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