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 #if INCLUDE_SHENANDOAHGC 86 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 87 #endif 88 89 90 // -------------------- Compile::mach_constant_base_node ----------------------- 91 // Constant table base node singleton. 92 MachConstantBaseNode* Compile::mach_constant_base_node() { 93 if (_mach_constant_base_node == NULL) { 94 _mach_constant_base_node = new MachConstantBaseNode(); 95 _mach_constant_base_node->add_req(C->root()); 96 } 97 return _mach_constant_base_node; 98 } 99 100 101 /// Support for intrinsics. 102 103 // Return the index at which m must be inserted (or already exists). 104 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 105 class IntrinsicDescPair { 106 private: 107 ciMethod* _m; 108 bool _is_virtual; 109 public: 110 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 111 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 112 ciMethod* m= elt->method(); 113 ciMethod* key_m = key->_m; 114 if (key_m < m) return -1; 115 else if (key_m > m) return 1; 116 else { 117 bool is_virtual = elt->is_virtual(); 118 bool key_virtual = key->_is_virtual; 119 if (key_virtual < is_virtual) return -1; 120 else if (key_virtual > is_virtual) return 1; 121 else return 0; 122 } 123 } 124 }; 125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 126 #ifdef ASSERT 127 for (int i = 1; i < _intrinsics->length(); i++) { 128 CallGenerator* cg1 = _intrinsics->at(i-1); 129 CallGenerator* cg2 = _intrinsics->at(i); 130 assert(cg1->method() != cg2->method() 131 ? cg1->method() < cg2->method() 132 : cg1->is_virtual() < cg2->is_virtual(), 133 "compiler intrinsics list must stay sorted"); 134 } 135 #endif 136 IntrinsicDescPair pair(m, is_virtual); 137 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 138 } 139 140 void Compile::register_intrinsic(CallGenerator* cg) { 141 if (_intrinsics == NULL) { 142 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 143 } 144 int len = _intrinsics->length(); 145 bool found = false; 146 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 147 assert(!found, "registering twice"); 148 _intrinsics->insert_before(index, cg); 149 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 150 } 151 152 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 153 assert(m->is_loaded(), "don't try this on unloaded methods"); 154 if (_intrinsics != NULL) { 155 bool found = false; 156 int index = intrinsic_insertion_index(m, is_virtual, found); 157 if (found) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = live_nodes(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless CastII nodes with range check dependency 406 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 407 Node* cast = range_check_cast_node(i); 408 if (!useful.member(cast)) { 409 remove_range_check_cast(cast); 410 } 411 } 412 // Remove useless expensive nodes 413 for (int i = C->expensive_count()-1; i >= 0; i--) { 414 Node* n = C->expensive_node(i); 415 if (!useful.member(n)) { 416 remove_expensive_node(n); 417 } 418 } 419 // Remove useless Opaque4 nodes 420 for (int i = opaque4_count() - 1; i >= 0; i--) { 421 Node* opaq = opaque4_node(i); 422 if (!useful.member(opaq)) { 423 remove_opaque4_node(opaq); 424 } 425 } 426 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 427 bs->eliminate_useless_gc_barriers(useful); 428 // clean up the late inline lists 429 remove_useless_late_inlines(&_string_late_inlines, useful); 430 remove_useless_late_inlines(&_boxing_late_inlines, useful); 431 remove_useless_late_inlines(&_late_inlines, useful); 432 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 433 } 434 435 //------------------------------frame_size_in_words----------------------------- 436 // frame_slots in units of words 437 int Compile::frame_size_in_words() const { 438 // shift is 0 in LP32 and 1 in LP64 439 const int shift = (LogBytesPerWord - LogBytesPerInt); 440 int words = _frame_slots >> shift; 441 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 442 return words; 443 } 444 445 // To bang the stack of this compiled method we use the stack size 446 // that the interpreter would need in case of a deoptimization. This 447 // removes the need to bang the stack in the deoptimization blob which 448 // in turn simplifies stack overflow handling. 449 int Compile::bang_size_in_bytes() const { 450 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 451 } 452 453 // ============================================================================ 454 //------------------------------CompileWrapper--------------------------------- 455 class CompileWrapper : public StackObj { 456 Compile *const _compile; 457 public: 458 CompileWrapper(Compile* compile); 459 460 ~CompileWrapper(); 461 }; 462 463 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 464 // the Compile* pointer is stored in the current ciEnv: 465 ciEnv* env = compile->env(); 466 assert(env == ciEnv::current(), "must already be a ciEnv active"); 467 assert(env->compiler_data() == NULL, "compile already active?"); 468 env->set_compiler_data(compile); 469 assert(compile == Compile::current(), "sanity"); 470 471 compile->set_type_dict(NULL); 472 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 473 compile->clone_map().set_clone_idx(0); 474 compile->set_type_hwm(NULL); 475 compile->set_type_last_size(0); 476 compile->set_last_tf(NULL, NULL); 477 compile->set_indexSet_arena(NULL); 478 compile->set_indexSet_free_block_list(NULL); 479 compile->init_type_arena(); 480 Type::Initialize(compile); 481 _compile->set_scratch_buffer_blob(NULL); 482 _compile->begin_method(); 483 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 484 } 485 CompileWrapper::~CompileWrapper() { 486 _compile->end_method(); 487 if (_compile->scratch_buffer_blob() != NULL) 488 BufferBlob::free(_compile->scratch_buffer_blob()); 489 _compile->env()->set_compiler_data(NULL); 490 } 491 492 493 //----------------------------print_compile_messages--------------------------- 494 void Compile::print_compile_messages() { 495 #ifndef PRODUCT 496 // Check if recompiling 497 if (_subsume_loads == false && PrintOpto) { 498 // Recompiling without allowing machine instructions to subsume loads 499 tty->print_cr("*********************************************************"); 500 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 501 tty->print_cr("*********************************************************"); 502 } 503 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 504 // Recompiling without escape analysis 505 tty->print_cr("*********************************************************"); 506 tty->print_cr("** Bailout: Recompile without escape analysis **"); 507 tty->print_cr("*********************************************************"); 508 } 509 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 510 // Recompiling without boxing elimination 511 tty->print_cr("*********************************************************"); 512 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 513 tty->print_cr("*********************************************************"); 514 } 515 if (C->directive()->BreakAtCompileOption) { 516 // Open the debugger when compiling this method. 517 tty->print("### Breaking when compiling: "); 518 method()->print_short_name(); 519 tty->cr(); 520 BREAKPOINT; 521 } 522 523 if( PrintOpto ) { 524 if (is_osr_compilation()) { 525 tty->print("[OSR]%3d", _compile_id); 526 } else { 527 tty->print("%3d", _compile_id); 528 } 529 } 530 #endif 531 } 532 533 534 //-----------------------init_scratch_buffer_blob------------------------------ 535 // Construct a temporary BufferBlob and cache it for this compile. 536 void Compile::init_scratch_buffer_blob(int const_size) { 537 // If there is already a scratch buffer blob allocated and the 538 // constant section is big enough, use it. Otherwise free the 539 // current and allocate a new one. 540 BufferBlob* blob = scratch_buffer_blob(); 541 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 542 // Use the current blob. 543 } else { 544 if (blob != NULL) { 545 BufferBlob::free(blob); 546 } 547 548 ResourceMark rm; 549 _scratch_const_size = const_size; 550 int size = C2Compiler::initial_code_buffer_size(const_size); 551 blob = BufferBlob::create("Compile::scratch_buffer", size); 552 // Record the buffer blob for next time. 553 set_scratch_buffer_blob(blob); 554 // Have we run out of code space? 555 if (scratch_buffer_blob() == NULL) { 556 // Let CompilerBroker disable further compilations. 557 record_failure("Not enough space for scratch buffer in CodeCache"); 558 return; 559 } 560 } 561 562 // Initialize the relocation buffers 563 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 564 set_scratch_locs_memory(locs_buf); 565 } 566 567 568 //-----------------------scratch_emit_size------------------------------------- 569 // Helper function that computes size by emitting code 570 uint Compile::scratch_emit_size(const Node* n) { 571 // Start scratch_emit_size section. 572 set_in_scratch_emit_size(true); 573 574 // Emit into a trash buffer and count bytes emitted. 575 // This is a pretty expensive way to compute a size, 576 // but it works well enough if seldom used. 577 // All common fixed-size instructions are given a size 578 // method by the AD file. 579 // Note that the scratch buffer blob and locs memory are 580 // allocated at the beginning of the compile task, and 581 // may be shared by several calls to scratch_emit_size. 582 // The allocation of the scratch buffer blob is particularly 583 // expensive, since it has to grab the code cache lock. 584 BufferBlob* blob = this->scratch_buffer_blob(); 585 assert(blob != NULL, "Initialize BufferBlob at start"); 586 assert(blob->size() > MAX_inst_size, "sanity"); 587 relocInfo* locs_buf = scratch_locs_memory(); 588 address blob_begin = blob->content_begin(); 589 address blob_end = (address)locs_buf; 590 assert(blob->contains(blob_end), "sanity"); 591 CodeBuffer buf(blob_begin, blob_end - blob_begin); 592 buf.initialize_consts_size(_scratch_const_size); 593 buf.initialize_stubs_size(MAX_stubs_size); 594 assert(locs_buf != NULL, "sanity"); 595 int lsize = MAX_locs_size / 3; 596 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 597 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 598 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 599 // Mark as scratch buffer. 600 buf.consts()->set_scratch_emit(); 601 buf.insts()->set_scratch_emit(); 602 buf.stubs()->set_scratch_emit(); 603 604 // Do the emission. 605 606 Label fakeL; // Fake label for branch instructions. 607 Label* saveL = NULL; 608 uint save_bnum = 0; 609 bool is_branch = n->is_MachBranch(); 610 if (is_branch) { 611 MacroAssembler masm(&buf); 612 masm.bind(fakeL); 613 n->as_MachBranch()->save_label(&saveL, &save_bnum); 614 n->as_MachBranch()->label_set(&fakeL, 0); 615 } 616 n->emit(buf, this->regalloc()); 617 618 // Emitting into the scratch buffer should not fail 619 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 620 621 if (is_branch) // Restore label. 622 n->as_MachBranch()->label_set(saveL, save_bnum); 623 624 // End scratch_emit_size section. 625 set_in_scratch_emit_size(false); 626 627 return buf.insts_size(); 628 } 629 630 631 // ============================================================================ 632 //------------------------------Compile standard------------------------------- 633 debug_only( int Compile::_debug_idx = 100000; ) 634 635 // Compile a method. entry_bci is -1 for normal compilations and indicates 636 // the continuation bci for on stack replacement. 637 638 639 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 640 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 641 : Phase(Compiler), 642 _env(ci_env), 643 _directive(directive), 644 _log(ci_env->log()), 645 _compile_id(ci_env->compile_id()), 646 _save_argument_registers(false), 647 _stub_name(NULL), 648 _stub_function(NULL), 649 _stub_entry_point(NULL), 650 _method(target), 651 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 652 _entry_bci(osr_bci), 653 _initial_gvn(NULL), 654 _for_igvn(NULL), 655 _warm_calls(NULL), 656 _subsume_loads(subsume_loads), 657 _do_escape_analysis(do_escape_analysis), 658 _eliminate_boxing(eliminate_boxing), 659 _failure_reason(NULL), 660 _code_buffer("Compile::Fill_buffer"), 661 _orig_pc_slot(0), 662 _orig_pc_slot_offset_in_bytes(0), 663 _has_method_handle_invokes(false), 664 _mach_constant_base_node(NULL), 665 _node_bundling_limit(0), 666 _node_bundling_base(NULL), 667 _java_calls(0), 668 _inner_loops(0), 669 _scratch_const_size(-1), 670 _in_scratch_emit_size(false), 671 _dead_node_list(comp_arena()), 672 _dead_node_count(0), 673 #ifndef PRODUCT 674 _trace_opto_output(directive->TraceOptoOutputOption), 675 _in_dump_cnt(0), 676 _printer(IdealGraphPrinter::printer()), 677 #endif 678 _congraph(NULL), 679 _comp_arena(mtCompiler), 680 _node_arena(mtCompiler), 681 _old_arena(mtCompiler), 682 _Compile_types(mtCompiler), 683 _replay_inline_data(NULL), 684 _late_inlines(comp_arena(), 2, 0, NULL), 685 _string_late_inlines(comp_arena(), 2, 0, NULL), 686 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 687 _late_inlines_pos(0), 688 _number_of_mh_late_inlines(0), 689 _inlining_progress(false), 690 _inlining_incrementally(false), 691 _print_inlining_list(NULL), 692 _print_inlining_stream(NULL), 693 _print_inlining_idx(0), 694 _print_inlining_output(NULL), 695 _interpreter_frame_size(0), 696 _max_node_limit(MaxNodeLimit), 697 _has_reserved_stack_access(target->has_reserved_stack_access()) { 698 C = this; 699 #ifndef PRODUCT 700 if (_printer != NULL) { 701 _printer->set_compile(this); 702 } 703 #endif 704 CompileWrapper cw(this); 705 706 if (CITimeVerbose) { 707 tty->print(" "); 708 target->holder()->name()->print(); 709 tty->print("."); 710 target->print_short_name(); 711 tty->print(" "); 712 } 713 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 714 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 715 716 #ifndef PRODUCT 717 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 718 if (!print_opto_assembly) { 719 bool print_assembly = directive->PrintAssemblyOption; 720 if (print_assembly && !Disassembler::can_decode()) { 721 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 722 print_opto_assembly = true; 723 } 724 } 725 set_print_assembly(print_opto_assembly); 726 set_parsed_irreducible_loop(false); 727 728 if (directive->ReplayInlineOption) { 729 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 730 } 731 #endif 732 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 733 set_print_intrinsics(directive->PrintIntrinsicsOption); 734 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 735 736 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 737 // Make sure the method being compiled gets its own MDO, 738 // so we can at least track the decompile_count(). 739 // Need MDO to record RTM code generation state. 740 method()->ensure_method_data(); 741 } 742 743 Init(::AliasLevel); 744 745 746 print_compile_messages(); 747 748 _ilt = InlineTree::build_inline_tree_root(); 749 750 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 751 assert(num_alias_types() >= AliasIdxRaw, ""); 752 753 #define MINIMUM_NODE_HASH 1023 754 // Node list that Iterative GVN will start with 755 Unique_Node_List for_igvn(comp_arena()); 756 set_for_igvn(&for_igvn); 757 758 // GVN that will be run immediately on new nodes 759 uint estimated_size = method()->code_size()*4+64; 760 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 761 PhaseGVN gvn(node_arena(), estimated_size); 762 set_initial_gvn(&gvn); 763 764 print_inlining_init(); 765 { // Scope for timing the parser 766 TracePhase tp("parse", &timers[_t_parser]); 767 768 // Put top into the hash table ASAP. 769 initial_gvn()->transform_no_reclaim(top()); 770 771 // Set up tf(), start(), and find a CallGenerator. 772 CallGenerator* cg = NULL; 773 if (is_osr_compilation()) { 774 const TypeTuple *domain = StartOSRNode::osr_domain(); 775 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 776 init_tf(TypeFunc::make(domain, range)); 777 StartNode* s = new StartOSRNode(root(), domain); 778 initial_gvn()->set_type_bottom(s); 779 init_start(s); 780 cg = CallGenerator::for_osr(method(), entry_bci()); 781 } else { 782 // Normal case. 783 init_tf(TypeFunc::make(method())); 784 StartNode* s = new StartNode(root(), tf()->domain()); 785 initial_gvn()->set_type_bottom(s); 786 init_start(s); 787 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 788 // With java.lang.ref.reference.get() we must go through the 789 // intrinsic - even when get() is the root 790 // method of the compile - so that, if necessary, the value in 791 // the referent field of the reference object gets recorded by 792 // the pre-barrier code. 793 cg = find_intrinsic(method(), false); 794 } 795 if (cg == NULL) { 796 float past_uses = method()->interpreter_invocation_count(); 797 float expected_uses = past_uses; 798 cg = CallGenerator::for_inline(method(), expected_uses); 799 } 800 } 801 if (failing()) return; 802 if (cg == NULL) { 803 record_method_not_compilable("cannot parse method"); 804 return; 805 } 806 JVMState* jvms = build_start_state(start(), tf()); 807 if ((jvms = cg->generate(jvms)) == NULL) { 808 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 809 record_method_not_compilable("method parse failed"); 810 } 811 return; 812 } 813 GraphKit kit(jvms); 814 815 if (!kit.stopped()) { 816 // Accept return values, and transfer control we know not where. 817 // This is done by a special, unique ReturnNode bound to root. 818 return_values(kit.jvms()); 819 } 820 821 if (kit.has_exceptions()) { 822 // Any exceptions that escape from this call must be rethrown 823 // to whatever caller is dynamically above us on the stack. 824 // This is done by a special, unique RethrowNode bound to root. 825 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 826 } 827 828 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 829 830 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 831 inline_string_calls(true); 832 } 833 834 if (failing()) return; 835 836 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 837 838 // Remove clutter produced by parsing. 839 if (!failing()) { 840 ResourceMark rm; 841 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 842 } 843 } 844 845 // Note: Large methods are capped off in do_one_bytecode(). 846 if (failing()) return; 847 848 // After parsing, node notes are no longer automagic. 849 // They must be propagated by register_new_node_with_optimizer(), 850 // clone(), or the like. 851 set_default_node_notes(NULL); 852 853 for (;;) { 854 int successes = Inline_Warm(); 855 if (failing()) return; 856 if (successes == 0) break; 857 } 858 859 // Drain the list. 860 Finish_Warm(); 861 #ifndef PRODUCT 862 if (_printer && _printer->should_print(1)) { 863 _printer->print_inlining(); 864 } 865 #endif 866 867 if (failing()) return; 868 NOT_PRODUCT( verify_graph_edges(); ) 869 870 // Now optimize 871 Optimize(); 872 if (failing()) return; 873 NOT_PRODUCT( verify_graph_edges(); ) 874 875 #ifndef PRODUCT 876 if (PrintIdeal) { 877 ttyLocker ttyl; // keep the following output all in one block 878 // This output goes directly to the tty, not the compiler log. 879 // To enable tools to match it up with the compilation activity, 880 // be sure to tag this tty output with the compile ID. 881 if (xtty != NULL) { 882 xtty->head("ideal compile_id='%d'%s", compile_id(), 883 is_osr_compilation() ? " compile_kind='osr'" : 884 ""); 885 } 886 root()->dump(9999); 887 if (xtty != NULL) { 888 xtty->tail("ideal"); 889 } 890 } 891 #endif 892 893 NOT_PRODUCT( verify_barriers(); ) 894 895 // Dump compilation data to replay it. 896 if (directive->DumpReplayOption) { 897 env()->dump_replay_data(_compile_id); 898 } 899 if (directive->DumpInlineOption && (ilt() != NULL)) { 900 env()->dump_inline_data(_compile_id); 901 } 902 903 // Now that we know the size of all the monitors we can add a fixed slot 904 // for the original deopt pc. 905 906 _orig_pc_slot = fixed_slots(); 907 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 908 set_fixed_slots(next_slot); 909 910 // Compute when to use implicit null checks. Used by matching trap based 911 // nodes and NullCheck optimization. 912 set_allowed_deopt_reasons(); 913 914 // Now generate code 915 Code_Gen(); 916 if (failing()) return; 917 918 // Check if we want to skip execution of all compiled code. 919 { 920 #ifndef PRODUCT 921 if (OptoNoExecute) { 922 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 923 return; 924 } 925 #endif 926 TracePhase tp("install_code", &timers[_t_registerMethod]); 927 928 if (is_osr_compilation()) { 929 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 930 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 931 } else { 932 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 933 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 934 } 935 936 env()->register_method(_method, _entry_bci, 937 &_code_offsets, 938 _orig_pc_slot_offset_in_bytes, 939 code_buffer(), 940 frame_size_in_words(), _oop_map_set, 941 &_handler_table, &_inc_table, 942 compiler, 943 has_unsafe_access(), 944 SharedRuntime::is_wide_vector(max_vector_size()), 945 rtm_state() 946 ); 947 948 if (log() != NULL) // Print code cache state into compiler log 949 log()->code_cache_state(); 950 } 951 } 952 953 //------------------------------Compile---------------------------------------- 954 // Compile a runtime stub 955 Compile::Compile( ciEnv* ci_env, 956 TypeFunc_generator generator, 957 address stub_function, 958 const char *stub_name, 959 int is_fancy_jump, 960 bool pass_tls, 961 bool save_arg_registers, 962 bool return_pc, 963 DirectiveSet* directive) 964 : Phase(Compiler), 965 _env(ci_env), 966 _directive(directive), 967 _log(ci_env->log()), 968 _compile_id(0), 969 _save_argument_registers(save_arg_registers), 970 _method(NULL), 971 _stub_name(stub_name), 972 _stub_function(stub_function), 973 _stub_entry_point(NULL), 974 _entry_bci(InvocationEntryBci), 975 _initial_gvn(NULL), 976 _for_igvn(NULL), 977 _warm_calls(NULL), 978 _orig_pc_slot(0), 979 _orig_pc_slot_offset_in_bytes(0), 980 _subsume_loads(true), 981 _do_escape_analysis(false), 982 _eliminate_boxing(false), 983 _failure_reason(NULL), 984 _code_buffer("Compile::Fill_buffer"), 985 _has_method_handle_invokes(false), 986 _mach_constant_base_node(NULL), 987 _node_bundling_limit(0), 988 _node_bundling_base(NULL), 989 _java_calls(0), 990 _inner_loops(0), 991 #ifndef PRODUCT 992 _trace_opto_output(directive->TraceOptoOutputOption), 993 _in_dump_cnt(0), 994 _printer(NULL), 995 #endif 996 _comp_arena(mtCompiler), 997 _node_arena(mtCompiler), 998 _old_arena(mtCompiler), 999 _Compile_types(mtCompiler), 1000 _dead_node_list(comp_arena()), 1001 _dead_node_count(0), 1002 _congraph(NULL), 1003 _replay_inline_data(NULL), 1004 _number_of_mh_late_inlines(0), 1005 _inlining_progress(false), 1006 _inlining_incrementally(false), 1007 _print_inlining_list(NULL), 1008 _print_inlining_stream(NULL), 1009 _print_inlining_idx(0), 1010 _print_inlining_output(NULL), 1011 _allowed_reasons(0), 1012 _interpreter_frame_size(0), 1013 _max_node_limit(MaxNodeLimit), 1014 _has_reserved_stack_access(false) { 1015 C = this; 1016 1017 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1018 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1019 1020 #ifndef PRODUCT 1021 set_print_assembly(PrintFrameConverterAssembly); 1022 set_parsed_irreducible_loop(false); 1023 #endif 1024 set_has_irreducible_loop(false); // no loops 1025 1026 CompileWrapper cw(this); 1027 Init(/*AliasLevel=*/ 0); 1028 init_tf((*generator)()); 1029 1030 { 1031 // The following is a dummy for the sake of GraphKit::gen_stub 1032 Unique_Node_List for_igvn(comp_arena()); 1033 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1034 PhaseGVN gvn(Thread::current()->resource_area(),255); 1035 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1036 gvn.transform_no_reclaim(top()); 1037 1038 GraphKit kit; 1039 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1040 } 1041 1042 NOT_PRODUCT( verify_graph_edges(); ) 1043 Code_Gen(); 1044 if (failing()) return; 1045 1046 1047 // Entry point will be accessed using compile->stub_entry_point(); 1048 if (code_buffer() == NULL) { 1049 Matcher::soft_match_failure(); 1050 } else { 1051 if (PrintAssembly && (WizardMode || Verbose)) 1052 tty->print_cr("### Stub::%s", stub_name); 1053 1054 if (!failing()) { 1055 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1056 1057 // Make the NMethod 1058 // For now we mark the frame as never safe for profile stackwalking 1059 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1060 code_buffer(), 1061 CodeOffsets::frame_never_safe, 1062 // _code_offsets.value(CodeOffsets::Frame_Complete), 1063 frame_size_in_words(), 1064 _oop_map_set, 1065 save_arg_registers); 1066 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1067 1068 _stub_entry_point = rs->entry_point(); 1069 } 1070 } 1071 } 1072 1073 //------------------------------Init------------------------------------------- 1074 // Prepare for a single compilation 1075 void Compile::Init(int aliaslevel) { 1076 _unique = 0; 1077 _regalloc = NULL; 1078 1079 _tf = NULL; // filled in later 1080 _top = NULL; // cached later 1081 _matcher = NULL; // filled in later 1082 _cfg = NULL; // filled in later 1083 1084 set_24_bit_selection_and_mode(Use24BitFP, false); 1085 1086 _node_note_array = NULL; 1087 _default_node_notes = NULL; 1088 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1089 1090 _immutable_memory = NULL; // filled in at first inquiry 1091 1092 // Globally visible Nodes 1093 // First set TOP to NULL to give safe behavior during creation of RootNode 1094 set_cached_top_node(NULL); 1095 set_root(new RootNode()); 1096 // Now that you have a Root to point to, create the real TOP 1097 set_cached_top_node( new ConNode(Type::TOP) ); 1098 set_recent_alloc(NULL, NULL); 1099 1100 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1101 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1102 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1103 env()->set_dependencies(new Dependencies(env())); 1104 1105 _fixed_slots = 0; 1106 set_has_split_ifs(false); 1107 set_has_loops(false); // first approximation 1108 set_has_stringbuilder(false); 1109 set_has_boxed_value(false); 1110 _trap_can_recompile = false; // no traps emitted yet 1111 _major_progress = true; // start out assuming good things will happen 1112 set_has_unsafe_access(false); 1113 set_max_vector_size(0); 1114 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1115 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1116 set_decompile_count(0); 1117 1118 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1119 set_num_loop_opts(LoopOptsCount); 1120 set_do_inlining(Inline); 1121 set_max_inline_size(MaxInlineSize); 1122 set_freq_inline_size(FreqInlineSize); 1123 set_do_scheduling(OptoScheduling); 1124 set_do_count_invocations(false); 1125 set_do_method_data_update(false); 1126 1127 set_do_vector_loop(false); 1128 1129 if (AllowVectorizeOnDemand) { 1130 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1131 set_do_vector_loop(true); 1132 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());}) 1133 } else if (has_method() && method()->name() != 0 && 1134 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1135 set_do_vector_loop(true); 1136 } 1137 } 1138 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1139 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());}) 1140 1141 set_age_code(has_method() && method()->profile_aging()); 1142 set_rtm_state(NoRTM); // No RTM lock eliding by default 1143 _max_node_limit = _directive->MaxNodeLimitOption; 1144 1145 #if INCLUDE_RTM_OPT 1146 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1147 int rtm_state = method()->method_data()->rtm_state(); 1148 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1149 // Don't generate RTM lock eliding code. 1150 set_rtm_state(NoRTM); 1151 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1152 // Generate RTM lock eliding code without abort ratio calculation code. 1153 set_rtm_state(UseRTM); 1154 } else if (UseRTMDeopt) { 1155 // Generate RTM lock eliding code and include abort ratio calculation 1156 // code if UseRTMDeopt is on. 1157 set_rtm_state(ProfileRTM); 1158 } 1159 } 1160 #endif 1161 if (debug_info()->recording_non_safepoints()) { 1162 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1163 (comp_arena(), 8, 0, NULL)); 1164 set_default_node_notes(Node_Notes::make(this)); 1165 } 1166 1167 // // -- Initialize types before each compile -- 1168 // // Update cached type information 1169 // if( _method && _method->constants() ) 1170 // Type::update_loaded_types(_method, _method->constants()); 1171 1172 // Init alias_type map. 1173 if (!_do_escape_analysis && aliaslevel == 3) 1174 aliaslevel = 2; // No unique types without escape analysis 1175 _AliasLevel = aliaslevel; 1176 const int grow_ats = 16; 1177 _max_alias_types = grow_ats; 1178 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1179 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1180 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1181 { 1182 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1183 } 1184 // Initialize the first few types. 1185 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1186 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1187 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1188 _num_alias_types = AliasIdxRaw+1; 1189 // Zero out the alias type cache. 1190 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1191 // A NULL adr_type hits in the cache right away. Preload the right answer. 1192 probe_alias_cache(NULL)->_index = AliasIdxTop; 1193 1194 _intrinsics = NULL; 1195 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1196 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1197 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1198 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1199 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1200 register_library_intrinsics(); 1201 #ifdef ASSERT 1202 _type_verify_symmetry = true; 1203 _exception_backedge = false; 1204 #endif 1205 } 1206 1207 //---------------------------init_start---------------------------------------- 1208 // Install the StartNode on this compile object. 1209 void Compile::init_start(StartNode* s) { 1210 if (failing()) 1211 return; // already failing 1212 assert(s == start(), ""); 1213 } 1214 1215 /** 1216 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1217 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1218 * the ideal graph. 1219 */ 1220 StartNode* Compile::start() const { 1221 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1222 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1223 Node* start = root()->fast_out(i); 1224 if (start->is_Start()) { 1225 return start->as_Start(); 1226 } 1227 } 1228 fatal("Did not find Start node!"); 1229 return NULL; 1230 } 1231 1232 //-------------------------------immutable_memory------------------------------------- 1233 // Access immutable memory 1234 Node* Compile::immutable_memory() { 1235 if (_immutable_memory != NULL) { 1236 return _immutable_memory; 1237 } 1238 StartNode* s = start(); 1239 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1240 Node *p = s->fast_out(i); 1241 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1242 _immutable_memory = p; 1243 return _immutable_memory; 1244 } 1245 } 1246 ShouldNotReachHere(); 1247 return NULL; 1248 } 1249 1250 //----------------------set_cached_top_node------------------------------------ 1251 // Install the cached top node, and make sure Node::is_top works correctly. 1252 void Compile::set_cached_top_node(Node* tn) { 1253 if (tn != NULL) verify_top(tn); 1254 Node* old_top = _top; 1255 _top = tn; 1256 // Calling Node::setup_is_top allows the nodes the chance to adjust 1257 // their _out arrays. 1258 if (_top != NULL) _top->setup_is_top(); 1259 if (old_top != NULL) old_top->setup_is_top(); 1260 assert(_top == NULL || top()->is_top(), ""); 1261 } 1262 1263 #ifdef ASSERT 1264 uint Compile::count_live_nodes_by_graph_walk() { 1265 Unique_Node_List useful(comp_arena()); 1266 // Get useful node list by walking the graph. 1267 identify_useful_nodes(useful); 1268 return useful.size(); 1269 } 1270 1271 void Compile::print_missing_nodes() { 1272 1273 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1274 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1275 return; 1276 } 1277 1278 // This is an expensive function. It is executed only when the user 1279 // specifies VerifyIdealNodeCount option or otherwise knows the 1280 // additional work that needs to be done to identify reachable nodes 1281 // by walking the flow graph and find the missing ones using 1282 // _dead_node_list. 1283 1284 Unique_Node_List useful(comp_arena()); 1285 // Get useful node list by walking the graph. 1286 identify_useful_nodes(useful); 1287 1288 uint l_nodes = C->live_nodes(); 1289 uint l_nodes_by_walk = useful.size(); 1290 1291 if (l_nodes != l_nodes_by_walk) { 1292 if (_log != NULL) { 1293 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1294 _log->stamp(); 1295 _log->end_head(); 1296 } 1297 VectorSet& useful_member_set = useful.member_set(); 1298 int last_idx = l_nodes_by_walk; 1299 for (int i = 0; i < last_idx; i++) { 1300 if (useful_member_set.test(i)) { 1301 if (_dead_node_list.test(i)) { 1302 if (_log != NULL) { 1303 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1304 } 1305 if (PrintIdealNodeCount) { 1306 // Print the log message to tty 1307 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1308 useful.at(i)->dump(); 1309 } 1310 } 1311 } 1312 else if (! _dead_node_list.test(i)) { 1313 if (_log != NULL) { 1314 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1315 } 1316 if (PrintIdealNodeCount) { 1317 // Print the log message to tty 1318 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1319 } 1320 } 1321 } 1322 if (_log != NULL) { 1323 _log->tail("mismatched_nodes"); 1324 } 1325 } 1326 } 1327 void Compile::record_modified_node(Node* n) { 1328 if (_modified_nodes != NULL && !_inlining_incrementally && 1329 n->outcnt() != 0 && !n->is_Con()) { 1330 _modified_nodes->push(n); 1331 } 1332 } 1333 1334 void Compile::remove_modified_node(Node* n) { 1335 if (_modified_nodes != NULL) { 1336 _modified_nodes->remove(n); 1337 } 1338 } 1339 #endif 1340 1341 #ifndef PRODUCT 1342 void Compile::verify_top(Node* tn) const { 1343 if (tn != NULL) { 1344 assert(tn->is_Con(), "top node must be a constant"); 1345 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1346 assert(tn->in(0) != NULL, "must have live top node"); 1347 } 1348 } 1349 #endif 1350 1351 1352 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1353 1354 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1355 guarantee(arr != NULL, ""); 1356 int num_blocks = arr->length(); 1357 if (grow_by < num_blocks) grow_by = num_blocks; 1358 int num_notes = grow_by * _node_notes_block_size; 1359 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1360 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1361 while (num_notes > 0) { 1362 arr->append(notes); 1363 notes += _node_notes_block_size; 1364 num_notes -= _node_notes_block_size; 1365 } 1366 assert(num_notes == 0, "exact multiple, please"); 1367 } 1368 1369 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1370 if (source == NULL || dest == NULL) return false; 1371 1372 if (dest->is_Con()) 1373 return false; // Do not push debug info onto constants. 1374 1375 #ifdef ASSERT 1376 // Leave a bread crumb trail pointing to the original node: 1377 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1378 dest->set_debug_orig(source); 1379 } 1380 #endif 1381 1382 if (node_note_array() == NULL) 1383 return false; // Not collecting any notes now. 1384 1385 // This is a copy onto a pre-existing node, which may already have notes. 1386 // If both nodes have notes, do not overwrite any pre-existing notes. 1387 Node_Notes* source_notes = node_notes_at(source->_idx); 1388 if (source_notes == NULL || source_notes->is_clear()) return false; 1389 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1390 if (dest_notes == NULL || dest_notes->is_clear()) { 1391 return set_node_notes_at(dest->_idx, source_notes); 1392 } 1393 1394 Node_Notes merged_notes = (*source_notes); 1395 // The order of operations here ensures that dest notes will win... 1396 merged_notes.update_from(dest_notes); 1397 return set_node_notes_at(dest->_idx, &merged_notes); 1398 } 1399 1400 1401 //--------------------------allow_range_check_smearing------------------------- 1402 // Gating condition for coalescing similar range checks. 1403 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1404 // single covering check that is at least as strong as any of them. 1405 // If the optimization succeeds, the simplified (strengthened) range check 1406 // will always succeed. If it fails, we will deopt, and then give up 1407 // on the optimization. 1408 bool Compile::allow_range_check_smearing() const { 1409 // If this method has already thrown a range-check, 1410 // assume it was because we already tried range smearing 1411 // and it failed. 1412 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1413 return !already_trapped; 1414 } 1415 1416 1417 //------------------------------flatten_alias_type----------------------------- 1418 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1419 int offset = tj->offset(); 1420 TypePtr::PTR ptr = tj->ptr(); 1421 1422 // Known instance (scalarizable allocation) alias only with itself. 1423 bool is_known_inst = tj->isa_oopptr() != NULL && 1424 tj->is_oopptr()->is_known_instance(); 1425 1426 // Process weird unsafe references. 1427 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1428 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1429 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1430 tj = TypeOopPtr::BOTTOM; 1431 ptr = tj->ptr(); 1432 offset = tj->offset(); 1433 } 1434 1435 // Array pointers need some flattening 1436 const TypeAryPtr *ta = tj->isa_aryptr(); 1437 if (ta && ta->is_stable()) { 1438 // Erase stability property for alias analysis. 1439 tj = ta = ta->cast_to_stable(false); 1440 } 1441 if( ta && is_known_inst ) { 1442 if ( offset != Type::OffsetBot && 1443 offset > arrayOopDesc::length_offset_in_bytes() ) { 1444 offset = Type::OffsetBot; // Flatten constant access into array body only 1445 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1446 } 1447 } else if( ta && _AliasLevel >= 2 ) { 1448 // For arrays indexed by constant indices, we flatten the alias 1449 // space to include all of the array body. Only the header, klass 1450 // and array length can be accessed un-aliased. 1451 if( offset != Type::OffsetBot ) { 1452 if( ta->const_oop() ) { // MethodData* or Method* 1453 offset = Type::OffsetBot; // Flatten constant access into array body 1454 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1455 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1456 // range is OK as-is. 1457 tj = ta = TypeAryPtr::RANGE; 1458 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1459 tj = TypeInstPtr::KLASS; // all klass loads look alike 1460 ta = TypeAryPtr::RANGE; // generic ignored junk 1461 ptr = TypePtr::BotPTR; 1462 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1463 tj = TypeInstPtr::MARK; 1464 ta = TypeAryPtr::RANGE; // generic ignored junk 1465 ptr = TypePtr::BotPTR; 1466 } else { // Random constant offset into array body 1467 offset = Type::OffsetBot; // Flatten constant access into array body 1468 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1469 } 1470 } 1471 // Arrays of fixed size alias with arrays of unknown size. 1472 if (ta->size() != TypeInt::POS) { 1473 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1474 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1475 } 1476 // Arrays of known objects become arrays of unknown objects. 1477 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1478 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1479 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1480 } 1481 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1482 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1483 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1484 } 1485 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1486 // cannot be distinguished by bytecode alone. 1487 if (ta->elem() == TypeInt::BOOL) { 1488 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1489 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1490 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1491 } 1492 // During the 2nd round of IterGVN, NotNull castings are removed. 1493 // Make sure the Bottom and NotNull variants alias the same. 1494 // Also, make sure exact and non-exact variants alias the same. 1495 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1496 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1497 } 1498 } 1499 1500 // Oop pointers need some flattening 1501 const TypeInstPtr *to = tj->isa_instptr(); 1502 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1503 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1504 if( ptr == TypePtr::Constant ) { 1505 if (to->klass() != ciEnv::current()->Class_klass() || 1506 offset < k->size_helper() * wordSize) { 1507 // No constant oop pointers (such as Strings); they alias with 1508 // unknown strings. 1509 assert(!is_known_inst, "not scalarizable allocation"); 1510 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1511 } 1512 } else if( is_known_inst ) { 1513 tj = to; // Keep NotNull and klass_is_exact for instance type 1514 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1515 // During the 2nd round of IterGVN, NotNull castings are removed. 1516 // Make sure the Bottom and NotNull variants alias the same. 1517 // Also, make sure exact and non-exact variants alias the same. 1518 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1519 } 1520 if (to->speculative() != NULL) { 1521 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1522 } 1523 // Canonicalize the holder of this field 1524 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1525 // First handle header references such as a LoadKlassNode, even if the 1526 // object's klass is unloaded at compile time (4965979). 1527 if (!is_known_inst) { // Do it only for non-instance types 1528 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1529 } 1530 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1531 // Static fields are in the space above the normal instance 1532 // fields in the java.lang.Class instance. 1533 if (to->klass() != ciEnv::current()->Class_klass()) { 1534 to = NULL; 1535 tj = TypeOopPtr::BOTTOM; 1536 offset = tj->offset(); 1537 } 1538 } else { 1539 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1540 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1541 if( is_known_inst ) { 1542 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1543 } else { 1544 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1545 } 1546 } 1547 } 1548 } 1549 1550 // Klass pointers to object array klasses need some flattening 1551 const TypeKlassPtr *tk = tj->isa_klassptr(); 1552 if( tk ) { 1553 // If we are referencing a field within a Klass, we need 1554 // to assume the worst case of an Object. Both exact and 1555 // inexact types must flatten to the same alias class so 1556 // use NotNull as the PTR. 1557 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1558 1559 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1560 TypeKlassPtr::OBJECT->klass(), 1561 offset); 1562 } 1563 1564 ciKlass* klass = tk->klass(); 1565 if( klass->is_obj_array_klass() ) { 1566 ciKlass* k = TypeAryPtr::OOPS->klass(); 1567 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1568 k = TypeInstPtr::BOTTOM->klass(); 1569 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1570 } 1571 1572 // Check for precise loads from the primary supertype array and force them 1573 // to the supertype cache alias index. Check for generic array loads from 1574 // the primary supertype array and also force them to the supertype cache 1575 // alias index. Since the same load can reach both, we need to merge 1576 // these 2 disparate memories into the same alias class. Since the 1577 // primary supertype array is read-only, there's no chance of confusion 1578 // where we bypass an array load and an array store. 1579 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1580 if (offset == Type::OffsetBot || 1581 (offset >= primary_supers_offset && 1582 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1583 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1584 offset = in_bytes(Klass::secondary_super_cache_offset()); 1585 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1586 } 1587 } 1588 1589 // Flatten all Raw pointers together. 1590 if (tj->base() == Type::RawPtr) 1591 tj = TypeRawPtr::BOTTOM; 1592 1593 if (tj->base() == Type::AnyPtr) 1594 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1595 1596 // Flatten all to bottom for now 1597 switch( _AliasLevel ) { 1598 case 0: 1599 tj = TypePtr::BOTTOM; 1600 break; 1601 case 1: // Flatten to: oop, static, field or array 1602 switch (tj->base()) { 1603 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1604 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1605 case Type::AryPtr: // do not distinguish arrays at all 1606 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1607 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1608 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1609 default: ShouldNotReachHere(); 1610 } 1611 break; 1612 case 2: // No collapsing at level 2; keep all splits 1613 case 3: // No collapsing at level 3; keep all splits 1614 break; 1615 default: 1616 Unimplemented(); 1617 } 1618 1619 offset = tj->offset(); 1620 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1621 1622 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1623 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1624 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1625 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1626 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1627 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1628 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1629 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1630 assert( tj->ptr() != TypePtr::TopPTR && 1631 tj->ptr() != TypePtr::AnyNull && 1632 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1633 // assert( tj->ptr() != TypePtr::Constant || 1634 // tj->base() == Type::RawPtr || 1635 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1636 1637 return tj; 1638 } 1639 1640 void Compile::AliasType::Init(int i, const TypePtr* at) { 1641 _index = i; 1642 _adr_type = at; 1643 _field = NULL; 1644 _element = NULL; 1645 _is_rewritable = true; // default 1646 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1647 if (atoop != NULL && atoop->is_known_instance()) { 1648 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1649 _general_index = Compile::current()->get_alias_index(gt); 1650 } else { 1651 _general_index = 0; 1652 } 1653 } 1654 1655 BasicType Compile::AliasType::basic_type() const { 1656 if (element() != NULL) { 1657 const Type* element = adr_type()->is_aryptr()->elem(); 1658 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1659 } if (field() != NULL) { 1660 return field()->layout_type(); 1661 } else { 1662 return T_ILLEGAL; // unknown 1663 } 1664 } 1665 1666 //---------------------------------print_on------------------------------------ 1667 #ifndef PRODUCT 1668 void Compile::AliasType::print_on(outputStream* st) { 1669 if (index() < 10) 1670 st->print("@ <%d> ", index()); 1671 else st->print("@ <%d>", index()); 1672 st->print(is_rewritable() ? " " : " RO"); 1673 int offset = adr_type()->offset(); 1674 if (offset == Type::OffsetBot) 1675 st->print(" +any"); 1676 else st->print(" +%-3d", offset); 1677 st->print(" in "); 1678 adr_type()->dump_on(st); 1679 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1680 if (field() != NULL && tjp) { 1681 if (tjp->klass() != field()->holder() || 1682 tjp->offset() != field()->offset_in_bytes()) { 1683 st->print(" != "); 1684 field()->print(); 1685 st->print(" ***"); 1686 } 1687 } 1688 } 1689 1690 void print_alias_types() { 1691 Compile* C = Compile::current(); 1692 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1693 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1694 C->alias_type(idx)->print_on(tty); 1695 tty->cr(); 1696 } 1697 } 1698 #endif 1699 1700 1701 //----------------------------probe_alias_cache-------------------------------- 1702 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1703 intptr_t key = (intptr_t) adr_type; 1704 key ^= key >> logAliasCacheSize; 1705 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1706 } 1707 1708 1709 //-----------------------------grow_alias_types-------------------------------- 1710 void Compile::grow_alias_types() { 1711 const int old_ats = _max_alias_types; // how many before? 1712 const int new_ats = old_ats; // how many more? 1713 const int grow_ats = old_ats+new_ats; // how many now? 1714 _max_alias_types = grow_ats; 1715 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1716 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1717 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1718 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1719 } 1720 1721 1722 //--------------------------------find_alias_type------------------------------ 1723 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1724 if (_AliasLevel == 0) 1725 return alias_type(AliasIdxBot); 1726 1727 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1728 if (ace->_adr_type == adr_type) { 1729 return alias_type(ace->_index); 1730 } 1731 1732 // Handle special cases. 1733 if (adr_type == NULL) return alias_type(AliasIdxTop); 1734 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1735 1736 // Do it the slow way. 1737 const TypePtr* flat = flatten_alias_type(adr_type); 1738 1739 #ifdef ASSERT 1740 { 1741 ResourceMark rm; 1742 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1743 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1744 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1745 Type::str(adr_type)); 1746 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1747 const TypeOopPtr* foop = flat->is_oopptr(); 1748 // Scalarizable allocations have exact klass always. 1749 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1750 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1751 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1752 Type::str(foop), Type::str(xoop)); 1753 } 1754 } 1755 #endif 1756 1757 int idx = AliasIdxTop; 1758 for (int i = 0; i < num_alias_types(); i++) { 1759 if (alias_type(i)->adr_type() == flat) { 1760 idx = i; 1761 break; 1762 } 1763 } 1764 1765 if (idx == AliasIdxTop) { 1766 if (no_create) return NULL; 1767 // Grow the array if necessary. 1768 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1769 // Add a new alias type. 1770 idx = _num_alias_types++; 1771 _alias_types[idx]->Init(idx, flat); 1772 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1773 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1774 if (flat->isa_instptr()) { 1775 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1776 && flat->is_instptr()->klass() == env()->Class_klass()) 1777 alias_type(idx)->set_rewritable(false); 1778 } 1779 if (flat->isa_aryptr()) { 1780 #ifdef ASSERT 1781 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1782 // (T_BYTE has the weakest alignment and size restrictions...) 1783 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1784 #endif 1785 if (flat->offset() == TypePtr::OffsetBot) { 1786 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1787 } 1788 } 1789 if (flat->isa_klassptr()) { 1790 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1791 alias_type(idx)->set_rewritable(false); 1792 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1793 alias_type(idx)->set_rewritable(false); 1794 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1795 alias_type(idx)->set_rewritable(false); 1796 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1797 alias_type(idx)->set_rewritable(false); 1798 } 1799 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1800 // but the base pointer type is not distinctive enough to identify 1801 // references into JavaThread.) 1802 1803 // Check for final fields. 1804 const TypeInstPtr* tinst = flat->isa_instptr(); 1805 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1806 ciField* field; 1807 if (tinst->const_oop() != NULL && 1808 tinst->klass() == ciEnv::current()->Class_klass() && 1809 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1810 // static field 1811 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1812 field = k->get_field_by_offset(tinst->offset(), true); 1813 } else { 1814 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1815 field = k->get_field_by_offset(tinst->offset(), false); 1816 } 1817 assert(field == NULL || 1818 original_field == NULL || 1819 (field->holder() == original_field->holder() && 1820 field->offset() == original_field->offset() && 1821 field->is_static() == original_field->is_static()), "wrong field?"); 1822 // Set field() and is_rewritable() attributes. 1823 if (field != NULL) alias_type(idx)->set_field(field); 1824 } 1825 } 1826 1827 // Fill the cache for next time. 1828 ace->_adr_type = adr_type; 1829 ace->_index = idx; 1830 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1831 1832 // Might as well try to fill the cache for the flattened version, too. 1833 AliasCacheEntry* face = probe_alias_cache(flat); 1834 if (face->_adr_type == NULL) { 1835 face->_adr_type = flat; 1836 face->_index = idx; 1837 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1838 } 1839 1840 return alias_type(idx); 1841 } 1842 1843 1844 Compile::AliasType* Compile::alias_type(ciField* field) { 1845 const TypeOopPtr* t; 1846 if (field->is_static()) 1847 t = TypeInstPtr::make(field->holder()->java_mirror()); 1848 else 1849 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1850 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1851 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1852 return atp; 1853 } 1854 1855 1856 //------------------------------have_alias_type-------------------------------- 1857 bool Compile::have_alias_type(const TypePtr* adr_type) { 1858 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1859 if (ace->_adr_type == adr_type) { 1860 return true; 1861 } 1862 1863 // Handle special cases. 1864 if (adr_type == NULL) return true; 1865 if (adr_type == TypePtr::BOTTOM) return true; 1866 1867 return find_alias_type(adr_type, true, NULL) != NULL; 1868 } 1869 1870 //-----------------------------must_alias-------------------------------------- 1871 // True if all values of the given address type are in the given alias category. 1872 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1873 if (alias_idx == AliasIdxBot) return true; // the universal category 1874 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1875 if (alias_idx == AliasIdxTop) return false; // the empty category 1876 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1877 1878 // the only remaining possible overlap is identity 1879 int adr_idx = get_alias_index(adr_type); 1880 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1881 assert(adr_idx == alias_idx || 1882 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1883 && adr_type != TypeOopPtr::BOTTOM), 1884 "should not be testing for overlap with an unsafe pointer"); 1885 return adr_idx == alias_idx; 1886 } 1887 1888 //------------------------------can_alias-------------------------------------- 1889 // True if any values of the given address type are in the given alias category. 1890 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1891 if (alias_idx == AliasIdxTop) return false; // the empty category 1892 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1893 // Known instance doesn't alias with bottom memory 1894 if (alias_idx == AliasIdxBot) return !adr_type->is_known_instance(); // the universal category 1895 if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins 1896 1897 // the only remaining possible overlap is identity 1898 int adr_idx = get_alias_index(adr_type); 1899 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1900 return adr_idx == alias_idx; 1901 } 1902 1903 1904 1905 //---------------------------pop_warm_call------------------------------------- 1906 WarmCallInfo* Compile::pop_warm_call() { 1907 WarmCallInfo* wci = _warm_calls; 1908 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1909 return wci; 1910 } 1911 1912 //----------------------------Inline_Warm-------------------------------------- 1913 int Compile::Inline_Warm() { 1914 // If there is room, try to inline some more warm call sites. 1915 // %%% Do a graph index compaction pass when we think we're out of space? 1916 if (!InlineWarmCalls) return 0; 1917 1918 int calls_made_hot = 0; 1919 int room_to_grow = NodeCountInliningCutoff - unique(); 1920 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1921 int amount_grown = 0; 1922 WarmCallInfo* call; 1923 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1924 int est_size = (int)call->size(); 1925 if (est_size > (room_to_grow - amount_grown)) { 1926 // This one won't fit anyway. Get rid of it. 1927 call->make_cold(); 1928 continue; 1929 } 1930 call->make_hot(); 1931 calls_made_hot++; 1932 amount_grown += est_size; 1933 amount_to_grow -= est_size; 1934 } 1935 1936 if (calls_made_hot > 0) set_major_progress(); 1937 return calls_made_hot; 1938 } 1939 1940 1941 //----------------------------Finish_Warm-------------------------------------- 1942 void Compile::Finish_Warm() { 1943 if (!InlineWarmCalls) return; 1944 if (failing()) return; 1945 if (warm_calls() == NULL) return; 1946 1947 // Clean up loose ends, if we are out of space for inlining. 1948 WarmCallInfo* call; 1949 while ((call = pop_warm_call()) != NULL) { 1950 call->make_cold(); 1951 } 1952 } 1953 1954 //---------------------cleanup_loop_predicates----------------------- 1955 // Remove the opaque nodes that protect the predicates so that all unused 1956 // checks and uncommon_traps will be eliminated from the ideal graph 1957 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1958 if (predicate_count()==0) return; 1959 for (int i = predicate_count(); i > 0; i--) { 1960 Node * n = predicate_opaque1_node(i-1); 1961 assert(n->Opcode() == Op_Opaque1, "must be"); 1962 igvn.replace_node(n, n->in(1)); 1963 } 1964 assert(predicate_count()==0, "should be clean!"); 1965 } 1966 1967 void Compile::add_range_check_cast(Node* n) { 1968 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1969 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1970 _range_check_casts->append(n); 1971 } 1972 1973 // Remove all range check dependent CastIINodes. 1974 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1975 for (int i = range_check_cast_count(); i > 0; i--) { 1976 Node* cast = range_check_cast_node(i-1); 1977 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1978 igvn.replace_node(cast, cast->in(1)); 1979 } 1980 assert(range_check_cast_count() == 0, "should be empty"); 1981 } 1982 1983 void Compile::add_opaque4_node(Node* n) { 1984 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1985 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1986 _opaque4_nodes->append(n); 1987 } 1988 1989 // Remove all Opaque4 nodes. 1990 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 1991 for (int i = opaque4_count(); i > 0; i--) { 1992 Node* opaq = opaque4_node(i-1); 1993 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 1994 // With Opaque4 nodes, the expectation is that the test of input 1 1995 // is always equal to the constant value of input 2. So we can 1996 // remove the Opaque4 and replace it by input 2. In debug builds, 1997 // leave the non constant test in instead to sanity check that it 1998 // never fails (if it does, that subgraph was constructed so, at 1999 // runtime, a Halt node is executed). 2000 #ifdef ASSERT 2001 igvn.replace_node(opaq, opaq->in(1)); 2002 #else 2003 igvn.replace_node(opaq, opaq->in(2)); 2004 #endif 2005 } 2006 assert(opaque4_count() == 0, "should be empty"); 2007 } 2008 2009 // StringOpts and late inlining of string methods 2010 void Compile::inline_string_calls(bool parse_time) { 2011 { 2012 // remove useless nodes to make the usage analysis simpler 2013 ResourceMark rm; 2014 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2015 } 2016 2017 { 2018 ResourceMark rm; 2019 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2020 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2021 print_method(PHASE_AFTER_STRINGOPTS, 3); 2022 } 2023 2024 // now inline anything that we skipped the first time around 2025 if (!parse_time) { 2026 _late_inlines_pos = _late_inlines.length(); 2027 } 2028 2029 while (_string_late_inlines.length() > 0) { 2030 CallGenerator* cg = _string_late_inlines.pop(); 2031 cg->do_late_inline(); 2032 if (failing()) return; 2033 } 2034 _string_late_inlines.trunc_to(0); 2035 } 2036 2037 // Late inlining of boxing methods 2038 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2039 if (_boxing_late_inlines.length() > 0) { 2040 assert(has_boxed_value(), "inconsistent"); 2041 2042 PhaseGVN* gvn = initial_gvn(); 2043 set_inlining_incrementally(true); 2044 2045 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2046 for_igvn()->clear(); 2047 gvn->replace_with(&igvn); 2048 2049 _late_inlines_pos = _late_inlines.length(); 2050 2051 while (_boxing_late_inlines.length() > 0) { 2052 CallGenerator* cg = _boxing_late_inlines.pop(); 2053 cg->do_late_inline(); 2054 if (failing()) return; 2055 } 2056 _boxing_late_inlines.trunc_to(0); 2057 2058 { 2059 ResourceMark rm; 2060 PhaseRemoveUseless pru(gvn, for_igvn()); 2061 } 2062 2063 igvn = PhaseIterGVN(gvn); 2064 igvn.optimize(); 2065 2066 set_inlining_progress(false); 2067 set_inlining_incrementally(false); 2068 } 2069 } 2070 2071 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2072 assert(IncrementalInline, "incremental inlining should be on"); 2073 PhaseGVN* gvn = initial_gvn(); 2074 2075 set_inlining_progress(false); 2076 for_igvn()->clear(); 2077 gvn->replace_with(&igvn); 2078 2079 { 2080 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2081 int i = 0; 2082 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2083 CallGenerator* cg = _late_inlines.at(i); 2084 _late_inlines_pos = i+1; 2085 cg->do_late_inline(); 2086 if (failing()) return; 2087 } 2088 int j = 0; 2089 for (; i < _late_inlines.length(); i++, j++) { 2090 _late_inlines.at_put(j, _late_inlines.at(i)); 2091 } 2092 _late_inlines.trunc_to(j); 2093 } 2094 2095 { 2096 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2097 ResourceMark rm; 2098 PhaseRemoveUseless pru(gvn, for_igvn()); 2099 } 2100 2101 { 2102 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2103 igvn = PhaseIterGVN(gvn); 2104 } 2105 } 2106 2107 // Perform incremental inlining until bound on number of live nodes is reached 2108 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2109 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2110 2111 PhaseGVN* gvn = initial_gvn(); 2112 2113 set_inlining_incrementally(true); 2114 set_inlining_progress(true); 2115 uint low_live_nodes = 0; 2116 2117 while(inlining_progress() && _late_inlines.length() > 0) { 2118 2119 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2120 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2121 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2122 // PhaseIdealLoop is expensive so we only try it once we are 2123 // out of live nodes and we only try it again if the previous 2124 // helped got the number of nodes down significantly 2125 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2126 if (failing()) return; 2127 low_live_nodes = live_nodes(); 2128 _major_progress = true; 2129 } 2130 2131 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2132 break; 2133 } 2134 } 2135 2136 inline_incrementally_one(igvn); 2137 2138 if (failing()) return; 2139 2140 { 2141 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2142 igvn.optimize(); 2143 } 2144 2145 if (failing()) return; 2146 } 2147 2148 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2149 2150 if (_string_late_inlines.length() > 0) { 2151 assert(has_stringbuilder(), "inconsistent"); 2152 for_igvn()->clear(); 2153 initial_gvn()->replace_with(&igvn); 2154 2155 inline_string_calls(false); 2156 2157 if (failing()) return; 2158 2159 { 2160 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2161 ResourceMark rm; 2162 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2163 } 2164 2165 { 2166 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2167 igvn = PhaseIterGVN(gvn); 2168 igvn.optimize(); 2169 } 2170 } 2171 2172 set_inlining_incrementally(false); 2173 } 2174 2175 2176 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) { 2177 if(loop_opts_cnt > 0) { 2178 debug_only( int cnt = 0; ); 2179 while(major_progress() && (loop_opts_cnt > 0)) { 2180 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2181 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2182 PhaseIdealLoop ideal_loop(igvn, mode); 2183 loop_opts_cnt--; 2184 if (failing()) return false; 2185 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2186 } 2187 } 2188 return true; 2189 } 2190 2191 // Remove edges from "root" to each SafePoint at a backward branch. 2192 // They were inserted during parsing (see add_safepoint()) to make 2193 // infinite loops without calls or exceptions visible to root, i.e., 2194 // useful. 2195 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2196 Node *r = root(); 2197 if (r != NULL) { 2198 for (uint i = r->req(); i < r->len(); ++i) { 2199 Node *n = r->in(i); 2200 if (n != NULL && n->is_SafePoint()) { 2201 r->rm_prec(i); 2202 if (n->outcnt() == 0) { 2203 igvn.remove_dead_node(n); 2204 } 2205 --i; 2206 } 2207 } 2208 // Parsing may have added top inputs to the root node (Path 2209 // leading to the Halt node proven dead). Make sure we get a 2210 // chance to clean them up. 2211 igvn._worklist.push(r); 2212 igvn.optimize(); 2213 } 2214 } 2215 2216 //------------------------------Optimize--------------------------------------- 2217 // Given a graph, optimize it. 2218 void Compile::Optimize() { 2219 TracePhase tp("optimizer", &timers[_t_optimizer]); 2220 2221 #ifndef PRODUCT 2222 if (_directive->BreakAtCompileOption) { 2223 BREAKPOINT; 2224 } 2225 2226 #endif 2227 2228 #ifdef ASSERT 2229 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2230 bs->verify_gc_barriers(true); 2231 #endif 2232 2233 ResourceMark rm; 2234 int loop_opts_cnt; 2235 2236 print_inlining_reinit(); 2237 2238 NOT_PRODUCT( verify_graph_edges(); ) 2239 2240 print_method(PHASE_AFTER_PARSING); 2241 2242 { 2243 // Iterative Global Value Numbering, including ideal transforms 2244 // Initialize IterGVN with types and values from parse-time GVN 2245 PhaseIterGVN igvn(initial_gvn()); 2246 #ifdef ASSERT 2247 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2248 #endif 2249 { 2250 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2251 igvn.optimize(); 2252 } 2253 2254 if (failing()) return; 2255 2256 print_method(PHASE_ITER_GVN1, 2); 2257 2258 inline_incrementally(igvn); 2259 2260 print_method(PHASE_INCREMENTAL_INLINE, 2); 2261 2262 if (failing()) return; 2263 2264 if (eliminate_boxing()) { 2265 // Inline valueOf() methods now. 2266 inline_boxing_calls(igvn); 2267 2268 if (AlwaysIncrementalInline) { 2269 inline_incrementally(igvn); 2270 } 2271 2272 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2273 2274 if (failing()) return; 2275 } 2276 2277 // Now that all inlining is over, cut edge from root to loop 2278 // safepoints 2279 remove_root_to_sfpts_edges(igvn); 2280 2281 // Remove the speculative part of types and clean up the graph from 2282 // the extra CastPP nodes whose only purpose is to carry them. Do 2283 // that early so that optimizations are not disrupted by the extra 2284 // CastPP nodes. 2285 remove_speculative_types(igvn); 2286 2287 // No more new expensive nodes will be added to the list from here 2288 // so keep only the actual candidates for optimizations. 2289 cleanup_expensive_nodes(igvn); 2290 2291 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2292 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2293 initial_gvn()->replace_with(&igvn); 2294 for_igvn()->clear(); 2295 Unique_Node_List new_worklist(C->comp_arena()); 2296 { 2297 ResourceMark rm; 2298 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2299 } 2300 set_for_igvn(&new_worklist); 2301 igvn = PhaseIterGVN(initial_gvn()); 2302 igvn.optimize(); 2303 } 2304 2305 // Perform escape analysis 2306 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2307 if (has_loops()) { 2308 // Cleanup graph (remove dead nodes). 2309 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2310 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2311 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2312 if (failing()) return; 2313 } 2314 ConnectionGraph::do_analysis(this, &igvn); 2315 2316 if (failing()) return; 2317 2318 // Optimize out fields loads from scalar replaceable allocations. 2319 igvn.optimize(); 2320 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2321 2322 if (failing()) return; 2323 2324 if (congraph() != NULL && macro_count() > 0) { 2325 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2326 PhaseMacroExpand mexp(igvn); 2327 mexp.eliminate_macro_nodes(); 2328 igvn.set_delay_transform(false); 2329 2330 igvn.optimize(); 2331 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2332 2333 if (failing()) return; 2334 } 2335 } 2336 2337 // Loop transforms on the ideal graph. Range Check Elimination, 2338 // peeling, unrolling, etc. 2339 2340 // Set loop opts counter 2341 loop_opts_cnt = num_loop_opts(); 2342 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2343 { 2344 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2345 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2346 loop_opts_cnt--; 2347 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2348 if (failing()) return; 2349 } 2350 // Loop opts pass if partial peeling occurred in previous pass 2351 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2352 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2353 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2354 loop_opts_cnt--; 2355 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2356 if (failing()) return; 2357 } 2358 // Loop opts pass for loop-unrolling before CCP 2359 if(major_progress() && (loop_opts_cnt > 0)) { 2360 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2361 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2362 loop_opts_cnt--; 2363 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2364 } 2365 if (!failing()) { 2366 // Verify that last round of loop opts produced a valid graph 2367 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2368 PhaseIdealLoop::verify(igvn); 2369 } 2370 } 2371 if (failing()) return; 2372 2373 // Conditional Constant Propagation; 2374 PhaseCCP ccp( &igvn ); 2375 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2376 { 2377 TracePhase tp("ccp", &timers[_t_ccp]); 2378 ccp.do_transform(); 2379 } 2380 print_method(PHASE_CPP1, 2); 2381 2382 assert( true, "Break here to ccp.dump_old2new_map()"); 2383 2384 // Iterative Global Value Numbering, including ideal transforms 2385 { 2386 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2387 igvn = ccp; 2388 igvn.optimize(); 2389 } 2390 2391 print_method(PHASE_ITER_GVN2, 2); 2392 2393 if (failing()) return; 2394 2395 // Loop transforms on the ideal graph. Range Check Elimination, 2396 // peeling, unrolling, etc. 2397 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2398 return; 2399 } 2400 2401 #if INCLUDE_ZGC 2402 if (UseZGC) { 2403 ZBarrierSetC2::find_dominating_barriers(igvn); 2404 } 2405 #endif 2406 2407 if (failing()) return; 2408 2409 // Ensure that major progress is now clear 2410 C->clear_major_progress(); 2411 2412 { 2413 // Verify that all previous optimizations produced a valid graph 2414 // at least to this point, even if no loop optimizations were done. 2415 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2416 PhaseIdealLoop::verify(igvn); 2417 } 2418 2419 if (range_check_cast_count() > 0) { 2420 // No more loop optimizations. Remove all range check dependent CastIINodes. 2421 C->remove_range_check_casts(igvn); 2422 igvn.optimize(); 2423 } 2424 2425 #ifdef ASSERT 2426 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2427 bs->verify_gc_barriers(false); 2428 #endif 2429 2430 { 2431 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2432 PhaseMacroExpand mex(igvn); 2433 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2434 if (mex.expand_macro_nodes()) { 2435 assert(failing(), "must bail out w/ explicit message"); 2436 return; 2437 } 2438 } 2439 2440 #if INCLUDE_SHENANDOAHGC 2441 if (UseShenandoahGC) { 2442 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2); 2443 if (((ShenandoahBarrierSetC2*)BarrierSet::barrier_set()->barrier_set_c2())->expand_barriers(this, igvn)) { 2444 assert(failing(), "must bail out w/ explicit message"); 2445 return; 2446 } 2447 } 2448 #endif 2449 2450 if (opaque4_count() > 0) { 2451 C->remove_opaque4_nodes(igvn); 2452 igvn.optimize(); 2453 } 2454 2455 DEBUG_ONLY( _modified_nodes = NULL; ) 2456 } // (End scope of igvn; run destructor if necessary for asserts.) 2457 2458 process_print_inlining(); 2459 // A method with only infinite loops has no edges entering loops from root 2460 { 2461 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2462 if (final_graph_reshaping()) { 2463 assert(failing(), "must bail out w/ explicit message"); 2464 return; 2465 } 2466 } 2467 2468 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2469 } 2470 2471 2472 //------------------------------Code_Gen--------------------------------------- 2473 // Given a graph, generate code for it 2474 void Compile::Code_Gen() { 2475 if (failing()) { 2476 return; 2477 } 2478 2479 // Perform instruction selection. You might think we could reclaim Matcher 2480 // memory PDQ, but actually the Matcher is used in generating spill code. 2481 // Internals of the Matcher (including some VectorSets) must remain live 2482 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2483 // set a bit in reclaimed memory. 2484 2485 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2486 // nodes. Mapping is only valid at the root of each matched subtree. 2487 NOT_PRODUCT( verify_graph_edges(); ) 2488 2489 Matcher matcher; 2490 _matcher = &matcher; 2491 { 2492 TracePhase tp("matcher", &timers[_t_matcher]); 2493 matcher.match(); 2494 } 2495 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2496 // nodes. Mapping is only valid at the root of each matched subtree. 2497 NOT_PRODUCT( verify_graph_edges(); ) 2498 2499 // If you have too many nodes, or if matching has failed, bail out 2500 check_node_count(0, "out of nodes matching instructions"); 2501 if (failing()) { 2502 return; 2503 } 2504 2505 print_method(PHASE_MATCHING, 2); 2506 2507 // Build a proper-looking CFG 2508 PhaseCFG cfg(node_arena(), root(), matcher); 2509 _cfg = &cfg; 2510 { 2511 TracePhase tp("scheduler", &timers[_t_scheduler]); 2512 bool success = cfg.do_global_code_motion(); 2513 if (!success) { 2514 return; 2515 } 2516 2517 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2518 NOT_PRODUCT( verify_graph_edges(); ) 2519 debug_only( cfg.verify(); ) 2520 } 2521 2522 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2523 _regalloc = ®alloc; 2524 { 2525 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2526 // Perform register allocation. After Chaitin, use-def chains are 2527 // no longer accurate (at spill code) and so must be ignored. 2528 // Node->LRG->reg mappings are still accurate. 2529 _regalloc->Register_Allocate(); 2530 2531 // Bail out if the allocator builds too many nodes 2532 if (failing()) { 2533 return; 2534 } 2535 } 2536 2537 // Prior to register allocation we kept empty basic blocks in case the 2538 // the allocator needed a place to spill. After register allocation we 2539 // are not adding any new instructions. If any basic block is empty, we 2540 // can now safely remove it. 2541 { 2542 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2543 cfg.remove_empty_blocks(); 2544 if (do_freq_based_layout()) { 2545 PhaseBlockLayout layout(cfg); 2546 } else { 2547 cfg.set_loop_alignment(); 2548 } 2549 cfg.fixup_flow(); 2550 } 2551 2552 // Apply peephole optimizations 2553 if( OptoPeephole ) { 2554 TracePhase tp("peephole", &timers[_t_peephole]); 2555 PhasePeephole peep( _regalloc, cfg); 2556 peep.do_transform(); 2557 } 2558 2559 // Do late expand if CPU requires this. 2560 if (Matcher::require_postalloc_expand) { 2561 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2562 cfg.postalloc_expand(_regalloc); 2563 } 2564 2565 // Convert Nodes to instruction bits in a buffer 2566 { 2567 TraceTime tp("output", &timers[_t_output], CITime); 2568 Output(); 2569 } 2570 2571 print_method(PHASE_FINAL_CODE); 2572 2573 // He's dead, Jim. 2574 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2575 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2576 } 2577 2578 2579 //------------------------------dump_asm--------------------------------------- 2580 // Dump formatted assembly 2581 #ifndef PRODUCT 2582 void Compile::dump_asm(int *pcs, uint pc_limit) { 2583 bool cut_short = false; 2584 tty->print_cr("#"); 2585 tty->print("# "); _tf->dump(); tty->cr(); 2586 tty->print_cr("#"); 2587 2588 // For all blocks 2589 int pc = 0x0; // Program counter 2590 char starts_bundle = ' '; 2591 _regalloc->dump_frame(); 2592 2593 Node *n = NULL; 2594 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2595 if (VMThread::should_terminate()) { 2596 cut_short = true; 2597 break; 2598 } 2599 Block* block = _cfg->get_block(i); 2600 if (block->is_connector() && !Verbose) { 2601 continue; 2602 } 2603 n = block->head(); 2604 if (pcs && n->_idx < pc_limit) { 2605 tty->print("%3.3x ", pcs[n->_idx]); 2606 } else { 2607 tty->print(" "); 2608 } 2609 block->dump_head(_cfg); 2610 if (block->is_connector()) { 2611 tty->print_cr(" # Empty connector block"); 2612 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2613 tty->print_cr(" # Block is sole successor of call"); 2614 } 2615 2616 // For all instructions 2617 Node *delay = NULL; 2618 for (uint j = 0; j < block->number_of_nodes(); j++) { 2619 if (VMThread::should_terminate()) { 2620 cut_short = true; 2621 break; 2622 } 2623 n = block->get_node(j); 2624 if (valid_bundle_info(n)) { 2625 Bundle* bundle = node_bundling(n); 2626 if (bundle->used_in_unconditional_delay()) { 2627 delay = n; 2628 continue; 2629 } 2630 if (bundle->starts_bundle()) { 2631 starts_bundle = '+'; 2632 } 2633 } 2634 2635 if (WizardMode) { 2636 n->dump(); 2637 } 2638 2639 if( !n->is_Region() && // Dont print in the Assembly 2640 !n->is_Phi() && // a few noisely useless nodes 2641 !n->is_Proj() && 2642 !n->is_MachTemp() && 2643 !n->is_SafePointScalarObject() && 2644 !n->is_Catch() && // Would be nice to print exception table targets 2645 !n->is_MergeMem() && // Not very interesting 2646 !n->is_top() && // Debug info table constants 2647 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2648 ) { 2649 if (pcs && n->_idx < pc_limit) 2650 tty->print("%3.3x", pcs[n->_idx]); 2651 else 2652 tty->print(" "); 2653 tty->print(" %c ", starts_bundle); 2654 starts_bundle = ' '; 2655 tty->print("\t"); 2656 n->format(_regalloc, tty); 2657 tty->cr(); 2658 } 2659 2660 // If we have an instruction with a delay slot, and have seen a delay, 2661 // then back up and print it 2662 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2663 assert(delay != NULL, "no unconditional delay instruction"); 2664 if (WizardMode) delay->dump(); 2665 2666 if (node_bundling(delay)->starts_bundle()) 2667 starts_bundle = '+'; 2668 if (pcs && n->_idx < pc_limit) 2669 tty->print("%3.3x", pcs[n->_idx]); 2670 else 2671 tty->print(" "); 2672 tty->print(" %c ", starts_bundle); 2673 starts_bundle = ' '; 2674 tty->print("\t"); 2675 delay->format(_regalloc, tty); 2676 tty->cr(); 2677 delay = NULL; 2678 } 2679 2680 // Dump the exception table as well 2681 if( n->is_Catch() && (Verbose || WizardMode) ) { 2682 // Print the exception table for this offset 2683 _handler_table.print_subtable_for(pc); 2684 } 2685 } 2686 2687 if (pcs && n->_idx < pc_limit) 2688 tty->print_cr("%3.3x", pcs[n->_idx]); 2689 else 2690 tty->cr(); 2691 2692 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2693 2694 } // End of per-block dump 2695 tty->cr(); 2696 2697 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2698 } 2699 #endif 2700 2701 //------------------------------Final_Reshape_Counts--------------------------- 2702 // This class defines counters to help identify when a method 2703 // may/must be executed using hardware with only 24-bit precision. 2704 struct Final_Reshape_Counts : public StackObj { 2705 int _call_count; // count non-inlined 'common' calls 2706 int _float_count; // count float ops requiring 24-bit precision 2707 int _double_count; // count double ops requiring more precision 2708 int _java_call_count; // count non-inlined 'java' calls 2709 int _inner_loop_count; // count loops which need alignment 2710 VectorSet _visited; // Visitation flags 2711 Node_List _tests; // Set of IfNodes & PCTableNodes 2712 2713 Final_Reshape_Counts() : 2714 _call_count(0), _float_count(0), _double_count(0), 2715 _java_call_count(0), _inner_loop_count(0), 2716 _visited( Thread::current()->resource_area() ) { } 2717 2718 void inc_call_count () { _call_count ++; } 2719 void inc_float_count () { _float_count ++; } 2720 void inc_double_count() { _double_count++; } 2721 void inc_java_call_count() { _java_call_count++; } 2722 void inc_inner_loop_count() { _inner_loop_count++; } 2723 2724 int get_call_count () const { return _call_count ; } 2725 int get_float_count () const { return _float_count ; } 2726 int get_double_count() const { return _double_count; } 2727 int get_java_call_count() const { return _java_call_count; } 2728 int get_inner_loop_count() const { return _inner_loop_count; } 2729 }; 2730 2731 #ifdef ASSERT 2732 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2733 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2734 // Make sure the offset goes inside the instance layout. 2735 return k->contains_field_offset(tp->offset()); 2736 // Note that OffsetBot and OffsetTop are very negative. 2737 } 2738 #endif 2739 2740 // Eliminate trivially redundant StoreCMs and accumulate their 2741 // precedence edges. 2742 void Compile::eliminate_redundant_card_marks(Node* n) { 2743 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2744 if (n->in(MemNode::Address)->outcnt() > 1) { 2745 // There are multiple users of the same address so it might be 2746 // possible to eliminate some of the StoreCMs 2747 Node* mem = n->in(MemNode::Memory); 2748 Node* adr = n->in(MemNode::Address); 2749 Node* val = n->in(MemNode::ValueIn); 2750 Node* prev = n; 2751 bool done = false; 2752 // Walk the chain of StoreCMs eliminating ones that match. As 2753 // long as it's a chain of single users then the optimization is 2754 // safe. Eliminating partially redundant StoreCMs would require 2755 // cloning copies down the other paths. 2756 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2757 if (adr == mem->in(MemNode::Address) && 2758 val == mem->in(MemNode::ValueIn)) { 2759 // redundant StoreCM 2760 if (mem->req() > MemNode::OopStore) { 2761 // Hasn't been processed by this code yet. 2762 n->add_prec(mem->in(MemNode::OopStore)); 2763 } else { 2764 // Already converted to precedence edge 2765 for (uint i = mem->req(); i < mem->len(); i++) { 2766 // Accumulate any precedence edges 2767 if (mem->in(i) != NULL) { 2768 n->add_prec(mem->in(i)); 2769 } 2770 } 2771 // Everything above this point has been processed. 2772 done = true; 2773 } 2774 // Eliminate the previous StoreCM 2775 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2776 assert(mem->outcnt() == 0, "should be dead"); 2777 mem->disconnect_inputs(NULL, this); 2778 } else { 2779 prev = mem; 2780 } 2781 mem = prev->in(MemNode::Memory); 2782 } 2783 } 2784 } 2785 2786 //------------------------------final_graph_reshaping_impl---------------------- 2787 // Implement items 1-5 from final_graph_reshaping below. 2788 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2789 2790 if ( n->outcnt() == 0 ) return; // dead node 2791 uint nop = n->Opcode(); 2792 2793 // Check for 2-input instruction with "last use" on right input. 2794 // Swap to left input. Implements item (2). 2795 if( n->req() == 3 && // two-input instruction 2796 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2797 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2798 n->in(2)->outcnt() == 1 &&// right use IS a last use 2799 !n->in(2)->is_Con() ) { // right use is not a constant 2800 // Check for commutative opcode 2801 switch( nop ) { 2802 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2803 case Op_MaxI: case Op_MinI: 2804 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2805 case Op_AndL: case Op_XorL: case Op_OrL: 2806 case Op_AndI: case Op_XorI: case Op_OrI: { 2807 // Move "last use" input to left by swapping inputs 2808 n->swap_edges(1, 2); 2809 break; 2810 } 2811 default: 2812 break; 2813 } 2814 } 2815 2816 #ifdef ASSERT 2817 if( n->is_Mem() ) { 2818 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2819 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2820 // oop will be recorded in oop map if load crosses safepoint 2821 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2822 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2823 "raw memory operations should have control edge"); 2824 } 2825 if (n->is_MemBar()) { 2826 MemBarNode* mb = n->as_MemBar(); 2827 if (mb->trailing_store() || mb->trailing_load_store()) { 2828 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2829 Node* mem = mb->in(MemBarNode::Precedent); 2830 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2831 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2832 } else if (mb->leading()) { 2833 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2834 } 2835 } 2836 #endif 2837 // Count FPU ops and common calls, implements item (3) 2838 switch( nop ) { 2839 // Count all float operations that may use FPU 2840 case Op_AddF: 2841 case Op_SubF: 2842 case Op_MulF: 2843 case Op_DivF: 2844 case Op_NegF: 2845 case Op_ModF: 2846 case Op_ConvI2F: 2847 case Op_ConF: 2848 case Op_CmpF: 2849 case Op_CmpF3: 2850 // case Op_ConvL2F: // longs are split into 32-bit halves 2851 frc.inc_float_count(); 2852 break; 2853 2854 case Op_ConvF2D: 2855 case Op_ConvD2F: 2856 frc.inc_float_count(); 2857 frc.inc_double_count(); 2858 break; 2859 2860 // Count all double operations that may use FPU 2861 case Op_AddD: 2862 case Op_SubD: 2863 case Op_MulD: 2864 case Op_DivD: 2865 case Op_NegD: 2866 case Op_ModD: 2867 case Op_ConvI2D: 2868 case Op_ConvD2I: 2869 // case Op_ConvL2D: // handled by leaf call 2870 // case Op_ConvD2L: // handled by leaf call 2871 case Op_ConD: 2872 case Op_CmpD: 2873 case Op_CmpD3: 2874 frc.inc_double_count(); 2875 break; 2876 case Op_Opaque1: // Remove Opaque Nodes before matching 2877 case Op_Opaque2: // Remove Opaque Nodes before matching 2878 case Op_Opaque3: 2879 n->subsume_by(n->in(1), this); 2880 break; 2881 case Op_CallStaticJava: 2882 case Op_CallJava: 2883 case Op_CallDynamicJava: 2884 frc.inc_java_call_count(); // Count java call site; 2885 case Op_CallRuntime: 2886 case Op_CallLeaf: 2887 case Op_CallLeafNoFP: { 2888 assert (n->is_Call(), ""); 2889 CallNode *call = n->as_Call(); 2890 #if INCLUDE_SHENANDOAHGC 2891 if (UseShenandoahGC && ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) { 2892 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt(); 2893 if (call->req() > cnt) { 2894 assert(call->req() == cnt+1, "only one extra input"); 2895 Node* addp = call->in(cnt); 2896 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?"); 2897 call->del_req(cnt); 2898 } 2899 } 2900 #endif 2901 // Count call sites where the FP mode bit would have to be flipped. 2902 // Do not count uncommon runtime calls: 2903 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2904 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2905 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2906 frc.inc_call_count(); // Count the call site 2907 } else { // See if uncommon argument is shared 2908 Node *n = call->in(TypeFunc::Parms); 2909 int nop = n->Opcode(); 2910 // Clone shared simple arguments to uncommon calls, item (1). 2911 if (n->outcnt() > 1 && 2912 !n->is_Proj() && 2913 nop != Op_CreateEx && 2914 nop != Op_CheckCastPP && 2915 nop != Op_DecodeN && 2916 nop != Op_DecodeNKlass && 2917 !n->is_Mem() && 2918 !n->is_Phi()) { 2919 Node *x = n->clone(); 2920 call->set_req(TypeFunc::Parms, x); 2921 } 2922 } 2923 break; 2924 } 2925 2926 case Op_StoreD: 2927 case Op_LoadD: 2928 case Op_LoadD_unaligned: 2929 frc.inc_double_count(); 2930 goto handle_mem; 2931 case Op_StoreF: 2932 case Op_LoadF: 2933 frc.inc_float_count(); 2934 goto handle_mem; 2935 2936 case Op_StoreCM: 2937 { 2938 // Convert OopStore dependence into precedence edge 2939 Node* prec = n->in(MemNode::OopStore); 2940 n->del_req(MemNode::OopStore); 2941 n->add_prec(prec); 2942 eliminate_redundant_card_marks(n); 2943 } 2944 2945 // fall through 2946 2947 case Op_StoreB: 2948 case Op_StoreC: 2949 case Op_StorePConditional: 2950 case Op_StoreI: 2951 case Op_StoreL: 2952 case Op_StoreIConditional: 2953 case Op_StoreLConditional: 2954 case Op_CompareAndSwapB: 2955 case Op_CompareAndSwapS: 2956 case Op_CompareAndSwapI: 2957 case Op_CompareAndSwapL: 2958 case Op_CompareAndSwapP: 2959 case Op_CompareAndSwapN: 2960 case Op_WeakCompareAndSwapB: 2961 case Op_WeakCompareAndSwapS: 2962 case Op_WeakCompareAndSwapI: 2963 case Op_WeakCompareAndSwapL: 2964 case Op_WeakCompareAndSwapP: 2965 case Op_WeakCompareAndSwapN: 2966 case Op_CompareAndExchangeB: 2967 case Op_CompareAndExchangeS: 2968 case Op_CompareAndExchangeI: 2969 case Op_CompareAndExchangeL: 2970 case Op_CompareAndExchangeP: 2971 case Op_CompareAndExchangeN: 2972 case Op_GetAndAddS: 2973 case Op_GetAndAddB: 2974 case Op_GetAndAddI: 2975 case Op_GetAndAddL: 2976 case Op_GetAndSetS: 2977 case Op_GetAndSetB: 2978 case Op_GetAndSetI: 2979 case Op_GetAndSetL: 2980 case Op_GetAndSetP: 2981 case Op_GetAndSetN: 2982 case Op_StoreP: 2983 case Op_StoreN: 2984 case Op_StoreNKlass: 2985 case Op_LoadB: 2986 case Op_LoadUB: 2987 case Op_LoadUS: 2988 case Op_LoadI: 2989 case Op_LoadKlass: 2990 case Op_LoadNKlass: 2991 case Op_LoadL: 2992 case Op_LoadL_unaligned: 2993 case Op_LoadPLocked: 2994 case Op_LoadP: 2995 #if INCLUDE_ZGC 2996 case Op_LoadBarrierSlowReg: 2997 case Op_LoadBarrierWeakSlowReg: 2998 #endif 2999 case Op_LoadN: 3000 case Op_LoadRange: 3001 case Op_LoadS: { 3002 handle_mem: 3003 #ifdef ASSERT 3004 if( VerifyOptoOopOffsets ) { 3005 assert( n->is_Mem(), "" ); 3006 MemNode *mem = (MemNode*)n; 3007 // Check to see if address types have grounded out somehow. 3008 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3009 assert( !tp || oop_offset_is_sane(tp), "" ); 3010 } 3011 #endif 3012 break; 3013 } 3014 3015 case Op_AddP: { // Assert sane base pointers 3016 Node *addp = n->in(AddPNode::Address); 3017 assert( !addp->is_AddP() || 3018 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3019 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3020 "Base pointers must match (addp %u)", addp->_idx ); 3021 #ifdef _LP64 3022 if ((UseCompressedOops || UseCompressedClassPointers) && 3023 addp->Opcode() == Op_ConP && 3024 addp == n->in(AddPNode::Base) && 3025 n->in(AddPNode::Offset)->is_Con()) { 3026 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3027 // on the platform and on the compressed oops mode. 3028 // Use addressing with narrow klass to load with offset on x86. 3029 // Some platforms can use the constant pool to load ConP. 3030 // Do this transformation here since IGVN will convert ConN back to ConP. 3031 const Type* t = addp->bottom_type(); 3032 bool is_oop = t->isa_oopptr() != NULL; 3033 bool is_klass = t->isa_klassptr() != NULL; 3034 3035 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3036 (is_klass && Matcher::const_klass_prefer_decode())) { 3037 Node* nn = NULL; 3038 3039 int op = is_oop ? Op_ConN : Op_ConNKlass; 3040 3041 // Look for existing ConN node of the same exact type. 3042 Node* r = root(); 3043 uint cnt = r->outcnt(); 3044 for (uint i = 0; i < cnt; i++) { 3045 Node* m = r->raw_out(i); 3046 if (m!= NULL && m->Opcode() == op && 3047 m->bottom_type()->make_ptr() == t) { 3048 nn = m; 3049 break; 3050 } 3051 } 3052 if (nn != NULL) { 3053 // Decode a narrow oop to match address 3054 // [R12 + narrow_oop_reg<<3 + offset] 3055 if (is_oop) { 3056 nn = new DecodeNNode(nn, t); 3057 } else { 3058 nn = new DecodeNKlassNode(nn, t); 3059 } 3060 // Check for succeeding AddP which uses the same Base. 3061 // Otherwise we will run into the assertion above when visiting that guy. 3062 for (uint i = 0; i < n->outcnt(); ++i) { 3063 Node *out_i = n->raw_out(i); 3064 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3065 out_i->set_req(AddPNode::Base, nn); 3066 #ifdef ASSERT 3067 for (uint j = 0; j < out_i->outcnt(); ++j) { 3068 Node *out_j = out_i->raw_out(j); 3069 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3070 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3071 } 3072 #endif 3073 } 3074 } 3075 n->set_req(AddPNode::Base, nn); 3076 n->set_req(AddPNode::Address, nn); 3077 if (addp->outcnt() == 0) { 3078 addp->disconnect_inputs(NULL, this); 3079 } 3080 } 3081 } 3082 } 3083 #endif 3084 // platform dependent reshaping of the address expression 3085 reshape_address(n->as_AddP()); 3086 break; 3087 } 3088 3089 case Op_CastPP: { 3090 // Remove CastPP nodes to gain more freedom during scheduling but 3091 // keep the dependency they encode as control or precedence edges 3092 // (if control is set already) on memory operations. Some CastPP 3093 // nodes don't have a control (don't carry a dependency): skip 3094 // those. 3095 if (n->in(0) != NULL) { 3096 ResourceMark rm; 3097 Unique_Node_List wq; 3098 wq.push(n); 3099 for (uint next = 0; next < wq.size(); ++next) { 3100 Node *m = wq.at(next); 3101 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3102 Node* use = m->fast_out(i); 3103 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3104 use->ensure_control_or_add_prec(n->in(0)); 3105 } else { 3106 switch(use->Opcode()) { 3107 case Op_AddP: 3108 case Op_DecodeN: 3109 case Op_DecodeNKlass: 3110 case Op_CheckCastPP: 3111 case Op_CastPP: 3112 wq.push(use); 3113 break; 3114 } 3115 } 3116 } 3117 } 3118 } 3119 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3120 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3121 Node* in1 = n->in(1); 3122 const Type* t = n->bottom_type(); 3123 Node* new_in1 = in1->clone(); 3124 new_in1->as_DecodeN()->set_type(t); 3125 3126 if (!Matcher::narrow_oop_use_complex_address()) { 3127 // 3128 // x86, ARM and friends can handle 2 adds in addressing mode 3129 // and Matcher can fold a DecodeN node into address by using 3130 // a narrow oop directly and do implicit NULL check in address: 3131 // 3132 // [R12 + narrow_oop_reg<<3 + offset] 3133 // NullCheck narrow_oop_reg 3134 // 3135 // On other platforms (Sparc) we have to keep new DecodeN node and 3136 // use it to do implicit NULL check in address: 3137 // 3138 // decode_not_null narrow_oop_reg, base_reg 3139 // [base_reg + offset] 3140 // NullCheck base_reg 3141 // 3142 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3143 // to keep the information to which NULL check the new DecodeN node 3144 // corresponds to use it as value in implicit_null_check(). 3145 // 3146 new_in1->set_req(0, n->in(0)); 3147 } 3148 3149 n->subsume_by(new_in1, this); 3150 if (in1->outcnt() == 0) { 3151 in1->disconnect_inputs(NULL, this); 3152 } 3153 } else { 3154 n->subsume_by(n->in(1), this); 3155 if (n->outcnt() == 0) { 3156 n->disconnect_inputs(NULL, this); 3157 } 3158 } 3159 break; 3160 } 3161 #ifdef _LP64 3162 case Op_CmpP: 3163 // Do this transformation here to preserve CmpPNode::sub() and 3164 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3165 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3166 Node* in1 = n->in(1); 3167 Node* in2 = n->in(2); 3168 if (!in1->is_DecodeNarrowPtr()) { 3169 in2 = in1; 3170 in1 = n->in(2); 3171 } 3172 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3173 3174 Node* new_in2 = NULL; 3175 if (in2->is_DecodeNarrowPtr()) { 3176 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3177 new_in2 = in2->in(1); 3178 } else if (in2->Opcode() == Op_ConP) { 3179 const Type* t = in2->bottom_type(); 3180 if (t == TypePtr::NULL_PTR) { 3181 assert(in1->is_DecodeN(), "compare klass to null?"); 3182 // Don't convert CmpP null check into CmpN if compressed 3183 // oops implicit null check is not generated. 3184 // This will allow to generate normal oop implicit null check. 3185 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3186 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3187 // 3188 // This transformation together with CastPP transformation above 3189 // will generated code for implicit NULL checks for compressed oops. 3190 // 3191 // The original code after Optimize() 3192 // 3193 // LoadN memory, narrow_oop_reg 3194 // decode narrow_oop_reg, base_reg 3195 // CmpP base_reg, NULL 3196 // CastPP base_reg // NotNull 3197 // Load [base_reg + offset], val_reg 3198 // 3199 // after these transformations will be 3200 // 3201 // LoadN memory, narrow_oop_reg 3202 // CmpN narrow_oop_reg, NULL 3203 // decode_not_null narrow_oop_reg, base_reg 3204 // Load [base_reg + offset], val_reg 3205 // 3206 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3207 // since narrow oops can be used in debug info now (see the code in 3208 // final_graph_reshaping_walk()). 3209 // 3210 // At the end the code will be matched to 3211 // on x86: 3212 // 3213 // Load_narrow_oop memory, narrow_oop_reg 3214 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3215 // NullCheck narrow_oop_reg 3216 // 3217 // and on sparc: 3218 // 3219 // Load_narrow_oop memory, narrow_oop_reg 3220 // decode_not_null narrow_oop_reg, base_reg 3221 // Load [base_reg + offset], val_reg 3222 // NullCheck base_reg 3223 // 3224 } else if (t->isa_oopptr()) { 3225 new_in2 = ConNode::make(t->make_narrowoop()); 3226 } else if (t->isa_klassptr()) { 3227 new_in2 = ConNode::make(t->make_narrowklass()); 3228 } 3229 } 3230 if (new_in2 != NULL) { 3231 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3232 n->subsume_by(cmpN, this); 3233 if (in1->outcnt() == 0) { 3234 in1->disconnect_inputs(NULL, this); 3235 } 3236 if (in2->outcnt() == 0) { 3237 in2->disconnect_inputs(NULL, this); 3238 } 3239 } 3240 } 3241 break; 3242 3243 case Op_DecodeN: 3244 case Op_DecodeNKlass: 3245 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3246 // DecodeN could be pinned when it can't be fold into 3247 // an address expression, see the code for Op_CastPP above. 3248 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3249 break; 3250 3251 case Op_EncodeP: 3252 case Op_EncodePKlass: { 3253 Node* in1 = n->in(1); 3254 if (in1->is_DecodeNarrowPtr()) { 3255 n->subsume_by(in1->in(1), this); 3256 } else if (in1->Opcode() == Op_ConP) { 3257 const Type* t = in1->bottom_type(); 3258 if (t == TypePtr::NULL_PTR) { 3259 assert(t->isa_oopptr(), "null klass?"); 3260 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3261 } else if (t->isa_oopptr()) { 3262 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3263 } else if (t->isa_klassptr()) { 3264 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3265 } 3266 } 3267 if (in1->outcnt() == 0) { 3268 in1->disconnect_inputs(NULL, this); 3269 } 3270 break; 3271 } 3272 3273 case Op_Proj: { 3274 if (OptimizeStringConcat) { 3275 ProjNode* p = n->as_Proj(); 3276 if (p->_is_io_use) { 3277 // Separate projections were used for the exception path which 3278 // are normally removed by a late inline. If it wasn't inlined 3279 // then they will hang around and should just be replaced with 3280 // the original one. 3281 Node* proj = NULL; 3282 // Replace with just one 3283 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3284 Node *use = i.get(); 3285 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3286 proj = use; 3287 break; 3288 } 3289 } 3290 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3291 if (proj != NULL) { 3292 p->subsume_by(proj, this); 3293 } 3294 } 3295 } 3296 break; 3297 } 3298 3299 case Op_Phi: 3300 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3301 // The EncodeP optimization may create Phi with the same edges 3302 // for all paths. It is not handled well by Register Allocator. 3303 Node* unique_in = n->in(1); 3304 assert(unique_in != NULL, ""); 3305 uint cnt = n->req(); 3306 for (uint i = 2; i < cnt; i++) { 3307 Node* m = n->in(i); 3308 assert(m != NULL, ""); 3309 if (unique_in != m) 3310 unique_in = NULL; 3311 } 3312 if (unique_in != NULL) { 3313 n->subsume_by(unique_in, this); 3314 } 3315 } 3316 break; 3317 3318 #endif 3319 3320 #ifdef ASSERT 3321 case Op_CastII: 3322 // Verify that all range check dependent CastII nodes were removed. 3323 if (n->isa_CastII()->has_range_check()) { 3324 n->dump(3); 3325 assert(false, "Range check dependent CastII node was not removed"); 3326 } 3327 break; 3328 #endif 3329 3330 case Op_ModI: 3331 if (UseDivMod) { 3332 // Check if a%b and a/b both exist 3333 Node* d = n->find_similar(Op_DivI); 3334 if (d) { 3335 // Replace them with a fused divmod if supported 3336 if (Matcher::has_match_rule(Op_DivModI)) { 3337 DivModINode* divmod = DivModINode::make(n); 3338 d->subsume_by(divmod->div_proj(), this); 3339 n->subsume_by(divmod->mod_proj(), this); 3340 } else { 3341 // replace a%b with a-((a/b)*b) 3342 Node* mult = new MulINode(d, d->in(2)); 3343 Node* sub = new SubINode(d->in(1), mult); 3344 n->subsume_by(sub, this); 3345 } 3346 } 3347 } 3348 break; 3349 3350 case Op_ModL: 3351 if (UseDivMod) { 3352 // Check if a%b and a/b both exist 3353 Node* d = n->find_similar(Op_DivL); 3354 if (d) { 3355 // Replace them with a fused divmod if supported 3356 if (Matcher::has_match_rule(Op_DivModL)) { 3357 DivModLNode* divmod = DivModLNode::make(n); 3358 d->subsume_by(divmod->div_proj(), this); 3359 n->subsume_by(divmod->mod_proj(), this); 3360 } else { 3361 // replace a%b with a-((a/b)*b) 3362 Node* mult = new MulLNode(d, d->in(2)); 3363 Node* sub = new SubLNode(d->in(1), mult); 3364 n->subsume_by(sub, this); 3365 } 3366 } 3367 } 3368 break; 3369 3370 case Op_LoadVector: 3371 case Op_StoreVector: 3372 break; 3373 3374 case Op_AddReductionVI: 3375 case Op_AddReductionVL: 3376 case Op_AddReductionVF: 3377 case Op_AddReductionVD: 3378 case Op_MulReductionVI: 3379 case Op_MulReductionVL: 3380 case Op_MulReductionVF: 3381 case Op_MulReductionVD: 3382 case Op_MinReductionV: 3383 case Op_MaxReductionV: 3384 break; 3385 3386 case Op_PackB: 3387 case Op_PackS: 3388 case Op_PackI: 3389 case Op_PackF: 3390 case Op_PackL: 3391 case Op_PackD: 3392 if (n->req()-1 > 2) { 3393 // Replace many operand PackNodes with a binary tree for matching 3394 PackNode* p = (PackNode*) n; 3395 Node* btp = p->binary_tree_pack(1, n->req()); 3396 n->subsume_by(btp, this); 3397 } 3398 break; 3399 case Op_Loop: 3400 case Op_CountedLoop: 3401 case Op_OuterStripMinedLoop: 3402 if (n->as_Loop()->is_inner_loop()) { 3403 frc.inc_inner_loop_count(); 3404 } 3405 n->as_Loop()->verify_strip_mined(0); 3406 break; 3407 case Op_LShiftI: 3408 case Op_RShiftI: 3409 case Op_URShiftI: 3410 case Op_LShiftL: 3411 case Op_RShiftL: 3412 case Op_URShiftL: 3413 if (Matcher::need_masked_shift_count) { 3414 // The cpu's shift instructions don't restrict the count to the 3415 // lower 5/6 bits. We need to do the masking ourselves. 3416 Node* in2 = n->in(2); 3417 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3418 const TypeInt* t = in2->find_int_type(); 3419 if (t != NULL && t->is_con()) { 3420 juint shift = t->get_con(); 3421 if (shift > mask) { // Unsigned cmp 3422 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3423 } 3424 } else { 3425 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3426 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3427 n->set_req(2, shift); 3428 } 3429 } 3430 if (in2->outcnt() == 0) { // Remove dead node 3431 in2->disconnect_inputs(NULL, this); 3432 } 3433 } 3434 break; 3435 case Op_MemBarStoreStore: 3436 case Op_MemBarRelease: 3437 // Break the link with AllocateNode: it is no longer useful and 3438 // confuses register allocation. 3439 if (n->req() > MemBarNode::Precedent) { 3440 n->set_req(MemBarNode::Precedent, top()); 3441 } 3442 break; 3443 case Op_MemBarAcquire: { 3444 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3445 // At parse time, the trailing MemBarAcquire for a volatile load 3446 // is created with an edge to the load. After optimizations, 3447 // that input may be a chain of Phis. If those phis have no 3448 // other use, then the MemBarAcquire keeps them alive and 3449 // register allocation can be confused. 3450 ResourceMark rm; 3451 Unique_Node_List wq; 3452 wq.push(n->in(MemBarNode::Precedent)); 3453 n->set_req(MemBarNode::Precedent, top()); 3454 while (wq.size() > 0) { 3455 Node* m = wq.pop(); 3456 if (m->outcnt() == 0) { 3457 for (uint j = 0; j < m->req(); j++) { 3458 Node* in = m->in(j); 3459 if (in != NULL) { 3460 wq.push(in); 3461 } 3462 } 3463 m->disconnect_inputs(NULL, this); 3464 } 3465 } 3466 } 3467 break; 3468 } 3469 #if INCLUDE_SHENANDOAHGC 3470 case Op_ShenandoahCompareAndSwapP: 3471 case Op_ShenandoahCompareAndSwapN: 3472 case Op_ShenandoahWeakCompareAndSwapN: 3473 case Op_ShenandoahWeakCompareAndSwapP: 3474 case Op_ShenandoahCompareAndExchangeP: 3475 case Op_ShenandoahCompareAndExchangeN: 3476 #ifdef ASSERT 3477 if( VerifyOptoOopOffsets ) { 3478 MemNode* mem = n->as_Mem(); 3479 // Check to see if address types have grounded out somehow. 3480 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3481 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 3482 bool oop_offset_is_sane = k->contains_field_offset(tp->offset()); 3483 assert( !tp || oop_offset_is_sane, "" ); 3484 } 3485 #endif 3486 break; 3487 case Op_ShenandoahLoadReferenceBarrier: 3488 assert(false, "should have been expanded already"); 3489 break; 3490 #endif 3491 case Op_RangeCheck: { 3492 RangeCheckNode* rc = n->as_RangeCheck(); 3493 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3494 n->subsume_by(iff, this); 3495 frc._tests.push(iff); 3496 break; 3497 } 3498 case Op_ConvI2L: { 3499 if (!Matcher::convi2l_type_required) { 3500 // Code generation on some platforms doesn't need accurate 3501 // ConvI2L types. Widening the type can help remove redundant 3502 // address computations. 3503 n->as_Type()->set_type(TypeLong::INT); 3504 ResourceMark rm; 3505 Unique_Node_List wq; 3506 wq.push(n); 3507 for (uint next = 0; next < wq.size(); next++) { 3508 Node *m = wq.at(next); 3509 3510 for(;;) { 3511 // Loop over all nodes with identical inputs edges as m 3512 Node* k = m->find_similar(m->Opcode()); 3513 if (k == NULL) { 3514 break; 3515 } 3516 // Push their uses so we get a chance to remove node made 3517 // redundant 3518 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3519 Node* u = k->fast_out(i); 3520 if (u->Opcode() == Op_LShiftL || 3521 u->Opcode() == Op_AddL || 3522 u->Opcode() == Op_SubL || 3523 u->Opcode() == Op_AddP) { 3524 wq.push(u); 3525 } 3526 } 3527 // Replace all nodes with identical edges as m with m 3528 k->subsume_by(m, this); 3529 } 3530 } 3531 } 3532 break; 3533 } 3534 case Op_CmpUL: { 3535 if (!Matcher::has_match_rule(Op_CmpUL)) { 3536 // No support for unsigned long comparisons 3537 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3538 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3539 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3540 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3541 Node* andl = new AndLNode(orl, remove_sign_mask); 3542 Node* cmp = new CmpLNode(andl, n->in(2)); 3543 n->subsume_by(cmp, this); 3544 } 3545 break; 3546 } 3547 default: 3548 assert( !n->is_Call(), "" ); 3549 assert( !n->is_Mem(), "" ); 3550 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3551 break; 3552 } 3553 3554 // Collect CFG split points 3555 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3556 frc._tests.push(n); 3557 } 3558 } 3559 3560 //------------------------------final_graph_reshaping_walk--------------------- 3561 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3562 // requires that the walk visits a node's inputs before visiting the node. 3563 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3564 ResourceArea *area = Thread::current()->resource_area(); 3565 Unique_Node_List sfpt(area); 3566 3567 frc._visited.set(root->_idx); // first, mark node as visited 3568 uint cnt = root->req(); 3569 Node *n = root; 3570 uint i = 0; 3571 while (true) { 3572 if (i < cnt) { 3573 // Place all non-visited non-null inputs onto stack 3574 Node* m = n->in(i); 3575 ++i; 3576 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3577 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3578 // compute worst case interpreter size in case of a deoptimization 3579 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3580 3581 sfpt.push(m); 3582 } 3583 cnt = m->req(); 3584 nstack.push(n, i); // put on stack parent and next input's index 3585 n = m; 3586 i = 0; 3587 } 3588 } else { 3589 // Now do post-visit work 3590 final_graph_reshaping_impl( n, frc ); 3591 if (nstack.is_empty()) 3592 break; // finished 3593 n = nstack.node(); // Get node from stack 3594 cnt = n->req(); 3595 i = nstack.index(); 3596 nstack.pop(); // Shift to the next node on stack 3597 } 3598 } 3599 3600 // Skip next transformation if compressed oops are not used. 3601 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3602 (!UseCompressedOops && !UseCompressedClassPointers)) 3603 return; 3604 3605 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3606 // It could be done for an uncommon traps or any safepoints/calls 3607 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3608 while (sfpt.size() > 0) { 3609 n = sfpt.pop(); 3610 JVMState *jvms = n->as_SafePoint()->jvms(); 3611 assert(jvms != NULL, "sanity"); 3612 int start = jvms->debug_start(); 3613 int end = n->req(); 3614 bool is_uncommon = (n->is_CallStaticJava() && 3615 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3616 for (int j = start; j < end; j++) { 3617 Node* in = n->in(j); 3618 if (in->is_DecodeNarrowPtr()) { 3619 bool safe_to_skip = true; 3620 if (!is_uncommon ) { 3621 // Is it safe to skip? 3622 for (uint i = 0; i < in->outcnt(); i++) { 3623 Node* u = in->raw_out(i); 3624 if (!u->is_SafePoint() || 3625 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3626 safe_to_skip = false; 3627 } 3628 } 3629 } 3630 if (safe_to_skip) { 3631 n->set_req(j, in->in(1)); 3632 } 3633 if (in->outcnt() == 0) { 3634 in->disconnect_inputs(NULL, this); 3635 } 3636 } 3637 } 3638 } 3639 } 3640 3641 //------------------------------final_graph_reshaping-------------------------- 3642 // Final Graph Reshaping. 3643 // 3644 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3645 // and not commoned up and forced early. Must come after regular 3646 // optimizations to avoid GVN undoing the cloning. Clone constant 3647 // inputs to Loop Phis; these will be split by the allocator anyways. 3648 // Remove Opaque nodes. 3649 // (2) Move last-uses by commutative operations to the left input to encourage 3650 // Intel update-in-place two-address operations and better register usage 3651 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3652 // calls canonicalizing them back. 3653 // (3) Count the number of double-precision FP ops, single-precision FP ops 3654 // and call sites. On Intel, we can get correct rounding either by 3655 // forcing singles to memory (requires extra stores and loads after each 3656 // FP bytecode) or we can set a rounding mode bit (requires setting and 3657 // clearing the mode bit around call sites). The mode bit is only used 3658 // if the relative frequency of single FP ops to calls is low enough. 3659 // This is a key transform for SPEC mpeg_audio. 3660 // (4) Detect infinite loops; blobs of code reachable from above but not 3661 // below. Several of the Code_Gen algorithms fail on such code shapes, 3662 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3663 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3664 // Detection is by looking for IfNodes where only 1 projection is 3665 // reachable from below or CatchNodes missing some targets. 3666 // (5) Assert for insane oop offsets in debug mode. 3667 3668 bool Compile::final_graph_reshaping() { 3669 // an infinite loop may have been eliminated by the optimizer, 3670 // in which case the graph will be empty. 3671 if (root()->req() == 1) { 3672 record_method_not_compilable("trivial infinite loop"); 3673 return true; 3674 } 3675 3676 // Expensive nodes have their control input set to prevent the GVN 3677 // from freely commoning them. There's no GVN beyond this point so 3678 // no need to keep the control input. We want the expensive nodes to 3679 // be freely moved to the least frequent code path by gcm. 3680 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3681 for (int i = 0; i < expensive_count(); i++) { 3682 _expensive_nodes->at(i)->set_req(0, NULL); 3683 } 3684 3685 Final_Reshape_Counts frc; 3686 3687 // Visit everybody reachable! 3688 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3689 Node_Stack nstack(live_nodes() >> 1); 3690 final_graph_reshaping_walk(nstack, root(), frc); 3691 3692 // Check for unreachable (from below) code (i.e., infinite loops). 3693 for( uint i = 0; i < frc._tests.size(); i++ ) { 3694 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3695 // Get number of CFG targets. 3696 // Note that PCTables include exception targets after calls. 3697 uint required_outcnt = n->required_outcnt(); 3698 if (n->outcnt() != required_outcnt) { 3699 // Check for a few special cases. Rethrow Nodes never take the 3700 // 'fall-thru' path, so expected kids is 1 less. 3701 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3702 if (n->in(0)->in(0)->is_Call()) { 3703 CallNode *call = n->in(0)->in(0)->as_Call(); 3704 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3705 required_outcnt--; // Rethrow always has 1 less kid 3706 } else if (call->req() > TypeFunc::Parms && 3707 call->is_CallDynamicJava()) { 3708 // Check for null receiver. In such case, the optimizer has 3709 // detected that the virtual call will always result in a null 3710 // pointer exception. The fall-through projection of this CatchNode 3711 // will not be populated. 3712 Node *arg0 = call->in(TypeFunc::Parms); 3713 if (arg0->is_Type() && 3714 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3715 required_outcnt--; 3716 } 3717 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3718 call->req() > TypeFunc::Parms+1 && 3719 call->is_CallStaticJava()) { 3720 // Check for negative array length. In such case, the optimizer has 3721 // detected that the allocation attempt will always result in an 3722 // exception. There is no fall-through projection of this CatchNode . 3723 Node *arg1 = call->in(TypeFunc::Parms+1); 3724 if (arg1->is_Type() && 3725 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3726 required_outcnt--; 3727 } 3728 } 3729 } 3730 } 3731 // Recheck with a better notion of 'required_outcnt' 3732 if (n->outcnt() != required_outcnt) { 3733 record_method_not_compilable("malformed control flow"); 3734 return true; // Not all targets reachable! 3735 } 3736 } 3737 // Check that I actually visited all kids. Unreached kids 3738 // must be infinite loops. 3739 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3740 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3741 record_method_not_compilable("infinite loop"); 3742 return true; // Found unvisited kid; must be unreach 3743 } 3744 3745 // Here so verification code in final_graph_reshaping_walk() 3746 // always see an OuterStripMinedLoopEnd 3747 if (n->is_OuterStripMinedLoopEnd()) { 3748 IfNode* init_iff = n->as_If(); 3749 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3750 n->subsume_by(iff, this); 3751 } 3752 } 3753 3754 // If original bytecodes contained a mixture of floats and doubles 3755 // check if the optimizer has made it homogenous, item (3). 3756 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3757 frc.get_float_count() > 32 && 3758 frc.get_double_count() == 0 && 3759 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3760 set_24_bit_selection_and_mode( false, true ); 3761 } 3762 3763 set_java_calls(frc.get_java_call_count()); 3764 set_inner_loops(frc.get_inner_loop_count()); 3765 3766 // No infinite loops, no reason to bail out. 3767 return false; 3768 } 3769 3770 //-----------------------------too_many_traps---------------------------------- 3771 // Report if there are too many traps at the current method and bci. 3772 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3773 bool Compile::too_many_traps(ciMethod* method, 3774 int bci, 3775 Deoptimization::DeoptReason reason) { 3776 ciMethodData* md = method->method_data(); 3777 if (md->is_empty()) { 3778 // Assume the trap has not occurred, or that it occurred only 3779 // because of a transient condition during start-up in the interpreter. 3780 return false; 3781 } 3782 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3783 if (md->has_trap_at(bci, m, reason) != 0) { 3784 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3785 // Also, if there are multiple reasons, or if there is no per-BCI record, 3786 // assume the worst. 3787 if (log()) 3788 log()->elem("observe trap='%s' count='%d'", 3789 Deoptimization::trap_reason_name(reason), 3790 md->trap_count(reason)); 3791 return true; 3792 } else { 3793 // Ignore method/bci and see if there have been too many globally. 3794 return too_many_traps(reason, md); 3795 } 3796 } 3797 3798 // Less-accurate variant which does not require a method and bci. 3799 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3800 ciMethodData* logmd) { 3801 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3802 // Too many traps globally. 3803 // Note that we use cumulative trap_count, not just md->trap_count. 3804 if (log()) { 3805 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3806 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3807 Deoptimization::trap_reason_name(reason), 3808 mcount, trap_count(reason)); 3809 } 3810 return true; 3811 } else { 3812 // The coast is clear. 3813 return false; 3814 } 3815 } 3816 3817 //--------------------------too_many_recompiles-------------------------------- 3818 // Report if there are too many recompiles at the current method and bci. 3819 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3820 // Is not eager to return true, since this will cause the compiler to use 3821 // Action_none for a trap point, to avoid too many recompilations. 3822 bool Compile::too_many_recompiles(ciMethod* method, 3823 int bci, 3824 Deoptimization::DeoptReason reason) { 3825 ciMethodData* md = method->method_data(); 3826 if (md->is_empty()) { 3827 // Assume the trap has not occurred, or that it occurred only 3828 // because of a transient condition during start-up in the interpreter. 3829 return false; 3830 } 3831 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3832 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3833 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3834 Deoptimization::DeoptReason per_bc_reason 3835 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3836 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3837 if ((per_bc_reason == Deoptimization::Reason_none 3838 || md->has_trap_at(bci, m, reason) != 0) 3839 // The trap frequency measure we care about is the recompile count: 3840 && md->trap_recompiled_at(bci, m) 3841 && md->overflow_recompile_count() >= bc_cutoff) { 3842 // Do not emit a trap here if it has already caused recompilations. 3843 // Also, if there are multiple reasons, or if there is no per-BCI record, 3844 // assume the worst. 3845 if (log()) 3846 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3847 Deoptimization::trap_reason_name(reason), 3848 md->trap_count(reason), 3849 md->overflow_recompile_count()); 3850 return true; 3851 } else if (trap_count(reason) != 0 3852 && decompile_count() >= m_cutoff) { 3853 // Too many recompiles globally, and we have seen this sort of trap. 3854 // Use cumulative decompile_count, not just md->decompile_count. 3855 if (log()) 3856 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3857 Deoptimization::trap_reason_name(reason), 3858 md->trap_count(reason), trap_count(reason), 3859 md->decompile_count(), decompile_count()); 3860 return true; 3861 } else { 3862 // The coast is clear. 3863 return false; 3864 } 3865 } 3866 3867 // Compute when not to trap. Used by matching trap based nodes and 3868 // NullCheck optimization. 3869 void Compile::set_allowed_deopt_reasons() { 3870 _allowed_reasons = 0; 3871 if (is_method_compilation()) { 3872 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3873 assert(rs < BitsPerInt, "recode bit map"); 3874 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3875 _allowed_reasons |= nth_bit(rs); 3876 } 3877 } 3878 } 3879 } 3880 3881 bool Compile::is_compiling_clinit_for(ciKlass* k) { 3882 ciMethod* root = method(); // the root method of compilation 3883 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit 3884 } 3885 3886 #ifndef PRODUCT 3887 //------------------------------verify_graph_edges--------------------------- 3888 // Walk the Graph and verify that there is a one-to-one correspondence 3889 // between Use-Def edges and Def-Use edges in the graph. 3890 void Compile::verify_graph_edges(bool no_dead_code) { 3891 if (VerifyGraphEdges) { 3892 ResourceArea *area = Thread::current()->resource_area(); 3893 Unique_Node_List visited(area); 3894 // Call recursive graph walk to check edges 3895 _root->verify_edges(visited); 3896 if (no_dead_code) { 3897 // Now make sure that no visited node is used by an unvisited node. 3898 bool dead_nodes = false; 3899 Unique_Node_List checked(area); 3900 while (visited.size() > 0) { 3901 Node* n = visited.pop(); 3902 checked.push(n); 3903 for (uint i = 0; i < n->outcnt(); i++) { 3904 Node* use = n->raw_out(i); 3905 if (checked.member(use)) continue; // already checked 3906 if (visited.member(use)) continue; // already in the graph 3907 if (use->is_Con()) continue; // a dead ConNode is OK 3908 // At this point, we have found a dead node which is DU-reachable. 3909 if (!dead_nodes) { 3910 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3911 dead_nodes = true; 3912 } 3913 use->dump(2); 3914 tty->print_cr("---"); 3915 checked.push(use); // No repeats; pretend it is now checked. 3916 } 3917 } 3918 assert(!dead_nodes, "using nodes must be reachable from root"); 3919 } 3920 } 3921 } 3922 3923 // Verify GC barriers consistency 3924 // Currently supported: 3925 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3926 void Compile::verify_barriers() { 3927 #if INCLUDE_G1GC || INCLUDE_SHENANDOAHGC 3928 if (UseG1GC SHENANDOAHGC_ONLY(|| UseShenandoahGC)) { 3929 // Verify G1 pre-barriers 3930 3931 #if INCLUDE_G1GC && INCLUDE_SHENANDOAHGC 3932 const int marking_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_active_offset() 3933 : ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 3934 #elif INCLUDE_G1GC 3935 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3936 #else 3937 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()); 3938 #endif 3939 3940 ResourceArea *area = Thread::current()->resource_area(); 3941 Unique_Node_List visited(area); 3942 Node_List worklist(area); 3943 // We're going to walk control flow backwards starting from the Root 3944 worklist.push(_root); 3945 while (worklist.size() > 0) { 3946 Node* x = worklist.pop(); 3947 if (x == NULL || x == top()) continue; 3948 if (visited.member(x)) { 3949 continue; 3950 } else { 3951 visited.push(x); 3952 } 3953 3954 if (x->is_Region()) { 3955 for (uint i = 1; i < x->req(); i++) { 3956 worklist.push(x->in(i)); 3957 } 3958 } else { 3959 worklist.push(x->in(0)); 3960 // We are looking for the pattern: 3961 // /->ThreadLocal 3962 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3963 // \->ConI(0) 3964 // We want to verify that the If and the LoadB have the same control 3965 // See GraphKit::g1_write_barrier_pre() 3966 if (x->is_If()) { 3967 IfNode *iff = x->as_If(); 3968 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3969 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3970 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3971 && cmp->in(1)->is_Load()) { 3972 LoadNode* load = cmp->in(1)->as_Load(); 3973 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3974 && load->in(2)->in(3)->is_Con() 3975 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3976 3977 Node* if_ctrl = iff->in(0); 3978 Node* load_ctrl = load->in(0); 3979 3980 if (if_ctrl != load_ctrl) { 3981 // Skip possible CProj->NeverBranch in infinite loops 3982 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3983 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3984 if_ctrl = if_ctrl->in(0)->in(0); 3985 } 3986 } 3987 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3988 } 3989 } 3990 } 3991 } 3992 } 3993 } 3994 } 3995 #endif 3996 } 3997 3998 #endif 3999 4000 // The Compile object keeps track of failure reasons separately from the ciEnv. 4001 // This is required because there is not quite a 1-1 relation between the 4002 // ciEnv and its compilation task and the Compile object. Note that one 4003 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 4004 // to backtrack and retry without subsuming loads. Other than this backtracking 4005 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 4006 // by the logic in C2Compiler. 4007 void Compile::record_failure(const char* reason) { 4008 if (log() != NULL) { 4009 log()->elem("failure reason='%s' phase='compile'", reason); 4010 } 4011 if (_failure_reason == NULL) { 4012 // Record the first failure reason. 4013 _failure_reason = reason; 4014 } 4015 4016 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 4017 C->print_method(PHASE_FAILURE); 4018 } 4019 _root = NULL; // flush the graph, too 4020 } 4021 4022 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 4023 : TraceTime(name, accumulator, CITime, CITimeVerbose), 4024 _phase_name(name), _dolog(CITimeVerbose) 4025 { 4026 if (_dolog) { 4027 C = Compile::current(); 4028 _log = C->log(); 4029 } else { 4030 C = NULL; 4031 _log = NULL; 4032 } 4033 if (_log != NULL) { 4034 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4035 _log->stamp(); 4036 _log->end_head(); 4037 } 4038 } 4039 4040 Compile::TracePhase::~TracePhase() { 4041 4042 C = Compile::current(); 4043 if (_dolog) { 4044 _log = C->log(); 4045 } else { 4046 _log = NULL; 4047 } 4048 4049 #ifdef ASSERT 4050 if (PrintIdealNodeCount) { 4051 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4052 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4053 } 4054 4055 if (VerifyIdealNodeCount) { 4056 Compile::current()->print_missing_nodes(); 4057 } 4058 #endif 4059 4060 if (_log != NULL) { 4061 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4062 } 4063 } 4064 4065 //============================================================================= 4066 // Two Constant's are equal when the type and the value are equal. 4067 bool Compile::Constant::operator==(const Constant& other) { 4068 if (type() != other.type() ) return false; 4069 if (can_be_reused() != other.can_be_reused()) return false; 4070 // For floating point values we compare the bit pattern. 4071 switch (type()) { 4072 case T_INT: 4073 case T_FLOAT: return (_v._value.i == other._v._value.i); 4074 case T_LONG: 4075 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4076 case T_OBJECT: 4077 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4078 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4079 case T_METADATA: return (_v._metadata == other._v._metadata); 4080 default: ShouldNotReachHere(); return false; 4081 } 4082 } 4083 4084 static int type_to_size_in_bytes(BasicType t) { 4085 switch (t) { 4086 case T_INT: return sizeof(jint ); 4087 case T_LONG: return sizeof(jlong ); 4088 case T_FLOAT: return sizeof(jfloat ); 4089 case T_DOUBLE: return sizeof(jdouble); 4090 case T_METADATA: return sizeof(Metadata*); 4091 // We use T_VOID as marker for jump-table entries (labels) which 4092 // need an internal word relocation. 4093 case T_VOID: 4094 case T_ADDRESS: 4095 case T_OBJECT: return sizeof(jobject); 4096 default: 4097 ShouldNotReachHere(); 4098 return -1; 4099 } 4100 } 4101 4102 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4103 // sort descending 4104 if (a->freq() > b->freq()) return -1; 4105 if (a->freq() < b->freq()) return 1; 4106 return 0; 4107 } 4108 4109 void Compile::ConstantTable::calculate_offsets_and_size() { 4110 // First, sort the array by frequencies. 4111 _constants.sort(qsort_comparator); 4112 4113 #ifdef ASSERT 4114 // Make sure all jump-table entries were sorted to the end of the 4115 // array (they have a negative frequency). 4116 bool found_void = false; 4117 for (int i = 0; i < _constants.length(); i++) { 4118 Constant con = _constants.at(i); 4119 if (con.type() == T_VOID) 4120 found_void = true; // jump-tables 4121 else 4122 assert(!found_void, "wrong sorting"); 4123 } 4124 #endif 4125 4126 int offset = 0; 4127 for (int i = 0; i < _constants.length(); i++) { 4128 Constant* con = _constants.adr_at(i); 4129 4130 // Align offset for type. 4131 int typesize = type_to_size_in_bytes(con->type()); 4132 offset = align_up(offset, typesize); 4133 con->set_offset(offset); // set constant's offset 4134 4135 if (con->type() == T_VOID) { 4136 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4137 offset = offset + typesize * n->outcnt(); // expand jump-table 4138 } else { 4139 offset = offset + typesize; 4140 } 4141 } 4142 4143 // Align size up to the next section start (which is insts; see 4144 // CodeBuffer::align_at_start). 4145 assert(_size == -1, "already set?"); 4146 _size = align_up(offset, (int)CodeEntryAlignment); 4147 } 4148 4149 bool Compile::ConstantTable::emit(CodeBuffer& cb) const { 4150 MacroAssembler _masm(&cb); 4151 for (int i = 0; i < _constants.length(); i++) { 4152 Constant con = _constants.at(i); 4153 address constant_addr = NULL; 4154 switch (con.type()) { 4155 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4156 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4157 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4158 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4159 case T_OBJECT: { 4160 jobject obj = con.get_jobject(); 4161 int oop_index = _masm.oop_recorder()->find_index(obj); 4162 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4163 break; 4164 } 4165 case T_ADDRESS: { 4166 address addr = (address) con.get_jobject(); 4167 constant_addr = _masm.address_constant(addr); 4168 break; 4169 } 4170 // We use T_VOID as marker for jump-table entries (labels) which 4171 // need an internal word relocation. 4172 case T_VOID: { 4173 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4174 // Fill the jump-table with a dummy word. The real value is 4175 // filled in later in fill_jump_table. 4176 address dummy = (address) n; 4177 constant_addr = _masm.address_constant(dummy); 4178 if (constant_addr == NULL) { 4179 return false; 4180 } 4181 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4182 "must be: %d == %d", (int)(constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4183 4184 // Expand jump-table 4185 address last_addr = NULL; 4186 for (uint j = 1; j < n->outcnt(); j++) { 4187 last_addr = _masm.address_constant(dummy + j); 4188 if (last_addr == NULL) { 4189 return false; 4190 } 4191 } 4192 #ifdef ASSERT 4193 address start = _masm.code()->consts()->start(); 4194 address new_constant_addr = last_addr - ((n->outcnt() - 1) * sizeof(address)); 4195 // Expanding the jump-table could result in an expansion of the const code section. 4196 // In that case, we need to check if the new constant address matches the offset. 4197 assert((constant_addr - start == con.offset()) || (new_constant_addr - start == con.offset()), 4198 "must be: %d == %d or %d == %d (after an expansion)", (int)(constant_addr - start), (int)(con.offset()), 4199 (int)(new_constant_addr - start), (int)(con.offset())); 4200 #endif 4201 continue; // Loop 4202 } 4203 case T_METADATA: { 4204 Metadata* obj = con.get_metadata(); 4205 int metadata_index = _masm.oop_recorder()->find_index(obj); 4206 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4207 break; 4208 } 4209 default: ShouldNotReachHere(); 4210 } 4211 4212 if (constant_addr == NULL) { 4213 return false; 4214 } 4215 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4216 "must be: %d == %d", (int)(constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4217 } 4218 return true; 4219 } 4220 4221 int Compile::ConstantTable::find_offset(Constant& con) const { 4222 int idx = _constants.find(con); 4223 guarantee(idx != -1, "constant must be in constant table"); 4224 int offset = _constants.at(idx).offset(); 4225 guarantee(offset != -1, "constant table not emitted yet?"); 4226 return offset; 4227 } 4228 4229 void Compile::ConstantTable::add(Constant& con) { 4230 if (con.can_be_reused()) { 4231 int idx = _constants.find(con); 4232 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4233 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4234 return; 4235 } 4236 } 4237 (void) _constants.append(con); 4238 } 4239 4240 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4241 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4242 Constant con(type, value, b->_freq); 4243 add(con); 4244 return con; 4245 } 4246 4247 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4248 Constant con(metadata); 4249 add(con); 4250 return con; 4251 } 4252 4253 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4254 jvalue value; 4255 BasicType type = oper->type()->basic_type(); 4256 switch (type) { 4257 case T_LONG: value.j = oper->constantL(); break; 4258 case T_FLOAT: value.f = oper->constantF(); break; 4259 case T_DOUBLE: value.d = oper->constantD(); break; 4260 case T_OBJECT: 4261 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4262 case T_METADATA: return add((Metadata*)oper->constant()); break; 4263 default: guarantee(false, "unhandled type: %s", type2name(type)); 4264 } 4265 return add(n, type, value); 4266 } 4267 4268 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4269 jvalue value; 4270 // We can use the node pointer here to identify the right jump-table 4271 // as this method is called from Compile::Fill_buffer right before 4272 // the MachNodes are emitted and the jump-table is filled (means the 4273 // MachNode pointers do not change anymore). 4274 value.l = (jobject) n; 4275 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4276 add(con); 4277 return con; 4278 } 4279 4280 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4281 // If called from Compile::scratch_emit_size do nothing. 4282 if (Compile::current()->in_scratch_emit_size()) return; 4283 4284 assert(labels.is_nonempty(), "must be"); 4285 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4286 4287 // Since MachConstantNode::constant_offset() also contains 4288 // table_base_offset() we need to subtract the table_base_offset() 4289 // to get the plain offset into the constant table. 4290 int offset = n->constant_offset() - table_base_offset(); 4291 4292 MacroAssembler _masm(&cb); 4293 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4294 4295 for (uint i = 0; i < n->outcnt(); i++) { 4296 address* constant_addr = &jump_table_base[i]; 4297 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)); 4298 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4299 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4300 } 4301 } 4302 4303 //----------------------------static_subtype_check----------------------------- 4304 // Shortcut important common cases when superklass is exact: 4305 // (0) superklass is java.lang.Object (can occur in reflective code) 4306 // (1) subklass is already limited to a subtype of superklass => always ok 4307 // (2) subklass does not overlap with superklass => always fail 4308 // (3) superklass has NO subtypes and we can check with a simple compare. 4309 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4310 if (StressReflectiveCode) { 4311 return SSC_full_test; // Let caller generate the general case. 4312 } 4313 4314 if (superk == env()->Object_klass()) { 4315 return SSC_always_true; // (0) this test cannot fail 4316 } 4317 4318 ciType* superelem = superk; 4319 if (superelem->is_array_klass()) 4320 superelem = superelem->as_array_klass()->base_element_type(); 4321 4322 if (!subk->is_interface()) { // cannot trust static interface types yet 4323 if (subk->is_subtype_of(superk)) { 4324 return SSC_always_true; // (1) false path dead; no dynamic test needed 4325 } 4326 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4327 !superk->is_subtype_of(subk)) { 4328 return SSC_always_false; 4329 } 4330 } 4331 4332 // If casting to an instance klass, it must have no subtypes 4333 if (superk->is_interface()) { 4334 // Cannot trust interfaces yet. 4335 // %%% S.B. superk->nof_implementors() == 1 4336 } else if (superelem->is_instance_klass()) { 4337 ciInstanceKlass* ik = superelem->as_instance_klass(); 4338 if (!ik->has_subklass() && !ik->is_interface()) { 4339 if (!ik->is_final()) { 4340 // Add a dependency if there is a chance of a later subclass. 4341 dependencies()->assert_leaf_type(ik); 4342 } 4343 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4344 } 4345 } else { 4346 // A primitive array type has no subtypes. 4347 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4348 } 4349 4350 return SSC_full_test; 4351 } 4352 4353 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4354 #ifdef _LP64 4355 // The scaled index operand to AddP must be a clean 64-bit value. 4356 // Java allows a 32-bit int to be incremented to a negative 4357 // value, which appears in a 64-bit register as a large 4358 // positive number. Using that large positive number as an 4359 // operand in pointer arithmetic has bad consequences. 4360 // On the other hand, 32-bit overflow is rare, and the possibility 4361 // can often be excluded, if we annotate the ConvI2L node with 4362 // a type assertion that its value is known to be a small positive 4363 // number. (The prior range check has ensured this.) 4364 // This assertion is used by ConvI2LNode::Ideal. 4365 int index_max = max_jint - 1; // array size is max_jint, index is one less 4366 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4367 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4368 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4369 #endif 4370 return idx; 4371 } 4372 4373 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4374 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) { 4375 if (ctrl != NULL) { 4376 // Express control dependency by a CastII node with a narrow type. 4377 value = new CastIINode(value, itype, carry_dependency, true /* range check dependency */); 4378 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4379 // node from floating above the range check during loop optimizations. Otherwise, the 4380 // ConvI2L node may be eliminated independently of the range check, causing the data path 4381 // to become TOP while the control path is still there (although it's unreachable). 4382 value->set_req(0, ctrl); 4383 // Save CastII node to remove it after loop optimizations. 4384 phase->C->add_range_check_cast(value); 4385 value = phase->transform(value); 4386 } 4387 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4388 return phase->transform(new ConvI2LNode(value, ltype)); 4389 } 4390 4391 void Compile::print_inlining_stream_free() { 4392 if (_print_inlining_stream != NULL) { 4393 _print_inlining_stream->~stringStream(); 4394 _print_inlining_stream = NULL; 4395 } 4396 } 4397 4398 // The message about the current inlining is accumulated in 4399 // _print_inlining_stream and transfered into the _print_inlining_list 4400 // once we know whether inlining succeeds or not. For regular 4401 // inlining, messages are appended to the buffer pointed by 4402 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4403 // a new buffer is added after _print_inlining_idx in the list. This 4404 // way we can update the inlining message for late inlining call site 4405 // when the inlining is attempted again. 4406 void Compile::print_inlining_init() { 4407 if (print_inlining() || print_intrinsics()) { 4408 // print_inlining_init is actually called several times. 4409 print_inlining_stream_free(); 4410 _print_inlining_stream = new stringStream(); 4411 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer()); 4412 } 4413 } 4414 4415 void Compile::print_inlining_reinit() { 4416 if (print_inlining() || print_intrinsics()) { 4417 print_inlining_stream_free(); 4418 // Re allocate buffer when we change ResourceMark 4419 _print_inlining_stream = new stringStream(); 4420 } 4421 } 4422 4423 void Compile::print_inlining_reset() { 4424 _print_inlining_stream->reset(); 4425 } 4426 4427 void Compile::print_inlining_commit() { 4428 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4429 // Transfer the message from _print_inlining_stream to the current 4430 // _print_inlining_list buffer and clear _print_inlining_stream. 4431 _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size()); 4432 print_inlining_reset(); 4433 } 4434 4435 void Compile::print_inlining_push() { 4436 // Add new buffer to the _print_inlining_list at current position 4437 _print_inlining_idx++; 4438 _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer()); 4439 } 4440 4441 Compile::PrintInliningBuffer* Compile::print_inlining_current() { 4442 return _print_inlining_list->at(_print_inlining_idx); 4443 } 4444 4445 void Compile::print_inlining_update(CallGenerator* cg) { 4446 if (print_inlining() || print_intrinsics()) { 4447 if (!cg->is_late_inline()) { 4448 if (print_inlining_current()->cg() != NULL) { 4449 print_inlining_push(); 4450 } 4451 print_inlining_commit(); 4452 } else { 4453 if (print_inlining_current()->cg() != cg && 4454 (print_inlining_current()->cg() != NULL || 4455 print_inlining_current()->ss()->size() != 0)) { 4456 print_inlining_push(); 4457 } 4458 print_inlining_commit(); 4459 print_inlining_current()->set_cg(cg); 4460 } 4461 } 4462 } 4463 4464 void Compile::print_inlining_move_to(CallGenerator* cg) { 4465 // We resume inlining at a late inlining call site. Locate the 4466 // corresponding inlining buffer so that we can update it. 4467 if (print_inlining()) { 4468 for (int i = 0; i < _print_inlining_list->length(); i++) { 4469 if (_print_inlining_list->at(i)->cg() == cg) { 4470 _print_inlining_idx = i; 4471 return; 4472 } 4473 } 4474 ShouldNotReachHere(); 4475 } 4476 } 4477 4478 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4479 if (print_inlining()) { 4480 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4481 assert(print_inlining_current()->cg() == cg, "wrong entry"); 4482 // replace message with new message 4483 _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer()); 4484 print_inlining_commit(); 4485 print_inlining_current()->set_cg(cg); 4486 } 4487 } 4488 4489 void Compile::print_inlining_assert_ready() { 4490 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4491 } 4492 4493 void Compile::process_print_inlining() { 4494 bool do_print_inlining = print_inlining() || print_intrinsics(); 4495 if (do_print_inlining || log() != NULL) { 4496 // Print inlining message for candidates that we couldn't inline 4497 // for lack of space 4498 for (int i = 0; i < _late_inlines.length(); i++) { 4499 CallGenerator* cg = _late_inlines.at(i); 4500 if (!cg->is_mh_late_inline()) { 4501 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4502 if (do_print_inlining) { 4503 cg->print_inlining_late(msg); 4504 } 4505 log_late_inline_failure(cg, msg); 4506 } 4507 } 4508 } 4509 if (do_print_inlining) { 4510 ResourceMark rm; 4511 stringStream ss; 4512 assert(_print_inlining_list != NULL, "process_print_inlining should be called only once."); 4513 for (int i = 0; i < _print_inlining_list->length(); i++) { 4514 PrintInliningBuffer* pib = _print_inlining_list->at(i); 4515 ss.print("%s", pib->ss()->as_string()); 4516 delete pib; 4517 DEBUG_ONLY(_print_inlining_list->at_put(i, NULL)); 4518 } 4519 // Reset _print_inlining_list, it only contains destructed objects. 4520 // It is on the arena, so it will be freed when the arena is reset. 4521 _print_inlining_list = NULL; 4522 // _print_inlining_stream won't be used anymore, either. 4523 print_inlining_stream_free(); 4524 size_t end = ss.size(); 4525 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4526 strncpy(_print_inlining_output, ss.base(), end+1); 4527 _print_inlining_output[end] = 0; 4528 } 4529 } 4530 4531 void Compile::dump_print_inlining() { 4532 if (_print_inlining_output != NULL) { 4533 tty->print_raw(_print_inlining_output); 4534 } 4535 } 4536 4537 void Compile::log_late_inline(CallGenerator* cg) { 4538 if (log() != NULL) { 4539 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4540 cg->unique_id()); 4541 JVMState* p = cg->call_node()->jvms(); 4542 while (p != NULL) { 4543 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4544 p = p->caller(); 4545 } 4546 log()->tail("late_inline"); 4547 } 4548 } 4549 4550 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4551 log_late_inline(cg); 4552 if (log() != NULL) { 4553 log()->inline_fail(msg); 4554 } 4555 } 4556 4557 void Compile::log_inline_id(CallGenerator* cg) { 4558 if (log() != NULL) { 4559 // The LogCompilation tool needs a unique way to identify late 4560 // inline call sites. This id must be unique for this call site in 4561 // this compilation. Try to have it unique across compilations as 4562 // well because it can be convenient when grepping through the log 4563 // file. 4564 // Distinguish OSR compilations from others in case CICountOSR is 4565 // on. 4566 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4567 cg->set_unique_id(id); 4568 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4569 } 4570 } 4571 4572 void Compile::log_inline_failure(const char* msg) { 4573 if (C->log() != NULL) { 4574 C->log()->inline_fail(msg); 4575 } 4576 } 4577 4578 4579 // Dump inlining replay data to the stream. 4580 // Don't change thread state and acquire any locks. 4581 void Compile::dump_inline_data(outputStream* out) { 4582 InlineTree* inl_tree = ilt(); 4583 if (inl_tree != NULL) { 4584 out->print(" inline %d", inl_tree->count()); 4585 inl_tree->dump_replay_data(out); 4586 } 4587 } 4588 4589 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4590 if (n1->Opcode() < n2->Opcode()) return -1; 4591 else if (n1->Opcode() > n2->Opcode()) return 1; 4592 4593 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4594 for (uint i = 1; i < n1->req(); i++) { 4595 if (n1->in(i) < n2->in(i)) return -1; 4596 else if (n1->in(i) > n2->in(i)) return 1; 4597 } 4598 4599 return 0; 4600 } 4601 4602 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4603 Node* n1 = *n1p; 4604 Node* n2 = *n2p; 4605 4606 return cmp_expensive_nodes(n1, n2); 4607 } 4608 4609 void Compile::sort_expensive_nodes() { 4610 if (!expensive_nodes_sorted()) { 4611 _expensive_nodes->sort(cmp_expensive_nodes); 4612 } 4613 } 4614 4615 bool Compile::expensive_nodes_sorted() const { 4616 for (int i = 1; i < _expensive_nodes->length(); i++) { 4617 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4618 return false; 4619 } 4620 } 4621 return true; 4622 } 4623 4624 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4625 if (_expensive_nodes->length() == 0) { 4626 return false; 4627 } 4628 4629 assert(OptimizeExpensiveOps, "optimization off?"); 4630 4631 // Take this opportunity to remove dead nodes from the list 4632 int j = 0; 4633 for (int i = 0; i < _expensive_nodes->length(); i++) { 4634 Node* n = _expensive_nodes->at(i); 4635 if (!n->is_unreachable(igvn)) { 4636 assert(n->is_expensive(), "should be expensive"); 4637 _expensive_nodes->at_put(j, n); 4638 j++; 4639 } 4640 } 4641 _expensive_nodes->trunc_to(j); 4642 4643 // Then sort the list so that similar nodes are next to each other 4644 // and check for at least two nodes of identical kind with same data 4645 // inputs. 4646 sort_expensive_nodes(); 4647 4648 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4649 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4650 return true; 4651 } 4652 } 4653 4654 return false; 4655 } 4656 4657 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4658 if (_expensive_nodes->length() == 0) { 4659 return; 4660 } 4661 4662 assert(OptimizeExpensiveOps, "optimization off?"); 4663 4664 // Sort to bring similar nodes next to each other and clear the 4665 // control input of nodes for which there's only a single copy. 4666 sort_expensive_nodes(); 4667 4668 int j = 0; 4669 int identical = 0; 4670 int i = 0; 4671 bool modified = false; 4672 for (; i < _expensive_nodes->length()-1; i++) { 4673 assert(j <= i, "can't write beyond current index"); 4674 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4675 identical++; 4676 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4677 continue; 4678 } 4679 if (identical > 0) { 4680 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4681 identical = 0; 4682 } else { 4683 Node* n = _expensive_nodes->at(i); 4684 igvn.replace_input_of(n, 0, NULL); 4685 igvn.hash_insert(n); 4686 modified = true; 4687 } 4688 } 4689 if (identical > 0) { 4690 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4691 } else if (_expensive_nodes->length() >= 1) { 4692 Node* n = _expensive_nodes->at(i); 4693 igvn.replace_input_of(n, 0, NULL); 4694 igvn.hash_insert(n); 4695 modified = true; 4696 } 4697 _expensive_nodes->trunc_to(j); 4698 if (modified) { 4699 igvn.optimize(); 4700 } 4701 } 4702 4703 void Compile::add_expensive_node(Node * n) { 4704 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4705 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4706 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4707 if (OptimizeExpensiveOps) { 4708 _expensive_nodes->append(n); 4709 } else { 4710 // Clear control input and let IGVN optimize expensive nodes if 4711 // OptimizeExpensiveOps is off. 4712 n->set_req(0, NULL); 4713 } 4714 } 4715 4716 /** 4717 * Remove the speculative part of types and clean up the graph 4718 */ 4719 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4720 if (UseTypeSpeculation) { 4721 Unique_Node_List worklist; 4722 worklist.push(root()); 4723 int modified = 0; 4724 // Go over all type nodes that carry a speculative type, drop the 4725 // speculative part of the type and enqueue the node for an igvn 4726 // which may optimize it out. 4727 for (uint next = 0; next < worklist.size(); ++next) { 4728 Node *n = worklist.at(next); 4729 if (n->is_Type()) { 4730 TypeNode* tn = n->as_Type(); 4731 const Type* t = tn->type(); 4732 const Type* t_no_spec = t->remove_speculative(); 4733 if (t_no_spec != t) { 4734 bool in_hash = igvn.hash_delete(n); 4735 assert(in_hash, "node should be in igvn hash table"); 4736 tn->set_type(t_no_spec); 4737 igvn.hash_insert(n); 4738 igvn._worklist.push(n); // give it a chance to go away 4739 modified++; 4740 } 4741 } 4742 uint max = n->len(); 4743 for( uint i = 0; i < max; ++i ) { 4744 Node *m = n->in(i); 4745 if (not_a_node(m)) continue; 4746 worklist.push(m); 4747 } 4748 } 4749 // Drop the speculative part of all types in the igvn's type table 4750 igvn.remove_speculative_types(); 4751 if (modified > 0) { 4752 igvn.optimize(); 4753 } 4754 #ifdef ASSERT 4755 // Verify that after the IGVN is over no speculative type has resurfaced 4756 worklist.clear(); 4757 worklist.push(root()); 4758 for (uint next = 0; next < worklist.size(); ++next) { 4759 Node *n = worklist.at(next); 4760 const Type* t = igvn.type_or_null(n); 4761 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4762 if (n->is_Type()) { 4763 t = n->as_Type()->type(); 4764 assert(t == t->remove_speculative(), "no more speculative types"); 4765 } 4766 uint max = n->len(); 4767 for( uint i = 0; i < max; ++i ) { 4768 Node *m = n->in(i); 4769 if (not_a_node(m)) continue; 4770 worklist.push(m); 4771 } 4772 } 4773 igvn.check_no_speculative_types(); 4774 #endif 4775 } 4776 } 4777 4778 // Auxiliary method to support randomized stressing/fuzzing. 4779 // 4780 // This method can be called the arbitrary number of times, with current count 4781 // as the argument. The logic allows selecting a single candidate from the 4782 // running list of candidates as follows: 4783 // int count = 0; 4784 // Cand* selected = null; 4785 // while(cand = cand->next()) { 4786 // if (randomized_select(++count)) { 4787 // selected = cand; 4788 // } 4789 // } 4790 // 4791 // Including count equalizes the chances any candidate is "selected". 4792 // This is useful when we don't have the complete list of candidates to choose 4793 // from uniformly. In this case, we need to adjust the randomicity of the 4794 // selection, or else we will end up biasing the selection towards the latter 4795 // candidates. 4796 // 4797 // Quick back-envelope calculation shows that for the list of n candidates 4798 // the equal probability for the candidate to persist as "best" can be 4799 // achieved by replacing it with "next" k-th candidate with the probability 4800 // of 1/k. It can be easily shown that by the end of the run, the 4801 // probability for any candidate is converged to 1/n, thus giving the 4802 // uniform distribution among all the candidates. 4803 // 4804 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4805 #define RANDOMIZED_DOMAIN_POW 29 4806 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4807 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4808 bool Compile::randomized_select(int count) { 4809 assert(count > 0, "only positive"); 4810 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4811 } 4812 4813 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) { 4814 if (type != NULL && phase->type(value)->higher_equal(type)) { 4815 return value; 4816 } 4817 Node* result = NULL; 4818 if (bt == T_BYTE) { 4819 result = phase->transform(new LShiftINode(value, phase->intcon(24))); 4820 result = new RShiftINode(result, phase->intcon(24)); 4821 } else if (bt == T_BOOLEAN) { 4822 result = new AndINode(value, phase->intcon(0xFF)); 4823 } else if (bt == T_CHAR) { 4824 result = new AndINode(value,phase->intcon(0xFFFF)); 4825 } else { 4826 assert(bt == T_SHORT, "unexpected narrow type"); 4827 result = phase->transform(new LShiftINode(value, phase->intcon(16))); 4828 result = new RShiftINode(result, phase->intcon(16)); 4829 } 4830 if (transform_res) { 4831 result = phase->transform(result); 4832 } 4833 return result; 4834 } 4835 4836 4837 CloneMap& Compile::clone_map() { return _clone_map; } 4838 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4839 4840 void NodeCloneInfo::dump() const { 4841 tty->print(" {%d:%d} ", idx(), gen()); 4842 } 4843 4844 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4845 uint64_t val = value(old->_idx); 4846 NodeCloneInfo cio(val); 4847 assert(val != 0, "old node should be in the map"); 4848 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4849 insert(nnn->_idx, cin.get()); 4850 #ifndef PRODUCT 4851 if (is_debug()) { 4852 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4853 } 4854 #endif 4855 } 4856 4857 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4858 NodeCloneInfo cio(value(old->_idx)); 4859 if (cio.get() == 0) { 4860 cio.set(old->_idx, 0); 4861 insert(old->_idx, cio.get()); 4862 #ifndef PRODUCT 4863 if (is_debug()) { 4864 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4865 } 4866 #endif 4867 } 4868 clone(old, nnn, gen); 4869 } 4870 4871 int CloneMap::max_gen() const { 4872 int g = 0; 4873 DictI di(_dict); 4874 for(; di.test(); ++di) { 4875 int t = gen(di._key); 4876 if (g < t) { 4877 g = t; 4878 #ifndef PRODUCT 4879 if (is_debug()) { 4880 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4881 } 4882 #endif 4883 } 4884 } 4885 return g; 4886 } 4887 4888 void CloneMap::dump(node_idx_t key) const { 4889 uint64_t val = value(key); 4890 if (val != 0) { 4891 NodeCloneInfo ni(val); 4892 ni.dump(); 4893 } 4894 }