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_last_size(0); 474 compile->set_last_tf(NULL, NULL); 475 compile->set_indexSet_arena(NULL); 476 compile->set_indexSet_free_block_list(NULL); 477 compile->init_type_arena(); 478 Type::Initialize(compile); 479 _compile->set_scratch_buffer_blob(NULL); 480 _compile->begin_method(); 481 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 482 } 483 CompileWrapper::~CompileWrapper() { 484 _compile->end_method(); 485 if (_compile->scratch_buffer_blob() != NULL) 486 BufferBlob::free(_compile->scratch_buffer_blob()); 487 _compile->env()->set_compiler_data(NULL); 488 } 489 490 491 //----------------------------print_compile_messages--------------------------- 492 void Compile::print_compile_messages() { 493 #ifndef PRODUCT 494 // Check if recompiling 495 if (_subsume_loads == false && PrintOpto) { 496 // Recompiling without allowing machine instructions to subsume loads 497 tty->print_cr("*********************************************************"); 498 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 499 tty->print_cr("*********************************************************"); 500 } 501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 502 // Recompiling without escape analysis 503 tty->print_cr("*********************************************************"); 504 tty->print_cr("** Bailout: Recompile without escape analysis **"); 505 tty->print_cr("*********************************************************"); 506 } 507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 508 // Recompiling without boxing elimination 509 tty->print_cr("*********************************************************"); 510 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 511 tty->print_cr("*********************************************************"); 512 } 513 if (C->directive()->BreakAtCompileOption) { 514 // Open the debugger when compiling this method. 515 tty->print("### Breaking when compiling: "); 516 method()->print_short_name(); 517 tty->cr(); 518 BREAKPOINT; 519 } 520 521 if( PrintOpto ) { 522 if (is_osr_compilation()) { 523 tty->print("[OSR]%3d", _compile_id); 524 } else { 525 tty->print("%3d", _compile_id); 526 } 527 } 528 #endif 529 } 530 531 532 //-----------------------init_scratch_buffer_blob------------------------------ 533 // Construct a temporary BufferBlob and cache it for this compile. 534 void Compile::init_scratch_buffer_blob(int const_size) { 535 // If there is already a scratch buffer blob allocated and the 536 // constant section is big enough, use it. Otherwise free the 537 // current and allocate a new one. 538 BufferBlob* blob = scratch_buffer_blob(); 539 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 540 // Use the current blob. 541 } else { 542 if (blob != NULL) { 543 BufferBlob::free(blob); 544 } 545 546 ResourceMark rm; 547 _scratch_const_size = const_size; 548 int size = C2Compiler::initial_code_buffer_size(const_size); 549 #ifdef ASSERT 550 if (C->has_scalarized_args()) { 551 // Oop verification for loading object fields from scalarized value types in the new entry point requires lots of space 552 size += 5120; 553 } 554 #endif 555 blob = BufferBlob::create("Compile::scratch_buffer", size); 556 // Record the buffer blob for next time. 557 set_scratch_buffer_blob(blob); 558 // Have we run out of code space? 559 if (scratch_buffer_blob() == NULL) { 560 // Let CompilerBroker disable further compilations. 561 record_failure("Not enough space for scratch buffer in CodeCache"); 562 return; 563 } 564 } 565 566 // Initialize the relocation buffers 567 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 568 set_scratch_locs_memory(locs_buf); 569 } 570 571 572 //-----------------------scratch_emit_size------------------------------------- 573 // Helper function that computes size by emitting code 574 uint Compile::scratch_emit_size(const Node* n) { 575 // Start scratch_emit_size section. 576 set_in_scratch_emit_size(true); 577 578 // Emit into a trash buffer and count bytes emitted. 579 // This is a pretty expensive way to compute a size, 580 // but it works well enough if seldom used. 581 // All common fixed-size instructions are given a size 582 // method by the AD file. 583 // Note that the scratch buffer blob and locs memory are 584 // allocated at the beginning of the compile task, and 585 // may be shared by several calls to scratch_emit_size. 586 // The allocation of the scratch buffer blob is particularly 587 // expensive, since it has to grab the code cache lock. 588 BufferBlob* blob = this->scratch_buffer_blob(); 589 assert(blob != NULL, "Initialize BufferBlob at start"); 590 assert(blob->size() > MAX_inst_size, "sanity"); 591 relocInfo* locs_buf = scratch_locs_memory(); 592 address blob_begin = blob->content_begin(); 593 address blob_end = (address)locs_buf; 594 assert(blob->contains(blob_end), "sanity"); 595 CodeBuffer buf(blob_begin, blob_end - blob_begin); 596 buf.initialize_consts_size(_scratch_const_size); 597 buf.initialize_stubs_size(MAX_stubs_size); 598 assert(locs_buf != NULL, "sanity"); 599 int lsize = MAX_locs_size / 3; 600 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 601 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 602 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 603 // Mark as scratch buffer. 604 buf.consts()->set_scratch_emit(); 605 buf.insts()->set_scratch_emit(); 606 buf.stubs()->set_scratch_emit(); 607 608 // Do the emission. 609 610 Label fakeL; // Fake label for branch instructions. 611 Label* saveL = NULL; 612 uint save_bnum = 0; 613 bool is_branch = n->is_MachBranch(); 614 if (is_branch) { 615 MacroAssembler masm(&buf); 616 masm.bind(fakeL); 617 n->as_MachBranch()->save_label(&saveL, &save_bnum); 618 n->as_MachBranch()->label_set(&fakeL, 0); 619 } else if (n->is_MachProlog()) { 620 saveL = ((MachPrologNode*)n)->_verified_entry; 621 ((MachPrologNode*)n)->_verified_entry = &fakeL; 622 } 623 n->emit(buf, this->regalloc()); 624 625 // Emitting into the scratch buffer should not fail 626 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 627 628 // Restore label. 629 if (is_branch) { 630 n->as_MachBranch()->label_set(saveL, save_bnum); 631 } else if (n->is_MachProlog()) { 632 ((MachPrologNode*)n)->_verified_entry = saveL; 633 } 634 635 // End scratch_emit_size section. 636 set_in_scratch_emit_size(false); 637 638 return buf.insts_size(); 639 } 640 641 642 // ============================================================================ 643 //------------------------------Compile standard------------------------------- 644 debug_only( int Compile::_debug_idx = 100000; ) 645 646 // Compile a method. entry_bci is -1 for normal compilations and indicates 647 // the continuation bci for on stack replacement. 648 649 650 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 651 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 652 : Phase(Compiler), 653 _compile_id(ci_env->compile_id()), 654 _save_argument_registers(false), 655 _subsume_loads(subsume_loads), 656 _do_escape_analysis(do_escape_analysis), 657 _eliminate_boxing(eliminate_boxing), 658 _method(target), 659 _entry_bci(osr_bci), 660 _stub_function(NULL), 661 _stub_name(NULL), 662 _stub_entry_point(NULL), 663 _max_node_limit(MaxNodeLimit), 664 _orig_pc_slot(0), 665 _orig_pc_slot_offset_in_bytes(0), 666 _sp_inc_slot(0), 667 _sp_inc_slot_offset_in_bytes(0), 668 _inlining_progress(false), 669 _inlining_incrementally(false), 670 _do_cleanup(false), 671 _has_reserved_stack_access(target->has_reserved_stack_access()), 672 #ifndef PRODUCT 673 _trace_opto_output(directive->TraceOptoOutputOption), 674 #endif 675 _has_method_handle_invokes(false), 676 _clinit_barrier_on_entry(false), 677 _comp_arena(mtCompiler), 678 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 679 _env(ci_env), 680 _directive(directive), 681 _log(ci_env->log()), 682 _failure_reason(NULL), 683 _congraph(NULL), 684 #ifndef PRODUCT 685 _printer(IdealGraphPrinter::printer()), 686 #endif 687 _dead_node_list(comp_arena()), 688 _dead_node_count(0), 689 _node_arena(mtCompiler), 690 _old_arena(mtCompiler), 691 _mach_constant_base_node(NULL), 692 _Compile_types(mtCompiler), 693 _initial_gvn(NULL), 694 _for_igvn(NULL), 695 _warm_calls(NULL), 696 _late_inlines(comp_arena(), 2, 0, NULL), 697 _string_late_inlines(comp_arena(), 2, 0, NULL), 698 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 699 _late_inlines_pos(0), 700 _number_of_mh_late_inlines(0), 701 _print_inlining_stream(NULL), 702 _print_inlining_list(NULL), 703 _print_inlining_idx(0), 704 _print_inlining_output(NULL), 705 _replay_inline_data(NULL), 706 _java_calls(0), 707 _inner_loops(0), 708 _interpreter_frame_size(0), 709 _node_bundling_limit(0), 710 _node_bundling_base(NULL), 711 _code_buffer("Compile::Fill_buffer"), 712 _scratch_const_size(-1), 713 _in_scratch_emit_size(false) 714 #ifndef PRODUCT 715 , _in_dump_cnt(0) 716 #endif 717 { 718 C = this; 719 #ifndef PRODUCT 720 if (_printer != NULL) { 721 _printer->set_compile(this); 722 } 723 #endif 724 CompileWrapper cw(this); 725 726 if (CITimeVerbose) { 727 tty->print(" "); 728 target->holder()->name()->print(); 729 tty->print("."); 730 target->print_short_name(); 731 tty->print(" "); 732 } 733 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 734 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 735 736 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 737 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 738 // We can always print a disassembly, either abstract (hex dump) or 739 // with the help of a suitable hsdis library. Thus, we should not 740 // couple print_assembly and print_opto_assembly controls. 741 // But: always print opto and regular assembly on compile command 'print'. 742 bool print_assembly = directive->PrintAssemblyOption; 743 set_print_assembly(print_opto_assembly || print_assembly); 744 #else 745 set_print_assembly(false); // must initialize. 746 #endif 747 748 #ifndef PRODUCT 749 set_parsed_irreducible_loop(false); 750 751 if (directive->ReplayInlineOption) { 752 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 753 } 754 #endif 755 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 756 set_print_intrinsics(directive->PrintIntrinsicsOption); 757 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 758 759 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 760 // Make sure the method being compiled gets its own MDO, 761 // so we can at least track the decompile_count(). 762 // Need MDO to record RTM code generation state. 763 method()->ensure_method_data(); 764 } 765 766 Init(::AliasLevel); 767 768 769 print_compile_messages(); 770 771 _ilt = InlineTree::build_inline_tree_root(); 772 773 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 774 assert(num_alias_types() >= AliasIdxRaw, ""); 775 776 #define MINIMUM_NODE_HASH 1023 777 // Node list that Iterative GVN will start with 778 Unique_Node_List for_igvn(comp_arena()); 779 set_for_igvn(&for_igvn); 780 781 // GVN that will be run immediately on new nodes 782 uint estimated_size = method()->code_size()*4+64; 783 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 784 PhaseGVN gvn(node_arena(), estimated_size); 785 set_initial_gvn(&gvn); 786 787 print_inlining_init(); 788 { // Scope for timing the parser 789 TracePhase tp("parse", &timers[_t_parser]); 790 791 // Put top into the hash table ASAP. 792 initial_gvn()->transform_no_reclaim(top()); 793 794 // Set up tf(), start(), and find a CallGenerator. 795 CallGenerator* cg = NULL; 796 if (is_osr_compilation()) { 797 init_tf(TypeFunc::make(method(), /* is_osr_compilation = */ true)); 798 StartNode* s = new StartOSRNode(root(), tf()->domain_sig()); 799 initial_gvn()->set_type_bottom(s); 800 init_start(s); 801 cg = CallGenerator::for_osr(method(), entry_bci()); 802 } else { 803 // Normal case. 804 init_tf(TypeFunc::make(method())); 805 StartNode* s = new StartNode(root(), tf()->domain_cc()); 806 initial_gvn()->set_type_bottom(s); 807 init_start(s); 808 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 809 // With java.lang.ref.reference.get() we must go through the 810 // intrinsic - even when get() is the root 811 // method of the compile - so that, if necessary, the value in 812 // the referent field of the reference object gets recorded by 813 // the pre-barrier code. 814 cg = find_intrinsic(method(), false); 815 } 816 if (cg == NULL) { 817 float past_uses = method()->interpreter_invocation_count(); 818 float expected_uses = past_uses; 819 cg = CallGenerator::for_inline(method(), expected_uses); 820 } 821 } 822 if (failing()) return; 823 if (cg == NULL) { 824 record_method_not_compilable("cannot parse method"); 825 return; 826 } 827 JVMState* jvms = build_start_state(start(), tf()); 828 if ((jvms = cg->generate(jvms)) == NULL) { 829 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 830 record_method_not_compilable("method parse failed"); 831 } 832 return; 833 } 834 GraphKit kit(jvms); 835 836 if (!kit.stopped()) { 837 // Accept return values, and transfer control we know not where. 838 // This is done by a special, unique ReturnNode bound to root. 839 return_values(kit.jvms()); 840 } 841 842 if (kit.has_exceptions()) { 843 // Any exceptions that escape from this call must be rethrown 844 // to whatever caller is dynamically above us on the stack. 845 // This is done by a special, unique RethrowNode bound to root. 846 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 847 } 848 849 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 850 851 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 852 inline_string_calls(true); 853 } 854 855 if (failing()) return; 856 857 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 858 859 // Remove clutter produced by parsing. 860 if (!failing()) { 861 ResourceMark rm; 862 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 863 } 864 } 865 866 // Note: Large methods are capped off in do_one_bytecode(). 867 if (failing()) return; 868 869 // After parsing, node notes are no longer automagic. 870 // They must be propagated by register_new_node_with_optimizer(), 871 // clone(), or the like. 872 set_default_node_notes(NULL); 873 874 for (;;) { 875 int successes = Inline_Warm(); 876 if (failing()) return; 877 if (successes == 0) break; 878 } 879 880 // Drain the list. 881 Finish_Warm(); 882 #ifndef PRODUCT 883 if (_printer && _printer->should_print(1)) { 884 _printer->print_inlining(); 885 } 886 #endif 887 888 if (failing()) return; 889 NOT_PRODUCT( verify_graph_edges(); ) 890 891 // Now optimize 892 Optimize(); 893 if (failing()) return; 894 NOT_PRODUCT( verify_graph_edges(); ) 895 896 #ifndef PRODUCT 897 if (PrintIdeal) { 898 ttyLocker ttyl; // keep the following output all in one block 899 // This output goes directly to the tty, not the compiler log. 900 // To enable tools to match it up with the compilation activity, 901 // be sure to tag this tty output with the compile ID. 902 if (xtty != NULL) { 903 xtty->head("ideal compile_id='%d'%s", compile_id(), 904 is_osr_compilation() ? " compile_kind='osr'" : 905 ""); 906 } 907 root()->dump(9999); 908 if (xtty != NULL) { 909 xtty->tail("ideal"); 910 } 911 } 912 #endif 913 914 #ifdef ASSERT 915 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 916 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 917 #endif 918 919 // Dump compilation data to replay it. 920 if (directive->DumpReplayOption) { 921 env()->dump_replay_data(_compile_id); 922 } 923 if (directive->DumpInlineOption && (ilt() != NULL)) { 924 env()->dump_inline_data(_compile_id); 925 } 926 927 // Now that we know the size of all the monitors we can add a fixed slot 928 // for the original deopt pc. 929 930 _orig_pc_slot = fixed_slots(); 931 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 932 933 if (needs_stack_repair()) { 934 // One extra slot for the special stack increment value 935 _sp_inc_slot = next_slot; 936 next_slot += 2; 937 } 938 939 set_fixed_slots(next_slot); 940 941 // Compute when to use implicit null checks. Used by matching trap based 942 // nodes and NullCheck optimization. 943 set_allowed_deopt_reasons(); 944 945 // Now generate code 946 Code_Gen(); 947 if (failing()) return; 948 949 // Check if we want to skip execution of all compiled code. 950 { 951 #ifndef PRODUCT 952 if (OptoNoExecute) { 953 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 954 return; 955 } 956 #endif 957 TracePhase tp("install_code", &timers[_t_registerMethod]); 958 959 if (is_osr_compilation()) { 960 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 961 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 962 } else { 963 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 964 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry) == -1) { 965 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, _first_block_size); 966 } 967 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry_RO) == -1) { 968 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, _first_block_size); 969 } 970 if (_code_offsets.value(CodeOffsets::Entry) == -1) { 971 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size); 972 } 973 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 974 } 975 976 env()->register_method(_method, _entry_bci, 977 &_code_offsets, 978 _orig_pc_slot_offset_in_bytes, 979 code_buffer(), 980 frame_size_in_words(), _oop_map_set, 981 &_handler_table, &_inc_table, 982 compiler, 983 has_unsafe_access(), 984 SharedRuntime::is_wide_vector(max_vector_size()), 985 rtm_state() 986 ); 987 988 if (log() != NULL) // Print code cache state into compiler log 989 log()->code_cache_state(); 990 } 991 } 992 993 //------------------------------Compile---------------------------------------- 994 // Compile a runtime stub 995 Compile::Compile( ciEnv* ci_env, 996 TypeFunc_generator generator, 997 address stub_function, 998 const char *stub_name, 999 int is_fancy_jump, 1000 bool pass_tls, 1001 bool save_arg_registers, 1002 bool return_pc, 1003 DirectiveSet* directive) 1004 : Phase(Compiler), 1005 _compile_id(0), 1006 _save_argument_registers(save_arg_registers), 1007 _subsume_loads(true), 1008 _do_escape_analysis(false), 1009 _eliminate_boxing(false), 1010 _method(NULL), 1011 _entry_bci(InvocationEntryBci), 1012 _stub_function(stub_function), 1013 _stub_name(stub_name), 1014 _stub_entry_point(NULL), 1015 _max_node_limit(MaxNodeLimit), 1016 _orig_pc_slot(0), 1017 _orig_pc_slot_offset_in_bytes(0), 1018 _sp_inc_slot(0), 1019 _sp_inc_slot_offset_in_bytes(0), 1020 _inlining_progress(false), 1021 _inlining_incrementally(false), 1022 _has_reserved_stack_access(false), 1023 #ifndef PRODUCT 1024 _trace_opto_output(directive->TraceOptoOutputOption), 1025 #endif 1026 _has_method_handle_invokes(false), 1027 _clinit_barrier_on_entry(false), 1028 _comp_arena(mtCompiler), 1029 _env(ci_env), 1030 _directive(directive), 1031 _log(ci_env->log()), 1032 _failure_reason(NULL), 1033 _congraph(NULL), 1034 #ifndef PRODUCT 1035 _printer(NULL), 1036 #endif 1037 _dead_node_list(comp_arena()), 1038 _dead_node_count(0), 1039 _node_arena(mtCompiler), 1040 _old_arena(mtCompiler), 1041 _mach_constant_base_node(NULL), 1042 _Compile_types(mtCompiler), 1043 _initial_gvn(NULL), 1044 _for_igvn(NULL), 1045 _warm_calls(NULL), 1046 _number_of_mh_late_inlines(0), 1047 _print_inlining_stream(NULL), 1048 _print_inlining_list(NULL), 1049 _print_inlining_idx(0), 1050 _print_inlining_output(NULL), 1051 _replay_inline_data(NULL), 1052 _java_calls(0), 1053 _inner_loops(0), 1054 _interpreter_frame_size(0), 1055 _node_bundling_limit(0), 1056 _node_bundling_base(NULL), 1057 _code_buffer("Compile::Fill_buffer"), 1058 #ifndef PRODUCT 1059 _in_dump_cnt(0), 1060 #endif 1061 _allowed_reasons(0) { 1062 C = this; 1063 1064 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1065 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1066 1067 #ifndef PRODUCT 1068 set_print_assembly(PrintFrameConverterAssembly); 1069 set_parsed_irreducible_loop(false); 1070 #else 1071 set_print_assembly(false); // Must initialize. 1072 #endif 1073 set_has_irreducible_loop(false); // no loops 1074 1075 CompileWrapper cw(this); 1076 Init(/*AliasLevel=*/ 0); 1077 init_tf((*generator)()); 1078 1079 { 1080 // The following is a dummy for the sake of GraphKit::gen_stub 1081 Unique_Node_List for_igvn(comp_arena()); 1082 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1083 PhaseGVN gvn(Thread::current()->resource_area(),255); 1084 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1085 gvn.transform_no_reclaim(top()); 1086 1087 GraphKit kit; 1088 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1089 } 1090 1091 NOT_PRODUCT( verify_graph_edges(); ) 1092 Code_Gen(); 1093 if (failing()) return; 1094 1095 1096 // Entry point will be accessed using compile->stub_entry_point(); 1097 if (code_buffer() == NULL) { 1098 Matcher::soft_match_failure(); 1099 } else { 1100 if (PrintAssembly && (WizardMode || Verbose)) 1101 tty->print_cr("### Stub::%s", stub_name); 1102 1103 if (!failing()) { 1104 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1105 1106 // Make the NMethod 1107 // For now we mark the frame as never safe for profile stackwalking 1108 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1109 code_buffer(), 1110 CodeOffsets::frame_never_safe, 1111 // _code_offsets.value(CodeOffsets::Frame_Complete), 1112 frame_size_in_words(), 1113 _oop_map_set, 1114 save_arg_registers); 1115 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1116 1117 _stub_entry_point = rs->entry_point(); 1118 } 1119 } 1120 } 1121 1122 //------------------------------Init------------------------------------------- 1123 // Prepare for a single compilation 1124 void Compile::Init(int aliaslevel) { 1125 _unique = 0; 1126 _regalloc = NULL; 1127 1128 _tf = NULL; // filled in later 1129 _top = NULL; // cached later 1130 _matcher = NULL; // filled in later 1131 _cfg = NULL; // filled in later 1132 1133 set_24_bit_selection_and_mode(Use24BitFP, false); 1134 1135 _node_note_array = NULL; 1136 _default_node_notes = NULL; 1137 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1138 1139 _immutable_memory = NULL; // filled in at first inquiry 1140 1141 // Globally visible Nodes 1142 // First set TOP to NULL to give safe behavior during creation of RootNode 1143 set_cached_top_node(NULL); 1144 set_root(new RootNode()); 1145 // Now that you have a Root to point to, create the real TOP 1146 set_cached_top_node( new ConNode(Type::TOP) ); 1147 set_recent_alloc(NULL, NULL); 1148 1149 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1150 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1151 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1152 env()->set_dependencies(new Dependencies(env())); 1153 1154 _fixed_slots = 0; 1155 set_has_split_ifs(false); 1156 set_has_loops(has_method() && method()->has_loops()); // first approximation 1157 set_has_stringbuilder(false); 1158 set_has_boxed_value(false); 1159 _trap_can_recompile = false; // no traps emitted yet 1160 _major_progress = true; // start out assuming good things will happen 1161 set_has_unsafe_access(false); 1162 set_max_vector_size(0); 1163 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1164 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1165 set_decompile_count(0); 1166 1167 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1168 _loop_opts_cnt = LoopOptsCount; 1169 _has_flattened_accesses = false; 1170 _flattened_accesses_share_alias = true; 1171 1172 set_do_inlining(Inline); 1173 set_max_inline_size(MaxInlineSize); 1174 set_freq_inline_size(FreqInlineSize); 1175 set_do_scheduling(OptoScheduling); 1176 set_do_count_invocations(false); 1177 set_do_method_data_update(false); 1178 1179 set_do_vector_loop(false); 1180 1181 if (AllowVectorizeOnDemand) { 1182 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1183 set_do_vector_loop(true); 1184 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());}) 1185 } else if (has_method() && method()->name() != 0 && 1186 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1187 set_do_vector_loop(true); 1188 } 1189 } 1190 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1191 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());}) 1192 1193 set_age_code(has_method() && method()->profile_aging()); 1194 set_rtm_state(NoRTM); // No RTM lock eliding by default 1195 _max_node_limit = _directive->MaxNodeLimitOption; 1196 1197 #if INCLUDE_RTM_OPT 1198 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1199 int rtm_state = method()->method_data()->rtm_state(); 1200 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1201 // Don't generate RTM lock eliding code. 1202 set_rtm_state(NoRTM); 1203 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1204 // Generate RTM lock eliding code without abort ratio calculation code. 1205 set_rtm_state(UseRTM); 1206 } else if (UseRTMDeopt) { 1207 // Generate RTM lock eliding code and include abort ratio calculation 1208 // code if UseRTMDeopt is on. 1209 set_rtm_state(ProfileRTM); 1210 } 1211 } 1212 #endif 1213 if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) { 1214 set_clinit_barrier_on_entry(true); 1215 } 1216 if (debug_info()->recording_non_safepoints()) { 1217 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1218 (comp_arena(), 8, 0, NULL)); 1219 set_default_node_notes(Node_Notes::make(this)); 1220 } 1221 1222 // // -- Initialize types before each compile -- 1223 // // Update cached type information 1224 // if( _method && _method->constants() ) 1225 // Type::update_loaded_types(_method, _method->constants()); 1226 1227 // Init alias_type map. 1228 if (!_do_escape_analysis && aliaslevel == 3) 1229 aliaslevel = 2; // No unique types without escape analysis 1230 _AliasLevel = aliaslevel; 1231 const int grow_ats = 16; 1232 _max_alias_types = grow_ats; 1233 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1234 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1235 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1236 { 1237 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1238 } 1239 // Initialize the first few types. 1240 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1241 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1242 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1243 _num_alias_types = AliasIdxRaw+1; 1244 // Zero out the alias type cache. 1245 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1246 // A NULL adr_type hits in the cache right away. Preload the right answer. 1247 probe_alias_cache(NULL)->_index = AliasIdxTop; 1248 1249 _intrinsics = NULL; 1250 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1251 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1252 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1253 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1254 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1255 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1256 register_library_intrinsics(); 1257 } 1258 1259 //---------------------------init_start---------------------------------------- 1260 // Install the StartNode on this compile object. 1261 void Compile::init_start(StartNode* s) { 1262 if (failing()) 1263 return; // already failing 1264 assert(s == start(), ""); 1265 } 1266 1267 /** 1268 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1269 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1270 * the ideal graph. 1271 */ 1272 StartNode* Compile::start() const { 1273 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1274 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1275 Node* start = root()->fast_out(i); 1276 if (start->is_Start()) { 1277 return start->as_Start(); 1278 } 1279 } 1280 fatal("Did not find Start node!"); 1281 return NULL; 1282 } 1283 1284 //-------------------------------immutable_memory------------------------------------- 1285 // Access immutable memory 1286 Node* Compile::immutable_memory() { 1287 if (_immutable_memory != NULL) { 1288 return _immutable_memory; 1289 } 1290 StartNode* s = start(); 1291 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1292 Node *p = s->fast_out(i); 1293 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1294 _immutable_memory = p; 1295 return _immutable_memory; 1296 } 1297 } 1298 ShouldNotReachHere(); 1299 return NULL; 1300 } 1301 1302 //----------------------set_cached_top_node------------------------------------ 1303 // Install the cached top node, and make sure Node::is_top works correctly. 1304 void Compile::set_cached_top_node(Node* tn) { 1305 if (tn != NULL) verify_top(tn); 1306 Node* old_top = _top; 1307 _top = tn; 1308 // Calling Node::setup_is_top allows the nodes the chance to adjust 1309 // their _out arrays. 1310 if (_top != NULL) _top->setup_is_top(); 1311 if (old_top != NULL) old_top->setup_is_top(); 1312 assert(_top == NULL || top()->is_top(), ""); 1313 } 1314 1315 #ifdef ASSERT 1316 uint Compile::count_live_nodes_by_graph_walk() { 1317 Unique_Node_List useful(comp_arena()); 1318 // Get useful node list by walking the graph. 1319 identify_useful_nodes(useful); 1320 return useful.size(); 1321 } 1322 1323 void Compile::print_missing_nodes() { 1324 1325 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1326 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1327 return; 1328 } 1329 1330 // This is an expensive function. It is executed only when the user 1331 // specifies VerifyIdealNodeCount option or otherwise knows the 1332 // additional work that needs to be done to identify reachable nodes 1333 // by walking the flow graph and find the missing ones using 1334 // _dead_node_list. 1335 1336 Unique_Node_List useful(comp_arena()); 1337 // Get useful node list by walking the graph. 1338 identify_useful_nodes(useful); 1339 1340 uint l_nodes = C->live_nodes(); 1341 uint l_nodes_by_walk = useful.size(); 1342 1343 if (l_nodes != l_nodes_by_walk) { 1344 if (_log != NULL) { 1345 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1346 _log->stamp(); 1347 _log->end_head(); 1348 } 1349 VectorSet& useful_member_set = useful.member_set(); 1350 int last_idx = l_nodes_by_walk; 1351 for (int i = 0; i < last_idx; i++) { 1352 if (useful_member_set.test(i)) { 1353 if (_dead_node_list.test(i)) { 1354 if (_log != NULL) { 1355 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1356 } 1357 if (PrintIdealNodeCount) { 1358 // Print the log message to tty 1359 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1360 useful.at(i)->dump(); 1361 } 1362 } 1363 } 1364 else if (! _dead_node_list.test(i)) { 1365 if (_log != NULL) { 1366 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1367 } 1368 if (PrintIdealNodeCount) { 1369 // Print the log message to tty 1370 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1371 } 1372 } 1373 } 1374 if (_log != NULL) { 1375 _log->tail("mismatched_nodes"); 1376 } 1377 } 1378 } 1379 void Compile::record_modified_node(Node* n) { 1380 if (_modified_nodes != NULL && !_inlining_incrementally && 1381 n->outcnt() != 0 && !n->is_Con()) { 1382 _modified_nodes->push(n); 1383 } 1384 } 1385 1386 void Compile::remove_modified_node(Node* n) { 1387 if (_modified_nodes != NULL) { 1388 _modified_nodes->remove(n); 1389 } 1390 } 1391 #endif 1392 1393 #ifndef PRODUCT 1394 void Compile::verify_top(Node* tn) const { 1395 if (tn != NULL) { 1396 assert(tn->is_Con(), "top node must be a constant"); 1397 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1398 assert(tn->in(0) != NULL, "must have live top node"); 1399 } 1400 } 1401 #endif 1402 1403 1404 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1405 1406 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1407 guarantee(arr != NULL, ""); 1408 int num_blocks = arr->length(); 1409 if (grow_by < num_blocks) grow_by = num_blocks; 1410 int num_notes = grow_by * _node_notes_block_size; 1411 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1412 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1413 while (num_notes > 0) { 1414 arr->append(notes); 1415 notes += _node_notes_block_size; 1416 num_notes -= _node_notes_block_size; 1417 } 1418 assert(num_notes == 0, "exact multiple, please"); 1419 } 1420 1421 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1422 if (source == NULL || dest == NULL) return false; 1423 1424 if (dest->is_Con()) 1425 return false; // Do not push debug info onto constants. 1426 1427 #ifdef ASSERT 1428 // Leave a bread crumb trail pointing to the original node: 1429 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1430 dest->set_debug_orig(source); 1431 } 1432 #endif 1433 1434 if (node_note_array() == NULL) 1435 return false; // Not collecting any notes now. 1436 1437 // This is a copy onto a pre-existing node, which may already have notes. 1438 // If both nodes have notes, do not overwrite any pre-existing notes. 1439 Node_Notes* source_notes = node_notes_at(source->_idx); 1440 if (source_notes == NULL || source_notes->is_clear()) return false; 1441 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1442 if (dest_notes == NULL || dest_notes->is_clear()) { 1443 return set_node_notes_at(dest->_idx, source_notes); 1444 } 1445 1446 Node_Notes merged_notes = (*source_notes); 1447 // The order of operations here ensures that dest notes will win... 1448 merged_notes.update_from(dest_notes); 1449 return set_node_notes_at(dest->_idx, &merged_notes); 1450 } 1451 1452 1453 //--------------------------allow_range_check_smearing------------------------- 1454 // Gating condition for coalescing similar range checks. 1455 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1456 // single covering check that is at least as strong as any of them. 1457 // If the optimization succeeds, the simplified (strengthened) range check 1458 // will always succeed. If it fails, we will deopt, and then give up 1459 // on the optimization. 1460 bool Compile::allow_range_check_smearing() const { 1461 // If this method has already thrown a range-check, 1462 // assume it was because we already tried range smearing 1463 // and it failed. 1464 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1465 return !already_trapped; 1466 } 1467 1468 1469 //------------------------------flatten_alias_type----------------------------- 1470 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1471 int offset = tj->offset(); 1472 TypePtr::PTR ptr = tj->ptr(); 1473 1474 // Known instance (scalarizable allocation) alias only with itself. 1475 bool is_known_inst = tj->isa_oopptr() != NULL && 1476 tj->is_oopptr()->is_known_instance(); 1477 1478 // Process weird unsafe references. 1479 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1480 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass(); 1481 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops"); 1482 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1483 tj = TypeOopPtr::BOTTOM; 1484 ptr = tj->ptr(); 1485 offset = tj->offset(); 1486 } 1487 1488 // Array pointers need some flattening 1489 const TypeAryPtr *ta = tj->isa_aryptr(); 1490 if (ta && ta->is_stable()) { 1491 // Erase stability property for alias analysis. 1492 tj = ta = ta->cast_to_stable(false); 1493 } 1494 if (ta && ta->is_not_flat()) { 1495 // Erase not flat property for alias analysis. 1496 tj = ta = ta->cast_to_not_flat(false); 1497 } 1498 if (ta && ta->is_not_null_free()) { 1499 // Erase not null free property for alias analysis. 1500 tj = ta = ta->cast_to_not_null_free(false); 1501 } 1502 1503 if( ta && is_known_inst ) { 1504 if ( offset != Type::OffsetBot && 1505 offset > arrayOopDesc::length_offset_in_bytes() ) { 1506 offset = Type::OffsetBot; // Flatten constant access into array body only 1507 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1508 } 1509 } else if( ta && _AliasLevel >= 2 ) { 1510 // For arrays indexed by constant indices, we flatten the alias 1511 // space to include all of the array body. Only the header, klass 1512 // and array length can be accessed un-aliased. 1513 // For flattened value type array, each field has its own slice so 1514 // we must include the field offset. 1515 if( offset != Type::OffsetBot ) { 1516 if( ta->const_oop() ) { // MethodData* or Method* 1517 offset = Type::OffsetBot; // Flatten constant access into array body 1518 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1519 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1520 // range is OK as-is. 1521 tj = ta = TypeAryPtr::RANGE; 1522 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1523 tj = TypeInstPtr::KLASS; // all klass loads look alike 1524 ta = TypeAryPtr::RANGE; // generic ignored junk 1525 ptr = TypePtr::BotPTR; 1526 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1527 tj = TypeInstPtr::MARK; 1528 ta = TypeAryPtr::RANGE; // generic ignored junk 1529 ptr = TypePtr::BotPTR; 1530 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1531 ta = tj->isa_aryptr(); 1532 } else { // Random constant offset into array body 1533 offset = Type::OffsetBot; // Flatten constant access into array body 1534 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1535 } 1536 } 1537 // Arrays of fixed size alias with arrays of unknown size. 1538 if (ta->size() != TypeInt::POS) { 1539 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1540 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1541 } 1542 // Arrays of known objects become arrays of unknown objects. 1543 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1544 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1545 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1546 } 1547 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1548 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1549 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1550 } 1551 // Initially all flattened array accesses share a single slice 1552 if (ta->elem()->isa_valuetype() && ta->elem() != TypeValueType::BOTTOM && _flattened_accesses_share_alias) { 1553 const TypeAry *tary = TypeAry::make(TypeValueType::BOTTOM, ta->size()); 1554 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), Type::Offset(Type::OffsetBot)); 1555 } 1556 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1557 // cannot be distinguished by bytecode alone. 1558 if (ta->elem() == TypeInt::BOOL) { 1559 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1560 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1561 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1562 } 1563 // During the 2nd round of IterGVN, NotNull castings are removed. 1564 // Make sure the Bottom and NotNull variants alias the same. 1565 // Also, make sure exact and non-exact variants alias the same. 1566 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1567 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1568 } 1569 } 1570 1571 // Oop pointers need some flattening 1572 const TypeInstPtr *to = tj->isa_instptr(); 1573 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1574 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1575 if( ptr == TypePtr::Constant ) { 1576 if (to->klass() != ciEnv::current()->Class_klass() || 1577 offset < k->size_helper() * wordSize) { 1578 // No constant oop pointers (such as Strings); they alias with 1579 // unknown strings. 1580 assert(!is_known_inst, "not scalarizable allocation"); 1581 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset), to->klass()->flatten_array()); 1582 } 1583 } else if( is_known_inst ) { 1584 tj = to; // Keep NotNull and klass_is_exact for instance type 1585 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1586 // During the 2nd round of IterGVN, NotNull castings are removed. 1587 // Make sure the Bottom and NotNull variants alias the same. 1588 // Also, make sure exact and non-exact variants alias the same. 1589 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset), to->klass()->flatten_array()); 1590 } 1591 if (to->speculative() != NULL) { 1592 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->klass()->flatten_array(), to->instance_id()); 1593 } 1594 // Canonicalize the holder of this field 1595 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1596 // First handle header references such as a LoadKlassNode, even if the 1597 // object's klass is unloaded at compile time (4965979). 1598 if (!is_known_inst) { // Do it only for non-instance types 1599 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset), false); 1600 } 1601 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1602 to = tj->is_instptr(); 1603 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1604 // Static fields are in the space above the normal instance 1605 // fields in the java.lang.Class instance. 1606 if (to->klass() != ciEnv::current()->Class_klass()) { 1607 to = NULL; 1608 tj = TypeOopPtr::BOTTOM; 1609 offset = tj->offset(); 1610 } 1611 } else { 1612 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1613 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1614 if( is_known_inst ) { 1615 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), canonical_holder->flatten_array(), to->instance_id()); 1616 } else { 1617 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset), canonical_holder->flatten_array()); 1618 } 1619 } 1620 } 1621 } 1622 1623 // Klass pointers to object array klasses need some flattening 1624 const TypeKlassPtr *tk = tj->isa_klassptr(); 1625 if( tk ) { 1626 // If we are referencing a field within a Klass, we need 1627 // to assume the worst case of an Object. Both exact and 1628 // inexact types must flatten to the same alias class so 1629 // use NotNull as the PTR. 1630 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1631 1632 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1633 TypeKlassPtr::OBJECT->klass(), 1634 Type::Offset(offset), 1635 false); 1636 } 1637 1638 ciKlass* klass = tk->klass(); 1639 if (klass != NULL && klass->is_obj_array_klass()) { 1640 ciKlass* k = TypeAryPtr::OOPS->klass(); 1641 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1642 k = TypeInstPtr::BOTTOM->klass(); 1643 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset), false); 1644 } 1645 1646 // Check for precise loads from the primary supertype array and force them 1647 // to the supertype cache alias index. Check for generic array loads from 1648 // the primary supertype array and also force them to the supertype cache 1649 // alias index. Since the same load can reach both, we need to merge 1650 // these 2 disparate memories into the same alias class. Since the 1651 // primary supertype array is read-only, there's no chance of confusion 1652 // where we bypass an array load and an array store. 1653 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1654 if (offset == Type::OffsetBot || 1655 (offset >= primary_supers_offset && 1656 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1657 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1658 offset = in_bytes(Klass::secondary_super_cache_offset()); 1659 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset), tk->flat_array()); 1660 } 1661 } 1662 1663 // Flatten all Raw pointers together. 1664 if (tj->base() == Type::RawPtr) 1665 tj = TypeRawPtr::BOTTOM; 1666 1667 if (tj->base() == Type::AnyPtr) 1668 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1669 1670 // Flatten all to bottom for now 1671 switch( _AliasLevel ) { 1672 case 0: 1673 tj = TypePtr::BOTTOM; 1674 break; 1675 case 1: // Flatten to: oop, static, field or array 1676 switch (tj->base()) { 1677 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1678 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1679 case Type::AryPtr: // do not distinguish arrays at all 1680 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1681 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1682 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1683 default: ShouldNotReachHere(); 1684 } 1685 break; 1686 case 2: // No collapsing at level 2; keep all splits 1687 case 3: // No collapsing at level 3; keep all splits 1688 break; 1689 default: 1690 Unimplemented(); 1691 } 1692 1693 offset = tj->offset(); 1694 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1695 1696 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1697 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1698 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1699 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1700 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1701 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1702 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1703 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), 1704 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1705 assert( tj->ptr() != TypePtr::TopPTR && 1706 tj->ptr() != TypePtr::AnyNull && 1707 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1708 // assert( tj->ptr() != TypePtr::Constant || 1709 // tj->base() == Type::RawPtr || 1710 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1711 1712 return tj; 1713 } 1714 1715 void Compile::AliasType::Init(int i, const TypePtr* at) { 1716 _index = i; 1717 _adr_type = at; 1718 _field = NULL; 1719 _element = NULL; 1720 _is_rewritable = true; // default 1721 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1722 if (atoop != NULL && atoop->is_known_instance()) { 1723 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1724 _general_index = Compile::current()->get_alias_index(gt); 1725 } else { 1726 _general_index = 0; 1727 } 1728 } 1729 1730 BasicType Compile::AliasType::basic_type() const { 1731 if (element() != NULL) { 1732 const Type* element = adr_type()->is_aryptr()->elem(); 1733 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1734 } if (field() != NULL) { 1735 return field()->layout_type(); 1736 } else { 1737 return T_ILLEGAL; // unknown 1738 } 1739 } 1740 1741 //---------------------------------print_on------------------------------------ 1742 #ifndef PRODUCT 1743 void Compile::AliasType::print_on(outputStream* st) { 1744 if (index() < 10) 1745 st->print("@ <%d> ", index()); 1746 else st->print("@ <%d>", index()); 1747 st->print(is_rewritable() ? " " : " RO"); 1748 int offset = adr_type()->offset(); 1749 if (offset == Type::OffsetBot) 1750 st->print(" +any"); 1751 else st->print(" +%-3d", offset); 1752 st->print(" in "); 1753 adr_type()->dump_on(st); 1754 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1755 if (field() != NULL && tjp) { 1756 if (tjp->klass() != field()->holder() || 1757 tjp->offset() != field()->offset_in_bytes()) { 1758 st->print(" != "); 1759 field()->print(); 1760 st->print(" ***"); 1761 } 1762 } 1763 } 1764 1765 void print_alias_types() { 1766 Compile* C = Compile::current(); 1767 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1768 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1769 C->alias_type(idx)->print_on(tty); 1770 tty->cr(); 1771 } 1772 } 1773 #endif 1774 1775 1776 //----------------------------probe_alias_cache-------------------------------- 1777 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1778 intptr_t key = (intptr_t) adr_type; 1779 key ^= key >> logAliasCacheSize; 1780 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1781 } 1782 1783 1784 //-----------------------------grow_alias_types-------------------------------- 1785 void Compile::grow_alias_types() { 1786 const int old_ats = _max_alias_types; // how many before? 1787 const int new_ats = old_ats; // how many more? 1788 const int grow_ats = old_ats+new_ats; // how many now? 1789 _max_alias_types = grow_ats; 1790 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1791 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1792 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1793 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1794 } 1795 1796 1797 //--------------------------------find_alias_type------------------------------ 1798 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field, bool uncached) { 1799 if (_AliasLevel == 0) 1800 return alias_type(AliasIdxBot); 1801 1802 AliasCacheEntry* ace = NULL; 1803 if (!uncached) { 1804 ace = probe_alias_cache(adr_type); 1805 if (ace->_adr_type == adr_type) { 1806 return alias_type(ace->_index); 1807 } 1808 } 1809 1810 // Handle special cases. 1811 if (adr_type == NULL) return alias_type(AliasIdxTop); 1812 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1813 1814 // Do it the slow way. 1815 const TypePtr* flat = flatten_alias_type(adr_type); 1816 1817 #ifdef ASSERT 1818 { 1819 ResourceMark rm; 1820 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1821 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1822 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1823 Type::str(adr_type)); 1824 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1825 const TypeOopPtr* foop = flat->is_oopptr(); 1826 // Scalarizable allocations have exact klass always. 1827 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1828 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1829 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1830 Type::str(foop), Type::str(xoop)); 1831 } 1832 } 1833 #endif 1834 1835 int idx = AliasIdxTop; 1836 for (int i = 0; i < num_alias_types(); i++) { 1837 if (alias_type(i)->adr_type() == flat) { 1838 idx = i; 1839 break; 1840 } 1841 } 1842 1843 if (idx == AliasIdxTop) { 1844 if (no_create) return NULL; 1845 // Grow the array if necessary. 1846 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1847 // Add a new alias type. 1848 idx = _num_alias_types++; 1849 _alias_types[idx]->Init(idx, flat); 1850 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1851 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1852 if (flat->isa_instptr()) { 1853 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1854 && flat->is_instptr()->klass() == env()->Class_klass()) 1855 alias_type(idx)->set_rewritable(false); 1856 } 1857 ciField* field = NULL; 1858 if (flat->isa_aryptr()) { 1859 #ifdef ASSERT 1860 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1861 // (T_BYTE has the weakest alignment and size restrictions...) 1862 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1863 #endif 1864 const Type* elemtype = flat->is_aryptr()->elem(); 1865 if (flat->offset() == TypePtr::OffsetBot) { 1866 alias_type(idx)->set_element(elemtype); 1867 } 1868 int field_offset = flat->is_aryptr()->field_offset().get(); 1869 if (elemtype->isa_valuetype() && 1870 elemtype->value_klass() != NULL && 1871 field_offset != Type::OffsetBot) { 1872 ciValueKlass* vk = elemtype->value_klass(); 1873 field_offset += vk->first_field_offset(); 1874 field = vk->get_field_by_offset(field_offset, false); 1875 } 1876 } 1877 if (flat->isa_klassptr()) { 1878 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1879 alias_type(idx)->set_rewritable(false); 1880 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1881 alias_type(idx)->set_rewritable(false); 1882 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1883 alias_type(idx)->set_rewritable(false); 1884 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1885 alias_type(idx)->set_rewritable(false); 1886 if (flat->offset() == in_bytes(Klass::layout_helper_offset())) 1887 alias_type(idx)->set_rewritable(false); 1888 } 1889 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1890 // but the base pointer type is not distinctive enough to identify 1891 // references into JavaThread.) 1892 1893 // Check for final fields. 1894 const TypeInstPtr* tinst = flat->isa_instptr(); 1895 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1896 if (tinst->const_oop() != NULL && 1897 tinst->klass() == ciEnv::current()->Class_klass() && 1898 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1899 // static field 1900 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1901 field = k->get_field_by_offset(tinst->offset(), true); 1902 } else if (tinst->klass()->is_valuetype()) { 1903 // Value type field 1904 ciValueKlass* vk = tinst->value_klass(); 1905 field = vk->get_field_by_offset(tinst->offset(), false); 1906 } else { 1907 ciInstanceKlass* k = tinst->klass()->as_instance_klass(); 1908 field = k->get_field_by_offset(tinst->offset(), false); 1909 } 1910 } 1911 assert(field == NULL || 1912 original_field == NULL || 1913 (field->holder() == original_field->holder() && 1914 field->offset() == original_field->offset() && 1915 field->is_static() == original_field->is_static()), "wrong field?"); 1916 // Set field() and is_rewritable() attributes. 1917 if (field != NULL) { 1918 alias_type(idx)->set_field(field); 1919 if (flat->isa_aryptr()) { 1920 // Fields of flattened inline type arrays are rewritable although they are declared final 1921 assert(flat->is_aryptr()->elem()->isa_valuetype(), "must be a flattened value array"); 1922 alias_type(idx)->set_rewritable(true); 1923 } 1924 } 1925 } 1926 1927 // Fill the cache for next time. 1928 if (!uncached) { 1929 ace->_adr_type = adr_type; 1930 ace->_index = idx; 1931 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1932 1933 // Might as well try to fill the cache for the flattened version, too. 1934 AliasCacheEntry* face = probe_alias_cache(flat); 1935 if (face->_adr_type == NULL) { 1936 face->_adr_type = flat; 1937 face->_index = idx; 1938 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1939 } 1940 } 1941 1942 return alias_type(idx); 1943 } 1944 1945 1946 Compile::AliasType* Compile::alias_type(ciField* field) { 1947 const TypeOopPtr* t; 1948 if (field->is_static()) 1949 t = TypeInstPtr::make(field->holder()->java_mirror()); 1950 else 1951 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1952 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1953 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1954 return atp; 1955 } 1956 1957 1958 //------------------------------have_alias_type-------------------------------- 1959 bool Compile::have_alias_type(const TypePtr* adr_type) { 1960 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1961 if (ace->_adr_type == adr_type) { 1962 return true; 1963 } 1964 1965 // Handle special cases. 1966 if (adr_type == NULL) return true; 1967 if (adr_type == TypePtr::BOTTOM) return true; 1968 1969 return find_alias_type(adr_type, true, NULL) != NULL; 1970 } 1971 1972 //-----------------------------must_alias-------------------------------------- 1973 // True if all values of the given address type are in the given alias category. 1974 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1975 if (alias_idx == AliasIdxBot) return true; // the universal category 1976 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1977 if (alias_idx == AliasIdxTop) return false; // the empty category 1978 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1979 1980 // the only remaining possible overlap is identity 1981 int adr_idx = get_alias_index(adr_type); 1982 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1983 assert(adr_idx == alias_idx || 1984 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1985 && adr_type != TypeOopPtr::BOTTOM), 1986 "should not be testing for overlap with an unsafe pointer"); 1987 return adr_idx == alias_idx; 1988 } 1989 1990 //------------------------------can_alias-------------------------------------- 1991 // True if any values of the given address type are in the given alias category. 1992 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1993 if (alias_idx == AliasIdxTop) return false; // the empty category 1994 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1995 if (alias_idx == AliasIdxBot) return true; // the universal category 1996 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1997 1998 // the only remaining possible overlap is identity 1999 int adr_idx = get_alias_index(adr_type); 2000 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 2001 return adr_idx == alias_idx; 2002 } 2003 2004 2005 2006 //---------------------------pop_warm_call------------------------------------- 2007 WarmCallInfo* Compile::pop_warm_call() { 2008 WarmCallInfo* wci = _warm_calls; 2009 if (wci != NULL) _warm_calls = wci->remove_from(wci); 2010 return wci; 2011 } 2012 2013 //----------------------------Inline_Warm-------------------------------------- 2014 int Compile::Inline_Warm() { 2015 // If there is room, try to inline some more warm call sites. 2016 // %%% Do a graph index compaction pass when we think we're out of space? 2017 if (!InlineWarmCalls) return 0; 2018 2019 int calls_made_hot = 0; 2020 int room_to_grow = NodeCountInliningCutoff - unique(); 2021 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 2022 int amount_grown = 0; 2023 WarmCallInfo* call; 2024 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 2025 int est_size = (int)call->size(); 2026 if (est_size > (room_to_grow - amount_grown)) { 2027 // This one won't fit anyway. Get rid of it. 2028 call->make_cold(); 2029 continue; 2030 } 2031 call->make_hot(); 2032 calls_made_hot++; 2033 amount_grown += est_size; 2034 amount_to_grow -= est_size; 2035 } 2036 2037 if (calls_made_hot > 0) set_major_progress(); 2038 return calls_made_hot; 2039 } 2040 2041 2042 //----------------------------Finish_Warm-------------------------------------- 2043 void Compile::Finish_Warm() { 2044 if (!InlineWarmCalls) return; 2045 if (failing()) return; 2046 if (warm_calls() == NULL) return; 2047 2048 // Clean up loose ends, if we are out of space for inlining. 2049 WarmCallInfo* call; 2050 while ((call = pop_warm_call()) != NULL) { 2051 call->make_cold(); 2052 } 2053 } 2054 2055 //---------------------cleanup_loop_predicates----------------------- 2056 // Remove the opaque nodes that protect the predicates so that all unused 2057 // checks and uncommon_traps will be eliminated from the ideal graph 2058 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 2059 if (predicate_count()==0) return; 2060 for (int i = predicate_count(); i > 0; i--) { 2061 Node * n = predicate_opaque1_node(i-1); 2062 assert(n->Opcode() == Op_Opaque1, "must be"); 2063 igvn.replace_node(n, n->in(1)); 2064 } 2065 assert(predicate_count()==0, "should be clean!"); 2066 } 2067 2068 void Compile::add_range_check_cast(Node* n) { 2069 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2070 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 2071 _range_check_casts->append(n); 2072 } 2073 2074 // Remove all range check dependent CastIINodes. 2075 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 2076 for (int i = range_check_cast_count(); i > 0; i--) { 2077 Node* cast = range_check_cast_node(i-1); 2078 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2079 igvn.replace_node(cast, cast->in(1)); 2080 } 2081 assert(range_check_cast_count() == 0, "should be empty"); 2082 } 2083 2084 void Compile::add_opaque4_node(Node* n) { 2085 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 2086 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 2087 _opaque4_nodes->append(n); 2088 } 2089 2090 // Remove all Opaque4 nodes. 2091 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 2092 for (int i = opaque4_count(); i > 0; i--) { 2093 Node* opaq = opaque4_node(i-1); 2094 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2095 igvn.replace_node(opaq, opaq->in(2)); 2096 } 2097 assert(opaque4_count() == 0, "should be empty"); 2098 } 2099 2100 void Compile::add_value_type(Node* n) { 2101 assert(n->is_ValueTypeBase(), "unexpected node"); 2102 if (_value_type_nodes != NULL) { 2103 _value_type_nodes->push(n); 2104 } 2105 } 2106 2107 void Compile::remove_value_type(Node* n) { 2108 assert(n->is_ValueTypeBase(), "unexpected node"); 2109 if (_value_type_nodes != NULL) { 2110 _value_type_nodes->remove(n); 2111 } 2112 } 2113 2114 // Does the return value keep otherwise useless value type allocations 2115 // alive? 2116 static bool return_val_keeps_allocations_alive(Node* ret_val) { 2117 ResourceMark rm; 2118 Unique_Node_List wq; 2119 wq.push(ret_val); 2120 bool some_allocations = false; 2121 for (uint i = 0; i < wq.size(); i++) { 2122 Node* n = wq.at(i); 2123 assert(!n->is_ValueTypeBase(), "chain of value type nodes"); 2124 if (n->outcnt() > 1) { 2125 // Some other use for the allocation 2126 return false; 2127 } else if (n->is_Phi()) { 2128 for (uint j = 1; j < n->req(); j++) { 2129 wq.push(n->in(j)); 2130 } 2131 } else if (n->is_CheckCastPP() && 2132 n->in(1)->is_Proj() && 2133 n->in(1)->in(0)->is_Allocate()) { 2134 some_allocations = true; 2135 } 2136 } 2137 return some_allocations; 2138 } 2139 2140 void Compile::process_value_types(PhaseIterGVN &igvn) { 2141 // Make value types scalar in safepoints 2142 while (_value_type_nodes->size() != 0) { 2143 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 2144 vt->make_scalar_in_safepoints(&igvn); 2145 if (vt->is_ValueTypePtr()) { 2146 igvn.replace_node(vt, vt->get_oop()); 2147 } else if (vt->outcnt() == 0) { 2148 igvn.remove_dead_node(vt); 2149 } 2150 } 2151 _value_type_nodes = NULL; 2152 if (tf()->returns_value_type_as_fields()) { 2153 Node* ret = NULL; 2154 for (uint i = 1; i < root()->req(); i++){ 2155 Node* in = root()->in(i); 2156 if (in->Opcode() == Op_Return) { 2157 assert(ret == NULL, "only one return"); 2158 ret = in; 2159 } 2160 } 2161 if (ret != NULL) { 2162 Node* ret_val = ret->in(TypeFunc::Parms); 2163 if (igvn.type(ret_val)->isa_oopptr() && 2164 return_val_keeps_allocations_alive(ret_val)) { 2165 igvn.replace_input_of(ret, TypeFunc::Parms, ValueTypeNode::tagged_klass(igvn.type(ret_val)->value_klass(), igvn)); 2166 assert(ret_val->outcnt() == 0, "should be dead now"); 2167 igvn.remove_dead_node(ret_val); 2168 } 2169 } 2170 } 2171 igvn.optimize(); 2172 } 2173 2174 void Compile::adjust_flattened_array_access_aliases(PhaseIterGVN& igvn) { 2175 if (!_has_flattened_accesses) { 2176 return; 2177 } 2178 // Initially, all flattened array accesses share the same slice to 2179 // keep dependencies with Object[] array accesses (that could be 2180 // to a flattened array) correct. We're done with parsing so we 2181 // now know all flattened array accesses in this compile 2182 // unit. Let's move flattened array accesses to their own slice, 2183 // one per element field. This should help memory access 2184 // optimizations. 2185 ResourceMark rm; 2186 Unique_Node_List wq; 2187 wq.push(root()); 2188 2189 Node_List mergememnodes; 2190 Node_List memnodes; 2191 2192 // Alias index currently shared by all flattened memory accesses 2193 int index = get_alias_index(TypeAryPtr::VALUES); 2194 2195 // Find MergeMem nodes and flattened array accesses 2196 for (uint i = 0; i < wq.size(); i++) { 2197 Node* n = wq.at(i); 2198 if (n->is_Mem()) { 2199 const TypePtr* adr_type = NULL; 2200 if (n->Opcode() == Op_StoreCM) { 2201 adr_type = get_adr_type(get_alias_index(n->in(MemNode::OopStore)->adr_type())); 2202 } else { 2203 adr_type = get_adr_type(get_alias_index(n->adr_type())); 2204 } 2205 if (adr_type == TypeAryPtr::VALUES) { 2206 memnodes.push(n); 2207 } 2208 } else if (n->is_MergeMem()) { 2209 MergeMemNode* mm = n->as_MergeMem(); 2210 if (mm->memory_at(index) != mm->base_memory()) { 2211 mergememnodes.push(n); 2212 } 2213 } 2214 for (uint j = 0; j < n->req(); j++) { 2215 Node* m = n->in(j); 2216 if (m != NULL) { 2217 wq.push(m); 2218 } 2219 } 2220 } 2221 2222 if (memnodes.size() > 0) { 2223 _flattened_accesses_share_alias = false; 2224 2225 // We are going to change the slice for the flattened array 2226 // accesses so we need to clear the cache entries that refer to 2227 // them. 2228 for (uint i = 0; i < AliasCacheSize; i++) { 2229 AliasCacheEntry* ace = &_alias_cache[i]; 2230 if (ace->_adr_type != NULL && 2231 ace->_adr_type->isa_aryptr() && 2232 ace->_adr_type->is_aryptr()->elem()->isa_valuetype()) { 2233 ace->_adr_type = NULL; 2234 ace->_index = 0; 2235 } 2236 } 2237 2238 // Find what aliases we are going to add 2239 int start_alias = num_alias_types()-1; 2240 int stop_alias = 0; 2241 2242 for (uint i = 0; i < memnodes.size(); i++) { 2243 Node* m = memnodes.at(i); 2244 const TypePtr* adr_type = NULL; 2245 if (m->Opcode() == Op_StoreCM) { 2246 adr_type = m->in(MemNode::OopStore)->adr_type(); 2247 Node* clone = new StoreCMNode(m->in(MemNode::Control), m->in(MemNode::Memory), m->in(MemNode::Address), 2248 m->adr_type(), m->in(MemNode::ValueIn), m->in(MemNode::OopStore), 2249 get_alias_index(adr_type)); 2250 igvn.register_new_node_with_optimizer(clone); 2251 igvn.replace_node(m, clone); 2252 } else { 2253 adr_type = m->adr_type(); 2254 #ifdef ASSERT 2255 m->as_Mem()->set_adr_type(adr_type); 2256 #endif 2257 } 2258 int idx = get_alias_index(adr_type); 2259 start_alias = MIN2(start_alias, idx); 2260 stop_alias = MAX2(stop_alias, idx); 2261 } 2262 2263 assert(stop_alias >= start_alias, "should have expanded aliases"); 2264 2265 Node_Stack stack(0); 2266 #ifdef ASSERT 2267 VectorSet seen(Thread::current()->resource_area()); 2268 #endif 2269 // Now let's fix the memory graph so each flattened array access 2270 // is moved to the right slice. Start from the MergeMem nodes. 2271 uint last = unique(); 2272 for (uint i = 0; i < mergememnodes.size(); i++) { 2273 MergeMemNode* current = mergememnodes.at(i)->as_MergeMem(); 2274 Node* n = current->memory_at(index); 2275 MergeMemNode* mm = NULL; 2276 do { 2277 // Follow memory edges through memory accesses, phis and 2278 // narrow membars and push nodes on the stack. Once we hit 2279 // bottom memory, we pop element off the stack one at a 2280 // time, in reverse order, and move them to the right slice 2281 // by changing their memory edges. 2282 if ((n->is_Phi() && n->adr_type() != TypePtr::BOTTOM) || n->is_Mem() || n->adr_type() == TypeAryPtr::VALUES) { 2283 assert(!seen.test_set(n->_idx), ""); 2284 // Uses (a load for instance) will need to be moved to the 2285 // right slice as well and will get a new memory state 2286 // that we don't know yet. The use could also be the 2287 // backedge of a loop. We put a place holder node between 2288 // the memory node and its uses. We replace that place 2289 // holder with the correct memory state once we know it, 2290 // i.e. when nodes are popped off the stack. Using the 2291 // place holder make the logic work in the presence of 2292 // loops. 2293 if (n->outcnt() > 1) { 2294 Node* place_holder = NULL; 2295 assert(!n->has_out_with(Op_Node), ""); 2296 for (DUIterator k = n->outs(); n->has_out(k); k++) { 2297 Node* u = n->out(k); 2298 if (u != current && u->_idx < last) { 2299 bool success = false; 2300 for (uint l = 0; l < u->req(); l++) { 2301 if (!stack.is_empty() && u == stack.node() && l == stack.index()) { 2302 continue; 2303 } 2304 Node* in = u->in(l); 2305 if (in == n) { 2306 if (place_holder == NULL) { 2307 place_holder = new Node(1); 2308 place_holder->init_req(0, n); 2309 } 2310 igvn.replace_input_of(u, l, place_holder); 2311 success = true; 2312 } 2313 } 2314 if (success) { 2315 --k; 2316 } 2317 } 2318 } 2319 } 2320 if (n->is_Phi()) { 2321 stack.push(n, 1); 2322 n = n->in(1); 2323 } else if (n->is_Mem()) { 2324 stack.push(n, n->req()); 2325 n = n->in(MemNode::Memory); 2326 } else { 2327 assert(n->is_Proj() && n->in(0)->Opcode() == Op_MemBarCPUOrder, ""); 2328 stack.push(n, n->req()); 2329 n = n->in(0)->in(TypeFunc::Memory); 2330 } 2331 } else { 2332 assert(n->adr_type() == TypePtr::BOTTOM || (n->Opcode() == Op_Node && n->_idx >= last) || (n->is_Proj() && n->in(0)->is_Initialize()), ""); 2333 // Build a new MergeMem node to carry the new memory state 2334 // as we build it. IGVN should fold extraneous MergeMem 2335 // nodes. 2336 mm = MergeMemNode::make(n); 2337 igvn.register_new_node_with_optimizer(mm); 2338 while (stack.size() > 0) { 2339 Node* m = stack.node(); 2340 uint idx = stack.index(); 2341 if (m->is_Mem()) { 2342 // Move memory node to its new slice 2343 const TypePtr* adr_type = m->adr_type(); 2344 int alias = get_alias_index(adr_type); 2345 Node* prev = mm->memory_at(alias); 2346 igvn.replace_input_of(m, MemNode::Memory, prev); 2347 mm->set_memory_at(alias, m); 2348 } else if (m->is_Phi()) { 2349 // We need as many new phis as there are new aliases 2350 igvn.replace_input_of(m, idx, mm); 2351 if (idx == m->req()-1) { 2352 Node* r = m->in(0); 2353 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2354 const Type* adr_type = get_adr_type(j); 2355 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) { 2356 continue; 2357 } 2358 Node* phi = new PhiNode(r, Type::MEMORY, get_adr_type(j)); 2359 igvn.register_new_node_with_optimizer(phi); 2360 for (uint k = 1; k < m->req(); k++) { 2361 phi->init_req(k, m->in(k)->as_MergeMem()->memory_at(j)); 2362 } 2363 mm->set_memory_at(j, phi); 2364 } 2365 Node* base_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2366 igvn.register_new_node_with_optimizer(base_phi); 2367 for (uint k = 1; k < m->req(); k++) { 2368 base_phi->init_req(k, m->in(k)->as_MergeMem()->base_memory()); 2369 } 2370 mm->set_base_memory(base_phi); 2371 } 2372 } else { 2373 // This is a MemBarCPUOrder node from 2374 // Parse::array_load()/Parse::array_store(), in the 2375 // branch that handles flattened arrays hidden under 2376 // an Object[] array. We also need one new membar per 2377 // new alias to keep the unknown access that the 2378 // membars protect properly ordered with accesses to 2379 // known flattened array. 2380 assert(m->is_Proj(), "projection expected"); 2381 Node* ctrl = m->in(0)->in(TypeFunc::Control); 2382 igvn.replace_input_of(m->in(0), TypeFunc::Control, top()); 2383 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2384 const Type* adr_type = get_adr_type(j); 2385 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) { 2386 continue; 2387 } 2388 MemBarNode* mb = new MemBarCPUOrderNode(this, j, NULL); 2389 igvn.register_new_node_with_optimizer(mb); 2390 Node* mem = mm->memory_at(j); 2391 mb->init_req(TypeFunc::Control, ctrl); 2392 mb->init_req(TypeFunc::Memory, mem); 2393 ctrl = new ProjNode(mb, TypeFunc::Control); 2394 igvn.register_new_node_with_optimizer(ctrl); 2395 mem = new ProjNode(mb, TypeFunc::Memory); 2396 igvn.register_new_node_with_optimizer(mem); 2397 mm->set_memory_at(j, mem); 2398 } 2399 igvn.replace_node(m->in(0)->as_Multi()->proj_out(TypeFunc::Control), ctrl); 2400 } 2401 if (idx < m->req()-1) { 2402 idx += 1; 2403 stack.set_index(idx); 2404 n = m->in(idx); 2405 break; 2406 } 2407 // Take care of place holder nodes 2408 if (m->has_out_with(Op_Node)) { 2409 Node* place_holder = m->find_out_with(Op_Node); 2410 if (place_holder != NULL) { 2411 Node* mm_clone = mm->clone(); 2412 igvn.register_new_node_with_optimizer(mm_clone); 2413 Node* hook = new Node(1); 2414 hook->init_req(0, mm); 2415 igvn.replace_node(place_holder, mm_clone); 2416 hook->destruct(); 2417 } 2418 assert(!m->has_out_with(Op_Node), "place holder should be gone now"); 2419 } 2420 stack.pop(); 2421 } 2422 } 2423 } while(stack.size() > 0); 2424 // Fix the memory state at the MergeMem we started from 2425 igvn.rehash_node_delayed(current); 2426 for (uint j = (uint)start_alias; j <= (uint)stop_alias; j++) { 2427 const Type* adr_type = get_adr_type(j); 2428 if (!adr_type->isa_aryptr() || !adr_type->is_aryptr()->elem()->isa_valuetype()) { 2429 continue; 2430 } 2431 current->set_memory_at(j, mm); 2432 } 2433 current->set_memory_at(index, current->base_memory()); 2434 } 2435 igvn.optimize(); 2436 } 2437 print_method(PHASE_SPLIT_VALUES_ARRAY, 2); 2438 } 2439 2440 2441 // StringOpts and late inlining of string methods 2442 void Compile::inline_string_calls(bool parse_time) { 2443 { 2444 // remove useless nodes to make the usage analysis simpler 2445 ResourceMark rm; 2446 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2447 } 2448 2449 { 2450 ResourceMark rm; 2451 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2452 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2453 print_method(PHASE_AFTER_STRINGOPTS, 3); 2454 } 2455 2456 // now inline anything that we skipped the first time around 2457 if (!parse_time) { 2458 _late_inlines_pos = _late_inlines.length(); 2459 } 2460 2461 while (_string_late_inlines.length() > 0) { 2462 CallGenerator* cg = _string_late_inlines.pop(); 2463 cg->do_late_inline(); 2464 if (failing()) return; 2465 } 2466 _string_late_inlines.trunc_to(0); 2467 } 2468 2469 // Late inlining of boxing methods 2470 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2471 if (_boxing_late_inlines.length() > 0) { 2472 assert(has_boxed_value(), "inconsistent"); 2473 2474 PhaseGVN* gvn = initial_gvn(); 2475 set_inlining_incrementally(true); 2476 2477 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2478 for_igvn()->clear(); 2479 gvn->replace_with(&igvn); 2480 2481 _late_inlines_pos = _late_inlines.length(); 2482 2483 while (_boxing_late_inlines.length() > 0) { 2484 CallGenerator* cg = _boxing_late_inlines.pop(); 2485 cg->do_late_inline(); 2486 if (failing()) return; 2487 } 2488 _boxing_late_inlines.trunc_to(0); 2489 2490 inline_incrementally_cleanup(igvn); 2491 2492 set_inlining_incrementally(false); 2493 } 2494 } 2495 2496 bool Compile::inline_incrementally_one() { 2497 assert(IncrementalInline, "incremental inlining should be on"); 2498 2499 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2500 set_inlining_progress(false); 2501 set_do_cleanup(false); 2502 int i = 0; 2503 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2504 CallGenerator* cg = _late_inlines.at(i); 2505 _late_inlines_pos = i+1; 2506 cg->do_late_inline(); 2507 if (failing()) return false; 2508 } 2509 int j = 0; 2510 for (; i < _late_inlines.length(); i++, j++) { 2511 _late_inlines.at_put(j, _late_inlines.at(i)); 2512 } 2513 _late_inlines.trunc_to(j); 2514 assert(inlining_progress() || _late_inlines.length() == 0, ""); 2515 2516 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2517 2518 set_inlining_progress(false); 2519 set_do_cleanup(false); 2520 return (_late_inlines.length() > 0) && !needs_cleanup; 2521 } 2522 2523 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2524 { 2525 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2526 ResourceMark rm; 2527 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2528 } 2529 { 2530 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2531 igvn = PhaseIterGVN(initial_gvn()); 2532 igvn.optimize(); 2533 } 2534 } 2535 2536 // Perform incremental inlining until bound on number of live nodes is reached 2537 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2538 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2539 2540 set_inlining_incrementally(true); 2541 uint low_live_nodes = 0; 2542 2543 while (_late_inlines.length() > 0) { 2544 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2545 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2546 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2547 // PhaseIdealLoop is expensive so we only try it once we are 2548 // out of live nodes and we only try it again if the previous 2549 // helped got the number of nodes down significantly 2550 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2551 if (failing()) return; 2552 low_live_nodes = live_nodes(); 2553 _major_progress = true; 2554 } 2555 2556 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2557 break; // finish 2558 } 2559 } 2560 2561 for_igvn()->clear(); 2562 initial_gvn()->replace_with(&igvn); 2563 2564 while (inline_incrementally_one()) { 2565 assert(!failing(), "inconsistent"); 2566 } 2567 2568 if (failing()) return; 2569 2570 inline_incrementally_cleanup(igvn); 2571 2572 if (failing()) return; 2573 } 2574 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2575 2576 if (_string_late_inlines.length() > 0) { 2577 assert(has_stringbuilder(), "inconsistent"); 2578 for_igvn()->clear(); 2579 initial_gvn()->replace_with(&igvn); 2580 2581 inline_string_calls(false); 2582 2583 if (failing()) return; 2584 2585 inline_incrementally_cleanup(igvn); 2586 } 2587 2588 set_inlining_incrementally(false); 2589 } 2590 2591 2592 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2593 if(_loop_opts_cnt > 0) { 2594 debug_only( int cnt = 0; ); 2595 while(major_progress() && (_loop_opts_cnt > 0)) { 2596 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2597 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2598 PhaseIdealLoop::optimize(igvn, mode); 2599 _loop_opts_cnt--; 2600 if (failing()) return false; 2601 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2602 } 2603 } 2604 return true; 2605 } 2606 2607 // Remove edges from "root" to each SafePoint at a backward branch. 2608 // They were inserted during parsing (see add_safepoint()) to make 2609 // infinite loops without calls or exceptions visible to root, i.e., 2610 // useful. 2611 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2612 Node *r = root(); 2613 if (r != NULL) { 2614 for (uint i = r->req(); i < r->len(); ++i) { 2615 Node *n = r->in(i); 2616 if (n != NULL && n->is_SafePoint()) { 2617 r->rm_prec(i); 2618 if (n->outcnt() == 0) { 2619 igvn.remove_dead_node(n); 2620 } 2621 --i; 2622 } 2623 } 2624 } 2625 } 2626 2627 //------------------------------Optimize--------------------------------------- 2628 // Given a graph, optimize it. 2629 void Compile::Optimize() { 2630 TracePhase tp("optimizer", &timers[_t_optimizer]); 2631 2632 #ifndef PRODUCT 2633 if (_directive->BreakAtCompileOption) { 2634 BREAKPOINT; 2635 } 2636 2637 #endif 2638 2639 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2640 #ifdef ASSERT 2641 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2642 #endif 2643 2644 ResourceMark rm; 2645 2646 print_inlining_reinit(); 2647 2648 NOT_PRODUCT( verify_graph_edges(); ) 2649 2650 print_method(PHASE_AFTER_PARSING); 2651 2652 { 2653 // Iterative Global Value Numbering, including ideal transforms 2654 // Initialize IterGVN with types and values from parse-time GVN 2655 PhaseIterGVN igvn(initial_gvn()); 2656 #ifdef ASSERT 2657 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2658 #endif 2659 { 2660 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2661 igvn.optimize(); 2662 } 2663 2664 if (failing()) return; 2665 2666 print_method(PHASE_ITER_GVN1, 2); 2667 2668 inline_incrementally(igvn); 2669 2670 print_method(PHASE_INCREMENTAL_INLINE, 2); 2671 2672 if (failing()) return; 2673 2674 if (eliminate_boxing()) { 2675 // Inline valueOf() methods now. 2676 inline_boxing_calls(igvn); 2677 2678 if (AlwaysIncrementalInline) { 2679 inline_incrementally(igvn); 2680 } 2681 2682 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2683 2684 if (failing()) return; 2685 } 2686 2687 // Now that all inlining is over, cut edge from root to loop 2688 // safepoints 2689 remove_root_to_sfpts_edges(igvn); 2690 2691 // Remove the speculative part of types and clean up the graph from 2692 // the extra CastPP nodes whose only purpose is to carry them. Do 2693 // that early so that optimizations are not disrupted by the extra 2694 // CastPP nodes. 2695 remove_speculative_types(igvn); 2696 2697 // No more new expensive nodes will be added to the list from here 2698 // so keep only the actual candidates for optimizations. 2699 cleanup_expensive_nodes(igvn); 2700 2701 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2702 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2703 initial_gvn()->replace_with(&igvn); 2704 for_igvn()->clear(); 2705 Unique_Node_List new_worklist(C->comp_arena()); 2706 { 2707 ResourceMark rm; 2708 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2709 } 2710 set_for_igvn(&new_worklist); 2711 igvn = PhaseIterGVN(initial_gvn()); 2712 igvn.optimize(); 2713 } 2714 2715 if (_value_type_nodes->size() > 0) { 2716 // Do this once all inlining is over to avoid getting inconsistent debug info 2717 process_value_types(igvn); 2718 } 2719 2720 adjust_flattened_array_access_aliases(igvn); 2721 2722 // Perform escape analysis 2723 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2724 if (has_loops()) { 2725 // Cleanup graph (remove dead nodes). 2726 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2727 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2728 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2729 if (failing()) return; 2730 } 2731 ConnectionGraph::do_analysis(this, &igvn); 2732 2733 if (failing()) return; 2734 2735 // Optimize out fields loads from scalar replaceable allocations. 2736 igvn.optimize(); 2737 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2738 2739 if (failing()) return; 2740 2741 if (congraph() != NULL && macro_count() > 0) { 2742 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2743 PhaseMacroExpand mexp(igvn); 2744 mexp.eliminate_macro_nodes(); 2745 igvn.set_delay_transform(false); 2746 2747 igvn.optimize(); 2748 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2749 2750 if (failing()) return; 2751 } 2752 } 2753 2754 // Loop transforms on the ideal graph. Range Check Elimination, 2755 // peeling, unrolling, etc. 2756 2757 // Set loop opts counter 2758 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2759 { 2760 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2761 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2762 _loop_opts_cnt--; 2763 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2764 if (failing()) return; 2765 } 2766 // Loop opts pass if partial peeling occurred in previous pass 2767 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2768 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2769 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2770 _loop_opts_cnt--; 2771 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2772 if (failing()) return; 2773 } 2774 // Loop opts pass for loop-unrolling before CCP 2775 if(major_progress() && (_loop_opts_cnt > 0)) { 2776 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2777 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2778 _loop_opts_cnt--; 2779 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2780 } 2781 if (!failing()) { 2782 // Verify that last round of loop opts produced a valid graph 2783 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2784 PhaseIdealLoop::verify(igvn); 2785 } 2786 } 2787 if (failing()) return; 2788 2789 // Conditional Constant Propagation; 2790 PhaseCCP ccp( &igvn ); 2791 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2792 { 2793 TracePhase tp("ccp", &timers[_t_ccp]); 2794 ccp.do_transform(); 2795 } 2796 print_method(PHASE_CPP1, 2); 2797 2798 assert( true, "Break here to ccp.dump_old2new_map()"); 2799 2800 // Iterative Global Value Numbering, including ideal transforms 2801 { 2802 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2803 igvn = ccp; 2804 igvn.optimize(); 2805 } 2806 print_method(PHASE_ITER_GVN2, 2); 2807 2808 if (failing()) return; 2809 2810 // Loop transforms on the ideal graph. Range Check Elimination, 2811 // peeling, unrolling, etc. 2812 if (!optimize_loops(igvn, LoopOptsDefault)) { 2813 return; 2814 } 2815 2816 if (failing()) return; 2817 2818 // Ensure that major progress is now clear 2819 C->clear_major_progress(); 2820 2821 { 2822 // Verify that all previous optimizations produced a valid graph 2823 // at least to this point, even if no loop optimizations were done. 2824 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2825 PhaseIdealLoop::verify(igvn); 2826 } 2827 2828 if (range_check_cast_count() > 0) { 2829 // No more loop optimizations. Remove all range check dependent CastIINodes. 2830 C->remove_range_check_casts(igvn); 2831 igvn.optimize(); 2832 } 2833 2834 #ifdef ASSERT 2835 bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand); 2836 #endif 2837 2838 { 2839 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2840 PhaseMacroExpand mex(igvn); 2841 if (mex.expand_macro_nodes()) { 2842 assert(failing(), "must bail out w/ explicit message"); 2843 return; 2844 } 2845 print_method(PHASE_MACRO_EXPANSION, 2); 2846 } 2847 2848 #ifdef ASSERT 2849 bs->verify_gc_barriers(this, BarrierSetC2::BeforeLateInsertion); 2850 #endif 2851 2852 bs->barrier_insertion_phase(C, igvn); 2853 if (failing()) return; 2854 2855 { 2856 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2857 if (bs->expand_barriers(this, igvn)) { 2858 assert(failing(), "must bail out w/ explicit message"); 2859 return; 2860 } 2861 print_method(PHASE_BARRIER_EXPANSION, 2); 2862 } 2863 2864 if (opaque4_count() > 0) { 2865 C->remove_opaque4_nodes(igvn); 2866 igvn.optimize(); 2867 } 2868 2869 DEBUG_ONLY( _modified_nodes = NULL; ) 2870 } // (End scope of igvn; run destructor if necessary for asserts.) 2871 2872 process_print_inlining(); 2873 // A method with only infinite loops has no edges entering loops from root 2874 { 2875 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2876 if (final_graph_reshaping()) { 2877 assert(failing(), "must bail out w/ explicit message"); 2878 return; 2879 } 2880 } 2881 2882 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2883 } 2884 2885 //------------------------------Code_Gen--------------------------------------- 2886 // Given a graph, generate code for it 2887 void Compile::Code_Gen() { 2888 if (failing()) { 2889 return; 2890 } 2891 2892 // Perform instruction selection. You might think we could reclaim Matcher 2893 // memory PDQ, but actually the Matcher is used in generating spill code. 2894 // Internals of the Matcher (including some VectorSets) must remain live 2895 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2896 // set a bit in reclaimed memory. 2897 2898 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2899 // nodes. Mapping is only valid at the root of each matched subtree. 2900 NOT_PRODUCT( verify_graph_edges(); ) 2901 2902 Matcher matcher; 2903 _matcher = &matcher; 2904 { 2905 TracePhase tp("matcher", &timers[_t_matcher]); 2906 matcher.match(); 2907 } 2908 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2909 // nodes. Mapping is only valid at the root of each matched subtree. 2910 NOT_PRODUCT( verify_graph_edges(); ) 2911 2912 // If you have too many nodes, or if matching has failed, bail out 2913 check_node_count(0, "out of nodes matching instructions"); 2914 if (failing()) { 2915 return; 2916 } 2917 2918 print_method(PHASE_MATCHING, 2); 2919 2920 // Build a proper-looking CFG 2921 PhaseCFG cfg(node_arena(), root(), matcher); 2922 _cfg = &cfg; 2923 { 2924 TracePhase tp("scheduler", &timers[_t_scheduler]); 2925 bool success = cfg.do_global_code_motion(); 2926 if (!success) { 2927 return; 2928 } 2929 2930 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2931 NOT_PRODUCT( verify_graph_edges(); ) 2932 debug_only( cfg.verify(); ) 2933 } 2934 2935 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2936 _regalloc = ®alloc; 2937 { 2938 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2939 // Perform register allocation. After Chaitin, use-def chains are 2940 // no longer accurate (at spill code) and so must be ignored. 2941 // Node->LRG->reg mappings are still accurate. 2942 _regalloc->Register_Allocate(); 2943 2944 // Bail out if the allocator builds too many nodes 2945 if (failing()) { 2946 return; 2947 } 2948 } 2949 2950 // Prior to register allocation we kept empty basic blocks in case the 2951 // the allocator needed a place to spill. After register allocation we 2952 // are not adding any new instructions. If any basic block is empty, we 2953 // can now safely remove it. 2954 { 2955 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2956 cfg.remove_empty_blocks(); 2957 if (do_freq_based_layout()) { 2958 PhaseBlockLayout layout(cfg); 2959 } else { 2960 cfg.set_loop_alignment(); 2961 } 2962 cfg.fixup_flow(); 2963 } 2964 2965 // Apply peephole optimizations 2966 if( OptoPeephole ) { 2967 TracePhase tp("peephole", &timers[_t_peephole]); 2968 PhasePeephole peep( _regalloc, cfg); 2969 peep.do_transform(); 2970 } 2971 2972 // Do late expand if CPU requires this. 2973 if (Matcher::require_postalloc_expand) { 2974 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2975 cfg.postalloc_expand(_regalloc); 2976 } 2977 2978 // Convert Nodes to instruction bits in a buffer 2979 { 2980 TraceTime tp("output", &timers[_t_output], CITime); 2981 Output(); 2982 } 2983 2984 print_method(PHASE_FINAL_CODE); 2985 2986 // He's dead, Jim. 2987 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2988 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2989 } 2990 2991 2992 //------------------------------dump_asm--------------------------------------- 2993 // Dump formatted assembly 2994 #if defined(SUPPORT_OPTO_ASSEMBLY) 2995 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) { 2996 2997 int pc_digits = 3; // #chars required for pc 2998 int sb_chars = 3; // #chars for "start bundle" indicator 2999 int tab_size = 8; 3000 if (pcs != NULL) { 3001 int max_pc = 0; 3002 for (uint i = 0; i < pc_limit; i++) { 3003 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc; 3004 } 3005 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc 3006 } 3007 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size; 3008 3009 bool cut_short = false; 3010 st->print_cr("#"); 3011 st->print("# "); _tf->dump_on(st); st->cr(); 3012 st->print_cr("#"); 3013 3014 // For all blocks 3015 int pc = 0x0; // Program counter 3016 char starts_bundle = ' '; 3017 _regalloc->dump_frame(); 3018 3019 Node *n = NULL; 3020 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 3021 if (VMThread::should_terminate()) { 3022 cut_short = true; 3023 break; 3024 } 3025 Block* block = _cfg->get_block(i); 3026 if (block->is_connector() && !Verbose) { 3027 continue; 3028 } 3029 n = block->head(); 3030 if ((pcs != NULL) && (n->_idx < pc_limit)) { 3031 pc = pcs[n->_idx]; 3032 st->print("%*.*x", pc_digits, pc_digits, pc); 3033 } 3034 st->fill_to(prefix_len); 3035 block->dump_head(_cfg, st); 3036 if (block->is_connector()) { 3037 st->fill_to(prefix_len); 3038 st->print_cr("# Empty connector block"); 3039 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 3040 st->fill_to(prefix_len); 3041 st->print_cr("# Block is sole successor of call"); 3042 } 3043 3044 // For all instructions 3045 Node *delay = NULL; 3046 for (uint j = 0; j < block->number_of_nodes(); j++) { 3047 if (VMThread::should_terminate()) { 3048 cut_short = true; 3049 break; 3050 } 3051 n = block->get_node(j); 3052 if (valid_bundle_info(n)) { 3053 Bundle* bundle = node_bundling(n); 3054 if (bundle->used_in_unconditional_delay()) { 3055 delay = n; 3056 continue; 3057 } 3058 if (bundle->starts_bundle()) { 3059 starts_bundle = '+'; 3060 } 3061 } 3062 3063 if (WizardMode) { 3064 n->dump(); 3065 } 3066 3067 if( !n->is_Region() && // Dont print in the Assembly 3068 !n->is_Phi() && // a few noisely useless nodes 3069 !n->is_Proj() && 3070 !n->is_MachTemp() && 3071 !n->is_SafePointScalarObject() && 3072 !n->is_Catch() && // Would be nice to print exception table targets 3073 !n->is_MergeMem() && // Not very interesting 3074 !n->is_top() && // Debug info table constants 3075 !(n->is_Con() && !n->is_Mach())// Debug info table constants 3076 ) { 3077 if ((pcs != NULL) && (n->_idx < pc_limit)) { 3078 pc = pcs[n->_idx]; 3079 st->print("%*.*x", pc_digits, pc_digits, pc); 3080 } else { 3081 st->fill_to(pc_digits); 3082 } 3083 st->print(" %c ", starts_bundle); 3084 starts_bundle = ' '; 3085 st->fill_to(prefix_len); 3086 n->format(_regalloc, st); 3087 st->cr(); 3088 } 3089 3090 // If we have an instruction with a delay slot, and have seen a delay, 3091 // then back up and print it 3092 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 3093 // Coverity finding - Explicit null dereferenced. 3094 guarantee(delay != NULL, "no unconditional delay instruction"); 3095 if (WizardMode) delay->dump(); 3096 3097 if (node_bundling(delay)->starts_bundle()) 3098 starts_bundle = '+'; 3099 if ((pcs != NULL) && (n->_idx < pc_limit)) { 3100 pc = pcs[n->_idx]; 3101 st->print("%*.*x", pc_digits, pc_digits, pc); 3102 } else { 3103 st->fill_to(pc_digits); 3104 } 3105 st->print(" %c ", starts_bundle); 3106 starts_bundle = ' '; 3107 st->fill_to(prefix_len); 3108 delay->format(_regalloc, st); 3109 st->cr(); 3110 delay = NULL; 3111 } 3112 3113 // Dump the exception table as well 3114 if( n->is_Catch() && (Verbose || WizardMode) ) { 3115 // Print the exception table for this offset 3116 _handler_table.print_subtable_for(pc); 3117 } 3118 st->bol(); // Make sure we start on a new line 3119 } 3120 st->cr(); // one empty line between blocks 3121 assert(cut_short || delay == NULL, "no unconditional delay branch"); 3122 } // End of per-block dump 3123 3124 if (cut_short) st->print_cr("*** disassembly is cut short ***"); 3125 } 3126 #endif 3127 3128 //------------------------------Final_Reshape_Counts--------------------------- 3129 // This class defines counters to help identify when a method 3130 // may/must be executed using hardware with only 24-bit precision. 3131 struct Final_Reshape_Counts : public StackObj { 3132 int _call_count; // count non-inlined 'common' calls 3133 int _float_count; // count float ops requiring 24-bit precision 3134 int _double_count; // count double ops requiring more precision 3135 int _java_call_count; // count non-inlined 'java' calls 3136 int _inner_loop_count; // count loops which need alignment 3137 VectorSet _visited; // Visitation flags 3138 Node_List _tests; // Set of IfNodes & PCTableNodes 3139 3140 Final_Reshape_Counts() : 3141 _call_count(0), _float_count(0), _double_count(0), 3142 _java_call_count(0), _inner_loop_count(0), 3143 _visited( Thread::current()->resource_area() ) { } 3144 3145 void inc_call_count () { _call_count ++; } 3146 void inc_float_count () { _float_count ++; } 3147 void inc_double_count() { _double_count++; } 3148 void inc_java_call_count() { _java_call_count++; } 3149 void inc_inner_loop_count() { _inner_loop_count++; } 3150 3151 int get_call_count () const { return _call_count ; } 3152 int get_float_count () const { return _float_count ; } 3153 int get_double_count() const { return _double_count; } 3154 int get_java_call_count() const { return _java_call_count; } 3155 int get_inner_loop_count() const { return _inner_loop_count; } 3156 }; 3157 3158 #ifdef ASSERT 3159 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 3160 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 3161 // Make sure the offset goes inside the instance layout. 3162 return k->contains_field_offset(tp->offset()); 3163 // Note that OffsetBot and OffsetTop are very negative. 3164 } 3165 #endif 3166 3167 // Eliminate trivially redundant StoreCMs and accumulate their 3168 // precedence edges. 3169 void Compile::eliminate_redundant_card_marks(Node* n) { 3170 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 3171 if (n->in(MemNode::Address)->outcnt() > 1) { 3172 // There are multiple users of the same address so it might be 3173 // possible to eliminate some of the StoreCMs 3174 Node* mem = n->in(MemNode::Memory); 3175 Node* adr = n->in(MemNode::Address); 3176 Node* val = n->in(MemNode::ValueIn); 3177 Node* prev = n; 3178 bool done = false; 3179 // Walk the chain of StoreCMs eliminating ones that match. As 3180 // long as it's a chain of single users then the optimization is 3181 // safe. Eliminating partially redundant StoreCMs would require 3182 // cloning copies down the other paths. 3183 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 3184 if (adr == mem->in(MemNode::Address) && 3185 val == mem->in(MemNode::ValueIn)) { 3186 // redundant StoreCM 3187 if (mem->req() > MemNode::OopStore) { 3188 // Hasn't been processed by this code yet. 3189 n->add_prec(mem->in(MemNode::OopStore)); 3190 } else { 3191 // Already converted to precedence edge 3192 for (uint i = mem->req(); i < mem->len(); i++) { 3193 // Accumulate any precedence edges 3194 if (mem->in(i) != NULL) { 3195 n->add_prec(mem->in(i)); 3196 } 3197 } 3198 // Everything above this point has been processed. 3199 done = true; 3200 } 3201 // Eliminate the previous StoreCM 3202 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 3203 assert(mem->outcnt() == 0, "should be dead"); 3204 mem->disconnect_inputs(NULL, this); 3205 } else { 3206 prev = mem; 3207 } 3208 mem = prev->in(MemNode::Memory); 3209 } 3210 } 3211 } 3212 3213 3214 //------------------------------final_graph_reshaping_impl---------------------- 3215 // Implement items 1-5 from final_graph_reshaping below. 3216 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 3217 3218 if ( n->outcnt() == 0 ) return; // dead node 3219 uint nop = n->Opcode(); 3220 3221 // Check for 2-input instruction with "last use" on right input. 3222 // Swap to left input. Implements item (2). 3223 if( n->req() == 3 && // two-input instruction 3224 n->in(1)->outcnt() > 1 && // left use is NOT a last use 3225 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 3226 n->in(2)->outcnt() == 1 &&// right use IS a last use 3227 !n->in(2)->is_Con() ) { // right use is not a constant 3228 // Check for commutative opcode 3229 switch( nop ) { 3230 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 3231 case Op_MaxI: case Op_MinI: 3232 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 3233 case Op_AndL: case Op_XorL: case Op_OrL: 3234 case Op_AndI: case Op_XorI: case Op_OrI: { 3235 // Move "last use" input to left by swapping inputs 3236 n->swap_edges(1, 2); 3237 break; 3238 } 3239 default: 3240 break; 3241 } 3242 } 3243 3244 #ifdef ASSERT 3245 if( n->is_Mem() ) { 3246 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 3247 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 3248 // oop will be recorded in oop map if load crosses safepoint 3249 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 3250 LoadNode::is_immutable_value(n->in(MemNode::Address))), 3251 "raw memory operations should have control edge"); 3252 } 3253 if (n->is_MemBar()) { 3254 MemBarNode* mb = n->as_MemBar(); 3255 if (mb->trailing_store() || mb->trailing_load_store()) { 3256 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 3257 Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent)); 3258 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 3259 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 3260 } else if (mb->leading()) { 3261 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 3262 } 3263 } 3264 #endif 3265 // Count FPU ops and common calls, implements item (3) 3266 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 3267 if (!gc_handled) { 3268 final_graph_reshaping_main_switch(n, frc, nop); 3269 } 3270 3271 // Collect CFG split points 3272 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3273 frc._tests.push(n); 3274 } 3275 } 3276 3277 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 3278 switch( nop ) { 3279 // Count all float operations that may use FPU 3280 case Op_AddF: 3281 case Op_SubF: 3282 case Op_MulF: 3283 case Op_DivF: 3284 case Op_NegF: 3285 case Op_ModF: 3286 case Op_ConvI2F: 3287 case Op_ConF: 3288 case Op_CmpF: 3289 case Op_CmpF3: 3290 // case Op_ConvL2F: // longs are split into 32-bit halves 3291 frc.inc_float_count(); 3292 break; 3293 3294 case Op_ConvF2D: 3295 case Op_ConvD2F: 3296 frc.inc_float_count(); 3297 frc.inc_double_count(); 3298 break; 3299 3300 // Count all double operations that may use FPU 3301 case Op_AddD: 3302 case Op_SubD: 3303 case Op_MulD: 3304 case Op_DivD: 3305 case Op_NegD: 3306 case Op_ModD: 3307 case Op_ConvI2D: 3308 case Op_ConvD2I: 3309 // case Op_ConvL2D: // handled by leaf call 3310 // case Op_ConvD2L: // handled by leaf call 3311 case Op_ConD: 3312 case Op_CmpD: 3313 case Op_CmpD3: 3314 frc.inc_double_count(); 3315 break; 3316 case Op_Opaque1: // Remove Opaque Nodes before matching 3317 case Op_Opaque2: // Remove Opaque Nodes before matching 3318 case Op_Opaque3: 3319 n->subsume_by(n->in(1), this); 3320 break; 3321 case Op_CallStaticJava: 3322 case Op_CallJava: 3323 case Op_CallDynamicJava: 3324 frc.inc_java_call_count(); // Count java call site; 3325 case Op_CallRuntime: 3326 case Op_CallLeaf: 3327 case Op_CallLeafNoFP: { 3328 assert (n->is_Call(), ""); 3329 CallNode *call = n->as_Call(); 3330 // Count call sites where the FP mode bit would have to be flipped. 3331 // Do not count uncommon runtime calls: 3332 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 3333 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 3334 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 3335 frc.inc_call_count(); // Count the call site 3336 } else { // See if uncommon argument is shared 3337 Node *n = call->in(TypeFunc::Parms); 3338 int nop = n->Opcode(); 3339 // Clone shared simple arguments to uncommon calls, item (1). 3340 if (n->outcnt() > 1 && 3341 !n->is_Proj() && 3342 nop != Op_CreateEx && 3343 nop != Op_CheckCastPP && 3344 nop != Op_DecodeN && 3345 nop != Op_DecodeNKlass && 3346 !n->is_Mem() && 3347 !n->is_Phi()) { 3348 Node *x = n->clone(); 3349 call->set_req(TypeFunc::Parms, x); 3350 } 3351 } 3352 break; 3353 } 3354 3355 case Op_StoreD: 3356 case Op_LoadD: 3357 case Op_LoadD_unaligned: 3358 frc.inc_double_count(); 3359 goto handle_mem; 3360 case Op_StoreF: 3361 case Op_LoadF: 3362 frc.inc_float_count(); 3363 goto handle_mem; 3364 3365 case Op_StoreCM: 3366 { 3367 // Convert OopStore dependence into precedence edge 3368 Node* prec = n->in(MemNode::OopStore); 3369 n->del_req(MemNode::OopStore); 3370 n->add_prec(prec); 3371 eliminate_redundant_card_marks(n); 3372 } 3373 3374 // fall through 3375 3376 case Op_StoreB: 3377 case Op_StoreC: 3378 case Op_StorePConditional: 3379 case Op_StoreI: 3380 case Op_StoreL: 3381 case Op_StoreIConditional: 3382 case Op_StoreLConditional: 3383 case Op_CompareAndSwapB: 3384 case Op_CompareAndSwapS: 3385 case Op_CompareAndSwapI: 3386 case Op_CompareAndSwapL: 3387 case Op_CompareAndSwapP: 3388 case Op_CompareAndSwapN: 3389 case Op_WeakCompareAndSwapB: 3390 case Op_WeakCompareAndSwapS: 3391 case Op_WeakCompareAndSwapI: 3392 case Op_WeakCompareAndSwapL: 3393 case Op_WeakCompareAndSwapP: 3394 case Op_WeakCompareAndSwapN: 3395 case Op_CompareAndExchangeB: 3396 case Op_CompareAndExchangeS: 3397 case Op_CompareAndExchangeI: 3398 case Op_CompareAndExchangeL: 3399 case Op_CompareAndExchangeP: 3400 case Op_CompareAndExchangeN: 3401 case Op_GetAndAddS: 3402 case Op_GetAndAddB: 3403 case Op_GetAndAddI: 3404 case Op_GetAndAddL: 3405 case Op_GetAndSetS: 3406 case Op_GetAndSetB: 3407 case Op_GetAndSetI: 3408 case Op_GetAndSetL: 3409 case Op_GetAndSetP: 3410 case Op_GetAndSetN: 3411 case Op_StoreP: 3412 case Op_StoreN: 3413 case Op_StoreNKlass: 3414 case Op_LoadB: 3415 case Op_LoadUB: 3416 case Op_LoadUS: 3417 case Op_LoadI: 3418 case Op_LoadKlass: 3419 case Op_LoadNKlass: 3420 case Op_LoadL: 3421 case Op_LoadL_unaligned: 3422 case Op_LoadPLocked: 3423 case Op_LoadP: 3424 case Op_LoadN: 3425 case Op_LoadRange: 3426 case Op_LoadS: { 3427 handle_mem: 3428 #ifdef ASSERT 3429 if( VerifyOptoOopOffsets ) { 3430 MemNode* mem = n->as_Mem(); 3431 // Check to see if address types have grounded out somehow. 3432 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3433 assert( !tp || oop_offset_is_sane(tp), "" ); 3434 } 3435 #endif 3436 if (EnableValhalla && (nop == Op_LoadKlass || nop == Op_LoadNKlass)) { 3437 const TypeKlassPtr* tk = n->bottom_type()->make_ptr()->is_klassptr(); 3438 assert(!tk->klass_is_exact(), "should have been folded"); 3439 if (tk->klass()->can_be_value_array_klass() && n->as_Mem()->adr_type()->offset() == oopDesc::klass_offset_in_bytes()) { 3440 // Array load klass needs to filter out property bits (but not 3441 // GetNullFreePropertyNode or GetFlattenedPropertyNode which 3442 // needs to extract the storage property bits) 3443 uint last = unique(); 3444 Node* pointer = NULL; 3445 if (nop == Op_LoadKlass) { 3446 Node* cast = new CastP2XNode(NULL, n); 3447 Node* masked = new LShiftXNode(cast, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits))); 3448 masked = new RShiftXNode(masked, new ConINode(TypeInt::make(oopDesc::storage_props_nof_bits))); 3449 pointer = new CastX2PNode(masked); 3450 pointer = new CheckCastPPNode(NULL, pointer, n->bottom_type()); 3451 } else { 3452 Node* cast = new CastN2INode(n); 3453 Node* masked = new AndINode(cast, new ConINode(TypeInt::make(oopDesc::compressed_klass_mask()))); 3454 pointer = new CastI2NNode(masked, n->bottom_type()); 3455 } 3456 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3457 Node* u = n->fast_out(i); 3458 if (u->_idx < last && u->Opcode() != Op_GetNullFreeProperty && u->Opcode() != Op_GetFlattenedProperty) { 3459 // If user is a comparison with a klass that can't be a value type 3460 // array klass, we don't need to clear the storage property bits. 3461 Node* cmp = (u->is_DecodeNKlass() && u->outcnt() == 1) ? u->unique_out() : u; 3462 if (cmp->is_Cmp()) { 3463 const TypeKlassPtr* kp1 = cmp->in(1)->bottom_type()->make_ptr()->isa_klassptr(); 3464 const TypeKlassPtr* kp2 = cmp->in(2)->bottom_type()->make_ptr()->isa_klassptr(); 3465 if ((kp1 != NULL && !kp1->klass()->can_be_value_array_klass()) || 3466 (kp2 != NULL && !kp2->klass()->can_be_value_array_klass())) { 3467 continue; 3468 } 3469 } 3470 int nb = u->replace_edge(n, pointer); 3471 --i, imax -= nb; 3472 } 3473 } 3474 } 3475 } 3476 break; 3477 } 3478 3479 case Op_AddP: { // Assert sane base pointers 3480 Node *addp = n->in(AddPNode::Address); 3481 assert( !addp->is_AddP() || 3482 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3483 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3484 "Base pointers must match (addp %u)", addp->_idx ); 3485 #ifdef _LP64 3486 if ((UseCompressedOops || UseCompressedClassPointers) && 3487 addp->Opcode() == Op_ConP && 3488 addp == n->in(AddPNode::Base) && 3489 n->in(AddPNode::Offset)->is_Con()) { 3490 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3491 // on the platform and on the compressed oops mode. 3492 // Use addressing with narrow klass to load with offset on x86. 3493 // Some platforms can use the constant pool to load ConP. 3494 // Do this transformation here since IGVN will convert ConN back to ConP. 3495 const Type* t = addp->bottom_type(); 3496 bool is_oop = t->isa_oopptr() != NULL; 3497 bool is_klass = t->isa_klassptr() != NULL; 3498 3499 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3500 (is_klass && Matcher::const_klass_prefer_decode())) { 3501 Node* nn = NULL; 3502 3503 int op = is_oop ? Op_ConN : Op_ConNKlass; 3504 3505 // Look for existing ConN node of the same exact type. 3506 Node* r = root(); 3507 uint cnt = r->outcnt(); 3508 for (uint i = 0; i < cnt; i++) { 3509 Node* m = r->raw_out(i); 3510 if (m!= NULL && m->Opcode() == op && 3511 m->bottom_type()->make_ptr() == t) { 3512 nn = m; 3513 break; 3514 } 3515 } 3516 if (nn != NULL) { 3517 // Decode a narrow oop to match address 3518 // [R12 + narrow_oop_reg<<3 + offset] 3519 if (is_oop) { 3520 nn = new DecodeNNode(nn, t); 3521 } else { 3522 nn = new DecodeNKlassNode(nn, t); 3523 } 3524 // Check for succeeding AddP which uses the same Base. 3525 // Otherwise we will run into the assertion above when visiting that guy. 3526 for (uint i = 0; i < n->outcnt(); ++i) { 3527 Node *out_i = n->raw_out(i); 3528 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3529 out_i->set_req(AddPNode::Base, nn); 3530 #ifdef ASSERT 3531 for (uint j = 0; j < out_i->outcnt(); ++j) { 3532 Node *out_j = out_i->raw_out(j); 3533 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3534 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3535 } 3536 #endif 3537 } 3538 } 3539 n->set_req(AddPNode::Base, nn); 3540 n->set_req(AddPNode::Address, nn); 3541 if (addp->outcnt() == 0) { 3542 addp->disconnect_inputs(NULL, this); 3543 } 3544 } 3545 } 3546 } 3547 #endif 3548 // platform dependent reshaping of the address expression 3549 reshape_address(n->as_AddP()); 3550 break; 3551 } 3552 3553 case Op_CastPP: { 3554 // Remove CastPP nodes to gain more freedom during scheduling but 3555 // keep the dependency they encode as control or precedence edges 3556 // (if control is set already) on memory operations. Some CastPP 3557 // nodes don't have a control (don't carry a dependency): skip 3558 // those. 3559 if (n->in(0) != NULL) { 3560 ResourceMark rm; 3561 Unique_Node_List wq; 3562 wq.push(n); 3563 for (uint next = 0; next < wq.size(); ++next) { 3564 Node *m = wq.at(next); 3565 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3566 Node* use = m->fast_out(i); 3567 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3568 use->ensure_control_or_add_prec(n->in(0)); 3569 } else { 3570 switch(use->Opcode()) { 3571 case Op_AddP: 3572 case Op_DecodeN: 3573 case Op_DecodeNKlass: 3574 case Op_CheckCastPP: 3575 case Op_CastPP: 3576 wq.push(use); 3577 break; 3578 } 3579 } 3580 } 3581 } 3582 } 3583 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3584 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3585 Node* in1 = n->in(1); 3586 const Type* t = n->bottom_type(); 3587 Node* new_in1 = in1->clone(); 3588 new_in1->as_DecodeN()->set_type(t); 3589 3590 if (!Matcher::narrow_oop_use_complex_address()) { 3591 // 3592 // x86, ARM and friends can handle 2 adds in addressing mode 3593 // and Matcher can fold a DecodeN node into address by using 3594 // a narrow oop directly and do implicit NULL check in address: 3595 // 3596 // [R12 + narrow_oop_reg<<3 + offset] 3597 // NullCheck narrow_oop_reg 3598 // 3599 // On other platforms (Sparc) we have to keep new DecodeN node and 3600 // use it to do implicit NULL check in address: 3601 // 3602 // decode_not_null narrow_oop_reg, base_reg 3603 // [base_reg + offset] 3604 // NullCheck base_reg 3605 // 3606 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3607 // to keep the information to which NULL check the new DecodeN node 3608 // corresponds to use it as value in implicit_null_check(). 3609 // 3610 new_in1->set_req(0, n->in(0)); 3611 } 3612 3613 n->subsume_by(new_in1, this); 3614 if (in1->outcnt() == 0) { 3615 in1->disconnect_inputs(NULL, this); 3616 } 3617 } else { 3618 n->subsume_by(n->in(1), this); 3619 if (n->outcnt() == 0) { 3620 n->disconnect_inputs(NULL, this); 3621 } 3622 } 3623 break; 3624 } 3625 #ifdef _LP64 3626 case Op_CmpP: 3627 // Do this transformation here to preserve CmpPNode::sub() and 3628 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3629 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3630 Node* in1 = n->in(1); 3631 Node* in2 = n->in(2); 3632 if (!in1->is_DecodeNarrowPtr()) { 3633 in2 = in1; 3634 in1 = n->in(2); 3635 } 3636 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3637 3638 Node* new_in2 = NULL; 3639 if (in2->is_DecodeNarrowPtr()) { 3640 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3641 new_in2 = in2->in(1); 3642 } else if (in2->Opcode() == Op_ConP) { 3643 const Type* t = in2->bottom_type(); 3644 if (t == TypePtr::NULL_PTR) { 3645 assert(in1->is_DecodeN(), "compare klass to null?"); 3646 // Don't convert CmpP null check into CmpN if compressed 3647 // oops implicit null check is not generated. 3648 // This will allow to generate normal oop implicit null check. 3649 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3650 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3651 // 3652 // This transformation together with CastPP transformation above 3653 // will generated code for implicit NULL checks for compressed oops. 3654 // 3655 // The original code after Optimize() 3656 // 3657 // LoadN memory, narrow_oop_reg 3658 // decode narrow_oop_reg, base_reg 3659 // CmpP base_reg, NULL 3660 // CastPP base_reg // NotNull 3661 // Load [base_reg + offset], val_reg 3662 // 3663 // after these transformations will be 3664 // 3665 // LoadN memory, narrow_oop_reg 3666 // CmpN narrow_oop_reg, NULL 3667 // decode_not_null narrow_oop_reg, base_reg 3668 // Load [base_reg + offset], val_reg 3669 // 3670 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3671 // since narrow oops can be used in debug info now (see the code in 3672 // final_graph_reshaping_walk()). 3673 // 3674 // At the end the code will be matched to 3675 // on x86: 3676 // 3677 // Load_narrow_oop memory, narrow_oop_reg 3678 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3679 // NullCheck narrow_oop_reg 3680 // 3681 // and on sparc: 3682 // 3683 // Load_narrow_oop memory, narrow_oop_reg 3684 // decode_not_null narrow_oop_reg, base_reg 3685 // Load [base_reg + offset], val_reg 3686 // NullCheck base_reg 3687 // 3688 } else if (t->isa_oopptr()) { 3689 new_in2 = ConNode::make(t->make_narrowoop()); 3690 } else if (t->isa_klassptr()) { 3691 new_in2 = ConNode::make(t->make_narrowklass()); 3692 } 3693 } 3694 if (new_in2 != NULL) { 3695 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3696 n->subsume_by(cmpN, this); 3697 if (in1->outcnt() == 0) { 3698 in1->disconnect_inputs(NULL, this); 3699 } 3700 if (in2->outcnt() == 0) { 3701 in2->disconnect_inputs(NULL, this); 3702 } 3703 } 3704 } 3705 break; 3706 3707 case Op_DecodeN: 3708 case Op_DecodeNKlass: 3709 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3710 // DecodeN could be pinned when it can't be fold into 3711 // an address expression, see the code for Op_CastPP above. 3712 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3713 break; 3714 3715 case Op_EncodeP: 3716 case Op_EncodePKlass: { 3717 Node* in1 = n->in(1); 3718 if (in1->is_DecodeNarrowPtr()) { 3719 n->subsume_by(in1->in(1), this); 3720 } else if (in1->Opcode() == Op_ConP) { 3721 const Type* t = in1->bottom_type(); 3722 if (t == TypePtr::NULL_PTR) { 3723 assert(t->isa_oopptr(), "null klass?"); 3724 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3725 } else if (t->isa_oopptr()) { 3726 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3727 } else if (t->isa_klassptr()) { 3728 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3729 } 3730 } 3731 if (in1->outcnt() == 0) { 3732 in1->disconnect_inputs(NULL, this); 3733 } 3734 break; 3735 } 3736 3737 case Op_Proj: { 3738 if (OptimizeStringConcat) { 3739 ProjNode* p = n->as_Proj(); 3740 if (p->_is_io_use) { 3741 // Separate projections were used for the exception path which 3742 // are normally removed by a late inline. If it wasn't inlined 3743 // then they will hang around and should just be replaced with 3744 // the original one. 3745 Node* proj = NULL; 3746 // Replace with just one 3747 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3748 Node *use = i.get(); 3749 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3750 proj = use; 3751 break; 3752 } 3753 } 3754 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3755 if (proj != NULL) { 3756 p->subsume_by(proj, this); 3757 } 3758 } 3759 } 3760 break; 3761 } 3762 3763 case Op_Phi: 3764 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3765 // The EncodeP optimization may create Phi with the same edges 3766 // for all paths. It is not handled well by Register Allocator. 3767 Node* unique_in = n->in(1); 3768 assert(unique_in != NULL, ""); 3769 uint cnt = n->req(); 3770 for (uint i = 2; i < cnt; i++) { 3771 Node* m = n->in(i); 3772 assert(m != NULL, ""); 3773 if (unique_in != m) 3774 unique_in = NULL; 3775 } 3776 if (unique_in != NULL) { 3777 n->subsume_by(unique_in, this); 3778 } 3779 } 3780 break; 3781 3782 #endif 3783 3784 #ifdef ASSERT 3785 case Op_CastII: 3786 // Verify that all range check dependent CastII nodes were removed. 3787 if (n->isa_CastII()->has_range_check()) { 3788 n->dump(3); 3789 assert(false, "Range check dependent CastII node was not removed"); 3790 } 3791 break; 3792 #endif 3793 3794 case Op_ModI: 3795 if (UseDivMod) { 3796 // Check if a%b and a/b both exist 3797 Node* d = n->find_similar(Op_DivI); 3798 if (d) { 3799 // Replace them with a fused divmod if supported 3800 if (Matcher::has_match_rule(Op_DivModI)) { 3801 DivModINode* divmod = DivModINode::make(n); 3802 d->subsume_by(divmod->div_proj(), this); 3803 n->subsume_by(divmod->mod_proj(), this); 3804 } else { 3805 // replace a%b with a-((a/b)*b) 3806 Node* mult = new MulINode(d, d->in(2)); 3807 Node* sub = new SubINode(d->in(1), mult); 3808 n->subsume_by(sub, this); 3809 } 3810 } 3811 } 3812 break; 3813 3814 case Op_ModL: 3815 if (UseDivMod) { 3816 // Check if a%b and a/b both exist 3817 Node* d = n->find_similar(Op_DivL); 3818 if (d) { 3819 // Replace them with a fused divmod if supported 3820 if (Matcher::has_match_rule(Op_DivModL)) { 3821 DivModLNode* divmod = DivModLNode::make(n); 3822 d->subsume_by(divmod->div_proj(), this); 3823 n->subsume_by(divmod->mod_proj(), this); 3824 } else { 3825 // replace a%b with a-((a/b)*b) 3826 Node* mult = new MulLNode(d, d->in(2)); 3827 Node* sub = new SubLNode(d->in(1), mult); 3828 n->subsume_by(sub, this); 3829 } 3830 } 3831 } 3832 break; 3833 3834 case Op_LoadVector: 3835 case Op_StoreVector: 3836 break; 3837 3838 case Op_AddReductionVI: 3839 case Op_AddReductionVL: 3840 case Op_AddReductionVF: 3841 case Op_AddReductionVD: 3842 case Op_MulReductionVI: 3843 case Op_MulReductionVL: 3844 case Op_MulReductionVF: 3845 case Op_MulReductionVD: 3846 case Op_MinReductionV: 3847 case Op_MaxReductionV: 3848 break; 3849 3850 case Op_PackB: 3851 case Op_PackS: 3852 case Op_PackI: 3853 case Op_PackF: 3854 case Op_PackL: 3855 case Op_PackD: 3856 if (n->req()-1 > 2) { 3857 // Replace many operand PackNodes with a binary tree for matching 3858 PackNode* p = (PackNode*) n; 3859 Node* btp = p->binary_tree_pack(1, n->req()); 3860 n->subsume_by(btp, this); 3861 } 3862 break; 3863 case Op_Loop: 3864 case Op_CountedLoop: 3865 case Op_OuterStripMinedLoop: 3866 if (n->as_Loop()->is_inner_loop()) { 3867 frc.inc_inner_loop_count(); 3868 } 3869 n->as_Loop()->verify_strip_mined(0); 3870 break; 3871 case Op_LShiftI: 3872 case Op_RShiftI: 3873 case Op_URShiftI: 3874 case Op_LShiftL: 3875 case Op_RShiftL: 3876 case Op_URShiftL: 3877 if (Matcher::need_masked_shift_count) { 3878 // The cpu's shift instructions don't restrict the count to the 3879 // lower 5/6 bits. We need to do the masking ourselves. 3880 Node* in2 = n->in(2); 3881 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3882 const TypeInt* t = in2->find_int_type(); 3883 if (t != NULL && t->is_con()) { 3884 juint shift = t->get_con(); 3885 if (shift > mask) { // Unsigned cmp 3886 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3887 } 3888 } else { 3889 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3890 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3891 n->set_req(2, shift); 3892 } 3893 } 3894 if (in2->outcnt() == 0) { // Remove dead node 3895 in2->disconnect_inputs(NULL, this); 3896 } 3897 } 3898 break; 3899 case Op_MemBarStoreStore: 3900 case Op_MemBarRelease: 3901 // Break the link with AllocateNode: it is no longer useful and 3902 // confuses register allocation. 3903 if (n->req() > MemBarNode::Precedent) { 3904 n->set_req(MemBarNode::Precedent, top()); 3905 } 3906 break; 3907 case Op_MemBarAcquire: { 3908 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3909 // At parse time, the trailing MemBarAcquire for a volatile load 3910 // is created with an edge to the load. After optimizations, 3911 // that input may be a chain of Phis. If those phis have no 3912 // other use, then the MemBarAcquire keeps them alive and 3913 // register allocation can be confused. 3914 ResourceMark rm; 3915 Unique_Node_List wq; 3916 wq.push(n->in(MemBarNode::Precedent)); 3917 n->set_req(MemBarNode::Precedent, top()); 3918 while (wq.size() > 0) { 3919 Node* m = wq.pop(); 3920 if (m->outcnt() == 0) { 3921 for (uint j = 0; j < m->req(); j++) { 3922 Node* in = m->in(j); 3923 if (in != NULL) { 3924 wq.push(in); 3925 } 3926 } 3927 m->disconnect_inputs(NULL, this); 3928 } 3929 } 3930 } 3931 break; 3932 } 3933 case Op_RangeCheck: { 3934 RangeCheckNode* rc = n->as_RangeCheck(); 3935 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3936 n->subsume_by(iff, this); 3937 frc._tests.push(iff); 3938 break; 3939 } 3940 case Op_ConvI2L: { 3941 if (!Matcher::convi2l_type_required) { 3942 // Code generation on some platforms doesn't need accurate 3943 // ConvI2L types. Widening the type can help remove redundant 3944 // address computations. 3945 n->as_Type()->set_type(TypeLong::INT); 3946 ResourceMark rm; 3947 Unique_Node_List wq; 3948 wq.push(n); 3949 for (uint next = 0; next < wq.size(); next++) { 3950 Node *m = wq.at(next); 3951 3952 for(;;) { 3953 // Loop over all nodes with identical inputs edges as m 3954 Node* k = m->find_similar(m->Opcode()); 3955 if (k == NULL) { 3956 break; 3957 } 3958 // Push their uses so we get a chance to remove node made 3959 // redundant 3960 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3961 Node* u = k->fast_out(i); 3962 if (u->Opcode() == Op_LShiftL || 3963 u->Opcode() == Op_AddL || 3964 u->Opcode() == Op_SubL || 3965 u->Opcode() == Op_AddP) { 3966 wq.push(u); 3967 } 3968 } 3969 // Replace all nodes with identical edges as m with m 3970 k->subsume_by(m, this); 3971 } 3972 } 3973 } 3974 break; 3975 } 3976 case Op_CmpUL: { 3977 if (!Matcher::has_match_rule(Op_CmpUL)) { 3978 // No support for unsigned long comparisons 3979 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3980 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3981 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3982 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3983 Node* andl = new AndLNode(orl, remove_sign_mask); 3984 Node* cmp = new CmpLNode(andl, n->in(2)); 3985 n->subsume_by(cmp, this); 3986 } 3987 break; 3988 } 3989 #ifdef ASSERT 3990 case Op_ValueTypePtr: 3991 case Op_ValueType: { 3992 n->dump(-1); 3993 assert(false, "value type node was not removed"); 3994 break; 3995 } 3996 #endif 3997 case Op_GetNullFreeProperty: 3998 case Op_GetFlattenedProperty: { 3999 // Extract the null free bits 4000 uint last = unique(); 4001 Node* null_free = NULL; 4002 int bit = nop == Op_GetNullFreeProperty ? ArrayStorageProperties::null_free_bit : ArrayStorageProperties::flattened_bit; 4003 if (n->in(1)->Opcode() == Op_LoadKlass) { 4004 Node* cast = new CastP2XNode(NULL, n->in(1)); 4005 null_free = new AndLNode(cast, new ConLNode(TypeLong::make(((jlong)1)<<(oopDesc::wide_storage_props_shift + bit)))); 4006 } else { 4007 assert(n->in(1)->Opcode() == Op_LoadNKlass, "not a compressed klass?"); 4008 Node* cast = new CastN2INode(n->in(1)); 4009 null_free = new AndINode(cast, new ConINode(TypeInt::make(1<<(oopDesc::narrow_storage_props_shift + bit)))); 4010 } 4011 n->subsume_by(null_free, this); 4012 break; 4013 } 4014 default: 4015 assert(!n->is_Call(), ""); 4016 assert(!n->is_Mem(), ""); 4017 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 4018 break; 4019 } 4020 } 4021 4022 //------------------------------final_graph_reshaping_walk--------------------- 4023 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 4024 // requires that the walk visits a node's inputs before visiting the node. 4025 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 4026 ResourceArea *area = Thread::current()->resource_area(); 4027 Unique_Node_List sfpt(area); 4028 4029 frc._visited.set(root->_idx); // first, mark node as visited 4030 uint cnt = root->req(); 4031 Node *n = root; 4032 uint i = 0; 4033 while (true) { 4034 if (i < cnt) { 4035 // Place all non-visited non-null inputs onto stack 4036 Node* m = n->in(i); 4037 ++i; 4038 if (m != NULL && !frc._visited.test_set(m->_idx)) { 4039 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 4040 // compute worst case interpreter size in case of a deoptimization 4041 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 4042 4043 sfpt.push(m); 4044 } 4045 cnt = m->req(); 4046 nstack.push(n, i); // put on stack parent and next input's index 4047 n = m; 4048 i = 0; 4049 } 4050 } else { 4051 // Now do post-visit work 4052 final_graph_reshaping_impl( n, frc ); 4053 if (nstack.is_empty()) 4054 break; // finished 4055 n = nstack.node(); // Get node from stack 4056 cnt = n->req(); 4057 i = nstack.index(); 4058 nstack.pop(); // Shift to the next node on stack 4059 } 4060 } 4061 4062 // Skip next transformation if compressed oops are not used. 4063 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 4064 (!UseCompressedOops && !UseCompressedClassPointers)) 4065 return; 4066 4067 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 4068 // It could be done for an uncommon traps or any safepoints/calls 4069 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 4070 while (sfpt.size() > 0) { 4071 n = sfpt.pop(); 4072 JVMState *jvms = n->as_SafePoint()->jvms(); 4073 assert(jvms != NULL, "sanity"); 4074 int start = jvms->debug_start(); 4075 int end = n->req(); 4076 bool is_uncommon = (n->is_CallStaticJava() && 4077 n->as_CallStaticJava()->uncommon_trap_request() != 0); 4078 for (int j = start; j < end; j++) { 4079 Node* in = n->in(j); 4080 if (in->is_DecodeNarrowPtr()) { 4081 bool safe_to_skip = true; 4082 if (!is_uncommon ) { 4083 // Is it safe to skip? 4084 for (uint i = 0; i < in->outcnt(); i++) { 4085 Node* u = in->raw_out(i); 4086 if (!u->is_SafePoint() || 4087 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 4088 safe_to_skip = false; 4089 } 4090 } 4091 } 4092 if (safe_to_skip) { 4093 n->set_req(j, in->in(1)); 4094 } 4095 if (in->outcnt() == 0) { 4096 in->disconnect_inputs(NULL, this); 4097 } 4098 } 4099 } 4100 } 4101 } 4102 4103 //------------------------------final_graph_reshaping-------------------------- 4104 // Final Graph Reshaping. 4105 // 4106 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 4107 // and not commoned up and forced early. Must come after regular 4108 // optimizations to avoid GVN undoing the cloning. Clone constant 4109 // inputs to Loop Phis; these will be split by the allocator anyways. 4110 // Remove Opaque nodes. 4111 // (2) Move last-uses by commutative operations to the left input to encourage 4112 // Intel update-in-place two-address operations and better register usage 4113 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 4114 // calls canonicalizing them back. 4115 // (3) Count the number of double-precision FP ops, single-precision FP ops 4116 // and call sites. On Intel, we can get correct rounding either by 4117 // forcing singles to memory (requires extra stores and loads after each 4118 // FP bytecode) or we can set a rounding mode bit (requires setting and 4119 // clearing the mode bit around call sites). The mode bit is only used 4120 // if the relative frequency of single FP ops to calls is low enough. 4121 // This is a key transform for SPEC mpeg_audio. 4122 // (4) Detect infinite loops; blobs of code reachable from above but not 4123 // below. Several of the Code_Gen algorithms fail on such code shapes, 4124 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 4125 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 4126 // Detection is by looking for IfNodes where only 1 projection is 4127 // reachable from below or CatchNodes missing some targets. 4128 // (5) Assert for insane oop offsets in debug mode. 4129 4130 bool Compile::final_graph_reshaping() { 4131 // an infinite loop may have been eliminated by the optimizer, 4132 // in which case the graph will be empty. 4133 if (root()->req() == 1) { 4134 record_method_not_compilable("trivial infinite loop"); 4135 return true; 4136 } 4137 4138 // Expensive nodes have their control input set to prevent the GVN 4139 // from freely commoning them. There's no GVN beyond this point so 4140 // no need to keep the control input. We want the expensive nodes to 4141 // be freely moved to the least frequent code path by gcm. 4142 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 4143 for (int i = 0; i < expensive_count(); i++) { 4144 _expensive_nodes->at(i)->set_req(0, NULL); 4145 } 4146 4147 Final_Reshape_Counts frc; 4148 4149 // Visit everybody reachable! 4150 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 4151 Node_Stack nstack(live_nodes() >> 1); 4152 final_graph_reshaping_walk(nstack, root(), frc); 4153 4154 // Check for unreachable (from below) code (i.e., infinite loops). 4155 for( uint i = 0; i < frc._tests.size(); i++ ) { 4156 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 4157 // Get number of CFG targets. 4158 // Note that PCTables include exception targets after calls. 4159 uint required_outcnt = n->required_outcnt(); 4160 if (n->outcnt() != required_outcnt) { 4161 // Check for a few special cases. Rethrow Nodes never take the 4162 // 'fall-thru' path, so expected kids is 1 less. 4163 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 4164 if (n->in(0)->in(0)->is_Call()) { 4165 CallNode *call = n->in(0)->in(0)->as_Call(); 4166 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 4167 required_outcnt--; // Rethrow always has 1 less kid 4168 } else if (call->req() > TypeFunc::Parms && 4169 call->is_CallDynamicJava()) { 4170 // Check for null receiver. In such case, the optimizer has 4171 // detected that the virtual call will always result in a null 4172 // pointer exception. The fall-through projection of this CatchNode 4173 // will not be populated. 4174 Node *arg0 = call->in(TypeFunc::Parms); 4175 if (arg0->is_Type() && 4176 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 4177 required_outcnt--; 4178 } 4179 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 4180 call->req() > TypeFunc::Parms+1 && 4181 call->is_CallStaticJava()) { 4182 // Check for negative array length. In such case, the optimizer has 4183 // detected that the allocation attempt will always result in an 4184 // exception. There is no fall-through projection of this CatchNode . 4185 Node *arg1 = call->in(TypeFunc::Parms+1); 4186 if (arg1->is_Type() && 4187 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 4188 required_outcnt--; 4189 } 4190 } 4191 } 4192 } 4193 // Recheck with a better notion of 'required_outcnt' 4194 if (n->outcnt() != required_outcnt) { 4195 record_method_not_compilable("malformed control flow"); 4196 return true; // Not all targets reachable! 4197 } 4198 } 4199 // Check that I actually visited all kids. Unreached kids 4200 // must be infinite loops. 4201 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 4202 if (!frc._visited.test(n->fast_out(j)->_idx)) { 4203 record_method_not_compilable("infinite loop"); 4204 return true; // Found unvisited kid; must be unreach 4205 } 4206 4207 // Here so verification code in final_graph_reshaping_walk() 4208 // always see an OuterStripMinedLoopEnd 4209 if (n->is_OuterStripMinedLoopEnd()) { 4210 IfNode* init_iff = n->as_If(); 4211 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 4212 n->subsume_by(iff, this); 4213 } 4214 } 4215 4216 // If original bytecodes contained a mixture of floats and doubles 4217 // check if the optimizer has made it homogenous, item (3). 4218 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 4219 frc.get_float_count() > 32 && 4220 frc.get_double_count() == 0 && 4221 (10 * frc.get_call_count() < frc.get_float_count()) ) { 4222 set_24_bit_selection_and_mode( false, true ); 4223 } 4224 4225 set_java_calls(frc.get_java_call_count()); 4226 set_inner_loops(frc.get_inner_loop_count()); 4227 4228 // No infinite loops, no reason to bail out. 4229 return false; 4230 } 4231 4232 //-----------------------------too_many_traps---------------------------------- 4233 // Report if there are too many traps at the current method and bci. 4234 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 4235 bool Compile::too_many_traps(ciMethod* method, 4236 int bci, 4237 Deoptimization::DeoptReason reason) { 4238 ciMethodData* md = method->method_data(); 4239 if (md->is_empty()) { 4240 // Assume the trap has not occurred, or that it occurred only 4241 // because of a transient condition during start-up in the interpreter. 4242 return false; 4243 } 4244 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 4245 if (md->has_trap_at(bci, m, reason) != 0) { 4246 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 4247 // Also, if there are multiple reasons, or if there is no per-BCI record, 4248 // assume the worst. 4249 if (log()) 4250 log()->elem("observe trap='%s' count='%d'", 4251 Deoptimization::trap_reason_name(reason), 4252 md->trap_count(reason)); 4253 return true; 4254 } else { 4255 // Ignore method/bci and see if there have been too many globally. 4256 return too_many_traps(reason, md); 4257 } 4258 } 4259 4260 // Less-accurate variant which does not require a method and bci. 4261 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 4262 ciMethodData* logmd) { 4263 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 4264 // Too many traps globally. 4265 // Note that we use cumulative trap_count, not just md->trap_count. 4266 if (log()) { 4267 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 4268 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 4269 Deoptimization::trap_reason_name(reason), 4270 mcount, trap_count(reason)); 4271 } 4272 return true; 4273 } else { 4274 // The coast is clear. 4275 return false; 4276 } 4277 } 4278 4279 //--------------------------too_many_recompiles-------------------------------- 4280 // Report if there are too many recompiles at the current method and bci. 4281 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 4282 // Is not eager to return true, since this will cause the compiler to use 4283 // Action_none for a trap point, to avoid too many recompilations. 4284 bool Compile::too_many_recompiles(ciMethod* method, 4285 int bci, 4286 Deoptimization::DeoptReason reason) { 4287 ciMethodData* md = method->method_data(); 4288 if (md->is_empty()) { 4289 // Assume the trap has not occurred, or that it occurred only 4290 // because of a transient condition during start-up in the interpreter. 4291 return false; 4292 } 4293 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 4294 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 4295 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 4296 Deoptimization::DeoptReason per_bc_reason 4297 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 4298 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 4299 if ((per_bc_reason == Deoptimization::Reason_none 4300 || md->has_trap_at(bci, m, reason) != 0) 4301 // The trap frequency measure we care about is the recompile count: 4302 && md->trap_recompiled_at(bci, m) 4303 && md->overflow_recompile_count() >= bc_cutoff) { 4304 // Do not emit a trap here if it has already caused recompilations. 4305 // Also, if there are multiple reasons, or if there is no per-BCI record, 4306 // assume the worst. 4307 if (log()) 4308 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 4309 Deoptimization::trap_reason_name(reason), 4310 md->trap_count(reason), 4311 md->overflow_recompile_count()); 4312 return true; 4313 } else if (trap_count(reason) != 0 4314 && decompile_count() >= m_cutoff) { 4315 // Too many recompiles globally, and we have seen this sort of trap. 4316 // Use cumulative decompile_count, not just md->decompile_count. 4317 if (log()) 4318 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 4319 Deoptimization::trap_reason_name(reason), 4320 md->trap_count(reason), trap_count(reason), 4321 md->decompile_count(), decompile_count()); 4322 return true; 4323 } else { 4324 // The coast is clear. 4325 return false; 4326 } 4327 } 4328 4329 // Compute when not to trap. Used by matching trap based nodes and 4330 // NullCheck optimization. 4331 void Compile::set_allowed_deopt_reasons() { 4332 _allowed_reasons = 0; 4333 if (is_method_compilation()) { 4334 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 4335 assert(rs < BitsPerInt, "recode bit map"); 4336 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 4337 _allowed_reasons |= nth_bit(rs); 4338 } 4339 } 4340 } 4341 } 4342 4343 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) { 4344 return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method); 4345 } 4346 4347 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) { 4348 return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method); 4349 } 4350 4351 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) { 4352 if (holder->is_initialized()) { 4353 return false; 4354 } 4355 if (holder->is_being_initialized()) { 4356 if (accessing_method->holder() == holder) { 4357 // Access inside a class. The barrier can be elided when access happens in <clinit>, 4358 // <init>, or a static method. In all those cases, there was an initialization 4359 // barrier on the holder klass passed. 4360 if (accessing_method->is_class_initializer() || 4361 accessing_method->is_object_constructor() || 4362 accessing_method->is_static()) { 4363 return false; 4364 } 4365 } else if (accessing_method->holder()->is_subclass_of(holder)) { 4366 // Access from a subclass. The barrier can be elided only when access happens in <clinit>. 4367 // In case of <init> or a static method, the barrier is on the subclass is not enough: 4368 // child class can become fully initialized while its parent class is still being initialized. 4369 if (accessing_method->is_class_initializer()) { 4370 return false; 4371 } 4372 } 4373 ciMethod* root = method(); // the root method of compilation 4374 if (root != accessing_method) { 4375 return needs_clinit_barrier(holder, root); // check access in the context of compilation root 4376 } 4377 } 4378 return true; 4379 } 4380 4381 #ifndef PRODUCT 4382 //------------------------------verify_graph_edges--------------------------- 4383 // Walk the Graph and verify that there is a one-to-one correspondence 4384 // between Use-Def edges and Def-Use edges in the graph. 4385 void Compile::verify_graph_edges(bool no_dead_code) { 4386 if (VerifyGraphEdges) { 4387 ResourceArea *area = Thread::current()->resource_area(); 4388 Unique_Node_List visited(area); 4389 // Call recursive graph walk to check edges 4390 _root->verify_edges(visited); 4391 if (no_dead_code) { 4392 // Now make sure that no visited node is used by an unvisited node. 4393 bool dead_nodes = false; 4394 Unique_Node_List checked(area); 4395 while (visited.size() > 0) { 4396 Node* n = visited.pop(); 4397 checked.push(n); 4398 for (uint i = 0; i < n->outcnt(); i++) { 4399 Node* use = n->raw_out(i); 4400 if (checked.member(use)) continue; // already checked 4401 if (visited.member(use)) continue; // already in the graph 4402 if (use->is_Con()) continue; // a dead ConNode is OK 4403 // At this point, we have found a dead node which is DU-reachable. 4404 if (!dead_nodes) { 4405 tty->print_cr("*** Dead nodes reachable via DU edges:"); 4406 dead_nodes = true; 4407 } 4408 use->dump(2); 4409 tty->print_cr("---"); 4410 checked.push(use); // No repeats; pretend it is now checked. 4411 } 4412 } 4413 assert(!dead_nodes, "using nodes must be reachable from root"); 4414 } 4415 } 4416 } 4417 #endif 4418 4419 // The Compile object keeps track of failure reasons separately from the ciEnv. 4420 // This is required because there is not quite a 1-1 relation between the 4421 // ciEnv and its compilation task and the Compile object. Note that one 4422 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4423 // to backtrack and retry without subsuming loads. Other than this backtracking 4424 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4425 // by the logic in C2Compiler. 4426 void Compile::record_failure(const char* reason) { 4427 if (log() != NULL) { 4428 log()->elem("failure reason='%s' phase='compile'", reason); 4429 } 4430 if (_failure_reason == NULL) { 4431 // Record the first failure reason. 4432 _failure_reason = reason; 4433 } 4434 4435 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4436 C->print_method(PHASE_FAILURE); 4437 } 4438 _root = NULL; // flush the graph, too 4439 } 4440 4441 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4442 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4443 _phase_name(name), _dolog(CITimeVerbose) 4444 { 4445 if (_dolog) { 4446 C = Compile::current(); 4447 _log = C->log(); 4448 } else { 4449 C = NULL; 4450 _log = NULL; 4451 } 4452 if (_log != NULL) { 4453 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4454 _log->stamp(); 4455 _log->end_head(); 4456 } 4457 } 4458 4459 Compile::TracePhase::~TracePhase() { 4460 4461 C = Compile::current(); 4462 if (_dolog) { 4463 _log = C->log(); 4464 } else { 4465 _log = NULL; 4466 } 4467 4468 #ifdef ASSERT 4469 if (PrintIdealNodeCount) { 4470 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4471 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4472 } 4473 4474 if (VerifyIdealNodeCount) { 4475 Compile::current()->print_missing_nodes(); 4476 } 4477 #endif 4478 4479 if (_log != NULL) { 4480 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4481 } 4482 } 4483 4484 //============================================================================= 4485 // Two Constant's are equal when the type and the value are equal. 4486 bool Compile::Constant::operator==(const Constant& other) { 4487 if (type() != other.type() ) return false; 4488 if (can_be_reused() != other.can_be_reused()) return false; 4489 // For floating point values we compare the bit pattern. 4490 switch (type()) { 4491 case T_INT: 4492 case T_FLOAT: return (_v._value.i == other._v._value.i); 4493 case T_LONG: 4494 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4495 case T_OBJECT: 4496 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4497 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4498 case T_METADATA: return (_v._metadata == other._v._metadata); 4499 default: ShouldNotReachHere(); return false; 4500 } 4501 } 4502 4503 static int type_to_size_in_bytes(BasicType t) { 4504 switch (t) { 4505 case T_INT: return sizeof(jint ); 4506 case T_LONG: return sizeof(jlong ); 4507 case T_FLOAT: return sizeof(jfloat ); 4508 case T_DOUBLE: return sizeof(jdouble); 4509 case T_METADATA: return sizeof(Metadata*); 4510 // We use T_VOID as marker for jump-table entries (labels) which 4511 // need an internal word relocation. 4512 case T_VOID: 4513 case T_ADDRESS: 4514 case T_OBJECT: return sizeof(jobject); 4515 default: 4516 ShouldNotReachHere(); 4517 return -1; 4518 } 4519 } 4520 4521 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4522 // sort descending 4523 if (a->freq() > b->freq()) return -1; 4524 if (a->freq() < b->freq()) return 1; 4525 return 0; 4526 } 4527 4528 void Compile::ConstantTable::calculate_offsets_and_size() { 4529 // First, sort the array by frequencies. 4530 _constants.sort(qsort_comparator); 4531 4532 #ifdef ASSERT 4533 // Make sure all jump-table entries were sorted to the end of the 4534 // array (they have a negative frequency). 4535 bool found_void = false; 4536 for (int i = 0; i < _constants.length(); i++) { 4537 Constant con = _constants.at(i); 4538 if (con.type() == T_VOID) 4539 found_void = true; // jump-tables 4540 else 4541 assert(!found_void, "wrong sorting"); 4542 } 4543 #endif 4544 4545 int offset = 0; 4546 for (int i = 0; i < _constants.length(); i++) { 4547 Constant* con = _constants.adr_at(i); 4548 4549 // Align offset for type. 4550 int typesize = type_to_size_in_bytes(con->type()); 4551 offset = align_up(offset, typesize); 4552 con->set_offset(offset); // set constant's offset 4553 4554 if (con->type() == T_VOID) { 4555 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4556 offset = offset + typesize * n->outcnt(); // expand jump-table 4557 } else { 4558 offset = offset + typesize; 4559 } 4560 } 4561 4562 // Align size up to the next section start (which is insts; see 4563 // CodeBuffer::align_at_start). 4564 assert(_size == -1, "already set?"); 4565 _size = align_up(offset, (int)CodeEntryAlignment); 4566 } 4567 4568 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4569 MacroAssembler _masm(&cb); 4570 for (int i = 0; i < _constants.length(); i++) { 4571 Constant con = _constants.at(i); 4572 address constant_addr = NULL; 4573 switch (con.type()) { 4574 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4575 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4576 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4577 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4578 case T_OBJECT: { 4579 jobject obj = con.get_jobject(); 4580 int oop_index = _masm.oop_recorder()->find_index(obj); 4581 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4582 break; 4583 } 4584 case T_ADDRESS: { 4585 address addr = (address) con.get_jobject(); 4586 constant_addr = _masm.address_constant(addr); 4587 break; 4588 } 4589 // We use T_VOID as marker for jump-table entries (labels) which 4590 // need an internal word relocation. 4591 case T_VOID: { 4592 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4593 // Fill the jump-table with a dummy word. The real value is 4594 // filled in later in fill_jump_table. 4595 address dummy = (address) n; 4596 constant_addr = _masm.address_constant(dummy); 4597 // Expand jump-table 4598 for (uint i = 1; i < n->outcnt(); i++) { 4599 address temp_addr = _masm.address_constant(dummy + i); 4600 assert(temp_addr, "consts section too small"); 4601 } 4602 break; 4603 } 4604 case T_METADATA: { 4605 Metadata* obj = con.get_metadata(); 4606 int metadata_index = _masm.oop_recorder()->find_index(obj); 4607 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4608 break; 4609 } 4610 default: ShouldNotReachHere(); 4611 } 4612 assert(constant_addr, "consts section too small"); 4613 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4614 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4615 } 4616 } 4617 4618 int Compile::ConstantTable::find_offset(Constant& con) const { 4619 int idx = _constants.find(con); 4620 guarantee(idx != -1, "constant must be in constant table"); 4621 int offset = _constants.at(idx).offset(); 4622 guarantee(offset != -1, "constant table not emitted yet?"); 4623 return offset; 4624 } 4625 4626 void Compile::ConstantTable::add(Constant& con) { 4627 if (con.can_be_reused()) { 4628 int idx = _constants.find(con); 4629 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4630 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4631 return; 4632 } 4633 } 4634 (void) _constants.append(con); 4635 } 4636 4637 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4638 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4639 Constant con(type, value, b->_freq); 4640 add(con); 4641 return con; 4642 } 4643 4644 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4645 Constant con(metadata); 4646 add(con); 4647 return con; 4648 } 4649 4650 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4651 jvalue value; 4652 BasicType type = oper->type()->basic_type(); 4653 switch (type) { 4654 case T_LONG: value.j = oper->constantL(); break; 4655 case T_FLOAT: value.f = oper->constantF(); break; 4656 case T_DOUBLE: value.d = oper->constantD(); break; 4657 case T_OBJECT: 4658 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4659 case T_METADATA: return add((Metadata*)oper->constant()); break; 4660 default: guarantee(false, "unhandled type: %s", type2name(type)); 4661 } 4662 return add(n, type, value); 4663 } 4664 4665 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4666 jvalue value; 4667 // We can use the node pointer here to identify the right jump-table 4668 // as this method is called from Compile::Fill_buffer right before 4669 // the MachNodes are emitted and the jump-table is filled (means the 4670 // MachNode pointers do not change anymore). 4671 value.l = (jobject) n; 4672 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4673 add(con); 4674 return con; 4675 } 4676 4677 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4678 // If called from Compile::scratch_emit_size do nothing. 4679 if (Compile::current()->in_scratch_emit_size()) return; 4680 4681 assert(labels.is_nonempty(), "must be"); 4682 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4683 4684 // Since MachConstantNode::constant_offset() also contains 4685 // table_base_offset() we need to subtract the table_base_offset() 4686 // to get the plain offset into the constant table. 4687 int offset = n->constant_offset() - table_base_offset(); 4688 4689 MacroAssembler _masm(&cb); 4690 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4691 4692 for (uint i = 0; i < n->outcnt(); i++) { 4693 address* constant_addr = &jump_table_base[i]; 4694 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)); 4695 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4696 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4697 } 4698 } 4699 4700 //----------------------------static_subtype_check----------------------------- 4701 // Shortcut important common cases when superklass is exact: 4702 // (0) superklass is java.lang.Object (can occur in reflective code) 4703 // (1) subklass is already limited to a subtype of superklass => always ok 4704 // (2) subklass does not overlap with superklass => always fail 4705 // (3) superklass has NO subtypes and we can check with a simple compare. 4706 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4707 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4708 return SSC_full_test; // Let caller generate the general case. 4709 } 4710 4711 if (superk == env()->Object_klass()) { 4712 return SSC_always_true; // (0) this test cannot fail 4713 } 4714 4715 ciType* superelem = superk; 4716 if (superelem->is_array_klass()) { 4717 ciArrayKlass* ak = superelem->as_array_klass(); 4718 superelem = superelem->as_array_klass()->base_element_type(); 4719 } 4720 4721 if (!subk->is_interface()) { // cannot trust static interface types yet 4722 if (subk->is_subtype_of(superk)) { 4723 return SSC_always_true; // (1) false path dead; no dynamic test needed 4724 } 4725 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4726 !superk->is_subtype_of(subk)) { 4727 return SSC_always_false; 4728 } 4729 } 4730 4731 // Do not fold the subtype check to an array klass pointer comparison for [V? arrays. 4732 // [V is a subtype of [V? but the klass for [V is not equal to the klass for [V?. Perform a full test. 4733 if (superk->is_obj_array_klass() && !superk->as_array_klass()->storage_properties().is_null_free() && superk->as_array_klass()->element_klass()->is_valuetype()) { 4734 return SSC_full_test; 4735 } 4736 // If casting to an instance klass, it must have no subtypes 4737 if (superk->is_interface()) { 4738 // Cannot trust interfaces yet. 4739 // %%% S.B. superk->nof_implementors() == 1 4740 } else if (superelem->is_instance_klass()) { 4741 ciInstanceKlass* ik = superelem->as_instance_klass(); 4742 if (!ik->has_subklass() && !ik->is_interface()) { 4743 if (!ik->is_final()) { 4744 // Add a dependency if there is a chance of a later subclass. 4745 dependencies()->assert_leaf_type(ik); 4746 } 4747 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4748 } 4749 } else { 4750 // A primitive array type has no subtypes. 4751 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4752 } 4753 4754 return SSC_full_test; 4755 } 4756 4757 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4758 #ifdef _LP64 4759 // The scaled index operand to AddP must be a clean 64-bit value. 4760 // Java allows a 32-bit int to be incremented to a negative 4761 // value, which appears in a 64-bit register as a large 4762 // positive number. Using that large positive number as an 4763 // operand in pointer arithmetic has bad consequences. 4764 // On the other hand, 32-bit overflow is rare, and the possibility 4765 // can often be excluded, if we annotate the ConvI2L node with 4766 // a type assertion that its value is known to be a small positive 4767 // number. (The prior range check has ensured this.) 4768 // This assertion is used by ConvI2LNode::Ideal. 4769 int index_max = max_jint - 1; // array size is max_jint, index is one less 4770 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4771 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4772 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4773 #endif 4774 return idx; 4775 } 4776 4777 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4778 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4779 if (ctrl != NULL) { 4780 // Express control dependency by a CastII node with a narrow type. 4781 value = new CastIINode(value, itype, false, true /* range check dependency */); 4782 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4783 // node from floating above the range check during loop optimizations. Otherwise, the 4784 // ConvI2L node may be eliminated independently of the range check, causing the data path 4785 // to become TOP while the control path is still there (although it's unreachable). 4786 value->set_req(0, ctrl); 4787 // Save CastII node to remove it after loop optimizations. 4788 phase->C->add_range_check_cast(value); 4789 value = phase->transform(value); 4790 } 4791 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4792 return phase->transform(new ConvI2LNode(value, ltype)); 4793 } 4794 4795 // The message about the current inlining is accumulated in 4796 // _print_inlining_stream and transfered into the _print_inlining_list 4797 // once we know whether inlining succeeds or not. For regular 4798 // inlining, messages are appended to the buffer pointed by 4799 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4800 // a new buffer is added after _print_inlining_idx in the list. This 4801 // way we can update the inlining message for late inlining call site 4802 // when the inlining is attempted again. 4803 void Compile::print_inlining_init() { 4804 if (print_inlining() || print_intrinsics()) { 4805 _print_inlining_stream = new stringStream(); 4806 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4807 } 4808 } 4809 4810 void Compile::print_inlining_reinit() { 4811 if (print_inlining() || print_intrinsics()) { 4812 // Re allocate buffer when we change ResourceMark 4813 _print_inlining_stream = new stringStream(); 4814 } 4815 } 4816 4817 void Compile::print_inlining_reset() { 4818 _print_inlining_stream->reset(); 4819 } 4820 4821 void Compile::print_inlining_commit() { 4822 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4823 // Transfer the message from _print_inlining_stream to the current 4824 // _print_inlining_list buffer and clear _print_inlining_stream. 4825 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4826 print_inlining_reset(); 4827 } 4828 4829 void Compile::print_inlining_push() { 4830 // Add new buffer to the _print_inlining_list at current position 4831 _print_inlining_idx++; 4832 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4833 } 4834 4835 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4836 return _print_inlining_list->at(_print_inlining_idx); 4837 } 4838 4839 void Compile::print_inlining_update(CallGenerator* cg) { 4840 if (print_inlining() || print_intrinsics()) { 4841 if (!cg->is_late_inline()) { 4842 if (print_inlining_current().cg() != NULL) { 4843 print_inlining_push(); 4844 } 4845 print_inlining_commit(); 4846 } else { 4847 if (print_inlining_current().cg() != cg && 4848 (print_inlining_current().cg() != NULL || 4849 print_inlining_current().ss()->size() != 0)) { 4850 print_inlining_push(); 4851 } 4852 print_inlining_commit(); 4853 print_inlining_current().set_cg(cg); 4854 } 4855 } 4856 } 4857 4858 void Compile::print_inlining_move_to(CallGenerator* cg) { 4859 // We resume inlining at a late inlining call site. Locate the 4860 // corresponding inlining buffer so that we can update it. 4861 if (print_inlining()) { 4862 for (int i = 0; i < _print_inlining_list->length(); i++) { 4863 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4864 _print_inlining_idx = i; 4865 return; 4866 } 4867 } 4868 ShouldNotReachHere(); 4869 } 4870 } 4871 4872 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4873 if (print_inlining()) { 4874 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4875 assert(print_inlining_current().cg() == cg, "wrong entry"); 4876 // replace message with new message 4877 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4878 print_inlining_commit(); 4879 print_inlining_current().set_cg(cg); 4880 } 4881 } 4882 4883 void Compile::print_inlining_assert_ready() { 4884 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4885 } 4886 4887 void Compile::process_print_inlining() { 4888 bool do_print_inlining = print_inlining() || print_intrinsics(); 4889 if (do_print_inlining || log() != NULL) { 4890 // Print inlining message for candidates that we couldn't inline 4891 // for lack of space 4892 for (int i = 0; i < _late_inlines.length(); i++) { 4893 CallGenerator* cg = _late_inlines.at(i); 4894 if (!cg->is_mh_late_inline()) { 4895 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4896 if (do_print_inlining) { 4897 cg->print_inlining_late(msg); 4898 } 4899 log_late_inline_failure(cg, msg); 4900 } 4901 } 4902 } 4903 if (do_print_inlining) { 4904 ResourceMark rm; 4905 stringStream ss; 4906 for (int i = 0; i < _print_inlining_list->length(); i++) { 4907 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4908 } 4909 size_t end = ss.size(); 4910 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4911 strncpy(_print_inlining_output, ss.base(), end+1); 4912 _print_inlining_output[end] = 0; 4913 } 4914 } 4915 4916 void Compile::dump_print_inlining() { 4917 if (_print_inlining_output != NULL) { 4918 tty->print_raw(_print_inlining_output); 4919 } 4920 } 4921 4922 void Compile::log_late_inline(CallGenerator* cg) { 4923 if (log() != NULL) { 4924 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4925 cg->unique_id()); 4926 JVMState* p = cg->call_node()->jvms(); 4927 while (p != NULL) { 4928 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4929 p = p->caller(); 4930 } 4931 log()->tail("late_inline"); 4932 } 4933 } 4934 4935 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4936 log_late_inline(cg); 4937 if (log() != NULL) { 4938 log()->inline_fail(msg); 4939 } 4940 } 4941 4942 void Compile::log_inline_id(CallGenerator* cg) { 4943 if (log() != NULL) { 4944 // The LogCompilation tool needs a unique way to identify late 4945 // inline call sites. This id must be unique for this call site in 4946 // this compilation. Try to have it unique across compilations as 4947 // well because it can be convenient when grepping through the log 4948 // file. 4949 // Distinguish OSR compilations from others in case CICountOSR is 4950 // on. 4951 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4952 cg->set_unique_id(id); 4953 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4954 } 4955 } 4956 4957 void Compile::log_inline_failure(const char* msg) { 4958 if (C->log() != NULL) { 4959 C->log()->inline_fail(msg); 4960 } 4961 } 4962 4963 4964 // Dump inlining replay data to the stream. 4965 // Don't change thread state and acquire any locks. 4966 void Compile::dump_inline_data(outputStream* out) { 4967 InlineTree* inl_tree = ilt(); 4968 if (inl_tree != NULL) { 4969 out->print(" inline %d", inl_tree->count()); 4970 inl_tree->dump_replay_data(out); 4971 } 4972 } 4973 4974 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4975 if (n1->Opcode() < n2->Opcode()) return -1; 4976 else if (n1->Opcode() > n2->Opcode()) return 1; 4977 4978 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4979 for (uint i = 1; i < n1->req(); i++) { 4980 if (n1->in(i) < n2->in(i)) return -1; 4981 else if (n1->in(i) > n2->in(i)) return 1; 4982 } 4983 4984 return 0; 4985 } 4986 4987 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4988 Node* n1 = *n1p; 4989 Node* n2 = *n2p; 4990 4991 return cmp_expensive_nodes(n1, n2); 4992 } 4993 4994 void Compile::sort_expensive_nodes() { 4995 if (!expensive_nodes_sorted()) { 4996 _expensive_nodes->sort(cmp_expensive_nodes); 4997 } 4998 } 4999 5000 bool Compile::expensive_nodes_sorted() const { 5001 for (int i = 1; i < _expensive_nodes->length(); i++) { 5002 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 5003 return false; 5004 } 5005 } 5006 return true; 5007 } 5008 5009 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 5010 if (_expensive_nodes->length() == 0) { 5011 return false; 5012 } 5013 5014 assert(OptimizeExpensiveOps, "optimization off?"); 5015 5016 // Take this opportunity to remove dead nodes from the list 5017 int j = 0; 5018 for (int i = 0; i < _expensive_nodes->length(); i++) { 5019 Node* n = _expensive_nodes->at(i); 5020 if (!n->is_unreachable(igvn)) { 5021 assert(n->is_expensive(), "should be expensive"); 5022 _expensive_nodes->at_put(j, n); 5023 j++; 5024 } 5025 } 5026 _expensive_nodes->trunc_to(j); 5027 5028 // Then sort the list so that similar nodes are next to each other 5029 // and check for at least two nodes of identical kind with same data 5030 // inputs. 5031 sort_expensive_nodes(); 5032 5033 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 5034 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 5035 return true; 5036 } 5037 } 5038 5039 return false; 5040 } 5041 5042 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 5043 if (_expensive_nodes->length() == 0) { 5044 return; 5045 } 5046 5047 assert(OptimizeExpensiveOps, "optimization off?"); 5048 5049 // Sort to bring similar nodes next to each other and clear the 5050 // control input of nodes for which there's only a single copy. 5051 sort_expensive_nodes(); 5052 5053 int j = 0; 5054 int identical = 0; 5055 int i = 0; 5056 bool modified = false; 5057 for (; i < _expensive_nodes->length()-1; i++) { 5058 assert(j <= i, "can't write beyond current index"); 5059 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 5060 identical++; 5061 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 5062 continue; 5063 } 5064 if (identical > 0) { 5065 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 5066 identical = 0; 5067 } else { 5068 Node* n = _expensive_nodes->at(i); 5069 igvn.replace_input_of(n, 0, NULL); 5070 igvn.hash_insert(n); 5071 modified = true; 5072 } 5073 } 5074 if (identical > 0) { 5075 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 5076 } else if (_expensive_nodes->length() >= 1) { 5077 Node* n = _expensive_nodes->at(i); 5078 igvn.replace_input_of(n, 0, NULL); 5079 igvn.hash_insert(n); 5080 modified = true; 5081 } 5082 _expensive_nodes->trunc_to(j); 5083 if (modified) { 5084 igvn.optimize(); 5085 } 5086 } 5087 5088 void Compile::add_expensive_node(Node * n) { 5089 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 5090 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 5091 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 5092 if (OptimizeExpensiveOps) { 5093 _expensive_nodes->append(n); 5094 } else { 5095 // Clear control input and let IGVN optimize expensive nodes if 5096 // OptimizeExpensiveOps is off. 5097 n->set_req(0, NULL); 5098 } 5099 } 5100 5101 /** 5102 * Remove the speculative part of types and clean up the graph 5103 */ 5104 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 5105 if (UseTypeSpeculation) { 5106 Unique_Node_List worklist; 5107 worklist.push(root()); 5108 int modified = 0; 5109 // Go over all type nodes that carry a speculative type, drop the 5110 // speculative part of the type and enqueue the node for an igvn 5111 // which may optimize it out. 5112 for (uint next = 0; next < worklist.size(); ++next) { 5113 Node *n = worklist.at(next); 5114 if (n->is_Type()) { 5115 TypeNode* tn = n->as_Type(); 5116 const Type* t = tn->type(); 5117 const Type* t_no_spec = t->remove_speculative(); 5118 if (t_no_spec != t) { 5119 bool in_hash = igvn.hash_delete(n); 5120 assert(in_hash, "node should be in igvn hash table"); 5121 tn->set_type(t_no_spec); 5122 igvn.hash_insert(n); 5123 igvn._worklist.push(n); // give it a chance to go away 5124 modified++; 5125 } 5126 } 5127 uint max = n->len(); 5128 for( uint i = 0; i < max; ++i ) { 5129 Node *m = n->in(i); 5130 if (not_a_node(m)) continue; 5131 worklist.push(m); 5132 } 5133 } 5134 // Drop the speculative part of all types in the igvn's type table 5135 igvn.remove_speculative_types(); 5136 if (modified > 0) { 5137 igvn.optimize(); 5138 } 5139 #ifdef ASSERT 5140 // Verify that after the IGVN is over no speculative type has resurfaced 5141 worklist.clear(); 5142 worklist.push(root()); 5143 for (uint next = 0; next < worklist.size(); ++next) { 5144 Node *n = worklist.at(next); 5145 const Type* t = igvn.type_or_null(n); 5146 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 5147 if (n->is_Type()) { 5148 t = n->as_Type()->type(); 5149 assert(t == t->remove_speculative(), "no more speculative types"); 5150 } 5151 uint max = n->len(); 5152 for( uint i = 0; i < max; ++i ) { 5153 Node *m = n->in(i); 5154 if (not_a_node(m)) continue; 5155 worklist.push(m); 5156 } 5157 } 5158 igvn.check_no_speculative_types(); 5159 #endif 5160 } 5161 } 5162 5163 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 5164 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 5165 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 5166 if (!EnableValhalla || ta == NULL || tb == NULL || 5167 ta->is_zero_type() || tb->is_zero_type() || 5168 !ta->can_be_value_type() || !tb->can_be_value_type()) { 5169 // Use old acmp if one operand is null or not a value type 5170 return new CmpPNode(a, b); 5171 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) { 5172 // We know that one operand is a value type. Therefore, 5173 // new acmp will only return true if both operands are NULL. 5174 // Check if both operands are null by or'ing the oops. 5175 a = phase->transform(new CastP2XNode(NULL, a)); 5176 b = phase->transform(new CastP2XNode(NULL, b)); 5177 a = phase->transform(new OrXNode(a, b)); 5178 return new CmpXNode(a, phase->MakeConX(0)); 5179 } 5180 // Use new acmp 5181 return NULL; 5182 } 5183 5184 // Auxiliary method to support randomized stressing/fuzzing. 5185 // 5186 // This method can be called the arbitrary number of times, with current count 5187 // as the argument. The logic allows selecting a single candidate from the 5188 // running list of candidates as follows: 5189 // int count = 0; 5190 // Cand* selected = null; 5191 // while(cand = cand->next()) { 5192 // if (randomized_select(++count)) { 5193 // selected = cand; 5194 // } 5195 // } 5196 // 5197 // Including count equalizes the chances any candidate is "selected". 5198 // This is useful when we don't have the complete list of candidates to choose 5199 // from uniformly. In this case, we need to adjust the randomicity of the 5200 // selection, or else we will end up biasing the selection towards the latter 5201 // candidates. 5202 // 5203 // Quick back-envelope calculation shows that for the list of n candidates 5204 // the equal probability for the candidate to persist as "best" can be 5205 // achieved by replacing it with "next" k-th candidate with the probability 5206 // of 1/k. It can be easily shown that by the end of the run, the 5207 // probability for any candidate is converged to 1/n, thus giving the 5208 // uniform distribution among all the candidates. 5209 // 5210 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 5211 #define RANDOMIZED_DOMAIN_POW 29 5212 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 5213 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 5214 bool Compile::randomized_select(int count) { 5215 assert(count > 0, "only positive"); 5216 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 5217 } 5218 5219 CloneMap& Compile::clone_map() { return _clone_map; } 5220 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 5221 5222 void NodeCloneInfo::dump() const { 5223 tty->print(" {%d:%d} ", idx(), gen()); 5224 } 5225 5226 void CloneMap::clone(Node* old, Node* nnn, int gen) { 5227 uint64_t val = value(old->_idx); 5228 NodeCloneInfo cio(val); 5229 assert(val != 0, "old node should be in the map"); 5230 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 5231 insert(nnn->_idx, cin.get()); 5232 #ifndef PRODUCT 5233 if (is_debug()) { 5234 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 5235 } 5236 #endif 5237 } 5238 5239 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 5240 NodeCloneInfo cio(value(old->_idx)); 5241 if (cio.get() == 0) { 5242 cio.set(old->_idx, 0); 5243 insert(old->_idx, cio.get()); 5244 #ifndef PRODUCT 5245 if (is_debug()) { 5246 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 5247 } 5248 #endif 5249 } 5250 clone(old, nnn, gen); 5251 } 5252 5253 int CloneMap::max_gen() const { 5254 int g = 0; 5255 DictI di(_dict); 5256 for(; di.test(); ++di) { 5257 int t = gen(di._key); 5258 if (g < t) { 5259 g = t; 5260 #ifndef PRODUCT 5261 if (is_debug()) { 5262 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 5263 } 5264 #endif 5265 } 5266 } 5267 return g; 5268 } 5269 5270 void CloneMap::dump(node_idx_t key) const { 5271 uint64_t val = value(key); 5272 if (val != 0) { 5273 NodeCloneInfo ni(val); 5274 ni.dump(); 5275 } 5276 }