1 /* 2 * Copyright (c) 1997, 2019, 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/valuetypenode.hpp" 71 #include "opto/vectornode.hpp" 72 #include "runtime/arguments.hpp" 73 #include "runtime/sharedRuntime.hpp" 74 #include "runtime/signature.hpp" 75 #include "runtime/stubRoutines.hpp" 76 #include "runtime/timer.hpp" 77 #include "utilities/align.hpp" 78 #include "utilities/copy.hpp" 79 #include "utilities/macros.hpp" 80 #if INCLUDE_ZGC 81 #include "gc/z/c2/zBarrierSetC2.hpp" 82 #endif 83 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == NULL) { 89 _mach_constant_base_node = new MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 class IntrinsicDescPair { 101 private: 102 ciMethod* _m; 103 bool _is_virtual; 104 public: 105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 107 ciMethod* m= elt->method(); 108 ciMethod* key_m = key->_m; 109 if (key_m < m) return -1; 110 else if (key_m > m) return 1; 111 else { 112 bool is_virtual = elt->is_virtual(); 113 bool key_virtual = key->_is_virtual; 114 if (key_virtual < is_virtual) return -1; 115 else if (key_virtual > is_virtual) return 1; 116 else return 0; 117 } 118 } 119 }; 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 121 #ifdef ASSERT 122 for (int i = 1; i < _intrinsics->length(); i++) { 123 CallGenerator* cg1 = _intrinsics->at(i-1); 124 CallGenerator* cg2 = _intrinsics->at(i); 125 assert(cg1->method() != cg2->method() 126 ? cg1->method() < cg2->method() 127 : cg1->is_virtual() < cg2->is_virtual(), 128 "compiler intrinsics list must stay sorted"); 129 } 130 #endif 131 IntrinsicDescPair pair(m, is_virtual); 132 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 if (_intrinsics == NULL) { 137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 138 } 139 int len = _intrinsics->length(); 140 bool found = false; 141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 142 assert(!found, "registering twice"); 143 _intrinsics->insert_before(index, cg); 144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 145 } 146 147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 148 assert(m->is_loaded(), "don't try this on unloaded methods"); 149 if (_intrinsics != NULL) { 150 bool found = false; 151 int index = intrinsic_insertion_index(m, is_virtual, found); 152 if (found) { 153 return _intrinsics->at(index); 154 } 155 } 156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 157 if (m->intrinsic_id() != vmIntrinsics::_none && 158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 160 if (cg != NULL) { 161 // Save it for next time: 162 register_intrinsic(cg); 163 return cg; 164 } else { 165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 166 } 167 } 168 return NULL; 169 } 170 171 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 172 // in library_call.cpp. 173 174 175 #ifndef PRODUCT 176 // statistics gathering... 177 178 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 180 181 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 182 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 183 int oflags = _intrinsic_hist_flags[id]; 184 assert(flags != 0, "what happened?"); 185 if (is_virtual) { 186 flags |= _intrinsic_virtual; 187 } 188 bool changed = (flags != oflags); 189 if ((flags & _intrinsic_worked) != 0) { 190 juint count = (_intrinsic_hist_count[id] += 1); 191 if (count == 1) { 192 changed = true; // first time 193 } 194 // increment the overall count also: 195 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 196 } 197 if (changed) { 198 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 199 // Something changed about the intrinsic's virtuality. 200 if ((flags & _intrinsic_virtual) != 0) { 201 // This is the first use of this intrinsic as a virtual call. 202 if (oflags != 0) { 203 // We already saw it as a non-virtual, so note both cases. 204 flags |= _intrinsic_both; 205 } 206 } else if ((oflags & _intrinsic_both) == 0) { 207 // This is the first use of this intrinsic as a non-virtual 208 flags |= _intrinsic_both; 209 } 210 } 211 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 212 } 213 // update the overall flags also: 214 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 215 return changed; 216 } 217 218 static char* format_flags(int flags, char* buf) { 219 buf[0] = 0; 220 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 221 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 222 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 223 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 224 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 225 if (buf[0] == 0) strcat(buf, ","); 226 assert(buf[0] == ',', "must be"); 227 return &buf[1]; 228 } 229 230 void Compile::print_intrinsic_statistics() { 231 char flagsbuf[100]; 232 ttyLocker ttyl; 233 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 234 tty->print_cr("Compiler intrinsic usage:"); 235 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 236 if (total == 0) total = 1; // avoid div0 in case of no successes 237 #define PRINT_STAT_LINE(name, c, f) \ 238 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 239 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 240 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 241 int flags = _intrinsic_hist_flags[id]; 242 juint count = _intrinsic_hist_count[id]; 243 if ((flags | count) != 0) { 244 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 245 } 246 } 247 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 248 if (xtty != NULL) xtty->tail("statistics"); 249 } 250 251 void Compile::print_statistics() { 252 { ttyLocker ttyl; 253 if (xtty != NULL) xtty->head("statistics type='opto'"); 254 Parse::print_statistics(); 255 PhaseCCP::print_statistics(); 256 PhaseRegAlloc::print_statistics(); 257 Scheduling::print_statistics(); 258 PhasePeephole::print_statistics(); 259 PhaseIdealLoop::print_statistics(); 260 if (xtty != NULL) xtty->tail("statistics"); 261 } 262 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 263 // put this under its own <statistics> element. 264 print_intrinsic_statistics(); 265 } 266 } 267 #endif //PRODUCT 268 269 // Support for bundling info 270 Bundle* Compile::node_bundling(const Node *n) { 271 assert(valid_bundle_info(n), "oob"); 272 return &_node_bundling_base[n->_idx]; 273 } 274 275 bool Compile::valid_bundle_info(const Node *n) { 276 return (_node_bundling_limit > n->_idx); 277 } 278 279 280 void Compile::gvn_replace_by(Node* n, Node* nn) { 281 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 282 Node* use = n->last_out(i); 283 bool is_in_table = initial_gvn()->hash_delete(use); 284 uint uses_found = 0; 285 for (uint j = 0; j < use->len(); j++) { 286 if (use->in(j) == n) { 287 if (j < use->req()) 288 use->set_req(j, nn); 289 else 290 use->set_prec(j, nn); 291 uses_found++; 292 } 293 } 294 if (is_in_table) { 295 // reinsert into table 296 initial_gvn()->hash_find_insert(use); 297 } 298 record_for_igvn(use); 299 i -= uses_found; // we deleted 1 or more copies of this edge 300 } 301 } 302 303 304 static inline bool not_a_node(const Node* n) { 305 if (n == NULL) return true; 306 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 307 if (*(address*)n == badAddress) return true; // kill by Node::destruct 308 return false; 309 } 310 311 // Identify all nodes that are reachable from below, useful. 312 // Use breadth-first pass that records state in a Unique_Node_List, 313 // recursive traversal is slower. 314 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 315 int estimated_worklist_size = live_nodes(); 316 useful.map( estimated_worklist_size, NULL ); // preallocate space 317 318 // Initialize worklist 319 if (root() != NULL) { useful.push(root()); } 320 // If 'top' is cached, declare it useful to preserve cached node 321 if( cached_top_node() ) { useful.push(cached_top_node()); } 322 323 // Push all useful nodes onto the list, breadthfirst 324 for( uint next = 0; next < useful.size(); ++next ) { 325 assert( next < unique(), "Unique useful nodes < total nodes"); 326 Node *n = useful.at(next); 327 uint max = n->len(); 328 for( uint i = 0; i < max; ++i ) { 329 Node *m = n->in(i); 330 if (not_a_node(m)) continue; 331 useful.push(m); 332 } 333 } 334 } 335 336 // Update dead_node_list with any missing dead nodes using useful 337 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 338 void Compile::update_dead_node_list(Unique_Node_List &useful) { 339 uint max_idx = unique(); 340 VectorSet& useful_node_set = useful.member_set(); 341 342 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 343 // If node with index node_idx is not in useful set, 344 // mark it as dead in dead node list. 345 if (! useful_node_set.test(node_idx) ) { 346 record_dead_node(node_idx); 347 } 348 } 349 } 350 351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 352 int shift = 0; 353 for (int i = 0; i < inlines->length(); i++) { 354 CallGenerator* cg = inlines->at(i); 355 CallNode* call = cg->call_node(); 356 if (shift > 0) { 357 inlines->at_put(i-shift, cg); 358 } 359 if (!useful.member(call)) { 360 shift++; 361 } 362 } 363 inlines->trunc_to(inlines->length()-shift); 364 } 365 366 // Disconnect all useless nodes by disconnecting those at the boundary. 367 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 368 uint next = 0; 369 while (next < useful.size()) { 370 Node *n = useful.at(next++); 371 if (n->is_SafePoint()) { 372 // We're done with a parsing phase. Replaced nodes are not valid 373 // beyond that point. 374 n->as_SafePoint()->delete_replaced_nodes(); 375 } 376 // Use raw traversal of out edges since this code removes out edges 377 int max = n->outcnt(); 378 for (int j = 0; j < max; ++j) { 379 Node* child = n->raw_out(j); 380 if (! useful.member(child)) { 381 assert(!child->is_top() || child != top(), 382 "If top is cached in Compile object it is in useful list"); 383 // Only need to remove this out-edge to the useless node 384 n->raw_del_out(j); 385 --j; 386 --max; 387 } 388 } 389 if (n->outcnt() == 1 && n->has_special_unique_user()) { 390 record_for_igvn(n->unique_out()); 391 } 392 } 393 // Remove useless macro and predicate opaq nodes 394 for (int i = C->macro_count()-1; i >= 0; i--) { 395 Node* n = C->macro_node(i); 396 if (!useful.member(n)) { 397 remove_macro_node(n); 398 } 399 } 400 // Remove useless CastII nodes with range check dependency 401 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 402 Node* cast = range_check_cast_node(i); 403 if (!useful.member(cast)) { 404 remove_range_check_cast(cast); 405 } 406 } 407 // Remove useless expensive nodes 408 for (int i = C->expensive_count()-1; i >= 0; i--) { 409 Node* n = C->expensive_node(i); 410 if (!useful.member(n)) { 411 remove_expensive_node(n); 412 } 413 } 414 // Remove useless Opaque4 nodes 415 for (int i = opaque4_count() - 1; i >= 0; i--) { 416 Node* opaq = opaque4_node(i); 417 if (!useful.member(opaq)) { 418 remove_opaque4_node(opaq); 419 } 420 } 421 // Remove useless value type nodes 422 if (_value_type_nodes != NULL) { 423 _value_type_nodes->remove_useless_nodes(useful.member_set()); 424 } 425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 426 bs->eliminate_useless_gc_barriers(useful, this); 427 // clean up the late inline lists 428 remove_useless_late_inlines(&_string_late_inlines, useful); 429 remove_useless_late_inlines(&_boxing_late_inlines, useful); 430 remove_useless_late_inlines(&_late_inlines, useful); 431 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 432 } 433 434 //------------------------------frame_size_in_words----------------------------- 435 // frame_slots in units of words 436 int Compile::frame_size_in_words() const { 437 // shift is 0 in LP32 and 1 in LP64 438 const int shift = (LogBytesPerWord - LogBytesPerInt); 439 int words = _frame_slots >> shift; 440 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 441 return words; 442 } 443 444 // To bang the stack of this compiled method we use the stack size 445 // that the interpreter would need in case of a deoptimization. This 446 // removes the need to bang the stack in the deoptimization blob which 447 // in turn simplifies stack overflow handling. 448 int Compile::bang_size_in_bytes() const { 449 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 450 } 451 452 // ============================================================================ 453 //------------------------------CompileWrapper--------------------------------- 454 class CompileWrapper : public StackObj { 455 Compile *const _compile; 456 public: 457 CompileWrapper(Compile* compile); 458 459 ~CompileWrapper(); 460 }; 461 462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 463 // the Compile* pointer is stored in the current ciEnv: 464 ciEnv* env = compile->env(); 465 assert(env == ciEnv::current(), "must already be a ciEnv active"); 466 assert(env->compiler_data() == NULL, "compile already active?"); 467 env->set_compiler_data(compile); 468 assert(compile == Compile::current(), "sanity"); 469 470 compile->set_type_dict(NULL); 471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 472 compile->clone_map().set_clone_idx(0); 473 compile->set_type_hwm(NULL); 474 compile->set_type_last_size(0); 475 compile->set_last_tf(NULL, NULL); 476 compile->set_indexSet_arena(NULL); 477 compile->set_indexSet_free_block_list(NULL); 478 compile->init_type_arena(); 479 Type::Initialize(compile); 480 _compile->set_scratch_buffer_blob(NULL); 481 _compile->begin_method(); 482 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 483 } 484 CompileWrapper::~CompileWrapper() { 485 _compile->end_method(); 486 if (_compile->scratch_buffer_blob() != NULL) 487 BufferBlob::free(_compile->scratch_buffer_blob()); 488 _compile->env()->set_compiler_data(NULL); 489 } 490 491 492 //----------------------------print_compile_messages--------------------------- 493 void Compile::print_compile_messages() { 494 #ifndef PRODUCT 495 // Check if recompiling 496 if (_subsume_loads == false && PrintOpto) { 497 // Recompiling without allowing machine instructions to subsume loads 498 tty->print_cr("*********************************************************"); 499 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 500 tty->print_cr("*********************************************************"); 501 } 502 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 503 // Recompiling without escape analysis 504 tty->print_cr("*********************************************************"); 505 tty->print_cr("** Bailout: Recompile without escape analysis **"); 506 tty->print_cr("*********************************************************"); 507 } 508 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 509 // Recompiling without boxing elimination 510 tty->print_cr("*********************************************************"); 511 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 512 tty->print_cr("*********************************************************"); 513 } 514 if (C->directive()->BreakAtCompileOption) { 515 // Open the debugger when compiling this method. 516 tty->print("### Breaking when compiling: "); 517 method()->print_short_name(); 518 tty->cr(); 519 BREAKPOINT; 520 } 521 522 if( PrintOpto ) { 523 if (is_osr_compilation()) { 524 tty->print("[OSR]%3d", _compile_id); 525 } else { 526 tty->print("%3d", _compile_id); 527 } 528 } 529 #endif 530 } 531 532 533 //-----------------------init_scratch_buffer_blob------------------------------ 534 // Construct a temporary BufferBlob and cache it for this compile. 535 void Compile::init_scratch_buffer_blob(int const_size) { 536 // If there is already a scratch buffer blob allocated and the 537 // constant section is big enough, use it. Otherwise free the 538 // current and allocate a new one. 539 BufferBlob* blob = scratch_buffer_blob(); 540 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 541 // Use the current blob. 542 } else { 543 if (blob != NULL) { 544 BufferBlob::free(blob); 545 } 546 547 ResourceMark rm; 548 _scratch_const_size = const_size; 549 int size = C2Compiler::initial_code_buffer_size(const_size); 550 #ifdef ASSERT 551 if (C->has_scalarized_args()) { 552 // Oop verification for loading object fields from scalarized value types in the new entry point requires lots of space 553 size += 5120; 554 } 555 #endif 556 blob = BufferBlob::create("Compile::scratch_buffer", size); 557 // Record the buffer blob for next time. 558 set_scratch_buffer_blob(blob); 559 // Have we run out of code space? 560 if (scratch_buffer_blob() == NULL) { 561 // Let CompilerBroker disable further compilations. 562 record_failure("Not enough space for scratch buffer in CodeCache"); 563 return; 564 } 565 } 566 567 // Initialize the relocation buffers 568 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 569 set_scratch_locs_memory(locs_buf); 570 } 571 572 573 //-----------------------scratch_emit_size------------------------------------- 574 // Helper function that computes size by emitting code 575 uint Compile::scratch_emit_size(const Node* n) { 576 // Start scratch_emit_size section. 577 set_in_scratch_emit_size(true); 578 579 // Emit into a trash buffer and count bytes emitted. 580 // This is a pretty expensive way to compute a size, 581 // but it works well enough if seldom used. 582 // All common fixed-size instructions are given a size 583 // method by the AD file. 584 // Note that the scratch buffer blob and locs memory are 585 // allocated at the beginning of the compile task, and 586 // may be shared by several calls to scratch_emit_size. 587 // The allocation of the scratch buffer blob is particularly 588 // expensive, since it has to grab the code cache lock. 589 BufferBlob* blob = this->scratch_buffer_blob(); 590 assert(blob != NULL, "Initialize BufferBlob at start"); 591 assert(blob->size() > MAX_inst_size, "sanity"); 592 relocInfo* locs_buf = scratch_locs_memory(); 593 address blob_begin = blob->content_begin(); 594 address blob_end = (address)locs_buf; 595 assert(blob->contains(blob_end), "sanity"); 596 CodeBuffer buf(blob_begin, blob_end - blob_begin); 597 buf.initialize_consts_size(_scratch_const_size); 598 buf.initialize_stubs_size(MAX_stubs_size); 599 assert(locs_buf != NULL, "sanity"); 600 int lsize = MAX_locs_size / 3; 601 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 602 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 603 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 604 // Mark as scratch buffer. 605 buf.consts()->set_scratch_emit(); 606 buf.insts()->set_scratch_emit(); 607 buf.stubs()->set_scratch_emit(); 608 609 // Do the emission. 610 611 Label fakeL; // Fake label for branch instructions. 612 Label* saveL = NULL; 613 uint save_bnum = 0; 614 bool is_branch = n->is_MachBranch(); 615 if (is_branch) { 616 MacroAssembler masm(&buf); 617 masm.bind(fakeL); 618 n->as_MachBranch()->save_label(&saveL, &save_bnum); 619 n->as_MachBranch()->label_set(&fakeL, 0); 620 } else if (n->is_MachProlog()) { 621 saveL = ((MachPrologNode*)n)->_verified_entry; 622 ((MachPrologNode*)n)->_verified_entry = &fakeL; 623 } 624 n->emit(buf, this->regalloc()); 625 626 // Emitting into the scratch buffer should not fail 627 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 628 629 // Restore label. 630 if (is_branch) { 631 n->as_MachBranch()->label_set(saveL, save_bnum); 632 } else if (n->is_MachProlog()) { 633 ((MachPrologNode*)n)->_verified_entry = saveL; 634 } 635 636 // End scratch_emit_size section. 637 set_in_scratch_emit_size(false); 638 639 return buf.insts_size(); 640 } 641 642 643 // ============================================================================ 644 //------------------------------Compile standard------------------------------- 645 debug_only( int Compile::_debug_idx = 100000; ) 646 647 // Compile a method. entry_bci is -1 for normal compilations and indicates 648 // the continuation bci for on stack replacement. 649 650 651 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 652 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 653 : Phase(Compiler), 654 _compile_id(ci_env->compile_id()), 655 _save_argument_registers(false), 656 _subsume_loads(subsume_loads), 657 _do_escape_analysis(do_escape_analysis), 658 _eliminate_boxing(eliminate_boxing), 659 _method(target), 660 _entry_bci(osr_bci), 661 _stub_function(NULL), 662 _stub_name(NULL), 663 _stub_entry_point(NULL), 664 _max_node_limit(MaxNodeLimit), 665 _orig_pc_slot(0), 666 _orig_pc_slot_offset_in_bytes(0), 667 _sp_inc_slot(0), 668 _sp_inc_slot_offset_in_bytes(0), 669 _inlining_progress(false), 670 _inlining_incrementally(false), 671 _do_cleanup(false), 672 _has_reserved_stack_access(target->has_reserved_stack_access()), 673 #ifndef PRODUCT 674 _trace_opto_output(directive->TraceOptoOutputOption), 675 #endif 676 _has_method_handle_invokes(false), 677 _clinit_barrier_on_entry(false), 678 _comp_arena(mtCompiler), 679 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 680 _env(ci_env), 681 _directive(directive), 682 _log(ci_env->log()), 683 _failure_reason(NULL), 684 _congraph(NULL), 685 #ifndef PRODUCT 686 _printer(IdealGraphPrinter::printer()), 687 #endif 688 _dead_node_list(comp_arena()), 689 _dead_node_count(0), 690 _node_arena(mtCompiler), 691 _old_arena(mtCompiler), 692 _mach_constant_base_node(NULL), 693 _Compile_types(mtCompiler), 694 _initial_gvn(NULL), 695 _for_igvn(NULL), 696 _warm_calls(NULL), 697 _late_inlines(comp_arena(), 2, 0, NULL), 698 _string_late_inlines(comp_arena(), 2, 0, NULL), 699 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 700 _late_inlines_pos(0), 701 _number_of_mh_late_inlines(0), 702 _print_inlining_stream(NULL), 703 _print_inlining_list(NULL), 704 _print_inlining_idx(0), 705 _print_inlining_output(NULL), 706 _replay_inline_data(NULL), 707 _java_calls(0), 708 _inner_loops(0), 709 _interpreter_frame_size(0), 710 _node_bundling_limit(0), 711 _node_bundling_base(NULL), 712 _code_buffer("Compile::Fill_buffer"), 713 _scratch_const_size(-1), 714 _in_scratch_emit_size(false) 715 #ifndef PRODUCT 716 , _in_dump_cnt(0) 717 #endif 718 { 719 C = this; 720 #ifndef PRODUCT 721 if (_printer != NULL) { 722 _printer->set_compile(this); 723 } 724 #endif 725 CompileWrapper cw(this); 726 727 if (CITimeVerbose) { 728 tty->print(" "); 729 target->holder()->name()->print(); 730 tty->print("."); 731 target->print_short_name(); 732 tty->print(" "); 733 } 734 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 735 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 736 737 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 738 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 739 // We can always print a disassembly, either abstract (hex dump) or 740 // with the help of a suitable hsdis library. Thus, we should not 741 // couple print_assembly and print_opto_assembly controls. 742 // But: always print opto and regular assembly on compile command 'print'. 743 bool print_assembly = directive->PrintAssemblyOption; 744 set_print_assembly(print_opto_assembly || print_assembly); 745 #else 746 set_print_assembly(false); // must initialize. 747 #endif 748 749 #ifndef PRODUCT 750 set_parsed_irreducible_loop(false); 751 752 if (directive->ReplayInlineOption) { 753 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 754 } 755 #endif 756 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 757 set_print_intrinsics(directive->PrintIntrinsicsOption); 758 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 759 760 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 761 // Make sure the method being compiled gets its own MDO, 762 // so we can at least track the decompile_count(). 763 // Need MDO to record RTM code generation state. 764 method()->ensure_method_data(); 765 } 766 767 Init(::AliasLevel); 768 769 770 print_compile_messages(); 771 772 _ilt = InlineTree::build_inline_tree_root(); 773 774 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 775 assert(num_alias_types() >= AliasIdxRaw, ""); 776 777 #define MINIMUM_NODE_HASH 1023 778 // Node list that Iterative GVN will start with 779 Unique_Node_List for_igvn(comp_arena()); 780 set_for_igvn(&for_igvn); 781 782 // GVN that will be run immediately on new nodes 783 uint estimated_size = method()->code_size()*4+64; 784 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 785 PhaseGVN gvn(node_arena(), estimated_size); 786 set_initial_gvn(&gvn); 787 788 print_inlining_init(); 789 { // Scope for timing the parser 790 TracePhase tp("parse", &timers[_t_parser]); 791 792 // Put top into the hash table ASAP. 793 initial_gvn()->transform_no_reclaim(top()); 794 795 // Set up tf(), start(), and find a CallGenerator. 796 CallGenerator* cg = NULL; 797 if (is_osr_compilation()) { 798 const TypeTuple *domain = StartOSRNode::osr_domain(); 799 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 800 init_tf(TypeFunc::make(domain, range)); 801 StartNode* s = new StartOSRNode(root(), domain); 802 initial_gvn()->set_type_bottom(s); 803 init_start(s); 804 cg = CallGenerator::for_osr(method(), entry_bci()); 805 } else { 806 // Normal case. 807 init_tf(TypeFunc::make(method())); 808 StartNode* s = new StartNode(root(), tf()->domain_cc()); 809 initial_gvn()->set_type_bottom(s); 810 init_start(s); 811 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 812 // With java.lang.ref.reference.get() we must go through the 813 // intrinsic - even when get() is the root 814 // method of the compile - so that, if necessary, the value in 815 // the referent field of the reference object gets recorded by 816 // the pre-barrier code. 817 cg = find_intrinsic(method(), false); 818 } 819 if (cg == NULL) { 820 float past_uses = method()->interpreter_invocation_count(); 821 float expected_uses = past_uses; 822 cg = CallGenerator::for_inline(method(), expected_uses); 823 } 824 } 825 if (failing()) return; 826 if (cg == NULL) { 827 record_method_not_compilable("cannot parse method"); 828 return; 829 } 830 JVMState* jvms = build_start_state(start(), tf()); 831 if ((jvms = cg->generate(jvms)) == NULL) { 832 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 833 record_method_not_compilable("method parse failed"); 834 } 835 return; 836 } 837 GraphKit kit(jvms); 838 839 if (!kit.stopped()) { 840 // Accept return values, and transfer control we know not where. 841 // This is done by a special, unique ReturnNode bound to root. 842 return_values(kit.jvms()); 843 } 844 845 if (kit.has_exceptions()) { 846 // Any exceptions that escape from this call must be rethrown 847 // to whatever caller is dynamically above us on the stack. 848 // This is done by a special, unique RethrowNode bound to root. 849 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 850 } 851 852 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 853 854 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 855 inline_string_calls(true); 856 } 857 858 if (failing()) return; 859 860 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 861 862 // Remove clutter produced by parsing. 863 if (!failing()) { 864 ResourceMark rm; 865 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 866 } 867 } 868 869 // Note: Large methods are capped off in do_one_bytecode(). 870 if (failing()) return; 871 872 // After parsing, node notes are no longer automagic. 873 // They must be propagated by register_new_node_with_optimizer(), 874 // clone(), or the like. 875 set_default_node_notes(NULL); 876 877 for (;;) { 878 int successes = Inline_Warm(); 879 if (failing()) return; 880 if (successes == 0) break; 881 } 882 883 // Drain the list. 884 Finish_Warm(); 885 #ifndef PRODUCT 886 if (_printer && _printer->should_print(1)) { 887 _printer->print_inlining(); 888 } 889 #endif 890 891 if (failing()) return; 892 NOT_PRODUCT( verify_graph_edges(); ) 893 894 // Now optimize 895 Optimize(); 896 if (failing()) return; 897 NOT_PRODUCT( verify_graph_edges(); ) 898 899 #ifndef PRODUCT 900 if (PrintIdeal) { 901 ttyLocker ttyl; // keep the following output all in one block 902 // This output goes directly to the tty, not the compiler log. 903 // To enable tools to match it up with the compilation activity, 904 // be sure to tag this tty output with the compile ID. 905 if (xtty != NULL) { 906 xtty->head("ideal compile_id='%d'%s", compile_id(), 907 is_osr_compilation() ? " compile_kind='osr'" : 908 ""); 909 } 910 root()->dump(9999); 911 if (xtty != NULL) { 912 xtty->tail("ideal"); 913 } 914 } 915 #endif 916 917 #ifdef ASSERT 918 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 919 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 920 #endif 921 922 // Dump compilation data to replay it. 923 if (directive->DumpReplayOption) { 924 env()->dump_replay_data(_compile_id); 925 } 926 if (directive->DumpInlineOption && (ilt() != NULL)) { 927 env()->dump_inline_data(_compile_id); 928 } 929 930 // Now that we know the size of all the monitors we can add a fixed slot 931 // for the original deopt pc. 932 933 _orig_pc_slot = fixed_slots(); 934 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 935 936 if (needs_stack_repair()) { 937 // One extra slot for the special stack increment value 938 _sp_inc_slot = next_slot; 939 next_slot += 2; 940 } 941 942 set_fixed_slots(next_slot); 943 944 // Compute when to use implicit null checks. Used by matching trap based 945 // nodes and NullCheck optimization. 946 set_allowed_deopt_reasons(); 947 948 // Now generate code 949 Code_Gen(); 950 if (failing()) return; 951 952 // Check if we want to skip execution of all compiled code. 953 { 954 #ifndef PRODUCT 955 if (OptoNoExecute) { 956 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 957 return; 958 } 959 #endif 960 TracePhase tp("install_code", &timers[_t_registerMethod]); 961 962 if (is_osr_compilation()) { 963 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 964 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 965 } else { 966 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 967 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry) == -1) { 968 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, _first_block_size); 969 } 970 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry_RO) == -1) { 971 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, _first_block_size); 972 } 973 if (_code_offsets.value(CodeOffsets::Entry) == -1) { 974 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size); 975 } 976 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 977 } 978 979 env()->register_method(_method, _entry_bci, 980 &_code_offsets, 981 _orig_pc_slot_offset_in_bytes, 982 code_buffer(), 983 frame_size_in_words(), _oop_map_set, 984 &_handler_table, &_inc_table, 985 compiler, 986 has_unsafe_access(), 987 SharedRuntime::is_wide_vector(max_vector_size()), 988 rtm_state() 989 ); 990 991 if (log() != NULL) // Print code cache state into compiler log 992 log()->code_cache_state(); 993 } 994 } 995 996 //------------------------------Compile---------------------------------------- 997 // Compile a runtime stub 998 Compile::Compile( ciEnv* ci_env, 999 TypeFunc_generator generator, 1000 address stub_function, 1001 const char *stub_name, 1002 int is_fancy_jump, 1003 bool pass_tls, 1004 bool save_arg_registers, 1005 bool return_pc, 1006 DirectiveSet* directive) 1007 : Phase(Compiler), 1008 _compile_id(0), 1009 _save_argument_registers(save_arg_registers), 1010 _subsume_loads(true), 1011 _do_escape_analysis(false), 1012 _eliminate_boxing(false), 1013 _method(NULL), 1014 _entry_bci(InvocationEntryBci), 1015 _stub_function(stub_function), 1016 _stub_name(stub_name), 1017 _stub_entry_point(NULL), 1018 _max_node_limit(MaxNodeLimit), 1019 _orig_pc_slot(0), 1020 _orig_pc_slot_offset_in_bytes(0), 1021 _sp_inc_slot(0), 1022 _sp_inc_slot_offset_in_bytes(0), 1023 _inlining_progress(false), 1024 _inlining_incrementally(false), 1025 _has_reserved_stack_access(false), 1026 #ifndef PRODUCT 1027 _trace_opto_output(directive->TraceOptoOutputOption), 1028 #endif 1029 _has_method_handle_invokes(false), 1030 _clinit_barrier_on_entry(false), 1031 _comp_arena(mtCompiler), 1032 _env(ci_env), 1033 _directive(directive), 1034 _log(ci_env->log()), 1035 _failure_reason(NULL), 1036 _congraph(NULL), 1037 #ifndef PRODUCT 1038 _printer(NULL), 1039 #endif 1040 _dead_node_list(comp_arena()), 1041 _dead_node_count(0), 1042 _node_arena(mtCompiler), 1043 _old_arena(mtCompiler), 1044 _mach_constant_base_node(NULL), 1045 _Compile_types(mtCompiler), 1046 _initial_gvn(NULL), 1047 _for_igvn(NULL), 1048 _warm_calls(NULL), 1049 _number_of_mh_late_inlines(0), 1050 _print_inlining_stream(NULL), 1051 _print_inlining_list(NULL), 1052 _print_inlining_idx(0), 1053 _print_inlining_output(NULL), 1054 _replay_inline_data(NULL), 1055 _java_calls(0), 1056 _inner_loops(0), 1057 _interpreter_frame_size(0), 1058 _node_bundling_limit(0), 1059 _node_bundling_base(NULL), 1060 _code_buffer("Compile::Fill_buffer"), 1061 #ifndef PRODUCT 1062 _in_dump_cnt(0), 1063 #endif 1064 _allowed_reasons(0) { 1065 C = this; 1066 1067 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1068 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1069 1070 #ifndef PRODUCT 1071 set_print_assembly(PrintFrameConverterAssembly); 1072 set_parsed_irreducible_loop(false); 1073 #else 1074 set_print_assembly(false); // Must initialize. 1075 #endif 1076 set_has_irreducible_loop(false); // no loops 1077 1078 CompileWrapper cw(this); 1079 Init(/*AliasLevel=*/ 0); 1080 init_tf((*generator)()); 1081 1082 { 1083 // The following is a dummy for the sake of GraphKit::gen_stub 1084 Unique_Node_List for_igvn(comp_arena()); 1085 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1086 PhaseGVN gvn(Thread::current()->resource_area(),255); 1087 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1088 gvn.transform_no_reclaim(top()); 1089 1090 GraphKit kit; 1091 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1092 } 1093 1094 NOT_PRODUCT( verify_graph_edges(); ) 1095 Code_Gen(); 1096 if (failing()) return; 1097 1098 1099 // Entry point will be accessed using compile->stub_entry_point(); 1100 if (code_buffer() == NULL) { 1101 Matcher::soft_match_failure(); 1102 } else { 1103 if (PrintAssembly && (WizardMode || Verbose)) 1104 tty->print_cr("### Stub::%s", stub_name); 1105 1106 if (!failing()) { 1107 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1108 1109 // Make the NMethod 1110 // For now we mark the frame as never safe for profile stackwalking 1111 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1112 code_buffer(), 1113 CodeOffsets::frame_never_safe, 1114 // _code_offsets.value(CodeOffsets::Frame_Complete), 1115 frame_size_in_words(), 1116 _oop_map_set, 1117 save_arg_registers); 1118 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1119 1120 _stub_entry_point = rs->entry_point(); 1121 } 1122 } 1123 } 1124 1125 //------------------------------Init------------------------------------------- 1126 // Prepare for a single compilation 1127 void Compile::Init(int aliaslevel) { 1128 _unique = 0; 1129 _regalloc = NULL; 1130 1131 _tf = NULL; // filled in later 1132 _top = NULL; // cached later 1133 _matcher = NULL; // filled in later 1134 _cfg = NULL; // filled in later 1135 1136 set_24_bit_selection_and_mode(Use24BitFP, false); 1137 1138 _node_note_array = NULL; 1139 _default_node_notes = NULL; 1140 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1141 1142 _immutable_memory = NULL; // filled in at first inquiry 1143 1144 // Globally visible Nodes 1145 // First set TOP to NULL to give safe behavior during creation of RootNode 1146 set_cached_top_node(NULL); 1147 set_root(new RootNode()); 1148 // Now that you have a Root to point to, create the real TOP 1149 set_cached_top_node( new ConNode(Type::TOP) ); 1150 set_recent_alloc(NULL, NULL); 1151 1152 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1153 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1154 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1155 env()->set_dependencies(new Dependencies(env())); 1156 1157 _fixed_slots = 0; 1158 set_has_split_ifs(false); 1159 set_has_loops(has_method() && method()->has_loops()); // first approximation 1160 set_has_stringbuilder(false); 1161 set_has_boxed_value(false); 1162 _trap_can_recompile = false; // no traps emitted yet 1163 _major_progress = true; // start out assuming good things will happen 1164 set_has_unsafe_access(false); 1165 set_max_vector_size(0); 1166 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1167 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1168 set_decompile_count(0); 1169 1170 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1171 _loop_opts_cnt = LoopOptsCount; 1172 _has_flattened_accesses = false; 1173 _flattened_accesses_share_alias = true; 1174 1175 set_do_inlining(Inline); 1176 set_max_inline_size(MaxInlineSize); 1177 set_freq_inline_size(FreqInlineSize); 1178 set_do_scheduling(OptoScheduling); 1179 set_do_count_invocations(false); 1180 set_do_method_data_update(false); 1181 1182 set_do_vector_loop(false); 1183 1184 if (AllowVectorizeOnDemand) { 1185 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1186 set_do_vector_loop(true); 1187 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());}) 1188 } else if (has_method() && method()->name() != 0 && 1189 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1190 set_do_vector_loop(true); 1191 } 1192 } 1193 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1194 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());}) 1195 1196 set_age_code(has_method() && method()->profile_aging()); 1197 set_rtm_state(NoRTM); // No RTM lock eliding by default 1198 _max_node_limit = _directive->MaxNodeLimitOption; 1199 1200 #if INCLUDE_RTM_OPT 1201 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1202 int rtm_state = method()->method_data()->rtm_state(); 1203 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1204 // Don't generate RTM lock eliding code. 1205 set_rtm_state(NoRTM); 1206 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1207 // Generate RTM lock eliding code without abort ratio calculation code. 1208 set_rtm_state(UseRTM); 1209 } else if (UseRTMDeopt) { 1210 // Generate RTM lock eliding code and include abort ratio calculation 1211 // code if UseRTMDeopt is on. 1212 set_rtm_state(ProfileRTM); 1213 } 1214 } 1215 #endif 1216 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1217 set_clinit_barrier_on_entry(true); 1218 } 1219 if (debug_info()->recording_non_safepoints()) { 1220 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1221 (comp_arena(), 8, 0, NULL)); 1222 set_default_node_notes(Node_Notes::make(this)); 1223 } 1224 1225 // // -- Initialize types before each compile -- 1226 // // Update cached type information 1227 // if( _method && _method->constants() ) 1228 // Type::update_loaded_types(_method, _method->constants()); 1229 1230 // Init alias_type map. 1231 if (!_do_escape_analysis && aliaslevel == 3) 1232 aliaslevel = 2; // No unique types without escape analysis 1233 _AliasLevel = aliaslevel; 1234 const int grow_ats = 16; 1235 _max_alias_types = grow_ats; 1236 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1237 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1238 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1239 { 1240 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1241 } 1242 // Initialize the first few types. 1243 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1244 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1245 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1246 _num_alias_types = AliasIdxRaw+1; 1247 // Zero out the alias type cache. 1248 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1249 // A NULL adr_type hits in the cache right away. Preload the right answer. 1250 probe_alias_cache(NULL)->_index = AliasIdxTop; 1251 1252 _intrinsics = NULL; 1253 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1254 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1255 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1256 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1257 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1258 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1259 register_library_intrinsics(); 1260 } 1261 1262 //---------------------------init_start---------------------------------------- 1263 // Install the StartNode on this compile object. 1264 void Compile::init_start(StartNode* s) { 1265 if (failing()) 1266 return; // already failing 1267 assert(s == start(), ""); 1268 } 1269 1270 /** 1271 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1272 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1273 * the ideal graph. 1274 */ 1275 StartNode* Compile::start() const { 1276 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1277 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1278 Node* start = root()->fast_out(i); 1279 if (start->is_Start()) { 1280 return start->as_Start(); 1281 } 1282 } 1283 fatal("Did not find Start node!"); 1284 return NULL; 1285 } 1286 1287 //-------------------------------immutable_memory------------------------------------- 1288 // Access immutable memory 1289 Node* Compile::immutable_memory() { 1290 if (_immutable_memory != NULL) { 1291 return _immutable_memory; 1292 } 1293 StartNode* s = start(); 1294 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1295 Node *p = s->fast_out(i); 1296 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1297 _immutable_memory = p; 1298 return _immutable_memory; 1299 } 1300 } 1301 ShouldNotReachHere(); 1302 return NULL; 1303 } 1304 1305 //----------------------set_cached_top_node------------------------------------ 1306 // Install the cached top node, and make sure Node::is_top works correctly. 1307 void Compile::set_cached_top_node(Node* tn) { 1308 if (tn != NULL) verify_top(tn); 1309 Node* old_top = _top; 1310 _top = tn; 1311 // Calling Node::setup_is_top allows the nodes the chance to adjust 1312 // their _out arrays. 1313 if (_top != NULL) _top->setup_is_top(); 1314 if (old_top != NULL) old_top->setup_is_top(); 1315 assert(_top == NULL || top()->is_top(), ""); 1316 } 1317 1318 #ifdef ASSERT 1319 uint Compile::count_live_nodes_by_graph_walk() { 1320 Unique_Node_List useful(comp_arena()); 1321 // Get useful node list by walking the graph. 1322 identify_useful_nodes(useful); 1323 return useful.size(); 1324 } 1325 1326 void Compile::print_missing_nodes() { 1327 1328 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1329 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1330 return; 1331 } 1332 1333 // This is an expensive function. It is executed only when the user 1334 // specifies VerifyIdealNodeCount option or otherwise knows the 1335 // additional work that needs to be done to identify reachable nodes 1336 // by walking the flow graph and find the missing ones using 1337 // _dead_node_list. 1338 1339 Unique_Node_List useful(comp_arena()); 1340 // Get useful node list by walking the graph. 1341 identify_useful_nodes(useful); 1342 1343 uint l_nodes = C->live_nodes(); 1344 uint l_nodes_by_walk = useful.size(); 1345 1346 if (l_nodes != l_nodes_by_walk) { 1347 if (_log != NULL) { 1348 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1349 _log->stamp(); 1350 _log->end_head(); 1351 } 1352 VectorSet& useful_member_set = useful.member_set(); 1353 int last_idx = l_nodes_by_walk; 1354 for (int i = 0; i < last_idx; i++) { 1355 if (useful_member_set.test(i)) { 1356 if (_dead_node_list.test(i)) { 1357 if (_log != NULL) { 1358 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1359 } 1360 if (PrintIdealNodeCount) { 1361 // Print the log message to tty 1362 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1363 useful.at(i)->dump(); 1364 } 1365 } 1366 } 1367 else if (! _dead_node_list.test(i)) { 1368 if (_log != NULL) { 1369 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1370 } 1371 if (PrintIdealNodeCount) { 1372 // Print the log message to tty 1373 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1374 } 1375 } 1376 } 1377 if (_log != NULL) { 1378 _log->tail("mismatched_nodes"); 1379 } 1380 } 1381 } 1382 void Compile::record_modified_node(Node* n) { 1383 if (_modified_nodes != NULL && !_inlining_incrementally && 1384 n->outcnt() != 0 && !n->is_Con()) { 1385 _modified_nodes->push(n); 1386 } 1387 } 1388 1389 void Compile::remove_modified_node(Node* n) { 1390 if (_modified_nodes != NULL) { 1391 _modified_nodes->remove(n); 1392 } 1393 } 1394 #endif 1395 1396 #ifndef PRODUCT 1397 void Compile::verify_top(Node* tn) const { 1398 if (tn != NULL) { 1399 assert(tn->is_Con(), "top node must be a constant"); 1400 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1401 assert(tn->in(0) != NULL, "must have live top node"); 1402 } 1403 } 1404 #endif 1405 1406 1407 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1408 1409 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1410 guarantee(arr != NULL, ""); 1411 int num_blocks = arr->length(); 1412 if (grow_by < num_blocks) grow_by = num_blocks; 1413 int num_notes = grow_by * _node_notes_block_size; 1414 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1415 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1416 while (num_notes > 0) { 1417 arr->append(notes); 1418 notes += _node_notes_block_size; 1419 num_notes -= _node_notes_block_size; 1420 } 1421 assert(num_notes == 0, "exact multiple, please"); 1422 } 1423 1424 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1425 if (source == NULL || dest == NULL) return false; 1426 1427 if (dest->is_Con()) 1428 return false; // Do not push debug info onto constants. 1429 1430 #ifdef ASSERT 1431 // Leave a bread crumb trail pointing to the original node: 1432 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1433 dest->set_debug_orig(source); 1434 } 1435 #endif 1436 1437 if (node_note_array() == NULL) 1438 return false; // Not collecting any notes now. 1439 1440 // This is a copy onto a pre-existing node, which may already have notes. 1441 // If both nodes have notes, do not overwrite any pre-existing notes. 1442 Node_Notes* source_notes = node_notes_at(source->_idx); 1443 if (source_notes == NULL || source_notes->is_clear()) return false; 1444 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1445 if (dest_notes == NULL || dest_notes->is_clear()) { 1446 return set_node_notes_at(dest->_idx, source_notes); 1447 } 1448 1449 Node_Notes merged_notes = (*source_notes); 1450 // The order of operations here ensures that dest notes will win... 1451 merged_notes.update_from(dest_notes); 1452 return set_node_notes_at(dest->_idx, &merged_notes); 1453 } 1454 1455 1456 //--------------------------allow_range_check_smearing------------------------- 1457 // Gating condition for coalescing similar range checks. 1458 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1459 // single covering check that is at least as strong as any of them. 1460 // If the optimization succeeds, the simplified (strengthened) range check 1461 // will always succeed. If it fails, we will deopt, and then give up 1462 // on the optimization. 1463 bool Compile::allow_range_check_smearing() const { 1464 // If this method has already thrown a range-check, 1465 // assume it was because we already tried range smearing 1466 // and it failed. 1467 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1468 return !already_trapped; 1469 } 1470 1471 1472 //------------------------------flatten_alias_type----------------------------- 1473 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1474 int offset = tj->offset(); 1475 TypePtr::PTR ptr = tj->ptr(); 1476 1477 // Known instance (scalarizable allocation) alias only with itself. 1478 bool is_known_inst = tj->isa_oopptr() != NULL && 1479 tj->is_oopptr()->is_known_instance(); 1480 1481 // Process weird unsafe references. 1482 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1483 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass(); 1484 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops"); 1485 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1486 tj = TypeOopPtr::BOTTOM; 1487 ptr = tj->ptr(); 1488 offset = tj->offset(); 1489 } 1490 1491 // Array pointers need some flattening 1492 const TypeAryPtr *ta = tj->isa_aryptr(); 1493 if (ta && ta->is_stable()) { 1494 // Erase stability property for alias analysis. 1495 tj = ta = ta->cast_to_stable(false); 1496 } 1497 if( ta && is_known_inst ) { 1498 if ( offset != Type::OffsetBot && 1499 offset > arrayOopDesc::length_offset_in_bytes() ) { 1500 offset = Type::OffsetBot; // Flatten constant access into array body only 1501 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1502 } 1503 } else if( ta && _AliasLevel >= 2 ) { 1504 // For arrays indexed by constant indices, we flatten the alias 1505 // space to include all of the array body. Only the header, klass 1506 // and array length can be accessed un-aliased. 1507 // For flattened value type array, each field has its own slice so 1508 // we must include the field offset. 1509 if( offset != Type::OffsetBot ) { 1510 if( ta->const_oop() ) { // MethodData* or Method* 1511 offset = Type::OffsetBot; // Flatten constant access into array body 1512 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1513 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1514 // range is OK as-is. 1515 tj = ta = TypeAryPtr::RANGE; 1516 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1517 tj = TypeInstPtr::KLASS; // all klass loads look alike 1518 ta = TypeAryPtr::RANGE; // generic ignored junk 1519 ptr = TypePtr::BotPTR; 1520 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1521 tj = TypeInstPtr::MARK; 1522 ta = TypeAryPtr::RANGE; // generic ignored junk 1523 ptr = TypePtr::BotPTR; 1524 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1525 ta = tj->isa_aryptr(); 1526 } else { // Random constant offset into array body 1527 offset = Type::OffsetBot; // Flatten constant access into array body 1528 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1529 } 1530 } 1531 // Arrays of fixed size alias with arrays of unknown size. 1532 if (ta->size() != TypeInt::POS) { 1533 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1534 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1535 } 1536 // Arrays of known objects become arrays of unknown objects. 1537 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1538 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1539 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1540 } 1541 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1542 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1543 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1544 } 1545 // Initially all flattened array accesses share a single slice 1546 if (ta->elem()->isa_valuetype() && ta->elem() != TypeValueType::BOTTOM && _flattened_accesses_share_alias) { 1547 const TypeAry *tary = TypeAry::make(TypeValueType::BOTTOM, ta->size()); 1548 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot)); 1549 } 1550 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1551 // cannot be distinguished by bytecode alone. 1552 if (ta->elem() == TypeInt::BOOL) { 1553 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1554 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1555 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1556 } 1557 // During the 2nd round of IterGVN, NotNull castings are removed. 1558 // Make sure the Bottom and NotNull variants alias the same. 1559 // Also, make sure exact and non-exact variants alias the same. 1560 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1561 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1562 } 1563 } 1564 1565 // Oop pointers need some flattening 1566 const TypeInstPtr *to = tj->isa_instptr(); 1567 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1568 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1569 if( ptr == TypePtr::Constant ) { 1570 if (to->klass() != ciEnv::current()->Class_klass() || 1571 offset < k->size_helper() * wordSize) { 1572 // No constant oop pointers (such as Strings); they alias with 1573 // unknown strings. 1574 assert(!is_known_inst, "not scalarizable allocation"); 1575 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1576 } 1577 } else if( is_known_inst ) { 1578 tj = to; // Keep NotNull and klass_is_exact for instance type 1579 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1580 // During the 2nd round of IterGVN, NotNull castings are removed. 1581 // Make sure the Bottom and NotNull variants alias the same. 1582 // Also, make sure exact and non-exact variants alias the same. 1583 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1584 } 1585 if (to->speculative() != NULL) { 1586 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1587 } 1588 // Canonicalize the holder of this field 1589 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1590 // First handle header references such as a LoadKlassNode, even if the 1591 // object's klass is unloaded at compile time (4965979). 1592 if (!is_known_inst) { // Do it only for non-instance types 1593 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1594 } 1595 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1596 to = tj->is_instptr(); 1597 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1598 // Static fields are in the space above the normal instance 1599 // fields in the java.lang.Class instance. 1600 if (to->klass() != ciEnv::current()->Class_klass()) { 1601 to = NULL; 1602 tj = TypeOopPtr::BOTTOM; 1603 offset = tj->offset(); 1604 } 1605 } else { 1606 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1607 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1608 if( is_known_inst ) { 1609 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1610 } else { 1611 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1612 } 1613 } 1614 } 1615 } 1616 1617 // Klass pointers to object array klasses need some flattening 1618 const TypeKlassPtr *tk = tj->isa_klassptr(); 1619 if( tk ) { 1620 // If we are referencing a field within a Klass, we need 1621 // to assume the worst case of an Object. Both exact and 1622 // inexact types must flatten to the same alias class so 1623 // use NotNull as the PTR. 1624 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1625 1626 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1627 TypeKlassPtr::OBJECT->klass(), 1628 Type::Offset(offset)); 1629 } 1630 1631 ciKlass* klass = tk->klass(); 1632 if (klass != NULL && klass->is_obj_array_klass()) { 1633 ciKlass* k = TypeAryPtr::OOPS->klass(); 1634 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1635 k = TypeInstPtr::BOTTOM->klass(); 1636 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1637 } 1638 1639 // Check for precise loads from the primary supertype array and force them 1640 // to the supertype cache alias index. Check for generic array loads from 1641 // the primary supertype array and also force them to the supertype cache 1642 // alias index. Since the same load can reach both, we need to merge 1643 // these 2 disparate memories into the same alias class. Since the 1644 // primary supertype array is read-only, there's no chance of confusion 1645 // where we bypass an array load and an array store. 1646 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1647 if (offset == Type::OffsetBot || 1648 (offset >= primary_supers_offset && 1649 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1650 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1651 offset = in_bytes(Klass::secondary_super_cache_offset()); 1652 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1653 } 1654 } 1655 1656 // Flatten all Raw pointers together. 1657 if (tj->base() == Type::RawPtr) 1658 tj = TypeRawPtr::BOTTOM; 1659 1660 if (tj->base() == Type::AnyPtr) 1661 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1662 1663 // Flatten all to bottom for now 1664 switch( _AliasLevel ) { 1665 case 0: 1666 tj = TypePtr::BOTTOM; 1667 break; 1668 case 1: // Flatten to: oop, static, field or array 1669 switch (tj->base()) { 1670 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1671 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1672 case Type::AryPtr: // do not distinguish arrays at all 1673 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1674 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1675 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1676 default: ShouldNotReachHere(); 1677 } 1678 break; 1679 case 2: // No collapsing at level 2; keep all splits 1680 case 3: // No collapsing at level 3; keep all splits 1681 break; 1682 default: 1683 Unimplemented(); 1684 } 1685 1686 offset = tj->offset(); 1687 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1688 1689 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1690 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1691 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1692 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1693 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1694 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1695 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1696 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), 1697 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1698 assert( tj->ptr() != TypePtr::TopPTR && 1699 tj->ptr() != TypePtr::AnyNull && 1700 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1701 // assert( tj->ptr() != TypePtr::Constant || 1702 // tj->base() == Type::RawPtr || 1703 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1704 1705 return tj; 1706 } 1707 1708 void Compile::AliasType::Init(int i, const TypePtr* at) { 1709 _index = i; 1710 _adr_type = at; 1711 _field = NULL; 1712 _element = NULL; 1713 _is_rewritable = true; // default 1714 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1715 if (atoop != NULL && atoop->is_known_instance()) { 1716 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1717 _general_index = Compile::current()->get_alias_index(gt); 1718 } else { 1719 _general_index = 0; 1720 } 1721 } 1722 1723 BasicType Compile::AliasType::basic_type() const { 1724 if (element() != NULL) { 1725 const Type* element = adr_type()->is_aryptr()->elem(); 1726 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1727 } if (field() != NULL) { 1728 return field()->layout_type(); 1729 } else { 1730 return T_ILLEGAL; // unknown 1731 } 1732 } 1733 1734 //---------------------------------print_on------------------------------------ 1735 #ifndef PRODUCT 1736 void Compile::AliasType::print_on(outputStream* st) { 1737 if (index() < 10) 1738 st->print("@ <%d> ", index()); 1739 else st->print("@ <%d>", index()); 1740 st->print(is_rewritable() ? " " : " RO"); 1741 int offset = adr_type()->offset(); 1742 if (offset == Type::OffsetBot) 1743 st->print(" +any"); 1744 else st->print(" +%-3d", offset); 1745 st->print(" in "); 1746 adr_type()->dump_on(st); 1747 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1748 if (field() != NULL && tjp) { 1749 if (tjp->klass() != field()->holder() || 1750 tjp->offset() != field()->offset_in_bytes()) { 1751 st->print(" != "); 1752 field()->print(); 1753 st->print(" ***"); 1754 } 1755 } 1756 } 1757 1758 void print_alias_types() { 1759 Compile* C = Compile::current(); 1760 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1761 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1762 C->alias_type(idx)->print_on(tty); 1763 tty->cr(); 1764 } 1765 } 1766 #endif 1767 1768 1769 //----------------------------probe_alias_cache-------------------------------- 1770 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1771 intptr_t key = (intptr_t) adr_type; 1772 key ^= key >> logAliasCacheSize; 1773 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1774 } 1775 1776 1777 //-----------------------------grow_alias_types-------------------------------- 1778 void Compile::grow_alias_types() { 1779 const int old_ats = _max_alias_types; // how many before? 1780 const int new_ats = old_ats; // how many more? 1781 const int grow_ats = old_ats+new_ats; // how many now? 1782 _max_alias_types = grow_ats; 1783 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1784 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1785 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1786 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1787 } 1788 1789 1790 //--------------------------------find_alias_type------------------------------ 1791 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) { 1792 if (_AliasLevel == 0) 1793 return alias_type(AliasIdxBot); 1794 1795 AliasCacheEntry* ace = NULL; 1796 if (!uncached) { 1797 ace = probe_alias_cache(adr_type); 1798 if (ace->_adr_type == adr_type) { 1799 return alias_type(ace->_index); 1800 } 1801 } 1802 1803 // Handle special cases. 1804 if (adr_type == NULL) return alias_type(AliasIdxTop); 1805 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1806 1807 // Do it the slow way. 1808 const TypePtr* flat = flatten_alias_type(adr_type); 1809 1810 #ifdef ASSERT 1811 { 1812 ResourceMark rm; 1813 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1814 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1815 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1816 Type::str(adr_type)); 1817 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1818 const TypeOopPtr* foop = flat->is_oopptr(); 1819 // Scalarizable allocations have exact klass always. 1820 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1821 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1822 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1823 Type::str(foop), Type::str(xoop)); 1824 } 1825 } 1826 #endif 1827 1828 int idx = AliasIdxTop; 1829 for (int i = 0; i < num_alias_types(); i++) { 1830 if (alias_type(i)->adr_type() == flat) { 1831 idx = i; 1832 break; 1833 } 1834 } 1835 1836 if (idx == AliasIdxTop) { 1837 if (no_create) return NULL; 1838 // Grow the array if necessary. 1839 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1840 // Add a new alias type. 1841 idx = _num_alias_types++; 1842 _alias_types[idx]->Init(idx, flat); 1843 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1844 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1845 if (flat->isa_instptr()) { 1846 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1847 && flat->is_instptr()->klass() == env()->Class_klass()) 1848 alias_type(idx)->set_rewritable(false); 1849 } 1850 ciField* field = NULL; 1851 if (flat->isa_aryptr()) { 1852 #ifdef ASSERT 1853 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1854 // (T_BYTE has the weakest alignment and size restrictions...) 1855 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1856 #endif 1857 const Type* elemtype = flat->is_aryptr()->elem(); 1858 if (flat->offset() == TypePtr::OffsetBot) { 1859 alias_type(idx)->set_element(elemtype); 1860 } 1861 int field_offset = flat->is_aryptr()->field_offset().get(); 1862 if (elemtype->isa_valuetype() && 1863 elemtype->value_klass() != NULL && 1864 field_offset != Type::OffsetBot) { 1865 ciValueKlass* vk = elemtype->value_klass(); 1866 field_offset += vk->first_field_offset(); 1867 field = vk->get_field_by_offset(field_offset, false); 1868 } 1869 } 1870 if (flat->isa_klassptr()) { 1871 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1872 alias_type(idx)->set_rewritable(false); 1873 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1874 alias_type(idx)->set_rewritable(false); 1875 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1876 alias_type(idx)->set_rewritable(false); 1877 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1878 alias_type(idx)->set_rewritable(false); 1879 if (flat->offset() == in_bytes(Klass::layout_helper_offset())) 1880 alias_type(idx)->set_rewritable(false); 1881 } 1882 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1883 // but the base pointer type is not distinctive enough to identify 1884 // references into JavaThread.) 1885 1886 // Check for final fields. 1887 const TypeInstPtr* tinst = flat->isa_instptr(); 1888 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1889 if (tinst->const_oop() != NULL && 1890 tinst->klass() == ciEnv::current()->Class_klass() && 1891 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1892 // static field 1893 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1894 field = k->get_field_by_offset(tinst->offset(), true); 1895 } else if (tinst->klass()->is_valuetype()) { 1896 // Value type field 1897 ciValueKlass* vk = tinst->value_klass(); 1898 field = vk->get_field_by_offset(tinst->offset(), false); 1899 } else { 1900 ciInstanceKlass* k = tinst->klass()->as_instance_klass(); 1901 field = k->get_field_by_offset(tinst->offset(), false); 1902 } 1903 } 1904 assert(field == NULL || 1905 original_field == NULL || 1906 (field->holder() == original_field->holder() && 1907 field->offset() == original_field->offset() && 1908 field->is_static() == original_field->is_static()), "wrong field?"); 1909 // Set field() and is_rewritable() attributes. 1910 if (field != NULL) alias_type(idx)->set_field(field); 1911 } 1912 1913 // Fill the cache for next time. 1914 if (!uncached) { 1915 ace->_adr_type = adr_type; 1916 ace->_index = idx; 1917 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1918 1919 // Might as well try to fill the cache for the flattened version, too. 1920 AliasCacheEntry* face = probe_alias_cache(flat); 1921 if (face->_adr_type == NULL) { 1922 face->_adr_type = flat; 1923 face->_index = idx; 1924 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1925 } 1926 } 1927 1928 return alias_type(idx); 1929 } 1930 1931 1932 Compile::AliasType* Compile::alias_type(ciField* field) { 1933 const TypeOopPtr* t; 1934 if (field->is_static()) 1935 t = TypeInstPtr::make(field->holder()->java_mirror()); 1936 else 1937 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1938 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1939 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1940 return atp; 1941 } 1942 1943 1944 //------------------------------have_alias_type-------------------------------- 1945 bool Compile::have_alias_type(const TypePtr* adr_type) { 1946 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1947 if (ace->_adr_type == adr_type) { 1948 return true; 1949 } 1950 1951 // Handle special cases. 1952 if (adr_type == NULL) return true; 1953 if (adr_type == TypePtr::BOTTOM) return true; 1954 1955 return find_alias_type(adr_type, true, NULL) != NULL; 1956 } 1957 1958 //-----------------------------must_alias-------------------------------------- 1959 // True if all values of the given address type are in the given alias category. 1960 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1961 if (alias_idx == AliasIdxBot) return true; // the universal category 1962 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1963 if (alias_idx == AliasIdxTop) return false; // the empty category 1964 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1965 1966 // the only remaining possible overlap is identity 1967 int adr_idx = get_alias_index(adr_type); 1968 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1969 assert(adr_idx == alias_idx || 1970 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1971 && adr_type != TypeOopPtr::BOTTOM), 1972 "should not be testing for overlap with an unsafe pointer"); 1973 return adr_idx == alias_idx; 1974 } 1975 1976 //------------------------------can_alias-------------------------------------- 1977 // True if any values of the given address type are in the given alias category. 1978 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1979 if (alias_idx == AliasIdxTop) return false; // the empty category 1980 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1981 if (alias_idx == AliasIdxBot) return true; // the universal category 1982 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1983 1984 // the only remaining possible overlap is identity 1985 int adr_idx = get_alias_index(adr_type); 1986 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1987 return adr_idx == alias_idx; 1988 } 1989 1990 1991 1992 //---------------------------pop_warm_call------------------------------------- 1993 WarmCallInfo* Compile::pop_warm_call() { 1994 WarmCallInfo* wci = _warm_calls; 1995 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1996 return wci; 1997 } 1998 1999 //----------------------------Inline_Warm-------------------------------------- 2000 int Compile::Inline_Warm() { 2001 // If there is room, try to inline some more warm call sites. 2002 // %%% Do a graph index compaction pass when we think we're out of space? 2003 if (!InlineWarmCalls) return 0; 2004 2005 int calls_made_hot = 0; 2006 int room_to_grow = NodeCountInliningCutoff - unique(); 2007 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 2008 int amount_grown = 0; 2009 WarmCallInfo* call; 2010 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 2011 int est_size = (int)call->size(); 2012 if (est_size > (room_to_grow - amount_grown)) { 2013 // This one won't fit anyway. Get rid of it. 2014 call->make_cold(); 2015 continue; 2016 } 2017 call->make_hot(); 2018 calls_made_hot++; 2019 amount_grown += est_size; 2020 amount_to_grow -= est_size; 2021 } 2022 2023 if (calls_made_hot > 0) set_major_progress(); 2024 return calls_made_hot; 2025 } 2026 2027 2028 //----------------------------Finish_Warm-------------------------------------- 2029 void Compile::Finish_Warm() { 2030 if (!InlineWarmCalls) return; 2031 if (failing()) return; 2032 if (warm_calls() == NULL) return; 2033 2034 // Clean up loose ends, if we are out of space for inlining. 2035 WarmCallInfo* call; 2036 while ((call = pop_warm_call()) != NULL) { 2037 call->make_cold(); 2038 } 2039 } 2040 2041 //---------------------cleanup_loop_predicates----------------------- 2042 // Remove the opaque nodes that protect the predicates so that all unused 2043 // checks and uncommon_traps will be eliminated from the ideal graph 2044 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 2045 if (predicate_count()==0) return; 2046 for (int i = predicate_count(); i > 0; i--) { 2047 Node * n = predicate_opaque1_node(i-1); 2048 assert(n->Opcode() == Op_Opaque1, "must be"); 2049 igvn.replace_node(n, n->in(1)); 2050 } 2051 assert(predicate_count()==0, "should be clean!"); 2052 } 2053 2054 void Compile::add_range_check_cast(Node* n) { 2055 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2056 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 2057 _range_check_casts->append(n); 2058 } 2059 2060 // Remove all range check dependent CastIINodes. 2061 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 2062 for (int i = range_check_cast_count(); i > 0; i--) { 2063 Node* cast = range_check_cast_node(i-1); 2064 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2065 igvn.replace_node(cast, cast->in(1)); 2066 } 2067 assert(range_check_cast_count() == 0, "should be empty"); 2068 } 2069 2070 void Compile::add_opaque4_node(Node* n) { 2071 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 2072 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 2073 _opaque4_nodes->append(n); 2074 } 2075 2076 // Remove all Opaque4 nodes. 2077 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 2078 for (int i = opaque4_count(); i > 0; i--) { 2079 Node* opaq = opaque4_node(i-1); 2080 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2081 igvn.replace_node(opaq, opaq->in(2)); 2082 } 2083 assert(opaque4_count() == 0, "should be empty"); 2084 } 2085 2086 void Compile::add_value_type(Node* n) { 2087 assert(n->is_ValueTypeBase(), "unexpected node"); 2088 if (_value_type_nodes != NULL) { 2089 _value_type_nodes->push(n); 2090 } 2091 } 2092 2093 void Compile::remove_value_type(Node* n) { 2094 assert(n->is_ValueTypeBase(), "unexpected node"); 2095 if (_value_type_nodes != NULL) { 2096 _value_type_nodes->remove(n); 2097 } 2098 } 2099 2100 // Does the return value keep otherwise useless value type allocations 2101 // alive? 2102 static bool return_val_keeps_allocations_alive(Node* ret_val) { 2103 ResourceMark rm; 2104 Unique_Node_List wq; 2105 wq.push(ret_val); 2106 bool some_allocations = false; 2107 for (uint i = 0; i < wq.size(); i++) { 2108 Node* n = wq.at(i); 2109 assert(!n->is_ValueTypeBase(), "chain of value type nodes"); 2110 if (n->outcnt() > 1) { 2111 // Some other use for the allocation 2112 return false; 2113 } else if (n->is_Phi()) { 2114 for (uint j = 1; j < n->req(); j++) { 2115 wq.push(n->in(j)); 2116 } 2117 } else if (n->is_CheckCastPP() && 2118 n->in(1)->is_Proj() && 2119 n->in(1)->in(0)->is_Allocate()) { 2120 some_allocations = true; 2121 } 2122 } 2123 return some_allocations; 2124 } 2125 2126 void Compile::process_value_types(PhaseIterGVN &igvn) { 2127 // Make value types scalar in safepoints 2128 while (_value_type_nodes->size() != 0) { 2129 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 2130 vt->make_scalar_in_safepoints(&igvn); 2131 if (vt->is_ValueTypePtr()) { 2132 igvn.replace_node(vt, vt->get_oop()); 2133 } else if (vt->outcnt() == 0) { 2134 igvn.remove_dead_node(vt); 2135 } 2136 } 2137 _value_type_nodes = NULL; 2138 if (tf()->returns_value_type_as_fields()) { 2139 Node* ret = NULL; 2140 for (uint i = 1; i < root()->req(); i++){ 2141 Node* in = root()->in(i); 2142 if (in->Opcode() == Op_Return) { 2143 assert(ret == NULL, "only one return"); 2144 ret = in; 2145 } 2146 } 2147 if (ret != NULL) { 2148 Node* ret_val = ret->in(TypeFunc::Parms); 2149 if (igvn.type(ret_val)->isa_oopptr() && 2150 return_val_keeps_allocations_alive(ret_val)) { 2151 igvn.replace_input_of(ret, TypeFunc::Parms, ValueTypeNode::tagged_klass(igvn.type(ret_val)->value_klass(), igvn)); 2152 assert(ret_val->outcnt() == 0, "should be dead now"); 2153 igvn.remove_dead_node(ret_val); 2154 } 2155 } 2156 } 2157 igvn.optimize(); 2158 } 2159 2160 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) { 2161 if (!_has_flattened_accesses) { 2162 return; 2163 } 2164 // Initially, all flattened array accesses share the same slice to 2165 // keep dependencies with Object[] array accesses (that could be 2166 // to a flattened array) correct. We're done with parsing so we 2167 // now know all flattened array accesses in this compile 2168 // unit. Let's move flattened array accesses to their own slice, 2169 // one per element field. This should help memory access 2170 // optimizations. 2171 ResourceMark rm; 2172 Unique_Node_List wq; 2173 wq.push(root()); 2174 2175 Node_List mergememnodes; 2176 Node_List memnodes; 2177 2178 // Alias index currently shared by all flattened memory accesses 2179 int index = get_alias_index(TypeAryPtr::VALUES); 2180 2181 // Find MergeMem nodes and flattened array accesses 2182 for (uint i = 0; i < wq.size(); i++) { 2183 Node* n = wq.at(i); 2184 if (n->is_Mem()) { 2185 const TypePtr* adr_type = get_adr_type(get_alias_index(n->adr_type())); 2186 if (adr_type == TypeAryPtr::VALUES) { 2187 memnodes.push(n); 2188 } 2189 } else if (n->is_MergeMem()) { 2190 MergeMemNode* mm = n->as_MergeMem(); 2191 if (mm->memory_at(index) != mm->base_memory()) { 2192 mergememnodes.push(n); 2193 } 2194 } 2195 for (uint j = 0; j < n->req(); j++) { 2196 Node* m = n->in(j); 2197 if (m != NULL) { 2198 wq.push(m); 2199 } 2200 } 2201 } 2202 2203 if (memnodes.size() > 0) { 2204 _flattened_accesses_share_alias = false; 2205 2206 // We are going to change the slice for the flattened array 2207 // accesses so we need to clear the cache entries that refer to 2208 // them. 2209 for (uint i = 0; i < AliasCacheSize; i++) { 2210 AliasCacheEntry* ace = &_alias_cache[i]; 2211 if (ace->_adr_type != NULL && 2212 ace->_adr_type->isa_aryptr() && 2213 ace->_adr_type->is_aryptr()->elem()->isa_valuetype()) { 2214 ace->_adr_type = NULL; 2215 ace->_index = 0; 2216 } 2217 } 2218 2219 // Find what aliases we are going to add 2220 int start_alias = num_alias_types()-1; 2221 int stop_alias = 0; 2222 2223 for (uint i = 0; i < memnodes.size(); i++) { 2224 Node* m = memnodes.at(i); 2225 const TypePtr* adr_type = m->adr_type(); 2226 #ifdef ASSERT 2227 m->as_Mem()->set_adr_type(adr_type); 2228 #endif 2229 int idx = get_alias_index(adr_type); 2230 start_alias = MIN2(start_alias, idx); 2231 stop_alias = MAX2(stop_alias, idx); 2232 } 2233 2234 assert(stop_alias >= start_alias, "should have expanded aliases"); 2235 2236 Node_Stack stack(0); 2237 #ifdef ASSERT 2238 VectorSet seen(Thread::current()->resource_area()); 2239 #endif 2240 // Now let's fix the memory graph so each flattened array access 2241 // is moved to the right slice. Start from the MergeMem nodes. 2242 uint last = unique(); 2243 for (uint i = 0; i < mergememnodes.size(); i++) { 2244 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem(); 2245 Node* n = current->memory_at(index); 2246 MergeMemNode* mm = NULL; 2247 do { 2248 // Follow memory edges through memory accesses, phis and 2249 // narrow membars and push nodes on the stack. Once we hit 2250 // bottom memory, we pop element off the stack one at a 2251 // time, in reverse order, and move them to the right slice 2252 // by changing their memory edges. 2253 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::VALUES) { 2254 assert(!seen.test_set(n->_idx), ""); 2255 // Uses (a load for instance) will need to be moved to the 2256 // right slice as well and will get a new memory state 2257 // that we don't know yet. The use could also be the 2258 // backedge of a loop. We put a place holder node between 2259 // the memory node and its uses. We replace that place 2260 // holder with the correct memory state once we know it, 2261 // i.e. when nodes are popped off the stack. Using the 2262 // place holder make the logic work in the presence of 2263 // loops. 2264 if (n->outcnt() > 1) { 2265 Node* place_holder = NULL; 2266 assert(!n->has_out_with(Op_Node), ""); 2267 for (DUIterator k = n->outs(); n->has_out(k); k++) { 2268 Node* u = n->out(k); 2269 if (u != current && u->_idx < last) { 2270 bool success = false; 2271 for (uint l = 0; l < u->req(); l++) { 2272 if (!stack.is_empty() && u == stack.node() && l == stack.index()) { 2273 continue; 2274 } 2275 Node* in = u->in(l); 2276 if (in == n) { 2277 if (place_holder == NULL) { 2278 place_holder = new Node(1); 2279 place_holder->init_req(0, n); 2280 } 2281 igvn.replace_input_of(u, l, place_holder); 2282 success = true; 2283 } 2284 } 2285 if (success) { 2286 --k; 2287 } 2288 } 2289 } 2290 } 2291 if (n->is_Phi()) { 2292 stack.push(n, 1); 2293 n = n->in(1); 2294 } else if (n->is_Mem()) { 2295 stack.push(n, n->req()); 2296 n = n->in(MemNode::Memory); 2297 } else { 2298 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, ""); 2299 stack.push(n, n->req()); 2300 n = n->in(0)->in(TypeFunc::Memory); 2301 } 2302 } else { 2303 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), ""); 2304 // Build a new MergeMem node to carry the new memory state 2305 // as we build it. IGVN should fold extraneous MergeMem 2306 // nodes. 2307 mm = MergeMemNode::make(n); 2308 igvn.register_new_node_with_optimizer(mm); 2309 while (stack.size() > 0) { 2310 Node* m = stack.node(); 2311 uint idx = stack.index(); 2312 if (m->is_Mem()) { 2313 // Move memory node to its new slice 2314 const TypePtr* adr_type = m->adr_type(); 2315 int alias = get_alias_index(adr_type); 2316 Node* prev = mm->memory_at(alias); 2317 igvn.replace_input_of(m, MemNode::Memory, prev); 2318 mm->set_memory_at(alias, m); 2319 } else if (m->is_Phi()) { 2320 // We need as many new phis as there are new aliases 2321 igvn.replace_input_of(m, idx, mm); 2322 if (idx == m->req()-1) { 2323 Node* r = m->in(0); 2324 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2325 const Type* adr_type = get_adr_type(j); 2326 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) { 2327 continue; 2328 } 2329 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j)); 2330 igvn.register_new_node_with_optimizer(phi); 2331 for (uint k = 1; k < m->req(); k++) { 2332 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j)); 2333 } 2334 mm->set_memory_at(j, phi); 2335 } 2336 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2337 igvn.register_new_node_with_optimizer(base_phi); 2338 for (uint k = 1; k < m->req(); k++) { 2339 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory()); 2340 } 2341 mm->set_base_memory(base_phi); 2342 } 2343 } else { 2344 // This is a MemBarCPUOrder node from 2345 // Parse::array_load()/Parse::array_store(), in the 2346 // branch that handles flattened arrays hidden under 2347 // an Object[] array. We also need one new membar per 2348 // new alias to keep the unknown access that the 2349 // membars protect properly ordered with accesses to 2350 // known flattened array. 2351 assert(m->is_Proj(), "projection expected"); 2352 Node* ctrl = m->in(0)->in(TypeFunc::Control); 2353 igvn.replace_input_of(m->in(0), TypeFunc::Control, top()); 2354 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2355 const Type* adr_type = get_adr_type(j); 2356 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) { 2357 continue; 2358 } 2359 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL); 2360 igvn.register_new_node_with_optimizer(mb); 2361 Node* mem = mm->memory_at(j); 2362 mb->init_req(TypeFunc::Control, ctrl); 2363 mb->init_req(TypeFunc::Memory, mem); 2364 ctrl = new ProjNode(mb, TypeFunc::Control); 2365 igvn.register_new_node_with_optimizer(ctrl); 2366 mem = new ProjNode(mb, TypeFunc::Memory); 2367 igvn.register_new_node_with_optimizer(mem); 2368 mm->set_memory_at(j, mem); 2369 } 2370 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl); 2371 } 2372 if (idx < m->req()-1) { 2373 idx += 1; 2374 stack.set_index(idx); 2375 n = m->in(idx); 2376 break; 2377 } 2378 // Take care of place holder nodes 2379 if (m->has_out_with(Op_Node)) { 2380 Node* place_holder = m->find_out_with(Op_Node); 2381 if (place_holder != NULL) { 2382 Node* mm_clone = mm->clone(); 2383 igvn.register_new_node_with_optimizer(mm_clone); 2384 Node* hook = new Node(1); 2385 hook->init_req(0, mm); 2386 igvn.replace_node(place_holder, mm_clone); 2387 hook->destruct(); 2388 } 2389 assert(!m->has_out_with(Op_Node), "place holder should be gone now"); 2390 } 2391 stack.pop(); 2392 } 2393 } 2394 } while(stack.size() > 0); 2395 // Fix the memory state at the MergeMem we started from 2396 igvn.rehash_node_delayed(current); 2397 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2398 const Type* adr_type = get_adr_type(j); 2399 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) { 2400 continue; 2401 } 2402 current->set_memory_at(j, mm); 2403 } 2404 current->set_memory_at(index, current->base_memory()); 2405 } 2406 igvn.optimize(); 2407 } 2408 print_method(PHASE_SPLIT_VALUES_ARRAY, 2); 2409 } 2410 2411 2412 // StringOpts and late inlining of string methods 2413 void Compile::inline_string_calls(bool parse_time) { 2414 { 2415 // remove useless nodes to make the usage analysis simpler 2416 ResourceMark rm; 2417 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2418 } 2419 2420 { 2421 ResourceMark rm; 2422 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2423 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2424 print_method(PHASE_AFTER_STRINGOPTS, 3); 2425 } 2426 2427 // now inline anything that we skipped the first time around 2428 if (!parse_time) { 2429 _late_inlines_pos = _late_inlines.length(); 2430 } 2431 2432 while (_string_late_inlines.length() > 0) { 2433 CallGenerator* cg = _string_late_inlines.pop(); 2434 cg->do_late_inline(); 2435 if (failing()) return; 2436 } 2437 _string_late_inlines.trunc_to(0); 2438 } 2439 2440 // Late inlining of boxing methods 2441 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2442 if (_boxing_late_inlines.length() > 0) { 2443 assert(has_boxed_value(), "inconsistent"); 2444 2445 PhaseGVN* gvn = initial_gvn(); 2446 set_inlining_incrementally(true); 2447 2448 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2449 for_igvn()->clear(); 2450 gvn->replace_with(&igvn); 2451 2452 _late_inlines_pos = _late_inlines.length(); 2453 2454 while (_boxing_late_inlines.length() > 0) { 2455 CallGenerator* cg = _boxing_late_inlines.pop(); 2456 cg->do_late_inline(); 2457 if (failing()) return; 2458 } 2459 _boxing_late_inlines.trunc_to(0); 2460 2461 inline_incrementally_cleanup(igvn); 2462 2463 set_inlining_incrementally(false); 2464 } 2465 } 2466 2467 bool Compile::inline_incrementally_one() { 2468 assert(IncrementalInline, "incremental inlining should be on"); 2469 2470 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2471 set_inlining_progress(false); 2472 set_do_cleanup(false); 2473 int i = 0; 2474 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2475 CallGenerator* cg = _late_inlines.at(i); 2476 _late_inlines_pos = i+1; 2477 cg->do_late_inline(); 2478 if (failing()) return false; 2479 } 2480 int j = 0; 2481 for (; i < _late_inlines.length(); i++, j++) { 2482 _late_inlines.at_put(j, _late_inlines.at(i)); 2483 } 2484 _late_inlines.trunc_to(j); 2485 assert(inlining_progress() || _late_inlines.length() == 0, ""); 2486 2487 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2488 2489 set_inlining_progress(false); 2490 set_do_cleanup(false); 2491 return (_late_inlines.length() > 0) && !needs_cleanup; 2492 } 2493 2494 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2495 { 2496 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2497 ResourceMark rm; 2498 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2499 } 2500 { 2501 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2502 igvn = PhaseIterGVN(initial_gvn()); 2503 igvn.optimize(); 2504 } 2505 } 2506 2507 // Perform incremental inlining until bound on number of live nodes is reached 2508 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2509 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2510 2511 set_inlining_incrementally(true); 2512 uint low_live_nodes = 0; 2513 2514 while (_late_inlines.length() > 0) { 2515 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2516 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2517 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2518 // PhaseIdealLoop is expensive so we only try it once we are 2519 // out of live nodes and we only try it again if the previous 2520 // helped got the number of nodes down significantly 2521 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2522 if (failing()) return; 2523 low_live_nodes = live_nodes(); 2524 _major_progress = true; 2525 } 2526 2527 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2528 break; // finish 2529 } 2530 } 2531 2532 for_igvn()->clear(); 2533 initial_gvn()->replace_with(&igvn); 2534 2535 while (inline_incrementally_one()) { 2536 assert(!failing(), "inconsistent"); 2537 } 2538 2539 if (failing()) return; 2540 2541 inline_incrementally_cleanup(igvn); 2542 2543 if (failing()) return; 2544 } 2545 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2546 2547 if (_string_late_inlines.length() > 0) { 2548 assert(has_stringbuilder(), "inconsistent"); 2549 for_igvn()->clear(); 2550 initial_gvn()->replace_with(&igvn); 2551 2552 inline_string_calls(false); 2553 2554 if (failing()) return; 2555 2556 inline_incrementally_cleanup(igvn); 2557 } 2558 2559 set_inlining_incrementally(false); 2560 } 2561 2562 2563 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2564 if(_loop_opts_cnt > 0) { 2565 debug_only( int cnt = 0; ); 2566 while(major_progress() && (_loop_opts_cnt > 0)) { 2567 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2568 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2569 PhaseIdealLoop::optimize(igvn, mode); 2570 _loop_opts_cnt--; 2571 if (failing()) return false; 2572 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2573 } 2574 } 2575 return true; 2576 } 2577 2578 // Remove edges from "root" to each SafePoint at a backward branch. 2579 // They were inserted during parsing (see add_safepoint()) to make 2580 // infinite loops without calls or exceptions visible to root, i.e., 2581 // useful. 2582 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2583 Node *r = root(); 2584 if (r != NULL) { 2585 for (uint i = r->req(); i < r->len(); ++i) { 2586 Node *n = r->in(i); 2587 if (n != NULL && n->is_SafePoint()) { 2588 r->rm_prec(i); 2589 if (n->outcnt() == 0) { 2590 igvn.remove_dead_node(n); 2591 } 2592 --i; 2593 } 2594 } 2595 } 2596 } 2597 2598 //------------------------------Optimize--------------------------------------- 2599 // Given a graph, optimize it. 2600 void Compile::Optimize() { 2601 TracePhase tp("optimizer", &timers[_t_optimizer]); 2602 2603 #ifndef PRODUCT 2604 if (_directive->BreakAtCompileOption) { 2605 BREAKPOINT; 2606 } 2607 2608 #endif 2609 2610 #ifdef ASSERT 2611 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2612 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2613 #endif 2614 2615 ResourceMark rm; 2616 2617 print_inlining_reinit(); 2618 2619 NOT_PRODUCT( verify_graph_edges(); ) 2620 2621 print_method(PHASE_AFTER_PARSING); 2622 2623 { 2624 // Iterative Global Value Numbering, including ideal transforms 2625 // Initialize IterGVN with types and values from parse-time GVN 2626 PhaseIterGVN igvn(initial_gvn()); 2627 #ifdef ASSERT 2628 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2629 #endif 2630 { 2631 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2632 igvn.optimize(); 2633 } 2634 2635 if (failing()) return; 2636 2637 print_method(PHASE_ITER_GVN1, 2); 2638 2639 inline_incrementally(igvn); 2640 2641 print_method(PHASE_INCREMENTAL_INLINE, 2); 2642 2643 if (failing()) return; 2644 2645 if (eliminate_boxing()) { 2646 // Inline valueOf() methods now. 2647 inline_boxing_calls(igvn); 2648 2649 if (AlwaysIncrementalInline) { 2650 inline_incrementally(igvn); 2651 } 2652 2653 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2654 2655 if (failing()) return; 2656 } 2657 2658 // Now that all inlining is over, cut edge from root to loop 2659 // safepoints 2660 remove_root_to_sfpts_edges(igvn); 2661 2662 // Remove the speculative part of types and clean up the graph from 2663 // the extra CastPP nodes whose only purpose is to carry them. Do 2664 // that early so that optimizations are not disrupted by the extra 2665 // CastPP nodes. 2666 remove_speculative_types(igvn); 2667 2668 // No more new expensive nodes will be added to the list from here 2669 // so keep only the actual candidates for optimizations. 2670 cleanup_expensive_nodes(igvn); 2671 2672 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2673 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2674 initial_gvn()->replace_with(&igvn); 2675 for_igvn()->clear(); 2676 Unique_Node_List new_worklist(C->comp_arena()); 2677 { 2678 ResourceMark rm; 2679 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2680 } 2681 set_for_igvn(&new_worklist); 2682 igvn = PhaseIterGVN(initial_gvn()); 2683 igvn.optimize(); 2684 } 2685 2686 if (_value_type_nodes->size() > 0) { 2687 // Do this once all inlining is over to avoid getting inconsistent debug info 2688 process_value_types(igvn); 2689 } 2690 2691 adjust_flattened_array_access_aliases(igvn); 2692 2693 // Perform escape analysis 2694 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2695 if (has_loops()) { 2696 // Cleanup graph (remove dead nodes). 2697 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2698 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2699 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2700 if (failing()) return; 2701 } 2702 ConnectionGraph::do_analysis(this, &igvn); 2703 2704 if (failing()) return; 2705 2706 // Optimize out fields loads from scalar replaceable allocations. 2707 igvn.optimize(); 2708 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2709 2710 if (failing()) return; 2711 2712 if (congraph() != NULL && macro_count() > 0) { 2713 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2714 PhaseMacroExpand mexp(igvn); 2715 mexp.eliminate_macro_nodes(); 2716 igvn.set_delay_transform(false); 2717 2718 igvn.optimize(); 2719 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2720 2721 if (failing()) return; 2722 } 2723 } 2724 2725 // Loop transforms on the ideal graph. Range Check Elimination, 2726 // peeling, unrolling, etc. 2727 2728 // Set loop opts counter 2729 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2730 { 2731 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2732 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2733 _loop_opts_cnt--; 2734 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2735 if (failing()) return; 2736 } 2737 // Loop opts pass if partial peeling occurred in previous pass 2738 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2739 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2740 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2741 _loop_opts_cnt--; 2742 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2743 if (failing()) return; 2744 } 2745 // Loop opts pass for loop-unrolling before CCP 2746 if(major_progress() && (_loop_opts_cnt > 0)) { 2747 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2748 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2749 _loop_opts_cnt--; 2750 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2751 } 2752 if (!failing()) { 2753 // Verify that last round of loop opts produced a valid graph 2754 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2755 PhaseIdealLoop::verify(igvn); 2756 } 2757 } 2758 if (failing()) return; 2759 2760 // Conditional Constant Propagation; 2761 PhaseCCP ccp( &igvn ); 2762 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2763 { 2764 TracePhase tp("ccp", &timers[_t_ccp]); 2765 ccp.do_transform(); 2766 } 2767 print_method(PHASE_CPP1, 2); 2768 2769 assert( true, "Break here to ccp.dump_old2new_map()"); 2770 2771 // Iterative Global Value Numbering, including ideal transforms 2772 { 2773 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2774 igvn = ccp; 2775 igvn.optimize(); 2776 } 2777 2778 print_method(PHASE_ITER_GVN2, 2); 2779 2780 if (failing()) return; 2781 2782 // Loop transforms on the ideal graph. Range Check Elimination, 2783 // peeling, unrolling, etc. 2784 if (!optimize_loops(igvn, LoopOptsDefault)) { 2785 return; 2786 } 2787 2788 #if INCLUDE_ZGC 2789 if (UseZGC) { 2790 ZBarrierSetC2::find_dominating_barriers(igvn); 2791 } 2792 #endif 2793 2794 if (failing()) return; 2795 2796 // Ensure that major progress is now clear 2797 C->clear_major_progress(); 2798 2799 { 2800 // Verify that all previous optimizations produced a valid graph 2801 // at least to this point, even if no loop optimizations were done. 2802 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2803 PhaseIdealLoop::verify(igvn); 2804 } 2805 2806 if (range_check_cast_count() > 0) { 2807 // No more loop optimizations. Remove all range check dependent CastIINodes. 2808 C->remove_range_check_casts(igvn); 2809 igvn.optimize(); 2810 } 2811 2812 #ifdef ASSERT 2813 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2814 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand); 2815 #endif 2816 2817 { 2818 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2819 PhaseMacroExpand mex(igvn); 2820 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2821 if (mex.expand_macro_nodes()) { 2822 assert(failing(), "must bail out w/ explicit message"); 2823 return; 2824 } 2825 } 2826 2827 { 2828 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2829 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2); 2830 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2831 if (bs->expand_barriers(this, igvn)) { 2832 assert(failing(), "must bail out w/ explicit message"); 2833 return; 2834 } 2835 } 2836 2837 if (opaque4_count() > 0) { 2838 C->remove_opaque4_nodes(igvn); 2839 igvn.optimize(); 2840 } 2841 2842 DEBUG_ONLY( _modified_nodes = NULL; ) 2843 } // (End scope of igvn; run destructor if necessary for asserts.) 2844 2845 process_print_inlining(); 2846 // A method with only infinite loops has no edges entering loops from root 2847 { 2848 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2849 if (final_graph_reshaping()) { 2850 assert(failing(), "must bail out w/ explicit message"); 2851 return; 2852 } 2853 } 2854 2855 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2856 } 2857 2858 //------------------------------Code_Gen--------------------------------------- 2859 // Given a graph, generate code for it 2860 void Compile::Code_Gen() { 2861 if (failing()) { 2862 return; 2863 } 2864 2865 // Perform instruction selection. You might think we could reclaim Matcher 2866 // memory PDQ, but actually the Matcher is used in generating spill code. 2867 // Internals of the Matcher (including some VectorSets) must remain live 2868 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2869 // set a bit in reclaimed memory. 2870 2871 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2872 // nodes. Mapping is only valid at the root of each matched subtree. 2873 NOT_PRODUCT( verify_graph_edges(); ) 2874 2875 Matcher matcher; 2876 _matcher = &matcher; 2877 { 2878 TracePhase tp("matcher", &timers[_t_matcher]); 2879 matcher.match(); 2880 } 2881 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2882 // nodes. Mapping is only valid at the root of each matched subtree. 2883 NOT_PRODUCT( verify_graph_edges(); ) 2884 2885 // If you have too many nodes, or if matching has failed, bail out 2886 check_node_count(0, "out of nodes matching instructions"); 2887 if (failing()) { 2888 return; 2889 } 2890 2891 print_method(PHASE_MATCHING, 2); 2892 2893 // Build a proper-looking CFG 2894 PhaseCFG cfg(node_arena(), root(), matcher); 2895 _cfg = &cfg; 2896 { 2897 TracePhase tp("scheduler", &timers[_t_scheduler]); 2898 bool success = cfg.do_global_code_motion(); 2899 if (!success) { 2900 return; 2901 } 2902 2903 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2904 NOT_PRODUCT( verify_graph_edges(); ) 2905 debug_only( cfg.verify(); ) 2906 } 2907 2908 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2909 _regalloc = ®alloc; 2910 { 2911 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2912 // Perform register allocation. After Chaitin, use-def chains are 2913 // no longer accurate (at spill code) and so must be ignored. 2914 // Node->LRG->reg mappings are still accurate. 2915 _regalloc->Register_Allocate(); 2916 2917 // Bail out if the allocator builds too many nodes 2918 if (failing()) { 2919 return; 2920 } 2921 } 2922 2923 // Prior to register allocation we kept empty basic blocks in case the 2924 // the allocator needed a place to spill. After register allocation we 2925 // are not adding any new instructions. If any basic block is empty, we 2926 // can now safely remove it. 2927 { 2928 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2929 cfg.remove_empty_blocks(); 2930 if (do_freq_based_layout()) { 2931 PhaseBlockLayout layout(cfg); 2932 } else { 2933 cfg.set_loop_alignment(); 2934 } 2935 cfg.fixup_flow(); 2936 } 2937 2938 // Apply peephole optimizations 2939 if( OptoPeephole ) { 2940 TracePhase tp("peephole", &timers[_t_peephole]); 2941 PhasePeephole peep( _regalloc, cfg); 2942 peep.do_transform(); 2943 } 2944 2945 // Do late expand if CPU requires this. 2946 if (Matcher::require_postalloc_expand) { 2947 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2948 cfg.postalloc_expand(_regalloc); 2949 } 2950 2951 // Convert Nodes to instruction bits in a buffer 2952 { 2953 TraceTime tp("output", &timers[_t_output], CITime); 2954 Output(); 2955 } 2956 2957 print_method(PHASE_FINAL_CODE); 2958 2959 // He's dead, Jim. 2960 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2961 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2962 } 2963 2964 2965 //------------------------------dump_asm--------------------------------------- 2966 // Dump formatted assembly 2967 #if defined(SUPPORT_OPTO_ASSEMBLY) 2968 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) { 2969 2970 int pc_digits = 3; // #chars required for pc 2971 int sb_chars = 3; // #chars for "start bundle" indicator 2972 int tab_size = 8; 2973 if (pcs != NULL) { 2974 int max_pc = 0; 2975 for (uint i = 0; i < pc_limit; i++) { 2976 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc; 2977 } 2978 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc 2979 } 2980 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size; 2981 2982 bool cut_short = false; 2983 st->print_cr("#"); 2984 st->print("# "); _tf->dump_on(st); st->cr(); 2985 st->print_cr("#"); 2986 2987 // For all blocks 2988 int pc = 0x0; // Program counter 2989 char starts_bundle = ' '; 2990 _regalloc->dump_frame(); 2991 2992 Node *n = NULL; 2993 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2994 if (VMThread::should_terminate()) { 2995 cut_short = true; 2996 break; 2997 } 2998 Block* block = _cfg->get_block(i); 2999 if (block->is_connector() && !Verbose) { 3000 continue; 3001 } 3002 n = block->head(); 3003 if ((pcs != NULL) && (n->_idx < pc_limit)) { 3004 pc = pcs[n->_idx]; 3005 st->print("%*.*x", pc_digits, pc_digits, pc); 3006 } 3007 st->fill_to(prefix_len); 3008 block->dump_head(_cfg, st); 3009 if (block->is_connector()) { 3010 st->fill_to(prefix_len); 3011 st->print_cr("# Empty connector block"); 3012 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 3013 st->fill_to(prefix_len); 3014 st->print_cr("# Block is sole successor of call"); 3015 } 3016 3017 // For all instructions 3018 Node *delay = NULL; 3019 for (uint j = 0; j < block->number_of_nodes(); j++) { 3020 if (VMThread::should_terminate()) { 3021 cut_short = true; 3022 break; 3023 } 3024 n = block->get_node(j); 3025 if (valid_bundle_info(n)) { 3026 Bundle* bundle = node_bundling(n); 3027 if (bundle->used_in_unconditional_delay()) { 3028 delay = n; 3029 continue; 3030 } 3031 if (bundle->starts_bundle()) { 3032 starts_bundle = '+'; 3033 } 3034 } 3035 3036 if (WizardMode) { 3037 n->dump(); 3038 } 3039 3040 if( !n->is_Region() && // Dont print in the Assembly 3041 !n->is_Phi() && // a few noisely useless nodes 3042 !n->is_Proj() && 3043 !n->is_MachTemp() && 3044 !n->is_SafePointScalarObject() && 3045 !n->is_Catch() && // Would be nice to print exception table targets 3046 !n->is_MergeMem() && // Not very interesting 3047 !n->is_top() && // Debug info table constants 3048 !(n->is_Con() && !n->is_Mach())// Debug info table constants 3049 ) { 3050 if ((pcs != NULL) && (n->_idx < pc_limit)) { 3051 pc = pcs[n->_idx]; 3052 st->print("%*.*x", pc_digits, pc_digits, pc); 3053 } else { 3054 st->fill_to(pc_digits); 3055 } 3056 st->print(" %c ", starts_bundle); 3057 starts_bundle = ' '; 3058 st->fill_to(prefix_len); 3059 n->format(_regalloc, st); 3060 st->cr(); 3061 } 3062 3063 // If we have an instruction with a delay slot, and have seen a delay, 3064 // then back up and print it 3065 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 3066 // Coverity finding - Explicit null dereferenced. 3067 guarantee(delay != NULL, "no unconditional delay instruction"); 3068 if (WizardMode) delay->dump(); 3069 3070 if (node_bundling(delay)->starts_bundle()) 3071 starts_bundle = '+'; 3072 if ((pcs != NULL) && (n->_idx < pc_limit)) { 3073 pc = pcs[n->_idx]; 3074 st->print("%*.*x", pc_digits, pc_digits, pc); 3075 } else { 3076 st->fill_to(pc_digits); 3077 } 3078 st->print(" %c ", starts_bundle); 3079 starts_bundle = ' '; 3080 st->fill_to(prefix_len); 3081 delay->format(_regalloc, st); 3082 st->cr(); 3083 delay = NULL; 3084 } 3085 3086 // Dump the exception table as well 3087 if( n->is_Catch() && (Verbose || WizardMode) ) { 3088 // Print the exception table for this offset 3089 _handler_table.print_subtable_for(pc); 3090 } 3091 st->bol(); // Make sure we start on a new line 3092 } 3093 st->cr(); // one empty line between blocks 3094 assert(cut_short || delay == NULL, "no unconditional delay branch"); 3095 } // End of per-block dump 3096 3097 if (cut_short) st->print_cr("*** disassembly is cut short ***"); 3098 } 3099 #endif 3100 3101 //------------------------------Final_Reshape_Counts--------------------------- 3102 // This class defines counters to help identify when a method 3103 // may/must be executed using hardware with only 24-bit precision. 3104 struct Final_Reshape_Counts : public StackObj { 3105 int _call_count; // count non-inlined 'common' calls 3106 int _float_count; // count float ops requiring 24-bit precision 3107 int _double_count; // count double ops requiring more precision 3108 int _java_call_count; // count non-inlined 'java' calls 3109 int _inner_loop_count; // count loops which need alignment 3110 VectorSet _visited; // Visitation flags 3111 Node_List _tests; // Set of IfNodes & PCTableNodes 3112 3113 Final_Reshape_Counts() : 3114 _call_count(0), _float_count(0), _double_count(0), 3115 _java_call_count(0), _inner_loop_count(0), 3116 _visited( Thread::current()->resource_area() ) { } 3117 3118 void inc_call_count () { _call_count ++; } 3119 void inc_float_count () { _float_count ++; } 3120 void inc_double_count() { _double_count++; } 3121 void inc_java_call_count() { _java_call_count++; } 3122 void inc_inner_loop_count() { _inner_loop_count++; } 3123 3124 int get_call_count () const { return _call_count ; } 3125 int get_float_count () const { return _float_count ; } 3126 int get_double_count() const { return _double_count; } 3127 int get_java_call_count() const { return _java_call_count; } 3128 int get_inner_loop_count() const { return _inner_loop_count; } 3129 }; 3130 3131 #ifdef ASSERT 3132 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 3133 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 3134 // Make sure the offset goes inside the instance layout. 3135 return k->contains_field_offset(tp->offset()); 3136 // Note that OffsetBot and OffsetTop are very negative. 3137 } 3138 #endif 3139 3140 // Eliminate trivially redundant StoreCMs and accumulate their 3141 // precedence edges. 3142 void Compile::eliminate_redundant_card_marks(Node* n) { 3143 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3144 if (n->in(MemNode::Address)->outcnt() > 1) { 3145 // There are multiple users of the same address so it might be 3146 // possible to eliminate some of the StoreCMs 3147 Node* mem = n->in(MemNode::Memory); 3148 Node* adr = n->in(MemNode::Address); 3149 Node* val = n->in(MemNode::ValueIn); 3150 Node* prev = n; 3151 bool done = false; 3152 // Walk the chain of StoreCMs eliminating ones that match. As 3153 // long as it's a chain of single users then the optimization is 3154 // safe. Eliminating partially redundant StoreCMs would require 3155 // cloning copies down the other paths. 3156 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3157 if (adr == mem->in(MemNode::Address) && 3158 val == mem->in(MemNode::ValueIn)) { 3159 // redundant StoreCM 3160 if (mem->req() > MemNode::OopStore) { 3161 // Hasn't been processed by this code yet. 3162 n->add_prec(mem->in(MemNode::OopStore)); 3163 } else { 3164 // Already converted to precedence edge 3165 for (uint i = mem->req(); i < mem->len(); i++) { 3166 // Accumulate any precedence edges 3167 if (mem->in(i) != NULL) { 3168 n->add_prec(mem->in(i)); 3169 } 3170 } 3171 // Everything above this point has been processed. 3172 done = true; 3173 } 3174 // Eliminate the previous StoreCM 3175 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3176 assert(mem->outcnt() == 0, "should be dead"); 3177 mem->disconnect_inputs(NULL, this); 3178 } else { 3179 prev = mem; 3180 } 3181 mem = prev->in(MemNode::Memory); 3182 } 3183 } 3184 } 3185 3186 3187 //------------------------------final_graph_reshaping_impl---------------------- 3188 // Implement items 1-5 from final_graph_reshaping below. 3189 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 3190 3191 if ( n->outcnt() == 0 ) return; // dead node 3192 uint nop = n->Opcode(); 3193 3194 // Check for 2-input instruction with "last use" on right input. 3195 // Swap to left input. Implements item (2). 3196 if( n->req() == 3 && // two-input instruction 3197 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3198 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3199 n->in(2)->outcnt() == 1 &&// right use IS a last use 3200 !n->in(2)->is_Con() ) { // right use is not a constant 3201 // Check for commutative opcode 3202 switch( nop ) { 3203 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3204 case Op_MaxI: case Op_MinI: 3205 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3206 case Op_AndL: case Op_XorL: case Op_OrL: 3207 case Op_AndI: case Op_XorI: case Op_OrI: { 3208 // Move "last use" input to left by swapping inputs 3209 n->swap_edges(1, 2); 3210 break; 3211 } 3212 default: 3213 break; 3214 } 3215 } 3216 3217 #ifdef ASSERT 3218 if( n->is_Mem() ) { 3219 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3220 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 3221 // oop will be recorded in oop map if load crosses safepoint 3222 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3223 LoadNode::is_immutable_value(n->in(MemNode::Address))), 3224 "raw memory operations should have control edge"); 3225 } 3226 if (n->is_MemBar()) { 3227 MemBarNode* mb = n->as_MemBar(); 3228 if (mb->trailing_store() || mb->trailing_load_store()) { 3229 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3230 Node* mem = mb->in(MemBarNode::Precedent); 3231 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3232 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3233 } else if (mb->leading()) { 3234 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3235 } 3236 } 3237 #endif 3238 // Count FPU ops and common calls, implements item (3) 3239 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 3240 if (!gc_handled) { 3241 final_graph_reshaping_main_switch(n, frc, nop); 3242 } 3243 3244 // Collect CFG split points 3245 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3246 frc._tests.push(n); 3247 } 3248 } 3249 3250 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 3251 switch( nop ) { 3252 // Count all float operations that may use FPU 3253 case Op_AddF: 3254 case Op_SubF: 3255 case Op_MulF: 3256 case Op_DivF: 3257 case Op_NegF: 3258 case Op_ModF: 3259 case Op_ConvI2F: 3260 case Op_ConF: 3261 case Op_CmpF: 3262 case Op_CmpF3: 3263 // case Op_ConvL2F: // longs are split into 32-bit halves 3264 frc.inc_float_count(); 3265 break; 3266 3267 case Op_ConvF2D: 3268 case Op_ConvD2F: 3269 frc.inc_float_count(); 3270 frc.inc_double_count(); 3271 break; 3272 3273 // Count all double operations that may use FPU 3274 case Op_AddD: 3275 case Op_SubD: 3276 case Op_MulD: 3277 case Op_DivD: 3278 case Op_NegD: 3279 case Op_ModD: 3280 case Op_ConvI2D: 3281 case Op_ConvD2I: 3282 // case Op_ConvL2D: // handled by leaf call 3283 // case Op_ConvD2L: // handled by leaf call 3284 case Op_ConD: 3285 case Op_CmpD: 3286 case Op_CmpD3: 3287 frc.inc_double_count(); 3288 break; 3289 case Op_Opaque1: // Remove Opaque Nodes before matching 3290 case Op_Opaque2: // Remove Opaque Nodes before matching 3291 case Op_Opaque3: 3292 n->subsume_by(n->in(1), this); 3293 break; 3294 case Op_CallStaticJava: 3295 case Op_CallJava: 3296 case Op_CallDynamicJava: 3297 frc.inc_java_call_count(); // Count java call site; 3298 case Op_CallRuntime: 3299 case Op_CallLeaf: 3300 case Op_CallLeafNoFP: { 3301 assert (n->is_Call(), ""); 3302 CallNode *call = n->as_Call(); 3303 // Count call sites where the FP mode bit would have to be flipped. 3304 // Do not count uncommon runtime calls: 3305 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3306 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3307 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3308 frc.inc_call_count(); // Count the call site 3309 } else { // See if uncommon argument is shared 3310 Node *n = call->in(TypeFunc::Parms); 3311 int nop = n->Opcode(); 3312 // Clone shared simple arguments to uncommon calls, item (1). 3313 if (n->outcnt() > 1 && 3314 !n->is_Proj() && 3315 nop != Op_CreateEx && 3316 nop != Op_CheckCastPP && 3317 nop != Op_DecodeN && 3318 nop != Op_DecodeNKlass && 3319 !n->is_Mem() && 3320 !n->is_Phi()) { 3321 Node *x = n->clone(); 3322 call->set_req(TypeFunc::Parms, x); 3323 } 3324 } 3325 break; 3326 } 3327 3328 case Op_StoreD: 3329 case Op_LoadD: 3330 case Op_LoadD_unaligned: 3331 frc.inc_double_count(); 3332 goto handle_mem; 3333 case Op_StoreF: 3334 case Op_LoadF: 3335 frc.inc_float_count(); 3336 goto handle_mem; 3337 3338 case Op_StoreCM: 3339 { 3340 // Convert OopStore dependence into precedence edge 3341 Node* prec = n->in(MemNode::OopStore); 3342 n->del_req(MemNode::OopStore); 3343 n->add_prec(prec); 3344 eliminate_redundant_card_marks(n); 3345 } 3346 3347 // fall through 3348 3349 case Op_StoreB: 3350 case Op_StoreC: 3351 case Op_StorePConditional: 3352 case Op_StoreI: 3353 case Op_StoreL: 3354 case Op_StoreIConditional: 3355 case Op_StoreLConditional: 3356 case Op_CompareAndSwapB: 3357 case Op_CompareAndSwapS: 3358 case Op_CompareAndSwapI: 3359 case Op_CompareAndSwapL: 3360 case Op_CompareAndSwapP: 3361 case Op_CompareAndSwapN: 3362 case Op_WeakCompareAndSwapB: 3363 case Op_WeakCompareAndSwapS: 3364 case Op_WeakCompareAndSwapI: 3365 case Op_WeakCompareAndSwapL: 3366 case Op_WeakCompareAndSwapP: 3367 case Op_WeakCompareAndSwapN: 3368 case Op_CompareAndExchangeB: 3369 case Op_CompareAndExchangeS: 3370 case Op_CompareAndExchangeI: 3371 case Op_CompareAndExchangeL: 3372 case Op_CompareAndExchangeP: 3373 case Op_CompareAndExchangeN: 3374 case Op_GetAndAddS: 3375 case Op_GetAndAddB: 3376 case Op_GetAndAddI: 3377 case Op_GetAndAddL: 3378 case Op_GetAndSetS: 3379 case Op_GetAndSetB: 3380 case Op_GetAndSetI: 3381 case Op_GetAndSetL: 3382 case Op_GetAndSetP: 3383 case Op_GetAndSetN: 3384 case Op_StoreP: 3385 case Op_StoreN: 3386 case Op_StoreNKlass: 3387 case Op_LoadB: 3388 case Op_LoadUB: 3389 case Op_LoadUS: 3390 case Op_LoadI: 3391 case Op_LoadKlass: 3392 case Op_LoadNKlass: 3393 case Op_LoadL: 3394 case Op_LoadL_unaligned: 3395 case Op_LoadPLocked: 3396 case Op_LoadP: 3397 case Op_LoadN: 3398 case Op_LoadRange: 3399 case Op_LoadS: { 3400 handle_mem: 3401 #ifdef ASSERT 3402 if( VerifyOptoOopOffsets ) { 3403 MemNode* mem = n->as_Mem(); 3404 // Check to see if address types have grounded out somehow. 3405 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3406 assert( !tp || oop_offset_is_sane(tp), "" ); 3407 } 3408 #endif 3409 if (nop == Op_LoadKlass || nop == Op_LoadNKlass) { 3410 const TypeKlassPtr* tk = n->bottom_type()->make_ptr()->is_klassptr(); 3411 assert(!tk->klass_is_exact(), "should have been folded"); 3412 if (tk->klass()->is_obj_array_klass() || tk->klass()->is_java_lang_Object()) { 3413 bool maybe_value_array = tk->klass()->is_java_lang_Object(); 3414 if (!maybe_value_array) { 3415 ciArrayKlass* ak = tk->klass()->as_array_klass(); 3416 ciKlass* elem = ak->element_klass(); 3417 maybe_value_array = elem->is_java_lang_Object() || elem->is_interface() || elem->is_valuetype(); 3418 } 3419 if (maybe_value_array) { 3420 // Array load klass needs to filter out property bits (but not 3421 // GetNullFreePropertyNode which needs to extract the null free 3422 // bits) 3423 uint last = unique(); 3424 Node* pointer = NULL; 3425 if (nop == Op_LoadKlass) { 3426 Node* cast = new CastP2XNode(NULL, n); 3427 Node* masked = new LShiftXNode(cast, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits))); 3428 masked = new RShiftXNode(masked, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits))); 3429 pointer = new CastX2PNode(masked); 3430 pointer = new CheckCastPPNode(NULL, pointer, n->bottom_type()); 3431 } else { 3432 Node* cast = new CastN2INode(n); 3433 Node* masked = new AndINode(cast, new ConINode(TypeInt::make(oopDesc::compressed_klass_mask()))); 3434 pointer = new CastI2NNode(masked, n->bottom_type()); 3435 } 3436 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3437 Node* u = n->fast_out(i); 3438 if (u->_idx < last && u->Opcode() != Op_GetNullFreeProperty) { 3439 int nb = u->replace_edge(n, pointer); 3440 --i, imax -= nb; 3441 } 3442 } 3443 } 3444 } 3445 } 3446 break; 3447 } 3448 3449 case Op_AddP: { // Assert sane base pointers 3450 Node *addp = n->in(AddPNode::Address); 3451 assert( !addp->is_AddP() || 3452 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3453 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3454 "Base pointers must match (addp %u)", addp->_idx ); 3455 #ifdef _LP64 3456 if ((UseCompressedOops || UseCompressedClassPointers) && 3457 addp->Opcode() == Op_ConP && 3458 addp == n->in(AddPNode::Base) && 3459 n->in(AddPNode::Offset)->is_Con()) { 3460 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3461 // on the platform and on the compressed oops mode. 3462 // Use addressing with narrow klass to load with offset on x86. 3463 // Some platforms can use the constant pool to load ConP. 3464 // Do this transformation here since IGVN will convert ConN back to ConP. 3465 const Type* t = addp->bottom_type(); 3466 bool is_oop = t->isa_oopptr() != NULL; 3467 bool is_klass = t->isa_klassptr() != NULL; 3468 3469 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3470 (is_klass && Matcher::const_klass_prefer_decode())) { 3471 Node* nn = NULL; 3472 3473 int op = is_oop ? Op_ConN : Op_ConNKlass; 3474 3475 // Look for existing ConN node of the same exact type. 3476 Node* r = root(); 3477 uint cnt = r->outcnt(); 3478 for (uint i = 0; i < cnt; i++) { 3479 Node* m = r->raw_out(i); 3480 if (m!= NULL && m->Opcode() == op && 3481 m->bottom_type()->make_ptr() == t) { 3482 nn = m; 3483 break; 3484 } 3485 } 3486 if (nn != NULL) { 3487 // Decode a narrow oop to match address 3488 // [R12 + narrow_oop_reg<<3 + offset] 3489 if (is_oop) { 3490 nn = new DecodeNNode(nn, t); 3491 } else { 3492 nn = new DecodeNKlassNode(nn, t); 3493 } 3494 // Check for succeeding AddP which uses the same Base. 3495 // Otherwise we will run into the assertion above when visiting that guy. 3496 for (uint i = 0; i < n->outcnt(); ++i) { 3497 Node *out_i = n->raw_out(i); 3498 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3499 out_i->set_req(AddPNode::Base, nn); 3500 #ifdef ASSERT 3501 for (uint j = 0; j < out_i->outcnt(); ++j) { 3502 Node *out_j = out_i->raw_out(j); 3503 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3504 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3505 } 3506 #endif 3507 } 3508 } 3509 n->set_req(AddPNode::Base, nn); 3510 n->set_req(AddPNode::Address, nn); 3511 if (addp->outcnt() == 0) { 3512 addp->disconnect_inputs(NULL, this); 3513 } 3514 } 3515 } 3516 } 3517 #endif 3518 // platform dependent reshaping of the address expression 3519 reshape_address(n->as_AddP()); 3520 break; 3521 } 3522 3523 case Op_CastPP: { 3524 // Remove CastPP nodes to gain more freedom during scheduling but 3525 // keep the dependency they encode as control or precedence edges 3526 // (if control is set already) on memory operations. Some CastPP 3527 // nodes don't have a control (don't carry a dependency): skip 3528 // those. 3529 if (n->in(0) != NULL) { 3530 ResourceMark rm; 3531 Unique_Node_List wq; 3532 wq.push(n); 3533 for (uint next = 0; next < wq.size(); ++next) { 3534 Node *m = wq.at(next); 3535 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3536 Node* use = m->fast_out(i); 3537 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3538 use->ensure_control_or_add_prec(n->in(0)); 3539 } else { 3540 switch(use->Opcode()) { 3541 case Op_AddP: 3542 case Op_DecodeN: 3543 case Op_DecodeNKlass: 3544 case Op_CheckCastPP: 3545 case Op_CastPP: 3546 wq.push(use); 3547 break; 3548 } 3549 } 3550 } 3551 } 3552 } 3553 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3554 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3555 Node* in1 = n->in(1); 3556 const Type* t = n->bottom_type(); 3557 Node* new_in1 = in1->clone(); 3558 new_in1->as_DecodeN()->set_type(t); 3559 3560 if (!Matcher::narrow_oop_use_complex_address()) { 3561 // 3562 // x86, ARM and friends can handle 2 adds in addressing mode 3563 // and Matcher can fold a DecodeN node into address by using 3564 // a narrow oop directly and do implicit NULL check in address: 3565 // 3566 // [R12 + narrow_oop_reg<<3 + offset] 3567 // NullCheck narrow_oop_reg 3568 // 3569 // On other platforms (Sparc) we have to keep new DecodeN node and 3570 // use it to do implicit NULL check in address: 3571 // 3572 // decode_not_null narrow_oop_reg, base_reg 3573 // [base_reg + offset] 3574 // NullCheck base_reg 3575 // 3576 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3577 // to keep the information to which NULL check the new DecodeN node 3578 // corresponds to use it as value in implicit_null_check(). 3579 // 3580 new_in1->set_req(0, n->in(0)); 3581 } 3582 3583 n->subsume_by(new_in1, this); 3584 if (in1->outcnt() == 0) { 3585 in1->disconnect_inputs(NULL, this); 3586 } 3587 } else { 3588 n->subsume_by(n->in(1), this); 3589 if (n->outcnt() == 0) { 3590 n->disconnect_inputs(NULL, this); 3591 } 3592 } 3593 break; 3594 } 3595 #ifdef _LP64 3596 case Op_CmpP: 3597 // Do this transformation here to preserve CmpPNode::sub() and 3598 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3599 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3600 Node* in1 = n->in(1); 3601 Node* in2 = n->in(2); 3602 if (!in1->is_DecodeNarrowPtr()) { 3603 in2 = in1; 3604 in1 = n->in(2); 3605 } 3606 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3607 3608 Node* new_in2 = NULL; 3609 if (in2->is_DecodeNarrowPtr()) { 3610 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3611 new_in2 = in2->in(1); 3612 } else if (in2->Opcode() == Op_ConP) { 3613 const Type* t = in2->bottom_type(); 3614 if (t == TypePtr::NULL_PTR) { 3615 assert(in1->is_DecodeN(), "compare klass to null?"); 3616 // Don't convert CmpP null check into CmpN if compressed 3617 // oops implicit null check is not generated. 3618 // This will allow to generate normal oop implicit null check. 3619 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3620 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3621 // 3622 // This transformation together with CastPP transformation above 3623 // will generated code for implicit NULL checks for compressed oops. 3624 // 3625 // The original code after Optimize() 3626 // 3627 // LoadN memory, narrow_oop_reg 3628 // decode narrow_oop_reg, base_reg 3629 // CmpP base_reg, NULL 3630 // CastPP base_reg // NotNull 3631 // Load [base_reg + offset], val_reg 3632 // 3633 // after these transformations will be 3634 // 3635 // LoadN memory, narrow_oop_reg 3636 // CmpN narrow_oop_reg, NULL 3637 // decode_not_null narrow_oop_reg, base_reg 3638 // Load [base_reg + offset], val_reg 3639 // 3640 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3641 // since narrow oops can be used in debug info now (see the code in 3642 // final_graph_reshaping_walk()). 3643 // 3644 // At the end the code will be matched to 3645 // on x86: 3646 // 3647 // Load_narrow_oop memory, narrow_oop_reg 3648 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3649 // NullCheck narrow_oop_reg 3650 // 3651 // and on sparc: 3652 // 3653 // Load_narrow_oop memory, narrow_oop_reg 3654 // decode_not_null narrow_oop_reg, base_reg 3655 // Load [base_reg + offset], val_reg 3656 // NullCheck base_reg 3657 // 3658 } else if (t->isa_oopptr()) { 3659 new_in2 = ConNode::make(t->make_narrowoop()); 3660 } else if (t->isa_klassptr()) { 3661 new_in2 = ConNode::make(t->make_narrowklass()); 3662 } 3663 } 3664 if (new_in2 != NULL) { 3665 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3666 n->subsume_by(cmpN, this); 3667 if (in1->outcnt() == 0) { 3668 in1->disconnect_inputs(NULL, this); 3669 } 3670 if (in2->outcnt() == 0) { 3671 in2->disconnect_inputs(NULL, this); 3672 } 3673 } 3674 } 3675 break; 3676 3677 case Op_DecodeN: 3678 case Op_DecodeNKlass: 3679 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3680 // DecodeN could be pinned when it can't be fold into 3681 // an address expression, see the code for Op_CastPP above. 3682 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3683 break; 3684 3685 case Op_EncodeP: 3686 case Op_EncodePKlass: { 3687 Node* in1 = n->in(1); 3688 if (in1->is_DecodeNarrowPtr()) { 3689 n->subsume_by(in1->in(1), this); 3690 } else if (in1->Opcode() == Op_ConP) { 3691 const Type* t = in1->bottom_type(); 3692 if (t == TypePtr::NULL_PTR) { 3693 assert(t->isa_oopptr(), "null klass?"); 3694 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3695 } else if (t->isa_oopptr()) { 3696 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3697 } else if (t->isa_klassptr()) { 3698 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3699 } 3700 } 3701 if (in1->outcnt() == 0) { 3702 in1->disconnect_inputs(NULL, this); 3703 } 3704 break; 3705 } 3706 3707 case Op_Proj: { 3708 if (OptimizeStringConcat) { 3709 ProjNode* p = n->as_Proj(); 3710 if (p->_is_io_use) { 3711 // Separate projections were used for the exception path which 3712 // are normally removed by a late inline. If it wasn't inlined 3713 // then they will hang around and should just be replaced with 3714 // the original one. 3715 Node* proj = NULL; 3716 // Replace with just one 3717 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3718 Node *use = i.get(); 3719 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3720 proj = use; 3721 break; 3722 } 3723 } 3724 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3725 if (proj != NULL) { 3726 p->subsume_by(proj, this); 3727 } 3728 } 3729 } 3730 break; 3731 } 3732 3733 case Op_Phi: 3734 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3735 // The EncodeP optimization may create Phi with the same edges 3736 // for all paths. It is not handled well by Register Allocator. 3737 Node* unique_in = n->in(1); 3738 assert(unique_in != NULL, ""); 3739 uint cnt = n->req(); 3740 for (uint i = 2; i < cnt; i++) { 3741 Node* m = n->in(i); 3742 assert(m != NULL, ""); 3743 if (unique_in != m) 3744 unique_in = NULL; 3745 } 3746 if (unique_in != NULL) { 3747 n->subsume_by(unique_in, this); 3748 } 3749 } 3750 break; 3751 3752 #endif 3753 3754 #ifdef ASSERT 3755 case Op_CastII: 3756 // Verify that all range check dependent CastII nodes were removed. 3757 if (n->isa_CastII()->has_range_check()) { 3758 n->dump(3); 3759 assert(false, "Range check dependent CastII node was not removed"); 3760 } 3761 break; 3762 #endif 3763 3764 case Op_ModI: 3765 if (UseDivMod) { 3766 // Check if a%b and a/b both exist 3767 Node* d = n->find_similar(Op_DivI); 3768 if (d) { 3769 // Replace them with a fused divmod if supported 3770 if (Matcher::has_match_rule(Op_DivModI)) { 3771 DivModINode* divmod = DivModINode::make(n); 3772 d->subsume_by(divmod->div_proj(), this); 3773 n->subsume_by(divmod->mod_proj(), this); 3774 } else { 3775 // replace a%b with a-((a/b)*b) 3776 Node* mult = new MulINode(d, d->in(2)); 3777 Node* sub = new SubINode(d->in(1), mult); 3778 n->subsume_by(sub, this); 3779 } 3780 } 3781 } 3782 break; 3783 3784 case Op_ModL: 3785 if (UseDivMod) { 3786 // Check if a%b and a/b both exist 3787 Node* d = n->find_similar(Op_DivL); 3788 if (d) { 3789 // Replace them with a fused divmod if supported 3790 if (Matcher::has_match_rule(Op_DivModL)) { 3791 DivModLNode* divmod = DivModLNode::make(n); 3792 d->subsume_by(divmod->div_proj(), this); 3793 n->subsume_by(divmod->mod_proj(), this); 3794 } else { 3795 // replace a%b with a-((a/b)*b) 3796 Node* mult = new MulLNode(d, d->in(2)); 3797 Node* sub = new SubLNode(d->in(1), mult); 3798 n->subsume_by(sub, this); 3799 } 3800 } 3801 } 3802 break; 3803 3804 case Op_LoadVector: 3805 case Op_StoreVector: 3806 break; 3807 3808 case Op_AddReductionVI: 3809 case Op_AddReductionVL: 3810 case Op_AddReductionVF: 3811 case Op_AddReductionVD: 3812 case Op_MulReductionVI: 3813 case Op_MulReductionVL: 3814 case Op_MulReductionVF: 3815 case Op_MulReductionVD: 3816 case Op_MinReductionV: 3817 case Op_MaxReductionV: 3818 break; 3819 3820 case Op_PackB: 3821 case Op_PackS: 3822 case Op_PackI: 3823 case Op_PackF: 3824 case Op_PackL: 3825 case Op_PackD: 3826 if (n->req()-1 > 2) { 3827 // Replace many operand PackNodes with a binary tree for matching 3828 PackNode* p = (PackNode*) n; 3829 Node* btp = p->binary_tree_pack(1, n->req()); 3830 n->subsume_by(btp, this); 3831 } 3832 break; 3833 case Op_Loop: 3834 case Op_CountedLoop: 3835 case Op_OuterStripMinedLoop: 3836 if (n->as_Loop()->is_inner_loop()) { 3837 frc.inc_inner_loop_count(); 3838 } 3839 n->as_Loop()->verify_strip_mined(0); 3840 break; 3841 case Op_LShiftI: 3842 case Op_RShiftI: 3843 case Op_URShiftI: 3844 case Op_LShiftL: 3845 case Op_RShiftL: 3846 case Op_URShiftL: 3847 if (Matcher::need_masked_shift_count) { 3848 // The cpu's shift instructions don't restrict the count to the 3849 // lower 5/6 bits. We need to do the masking ourselves. 3850 Node* in2 = n->in(2); 3851 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3852 const TypeInt* t = in2->find_int_type(); 3853 if (t != NULL && t->is_con()) { 3854 juint shift = t->get_con(); 3855 if (shift > mask) { // Unsigned cmp 3856 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3857 } 3858 } else { 3859 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3860 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3861 n->set_req(2, shift); 3862 } 3863 } 3864 if (in2->outcnt() == 0) { // Remove dead node 3865 in2->disconnect_inputs(NULL, this); 3866 } 3867 } 3868 break; 3869 case Op_MemBarStoreStore: 3870 case Op_MemBarRelease: 3871 // Break the link with AllocateNode: it is no longer useful and 3872 // confuses register allocation. 3873 if (n->req() > MemBarNode::Precedent) { 3874 n->set_req(MemBarNode::Precedent, top()); 3875 } 3876 break; 3877 case Op_MemBarAcquire: { 3878 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3879 // At parse time, the trailing MemBarAcquire for a volatile load 3880 // is created with an edge to the load. After optimizations, 3881 // that input may be a chain of Phis. If those phis have no 3882 // other use, then the MemBarAcquire keeps them alive and 3883 // register allocation can be confused. 3884 ResourceMark rm; 3885 Unique_Node_List wq; 3886 wq.push(n->in(MemBarNode::Precedent)); 3887 n->set_req(MemBarNode::Precedent, top()); 3888 while (wq.size() > 0) { 3889 Node* m = wq.pop(); 3890 if (m->outcnt() == 0) { 3891 for (uint j = 0; j < m->req(); j++) { 3892 Node* in = m->in(j); 3893 if (in != NULL) { 3894 wq.push(in); 3895 } 3896 } 3897 m->disconnect_inputs(NULL, this); 3898 } 3899 } 3900 } 3901 break; 3902 } 3903 case Op_RangeCheck: { 3904 RangeCheckNode* rc = n->as_RangeCheck(); 3905 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3906 n->subsume_by(iff, this); 3907 frc._tests.push(iff); 3908 break; 3909 } 3910 case Op_ConvI2L: { 3911 if (!Matcher::convi2l_type_required) { 3912 // Code generation on some platforms doesn't need accurate 3913 // ConvI2L types. Widening the type can help remove redundant 3914 // address computations. 3915 n->as_Type()->set_type(TypeLong::INT); 3916 ResourceMark rm; 3917 Node_List wq; 3918 wq.push(n); 3919 for (uint next = 0; next < wq.size(); next++) { 3920 Node *m = wq.at(next); 3921 3922 for(;;) { 3923 // Loop over all nodes with identical inputs edges as m 3924 Node* k = m->find_similar(m->Opcode()); 3925 if (k == NULL) { 3926 break; 3927 } 3928 // Push their uses so we get a chance to remove node made 3929 // redundant 3930 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3931 Node* u = k->fast_out(i); 3932 assert(!wq.contains(u), "shouldn't process one node several times"); 3933 if (u->Opcode() == Op_LShiftL || 3934 u->Opcode() == Op_AddL || 3935 u->Opcode() == Op_SubL || 3936 u->Opcode() == Op_AddP) { 3937 wq.push(u); 3938 } 3939 } 3940 // Replace all nodes with identical edges as m with m 3941 k->subsume_by(m, this); 3942 } 3943 } 3944 } 3945 break; 3946 } 3947 case Op_CmpUL: { 3948 if (!Matcher::has_match_rule(Op_CmpUL)) { 3949 // No support for unsigned long comparisons 3950 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3951 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3952 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3953 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3954 Node* andl = new AndLNode(orl, remove_sign_mask); 3955 Node* cmp = new CmpLNode(andl, n->in(2)); 3956 n->subsume_by(cmp, this); 3957 } 3958 break; 3959 } 3960 #ifdef ASSERT 3961 case Op_ValueTypePtr: 3962 case Op_ValueType: { 3963 n->dump(-1); 3964 assert(false, "value type node was not removed"); 3965 break; 3966 } 3967 #endif 3968 case Op_GetNullFreeProperty: { 3969 // Extract the null free bits 3970 uint last = unique(); 3971 Node* null_free = NULL; 3972 if (n->in(1)->Opcode() == Op_LoadKlass) { 3973 Node* cast = new CastP2XNode(NULL, n->in(1)); 3974 null_free = new AndLNode(cast, new ConLNode(TypeLong::make(((jlong)1)<<(oopDesc::wide_storage_props_shift + ArrayStorageProperties::null_free_bit)))); 3975 } else { 3976 assert(n->in(1)->Opcode() == Op_LoadNKlass, "not a compressed klass?"); 3977 Node* cast = new CastN2INode(n->in(1)); 3978 null_free = new AndINode(cast, new ConINode(TypeInt::make(1<<(oopDesc::narrow_storage_props_shift + ArrayStorageProperties::null_free_bit)))); 3979 } 3980 n->replace_by(null_free); 3981 break; 3982 } 3983 default: 3984 assert(!n->is_Call(), ""); 3985 assert(!n->is_Mem(), ""); 3986 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3987 break; 3988 } 3989 } 3990 3991 //------------------------------final_graph_reshaping_walk--------------------- 3992 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3993 // requires that the walk visits a node's inputs before visiting the node. 3994 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3995 ResourceArea *area = Thread::current()->resource_area(); 3996 Unique_Node_List sfpt(area); 3997 3998 frc._visited.set(root->_idx); // first, mark node as visited 3999 uint cnt = root->req(); 4000 Node *n = root; 4001 uint i = 0; 4002 while (true) { 4003 if (i < cnt) { 4004 // Place all non-visited non-null inputs onto stack 4005 Node* m = n->in(i); 4006 ++i; 4007 if (m != NULL && !frc._visited.test_set(m->_idx)) { 4008 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 4009 // compute worst case interpreter size in case of a deoptimization 4010 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 4011 4012 sfpt.push(m); 4013 } 4014 cnt = m->req(); 4015 nstack.push(n, i); // put on stack parent and next input's index 4016 n = m; 4017 i = 0; 4018 } 4019 } else { 4020 // Now do post-visit work 4021 final_graph_reshaping_impl( n, frc ); 4022 if (nstack.is_empty()) 4023 break; // finished 4024 n = nstack.node(); // Get node from stack 4025 cnt = n->req(); 4026 i = nstack.index(); 4027 nstack.pop(); // Shift to the next node on stack 4028 } 4029 } 4030 4031 // Skip next transformation if compressed oops are not used. 4032 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 4033 (!UseCompressedOops && !UseCompressedClassPointers)) 4034 return; 4035 4036 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 4037 // It could be done for an uncommon traps or any safepoints/calls 4038 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 4039 while (sfpt.size() > 0) { 4040 n = sfpt.pop(); 4041 JVMState *jvms = n->as_SafePoint()->jvms(); 4042 assert(jvms != NULL, "sanity"); 4043 int start = jvms->debug_start(); 4044 int end = n->req(); 4045 bool is_uncommon = (n->is_CallStaticJava() && 4046 n->as_CallStaticJava()->uncommon_trap_request() != 0); 4047 for (int j = start; j < end; j++) { 4048 Node* in = n->in(j); 4049 if (in->is_DecodeNarrowPtr()) { 4050 bool safe_to_skip = true; 4051 if (!is_uncommon ) { 4052 // Is it safe to skip? 4053 for (uint i = 0; i < in->outcnt(); i++) { 4054 Node* u = in->raw_out(i); 4055 if (!u->is_SafePoint() || 4056 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 4057 safe_to_skip = false; 4058 } 4059 } 4060 } 4061 if (safe_to_skip) { 4062 n->set_req(j, in->in(1)); 4063 } 4064 if (in->outcnt() == 0) { 4065 in->disconnect_inputs(NULL, this); 4066 } 4067 } 4068 } 4069 } 4070 } 4071 4072 //------------------------------final_graph_reshaping-------------------------- 4073 // Final Graph Reshaping. 4074 // 4075 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 4076 // and not commoned up and forced early. Must come after regular 4077 // optimizations to avoid GVN undoing the cloning. Clone constant 4078 // inputs to Loop Phis; these will be split by the allocator anyways. 4079 // Remove Opaque nodes. 4080 // (2) Move last-uses by commutative operations to the left input to encourage 4081 // Intel update-in-place two-address operations and better register usage 4082 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 4083 // calls canonicalizing them back. 4084 // (3) Count the number of double-precision FP ops, single-precision FP ops 4085 // and call sites. On Intel, we can get correct rounding either by 4086 // forcing singles to memory (requires extra stores and loads after each 4087 // FP bytecode) or we can set a rounding mode bit (requires setting and 4088 // clearing the mode bit around call sites). The mode bit is only used 4089 // if the relative frequency of single FP ops to calls is low enough. 4090 // This is a key transform for SPEC mpeg_audio. 4091 // (4) Detect infinite loops; blobs of code reachable from above but not 4092 // below. Several of the Code_Gen algorithms fail on such code shapes, 4093 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4094 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4095 // Detection is by looking for IfNodes where only 1 projection is 4096 // reachable from below or CatchNodes missing some targets. 4097 // (5) Assert for insane oop offsets in debug mode. 4098 4099 bool Compile::final_graph_reshaping() { 4100 // an infinite loop may have been eliminated by the optimizer, 4101 // in which case the graph will be empty. 4102 if (root()->req() == 1) { 4103 record_method_not_compilable("trivial infinite loop"); 4104 return true; 4105 } 4106 4107 // Expensive nodes have their control input set to prevent the GVN 4108 // from freely commoning them. There's no GVN beyond this point so 4109 // no need to keep the control input. We want the expensive nodes to 4110 // be freely moved to the least frequent code path by gcm. 4111 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4112 for (int i = 0; i < expensive_count(); i++) { 4113 _expensive_nodes->at(i)->set_req(0, NULL); 4114 } 4115 4116 Final_Reshape_Counts frc; 4117 4118 // Visit everybody reachable! 4119 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4120 Node_Stack nstack(live_nodes() >> 1); 4121 final_graph_reshaping_walk(nstack, root(), frc); 4122 4123 // Check for unreachable (from below) code (i.e., infinite loops). 4124 for( uint i = 0; i < frc._tests.size(); i++ ) { 4125 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4126 // Get number of CFG targets. 4127 // Note that PCTables include exception targets after calls. 4128 uint required_outcnt = n->required_outcnt(); 4129 if (n->outcnt() != required_outcnt) { 4130 // Check for a few special cases. Rethrow Nodes never take the 4131 // 'fall-thru' path, so expected kids is 1 less. 4132 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4133 if (n->in(0)->in(0)->is_Call()) { 4134 CallNode *call = n->in(0)->in(0)->as_Call(); 4135 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4136 required_outcnt--; // Rethrow always has 1 less kid 4137 } else if (call->req() > TypeFunc::Parms && 4138 call->is_CallDynamicJava()) { 4139 // Check for null receiver. In such case, the optimizer has 4140 // detected that the virtual call will always result in a null 4141 // pointer exception. The fall-through projection of this CatchNode 4142 // will not be populated. 4143 Node *arg0 = call->in(TypeFunc::Parms); 4144 if (arg0->is_Type() && 4145 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4146 required_outcnt--; 4147 } 4148 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 4149 call->req() > TypeFunc::Parms+1 && 4150 call->is_CallStaticJava()) { 4151 // Check for negative array length. In such case, the optimizer has 4152 // detected that the allocation attempt will always result in an 4153 // exception. There is no fall-through projection of this CatchNode . 4154 Node *arg1 = call->in(TypeFunc::Parms+1); 4155 if (arg1->is_Type() && 4156 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 4157 required_outcnt--; 4158 } 4159 } 4160 } 4161 } 4162 // Recheck with a better notion of 'required_outcnt' 4163 if (n->outcnt() != required_outcnt) { 4164 record_method_not_compilable("malformed control flow"); 4165 return true; // Not all targets reachable! 4166 } 4167 } 4168 // Check that I actually visited all kids. Unreached kids 4169 // must be infinite loops. 4170 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4171 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4172 record_method_not_compilable("infinite loop"); 4173 return true; // Found unvisited kid; must be unreach 4174 } 4175 4176 // Here so verification code in final_graph_reshaping_walk() 4177 // always see an OuterStripMinedLoopEnd 4178 if (n->is_OuterStripMinedLoopEnd()) { 4179 IfNode* init_iff = n->as_If(); 4180 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4181 n->subsume_by(iff, this); 4182 } 4183 } 4184 4185 // If original bytecodes contained a mixture of floats and doubles 4186 // check if the optimizer has made it homogenous, item (3). 4187 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 4188 frc.get_float_count() > 32 && 4189 frc.get_double_count() == 0 && 4190 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4191 set_24_bit_selection_and_mode( false, true ); 4192 } 4193 4194 set_java_calls(frc.get_java_call_count()); 4195 set_inner_loops(frc.get_inner_loop_count()); 4196 4197 // No infinite loops, no reason to bail out. 4198 return false; 4199 } 4200 4201 //-----------------------------too_many_traps---------------------------------- 4202 // Report if there are too many traps at the current method and bci. 4203 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4204 bool Compile::too_many_traps(ciMethod* method, 4205 int bci, 4206 Deoptimization::DeoptReason reason) { 4207 ciMethodData* md = method->method_data(); 4208 if (md->is_empty()) { 4209 // Assume the trap has not occurred, or that it occurred only 4210 // because of a transient condition during start-up in the interpreter. 4211 return false; 4212 } 4213 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 4214 if (md->has_trap_at(bci, m, reason) != 0) { 4215 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4216 // Also, if there are multiple reasons, or if there is no per-BCI record, 4217 // assume the worst. 4218 if (log()) 4219 log()->elem("observe trap='%s' count='%d'", 4220 Deoptimization::trap_reason_name(reason), 4221 md->trap_count(reason)); 4222 return true; 4223 } else { 4224 // Ignore method/bci and see if there have been too many globally. 4225 return too_many_traps(reason, md); 4226 } 4227 } 4228 4229 // Less-accurate variant which does not require a method and bci. 4230 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4231 ciMethodData* logmd) { 4232 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4233 // Too many traps globally. 4234 // Note that we use cumulative trap_count, not just md->trap_count. 4235 if (log()) { 4236 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 4237 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4238 Deoptimization::trap_reason_name(reason), 4239 mcount, trap_count(reason)); 4240 } 4241 return true; 4242 } else { 4243 // The coast is clear. 4244 return false; 4245 } 4246 } 4247 4248 //--------------------------too_many_recompiles-------------------------------- 4249 // Report if there are too many recompiles at the current method and bci. 4250 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4251 // Is not eager to return true, since this will cause the compiler to use 4252 // Action_none for a trap point, to avoid too many recompilations. 4253 bool Compile::too_many_recompiles(ciMethod* method, 4254 int bci, 4255 Deoptimization::DeoptReason reason) { 4256 ciMethodData* md = method->method_data(); 4257 if (md->is_empty()) { 4258 // Assume the trap has not occurred, or that it occurred only 4259 // because of a transient condition during start-up in the interpreter. 4260 return false; 4261 } 4262 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4263 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4264 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4265 Deoptimization::DeoptReason per_bc_reason 4266 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4267 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 4268 if ((per_bc_reason == Deoptimization::Reason_none 4269 || md->has_trap_at(bci, m, reason) != 0) 4270 // The trap frequency measure we care about is the recompile count: 4271 && md->trap_recompiled_at(bci, m) 4272 && md->overflow_recompile_count() >= bc_cutoff) { 4273 // Do not emit a trap here if it has already caused recompilations. 4274 // Also, if there are multiple reasons, or if there is no per-BCI record, 4275 // assume the worst. 4276 if (log()) 4277 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4278 Deoptimization::trap_reason_name(reason), 4279 md->trap_count(reason), 4280 md->overflow_recompile_count()); 4281 return true; 4282 } else if (trap_count(reason) != 0 4283 && decompile_count() >= m_cutoff) { 4284 // Too many recompiles globally, and we have seen this sort of trap. 4285 // Use cumulative decompile_count, not just md->decompile_count. 4286 if (log()) 4287 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4288 Deoptimization::trap_reason_name(reason), 4289 md->trap_count(reason), trap_count(reason), 4290 md->decompile_count(), decompile_count()); 4291 return true; 4292 } else { 4293 // The coast is clear. 4294 return false; 4295 } 4296 } 4297 4298 // Compute when not to trap. Used by matching trap based nodes and 4299 // NullCheck optimization. 4300 void Compile::set_allowed_deopt_reasons() { 4301 _allowed_reasons = 0; 4302 if (is_method_compilation()) { 4303 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4304 assert(rs < BitsPerInt, "recode bit map"); 4305 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4306 _allowed_reasons |= nth_bit(rs); 4307 } 4308 } 4309 } 4310 } 4311 4312 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4313 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4314 } 4315 4316 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4317 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4318 } 4319 4320 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4321 if (holder->is_initialized()) { 4322 return false; 4323 } 4324 if (holder->is_being_initialized()) { 4325 if (accessing_method->holder() == holder) { 4326 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4327 // <init>, or a static method. In all those cases, there was an initialization 4328 // barrier on the holder klass passed. 4329 if (accessing_method->is_class_initializer() || 4330 accessing_method->is_object_constructor() || 4331 accessing_method->is_static()) { 4332 return false; 4333 } 4334 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4335 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4336 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4337 // child class can become fully initialized while its parent class is still being initialized. 4338 if (accessing_method->is_class_initializer()) { 4339 return false; 4340 } 4341 } 4342 ciMethod* root = method(); // the root method of compilation 4343 if (root != accessing_method) { 4344 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4345 } 4346 } 4347 return true; 4348 } 4349 4350 #ifndef PRODUCT 4351 //------------------------------verify_graph_edges--------------------------- 4352 // Walk the Graph and verify that there is a one-to-one correspondence 4353 // between Use-Def edges and Def-Use edges in the graph. 4354 void Compile::verify_graph_edges(bool no_dead_code) { 4355 if (VerifyGraphEdges) { 4356 ResourceArea *area = Thread::current()->resource_area(); 4357 Unique_Node_List visited(area); 4358 // Call recursive graph walk to check edges 4359 _root->verify_edges(visited); 4360 if (no_dead_code) { 4361 // Now make sure that no visited node is used by an unvisited node. 4362 bool dead_nodes = false; 4363 Unique_Node_List checked(area); 4364 while (visited.size() > 0) { 4365 Node* n = visited.pop(); 4366 checked.push(n); 4367 for (uint i = 0; i < n->outcnt(); i++) { 4368 Node* use = n->raw_out(i); 4369 if (checked.member(use)) continue; // already checked 4370 if (visited.member(use)) continue; // already in the graph 4371 if (use->is_Con()) continue; // a dead ConNode is OK 4372 // At this point, we have found a dead node which is DU-reachable. 4373 if (!dead_nodes) { 4374 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4375 dead_nodes = true; 4376 } 4377 use->dump(2); 4378 tty->print_cr("---"); 4379 checked.push(use); // No repeats; pretend it is now checked. 4380 } 4381 } 4382 assert(!dead_nodes, "using nodes must be reachable from root"); 4383 } 4384 } 4385 } 4386 #endif 4387 4388 // The Compile object keeps track of failure reasons separately from the ciEnv. 4389 // This is required because there is not quite a 1-1 relation between the 4390 // ciEnv and its compilation task and the Compile object. Note that one 4391 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4392 // to backtrack and retry without subsuming loads. Other than this backtracking 4393 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4394 // by the logic in C2Compiler. 4395 void Compile::record_failure(const char* reason) { 4396 if (log() != NULL) { 4397 log()->elem("failure reason='%s' phase='compile'", reason); 4398 } 4399 if (_failure_reason == NULL) { 4400 // Record the first failure reason. 4401 _failure_reason = reason; 4402 } 4403 4404 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4405 C->print_method(PHASE_FAILURE); 4406 } 4407 _root = NULL; // flush the graph, too 4408 } 4409 4410 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4411 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4412 _phase_name(name), _dolog(CITimeVerbose) 4413 { 4414 if (_dolog) { 4415 C = Compile::current(); 4416 _log = C->log(); 4417 } else { 4418 C = NULL; 4419 _log = NULL; 4420 } 4421 if (_log != NULL) { 4422 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4423 _log->stamp(); 4424 _log->end_head(); 4425 } 4426 } 4427 4428 Compile::TracePhase::~TracePhase() { 4429 4430 C = Compile::current(); 4431 if (_dolog) { 4432 _log = C->log(); 4433 } else { 4434 _log = NULL; 4435 } 4436 4437 #ifdef ASSERT 4438 if (PrintIdealNodeCount) { 4439 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4440 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4441 } 4442 4443 if (VerifyIdealNodeCount) { 4444 Compile::current()->print_missing_nodes(); 4445 } 4446 #endif 4447 4448 if (_log != NULL) { 4449 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4450 } 4451 } 4452 4453 //============================================================================= 4454 // Two Constant's are equal when the type and the value are equal. 4455 bool Compile::Constant::operator==(const Constant& other) { 4456 if (type() != other.type() ) return false; 4457 if (can_be_reused() != other.can_be_reused()) return false; 4458 // For floating point values we compare the bit pattern. 4459 switch (type()) { 4460 case T_INT: 4461 case T_FLOAT: return (_v._value.i == other._v._value.i); 4462 case T_LONG: 4463 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4464 case T_OBJECT: 4465 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4466 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4467 case T_METADATA: return (_v._metadata == other._v._metadata); 4468 default: ShouldNotReachHere(); return false; 4469 } 4470 } 4471 4472 static int type_to_size_in_bytes(BasicType t) { 4473 switch (t) { 4474 case T_INT: return sizeof(jint ); 4475 case T_LONG: return sizeof(jlong ); 4476 case T_FLOAT: return sizeof(jfloat ); 4477 case T_DOUBLE: return sizeof(jdouble); 4478 case T_METADATA: return sizeof(Metadata*); 4479 // We use T_VOID as marker for jump-table entries (labels) which 4480 // need an internal word relocation. 4481 case T_VOID: 4482 case T_ADDRESS: 4483 case T_OBJECT: return sizeof(jobject); 4484 default: 4485 ShouldNotReachHere(); 4486 return -1; 4487 } 4488 } 4489 4490 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4491 // sort descending 4492 if (a->freq() > b->freq()) return -1; 4493 if (a->freq() < b->freq()) return 1; 4494 return 0; 4495 } 4496 4497 void Compile::ConstantTable::calculate_offsets_and_size() { 4498 // First, sort the array by frequencies. 4499 _constants.sort(qsort_comparator); 4500 4501 #ifdef ASSERT 4502 // Make sure all jump-table entries were sorted to the end of the 4503 // array (they have a negative frequency). 4504 bool found_void = false; 4505 for (int i = 0; i < _constants.length(); i++) { 4506 Constant con = _constants.at(i); 4507 if (con.type() == T_VOID) 4508 found_void = true; // jump-tables 4509 else 4510 assert(!found_void, "wrong sorting"); 4511 } 4512 #endif 4513 4514 int offset = 0; 4515 for (int i = 0; i < _constants.length(); i++) { 4516 Constant* con = _constants.adr_at(i); 4517 4518 // Align offset for type. 4519 int typesize = type_to_size_in_bytes(con->type()); 4520 offset = align_up(offset, typesize); 4521 con->set_offset(offset); // set constant's offset 4522 4523 if (con->type() == T_VOID) { 4524 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4525 offset = offset + typesize * n->outcnt(); // expand jump-table 4526 } else { 4527 offset = offset + typesize; 4528 } 4529 } 4530 4531 // Align size up to the next section start (which is insts; see 4532 // CodeBuffer::align_at_start). 4533 assert(_size == -1, "already set?"); 4534 _size = align_up(offset, (int)CodeEntryAlignment); 4535 } 4536 4537 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4538 MacroAssembler _masm(&cb); 4539 for (int i = 0; i < _constants.length(); i++) { 4540 Constant con = _constants.at(i); 4541 address constant_addr = NULL; 4542 switch (con.type()) { 4543 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4544 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4545 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4546 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4547 case T_OBJECT: { 4548 jobject obj = con.get_jobject(); 4549 int oop_index = _masm.oop_recorder()->find_index(obj); 4550 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4551 break; 4552 } 4553 case T_ADDRESS: { 4554 address addr = (address) con.get_jobject(); 4555 constant_addr = _masm.address_constant(addr); 4556 break; 4557 } 4558 // We use T_VOID as marker for jump-table entries (labels) which 4559 // need an internal word relocation. 4560 case T_VOID: { 4561 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4562 // Fill the jump-table with a dummy word. The real value is 4563 // filled in later in fill_jump_table. 4564 address dummy = (address) n; 4565 constant_addr = _masm.address_constant(dummy); 4566 // Expand jump-table 4567 for (uint i = 1; i < n->outcnt(); i++) { 4568 address temp_addr = _masm.address_constant(dummy + i); 4569 assert(temp_addr, "consts section too small"); 4570 } 4571 break; 4572 } 4573 case T_METADATA: { 4574 Metadata* obj = con.get_metadata(); 4575 int metadata_index = _masm.oop_recorder()->find_index(obj); 4576 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4577 break; 4578 } 4579 default: ShouldNotReachHere(); 4580 } 4581 assert(constant_addr, "consts section too small"); 4582 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4583 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4584 } 4585 } 4586 4587 int Compile::ConstantTable::find_offset(Constant& con) const { 4588 int idx = _constants.find(con); 4589 guarantee(idx != -1, "constant must be in constant table"); 4590 int offset = _constants.at(idx).offset(); 4591 guarantee(offset != -1, "constant table not emitted yet?"); 4592 return offset; 4593 } 4594 4595 void Compile::ConstantTable::add(Constant& con) { 4596 if (con.can_be_reused()) { 4597 int idx = _constants.find(con); 4598 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4599 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4600 return; 4601 } 4602 } 4603 (void) _constants.append(con); 4604 } 4605 4606 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4607 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4608 Constant con(type, value, b->_freq); 4609 add(con); 4610 return con; 4611 } 4612 4613 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4614 Constant con(metadata); 4615 add(con); 4616 return con; 4617 } 4618 4619 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4620 jvalue value; 4621 BasicType type = oper->type()->basic_type(); 4622 switch (type) { 4623 case T_LONG: value.j = oper->constantL(); break; 4624 case T_FLOAT: value.f = oper->constantF(); break; 4625 case T_DOUBLE: value.d = oper->constantD(); break; 4626 case T_OBJECT: 4627 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4628 case T_METADATA: return add((Metadata*)oper->constant()); break; 4629 default: guarantee(false, "unhandled type: %s", type2name(type)); 4630 } 4631 return add(n, type, value); 4632 } 4633 4634 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4635 jvalue value; 4636 // We can use the node pointer here to identify the right jump-table 4637 // as this method is called from Compile::Fill_buffer right before 4638 // the MachNodes are emitted and the jump-table is filled (means the 4639 // MachNode pointers do not change anymore). 4640 value.l = (jobject) n; 4641 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4642 add(con); 4643 return con; 4644 } 4645 4646 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4647 // If called from Compile::scratch_emit_size do nothing. 4648 if (Compile::current()->in_scratch_emit_size()) return; 4649 4650 assert(labels.is_nonempty(), "must be"); 4651 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4652 4653 // Since MachConstantNode::constant_offset() also contains 4654 // table_base_offset() we need to subtract the table_base_offset() 4655 // to get the plain offset into the constant table. 4656 int offset = n->constant_offset() - table_base_offset(); 4657 4658 MacroAssembler _masm(&cb); 4659 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4660 4661 for (uint i = 0; i < n->outcnt(); i++) { 4662 address* constant_addr = &jump_table_base[i]; 4663 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)); 4664 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4665 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4666 } 4667 } 4668 4669 //----------------------------static_subtype_check----------------------------- 4670 // Shortcut important common cases when superklass is exact: 4671 // (0) superklass is java.lang.Object (can occur in reflective code) 4672 // (1) subklass is already limited to a subtype of superklass => always ok 4673 // (2) subklass does not overlap with superklass => always fail 4674 // (3) superklass has NO subtypes and we can check with a simple compare. 4675 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4676 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4677 return SSC_full_test; // Let caller generate the general case. 4678 } 4679 4680 if (superk == env()->Object_klass()) { 4681 return SSC_always_true; // (0) this test cannot fail 4682 } 4683 4684 ciType* superelem = superk; 4685 if (superelem->is_array_klass()) { 4686 ciArrayKlass* ak = superelem->as_array_klass(); 4687 superelem = superelem->as_array_klass()->base_element_type(); 4688 } 4689 4690 if (!subk->is_interface()) { // cannot trust static interface types yet 4691 if (subk->is_subtype_of(superk)) { 4692 return SSC_always_true; // (1) false path dead; no dynamic test needed 4693 } 4694 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4695 !superk->is_subtype_of(subk)) { 4696 return SSC_always_false; 4697 } 4698 } 4699 4700 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays. 4701 // [V is a subtype of [V? but the klass for [V is not equal to the klass for [V?. Perform a full test. 4702 if (superk->is_obj_array_klass() && !superk->as_array_klass()->storage_properties().is_null_free() && superk->as_array_klass()->element_klass()->is_valuetype()) { 4703 return SSC_full_test; 4704 } 4705 // If casting to an instance klass, it must have no subtypes 4706 if (superk->is_interface()) { 4707 // Cannot trust interfaces yet. 4708 // %%% S.B. superk->nof_implementors() == 1 4709 } else if (superelem->is_instance_klass()) { 4710 ciInstanceKlass* ik = superelem->as_instance_klass(); 4711 if (!ik->has_subklass() && !ik->is_interface()) { 4712 if (!ik->is_final()) { 4713 // Add a dependency if there is a chance of a later subclass. 4714 dependencies()->assert_leaf_type(ik); 4715 } 4716 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4717 } 4718 } else { 4719 // A primitive array type has no subtypes. 4720 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4721 } 4722 4723 return SSC_full_test; 4724 } 4725 4726 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4727 #ifdef _LP64 4728 // The scaled index operand to AddP must be a clean 64-bit value. 4729 // Java allows a 32-bit int to be incremented to a negative 4730 // value, which appears in a 64-bit register as a large 4731 // positive number. Using that large positive number as an 4732 // operand in pointer arithmetic has bad consequences. 4733 // On the other hand, 32-bit overflow is rare, and the possibility 4734 // can often be excluded, if we annotate the ConvI2L node with 4735 // a type assertion that its value is known to be a small positive 4736 // number. (The prior range check has ensured this.) 4737 // This assertion is used by ConvI2LNode::Ideal. 4738 int index_max = max_jint - 1; // array size is max_jint, index is one less 4739 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4740 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4741 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4742 #endif 4743 return idx; 4744 } 4745 4746 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4747 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4748 if (ctrl != NULL) { 4749 // Express control dependency by a CastII node with a narrow type. 4750 value = new CastIINode(value, itype, false, true /* range check dependency */); 4751 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4752 // node from floating above the range check during loop optimizations. Otherwise, the 4753 // ConvI2L node may be eliminated independently of the range check, causing the data path 4754 // to become TOP while the control path is still there (although it's unreachable). 4755 value->set_req(0, ctrl); 4756 // Save CastII node to remove it after loop optimizations. 4757 phase->C->add_range_check_cast(value); 4758 value = phase->transform(value); 4759 } 4760 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4761 return phase->transform(new ConvI2LNode(value, ltype)); 4762 } 4763 4764 // The message about the current inlining is accumulated in 4765 // _print_inlining_stream and transfered into the _print_inlining_list 4766 // once we know whether inlining succeeds or not. For regular 4767 // inlining, messages are appended to the buffer pointed by 4768 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4769 // a new buffer is added after _print_inlining_idx in the list. This 4770 // way we can update the inlining message for late inlining call site 4771 // when the inlining is attempted again. 4772 void Compile::print_inlining_init() { 4773 if (print_inlining() || print_intrinsics()) { 4774 _print_inlining_stream = new stringStream(); 4775 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4776 } 4777 } 4778 4779 void Compile::print_inlining_reinit() { 4780 if (print_inlining() || print_intrinsics()) { 4781 // Re allocate buffer when we change ResourceMark 4782 _print_inlining_stream = new stringStream(); 4783 } 4784 } 4785 4786 void Compile::print_inlining_reset() { 4787 _print_inlining_stream->reset(); 4788 } 4789 4790 void Compile::print_inlining_commit() { 4791 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4792 // Transfer the message from _print_inlining_stream to the current 4793 // _print_inlining_list buffer and clear _print_inlining_stream. 4794 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4795 print_inlining_reset(); 4796 } 4797 4798 void Compile::print_inlining_push() { 4799 // Add new buffer to the _print_inlining_list at current position 4800 _print_inlining_idx++; 4801 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4802 } 4803 4804 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4805 return _print_inlining_list->at(_print_inlining_idx); 4806 } 4807 4808 void Compile::print_inlining_update(CallGenerator* cg) { 4809 if (print_inlining() || print_intrinsics()) { 4810 if (!cg->is_late_inline()) { 4811 if (print_inlining_current().cg() != NULL) { 4812 print_inlining_push(); 4813 } 4814 print_inlining_commit(); 4815 } else { 4816 if (print_inlining_current().cg() != cg && 4817 (print_inlining_current().cg() != NULL || 4818 print_inlining_current().ss()->size() != 0)) { 4819 print_inlining_push(); 4820 } 4821 print_inlining_commit(); 4822 print_inlining_current().set_cg(cg); 4823 } 4824 } 4825 } 4826 4827 void Compile::print_inlining_move_to(CallGenerator* cg) { 4828 // We resume inlining at a late inlining call site. Locate the 4829 // corresponding inlining buffer so that we can update it. 4830 if (print_inlining()) { 4831 for (int i = 0; i < _print_inlining_list->length(); i++) { 4832 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4833 _print_inlining_idx = i; 4834 return; 4835 } 4836 } 4837 ShouldNotReachHere(); 4838 } 4839 } 4840 4841 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4842 if (print_inlining()) { 4843 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4844 assert(print_inlining_current().cg() == cg, "wrong entry"); 4845 // replace message with new message 4846 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4847 print_inlining_commit(); 4848 print_inlining_current().set_cg(cg); 4849 } 4850 } 4851 4852 void Compile::print_inlining_assert_ready() { 4853 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4854 } 4855 4856 void Compile::process_print_inlining() { 4857 bool do_print_inlining = print_inlining() || print_intrinsics(); 4858 if (do_print_inlining || log() != NULL) { 4859 // Print inlining message for candidates that we couldn't inline 4860 // for lack of space 4861 for (int i = 0; i < _late_inlines.length(); i++) { 4862 CallGenerator* cg = _late_inlines.at(i); 4863 if (!cg->is_mh_late_inline()) { 4864 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4865 if (do_print_inlining) { 4866 cg->print_inlining_late(msg); 4867 } 4868 log_late_inline_failure(cg, msg); 4869 } 4870 } 4871 } 4872 if (do_print_inlining) { 4873 ResourceMark rm; 4874 stringStream ss; 4875 for (int i = 0; i < _print_inlining_list->length(); i++) { 4876 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4877 } 4878 size_t end = ss.size(); 4879 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4880 strncpy(_print_inlining_output, ss.base(), end+1); 4881 _print_inlining_output[end] = 0; 4882 } 4883 } 4884 4885 void Compile::dump_print_inlining() { 4886 if (_print_inlining_output != NULL) { 4887 tty->print_raw(_print_inlining_output); 4888 } 4889 } 4890 4891 void Compile::log_late_inline(CallGenerator* cg) { 4892 if (log() != NULL) { 4893 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4894 cg->unique_id()); 4895 JVMState* p = cg->call_node()->jvms(); 4896 while (p != NULL) { 4897 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4898 p = p->caller(); 4899 } 4900 log()->tail("late_inline"); 4901 } 4902 } 4903 4904 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4905 log_late_inline(cg); 4906 if (log() != NULL) { 4907 log()->inline_fail(msg); 4908 } 4909 } 4910 4911 void Compile::log_inline_id(CallGenerator* cg) { 4912 if (log() != NULL) { 4913 // The LogCompilation tool needs a unique way to identify late 4914 // inline call sites. This id must be unique for this call site in 4915 // this compilation. Try to have it unique across compilations as 4916 // well because it can be convenient when grepping through the log 4917 // file. 4918 // Distinguish OSR compilations from others in case CICountOSR is 4919 // on. 4920 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4921 cg->set_unique_id(id); 4922 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4923 } 4924 } 4925 4926 void Compile::log_inline_failure(const char* msg) { 4927 if (C->log() != NULL) { 4928 C->log()->inline_fail(msg); 4929 } 4930 } 4931 4932 4933 // Dump inlining replay data to the stream. 4934 // Don't change thread state and acquire any locks. 4935 void Compile::dump_inline_data(outputStream* out) { 4936 InlineTree* inl_tree = ilt(); 4937 if (inl_tree != NULL) { 4938 out->print(" inline %d", inl_tree->count()); 4939 inl_tree->dump_replay_data(out); 4940 } 4941 } 4942 4943 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4944 if (n1->Opcode() < n2->Opcode()) return -1; 4945 else if (n1->Opcode() > n2->Opcode()) return 1; 4946 4947 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4948 for (uint i = 1; i < n1->req(); i++) { 4949 if (n1->in(i) < n2->in(i)) return -1; 4950 else if (n1->in(i) > n2->in(i)) return 1; 4951 } 4952 4953 return 0; 4954 } 4955 4956 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4957 Node* n1 = *n1p; 4958 Node* n2 = *n2p; 4959 4960 return cmp_expensive_nodes(n1, n2); 4961 } 4962 4963 void Compile::sort_expensive_nodes() { 4964 if (!expensive_nodes_sorted()) { 4965 _expensive_nodes->sort(cmp_expensive_nodes); 4966 } 4967 } 4968 4969 bool Compile::expensive_nodes_sorted() const { 4970 for (int i = 1; i < _expensive_nodes->length(); i++) { 4971 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4972 return false; 4973 } 4974 } 4975 return true; 4976 } 4977 4978 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4979 if (_expensive_nodes->length() == 0) { 4980 return false; 4981 } 4982 4983 assert(OptimizeExpensiveOps, "optimization off?"); 4984 4985 // Take this opportunity to remove dead nodes from the list 4986 int j = 0; 4987 for (int i = 0; i < _expensive_nodes->length(); i++) { 4988 Node* n = _expensive_nodes->at(i); 4989 if (!n->is_unreachable(igvn)) { 4990 assert(n->is_expensive(), "should be expensive"); 4991 _expensive_nodes->at_put(j, n); 4992 j++; 4993 } 4994 } 4995 _expensive_nodes->trunc_to(j); 4996 4997 // Then sort the list so that similar nodes are next to each other 4998 // and check for at least two nodes of identical kind with same data 4999 // inputs. 5000 sort_expensive_nodes(); 5001 5002 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 5003 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 5004 return true; 5005 } 5006 } 5007 5008 return false; 5009 } 5010 5011 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 5012 if (_expensive_nodes->length() == 0) { 5013 return; 5014 } 5015 5016 assert(OptimizeExpensiveOps, "optimization off?"); 5017 5018 // Sort to bring similar nodes next to each other and clear the 5019 // control input of nodes for which there's only a single copy. 5020 sort_expensive_nodes(); 5021 5022 int j = 0; 5023 int identical = 0; 5024 int i = 0; 5025 bool modified = false; 5026 for (; i < _expensive_nodes->length()-1; i++) { 5027 assert(j <= i, "can't write beyond current index"); 5028 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 5029 identical++; 5030 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 5031 continue; 5032 } 5033 if (identical > 0) { 5034 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 5035 identical = 0; 5036 } else { 5037 Node* n = _expensive_nodes->at(i); 5038 igvn.replace_input_of(n, 0, NULL); 5039 igvn.hash_insert(n); 5040 modified = true; 5041 } 5042 } 5043 if (identical > 0) { 5044 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 5045 } else if (_expensive_nodes->length() >= 1) { 5046 Node* n = _expensive_nodes->at(i); 5047 igvn.replace_input_of(n, 0, NULL); 5048 igvn.hash_insert(n); 5049 modified = true; 5050 } 5051 _expensive_nodes->trunc_to(j); 5052 if (modified) { 5053 igvn.optimize(); 5054 } 5055 } 5056 5057 void Compile::add_expensive_node(Node * n) { 5058 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 5059 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 5060 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 5061 if (OptimizeExpensiveOps) { 5062 _expensive_nodes->append(n); 5063 } else { 5064 // Clear control input and let IGVN optimize expensive nodes if 5065 // OptimizeExpensiveOps is off. 5066 n->set_req(0, NULL); 5067 } 5068 } 5069 5070 /** 5071 * Remove the speculative part of types and clean up the graph 5072 */ 5073 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 5074 if (UseTypeSpeculation) { 5075 Unique_Node_List worklist; 5076 worklist.push(root()); 5077 int modified = 0; 5078 // Go over all type nodes that carry a speculative type, drop the 5079 // speculative part of the type and enqueue the node for an igvn 5080 // which may optimize it out. 5081 for (uint next = 0; next < worklist.size(); ++next) { 5082 Node *n = worklist.at(next); 5083 if (n->is_Type()) { 5084 TypeNode* tn = n->as_Type(); 5085 const Type* t = tn->type(); 5086 const Type* t_no_spec = t->remove_speculative(); 5087 if (t_no_spec != t) { 5088 bool in_hash = igvn.hash_delete(n); 5089 assert(in_hash, "node should be in igvn hash table"); 5090 tn->set_type(t_no_spec); 5091 igvn.hash_insert(n); 5092 igvn._worklist.push(n); // give it a chance to go away 5093 modified++; 5094 } 5095 } 5096 uint max = n->len(); 5097 for( uint i = 0; i < max; ++i ) { 5098 Node *m = n->in(i); 5099 if (not_a_node(m)) continue; 5100 worklist.push(m); 5101 } 5102 } 5103 // Drop the speculative part of all types in the igvn's type table 5104 igvn.remove_speculative_types(); 5105 if (modified > 0) { 5106 igvn.optimize(); 5107 } 5108 #ifdef ASSERT 5109 // Verify that after the IGVN is over no speculative type has resurfaced 5110 worklist.clear(); 5111 worklist.push(root()); 5112 for (uint next = 0; next < worklist.size(); ++next) { 5113 Node *n = worklist.at(next); 5114 const Type* t = igvn.type_or_null(n); 5115 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 5116 if (n->is_Type()) { 5117 t = n->as_Type()->type(); 5118 assert(t == t->remove_speculative(), "no more speculative types"); 5119 } 5120 uint max = n->len(); 5121 for( uint i = 0; i < max; ++i ) { 5122 Node *m = n->in(i); 5123 if (not_a_node(m)) continue; 5124 worklist.push(m); 5125 } 5126 } 5127 igvn.check_no_speculative_types(); 5128 #endif 5129 } 5130 } 5131 5132 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 5133 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 5134 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 5135 if (!EnableValhalla || ta == NULL || tb == NULL || 5136 ta->is_zero_type() || tb->is_zero_type() || 5137 !ta->can_be_value_type() || !tb->can_be_value_type()) { 5138 // Use old acmp if one operand is null or not a value type 5139 return new CmpPNode(a, b); 5140 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) { 5141 // We know that one operand is a value type. Therefore, 5142 // new acmp will only return true if both operands are NULL. 5143 // Check if both operands are null by or'ing the oops. 5144 a = phase->transform(new CastP2XNode(NULL, a)); 5145 b = phase->transform(new CastP2XNode(NULL, b)); 5146 a = phase->transform(new OrXNode(a, b)); 5147 return new CmpXNode(a, phase->MakeConX(0)); 5148 } 5149 // Use new acmp 5150 return NULL; 5151 } 5152 5153 // Auxiliary method to support randomized stressing/fuzzing. 5154 // 5155 // This method can be called the arbitrary number of times, with current count 5156 // as the argument. The logic allows selecting a single candidate from the 5157 // running list of candidates as follows: 5158 // int count = 0; 5159 // Cand* selected = null; 5160 // while(cand = cand->next()) { 5161 // if (randomized_select(++count)) { 5162 // selected = cand; 5163 // } 5164 // } 5165 // 5166 // Including count equalizes the chances any candidate is "selected". 5167 // This is useful when we don't have the complete list of candidates to choose 5168 // from uniformly. In this case, we need to adjust the randomicity of the 5169 // selection, or else we will end up biasing the selection towards the latter 5170 // candidates. 5171 // 5172 // Quick back-envelope calculation shows that for the list of n candidates 5173 // the equal probability for the candidate to persist as "best" can be 5174 // achieved by replacing it with "next" k-th candidate with the probability 5175 // of 1/k. It can be easily shown that by the end of the run, the 5176 // probability for any candidate is converged to 1/n, thus giving the 5177 // uniform distribution among all the candidates. 5178 // 5179 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5180 #define RANDOMIZED_DOMAIN_POW 29 5181 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5182 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5183 bool Compile::randomized_select(int count) { 5184 assert(count > 0, "only positive"); 5185 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5186 } 5187 5188 CloneMap& Compile::clone_map() { return _clone_map; } 5189 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5190 5191 void NodeCloneInfo::dump() const { 5192 tty->print(" {%d:%d} ", idx(), gen()); 5193 } 5194 5195 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5196 uint64_t val = value(old->_idx); 5197 NodeCloneInfo cio(val); 5198 assert(val != 0, "old node should be in the map"); 5199 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5200 insert(nnn->_idx, cin.get()); 5201 #ifndef PRODUCT 5202 if (is_debug()) { 5203 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5204 } 5205 #endif 5206 } 5207 5208 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5209 NodeCloneInfo cio(value(old->_idx)); 5210 if (cio.get() == 0) { 5211 cio.set(old->_idx, 0); 5212 insert(old->_idx, cio.get()); 5213 #ifndef PRODUCT 5214 if (is_debug()) { 5215 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5216 } 5217 #endif 5218 } 5219 clone(old, nnn, gen); 5220 } 5221 5222 int CloneMap::max_gen() const { 5223 int g = 0; 5224 DictI di(_dict); 5225 for(; di.test(); ++di) { 5226 int t = gen(di._key); 5227 if (g < t) { 5228 g = t; 5229 #ifndef PRODUCT 5230 if (is_debug()) { 5231 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5232 } 5233 #endif 5234 } 5235 } 5236 return g; 5237 } 5238 5239 void CloneMap::dump(node_idx_t key) const { 5240 uint64_t val = value(key); 5241 if (val != 0) { 5242 NodeCloneInfo ni(val); 5243 ni.dump(); 5244 } 5245 }