1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "opto/addnode.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/callGenerator.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/castnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/chaitin.hpp" 47 #include "opto/compile.hpp" 48 #include "opto/connode.hpp" 49 #include "opto/convertnode.hpp" 50 #include "opto/divnode.hpp" 51 #include "opto/escape.hpp" 52 #include "opto/idealGraphPrinter.hpp" 53 #include "opto/loopnode.hpp" 54 #include "opto/machnode.hpp" 55 #include "opto/macro.hpp" 56 #include "opto/matcher.hpp" 57 #include "opto/mathexactnode.hpp" 58 #include "opto/memnode.hpp" 59 #include "opto/mulnode.hpp" 60 #include "opto/narrowptrnode.hpp" 61 #include "opto/node.hpp" 62 #include "opto/opcodes.hpp" 63 #include "opto/output.hpp" 64 #include "opto/parse.hpp" 65 #include "opto/phaseX.hpp" 66 #include "opto/rootnode.hpp" 67 #include "opto/runtime.hpp" 68 #include "opto/stringopts.hpp" 69 #include "opto/type.hpp" 70 #include "opto/vectornode.hpp" 71 #include "runtime/arguments.hpp" 72 #include "runtime/sharedRuntime.hpp" 73 #include "runtime/signature.hpp" 74 #include "runtime/stubRoutines.hpp" 75 #include "runtime/timer.hpp" 76 #include "utilities/align.hpp" 77 #include "utilities/copy.hpp" 78 #include "utilities/macros.hpp" 79 #if INCLUDE_G1GC 80 #include "gc/g1/g1ThreadLocalData.hpp" 81 #endif // INCLUDE_G1GC 82 #if INCLUDE_ZGC 83 #include "gc/z/c2/zBarrierSetC2.hpp" 84 #endif 85 86 87 // -------------------- Compile::mach_constant_base_node ----------------------- 88 // Constant table base node singleton. 89 MachConstantBaseNode* Compile::mach_constant_base_node() { 90 if (_mach_constant_base_node == NULL) { 91 _mach_constant_base_node = new MachConstantBaseNode(); 92 _mach_constant_base_node->add_req(C->root()); 93 } 94 return _mach_constant_base_node; 95 } 96 97 98 /// Support for intrinsics. 99 100 // Return the index at which m must be inserted (or already exists). 101 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 102 class IntrinsicDescPair { 103 private: 104 ciMethod* _m; 105 bool _is_virtual; 106 public: 107 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 108 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 109 ciMethod* m= elt->method(); 110 ciMethod* key_m = key->_m; 111 if (key_m < m) return -1; 112 else if (key_m > m) return 1; 113 else { 114 bool is_virtual = elt->is_virtual(); 115 bool key_virtual = key->_is_virtual; 116 if (key_virtual < is_virtual) return -1; 117 else if (key_virtual > is_virtual) return 1; 118 else return 0; 119 } 120 } 121 }; 122 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 123 #ifdef ASSERT 124 for (int i = 1; i < _intrinsics->length(); i++) { 125 CallGenerator* cg1 = _intrinsics->at(i-1); 126 CallGenerator* cg2 = _intrinsics->at(i); 127 assert(cg1->method() != cg2->method() 128 ? cg1->method() < cg2->method() 129 : cg1->is_virtual() < cg2->is_virtual(), 130 "compiler intrinsics list must stay sorted"); 131 } 132 #endif 133 IntrinsicDescPair pair(m, is_virtual); 134 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 135 } 136 137 void Compile::register_intrinsic(CallGenerator* cg) { 138 if (_intrinsics == NULL) { 139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 140 } 141 int len = _intrinsics->length(); 142 bool found = false; 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 144 assert(!found, "registering twice"); 145 _intrinsics->insert_before(index, cg); 146 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 147 } 148 149 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 150 assert(m->is_loaded(), "don't try this on unloaded methods"); 151 if (_intrinsics != NULL) { 152 bool found = false; 153 int index = intrinsic_insertion_index(m, is_virtual, found); 154 if (found) { 155 return _intrinsics->at(index); 156 } 157 } 158 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 159 if (m->intrinsic_id() != vmIntrinsics::_none && 160 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 161 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 162 if (cg != NULL) { 163 // Save it for next time: 164 register_intrinsic(cg); 165 return cg; 166 } else { 167 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 168 } 169 } 170 return NULL; 171 } 172 173 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 174 // in library_call.cpp. 175 176 177 #ifndef PRODUCT 178 // statistics gathering... 179 180 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 181 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 182 183 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 184 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 185 int oflags = _intrinsic_hist_flags[id]; 186 assert(flags != 0, "what happened?"); 187 if (is_virtual) { 188 flags |= _intrinsic_virtual; 189 } 190 bool changed = (flags != oflags); 191 if ((flags & _intrinsic_worked) != 0) { 192 juint count = (_intrinsic_hist_count[id] += 1); 193 if (count == 1) { 194 changed = true; // first time 195 } 196 // increment the overall count also: 197 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 198 } 199 if (changed) { 200 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 201 // Something changed about the intrinsic's virtuality. 202 if ((flags & _intrinsic_virtual) != 0) { 203 // This is the first use of this intrinsic as a virtual call. 204 if (oflags != 0) { 205 // We already saw it as a non-virtual, so note both cases. 206 flags |= _intrinsic_both; 207 } 208 } else if ((oflags & _intrinsic_both) == 0) { 209 // This is the first use of this intrinsic as a non-virtual 210 flags |= _intrinsic_both; 211 } 212 } 213 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 214 } 215 // update the overall flags also: 216 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 217 return changed; 218 } 219 220 static char* format_flags(int flags, char* buf) { 221 buf[0] = 0; 222 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 223 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 224 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 225 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 226 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 227 if (buf[0] == 0) strcat(buf, ","); 228 assert(buf[0] == ',', "must be"); 229 return &buf[1]; 230 } 231 232 void Compile::print_intrinsic_statistics() { 233 char flagsbuf[100]; 234 ttyLocker ttyl; 235 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 236 tty->print_cr("Compiler intrinsic usage:"); 237 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 238 if (total == 0) total = 1; // avoid div0 in case of no successes 239 #define PRINT_STAT_LINE(name, c, f) \ 240 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 241 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 242 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 243 int flags = _intrinsic_hist_flags[id]; 244 juint count = _intrinsic_hist_count[id]; 245 if ((flags | count) != 0) { 246 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 247 } 248 } 249 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 250 if (xtty != NULL) xtty->tail("statistics"); 251 } 252 253 void Compile::print_statistics() { 254 { ttyLocker ttyl; 255 if (xtty != NULL) xtty->head("statistics type='opto'"); 256 Parse::print_statistics(); 257 PhaseCCP::print_statistics(); 258 PhaseRegAlloc::print_statistics(); 259 Scheduling::print_statistics(); 260 PhasePeephole::print_statistics(); 261 PhaseIdealLoop::print_statistics(); 262 if (xtty != NULL) xtty->tail("statistics"); 263 } 264 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 265 // put this under its own <statistics> element. 266 print_intrinsic_statistics(); 267 } 268 } 269 #endif //PRODUCT 270 271 // Support for bundling info 272 Bundle* Compile::node_bundling(const Node *n) { 273 assert(valid_bundle_info(n), "oob"); 274 return &_node_bundling_base[n->_idx]; 275 } 276 277 bool Compile::valid_bundle_info(const Node *n) { 278 return (_node_bundling_limit > n->_idx); 279 } 280 281 282 void Compile::gvn_replace_by(Node* n, Node* nn) { 283 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 284 Node* use = n->last_out(i); 285 bool is_in_table = initial_gvn()->hash_delete(use); 286 uint uses_found = 0; 287 for (uint j = 0; j < use->len(); j++) { 288 if (use->in(j) == n) { 289 if (j < use->req()) 290 use->set_req(j, nn); 291 else 292 use->set_prec(j, nn); 293 uses_found++; 294 } 295 } 296 if (is_in_table) { 297 // reinsert into table 298 initial_gvn()->hash_find_insert(use); 299 } 300 record_for_igvn(use); 301 i -= uses_found; // we deleted 1 or more copies of this edge 302 } 303 } 304 305 306 static inline bool not_a_node(const Node* n) { 307 if (n == NULL) return true; 308 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 309 if (*(address*)n == badAddress) return true; // kill by Node::destruct 310 return false; 311 } 312 313 // Identify all nodes that are reachable from below, useful. 314 // Use breadth-first pass that records state in a Unique_Node_List, 315 // recursive traversal is slower. 316 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 317 int estimated_worklist_size = live_nodes(); 318 useful.map( estimated_worklist_size, NULL ); // preallocate space 319 320 // Initialize worklist 321 if (root() != NULL) { useful.push(root()); } 322 // If 'top' is cached, declare it useful to preserve cached node 323 if( cached_top_node() ) { useful.push(cached_top_node()); } 324 325 // Push all useful nodes onto the list, breadthfirst 326 for( uint next = 0; next < useful.size(); ++next ) { 327 assert( next < unique(), "Unique useful nodes < total nodes"); 328 Node *n = useful.at(next); 329 uint max = n->len(); 330 for( uint i = 0; i < max; ++i ) { 331 Node *m = n->in(i); 332 if (not_a_node(m)) continue; 333 useful.push(m); 334 } 335 } 336 } 337 338 // Update dead_node_list with any missing dead nodes using useful 339 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 340 void Compile::update_dead_node_list(Unique_Node_List &useful) { 341 uint max_idx = unique(); 342 VectorSet& useful_node_set = useful.member_set(); 343 344 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 345 // If node with index node_idx is not in useful set, 346 // mark it as dead in dead node list. 347 if (! useful_node_set.test(node_idx) ) { 348 record_dead_node(node_idx); 349 } 350 } 351 } 352 353 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 354 int shift = 0; 355 for (int i = 0; i < inlines->length(); i++) { 356 CallGenerator* cg = inlines->at(i); 357 CallNode* call = cg->call_node(); 358 if (shift > 0) { 359 inlines->at_put(i-shift, cg); 360 } 361 if (!useful.member(call)) { 362 shift++; 363 } 364 } 365 inlines->trunc_to(inlines->length()-shift); 366 } 367 368 // Disconnect all useless nodes by disconnecting those at the boundary. 369 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 370 uint next = 0; 371 while (next < useful.size()) { 372 Node *n = useful.at(next++); 373 if (n->is_SafePoint()) { 374 // We're done with a parsing phase. Replaced nodes are not valid 375 // beyond that point. 376 n->as_SafePoint()->delete_replaced_nodes(); 377 } 378 // Use raw traversal of out edges since this code removes out edges 379 int max = n->outcnt(); 380 for (int j = 0; j < max; ++j) { 381 Node* child = n->raw_out(j); 382 if (! useful.member(child)) { 383 assert(!child->is_top() || child != top(), 384 "If top is cached in Compile object it is in useful list"); 385 // Only need to remove this out-edge to the useless node 386 n->raw_del_out(j); 387 --j; 388 --max; 389 } 390 } 391 if (n->outcnt() == 1 && n->has_special_unique_user()) { 392 record_for_igvn(n->unique_out()); 393 } 394 } 395 // Remove useless macro and predicate opaq nodes 396 for (int i = C->macro_count()-1; i >= 0; i--) { 397 Node* n = C->macro_node(i); 398 if (!useful.member(n)) { 399 remove_macro_node(n); 400 } 401 } 402 // Remove useless CastII nodes with range check dependency 403 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 404 Node* cast = range_check_cast_node(i); 405 if (!useful.member(cast)) { 406 remove_range_check_cast(cast); 407 } 408 } 409 // Remove useless expensive nodes 410 for (int i = C->expensive_count()-1; i >= 0; i--) { 411 Node* n = C->expensive_node(i); 412 if (!useful.member(n)) { 413 remove_expensive_node(n); 414 } 415 } 416 // Remove useless Opaque4 nodes 417 for (int i = opaque4_count() - 1; i >= 0; i--) { 418 Node* opaq = opaque4_node(i); 419 if (!useful.member(opaq)) { 420 remove_opaque4_node(opaq); 421 } 422 } 423 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 424 bs->eliminate_useless_gc_barriers(useful, this); 425 // clean up the late inline lists 426 remove_useless_late_inlines(&_string_late_inlines, useful); 427 remove_useless_late_inlines(&_boxing_late_inlines, useful); 428 remove_useless_late_inlines(&_late_inlines, useful); 429 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 430 } 431 432 //------------------------------frame_size_in_words----------------------------- 433 // frame_slots in units of words 434 int Compile::frame_size_in_words() const { 435 // shift is 0 in LP32 and 1 in LP64 436 const int shift = (LogBytesPerWord - LogBytesPerInt); 437 int words = _frame_slots >> shift; 438 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 439 return words; 440 } 441 442 // To bang the stack of this compiled method we use the stack size 443 // that the interpreter would need in case of a deoptimization. This 444 // removes the need to bang the stack in the deoptimization blob which 445 // in turn simplifies stack overflow handling. 446 int Compile::bang_size_in_bytes() const { 447 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 448 } 449 450 // ============================================================================ 451 //------------------------------CompileWrapper--------------------------------- 452 class CompileWrapper : public StackObj { 453 Compile *const _compile; 454 public: 455 CompileWrapper(Compile* compile); 456 457 ~CompileWrapper(); 458 }; 459 460 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 461 // the Compile* pointer is stored in the current ciEnv: 462 ciEnv* env = compile->env(); 463 assert(env == ciEnv::current(), "must already be a ciEnv active"); 464 assert(env->compiler_data() == NULL, "compile already active?"); 465 env->set_compiler_data(compile); 466 assert(compile == Compile::current(), "sanity"); 467 468 compile->set_type_dict(NULL); 469 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 470 compile->clone_map().set_clone_idx(0); 471 compile->set_type_hwm(NULL); 472 compile->set_type_last_size(0); 473 compile->set_last_tf(NULL, NULL); 474 compile->set_indexSet_arena(NULL); 475 compile->set_indexSet_free_block_list(NULL); 476 compile->init_type_arena(); 477 Type::Initialize(compile); 478 _compile->set_scratch_buffer_blob(NULL); 479 _compile->begin_method(); 480 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 481 } 482 CompileWrapper::~CompileWrapper() { 483 _compile->end_method(); 484 if (_compile->scratch_buffer_blob() != NULL) 485 BufferBlob::free(_compile->scratch_buffer_blob()); 486 _compile->env()->set_compiler_data(NULL); 487 } 488 489 490 //----------------------------print_compile_messages--------------------------- 491 void Compile::print_compile_messages() { 492 #ifndef PRODUCT 493 // Check if recompiling 494 if (_subsume_loads == false && PrintOpto) { 495 // Recompiling without allowing machine instructions to subsume loads 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 501 // Recompiling without escape analysis 502 tty->print_cr("*********************************************************"); 503 tty->print_cr("** Bailout: Recompile without escape analysis **"); 504 tty->print_cr("*********************************************************"); 505 } 506 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 507 // Recompiling without boxing elimination 508 tty->print_cr("*********************************************************"); 509 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 510 tty->print_cr("*********************************************************"); 511 } 512 if (C->directive()->BreakAtCompileOption) { 513 // Open the debugger when compiling this method. 514 tty->print("### Breaking when compiling: "); 515 method()->print_short_name(); 516 tty->cr(); 517 BREAKPOINT; 518 } 519 520 if( PrintOpto ) { 521 if (is_osr_compilation()) { 522 tty->print("[OSR]%3d", _compile_id); 523 } else { 524 tty->print("%3d", _compile_id); 525 } 526 } 527 #endif 528 } 529 530 531 //-----------------------init_scratch_buffer_blob------------------------------ 532 // Construct a temporary BufferBlob and cache it for this compile. 533 void Compile::init_scratch_buffer_blob(int const_size) { 534 // If there is already a scratch buffer blob allocated and the 535 // constant section is big enough, use it. Otherwise free the 536 // current and allocate a new one. 537 BufferBlob* blob = scratch_buffer_blob(); 538 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 539 // Use the current blob. 540 } else { 541 if (blob != NULL) { 542 BufferBlob::free(blob); 543 } 544 545 ResourceMark rm; 546 _scratch_const_size = const_size; 547 int locs_size = sizeof(relocInfo) * MAX_locs_size; 548 int slop = 2 * CodeSection::end_slop(); // space between sections 549 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size + slop + locs_size); 550 blob = BufferBlob::create("Compile::scratch_buffer", size); 551 // Record the buffer blob for next time. 552 set_scratch_buffer_blob(blob); 553 // Have we run out of code space? 554 if (scratch_buffer_blob() == NULL) { 555 // Let CompilerBroker disable further compilations. 556 record_failure("Not enough space for scratch buffer in CodeCache"); 557 return; 558 } 559 } 560 561 // Initialize the relocation buffers 562 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 563 set_scratch_locs_memory(locs_buf); 564 } 565 566 567 //-----------------------scratch_emit_size------------------------------------- 568 // Helper function that computes size by emitting code 569 uint Compile::scratch_emit_size(const Node* n) { 570 // Start scratch_emit_size section. 571 set_in_scratch_emit_size(true); 572 573 // Emit into a trash buffer and count bytes emitted. 574 // This is a pretty expensive way to compute a size, 575 // but it works well enough if seldom used. 576 // All common fixed-size instructions are given a size 577 // method by the AD file. 578 // Note that the scratch buffer blob and locs memory are 579 // allocated at the beginning of the compile task, and 580 // may be shared by several calls to scratch_emit_size. 581 // The allocation of the scratch buffer blob is particularly 582 // expensive, since it has to grab the code cache lock. 583 BufferBlob* blob = this->scratch_buffer_blob(); 584 assert(blob != NULL, "Initialize BufferBlob at start"); 585 assert(blob->size() > MAX_inst_size, "sanity"); 586 relocInfo* locs_buf = scratch_locs_memory(); 587 address blob_begin = blob->content_begin(); 588 address blob_end = (address)locs_buf; 589 assert(blob->contains(blob_end), "sanity"); 590 CodeBuffer buf(blob_begin, blob_end - blob_begin); 591 buf.initialize_consts_size(_scratch_const_size); 592 buf.initialize_stubs_size(MAX_stubs_size); 593 assert(locs_buf != NULL, "sanity"); 594 int lsize = MAX_locs_size / 3; 595 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 596 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 597 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 598 // Mark as scratch buffer. 599 buf.consts()->set_scratch_emit(); 600 buf.insts()->set_scratch_emit(); 601 buf.stubs()->set_scratch_emit(); 602 603 // Do the emission. 604 605 Label fakeL; // Fake label for branch instructions. 606 Label* saveL = NULL; 607 uint save_bnum = 0; 608 bool is_branch = n->is_MachBranch(); 609 if (is_branch) { 610 MacroAssembler masm(&buf); 611 masm.bind(fakeL); 612 n->as_MachBranch()->save_label(&saveL, &save_bnum); 613 n->as_MachBranch()->label_set(&fakeL, 0); 614 } 615 n->emit(buf, this->regalloc()); 616 617 // Emitting into the scratch buffer should not fail 618 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 619 620 if (is_branch) // Restore label. 621 n->as_MachBranch()->label_set(saveL, save_bnum); 622 623 // End scratch_emit_size section. 624 set_in_scratch_emit_size(false); 625 626 return buf.insts_size(); 627 } 628 629 630 // ============================================================================ 631 //------------------------------Compile standard------------------------------- 632 debug_only( int Compile::_debug_idx = 100000; ) 633 634 // Compile a method. entry_bci is -1 for normal compilations and indicates 635 // the continuation bci for on stack replacement. 636 637 638 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 639 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 640 : Phase(Compiler), 641 _compile_id(ci_env->compile_id()), 642 _save_argument_registers(false), 643 _subsume_loads(subsume_loads), 644 _do_escape_analysis(do_escape_analysis), 645 _eliminate_boxing(eliminate_boxing), 646 _method(target), 647 _entry_bci(osr_bci), 648 _stub_function(NULL), 649 _stub_name(NULL), 650 _stub_entry_point(NULL), 651 _max_node_limit(MaxNodeLimit), 652 _orig_pc_slot(0), 653 _orig_pc_slot_offset_in_bytes(0), 654 _inlining_progress(false), 655 _inlining_incrementally(false), 656 _has_reserved_stack_access(target->has_reserved_stack_access()), 657 #ifndef PRODUCT 658 _trace_opto_output(directive->TraceOptoOutputOption), 659 #endif 660 _has_method_handle_invokes(false), 661 _comp_arena(mtCompiler), 662 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 663 _env(ci_env), 664 _directive(directive), 665 _log(ci_env->log()), 666 _failure_reason(NULL), 667 _congraph(NULL), 668 #ifndef PRODUCT 669 _printer(IdealGraphPrinter::printer()), 670 #endif 671 _dead_node_list(comp_arena()), 672 _dead_node_count(0), 673 _node_arena(mtCompiler), 674 _old_arena(mtCompiler), 675 _mach_constant_base_node(NULL), 676 _Compile_types(mtCompiler), 677 _initial_gvn(NULL), 678 _for_igvn(NULL), 679 _warm_calls(NULL), 680 _late_inlines(comp_arena(), 2, 0, NULL), 681 _string_late_inlines(comp_arena(), 2, 0, NULL), 682 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 683 _late_inlines_pos(0), 684 _number_of_mh_late_inlines(0), 685 _print_inlining_stream(NULL), 686 _print_inlining_list(NULL), 687 _print_inlining_idx(0), 688 _print_inlining_output(NULL), 689 _replay_inline_data(NULL), 690 _java_calls(0), 691 _inner_loops(0), 692 _interpreter_frame_size(0), 693 _node_bundling_limit(0), 694 _node_bundling_base(NULL), 695 _code_buffer("Compile::Fill_buffer"), 696 _scratch_const_size(-1), 697 _in_scratch_emit_size(false) 698 #ifndef PRODUCT 699 , _in_dump_cnt(0) 700 #endif 701 { 702 C = this; 703 #ifndef PRODUCT 704 if (_printer != NULL) { 705 _printer->set_compile(this); 706 } 707 #endif 708 CompileWrapper cw(this); 709 710 if (CITimeVerbose) { 711 tty->print(" "); 712 target->holder()->name()->print(); 713 tty->print("."); 714 target->print_short_name(); 715 tty->print(" "); 716 } 717 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 718 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 719 720 #ifndef PRODUCT 721 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 722 if (!print_opto_assembly) { 723 bool print_assembly = directive->PrintAssemblyOption; 724 if (print_assembly && !Disassembler::can_decode()) { 725 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 726 print_opto_assembly = true; 727 } 728 } 729 set_print_assembly(print_opto_assembly); 730 set_parsed_irreducible_loop(false); 731 732 if (directive->ReplayInlineOption) { 733 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 734 } 735 #endif 736 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 737 set_print_intrinsics(directive->PrintIntrinsicsOption); 738 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 739 740 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 741 // Make sure the method being compiled gets its own MDO, 742 // so we can at least track the decompile_count(). 743 // Need MDO to record RTM code generation state. 744 method()->ensure_method_data(); 745 } 746 747 Init(::AliasLevel); 748 749 750 print_compile_messages(); 751 752 _ilt = InlineTree::build_inline_tree_root(); 753 754 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 755 assert(num_alias_types() >= AliasIdxRaw, ""); 756 757 #define MINIMUM_NODE_HASH 1023 758 // Node list that Iterative GVN will start with 759 Unique_Node_List for_igvn(comp_arena()); 760 set_for_igvn(&for_igvn); 761 762 // GVN that will be run immediately on new nodes 763 uint estimated_size = method()->code_size()*4+64; 764 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 765 PhaseGVN gvn(node_arena(), estimated_size); 766 set_initial_gvn(&gvn); 767 768 print_inlining_init(); 769 { // Scope for timing the parser 770 TracePhase tp("parse", &timers[_t_parser]); 771 772 // Put top into the hash table ASAP. 773 initial_gvn()->transform_no_reclaim(top()); 774 775 // Set up tf(), start(), and find a CallGenerator. 776 CallGenerator* cg = NULL; 777 if (is_osr_compilation()) { 778 const TypeTuple *domain = StartOSRNode::osr_domain(); 779 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 780 init_tf(TypeFunc::make(domain, range)); 781 StartNode* s = new StartOSRNode(root(), domain); 782 initial_gvn()->set_type_bottom(s); 783 init_start(s); 784 cg = CallGenerator::for_osr(method(), entry_bci()); 785 } else { 786 // Normal case. 787 init_tf(TypeFunc::make(method())); 788 StartNode* s = new StartNode(root(), tf()->domain()); 789 initial_gvn()->set_type_bottom(s); 790 init_start(s); 791 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 792 // With java.lang.ref.reference.get() we must go through the 793 // intrinsic - even when get() is the root 794 // method of the compile - so that, if necessary, the value in 795 // the referent field of the reference object gets recorded by 796 // the pre-barrier code. 797 cg = find_intrinsic(method(), false); 798 } 799 if (cg == NULL) { 800 float past_uses = method()->interpreter_invocation_count(); 801 float expected_uses = past_uses; 802 cg = CallGenerator::for_inline(method(), expected_uses); 803 } 804 } 805 if (failing()) return; 806 if (cg == NULL) { 807 record_method_not_compilable("cannot parse method"); 808 return; 809 } 810 JVMState* jvms = build_start_state(start(), tf()); 811 if ((jvms = cg->generate(jvms)) == NULL) { 812 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 813 record_method_not_compilable("method parse failed"); 814 } 815 return; 816 } 817 GraphKit kit(jvms); 818 819 if (!kit.stopped()) { 820 // Accept return values, and transfer control we know not where. 821 // This is done by a special, unique ReturnNode bound to root. 822 return_values(kit.jvms()); 823 } 824 825 if (kit.has_exceptions()) { 826 // Any exceptions that escape from this call must be rethrown 827 // to whatever caller is dynamically above us on the stack. 828 // This is done by a special, unique RethrowNode bound to root. 829 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 830 } 831 832 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 833 834 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 835 inline_string_calls(true); 836 } 837 838 if (failing()) return; 839 840 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 841 842 // Remove clutter produced by parsing. 843 if (!failing()) { 844 ResourceMark rm; 845 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 846 } 847 } 848 849 // Note: Large methods are capped off in do_one_bytecode(). 850 if (failing()) return; 851 852 // After parsing, node notes are no longer automagic. 853 // They must be propagated by register_new_node_with_optimizer(), 854 // clone(), or the like. 855 set_default_node_notes(NULL); 856 857 for (;;) { 858 int successes = Inline_Warm(); 859 if (failing()) return; 860 if (successes == 0) break; 861 } 862 863 // Drain the list. 864 Finish_Warm(); 865 #ifndef PRODUCT 866 if (_printer && _printer->should_print(1)) { 867 _printer->print_inlining(); 868 } 869 #endif 870 871 if (failing()) return; 872 NOT_PRODUCT( verify_graph_edges(); ) 873 874 // Now optimize 875 Optimize(); 876 if (failing()) return; 877 NOT_PRODUCT( verify_graph_edges(); ) 878 879 #ifndef PRODUCT 880 if (PrintIdeal) { 881 ttyLocker ttyl; // keep the following output all in one block 882 // This output goes directly to the tty, not the compiler log. 883 // To enable tools to match it up with the compilation activity, 884 // be sure to tag this tty output with the compile ID. 885 if (xtty != NULL) { 886 xtty->head("ideal compile_id='%d'%s", compile_id(), 887 is_osr_compilation() ? " compile_kind='osr'" : 888 ""); 889 } 890 root()->dump(9999); 891 if (xtty != NULL) { 892 xtty->tail("ideal"); 893 } 894 } 895 #endif 896 897 NOT_PRODUCT( verify_barriers(); ) 898 899 // Dump compilation data to replay it. 900 if (directive->DumpReplayOption) { 901 env()->dump_replay_data(_compile_id); 902 } 903 if (directive->DumpInlineOption && (ilt() != NULL)) { 904 env()->dump_inline_data(_compile_id); 905 } 906 907 // Now that we know the size of all the monitors we can add a fixed slot 908 // for the original deopt pc. 909 910 _orig_pc_slot = fixed_slots(); 911 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 912 set_fixed_slots(next_slot); 913 914 // Compute when to use implicit null checks. Used by matching trap based 915 // nodes and NullCheck optimization. 916 set_allowed_deopt_reasons(); 917 918 // Now generate code 919 Code_Gen(); 920 if (failing()) return; 921 922 // Check if we want to skip execution of all compiled code. 923 { 924 #ifndef PRODUCT 925 if (OptoNoExecute) { 926 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 927 return; 928 } 929 #endif 930 TracePhase tp("install_code", &timers[_t_registerMethod]); 931 932 if (is_osr_compilation()) { 933 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 934 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 935 } else { 936 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 937 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 938 } 939 940 env()->register_method(_method, _entry_bci, 941 &_code_offsets, 942 _orig_pc_slot_offset_in_bytes, 943 code_buffer(), 944 frame_size_in_words(), _oop_map_set, 945 &_handler_table, &_inc_table, 946 compiler, 947 has_unsafe_access(), 948 SharedRuntime::is_wide_vector(max_vector_size()), 949 rtm_state() 950 ); 951 952 if (log() != NULL) // Print code cache state into compiler log 953 log()->code_cache_state(); 954 } 955 } 956 957 //------------------------------Compile---------------------------------------- 958 // Compile a runtime stub 959 Compile::Compile( ciEnv* ci_env, 960 TypeFunc_generator generator, 961 address stub_function, 962 const char *stub_name, 963 int is_fancy_jump, 964 bool pass_tls, 965 bool save_arg_registers, 966 bool return_pc, 967 DirectiveSet* directive) 968 : Phase(Compiler), 969 _compile_id(0), 970 _save_argument_registers(save_arg_registers), 971 _subsume_loads(true), 972 _do_escape_analysis(false), 973 _eliminate_boxing(false), 974 _method(NULL), 975 _entry_bci(InvocationEntryBci), 976 _stub_function(stub_function), 977 _stub_name(stub_name), 978 _stub_entry_point(NULL), 979 _max_node_limit(MaxNodeLimit), 980 _orig_pc_slot(0), 981 _orig_pc_slot_offset_in_bytes(0), 982 _inlining_progress(false), 983 _inlining_incrementally(false), 984 _has_reserved_stack_access(false), 985 #ifndef PRODUCT 986 _trace_opto_output(directive->TraceOptoOutputOption), 987 #endif 988 _has_method_handle_invokes(false), 989 _comp_arena(mtCompiler), 990 _env(ci_env), 991 _directive(directive), 992 _log(ci_env->log()), 993 _failure_reason(NULL), 994 _congraph(NULL), 995 #ifndef PRODUCT 996 _printer(NULL), 997 #endif 998 _dead_node_list(comp_arena()), 999 _dead_node_count(0), 1000 _node_arena(mtCompiler), 1001 _old_arena(mtCompiler), 1002 _mach_constant_base_node(NULL), 1003 _Compile_types(mtCompiler), 1004 _initial_gvn(NULL), 1005 _for_igvn(NULL), 1006 _warm_calls(NULL), 1007 _number_of_mh_late_inlines(0), 1008 _print_inlining_stream(NULL), 1009 _print_inlining_list(NULL), 1010 _print_inlining_idx(0), 1011 _print_inlining_output(NULL), 1012 _replay_inline_data(NULL), 1013 _java_calls(0), 1014 _inner_loops(0), 1015 _interpreter_frame_size(0), 1016 _node_bundling_limit(0), 1017 _node_bundling_base(NULL), 1018 _code_buffer("Compile::Fill_buffer"), 1019 #ifndef PRODUCT 1020 _in_dump_cnt(0), 1021 #endif 1022 _allowed_reasons(0) { 1023 C = this; 1024 1025 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1026 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1027 1028 #ifndef PRODUCT 1029 set_print_assembly(PrintFrameConverterAssembly); 1030 set_parsed_irreducible_loop(false); 1031 #endif 1032 set_has_irreducible_loop(false); // no loops 1033 1034 CompileWrapper cw(this); 1035 Init(/*AliasLevel=*/ 0); 1036 init_tf((*generator)()); 1037 1038 { 1039 // The following is a dummy for the sake of GraphKit::gen_stub 1040 Unique_Node_List for_igvn(comp_arena()); 1041 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1042 PhaseGVN gvn(Thread::current()->resource_area(),255); 1043 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1044 gvn.transform_no_reclaim(top()); 1045 1046 GraphKit kit; 1047 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1048 } 1049 1050 NOT_PRODUCT( verify_graph_edges(); ) 1051 Code_Gen(); 1052 if (failing()) return; 1053 1054 1055 // Entry point will be accessed using compile->stub_entry_point(); 1056 if (code_buffer() == NULL) { 1057 Matcher::soft_match_failure(); 1058 } else { 1059 if (PrintAssembly && (WizardMode || Verbose)) 1060 tty->print_cr("### Stub::%s", stub_name); 1061 1062 if (!failing()) { 1063 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1064 1065 // Make the NMethod 1066 // For now we mark the frame as never safe for profile stackwalking 1067 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1068 code_buffer(), 1069 CodeOffsets::frame_never_safe, 1070 // _code_offsets.value(CodeOffsets::Frame_Complete), 1071 frame_size_in_words(), 1072 _oop_map_set, 1073 save_arg_registers); 1074 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1075 1076 _stub_entry_point = rs->entry_point(); 1077 } 1078 } 1079 } 1080 1081 //------------------------------Init------------------------------------------- 1082 // Prepare for a single compilation 1083 void Compile::Init(int aliaslevel) { 1084 _unique = 0; 1085 _regalloc = NULL; 1086 1087 _tf = NULL; // filled in later 1088 _top = NULL; // cached later 1089 _matcher = NULL; // filled in later 1090 _cfg = NULL; // filled in later 1091 1092 set_24_bit_selection_and_mode(Use24BitFP, false); 1093 1094 _node_note_array = NULL; 1095 _default_node_notes = NULL; 1096 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1097 1098 _immutable_memory = NULL; // filled in at first inquiry 1099 1100 // Globally visible Nodes 1101 // First set TOP to NULL to give safe behavior during creation of RootNode 1102 set_cached_top_node(NULL); 1103 set_root(new RootNode()); 1104 // Now that you have a Root to point to, create the real TOP 1105 set_cached_top_node( new ConNode(Type::TOP) ); 1106 set_recent_alloc(NULL, NULL); 1107 1108 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1109 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1110 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1111 env()->set_dependencies(new Dependencies(env())); 1112 1113 _fixed_slots = 0; 1114 set_has_split_ifs(false); 1115 set_has_loops(has_method() && method()->has_loops()); // first approximation 1116 set_has_stringbuilder(false); 1117 set_has_boxed_value(false); 1118 _trap_can_recompile = false; // no traps emitted yet 1119 _major_progress = true; // start out assuming good things will happen 1120 set_has_unsafe_access(false); 1121 set_max_vector_size(0); 1122 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1123 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1124 set_decompile_count(0); 1125 1126 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1127 set_num_loop_opts(LoopOptsCount); 1128 set_do_inlining(Inline); 1129 set_max_inline_size(MaxInlineSize); 1130 set_freq_inline_size(FreqInlineSize); 1131 set_do_scheduling(OptoScheduling); 1132 set_do_count_invocations(false); 1133 set_do_method_data_update(false); 1134 1135 set_do_vector_loop(false); 1136 1137 if (AllowVectorizeOnDemand) { 1138 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1139 set_do_vector_loop(true); 1140 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());}) 1141 } else if (has_method() && method()->name() != 0 && 1142 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1143 set_do_vector_loop(true); 1144 } 1145 } 1146 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1147 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());}) 1148 1149 set_age_code(has_method() && method()->profile_aging()); 1150 set_rtm_state(NoRTM); // No RTM lock eliding by default 1151 _max_node_limit = _directive->MaxNodeLimitOption; 1152 1153 #if INCLUDE_RTM_OPT 1154 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1155 int rtm_state = method()->method_data()->rtm_state(); 1156 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1157 // Don't generate RTM lock eliding code. 1158 set_rtm_state(NoRTM); 1159 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1160 // Generate RTM lock eliding code without abort ratio calculation code. 1161 set_rtm_state(UseRTM); 1162 } else if (UseRTMDeopt) { 1163 // Generate RTM lock eliding code and include abort ratio calculation 1164 // code if UseRTMDeopt is on. 1165 set_rtm_state(ProfileRTM); 1166 } 1167 } 1168 #endif 1169 if (debug_info()->recording_non_safepoints()) { 1170 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1171 (comp_arena(), 8, 0, NULL)); 1172 set_default_node_notes(Node_Notes::make(this)); 1173 } 1174 1175 // // -- Initialize types before each compile -- 1176 // // Update cached type information 1177 // if( _method && _method->constants() ) 1178 // Type::update_loaded_types(_method, _method->constants()); 1179 1180 // Init alias_type map. 1181 if (!_do_escape_analysis && aliaslevel == 3) 1182 aliaslevel = 2; // No unique types without escape analysis 1183 _AliasLevel = aliaslevel; 1184 const int grow_ats = 16; 1185 _max_alias_types = grow_ats; 1186 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1187 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1188 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1189 { 1190 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1191 } 1192 // Initialize the first few types. 1193 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1194 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1195 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1196 _num_alias_types = AliasIdxRaw+1; 1197 // Zero out the alias type cache. 1198 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1199 // A NULL adr_type hits in the cache right away. Preload the right answer. 1200 probe_alias_cache(NULL)->_index = AliasIdxTop; 1201 1202 _intrinsics = NULL; 1203 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1204 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1205 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1206 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1207 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1208 register_library_intrinsics(); 1209 } 1210 1211 //---------------------------init_start---------------------------------------- 1212 // Install the StartNode on this compile object. 1213 void Compile::init_start(StartNode* s) { 1214 if (failing()) 1215 return; // already failing 1216 assert(s == start(), ""); 1217 } 1218 1219 /** 1220 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1221 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1222 * the ideal graph. 1223 */ 1224 StartNode* Compile::start() const { 1225 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1226 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1227 Node* start = root()->fast_out(i); 1228 if (start->is_Start()) { 1229 return start->as_Start(); 1230 } 1231 } 1232 fatal("Did not find Start node!"); 1233 return NULL; 1234 } 1235 1236 //-------------------------------immutable_memory------------------------------------- 1237 // Access immutable memory 1238 Node* Compile::immutable_memory() { 1239 if (_immutable_memory != NULL) { 1240 return _immutable_memory; 1241 } 1242 StartNode* s = start(); 1243 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1244 Node *p = s->fast_out(i); 1245 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1246 _immutable_memory = p; 1247 return _immutable_memory; 1248 } 1249 } 1250 ShouldNotReachHere(); 1251 return NULL; 1252 } 1253 1254 //----------------------set_cached_top_node------------------------------------ 1255 // Install the cached top node, and make sure Node::is_top works correctly. 1256 void Compile::set_cached_top_node(Node* tn) { 1257 if (tn != NULL) verify_top(tn); 1258 Node* old_top = _top; 1259 _top = tn; 1260 // Calling Node::setup_is_top allows the nodes the chance to adjust 1261 // their _out arrays. 1262 if (_top != NULL) _top->setup_is_top(); 1263 if (old_top != NULL) old_top->setup_is_top(); 1264 assert(_top == NULL || top()->is_top(), ""); 1265 } 1266 1267 #ifdef ASSERT 1268 uint Compile::count_live_nodes_by_graph_walk() { 1269 Unique_Node_List useful(comp_arena()); 1270 // Get useful node list by walking the graph. 1271 identify_useful_nodes(useful); 1272 return useful.size(); 1273 } 1274 1275 void Compile::print_missing_nodes() { 1276 1277 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1278 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1279 return; 1280 } 1281 1282 // This is an expensive function. It is executed only when the user 1283 // specifies VerifyIdealNodeCount option or otherwise knows the 1284 // additional work that needs to be done to identify reachable nodes 1285 // by walking the flow graph and find the missing ones using 1286 // _dead_node_list. 1287 1288 Unique_Node_List useful(comp_arena()); 1289 // Get useful node list by walking the graph. 1290 identify_useful_nodes(useful); 1291 1292 uint l_nodes = C->live_nodes(); 1293 uint l_nodes_by_walk = useful.size(); 1294 1295 if (l_nodes != l_nodes_by_walk) { 1296 if (_log != NULL) { 1297 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1298 _log->stamp(); 1299 _log->end_head(); 1300 } 1301 VectorSet& useful_member_set = useful.member_set(); 1302 int last_idx = l_nodes_by_walk; 1303 for (int i = 0; i < last_idx; i++) { 1304 if (useful_member_set.test(i)) { 1305 if (_dead_node_list.test(i)) { 1306 if (_log != NULL) { 1307 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1308 } 1309 if (PrintIdealNodeCount) { 1310 // Print the log message to tty 1311 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1312 useful.at(i)->dump(); 1313 } 1314 } 1315 } 1316 else if (! _dead_node_list.test(i)) { 1317 if (_log != NULL) { 1318 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1319 } 1320 if (PrintIdealNodeCount) { 1321 // Print the log message to tty 1322 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1323 } 1324 } 1325 } 1326 if (_log != NULL) { 1327 _log->tail("mismatched_nodes"); 1328 } 1329 } 1330 } 1331 void Compile::record_modified_node(Node* n) { 1332 if (_modified_nodes != NULL && !_inlining_incrementally && 1333 n->outcnt() != 0 && !n->is_Con()) { 1334 _modified_nodes->push(n); 1335 } 1336 } 1337 1338 void Compile::remove_modified_node(Node* n) { 1339 if (_modified_nodes != NULL) { 1340 _modified_nodes->remove(n); 1341 } 1342 } 1343 #endif 1344 1345 #ifndef PRODUCT 1346 void Compile::verify_top(Node* tn) const { 1347 if (tn != NULL) { 1348 assert(tn->is_Con(), "top node must be a constant"); 1349 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1350 assert(tn->in(0) != NULL, "must have live top node"); 1351 } 1352 } 1353 #endif 1354 1355 1356 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1357 1358 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1359 guarantee(arr != NULL, ""); 1360 int num_blocks = arr->length(); 1361 if (grow_by < num_blocks) grow_by = num_blocks; 1362 int num_notes = grow_by * _node_notes_block_size; 1363 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1364 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1365 while (num_notes > 0) { 1366 arr->append(notes); 1367 notes += _node_notes_block_size; 1368 num_notes -= _node_notes_block_size; 1369 } 1370 assert(num_notes == 0, "exact multiple, please"); 1371 } 1372 1373 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1374 if (source == NULL || dest == NULL) return false; 1375 1376 if (dest->is_Con()) 1377 return false; // Do not push debug info onto constants. 1378 1379 #ifdef ASSERT 1380 // Leave a bread crumb trail pointing to the original node: 1381 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1382 dest->set_debug_orig(source); 1383 } 1384 #endif 1385 1386 if (node_note_array() == NULL) 1387 return false; // Not collecting any notes now. 1388 1389 // This is a copy onto a pre-existing node, which may already have notes. 1390 // If both nodes have notes, do not overwrite any pre-existing notes. 1391 Node_Notes* source_notes = node_notes_at(source->_idx); 1392 if (source_notes == NULL || source_notes->is_clear()) return false; 1393 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1394 if (dest_notes == NULL || dest_notes->is_clear()) { 1395 return set_node_notes_at(dest->_idx, source_notes); 1396 } 1397 1398 Node_Notes merged_notes = (*source_notes); 1399 // The order of operations here ensures that dest notes will win... 1400 merged_notes.update_from(dest_notes); 1401 return set_node_notes_at(dest->_idx, &merged_notes); 1402 } 1403 1404 1405 //--------------------------allow_range_check_smearing------------------------- 1406 // Gating condition for coalescing similar range checks. 1407 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1408 // single covering check that is at least as strong as any of them. 1409 // If the optimization succeeds, the simplified (strengthened) range check 1410 // will always succeed. If it fails, we will deopt, and then give up 1411 // on the optimization. 1412 bool Compile::allow_range_check_smearing() const { 1413 // If this method has already thrown a range-check, 1414 // assume it was because we already tried range smearing 1415 // and it failed. 1416 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1417 return !already_trapped; 1418 } 1419 1420 1421 //------------------------------flatten_alias_type----------------------------- 1422 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1423 int offset = tj->offset(); 1424 TypePtr::PTR ptr = tj->ptr(); 1425 1426 // Known instance (scalarizable allocation) alias only with itself. 1427 bool is_known_inst = tj->isa_oopptr() != NULL && 1428 tj->is_oopptr()->is_known_instance(); 1429 1430 // Process weird unsafe references. 1431 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1432 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1433 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1434 tj = TypeOopPtr::BOTTOM; 1435 ptr = tj->ptr(); 1436 offset = tj->offset(); 1437 } 1438 1439 // Array pointers need some flattening 1440 const TypeAryPtr *ta = tj->isa_aryptr(); 1441 if (ta && ta->is_stable()) { 1442 // Erase stability property for alias analysis. 1443 tj = ta = ta->cast_to_stable(false); 1444 } 1445 if( ta && is_known_inst ) { 1446 if ( offset != Type::OffsetBot && 1447 offset > arrayOopDesc::length_offset_in_bytes() ) { 1448 offset = Type::OffsetBot; // Flatten constant access into array body only 1449 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1450 } 1451 } else if( ta && _AliasLevel >= 2 ) { 1452 // For arrays indexed by constant indices, we flatten the alias 1453 // space to include all of the array body. Only the header, klass 1454 // and array length can be accessed un-aliased. 1455 if( offset != Type::OffsetBot ) { 1456 if( ta->const_oop() ) { // MethodData* or Method* 1457 offset = Type::OffsetBot; // Flatten constant access into array body 1458 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1459 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1460 // range is OK as-is. 1461 tj = ta = TypeAryPtr::RANGE; 1462 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1463 tj = TypeInstPtr::KLASS; // all klass loads look alike 1464 ta = TypeAryPtr::RANGE; // generic ignored junk 1465 ptr = TypePtr::BotPTR; 1466 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1467 tj = TypeInstPtr::MARK; 1468 ta = TypeAryPtr::RANGE; // generic ignored junk 1469 ptr = TypePtr::BotPTR; 1470 } else { // Random constant offset into array body 1471 offset = Type::OffsetBot; // Flatten constant access into array body 1472 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1473 } 1474 } 1475 // Arrays of fixed size alias with arrays of unknown size. 1476 if (ta->size() != TypeInt::POS) { 1477 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1478 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1479 } 1480 // Arrays of known objects become arrays of unknown objects. 1481 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1482 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1483 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1484 } 1485 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1486 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1487 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1488 } 1489 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1490 // cannot be distinguished by bytecode alone. 1491 if (ta->elem() == TypeInt::BOOL) { 1492 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1493 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1494 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1495 } 1496 // During the 2nd round of IterGVN, NotNull castings are removed. 1497 // Make sure the Bottom and NotNull variants alias the same. 1498 // Also, make sure exact and non-exact variants alias the same. 1499 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1500 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1501 } 1502 } 1503 1504 // Oop pointers need some flattening 1505 const TypeInstPtr *to = tj->isa_instptr(); 1506 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1507 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1508 if( ptr == TypePtr::Constant ) { 1509 if (to->klass() != ciEnv::current()->Class_klass() || 1510 offset < k->size_helper() * wordSize) { 1511 // No constant oop pointers (such as Strings); they alias with 1512 // unknown strings. 1513 assert(!is_known_inst, "not scalarizable allocation"); 1514 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1515 } 1516 } else if( is_known_inst ) { 1517 tj = to; // Keep NotNull and klass_is_exact for instance type 1518 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1519 // During the 2nd round of IterGVN, NotNull castings are removed. 1520 // Make sure the Bottom and NotNull variants alias the same. 1521 // Also, make sure exact and non-exact variants alias the same. 1522 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1523 } 1524 if (to->speculative() != NULL) { 1525 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1526 } 1527 // Canonicalize the holder of this field 1528 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1529 // First handle header references such as a LoadKlassNode, even if the 1530 // object's klass is unloaded at compile time (4965979). 1531 if (!is_known_inst) { // Do it only for non-instance types 1532 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1533 } 1534 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1535 // Static fields are in the space above the normal instance 1536 // fields in the java.lang.Class instance. 1537 if (to->klass() != ciEnv::current()->Class_klass()) { 1538 to = NULL; 1539 tj = TypeOopPtr::BOTTOM; 1540 offset = tj->offset(); 1541 } 1542 } else { 1543 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1544 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1545 if( is_known_inst ) { 1546 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1547 } else { 1548 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1549 } 1550 } 1551 } 1552 } 1553 1554 // Klass pointers to object array klasses need some flattening 1555 const TypeKlassPtr *tk = tj->isa_klassptr(); 1556 if( tk ) { 1557 // If we are referencing a field within a Klass, we need 1558 // to assume the worst case of an Object. Both exact and 1559 // inexact types must flatten to the same alias class so 1560 // use NotNull as the PTR. 1561 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1562 1563 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1564 TypeKlassPtr::OBJECT->klass(), 1565 offset); 1566 } 1567 1568 ciKlass* klass = tk->klass(); 1569 if( klass->is_obj_array_klass() ) { 1570 ciKlass* k = TypeAryPtr::OOPS->klass(); 1571 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1572 k = TypeInstPtr::BOTTOM->klass(); 1573 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1574 } 1575 1576 // Check for precise loads from the primary supertype array and force them 1577 // to the supertype cache alias index. Check for generic array loads from 1578 // the primary supertype array and also force them to the supertype cache 1579 // alias index. Since the same load can reach both, we need to merge 1580 // these 2 disparate memories into the same alias class. Since the 1581 // primary supertype array is read-only, there's no chance of confusion 1582 // where we bypass an array load and an array store. 1583 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1584 if (offset == Type::OffsetBot || 1585 (offset >= primary_supers_offset && 1586 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1587 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1588 offset = in_bytes(Klass::secondary_super_cache_offset()); 1589 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1590 } 1591 } 1592 1593 // Flatten all Raw pointers together. 1594 if (tj->base() == Type::RawPtr) 1595 tj = TypeRawPtr::BOTTOM; 1596 1597 if (tj->base() == Type::AnyPtr) 1598 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1599 1600 // Flatten all to bottom for now 1601 switch( _AliasLevel ) { 1602 case 0: 1603 tj = TypePtr::BOTTOM; 1604 break; 1605 case 1: // Flatten to: oop, static, field or array 1606 switch (tj->base()) { 1607 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1608 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1609 case Type::AryPtr: // do not distinguish arrays at all 1610 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1611 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1612 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1613 default: ShouldNotReachHere(); 1614 } 1615 break; 1616 case 2: // No collapsing at level 2; keep all splits 1617 case 3: // No collapsing at level 3; keep all splits 1618 break; 1619 default: 1620 Unimplemented(); 1621 } 1622 1623 offset = tj->offset(); 1624 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1625 1626 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1627 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1628 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1629 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1630 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1631 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1632 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1633 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1634 assert( tj->ptr() != TypePtr::TopPTR && 1635 tj->ptr() != TypePtr::AnyNull && 1636 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1637 // assert( tj->ptr() != TypePtr::Constant || 1638 // tj->base() == Type::RawPtr || 1639 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1640 1641 return tj; 1642 } 1643 1644 void Compile::AliasType::Init(int i, const TypePtr* at) { 1645 _index = i; 1646 _adr_type = at; 1647 _field = NULL; 1648 _element = NULL; 1649 _is_rewritable = true; // default 1650 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1651 if (atoop != NULL && atoop->is_known_instance()) { 1652 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1653 _general_index = Compile::current()->get_alias_index(gt); 1654 } else { 1655 _general_index = 0; 1656 } 1657 } 1658 1659 BasicType Compile::AliasType::basic_type() const { 1660 if (element() != NULL) { 1661 const Type* element = adr_type()->is_aryptr()->elem(); 1662 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1663 } if (field() != NULL) { 1664 return field()->layout_type(); 1665 } else { 1666 return T_ILLEGAL; // unknown 1667 } 1668 } 1669 1670 //---------------------------------print_on------------------------------------ 1671 #ifndef PRODUCT 1672 void Compile::AliasType::print_on(outputStream* st) { 1673 if (index() < 10) 1674 st->print("@ <%d> ", index()); 1675 else st->print("@ <%d>", index()); 1676 st->print(is_rewritable() ? " " : " RO"); 1677 int offset = adr_type()->offset(); 1678 if (offset == Type::OffsetBot) 1679 st->print(" +any"); 1680 else st->print(" +%-3d", offset); 1681 st->print(" in "); 1682 adr_type()->dump_on(st); 1683 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1684 if (field() != NULL && tjp) { 1685 if (tjp->klass() != field()->holder() || 1686 tjp->offset() != field()->offset_in_bytes()) { 1687 st->print(" != "); 1688 field()->print(); 1689 st->print(" ***"); 1690 } 1691 } 1692 } 1693 1694 void print_alias_types() { 1695 Compile* C = Compile::current(); 1696 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1697 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1698 C->alias_type(idx)->print_on(tty); 1699 tty->cr(); 1700 } 1701 } 1702 #endif 1703 1704 1705 //----------------------------probe_alias_cache-------------------------------- 1706 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1707 intptr_t key = (intptr_t) adr_type; 1708 key ^= key >> logAliasCacheSize; 1709 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1710 } 1711 1712 1713 //-----------------------------grow_alias_types-------------------------------- 1714 void Compile::grow_alias_types() { 1715 const int old_ats = _max_alias_types; // how many before? 1716 const int new_ats = old_ats; // how many more? 1717 const int grow_ats = old_ats+new_ats; // how many now? 1718 _max_alias_types = grow_ats; 1719 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1720 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1721 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1722 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1723 } 1724 1725 1726 //--------------------------------find_alias_type------------------------------ 1727 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1728 if (_AliasLevel == 0) 1729 return alias_type(AliasIdxBot); 1730 1731 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1732 if (ace->_adr_type == adr_type) { 1733 return alias_type(ace->_index); 1734 } 1735 1736 // Handle special cases. 1737 if (adr_type == NULL) return alias_type(AliasIdxTop); 1738 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1739 1740 // Do it the slow way. 1741 const TypePtr* flat = flatten_alias_type(adr_type); 1742 1743 #ifdef ASSERT 1744 { 1745 ResourceMark rm; 1746 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1747 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1748 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1749 Type::str(adr_type)); 1750 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1751 const TypeOopPtr* foop = flat->is_oopptr(); 1752 // Scalarizable allocations have exact klass always. 1753 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1754 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1755 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1756 Type::str(foop), Type::str(xoop)); 1757 } 1758 } 1759 #endif 1760 1761 int idx = AliasIdxTop; 1762 for (int i = 0; i < num_alias_types(); i++) { 1763 if (alias_type(i)->adr_type() == flat) { 1764 idx = i; 1765 break; 1766 } 1767 } 1768 1769 if (idx == AliasIdxTop) { 1770 if (no_create) return NULL; 1771 // Grow the array if necessary. 1772 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1773 // Add a new alias type. 1774 idx = _num_alias_types++; 1775 _alias_types[idx]->Init(idx, flat); 1776 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1777 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1778 if (flat->isa_instptr()) { 1779 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1780 && flat->is_instptr()->klass() == env()->Class_klass()) 1781 alias_type(idx)->set_rewritable(false); 1782 } 1783 if (flat->isa_aryptr()) { 1784 #ifdef ASSERT 1785 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1786 // (T_BYTE has the weakest alignment and size restrictions...) 1787 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1788 #endif 1789 if (flat->offset() == TypePtr::OffsetBot) { 1790 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1791 } 1792 } 1793 if (flat->isa_klassptr()) { 1794 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1795 alias_type(idx)->set_rewritable(false); 1796 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1797 alias_type(idx)->set_rewritable(false); 1798 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1799 alias_type(idx)->set_rewritable(false); 1800 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1801 alias_type(idx)->set_rewritable(false); 1802 } 1803 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1804 // but the base pointer type is not distinctive enough to identify 1805 // references into JavaThread.) 1806 1807 // Check for final fields. 1808 const TypeInstPtr* tinst = flat->isa_instptr(); 1809 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1810 ciField* field; 1811 if (tinst->const_oop() != NULL && 1812 tinst->klass() == ciEnv::current()->Class_klass() && 1813 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1814 // static field 1815 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1816 field = k->get_field_by_offset(tinst->offset(), true); 1817 } else { 1818 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1819 field = k->get_field_by_offset(tinst->offset(), false); 1820 } 1821 assert(field == NULL || 1822 original_field == NULL || 1823 (field->holder() == original_field->holder() && 1824 field->offset() == original_field->offset() && 1825 field->is_static() == original_field->is_static()), "wrong field?"); 1826 // Set field() and is_rewritable() attributes. 1827 if (field != NULL) alias_type(idx)->set_field(field); 1828 } 1829 } 1830 1831 // Fill the cache for next time. 1832 ace->_adr_type = adr_type; 1833 ace->_index = idx; 1834 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1835 1836 // Might as well try to fill the cache for the flattened version, too. 1837 AliasCacheEntry* face = probe_alias_cache(flat); 1838 if (face->_adr_type == NULL) { 1839 face->_adr_type = flat; 1840 face->_index = idx; 1841 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1842 } 1843 1844 return alias_type(idx); 1845 } 1846 1847 1848 Compile::AliasType* Compile::alias_type(ciField* field) { 1849 const TypeOopPtr* t; 1850 if (field->is_static()) 1851 t = TypeInstPtr::make(field->holder()->java_mirror()); 1852 else 1853 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1854 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1855 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1856 return atp; 1857 } 1858 1859 1860 //------------------------------have_alias_type-------------------------------- 1861 bool Compile::have_alias_type(const TypePtr* adr_type) { 1862 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1863 if (ace->_adr_type == adr_type) { 1864 return true; 1865 } 1866 1867 // Handle special cases. 1868 if (adr_type == NULL) return true; 1869 if (adr_type == TypePtr::BOTTOM) return true; 1870 1871 return find_alias_type(adr_type, true, NULL) != NULL; 1872 } 1873 1874 //-----------------------------must_alias-------------------------------------- 1875 // True if all values of the given address type are in the given alias category. 1876 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1877 if (alias_idx == AliasIdxBot) return true; // the universal category 1878 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1879 if (alias_idx == AliasIdxTop) return false; // the empty category 1880 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1881 1882 // the only remaining possible overlap is identity 1883 int adr_idx = get_alias_index(adr_type); 1884 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1885 assert(adr_idx == alias_idx || 1886 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1887 && adr_type != TypeOopPtr::BOTTOM), 1888 "should not be testing for overlap with an unsafe pointer"); 1889 return adr_idx == alias_idx; 1890 } 1891 1892 //------------------------------can_alias-------------------------------------- 1893 // True if any values of the given address type are in the given alias category. 1894 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1895 if (alias_idx == AliasIdxTop) return false; // the empty category 1896 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1897 if (alias_idx == AliasIdxBot) return true; // the universal category 1898 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1899 1900 // the only remaining possible overlap is identity 1901 int adr_idx = get_alias_index(adr_type); 1902 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1903 return adr_idx == alias_idx; 1904 } 1905 1906 1907 1908 //---------------------------pop_warm_call------------------------------------- 1909 WarmCallInfo* Compile::pop_warm_call() { 1910 WarmCallInfo* wci = _warm_calls; 1911 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1912 return wci; 1913 } 1914 1915 //----------------------------Inline_Warm-------------------------------------- 1916 int Compile::Inline_Warm() { 1917 // If there is room, try to inline some more warm call sites. 1918 // %%% Do a graph index compaction pass when we think we're out of space? 1919 if (!InlineWarmCalls) return 0; 1920 1921 int calls_made_hot = 0; 1922 int room_to_grow = NodeCountInliningCutoff - unique(); 1923 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1924 int amount_grown = 0; 1925 WarmCallInfo* call; 1926 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1927 int est_size = (int)call->size(); 1928 if (est_size > (room_to_grow - amount_grown)) { 1929 // This one won't fit anyway. Get rid of it. 1930 call->make_cold(); 1931 continue; 1932 } 1933 call->make_hot(); 1934 calls_made_hot++; 1935 amount_grown += est_size; 1936 amount_to_grow -= est_size; 1937 } 1938 1939 if (calls_made_hot > 0) set_major_progress(); 1940 return calls_made_hot; 1941 } 1942 1943 1944 //----------------------------Finish_Warm-------------------------------------- 1945 void Compile::Finish_Warm() { 1946 if (!InlineWarmCalls) return; 1947 if (failing()) return; 1948 if (warm_calls() == NULL) return; 1949 1950 // Clean up loose ends, if we are out of space for inlining. 1951 WarmCallInfo* call; 1952 while ((call = pop_warm_call()) != NULL) { 1953 call->make_cold(); 1954 } 1955 } 1956 1957 //---------------------cleanup_loop_predicates----------------------- 1958 // Remove the opaque nodes that protect the predicates so that all unused 1959 // checks and uncommon_traps will be eliminated from the ideal graph 1960 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1961 if (predicate_count()==0) return; 1962 for (int i = predicate_count(); i > 0; i--) { 1963 Node * n = predicate_opaque1_node(i-1); 1964 assert(n->Opcode() == Op_Opaque1, "must be"); 1965 igvn.replace_node(n, n->in(1)); 1966 } 1967 assert(predicate_count()==0, "should be clean!"); 1968 } 1969 1970 void Compile::add_range_check_cast(Node* n) { 1971 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1972 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1973 _range_check_casts->append(n); 1974 } 1975 1976 // Remove all range check dependent CastIINodes. 1977 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1978 for (int i = range_check_cast_count(); i > 0; i--) { 1979 Node* cast = range_check_cast_node(i-1); 1980 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1981 igvn.replace_node(cast, cast->in(1)); 1982 } 1983 assert(range_check_cast_count() == 0, "should be empty"); 1984 } 1985 1986 void Compile::add_opaque4_node(Node* n) { 1987 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1988 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1989 _opaque4_nodes->append(n); 1990 } 1991 1992 // Remove all Opaque4 nodes. 1993 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1994 for (int i = opaque4_count(); i > 0; i--) { 1995 Node* opaq = opaque4_node(i-1); 1996 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 1997 igvn.replace_node(opaq, opaq->in(2)); 1998 } 1999 assert(opaque4_count() == 0, "should be empty"); 2000 } 2001 2002 // StringOpts and late inlining of string methods 2003 void Compile::inline_string_calls(bool parse_time) { 2004 { 2005 // remove useless nodes to make the usage analysis simpler 2006 ResourceMark rm; 2007 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2008 } 2009 2010 { 2011 ResourceMark rm; 2012 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2013 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2014 print_method(PHASE_AFTER_STRINGOPTS, 3); 2015 } 2016 2017 // now inline anything that we skipped the first time around 2018 if (!parse_time) { 2019 _late_inlines_pos = _late_inlines.length(); 2020 } 2021 2022 while (_string_late_inlines.length() > 0) { 2023 CallGenerator* cg = _string_late_inlines.pop(); 2024 cg->do_late_inline(); 2025 if (failing()) return; 2026 } 2027 _string_late_inlines.trunc_to(0); 2028 } 2029 2030 // Late inlining of boxing methods 2031 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2032 if (_boxing_late_inlines.length() > 0) { 2033 assert(has_boxed_value(), "inconsistent"); 2034 2035 PhaseGVN* gvn = initial_gvn(); 2036 set_inlining_incrementally(true); 2037 2038 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2039 for_igvn()->clear(); 2040 gvn->replace_with(&igvn); 2041 2042 _late_inlines_pos = _late_inlines.length(); 2043 2044 while (_boxing_late_inlines.length() > 0) { 2045 CallGenerator* cg = _boxing_late_inlines.pop(); 2046 cg->do_late_inline(); 2047 if (failing()) return; 2048 } 2049 _boxing_late_inlines.trunc_to(0); 2050 2051 { 2052 ResourceMark rm; 2053 PhaseRemoveUseless pru(gvn, for_igvn()); 2054 } 2055 2056 igvn = PhaseIterGVN(gvn); 2057 igvn.optimize(); 2058 2059 set_inlining_progress(false); 2060 set_inlining_incrementally(false); 2061 } 2062 } 2063 2064 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2065 assert(IncrementalInline, "incremental inlining should be on"); 2066 PhaseGVN* gvn = initial_gvn(); 2067 2068 set_inlining_progress(false); 2069 for_igvn()->clear(); 2070 gvn->replace_with(&igvn); 2071 2072 { 2073 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2074 int i = 0; 2075 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2076 CallGenerator* cg = _late_inlines.at(i); 2077 _late_inlines_pos = i+1; 2078 cg->do_late_inline(); 2079 if (failing()) return; 2080 } 2081 int j = 0; 2082 for (; i < _late_inlines.length(); i++, j++) { 2083 _late_inlines.at_put(j, _late_inlines.at(i)); 2084 } 2085 _late_inlines.trunc_to(j); 2086 } 2087 2088 { 2089 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2090 ResourceMark rm; 2091 PhaseRemoveUseless pru(gvn, for_igvn()); 2092 } 2093 2094 { 2095 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2096 igvn = PhaseIterGVN(gvn); 2097 } 2098 } 2099 2100 // Perform incremental inlining until bound on number of live nodes is reached 2101 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2102 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2103 2104 PhaseGVN* gvn = initial_gvn(); 2105 2106 set_inlining_incrementally(true); 2107 set_inlining_progress(true); 2108 uint low_live_nodes = 0; 2109 2110 while(inlining_progress() && _late_inlines.length() > 0) { 2111 2112 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2113 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2114 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2115 // PhaseIdealLoop is expensive so we only try it once we are 2116 // out of live nodes and we only try it again if the previous 2117 // helped got the number of nodes down significantly 2118 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2119 if (failing()) return; 2120 low_live_nodes = live_nodes(); 2121 _major_progress = true; 2122 } 2123 2124 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2125 break; 2126 } 2127 } 2128 2129 inline_incrementally_one(igvn); 2130 2131 if (failing()) return; 2132 2133 { 2134 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2135 igvn.optimize(); 2136 } 2137 2138 if (failing()) return; 2139 } 2140 2141 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2142 2143 if (_string_late_inlines.length() > 0) { 2144 assert(has_stringbuilder(), "inconsistent"); 2145 for_igvn()->clear(); 2146 initial_gvn()->replace_with(&igvn); 2147 2148 inline_string_calls(false); 2149 2150 if (failing()) return; 2151 2152 { 2153 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2154 ResourceMark rm; 2155 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2156 } 2157 2158 { 2159 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2160 igvn = PhaseIterGVN(gvn); 2161 igvn.optimize(); 2162 } 2163 } 2164 2165 set_inlining_incrementally(false); 2166 } 2167 2168 2169 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) { 2170 if(loop_opts_cnt > 0) { 2171 debug_only( int cnt = 0; ); 2172 while(major_progress() && (loop_opts_cnt > 0)) { 2173 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2174 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2175 PhaseIdealLoop ideal_loop(igvn, mode); 2176 loop_opts_cnt--; 2177 if (failing()) return false; 2178 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2179 } 2180 } 2181 return true; 2182 } 2183 2184 //------------------------------Optimize--------------------------------------- 2185 // Given a graph, optimize it. 2186 void Compile::Optimize() { 2187 TracePhase tp("optimizer", &timers[_t_optimizer]); 2188 2189 #ifndef PRODUCT 2190 if (_directive->BreakAtCompileOption) { 2191 BREAKPOINT; 2192 } 2193 2194 #endif 2195 2196 #ifdef ASSERT 2197 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2198 bs->verify_gc_barriers(true); 2199 #endif 2200 2201 ResourceMark rm; 2202 int loop_opts_cnt; 2203 2204 print_inlining_reinit(); 2205 2206 NOT_PRODUCT( verify_graph_edges(); ) 2207 2208 print_method(PHASE_AFTER_PARSING); 2209 2210 { 2211 // Iterative Global Value Numbering, including ideal transforms 2212 // Initialize IterGVN with types and values from parse-time GVN 2213 PhaseIterGVN igvn(initial_gvn()); 2214 #ifdef ASSERT 2215 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2216 #endif 2217 { 2218 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2219 igvn.optimize(); 2220 } 2221 2222 if (failing()) return; 2223 2224 print_method(PHASE_ITER_GVN1, 2); 2225 2226 inline_incrementally(igvn); 2227 2228 print_method(PHASE_INCREMENTAL_INLINE, 2); 2229 2230 if (failing()) return; 2231 2232 if (eliminate_boxing()) { 2233 // Inline valueOf() methods now. 2234 inline_boxing_calls(igvn); 2235 2236 if (AlwaysIncrementalInline) { 2237 inline_incrementally(igvn); 2238 } 2239 2240 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2241 2242 if (failing()) return; 2243 } 2244 2245 // Remove the speculative part of types and clean up the graph from 2246 // the extra CastPP nodes whose only purpose is to carry them. Do 2247 // that early so that optimizations are not disrupted by the extra 2248 // CastPP nodes. 2249 remove_speculative_types(igvn); 2250 2251 // No more new expensive nodes will be added to the list from here 2252 // so keep only the actual candidates for optimizations. 2253 cleanup_expensive_nodes(igvn); 2254 2255 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2256 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2257 initial_gvn()->replace_with(&igvn); 2258 for_igvn()->clear(); 2259 Unique_Node_List new_worklist(C->comp_arena()); 2260 { 2261 ResourceMark rm; 2262 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2263 } 2264 set_for_igvn(&new_worklist); 2265 igvn = PhaseIterGVN(initial_gvn()); 2266 igvn.optimize(); 2267 } 2268 2269 // Perform escape analysis 2270 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2271 if (has_loops()) { 2272 // Cleanup graph (remove dead nodes). 2273 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2274 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2275 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2276 if (failing()) return; 2277 } 2278 ConnectionGraph::do_analysis(this, &igvn); 2279 2280 if (failing()) return; 2281 2282 // Optimize out fields loads from scalar replaceable allocations. 2283 igvn.optimize(); 2284 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2285 2286 if (failing()) return; 2287 2288 if (congraph() != NULL && macro_count() > 0) { 2289 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2290 PhaseMacroExpand mexp(igvn); 2291 mexp.eliminate_macro_nodes(); 2292 igvn.set_delay_transform(false); 2293 2294 igvn.optimize(); 2295 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2296 2297 if (failing()) return; 2298 } 2299 } 2300 2301 // Loop transforms on the ideal graph. Range Check Elimination, 2302 // peeling, unrolling, etc. 2303 2304 // Set loop opts counter 2305 loop_opts_cnt = num_loop_opts(); 2306 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2307 { 2308 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2309 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2310 loop_opts_cnt--; 2311 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2312 if (failing()) return; 2313 } 2314 // Loop opts pass if partial peeling occurred in previous pass 2315 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2316 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2317 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2318 loop_opts_cnt--; 2319 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2320 if (failing()) return; 2321 } 2322 // Loop opts pass for loop-unrolling before CCP 2323 if(major_progress() && (loop_opts_cnt > 0)) { 2324 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2325 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2326 loop_opts_cnt--; 2327 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2328 } 2329 if (!failing()) { 2330 // Verify that last round of loop opts produced a valid graph 2331 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2332 PhaseIdealLoop::verify(igvn); 2333 } 2334 } 2335 if (failing()) return; 2336 2337 // Conditional Constant Propagation; 2338 PhaseCCP ccp( &igvn ); 2339 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2340 { 2341 TracePhase tp("ccp", &timers[_t_ccp]); 2342 ccp.do_transform(); 2343 } 2344 print_method(PHASE_CPP1, 2); 2345 2346 assert( true, "Break here to ccp.dump_old2new_map()"); 2347 2348 // Iterative Global Value Numbering, including ideal transforms 2349 { 2350 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2351 igvn = ccp; 2352 igvn.optimize(); 2353 } 2354 2355 print_method(PHASE_ITER_GVN2, 2); 2356 2357 if (failing()) return; 2358 2359 // Loop transforms on the ideal graph. Range Check Elimination, 2360 // peeling, unrolling, etc. 2361 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2362 return; 2363 } 2364 2365 #if INCLUDE_ZGC 2366 if (UseZGC) { 2367 ZBarrierSetC2::find_dominating_barriers(igvn); 2368 } 2369 #endif 2370 2371 if (failing()) return; 2372 2373 // Ensure that major progress is now clear 2374 C->clear_major_progress(); 2375 2376 { 2377 // Verify that all previous optimizations produced a valid graph 2378 // at least to this point, even if no loop optimizations were done. 2379 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2380 PhaseIdealLoop::verify(igvn); 2381 } 2382 2383 if (range_check_cast_count() > 0) { 2384 // No more loop optimizations. Remove all range check dependent CastIINodes. 2385 C->remove_range_check_casts(igvn); 2386 igvn.optimize(); 2387 } 2388 2389 #ifdef ASSERT 2390 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2391 bs->verify_gc_barriers(false); 2392 #endif 2393 2394 { 2395 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2396 PhaseMacroExpand mex(igvn); 2397 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2398 if (mex.expand_macro_nodes()) { 2399 assert(failing(), "must bail out w/ explicit message"); 2400 return; 2401 } 2402 } 2403 2404 if (opaque4_count() > 0) { 2405 C->remove_opaque4_nodes(igvn); 2406 igvn.optimize(); 2407 } 2408 2409 DEBUG_ONLY( _modified_nodes = NULL; ) 2410 } // (End scope of igvn; run destructor if necessary for asserts.) 2411 2412 process_print_inlining(); 2413 // A method with only infinite loops has no edges entering loops from root 2414 { 2415 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2416 if (final_graph_reshaping()) { 2417 assert(failing(), "must bail out w/ explicit message"); 2418 return; 2419 } 2420 } 2421 2422 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2423 } 2424 2425 2426 //------------------------------Code_Gen--------------------------------------- 2427 // Given a graph, generate code for it 2428 void Compile::Code_Gen() { 2429 if (failing()) { 2430 return; 2431 } 2432 2433 // Perform instruction selection. You might think we could reclaim Matcher 2434 // memory PDQ, but actually the Matcher is used in generating spill code. 2435 // Internals of the Matcher (including some VectorSets) must remain live 2436 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2437 // set a bit in reclaimed memory. 2438 2439 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2440 // nodes. Mapping is only valid at the root of each matched subtree. 2441 NOT_PRODUCT( verify_graph_edges(); ) 2442 2443 Matcher matcher; 2444 _matcher = &matcher; 2445 { 2446 TracePhase tp("matcher", &timers[_t_matcher]); 2447 matcher.match(); 2448 } 2449 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2450 // nodes. Mapping is only valid at the root of each matched subtree. 2451 NOT_PRODUCT( verify_graph_edges(); ) 2452 2453 // If you have too many nodes, or if matching has failed, bail out 2454 check_node_count(0, "out of nodes matching instructions"); 2455 if (failing()) { 2456 return; 2457 } 2458 2459 print_method(PHASE_MATCHING, 2); 2460 2461 // Build a proper-looking CFG 2462 PhaseCFG cfg(node_arena(), root(), matcher); 2463 _cfg = &cfg; 2464 { 2465 TracePhase tp("scheduler", &timers[_t_scheduler]); 2466 bool success = cfg.do_global_code_motion(); 2467 if (!success) { 2468 return; 2469 } 2470 2471 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2472 NOT_PRODUCT( verify_graph_edges(); ) 2473 debug_only( cfg.verify(); ) 2474 } 2475 2476 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2477 _regalloc = ®alloc; 2478 { 2479 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2480 // Perform register allocation. After Chaitin, use-def chains are 2481 // no longer accurate (at spill code) and so must be ignored. 2482 // Node->LRG->reg mappings are still accurate. 2483 _regalloc->Register_Allocate(); 2484 2485 // Bail out if the allocator builds too many nodes 2486 if (failing()) { 2487 return; 2488 } 2489 } 2490 2491 // Prior to register allocation we kept empty basic blocks in case the 2492 // the allocator needed a place to spill. After register allocation we 2493 // are not adding any new instructions. If any basic block is empty, we 2494 // can now safely remove it. 2495 { 2496 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2497 cfg.remove_empty_blocks(); 2498 if (do_freq_based_layout()) { 2499 PhaseBlockLayout layout(cfg); 2500 } else { 2501 cfg.set_loop_alignment(); 2502 } 2503 cfg.fixup_flow(); 2504 } 2505 2506 // Apply peephole optimizations 2507 if( OptoPeephole ) { 2508 TracePhase tp("peephole", &timers[_t_peephole]); 2509 PhasePeephole peep( _regalloc, cfg); 2510 peep.do_transform(); 2511 } 2512 2513 // Do late expand if CPU requires this. 2514 if (Matcher::require_postalloc_expand) { 2515 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2516 cfg.postalloc_expand(_regalloc); 2517 } 2518 2519 // Convert Nodes to instruction bits in a buffer 2520 { 2521 TraceTime tp("output", &timers[_t_output], CITime); 2522 Output(); 2523 } 2524 2525 print_method(PHASE_FINAL_CODE); 2526 2527 // He's dead, Jim. 2528 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2529 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2530 } 2531 2532 2533 //------------------------------dump_asm--------------------------------------- 2534 // Dump formatted assembly 2535 #ifndef PRODUCT 2536 void Compile::dump_asm(int *pcs, uint pc_limit) { 2537 bool cut_short = false; 2538 tty->print_cr("#"); 2539 tty->print("# "); _tf->dump(); tty->cr(); 2540 tty->print_cr("#"); 2541 2542 // For all blocks 2543 int pc = 0x0; // Program counter 2544 char starts_bundle = ' '; 2545 _regalloc->dump_frame(); 2546 2547 Node *n = NULL; 2548 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2549 if (VMThread::should_terminate()) { 2550 cut_short = true; 2551 break; 2552 } 2553 Block* block = _cfg->get_block(i); 2554 if (block->is_connector() && !Verbose) { 2555 continue; 2556 } 2557 n = block->head(); 2558 if (pcs && n->_idx < pc_limit) { 2559 tty->print("%3.3x ", pcs[n->_idx]); 2560 } else { 2561 tty->print(" "); 2562 } 2563 block->dump_head(_cfg); 2564 if (block->is_connector()) { 2565 tty->print_cr(" # Empty connector block"); 2566 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2567 tty->print_cr(" # Block is sole successor of call"); 2568 } 2569 2570 // For all instructions 2571 Node *delay = NULL; 2572 for (uint j = 0; j < block->number_of_nodes(); j++) { 2573 if (VMThread::should_terminate()) { 2574 cut_short = true; 2575 break; 2576 } 2577 n = block->get_node(j); 2578 if (valid_bundle_info(n)) { 2579 Bundle* bundle = node_bundling(n); 2580 if (bundle->used_in_unconditional_delay()) { 2581 delay = n; 2582 continue; 2583 } 2584 if (bundle->starts_bundle()) { 2585 starts_bundle = '+'; 2586 } 2587 } 2588 2589 if (WizardMode) { 2590 n->dump(); 2591 } 2592 2593 if( !n->is_Region() && // Dont print in the Assembly 2594 !n->is_Phi() && // a few noisely useless nodes 2595 !n->is_Proj() && 2596 !n->is_MachTemp() && 2597 !n->is_SafePointScalarObject() && 2598 !n->is_Catch() && // Would be nice to print exception table targets 2599 !n->is_MergeMem() && // Not very interesting 2600 !n->is_top() && // Debug info table constants 2601 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2602 ) { 2603 if (pcs && n->_idx < pc_limit) 2604 tty->print("%3.3x", pcs[n->_idx]); 2605 else 2606 tty->print(" "); 2607 tty->print(" %c ", starts_bundle); 2608 starts_bundle = ' '; 2609 tty->print("\t"); 2610 n->format(_regalloc, tty); 2611 tty->cr(); 2612 } 2613 2614 // If we have an instruction with a delay slot, and have seen a delay, 2615 // then back up and print it 2616 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2617 assert(delay != NULL, "no unconditional delay instruction"); 2618 if (WizardMode) delay->dump(); 2619 2620 if (node_bundling(delay)->starts_bundle()) 2621 starts_bundle = '+'; 2622 if (pcs && n->_idx < pc_limit) 2623 tty->print("%3.3x", pcs[n->_idx]); 2624 else 2625 tty->print(" "); 2626 tty->print(" %c ", starts_bundle); 2627 starts_bundle = ' '; 2628 tty->print("\t"); 2629 delay->format(_regalloc, tty); 2630 tty->cr(); 2631 delay = NULL; 2632 } 2633 2634 // Dump the exception table as well 2635 if( n->is_Catch() && (Verbose || WizardMode) ) { 2636 // Print the exception table for this offset 2637 _handler_table.print_subtable_for(pc); 2638 } 2639 } 2640 2641 if (pcs && n->_idx < pc_limit) 2642 tty->print_cr("%3.3x", pcs[n->_idx]); 2643 else 2644 tty->cr(); 2645 2646 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2647 2648 } // End of per-block dump 2649 tty->cr(); 2650 2651 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2652 } 2653 #endif 2654 2655 //------------------------------Final_Reshape_Counts--------------------------- 2656 // This class defines counters to help identify when a method 2657 // may/must be executed using hardware with only 24-bit precision. 2658 struct Final_Reshape_Counts : public StackObj { 2659 int _call_count; // count non-inlined 'common' calls 2660 int _float_count; // count float ops requiring 24-bit precision 2661 int _double_count; // count double ops requiring more precision 2662 int _java_call_count; // count non-inlined 'java' calls 2663 int _inner_loop_count; // count loops which need alignment 2664 VectorSet _visited; // Visitation flags 2665 Node_List _tests; // Set of IfNodes & PCTableNodes 2666 2667 Final_Reshape_Counts() : 2668 _call_count(0), _float_count(0), _double_count(0), 2669 _java_call_count(0), _inner_loop_count(0), 2670 _visited( Thread::current()->resource_area() ) { } 2671 2672 void inc_call_count () { _call_count ++; } 2673 void inc_float_count () { _float_count ++; } 2674 void inc_double_count() { _double_count++; } 2675 void inc_java_call_count() { _java_call_count++; } 2676 void inc_inner_loop_count() { _inner_loop_count++; } 2677 2678 int get_call_count () const { return _call_count ; } 2679 int get_float_count () const { return _float_count ; } 2680 int get_double_count() const { return _double_count; } 2681 int get_java_call_count() const { return _java_call_count; } 2682 int get_inner_loop_count() const { return _inner_loop_count; } 2683 }; 2684 2685 #ifdef ASSERT 2686 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2687 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2688 // Make sure the offset goes inside the instance layout. 2689 return k->contains_field_offset(tp->offset()); 2690 // Note that OffsetBot and OffsetTop are very negative. 2691 } 2692 #endif 2693 2694 // Eliminate trivially redundant StoreCMs and accumulate their 2695 // precedence edges. 2696 void Compile::eliminate_redundant_card_marks(Node* n) { 2697 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2698 if (n->in(MemNode::Address)->outcnt() > 1) { 2699 // There are multiple users of the same address so it might be 2700 // possible to eliminate some of the StoreCMs 2701 Node* mem = n->in(MemNode::Memory); 2702 Node* adr = n->in(MemNode::Address); 2703 Node* val = n->in(MemNode::ValueIn); 2704 Node* prev = n; 2705 bool done = false; 2706 // Walk the chain of StoreCMs eliminating ones that match. As 2707 // long as it's a chain of single users then the optimization is 2708 // safe. Eliminating partially redundant StoreCMs would require 2709 // cloning copies down the other paths. 2710 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2711 if (adr == mem->in(MemNode::Address) && 2712 val == mem->in(MemNode::ValueIn)) { 2713 // redundant StoreCM 2714 if (mem->req() > MemNode::OopStore) { 2715 // Hasn't been processed by this code yet. 2716 n->add_prec(mem->in(MemNode::OopStore)); 2717 } else { 2718 // Already converted to precedence edge 2719 for (uint i = mem->req(); i < mem->len(); i++) { 2720 // Accumulate any precedence edges 2721 if (mem->in(i) != NULL) { 2722 n->add_prec(mem->in(i)); 2723 } 2724 } 2725 // Everything above this point has been processed. 2726 done = true; 2727 } 2728 // Eliminate the previous StoreCM 2729 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2730 assert(mem->outcnt() == 0, "should be dead"); 2731 mem->disconnect_inputs(NULL, this); 2732 } else { 2733 prev = mem; 2734 } 2735 mem = prev->in(MemNode::Memory); 2736 } 2737 } 2738 } 2739 2740 //------------------------------final_graph_reshaping_impl---------------------- 2741 // Implement items 1-5 from final_graph_reshaping below. 2742 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2743 2744 if ( n->outcnt() == 0 ) return; // dead node 2745 uint nop = n->Opcode(); 2746 2747 // Check for 2-input instruction with "last use" on right input. 2748 // Swap to left input. Implements item (2). 2749 if( n->req() == 3 && // two-input instruction 2750 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2751 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2752 n->in(2)->outcnt() == 1 &&// right use IS a last use 2753 !n->in(2)->is_Con() ) { // right use is not a constant 2754 // Check for commutative opcode 2755 switch( nop ) { 2756 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2757 case Op_MaxI: case Op_MinI: 2758 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2759 case Op_AndL: case Op_XorL: case Op_OrL: 2760 case Op_AndI: case Op_XorI: case Op_OrI: { 2761 // Move "last use" input to left by swapping inputs 2762 n->swap_edges(1, 2); 2763 break; 2764 } 2765 default: 2766 break; 2767 } 2768 } 2769 2770 #ifdef ASSERT 2771 if( n->is_Mem() ) { 2772 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2773 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2774 // oop will be recorded in oop map if load crosses safepoint 2775 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2776 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2777 "raw memory operations should have control edge"); 2778 } 2779 if (n->is_MemBar()) { 2780 MemBarNode* mb = n->as_MemBar(); 2781 if (mb->trailing_store() || mb->trailing_load_store()) { 2782 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2783 Node* mem = mb->in(MemBarNode::Precedent); 2784 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2785 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2786 } else if (mb->leading()) { 2787 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2788 } 2789 } 2790 #endif 2791 // Count FPU ops and common calls, implements item (3) 2792 switch( nop ) { 2793 // Count all float operations that may use FPU 2794 case Op_AddF: 2795 case Op_SubF: 2796 case Op_MulF: 2797 case Op_DivF: 2798 case Op_NegF: 2799 case Op_ModF: 2800 case Op_ConvI2F: 2801 case Op_ConF: 2802 case Op_CmpF: 2803 case Op_CmpF3: 2804 // case Op_ConvL2F: // longs are split into 32-bit halves 2805 frc.inc_float_count(); 2806 break; 2807 2808 case Op_ConvF2D: 2809 case Op_ConvD2F: 2810 frc.inc_float_count(); 2811 frc.inc_double_count(); 2812 break; 2813 2814 // Count all double operations that may use FPU 2815 case Op_AddD: 2816 case Op_SubD: 2817 case Op_MulD: 2818 case Op_DivD: 2819 case Op_NegD: 2820 case Op_ModD: 2821 case Op_ConvI2D: 2822 case Op_ConvD2I: 2823 // case Op_ConvL2D: // handled by leaf call 2824 // case Op_ConvD2L: // handled by leaf call 2825 case Op_ConD: 2826 case Op_CmpD: 2827 case Op_CmpD3: 2828 frc.inc_double_count(); 2829 break; 2830 case Op_Opaque1: // Remove Opaque Nodes before matching 2831 case Op_Opaque2: // Remove Opaque Nodes before matching 2832 case Op_Opaque3: 2833 n->subsume_by(n->in(1), this); 2834 break; 2835 case Op_CallStaticJava: 2836 case Op_CallJava: 2837 case Op_CallDynamicJava: 2838 frc.inc_java_call_count(); // Count java call site; 2839 case Op_CallRuntime: 2840 case Op_CallLeaf: 2841 case Op_CallLeafNoFP: { 2842 assert (n->is_Call(), ""); 2843 CallNode *call = n->as_Call(); 2844 // Count call sites where the FP mode bit would have to be flipped. 2845 // Do not count uncommon runtime calls: 2846 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2847 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2848 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2849 frc.inc_call_count(); // Count the call site 2850 } else { // See if uncommon argument is shared 2851 Node *n = call->in(TypeFunc::Parms); 2852 int nop = n->Opcode(); 2853 // Clone shared simple arguments to uncommon calls, item (1). 2854 if (n->outcnt() > 1 && 2855 !n->is_Proj() && 2856 nop != Op_CreateEx && 2857 nop != Op_CheckCastPP && 2858 nop != Op_DecodeN && 2859 nop != Op_DecodeNKlass && 2860 !n->is_Mem() && 2861 !n->is_Phi()) { 2862 Node *x = n->clone(); 2863 call->set_req(TypeFunc::Parms, x); 2864 } 2865 } 2866 break; 2867 } 2868 2869 case Op_StoreD: 2870 case Op_LoadD: 2871 case Op_LoadD_unaligned: 2872 frc.inc_double_count(); 2873 goto handle_mem; 2874 case Op_StoreF: 2875 case Op_LoadF: 2876 frc.inc_float_count(); 2877 goto handle_mem; 2878 2879 case Op_StoreCM: 2880 { 2881 // Convert OopStore dependence into precedence edge 2882 Node* prec = n->in(MemNode::OopStore); 2883 n->del_req(MemNode::OopStore); 2884 n->add_prec(prec); 2885 eliminate_redundant_card_marks(n); 2886 } 2887 2888 // fall through 2889 2890 case Op_StoreB: 2891 case Op_StoreC: 2892 case Op_StorePConditional: 2893 case Op_StoreI: 2894 case Op_StoreL: 2895 case Op_StoreIConditional: 2896 case Op_StoreLConditional: 2897 case Op_CompareAndSwapB: 2898 case Op_CompareAndSwapS: 2899 case Op_CompareAndSwapI: 2900 case Op_CompareAndSwapL: 2901 case Op_CompareAndSwapP: 2902 case Op_CompareAndSwapN: 2903 case Op_WeakCompareAndSwapB: 2904 case Op_WeakCompareAndSwapS: 2905 case Op_WeakCompareAndSwapI: 2906 case Op_WeakCompareAndSwapL: 2907 case Op_WeakCompareAndSwapP: 2908 case Op_WeakCompareAndSwapN: 2909 case Op_CompareAndExchangeB: 2910 case Op_CompareAndExchangeS: 2911 case Op_CompareAndExchangeI: 2912 case Op_CompareAndExchangeL: 2913 case Op_CompareAndExchangeP: 2914 case Op_CompareAndExchangeN: 2915 case Op_GetAndAddS: 2916 case Op_GetAndAddB: 2917 case Op_GetAndAddI: 2918 case Op_GetAndAddL: 2919 case Op_GetAndSetS: 2920 case Op_GetAndSetB: 2921 case Op_GetAndSetI: 2922 case Op_GetAndSetL: 2923 case Op_GetAndSetP: 2924 case Op_GetAndSetN: 2925 case Op_StoreP: 2926 case Op_StoreN: 2927 case Op_StoreNKlass: 2928 case Op_LoadB: 2929 case Op_LoadUB: 2930 case Op_LoadUS: 2931 case Op_LoadI: 2932 case Op_LoadKlass: 2933 case Op_LoadNKlass: 2934 case Op_LoadL: 2935 case Op_LoadL_unaligned: 2936 case Op_LoadPLocked: 2937 case Op_LoadP: 2938 #if INCLUDE_ZGC 2939 case Op_LoadBarrierSlowReg: 2940 case Op_LoadBarrierWeakSlowReg: 2941 #endif 2942 case Op_LoadN: 2943 case Op_LoadRange: 2944 case Op_LoadS: { 2945 handle_mem: 2946 #ifdef ASSERT 2947 if( VerifyOptoOopOffsets ) { 2948 assert( n->is_Mem(), "" ); 2949 MemNode *mem = (MemNode*)n; 2950 // Check to see if address types have grounded out somehow. 2951 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2952 assert( !tp || oop_offset_is_sane(tp), "" ); 2953 } 2954 #endif 2955 break; 2956 } 2957 2958 case Op_AddP: { // Assert sane base pointers 2959 Node *addp = n->in(AddPNode::Address); 2960 assert( !addp->is_AddP() || 2961 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2962 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2963 "Base pointers must match (addp %u)", addp->_idx ); 2964 #ifdef _LP64 2965 if ((UseCompressedOops || UseCompressedClassPointers) && 2966 addp->Opcode() == Op_ConP && 2967 addp == n->in(AddPNode::Base) && 2968 n->in(AddPNode::Offset)->is_Con()) { 2969 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2970 // on the platform and on the compressed oops mode. 2971 // Use addressing with narrow klass to load with offset on x86. 2972 // Some platforms can use the constant pool to load ConP. 2973 // Do this transformation here since IGVN will convert ConN back to ConP. 2974 const Type* t = addp->bottom_type(); 2975 bool is_oop = t->isa_oopptr() != NULL; 2976 bool is_klass = t->isa_klassptr() != NULL; 2977 2978 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 2979 (is_klass && Matcher::const_klass_prefer_decode())) { 2980 Node* nn = NULL; 2981 2982 int op = is_oop ? Op_ConN : Op_ConNKlass; 2983 2984 // Look for existing ConN node of the same exact type. 2985 Node* r = root(); 2986 uint cnt = r->outcnt(); 2987 for (uint i = 0; i < cnt; i++) { 2988 Node* m = r->raw_out(i); 2989 if (m!= NULL && m->Opcode() == op && 2990 m->bottom_type()->make_ptr() == t) { 2991 nn = m; 2992 break; 2993 } 2994 } 2995 if (nn != NULL) { 2996 // Decode a narrow oop to match address 2997 // [R12 + narrow_oop_reg<<3 + offset] 2998 if (is_oop) { 2999 nn = new DecodeNNode(nn, t); 3000 } else { 3001 nn = new DecodeNKlassNode(nn, t); 3002 } 3003 // Check for succeeding AddP which uses the same Base. 3004 // Otherwise we will run into the assertion above when visiting that guy. 3005 for (uint i = 0; i < n->outcnt(); ++i) { 3006 Node *out_i = n->raw_out(i); 3007 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3008 out_i->set_req(AddPNode::Base, nn); 3009 #ifdef ASSERT 3010 for (uint j = 0; j < out_i->outcnt(); ++j) { 3011 Node *out_j = out_i->raw_out(j); 3012 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3013 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3014 } 3015 #endif 3016 } 3017 } 3018 n->set_req(AddPNode::Base, nn); 3019 n->set_req(AddPNode::Address, nn); 3020 if (addp->outcnt() == 0) { 3021 addp->disconnect_inputs(NULL, this); 3022 } 3023 } 3024 } 3025 } 3026 #endif 3027 // platform dependent reshaping of the address expression 3028 reshape_address(n->as_AddP()); 3029 break; 3030 } 3031 3032 case Op_CastPP: { 3033 // Remove CastPP nodes to gain more freedom during scheduling but 3034 // keep the dependency they encode as control or precedence edges 3035 // (if control is set already) on memory operations. Some CastPP 3036 // nodes don't have a control (don't carry a dependency): skip 3037 // those. 3038 if (n->in(0) != NULL) { 3039 ResourceMark rm; 3040 Unique_Node_List wq; 3041 wq.push(n); 3042 for (uint next = 0; next < wq.size(); ++next) { 3043 Node *m = wq.at(next); 3044 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3045 Node* use = m->fast_out(i); 3046 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3047 use->ensure_control_or_add_prec(n->in(0)); 3048 } else { 3049 switch(use->Opcode()) { 3050 case Op_AddP: 3051 case Op_DecodeN: 3052 case Op_DecodeNKlass: 3053 case Op_CheckCastPP: 3054 case Op_CastPP: 3055 wq.push(use); 3056 break; 3057 } 3058 } 3059 } 3060 } 3061 } 3062 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3063 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3064 Node* in1 = n->in(1); 3065 const Type* t = n->bottom_type(); 3066 Node* new_in1 = in1->clone(); 3067 new_in1->as_DecodeN()->set_type(t); 3068 3069 if (!Matcher::narrow_oop_use_complex_address()) { 3070 // 3071 // x86, ARM and friends can handle 2 adds in addressing mode 3072 // and Matcher can fold a DecodeN node into address by using 3073 // a narrow oop directly and do implicit NULL check in address: 3074 // 3075 // [R12 + narrow_oop_reg<<3 + offset] 3076 // NullCheck narrow_oop_reg 3077 // 3078 // On other platforms (Sparc) we have to keep new DecodeN node and 3079 // use it to do implicit NULL check in address: 3080 // 3081 // decode_not_null narrow_oop_reg, base_reg 3082 // [base_reg + offset] 3083 // NullCheck base_reg 3084 // 3085 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3086 // to keep the information to which NULL check the new DecodeN node 3087 // corresponds to use it as value in implicit_null_check(). 3088 // 3089 new_in1->set_req(0, n->in(0)); 3090 } 3091 3092 n->subsume_by(new_in1, this); 3093 if (in1->outcnt() == 0) { 3094 in1->disconnect_inputs(NULL, this); 3095 } 3096 } else { 3097 n->subsume_by(n->in(1), this); 3098 if (n->outcnt() == 0) { 3099 n->disconnect_inputs(NULL, this); 3100 } 3101 } 3102 break; 3103 } 3104 #ifdef _LP64 3105 case Op_CmpP: 3106 // Do this transformation here to preserve CmpPNode::sub() and 3107 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3108 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3109 Node* in1 = n->in(1); 3110 Node* in2 = n->in(2); 3111 if (!in1->is_DecodeNarrowPtr()) { 3112 in2 = in1; 3113 in1 = n->in(2); 3114 } 3115 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3116 3117 Node* new_in2 = NULL; 3118 if (in2->is_DecodeNarrowPtr()) { 3119 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3120 new_in2 = in2->in(1); 3121 } else if (in2->Opcode() == Op_ConP) { 3122 const Type* t = in2->bottom_type(); 3123 if (t == TypePtr::NULL_PTR) { 3124 assert(in1->is_DecodeN(), "compare klass to null?"); 3125 // Don't convert CmpP null check into CmpN if compressed 3126 // oops implicit null check is not generated. 3127 // This will allow to generate normal oop implicit null check. 3128 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3129 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3130 // 3131 // This transformation together with CastPP transformation above 3132 // will generated code for implicit NULL checks for compressed oops. 3133 // 3134 // The original code after Optimize() 3135 // 3136 // LoadN memory, narrow_oop_reg 3137 // decode narrow_oop_reg, base_reg 3138 // CmpP base_reg, NULL 3139 // CastPP base_reg // NotNull 3140 // Load [base_reg + offset], val_reg 3141 // 3142 // after these transformations will be 3143 // 3144 // LoadN memory, narrow_oop_reg 3145 // CmpN narrow_oop_reg, NULL 3146 // decode_not_null narrow_oop_reg, base_reg 3147 // Load [base_reg + offset], val_reg 3148 // 3149 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3150 // since narrow oops can be used in debug info now (see the code in 3151 // final_graph_reshaping_walk()). 3152 // 3153 // At the end the code will be matched to 3154 // on x86: 3155 // 3156 // Load_narrow_oop memory, narrow_oop_reg 3157 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3158 // NullCheck narrow_oop_reg 3159 // 3160 // and on sparc: 3161 // 3162 // Load_narrow_oop memory, narrow_oop_reg 3163 // decode_not_null narrow_oop_reg, base_reg 3164 // Load [base_reg + offset], val_reg 3165 // NullCheck base_reg 3166 // 3167 } else if (t->isa_oopptr()) { 3168 new_in2 = ConNode::make(t->make_narrowoop()); 3169 } else if (t->isa_klassptr()) { 3170 new_in2 = ConNode::make(t->make_narrowklass()); 3171 } 3172 } 3173 if (new_in2 != NULL) { 3174 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3175 n->subsume_by(cmpN, this); 3176 if (in1->outcnt() == 0) { 3177 in1->disconnect_inputs(NULL, this); 3178 } 3179 if (in2->outcnt() == 0) { 3180 in2->disconnect_inputs(NULL, this); 3181 } 3182 } 3183 } 3184 break; 3185 3186 case Op_DecodeN: 3187 case Op_DecodeNKlass: 3188 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3189 // DecodeN could be pinned when it can't be fold into 3190 // an address expression, see the code for Op_CastPP above. 3191 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3192 break; 3193 3194 case Op_EncodeP: 3195 case Op_EncodePKlass: { 3196 Node* in1 = n->in(1); 3197 if (in1->is_DecodeNarrowPtr()) { 3198 n->subsume_by(in1->in(1), this); 3199 } else if (in1->Opcode() == Op_ConP) { 3200 const Type* t = in1->bottom_type(); 3201 if (t == TypePtr::NULL_PTR) { 3202 assert(t->isa_oopptr(), "null klass?"); 3203 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3204 } else if (t->isa_oopptr()) { 3205 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3206 } else if (t->isa_klassptr()) { 3207 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3208 } 3209 } 3210 if (in1->outcnt() == 0) { 3211 in1->disconnect_inputs(NULL, this); 3212 } 3213 break; 3214 } 3215 3216 case Op_Proj: { 3217 if (OptimizeStringConcat) { 3218 ProjNode* p = n->as_Proj(); 3219 if (p->_is_io_use) { 3220 // Separate projections were used for the exception path which 3221 // are normally removed by a late inline. If it wasn't inlined 3222 // then they will hang around and should just be replaced with 3223 // the original one. 3224 Node* proj = NULL; 3225 // Replace with just one 3226 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3227 Node *use = i.get(); 3228 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3229 proj = use; 3230 break; 3231 } 3232 } 3233 assert(proj != NULL, "must be found"); 3234 p->subsume_by(proj, this); 3235 } 3236 } 3237 break; 3238 } 3239 3240 case Op_Phi: 3241 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3242 // The EncodeP optimization may create Phi with the same edges 3243 // for all paths. It is not handled well by Register Allocator. 3244 Node* unique_in = n->in(1); 3245 assert(unique_in != NULL, ""); 3246 uint cnt = n->req(); 3247 for (uint i = 2; i < cnt; i++) { 3248 Node* m = n->in(i); 3249 assert(m != NULL, ""); 3250 if (unique_in != m) 3251 unique_in = NULL; 3252 } 3253 if (unique_in != NULL) { 3254 n->subsume_by(unique_in, this); 3255 } 3256 } 3257 break; 3258 3259 #endif 3260 3261 #ifdef ASSERT 3262 case Op_CastII: 3263 // Verify that all range check dependent CastII nodes were removed. 3264 if (n->isa_CastII()->has_range_check()) { 3265 n->dump(3); 3266 assert(false, "Range check dependent CastII node was not removed"); 3267 } 3268 break; 3269 #endif 3270 3271 case Op_ModI: 3272 if (UseDivMod) { 3273 // Check if a%b and a/b both exist 3274 Node* d = n->find_similar(Op_DivI); 3275 if (d) { 3276 // Replace them with a fused divmod if supported 3277 if (Matcher::has_match_rule(Op_DivModI)) { 3278 DivModINode* divmod = DivModINode::make(n); 3279 d->subsume_by(divmod->div_proj(), this); 3280 n->subsume_by(divmod->mod_proj(), this); 3281 } else { 3282 // replace a%b with a-((a/b)*b) 3283 Node* mult = new MulINode(d, d->in(2)); 3284 Node* sub = new SubINode(d->in(1), mult); 3285 n->subsume_by(sub, this); 3286 } 3287 } 3288 } 3289 break; 3290 3291 case Op_ModL: 3292 if (UseDivMod) { 3293 // Check if a%b and a/b both exist 3294 Node* d = n->find_similar(Op_DivL); 3295 if (d) { 3296 // Replace them with a fused divmod if supported 3297 if (Matcher::has_match_rule(Op_DivModL)) { 3298 DivModLNode* divmod = DivModLNode::make(n); 3299 d->subsume_by(divmod->div_proj(), this); 3300 n->subsume_by(divmod->mod_proj(), this); 3301 } else { 3302 // replace a%b with a-((a/b)*b) 3303 Node* mult = new MulLNode(d, d->in(2)); 3304 Node* sub = new SubLNode(d->in(1), mult); 3305 n->subsume_by(sub, this); 3306 } 3307 } 3308 } 3309 break; 3310 3311 case Op_LoadVector: 3312 case Op_StoreVector: 3313 break; 3314 3315 case Op_AddReductionVI: 3316 case Op_AddReductionVL: 3317 case Op_AddReductionVF: 3318 case Op_AddReductionVD: 3319 case Op_MulReductionVI: 3320 case Op_MulReductionVL: 3321 case Op_MulReductionVF: 3322 case Op_MulReductionVD: 3323 break; 3324 3325 case Op_PackB: 3326 case Op_PackS: 3327 case Op_PackI: 3328 case Op_PackF: 3329 case Op_PackL: 3330 case Op_PackD: 3331 if (n->req()-1 > 2) { 3332 // Replace many operand PackNodes with a binary tree for matching 3333 PackNode* p = (PackNode*) n; 3334 Node* btp = p->binary_tree_pack(1, n->req()); 3335 n->subsume_by(btp, this); 3336 } 3337 break; 3338 case Op_Loop: 3339 case Op_CountedLoop: 3340 case Op_OuterStripMinedLoop: 3341 if (n->as_Loop()->is_inner_loop()) { 3342 frc.inc_inner_loop_count(); 3343 } 3344 n->as_Loop()->verify_strip_mined(0); 3345 break; 3346 case Op_LShiftI: 3347 case Op_RShiftI: 3348 case Op_URShiftI: 3349 case Op_LShiftL: 3350 case Op_RShiftL: 3351 case Op_URShiftL: 3352 if (Matcher::need_masked_shift_count) { 3353 // The cpu's shift instructions don't restrict the count to the 3354 // lower 5/6 bits. We need to do the masking ourselves. 3355 Node* in2 = n->in(2); 3356 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3357 const TypeInt* t = in2->find_int_type(); 3358 if (t != NULL && t->is_con()) { 3359 juint shift = t->get_con(); 3360 if (shift > mask) { // Unsigned cmp 3361 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3362 } 3363 } else { 3364 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3365 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3366 n->set_req(2, shift); 3367 } 3368 } 3369 if (in2->outcnt() == 0) { // Remove dead node 3370 in2->disconnect_inputs(NULL, this); 3371 } 3372 } 3373 break; 3374 case Op_MemBarStoreStore: 3375 case Op_MemBarRelease: 3376 // Break the link with AllocateNode: it is no longer useful and 3377 // confuses register allocation. 3378 if (n->req() > MemBarNode::Precedent) { 3379 n->set_req(MemBarNode::Precedent, top()); 3380 } 3381 break; 3382 case Op_MemBarAcquire: { 3383 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3384 // At parse time, the trailing MemBarAcquire for a volatile load 3385 // is created with an edge to the load. After optimizations, 3386 // that input may be a chain of Phis. If those phis have no 3387 // other use, then the MemBarAcquire keeps them alive and 3388 // register allocation can be confused. 3389 ResourceMark rm; 3390 Unique_Node_List wq; 3391 wq.push(n->in(MemBarNode::Precedent)); 3392 n->set_req(MemBarNode::Precedent, top()); 3393 while (wq.size() > 0) { 3394 Node* m = wq.pop(); 3395 if (m->outcnt() == 0) { 3396 for (uint j = 0; j < m->req(); j++) { 3397 Node* in = m->in(j); 3398 if (in != NULL) { 3399 wq.push(in); 3400 } 3401 } 3402 m->disconnect_inputs(NULL, this); 3403 } 3404 } 3405 } 3406 break; 3407 } 3408 case Op_RangeCheck: { 3409 RangeCheckNode* rc = n->as_RangeCheck(); 3410 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3411 n->subsume_by(iff, this); 3412 frc._tests.push(iff); 3413 break; 3414 } 3415 case Op_ConvI2L: { 3416 if (!Matcher::convi2l_type_required) { 3417 // Code generation on some platforms doesn't need accurate 3418 // ConvI2L types. Widening the type can help remove redundant 3419 // address computations. 3420 n->as_Type()->set_type(TypeLong::INT); 3421 ResourceMark rm; 3422 Node_List wq; 3423 wq.push(n); 3424 for (uint next = 0; next < wq.size(); next++) { 3425 Node *m = wq.at(next); 3426 3427 for(;;) { 3428 // Loop over all nodes with identical inputs edges as m 3429 Node* k = m->find_similar(m->Opcode()); 3430 if (k == NULL) { 3431 break; 3432 } 3433 // Push their uses so we get a chance to remove node made 3434 // redundant 3435 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3436 Node* u = k->fast_out(i); 3437 assert(!wq.contains(u), "shouldn't process one node several times"); 3438 if (u->Opcode() == Op_LShiftL || 3439 u->Opcode() == Op_AddL || 3440 u->Opcode() == Op_SubL || 3441 u->Opcode() == Op_AddP) { 3442 wq.push(u); 3443 } 3444 } 3445 // Replace all nodes with identical edges as m with m 3446 k->subsume_by(m, this); 3447 } 3448 } 3449 } 3450 break; 3451 } 3452 case Op_CmpUL: { 3453 if (!Matcher::has_match_rule(Op_CmpUL)) { 3454 // We don't support unsigned long comparisons. Set 'max_idx_expr' 3455 // to max_julong if < 0 to make the signed comparison fail. 3456 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3457 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3458 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3459 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3460 Node* andl = new AndLNode(orl, remove_sign_mask); 3461 Node* cmp = new CmpLNode(andl, n->in(2)); 3462 n->subsume_by(cmp, this); 3463 } 3464 break; 3465 } 3466 default: 3467 assert( !n->is_Call(), "" ); 3468 assert( !n->is_Mem(), "" ); 3469 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3470 break; 3471 } 3472 3473 // Collect CFG split points 3474 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3475 frc._tests.push(n); 3476 } 3477 } 3478 3479 //------------------------------final_graph_reshaping_walk--------------------- 3480 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3481 // requires that the walk visits a node's inputs before visiting the node. 3482 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3483 ResourceArea *area = Thread::current()->resource_area(); 3484 Unique_Node_List sfpt(area); 3485 3486 frc._visited.set(root->_idx); // first, mark node as visited 3487 uint cnt = root->req(); 3488 Node *n = root; 3489 uint i = 0; 3490 while (true) { 3491 if (i < cnt) { 3492 // Place all non-visited non-null inputs onto stack 3493 Node* m = n->in(i); 3494 ++i; 3495 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3496 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3497 // compute worst case interpreter size in case of a deoptimization 3498 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3499 3500 sfpt.push(m); 3501 } 3502 cnt = m->req(); 3503 nstack.push(n, i); // put on stack parent and next input's index 3504 n = m; 3505 i = 0; 3506 } 3507 } else { 3508 // Now do post-visit work 3509 final_graph_reshaping_impl( n, frc ); 3510 if (nstack.is_empty()) 3511 break; // finished 3512 n = nstack.node(); // Get node from stack 3513 cnt = n->req(); 3514 i = nstack.index(); 3515 nstack.pop(); // Shift to the next node on stack 3516 } 3517 } 3518 3519 // Skip next transformation if compressed oops are not used. 3520 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3521 (!UseCompressedOops && !UseCompressedClassPointers)) 3522 return; 3523 3524 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3525 // It could be done for an uncommon traps or any safepoints/calls 3526 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3527 while (sfpt.size() > 0) { 3528 n = sfpt.pop(); 3529 JVMState *jvms = n->as_SafePoint()->jvms(); 3530 assert(jvms != NULL, "sanity"); 3531 int start = jvms->debug_start(); 3532 int end = n->req(); 3533 bool is_uncommon = (n->is_CallStaticJava() && 3534 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3535 for (int j = start; j < end; j++) { 3536 Node* in = n->in(j); 3537 if (in->is_DecodeNarrowPtr()) { 3538 bool safe_to_skip = true; 3539 if (!is_uncommon ) { 3540 // Is it safe to skip? 3541 for (uint i = 0; i < in->outcnt(); i++) { 3542 Node* u = in->raw_out(i); 3543 if (!u->is_SafePoint() || 3544 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3545 safe_to_skip = false; 3546 } 3547 } 3548 } 3549 if (safe_to_skip) { 3550 n->set_req(j, in->in(1)); 3551 } 3552 if (in->outcnt() == 0) { 3553 in->disconnect_inputs(NULL, this); 3554 } 3555 } 3556 } 3557 } 3558 } 3559 3560 //------------------------------final_graph_reshaping-------------------------- 3561 // Final Graph Reshaping. 3562 // 3563 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3564 // and not commoned up and forced early. Must come after regular 3565 // optimizations to avoid GVN undoing the cloning. Clone constant 3566 // inputs to Loop Phis; these will be split by the allocator anyways. 3567 // Remove Opaque nodes. 3568 // (2) Move last-uses by commutative operations to the left input to encourage 3569 // Intel update-in-place two-address operations and better register usage 3570 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3571 // calls canonicalizing them back. 3572 // (3) Count the number of double-precision FP ops, single-precision FP ops 3573 // and call sites. On Intel, we can get correct rounding either by 3574 // forcing singles to memory (requires extra stores and loads after each 3575 // FP bytecode) or we can set a rounding mode bit (requires setting and 3576 // clearing the mode bit around call sites). The mode bit is only used 3577 // if the relative frequency of single FP ops to calls is low enough. 3578 // This is a key transform for SPEC mpeg_audio. 3579 // (4) Detect infinite loops; blobs of code reachable from above but not 3580 // below. Several of the Code_Gen algorithms fail on such code shapes, 3581 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3582 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3583 // Detection is by looking for IfNodes where only 1 projection is 3584 // reachable from below or CatchNodes missing some targets. 3585 // (5) Assert for insane oop offsets in debug mode. 3586 3587 bool Compile::final_graph_reshaping() { 3588 // an infinite loop may have been eliminated by the optimizer, 3589 // in which case the graph will be empty. 3590 if (root()->req() == 1) { 3591 record_method_not_compilable("trivial infinite loop"); 3592 return true; 3593 } 3594 3595 // Expensive nodes have their control input set to prevent the GVN 3596 // from freely commoning them. There's no GVN beyond this point so 3597 // no need to keep the control input. We want the expensive nodes to 3598 // be freely moved to the least frequent code path by gcm. 3599 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3600 for (int i = 0; i < expensive_count(); i++) { 3601 _expensive_nodes->at(i)->set_req(0, NULL); 3602 } 3603 3604 Final_Reshape_Counts frc; 3605 3606 // Visit everybody reachable! 3607 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3608 Node_Stack nstack(live_nodes() >> 1); 3609 final_graph_reshaping_walk(nstack, root(), frc); 3610 3611 // Check for unreachable (from below) code (i.e., infinite loops). 3612 for( uint i = 0; i < frc._tests.size(); i++ ) { 3613 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3614 // Get number of CFG targets. 3615 // Note that PCTables include exception targets after calls. 3616 uint required_outcnt = n->required_outcnt(); 3617 if (n->outcnt() != required_outcnt) { 3618 // Check for a few special cases. Rethrow Nodes never take the 3619 // 'fall-thru' path, so expected kids is 1 less. 3620 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3621 if (n->in(0)->in(0)->is_Call()) { 3622 CallNode *call = n->in(0)->in(0)->as_Call(); 3623 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3624 required_outcnt--; // Rethrow always has 1 less kid 3625 } else if (call->req() > TypeFunc::Parms && 3626 call->is_CallDynamicJava()) { 3627 // Check for null receiver. In such case, the optimizer has 3628 // detected that the virtual call will always result in a null 3629 // pointer exception. The fall-through projection of this CatchNode 3630 // will not be populated. 3631 Node *arg0 = call->in(TypeFunc::Parms); 3632 if (arg0->is_Type() && 3633 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3634 required_outcnt--; 3635 } 3636 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3637 call->req() > TypeFunc::Parms+1 && 3638 call->is_CallStaticJava()) { 3639 // Check for negative array length. In such case, the optimizer has 3640 // detected that the allocation attempt will always result in an 3641 // exception. There is no fall-through projection of this CatchNode . 3642 Node *arg1 = call->in(TypeFunc::Parms+1); 3643 if (arg1->is_Type() && 3644 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3645 required_outcnt--; 3646 } 3647 } 3648 } 3649 } 3650 // Recheck with a better notion of 'required_outcnt' 3651 if (n->outcnt() != required_outcnt) { 3652 record_method_not_compilable("malformed control flow"); 3653 return true; // Not all targets reachable! 3654 } 3655 } 3656 // Check that I actually visited all kids. Unreached kids 3657 // must be infinite loops. 3658 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3659 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3660 record_method_not_compilable("infinite loop"); 3661 return true; // Found unvisited kid; must be unreach 3662 } 3663 3664 // Here so verification code in final_graph_reshaping_walk() 3665 // always see an OuterStripMinedLoopEnd 3666 if (n->is_OuterStripMinedLoopEnd()) { 3667 IfNode* init_iff = n->as_If(); 3668 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3669 n->subsume_by(iff, this); 3670 } 3671 } 3672 3673 // If original bytecodes contained a mixture of floats and doubles 3674 // check if the optimizer has made it homogenous, item (3). 3675 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3676 frc.get_float_count() > 32 && 3677 frc.get_double_count() == 0 && 3678 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3679 set_24_bit_selection_and_mode( false, true ); 3680 } 3681 3682 set_java_calls(frc.get_java_call_count()); 3683 set_inner_loops(frc.get_inner_loop_count()); 3684 3685 // No infinite loops, no reason to bail out. 3686 return false; 3687 } 3688 3689 //-----------------------------too_many_traps---------------------------------- 3690 // Report if there are too many traps at the current method and bci. 3691 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3692 bool Compile::too_many_traps(ciMethod* method, 3693 int bci, 3694 Deoptimization::DeoptReason reason) { 3695 ciMethodData* md = method->method_data(); 3696 if (md->is_empty()) { 3697 // Assume the trap has not occurred, or that it occurred only 3698 // because of a transient condition during start-up in the interpreter. 3699 return false; 3700 } 3701 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3702 if (md->has_trap_at(bci, m, reason) != 0) { 3703 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3704 // Also, if there are multiple reasons, or if there is no per-BCI record, 3705 // assume the worst. 3706 if (log()) 3707 log()->elem("observe trap='%s' count='%d'", 3708 Deoptimization::trap_reason_name(reason), 3709 md->trap_count(reason)); 3710 return true; 3711 } else { 3712 // Ignore method/bci and see if there have been too many globally. 3713 return too_many_traps(reason, md); 3714 } 3715 } 3716 3717 // Less-accurate variant which does not require a method and bci. 3718 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3719 ciMethodData* logmd) { 3720 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3721 // Too many traps globally. 3722 // Note that we use cumulative trap_count, not just md->trap_count. 3723 if (log()) { 3724 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3725 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3726 Deoptimization::trap_reason_name(reason), 3727 mcount, trap_count(reason)); 3728 } 3729 return true; 3730 } else { 3731 // The coast is clear. 3732 return false; 3733 } 3734 } 3735 3736 //--------------------------too_many_recompiles-------------------------------- 3737 // Report if there are too many recompiles at the current method and bci. 3738 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3739 // Is not eager to return true, since this will cause the compiler to use 3740 // Action_none for a trap point, to avoid too many recompilations. 3741 bool Compile::too_many_recompiles(ciMethod* method, 3742 int bci, 3743 Deoptimization::DeoptReason reason) { 3744 ciMethodData* md = method->method_data(); 3745 if (md->is_empty()) { 3746 // Assume the trap has not occurred, or that it occurred only 3747 // because of a transient condition during start-up in the interpreter. 3748 return false; 3749 } 3750 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3751 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3752 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3753 Deoptimization::DeoptReason per_bc_reason 3754 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3755 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3756 if ((per_bc_reason == Deoptimization::Reason_none 3757 || md->has_trap_at(bci, m, reason) != 0) 3758 // The trap frequency measure we care about is the recompile count: 3759 && md->trap_recompiled_at(bci, m) 3760 && md->overflow_recompile_count() >= bc_cutoff) { 3761 // Do not emit a trap here if it has already caused recompilations. 3762 // Also, if there are multiple reasons, or if there is no per-BCI record, 3763 // assume the worst. 3764 if (log()) 3765 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3766 Deoptimization::trap_reason_name(reason), 3767 md->trap_count(reason), 3768 md->overflow_recompile_count()); 3769 return true; 3770 } else if (trap_count(reason) != 0 3771 && decompile_count() >= m_cutoff) { 3772 // Too many recompiles globally, and we have seen this sort of trap. 3773 // Use cumulative decompile_count, not just md->decompile_count. 3774 if (log()) 3775 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3776 Deoptimization::trap_reason_name(reason), 3777 md->trap_count(reason), trap_count(reason), 3778 md->decompile_count(), decompile_count()); 3779 return true; 3780 } else { 3781 // The coast is clear. 3782 return false; 3783 } 3784 } 3785 3786 // Compute when not to trap. Used by matching trap based nodes and 3787 // NullCheck optimization. 3788 void Compile::set_allowed_deopt_reasons() { 3789 _allowed_reasons = 0; 3790 if (is_method_compilation()) { 3791 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3792 assert(rs < BitsPerInt, "recode bit map"); 3793 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3794 _allowed_reasons |= nth_bit(rs); 3795 } 3796 } 3797 } 3798 } 3799 3800 #ifndef PRODUCT 3801 //------------------------------verify_graph_edges--------------------------- 3802 // Walk the Graph and verify that there is a one-to-one correspondence 3803 // between Use-Def edges and Def-Use edges in the graph. 3804 void Compile::verify_graph_edges(bool no_dead_code) { 3805 if (VerifyGraphEdges) { 3806 ResourceArea *area = Thread::current()->resource_area(); 3807 Unique_Node_List visited(area); 3808 // Call recursive graph walk to check edges 3809 _root->verify_edges(visited); 3810 if (no_dead_code) { 3811 // Now make sure that no visited node is used by an unvisited node. 3812 bool dead_nodes = false; 3813 Unique_Node_List checked(area); 3814 while (visited.size() > 0) { 3815 Node* n = visited.pop(); 3816 checked.push(n); 3817 for (uint i = 0; i < n->outcnt(); i++) { 3818 Node* use = n->raw_out(i); 3819 if (checked.member(use)) continue; // already checked 3820 if (visited.member(use)) continue; // already in the graph 3821 if (use->is_Con()) continue; // a dead ConNode is OK 3822 // At this point, we have found a dead node which is DU-reachable. 3823 if (!dead_nodes) { 3824 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3825 dead_nodes = true; 3826 } 3827 use->dump(2); 3828 tty->print_cr("---"); 3829 checked.push(use); // No repeats; pretend it is now checked. 3830 } 3831 } 3832 assert(!dead_nodes, "using nodes must be reachable from root"); 3833 } 3834 } 3835 } 3836 3837 // Verify GC barriers consistency 3838 // Currently supported: 3839 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3840 void Compile::verify_barriers() { 3841 #if INCLUDE_G1GC 3842 if (UseG1GC) { 3843 // Verify G1 pre-barriers 3844 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3845 3846 ResourceArea *area = Thread::current()->resource_area(); 3847 Unique_Node_List visited(area); 3848 Node_List worklist(area); 3849 // We're going to walk control flow backwards starting from the Root 3850 worklist.push(_root); 3851 while (worklist.size() > 0) { 3852 Node* x = worklist.pop(); 3853 if (x == NULL || x == top()) continue; 3854 if (visited.member(x)) { 3855 continue; 3856 } else { 3857 visited.push(x); 3858 } 3859 3860 if (x->is_Region()) { 3861 for (uint i = 1; i < x->req(); i++) { 3862 worklist.push(x->in(i)); 3863 } 3864 } else { 3865 worklist.push(x->in(0)); 3866 // We are looking for the pattern: 3867 // /->ThreadLocal 3868 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3869 // \->ConI(0) 3870 // We want to verify that the If and the LoadB have the same control 3871 // See GraphKit::g1_write_barrier_pre() 3872 if (x->is_If()) { 3873 IfNode *iff = x->as_If(); 3874 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3875 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3876 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3877 && cmp->in(1)->is_Load()) { 3878 LoadNode* load = cmp->in(1)->as_Load(); 3879 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3880 && load->in(2)->in(3)->is_Con() 3881 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3882 3883 Node* if_ctrl = iff->in(0); 3884 Node* load_ctrl = load->in(0); 3885 3886 if (if_ctrl != load_ctrl) { 3887 // Skip possible CProj->NeverBranch in infinite loops 3888 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3889 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3890 if_ctrl = if_ctrl->in(0)->in(0); 3891 } 3892 } 3893 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3894 } 3895 } 3896 } 3897 } 3898 } 3899 } 3900 } 3901 #endif 3902 } 3903 3904 #endif 3905 3906 // The Compile object keeps track of failure reasons separately from the ciEnv. 3907 // This is required because there is not quite a 1-1 relation between the 3908 // ciEnv and its compilation task and the Compile object. Note that one 3909 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3910 // to backtrack and retry without subsuming loads. Other than this backtracking 3911 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3912 // by the logic in C2Compiler. 3913 void Compile::record_failure(const char* reason) { 3914 if (log() != NULL) { 3915 log()->elem("failure reason='%s' phase='compile'", reason); 3916 } 3917 if (_failure_reason == NULL) { 3918 // Record the first failure reason. 3919 _failure_reason = reason; 3920 } 3921 3922 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3923 C->print_method(PHASE_FAILURE); 3924 } 3925 _root = NULL; // flush the graph, too 3926 } 3927 3928 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3929 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3930 _phase_name(name), _dolog(CITimeVerbose) 3931 { 3932 if (_dolog) { 3933 C = Compile::current(); 3934 _log = C->log(); 3935 } else { 3936 C = NULL; 3937 _log = NULL; 3938 } 3939 if (_log != NULL) { 3940 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3941 _log->stamp(); 3942 _log->end_head(); 3943 } 3944 } 3945 3946 Compile::TracePhase::~TracePhase() { 3947 3948 C = Compile::current(); 3949 if (_dolog) { 3950 _log = C->log(); 3951 } else { 3952 _log = NULL; 3953 } 3954 3955 #ifdef ASSERT 3956 if (PrintIdealNodeCount) { 3957 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3958 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3959 } 3960 3961 if (VerifyIdealNodeCount) { 3962 Compile::current()->print_missing_nodes(); 3963 } 3964 #endif 3965 3966 if (_log != NULL) { 3967 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3968 } 3969 } 3970 3971 //============================================================================= 3972 // Two Constant's are equal when the type and the value are equal. 3973 bool Compile::Constant::operator==(const Constant& other) { 3974 if (type() != other.type() ) return false; 3975 if (can_be_reused() != other.can_be_reused()) return false; 3976 // For floating point values we compare the bit pattern. 3977 switch (type()) { 3978 case T_INT: 3979 case T_FLOAT: return (_v._value.i == other._v._value.i); 3980 case T_LONG: 3981 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3982 case T_OBJECT: 3983 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3984 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3985 case T_METADATA: return (_v._metadata == other._v._metadata); 3986 default: ShouldNotReachHere(); return false; 3987 } 3988 } 3989 3990 static int type_to_size_in_bytes(BasicType t) { 3991 switch (t) { 3992 case T_INT: return sizeof(jint ); 3993 case T_LONG: return sizeof(jlong ); 3994 case T_FLOAT: return sizeof(jfloat ); 3995 case T_DOUBLE: return sizeof(jdouble); 3996 case T_METADATA: return sizeof(Metadata*); 3997 // We use T_VOID as marker for jump-table entries (labels) which 3998 // need an internal word relocation. 3999 case T_VOID: 4000 case T_ADDRESS: 4001 case T_OBJECT: return sizeof(jobject); 4002 default: 4003 ShouldNotReachHere(); 4004 return -1; 4005 } 4006 } 4007 4008 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4009 // sort descending 4010 if (a->freq() > b->freq()) return -1; 4011 if (a->freq() < b->freq()) return 1; 4012 return 0; 4013 } 4014 4015 void Compile::ConstantTable::calculate_offsets_and_size() { 4016 // First, sort the array by frequencies. 4017 _constants.sort(qsort_comparator); 4018 4019 #ifdef ASSERT 4020 // Make sure all jump-table entries were sorted to the end of the 4021 // array (they have a negative frequency). 4022 bool found_void = false; 4023 for (int i = 0; i < _constants.length(); i++) { 4024 Constant con = _constants.at(i); 4025 if (con.type() == T_VOID) 4026 found_void = true; // jump-tables 4027 else 4028 assert(!found_void, "wrong sorting"); 4029 } 4030 #endif 4031 4032 int offset = 0; 4033 for (int i = 0; i < _constants.length(); i++) { 4034 Constant* con = _constants.adr_at(i); 4035 4036 // Align offset for type. 4037 int typesize = type_to_size_in_bytes(con->type()); 4038 offset = align_up(offset, typesize); 4039 con->set_offset(offset); // set constant's offset 4040 4041 if (con->type() == T_VOID) { 4042 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4043 offset = offset + typesize * n->outcnt(); // expand jump-table 4044 } else { 4045 offset = offset + typesize; 4046 } 4047 } 4048 4049 // Align size up to the next section start (which is insts; see 4050 // CodeBuffer::align_at_start). 4051 assert(_size == -1, "already set?"); 4052 _size = align_up(offset, (int)CodeEntryAlignment); 4053 } 4054 4055 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4056 MacroAssembler _masm(&cb); 4057 for (int i = 0; i < _constants.length(); i++) { 4058 Constant con = _constants.at(i); 4059 address constant_addr = NULL; 4060 switch (con.type()) { 4061 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4062 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4063 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4064 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4065 case T_OBJECT: { 4066 jobject obj = con.get_jobject(); 4067 int oop_index = _masm.oop_recorder()->find_index(obj); 4068 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4069 break; 4070 } 4071 case T_ADDRESS: { 4072 address addr = (address) con.get_jobject(); 4073 constant_addr = _masm.address_constant(addr); 4074 break; 4075 } 4076 // We use T_VOID as marker for jump-table entries (labels) which 4077 // need an internal word relocation. 4078 case T_VOID: { 4079 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4080 // Fill the jump-table with a dummy word. The real value is 4081 // filled in later in fill_jump_table. 4082 address dummy = (address) n; 4083 constant_addr = _masm.address_constant(dummy); 4084 // Expand jump-table 4085 for (uint i = 1; i < n->outcnt(); i++) { 4086 address temp_addr = _masm.address_constant(dummy + i); 4087 assert(temp_addr, "consts section too small"); 4088 } 4089 break; 4090 } 4091 case T_METADATA: { 4092 Metadata* obj = con.get_metadata(); 4093 int metadata_index = _masm.oop_recorder()->find_index(obj); 4094 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4095 break; 4096 } 4097 default: ShouldNotReachHere(); 4098 } 4099 assert(constant_addr, "consts section too small"); 4100 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4101 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4102 } 4103 } 4104 4105 int Compile::ConstantTable::find_offset(Constant& con) const { 4106 int idx = _constants.find(con); 4107 guarantee(idx != -1, "constant must be in constant table"); 4108 int offset = _constants.at(idx).offset(); 4109 guarantee(offset != -1, "constant table not emitted yet?"); 4110 return offset; 4111 } 4112 4113 void Compile::ConstantTable::add(Constant& con) { 4114 if (con.can_be_reused()) { 4115 int idx = _constants.find(con); 4116 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4117 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4118 return; 4119 } 4120 } 4121 (void) _constants.append(con); 4122 } 4123 4124 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4125 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4126 Constant con(type, value, b->_freq); 4127 add(con); 4128 return con; 4129 } 4130 4131 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4132 Constant con(metadata); 4133 add(con); 4134 return con; 4135 } 4136 4137 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4138 jvalue value; 4139 BasicType type = oper->type()->basic_type(); 4140 switch (type) { 4141 case T_LONG: value.j = oper->constantL(); break; 4142 case T_FLOAT: value.f = oper->constantF(); break; 4143 case T_DOUBLE: value.d = oper->constantD(); break; 4144 case T_OBJECT: 4145 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4146 case T_METADATA: return add((Metadata*)oper->constant()); break; 4147 default: guarantee(false, "unhandled type: %s", type2name(type)); 4148 } 4149 return add(n, type, value); 4150 } 4151 4152 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4153 jvalue value; 4154 // We can use the node pointer here to identify the right jump-table 4155 // as this method is called from Compile::Fill_buffer right before 4156 // the MachNodes are emitted and the jump-table is filled (means the 4157 // MachNode pointers do not change anymore). 4158 value.l = (jobject) n; 4159 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4160 add(con); 4161 return con; 4162 } 4163 4164 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4165 // If called from Compile::scratch_emit_size do nothing. 4166 if (Compile::current()->in_scratch_emit_size()) return; 4167 4168 assert(labels.is_nonempty(), "must be"); 4169 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4170 4171 // Since MachConstantNode::constant_offset() also contains 4172 // table_base_offset() we need to subtract the table_base_offset() 4173 // to get the plain offset into the constant table. 4174 int offset = n->constant_offset() - table_base_offset(); 4175 4176 MacroAssembler _masm(&cb); 4177 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4178 4179 for (uint i = 0; i < n->outcnt(); i++) { 4180 address* constant_addr = &jump_table_base[i]; 4181 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)); 4182 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4183 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4184 } 4185 } 4186 4187 //----------------------------static_subtype_check----------------------------- 4188 // Shortcut important common cases when superklass is exact: 4189 // (0) superklass is java.lang.Object (can occur in reflective code) 4190 // (1) subklass is already limited to a subtype of superklass => always ok 4191 // (2) subklass does not overlap with superklass => always fail 4192 // (3) superklass has NO subtypes and we can check with a simple compare. 4193 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4194 if (StressReflectiveCode) { 4195 return SSC_full_test; // Let caller generate the general case. 4196 } 4197 4198 if (superk == env()->Object_klass()) { 4199 return SSC_always_true; // (0) this test cannot fail 4200 } 4201 4202 ciType* superelem = superk; 4203 if (superelem->is_array_klass()) 4204 superelem = superelem->as_array_klass()->base_element_type(); 4205 4206 if (!subk->is_interface()) { // cannot trust static interface types yet 4207 if (subk->is_subtype_of(superk)) { 4208 return SSC_always_true; // (1) false path dead; no dynamic test needed 4209 } 4210 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4211 !superk->is_subtype_of(subk)) { 4212 return SSC_always_false; 4213 } 4214 } 4215 4216 // If casting to an instance klass, it must have no subtypes 4217 if (superk->is_interface()) { 4218 // Cannot trust interfaces yet. 4219 // %%% S.B. superk->nof_implementors() == 1 4220 } else if (superelem->is_instance_klass()) { 4221 ciInstanceKlass* ik = superelem->as_instance_klass(); 4222 if (!ik->has_subklass() && !ik->is_interface()) { 4223 if (!ik->is_final()) { 4224 // Add a dependency if there is a chance of a later subclass. 4225 dependencies()->assert_leaf_type(ik); 4226 } 4227 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4228 } 4229 } else { 4230 // A primitive array type has no subtypes. 4231 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4232 } 4233 4234 return SSC_full_test; 4235 } 4236 4237 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4238 #ifdef _LP64 4239 // The scaled index operand to AddP must be a clean 64-bit value. 4240 // Java allows a 32-bit int to be incremented to a negative 4241 // value, which appears in a 64-bit register as a large 4242 // positive number. Using that large positive number as an 4243 // operand in pointer arithmetic has bad consequences. 4244 // On the other hand, 32-bit overflow is rare, and the possibility 4245 // can often be excluded, if we annotate the ConvI2L node with 4246 // a type assertion that its value is known to be a small positive 4247 // number. (The prior range check has ensured this.) 4248 // This assertion is used by ConvI2LNode::Ideal. 4249 int index_max = max_jint - 1; // array size is max_jint, index is one less 4250 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4251 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4252 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4253 #endif 4254 return idx; 4255 } 4256 4257 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4258 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4259 if (ctrl != NULL) { 4260 // Express control dependency by a CastII node with a narrow type. 4261 value = new CastIINode(value, itype, false, true /* range check dependency */); 4262 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4263 // node from floating above the range check during loop optimizations. Otherwise, the 4264 // ConvI2L node may be eliminated independently of the range check, causing the data path 4265 // to become TOP while the control path is still there (although it's unreachable). 4266 value->set_req(0, ctrl); 4267 // Save CastII node to remove it after loop optimizations. 4268 phase->C->add_range_check_cast(value); 4269 value = phase->transform(value); 4270 } 4271 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4272 return phase->transform(new ConvI2LNode(value, ltype)); 4273 } 4274 4275 // The message about the current inlining is accumulated in 4276 // _print_inlining_stream and transfered into the _print_inlining_list 4277 // once we know whether inlining succeeds or not. For regular 4278 // inlining, messages are appended to the buffer pointed by 4279 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4280 // a new buffer is added after _print_inlining_idx in the list. This 4281 // way we can update the inlining message for late inlining call site 4282 // when the inlining is attempted again. 4283 void Compile::print_inlining_init() { 4284 if (print_inlining() || print_intrinsics()) { 4285 _print_inlining_stream = new stringStream(); 4286 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4287 } 4288 } 4289 4290 void Compile::print_inlining_reinit() { 4291 if (print_inlining() || print_intrinsics()) { 4292 // Re allocate buffer when we change ResourceMark 4293 _print_inlining_stream = new stringStream(); 4294 } 4295 } 4296 4297 void Compile::print_inlining_reset() { 4298 _print_inlining_stream->reset(); 4299 } 4300 4301 void Compile::print_inlining_commit() { 4302 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4303 // Transfer the message from _print_inlining_stream to the current 4304 // _print_inlining_list buffer and clear _print_inlining_stream. 4305 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4306 print_inlining_reset(); 4307 } 4308 4309 void Compile::print_inlining_push() { 4310 // Add new buffer to the _print_inlining_list at current position 4311 _print_inlining_idx++; 4312 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4313 } 4314 4315 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4316 return _print_inlining_list->at(_print_inlining_idx); 4317 } 4318 4319 void Compile::print_inlining_update(CallGenerator* cg) { 4320 if (print_inlining() || print_intrinsics()) { 4321 if (!cg->is_late_inline()) { 4322 if (print_inlining_current().cg() != NULL) { 4323 print_inlining_push(); 4324 } 4325 print_inlining_commit(); 4326 } else { 4327 if (print_inlining_current().cg() != cg && 4328 (print_inlining_current().cg() != NULL || 4329 print_inlining_current().ss()->size() != 0)) { 4330 print_inlining_push(); 4331 } 4332 print_inlining_commit(); 4333 print_inlining_current().set_cg(cg); 4334 } 4335 } 4336 } 4337 4338 void Compile::print_inlining_move_to(CallGenerator* cg) { 4339 // We resume inlining at a late inlining call site. Locate the 4340 // corresponding inlining buffer so that we can update it. 4341 if (print_inlining()) { 4342 for (int i = 0; i < _print_inlining_list->length(); i++) { 4343 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4344 _print_inlining_idx = i; 4345 return; 4346 } 4347 } 4348 ShouldNotReachHere(); 4349 } 4350 } 4351 4352 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4353 if (print_inlining()) { 4354 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4355 assert(print_inlining_current().cg() == cg, "wrong entry"); 4356 // replace message with new message 4357 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4358 print_inlining_commit(); 4359 print_inlining_current().set_cg(cg); 4360 } 4361 } 4362 4363 void Compile::print_inlining_assert_ready() { 4364 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4365 } 4366 4367 void Compile::process_print_inlining() { 4368 bool do_print_inlining = print_inlining() || print_intrinsics(); 4369 if (do_print_inlining || log() != NULL) { 4370 // Print inlining message for candidates that we couldn't inline 4371 // for lack of space 4372 for (int i = 0; i < _late_inlines.length(); i++) { 4373 CallGenerator* cg = _late_inlines.at(i); 4374 if (!cg->is_mh_late_inline()) { 4375 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4376 if (do_print_inlining) { 4377 cg->print_inlining_late(msg); 4378 } 4379 log_late_inline_failure(cg, msg); 4380 } 4381 } 4382 } 4383 if (do_print_inlining) { 4384 ResourceMark rm; 4385 stringStream ss; 4386 for (int i = 0; i < _print_inlining_list->length(); i++) { 4387 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4388 } 4389 size_t end = ss.size(); 4390 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4391 strncpy(_print_inlining_output, ss.base(), end+1); 4392 _print_inlining_output[end] = 0; 4393 } 4394 } 4395 4396 void Compile::dump_print_inlining() { 4397 if (_print_inlining_output != NULL) { 4398 tty->print_raw(_print_inlining_output); 4399 } 4400 } 4401 4402 void Compile::log_late_inline(CallGenerator* cg) { 4403 if (log() != NULL) { 4404 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4405 cg->unique_id()); 4406 JVMState* p = cg->call_node()->jvms(); 4407 while (p != NULL) { 4408 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4409 p = p->caller(); 4410 } 4411 log()->tail("late_inline"); 4412 } 4413 } 4414 4415 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4416 log_late_inline(cg); 4417 if (log() != NULL) { 4418 log()->inline_fail(msg); 4419 } 4420 } 4421 4422 void Compile::log_inline_id(CallGenerator* cg) { 4423 if (log() != NULL) { 4424 // The LogCompilation tool needs a unique way to identify late 4425 // inline call sites. This id must be unique for this call site in 4426 // this compilation. Try to have it unique across compilations as 4427 // well because it can be convenient when grepping through the log 4428 // file. 4429 // Distinguish OSR compilations from others in case CICountOSR is 4430 // on. 4431 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4432 cg->set_unique_id(id); 4433 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4434 } 4435 } 4436 4437 void Compile::log_inline_failure(const char* msg) { 4438 if (C->log() != NULL) { 4439 C->log()->inline_fail(msg); 4440 } 4441 } 4442 4443 4444 // Dump inlining replay data to the stream. 4445 // Don't change thread state and acquire any locks. 4446 void Compile::dump_inline_data(outputStream* out) { 4447 InlineTree* inl_tree = ilt(); 4448 if (inl_tree != NULL) { 4449 out->print(" inline %d", inl_tree->count()); 4450 inl_tree->dump_replay_data(out); 4451 } 4452 } 4453 4454 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4455 if (n1->Opcode() < n2->Opcode()) return -1; 4456 else if (n1->Opcode() > n2->Opcode()) return 1; 4457 4458 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4459 for (uint i = 1; i < n1->req(); i++) { 4460 if (n1->in(i) < n2->in(i)) return -1; 4461 else if (n1->in(i) > n2->in(i)) return 1; 4462 } 4463 4464 return 0; 4465 } 4466 4467 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4468 Node* n1 = *n1p; 4469 Node* n2 = *n2p; 4470 4471 return cmp_expensive_nodes(n1, n2); 4472 } 4473 4474 void Compile::sort_expensive_nodes() { 4475 if (!expensive_nodes_sorted()) { 4476 _expensive_nodes->sort(cmp_expensive_nodes); 4477 } 4478 } 4479 4480 bool Compile::expensive_nodes_sorted() const { 4481 for (int i = 1; i < _expensive_nodes->length(); i++) { 4482 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4483 return false; 4484 } 4485 } 4486 return true; 4487 } 4488 4489 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4490 if (_expensive_nodes->length() == 0) { 4491 return false; 4492 } 4493 4494 assert(OptimizeExpensiveOps, "optimization off?"); 4495 4496 // Take this opportunity to remove dead nodes from the list 4497 int j = 0; 4498 for (int i = 0; i < _expensive_nodes->length(); i++) { 4499 Node* n = _expensive_nodes->at(i); 4500 if (!n->is_unreachable(igvn)) { 4501 assert(n->is_expensive(), "should be expensive"); 4502 _expensive_nodes->at_put(j, n); 4503 j++; 4504 } 4505 } 4506 _expensive_nodes->trunc_to(j); 4507 4508 // Then sort the list so that similar nodes are next to each other 4509 // and check for at least two nodes of identical kind with same data 4510 // inputs. 4511 sort_expensive_nodes(); 4512 4513 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4514 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4515 return true; 4516 } 4517 } 4518 4519 return false; 4520 } 4521 4522 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4523 if (_expensive_nodes->length() == 0) { 4524 return; 4525 } 4526 4527 assert(OptimizeExpensiveOps, "optimization off?"); 4528 4529 // Sort to bring similar nodes next to each other and clear the 4530 // control input of nodes for which there's only a single copy. 4531 sort_expensive_nodes(); 4532 4533 int j = 0; 4534 int identical = 0; 4535 int i = 0; 4536 bool modified = false; 4537 for (; i < _expensive_nodes->length()-1; i++) { 4538 assert(j <= i, "can't write beyond current index"); 4539 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4540 identical++; 4541 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4542 continue; 4543 } 4544 if (identical > 0) { 4545 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4546 identical = 0; 4547 } else { 4548 Node* n = _expensive_nodes->at(i); 4549 igvn.replace_input_of(n, 0, NULL); 4550 igvn.hash_insert(n); 4551 modified = true; 4552 } 4553 } 4554 if (identical > 0) { 4555 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4556 } else if (_expensive_nodes->length() >= 1) { 4557 Node* n = _expensive_nodes->at(i); 4558 igvn.replace_input_of(n, 0, NULL); 4559 igvn.hash_insert(n); 4560 modified = true; 4561 } 4562 _expensive_nodes->trunc_to(j); 4563 if (modified) { 4564 igvn.optimize(); 4565 } 4566 } 4567 4568 void Compile::add_expensive_node(Node * n) { 4569 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4570 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4571 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4572 if (OptimizeExpensiveOps) { 4573 _expensive_nodes->append(n); 4574 } else { 4575 // Clear control input and let IGVN optimize expensive nodes if 4576 // OptimizeExpensiveOps is off. 4577 n->set_req(0, NULL); 4578 } 4579 } 4580 4581 /** 4582 * Remove the speculative part of types and clean up the graph 4583 */ 4584 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4585 if (UseTypeSpeculation) { 4586 Unique_Node_List worklist; 4587 worklist.push(root()); 4588 int modified = 0; 4589 // Go over all type nodes that carry a speculative type, drop the 4590 // speculative part of the type and enqueue the node for an igvn 4591 // which may optimize it out. 4592 for (uint next = 0; next < worklist.size(); ++next) { 4593 Node *n = worklist.at(next); 4594 if (n->is_Type()) { 4595 TypeNode* tn = n->as_Type(); 4596 const Type* t = tn->type(); 4597 const Type* t_no_spec = t->remove_speculative(); 4598 if (t_no_spec != t) { 4599 bool in_hash = igvn.hash_delete(n); 4600 assert(in_hash, "node should be in igvn hash table"); 4601 tn->set_type(t_no_spec); 4602 igvn.hash_insert(n); 4603 igvn._worklist.push(n); // give it a chance to go away 4604 modified++; 4605 } 4606 } 4607 uint max = n->len(); 4608 for( uint i = 0; i < max; ++i ) { 4609 Node *m = n->in(i); 4610 if (not_a_node(m)) continue; 4611 worklist.push(m); 4612 } 4613 } 4614 // Drop the speculative part of all types in the igvn's type table 4615 igvn.remove_speculative_types(); 4616 if (modified > 0) { 4617 igvn.optimize(); 4618 } 4619 #ifdef ASSERT 4620 // Verify that after the IGVN is over no speculative type has resurfaced 4621 worklist.clear(); 4622 worklist.push(root()); 4623 for (uint next = 0; next < worklist.size(); ++next) { 4624 Node *n = worklist.at(next); 4625 const Type* t = igvn.type_or_null(n); 4626 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4627 if (n->is_Type()) { 4628 t = n->as_Type()->type(); 4629 assert(t == t->remove_speculative(), "no more speculative types"); 4630 } 4631 uint max = n->len(); 4632 for( uint i = 0; i < max; ++i ) { 4633 Node *m = n->in(i); 4634 if (not_a_node(m)) continue; 4635 worklist.push(m); 4636 } 4637 } 4638 igvn.check_no_speculative_types(); 4639 #endif 4640 } 4641 } 4642 4643 // Auxiliary method to support randomized stressing/fuzzing. 4644 // 4645 // This method can be called the arbitrary number of times, with current count 4646 // as the argument. The logic allows selecting a single candidate from the 4647 // running list of candidates as follows: 4648 // int count = 0; 4649 // Cand* selected = null; 4650 // while(cand = cand->next()) { 4651 // if (randomized_select(++count)) { 4652 // selected = cand; 4653 // } 4654 // } 4655 // 4656 // Including count equalizes the chances any candidate is "selected". 4657 // This is useful when we don't have the complete list of candidates to choose 4658 // from uniformly. In this case, we need to adjust the randomicity of the 4659 // selection, or else we will end up biasing the selection towards the latter 4660 // candidates. 4661 // 4662 // Quick back-envelope calculation shows that for the list of n candidates 4663 // the equal probability for the candidate to persist as "best" can be 4664 // achieved by replacing it with "next" k-th candidate with the probability 4665 // of 1/k. It can be easily shown that by the end of the run, the 4666 // probability for any candidate is converged to 1/n, thus giving the 4667 // uniform distribution among all the candidates. 4668 // 4669 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4670 #define RANDOMIZED_DOMAIN_POW 29 4671 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4672 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4673 bool Compile::randomized_select(int count) { 4674 assert(count > 0, "only positive"); 4675 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4676 } 4677 4678 CloneMap& Compile::clone_map() { return _clone_map; } 4679 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4680 4681 void NodeCloneInfo::dump() const { 4682 tty->print(" {%d:%d} ", idx(), gen()); 4683 } 4684 4685 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4686 uint64_t val = value(old->_idx); 4687 NodeCloneInfo cio(val); 4688 assert(val != 0, "old node should be in the map"); 4689 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4690 insert(nnn->_idx, cin.get()); 4691 #ifndef PRODUCT 4692 if (is_debug()) { 4693 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4694 } 4695 #endif 4696 } 4697 4698 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4699 NodeCloneInfo cio(value(old->_idx)); 4700 if (cio.get() == 0) { 4701 cio.set(old->_idx, 0); 4702 insert(old->_idx, cio.get()); 4703 #ifndef PRODUCT 4704 if (is_debug()) { 4705 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4706 } 4707 #endif 4708 } 4709 clone(old, nnn, gen); 4710 } 4711 4712 int CloneMap::max_gen() const { 4713 int g = 0; 4714 DictI di(_dict); 4715 for(; di.test(); ++di) { 4716 int t = gen(di._key); 4717 if (g < t) { 4718 g = t; 4719 #ifndef PRODUCT 4720 if (is_debug()) { 4721 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4722 } 4723 #endif 4724 } 4725 } 4726 return g; 4727 } 4728 4729 void CloneMap::dump(node_idx_t key) const { 4730 uint64_t val = value(key); 4731 if (val != 0) { 4732 NodeCloneInfo ni(val); 4733 ni.dump(); 4734 } 4735 }