1 /* 2 * Copyright (c) 1997, 2018, 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/systemDictionary.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 "memory/resourceArea.hpp" 39 #include "opto/addnode.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/callGenerator.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/castnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/chaitin.hpp" 47 #include "opto/compile.hpp" 48 #include "opto/connode.hpp" 49 #include "opto/convertnode.hpp" 50 #include "opto/divnode.hpp" 51 #include "opto/escape.hpp" 52 #include "opto/idealGraphPrinter.hpp" 53 #include "opto/loopnode.hpp" 54 #include "opto/machnode.hpp" 55 #include "opto/macro.hpp" 56 #include "opto/matcher.hpp" 57 #include "opto/mathexactnode.hpp" 58 #include "opto/memnode.hpp" 59 #include "opto/mulnode.hpp" 60 #include "opto/narrowptrnode.hpp" 61 #include "opto/node.hpp" 62 #include "opto/opcodes.hpp" 63 #include "opto/output.hpp" 64 #include "opto/parse.hpp" 65 #include "opto/phaseX.hpp" 66 #include "opto/rootnode.hpp" 67 #include "opto/runtime.hpp" 68 #include "opto/stringopts.hpp" 69 #include "opto/type.hpp" 70 #include "opto/vectornode.hpp" 71 #include "runtime/arguments.hpp" 72 #include "runtime/sharedRuntime.hpp" 73 #include "runtime/signature.hpp" 74 #include "runtime/stubRoutines.hpp" 75 #include "runtime/timer.hpp" 76 #include "utilities/align.hpp" 77 #include "utilities/copy.hpp" 78 #include "utilities/macros.hpp" 79 #if INCLUDE_ZGC 80 #include "gc/z/c2/zBarrierSetC2.hpp" 81 #endif 82 83 84 // -------------------- Compile::mach_constant_base_node ----------------------- 85 // Constant table base node singleton. 86 MachConstantBaseNode* Compile::mach_constant_base_node() { 87 if (_mach_constant_base_node == NULL) { 88 _mach_constant_base_node = new MachConstantBaseNode(); 89 _mach_constant_base_node->add_req(C->root()); 90 } 91 return _mach_constant_base_node; 92 } 93 94 95 /// Support for intrinsics. 96 97 // Return the index at which m must be inserted (or already exists). 98 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 99 class IntrinsicDescPair { 100 private: 101 ciMethod* _m; 102 bool _is_virtual; 103 public: 104 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 105 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 106 ciMethod* m= elt->method(); 107 ciMethod* key_m = key->_m; 108 if (key_m < m) return -1; 109 else if (key_m > m) return 1; 110 else { 111 bool is_virtual = elt->is_virtual(); 112 bool key_virtual = key->_is_virtual; 113 if (key_virtual < is_virtual) return -1; 114 else if (key_virtual > is_virtual) return 1; 115 else return 0; 116 } 117 } 118 }; 119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 120 #ifdef ASSERT 121 for (int i = 1; i < _intrinsics->length(); i++) { 122 CallGenerator* cg1 = _intrinsics->at(i-1); 123 CallGenerator* cg2 = _intrinsics->at(i); 124 assert(cg1->method() != cg2->method() 125 ? cg1->method() < cg2->method() 126 : cg1->is_virtual() < cg2->is_virtual(), 127 "compiler intrinsics list must stay sorted"); 128 } 129 #endif 130 IntrinsicDescPair pair(m, is_virtual); 131 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 132 } 133 134 void Compile::register_intrinsic(CallGenerator* cg) { 135 if (_intrinsics == NULL) { 136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 137 } 138 int len = _intrinsics->length(); 139 bool found = false; 140 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 141 assert(!found, "registering twice"); 142 _intrinsics->insert_before(index, cg); 143 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 144 } 145 146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 147 assert(m->is_loaded(), "don't try this on unloaded methods"); 148 if (_intrinsics != NULL) { 149 bool found = false; 150 int index = intrinsic_insertion_index(m, is_virtual, found); 151 if (found) { 152 return _intrinsics->at(index); 153 } 154 } 155 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 156 if (m->intrinsic_id() != vmIntrinsics::_none && 157 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 158 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 159 if (cg != NULL) { 160 // Save it for next time: 161 register_intrinsic(cg); 162 return cg; 163 } else { 164 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 165 } 166 } 167 return NULL; 168 } 169 170 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 171 // in library_call.cpp. 172 173 174 #ifndef PRODUCT 175 // statistics gathering... 176 177 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 179 180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 182 int oflags = _intrinsic_hist_flags[id]; 183 assert(flags != 0, "what happened?"); 184 if (is_virtual) { 185 flags |= _intrinsic_virtual; 186 } 187 bool changed = (flags != oflags); 188 if ((flags & _intrinsic_worked) != 0) { 189 juint count = (_intrinsic_hist_count[id] += 1); 190 if (count == 1) { 191 changed = true; // first time 192 } 193 // increment the overall count also: 194 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 195 } 196 if (changed) { 197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 198 // Something changed about the intrinsic's virtuality. 199 if ((flags & _intrinsic_virtual) != 0) { 200 // This is the first use of this intrinsic as a virtual call. 201 if (oflags != 0) { 202 // We already saw it as a non-virtual, so note both cases. 203 flags |= _intrinsic_both; 204 } 205 } else if ((oflags & _intrinsic_both) == 0) { 206 // This is the first use of this intrinsic as a non-virtual 207 flags |= _intrinsic_both; 208 } 209 } 210 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 211 } 212 // update the overall flags also: 213 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 214 return changed; 215 } 216 217 static char* format_flags(int flags, char* buf) { 218 buf[0] = 0; 219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 224 if (buf[0] == 0) strcat(buf, ","); 225 assert(buf[0] == ',', "must be"); 226 return &buf[1]; 227 } 228 229 void Compile::print_intrinsic_statistics() { 230 char flagsbuf[100]; 231 ttyLocker ttyl; 232 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 233 tty->print_cr("Compiler intrinsic usage:"); 234 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 235 if (total == 0) total = 1; // avoid div0 in case of no successes 236 #define PRINT_STAT_LINE(name, c, f) \ 237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 238 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 239 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 240 int flags = _intrinsic_hist_flags[id]; 241 juint count = _intrinsic_hist_count[id]; 242 if ((flags | count) != 0) { 243 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 244 } 245 } 246 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 247 if (xtty != NULL) xtty->tail("statistics"); 248 } 249 250 void Compile::print_statistics() { 251 { ttyLocker ttyl; 252 if (xtty != NULL) xtty->head("statistics type='opto'"); 253 Parse::print_statistics(); 254 PhaseCCP::print_statistics(); 255 PhaseRegAlloc::print_statistics(); 256 Scheduling::print_statistics(); 257 PhasePeephole::print_statistics(); 258 PhaseIdealLoop::print_statistics(); 259 if (xtty != NULL) xtty->tail("statistics"); 260 } 261 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 262 // put this under its own <statistics> element. 263 print_intrinsic_statistics(); 264 } 265 } 266 #endif //PRODUCT 267 268 // Support for bundling info 269 Bundle* Compile::node_bundling(const Node *n) { 270 assert(valid_bundle_info(n), "oob"); 271 return &_node_bundling_base[n->_idx]; 272 } 273 274 bool Compile::valid_bundle_info(const Node *n) { 275 return (_node_bundling_limit > n->_idx); 276 } 277 278 279 void Compile::gvn_replace_by(Node* n, Node* nn) { 280 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 281 Node* use = n->last_out(i); 282 bool is_in_table = initial_gvn()->hash_delete(use); 283 uint uses_found = 0; 284 for (uint j = 0; j < use->len(); j++) { 285 if (use->in(j) == n) { 286 if (j < use->req()) 287 use->set_req(j, nn); 288 else 289 use->set_prec(j, nn); 290 uses_found++; 291 } 292 } 293 if (is_in_table) { 294 // reinsert into table 295 initial_gvn()->hash_find_insert(use); 296 } 297 record_for_igvn(use); 298 i -= uses_found; // we deleted 1 or more copies of this edge 299 } 300 } 301 302 303 static inline bool not_a_node(const Node* n) { 304 if (n == NULL) return true; 305 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 306 if (*(address*)n == badAddress) return true; // kill by Node::destruct 307 return false; 308 } 309 310 // Identify all nodes that are reachable from below, useful. 311 // Use breadth-first pass that records state in a Unique_Node_List, 312 // recursive traversal is slower. 313 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 314 int estimated_worklist_size = live_nodes(); 315 useful.map( estimated_worklist_size, NULL ); // preallocate space 316 317 // Initialize worklist 318 if (root() != NULL) { useful.push(root()); } 319 // If 'top' is cached, declare it useful to preserve cached node 320 if( cached_top_node() ) { useful.push(cached_top_node()); } 321 322 // Push all useful nodes onto the list, breadthfirst 323 for( uint next = 0; next < useful.size(); ++next ) { 324 assert( next < unique(), "Unique useful nodes < total nodes"); 325 Node *n = useful.at(next); 326 uint max = n->len(); 327 for( uint i = 0; i < max; ++i ) { 328 Node *m = n->in(i); 329 if (not_a_node(m)) continue; 330 useful.push(m); 331 } 332 } 333 } 334 335 // Update dead_node_list with any missing dead nodes using useful 336 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 337 void Compile::update_dead_node_list(Unique_Node_List &useful) { 338 uint max_idx = unique(); 339 VectorSet& useful_node_set = useful.member_set(); 340 341 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 342 // If node with index node_idx is not in useful set, 343 // mark it as dead in dead node list. 344 if (! useful_node_set.test(node_idx) ) { 345 record_dead_node(node_idx); 346 } 347 } 348 } 349 350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 351 int shift = 0; 352 for (int i = 0; i < inlines->length(); i++) { 353 CallGenerator* cg = inlines->at(i); 354 CallNode* call = cg->call_node(); 355 if (shift > 0) { 356 inlines->at_put(i-shift, cg); 357 } 358 if (!useful.member(call)) { 359 shift++; 360 } 361 } 362 inlines->trunc_to(inlines->length()-shift); 363 } 364 365 // Disconnect all useless nodes by disconnecting those at the boundary. 366 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 367 uint next = 0; 368 while (next < useful.size()) { 369 Node *n = useful.at(next++); 370 if (n->is_SafePoint()) { 371 // We're done with a parsing phase. Replaced nodes are not valid 372 // beyond that point. 373 n->as_SafePoint()->delete_replaced_nodes(); 374 } 375 // Use raw traversal of out edges since this code removes out edges 376 int max = n->outcnt(); 377 for (int j = 0; j < max; ++j) { 378 Node* child = n->raw_out(j); 379 if (! useful.member(child)) { 380 assert(!child->is_top() || child != top(), 381 "If top is cached in Compile object it is in useful list"); 382 // Only need to remove this out-edge to the useless node 383 n->raw_del_out(j); 384 --j; 385 --max; 386 } 387 } 388 if (n->outcnt() == 1 && n->has_special_unique_user()) { 389 record_for_igvn(n->unique_out()); 390 } 391 } 392 // Remove useless macro and predicate opaq nodes 393 for (int i = C->macro_count()-1; i >= 0; i--) { 394 Node* n = C->macro_node(i); 395 if (!useful.member(n)) { 396 remove_macro_node(n); 397 } 398 } 399 // Remove useless CastII nodes with range check dependency 400 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 401 Node* cast = range_check_cast_node(i); 402 if (!useful.member(cast)) { 403 remove_range_check_cast(cast); 404 } 405 } 406 // Remove useless expensive nodes 407 for (int i = C->expensive_count()-1; i >= 0; i--) { 408 Node* n = C->expensive_node(i); 409 if (!useful.member(n)) { 410 remove_expensive_node(n); 411 } 412 } 413 // Remove useless Opaque4 nodes 414 for (int i = opaque4_count() - 1; i >= 0; i--) { 415 Node* opaq = opaque4_node(i); 416 if (!useful.member(opaq)) { 417 remove_opaque4_node(opaq); 418 } 419 } 420 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 421 bs->eliminate_useless_gc_barriers(useful, this); 422 // clean up the late inline lists 423 remove_useless_late_inlines(&_string_late_inlines, useful); 424 remove_useless_late_inlines(&_boxing_late_inlines, useful); 425 remove_useless_late_inlines(&_late_inlines, useful); 426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 427 } 428 429 //------------------------------frame_size_in_words----------------------------- 430 // frame_slots in units of words 431 int Compile::frame_size_in_words() const { 432 // shift is 0 in LP32 and 1 in LP64 433 const int shift = (LogBytesPerWord - LogBytesPerInt); 434 int words = _frame_slots >> shift; 435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 436 return words; 437 } 438 439 // To bang the stack of this compiled method we use the stack size 440 // that the interpreter would need in case of a deoptimization. This 441 // removes the need to bang the stack in the deoptimization blob which 442 // in turn simplifies stack overflow handling. 443 int Compile::bang_size_in_bytes() const { 444 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 445 } 446 447 // ============================================================================ 448 //------------------------------CompileWrapper--------------------------------- 449 class CompileWrapper : public StackObj { 450 Compile *const _compile; 451 public: 452 CompileWrapper(Compile* compile); 453 454 ~CompileWrapper(); 455 }; 456 457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 458 // the Compile* pointer is stored in the current ciEnv: 459 ciEnv* env = compile->env(); 460 assert(env == ciEnv::current(), "must already be a ciEnv active"); 461 assert(env->compiler_data() == NULL, "compile already active?"); 462 env->set_compiler_data(compile); 463 assert(compile == Compile::current(), "sanity"); 464 465 compile->set_type_dict(NULL); 466 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 467 compile->clone_map().set_clone_idx(0); 468 compile->set_type_hwm(NULL); 469 compile->set_type_last_size(0); 470 compile->set_last_tf(NULL, NULL); 471 compile->set_indexSet_arena(NULL); 472 compile->set_indexSet_free_block_list(NULL); 473 compile->init_type_arena(); 474 Type::Initialize(compile); 475 _compile->set_scratch_buffer_blob(NULL); 476 _compile->begin_method(); 477 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 478 } 479 CompileWrapper::~CompileWrapper() { 480 _compile->end_method(); 481 if (_compile->scratch_buffer_blob() != NULL) 482 BufferBlob::free(_compile->scratch_buffer_blob()); 483 _compile->env()->set_compiler_data(NULL); 484 } 485 486 487 //----------------------------print_compile_messages--------------------------- 488 void Compile::print_compile_messages() { 489 #ifndef PRODUCT 490 // Check if recompiling 491 if (_subsume_loads == false && PrintOpto) { 492 // Recompiling without allowing machine instructions to subsume loads 493 tty->print_cr("*********************************************************"); 494 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 495 tty->print_cr("*********************************************************"); 496 } 497 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 498 // Recompiling without escape analysis 499 tty->print_cr("*********************************************************"); 500 tty->print_cr("** Bailout: Recompile without escape analysis **"); 501 tty->print_cr("*********************************************************"); 502 } 503 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 504 // Recompiling without boxing elimination 505 tty->print_cr("*********************************************************"); 506 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 507 tty->print_cr("*********************************************************"); 508 } 509 if (C->directive()->BreakAtCompileOption) { 510 // Open the debugger when compiling this method. 511 tty->print("### Breaking when compiling: "); 512 method()->print_short_name(); 513 tty->cr(); 514 BREAKPOINT; 515 } 516 517 if( PrintOpto ) { 518 if (is_osr_compilation()) { 519 tty->print("[OSR]%3d", _compile_id); 520 } else { 521 tty->print("%3d", _compile_id); 522 } 523 } 524 #endif 525 } 526 527 528 //-----------------------init_scratch_buffer_blob------------------------------ 529 // Construct a temporary BufferBlob and cache it for this compile. 530 void Compile::init_scratch_buffer_blob(int const_size) { 531 // If there is already a scratch buffer blob allocated and the 532 // constant section is big enough, use it. Otherwise free the 533 // current and allocate a new one. 534 BufferBlob* blob = scratch_buffer_blob(); 535 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 536 // Use the current blob. 537 } else { 538 if (blob != NULL) { 539 BufferBlob::free(blob); 540 } 541 542 ResourceMark rm; 543 _scratch_const_size = const_size; 544 int size = C2Compiler::initial_code_buffer_size(const_size); 545 blob = BufferBlob::create("Compile::scratch_buffer", size); 546 // Record the buffer blob for next time. 547 set_scratch_buffer_blob(blob); 548 // Have we run out of code space? 549 if (scratch_buffer_blob() == NULL) { 550 // Let CompilerBroker disable further compilations. 551 record_failure("Not enough space for scratch buffer in CodeCache"); 552 return; 553 } 554 } 555 556 // Initialize the relocation buffers 557 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 558 set_scratch_locs_memory(locs_buf); 559 } 560 561 562 //-----------------------scratch_emit_size------------------------------------- 563 // Helper function that computes size by emitting code 564 uint Compile::scratch_emit_size(const Node* n) { 565 // Start scratch_emit_size section. 566 set_in_scratch_emit_size(true); 567 568 // Emit into a trash buffer and count bytes emitted. 569 // This is a pretty expensive way to compute a size, 570 // but it works well enough if seldom used. 571 // All common fixed-size instructions are given a size 572 // method by the AD file. 573 // Note that the scratch buffer blob and locs memory are 574 // allocated at the beginning of the compile task, and 575 // may be shared by several calls to scratch_emit_size. 576 // The allocation of the scratch buffer blob is particularly 577 // expensive, since it has to grab the code cache lock. 578 BufferBlob* blob = this->scratch_buffer_blob(); 579 assert(blob != NULL, "Initialize BufferBlob at start"); 580 assert(blob->size() > MAX_inst_size, "sanity"); 581 relocInfo* locs_buf = scratch_locs_memory(); 582 address blob_begin = blob->content_begin(); 583 address blob_end = (address)locs_buf; 584 assert(blob->contains(blob_end), "sanity"); 585 CodeBuffer buf(blob_begin, blob_end - blob_begin); 586 buf.initialize_consts_size(_scratch_const_size); 587 buf.initialize_stubs_size(MAX_stubs_size); 588 assert(locs_buf != NULL, "sanity"); 589 int lsize = MAX_locs_size / 3; 590 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 591 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 592 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 593 // Mark as scratch buffer. 594 buf.consts()->set_scratch_emit(); 595 buf.insts()->set_scratch_emit(); 596 buf.stubs()->set_scratch_emit(); 597 598 // Do the emission. 599 600 Label fakeL; // Fake label for branch instructions. 601 Label* saveL = NULL; 602 uint save_bnum = 0; 603 bool is_branch = n->is_MachBranch(); 604 if (is_branch) { 605 MacroAssembler masm(&buf); 606 masm.bind(fakeL); 607 n->as_MachBranch()->save_label(&saveL, &save_bnum); 608 n->as_MachBranch()->label_set(&fakeL, 0); 609 } 610 n->emit(buf, this->regalloc()); 611 612 // Emitting into the scratch buffer should not fail 613 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 614 615 if (is_branch) // Restore label. 616 n->as_MachBranch()->label_set(saveL, save_bnum); 617 618 // End scratch_emit_size section. 619 set_in_scratch_emit_size(false); 620 621 return buf.insts_size(); 622 } 623 624 625 // ============================================================================ 626 //------------------------------Compile standard------------------------------- 627 debug_only( int Compile::_debug_idx = 100000; ) 628 629 // Compile a method. entry_bci is -1 for normal compilations and indicates 630 // the continuation bci for on stack replacement. 631 632 633 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 634 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 635 : Phase(Compiler), 636 _compile_id(ci_env->compile_id()), 637 _save_argument_registers(false), 638 _subsume_loads(subsume_loads), 639 _do_escape_analysis(do_escape_analysis), 640 _eliminate_boxing(eliminate_boxing), 641 _method(target), 642 _entry_bci(osr_bci), 643 _stub_function(NULL), 644 _stub_name(NULL), 645 _stub_entry_point(NULL), 646 _max_node_limit(MaxNodeLimit), 647 _orig_pc_slot(0), 648 _orig_pc_slot_offset_in_bytes(0), 649 _inlining_progress(false), 650 _inlining_incrementally(false), 651 _do_cleanup(false), 652 _has_reserved_stack_access(target->has_reserved_stack_access()), 653 #ifndef PRODUCT 654 _trace_opto_output(directive->TraceOptoOutputOption), 655 #endif 656 _has_method_handle_invokes(false), 657 _comp_arena(mtCompiler), 658 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 659 _env(ci_env), 660 _directive(directive), 661 _log(ci_env->log()), 662 _failure_reason(NULL), 663 _congraph(NULL), 664 #ifndef PRODUCT 665 _printer(IdealGraphPrinter::printer()), 666 #endif 667 _dead_node_list(comp_arena()), 668 _dead_node_count(0), 669 _node_arena(mtCompiler), 670 _old_arena(mtCompiler), 671 _mach_constant_base_node(NULL), 672 _Compile_types(mtCompiler), 673 _initial_gvn(NULL), 674 _for_igvn(NULL), 675 _warm_calls(NULL), 676 _late_inlines(comp_arena(), 2, 0, NULL), 677 _string_late_inlines(comp_arena(), 2, 0, NULL), 678 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 679 _late_inlines_pos(0), 680 _number_of_mh_late_inlines(0), 681 _print_inlining_stream(NULL), 682 _print_inlining_list(NULL), 683 _print_inlining_idx(0), 684 _print_inlining_output(NULL), 685 _replay_inline_data(NULL), 686 _java_calls(0), 687 _inner_loops(0), 688 _interpreter_frame_size(0), 689 _node_bundling_limit(0), 690 _node_bundling_base(NULL), 691 _code_buffer("Compile::Fill_buffer"), 692 _scratch_const_size(-1), 693 _in_scratch_emit_size(false) 694 #ifndef PRODUCT 695 , _in_dump_cnt(0) 696 #endif 697 { 698 C = this; 699 #ifndef PRODUCT 700 if (_printer != NULL) { 701 _printer->set_compile(this); 702 } 703 #endif 704 CompileWrapper cw(this); 705 706 if (CITimeVerbose) { 707 tty->print(" "); 708 target->holder()->name()->print(); 709 tty->print("."); 710 target->print_short_name(); 711 tty->print(" "); 712 } 713 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 714 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 715 716 #ifndef PRODUCT 717 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 718 if (!print_opto_assembly) { 719 bool print_assembly = directive->PrintAssemblyOption; 720 if (print_assembly && !Disassembler::can_decode()) { 721 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 722 print_opto_assembly = true; 723 } 724 } 725 set_print_assembly(print_opto_assembly); 726 set_parsed_irreducible_loop(false); 727 728 if (directive->ReplayInlineOption) { 729 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 730 } 731 #endif 732 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 733 set_print_intrinsics(directive->PrintIntrinsicsOption); 734 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 735 736 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 737 // Make sure the method being compiled gets its own MDO, 738 // so we can at least track the decompile_count(). 739 // Need MDO to record RTM code generation state. 740 method()->ensure_method_data(); 741 } 742 743 Init(::AliasLevel); 744 745 746 print_compile_messages(); 747 748 _ilt = InlineTree::build_inline_tree_root(); 749 750 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 751 assert(num_alias_types() >= AliasIdxRaw, ""); 752 753 #define MINIMUM_NODE_HASH 1023 754 // Node list that Iterative GVN will start with 755 Unique_Node_List for_igvn(comp_arena()); 756 set_for_igvn(&for_igvn); 757 758 // GVN that will be run immediately on new nodes 759 uint estimated_size = method()->code_size()*4+64; 760 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 761 PhaseGVN gvn(node_arena(), estimated_size); 762 set_initial_gvn(&gvn); 763 764 print_inlining_init(); 765 { // Scope for timing the parser 766 TracePhase tp("parse", &timers[_t_parser]); 767 768 // Put top into the hash table ASAP. 769 initial_gvn()->transform_no_reclaim(top()); 770 771 // Set up tf(), start(), and find a CallGenerator. 772 CallGenerator* cg = NULL; 773 if (is_osr_compilation()) { 774 const TypeTuple *domain = StartOSRNode::osr_domain(); 775 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 776 init_tf(TypeFunc::make(domain, range)); 777 StartNode* s = new StartOSRNode(root(), domain); 778 initial_gvn()->set_type_bottom(s); 779 init_start(s); 780 cg = CallGenerator::for_osr(method(), entry_bci()); 781 } else { 782 // Normal case. 783 init_tf(TypeFunc::make(method())); 784 StartNode* s = new StartNode(root(), tf()->domain()); 785 initial_gvn()->set_type_bottom(s); 786 init_start(s); 787 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 788 // With java.lang.ref.reference.get() we must go through the 789 // intrinsic - even when get() is the root 790 // method of the compile - so that, if necessary, the value in 791 // the referent field of the reference object gets recorded by 792 // the pre-barrier code. 793 cg = find_intrinsic(method(), false); 794 } 795 if (cg == NULL) { 796 float past_uses = method()->interpreter_invocation_count(); 797 float expected_uses = past_uses; 798 cg = CallGenerator::for_inline(method(), expected_uses); 799 } 800 } 801 if (failing()) return; 802 if (cg == NULL) { 803 record_method_not_compilable("cannot parse method"); 804 return; 805 } 806 JVMState* jvms = build_start_state(start(), tf()); 807 if ((jvms = cg->generate(jvms)) == NULL) { 808 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 809 record_method_not_compilable("method parse failed"); 810 } 811 return; 812 } 813 GraphKit kit(jvms); 814 815 if (!kit.stopped()) { 816 // Accept return values, and transfer control we know not where. 817 // This is done by a special, unique ReturnNode bound to root. 818 return_values(kit.jvms()); 819 } 820 821 if (kit.has_exceptions()) { 822 // Any exceptions that escape from this call must be rethrown 823 // to whatever caller is dynamically above us on the stack. 824 // This is done by a special, unique RethrowNode bound to root. 825 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 826 } 827 828 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 829 830 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 831 inline_string_calls(true); 832 } 833 834 if (failing()) return; 835 836 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 837 838 // Remove clutter produced by parsing. 839 if (!failing()) { 840 ResourceMark rm; 841 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 842 } 843 } 844 845 // Note: Large methods are capped off in do_one_bytecode(). 846 if (failing()) return; 847 848 // After parsing, node notes are no longer automagic. 849 // They must be propagated by register_new_node_with_optimizer(), 850 // clone(), or the like. 851 set_default_node_notes(NULL); 852 853 for (;;) { 854 int successes = Inline_Warm(); 855 if (failing()) return; 856 if (successes == 0) break; 857 } 858 859 // Drain the list. 860 Finish_Warm(); 861 #ifndef PRODUCT 862 if (_printer && _printer->should_print(1)) { 863 _printer->print_inlining(); 864 } 865 #endif 866 867 if (failing()) return; 868 NOT_PRODUCT( verify_graph_edges(); ) 869 870 // Now optimize 871 Optimize(); 872 if (failing()) return; 873 NOT_PRODUCT( verify_graph_edges(); ) 874 875 #ifndef PRODUCT 876 if (PrintIdeal) { 877 ttyLocker ttyl; // keep the following output all in one block 878 // This output goes directly to the tty, not the compiler log. 879 // To enable tools to match it up with the compilation activity, 880 // be sure to tag this tty output with the compile ID. 881 if (xtty != NULL) { 882 xtty->head("ideal compile_id='%d'%s", compile_id(), 883 is_osr_compilation() ? " compile_kind='osr'" : 884 ""); 885 } 886 root()->dump(9999); 887 if (xtty != NULL) { 888 xtty->tail("ideal"); 889 } 890 } 891 #endif 892 893 #ifdef ASSERT 894 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 895 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 896 #endif 897 898 // Dump compilation data to replay it. 899 if (directive->DumpReplayOption) { 900 env()->dump_replay_data(_compile_id); 901 } 902 if (directive->DumpInlineOption && (ilt() != NULL)) { 903 env()->dump_inline_data(_compile_id); 904 } 905 906 // Now that we know the size of all the monitors we can add a fixed slot 907 // for the original deopt pc. 908 909 _orig_pc_slot = fixed_slots(); 910 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 911 set_fixed_slots(next_slot); 912 913 // Compute when to use implicit null checks. Used by matching trap based 914 // nodes and NullCheck optimization. 915 set_allowed_deopt_reasons(); 916 917 // Now generate code 918 Code_Gen(); 919 if (failing()) return; 920 921 // Check if we want to skip execution of all compiled code. 922 { 923 #ifndef PRODUCT 924 if (OptoNoExecute) { 925 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 926 return; 927 } 928 #endif 929 TracePhase tp("install_code", &timers[_t_registerMethod]); 930 931 if (is_osr_compilation()) { 932 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 933 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 934 } else { 935 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 936 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 937 } 938 939 env()->register_method(_method, _entry_bci, 940 &_code_offsets, 941 _orig_pc_slot_offset_in_bytes, 942 code_buffer(), 943 frame_size_in_words(), _oop_map_set, 944 &_handler_table, &_inc_table, 945 compiler, 946 has_unsafe_access(), 947 SharedRuntime::is_wide_vector(max_vector_size()), 948 rtm_state() 949 ); 950 951 if (log() != NULL) // Print code cache state into compiler log 952 log()->code_cache_state(); 953 } 954 } 955 956 //------------------------------Compile---------------------------------------- 957 // Compile a runtime stub 958 Compile::Compile( ciEnv* ci_env, 959 TypeFunc_generator generator, 960 address stub_function, 961 const char *stub_name, 962 int is_fancy_jump, 963 bool pass_tls, 964 bool save_arg_registers, 965 bool return_pc, 966 DirectiveSet* directive) 967 : Phase(Compiler), 968 _compile_id(0), 969 _save_argument_registers(save_arg_registers), 970 _subsume_loads(true), 971 _do_escape_analysis(false), 972 _eliminate_boxing(false), 973 _method(NULL), 974 _entry_bci(InvocationEntryBci), 975 _stub_function(stub_function), 976 _stub_name(stub_name), 977 _stub_entry_point(NULL), 978 _max_node_limit(MaxNodeLimit), 979 _orig_pc_slot(0), 980 _orig_pc_slot_offset_in_bytes(0), 981 _inlining_progress(false), 982 _inlining_incrementally(false), 983 _has_reserved_stack_access(false), 984 #ifndef PRODUCT 985 _trace_opto_output(directive->TraceOptoOutputOption), 986 #endif 987 _has_method_handle_invokes(false), 988 _comp_arena(mtCompiler), 989 _env(ci_env), 990 _directive(directive), 991 _log(ci_env->log()), 992 _failure_reason(NULL), 993 _congraph(NULL), 994 #ifndef PRODUCT 995 _printer(NULL), 996 #endif 997 _dead_node_list(comp_arena()), 998 _dead_node_count(0), 999 _node_arena(mtCompiler), 1000 _old_arena(mtCompiler), 1001 _mach_constant_base_node(NULL), 1002 _Compile_types(mtCompiler), 1003 _initial_gvn(NULL), 1004 _for_igvn(NULL), 1005 _warm_calls(NULL), 1006 _number_of_mh_late_inlines(0), 1007 _print_inlining_stream(NULL), 1008 _print_inlining_list(NULL), 1009 _print_inlining_idx(0), 1010 _print_inlining_output(NULL), 1011 _replay_inline_data(NULL), 1012 _java_calls(0), 1013 _inner_loops(0), 1014 _interpreter_frame_size(0), 1015 _node_bundling_limit(0), 1016 _node_bundling_base(NULL), 1017 _code_buffer("Compile::Fill_buffer"), 1018 #ifndef PRODUCT 1019 _in_dump_cnt(0), 1020 #endif 1021 _allowed_reasons(0) { 1022 C = this; 1023 1024 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1025 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1026 1027 #ifndef PRODUCT 1028 set_print_assembly(PrintFrameConverterAssembly); 1029 set_parsed_irreducible_loop(false); 1030 #endif 1031 set_has_irreducible_loop(false); // no loops 1032 1033 CompileWrapper cw(this); 1034 Init(/*AliasLevel=*/ 0); 1035 init_tf((*generator)()); 1036 1037 { 1038 // The following is a dummy for the sake of GraphKit::gen_stub 1039 Unique_Node_List for_igvn(comp_arena()); 1040 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1041 PhaseGVN gvn(Thread::current()->resource_area(),255); 1042 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1043 gvn.transform_no_reclaim(top()); 1044 1045 GraphKit kit; 1046 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1047 } 1048 1049 NOT_PRODUCT( verify_graph_edges(); ) 1050 Code_Gen(); 1051 if (failing()) return; 1052 1053 1054 // Entry point will be accessed using compile->stub_entry_point(); 1055 if (code_buffer() == NULL) { 1056 Matcher::soft_match_failure(); 1057 } else { 1058 if (PrintAssembly && (WizardMode || Verbose)) 1059 tty->print_cr("### Stub::%s", stub_name); 1060 1061 if (!failing()) { 1062 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1063 1064 // Make the NMethod 1065 // For now we mark the frame as never safe for profile stackwalking 1066 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1067 code_buffer(), 1068 CodeOffsets::frame_never_safe, 1069 // _code_offsets.value(CodeOffsets::Frame_Complete), 1070 frame_size_in_words(), 1071 _oop_map_set, 1072 save_arg_registers); 1073 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1074 1075 _stub_entry_point = rs->entry_point(); 1076 } 1077 } 1078 } 1079 1080 //------------------------------Init------------------------------------------- 1081 // Prepare for a single compilation 1082 void Compile::Init(int aliaslevel) { 1083 _unique = 0; 1084 _regalloc = NULL; 1085 1086 _tf = NULL; // filled in later 1087 _top = NULL; // cached later 1088 _matcher = NULL; // filled in later 1089 _cfg = NULL; // filled in later 1090 1091 set_24_bit_selection_and_mode(Use24BitFP, false); 1092 1093 _node_note_array = NULL; 1094 _default_node_notes = NULL; 1095 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1096 1097 _immutable_memory = NULL; // filled in at first inquiry 1098 1099 // Globally visible Nodes 1100 // First set TOP to NULL to give safe behavior during creation of RootNode 1101 set_cached_top_node(NULL); 1102 set_root(new RootNode()); 1103 // Now that you have a Root to point to, create the real TOP 1104 set_cached_top_node( new ConNode(Type::TOP) ); 1105 set_recent_alloc(NULL, NULL); 1106 1107 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1108 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1109 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1110 env()->set_dependencies(new Dependencies(env())); 1111 1112 _fixed_slots = 0; 1113 set_has_split_ifs(false); 1114 set_has_loops(has_method() && method()->has_loops()); // first approximation 1115 set_has_stringbuilder(false); 1116 set_has_boxed_value(false); 1117 _trap_can_recompile = false; // no traps emitted yet 1118 _major_progress = true; // start out assuming good things will happen 1119 set_has_unsafe_access(false); 1120 set_max_vector_size(0); 1121 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1122 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1123 set_decompile_count(0); 1124 1125 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1126 _loop_opts_cnt = LoopOptsCount; 1127 set_do_inlining(Inline); 1128 set_max_inline_size(MaxInlineSize); 1129 set_freq_inline_size(FreqInlineSize); 1130 set_do_scheduling(OptoScheduling); 1131 set_do_count_invocations(false); 1132 set_do_method_data_update(false); 1133 1134 set_do_vector_loop(false); 1135 1136 if (AllowVectorizeOnDemand) { 1137 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1138 set_do_vector_loop(true); 1139 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());}) 1140 } else if (has_method() && method()->name() != 0 && 1141 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1142 set_do_vector_loop(true); 1143 } 1144 } 1145 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1146 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());}) 1147 1148 set_age_code(has_method() && method()->profile_aging()); 1149 set_rtm_state(NoRTM); // No RTM lock eliding by default 1150 _max_node_limit = _directive->MaxNodeLimitOption; 1151 1152 #if INCLUDE_RTM_OPT 1153 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1154 int rtm_state = method()->method_data()->rtm_state(); 1155 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1156 // Don't generate RTM lock eliding code. 1157 set_rtm_state(NoRTM); 1158 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1159 // Generate RTM lock eliding code without abort ratio calculation code. 1160 set_rtm_state(UseRTM); 1161 } else if (UseRTMDeopt) { 1162 // Generate RTM lock eliding code and include abort ratio calculation 1163 // code if UseRTMDeopt is on. 1164 set_rtm_state(ProfileRTM); 1165 } 1166 } 1167 #endif 1168 if (debug_info()->recording_non_safepoints()) { 1169 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1170 (comp_arena(), 8, 0, NULL)); 1171 set_default_node_notes(Node_Notes::make(this)); 1172 } 1173 1174 // // -- Initialize types before each compile -- 1175 // // Update cached type information 1176 // if( _method && _method->constants() ) 1177 // Type::update_loaded_types(_method, _method->constants()); 1178 1179 // Init alias_type map. 1180 if (!_do_escape_analysis && aliaslevel == 3) 1181 aliaslevel = 2; // No unique types without escape analysis 1182 _AliasLevel = aliaslevel; 1183 const int grow_ats = 16; 1184 _max_alias_types = grow_ats; 1185 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1186 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1187 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1188 { 1189 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1190 } 1191 // Initialize the first few types. 1192 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1193 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1194 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1195 _num_alias_types = AliasIdxRaw+1; 1196 // Zero out the alias type cache. 1197 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1198 // A NULL adr_type hits in the cache right away. Preload the right answer. 1199 probe_alias_cache(NULL)->_index = AliasIdxTop; 1200 1201 _intrinsics = NULL; 1202 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1203 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1204 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1205 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1206 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1207 register_library_intrinsics(); 1208 } 1209 1210 //---------------------------init_start---------------------------------------- 1211 // Install the StartNode on this compile object. 1212 void Compile::init_start(StartNode* s) { 1213 if (failing()) 1214 return; // already failing 1215 assert(s == start(), ""); 1216 } 1217 1218 /** 1219 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1220 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1221 * the ideal graph. 1222 */ 1223 StartNode* Compile::start() const { 1224 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1225 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1226 Node* start = root()->fast_out(i); 1227 if (start->is_Start()) { 1228 return start->as_Start(); 1229 } 1230 } 1231 fatal("Did not find Start node!"); 1232 return NULL; 1233 } 1234 1235 //-------------------------------immutable_memory------------------------------------- 1236 // Access immutable memory 1237 Node* Compile::immutable_memory() { 1238 if (_immutable_memory != NULL) { 1239 return _immutable_memory; 1240 } 1241 StartNode* s = start(); 1242 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1243 Node *p = s->fast_out(i); 1244 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1245 _immutable_memory = p; 1246 return _immutable_memory; 1247 } 1248 } 1249 ShouldNotReachHere(); 1250 return NULL; 1251 } 1252 1253 //----------------------set_cached_top_node------------------------------------ 1254 // Install the cached top node, and make sure Node::is_top works correctly. 1255 void Compile::set_cached_top_node(Node* tn) { 1256 if (tn != NULL) verify_top(tn); 1257 Node* old_top = _top; 1258 _top = tn; 1259 // Calling Node::setup_is_top allows the nodes the chance to adjust 1260 // their _out arrays. 1261 if (_top != NULL) _top->setup_is_top(); 1262 if (old_top != NULL) old_top->setup_is_top(); 1263 assert(_top == NULL || top()->is_top(), ""); 1264 } 1265 1266 #ifdef ASSERT 1267 uint Compile::count_live_nodes_by_graph_walk() { 1268 Unique_Node_List useful(comp_arena()); 1269 // Get useful node list by walking the graph. 1270 identify_useful_nodes(useful); 1271 return useful.size(); 1272 } 1273 1274 void Compile::print_missing_nodes() { 1275 1276 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1277 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1278 return; 1279 } 1280 1281 // This is an expensive function. It is executed only when the user 1282 // specifies VerifyIdealNodeCount option or otherwise knows the 1283 // additional work that needs to be done to identify reachable nodes 1284 // by walking the flow graph and find the missing ones using 1285 // _dead_node_list. 1286 1287 Unique_Node_List useful(comp_arena()); 1288 // Get useful node list by walking the graph. 1289 identify_useful_nodes(useful); 1290 1291 uint l_nodes = C->live_nodes(); 1292 uint l_nodes_by_walk = useful.size(); 1293 1294 if (l_nodes != l_nodes_by_walk) { 1295 if (_log != NULL) { 1296 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1297 _log->stamp(); 1298 _log->end_head(); 1299 } 1300 VectorSet& useful_member_set = useful.member_set(); 1301 int last_idx = l_nodes_by_walk; 1302 for (int i = 0; i < last_idx; i++) { 1303 if (useful_member_set.test(i)) { 1304 if (_dead_node_list.test(i)) { 1305 if (_log != NULL) { 1306 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1307 } 1308 if (PrintIdealNodeCount) { 1309 // Print the log message to tty 1310 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1311 useful.at(i)->dump(); 1312 } 1313 } 1314 } 1315 else if (! _dead_node_list.test(i)) { 1316 if (_log != NULL) { 1317 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1318 } 1319 if (PrintIdealNodeCount) { 1320 // Print the log message to tty 1321 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1322 } 1323 } 1324 } 1325 if (_log != NULL) { 1326 _log->tail("mismatched_nodes"); 1327 } 1328 } 1329 } 1330 void Compile::record_modified_node(Node* n) { 1331 if (_modified_nodes != NULL && !_inlining_incrementally && 1332 n->outcnt() != 0 && !n->is_Con()) { 1333 _modified_nodes->push(n); 1334 } 1335 } 1336 1337 void Compile::remove_modified_node(Node* n) { 1338 if (_modified_nodes != NULL) { 1339 _modified_nodes->remove(n); 1340 } 1341 } 1342 #endif 1343 1344 #ifndef PRODUCT 1345 void Compile::verify_top(Node* tn) const { 1346 if (tn != NULL) { 1347 assert(tn->is_Con(), "top node must be a constant"); 1348 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1349 assert(tn->in(0) != NULL, "must have live top node"); 1350 } 1351 } 1352 #endif 1353 1354 1355 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1356 1357 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1358 guarantee(arr != NULL, ""); 1359 int num_blocks = arr->length(); 1360 if (grow_by < num_blocks) grow_by = num_blocks; 1361 int num_notes = grow_by * _node_notes_block_size; 1362 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1363 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1364 while (num_notes > 0) { 1365 arr->append(notes); 1366 notes += _node_notes_block_size; 1367 num_notes -= _node_notes_block_size; 1368 } 1369 assert(num_notes == 0, "exact multiple, please"); 1370 } 1371 1372 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1373 if (source == NULL || dest == NULL) return false; 1374 1375 if (dest->is_Con()) 1376 return false; // Do not push debug info onto constants. 1377 1378 #ifdef ASSERT 1379 // Leave a bread crumb trail pointing to the original node: 1380 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1381 dest->set_debug_orig(source); 1382 } 1383 #endif 1384 1385 if (node_note_array() == NULL) 1386 return false; // Not collecting any notes now. 1387 1388 // This is a copy onto a pre-existing node, which may already have notes. 1389 // If both nodes have notes, do not overwrite any pre-existing notes. 1390 Node_Notes* source_notes = node_notes_at(source->_idx); 1391 if (source_notes == NULL || source_notes->is_clear()) return false; 1392 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1393 if (dest_notes == NULL || dest_notes->is_clear()) { 1394 return set_node_notes_at(dest->_idx, source_notes); 1395 } 1396 1397 Node_Notes merged_notes = (*source_notes); 1398 // The order of operations here ensures that dest notes will win... 1399 merged_notes.update_from(dest_notes); 1400 return set_node_notes_at(dest->_idx, &merged_notes); 1401 } 1402 1403 1404 //--------------------------allow_range_check_smearing------------------------- 1405 // Gating condition for coalescing similar range checks. 1406 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1407 // single covering check that is at least as strong as any of them. 1408 // If the optimization succeeds, the simplified (strengthened) range check 1409 // will always succeed. If it fails, we will deopt, and then give up 1410 // on the optimization. 1411 bool Compile::allow_range_check_smearing() const { 1412 // If this method has already thrown a range-check, 1413 // assume it was because we already tried range smearing 1414 // and it failed. 1415 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1416 return !already_trapped; 1417 } 1418 1419 1420 //------------------------------flatten_alias_type----------------------------- 1421 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1422 int offset = tj->offset(); 1423 TypePtr::PTR ptr = tj->ptr(); 1424 1425 // Known instance (scalarizable allocation) alias only with itself. 1426 bool is_known_inst = tj->isa_oopptr() != NULL && 1427 tj->is_oopptr()->is_known_instance(); 1428 1429 // Process weird unsafe references. 1430 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1431 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1432 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1433 tj = TypeOopPtr::BOTTOM; 1434 ptr = tj->ptr(); 1435 offset = tj->offset(); 1436 } 1437 1438 // Array pointers need some flattening 1439 const TypeAryPtr *ta = tj->isa_aryptr(); 1440 if (ta && ta->is_stable()) { 1441 // Erase stability property for alias analysis. 1442 tj = ta = ta->cast_to_stable(false); 1443 } 1444 if( ta && is_known_inst ) { 1445 if ( offset != Type::OffsetBot && 1446 offset > arrayOopDesc::length_offset_in_bytes() ) { 1447 offset = Type::OffsetBot; // Flatten constant access into array body only 1448 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1449 } 1450 } else if( ta && _AliasLevel >= 2 ) { 1451 // For arrays indexed by constant indices, we flatten the alias 1452 // space to include all of the array body. Only the header, klass 1453 // and array length can be accessed un-aliased. 1454 if( offset != Type::OffsetBot ) { 1455 if( ta->const_oop() ) { // MethodData* or Method* 1456 offset = Type::OffsetBot; // Flatten constant access into array body 1457 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1458 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1459 // range is OK as-is. 1460 tj = ta = TypeAryPtr::RANGE; 1461 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1462 tj = TypeInstPtr::KLASS; // all klass loads look alike 1463 ta = TypeAryPtr::RANGE; // generic ignored junk 1464 ptr = TypePtr::BotPTR; 1465 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1466 tj = TypeInstPtr::MARK; 1467 ta = TypeAryPtr::RANGE; // generic ignored junk 1468 ptr = TypePtr::BotPTR; 1469 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1470 ta = tj->isa_aryptr(); 1471 } else { // Random constant offset into array body 1472 offset = Type::OffsetBot; // Flatten constant access into array body 1473 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1474 } 1475 } 1476 // Arrays of fixed size alias with arrays of unknown size. 1477 if (ta->size() != TypeInt::POS) { 1478 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1479 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1480 } 1481 // Arrays of known objects become arrays of unknown objects. 1482 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1483 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1484 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1485 } 1486 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1487 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1488 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1489 } 1490 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1491 // cannot be distinguished by bytecode alone. 1492 if (ta->elem() == TypeInt::BOOL) { 1493 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1494 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1495 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1496 } 1497 // During the 2nd round of IterGVN, NotNull castings are removed. 1498 // Make sure the Bottom and NotNull variants alias the same. 1499 // Also, make sure exact and non-exact variants alias the same. 1500 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1501 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1502 } 1503 } 1504 1505 // Oop pointers need some flattening 1506 const TypeInstPtr *to = tj->isa_instptr(); 1507 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1508 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1509 if( ptr == TypePtr::Constant ) { 1510 if (to->klass() != ciEnv::current()->Class_klass() || 1511 offset < k->size_helper() * wordSize) { 1512 // No constant oop pointers (such as Strings); they alias with 1513 // unknown strings. 1514 assert(!is_known_inst, "not scalarizable allocation"); 1515 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1516 } 1517 } else if( is_known_inst ) { 1518 tj = to; // Keep NotNull and klass_is_exact for instance type 1519 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1520 // During the 2nd round of IterGVN, NotNull castings are removed. 1521 // Make sure the Bottom and NotNull variants alias the same. 1522 // Also, make sure exact and non-exact variants alias the same. 1523 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1524 } 1525 if (to->speculative() != NULL) { 1526 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1527 } 1528 // Canonicalize the holder of this field 1529 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1530 // First handle header references such as a LoadKlassNode, even if the 1531 // object's klass is unloaded at compile time (4965979). 1532 if (!is_known_inst) { // Do it only for non-instance types 1533 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1534 } 1535 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1536 to = tj->is_instptr(); 1537 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1538 // Static fields are in the space above the normal instance 1539 // fields in the java.lang.Class instance. 1540 if (to->klass() != ciEnv::current()->Class_klass()) { 1541 to = NULL; 1542 tj = TypeOopPtr::BOTTOM; 1543 offset = tj->offset(); 1544 } 1545 } else { 1546 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1547 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1548 if( is_known_inst ) { 1549 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1550 } else { 1551 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1552 } 1553 } 1554 } 1555 } 1556 1557 // Klass pointers to object array klasses need some flattening 1558 const TypeKlassPtr *tk = tj->isa_klassptr(); 1559 if( tk ) { 1560 // If we are referencing a field within a Klass, we need 1561 // to assume the worst case of an Object. Both exact and 1562 // inexact types must flatten to the same alias class so 1563 // use NotNull as the PTR. 1564 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1565 1566 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1567 TypeKlassPtr::OBJECT->klass(), 1568 offset); 1569 } 1570 1571 ciKlass* klass = tk->klass(); 1572 if( klass->is_obj_array_klass() ) { 1573 ciKlass* k = TypeAryPtr::OOPS->klass(); 1574 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1575 k = TypeInstPtr::BOTTOM->klass(); 1576 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1577 } 1578 1579 // Check for precise loads from the primary supertype array and force them 1580 // to the supertype cache alias index. Check for generic array loads from 1581 // the primary supertype array and also force them to the supertype cache 1582 // alias index. Since the same load can reach both, we need to merge 1583 // these 2 disparate memories into the same alias class. Since the 1584 // primary supertype array is read-only, there's no chance of confusion 1585 // where we bypass an array load and an array store. 1586 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1587 if (offset == Type::OffsetBot || 1588 (offset >= primary_supers_offset && 1589 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1590 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1591 offset = in_bytes(Klass::secondary_super_cache_offset()); 1592 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1593 } 1594 } 1595 1596 // Flatten all Raw pointers together. 1597 if (tj->base() == Type::RawPtr) 1598 tj = TypeRawPtr::BOTTOM; 1599 1600 if (tj->base() == Type::AnyPtr) 1601 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1602 1603 // Flatten all to bottom for now 1604 switch( _AliasLevel ) { 1605 case 0: 1606 tj = TypePtr::BOTTOM; 1607 break; 1608 case 1: // Flatten to: oop, static, field or array 1609 switch (tj->base()) { 1610 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1611 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1612 case Type::AryPtr: // do not distinguish arrays at all 1613 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1614 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1615 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1616 default: ShouldNotReachHere(); 1617 } 1618 break; 1619 case 2: // No collapsing at level 2; keep all splits 1620 case 3: // No collapsing at level 3; keep all splits 1621 break; 1622 default: 1623 Unimplemented(); 1624 } 1625 1626 offset = tj->offset(); 1627 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1628 1629 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1630 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1631 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1632 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1633 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1634 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1635 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1636 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), 1637 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1638 assert( tj->ptr() != TypePtr::TopPTR && 1639 tj->ptr() != TypePtr::AnyNull && 1640 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1641 // assert( tj->ptr() != TypePtr::Constant || 1642 // tj->base() == Type::RawPtr || 1643 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1644 1645 return tj; 1646 } 1647 1648 void Compile::AliasType::Init(int i, const TypePtr* at) { 1649 _index = i; 1650 _adr_type = at; 1651 _field = NULL; 1652 _element = NULL; 1653 _is_rewritable = true; // default 1654 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1655 if (atoop != NULL && atoop->is_known_instance()) { 1656 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1657 _general_index = Compile::current()->get_alias_index(gt); 1658 } else { 1659 _general_index = 0; 1660 } 1661 } 1662 1663 BasicType Compile::AliasType::basic_type() const { 1664 if (element() != NULL) { 1665 const Type* element = adr_type()->is_aryptr()->elem(); 1666 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1667 } if (field() != NULL) { 1668 return field()->layout_type(); 1669 } else { 1670 return T_ILLEGAL; // unknown 1671 } 1672 } 1673 1674 //---------------------------------print_on------------------------------------ 1675 #ifndef PRODUCT 1676 void Compile::AliasType::print_on(outputStream* st) { 1677 if (index() < 10) 1678 st->print("@ <%d> ", index()); 1679 else st->print("@ <%d>", index()); 1680 st->print(is_rewritable() ? " " : " RO"); 1681 int offset = adr_type()->offset(); 1682 if (offset == Type::OffsetBot) 1683 st->print(" +any"); 1684 else st->print(" +%-3d", offset); 1685 st->print(" in "); 1686 adr_type()->dump_on(st); 1687 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1688 if (field() != NULL && tjp) { 1689 if (tjp->klass() != field()->holder() || 1690 tjp->offset() != field()->offset_in_bytes()) { 1691 st->print(" != "); 1692 field()->print(); 1693 st->print(" ***"); 1694 } 1695 } 1696 } 1697 1698 void print_alias_types() { 1699 Compile* C = Compile::current(); 1700 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1701 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1702 C->alias_type(idx)->print_on(tty); 1703 tty->cr(); 1704 } 1705 } 1706 #endif 1707 1708 1709 //----------------------------probe_alias_cache-------------------------------- 1710 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1711 intptr_t key = (intptr_t) adr_type; 1712 key ^= key >> logAliasCacheSize; 1713 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1714 } 1715 1716 1717 //-----------------------------grow_alias_types-------------------------------- 1718 void Compile::grow_alias_types() { 1719 const int old_ats = _max_alias_types; // how many before? 1720 const int new_ats = old_ats; // how many more? 1721 const int grow_ats = old_ats+new_ats; // how many now? 1722 _max_alias_types = grow_ats; 1723 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1724 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1725 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1726 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1727 } 1728 1729 1730 //--------------------------------find_alias_type------------------------------ 1731 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1732 if (_AliasLevel == 0) 1733 return alias_type(AliasIdxBot); 1734 1735 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1736 if (ace->_adr_type == adr_type) { 1737 return alias_type(ace->_index); 1738 } 1739 1740 // Handle special cases. 1741 if (adr_type == NULL) return alias_type(AliasIdxTop); 1742 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1743 1744 // Do it the slow way. 1745 const TypePtr* flat = flatten_alias_type(adr_type); 1746 1747 #ifdef ASSERT 1748 { 1749 ResourceMark rm; 1750 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1751 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1752 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1753 Type::str(adr_type)); 1754 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1755 const TypeOopPtr* foop = flat->is_oopptr(); 1756 // Scalarizable allocations have exact klass always. 1757 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1758 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1759 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1760 Type::str(foop), Type::str(xoop)); 1761 } 1762 } 1763 #endif 1764 1765 int idx = AliasIdxTop; 1766 for (int i = 0; i < num_alias_types(); i++) { 1767 if (alias_type(i)->adr_type() == flat) { 1768 idx = i; 1769 break; 1770 } 1771 } 1772 1773 if (idx == AliasIdxTop) { 1774 if (no_create) return NULL; 1775 // Grow the array if necessary. 1776 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1777 // Add a new alias type. 1778 idx = _num_alias_types++; 1779 _alias_types[idx]->Init(idx, flat); 1780 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1781 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1782 if (flat->isa_instptr()) { 1783 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1784 && flat->is_instptr()->klass() == env()->Class_klass()) 1785 alias_type(idx)->set_rewritable(false); 1786 } 1787 if (flat->isa_aryptr()) { 1788 #ifdef ASSERT 1789 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1790 // (T_BYTE has the weakest alignment and size restrictions...) 1791 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1792 #endif 1793 if (flat->offset() == TypePtr::OffsetBot) { 1794 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1795 } 1796 } 1797 if (flat->isa_klassptr()) { 1798 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1799 alias_type(idx)->set_rewritable(false); 1800 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1801 alias_type(idx)->set_rewritable(false); 1802 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1803 alias_type(idx)->set_rewritable(false); 1804 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1805 alias_type(idx)->set_rewritable(false); 1806 } 1807 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1808 // but the base pointer type is not distinctive enough to identify 1809 // references into JavaThread.) 1810 1811 // Check for final fields. 1812 const TypeInstPtr* tinst = flat->isa_instptr(); 1813 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1814 ciField* field; 1815 if (tinst->const_oop() != NULL && 1816 tinst->klass() == ciEnv::current()->Class_klass() && 1817 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1818 // static field 1819 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1820 field = k->get_field_by_offset(tinst->offset(), true); 1821 } else { 1822 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1823 field = k->get_field_by_offset(tinst->offset(), false); 1824 } 1825 assert(field == NULL || 1826 original_field == NULL || 1827 (field->holder() == original_field->holder() && 1828 field->offset() == original_field->offset() && 1829 field->is_static() == original_field->is_static()), "wrong field?"); 1830 // Set field() and is_rewritable() attributes. 1831 if (field != NULL) alias_type(idx)->set_field(field); 1832 } 1833 } 1834 1835 // Fill the cache for next time. 1836 ace->_adr_type = adr_type; 1837 ace->_index = idx; 1838 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1839 1840 // Might as well try to fill the cache for the flattened version, too. 1841 AliasCacheEntry* face = probe_alias_cache(flat); 1842 if (face->_adr_type == NULL) { 1843 face->_adr_type = flat; 1844 face->_index = idx; 1845 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1846 } 1847 1848 return alias_type(idx); 1849 } 1850 1851 1852 Compile::AliasType* Compile::alias_type(ciField* field) { 1853 const TypeOopPtr* t; 1854 if (field->is_static()) 1855 t = TypeInstPtr::make(field->holder()->java_mirror()); 1856 else 1857 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1858 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1859 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1860 return atp; 1861 } 1862 1863 1864 //------------------------------have_alias_type-------------------------------- 1865 bool Compile::have_alias_type(const TypePtr* adr_type) { 1866 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1867 if (ace->_adr_type == adr_type) { 1868 return true; 1869 } 1870 1871 // Handle special cases. 1872 if (adr_type == NULL) return true; 1873 if (adr_type == TypePtr::BOTTOM) return true; 1874 1875 return find_alias_type(adr_type, true, NULL) != NULL; 1876 } 1877 1878 //-----------------------------must_alias-------------------------------------- 1879 // True if all values of the given address type are in the given alias category. 1880 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1881 if (alias_idx == AliasIdxBot) return true; // the universal category 1882 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1883 if (alias_idx == AliasIdxTop) return false; // the empty category 1884 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1885 1886 // the only remaining possible overlap is identity 1887 int adr_idx = get_alias_index(adr_type); 1888 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1889 assert(adr_idx == alias_idx || 1890 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1891 && adr_type != TypeOopPtr::BOTTOM), 1892 "should not be testing for overlap with an unsafe pointer"); 1893 return adr_idx == alias_idx; 1894 } 1895 1896 //------------------------------can_alias-------------------------------------- 1897 // True if any values of the given address type are in the given alias category. 1898 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1899 if (alias_idx == AliasIdxTop) return false; // the empty category 1900 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1901 if (alias_idx == AliasIdxBot) return true; // the universal category 1902 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1903 1904 // the only remaining possible overlap is identity 1905 int adr_idx = get_alias_index(adr_type); 1906 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1907 return adr_idx == alias_idx; 1908 } 1909 1910 1911 1912 //---------------------------pop_warm_call------------------------------------- 1913 WarmCallInfo* Compile::pop_warm_call() { 1914 WarmCallInfo* wci = _warm_calls; 1915 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1916 return wci; 1917 } 1918 1919 //----------------------------Inline_Warm-------------------------------------- 1920 int Compile::Inline_Warm() { 1921 // If there is room, try to inline some more warm call sites. 1922 // %%% Do a graph index compaction pass when we think we're out of space? 1923 if (!InlineWarmCalls) return 0; 1924 1925 int calls_made_hot = 0; 1926 int room_to_grow = NodeCountInliningCutoff - unique(); 1927 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1928 int amount_grown = 0; 1929 WarmCallInfo* call; 1930 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1931 int est_size = (int)call->size(); 1932 if (est_size > (room_to_grow - amount_grown)) { 1933 // This one won't fit anyway. Get rid of it. 1934 call->make_cold(); 1935 continue; 1936 } 1937 call->make_hot(); 1938 calls_made_hot++; 1939 amount_grown += est_size; 1940 amount_to_grow -= est_size; 1941 } 1942 1943 if (calls_made_hot > 0) set_major_progress(); 1944 return calls_made_hot; 1945 } 1946 1947 1948 //----------------------------Finish_Warm-------------------------------------- 1949 void Compile::Finish_Warm() { 1950 if (!InlineWarmCalls) return; 1951 if (failing()) return; 1952 if (warm_calls() == NULL) return; 1953 1954 // Clean up loose ends, if we are out of space for inlining. 1955 WarmCallInfo* call; 1956 while ((call = pop_warm_call()) != NULL) { 1957 call->make_cold(); 1958 } 1959 } 1960 1961 //---------------------cleanup_loop_predicates----------------------- 1962 // Remove the opaque nodes that protect the predicates so that all unused 1963 // checks and uncommon_traps will be eliminated from the ideal graph 1964 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1965 if (predicate_count()==0) return; 1966 for (int i = predicate_count(); i > 0; i--) { 1967 Node * n = predicate_opaque1_node(i-1); 1968 assert(n->Opcode() == Op_Opaque1, "must be"); 1969 igvn.replace_node(n, n->in(1)); 1970 } 1971 assert(predicate_count()==0, "should be clean!"); 1972 } 1973 1974 void Compile::add_range_check_cast(Node* n) { 1975 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1976 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1977 _range_check_casts->append(n); 1978 } 1979 1980 // Remove all range check dependent CastIINodes. 1981 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1982 for (int i = range_check_cast_count(); i > 0; i--) { 1983 Node* cast = range_check_cast_node(i-1); 1984 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1985 igvn.replace_node(cast, cast->in(1)); 1986 } 1987 assert(range_check_cast_count() == 0, "should be empty"); 1988 } 1989 1990 void Compile::add_opaque4_node(Node* n) { 1991 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1992 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1993 _opaque4_nodes->append(n); 1994 } 1995 1996 // Remove all Opaque4 nodes. 1997 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1998 for (int i = opaque4_count(); i > 0; i--) { 1999 Node* opaq = opaque4_node(i-1); 2000 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2001 igvn.replace_node(opaq, opaq->in(2)); 2002 } 2003 assert(opaque4_count() == 0, "should be empty"); 2004 } 2005 2006 // StringOpts and late inlining of string methods 2007 void Compile::inline_string_calls(bool parse_time) { 2008 { 2009 // remove useless nodes to make the usage analysis simpler 2010 ResourceMark rm; 2011 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2012 } 2013 2014 { 2015 ResourceMark rm; 2016 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2017 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2018 print_method(PHASE_AFTER_STRINGOPTS, 3); 2019 } 2020 2021 // now inline anything that we skipped the first time around 2022 if (!parse_time) { 2023 _late_inlines_pos = _late_inlines.length(); 2024 } 2025 2026 while (_string_late_inlines.length() > 0) { 2027 CallGenerator* cg = _string_late_inlines.pop(); 2028 cg->do_late_inline(); 2029 if (failing()) return; 2030 } 2031 _string_late_inlines.trunc_to(0); 2032 } 2033 2034 // Late inlining of boxing methods 2035 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2036 if (_boxing_late_inlines.length() > 0) { 2037 assert(has_boxed_value(), "inconsistent"); 2038 2039 PhaseGVN* gvn = initial_gvn(); 2040 set_inlining_incrementally(true); 2041 2042 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2043 for_igvn()->clear(); 2044 gvn->replace_with(&igvn); 2045 2046 _late_inlines_pos = _late_inlines.length(); 2047 2048 while (_boxing_late_inlines.length() > 0) { 2049 CallGenerator* cg = _boxing_late_inlines.pop(); 2050 cg->do_late_inline(); 2051 if (failing()) return; 2052 } 2053 _boxing_late_inlines.trunc_to(0); 2054 2055 inline_incrementally_cleanup(igvn); 2056 2057 set_inlining_incrementally(false); 2058 } 2059 } 2060 2061 bool Compile::inline_incrementally_one() { 2062 assert(IncrementalInline, "incremental inlining should be on"); 2063 2064 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2065 set_inlining_progress(false); 2066 set_do_cleanup(false); 2067 int i = 0; 2068 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2069 CallGenerator* cg = _late_inlines.at(i); 2070 _late_inlines_pos = i+1; 2071 cg->do_late_inline(); 2072 if (failing()) return false; 2073 } 2074 int j = 0; 2075 for (; i < _late_inlines.length(); i++, j++) { 2076 _late_inlines.at_put(j, _late_inlines.at(i)); 2077 } 2078 _late_inlines.trunc_to(j); 2079 assert(inlining_progress() || _late_inlines.length() == 0, ""); 2080 2081 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2082 2083 set_inlining_progress(false); 2084 set_do_cleanup(false); 2085 return (_late_inlines.length() > 0) && !needs_cleanup; 2086 } 2087 2088 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2089 { 2090 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2091 ResourceMark rm; 2092 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2093 } 2094 { 2095 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2096 igvn = PhaseIterGVN(initial_gvn()); 2097 igvn.optimize(); 2098 } 2099 } 2100 2101 // Perform incremental inlining until bound on number of live nodes is reached 2102 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2103 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2104 2105 set_inlining_incrementally(true); 2106 uint low_live_nodes = 0; 2107 2108 while (_late_inlines.length() > 0) { 2109 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2110 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2111 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2112 // PhaseIdealLoop is expensive so we only try it once we are 2113 // out of live nodes and we only try it again if the previous 2114 // helped got the number of nodes down significantly 2115 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2116 if (failing()) return; 2117 low_live_nodes = live_nodes(); 2118 _major_progress = true; 2119 } 2120 2121 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2122 break; // finish 2123 } 2124 } 2125 2126 for_igvn()->clear(); 2127 initial_gvn()->replace_with(&igvn); 2128 2129 while (inline_incrementally_one()) { 2130 assert(!failing(), "inconsistent"); 2131 } 2132 2133 if (failing()) return; 2134 2135 inline_incrementally_cleanup(igvn); 2136 2137 if (failing()) return; 2138 } 2139 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2140 2141 if (_string_late_inlines.length() > 0) { 2142 assert(has_stringbuilder(), "inconsistent"); 2143 for_igvn()->clear(); 2144 initial_gvn()->replace_with(&igvn); 2145 2146 inline_string_calls(false); 2147 2148 if (failing()) return; 2149 2150 inline_incrementally_cleanup(igvn); 2151 } 2152 2153 set_inlining_incrementally(false); 2154 } 2155 2156 2157 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2158 if(_loop_opts_cnt > 0) { 2159 debug_only( int cnt = 0; ); 2160 while(major_progress() && (_loop_opts_cnt > 0)) { 2161 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2162 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2163 PhaseIdealLoop::optimize(igvn, mode); 2164 _loop_opts_cnt--; 2165 if (failing()) return false; 2166 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2167 } 2168 } 2169 return true; 2170 } 2171 2172 // Remove edges from "root" to each SafePoint at a backward branch. 2173 // They were inserted during parsing (see add_safepoint()) to make 2174 // infinite loops without calls or exceptions visible to root, i.e., 2175 // useful. 2176 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2177 Node *r = root(); 2178 if (r != NULL) { 2179 for (uint i = r->req(); i < r->len(); ++i) { 2180 Node *n = r->in(i); 2181 if (n != NULL && n->is_SafePoint()) { 2182 r->rm_prec(i); 2183 if (n->outcnt() == 0) { 2184 igvn.remove_dead_node(n); 2185 } 2186 --i; 2187 } 2188 } 2189 } 2190 } 2191 2192 //------------------------------Optimize--------------------------------------- 2193 // Given a graph, optimize it. 2194 void Compile::Optimize() { 2195 TracePhase tp("optimizer", &timers[_t_optimizer]); 2196 2197 #ifndef PRODUCT 2198 if (_directive->BreakAtCompileOption) { 2199 BREAKPOINT; 2200 } 2201 2202 #endif 2203 2204 #ifdef ASSERT 2205 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2206 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2207 #endif 2208 2209 ResourceMark rm; 2210 2211 print_inlining_reinit(); 2212 2213 NOT_PRODUCT( verify_graph_edges(); ) 2214 2215 print_method(PHASE_AFTER_PARSING); 2216 2217 { 2218 // Iterative Global Value Numbering, including ideal transforms 2219 // Initialize IterGVN with types and values from parse-time GVN 2220 PhaseIterGVN igvn(initial_gvn()); 2221 #ifdef ASSERT 2222 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2223 #endif 2224 { 2225 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2226 igvn.optimize(); 2227 } 2228 2229 if (failing()) return; 2230 2231 print_method(PHASE_ITER_GVN1, 2); 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 // Now that all inlining is over, cut edge from root to loop 2253 // safepoints 2254 remove_root_to_sfpts_edges(igvn); 2255 2256 // Remove the speculative part of types and clean up the graph from 2257 // the extra CastPP nodes whose only purpose is to carry them. Do 2258 // that early so that optimizations are not disrupted by the extra 2259 // CastPP nodes. 2260 remove_speculative_types(igvn); 2261 2262 // No more new expensive nodes will be added to the list from here 2263 // so keep only the actual candidates for optimizations. 2264 cleanup_expensive_nodes(igvn); 2265 2266 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2267 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2268 initial_gvn()->replace_with(&igvn); 2269 for_igvn()->clear(); 2270 Unique_Node_List new_worklist(C->comp_arena()); 2271 { 2272 ResourceMark rm; 2273 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2274 } 2275 set_for_igvn(&new_worklist); 2276 igvn = PhaseIterGVN(initial_gvn()); 2277 igvn.optimize(); 2278 } 2279 2280 // Perform escape analysis 2281 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2282 if (has_loops()) { 2283 // Cleanup graph (remove dead nodes). 2284 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2285 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2286 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2287 if (failing()) return; 2288 } 2289 ConnectionGraph::do_analysis(this, &igvn); 2290 2291 if (failing()) return; 2292 2293 // Optimize out fields loads from scalar replaceable allocations. 2294 igvn.optimize(); 2295 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2296 2297 if (failing()) return; 2298 2299 if (congraph() != NULL && macro_count() > 0) { 2300 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2301 PhaseMacroExpand mexp(igvn); 2302 mexp.eliminate_macro_nodes(); 2303 igvn.set_delay_transform(false); 2304 2305 igvn.optimize(); 2306 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2307 2308 if (failing()) return; 2309 } 2310 } 2311 2312 // Loop transforms on the ideal graph. Range Check Elimination, 2313 // peeling, unrolling, etc. 2314 2315 // Set loop opts counter 2316 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2317 { 2318 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2319 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2320 _loop_opts_cnt--; 2321 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2322 if (failing()) return; 2323 } 2324 // Loop opts pass if partial peeling occurred in previous pass 2325 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2326 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2327 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2328 _loop_opts_cnt--; 2329 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2330 if (failing()) return; 2331 } 2332 // Loop opts pass for loop-unrolling before CCP 2333 if(major_progress() && (_loop_opts_cnt > 0)) { 2334 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2335 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2336 _loop_opts_cnt--; 2337 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2338 } 2339 if (!failing()) { 2340 // Verify that last round of loop opts produced a valid graph 2341 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2342 PhaseIdealLoop::verify(igvn); 2343 } 2344 } 2345 if (failing()) return; 2346 2347 // Conditional Constant Propagation; 2348 PhaseCCP ccp( &igvn ); 2349 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2350 { 2351 TracePhase tp("ccp", &timers[_t_ccp]); 2352 ccp.do_transform(); 2353 } 2354 print_method(PHASE_CPP1, 2); 2355 2356 assert( true, "Break here to ccp.dump_old2new_map()"); 2357 2358 // Iterative Global Value Numbering, including ideal transforms 2359 { 2360 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2361 igvn = ccp; 2362 igvn.optimize(); 2363 } 2364 2365 print_method(PHASE_ITER_GVN2, 2); 2366 2367 if (failing()) return; 2368 2369 // Loop transforms on the ideal graph. Range Check Elimination, 2370 // peeling, unrolling, etc. 2371 if (!optimize_loops(igvn, LoopOptsDefault)) { 2372 return; 2373 } 2374 2375 #if INCLUDE_ZGC 2376 if (UseZGC) { 2377 ZBarrierSetC2::find_dominating_barriers(igvn); 2378 } 2379 #endif 2380 2381 if (failing()) return; 2382 2383 // Ensure that major progress is now clear 2384 C->clear_major_progress(); 2385 2386 { 2387 // Verify that all previous optimizations produced a valid graph 2388 // at least to this point, even if no loop optimizations were done. 2389 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2390 PhaseIdealLoop::verify(igvn); 2391 } 2392 2393 if (range_check_cast_count() > 0) { 2394 // No more loop optimizations. Remove all range check dependent CastIINodes. 2395 C->remove_range_check_casts(igvn); 2396 igvn.optimize(); 2397 } 2398 2399 #ifdef ASSERT 2400 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2401 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand); 2402 #endif 2403 2404 { 2405 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2406 PhaseMacroExpand mex(igvn); 2407 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2408 if (mex.expand_macro_nodes()) { 2409 assert(failing(), "must bail out w/ explicit message"); 2410 return; 2411 } 2412 } 2413 2414 { 2415 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2416 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2); 2417 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2418 if (bs->expand_barriers(this, igvn)) { 2419 assert(failing(), "must bail out w/ explicit message"); 2420 return; 2421 } 2422 } 2423 2424 if (opaque4_count() > 0) { 2425 C->remove_opaque4_nodes(igvn); 2426 igvn.optimize(); 2427 } 2428 2429 DEBUG_ONLY( _modified_nodes = NULL; ) 2430 } // (End scope of igvn; run destructor if necessary for asserts.) 2431 2432 process_print_inlining(); 2433 // A method with only infinite loops has no edges entering loops from root 2434 { 2435 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2436 if (final_graph_reshaping()) { 2437 assert(failing(), "must bail out w/ explicit message"); 2438 return; 2439 } 2440 } 2441 2442 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2443 } 2444 2445 2446 //------------------------------Code_Gen--------------------------------------- 2447 // Given a graph, generate code for it 2448 void Compile::Code_Gen() { 2449 if (failing()) { 2450 return; 2451 } 2452 2453 // Perform instruction selection. You might think we could reclaim Matcher 2454 // memory PDQ, but actually the Matcher is used in generating spill code. 2455 // Internals of the Matcher (including some VectorSets) must remain live 2456 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2457 // set a bit in reclaimed memory. 2458 2459 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2460 // nodes. Mapping is only valid at the root of each matched subtree. 2461 NOT_PRODUCT( verify_graph_edges(); ) 2462 2463 Matcher matcher; 2464 _matcher = &matcher; 2465 { 2466 TracePhase tp("matcher", &timers[_t_matcher]); 2467 matcher.match(); 2468 } 2469 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2470 // nodes. Mapping is only valid at the root of each matched subtree. 2471 NOT_PRODUCT( verify_graph_edges(); ) 2472 2473 // If you have too many nodes, or if matching has failed, bail out 2474 check_node_count(0, "out of nodes matching instructions"); 2475 if (failing()) { 2476 return; 2477 } 2478 2479 print_method(PHASE_MATCHING, 2); 2480 2481 // Build a proper-looking CFG 2482 PhaseCFG cfg(node_arena(), root(), matcher); 2483 _cfg = &cfg; 2484 { 2485 TracePhase tp("scheduler", &timers[_t_scheduler]); 2486 bool success = cfg.do_global_code_motion(); 2487 if (!success) { 2488 return; 2489 } 2490 2491 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2492 NOT_PRODUCT( verify_graph_edges(); ) 2493 debug_only( cfg.verify(); ) 2494 } 2495 2496 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2497 _regalloc = ®alloc; 2498 { 2499 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2500 // Perform register allocation. After Chaitin, use-def chains are 2501 // no longer accurate (at spill code) and so must be ignored. 2502 // Node->LRG->reg mappings are still accurate. 2503 _regalloc->Register_Allocate(); 2504 2505 // Bail out if the allocator builds too many nodes 2506 if (failing()) { 2507 return; 2508 } 2509 } 2510 2511 // Prior to register allocation we kept empty basic blocks in case the 2512 // the allocator needed a place to spill. After register allocation we 2513 // are not adding any new instructions. If any basic block is empty, we 2514 // can now safely remove it. 2515 { 2516 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2517 cfg.remove_empty_blocks(); 2518 if (do_freq_based_layout()) { 2519 PhaseBlockLayout layout(cfg); 2520 } else { 2521 cfg.set_loop_alignment(); 2522 } 2523 cfg.fixup_flow(); 2524 } 2525 2526 // Apply peephole optimizations 2527 if( OptoPeephole ) { 2528 TracePhase tp("peephole", &timers[_t_peephole]); 2529 PhasePeephole peep( _regalloc, cfg); 2530 peep.do_transform(); 2531 } 2532 2533 // Do late expand if CPU requires this. 2534 if (Matcher::require_postalloc_expand) { 2535 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2536 cfg.postalloc_expand(_regalloc); 2537 } 2538 2539 // Convert Nodes to instruction bits in a buffer 2540 { 2541 TraceTime tp("output", &timers[_t_output], CITime); 2542 Output(); 2543 } 2544 2545 print_method(PHASE_FINAL_CODE); 2546 2547 // He's dead, Jim. 2548 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2549 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2550 } 2551 2552 2553 //------------------------------dump_asm--------------------------------------- 2554 // Dump formatted assembly 2555 #ifndef PRODUCT 2556 void Compile::dump_asm(int *pcs, uint pc_limit) { 2557 bool cut_short = false; 2558 tty->print_cr("#"); 2559 tty->print("# "); _tf->dump(); tty->cr(); 2560 tty->print_cr("#"); 2561 2562 // For all blocks 2563 int pc = 0x0; // Program counter 2564 char starts_bundle = ' '; 2565 _regalloc->dump_frame(); 2566 2567 Node *n = NULL; 2568 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2569 if (VMThread::should_terminate()) { 2570 cut_short = true; 2571 break; 2572 } 2573 Block* block = _cfg->get_block(i); 2574 if (block->is_connector() && !Verbose) { 2575 continue; 2576 } 2577 n = block->head(); 2578 if (pcs && n->_idx < pc_limit) { 2579 tty->print("%3.3x ", pcs[n->_idx]); 2580 } else { 2581 tty->print(" "); 2582 } 2583 block->dump_head(_cfg); 2584 if (block->is_connector()) { 2585 tty->print_cr(" # Empty connector block"); 2586 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2587 tty->print_cr(" # Block is sole successor of call"); 2588 } 2589 2590 // For all instructions 2591 Node *delay = NULL; 2592 for (uint j = 0; j < block->number_of_nodes(); j++) { 2593 if (VMThread::should_terminate()) { 2594 cut_short = true; 2595 break; 2596 } 2597 n = block->get_node(j); 2598 if (valid_bundle_info(n)) { 2599 Bundle* bundle = node_bundling(n); 2600 if (bundle->used_in_unconditional_delay()) { 2601 delay = n; 2602 continue; 2603 } 2604 if (bundle->starts_bundle()) { 2605 starts_bundle = '+'; 2606 } 2607 } 2608 2609 if (WizardMode) { 2610 n->dump(); 2611 } 2612 2613 if( !n->is_Region() && // Dont print in the Assembly 2614 !n->is_Phi() && // a few noisely useless nodes 2615 !n->is_Proj() && 2616 !n->is_MachTemp() && 2617 !n->is_SafePointScalarObject() && 2618 !n->is_Catch() && // Would be nice to print exception table targets 2619 !n->is_MergeMem() && // Not very interesting 2620 !n->is_top() && // Debug info table constants 2621 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2622 ) { 2623 if (pcs && n->_idx < pc_limit) 2624 tty->print("%3.3x", pcs[n->_idx]); 2625 else 2626 tty->print(" "); 2627 tty->print(" %c ", starts_bundle); 2628 starts_bundle = ' '; 2629 tty->print("\t"); 2630 n->format(_regalloc, tty); 2631 tty->cr(); 2632 } 2633 2634 // If we have an instruction with a delay slot, and have seen a delay, 2635 // then back up and print it 2636 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2637 assert(delay != NULL, "no unconditional delay instruction"); 2638 if (WizardMode) delay->dump(); 2639 2640 if (node_bundling(delay)->starts_bundle()) 2641 starts_bundle = '+'; 2642 if (pcs && n->_idx < pc_limit) 2643 tty->print("%3.3x", pcs[n->_idx]); 2644 else 2645 tty->print(" "); 2646 tty->print(" %c ", starts_bundle); 2647 starts_bundle = ' '; 2648 tty->print("\t"); 2649 delay->format(_regalloc, tty); 2650 tty->cr(); 2651 delay = NULL; 2652 } 2653 2654 // Dump the exception table as well 2655 if( n->is_Catch() && (Verbose || WizardMode) ) { 2656 // Print the exception table for this offset 2657 _handler_table.print_subtable_for(pc); 2658 } 2659 } 2660 2661 if (pcs && n->_idx < pc_limit) 2662 tty->print_cr("%3.3x", pcs[n->_idx]); 2663 else 2664 tty->cr(); 2665 2666 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2667 2668 } // End of per-block dump 2669 tty->cr(); 2670 2671 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2672 } 2673 #endif 2674 2675 //------------------------------Final_Reshape_Counts--------------------------- 2676 // This class defines counters to help identify when a method 2677 // may/must be executed using hardware with only 24-bit precision. 2678 struct Final_Reshape_Counts : public StackObj { 2679 int _call_count; // count non-inlined 'common' calls 2680 int _float_count; // count float ops requiring 24-bit precision 2681 int _double_count; // count double ops requiring more precision 2682 int _java_call_count; // count non-inlined 'java' calls 2683 int _inner_loop_count; // count loops which need alignment 2684 VectorSet _visited; // Visitation flags 2685 Node_List _tests; // Set of IfNodes & PCTableNodes 2686 2687 Final_Reshape_Counts() : 2688 _call_count(0), _float_count(0), _double_count(0), 2689 _java_call_count(0), _inner_loop_count(0), 2690 _visited( Thread::current()->resource_area() ) { } 2691 2692 void inc_call_count () { _call_count ++; } 2693 void inc_float_count () { _float_count ++; } 2694 void inc_double_count() { _double_count++; } 2695 void inc_java_call_count() { _java_call_count++; } 2696 void inc_inner_loop_count() { _inner_loop_count++; } 2697 2698 int get_call_count () const { return _call_count ; } 2699 int get_float_count () const { return _float_count ; } 2700 int get_double_count() const { return _double_count; } 2701 int get_java_call_count() const { return _java_call_count; } 2702 int get_inner_loop_count() const { return _inner_loop_count; } 2703 }; 2704 2705 #ifdef ASSERT 2706 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2707 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2708 // Make sure the offset goes inside the instance layout. 2709 return k->contains_field_offset(tp->offset()); 2710 // Note that OffsetBot and OffsetTop are very negative. 2711 } 2712 #endif 2713 2714 // Eliminate trivially redundant StoreCMs and accumulate their 2715 // precedence edges. 2716 void Compile::eliminate_redundant_card_marks(Node* n) { 2717 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2718 if (n->in(MemNode::Address)->outcnt() > 1) { 2719 // There are multiple users of the same address so it might be 2720 // possible to eliminate some of the StoreCMs 2721 Node* mem = n->in(MemNode::Memory); 2722 Node* adr = n->in(MemNode::Address); 2723 Node* val = n->in(MemNode::ValueIn); 2724 Node* prev = n; 2725 bool done = false; 2726 // Walk the chain of StoreCMs eliminating ones that match. As 2727 // long as it's a chain of single users then the optimization is 2728 // safe. Eliminating partially redundant StoreCMs would require 2729 // cloning copies down the other paths. 2730 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2731 if (adr == mem->in(MemNode::Address) && 2732 val == mem->in(MemNode::ValueIn)) { 2733 // redundant StoreCM 2734 if (mem->req() > MemNode::OopStore) { 2735 // Hasn't been processed by this code yet. 2736 n->add_prec(mem->in(MemNode::OopStore)); 2737 } else { 2738 // Already converted to precedence edge 2739 for (uint i = mem->req(); i < mem->len(); i++) { 2740 // Accumulate any precedence edges 2741 if (mem->in(i) != NULL) { 2742 n->add_prec(mem->in(i)); 2743 } 2744 } 2745 // Everything above this point has been processed. 2746 done = true; 2747 } 2748 // Eliminate the previous StoreCM 2749 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2750 assert(mem->outcnt() == 0, "should be dead"); 2751 mem->disconnect_inputs(NULL, this); 2752 } else { 2753 prev = mem; 2754 } 2755 mem = prev->in(MemNode::Memory); 2756 } 2757 } 2758 } 2759 2760 //------------------------------final_graph_reshaping_impl---------------------- 2761 // Implement items 1-5 from final_graph_reshaping below. 2762 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2763 2764 if ( n->outcnt() == 0 ) return; // dead node 2765 uint nop = n->Opcode(); 2766 2767 // Check for 2-input instruction with "last use" on right input. 2768 // Swap to left input. Implements item (2). 2769 if( n->req() == 3 && // two-input instruction 2770 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2771 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2772 n->in(2)->outcnt() == 1 &&// right use IS a last use 2773 !n->in(2)->is_Con() ) { // right use is not a constant 2774 // Check for commutative opcode 2775 switch( nop ) { 2776 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2777 case Op_MaxI: case Op_MinI: 2778 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2779 case Op_AndL: case Op_XorL: case Op_OrL: 2780 case Op_AndI: case Op_XorI: case Op_OrI: { 2781 // Move "last use" input to left by swapping inputs 2782 n->swap_edges(1, 2); 2783 break; 2784 } 2785 default: 2786 break; 2787 } 2788 } 2789 2790 #ifdef ASSERT 2791 if( n->is_Mem() ) { 2792 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2793 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2794 // oop will be recorded in oop map if load crosses safepoint 2795 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2796 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2797 "raw memory operations should have control edge"); 2798 } 2799 if (n->is_MemBar()) { 2800 MemBarNode* mb = n->as_MemBar(); 2801 if (mb->trailing_store() || mb->trailing_load_store()) { 2802 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2803 Node* mem = mb->in(MemBarNode::Precedent); 2804 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2805 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2806 } else if (mb->leading()) { 2807 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2808 } 2809 } 2810 #endif 2811 // Count FPU ops and common calls, implements item (3) 2812 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2813 if (!gc_handled) { 2814 final_graph_reshaping_main_switch(n, frc, nop); 2815 } 2816 2817 // Collect CFG split points 2818 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2819 frc._tests.push(n); 2820 } 2821 } 2822 2823 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2824 switch( nop ) { 2825 // Count all float operations that may use FPU 2826 case Op_AddF: 2827 case Op_SubF: 2828 case Op_MulF: 2829 case Op_DivF: 2830 case Op_NegF: 2831 case Op_ModF: 2832 case Op_ConvI2F: 2833 case Op_ConF: 2834 case Op_CmpF: 2835 case Op_CmpF3: 2836 // case Op_ConvL2F: // longs are split into 32-bit halves 2837 frc.inc_float_count(); 2838 break; 2839 2840 case Op_ConvF2D: 2841 case Op_ConvD2F: 2842 frc.inc_float_count(); 2843 frc.inc_double_count(); 2844 break; 2845 2846 // Count all double operations that may use FPU 2847 case Op_AddD: 2848 case Op_SubD: 2849 case Op_MulD: 2850 case Op_DivD: 2851 case Op_NegD: 2852 case Op_ModD: 2853 case Op_ConvI2D: 2854 case Op_ConvD2I: 2855 // case Op_ConvL2D: // handled by leaf call 2856 // case Op_ConvD2L: // handled by leaf call 2857 case Op_ConD: 2858 case Op_CmpD: 2859 case Op_CmpD3: 2860 frc.inc_double_count(); 2861 break; 2862 case Op_Opaque1: // Remove Opaque Nodes before matching 2863 case Op_Opaque2: // Remove Opaque Nodes before matching 2864 case Op_Opaque3: 2865 n->subsume_by(n->in(1), this); 2866 break; 2867 case Op_CallStaticJava: 2868 case Op_CallJava: 2869 case Op_CallDynamicJava: 2870 frc.inc_java_call_count(); // Count java call site; 2871 case Op_CallRuntime: 2872 case Op_CallLeaf: 2873 case Op_CallLeafNoFP: { 2874 assert (n->is_Call(), ""); 2875 CallNode *call = n->as_Call(); 2876 // Count call sites where the FP mode bit would have to be flipped. 2877 // Do not count uncommon runtime calls: 2878 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2879 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2880 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2881 frc.inc_call_count(); // Count the call site 2882 } else { // See if uncommon argument is shared 2883 Node *n = call->in(TypeFunc::Parms); 2884 int nop = n->Opcode(); 2885 // Clone shared simple arguments to uncommon calls, item (1). 2886 if (n->outcnt() > 1 && 2887 !n->is_Proj() && 2888 nop != Op_CreateEx && 2889 nop != Op_CheckCastPP && 2890 nop != Op_DecodeN && 2891 nop != Op_DecodeNKlass && 2892 !n->is_Mem() && 2893 !n->is_Phi()) { 2894 Node *x = n->clone(); 2895 call->set_req(TypeFunc::Parms, x); 2896 } 2897 } 2898 break; 2899 } 2900 2901 case Op_StoreD: 2902 case Op_LoadD: 2903 case Op_LoadD_unaligned: 2904 frc.inc_double_count(); 2905 goto handle_mem; 2906 case Op_StoreF: 2907 case Op_LoadF: 2908 frc.inc_float_count(); 2909 goto handle_mem; 2910 2911 case Op_StoreCM: 2912 { 2913 // Convert OopStore dependence into precedence edge 2914 Node* prec = n->in(MemNode::OopStore); 2915 n->del_req(MemNode::OopStore); 2916 n->add_prec(prec); 2917 eliminate_redundant_card_marks(n); 2918 } 2919 2920 // fall through 2921 2922 case Op_StoreB: 2923 case Op_StoreC: 2924 case Op_StorePConditional: 2925 case Op_StoreI: 2926 case Op_StoreL: 2927 case Op_StoreIConditional: 2928 case Op_StoreLConditional: 2929 case Op_CompareAndSwapB: 2930 case Op_CompareAndSwapS: 2931 case Op_CompareAndSwapI: 2932 case Op_CompareAndSwapL: 2933 case Op_CompareAndSwapP: 2934 case Op_CompareAndSwapN: 2935 case Op_WeakCompareAndSwapB: 2936 case Op_WeakCompareAndSwapS: 2937 case Op_WeakCompareAndSwapI: 2938 case Op_WeakCompareAndSwapL: 2939 case Op_WeakCompareAndSwapP: 2940 case Op_WeakCompareAndSwapN: 2941 case Op_CompareAndExchangeB: 2942 case Op_CompareAndExchangeS: 2943 case Op_CompareAndExchangeI: 2944 case Op_CompareAndExchangeL: 2945 case Op_CompareAndExchangeP: 2946 case Op_CompareAndExchangeN: 2947 case Op_GetAndAddS: 2948 case Op_GetAndAddB: 2949 case Op_GetAndAddI: 2950 case Op_GetAndAddL: 2951 case Op_GetAndSetS: 2952 case Op_GetAndSetB: 2953 case Op_GetAndSetI: 2954 case Op_GetAndSetL: 2955 case Op_GetAndSetP: 2956 case Op_GetAndSetN: 2957 case Op_StoreP: 2958 case Op_StoreN: 2959 case Op_StoreNKlass: 2960 case Op_LoadB: 2961 case Op_LoadUB: 2962 case Op_LoadUS: 2963 case Op_LoadI: 2964 case Op_LoadKlass: 2965 case Op_LoadNKlass: 2966 case Op_LoadL: 2967 case Op_LoadL_unaligned: 2968 case Op_LoadPLocked: 2969 case Op_LoadP: 2970 case Op_LoadN: 2971 case Op_LoadRange: 2972 case Op_LoadS: { 2973 handle_mem: 2974 #ifdef ASSERT 2975 if( VerifyOptoOopOffsets ) { 2976 MemNode* mem = n->as_Mem(); 2977 // Check to see if address types have grounded out somehow. 2978 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2979 assert( !tp || oop_offset_is_sane(tp), "" ); 2980 } 2981 #endif 2982 break; 2983 } 2984 2985 case Op_AddP: { // Assert sane base pointers 2986 Node *addp = n->in(AddPNode::Address); 2987 assert( !addp->is_AddP() || 2988 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2989 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2990 "Base pointers must match (addp %u)", addp->_idx ); 2991 #ifdef _LP64 2992 if ((UseCompressedOops || UseCompressedClassPointers) && 2993 addp->Opcode() == Op_ConP && 2994 addp == n->in(AddPNode::Base) && 2995 n->in(AddPNode::Offset)->is_Con()) { 2996 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2997 // on the platform and on the compressed oops mode. 2998 // Use addressing with narrow klass to load with offset on x86. 2999 // Some platforms can use the constant pool to load ConP. 3000 // Do this transformation here since IGVN will convert ConN back to ConP. 3001 const Type* t = addp->bottom_type(); 3002 bool is_oop = t->isa_oopptr() != NULL; 3003 bool is_klass = t->isa_klassptr() != NULL; 3004 3005 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3006 (is_klass && Matcher::const_klass_prefer_decode())) { 3007 Node* nn = NULL; 3008 3009 int op = is_oop ? Op_ConN : Op_ConNKlass; 3010 3011 // Look for existing ConN node of the same exact type. 3012 Node* r = root(); 3013 uint cnt = r->outcnt(); 3014 for (uint i = 0; i < cnt; i++) { 3015 Node* m = r->raw_out(i); 3016 if (m!= NULL && m->Opcode() == op && 3017 m->bottom_type()->make_ptr() == t) { 3018 nn = m; 3019 break; 3020 } 3021 } 3022 if (nn != NULL) { 3023 // Decode a narrow oop to match address 3024 // [R12 + narrow_oop_reg<<3 + offset] 3025 if (is_oop) { 3026 nn = new DecodeNNode(nn, t); 3027 } else { 3028 nn = new DecodeNKlassNode(nn, t); 3029 } 3030 // Check for succeeding AddP which uses the same Base. 3031 // Otherwise we will run into the assertion above when visiting that guy. 3032 for (uint i = 0; i < n->outcnt(); ++i) { 3033 Node *out_i = n->raw_out(i); 3034 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3035 out_i->set_req(AddPNode::Base, nn); 3036 #ifdef ASSERT 3037 for (uint j = 0; j < out_i->outcnt(); ++j) { 3038 Node *out_j = out_i->raw_out(j); 3039 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3040 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3041 } 3042 #endif 3043 } 3044 } 3045 n->set_req(AddPNode::Base, nn); 3046 n->set_req(AddPNode::Address, nn); 3047 if (addp->outcnt() == 0) { 3048 addp->disconnect_inputs(NULL, this); 3049 } 3050 } 3051 } 3052 } 3053 #endif 3054 // platform dependent reshaping of the address expression 3055 reshape_address(n->as_AddP()); 3056 break; 3057 } 3058 3059 case Op_CastPP: { 3060 // Remove CastPP nodes to gain more freedom during scheduling but 3061 // keep the dependency they encode as control or precedence edges 3062 // (if control is set already) on memory operations. Some CastPP 3063 // nodes don't have a control (don't carry a dependency): skip 3064 // those. 3065 if (n->in(0) != NULL) { 3066 ResourceMark rm; 3067 Unique_Node_List wq; 3068 wq.push(n); 3069 for (uint next = 0; next < wq.size(); ++next) { 3070 Node *m = wq.at(next); 3071 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3072 Node* use = m->fast_out(i); 3073 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) { 3074 use->ensure_control_or_add_prec(n->in(0)); 3075 } else { 3076 switch(use->Opcode()) { 3077 case Op_AddP: 3078 case Op_DecodeN: 3079 case Op_DecodeNKlass: 3080 case Op_CheckCastPP: 3081 case Op_CastPP: 3082 wq.push(use); 3083 break; 3084 } 3085 } 3086 } 3087 } 3088 } 3089 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3090 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3091 Node* in1 = n->in(1); 3092 const Type* t = n->bottom_type(); 3093 Node* new_in1 = in1->clone(); 3094 new_in1->as_DecodeN()->set_type(t); 3095 3096 if (!Matcher::narrow_oop_use_complex_address()) { 3097 // 3098 // x86, ARM and friends can handle 2 adds in addressing mode 3099 // and Matcher can fold a DecodeN node into address by using 3100 // a narrow oop directly and do implicit NULL check in address: 3101 // 3102 // [R12 + narrow_oop_reg<<3 + offset] 3103 // NullCheck narrow_oop_reg 3104 // 3105 // On other platforms (Sparc) we have to keep new DecodeN node and 3106 // use it to do implicit NULL check in address: 3107 // 3108 // decode_not_null narrow_oop_reg, base_reg 3109 // [base_reg + offset] 3110 // NullCheck base_reg 3111 // 3112 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3113 // to keep the information to which NULL check the new DecodeN node 3114 // corresponds to use it as value in implicit_null_check(). 3115 // 3116 new_in1->set_req(0, n->in(0)); 3117 } 3118 3119 n->subsume_by(new_in1, this); 3120 if (in1->outcnt() == 0) { 3121 in1->disconnect_inputs(NULL, this); 3122 } 3123 } else { 3124 n->subsume_by(n->in(1), this); 3125 if (n->outcnt() == 0) { 3126 n->disconnect_inputs(NULL, this); 3127 } 3128 } 3129 break; 3130 } 3131 #ifdef _LP64 3132 case Op_CmpP: 3133 // Do this transformation here to preserve CmpPNode::sub() and 3134 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3135 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3136 Node* in1 = n->in(1); 3137 Node* in2 = n->in(2); 3138 if (!in1->is_DecodeNarrowPtr()) { 3139 in2 = in1; 3140 in1 = n->in(2); 3141 } 3142 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3143 3144 Node* new_in2 = NULL; 3145 if (in2->is_DecodeNarrowPtr()) { 3146 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3147 new_in2 = in2->in(1); 3148 } else if (in2->Opcode() == Op_ConP) { 3149 const Type* t = in2->bottom_type(); 3150 if (t == TypePtr::NULL_PTR) { 3151 assert(in1->is_DecodeN(), "compare klass to null?"); 3152 // Don't convert CmpP null check into CmpN if compressed 3153 // oops implicit null check is not generated. 3154 // This will allow to generate normal oop implicit null check. 3155 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3156 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3157 // 3158 // This transformation together with CastPP transformation above 3159 // will generated code for implicit NULL checks for compressed oops. 3160 // 3161 // The original code after Optimize() 3162 // 3163 // LoadN memory, narrow_oop_reg 3164 // decode narrow_oop_reg, base_reg 3165 // CmpP base_reg, NULL 3166 // CastPP base_reg // NotNull 3167 // Load [base_reg + offset], val_reg 3168 // 3169 // after these transformations will be 3170 // 3171 // LoadN memory, narrow_oop_reg 3172 // CmpN narrow_oop_reg, NULL 3173 // decode_not_null narrow_oop_reg, base_reg 3174 // Load [base_reg + offset], val_reg 3175 // 3176 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3177 // since narrow oops can be used in debug info now (see the code in 3178 // final_graph_reshaping_walk()). 3179 // 3180 // At the end the code will be matched to 3181 // on x86: 3182 // 3183 // Load_narrow_oop memory, narrow_oop_reg 3184 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3185 // NullCheck narrow_oop_reg 3186 // 3187 // and on sparc: 3188 // 3189 // Load_narrow_oop memory, narrow_oop_reg 3190 // decode_not_null narrow_oop_reg, base_reg 3191 // Load [base_reg + offset], val_reg 3192 // NullCheck base_reg 3193 // 3194 } else if (t->isa_oopptr()) { 3195 new_in2 = ConNode::make(t->make_narrowoop()); 3196 } else if (t->isa_klassptr()) { 3197 new_in2 = ConNode::make(t->make_narrowklass()); 3198 } 3199 } 3200 if (new_in2 != NULL) { 3201 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3202 n->subsume_by(cmpN, this); 3203 if (in1->outcnt() == 0) { 3204 in1->disconnect_inputs(NULL, this); 3205 } 3206 if (in2->outcnt() == 0) { 3207 in2->disconnect_inputs(NULL, this); 3208 } 3209 } 3210 } 3211 break; 3212 3213 case Op_DecodeN: 3214 case Op_DecodeNKlass: 3215 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3216 // DecodeN could be pinned when it can't be fold into 3217 // an address expression, see the code for Op_CastPP above. 3218 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3219 break; 3220 3221 case Op_EncodeP: 3222 case Op_EncodePKlass: { 3223 Node* in1 = n->in(1); 3224 if (in1->is_DecodeNarrowPtr()) { 3225 n->subsume_by(in1->in(1), this); 3226 } else if (in1->Opcode() == Op_ConP) { 3227 const Type* t = in1->bottom_type(); 3228 if (t == TypePtr::NULL_PTR) { 3229 assert(t->isa_oopptr(), "null klass?"); 3230 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3231 } else if (t->isa_oopptr()) { 3232 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3233 } else if (t->isa_klassptr()) { 3234 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3235 } 3236 } 3237 if (in1->outcnt() == 0) { 3238 in1->disconnect_inputs(NULL, this); 3239 } 3240 break; 3241 } 3242 3243 case Op_Proj: { 3244 if (OptimizeStringConcat) { 3245 ProjNode* p = n->as_Proj(); 3246 if (p->_is_io_use) { 3247 // Separate projections were used for the exception path which 3248 // are normally removed by a late inline. If it wasn't inlined 3249 // then they will hang around and should just be replaced with 3250 // the original one. 3251 Node* proj = NULL; 3252 // Replace with just one 3253 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3254 Node *use = i.get(); 3255 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3256 proj = use; 3257 break; 3258 } 3259 } 3260 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3261 if (proj != NULL) { 3262 p->subsume_by(proj, this); 3263 } 3264 } 3265 } 3266 break; 3267 } 3268 3269 case Op_Phi: 3270 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3271 // The EncodeP optimization may create Phi with the same edges 3272 // for all paths. It is not handled well by Register Allocator. 3273 Node* unique_in = n->in(1); 3274 assert(unique_in != NULL, ""); 3275 uint cnt = n->req(); 3276 for (uint i = 2; i < cnt; i++) { 3277 Node* m = n->in(i); 3278 assert(m != NULL, ""); 3279 if (unique_in != m) 3280 unique_in = NULL; 3281 } 3282 if (unique_in != NULL) { 3283 n->subsume_by(unique_in, this); 3284 } 3285 } 3286 break; 3287 3288 #endif 3289 3290 #ifdef ASSERT 3291 case Op_CastII: 3292 // Verify that all range check dependent CastII nodes were removed. 3293 if (n->isa_CastII()->has_range_check()) { 3294 n->dump(3); 3295 assert(false, "Range check dependent CastII node was not removed"); 3296 } 3297 break; 3298 #endif 3299 3300 case Op_ModI: 3301 if (UseDivMod) { 3302 // Check if a%b and a/b both exist 3303 Node* d = n->find_similar(Op_DivI); 3304 if (d) { 3305 // Replace them with a fused divmod if supported 3306 if (Matcher::has_match_rule(Op_DivModI)) { 3307 DivModINode* divmod = DivModINode::make(n); 3308 d->subsume_by(divmod->div_proj(), this); 3309 n->subsume_by(divmod->mod_proj(), this); 3310 } else { 3311 // replace a%b with a-((a/b)*b) 3312 Node* mult = new MulINode(d, d->in(2)); 3313 Node* sub = new SubINode(d->in(1), mult); 3314 n->subsume_by(sub, this); 3315 } 3316 } 3317 } 3318 break; 3319 3320 case Op_ModL: 3321 if (UseDivMod) { 3322 // Check if a%b and a/b both exist 3323 Node* d = n->find_similar(Op_DivL); 3324 if (d) { 3325 // Replace them with a fused divmod if supported 3326 if (Matcher::has_match_rule(Op_DivModL)) { 3327 DivModLNode* divmod = DivModLNode::make(n); 3328 d->subsume_by(divmod->div_proj(), this); 3329 n->subsume_by(divmod->mod_proj(), this); 3330 } else { 3331 // replace a%b with a-((a/b)*b) 3332 Node* mult = new MulLNode(d, d->in(2)); 3333 Node* sub = new SubLNode(d->in(1), mult); 3334 n->subsume_by(sub, this); 3335 } 3336 } 3337 } 3338 break; 3339 3340 case Op_LoadVector: 3341 case Op_StoreVector: 3342 break; 3343 3344 case Op_AddReductionVI: 3345 case Op_AddReductionVL: 3346 case Op_AddReductionVF: 3347 case Op_AddReductionVD: 3348 case Op_MulReductionVI: 3349 case Op_MulReductionVL: 3350 case Op_MulReductionVF: 3351 case Op_MulReductionVD: 3352 break; 3353 3354 case Op_PackB: 3355 case Op_PackS: 3356 case Op_PackI: 3357 case Op_PackF: 3358 case Op_PackL: 3359 case Op_PackD: 3360 if (n->req()-1 > 2) { 3361 // Replace many operand PackNodes with a binary tree for matching 3362 PackNode* p = (PackNode*) n; 3363 Node* btp = p->binary_tree_pack(1, n->req()); 3364 n->subsume_by(btp, this); 3365 } 3366 break; 3367 case Op_Loop: 3368 case Op_CountedLoop: 3369 case Op_OuterStripMinedLoop: 3370 if (n->as_Loop()->is_inner_loop()) { 3371 frc.inc_inner_loop_count(); 3372 } 3373 n->as_Loop()->verify_strip_mined(0); 3374 break; 3375 case Op_LShiftI: 3376 case Op_RShiftI: 3377 case Op_URShiftI: 3378 case Op_LShiftL: 3379 case Op_RShiftL: 3380 case Op_URShiftL: 3381 if (Matcher::need_masked_shift_count) { 3382 // The cpu's shift instructions don't restrict the count to the 3383 // lower 5/6 bits. We need to do the masking ourselves. 3384 Node* in2 = n->in(2); 3385 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3386 const TypeInt* t = in2->find_int_type(); 3387 if (t != NULL && t->is_con()) { 3388 juint shift = t->get_con(); 3389 if (shift > mask) { // Unsigned cmp 3390 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3391 } 3392 } else { 3393 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3394 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3395 n->set_req(2, shift); 3396 } 3397 } 3398 if (in2->outcnt() == 0) { // Remove dead node 3399 in2->disconnect_inputs(NULL, this); 3400 } 3401 } 3402 break; 3403 case Op_MemBarStoreStore: 3404 case Op_MemBarRelease: 3405 // Break the link with AllocateNode: it is no longer useful and 3406 // confuses register allocation. 3407 if (n->req() > MemBarNode::Precedent) { 3408 n->set_req(MemBarNode::Precedent, top()); 3409 } 3410 break; 3411 case Op_MemBarAcquire: { 3412 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3413 // At parse time, the trailing MemBarAcquire for a volatile load 3414 // is created with an edge to the load. After optimizations, 3415 // that input may be a chain of Phis. If those phis have no 3416 // other use, then the MemBarAcquire keeps them alive and 3417 // register allocation can be confused. 3418 ResourceMark rm; 3419 Unique_Node_List wq; 3420 wq.push(n->in(MemBarNode::Precedent)); 3421 n->set_req(MemBarNode::Precedent, top()); 3422 while (wq.size() > 0) { 3423 Node* m = wq.pop(); 3424 if (m->outcnt() == 0) { 3425 for (uint j = 0; j < m->req(); j++) { 3426 Node* in = m->in(j); 3427 if (in != NULL) { 3428 wq.push(in); 3429 } 3430 } 3431 m->disconnect_inputs(NULL, this); 3432 } 3433 } 3434 } 3435 break; 3436 } 3437 case Op_RangeCheck: { 3438 RangeCheckNode* rc = n->as_RangeCheck(); 3439 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3440 n->subsume_by(iff, this); 3441 frc._tests.push(iff); 3442 break; 3443 } 3444 case Op_ConvI2L: { 3445 if (!Matcher::convi2l_type_required) { 3446 // Code generation on some platforms doesn't need accurate 3447 // ConvI2L types. Widening the type can help remove redundant 3448 // address computations. 3449 n->as_Type()->set_type(TypeLong::INT); 3450 ResourceMark rm; 3451 Node_List wq; 3452 wq.push(n); 3453 for (uint next = 0; next < wq.size(); next++) { 3454 Node *m = wq.at(next); 3455 3456 for(;;) { 3457 // Loop over all nodes with identical inputs edges as m 3458 Node* k = m->find_similar(m->Opcode()); 3459 if (k == NULL) { 3460 break; 3461 } 3462 // Push their uses so we get a chance to remove node made 3463 // redundant 3464 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3465 Node* u = k->fast_out(i); 3466 assert(!wq.contains(u), "shouldn't process one node several times"); 3467 if (u->Opcode() == Op_LShiftL || 3468 u->Opcode() == Op_AddL || 3469 u->Opcode() == Op_SubL || 3470 u->Opcode() == Op_AddP) { 3471 wq.push(u); 3472 } 3473 } 3474 // Replace all nodes with identical edges as m with m 3475 k->subsume_by(m, this); 3476 } 3477 } 3478 } 3479 break; 3480 } 3481 case Op_CmpUL: { 3482 if (!Matcher::has_match_rule(Op_CmpUL)) { 3483 // No support for unsigned long comparisons 3484 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3485 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3486 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3487 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3488 Node* andl = new AndLNode(orl, remove_sign_mask); 3489 Node* cmp = new CmpLNode(andl, n->in(2)); 3490 n->subsume_by(cmp, this); 3491 } 3492 break; 3493 } 3494 default: 3495 assert(!n->is_Call(), ""); 3496 assert(!n->is_Mem(), ""); 3497 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3498 break; 3499 } 3500 } 3501 3502 //------------------------------final_graph_reshaping_walk--------------------- 3503 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3504 // requires that the walk visits a node's inputs before visiting the node. 3505 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3506 ResourceArea *area = Thread::current()->resource_area(); 3507 Unique_Node_List sfpt(area); 3508 3509 frc._visited.set(root->_idx); // first, mark node as visited 3510 uint cnt = root->req(); 3511 Node *n = root; 3512 uint i = 0; 3513 while (true) { 3514 if (i < cnt) { 3515 // Place all non-visited non-null inputs onto stack 3516 Node* m = n->in(i); 3517 ++i; 3518 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3519 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3520 // compute worst case interpreter size in case of a deoptimization 3521 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3522 3523 sfpt.push(m); 3524 } 3525 cnt = m->req(); 3526 nstack.push(n, i); // put on stack parent and next input's index 3527 n = m; 3528 i = 0; 3529 } 3530 } else { 3531 // Now do post-visit work 3532 final_graph_reshaping_impl( n, frc ); 3533 if (nstack.is_empty()) 3534 break; // finished 3535 n = nstack.node(); // Get node from stack 3536 cnt = n->req(); 3537 i = nstack.index(); 3538 nstack.pop(); // Shift to the next node on stack 3539 } 3540 } 3541 3542 // Skip next transformation if compressed oops are not used. 3543 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3544 (!UseCompressedOops && !UseCompressedClassPointers)) 3545 return; 3546 3547 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3548 // It could be done for an uncommon traps or any safepoints/calls 3549 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3550 while (sfpt.size() > 0) { 3551 n = sfpt.pop(); 3552 JVMState *jvms = n->as_SafePoint()->jvms(); 3553 assert(jvms != NULL, "sanity"); 3554 int start = jvms->debug_start(); 3555 int end = n->req(); 3556 bool is_uncommon = (n->is_CallStaticJava() && 3557 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3558 for (int j = start; j < end; j++) { 3559 Node* in = n->in(j); 3560 if (in->is_DecodeNarrowPtr()) { 3561 bool safe_to_skip = true; 3562 if (!is_uncommon ) { 3563 // Is it safe to skip? 3564 for (uint i = 0; i < in->outcnt(); i++) { 3565 Node* u = in->raw_out(i); 3566 if (!u->is_SafePoint() || 3567 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3568 safe_to_skip = false; 3569 } 3570 } 3571 } 3572 if (safe_to_skip) { 3573 n->set_req(j, in->in(1)); 3574 } 3575 if (in->outcnt() == 0) { 3576 in->disconnect_inputs(NULL, this); 3577 } 3578 } 3579 } 3580 } 3581 } 3582 3583 //------------------------------final_graph_reshaping-------------------------- 3584 // Final Graph Reshaping. 3585 // 3586 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3587 // and not commoned up and forced early. Must come after regular 3588 // optimizations to avoid GVN undoing the cloning. Clone constant 3589 // inputs to Loop Phis; these will be split by the allocator anyways. 3590 // Remove Opaque nodes. 3591 // (2) Move last-uses by commutative operations to the left input to encourage 3592 // Intel update-in-place two-address operations and better register usage 3593 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3594 // calls canonicalizing them back. 3595 // (3) Count the number of double-precision FP ops, single-precision FP ops 3596 // and call sites. On Intel, we can get correct rounding either by 3597 // forcing singles to memory (requires extra stores and loads after each 3598 // FP bytecode) or we can set a rounding mode bit (requires setting and 3599 // clearing the mode bit around call sites). The mode bit is only used 3600 // if the relative frequency of single FP ops to calls is low enough. 3601 // This is a key transform for SPEC mpeg_audio. 3602 // (4) Detect infinite loops; blobs of code reachable from above but not 3603 // below. Several of the Code_Gen algorithms fail on such code shapes, 3604 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3605 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3606 // Detection is by looking for IfNodes where only 1 projection is 3607 // reachable from below or CatchNodes missing some targets. 3608 // (5) Assert for insane oop offsets in debug mode. 3609 3610 bool Compile::final_graph_reshaping() { 3611 // an infinite loop may have been eliminated by the optimizer, 3612 // in which case the graph will be empty. 3613 if (root()->req() == 1) { 3614 record_method_not_compilable("trivial infinite loop"); 3615 return true; 3616 } 3617 3618 // Expensive nodes have their control input set to prevent the GVN 3619 // from freely commoning them. There's no GVN beyond this point so 3620 // no need to keep the control input. We want the expensive nodes to 3621 // be freely moved to the least frequent code path by gcm. 3622 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3623 for (int i = 0; i < expensive_count(); i++) { 3624 _expensive_nodes->at(i)->set_req(0, NULL); 3625 } 3626 3627 Final_Reshape_Counts frc; 3628 3629 // Visit everybody reachable! 3630 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3631 Node_Stack nstack(live_nodes() >> 1); 3632 final_graph_reshaping_walk(nstack, root(), frc); 3633 3634 // Check for unreachable (from below) code (i.e., infinite loops). 3635 for( uint i = 0; i < frc._tests.size(); i++ ) { 3636 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3637 // Get number of CFG targets. 3638 // Note that PCTables include exception targets after calls. 3639 uint required_outcnt = n->required_outcnt(); 3640 if (n->outcnt() != required_outcnt) { 3641 // Check for a few special cases. Rethrow Nodes never take the 3642 // 'fall-thru' path, so expected kids is 1 less. 3643 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3644 if (n->in(0)->in(0)->is_Call()) { 3645 CallNode *call = n->in(0)->in(0)->as_Call(); 3646 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3647 required_outcnt--; // Rethrow always has 1 less kid 3648 } else if (call->req() > TypeFunc::Parms && 3649 call->is_CallDynamicJava()) { 3650 // Check for null receiver. In such case, the optimizer has 3651 // detected that the virtual call will always result in a null 3652 // pointer exception. The fall-through projection of this CatchNode 3653 // will not be populated. 3654 Node *arg0 = call->in(TypeFunc::Parms); 3655 if (arg0->is_Type() && 3656 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3657 required_outcnt--; 3658 } 3659 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3660 call->req() > TypeFunc::Parms+1 && 3661 call->is_CallStaticJava()) { 3662 // Check for negative array length. In such case, the optimizer has 3663 // detected that the allocation attempt will always result in an 3664 // exception. There is no fall-through projection of this CatchNode . 3665 Node *arg1 = call->in(TypeFunc::Parms+1); 3666 if (arg1->is_Type() && 3667 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3668 required_outcnt--; 3669 } 3670 } 3671 } 3672 } 3673 // Recheck with a better notion of 'required_outcnt' 3674 if (n->outcnt() != required_outcnt) { 3675 record_method_not_compilable("malformed control flow"); 3676 return true; // Not all targets reachable! 3677 } 3678 } 3679 // Check that I actually visited all kids. Unreached kids 3680 // must be infinite loops. 3681 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3682 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3683 record_method_not_compilable("infinite loop"); 3684 return true; // Found unvisited kid; must be unreach 3685 } 3686 3687 // Here so verification code in final_graph_reshaping_walk() 3688 // always see an OuterStripMinedLoopEnd 3689 if (n->is_OuterStripMinedLoopEnd()) { 3690 IfNode* init_iff = n->as_If(); 3691 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3692 n->subsume_by(iff, this); 3693 } 3694 } 3695 3696 // If original bytecodes contained a mixture of floats and doubles 3697 // check if the optimizer has made it homogenous, item (3). 3698 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3699 frc.get_float_count() > 32 && 3700 frc.get_double_count() == 0 && 3701 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3702 set_24_bit_selection_and_mode( false, true ); 3703 } 3704 3705 set_java_calls(frc.get_java_call_count()); 3706 set_inner_loops(frc.get_inner_loop_count()); 3707 3708 // No infinite loops, no reason to bail out. 3709 return false; 3710 } 3711 3712 //-----------------------------too_many_traps---------------------------------- 3713 // Report if there are too many traps at the current method and bci. 3714 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3715 bool Compile::too_many_traps(ciMethod* method, 3716 int bci, 3717 Deoptimization::DeoptReason reason) { 3718 ciMethodData* md = method->method_data(); 3719 if (md->is_empty()) { 3720 // Assume the trap has not occurred, or that it occurred only 3721 // because of a transient condition during start-up in the interpreter. 3722 return false; 3723 } 3724 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3725 if (md->has_trap_at(bci, m, reason) != 0) { 3726 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3727 // Also, if there are multiple reasons, or if there is no per-BCI record, 3728 // assume the worst. 3729 if (log()) 3730 log()->elem("observe trap='%s' count='%d'", 3731 Deoptimization::trap_reason_name(reason), 3732 md->trap_count(reason)); 3733 return true; 3734 } else { 3735 // Ignore method/bci and see if there have been too many globally. 3736 return too_many_traps(reason, md); 3737 } 3738 } 3739 3740 // Less-accurate variant which does not require a method and bci. 3741 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3742 ciMethodData* logmd) { 3743 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3744 // Too many traps globally. 3745 // Note that we use cumulative trap_count, not just md->trap_count. 3746 if (log()) { 3747 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3748 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3749 Deoptimization::trap_reason_name(reason), 3750 mcount, trap_count(reason)); 3751 } 3752 return true; 3753 } else { 3754 // The coast is clear. 3755 return false; 3756 } 3757 } 3758 3759 //--------------------------too_many_recompiles-------------------------------- 3760 // Report if there are too many recompiles at the current method and bci. 3761 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3762 // Is not eager to return true, since this will cause the compiler to use 3763 // Action_none for a trap point, to avoid too many recompilations. 3764 bool Compile::too_many_recompiles(ciMethod* method, 3765 int bci, 3766 Deoptimization::DeoptReason reason) { 3767 ciMethodData* md = method->method_data(); 3768 if (md->is_empty()) { 3769 // Assume the trap has not occurred, or that it occurred only 3770 // because of a transient condition during start-up in the interpreter. 3771 return false; 3772 } 3773 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3774 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3775 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3776 Deoptimization::DeoptReason per_bc_reason 3777 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3778 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3779 if ((per_bc_reason == Deoptimization::Reason_none 3780 || md->has_trap_at(bci, m, reason) != 0) 3781 // The trap frequency measure we care about is the recompile count: 3782 && md->trap_recompiled_at(bci, m) 3783 && md->overflow_recompile_count() >= bc_cutoff) { 3784 // Do not emit a trap here if it has already caused recompilations. 3785 // Also, if there are multiple reasons, or if there is no per-BCI record, 3786 // assume the worst. 3787 if (log()) 3788 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3789 Deoptimization::trap_reason_name(reason), 3790 md->trap_count(reason), 3791 md->overflow_recompile_count()); 3792 return true; 3793 } else if (trap_count(reason) != 0 3794 && decompile_count() >= m_cutoff) { 3795 // Too many recompiles globally, and we have seen this sort of trap. 3796 // Use cumulative decompile_count, not just md->decompile_count. 3797 if (log()) 3798 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3799 Deoptimization::trap_reason_name(reason), 3800 md->trap_count(reason), trap_count(reason), 3801 md->decompile_count(), decompile_count()); 3802 return true; 3803 } else { 3804 // The coast is clear. 3805 return false; 3806 } 3807 } 3808 3809 // Compute when not to trap. Used by matching trap based nodes and 3810 // NullCheck optimization. 3811 void Compile::set_allowed_deopt_reasons() { 3812 _allowed_reasons = 0; 3813 if (is_method_compilation()) { 3814 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3815 assert(rs < BitsPerInt, "recode bit map"); 3816 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3817 _allowed_reasons |= nth_bit(rs); 3818 } 3819 } 3820 } 3821 } 3822 3823 bool Compile::is_compiling_clinit_for(ciKlass* k) { 3824 ciMethod* root = method(); // the root method of compilation 3825 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit 3826 } 3827 3828 #ifndef PRODUCT 3829 //------------------------------verify_graph_edges--------------------------- 3830 // Walk the Graph and verify that there is a one-to-one correspondence 3831 // between Use-Def edges and Def-Use edges in the graph. 3832 void Compile::verify_graph_edges(bool no_dead_code) { 3833 if (VerifyGraphEdges) { 3834 ResourceArea *area = Thread::current()->resource_area(); 3835 Unique_Node_List visited(area); 3836 // Call recursive graph walk to check edges 3837 _root->verify_edges(visited); 3838 if (no_dead_code) { 3839 // Now make sure that no visited node is used by an unvisited node. 3840 bool dead_nodes = false; 3841 Unique_Node_List checked(area); 3842 while (visited.size() > 0) { 3843 Node* n = visited.pop(); 3844 checked.push(n); 3845 for (uint i = 0; i < n->outcnt(); i++) { 3846 Node* use = n->raw_out(i); 3847 if (checked.member(use)) continue; // already checked 3848 if (visited.member(use)) continue; // already in the graph 3849 if (use->is_Con()) continue; // a dead ConNode is OK 3850 // At this point, we have found a dead node which is DU-reachable. 3851 if (!dead_nodes) { 3852 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3853 dead_nodes = true; 3854 } 3855 use->dump(2); 3856 tty->print_cr("---"); 3857 checked.push(use); // No repeats; pretend it is now checked. 3858 } 3859 } 3860 assert(!dead_nodes, "using nodes must be reachable from root"); 3861 } 3862 } 3863 } 3864 #endif 3865 3866 // The Compile object keeps track of failure reasons separately from the ciEnv. 3867 // This is required because there is not quite a 1-1 relation between the 3868 // ciEnv and its compilation task and the Compile object. Note that one 3869 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3870 // to backtrack and retry without subsuming loads. Other than this backtracking 3871 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3872 // by the logic in C2Compiler. 3873 void Compile::record_failure(const char* reason) { 3874 if (log() != NULL) { 3875 log()->elem("failure reason='%s' phase='compile'", reason); 3876 } 3877 if (_failure_reason == NULL) { 3878 // Record the first failure reason. 3879 _failure_reason = reason; 3880 } 3881 3882 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3883 C->print_method(PHASE_FAILURE); 3884 } 3885 _root = NULL; // flush the graph, too 3886 } 3887 3888 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3889 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3890 _phase_name(name), _dolog(CITimeVerbose) 3891 { 3892 if (_dolog) { 3893 C = Compile::current(); 3894 _log = C->log(); 3895 } else { 3896 C = NULL; 3897 _log = NULL; 3898 } 3899 if (_log != NULL) { 3900 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3901 _log->stamp(); 3902 _log->end_head(); 3903 } 3904 } 3905 3906 Compile::TracePhase::~TracePhase() { 3907 3908 C = Compile::current(); 3909 if (_dolog) { 3910 _log = C->log(); 3911 } else { 3912 _log = NULL; 3913 } 3914 3915 #ifdef ASSERT 3916 if (PrintIdealNodeCount) { 3917 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3918 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3919 } 3920 3921 if (VerifyIdealNodeCount) { 3922 Compile::current()->print_missing_nodes(); 3923 } 3924 #endif 3925 3926 if (_log != NULL) { 3927 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3928 } 3929 } 3930 3931 //============================================================================= 3932 // Two Constant's are equal when the type and the value are equal. 3933 bool Compile::Constant::operator==(const Constant& other) { 3934 if (type() != other.type() ) return false; 3935 if (can_be_reused() != other.can_be_reused()) return false; 3936 // For floating point values we compare the bit pattern. 3937 switch (type()) { 3938 case T_INT: 3939 case T_FLOAT: return (_v._value.i == other._v._value.i); 3940 case T_LONG: 3941 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3942 case T_OBJECT: 3943 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3944 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3945 case T_METADATA: return (_v._metadata == other._v._metadata); 3946 default: ShouldNotReachHere(); return false; 3947 } 3948 } 3949 3950 static int type_to_size_in_bytes(BasicType t) { 3951 switch (t) { 3952 case T_INT: return sizeof(jint ); 3953 case T_LONG: return sizeof(jlong ); 3954 case T_FLOAT: return sizeof(jfloat ); 3955 case T_DOUBLE: return sizeof(jdouble); 3956 case T_METADATA: return sizeof(Metadata*); 3957 // We use T_VOID as marker for jump-table entries (labels) which 3958 // need an internal word relocation. 3959 case T_VOID: 3960 case T_ADDRESS: 3961 case T_OBJECT: return sizeof(jobject); 3962 default: 3963 ShouldNotReachHere(); 3964 return -1; 3965 } 3966 } 3967 3968 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3969 // sort descending 3970 if (a->freq() > b->freq()) return -1; 3971 if (a->freq() < b->freq()) return 1; 3972 return 0; 3973 } 3974 3975 void Compile::ConstantTable::calculate_offsets_and_size() { 3976 // First, sort the array by frequencies. 3977 _constants.sort(qsort_comparator); 3978 3979 #ifdef ASSERT 3980 // Make sure all jump-table entries were sorted to the end of the 3981 // array (they have a negative frequency). 3982 bool found_void = false; 3983 for (int i = 0; i < _constants.length(); i++) { 3984 Constant con = _constants.at(i); 3985 if (con.type() == T_VOID) 3986 found_void = true; // jump-tables 3987 else 3988 assert(!found_void, "wrong sorting"); 3989 } 3990 #endif 3991 3992 int offset = 0; 3993 for (int i = 0; i < _constants.length(); i++) { 3994 Constant* con = _constants.adr_at(i); 3995 3996 // Align offset for type. 3997 int typesize = type_to_size_in_bytes(con->type()); 3998 offset = align_up(offset, typesize); 3999 con->set_offset(offset); // set constant's offset 4000 4001 if (con->type() == T_VOID) { 4002 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4003 offset = offset + typesize * n->outcnt(); // expand jump-table 4004 } else { 4005 offset = offset + typesize; 4006 } 4007 } 4008 4009 // Align size up to the next section start (which is insts; see 4010 // CodeBuffer::align_at_start). 4011 assert(_size == -1, "already set?"); 4012 _size = align_up(offset, (int)CodeEntryAlignment); 4013 } 4014 4015 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4016 MacroAssembler _masm(&cb); 4017 for (int i = 0; i < _constants.length(); i++) { 4018 Constant con = _constants.at(i); 4019 address constant_addr = NULL; 4020 switch (con.type()) { 4021 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4022 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4023 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4024 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4025 case T_OBJECT: { 4026 jobject obj = con.get_jobject(); 4027 int oop_index = _masm.oop_recorder()->find_index(obj); 4028 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4029 break; 4030 } 4031 case T_ADDRESS: { 4032 address addr = (address) con.get_jobject(); 4033 constant_addr = _masm.address_constant(addr); 4034 break; 4035 } 4036 // We use T_VOID as marker for jump-table entries (labels) which 4037 // need an internal word relocation. 4038 case T_VOID: { 4039 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4040 // Fill the jump-table with a dummy word. The real value is 4041 // filled in later in fill_jump_table. 4042 address dummy = (address) n; 4043 constant_addr = _masm.address_constant(dummy); 4044 // Expand jump-table 4045 for (uint i = 1; i < n->outcnt(); i++) { 4046 address temp_addr = _masm.address_constant(dummy + i); 4047 assert(temp_addr, "consts section too small"); 4048 } 4049 break; 4050 } 4051 case T_METADATA: { 4052 Metadata* obj = con.get_metadata(); 4053 int metadata_index = _masm.oop_recorder()->find_index(obj); 4054 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4055 break; 4056 } 4057 default: ShouldNotReachHere(); 4058 } 4059 assert(constant_addr, "consts section too small"); 4060 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4061 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4062 } 4063 } 4064 4065 int Compile::ConstantTable::find_offset(Constant& con) const { 4066 int idx = _constants.find(con); 4067 guarantee(idx != -1, "constant must be in constant table"); 4068 int offset = _constants.at(idx).offset(); 4069 guarantee(offset != -1, "constant table not emitted yet?"); 4070 return offset; 4071 } 4072 4073 void Compile::ConstantTable::add(Constant& con) { 4074 if (con.can_be_reused()) { 4075 int idx = _constants.find(con); 4076 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4077 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4078 return; 4079 } 4080 } 4081 (void) _constants.append(con); 4082 } 4083 4084 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4085 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4086 Constant con(type, value, b->_freq); 4087 add(con); 4088 return con; 4089 } 4090 4091 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4092 Constant con(metadata); 4093 add(con); 4094 return con; 4095 } 4096 4097 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4098 jvalue value; 4099 BasicType type = oper->type()->basic_type(); 4100 switch (type) { 4101 case T_LONG: value.j = oper->constantL(); break; 4102 case T_FLOAT: value.f = oper->constantF(); break; 4103 case T_DOUBLE: value.d = oper->constantD(); break; 4104 case T_OBJECT: 4105 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4106 case T_METADATA: return add((Metadata*)oper->constant()); break; 4107 default: guarantee(false, "unhandled type: %s", type2name(type)); 4108 } 4109 return add(n, type, value); 4110 } 4111 4112 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4113 jvalue value; 4114 // We can use the node pointer here to identify the right jump-table 4115 // as this method is called from Compile::Fill_buffer right before 4116 // the MachNodes are emitted and the jump-table is filled (means the 4117 // MachNode pointers do not change anymore). 4118 value.l = (jobject) n; 4119 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4120 add(con); 4121 return con; 4122 } 4123 4124 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4125 // If called from Compile::scratch_emit_size do nothing. 4126 if (Compile::current()->in_scratch_emit_size()) return; 4127 4128 assert(labels.is_nonempty(), "must be"); 4129 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4130 4131 // Since MachConstantNode::constant_offset() also contains 4132 // table_base_offset() we need to subtract the table_base_offset() 4133 // to get the plain offset into the constant table. 4134 int offset = n->constant_offset() - table_base_offset(); 4135 4136 MacroAssembler _masm(&cb); 4137 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4138 4139 for (uint i = 0; i < n->outcnt(); i++) { 4140 address* constant_addr = &jump_table_base[i]; 4141 assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)); 4142 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4143 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4144 } 4145 } 4146 4147 //----------------------------static_subtype_check----------------------------- 4148 // Shortcut important common cases when superklass is exact: 4149 // (0) superklass is java.lang.Object (can occur in reflective code) 4150 // (1) subklass is already limited to a subtype of superklass => always ok 4151 // (2) subklass does not overlap with superklass => always fail 4152 // (3) superklass has NO subtypes and we can check with a simple compare. 4153 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4154 if (StressReflectiveCode) { 4155 return SSC_full_test; // Let caller generate the general case. 4156 } 4157 4158 if (superk == env()->Object_klass()) { 4159 return SSC_always_true; // (0) this test cannot fail 4160 } 4161 4162 ciType* superelem = superk; 4163 if (superelem->is_array_klass()) 4164 superelem = superelem->as_array_klass()->base_element_type(); 4165 4166 if (!subk->is_interface()) { // cannot trust static interface types yet 4167 if (subk->is_subtype_of(superk)) { 4168 return SSC_always_true; // (1) false path dead; no dynamic test needed 4169 } 4170 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4171 !superk->is_subtype_of(subk)) { 4172 return SSC_always_false; 4173 } 4174 } 4175 4176 // If casting to an instance klass, it must have no subtypes 4177 if (superk->is_interface()) { 4178 // Cannot trust interfaces yet. 4179 // %%% S.B. superk->nof_implementors() == 1 4180 } else if (superelem->is_instance_klass()) { 4181 ciInstanceKlass* ik = superelem->as_instance_klass(); 4182 if (!ik->has_subklass() && !ik->is_interface()) { 4183 if (!ik->is_final()) { 4184 // Add a dependency if there is a chance of a later subclass. 4185 dependencies()->assert_leaf_type(ik); 4186 } 4187 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4188 } 4189 } else { 4190 // A primitive array type has no subtypes. 4191 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4192 } 4193 4194 return SSC_full_test; 4195 } 4196 4197 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4198 #ifdef _LP64 4199 // The scaled index operand to AddP must be a clean 64-bit value. 4200 // Java allows a 32-bit int to be incremented to a negative 4201 // value, which appears in a 64-bit register as a large 4202 // positive number. Using that large positive number as an 4203 // operand in pointer arithmetic has bad consequences. 4204 // On the other hand, 32-bit overflow is rare, and the possibility 4205 // can often be excluded, if we annotate the ConvI2L node with 4206 // a type assertion that its value is known to be a small positive 4207 // number. (The prior range check has ensured this.) 4208 // This assertion is used by ConvI2LNode::Ideal. 4209 int index_max = max_jint - 1; // array size is max_jint, index is one less 4210 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4211 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4212 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4213 #endif 4214 return idx; 4215 } 4216 4217 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4218 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4219 if (ctrl != NULL) { 4220 // Express control dependency by a CastII node with a narrow type. 4221 value = new CastIINode(value, itype, false, true /* range check dependency */); 4222 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4223 // node from floating above the range check during loop optimizations. Otherwise, the 4224 // ConvI2L node may be eliminated independently of the range check, causing the data path 4225 // to become TOP while the control path is still there (although it's unreachable). 4226 value->set_req(0, ctrl); 4227 // Save CastII node to remove it after loop optimizations. 4228 phase->C->add_range_check_cast(value); 4229 value = phase->transform(value); 4230 } 4231 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4232 return phase->transform(new ConvI2LNode(value, ltype)); 4233 } 4234 4235 // The message about the current inlining is accumulated in 4236 // _print_inlining_stream and transfered into the _print_inlining_list 4237 // once we know whether inlining succeeds or not. For regular 4238 // inlining, messages are appended to the buffer pointed by 4239 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4240 // a new buffer is added after _print_inlining_idx in the list. This 4241 // way we can update the inlining message for late inlining call site 4242 // when the inlining is attempted again. 4243 void Compile::print_inlining_init() { 4244 if (print_inlining() || print_intrinsics()) { 4245 _print_inlining_stream = new stringStream(); 4246 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4247 } 4248 } 4249 4250 void Compile::print_inlining_reinit() { 4251 if (print_inlining() || print_intrinsics()) { 4252 // Re allocate buffer when we change ResourceMark 4253 _print_inlining_stream = new stringStream(); 4254 } 4255 } 4256 4257 void Compile::print_inlining_reset() { 4258 _print_inlining_stream->reset(); 4259 } 4260 4261 void Compile::print_inlining_commit() { 4262 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4263 // Transfer the message from _print_inlining_stream to the current 4264 // _print_inlining_list buffer and clear _print_inlining_stream. 4265 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4266 print_inlining_reset(); 4267 } 4268 4269 void Compile::print_inlining_push() { 4270 // Add new buffer to the _print_inlining_list at current position 4271 _print_inlining_idx++; 4272 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4273 } 4274 4275 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4276 return _print_inlining_list->at(_print_inlining_idx); 4277 } 4278 4279 void Compile::print_inlining_update(CallGenerator* cg) { 4280 if (print_inlining() || print_intrinsics()) { 4281 if (!cg->is_late_inline()) { 4282 if (print_inlining_current().cg() != NULL) { 4283 print_inlining_push(); 4284 } 4285 print_inlining_commit(); 4286 } else { 4287 if (print_inlining_current().cg() != cg && 4288 (print_inlining_current().cg() != NULL || 4289 print_inlining_current().ss()->size() != 0)) { 4290 print_inlining_push(); 4291 } 4292 print_inlining_commit(); 4293 print_inlining_current().set_cg(cg); 4294 } 4295 } 4296 } 4297 4298 void Compile::print_inlining_move_to(CallGenerator* cg) { 4299 // We resume inlining at a late inlining call site. Locate the 4300 // corresponding inlining buffer so that we can update it. 4301 if (print_inlining()) { 4302 for (int i = 0; i < _print_inlining_list->length(); i++) { 4303 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4304 _print_inlining_idx = i; 4305 return; 4306 } 4307 } 4308 ShouldNotReachHere(); 4309 } 4310 } 4311 4312 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4313 if (print_inlining()) { 4314 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4315 assert(print_inlining_current().cg() == cg, "wrong entry"); 4316 // replace message with new message 4317 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4318 print_inlining_commit(); 4319 print_inlining_current().set_cg(cg); 4320 } 4321 } 4322 4323 void Compile::print_inlining_assert_ready() { 4324 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4325 } 4326 4327 void Compile::process_print_inlining() { 4328 bool do_print_inlining = print_inlining() || print_intrinsics(); 4329 if (do_print_inlining || log() != NULL) { 4330 // Print inlining message for candidates that we couldn't inline 4331 // for lack of space 4332 for (int i = 0; i < _late_inlines.length(); i++) { 4333 CallGenerator* cg = _late_inlines.at(i); 4334 if (!cg->is_mh_late_inline()) { 4335 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4336 if (do_print_inlining) { 4337 cg->print_inlining_late(msg); 4338 } 4339 log_late_inline_failure(cg, msg); 4340 } 4341 } 4342 } 4343 if (do_print_inlining) { 4344 ResourceMark rm; 4345 stringStream ss; 4346 for (int i = 0; i < _print_inlining_list->length(); i++) { 4347 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4348 } 4349 size_t end = ss.size(); 4350 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4351 strncpy(_print_inlining_output, ss.base(), end+1); 4352 _print_inlining_output[end] = 0; 4353 } 4354 } 4355 4356 void Compile::dump_print_inlining() { 4357 if (_print_inlining_output != NULL) { 4358 tty->print_raw(_print_inlining_output); 4359 } 4360 } 4361 4362 void Compile::log_late_inline(CallGenerator* cg) { 4363 if (log() != NULL) { 4364 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4365 cg->unique_id()); 4366 JVMState* p = cg->call_node()->jvms(); 4367 while (p != NULL) { 4368 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4369 p = p->caller(); 4370 } 4371 log()->tail("late_inline"); 4372 } 4373 } 4374 4375 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4376 log_late_inline(cg); 4377 if (log() != NULL) { 4378 log()->inline_fail(msg); 4379 } 4380 } 4381 4382 void Compile::log_inline_id(CallGenerator* cg) { 4383 if (log() != NULL) { 4384 // The LogCompilation tool needs a unique way to identify late 4385 // inline call sites. This id must be unique for this call site in 4386 // this compilation. Try to have it unique across compilations as 4387 // well because it can be convenient when grepping through the log 4388 // file. 4389 // Distinguish OSR compilations from others in case CICountOSR is 4390 // on. 4391 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4392 cg->set_unique_id(id); 4393 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4394 } 4395 } 4396 4397 void Compile::log_inline_failure(const char* msg) { 4398 if (C->log() != NULL) { 4399 C->log()->inline_fail(msg); 4400 } 4401 } 4402 4403 4404 // Dump inlining replay data to the stream. 4405 // Don't change thread state and acquire any locks. 4406 void Compile::dump_inline_data(outputStream* out) { 4407 InlineTree* inl_tree = ilt(); 4408 if (inl_tree != NULL) { 4409 out->print(" inline %d", inl_tree->count()); 4410 inl_tree->dump_replay_data(out); 4411 } 4412 } 4413 4414 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4415 if (n1->Opcode() < n2->Opcode()) return -1; 4416 else if (n1->Opcode() > n2->Opcode()) return 1; 4417 4418 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4419 for (uint i = 1; i < n1->req(); i++) { 4420 if (n1->in(i) < n2->in(i)) return -1; 4421 else if (n1->in(i) > n2->in(i)) return 1; 4422 } 4423 4424 return 0; 4425 } 4426 4427 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4428 Node* n1 = *n1p; 4429 Node* n2 = *n2p; 4430 4431 return cmp_expensive_nodes(n1, n2); 4432 } 4433 4434 void Compile::sort_expensive_nodes() { 4435 if (!expensive_nodes_sorted()) { 4436 _expensive_nodes->sort(cmp_expensive_nodes); 4437 } 4438 } 4439 4440 bool Compile::expensive_nodes_sorted() const { 4441 for (int i = 1; i < _expensive_nodes->length(); i++) { 4442 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4443 return false; 4444 } 4445 } 4446 return true; 4447 } 4448 4449 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4450 if (_expensive_nodes->length() == 0) { 4451 return false; 4452 } 4453 4454 assert(OptimizeExpensiveOps, "optimization off?"); 4455 4456 // Take this opportunity to remove dead nodes from the list 4457 int j = 0; 4458 for (int i = 0; i < _expensive_nodes->length(); i++) { 4459 Node* n = _expensive_nodes->at(i); 4460 if (!n->is_unreachable(igvn)) { 4461 assert(n->is_expensive(), "should be expensive"); 4462 _expensive_nodes->at_put(j, n); 4463 j++; 4464 } 4465 } 4466 _expensive_nodes->trunc_to(j); 4467 4468 // Then sort the list so that similar nodes are next to each other 4469 // and check for at least two nodes of identical kind with same data 4470 // inputs. 4471 sort_expensive_nodes(); 4472 4473 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4474 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4475 return true; 4476 } 4477 } 4478 4479 return false; 4480 } 4481 4482 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4483 if (_expensive_nodes->length() == 0) { 4484 return; 4485 } 4486 4487 assert(OptimizeExpensiveOps, "optimization off?"); 4488 4489 // Sort to bring similar nodes next to each other and clear the 4490 // control input of nodes for which there's only a single copy. 4491 sort_expensive_nodes(); 4492 4493 int j = 0; 4494 int identical = 0; 4495 int i = 0; 4496 bool modified = false; 4497 for (; i < _expensive_nodes->length()-1; i++) { 4498 assert(j <= i, "can't write beyond current index"); 4499 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4500 identical++; 4501 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4502 continue; 4503 } 4504 if (identical > 0) { 4505 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4506 identical = 0; 4507 } else { 4508 Node* n = _expensive_nodes->at(i); 4509 igvn.replace_input_of(n, 0, NULL); 4510 igvn.hash_insert(n); 4511 modified = true; 4512 } 4513 } 4514 if (identical > 0) { 4515 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4516 } else if (_expensive_nodes->length() >= 1) { 4517 Node* n = _expensive_nodes->at(i); 4518 igvn.replace_input_of(n, 0, NULL); 4519 igvn.hash_insert(n); 4520 modified = true; 4521 } 4522 _expensive_nodes->trunc_to(j); 4523 if (modified) { 4524 igvn.optimize(); 4525 } 4526 } 4527 4528 void Compile::add_expensive_node(Node * n) { 4529 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4530 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4531 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4532 if (OptimizeExpensiveOps) { 4533 _expensive_nodes->append(n); 4534 } else { 4535 // Clear control input and let IGVN optimize expensive nodes if 4536 // OptimizeExpensiveOps is off. 4537 n->set_req(0, NULL); 4538 } 4539 } 4540 4541 /** 4542 * Remove the speculative part of types and clean up the graph 4543 */ 4544 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4545 if (UseTypeSpeculation) { 4546 Unique_Node_List worklist; 4547 worklist.push(root()); 4548 int modified = 0; 4549 // Go over all type nodes that carry a speculative type, drop the 4550 // speculative part of the type and enqueue the node for an igvn 4551 // which may optimize it out. 4552 for (uint next = 0; next < worklist.size(); ++next) { 4553 Node *n = worklist.at(next); 4554 if (n->is_Type()) { 4555 TypeNode* tn = n->as_Type(); 4556 const Type* t = tn->type(); 4557 const Type* t_no_spec = t->remove_speculative(); 4558 if (t_no_spec != t) { 4559 bool in_hash = igvn.hash_delete(n); 4560 assert(in_hash, "node should be in igvn hash table"); 4561 tn->set_type(t_no_spec); 4562 igvn.hash_insert(n); 4563 igvn._worklist.push(n); // give it a chance to go away 4564 modified++; 4565 } 4566 } 4567 uint max = n->len(); 4568 for( uint i = 0; i < max; ++i ) { 4569 Node *m = n->in(i); 4570 if (not_a_node(m)) continue; 4571 worklist.push(m); 4572 } 4573 } 4574 // Drop the speculative part of all types in the igvn's type table 4575 igvn.remove_speculative_types(); 4576 if (modified > 0) { 4577 igvn.optimize(); 4578 } 4579 #ifdef ASSERT 4580 // Verify that after the IGVN is over no speculative type has resurfaced 4581 worklist.clear(); 4582 worklist.push(root()); 4583 for (uint next = 0; next < worklist.size(); ++next) { 4584 Node *n = worklist.at(next); 4585 const Type* t = igvn.type_or_null(n); 4586 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4587 if (n->is_Type()) { 4588 t = n->as_Type()->type(); 4589 assert(t == t->remove_speculative(), "no more speculative types"); 4590 } 4591 uint max = n->len(); 4592 for( uint i = 0; i < max; ++i ) { 4593 Node *m = n->in(i); 4594 if (not_a_node(m)) continue; 4595 worklist.push(m); 4596 } 4597 } 4598 igvn.check_no_speculative_types(); 4599 #endif 4600 } 4601 } 4602 4603 // Auxiliary method to support randomized stressing/fuzzing. 4604 // 4605 // This method can be called the arbitrary number of times, with current count 4606 // as the argument. The logic allows selecting a single candidate from the 4607 // running list of candidates as follows: 4608 // int count = 0; 4609 // Cand* selected = null; 4610 // while(cand = cand->next()) { 4611 // if (randomized_select(++count)) { 4612 // selected = cand; 4613 // } 4614 // } 4615 // 4616 // Including count equalizes the chances any candidate is "selected". 4617 // This is useful when we don't have the complete list of candidates to choose 4618 // from uniformly. In this case, we need to adjust the randomicity of the 4619 // selection, or else we will end up biasing the selection towards the latter 4620 // candidates. 4621 // 4622 // Quick back-envelope calculation shows that for the list of n candidates 4623 // the equal probability for the candidate to persist as "best" can be 4624 // achieved by replacing it with "next" k-th candidate with the probability 4625 // of 1/k. It can be easily shown that by the end of the run, the 4626 // probability for any candidate is converged to 1/n, thus giving the 4627 // uniform distribution among all the candidates. 4628 // 4629 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4630 #define RANDOMIZED_DOMAIN_POW 29 4631 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4632 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4633 bool Compile::randomized_select(int count) { 4634 assert(count > 0, "only positive"); 4635 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4636 } 4637 4638 CloneMap& Compile::clone_map() { return _clone_map; } 4639 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4640 4641 void NodeCloneInfo::dump() const { 4642 tty->print(" {%d:%d} ", idx(), gen()); 4643 } 4644 4645 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4646 uint64_t val = value(old->_idx); 4647 NodeCloneInfo cio(val); 4648 assert(val != 0, "old node should be in the map"); 4649 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4650 insert(nnn->_idx, cin.get()); 4651 #ifndef PRODUCT 4652 if (is_debug()) { 4653 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4654 } 4655 #endif 4656 } 4657 4658 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4659 NodeCloneInfo cio(value(old->_idx)); 4660 if (cio.get() == 0) { 4661 cio.set(old->_idx, 0); 4662 insert(old->_idx, cio.get()); 4663 #ifndef PRODUCT 4664 if (is_debug()) { 4665 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4666 } 4667 #endif 4668 } 4669 clone(old, nnn, gen); 4670 } 4671 4672 int CloneMap::max_gen() const { 4673 int g = 0; 4674 DictI di(_dict); 4675 for(; di.test(); ++di) { 4676 int t = gen(di._key); 4677 if (g < t) { 4678 g = t; 4679 #ifndef PRODUCT 4680 if (is_debug()) { 4681 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4682 } 4683 #endif 4684 } 4685 } 4686 return g; 4687 } 4688 4689 void CloneMap::dump(node_idx_t key) const { 4690 uint64_t val = value(key); 4691 if (val != 0) { 4692 NodeCloneInfo ni(val); 4693 ni.dump(); 4694 } 4695 }