list->at_put(i, nullptr)); 4556 } 4557 // Reset _print_inlining_list, it only contains destructed objects. 4558 // It is on the arena, so it will be freed when the arena is reset. 4559 _print_inlining_list = nullptr; 4560 // _print_inlining_stream won't be used anymore, either. 4561 print_inlining_reset(); 4562 size_t end = ss.size(); 4563 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4564 strncpy(_print_inlining_output, ss.freeze(), end+1); 4565 _print_inlining_output[end] = 0; 4566 } 4567 } 4568 4569 void Compile::dump_print_inlining() { 4570 if (_print_inlining_output != nullptr) { 4571 tty->print_raw(_print_inlining_output); 4572 } 4573 } 4574 4575 void Compile::log_late_inline(CallGenerator* cg) { 4576 if (log() != nullptr) { 4577 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4578 cg->unique_id()); 4579 JVMState* p = cg->call_node()->jvms(); 4580 while (p != nullptr) { 4581 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4582 p = p->caller(); 4583 } 4584 log()->tail("late_inline"); 4585 } 4586 } 4587 4588 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4589 log_late_inline(cg); 4590 if (log() != nullptr) { 4591 log()->inline_fail(msg); 4592 } 4593 } 4594 4595 void Compile::log_inline_id(CallGenerator* cg) { 4596 if (log() != nullptr) { 4597 // The LogCompilation tool needs a unique way to identify late 4598 // inline call sites. This id must be unique for this call site in 4599 // this compilation. Try to have it unique across compilations as 4600 // well because it can be convenient when grepping through the log 4601 // file. 4602 // Distinguish OSR compilations from others in case CICountOSR is 4603 // on. 4604 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4605 cg->set_unique_id(id); 4606 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4607 } 4608 } 4609 4610 void Compile::log_inline_failure(const char* msg) { 4611 if (C->log() != nullptr) { 4612 C->log()->inline_fail(msg); 4613 } 4614 } 4615 4616 4617 // Dump inlining replay data to the stream. 4618 // Don't change thread state and acquire any locks. 4619 void Compile::dump_inline_data(outputStream* out) { 4620 InlineTree* inl_tree = ilt(); 4621 if (inl_tree != nullptr) { 4622 out->print(" inline %d", inl_tree->count()); 4623 inl_tree->dump_replay_data(out); 4624 } 4625 } 4626 4627 void Compile::dump_inline_data_reduced(outputStream* out) { 4628 assert(ReplayReduce, ""); 4629 4630 InlineTree* inl_tree = ilt(); 4631 if (inl_tree == nullptr) { 4632 return; 4633 } 4634 // Enable iterative replay file reduction 4635 // Output "compile" lines for depth 1 subtrees, 4636 // simulating that those trees were compiled 4637 // instead of inlined. 4638 for (int i = 0; i < inl_tree->subtrees().length(); ++i) { 4639 InlineTree* sub = inl_tree->subtrees().at(i); 4640 if (sub->inline_level() != 1) { 4641 continue; 4642 } 4643 4644 ciMethod* method = sub->method(); 4645 int entry_bci = -1; 4646 int comp_level = env()->task()->comp_level(); 4647 out->print("compile "); 4648 method->dump_name_as_ascii(out); 4649 out->print(" %d %d", entry_bci, comp_level); 4650 out->print(" inline %d", sub->count()); 4651 sub->dump_replay_data(out, -1); 4652 out->cr(); 4653 } 4654 } 4655 4656 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4657 if (n1->Opcode() < n2->Opcode()) return -1; 4658 else if (n1->Opcode() > n2->Opcode()) return 1; 4659 4660 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4661 for (uint i = 1; i < n1->req(); i++) { 4662 if (n1->in(i) < n2->in(i)) return -1; 4663 else if (n1->in(i) > n2->in(i)) return 1; 4664 } 4665 4666 return 0; 4667 } 4668 4669 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4670 Node* n1 = *n1p; 4671 Node* n2 = *n2p; 4672 4673 return cmp_expensive_nodes(n1, n2); 4674 } 4675 4676 void Compile::sort_expensive_nodes() { 4677 if (!expensive_nodes_sorted()) { 4678 _expensive_nodes.sort(cmp_expensive_nodes); 4679 } 4680 } 4681 4682 bool Compile::expensive_nodes_sorted() const { 4683 for (int i = 1; i < _expensive_nodes.length(); i++) { 4684 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) { 4685 return false; 4686 } 4687 } 4688 return true; 4689 } 4690 4691 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4692 if (_expensive_nodes.length() == 0) { 4693 return false; 4694 } 4695 4696 assert(OptimizeExpensiveOps, "optimization off?"); 4697 4698 // Take this opportunity to remove dead nodes from the list 4699 int j = 0; 4700 for (int i = 0; i < _expensive_nodes.length(); i++) { 4701 Node* n = _expensive_nodes.at(i); 4702 if (!n->is_unreachable(igvn)) { 4703 assert(n->is_expensive(), "should be expensive"); 4704 _expensive_nodes.at_put(j, n); 4705 j++; 4706 } 4707 } 4708 _expensive_nodes.trunc_to(j); 4709 4710 // Then sort the list so that similar nodes are next to each other 4711 // and check for at least two nodes of identical kind with same data 4712 // inputs. 4713 sort_expensive_nodes(); 4714 4715 for (int i = 0; i < _expensive_nodes.length()-1; i++) { 4716 if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) { 4717 return true; 4718 } 4719 } 4720 4721 return false; 4722 } 4723 4724 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4725 if (_expensive_nodes.length() == 0) { 4726 return; 4727 } 4728 4729 assert(OptimizeExpensiveOps, "optimization off?"); 4730 4731 // Sort to bring similar nodes next to each other and clear the 4732 // control input of nodes for which there's only a single copy. 4733 sort_expensive_nodes(); 4734 4735 int j = 0; 4736 int identical = 0; 4737 int i = 0; 4738 bool modified = false; 4739 for (; i < _expensive_nodes.length()-1; i++) { 4740 assert(j <= i, "can't write beyond current index"); 4741 if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) { 4742 identical++; 4743 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4744 continue; 4745 } 4746 if (identical > 0) { 4747 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4748 identical = 0; 4749 } else { 4750 Node* n = _expensive_nodes.at(i); 4751 igvn.replace_input_of(n, 0, nullptr); 4752 igvn.hash_insert(n); 4753 modified = true; 4754 } 4755 } 4756 if (identical > 0) { 4757 _expensive_nodes.at_put(j++, _expensive_nodes.at(i)); 4758 } else if (_expensive_nodes.length() >= 1) { 4759 Node* n = _expensive_nodes.at(i); 4760 igvn.replace_input_of(n, 0, nullptr); 4761 igvn.hash_insert(n); 4762 modified = true; 4763 } 4764 _expensive_nodes.trunc_to(j); 4765 if (modified) { 4766 igvn.optimize(); 4767 } 4768 } 4769 4770 void Compile::add_expensive_node(Node * n) { 4771 assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list"); 4772 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4773 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4774 if (OptimizeExpensiveOps) { 4775 _expensive_nodes.append(n); 4776 } else { 4777 // Clear control input and let IGVN optimize expensive nodes if 4778 // OptimizeExpensiveOps is off. 4779 n->set_req(0, nullptr); 4780 } 4781 } 4782 4783 /** 4784 * Track coarsened Lock and Unlock nodes. 4785 */ 4786 4787 class Lock_List : public Node_List { 4788 uint _origin_cnt; 4789 public: 4790 Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {} 4791 uint origin_cnt() const { return _origin_cnt; } 4792 }; 4793 4794 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) { 4795 int length = locks.length(); 4796 if (length > 0) { 4797 // Have to keep this list until locks elimination during Macro nodes elimination. 4798 Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length); 4799 for (int i = 0; i < length; i++) { 4800 AbstractLockNode* lock = locks.at(i); 4801 assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx); 4802 locks_list->push(lock); 4803 } 4804 _coarsened_locks.append(locks_list); 4805 } 4806 } 4807 4808 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) { 4809 int count = coarsened_count(); 4810 for (int i = 0; i < count; i++) { 4811 Node_List* locks_list = _coarsened_locks.at(i); 4812 for (uint j = 0; j < locks_list->size(); j++) { 4813 Node* lock = locks_list->at(j); 4814 assert(lock->is_AbstractLock(), "sanity"); 4815 if (!useful.member(lock)) { 4816 locks_list->yank(lock); 4817 } 4818 } 4819 } 4820 } 4821 4822 void Compile::remove_coarsened_lock(Node* n) { 4823 if (n->is_AbstractLock()) { 4824 int count = coarsened_count(); 4825 for (int i = 0; i < count; i++) { 4826 Node_List* locks_list = _coarsened_locks.at(i); 4827 locks_list->yank(n); 4828 } 4829 } 4830 } 4831 4832 bool Compile::coarsened_locks_consistent() { 4833 int count = coarsened_count(); 4834 for (int i = 0; i < count; i++) { 4835 bool unbalanced = false; 4836 bool modified = false; // track locks kind modifications 4837 Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i); 4838 uint size = locks_list->size(); 4839 if (size == 0) { 4840 unbalanced = false; // All locks were eliminated - good 4841 } else if (size != locks_list->origin_cnt()) { 4842 unbalanced = true; // Some locks were removed from list 4843 } else { 4844 for (uint j = 0; j < size; j++) { 4845 Node* lock = locks_list->at(j); 4846 // All nodes in group should have the same state (modified or not) 4847 if (!lock->as_AbstractLock()->is_coarsened()) { 4848 if (j == 0) { 4849 // first on list was modified, the rest should be too for consistency 4850 modified = true; 4851 } else if (!modified) { 4852 // this lock was modified but previous locks on the list were not 4853 unbalanced = true; 4854 break; 4855 } 4856 } else if (modified) { 4857 // previous locks on list were modified but not this lock 4858 unbalanced = true; 4859 break; 4860 } 4861 } 4862 } 4863 if (unbalanced) { 4864 // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified 4865 #ifdef ASSERT 4866 if (PrintEliminateLocks) { 4867 tty->print_cr("=== unbalanced coarsened locks ==="); 4868 for (uint l = 0; l < size; l++) { 4869 locks_list->at(l)->dump(); 4870 } 4871 } 4872 #endif 4873 record_failure(C2Compiler::retry_no_locks_coarsening()); 4874 return false; 4875 } 4876 } 4877 return true; 4878 } 4879 4880 /** 4881 * Remove the speculative part of types and clean up the graph 4882 */ 4883 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4884 if (UseTypeSpeculation) { 4885 Unique_Node_List worklist; 4886 worklist.push(root()); 4887 int modified = 0; 4888 // Go over all type nodes that carry a speculative type, drop the 4889 // speculative part of the type and enqueue the node for an igvn 4890 // which may optimize it out. 4891 for (uint next = 0; next < worklist.size(); ++next) { 4892 Node *n = worklist.at(next); 4893 if (n->is_Type()) { 4894 TypeNode* tn = n->as_Type(); 4895 const Type* t = tn->type(); 4896 const Type* t_no_spec = t->remove_speculative(); 4897 if (t_no_spec != t) { 4898 bool in_hash = igvn.hash_delete(n); 4899 assert(in_hash, "node should be in igvn hash table"); 4900 tn->set_type(t_no_spec); 4901 igvn.hash_insert(n); 4902 igvn._worklist.push(n); // give it a chance to go away 4903 modified++; 4904 } 4905 } 4906 // Iterate over outs - endless loops is unreachable from below 4907 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4908 Node *m = n->fast_out(i); 4909 if (not_a_node(m)) { 4910 continue; 4911 } 4912 worklist.push(m); 4913 } 4914 } 4915 // Drop the speculative part of all types in the igvn's type table 4916 igvn.remove_speculative_types(); 4917 if (modified > 0) { 4918 igvn.optimize(); 4919 } 4920 #ifdef ASSERT 4921 // Verify that after the IGVN is over no speculative type has resurfaced 4922 worklist.clear(); 4923 worklist.push(root()); 4924 for (uint next = 0; next < worklist.size(); ++next) { 4925 Node *n = worklist.at(next); 4926 const Type* t = igvn.type_or_null(n); 4927 assert((t == nullptr) || (t == t->remove_speculative()), "no more speculative types"); 4928 if (n->is_Type()) { 4929 t = n->as_Type()->type(); 4930 assert(t == t->remove_speculative(), "no more speculative types"); 4931 } 4932 // Iterate over outs - endless loops is unreachable from below 4933 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 4934 Node *m = n->fast_out(i); 4935 if (not_a_node(m)) { 4936 continue; 4937 } 4938 worklist.push(m); 4939 } 4940 } 4941 igvn.check_no_speculative_types(); 4942 #endif 4943 } 4944 } 4945 4946 // Auxiliary methods to support randomized stressing/fuzzing. 4947 4948 int Compile::random() { 4949 _stress_seed = os::next_random(_stress_seed); 4950 return static_cast<int>(_stress_seed); 4951 } 4952 4953 // This method can be called the arbitrary number of times, with current count 4954 // as the argument. The logic allows selecting a single candidate from the 4955 // running list of candidates as follows: 4956 // int count = 0; 4957 // Cand* selected = null; 4958 // while(cand = cand->next()) { 4959 // if (randomized_select(++count)) { 4960 // selected = cand; 4961 // } 4962 // } 4963 // 4964 // Including count equalizes the chances any candidate is "selected". 4965 // This is useful when we don't have the complete list of candidates to choose 4966 // from uniformly. In this case, we need to adjust the randomicity of the 4967 // selection, or else we will end up biasing the selection towards the latter 4968 // candidates. 4969 // 4970 // Quick back-envelope calculation shows that for the list of n candidates 4971 // the equal probability for the candidate to persist as "best" can be 4972 // achieved by replacing it with "next" k-th candidate with the probability 4973 // of 1/k. It can be easily shown that by the end of the run, the 4974 // probability for any candidate is converged to 1/n, thus giving the 4975 // uniform distribution among all the candidates. 4976 // 4977 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4978 #define RANDOMIZED_DOMAIN_POW 29 4979 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4980 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4981 bool Compile::randomized_select(int count) { 4982 assert(count > 0, "only positive"); 4983 return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4984 } 4985 4986 CloneMap& Compile::clone_map() { return _clone_map; } 4987 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4988 4989 void NodeCloneInfo::dump() const { 4990 tty->print(" {%d:%d} ", idx(), gen()); 4991 } 4992 4993 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4994 uint64_t val = value(old->_idx); 4995 NodeCloneInfo cio(val); 4996 assert(val != 0, "old node should be in the map"); 4997 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4998 insert(nnn->_idx, cin.get()); 4999 #ifndef PRODUCT 5000 if (is_debug()) { 5001 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5002 } 5003 #endif 5004 } 5005 5006 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5007 NodeCloneInfo cio(value(old->_idx)); 5008 if (cio.get() == 0) { 5009 cio.set(old->_idx, 0); 5010 insert(old->_idx, cio.get()); 5011 #ifndef PRODUCT 5012 if (is_debug()) { 5013 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5014 } 5015 #endif 5016 } 5017 clone(old, nnn, gen); 5018 } 5019 5020 int CloneMap::max_gen() const { 5021 int g = 0; 5022 DictI di(_dict); 5023 for(; di.test(); ++di) { 5024 int t = gen(di._key); 5025 if (g < t) { 5026 g = t; 5027 #ifndef PRODUCT 5028 if (is_debug()) { 5029 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5030 } 5031 #endif 5032 } 5033 } 5034 return g; 5035 } 5036 5037 void CloneMap::dump(node_idx_t key) const { 5038 uint64_t val = value(key); 5039 if (val != 0) { 5040 NodeCloneInfo ni(val); 5041 ni.dump(); 5042 } 5043 } 5044 5045 // Move Allocate nodes to the start of the list 5046 void Compile::sort_macro_nodes() { 5047 int count = macro_count(); 5048 int allocates = 0; 5049 for (int i = 0; i < count; i++) { 5050 Node* n = macro_node(i); 5051 if (n->is_Allocate()) { 5052 if (i != allocates) { 5053 Node* tmp = macro_node(allocates); 5054 _macro_nodes.at_put(allocates, n); 5055 _macro_nodes.at_put(i, tmp); 5056 } 5057 allocates++; 5058 } 5059 } 5060 } 5061 5062 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) { 5063 EventCompilerPhase event; 5064 if (event.should_commit()) { 5065 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level); 5066 } 5067 #ifndef PRODUCT 5068 ResourceMark rm; 5069 stringStream ss; 5070 ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt)); 5071 if (n != nullptr) { 5072 ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]); 5073 } 5074 5075 const char* name = ss.as_string(); 5076 if (should_print_igv(level)) { 5077 _igv_printer->print_method(name, level); 5078 } 5079 if (should_print_phase(cpt)) { 5080 print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt)); 5081 } 5082 #endif 5083 C->_latest_stage_start_counter.stamp(); 5084 } 5085 5086 // Only used from CompileWrapper 5087 void Compile::begin_method() { 5088 #ifndef PRODUCT 5089 if (_method != nullptr && should_print_igv(1)) { 5090 _igv_printer->begin_method(); 5091 } 5092 #endif 5093 C->_latest_stage_start_counter.stamp(); 5094 } 5095 5096 // Only used from CompileWrapper 5097 void Compile::end_method() { 5098 EventCompilerPhase event; 5099 if (event.should_commit()) { 5100 CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1); 5101 } 5102 5103 #ifndef PRODUCT 5104 if (_method != nullptr && should_print_igv(1)) { 5105 _igv_printer->end_method(); 5106 } 5107 #endif 5108 } 5109 5110 bool Compile::should_print_phase(CompilerPhaseType cpt) { 5111 #ifndef PRODUCT 5112 if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) { 5113 return true; 5114 } 5115 #endif 5116 return false; 5117 } 5118 5119 bool Compile::should_print_igv(int level) { 5120 #ifndef PRODUCT 5121 if (PrintIdealGraphLevel < 0) { // disabled by the user 5122 return false; 5123 } 5124 5125 bool need = directive()->IGVPrintLevelOption >= level; 5126 if (need && !_igv_printer) { 5127 _igv_printer = IdealGraphPrinter::printer(); 5128 _igv_printer->set_compile(this); 5129 } 5130 return need; 5131 #else 5132 return false; 5133 #endif 5134 } 5135 5136 #ifndef PRODUCT 5137 IdealGraphPrinter* Compile::_debug_file_printer = nullptr; 5138 IdealGraphPrinter* Compile::_debug_network_printer = nullptr; 5139 5140 // Called from debugger. Prints method to the default file with the default phase name. 5141 // This works regardless of any Ideal Graph Visualizer flags set or not. 5142 void igv_print() { 5143 Compile::current()->igv_print_method_to_file(); 5144 } 5145 5146 // Same as igv_print() above but with a specified phase name. 5147 void igv_print(const char* phase_name) { 5148 Compile::current()->igv_print_method_to_file(phase_name); 5149 } 5150 5151 // Called from debugger. Prints method with the default phase name to the default network or the one specified with 5152 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument. 5153 // This works regardless of any Ideal Graph Visualizer flags set or not. 5154 void igv_print(bool network) { 5155 if (network) { 5156 Compile::current()->igv_print_method_to_network(); 5157 } else { 5158 Compile::current()->igv_print_method_to_file(); 5159 } 5160 } 5161 5162 // Same as igv_print(bool network) above but with a specified phase name. 5163 void igv_print(bool network, const char* phase_name) { 5164 if (network) { 5165 Compile::current()->igv_print_method_to_network(phase_name); 5166 } else { 5167 Compile::current()->igv_print_method_to_file(phase_name); 5168 } 5169 } 5170 5171 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set. 5172 void igv_print_default() { 5173 Compile::current()->print_method(PHASE_DEBUG, 0); 5174 } 5175 5176 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay. 5177 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow 5178 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not. 5179 void igv_append() { 5180 Compile::current()->igv_print_method_to_file("Debug", true); 5181 } 5182 5183 // Same as igv_append() above but with a specified phase name. 5184 void igv_append(const char* phase_name) { 5185 Compile::current()->igv_print_method_to_file(phase_name, true); 5186 } 5187 5188 void Compile::igv_print_method_to_file(const char* phase_name, bool append) { 5189 const char* file_name = "custom_debug.xml"; 5190 if (_debug_file_printer == nullptr) { 5191 _debug_file_printer = new IdealGraphPrinter(C, file_name, append); 5192 } else { 5193 _debug_file_printer->update_compiled_method(C->method()); 5194 } 5195 tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name); 5196 _debug_file_printer->print(phase_name, (Node*)C->root()); 5197 } 5198 5199 void Compile::igv_print_method_to_network(const char* phase_name) { 5200 if (_debug_network_printer == nullptr) { 5201 _debug_network_printer = new IdealGraphPrinter(C); 5202 } else { 5203 _debug_network_printer->update_compiled_method(C->method()); 5204 } 5205 tty->print_cr("Method printed over network stream to IGV"); 5206 _debug_network_printer->print(phase_name, (Node*)C->root()); 5207 } 5208 #endif 5209 5210 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 5211 if (type != nullptr && phase->type(value)->higher_equal(type)) { 5212 return value; 5213 } 5214 Node* result = nullptr; 5215 if (bt == T_BYTE) { 5216 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 5217 result = new RShiftINode(result, phase->intcon(24)); 5218 } else if (bt == T_BOOLEAN) { 5219 result = new AndINode(value, phase->intcon(0xFF)); 5220 } else if (bt == T_CHAR) { 5221 result = new AndINode(value,phase->intcon(0xFFFF)); 5222 } else { 5223 assert(bt == T_SHORT, "unexpected narrow type"); 5224 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 5225 result = new RShiftINode(result, phase->intcon(16)); 5226 } 5227 if (transform_res) { 5228 result = phase->transform(result); 5229 } 5230 return result; 5231 } 5232