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