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_G1GC 80 #include "gc/g1/g1ThreadLocalData.hpp" 81 #endif // INCLUDE_G1GC 82 #if INCLUDE_ZGC 83 #include "gc/z/c2/zBarrierSetC2.hpp" 84 #endif 85 86 87 // -------------------- Compile::mach_constant_base_node ----------------------- 88 // Constant table base node singleton. 89 MachConstantBaseNode* Compile::mach_constant_base_node() { 90 if (_mach_constant_base_node == NULL) { 91 _mach_constant_base_node = new MachConstantBaseNode(); 92 _mach_constant_base_node->add_req(C->root()); 93 } 94 return _mach_constant_base_node; 95 } 96 97 98 /// Support for intrinsics. 99 100 // Return the index at which m must be inserted (or already exists). 101 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 102 class IntrinsicDescPair { 103 private: 104 ciMethod* _m; 105 bool _is_virtual; 106 public: 107 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 108 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 109 ciMethod* m= elt->method(); 110 ciMethod* key_m = key->_m; 111 if (key_m < m) return -1; 112 else if (key_m > m) return 1; 113 else { 114 bool is_virtual = elt->is_virtual(); 115 bool key_virtual = key->_is_virtual; 116 if (key_virtual < is_virtual) return -1; 117 else if (key_virtual > is_virtual) return 1; 118 else return 0; 119 } 120 } 121 }; 122 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 123 #ifdef ASSERT 124 for (int i = 1; i < _intrinsics->length(); i++) { 125 CallGenerator* cg1 = _intrinsics->at(i-1); 126 CallGenerator* cg2 = _intrinsics->at(i); 127 assert(cg1->method() != cg2->method() 128 ? cg1->method() < cg2->method() 129 : cg1->is_virtual() < cg2->is_virtual(), 130 "compiler intrinsics list must stay sorted"); 131 } 132 #endif 133 IntrinsicDescPair pair(m, is_virtual); 134 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 135 } 136 137 void Compile::register_intrinsic(CallGenerator* cg) { 138 if (_intrinsics == NULL) { 139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 140 } 141 int len = _intrinsics->length(); 142 bool found = false; 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 144 assert(!found, "registering twice"); 145 _intrinsics->insert_before(index, cg); 146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 147 } 148 149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 150 assert(m->is_loaded(), "don't try this on unloaded methods"); 151 if (_intrinsics != NULL) { 152 bool found = false; 153 int index = intrinsic_insertion_index(m, is_virtual, found); 154 if (found) { 155 return _intrinsics->at(index); 156 } 157 } 158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 159 if (m->intrinsic_id() != vmIntrinsics::_none && 160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 162 if (cg != NULL) { 163 // Save it for next time: 164 register_intrinsic(cg); 165 return cg; 166 } else { 167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 168 } 169 } 170 return NULL; 171 } 172 173 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 174 // in library_call.cpp. 175 176 177 #ifndef PRODUCT 178 // statistics gathering... 179 180 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 182 183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 185 int oflags = _intrinsic_hist_flags[id]; 186 assert(flags != 0, "what happened?"); 187 if (is_virtual) { 188 flags |= _intrinsic_virtual; 189 } 190 bool changed = (flags != oflags); 191 if ((flags & _intrinsic_worked) != 0) { 192 juint count = (_intrinsic_hist_count[id] += 1); 193 if (count == 1) { 194 changed = true; // first time 195 } 196 // increment the overall count also: 197 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 198 } 199 if (changed) { 200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 201 // Something changed about the intrinsic's virtuality. 202 if ((flags & _intrinsic_virtual) != 0) { 203 // This is the first use of this intrinsic as a virtual call. 204 if (oflags != 0) { 205 // We already saw it as a non-virtual, so note both cases. 206 flags |= _intrinsic_both; 207 } 208 } else if ((oflags & _intrinsic_both) == 0) { 209 // This is the first use of this intrinsic as a non-virtual 210 flags |= _intrinsic_both; 211 } 212 } 213 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 214 } 215 // update the overall flags also: 216 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 217 return changed; 218 } 219 220 static char* format_flags(int flags, char* buf) { 221 buf[0] = 0; 222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 227 if (buf[0] == 0) strcat(buf, ","); 228 assert(buf[0] == ',', "must be"); 229 return &buf[1]; 230 } 231 232 void Compile::print_intrinsic_statistics() { 233 char flagsbuf[100]; 234 ttyLocker ttyl; 235 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 236 tty->print_cr("Compiler intrinsic usage:"); 237 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 238 if (total == 0) total = 1; // avoid div0 in case of no successes 239 #define PRINT_STAT_LINE(name, c, f) \ 240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 241 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 242 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 243 int flags = _intrinsic_hist_flags[id]; 244 juint count = _intrinsic_hist_count[id]; 245 if ((flags | count) != 0) { 246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 247 } 248 } 249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 250 if (xtty != NULL) xtty->tail("statistics"); 251 } 252 253 void Compile::print_statistics() { 254 { ttyLocker ttyl; 255 if (xtty != NULL) xtty->head("statistics type='opto'"); 256 Parse::print_statistics(); 257 PhaseCCP::print_statistics(); 258 PhaseRegAlloc::print_statistics(); 259 Scheduling::print_statistics(); 260 PhasePeephole::print_statistics(); 261 PhaseIdealLoop::print_statistics(); 262 if (xtty != NULL) xtty->tail("statistics"); 263 } 264 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 265 // put this under its own <statistics> element. 266 print_intrinsic_statistics(); 267 } 268 } 269 #endif //PRODUCT 270 271 // Support for bundling info 272 Bundle* Compile::node_bundling(const Node *n) { 273 assert(valid_bundle_info(n), "oob"); 274 return &_node_bundling_base[n->_idx]; 275 } 276 277 bool Compile::valid_bundle_info(const Node *n) { 278 return (_node_bundling_limit > n->_idx); 279 } 280 281 282 void Compile::gvn_replace_by(Node* n, Node* nn) { 283 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 284 Node* use = n->last_out(i); 285 bool is_in_table = initial_gvn()->hash_delete(use); 286 uint uses_found = 0; 287 for (uint j = 0; j < use->len(); j++) { 288 if (use->in(j) == n) { 289 if (j < use->req()) 290 use->set_req(j, nn); 291 else 292 use->set_prec(j, nn); 293 uses_found++; 294 } 295 } 296 if (is_in_table) { 297 // reinsert into table 298 initial_gvn()->hash_find_insert(use); 299 } 300 record_for_igvn(use); 301 i -= uses_found; // we deleted 1 or more copies of this edge 302 } 303 } 304 305 306 static inline bool not_a_node(const Node* n) { 307 if (n == NULL) return true; 308 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 309 if (*(address*)n == badAddress) return true; // kill by Node::destruct 310 return false; 311 } 312 313 // Identify all nodes that are reachable from below, useful. 314 // Use breadth-first pass that records state in a Unique_Node_List, 315 // recursive traversal is slower. 316 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 317 int estimated_worklist_size = live_nodes(); 318 useful.map( estimated_worklist_size, NULL ); // preallocate space 319 320 // Initialize worklist 321 if (root() != NULL) { useful.push(root()); } 322 // If 'top' is cached, declare it useful to preserve cached node 323 if( cached_top_node() ) { useful.push(cached_top_node()); } 324 325 // Push all useful nodes onto the list, breadthfirst 326 for( uint next = 0; next < useful.size(); ++next ) { 327 assert( next < unique(), "Unique useful nodes < total nodes"); 328 Node *n = useful.at(next); 329 uint max = n->len(); 330 for( uint i = 0; i < max; ++i ) { 331 Node *m = n->in(i); 332 if (not_a_node(m)) continue; 333 useful.push(m); 334 } 335 } 336 } 337 338 // Update dead_node_list with any missing dead nodes using useful 339 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 340 void Compile::update_dead_node_list(Unique_Node_List &useful) { 341 uint max_idx = unique(); 342 VectorSet& useful_node_set = useful.member_set(); 343 344 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 345 // If node with index node_idx is not in useful set, 346 // mark it as dead in dead node list. 347 if (! useful_node_set.test(node_idx) ) { 348 record_dead_node(node_idx); 349 } 350 } 351 } 352 353 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 354 int shift = 0; 355 for (int i = 0; i < inlines->length(); i++) { 356 CallGenerator* cg = inlines->at(i); 357 CallNode* call = cg->call_node(); 358 if (shift > 0) { 359 inlines->at_put(i-shift, cg); 360 } 361 if (!useful.member(call)) { 362 shift++; 363 } 364 } 365 inlines->trunc_to(inlines->length()-shift); 366 } 367 368 // Disconnect all useless nodes by disconnecting those at the boundary. 369 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 370 uint next = 0; 371 while (next < useful.size()) { 372 Node *n = useful.at(next++); 373 if (n->is_SafePoint()) { 374 // We're done with a parsing phase. Replaced nodes are not valid 375 // beyond that point. 376 n->as_SafePoint()->delete_replaced_nodes(); 377 } 378 // Use raw traversal of out edges since this code removes out edges 379 int max = n->outcnt(); 380 for (int j = 0; j < max; ++j) { 381 Node* child = n->raw_out(j); 382 if (! useful.member(child)) { 383 assert(!child->is_top() || child != top(), 384 "If top is cached in Compile object it is in useful list"); 385 // Only need to remove this out-edge to the useless node 386 n->raw_del_out(j); 387 --j; 388 --max; 389 } 390 } 391 if (n->outcnt() == 1 && n->has_special_unique_user()) { 392 record_for_igvn(n->unique_out()); 393 } 394 } 395 // Remove useless macro and predicate opaq nodes 396 for (int i = C->macro_count()-1; i >= 0; i--) { 397 Node* n = C->macro_node(i); 398 if (!useful.member(n)) { 399 remove_macro_node(n); 400 } 401 } 402 // Remove useless CastII nodes with range check dependency 403 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 404 Node* cast = range_check_cast_node(i); 405 if (!useful.member(cast)) { 406 remove_range_check_cast(cast); 407 } 408 } 409 // Remove useless expensive nodes 410 for (int i = C->expensive_count()-1; i >= 0; i--) { 411 Node* n = C->expensive_node(i); 412 if (!useful.member(n)) { 413 remove_expensive_node(n); 414 } 415 } 416 // Remove useless Opaque4 nodes 417 for (int i = opaque4_count() - 1; i >= 0; i--) { 418 Node* opaq = opaque4_node(i); 419 if (!useful.member(opaq)) { 420 remove_opaque4_node(opaq); 421 } 422 } 423 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 424 bs->eliminate_useless_gc_barriers(useful, this); 425 // clean up the late inline lists 426 remove_useless_late_inlines(&_string_late_inlines, useful); 427 remove_useless_late_inlines(&_boxing_late_inlines, useful); 428 remove_useless_late_inlines(&_late_inlines, useful); 429 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 430 } 431 432 //------------------------------frame_size_in_words----------------------------- 433 // frame_slots in units of words 434 int Compile::frame_size_in_words() const { 435 // shift is 0 in LP32 and 1 in LP64 436 const int shift = (LogBytesPerWord - LogBytesPerInt); 437 int words = _frame_slots >> shift; 438 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 439 return words; 440 } 441 442 // To bang the stack of this compiled method we use the stack size 443 // that the interpreter would need in case of a deoptimization. This 444 // removes the need to bang the stack in the deoptimization blob which 445 // in turn simplifies stack overflow handling. 446 int Compile::bang_size_in_bytes() const { 447 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 448 } 449 450 // ============================================================================ 451 //------------------------------CompileWrapper--------------------------------- 452 class CompileWrapper : public StackObj { 453 Compile *const _compile; 454 public: 455 CompileWrapper(Compile* compile); 456 457 ~CompileWrapper(); 458 }; 459 460 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 461 // the Compile* pointer is stored in the current ciEnv: 462 ciEnv* env = compile->env(); 463 assert(env == ciEnv::current(), "must already be a ciEnv active"); 464 assert(env->compiler_data() == NULL, "compile already active?"); 465 env->set_compiler_data(compile); 466 assert(compile == Compile::current(), "sanity"); 467 468 compile->set_type_dict(NULL); 469 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 470 compile->clone_map().set_clone_idx(0); 471 compile->set_type_hwm(NULL); 472 compile->set_type_last_size(0); 473 compile->set_last_tf(NULL, NULL); 474 compile->set_indexSet_arena(NULL); 475 compile->set_indexSet_free_block_list(NULL); 476 compile->init_type_arena(); 477 Type::Initialize(compile); 478 _compile->set_scratch_buffer_blob(NULL); 479 _compile->begin_method(); 480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 481 } 482 CompileWrapper::~CompileWrapper() { 483 _compile->end_method(); 484 if (_compile->scratch_buffer_blob() != NULL) 485 BufferBlob::free(_compile->scratch_buffer_blob()); 486 _compile->env()->set_compiler_data(NULL); 487 } 488 489 490 //----------------------------print_compile_messages--------------------------- 491 void Compile::print_compile_messages() { 492 #ifndef PRODUCT 493 // Check if recompiling 494 if (_subsume_loads == false && PrintOpto) { 495 // Recompiling without allowing machine instructions to subsume loads 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 501 // Recompiling without escape analysis 502 tty->print_cr("*********************************************************"); 503 tty->print_cr("** Bailout: Recompile without escape analysis **"); 504 tty->print_cr("*********************************************************"); 505 } 506 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 507 // Recompiling without boxing elimination 508 tty->print_cr("*********************************************************"); 509 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 510 tty->print_cr("*********************************************************"); 511 } 512 if (C->directive()->BreakAtCompileOption) { 513 // Open the debugger when compiling this method. 514 tty->print("### Breaking when compiling: "); 515 method()->print_short_name(); 516 tty->cr(); 517 BREAKPOINT; 518 } 519 520 if( PrintOpto ) { 521 if (is_osr_compilation()) { 522 tty->print("[OSR]%3d", _compile_id); 523 } else { 524 tty->print("%3d", _compile_id); 525 } 526 } 527 #endif 528 } 529 530 531 //-----------------------init_scratch_buffer_blob------------------------------ 532 // Construct a temporary BufferBlob and cache it for this compile. 533 void Compile::init_scratch_buffer_blob(int const_size) { 534 // If there is already a scratch buffer blob allocated and the 535 // constant section is big enough, use it. Otherwise free the 536 // current and allocate a new one. 537 BufferBlob* blob = scratch_buffer_blob(); 538 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 539 // Use the current blob. 540 } else { 541 if (blob != NULL) { 542 BufferBlob::free(blob); 543 } 544 545 ResourceMark rm; 546 _scratch_const_size = const_size; 547 int size = C2Compiler::initial_code_buffer_size(const_size); 548 blob = BufferBlob::create("Compile::scratch_buffer", size); 549 // Record the buffer blob for next time. 550 set_scratch_buffer_blob(blob); 551 // Have we run out of code space? 552 if (scratch_buffer_blob() == NULL) { 553 // Let CompilerBroker disable further compilations. 554 record_failure("Not enough space for scratch buffer in CodeCache"); 555 return; 556 } 557 } 558 559 // Initialize the relocation buffers 560 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 561 set_scratch_locs_memory(locs_buf); 562 } 563 564 565 //-----------------------scratch_emit_size------------------------------------- 566 // Helper function that computes size by emitting code 567 uint Compile::scratch_emit_size(const Node* n) { 568 // Start scratch_emit_size section. 569 set_in_scratch_emit_size(true); 570 571 // Emit into a trash buffer and count bytes emitted. 572 // This is a pretty expensive way to compute a size, 573 // but it works well enough if seldom used. 574 // All common fixed-size instructions are given a size 575 // method by the AD file. 576 // Note that the scratch buffer blob and locs memory are 577 // allocated at the beginning of the compile task, and 578 // may be shared by several calls to scratch_emit_size. 579 // The allocation of the scratch buffer blob is particularly 580 // expensive, since it has to grab the code cache lock. 581 BufferBlob* blob = this->scratch_buffer_blob(); 582 assert(blob != NULL, "Initialize BufferBlob at start"); 583 assert(blob->size() > MAX_inst_size, "sanity"); 584 relocInfo* locs_buf = scratch_locs_memory(); 585 address blob_begin = blob->content_begin(); 586 address blob_end = (address)locs_buf; 587 assert(blob->contains(blob_end), "sanity"); 588 CodeBuffer buf(blob_begin, blob_end - blob_begin); 589 buf.initialize_consts_size(_scratch_const_size); 590 buf.initialize_stubs_size(MAX_stubs_size); 591 assert(locs_buf != NULL, "sanity"); 592 int lsize = MAX_locs_size / 3; 593 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 594 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 595 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 596 // Mark as scratch buffer. 597 buf.consts()->set_scratch_emit(); 598 buf.insts()->set_scratch_emit(); 599 buf.stubs()->set_scratch_emit(); 600 601 // Do the emission. 602 603 Label fakeL; // Fake label for branch instructions. 604 Label* saveL = NULL; 605 uint save_bnum = 0; 606 bool is_branch = n->is_MachBranch(); 607 if (is_branch) { 608 MacroAssembler masm(&buf); 609 masm.bind(fakeL); 610 n->as_MachBranch()->save_label(&saveL, &save_bnum); 611 n->as_MachBranch()->label_set(&fakeL, 0); 612 } 613 n->emit(buf, this->regalloc()); 614 615 // Emitting into the scratch buffer should not fail 616 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 617 618 if (is_branch) // Restore label. 619 n->as_MachBranch()->label_set(saveL, save_bnum); 620 621 // End scratch_emit_size section. 622 set_in_scratch_emit_size(false); 623 624 return buf.insts_size(); 625 } 626 627 628 // ============================================================================ 629 //------------------------------Compile standard------------------------------- 630 debug_only( int Compile::_debug_idx = 100000; ) 631 632 // Compile a method. entry_bci is -1 for normal compilations and indicates 633 // the continuation bci for on stack replacement. 634 635 636 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 637 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 638 : Phase(Compiler), 639 _compile_id(ci_env->compile_id()), 640 _save_argument_registers(false), 641 _subsume_loads(subsume_loads), 642 _do_escape_analysis(do_escape_analysis), 643 _eliminate_boxing(eliminate_boxing), 644 _method(target), 645 _entry_bci(osr_bci), 646 _stub_function(NULL), 647 _stub_name(NULL), 648 _stub_entry_point(NULL), 649 _max_node_limit(MaxNodeLimit), 650 _orig_pc_slot(0), 651 _orig_pc_slot_offset_in_bytes(0), 652 _inlining_progress(false), 653 _inlining_incrementally(false), 654 _has_reserved_stack_access(target->has_reserved_stack_access()), 655 #ifndef PRODUCT 656 _trace_opto_output(directive->TraceOptoOutputOption), 657 #endif 658 _has_method_handle_invokes(false), 659 _comp_arena(mtCompiler), 660 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 661 _env(ci_env), 662 _directive(directive), 663 _log(ci_env->log()), 664 _failure_reason(NULL), 665 _congraph(NULL), 666 #ifndef PRODUCT 667 _printer(IdealGraphPrinter::printer()), 668 #endif 669 _dead_node_list(comp_arena()), 670 _dead_node_count(0), 671 _node_arena(mtCompiler), 672 _old_arena(mtCompiler), 673 _mach_constant_base_node(NULL), 674 _Compile_types(mtCompiler), 675 _initial_gvn(NULL), 676 _for_igvn(NULL), 677 _warm_calls(NULL), 678 _late_inlines(comp_arena(), 2, 0, NULL), 679 _string_late_inlines(comp_arena(), 2, 0, NULL), 680 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 681 _late_inlines_pos(0), 682 _number_of_mh_late_inlines(0), 683 _print_inlining_stream(NULL), 684 _print_inlining_list(NULL), 685 _print_inlining_idx(0), 686 _print_inlining_output(NULL), 687 _replay_inline_data(NULL), 688 _java_calls(0), 689 _inner_loops(0), 690 _interpreter_frame_size(0), 691 _node_bundling_limit(0), 692 _node_bundling_base(NULL), 693 _code_buffer("Compile::Fill_buffer"), 694 _scratch_const_size(-1), 695 _in_scratch_emit_size(false) 696 #ifndef PRODUCT 697 , _in_dump_cnt(0) 698 #endif 699 { 700 C = this; 701 #ifndef PRODUCT 702 if (_printer != NULL) { 703 _printer->set_compile(this); 704 } 705 #endif 706 CompileWrapper cw(this); 707 708 if (CITimeVerbose) { 709 tty->print(" "); 710 target->holder()->name()->print(); 711 tty->print("."); 712 target->print_short_name(); 713 tty->print(" "); 714 } 715 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 716 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 717 718 #ifndef PRODUCT 719 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 720 if (!print_opto_assembly) { 721 bool print_assembly = directive->PrintAssemblyOption; 722 if (print_assembly && !Disassembler::can_decode()) { 723 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 724 print_opto_assembly = true; 725 } 726 } 727 set_print_assembly(print_opto_assembly); 728 set_parsed_irreducible_loop(false); 729 730 if (directive->ReplayInlineOption) { 731 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 732 } 733 #endif 734 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 735 set_print_intrinsics(directive->PrintIntrinsicsOption); 736 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 737 738 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 739 // Make sure the method being compiled gets its own MDO, 740 // so we can at least track the decompile_count(). 741 // Need MDO to record RTM code generation state. 742 method()->ensure_method_data(); 743 } 744 745 Init(::AliasLevel); 746 747 748 print_compile_messages(); 749 750 _ilt = InlineTree::build_inline_tree_root(); 751 752 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 753 assert(num_alias_types() >= AliasIdxRaw, ""); 754 755 #define MINIMUM_NODE_HASH 1023 756 // Node list that Iterative GVN will start with 757 Unique_Node_List for_igvn(comp_arena()); 758 set_for_igvn(&for_igvn); 759 760 // GVN that will be run immediately on new nodes 761 uint estimated_size = method()->code_size()*4+64; 762 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 763 PhaseGVN gvn(node_arena(), estimated_size); 764 set_initial_gvn(&gvn); 765 766 print_inlining_init(); 767 { // Scope for timing the parser 768 TracePhase tp("parse", &timers[_t_parser]); 769 770 // Put top into the hash table ASAP. 771 initial_gvn()->transform_no_reclaim(top()); 772 773 // Set up tf(), start(), and find a CallGenerator. 774 CallGenerator* cg = NULL; 775 if (is_osr_compilation()) { 776 const TypeTuple *domain = StartOSRNode::osr_domain(); 777 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 778 init_tf(TypeFunc::make(domain, range)); 779 StartNode* s = new StartOSRNode(root(), domain); 780 initial_gvn()->set_type_bottom(s); 781 init_start(s); 782 cg = CallGenerator::for_osr(method(), entry_bci()); 783 } else { 784 // Normal case. 785 init_tf(TypeFunc::make(method())); 786 StartNode* s = new StartNode(root(), tf()->domain()); 787 initial_gvn()->set_type_bottom(s); 788 init_start(s); 789 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 790 // With java.lang.ref.reference.get() we must go through the 791 // intrinsic - even when get() is the root 792 // method of the compile - so that, if necessary, the value in 793 // the referent field of the reference object gets recorded by 794 // the pre-barrier code. 795 cg = find_intrinsic(method(), false); 796 } 797 if (cg == NULL) { 798 float past_uses = method()->interpreter_invocation_count(); 799 float expected_uses = past_uses; 800 cg = CallGenerator::for_inline(method(), expected_uses); 801 } 802 } 803 if (failing()) return; 804 if (cg == NULL) { 805 record_method_not_compilable("cannot parse method"); 806 return; 807 } 808 JVMState* jvms = build_start_state(start(), tf()); 809 if ((jvms = cg->generate(jvms)) == NULL) { 810 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 811 record_method_not_compilable("method parse failed"); 812 } 813 return; 814 } 815 GraphKit kit(jvms); 816 817 if (!kit.stopped()) { 818 // Accept return values, and transfer control we know not where. 819 // This is done by a special, unique ReturnNode bound to root. 820 return_values(kit.jvms()); 821 } 822 823 if (kit.has_exceptions()) { 824 // Any exceptions that escape from this call must be rethrown 825 // to whatever caller is dynamically above us on the stack. 826 // This is done by a special, unique RethrowNode bound to root. 827 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 828 } 829 830 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 831 832 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 833 inline_string_calls(true); 834 } 835 836 if (failing()) return; 837 838 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 839 840 // Remove clutter produced by parsing. 841 if (!failing()) { 842 ResourceMark rm; 843 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 844 } 845 } 846 847 // Note: Large methods are capped off in do_one_bytecode(). 848 if (failing()) return; 849 850 // After parsing, node notes are no longer automagic. 851 // They must be propagated by register_new_node_with_optimizer(), 852 // clone(), or the like. 853 set_default_node_notes(NULL); 854 855 for (;;) { 856 int successes = Inline_Warm(); 857 if (failing()) return; 858 if (successes == 0) break; 859 } 860 861 // Drain the list. 862 Finish_Warm(); 863 #ifndef PRODUCT 864 if (_printer && _printer->should_print(1)) { 865 _printer->print_inlining(); 866 } 867 #endif 868 869 if (failing()) return; 870 NOT_PRODUCT( verify_graph_edges(); ) 871 872 // Now optimize 873 Optimize(); 874 if (failing()) return; 875 NOT_PRODUCT( verify_graph_edges(); ) 876 877 #ifndef PRODUCT 878 if (PrintIdeal) { 879 ttyLocker ttyl; // keep the following output all in one block 880 // This output goes directly to the tty, not the compiler log. 881 // To enable tools to match it up with the compilation activity, 882 // be sure to tag this tty output with the compile ID. 883 if (xtty != NULL) { 884 xtty->head("ideal compile_id='%d'%s", compile_id(), 885 is_osr_compilation() ? " compile_kind='osr'" : 886 ""); 887 } 888 root()->dump(9999); 889 if (xtty != NULL) { 890 xtty->tail("ideal"); 891 } 892 } 893 #endif 894 895 NOT_PRODUCT( verify_barriers(); ) 896 897 // Dump compilation data to replay it. 898 if (directive->DumpReplayOption) { 899 env()->dump_replay_data(_compile_id); 900 } 901 if (directive->DumpInlineOption && (ilt() != NULL)) { 902 env()->dump_inline_data(_compile_id); 903 } 904 905 // Now that we know the size of all the monitors we can add a fixed slot 906 // for the original deopt pc. 907 908 _orig_pc_slot = fixed_slots(); 909 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 910 set_fixed_slots(next_slot); 911 912 // Compute when to use implicit null checks. Used by matching trap based 913 // nodes and NullCheck optimization. 914 set_allowed_deopt_reasons(); 915 916 // Now generate code 917 Code_Gen(); 918 if (failing()) return; 919 920 // Check if we want to skip execution of all compiled code. 921 { 922 #ifndef PRODUCT 923 if (OptoNoExecute) { 924 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 925 return; 926 } 927 #endif 928 TracePhase tp("install_code", &timers[_t_registerMethod]); 929 930 if (is_osr_compilation()) { 931 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 932 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 933 } else { 934 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 935 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 936 } 937 938 env()->register_method(_method, _entry_bci, 939 &_code_offsets, 940 _orig_pc_slot_offset_in_bytes, 941 code_buffer(), 942 frame_size_in_words(), _oop_map_set, 943 &_handler_table, &_inc_table, 944 compiler, 945 has_unsafe_access(), 946 SharedRuntime::is_wide_vector(max_vector_size()), 947 rtm_state() 948 ); 949 950 if (log() != NULL) // Print code cache state into compiler log 951 log()->code_cache_state(); 952 } 953 } 954 955 //------------------------------Compile---------------------------------------- 956 // Compile a runtime stub 957 Compile::Compile( ciEnv* ci_env, 958 TypeFunc_generator generator, 959 address stub_function, 960 const char *stub_name, 961 int is_fancy_jump, 962 bool pass_tls, 963 bool save_arg_registers, 964 bool return_pc, 965 DirectiveSet* directive) 966 : Phase(Compiler), 967 _compile_id(0), 968 _save_argument_registers(save_arg_registers), 969 _subsume_loads(true), 970 _do_escape_analysis(false), 971 _eliminate_boxing(false), 972 _method(NULL), 973 _entry_bci(InvocationEntryBci), 974 _stub_function(stub_function), 975 _stub_name(stub_name), 976 _stub_entry_point(NULL), 977 _max_node_limit(MaxNodeLimit), 978 _orig_pc_slot(0), 979 _orig_pc_slot_offset_in_bytes(0), 980 _inlining_progress(false), 981 _inlining_incrementally(false), 982 _has_reserved_stack_access(false), 983 #ifndef PRODUCT 984 _trace_opto_output(directive->TraceOptoOutputOption), 985 #endif 986 _has_method_handle_invokes(false), 987 _comp_arena(mtCompiler), 988 _env(ci_env), 989 _directive(directive), 990 _log(ci_env->log()), 991 _failure_reason(NULL), 992 _congraph(NULL), 993 #ifndef PRODUCT 994 _printer(NULL), 995 #endif 996 _dead_node_list(comp_arena()), 997 _dead_node_count(0), 998 _node_arena(mtCompiler), 999 _old_arena(mtCompiler), 1000 _mach_constant_base_node(NULL), 1001 _Compile_types(mtCompiler), 1002 _initial_gvn(NULL), 1003 _for_igvn(NULL), 1004 _warm_calls(NULL), 1005 _number_of_mh_late_inlines(0), 1006 _print_inlining_stream(NULL), 1007 _print_inlining_list(NULL), 1008 _print_inlining_idx(0), 1009 _print_inlining_output(NULL), 1010 _replay_inline_data(NULL), 1011 _java_calls(0), 1012 _inner_loops(0), 1013 _interpreter_frame_size(0), 1014 _node_bundling_limit(0), 1015 _node_bundling_base(NULL), 1016 _code_buffer("Compile::Fill_buffer"), 1017 #ifndef PRODUCT 1018 _in_dump_cnt(0), 1019 #endif 1020 _allowed_reasons(0) { 1021 C = this; 1022 1023 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1024 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1025 1026 #ifndef PRODUCT 1027 set_print_assembly(PrintFrameConverterAssembly); 1028 set_parsed_irreducible_loop(false); 1029 #endif 1030 set_has_irreducible_loop(false); // no loops 1031 1032 CompileWrapper cw(this); 1033 Init(/*AliasLevel=*/ 0); 1034 init_tf((*generator)()); 1035 1036 { 1037 // The following is a dummy for the sake of GraphKit::gen_stub 1038 Unique_Node_List for_igvn(comp_arena()); 1039 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1040 PhaseGVN gvn(Thread::current()->resource_area(),255); 1041 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1042 gvn.transform_no_reclaim(top()); 1043 1044 GraphKit kit; 1045 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1046 } 1047 1048 NOT_PRODUCT( verify_graph_edges(); ) 1049 Code_Gen(); 1050 if (failing()) return; 1051 1052 1053 // Entry point will be accessed using compile->stub_entry_point(); 1054 if (code_buffer() == NULL) { 1055 Matcher::soft_match_failure(); 1056 } else { 1057 if (PrintAssembly && (WizardMode || Verbose)) 1058 tty->print_cr("### Stub::%s", stub_name); 1059 1060 if (!failing()) { 1061 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1062 1063 // Make the NMethod 1064 // For now we mark the frame as never safe for profile stackwalking 1065 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1066 code_buffer(), 1067 CodeOffsets::frame_never_safe, 1068 // _code_offsets.value(CodeOffsets::Frame_Complete), 1069 frame_size_in_words(), 1070 _oop_map_set, 1071 save_arg_registers); 1072 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1073 1074 _stub_entry_point = rs->entry_point(); 1075 } 1076 } 1077 } 1078 1079 //------------------------------Init------------------------------------------- 1080 // Prepare for a single compilation 1081 void Compile::Init(int aliaslevel) { 1082 _unique = 0; 1083 _regalloc = NULL; 1084 1085 _tf = NULL; // filled in later 1086 _top = NULL; // cached later 1087 _matcher = NULL; // filled in later 1088 _cfg = NULL; // filled in later 1089 1090 set_24_bit_selection_and_mode(Use24BitFP, false); 1091 1092 _node_note_array = NULL; 1093 _default_node_notes = NULL; 1094 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1095 1096 _immutable_memory = NULL; // filled in at first inquiry 1097 1098 // Globally visible Nodes 1099 // First set TOP to NULL to give safe behavior during creation of RootNode 1100 set_cached_top_node(NULL); 1101 set_root(new RootNode()); 1102 // Now that you have a Root to point to, create the real TOP 1103 set_cached_top_node( new ConNode(Type::TOP) ); 1104 set_recent_alloc(NULL, NULL); 1105 1106 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1107 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1108 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1109 env()->set_dependencies(new Dependencies(env())); 1110 1111 _fixed_slots = 0; 1112 set_has_split_ifs(false); 1113 set_has_loops(has_method() && method()->has_loops()); // first approximation 1114 set_has_stringbuilder(false); 1115 set_has_boxed_value(false); 1116 _trap_can_recompile = false; // no traps emitted yet 1117 _major_progress = true; // start out assuming good things will happen 1118 set_has_unsafe_access(false); 1119 set_max_vector_size(0); 1120 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1121 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1122 set_decompile_count(0); 1123 1124 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1125 set_num_loop_opts(LoopOptsCount); 1126 set_do_inlining(Inline); 1127 set_max_inline_size(MaxInlineSize); 1128 set_freq_inline_size(FreqInlineSize); 1129 set_do_scheduling(OptoScheduling); 1130 set_do_count_invocations(false); 1131 set_do_method_data_update(false); 1132 1133 set_do_vector_loop(false); 1134 1135 if (AllowVectorizeOnDemand) { 1136 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1137 set_do_vector_loop(true); 1138 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());}) 1139 } else if (has_method() && method()->name() != 0 && 1140 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1141 set_do_vector_loop(true); 1142 } 1143 } 1144 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1145 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());}) 1146 1147 set_age_code(has_method() && method()->profile_aging()); 1148 set_rtm_state(NoRTM); // No RTM lock eliding by default 1149 _max_node_limit = _directive->MaxNodeLimitOption; 1150 1151 #if INCLUDE_RTM_OPT 1152 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1153 int rtm_state = method()->method_data()->rtm_state(); 1154 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1155 // Don't generate RTM lock eliding code. 1156 set_rtm_state(NoRTM); 1157 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1158 // Generate RTM lock eliding code without abort ratio calculation code. 1159 set_rtm_state(UseRTM); 1160 } else if (UseRTMDeopt) { 1161 // Generate RTM lock eliding code and include abort ratio calculation 1162 // code if UseRTMDeopt is on. 1163 set_rtm_state(ProfileRTM); 1164 } 1165 } 1166 #endif 1167 if (debug_info()->recording_non_safepoints()) { 1168 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1169 (comp_arena(), 8, 0, NULL)); 1170 set_default_node_notes(Node_Notes::make(this)); 1171 } 1172 1173 // // -- Initialize types before each compile -- 1174 // // Update cached type information 1175 // if( _method && _method->constants() ) 1176 // Type::update_loaded_types(_method, _method->constants()); 1177 1178 // Init alias_type map. 1179 if (!_do_escape_analysis && aliaslevel == 3) 1180 aliaslevel = 2; // No unique types without escape analysis 1181 _AliasLevel = aliaslevel; 1182 const int grow_ats = 16; 1183 _max_alias_types = grow_ats; 1184 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1185 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1186 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1187 { 1188 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1189 } 1190 // Initialize the first few types. 1191 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1192 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1193 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1194 _num_alias_types = AliasIdxRaw+1; 1195 // Zero out the alias type cache. 1196 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1197 // A NULL adr_type hits in the cache right away. Preload the right answer. 1198 probe_alias_cache(NULL)->_index = AliasIdxTop; 1199 1200 _intrinsics = NULL; 1201 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1202 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1203 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1204 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1205 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1206 register_library_intrinsics(); 1207 } 1208 1209 //---------------------------init_start---------------------------------------- 1210 // Install the StartNode on this compile object. 1211 void Compile::init_start(StartNode* s) { 1212 if (failing()) 1213 return; // already failing 1214 assert(s == start(), ""); 1215 } 1216 1217 /** 1218 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1219 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1220 * the ideal graph. 1221 */ 1222 StartNode* Compile::start() const { 1223 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1224 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1225 Node* start = root()->fast_out(i); 1226 if (start->is_Start()) { 1227 return start->as_Start(); 1228 } 1229 } 1230 fatal("Did not find Start node!"); 1231 return NULL; 1232 } 1233 1234 //-------------------------------immutable_memory------------------------------------- 1235 // Access immutable memory 1236 Node* Compile::immutable_memory() { 1237 if (_immutable_memory != NULL) { 1238 return _immutable_memory; 1239 } 1240 StartNode* s = start(); 1241 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1242 Node *p = s->fast_out(i); 1243 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1244 _immutable_memory = p; 1245 return _immutable_memory; 1246 } 1247 } 1248 ShouldNotReachHere(); 1249 return NULL; 1250 } 1251 1252 //----------------------set_cached_top_node------------------------------------ 1253 // Install the cached top node, and make sure Node::is_top works correctly. 1254 void Compile::set_cached_top_node(Node* tn) { 1255 if (tn != NULL) verify_top(tn); 1256 Node* old_top = _top; 1257 _top = tn; 1258 // Calling Node::setup_is_top allows the nodes the chance to adjust 1259 // their _out arrays. 1260 if (_top != NULL) _top->setup_is_top(); 1261 if (old_top != NULL) old_top->setup_is_top(); 1262 assert(_top == NULL || top()->is_top(), ""); 1263 } 1264 1265 #ifdef ASSERT 1266 uint Compile::count_live_nodes_by_graph_walk() { 1267 Unique_Node_List useful(comp_arena()); 1268 // Get useful node list by walking the graph. 1269 identify_useful_nodes(useful); 1270 return useful.size(); 1271 } 1272 1273 void Compile::print_missing_nodes() { 1274 1275 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1276 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1277 return; 1278 } 1279 1280 // This is an expensive function. It is executed only when the user 1281 // specifies VerifyIdealNodeCount option or otherwise knows the 1282 // additional work that needs to be done to identify reachable nodes 1283 // by walking the flow graph and find the missing ones using 1284 // _dead_node_list. 1285 1286 Unique_Node_List useful(comp_arena()); 1287 // Get useful node list by walking the graph. 1288 identify_useful_nodes(useful); 1289 1290 uint l_nodes = C->live_nodes(); 1291 uint l_nodes_by_walk = useful.size(); 1292 1293 if (l_nodes != l_nodes_by_walk) { 1294 if (_log != NULL) { 1295 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1296 _log->stamp(); 1297 _log->end_head(); 1298 } 1299 VectorSet& useful_member_set = useful.member_set(); 1300 int last_idx = l_nodes_by_walk; 1301 for (int i = 0; i < last_idx; i++) { 1302 if (useful_member_set.test(i)) { 1303 if (_dead_node_list.test(i)) { 1304 if (_log != NULL) { 1305 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1306 } 1307 if (PrintIdealNodeCount) { 1308 // Print the log message to tty 1309 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1310 useful.at(i)->dump(); 1311 } 1312 } 1313 } 1314 else if (! _dead_node_list.test(i)) { 1315 if (_log != NULL) { 1316 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1317 } 1318 if (PrintIdealNodeCount) { 1319 // Print the log message to tty 1320 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1321 } 1322 } 1323 } 1324 if (_log != NULL) { 1325 _log->tail("mismatched_nodes"); 1326 } 1327 } 1328 } 1329 void Compile::record_modified_node(Node* n) { 1330 if (_modified_nodes != NULL && !_inlining_incrementally && 1331 n->outcnt() != 0 && !n->is_Con()) { 1332 _modified_nodes->push(n); 1333 } 1334 } 1335 1336 void Compile::remove_modified_node(Node* n) { 1337 if (_modified_nodes != NULL) { 1338 _modified_nodes->remove(n); 1339 } 1340 } 1341 #endif 1342 1343 #ifndef PRODUCT 1344 void Compile::verify_top(Node* tn) const { 1345 if (tn != NULL) { 1346 assert(tn->is_Con(), "top node must be a constant"); 1347 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1348 assert(tn->in(0) != NULL, "must have live top node"); 1349 } 1350 } 1351 #endif 1352 1353 1354 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1355 1356 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1357 guarantee(arr != NULL, ""); 1358 int num_blocks = arr->length(); 1359 if (grow_by < num_blocks) grow_by = num_blocks; 1360 int num_notes = grow_by * _node_notes_block_size; 1361 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1362 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1363 while (num_notes > 0) { 1364 arr->append(notes); 1365 notes += _node_notes_block_size; 1366 num_notes -= _node_notes_block_size; 1367 } 1368 assert(num_notes == 0, "exact multiple, please"); 1369 } 1370 1371 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1372 if (source == NULL || dest == NULL) return false; 1373 1374 if (dest->is_Con()) 1375 return false; // Do not push debug info onto constants. 1376 1377 #ifdef ASSERT 1378 // Leave a bread crumb trail pointing to the original node: 1379 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1380 dest->set_debug_orig(source); 1381 } 1382 #endif 1383 1384 if (node_note_array() == NULL) 1385 return false; // Not collecting any notes now. 1386 1387 // This is a copy onto a pre-existing node, which may already have notes. 1388 // If both nodes have notes, do not overwrite any pre-existing notes. 1389 Node_Notes* source_notes = node_notes_at(source->_idx); 1390 if (source_notes == NULL || source_notes->is_clear()) return false; 1391 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1392 if (dest_notes == NULL || dest_notes->is_clear()) { 1393 return set_node_notes_at(dest->_idx, source_notes); 1394 } 1395 1396 Node_Notes merged_notes = (*source_notes); 1397 // The order of operations here ensures that dest notes will win... 1398 merged_notes.update_from(dest_notes); 1399 return set_node_notes_at(dest->_idx, &merged_notes); 1400 } 1401 1402 1403 //--------------------------allow_range_check_smearing------------------------- 1404 // Gating condition for coalescing similar range checks. 1405 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1406 // single covering check that is at least as strong as any of them. 1407 // If the optimization succeeds, the simplified (strengthened) range check 1408 // will always succeed. If it fails, we will deopt, and then give up 1409 // on the optimization. 1410 bool Compile::allow_range_check_smearing() const { 1411 // If this method has already thrown a range-check, 1412 // assume it was because we already tried range smearing 1413 // and it failed. 1414 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1415 return !already_trapped; 1416 } 1417 1418 1419 //------------------------------flatten_alias_type----------------------------- 1420 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1421 int offset = tj->offset(); 1422 TypePtr::PTR ptr = tj->ptr(); 1423 1424 // Known instance (scalarizable allocation) alias only with itself. 1425 bool is_known_inst = tj->isa_oopptr() != NULL && 1426 tj->is_oopptr()->is_known_instance(); 1427 1428 // Process weird unsafe references. 1429 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1430 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1431 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1432 tj = TypeOopPtr::BOTTOM; 1433 ptr = tj->ptr(); 1434 offset = tj->offset(); 1435 } 1436 1437 // Array pointers need some flattening 1438 const TypeAryPtr *ta = tj->isa_aryptr(); 1439 if (ta && ta->is_stable()) { 1440 // Erase stability property for alias analysis. 1441 tj = ta = ta->cast_to_stable(false); 1442 } 1443 if( ta && is_known_inst ) { 1444 if ( offset != Type::OffsetBot && 1445 offset > arrayOopDesc::length_offset_in_bytes() ) { 1446 offset = Type::OffsetBot; // Flatten constant access into array body only 1447 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1448 } 1449 } else if( ta && _AliasLevel >= 2 ) { 1450 // For arrays indexed by constant indices, we flatten the alias 1451 // space to include all of the array body. Only the header, klass 1452 // and array length can be accessed un-aliased. 1453 if( offset != Type::OffsetBot ) { 1454 if( ta->const_oop() ) { // MethodData* or Method* 1455 offset = Type::OffsetBot; // Flatten constant access into array body 1456 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1457 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1458 // range is OK as-is. 1459 tj = ta = TypeAryPtr::RANGE; 1460 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1461 tj = TypeInstPtr::KLASS; // all klass loads look alike 1462 ta = TypeAryPtr::RANGE; // generic ignored junk 1463 ptr = TypePtr::BotPTR; 1464 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1465 tj = TypeInstPtr::MARK; 1466 ta = TypeAryPtr::RANGE; // generic ignored junk 1467 ptr = TypePtr::BotPTR; 1468 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1469 ta = tj->isa_aryptr(); 1470 } else { // Random constant offset into array body 1471 offset = Type::OffsetBot; // Flatten constant access into array body 1472 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1473 } 1474 } 1475 // Arrays of fixed size alias with arrays of unknown size. 1476 if (ta->size() != TypeInt::POS) { 1477 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1478 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1479 } 1480 // Arrays of known objects become arrays of unknown objects. 1481 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1482 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1483 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1484 } 1485 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1486 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1487 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1488 } 1489 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1490 // cannot be distinguished by bytecode alone. 1491 if (ta->elem() == TypeInt::BOOL) { 1492 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1493 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1494 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1495 } 1496 // During the 2nd round of IterGVN, NotNull castings are removed. 1497 // Make sure the Bottom and NotNull variants alias the same. 1498 // Also, make sure exact and non-exact variants alias the same. 1499 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1500 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1501 } 1502 } 1503 1504 // Oop pointers need some flattening 1505 const TypeInstPtr *to = tj->isa_instptr(); 1506 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1507 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1508 if( ptr == TypePtr::Constant ) { 1509 if (to->klass() != ciEnv::current()->Class_klass() || 1510 offset < k->size_helper() * wordSize) { 1511 // No constant oop pointers (such as Strings); they alias with 1512 // unknown strings. 1513 assert(!is_known_inst, "not scalarizable allocation"); 1514 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1515 } 1516 } else if( is_known_inst ) { 1517 tj = to; // Keep NotNull and klass_is_exact for instance type 1518 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1519 // During the 2nd round of IterGVN, NotNull castings are removed. 1520 // Make sure the Bottom and NotNull variants alias the same. 1521 // Also, make sure exact and non-exact variants alias the same. 1522 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1523 } 1524 if (to->speculative() != NULL) { 1525 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1526 } 1527 // Canonicalize the holder of this field 1528 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1529 // First handle header references such as a LoadKlassNode, even if the 1530 // object's klass is unloaded at compile time (4965979). 1531 if (!is_known_inst) { // Do it only for non-instance types 1532 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1533 } 1534 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1535 to = tj->is_instptr(); 1536 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1537 // Static fields are in the space above the normal instance 1538 // fields in the java.lang.Class instance. 1539 if (to->klass() != ciEnv::current()->Class_klass()) { 1540 to = NULL; 1541 tj = TypeOopPtr::BOTTOM; 1542 offset = tj->offset(); 1543 } 1544 } else { 1545 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1546 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1547 if( is_known_inst ) { 1548 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1549 } else { 1550 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1551 } 1552 } 1553 } 1554 } 1555 1556 // Klass pointers to object array klasses need some flattening 1557 const TypeKlassPtr *tk = tj->isa_klassptr(); 1558 if( tk ) { 1559 // If we are referencing a field within a Klass, we need 1560 // to assume the worst case of an Object. Both exact and 1561 // inexact types must flatten to the same alias class so 1562 // use NotNull as the PTR. 1563 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1564 1565 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1566 TypeKlassPtr::OBJECT->klass(), 1567 offset); 1568 } 1569 1570 ciKlass* klass = tk->klass(); 1571 if( klass->is_obj_array_klass() ) { 1572 ciKlass* k = TypeAryPtr::OOPS->klass(); 1573 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1574 k = TypeInstPtr::BOTTOM->klass(); 1575 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1576 } 1577 1578 // Check for precise loads from the primary supertype array and force them 1579 // to the supertype cache alias index. Check for generic array loads from 1580 // the primary supertype array and also force them to the supertype cache 1581 // alias index. Since the same load can reach both, we need to merge 1582 // these 2 disparate memories into the same alias class. Since the 1583 // primary supertype array is read-only, there's no chance of confusion 1584 // where we bypass an array load and an array store. 1585 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1586 if (offset == Type::OffsetBot || 1587 (offset >= primary_supers_offset && 1588 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1589 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1590 offset = in_bytes(Klass::secondary_super_cache_offset()); 1591 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1592 } 1593 } 1594 1595 // Flatten all Raw pointers together. 1596 if (tj->base() == Type::RawPtr) 1597 tj = TypeRawPtr::BOTTOM; 1598 1599 if (tj->base() == Type::AnyPtr) 1600 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1601 1602 // Flatten all to bottom for now 1603 switch( _AliasLevel ) { 1604 case 0: 1605 tj = TypePtr::BOTTOM; 1606 break; 1607 case 1: // Flatten to: oop, static, field or array 1608 switch (tj->base()) { 1609 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1610 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1611 case Type::AryPtr: // do not distinguish arrays at all 1612 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1613 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1614 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1615 default: ShouldNotReachHere(); 1616 } 1617 break; 1618 case 2: // No collapsing at level 2; keep all splits 1619 case 3: // No collapsing at level 3; keep all splits 1620 break; 1621 default: 1622 Unimplemented(); 1623 } 1624 1625 offset = tj->offset(); 1626 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1627 1628 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1629 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1630 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1631 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1632 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1633 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1634 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1635 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), 1636 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1637 assert( tj->ptr() != TypePtr::TopPTR && 1638 tj->ptr() != TypePtr::AnyNull && 1639 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1640 // assert( tj->ptr() != TypePtr::Constant || 1641 // tj->base() == Type::RawPtr || 1642 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1643 1644 return tj; 1645 } 1646 1647 void Compile::AliasType::Init(int i, const TypePtr* at) { 1648 _index = i; 1649 _adr_type = at; 1650 _field = NULL; 1651 _element = NULL; 1652 _is_rewritable = true; // default 1653 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1654 if (atoop != NULL && atoop->is_known_instance()) { 1655 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1656 _general_index = Compile::current()->get_alias_index(gt); 1657 } else { 1658 _general_index = 0; 1659 } 1660 } 1661 1662 BasicType Compile::AliasType::basic_type() const { 1663 if (element() != NULL) { 1664 const Type* element = adr_type()->is_aryptr()->elem(); 1665 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1666 } if (field() != NULL) { 1667 return field()->layout_type(); 1668 } else { 1669 return T_ILLEGAL; // unknown 1670 } 1671 } 1672 1673 //---------------------------------print_on------------------------------------ 1674 #ifndef PRODUCT 1675 void Compile::AliasType::print_on(outputStream* st) { 1676 if (index() < 10) 1677 st->print("@ <%d> ", index()); 1678 else st->print("@ <%d>", index()); 1679 st->print(is_rewritable() ? " " : " RO"); 1680 int offset = adr_type()->offset(); 1681 if (offset == Type::OffsetBot) 1682 st->print(" +any"); 1683 else st->print(" +%-3d", offset); 1684 st->print(" in "); 1685 adr_type()->dump_on(st); 1686 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1687 if (field() != NULL && tjp) { 1688 if (tjp->klass() != field()->holder() || 1689 tjp->offset() != field()->offset_in_bytes()) { 1690 st->print(" != "); 1691 field()->print(); 1692 st->print(" ***"); 1693 } 1694 } 1695 } 1696 1697 void print_alias_types() { 1698 Compile* C = Compile::current(); 1699 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1700 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1701 C->alias_type(idx)->print_on(tty); 1702 tty->cr(); 1703 } 1704 } 1705 #endif 1706 1707 1708 //----------------------------probe_alias_cache-------------------------------- 1709 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1710 intptr_t key = (intptr_t) adr_type; 1711 key ^= key >> logAliasCacheSize; 1712 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1713 } 1714 1715 1716 //-----------------------------grow_alias_types-------------------------------- 1717 void Compile::grow_alias_types() { 1718 const int old_ats = _max_alias_types; // how many before? 1719 const int new_ats = old_ats; // how many more? 1720 const int grow_ats = old_ats+new_ats; // how many now? 1721 _max_alias_types = grow_ats; 1722 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1723 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1724 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1725 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1726 } 1727 1728 1729 //--------------------------------find_alias_type------------------------------ 1730 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1731 if (_AliasLevel == 0) 1732 return alias_type(AliasIdxBot); 1733 1734 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1735 if (ace->_adr_type == adr_type) { 1736 return alias_type(ace->_index); 1737 } 1738 1739 // Handle special cases. 1740 if (adr_type == NULL) return alias_type(AliasIdxTop); 1741 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1742 1743 // Do it the slow way. 1744 const TypePtr* flat = flatten_alias_type(adr_type); 1745 1746 #ifdef ASSERT 1747 { 1748 ResourceMark rm; 1749 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1750 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1751 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1752 Type::str(adr_type)); 1753 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1754 const TypeOopPtr* foop = flat->is_oopptr(); 1755 // Scalarizable allocations have exact klass always. 1756 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1757 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1758 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1759 Type::str(foop), Type::str(xoop)); 1760 } 1761 } 1762 #endif 1763 1764 int idx = AliasIdxTop; 1765 for (int i = 0; i < num_alias_types(); i++) { 1766 if (alias_type(i)->adr_type() == flat) { 1767 idx = i; 1768 break; 1769 } 1770 } 1771 1772 if (idx == AliasIdxTop) { 1773 if (no_create) return NULL; 1774 // Grow the array if necessary. 1775 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1776 // Add a new alias type. 1777 idx = _num_alias_types++; 1778 _alias_types[idx]->Init(idx, flat); 1779 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1780 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1781 if (flat->isa_instptr()) { 1782 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1783 && flat->is_instptr()->klass() == env()->Class_klass()) 1784 alias_type(idx)->set_rewritable(false); 1785 } 1786 if (flat->isa_aryptr()) { 1787 #ifdef ASSERT 1788 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1789 // (T_BYTE has the weakest alignment and size restrictions...) 1790 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1791 #endif 1792 if (flat->offset() == TypePtr::OffsetBot) { 1793 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1794 } 1795 } 1796 if (flat->isa_klassptr()) { 1797 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1798 alias_type(idx)->set_rewritable(false); 1799 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1800 alias_type(idx)->set_rewritable(false); 1801 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1802 alias_type(idx)->set_rewritable(false); 1803 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1804 alias_type(idx)->set_rewritable(false); 1805 } 1806 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1807 // but the base pointer type is not distinctive enough to identify 1808 // references into JavaThread.) 1809 1810 // Check for final fields. 1811 const TypeInstPtr* tinst = flat->isa_instptr(); 1812 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1813 ciField* field; 1814 if (tinst->const_oop() != NULL && 1815 tinst->klass() == ciEnv::current()->Class_klass() && 1816 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1817 // static field 1818 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1819 field = k->get_field_by_offset(tinst->offset(), true); 1820 } else { 1821 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1822 field = k->get_field_by_offset(tinst->offset(), false); 1823 } 1824 assert(field == NULL || 1825 original_field == NULL || 1826 (field->holder() == original_field->holder() && 1827 field->offset() == original_field->offset() && 1828 field->is_static() == original_field->is_static()), "wrong field?"); 1829 // Set field() and is_rewritable() attributes. 1830 if (field != NULL) alias_type(idx)->set_field(field); 1831 } 1832 } 1833 1834 // Fill the cache for next time. 1835 ace->_adr_type = adr_type; 1836 ace->_index = idx; 1837 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1838 1839 // Might as well try to fill the cache for the flattened version, too. 1840 AliasCacheEntry* face = probe_alias_cache(flat); 1841 if (face->_adr_type == NULL) { 1842 face->_adr_type = flat; 1843 face->_index = idx; 1844 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1845 } 1846 1847 return alias_type(idx); 1848 } 1849 1850 1851 Compile::AliasType* Compile::alias_type(ciField* field) { 1852 const TypeOopPtr* t; 1853 if (field->is_static()) 1854 t = TypeInstPtr::make(field->holder()->java_mirror()); 1855 else 1856 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1857 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1858 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1859 return atp; 1860 } 1861 1862 1863 //------------------------------have_alias_type-------------------------------- 1864 bool Compile::have_alias_type(const TypePtr* adr_type) { 1865 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1866 if (ace->_adr_type == adr_type) { 1867 return true; 1868 } 1869 1870 // Handle special cases. 1871 if (adr_type == NULL) return true; 1872 if (adr_type == TypePtr::BOTTOM) return true; 1873 1874 return find_alias_type(adr_type, true, NULL) != NULL; 1875 } 1876 1877 //-----------------------------must_alias-------------------------------------- 1878 // True if all values of the given address type are in the given alias category. 1879 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1880 if (alias_idx == AliasIdxBot) return true; // the universal category 1881 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1882 if (alias_idx == AliasIdxTop) return false; // the empty category 1883 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1884 1885 // the only remaining possible overlap is identity 1886 int adr_idx = get_alias_index(adr_type); 1887 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1888 assert(adr_idx == alias_idx || 1889 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1890 && adr_type != TypeOopPtr::BOTTOM), 1891 "should not be testing for overlap with an unsafe pointer"); 1892 return adr_idx == alias_idx; 1893 } 1894 1895 //------------------------------can_alias-------------------------------------- 1896 // True if any values of the given address type are in the given alias category. 1897 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1898 if (alias_idx == AliasIdxTop) return false; // the empty category 1899 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1900 if (alias_idx == AliasIdxBot) return true; // the universal category 1901 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1902 1903 // the only remaining possible overlap is identity 1904 int adr_idx = get_alias_index(adr_type); 1905 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1906 return adr_idx == alias_idx; 1907 } 1908 1909 1910 1911 //---------------------------pop_warm_call------------------------------------- 1912 WarmCallInfo* Compile::pop_warm_call() { 1913 WarmCallInfo* wci = _warm_calls; 1914 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1915 return wci; 1916 } 1917 1918 //----------------------------Inline_Warm-------------------------------------- 1919 int Compile::Inline_Warm() { 1920 // If there is room, try to inline some more warm call sites. 1921 // %%% Do a graph index compaction pass when we think we're out of space? 1922 if (!InlineWarmCalls) return 0; 1923 1924 int calls_made_hot = 0; 1925 int room_to_grow = NodeCountInliningCutoff - unique(); 1926 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1927 int amount_grown = 0; 1928 WarmCallInfo* call; 1929 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1930 int est_size = (int)call->size(); 1931 if (est_size > (room_to_grow - amount_grown)) { 1932 // This one won't fit anyway. Get rid of it. 1933 call->make_cold(); 1934 continue; 1935 } 1936 call->make_hot(); 1937 calls_made_hot++; 1938 amount_grown += est_size; 1939 amount_to_grow -= est_size; 1940 } 1941 1942 if (calls_made_hot > 0) set_major_progress(); 1943 return calls_made_hot; 1944 } 1945 1946 1947 //----------------------------Finish_Warm-------------------------------------- 1948 void Compile::Finish_Warm() { 1949 if (!InlineWarmCalls) return; 1950 if (failing()) return; 1951 if (warm_calls() == NULL) return; 1952 1953 // Clean up loose ends, if we are out of space for inlining. 1954 WarmCallInfo* call; 1955 while ((call = pop_warm_call()) != NULL) { 1956 call->make_cold(); 1957 } 1958 } 1959 1960 //---------------------cleanup_loop_predicates----------------------- 1961 // Remove the opaque nodes that protect the predicates so that all unused 1962 // checks and uncommon_traps will be eliminated from the ideal graph 1963 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1964 if (predicate_count()==0) return; 1965 for (int i = predicate_count(); i > 0; i--) { 1966 Node * n = predicate_opaque1_node(i-1); 1967 assert(n->Opcode() == Op_Opaque1, "must be"); 1968 igvn.replace_node(n, n->in(1)); 1969 } 1970 assert(predicate_count()==0, "should be clean!"); 1971 } 1972 1973 void Compile::add_range_check_cast(Node* n) { 1974 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1975 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1976 _range_check_casts->append(n); 1977 } 1978 1979 // Remove all range check dependent CastIINodes. 1980 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1981 for (int i = range_check_cast_count(); i > 0; i--) { 1982 Node* cast = range_check_cast_node(i-1); 1983 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1984 igvn.replace_node(cast, cast->in(1)); 1985 } 1986 assert(range_check_cast_count() == 0, "should be empty"); 1987 } 1988 1989 void Compile::add_opaque4_node(Node* n) { 1990 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1991 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1992 _opaque4_nodes->append(n); 1993 } 1994 1995 // Remove all Opaque4 nodes. 1996 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1997 for (int i = opaque4_count(); i > 0; i--) { 1998 Node* opaq = opaque4_node(i-1); 1999 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2000 igvn.replace_node(opaq, opaq->in(2)); 2001 } 2002 assert(opaque4_count() == 0, "should be empty"); 2003 } 2004 2005 // StringOpts and late inlining of string methods 2006 void Compile::inline_string_calls(bool parse_time) { 2007 { 2008 // remove useless nodes to make the usage analysis simpler 2009 ResourceMark rm; 2010 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2011 } 2012 2013 { 2014 ResourceMark rm; 2015 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2016 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2017 print_method(PHASE_AFTER_STRINGOPTS, 3); 2018 } 2019 2020 // now inline anything that we skipped the first time around 2021 if (!parse_time) { 2022 _late_inlines_pos = _late_inlines.length(); 2023 } 2024 2025 while (_string_late_inlines.length() > 0) { 2026 CallGenerator* cg = _string_late_inlines.pop(); 2027 cg->do_late_inline(); 2028 if (failing()) return; 2029 } 2030 _string_late_inlines.trunc_to(0); 2031 } 2032 2033 // Late inlining of boxing methods 2034 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2035 if (_boxing_late_inlines.length() > 0) { 2036 assert(has_boxed_value(), "inconsistent"); 2037 2038 PhaseGVN* gvn = initial_gvn(); 2039 set_inlining_incrementally(true); 2040 2041 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2042 for_igvn()->clear(); 2043 gvn->replace_with(&igvn); 2044 2045 _late_inlines_pos = _late_inlines.length(); 2046 2047 while (_boxing_late_inlines.length() > 0) { 2048 CallGenerator* cg = _boxing_late_inlines.pop(); 2049 cg->do_late_inline(); 2050 if (failing()) return; 2051 } 2052 _boxing_late_inlines.trunc_to(0); 2053 2054 { 2055 ResourceMark rm; 2056 PhaseRemoveUseless pru(gvn, for_igvn()); 2057 } 2058 2059 igvn = PhaseIterGVN(gvn); 2060 igvn.optimize(); 2061 2062 set_inlining_progress(false); 2063 set_inlining_incrementally(false); 2064 } 2065 } 2066 2067 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2068 assert(IncrementalInline, "incremental inlining should be on"); 2069 PhaseGVN* gvn = initial_gvn(); 2070 2071 set_inlining_progress(false); 2072 for_igvn()->clear(); 2073 gvn->replace_with(&igvn); 2074 2075 { 2076 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2077 int i = 0; 2078 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2079 CallGenerator* cg = _late_inlines.at(i); 2080 _late_inlines_pos = i+1; 2081 cg->do_late_inline(); 2082 if (failing()) return; 2083 } 2084 int j = 0; 2085 for (; i < _late_inlines.length(); i++, j++) { 2086 _late_inlines.at_put(j, _late_inlines.at(i)); 2087 } 2088 _late_inlines.trunc_to(j); 2089 } 2090 2091 { 2092 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2093 ResourceMark rm; 2094 PhaseRemoveUseless pru(gvn, for_igvn()); 2095 } 2096 2097 { 2098 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2099 igvn = PhaseIterGVN(gvn); 2100 } 2101 } 2102 2103 // Perform incremental inlining until bound on number of live nodes is reached 2104 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2105 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2106 2107 PhaseGVN* gvn = initial_gvn(); 2108 2109 set_inlining_incrementally(true); 2110 set_inlining_progress(true); 2111 uint low_live_nodes = 0; 2112 2113 while(inlining_progress() && _late_inlines.length() > 0) { 2114 2115 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2116 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2117 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2118 // PhaseIdealLoop is expensive so we only try it once we are 2119 // out of live nodes and we only try it again if the previous 2120 // helped got the number of nodes down significantly 2121 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2122 if (failing()) return; 2123 low_live_nodes = live_nodes(); 2124 _major_progress = true; 2125 } 2126 2127 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2128 break; 2129 } 2130 } 2131 2132 inline_incrementally_one(igvn); 2133 2134 if (failing()) return; 2135 2136 { 2137 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2138 igvn.optimize(); 2139 } 2140 2141 if (failing()) return; 2142 } 2143 2144 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2145 2146 if (_string_late_inlines.length() > 0) { 2147 assert(has_stringbuilder(), "inconsistent"); 2148 for_igvn()->clear(); 2149 initial_gvn()->replace_with(&igvn); 2150 2151 inline_string_calls(false); 2152 2153 if (failing()) return; 2154 2155 { 2156 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2157 ResourceMark rm; 2158 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2159 } 2160 2161 { 2162 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2163 igvn = PhaseIterGVN(gvn); 2164 igvn.optimize(); 2165 } 2166 } 2167 2168 set_inlining_incrementally(false); 2169 } 2170 2171 2172 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) { 2173 if(loop_opts_cnt > 0) { 2174 debug_only( int cnt = 0; ); 2175 while(major_progress() && (loop_opts_cnt > 0)) { 2176 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2177 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2178 PhaseIdealLoop ideal_loop(igvn, mode); 2179 loop_opts_cnt--; 2180 if (failing()) return false; 2181 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2182 } 2183 } 2184 return true; 2185 } 2186 2187 //------------------------------Optimize--------------------------------------- 2188 // Given a graph, optimize it. 2189 void Compile::Optimize() { 2190 TracePhase tp("optimizer", &timers[_t_optimizer]); 2191 2192 #ifndef PRODUCT 2193 if (_directive->BreakAtCompileOption) { 2194 BREAKPOINT; 2195 } 2196 2197 #endif 2198 2199 #ifdef ASSERT 2200 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2201 bs->verify_gc_barriers(true); 2202 #endif 2203 2204 ResourceMark rm; 2205 int loop_opts_cnt; 2206 2207 print_inlining_reinit(); 2208 2209 NOT_PRODUCT( verify_graph_edges(); ) 2210 2211 print_method(PHASE_AFTER_PARSING); 2212 2213 { 2214 // Iterative Global Value Numbering, including ideal transforms 2215 // Initialize IterGVN with types and values from parse-time GVN 2216 PhaseIterGVN igvn(initial_gvn()); 2217 #ifdef ASSERT 2218 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2219 #endif 2220 { 2221 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2222 igvn.optimize(); 2223 } 2224 2225 if (failing()) return; 2226 2227 print_method(PHASE_ITER_GVN1, 2); 2228 2229 inline_incrementally(igvn); 2230 2231 print_method(PHASE_INCREMENTAL_INLINE, 2); 2232 2233 if (failing()) return; 2234 2235 if (eliminate_boxing()) { 2236 // Inline valueOf() methods now. 2237 inline_boxing_calls(igvn); 2238 2239 if (AlwaysIncrementalInline) { 2240 inline_incrementally(igvn); 2241 } 2242 2243 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2244 2245 if (failing()) return; 2246 } 2247 2248 // Remove the speculative part of types and clean up the graph from 2249 // the extra CastPP nodes whose only purpose is to carry them. Do 2250 // that early so that optimizations are not disrupted by the extra 2251 // CastPP nodes. 2252 remove_speculative_types(igvn); 2253 2254 // No more new expensive nodes will be added to the list from here 2255 // so keep only the actual candidates for optimizations. 2256 cleanup_expensive_nodes(igvn); 2257 2258 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2259 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2260 initial_gvn()->replace_with(&igvn); 2261 for_igvn()->clear(); 2262 Unique_Node_List new_worklist(C->comp_arena()); 2263 { 2264 ResourceMark rm; 2265 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2266 } 2267 set_for_igvn(&new_worklist); 2268 igvn = PhaseIterGVN(initial_gvn()); 2269 igvn.optimize(); 2270 } 2271 2272 // Perform escape analysis 2273 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2274 if (has_loops()) { 2275 // Cleanup graph (remove dead nodes). 2276 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2277 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2278 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2279 if (failing()) return; 2280 } 2281 ConnectionGraph::do_analysis(this, &igvn); 2282 2283 if (failing()) return; 2284 2285 // Optimize out fields loads from scalar replaceable allocations. 2286 igvn.optimize(); 2287 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2288 2289 if (failing()) return; 2290 2291 if (congraph() != NULL && macro_count() > 0) { 2292 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2293 PhaseMacroExpand mexp(igvn); 2294 mexp.eliminate_macro_nodes(); 2295 igvn.set_delay_transform(false); 2296 2297 igvn.optimize(); 2298 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2299 2300 if (failing()) return; 2301 } 2302 } 2303 2304 // Loop transforms on the ideal graph. Range Check Elimination, 2305 // peeling, unrolling, etc. 2306 2307 // Set loop opts counter 2308 loop_opts_cnt = num_loop_opts(); 2309 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2310 { 2311 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2312 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2313 loop_opts_cnt--; 2314 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2315 if (failing()) return; 2316 } 2317 // Loop opts pass if partial peeling occurred in previous pass 2318 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2319 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2320 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2321 loop_opts_cnt--; 2322 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2323 if (failing()) return; 2324 } 2325 // Loop opts pass for loop-unrolling before CCP 2326 if(major_progress() && (loop_opts_cnt > 0)) { 2327 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2328 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2329 loop_opts_cnt--; 2330 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2331 } 2332 if (!failing()) { 2333 // Verify that last round of loop opts produced a valid graph 2334 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2335 PhaseIdealLoop::verify(igvn); 2336 } 2337 } 2338 if (failing()) return; 2339 2340 // Conditional Constant Propagation; 2341 PhaseCCP ccp( &igvn ); 2342 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2343 { 2344 TracePhase tp("ccp", &timers[_t_ccp]); 2345 ccp.do_transform(); 2346 } 2347 print_method(PHASE_CPP1, 2); 2348 2349 assert( true, "Break here to ccp.dump_old2new_map()"); 2350 2351 // Iterative Global Value Numbering, including ideal transforms 2352 { 2353 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2354 igvn = ccp; 2355 igvn.optimize(); 2356 } 2357 2358 print_method(PHASE_ITER_GVN2, 2); 2359 2360 if (failing()) return; 2361 2362 // Loop transforms on the ideal graph. Range Check Elimination, 2363 // peeling, unrolling, etc. 2364 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2365 return; 2366 } 2367 2368 #if INCLUDE_ZGC 2369 if (UseZGC) { 2370 ZBarrierSetC2::find_dominating_barriers(igvn); 2371 } 2372 #endif 2373 2374 if (failing()) return; 2375 2376 // Ensure that major progress is now clear 2377 C->clear_major_progress(); 2378 2379 { 2380 // Verify that all previous optimizations produced a valid graph 2381 // at least to this point, even if no loop optimizations were done. 2382 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2383 PhaseIdealLoop::verify(igvn); 2384 } 2385 2386 if (range_check_cast_count() > 0) { 2387 // No more loop optimizations. Remove all range check dependent CastIINodes. 2388 C->remove_range_check_casts(igvn); 2389 igvn.optimize(); 2390 } 2391 2392 #ifdef ASSERT 2393 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2394 bs->verify_gc_barriers(false); 2395 #endif 2396 2397 { 2398 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2399 PhaseMacroExpand mex(igvn); 2400 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2401 if (mex.expand_macro_nodes()) { 2402 assert(failing(), "must bail out w/ explicit message"); 2403 return; 2404 } 2405 } 2406 2407 if (opaque4_count() > 0) { 2408 C->remove_opaque4_nodes(igvn); 2409 igvn.optimize(); 2410 } 2411 2412 DEBUG_ONLY( _modified_nodes = NULL; ) 2413 } // (End scope of igvn; run destructor if necessary for asserts.) 2414 2415 process_print_inlining(); 2416 // A method with only infinite loops has no edges entering loops from root 2417 { 2418 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2419 if (final_graph_reshaping()) { 2420 assert(failing(), "must bail out w/ explicit message"); 2421 return; 2422 } 2423 } 2424 2425 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2426 } 2427 2428 2429 //------------------------------Code_Gen--------------------------------------- 2430 // Given a graph, generate code for it 2431 void Compile::Code_Gen() { 2432 if (failing()) { 2433 return; 2434 } 2435 2436 // Perform instruction selection. You might think we could reclaim Matcher 2437 // memory PDQ, but actually the Matcher is used in generating spill code. 2438 // Internals of the Matcher (including some VectorSets) must remain live 2439 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2440 // set a bit in reclaimed memory. 2441 2442 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2443 // nodes. Mapping is only valid at the root of each matched subtree. 2444 NOT_PRODUCT( verify_graph_edges(); ) 2445 2446 Matcher matcher; 2447 _matcher = &matcher; 2448 { 2449 TracePhase tp("matcher", &timers[_t_matcher]); 2450 matcher.match(); 2451 } 2452 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2453 // nodes. Mapping is only valid at the root of each matched subtree. 2454 NOT_PRODUCT( verify_graph_edges(); ) 2455 2456 // If you have too many nodes, or if matching has failed, bail out 2457 check_node_count(0, "out of nodes matching instructions"); 2458 if (failing()) { 2459 return; 2460 } 2461 2462 print_method(PHASE_MATCHING, 2); 2463 2464 // Build a proper-looking CFG 2465 PhaseCFG cfg(node_arena(), root(), matcher); 2466 _cfg = &cfg; 2467 { 2468 TracePhase tp("scheduler", &timers[_t_scheduler]); 2469 bool success = cfg.do_global_code_motion(); 2470 if (!success) { 2471 return; 2472 } 2473 2474 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2475 NOT_PRODUCT( verify_graph_edges(); ) 2476 debug_only( cfg.verify(); ) 2477 } 2478 2479 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2480 _regalloc = ®alloc; 2481 { 2482 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2483 // Perform register allocation. After Chaitin, use-def chains are 2484 // no longer accurate (at spill code) and so must be ignored. 2485 // Node->LRG->reg mappings are still accurate. 2486 _regalloc->Register_Allocate(); 2487 2488 // Bail out if the allocator builds too many nodes 2489 if (failing()) { 2490 return; 2491 } 2492 } 2493 2494 // Prior to register allocation we kept empty basic blocks in case the 2495 // the allocator needed a place to spill. After register allocation we 2496 // are not adding any new instructions. If any basic block is empty, we 2497 // can now safely remove it. 2498 { 2499 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2500 cfg.remove_empty_blocks(); 2501 if (do_freq_based_layout()) { 2502 PhaseBlockLayout layout(cfg); 2503 } else { 2504 cfg.set_loop_alignment(); 2505 } 2506 cfg.fixup_flow(); 2507 } 2508 2509 // Apply peephole optimizations 2510 if( OptoPeephole ) { 2511 TracePhase tp("peephole", &timers[_t_peephole]); 2512 PhasePeephole peep( _regalloc, cfg); 2513 peep.do_transform(); 2514 } 2515 2516 // Do late expand if CPU requires this. 2517 if (Matcher::require_postalloc_expand) { 2518 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2519 cfg.postalloc_expand(_regalloc); 2520 } 2521 2522 // Convert Nodes to instruction bits in a buffer 2523 { 2524 TraceTime tp("output", &timers[_t_output], CITime); 2525 Output(); 2526 } 2527 2528 print_method(PHASE_FINAL_CODE); 2529 2530 // He's dead, Jim. 2531 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2532 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2533 } 2534 2535 2536 //------------------------------dump_asm--------------------------------------- 2537 // Dump formatted assembly 2538 #ifndef PRODUCT 2539 void Compile::dump_asm(int *pcs, uint pc_limit) { 2540 bool cut_short = false; 2541 tty->print_cr("#"); 2542 tty->print("# "); _tf->dump(); tty->cr(); 2543 tty->print_cr("#"); 2544 2545 // For all blocks 2546 int pc = 0x0; // Program counter 2547 char starts_bundle = ' '; 2548 _regalloc->dump_frame(); 2549 2550 Node *n = NULL; 2551 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2552 if (VMThread::should_terminate()) { 2553 cut_short = true; 2554 break; 2555 } 2556 Block* block = _cfg->get_block(i); 2557 if (block->is_connector() && !Verbose) { 2558 continue; 2559 } 2560 n = block->head(); 2561 if (pcs && n->_idx < pc_limit) { 2562 tty->print("%3.3x ", pcs[n->_idx]); 2563 } else { 2564 tty->print(" "); 2565 } 2566 block->dump_head(_cfg); 2567 if (block->is_connector()) { 2568 tty->print_cr(" # Empty connector block"); 2569 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2570 tty->print_cr(" # Block is sole successor of call"); 2571 } 2572 2573 // For all instructions 2574 Node *delay = NULL; 2575 for (uint j = 0; j < block->number_of_nodes(); j++) { 2576 if (VMThread::should_terminate()) { 2577 cut_short = true; 2578 break; 2579 } 2580 n = block->get_node(j); 2581 if (valid_bundle_info(n)) { 2582 Bundle* bundle = node_bundling(n); 2583 if (bundle->used_in_unconditional_delay()) { 2584 delay = n; 2585 continue; 2586 } 2587 if (bundle->starts_bundle()) { 2588 starts_bundle = '+'; 2589 } 2590 } 2591 2592 if (WizardMode) { 2593 n->dump(); 2594 } 2595 2596 if( !n->is_Region() && // Dont print in the Assembly 2597 !n->is_Phi() && // a few noisely useless nodes 2598 !n->is_Proj() && 2599 !n->is_MachTemp() && 2600 !n->is_SafePointScalarObject() && 2601 !n->is_Catch() && // Would be nice to print exception table targets 2602 !n->is_MergeMem() && // Not very interesting 2603 !n->is_top() && // Debug info table constants 2604 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2605 ) { 2606 if (pcs && n->_idx < pc_limit) 2607 tty->print("%3.3x", pcs[n->_idx]); 2608 else 2609 tty->print(" "); 2610 tty->print(" %c ", starts_bundle); 2611 starts_bundle = ' '; 2612 tty->print("\t"); 2613 n->format(_regalloc, tty); 2614 tty->cr(); 2615 } 2616 2617 // If we have an instruction with a delay slot, and have seen a delay, 2618 // then back up and print it 2619 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2620 assert(delay != NULL, "no unconditional delay instruction"); 2621 if (WizardMode) delay->dump(); 2622 2623 if (node_bundling(delay)->starts_bundle()) 2624 starts_bundle = '+'; 2625 if (pcs && n->_idx < pc_limit) 2626 tty->print("%3.3x", pcs[n->_idx]); 2627 else 2628 tty->print(" "); 2629 tty->print(" %c ", starts_bundle); 2630 starts_bundle = ' '; 2631 tty->print("\t"); 2632 delay->format(_regalloc, tty); 2633 tty->cr(); 2634 delay = NULL; 2635 } 2636 2637 // Dump the exception table as well 2638 if( n->is_Catch() && (Verbose || WizardMode) ) { 2639 // Print the exception table for this offset 2640 _handler_table.print_subtable_for(pc); 2641 } 2642 } 2643 2644 if (pcs && n->_idx < pc_limit) 2645 tty->print_cr("%3.3x", pcs[n->_idx]); 2646 else 2647 tty->cr(); 2648 2649 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2650 2651 } // End of per-block dump 2652 tty->cr(); 2653 2654 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2655 } 2656 #endif 2657 2658 //------------------------------Final_Reshape_Counts--------------------------- 2659 // This class defines counters to help identify when a method 2660 // may/must be executed using hardware with only 24-bit precision. 2661 struct Final_Reshape_Counts : public StackObj { 2662 int _call_count; // count non-inlined 'common' calls 2663 int _float_count; // count float ops requiring 24-bit precision 2664 int _double_count; // count double ops requiring more precision 2665 int _java_call_count; // count non-inlined 'java' calls 2666 int _inner_loop_count; // count loops which need alignment 2667 VectorSet _visited; // Visitation flags 2668 Node_List _tests; // Set of IfNodes & PCTableNodes 2669 2670 Final_Reshape_Counts() : 2671 _call_count(0), _float_count(0), _double_count(0), 2672 _java_call_count(0), _inner_loop_count(0), 2673 _visited( Thread::current()->resource_area() ) { } 2674 2675 void inc_call_count () { _call_count ++; } 2676 void inc_float_count () { _float_count ++; } 2677 void inc_double_count() { _double_count++; } 2678 void inc_java_call_count() { _java_call_count++; } 2679 void inc_inner_loop_count() { _inner_loop_count++; } 2680 2681 int get_call_count () const { return _call_count ; } 2682 int get_float_count () const { return _float_count ; } 2683 int get_double_count() const { return _double_count; } 2684 int get_java_call_count() const { return _java_call_count; } 2685 int get_inner_loop_count() const { return _inner_loop_count; } 2686 }; 2687 2688 #ifdef ASSERT 2689 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2690 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2691 // Make sure the offset goes inside the instance layout. 2692 return k->contains_field_offset(tp->offset()); 2693 // Note that OffsetBot and OffsetTop are very negative. 2694 } 2695 #endif 2696 2697 // Eliminate trivially redundant StoreCMs and accumulate their 2698 // precedence edges. 2699 void Compile::eliminate_redundant_card_marks(Node* n) { 2700 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2701 if (n->in(MemNode::Address)->outcnt() > 1) { 2702 // There are multiple users of the same address so it might be 2703 // possible to eliminate some of the StoreCMs 2704 Node* mem = n->in(MemNode::Memory); 2705 Node* adr = n->in(MemNode::Address); 2706 Node* val = n->in(MemNode::ValueIn); 2707 Node* prev = n; 2708 bool done = false; 2709 // Walk the chain of StoreCMs eliminating ones that match. As 2710 // long as it's a chain of single users then the optimization is 2711 // safe. Eliminating partially redundant StoreCMs would require 2712 // cloning copies down the other paths. 2713 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2714 if (adr == mem->in(MemNode::Address) && 2715 val == mem->in(MemNode::ValueIn)) { 2716 // redundant StoreCM 2717 if (mem->req() > MemNode::OopStore) { 2718 // Hasn't been processed by this code yet. 2719 n->add_prec(mem->in(MemNode::OopStore)); 2720 } else { 2721 // Already converted to precedence edge 2722 for (uint i = mem->req(); i < mem->len(); i++) { 2723 // Accumulate any precedence edges 2724 if (mem->in(i) != NULL) { 2725 n->add_prec(mem->in(i)); 2726 } 2727 } 2728 // Everything above this point has been processed. 2729 done = true; 2730 } 2731 // Eliminate the previous StoreCM 2732 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2733 assert(mem->outcnt() == 0, "should be dead"); 2734 mem->disconnect_inputs(NULL, this); 2735 } else { 2736 prev = mem; 2737 } 2738 mem = prev->in(MemNode::Memory); 2739 } 2740 } 2741 } 2742 2743 //------------------------------final_graph_reshaping_impl---------------------- 2744 // Implement items 1-5 from final_graph_reshaping below. 2745 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2746 2747 if ( n->outcnt() == 0 ) return; // dead node 2748 uint nop = n->Opcode(); 2749 2750 // Check for 2-input instruction with "last use" on right input. 2751 // Swap to left input. Implements item (2). 2752 if( n->req() == 3 && // two-input instruction 2753 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2754 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2755 n->in(2)->outcnt() == 1 &&// right use IS a last use 2756 !n->in(2)->is_Con() ) { // right use is not a constant 2757 // Check for commutative opcode 2758 switch( nop ) { 2759 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2760 case Op_MaxI: case Op_MinI: 2761 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2762 case Op_AndL: case Op_XorL: case Op_OrL: 2763 case Op_AndI: case Op_XorI: case Op_OrI: { 2764 // Move "last use" input to left by swapping inputs 2765 n->swap_edges(1, 2); 2766 break; 2767 } 2768 default: 2769 break; 2770 } 2771 } 2772 2773 #ifdef ASSERT 2774 if( n->is_Mem() ) { 2775 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2776 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2777 // oop will be recorded in oop map if load crosses safepoint 2778 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2779 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2780 "raw memory operations should have control edge"); 2781 } 2782 if (n->is_MemBar()) { 2783 MemBarNode* mb = n->as_MemBar(); 2784 if (mb->trailing_store() || mb->trailing_load_store()) { 2785 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2786 Node* mem = mb->in(MemBarNode::Precedent); 2787 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2788 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2789 } else if (mb->leading()) { 2790 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2791 } 2792 } 2793 #endif 2794 // Count FPU ops and common calls, implements item (3) 2795 switch( nop ) { 2796 // Count all float operations that may use FPU 2797 case Op_AddF: 2798 case Op_SubF: 2799 case Op_MulF: 2800 case Op_DivF: 2801 case Op_NegF: 2802 case Op_ModF: 2803 case Op_ConvI2F: 2804 case Op_ConF: 2805 case Op_CmpF: 2806 case Op_CmpF3: 2807 // case Op_ConvL2F: // longs are split into 32-bit halves 2808 frc.inc_float_count(); 2809 break; 2810 2811 case Op_ConvF2D: 2812 case Op_ConvD2F: 2813 frc.inc_float_count(); 2814 frc.inc_double_count(); 2815 break; 2816 2817 // Count all double operations that may use FPU 2818 case Op_AddD: 2819 case Op_SubD: 2820 case Op_MulD: 2821 case Op_DivD: 2822 case Op_NegD: 2823 case Op_ModD: 2824 case Op_ConvI2D: 2825 case Op_ConvD2I: 2826 // case Op_ConvL2D: // handled by leaf call 2827 // case Op_ConvD2L: // handled by leaf call 2828 case Op_ConD: 2829 case Op_CmpD: 2830 case Op_CmpD3: 2831 frc.inc_double_count(); 2832 break; 2833 case Op_Opaque1: // Remove Opaque Nodes before matching 2834 case Op_Opaque2: // Remove Opaque Nodes before matching 2835 case Op_Opaque3: 2836 n->subsume_by(n->in(1), this); 2837 break; 2838 case Op_CallStaticJava: 2839 case Op_CallJava: 2840 case Op_CallDynamicJava: 2841 frc.inc_java_call_count(); // Count java call site; 2842 case Op_CallRuntime: 2843 case Op_CallLeaf: 2844 case Op_CallLeafNoFP: { 2845 assert (n->is_Call(), ""); 2846 CallNode *call = n->as_Call(); 2847 // Count call sites where the FP mode bit would have to be flipped. 2848 // Do not count uncommon runtime calls: 2849 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2850 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2851 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2852 frc.inc_call_count(); // Count the call site 2853 } else { // See if uncommon argument is shared 2854 Node *n = call->in(TypeFunc::Parms); 2855 int nop = n->Opcode(); 2856 // Clone shared simple arguments to uncommon calls, item (1). 2857 if (n->outcnt() > 1 && 2858 !n->is_Proj() && 2859 nop != Op_CreateEx && 2860 nop != Op_CheckCastPP && 2861 nop != Op_DecodeN && 2862 nop != Op_DecodeNKlass && 2863 !n->is_Mem() && 2864 !n->is_Phi()) { 2865 Node *x = n->clone(); 2866 call->set_req(TypeFunc::Parms, x); 2867 } 2868 } 2869 break; 2870 } 2871 2872 case Op_StoreD: 2873 case Op_LoadD: 2874 case Op_LoadD_unaligned: 2875 frc.inc_double_count(); 2876 goto handle_mem; 2877 case Op_StoreF: 2878 case Op_LoadF: 2879 frc.inc_float_count(); 2880 goto handle_mem; 2881 2882 case Op_StoreCM: 2883 { 2884 // Convert OopStore dependence into precedence edge 2885 Node* prec = n->in(MemNode::OopStore); 2886 n->del_req(MemNode::OopStore); 2887 n->add_prec(prec); 2888 eliminate_redundant_card_marks(n); 2889 } 2890 2891 // fall through 2892 2893 case Op_StoreB: 2894 case Op_StoreC: 2895 case Op_StorePConditional: 2896 case Op_StoreI: 2897 case Op_StoreL: 2898 case Op_StoreIConditional: 2899 case Op_StoreLConditional: 2900 case Op_CompareAndSwapB: 2901 case Op_CompareAndSwapS: 2902 case Op_CompareAndSwapI: 2903 case Op_CompareAndSwapL: 2904 case Op_CompareAndSwapP: 2905 case Op_CompareAndSwapN: 2906 case Op_WeakCompareAndSwapB: 2907 case Op_WeakCompareAndSwapS: 2908 case Op_WeakCompareAndSwapI: 2909 case Op_WeakCompareAndSwapL: 2910 case Op_WeakCompareAndSwapP: 2911 case Op_WeakCompareAndSwapN: 2912 case Op_CompareAndExchangeB: 2913 case Op_CompareAndExchangeS: 2914 case Op_CompareAndExchangeI: 2915 case Op_CompareAndExchangeL: 2916 case Op_CompareAndExchangeP: 2917 case Op_CompareAndExchangeN: 2918 case Op_GetAndAddS: 2919 case Op_GetAndAddB: 2920 case Op_GetAndAddI: 2921 case Op_GetAndAddL: 2922 case Op_GetAndSetS: 2923 case Op_GetAndSetB: 2924 case Op_GetAndSetI: 2925 case Op_GetAndSetL: 2926 case Op_GetAndSetP: 2927 case Op_GetAndSetN: 2928 case Op_StoreP: 2929 case Op_StoreN: 2930 case Op_StoreNKlass: 2931 case Op_LoadB: 2932 case Op_LoadUB: 2933 case Op_LoadUS: 2934 case Op_LoadI: 2935 case Op_LoadKlass: 2936 case Op_LoadNKlass: 2937 case Op_LoadL: 2938 case Op_LoadL_unaligned: 2939 case Op_LoadPLocked: 2940 case Op_LoadP: 2941 #if INCLUDE_ZGC 2942 case Op_LoadBarrierSlowReg: 2943 case Op_LoadBarrierWeakSlowReg: 2944 #endif 2945 case Op_LoadN: 2946 case Op_LoadRange: 2947 case Op_LoadS: { 2948 handle_mem: 2949 #ifdef ASSERT 2950 if( VerifyOptoOopOffsets ) { 2951 assert( n->is_Mem(), "" ); 2952 MemNode *mem = (MemNode*)n; 2953 // Check to see if address types have grounded out somehow. 2954 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2955 assert( !tp || oop_offset_is_sane(tp), "" ); 2956 } 2957 #endif 2958 break; 2959 } 2960 2961 case Op_AddP: { // Assert sane base pointers 2962 Node *addp = n->in(AddPNode::Address); 2963 assert( !addp->is_AddP() || 2964 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2965 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2966 "Base pointers must match (addp %u)", addp->_idx ); 2967 #ifdef _LP64 2968 if ((UseCompressedOops || UseCompressedClassPointers) && 2969 addp->Opcode() == Op_ConP && 2970 addp == n->in(AddPNode::Base) && 2971 n->in(AddPNode::Offset)->is_Con()) { 2972 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2973 // on the platform and on the compressed oops mode. 2974 // Use addressing with narrow klass to load with offset on x86. 2975 // Some platforms can use the constant pool to load ConP. 2976 // Do this transformation here since IGVN will convert ConN back to ConP. 2977 const Type* t = addp->bottom_type(); 2978 bool is_oop = t->isa_oopptr() != NULL; 2979 bool is_klass = t->isa_klassptr() != NULL; 2980 2981 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 2982 (is_klass && Matcher::const_klass_prefer_decode())) { 2983 Node* nn = NULL; 2984 2985 int op = is_oop ? Op_ConN : Op_ConNKlass; 2986 2987 // Look for existing ConN node of the same exact type. 2988 Node* r = root(); 2989 uint cnt = r->outcnt(); 2990 for (uint i = 0; i < cnt; i++) { 2991 Node* m = r->raw_out(i); 2992 if (m!= NULL && m->Opcode() == op && 2993 m->bottom_type()->make_ptr() == t) { 2994 nn = m; 2995 break; 2996 } 2997 } 2998 if (nn != NULL) { 2999 // Decode a narrow oop to match address 3000 // [R12 + narrow_oop_reg<<3 + offset] 3001 if (is_oop) { 3002 nn = new DecodeNNode(nn, t); 3003 } else { 3004 nn = new DecodeNKlassNode(nn, t); 3005 } 3006 // Check for succeeding AddP which uses the same Base. 3007 // Otherwise we will run into the assertion above when visiting that guy. 3008 for (uint i = 0; i < n->outcnt(); ++i) { 3009 Node *out_i = n->raw_out(i); 3010 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3011 out_i->set_req(AddPNode::Base, nn); 3012 #ifdef ASSERT 3013 for (uint j = 0; j < out_i->outcnt(); ++j) { 3014 Node *out_j = out_i->raw_out(j); 3015 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3016 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3017 } 3018 #endif 3019 } 3020 } 3021 n->set_req(AddPNode::Base, nn); 3022 n->set_req(AddPNode::Address, nn); 3023 if (addp->outcnt() == 0) { 3024 addp->disconnect_inputs(NULL, this); 3025 } 3026 } 3027 } 3028 } 3029 #endif 3030 // platform dependent reshaping of the address expression 3031 reshape_address(n->as_AddP()); 3032 break; 3033 } 3034 3035 case Op_CastPP: { 3036 // Remove CastPP nodes to gain more freedom during scheduling but 3037 // keep the dependency they encode as control or precedence edges 3038 // (if control is set already) on memory operations. Some CastPP 3039 // nodes don't have a control (don't carry a dependency): skip 3040 // those. 3041 if (n->in(0) != NULL) { 3042 ResourceMark rm; 3043 Unique_Node_List wq; 3044 wq.push(n); 3045 for (uint next = 0; next < wq.size(); ++next) { 3046 Node *m = wq.at(next); 3047 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3048 Node* use = m->fast_out(i); 3049 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3050 use->ensure_control_or_add_prec(n->in(0)); 3051 } else { 3052 switch(use->Opcode()) { 3053 case Op_AddP: 3054 case Op_DecodeN: 3055 case Op_DecodeNKlass: 3056 case Op_CheckCastPP: 3057 case Op_CastPP: 3058 wq.push(use); 3059 break; 3060 } 3061 } 3062 } 3063 } 3064 } 3065 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3066 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3067 Node* in1 = n->in(1); 3068 const Type* t = n->bottom_type(); 3069 Node* new_in1 = in1->clone(); 3070 new_in1->as_DecodeN()->set_type(t); 3071 3072 if (!Matcher::narrow_oop_use_complex_address()) { 3073 // 3074 // x86, ARM and friends can handle 2 adds in addressing mode 3075 // and Matcher can fold a DecodeN node into address by using 3076 // a narrow oop directly and do implicit NULL check in address: 3077 // 3078 // [R12 + narrow_oop_reg<<3 + offset] 3079 // NullCheck narrow_oop_reg 3080 // 3081 // On other platforms (Sparc) we have to keep new DecodeN node and 3082 // use it to do implicit NULL check in address: 3083 // 3084 // decode_not_null narrow_oop_reg, base_reg 3085 // [base_reg + offset] 3086 // NullCheck base_reg 3087 // 3088 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3089 // to keep the information to which NULL check the new DecodeN node 3090 // corresponds to use it as value in implicit_null_check(). 3091 // 3092 new_in1->set_req(0, n->in(0)); 3093 } 3094 3095 n->subsume_by(new_in1, this); 3096 if (in1->outcnt() == 0) { 3097 in1->disconnect_inputs(NULL, this); 3098 } 3099 } else { 3100 n->subsume_by(n->in(1), this); 3101 if (n->outcnt() == 0) { 3102 n->disconnect_inputs(NULL, this); 3103 } 3104 } 3105 break; 3106 } 3107 #ifdef _LP64 3108 case Op_CmpP: 3109 // Do this transformation here to preserve CmpPNode::sub() and 3110 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3111 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3112 Node* in1 = n->in(1); 3113 Node* in2 = n->in(2); 3114 if (!in1->is_DecodeNarrowPtr()) { 3115 in2 = in1; 3116 in1 = n->in(2); 3117 } 3118 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3119 3120 Node* new_in2 = NULL; 3121 if (in2->is_DecodeNarrowPtr()) { 3122 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3123 new_in2 = in2->in(1); 3124 } else if (in2->Opcode() == Op_ConP) { 3125 const Type* t = in2->bottom_type(); 3126 if (t == TypePtr::NULL_PTR) { 3127 assert(in1->is_DecodeN(), "compare klass to null?"); 3128 // Don't convert CmpP null check into CmpN if compressed 3129 // oops implicit null check is not generated. 3130 // This will allow to generate normal oop implicit null check. 3131 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3132 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3133 // 3134 // This transformation together with CastPP transformation above 3135 // will generated code for implicit NULL checks for compressed oops. 3136 // 3137 // The original code after Optimize() 3138 // 3139 // LoadN memory, narrow_oop_reg 3140 // decode narrow_oop_reg, base_reg 3141 // CmpP base_reg, NULL 3142 // CastPP base_reg // NotNull 3143 // Load [base_reg + offset], val_reg 3144 // 3145 // after these transformations will be 3146 // 3147 // LoadN memory, narrow_oop_reg 3148 // CmpN narrow_oop_reg, NULL 3149 // decode_not_null narrow_oop_reg, base_reg 3150 // Load [base_reg + offset], val_reg 3151 // 3152 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3153 // since narrow oops can be used in debug info now (see the code in 3154 // final_graph_reshaping_walk()). 3155 // 3156 // At the end the code will be matched to 3157 // on x86: 3158 // 3159 // Load_narrow_oop memory, narrow_oop_reg 3160 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3161 // NullCheck narrow_oop_reg 3162 // 3163 // and on sparc: 3164 // 3165 // Load_narrow_oop memory, narrow_oop_reg 3166 // decode_not_null narrow_oop_reg, base_reg 3167 // Load [base_reg + offset], val_reg 3168 // NullCheck base_reg 3169 // 3170 } else if (t->isa_oopptr()) { 3171 new_in2 = ConNode::make(t->make_narrowoop()); 3172 } else if (t->isa_klassptr()) { 3173 new_in2 = ConNode::make(t->make_narrowklass()); 3174 } 3175 } 3176 if (new_in2 != NULL) { 3177 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3178 n->subsume_by(cmpN, this); 3179 if (in1->outcnt() == 0) { 3180 in1->disconnect_inputs(NULL, this); 3181 } 3182 if (in2->outcnt() == 0) { 3183 in2->disconnect_inputs(NULL, this); 3184 } 3185 } 3186 } 3187 break; 3188 3189 case Op_DecodeN: 3190 case Op_DecodeNKlass: 3191 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3192 // DecodeN could be pinned when it can't be fold into 3193 // an address expression, see the code for Op_CastPP above. 3194 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3195 break; 3196 3197 case Op_EncodeP: 3198 case Op_EncodePKlass: { 3199 Node* in1 = n->in(1); 3200 if (in1->is_DecodeNarrowPtr()) { 3201 n->subsume_by(in1->in(1), this); 3202 } else if (in1->Opcode() == Op_ConP) { 3203 const Type* t = in1->bottom_type(); 3204 if (t == TypePtr::NULL_PTR) { 3205 assert(t->isa_oopptr(), "null klass?"); 3206 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3207 } else if (t->isa_oopptr()) { 3208 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3209 } else if (t->isa_klassptr()) { 3210 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3211 } 3212 } 3213 if (in1->outcnt() == 0) { 3214 in1->disconnect_inputs(NULL, this); 3215 } 3216 break; 3217 } 3218 3219 case Op_Proj: { 3220 if (OptimizeStringConcat) { 3221 ProjNode* p = n->as_Proj(); 3222 if (p->_is_io_use) { 3223 // Separate projections were used for the exception path which 3224 // are normally removed by a late inline. If it wasn't inlined 3225 // then they will hang around and should just be replaced with 3226 // the original one. 3227 Node* proj = NULL; 3228 // Replace with just one 3229 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3230 Node *use = i.get(); 3231 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3232 proj = use; 3233 break; 3234 } 3235 } 3236 assert(proj != NULL, "must be found"); 3237 p->subsume_by(proj, this); 3238 } 3239 } 3240 break; 3241 } 3242 3243 case Op_Phi: 3244 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3245 // The EncodeP optimization may create Phi with the same edges 3246 // for all paths. It is not handled well by Register Allocator. 3247 Node* unique_in = n->in(1); 3248 assert(unique_in != NULL, ""); 3249 uint cnt = n->req(); 3250 for (uint i = 2; i < cnt; i++) { 3251 Node* m = n->in(i); 3252 assert(m != NULL, ""); 3253 if (unique_in != m) 3254 unique_in = NULL; 3255 } 3256 if (unique_in != NULL) { 3257 n->subsume_by(unique_in, this); 3258 } 3259 } 3260 break; 3261 3262 #endif 3263 3264 #ifdef ASSERT 3265 case Op_CastII: 3266 // Verify that all range check dependent CastII nodes were removed. 3267 if (n->isa_CastII()->has_range_check()) { 3268 n->dump(3); 3269 assert(false, "Range check dependent CastII node was not removed"); 3270 } 3271 break; 3272 #endif 3273 3274 case Op_ModI: 3275 if (UseDivMod) { 3276 // Check if a%b and a/b both exist 3277 Node* d = n->find_similar(Op_DivI); 3278 if (d) { 3279 // Replace them with a fused divmod if supported 3280 if (Matcher::has_match_rule(Op_DivModI)) { 3281 DivModINode* divmod = DivModINode::make(n); 3282 d->subsume_by(divmod->div_proj(), this); 3283 n->subsume_by(divmod->mod_proj(), this); 3284 } else { 3285 // replace a%b with a-((a/b)*b) 3286 Node* mult = new MulINode(d, d->in(2)); 3287 Node* sub = new SubINode(d->in(1), mult); 3288 n->subsume_by(sub, this); 3289 } 3290 } 3291 } 3292 break; 3293 3294 case Op_ModL: 3295 if (UseDivMod) { 3296 // Check if a%b and a/b both exist 3297 Node* d = n->find_similar(Op_DivL); 3298 if (d) { 3299 // Replace them with a fused divmod if supported 3300 if (Matcher::has_match_rule(Op_DivModL)) { 3301 DivModLNode* divmod = DivModLNode::make(n); 3302 d->subsume_by(divmod->div_proj(), this); 3303 n->subsume_by(divmod->mod_proj(), this); 3304 } else { 3305 // replace a%b with a-((a/b)*b) 3306 Node* mult = new MulLNode(d, d->in(2)); 3307 Node* sub = new SubLNode(d->in(1), mult); 3308 n->subsume_by(sub, this); 3309 } 3310 } 3311 } 3312 break; 3313 3314 case Op_LoadVector: 3315 case Op_StoreVector: 3316 break; 3317 3318 case Op_AddReductionVI: 3319 case Op_AddReductionVL: 3320 case Op_AddReductionVF: 3321 case Op_AddReductionVD: 3322 case Op_MulReductionVI: 3323 case Op_MulReductionVL: 3324 case Op_MulReductionVF: 3325 case Op_MulReductionVD: 3326 break; 3327 3328 case Op_PackB: 3329 case Op_PackS: 3330 case Op_PackI: 3331 case Op_PackF: 3332 case Op_PackL: 3333 case Op_PackD: 3334 if (n->req()-1 > 2) { 3335 // Replace many operand PackNodes with a binary tree for matching 3336 PackNode* p = (PackNode*) n; 3337 Node* btp = p->binary_tree_pack(1, n->req()); 3338 n->subsume_by(btp, this); 3339 } 3340 break; 3341 case Op_Loop: 3342 case Op_CountedLoop: 3343 case Op_OuterStripMinedLoop: 3344 if (n->as_Loop()->is_inner_loop()) { 3345 frc.inc_inner_loop_count(); 3346 } 3347 n->as_Loop()->verify_strip_mined(0); 3348 break; 3349 case Op_LShiftI: 3350 case Op_RShiftI: 3351 case Op_URShiftI: 3352 case Op_LShiftL: 3353 case Op_RShiftL: 3354 case Op_URShiftL: 3355 if (Matcher::need_masked_shift_count) { 3356 // The cpu's shift instructions don't restrict the count to the 3357 // lower 5/6 bits. We need to do the masking ourselves. 3358 Node* in2 = n->in(2); 3359 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3360 const TypeInt* t = in2->find_int_type(); 3361 if (t != NULL && t->is_con()) { 3362 juint shift = t->get_con(); 3363 if (shift > mask) { // Unsigned cmp 3364 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3365 } 3366 } else { 3367 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3368 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3369 n->set_req(2, shift); 3370 } 3371 } 3372 if (in2->outcnt() == 0) { // Remove dead node 3373 in2->disconnect_inputs(NULL, this); 3374 } 3375 } 3376 break; 3377 case Op_MemBarStoreStore: 3378 case Op_MemBarRelease: 3379 // Break the link with AllocateNode: it is no longer useful and 3380 // confuses register allocation. 3381 if (n->req() > MemBarNode::Precedent) { 3382 n->set_req(MemBarNode::Precedent, top()); 3383 } 3384 break; 3385 case Op_MemBarAcquire: { 3386 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3387 // At parse time, the trailing MemBarAcquire for a volatile load 3388 // is created with an edge to the load. After optimizations, 3389 // that input may be a chain of Phis. If those phis have no 3390 // other use, then the MemBarAcquire keeps them alive and 3391 // register allocation can be confused. 3392 ResourceMark rm; 3393 Unique_Node_List wq; 3394 wq.push(n->in(MemBarNode::Precedent)); 3395 n->set_req(MemBarNode::Precedent, top()); 3396 while (wq.size() > 0) { 3397 Node* m = wq.pop(); 3398 if (m->outcnt() == 0) { 3399 for (uint j = 0; j < m->req(); j++) { 3400 Node* in = m->in(j); 3401 if (in != NULL) { 3402 wq.push(in); 3403 } 3404 } 3405 m->disconnect_inputs(NULL, this); 3406 } 3407 } 3408 } 3409 break; 3410 } 3411 case Op_RangeCheck: { 3412 RangeCheckNode* rc = n->as_RangeCheck(); 3413 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3414 n->subsume_by(iff, this); 3415 frc._tests.push(iff); 3416 break; 3417 } 3418 case Op_ConvI2L: { 3419 if (!Matcher::convi2l_type_required) { 3420 // Code generation on some platforms doesn't need accurate 3421 // ConvI2L types. Widening the type can help remove redundant 3422 // address computations. 3423 n->as_Type()->set_type(TypeLong::INT); 3424 ResourceMark rm; 3425 Node_List wq; 3426 wq.push(n); 3427 for (uint next = 0; next < wq.size(); next++) { 3428 Node *m = wq.at(next); 3429 3430 for(;;) { 3431 // Loop over all nodes with identical inputs edges as m 3432 Node* k = m->find_similar(m->Opcode()); 3433 if (k == NULL) { 3434 break; 3435 } 3436 // Push their uses so we get a chance to remove node made 3437 // redundant 3438 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3439 Node* u = k->fast_out(i); 3440 assert(!wq.contains(u), "shouldn't process one node several times"); 3441 if (u->Opcode() == Op_LShiftL || 3442 u->Opcode() == Op_AddL || 3443 u->Opcode() == Op_SubL || 3444 u->Opcode() == Op_AddP) { 3445 wq.push(u); 3446 } 3447 } 3448 // Replace all nodes with identical edges as m with m 3449 k->subsume_by(m, this); 3450 } 3451 } 3452 } 3453 break; 3454 } 3455 case Op_CmpUL: { 3456 if (!Matcher::has_match_rule(Op_CmpUL)) { 3457 // We don't support unsigned long comparisons. Set 'max_idx_expr' 3458 // to max_julong if < 0 to make the signed comparison fail. 3459 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3460 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3461 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3462 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3463 Node* andl = new AndLNode(orl, remove_sign_mask); 3464 Node* cmp = new CmpLNode(andl, n->in(2)); 3465 n->subsume_by(cmp, this); 3466 } 3467 break; 3468 } 3469 default: 3470 assert( !n->is_Call(), "" ); 3471 assert( !n->is_Mem(), "" ); 3472 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3473 break; 3474 } 3475 3476 // Collect CFG split points 3477 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3478 frc._tests.push(n); 3479 } 3480 } 3481 3482 //------------------------------final_graph_reshaping_walk--------------------- 3483 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3484 // requires that the walk visits a node's inputs before visiting the node. 3485 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3486 ResourceArea *area = Thread::current()->resource_area(); 3487 Unique_Node_List sfpt(area); 3488 3489 frc._visited.set(root->_idx); // first, mark node as visited 3490 uint cnt = root->req(); 3491 Node *n = root; 3492 uint i = 0; 3493 while (true) { 3494 if (i < cnt) { 3495 // Place all non-visited non-null inputs onto stack 3496 Node* m = n->in(i); 3497 ++i; 3498 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3499 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3500 // compute worst case interpreter size in case of a deoptimization 3501 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3502 3503 sfpt.push(m); 3504 } 3505 cnt = m->req(); 3506 nstack.push(n, i); // put on stack parent and next input's index 3507 n = m; 3508 i = 0; 3509 } 3510 } else { 3511 // Now do post-visit work 3512 final_graph_reshaping_impl( n, frc ); 3513 if (nstack.is_empty()) 3514 break; // finished 3515 n = nstack.node(); // Get node from stack 3516 cnt = n->req(); 3517 i = nstack.index(); 3518 nstack.pop(); // Shift to the next node on stack 3519 } 3520 } 3521 3522 // Skip next transformation if compressed oops are not used. 3523 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3524 (!UseCompressedOops && !UseCompressedClassPointers)) 3525 return; 3526 3527 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3528 // It could be done for an uncommon traps or any safepoints/calls 3529 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3530 while (sfpt.size() > 0) { 3531 n = sfpt.pop(); 3532 JVMState *jvms = n->as_SafePoint()->jvms(); 3533 assert(jvms != NULL, "sanity"); 3534 int start = jvms->debug_start(); 3535 int end = n->req(); 3536 bool is_uncommon = (n->is_CallStaticJava() && 3537 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3538 for (int j = start; j < end; j++) { 3539 Node* in = n->in(j); 3540 if (in->is_DecodeNarrowPtr()) { 3541 bool safe_to_skip = true; 3542 if (!is_uncommon ) { 3543 // Is it safe to skip? 3544 for (uint i = 0; i < in->outcnt(); i++) { 3545 Node* u = in->raw_out(i); 3546 if (!u->is_SafePoint() || 3547 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3548 safe_to_skip = false; 3549 } 3550 } 3551 } 3552 if (safe_to_skip) { 3553 n->set_req(j, in->in(1)); 3554 } 3555 if (in->outcnt() == 0) { 3556 in->disconnect_inputs(NULL, this); 3557 } 3558 } 3559 } 3560 } 3561 } 3562 3563 //------------------------------final_graph_reshaping-------------------------- 3564 // Final Graph Reshaping. 3565 // 3566 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3567 // and not commoned up and forced early. Must come after regular 3568 // optimizations to avoid GVN undoing the cloning. Clone constant 3569 // inputs to Loop Phis; these will be split by the allocator anyways. 3570 // Remove Opaque nodes. 3571 // (2) Move last-uses by commutative operations to the left input to encourage 3572 // Intel update-in-place two-address operations and better register usage 3573 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3574 // calls canonicalizing them back. 3575 // (3) Count the number of double-precision FP ops, single-precision FP ops 3576 // and call sites. On Intel, we can get correct rounding either by 3577 // forcing singles to memory (requires extra stores and loads after each 3578 // FP bytecode) or we can set a rounding mode bit (requires setting and 3579 // clearing the mode bit around call sites). The mode bit is only used 3580 // if the relative frequency of single FP ops to calls is low enough. 3581 // This is a key transform for SPEC mpeg_audio. 3582 // (4) Detect infinite loops; blobs of code reachable from above but not 3583 // below. Several of the Code_Gen algorithms fail on such code shapes, 3584 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3585 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3586 // Detection is by looking for IfNodes where only 1 projection is 3587 // reachable from below or CatchNodes missing some targets. 3588 // (5) Assert for insane oop offsets in debug mode. 3589 3590 bool Compile::final_graph_reshaping() { 3591 // an infinite loop may have been eliminated by the optimizer, 3592 // in which case the graph will be empty. 3593 if (root()->req() == 1) { 3594 record_method_not_compilable("trivial infinite loop"); 3595 return true; 3596 } 3597 3598 // Expensive nodes have their control input set to prevent the GVN 3599 // from freely commoning them. There's no GVN beyond this point so 3600 // no need to keep the control input. We want the expensive nodes to 3601 // be freely moved to the least frequent code path by gcm. 3602 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3603 for (int i = 0; i < expensive_count(); i++) { 3604 _expensive_nodes->at(i)->set_req(0, NULL); 3605 } 3606 3607 Final_Reshape_Counts frc; 3608 3609 // Visit everybody reachable! 3610 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3611 Node_Stack nstack(live_nodes() >> 1); 3612 final_graph_reshaping_walk(nstack, root(), frc); 3613 3614 // Check for unreachable (from below) code (i.e., infinite loops). 3615 for( uint i = 0; i < frc._tests.size(); i++ ) { 3616 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3617 // Get number of CFG targets. 3618 // Note that PCTables include exception targets after calls. 3619 uint required_outcnt = n->required_outcnt(); 3620 if (n->outcnt() != required_outcnt) { 3621 // Check for a few special cases. Rethrow Nodes never take the 3622 // 'fall-thru' path, so expected kids is 1 less. 3623 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3624 if (n->in(0)->in(0)->is_Call()) { 3625 CallNode *call = n->in(0)->in(0)->as_Call(); 3626 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3627 required_outcnt--; // Rethrow always has 1 less kid 3628 } else if (call->req() > TypeFunc::Parms && 3629 call->is_CallDynamicJava()) { 3630 // Check for null receiver. In such case, the optimizer has 3631 // detected that the virtual call will always result in a null 3632 // pointer exception. The fall-through projection of this CatchNode 3633 // will not be populated. 3634 Node *arg0 = call->in(TypeFunc::Parms); 3635 if (arg0->is_Type() && 3636 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3637 required_outcnt--; 3638 } 3639 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3640 call->req() > TypeFunc::Parms+1 && 3641 call->is_CallStaticJava()) { 3642 // Check for negative array length. In such case, the optimizer has 3643 // detected that the allocation attempt will always result in an 3644 // exception. There is no fall-through projection of this CatchNode . 3645 Node *arg1 = call->in(TypeFunc::Parms+1); 3646 if (arg1->is_Type() && 3647 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3648 required_outcnt--; 3649 } 3650 } 3651 } 3652 } 3653 // Recheck with a better notion of 'required_outcnt' 3654 if (n->outcnt() != required_outcnt) { 3655 record_method_not_compilable("malformed control flow"); 3656 return true; // Not all targets reachable! 3657 } 3658 } 3659 // Check that I actually visited all kids. Unreached kids 3660 // must be infinite loops. 3661 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3662 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3663 record_method_not_compilable("infinite loop"); 3664 return true; // Found unvisited kid; must be unreach 3665 } 3666 3667 // Here so verification code in final_graph_reshaping_walk() 3668 // always see an OuterStripMinedLoopEnd 3669 if (n->is_OuterStripMinedLoopEnd()) { 3670 IfNode* init_iff = n->as_If(); 3671 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3672 n->subsume_by(iff, this); 3673 } 3674 } 3675 3676 // If original bytecodes contained a mixture of floats and doubles 3677 // check if the optimizer has made it homogenous, item (3). 3678 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3679 frc.get_float_count() > 32 && 3680 frc.get_double_count() == 0 && 3681 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3682 set_24_bit_selection_and_mode( false, true ); 3683 } 3684 3685 set_java_calls(frc.get_java_call_count()); 3686 set_inner_loops(frc.get_inner_loop_count()); 3687 3688 // No infinite loops, no reason to bail out. 3689 return false; 3690 } 3691 3692 //-----------------------------too_many_traps---------------------------------- 3693 // Report if there are too many traps at the current method and bci. 3694 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3695 bool Compile::too_many_traps(ciMethod* method, 3696 int bci, 3697 Deoptimization::DeoptReason reason) { 3698 ciMethodData* md = method->method_data(); 3699 if (md->is_empty()) { 3700 // Assume the trap has not occurred, or that it occurred only 3701 // because of a transient condition during start-up in the interpreter. 3702 return false; 3703 } 3704 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3705 if (md->has_trap_at(bci, m, reason) != 0) { 3706 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3707 // Also, if there are multiple reasons, or if there is no per-BCI record, 3708 // assume the worst. 3709 if (log()) 3710 log()->elem("observe trap='%s' count='%d'", 3711 Deoptimization::trap_reason_name(reason), 3712 md->trap_count(reason)); 3713 return true; 3714 } else { 3715 // Ignore method/bci and see if there have been too many globally. 3716 return too_many_traps(reason, md); 3717 } 3718 } 3719 3720 // Less-accurate variant which does not require a method and bci. 3721 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3722 ciMethodData* logmd) { 3723 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3724 // Too many traps globally. 3725 // Note that we use cumulative trap_count, not just md->trap_count. 3726 if (log()) { 3727 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3728 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3729 Deoptimization::trap_reason_name(reason), 3730 mcount, trap_count(reason)); 3731 } 3732 return true; 3733 } else { 3734 // The coast is clear. 3735 return false; 3736 } 3737 } 3738 3739 //--------------------------too_many_recompiles-------------------------------- 3740 // Report if there are too many recompiles at the current method and bci. 3741 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3742 // Is not eager to return true, since this will cause the compiler to use 3743 // Action_none for a trap point, to avoid too many recompilations. 3744 bool Compile::too_many_recompiles(ciMethod* method, 3745 int bci, 3746 Deoptimization::DeoptReason reason) { 3747 ciMethodData* md = method->method_data(); 3748 if (md->is_empty()) { 3749 // Assume the trap has not occurred, or that it occurred only 3750 // because of a transient condition during start-up in the interpreter. 3751 return false; 3752 } 3753 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3754 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3755 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3756 Deoptimization::DeoptReason per_bc_reason 3757 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3758 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3759 if ((per_bc_reason == Deoptimization::Reason_none 3760 || md->has_trap_at(bci, m, reason) != 0) 3761 // The trap frequency measure we care about is the recompile count: 3762 && md->trap_recompiled_at(bci, m) 3763 && md->overflow_recompile_count() >= bc_cutoff) { 3764 // Do not emit a trap here if it has already caused recompilations. 3765 // Also, if there are multiple reasons, or if there is no per-BCI record, 3766 // assume the worst. 3767 if (log()) 3768 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3769 Deoptimization::trap_reason_name(reason), 3770 md->trap_count(reason), 3771 md->overflow_recompile_count()); 3772 return true; 3773 } else if (trap_count(reason) != 0 3774 && decompile_count() >= m_cutoff) { 3775 // Too many recompiles globally, and we have seen this sort of trap. 3776 // Use cumulative decompile_count, not just md->decompile_count. 3777 if (log()) 3778 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3779 Deoptimization::trap_reason_name(reason), 3780 md->trap_count(reason), trap_count(reason), 3781 md->decompile_count(), decompile_count()); 3782 return true; 3783 } else { 3784 // The coast is clear. 3785 return false; 3786 } 3787 } 3788 3789 // Compute when not to trap. Used by matching trap based nodes and 3790 // NullCheck optimization. 3791 void Compile::set_allowed_deopt_reasons() { 3792 _allowed_reasons = 0; 3793 if (is_method_compilation()) { 3794 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3795 assert(rs < BitsPerInt, "recode bit map"); 3796 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3797 _allowed_reasons |= nth_bit(rs); 3798 } 3799 } 3800 } 3801 } 3802 3803 #ifndef PRODUCT 3804 //------------------------------verify_graph_edges--------------------------- 3805 // Walk the Graph and verify that there is a one-to-one correspondence 3806 // between Use-Def edges and Def-Use edges in the graph. 3807 void Compile::verify_graph_edges(bool no_dead_code) { 3808 if (VerifyGraphEdges) { 3809 ResourceArea *area = Thread::current()->resource_area(); 3810 Unique_Node_List visited(area); 3811 // Call recursive graph walk to check edges 3812 _root->verify_edges(visited); 3813 if (no_dead_code) { 3814 // Now make sure that no visited node is used by an unvisited node. 3815 bool dead_nodes = false; 3816 Unique_Node_List checked(area); 3817 while (visited.size() > 0) { 3818 Node* n = visited.pop(); 3819 checked.push(n); 3820 for (uint i = 0; i < n->outcnt(); i++) { 3821 Node* use = n->raw_out(i); 3822 if (checked.member(use)) continue; // already checked 3823 if (visited.member(use)) continue; // already in the graph 3824 if (use->is_Con()) continue; // a dead ConNode is OK 3825 // At this point, we have found a dead node which is DU-reachable. 3826 if (!dead_nodes) { 3827 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3828 dead_nodes = true; 3829 } 3830 use->dump(2); 3831 tty->print_cr("---"); 3832 checked.push(use); // No repeats; pretend it is now checked. 3833 } 3834 } 3835 assert(!dead_nodes, "using nodes must be reachable from root"); 3836 } 3837 } 3838 } 3839 3840 // Verify GC barriers consistency 3841 // Currently supported: 3842 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3843 void Compile::verify_barriers() { 3844 #if INCLUDE_G1GC 3845 if (UseG1GC) { 3846 // Verify G1 pre-barriers 3847 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3848 3849 ResourceArea *area = Thread::current()->resource_area(); 3850 Unique_Node_List visited(area); 3851 Node_List worklist(area); 3852 // We're going to walk control flow backwards starting from the Root 3853 worklist.push(_root); 3854 while (worklist.size() > 0) { 3855 Node* x = worklist.pop(); 3856 if (x == NULL || x == top()) continue; 3857 if (visited.member(x)) { 3858 continue; 3859 } else { 3860 visited.push(x); 3861 } 3862 3863 if (x->is_Region()) { 3864 for (uint i = 1; i < x->req(); i++) { 3865 worklist.push(x->in(i)); 3866 } 3867 } else { 3868 worklist.push(x->in(0)); 3869 // We are looking for the pattern: 3870 // /->ThreadLocal 3871 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3872 // \->ConI(0) 3873 // We want to verify that the If and the LoadB have the same control 3874 // See GraphKit::g1_write_barrier_pre() 3875 if (x->is_If()) { 3876 IfNode *iff = x->as_If(); 3877 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3878 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3879 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3880 && cmp->in(1)->is_Load()) { 3881 LoadNode* load = cmp->in(1)->as_Load(); 3882 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3883 && load->in(2)->in(3)->is_Con() 3884 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3885 3886 Node* if_ctrl = iff->in(0); 3887 Node* load_ctrl = load->in(0); 3888 3889 if (if_ctrl != load_ctrl) { 3890 // Skip possible CProj->NeverBranch in infinite loops 3891 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3892 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3893 if_ctrl = if_ctrl->in(0)->in(0); 3894 } 3895 } 3896 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3897 } 3898 } 3899 } 3900 } 3901 } 3902 } 3903 } 3904 #endif 3905 } 3906 3907 #endif 3908 3909 // The Compile object keeps track of failure reasons separately from the ciEnv. 3910 // This is required because there is not quite a 1-1 relation between the 3911 // ciEnv and its compilation task and the Compile object. Note that one 3912 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3913 // to backtrack and retry without subsuming loads. Other than this backtracking 3914 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3915 // by the logic in C2Compiler. 3916 void Compile::record_failure(const char* reason) { 3917 if (log() != NULL) { 3918 log()->elem("failure reason='%s' phase='compile'", reason); 3919 } 3920 if (_failure_reason == NULL) { 3921 // Record the first failure reason. 3922 _failure_reason = reason; 3923 } 3924 3925 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3926 C->print_method(PHASE_FAILURE); 3927 } 3928 _root = NULL; // flush the graph, too 3929 } 3930 3931 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3932 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3933 _phase_name(name), _dolog(CITimeVerbose) 3934 { 3935 if (_dolog) { 3936 C = Compile::current(); 3937 _log = C->log(); 3938 } else { 3939 C = NULL; 3940 _log = NULL; 3941 } 3942 if (_log != NULL) { 3943 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3944 _log->stamp(); 3945 _log->end_head(); 3946 } 3947 } 3948 3949 Compile::TracePhase::~TracePhase() { 3950 3951 C = Compile::current(); 3952 if (_dolog) { 3953 _log = C->log(); 3954 } else { 3955 _log = NULL; 3956 } 3957 3958 #ifdef ASSERT 3959 if (PrintIdealNodeCount) { 3960 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3961 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3962 } 3963 3964 if (VerifyIdealNodeCount) { 3965 Compile::current()->print_missing_nodes(); 3966 } 3967 #endif 3968 3969 if (_log != NULL) { 3970 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3971 } 3972 } 3973 3974 //============================================================================= 3975 // Two Constant's are equal when the type and the value are equal. 3976 bool Compile::Constant::operator==(const Constant& other) { 3977 if (type() != other.type() ) return false; 3978 if (can_be_reused() != other.can_be_reused()) return false; 3979 // For floating point values we compare the bit pattern. 3980 switch (type()) { 3981 case T_INT: 3982 case T_FLOAT: return (_v._value.i == other._v._value.i); 3983 case T_LONG: 3984 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3985 case T_OBJECT: 3986 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3987 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3988 case T_METADATA: return (_v._metadata == other._v._metadata); 3989 default: ShouldNotReachHere(); return false; 3990 } 3991 } 3992 3993 static int type_to_size_in_bytes(BasicType t) { 3994 switch (t) { 3995 case T_INT: return sizeof(jint ); 3996 case T_LONG: return sizeof(jlong ); 3997 case T_FLOAT: return sizeof(jfloat ); 3998 case T_DOUBLE: return sizeof(jdouble); 3999 case T_METADATA: return sizeof(Metadata*); 4000 // We use T_VOID as marker for jump-table entries (labels) which 4001 // need an internal word relocation. 4002 case T_VOID: 4003 case T_ADDRESS: 4004 case T_OBJECT: return sizeof(jobject); 4005 default: 4006 ShouldNotReachHere(); 4007 return -1; 4008 } 4009 } 4010 4011 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4012 // sort descending 4013 if (a->freq() > b->freq()) return -1; 4014 if (a->freq() < b->freq()) return 1; 4015 return 0; 4016 } 4017 4018 void Compile::ConstantTable::calculate_offsets_and_size() { 4019 // First, sort the array by frequencies. 4020 _constants.sort(qsort_comparator); 4021 4022 #ifdef ASSERT 4023 // Make sure all jump-table entries were sorted to the end of the 4024 // array (they have a negative frequency). 4025 bool found_void = false; 4026 for (int i = 0; i < _constants.length(); i++) { 4027 Constant con = _constants.at(i); 4028 if (con.type() == T_VOID) 4029 found_void = true; // jump-tables 4030 else 4031 assert(!found_void, "wrong sorting"); 4032 } 4033 #endif 4034 4035 int offset = 0; 4036 for (int i = 0; i < _constants.length(); i++) { 4037 Constant* con = _constants.adr_at(i); 4038 4039 // Align offset for type. 4040 int typesize = type_to_size_in_bytes(con->type()); 4041 offset = align_up(offset, typesize); 4042 con->set_offset(offset); // set constant's offset 4043 4044 if (con->type() == T_VOID) { 4045 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4046 offset = offset + typesize * n->outcnt(); // expand jump-table 4047 } else { 4048 offset = offset + typesize; 4049 } 4050 } 4051 4052 // Align size up to the next section start (which is insts; see 4053 // CodeBuffer::align_at_start). 4054 assert(_size == -1, "already set?"); 4055 _size = align_up(offset, (int)CodeEntryAlignment); 4056 } 4057 4058 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4059 MacroAssembler _masm(&cb); 4060 for (int i = 0; i < _constants.length(); i++) { 4061 Constant con = _constants.at(i); 4062 address constant_addr = NULL; 4063 switch (con.type()) { 4064 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4065 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4066 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4067 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4068 case T_OBJECT: { 4069 jobject obj = con.get_jobject(); 4070 int oop_index = _masm.oop_recorder()->find_index(obj); 4071 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4072 break; 4073 } 4074 case T_ADDRESS: { 4075 address addr = (address) con.get_jobject(); 4076 constant_addr = _masm.address_constant(addr); 4077 break; 4078 } 4079 // We use T_VOID as marker for jump-table entries (labels) which 4080 // need an internal word relocation. 4081 case T_VOID: { 4082 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4083 // Fill the jump-table with a dummy word. The real value is 4084 // filled in later in fill_jump_table. 4085 address dummy = (address) n; 4086 constant_addr = _masm.address_constant(dummy); 4087 // Expand jump-table 4088 for (uint i = 1; i < n->outcnt(); i++) { 4089 address temp_addr = _masm.address_constant(dummy + i); 4090 assert(temp_addr, "consts section too small"); 4091 } 4092 break; 4093 } 4094 case T_METADATA: { 4095 Metadata* obj = con.get_metadata(); 4096 int metadata_index = _masm.oop_recorder()->find_index(obj); 4097 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4098 break; 4099 } 4100 default: ShouldNotReachHere(); 4101 } 4102 assert(constant_addr, "consts section too small"); 4103 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4104 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4105 } 4106 } 4107 4108 int Compile::ConstantTable::find_offset(Constant& con) const { 4109 int idx = _constants.find(con); 4110 guarantee(idx != -1, "constant must be in constant table"); 4111 int offset = _constants.at(idx).offset(); 4112 guarantee(offset != -1, "constant table not emitted yet?"); 4113 return offset; 4114 } 4115 4116 void Compile::ConstantTable::add(Constant& con) { 4117 if (con.can_be_reused()) { 4118 int idx = _constants.find(con); 4119 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4120 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4121 return; 4122 } 4123 } 4124 (void) _constants.append(con); 4125 } 4126 4127 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4128 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4129 Constant con(type, value, b->_freq); 4130 add(con); 4131 return con; 4132 } 4133 4134 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4135 Constant con(metadata); 4136 add(con); 4137 return con; 4138 } 4139 4140 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4141 jvalue value; 4142 BasicType type = oper->type()->basic_type(); 4143 switch (type) { 4144 case T_LONG: value.j = oper->constantL(); break; 4145 case T_FLOAT: value.f = oper->constantF(); break; 4146 case T_DOUBLE: value.d = oper->constantD(); break; 4147 case T_OBJECT: 4148 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4149 case T_METADATA: return add((Metadata*)oper->constant()); break; 4150 default: guarantee(false, "unhandled type: %s", type2name(type)); 4151 } 4152 return add(n, type, value); 4153 } 4154 4155 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4156 jvalue value; 4157 // We can use the node pointer here to identify the right jump-table 4158 // as this method is called from Compile::Fill_buffer right before 4159 // the MachNodes are emitted and the jump-table is filled (means the 4160 // MachNode pointers do not change anymore). 4161 value.l = (jobject) n; 4162 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4163 add(con); 4164 return con; 4165 } 4166 4167 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4168 // If called from Compile::scratch_emit_size do nothing. 4169 if (Compile::current()->in_scratch_emit_size()) return; 4170 4171 assert(labels.is_nonempty(), "must be"); 4172 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4173 4174 // Since MachConstantNode::constant_offset() also contains 4175 // table_base_offset() we need to subtract the table_base_offset() 4176 // to get the plain offset into the constant table. 4177 int offset = n->constant_offset() - table_base_offset(); 4178 4179 MacroAssembler _masm(&cb); 4180 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4181 4182 for (uint i = 0; i < n->outcnt(); i++) { 4183 address* constant_addr = &jump_table_base[i]; 4184 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)); 4185 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4186 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4187 } 4188 } 4189 4190 //----------------------------static_subtype_check----------------------------- 4191 // Shortcut important common cases when superklass is exact: 4192 // (0) superklass is java.lang.Object (can occur in reflective code) 4193 // (1) subklass is already limited to a subtype of superklass => always ok 4194 // (2) subklass does not overlap with superklass => always fail 4195 // (3) superklass has NO subtypes and we can check with a simple compare. 4196 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4197 if (StressReflectiveCode) { 4198 return SSC_full_test; // Let caller generate the general case. 4199 } 4200 4201 if (superk == env()->Object_klass()) { 4202 return SSC_always_true; // (0) this test cannot fail 4203 } 4204 4205 ciType* superelem = superk; 4206 if (superelem->is_array_klass()) 4207 superelem = superelem->as_array_klass()->base_element_type(); 4208 4209 if (!subk->is_interface()) { // cannot trust static interface types yet 4210 if (subk->is_subtype_of(superk)) { 4211 return SSC_always_true; // (1) false path dead; no dynamic test needed 4212 } 4213 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4214 !superk->is_subtype_of(subk)) { 4215 return SSC_always_false; 4216 } 4217 } 4218 4219 // If casting to an instance klass, it must have no subtypes 4220 if (superk->is_interface()) { 4221 // Cannot trust interfaces yet. 4222 // %%% S.B. superk->nof_implementors() == 1 4223 } else if (superelem->is_instance_klass()) { 4224 ciInstanceKlass* ik = superelem->as_instance_klass(); 4225 if (!ik->has_subklass() && !ik->is_interface()) { 4226 if (!ik->is_final()) { 4227 // Add a dependency if there is a chance of a later subclass. 4228 dependencies()->assert_leaf_type(ik); 4229 } 4230 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4231 } 4232 } else { 4233 // A primitive array type has no subtypes. 4234 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4235 } 4236 4237 return SSC_full_test; 4238 } 4239 4240 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4241 #ifdef _LP64 4242 // The scaled index operand to AddP must be a clean 64-bit value. 4243 // Java allows a 32-bit int to be incremented to a negative 4244 // value, which appears in a 64-bit register as a large 4245 // positive number. Using that large positive number as an 4246 // operand in pointer arithmetic has bad consequences. 4247 // On the other hand, 32-bit overflow is rare, and the possibility 4248 // can often be excluded, if we annotate the ConvI2L node with 4249 // a type assertion that its value is known to be a small positive 4250 // number. (The prior range check has ensured this.) 4251 // This assertion is used by ConvI2LNode::Ideal. 4252 int index_max = max_jint - 1; // array size is max_jint, index is one less 4253 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4254 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4255 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4256 #endif 4257 return idx; 4258 } 4259 4260 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4261 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4262 if (ctrl != NULL) { 4263 // Express control dependency by a CastII node with a narrow type. 4264 value = new CastIINode(value, itype, false, true /* range check dependency */); 4265 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4266 // node from floating above the range check during loop optimizations. Otherwise, the 4267 // ConvI2L node may be eliminated independently of the range check, causing the data path 4268 // to become TOP while the control path is still there (although it's unreachable). 4269 value->set_req(0, ctrl); 4270 // Save CastII node to remove it after loop optimizations. 4271 phase->C->add_range_check_cast(value); 4272 value = phase->transform(value); 4273 } 4274 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4275 return phase->transform(new ConvI2LNode(value, ltype)); 4276 } 4277 4278 // The message about the current inlining is accumulated in 4279 // _print_inlining_stream and transfered into the _print_inlining_list 4280 // once we know whether inlining succeeds or not. For regular 4281 // inlining, messages are appended to the buffer pointed by 4282 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4283 // a new buffer is added after _print_inlining_idx in the list. This 4284 // way we can update the inlining message for late inlining call site 4285 // when the inlining is attempted again. 4286 void Compile::print_inlining_init() { 4287 if (print_inlining() || print_intrinsics()) { 4288 _print_inlining_stream = new stringStream(); 4289 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4290 } 4291 } 4292 4293 void Compile::print_inlining_reinit() { 4294 if (print_inlining() || print_intrinsics()) { 4295 // Re allocate buffer when we change ResourceMark 4296 _print_inlining_stream = new stringStream(); 4297 } 4298 } 4299 4300 void Compile::print_inlining_reset() { 4301 _print_inlining_stream->reset(); 4302 } 4303 4304 void Compile::print_inlining_commit() { 4305 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4306 // Transfer the message from _print_inlining_stream to the current 4307 // _print_inlining_list buffer and clear _print_inlining_stream. 4308 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4309 print_inlining_reset(); 4310 } 4311 4312 void Compile::print_inlining_push() { 4313 // Add new buffer to the _print_inlining_list at current position 4314 _print_inlining_idx++; 4315 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4316 } 4317 4318 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4319 return _print_inlining_list->at(_print_inlining_idx); 4320 } 4321 4322 void Compile::print_inlining_update(CallGenerator* cg) { 4323 if (print_inlining() || print_intrinsics()) { 4324 if (!cg->is_late_inline()) { 4325 if (print_inlining_current().cg() != NULL) { 4326 print_inlining_push(); 4327 } 4328 print_inlining_commit(); 4329 } else { 4330 if (print_inlining_current().cg() != cg && 4331 (print_inlining_current().cg() != NULL || 4332 print_inlining_current().ss()->size() != 0)) { 4333 print_inlining_push(); 4334 } 4335 print_inlining_commit(); 4336 print_inlining_current().set_cg(cg); 4337 } 4338 } 4339 } 4340 4341 void Compile::print_inlining_move_to(CallGenerator* cg) { 4342 // We resume inlining at a late inlining call site. Locate the 4343 // corresponding inlining buffer so that we can update it. 4344 if (print_inlining()) { 4345 for (int i = 0; i < _print_inlining_list->length(); i++) { 4346 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4347 _print_inlining_idx = i; 4348 return; 4349 } 4350 } 4351 ShouldNotReachHere(); 4352 } 4353 } 4354 4355 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4356 if (print_inlining()) { 4357 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4358 assert(print_inlining_current().cg() == cg, "wrong entry"); 4359 // replace message with new message 4360 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4361 print_inlining_commit(); 4362 print_inlining_current().set_cg(cg); 4363 } 4364 } 4365 4366 void Compile::print_inlining_assert_ready() { 4367 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4368 } 4369 4370 void Compile::process_print_inlining() { 4371 bool do_print_inlining = print_inlining() || print_intrinsics(); 4372 if (do_print_inlining || log() != NULL) { 4373 // Print inlining message for candidates that we couldn't inline 4374 // for lack of space 4375 for (int i = 0; i < _late_inlines.length(); i++) { 4376 CallGenerator* cg = _late_inlines.at(i); 4377 if (!cg->is_mh_late_inline()) { 4378 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4379 if (do_print_inlining) { 4380 cg->print_inlining_late(msg); 4381 } 4382 log_late_inline_failure(cg, msg); 4383 } 4384 } 4385 } 4386 if (do_print_inlining) { 4387 ResourceMark rm; 4388 stringStream ss; 4389 for (int i = 0; i < _print_inlining_list->length(); i++) { 4390 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4391 } 4392 size_t end = ss.size(); 4393 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4394 strncpy(_print_inlining_output, ss.base(), end+1); 4395 _print_inlining_output[end] = 0; 4396 } 4397 } 4398 4399 void Compile::dump_print_inlining() { 4400 if (_print_inlining_output != NULL) { 4401 tty->print_raw(_print_inlining_output); 4402 } 4403 } 4404 4405 void Compile::log_late_inline(CallGenerator* cg) { 4406 if (log() != NULL) { 4407 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4408 cg->unique_id()); 4409 JVMState* p = cg->call_node()->jvms(); 4410 while (p != NULL) { 4411 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4412 p = p->caller(); 4413 } 4414 log()->tail("late_inline"); 4415 } 4416 } 4417 4418 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4419 log_late_inline(cg); 4420 if (log() != NULL) { 4421 log()->inline_fail(msg); 4422 } 4423 } 4424 4425 void Compile::log_inline_id(CallGenerator* cg) { 4426 if (log() != NULL) { 4427 // The LogCompilation tool needs a unique way to identify late 4428 // inline call sites. This id must be unique for this call site in 4429 // this compilation. Try to have it unique across compilations as 4430 // well because it can be convenient when grepping through the log 4431 // file. 4432 // Distinguish OSR compilations from others in case CICountOSR is 4433 // on. 4434 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4435 cg->set_unique_id(id); 4436 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4437 } 4438 } 4439 4440 void Compile::log_inline_failure(const char* msg) { 4441 if (C->log() != NULL) { 4442 C->log()->inline_fail(msg); 4443 } 4444 } 4445 4446 4447 // Dump inlining replay data to the stream. 4448 // Don't change thread state and acquire any locks. 4449 void Compile::dump_inline_data(outputStream* out) { 4450 InlineTree* inl_tree = ilt(); 4451 if (inl_tree != NULL) { 4452 out->print(" inline %d", inl_tree->count()); 4453 inl_tree->dump_replay_data(out); 4454 } 4455 } 4456 4457 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4458 if (n1->Opcode() < n2->Opcode()) return -1; 4459 else if (n1->Opcode() > n2->Opcode()) return 1; 4460 4461 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4462 for (uint i = 1; i < n1->req(); i++) { 4463 if (n1->in(i) < n2->in(i)) return -1; 4464 else if (n1->in(i) > n2->in(i)) return 1; 4465 } 4466 4467 return 0; 4468 } 4469 4470 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4471 Node* n1 = *n1p; 4472 Node* n2 = *n2p; 4473 4474 return cmp_expensive_nodes(n1, n2); 4475 } 4476 4477 void Compile::sort_expensive_nodes() { 4478 if (!expensive_nodes_sorted()) { 4479 _expensive_nodes->sort(cmp_expensive_nodes); 4480 } 4481 } 4482 4483 bool Compile::expensive_nodes_sorted() const { 4484 for (int i = 1; i < _expensive_nodes->length(); i++) { 4485 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4486 return false; 4487 } 4488 } 4489 return true; 4490 } 4491 4492 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4493 if (_expensive_nodes->length() == 0) { 4494 return false; 4495 } 4496 4497 assert(OptimizeExpensiveOps, "optimization off?"); 4498 4499 // Take this opportunity to remove dead nodes from the list 4500 int j = 0; 4501 for (int i = 0; i < _expensive_nodes->length(); i++) { 4502 Node* n = _expensive_nodes->at(i); 4503 if (!n->is_unreachable(igvn)) { 4504 assert(n->is_expensive(), "should be expensive"); 4505 _expensive_nodes->at_put(j, n); 4506 j++; 4507 } 4508 } 4509 _expensive_nodes->trunc_to(j); 4510 4511 // Then sort the list so that similar nodes are next to each other 4512 // and check for at least two nodes of identical kind with same data 4513 // inputs. 4514 sort_expensive_nodes(); 4515 4516 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4517 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4518 return true; 4519 } 4520 } 4521 4522 return false; 4523 } 4524 4525 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4526 if (_expensive_nodes->length() == 0) { 4527 return; 4528 } 4529 4530 assert(OptimizeExpensiveOps, "optimization off?"); 4531 4532 // Sort to bring similar nodes next to each other and clear the 4533 // control input of nodes for which there's only a single copy. 4534 sort_expensive_nodes(); 4535 4536 int j = 0; 4537 int identical = 0; 4538 int i = 0; 4539 bool modified = false; 4540 for (; i < _expensive_nodes->length()-1; i++) { 4541 assert(j <= i, "can't write beyond current index"); 4542 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4543 identical++; 4544 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4545 continue; 4546 } 4547 if (identical > 0) { 4548 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4549 identical = 0; 4550 } else { 4551 Node* n = _expensive_nodes->at(i); 4552 igvn.replace_input_of(n, 0, NULL); 4553 igvn.hash_insert(n); 4554 modified = true; 4555 } 4556 } 4557 if (identical > 0) { 4558 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4559 } else if (_expensive_nodes->length() >= 1) { 4560 Node* n = _expensive_nodes->at(i); 4561 igvn.replace_input_of(n, 0, NULL); 4562 igvn.hash_insert(n); 4563 modified = true; 4564 } 4565 _expensive_nodes->trunc_to(j); 4566 if (modified) { 4567 igvn.optimize(); 4568 } 4569 } 4570 4571 void Compile::add_expensive_node(Node * n) { 4572 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4573 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4574 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4575 if (OptimizeExpensiveOps) { 4576 _expensive_nodes->append(n); 4577 } else { 4578 // Clear control input and let IGVN optimize expensive nodes if 4579 // OptimizeExpensiveOps is off. 4580 n->set_req(0, NULL); 4581 } 4582 } 4583 4584 /** 4585 * Remove the speculative part of types and clean up the graph 4586 */ 4587 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4588 if (UseTypeSpeculation) { 4589 Unique_Node_List worklist; 4590 worklist.push(root()); 4591 int modified = 0; 4592 // Go over all type nodes that carry a speculative type, drop the 4593 // speculative part of the type and enqueue the node for an igvn 4594 // which may optimize it out. 4595 for (uint next = 0; next < worklist.size(); ++next) { 4596 Node *n = worklist.at(next); 4597 if (n->is_Type()) { 4598 TypeNode* tn = n->as_Type(); 4599 const Type* t = tn->type(); 4600 const Type* t_no_spec = t->remove_speculative(); 4601 if (t_no_spec != t) { 4602 bool in_hash = igvn.hash_delete(n); 4603 assert(in_hash, "node should be in igvn hash table"); 4604 tn->set_type(t_no_spec); 4605 igvn.hash_insert(n); 4606 igvn._worklist.push(n); // give it a chance to go away 4607 modified++; 4608 } 4609 } 4610 uint max = n->len(); 4611 for( uint i = 0; i < max; ++i ) { 4612 Node *m = n->in(i); 4613 if (not_a_node(m)) continue; 4614 worklist.push(m); 4615 } 4616 } 4617 // Drop the speculative part of all types in the igvn's type table 4618 igvn.remove_speculative_types(); 4619 if (modified > 0) { 4620 igvn.optimize(); 4621 } 4622 #ifdef ASSERT 4623 // Verify that after the IGVN is over no speculative type has resurfaced 4624 worklist.clear(); 4625 worklist.push(root()); 4626 for (uint next = 0; next < worklist.size(); ++next) { 4627 Node *n = worklist.at(next); 4628 const Type* t = igvn.type_or_null(n); 4629 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4630 if (n->is_Type()) { 4631 t = n->as_Type()->type(); 4632 assert(t == t->remove_speculative(), "no more speculative types"); 4633 } 4634 uint max = n->len(); 4635 for( uint i = 0; i < max; ++i ) { 4636 Node *m = n->in(i); 4637 if (not_a_node(m)) continue; 4638 worklist.push(m); 4639 } 4640 } 4641 igvn.check_no_speculative_types(); 4642 #endif 4643 } 4644 } 4645 4646 // Auxiliary method to support randomized stressing/fuzzing. 4647 // 4648 // This method can be called the arbitrary number of times, with current count 4649 // as the argument. The logic allows selecting a single candidate from the 4650 // running list of candidates as follows: 4651 // int count = 0; 4652 // Cand* selected = null; 4653 // while(cand = cand->next()) { 4654 // if (randomized_select(++count)) { 4655 // selected = cand; 4656 // } 4657 // } 4658 // 4659 // Including count equalizes the chances any candidate is "selected". 4660 // This is useful when we don't have the complete list of candidates to choose 4661 // from uniformly. In this case, we need to adjust the randomicity of the 4662 // selection, or else we will end up biasing the selection towards the latter 4663 // candidates. 4664 // 4665 // Quick back-envelope calculation shows that for the list of n candidates 4666 // the equal probability for the candidate to persist as "best" can be 4667 // achieved by replacing it with "next" k-th candidate with the probability 4668 // of 1/k. It can be easily shown that by the end of the run, the 4669 // probability for any candidate is converged to 1/n, thus giving the 4670 // uniform distribution among all the candidates. 4671 // 4672 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4673 #define RANDOMIZED_DOMAIN_POW 29 4674 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4675 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4676 bool Compile::randomized_select(int count) { 4677 assert(count > 0, "only positive"); 4678 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4679 } 4680 4681 CloneMap& Compile::clone_map() { return _clone_map; } 4682 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4683 4684 void NodeCloneInfo::dump() const { 4685 tty->print(" {%d:%d} ", idx(), gen()); 4686 } 4687 4688 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4689 uint64_t val = value(old->_idx); 4690 NodeCloneInfo cio(val); 4691 assert(val != 0, "old node should be in the map"); 4692 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4693 insert(nnn->_idx, cin.get()); 4694 #ifndef PRODUCT 4695 if (is_debug()) { 4696 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4697 } 4698 #endif 4699 } 4700 4701 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4702 NodeCloneInfo cio(value(old->_idx)); 4703 if (cio.get() == 0) { 4704 cio.set(old->_idx, 0); 4705 insert(old->_idx, cio.get()); 4706 #ifndef PRODUCT 4707 if (is_debug()) { 4708 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4709 } 4710 #endif 4711 } 4712 clone(old, nnn, gen); 4713 } 4714 4715 int CloneMap::max_gen() const { 4716 int g = 0; 4717 DictI di(_dict); 4718 for(; di.test(); ++di) { 4719 int t = gen(di._key); 4720 if (g < t) { 4721 g = t; 4722 #ifndef PRODUCT 4723 if (is_debug()) { 4724 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4725 } 4726 #endif 4727 } 4728 } 4729 return g; 4730 } 4731 4732 void CloneMap::dump(node_idx_t key) const { 4733 uint64_t val = value(key); 4734 if (val != 0) { 4735 NodeCloneInfo ni(val); 4736 ni.dump(); 4737 } 4738 }