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/valuetypenode.hpp" 71 #include "opto/vectornode.hpp" 72 #include "runtime/arguments.hpp" 73 #include "runtime/sharedRuntime.hpp" 74 #include "runtime/signature.hpp" 75 #include "runtime/stubRoutines.hpp" 76 #include "runtime/timer.hpp" 77 #include "utilities/align.hpp" 78 #include "utilities/copy.hpp" 79 #include "utilities/macros.hpp" 80 #if INCLUDE_ZGC 81 #include "gc/z/c2/zBarrierSetC2.hpp" 82 #endif 83 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == NULL) { 89 _mach_constant_base_node = new MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 class IntrinsicDescPair { 101 private: 102 ciMethod* _m; 103 bool _is_virtual; 104 public: 105 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 106 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 107 ciMethod* m= elt->method(); 108 ciMethod* key_m = key->_m; 109 if (key_m < m) return -1; 110 else if (key_m > m) return 1; 111 else { 112 bool is_virtual = elt->is_virtual(); 113 bool key_virtual = key->_is_virtual; 114 if (key_virtual < is_virtual) return -1; 115 else if (key_virtual > is_virtual) return 1; 116 else return 0; 117 } 118 } 119 }; 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 121 #ifdef ASSERT 122 for (int i = 1; i < _intrinsics->length(); i++) { 123 CallGenerator* cg1 = _intrinsics->at(i-1); 124 CallGenerator* cg2 = _intrinsics->at(i); 125 assert(cg1->method() != cg2->method() 126 ? cg1->method() < cg2->method() 127 : cg1->is_virtual() < cg2->is_virtual(), 128 "compiler intrinsics list must stay sorted"); 129 } 130 #endif 131 IntrinsicDescPair pair(m, is_virtual); 132 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 if (_intrinsics == NULL) { 137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 138 } 139 int len = _intrinsics->length(); 140 bool found = false; 141 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 142 assert(!found, "registering twice"); 143 _intrinsics->insert_before(index, cg); 144 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 145 } 146 147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 148 assert(m->is_loaded(), "don't try this on unloaded methods"); 149 if (_intrinsics != NULL) { 150 bool found = false; 151 int index = intrinsic_insertion_index(m, is_virtual, found); 152 if (found) { 153 return _intrinsics->at(index); 154 } 155 } 156 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 157 if (m->intrinsic_id() != vmIntrinsics::_none && 158 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 159 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 160 if (cg != NULL) { 161 // Save it for next time: 162 register_intrinsic(cg); 163 return cg; 164 } else { 165 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 166 } 167 } 168 return NULL; 169 } 170 171 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 172 // in library_call.cpp. 173 174 175 #ifndef PRODUCT 176 // statistics gathering... 177 178 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 180 181 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 182 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 183 int oflags = _intrinsic_hist_flags[id]; 184 assert(flags != 0, "what happened?"); 185 if (is_virtual) { 186 flags |= _intrinsic_virtual; 187 } 188 bool changed = (flags != oflags); 189 if ((flags & _intrinsic_worked) != 0) { 190 juint count = (_intrinsic_hist_count[id] += 1); 191 if (count == 1) { 192 changed = true; // first time 193 } 194 // increment the overall count also: 195 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 196 } 197 if (changed) { 198 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 199 // Something changed about the intrinsic's virtuality. 200 if ((flags & _intrinsic_virtual) != 0) { 201 // This is the first use of this intrinsic as a virtual call. 202 if (oflags != 0) { 203 // We already saw it as a non-virtual, so note both cases. 204 flags |= _intrinsic_both; 205 } 206 } else if ((oflags & _intrinsic_both) == 0) { 207 // This is the first use of this intrinsic as a non-virtual 208 flags |= _intrinsic_both; 209 } 210 } 211 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 212 } 213 // update the overall flags also: 214 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 215 return changed; 216 } 217 218 static char* format_flags(int flags, char* buf) { 219 buf[0] = 0; 220 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 221 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 222 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 223 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 224 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 225 if (buf[0] == 0) strcat(buf, ","); 226 assert(buf[0] == ',', "must be"); 227 return &buf[1]; 228 } 229 230 void Compile::print_intrinsic_statistics() { 231 char flagsbuf[100]; 232 ttyLocker ttyl; 233 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 234 tty->print_cr("Compiler intrinsic usage:"); 235 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 236 if (total == 0) total = 1; // avoid div0 in case of no successes 237 #define PRINT_STAT_LINE(name, c, f) \ 238 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 239 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 240 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 241 int flags = _intrinsic_hist_flags[id]; 242 juint count = _intrinsic_hist_count[id]; 243 if ((flags | count) != 0) { 244 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 245 } 246 } 247 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 248 if (xtty != NULL) xtty->tail("statistics"); 249 } 250 251 void Compile::print_statistics() { 252 { ttyLocker ttyl; 253 if (xtty != NULL) xtty->head("statistics type='opto'"); 254 Parse::print_statistics(); 255 PhaseCCP::print_statistics(); 256 PhaseRegAlloc::print_statistics(); 257 Scheduling::print_statistics(); 258 PhasePeephole::print_statistics(); 259 PhaseIdealLoop::print_statistics(); 260 if (xtty != NULL) xtty->tail("statistics"); 261 } 262 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 263 // put this under its own <statistics> element. 264 print_intrinsic_statistics(); 265 } 266 } 267 #endif //PRODUCT 268 269 // Support for bundling info 270 Bundle* Compile::node_bundling(const Node *n) { 271 assert(valid_bundle_info(n), "oob"); 272 return &_node_bundling_base[n->_idx]; 273 } 274 275 bool Compile::valid_bundle_info(const Node *n) { 276 return (_node_bundling_limit > n->_idx); 277 } 278 279 280 void Compile::gvn_replace_by(Node* n, Node* nn) { 281 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 282 Node* use = n->last_out(i); 283 bool is_in_table = initial_gvn()->hash_delete(use); 284 uint uses_found = 0; 285 for (uint j = 0; j < use->len(); j++) { 286 if (use->in(j) == n) { 287 if (j < use->req()) 288 use->set_req(j, nn); 289 else 290 use->set_prec(j, nn); 291 uses_found++; 292 } 293 } 294 if (is_in_table) { 295 // reinsert into table 296 initial_gvn()->hash_find_insert(use); 297 } 298 record_for_igvn(use); 299 i -= uses_found; // we deleted 1 or more copies of this edge 300 } 301 } 302 303 304 static inline bool not_a_node(const Node* n) { 305 if (n == NULL) return true; 306 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 307 if (*(address*)n == badAddress) return true; // kill by Node::destruct 308 return false; 309 } 310 311 // Identify all nodes that are reachable from below, useful. 312 // Use breadth-first pass that records state in a Unique_Node_List, 313 // recursive traversal is slower. 314 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 315 int estimated_worklist_size = live_nodes(); 316 useful.map( estimated_worklist_size, NULL ); // preallocate space 317 318 // Initialize worklist 319 if (root() != NULL) { useful.push(root()); } 320 // If 'top' is cached, declare it useful to preserve cached node 321 if( cached_top_node() ) { useful.push(cached_top_node()); } 322 323 // Push all useful nodes onto the list, breadthfirst 324 for( uint next = 0; next < useful.size(); ++next ) { 325 assert( next < unique(), "Unique useful nodes < total nodes"); 326 Node *n = useful.at(next); 327 uint max = n->len(); 328 for( uint i = 0; i < max; ++i ) { 329 Node *m = n->in(i); 330 if (not_a_node(m)) continue; 331 useful.push(m); 332 } 333 } 334 } 335 336 // Update dead_node_list with any missing dead nodes using useful 337 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 338 void Compile::update_dead_node_list(Unique_Node_List &useful) { 339 uint max_idx = unique(); 340 VectorSet& useful_node_set = useful.member_set(); 341 342 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 343 // If node with index node_idx is not in useful set, 344 // mark it as dead in dead node list. 345 if (! useful_node_set.test(node_idx) ) { 346 record_dead_node(node_idx); 347 } 348 } 349 } 350 351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 352 int shift = 0; 353 for (int i = 0; i < inlines->length(); i++) { 354 CallGenerator* cg = inlines->at(i); 355 CallNode* call = cg->call_node(); 356 if (shift > 0) { 357 inlines->at_put(i-shift, cg); 358 } 359 if (!useful.member(call)) { 360 shift++; 361 } 362 } 363 inlines->trunc_to(inlines->length()-shift); 364 } 365 366 // Disconnect all useless nodes by disconnecting those at the boundary. 367 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 368 uint next = 0; 369 while (next < useful.size()) { 370 Node *n = useful.at(next++); 371 if (n->is_SafePoint()) { 372 // We're done with a parsing phase. Replaced nodes are not valid 373 // beyond that point. 374 n->as_SafePoint()->delete_replaced_nodes(); 375 } 376 // Use raw traversal of out edges since this code removes out edges 377 int max = n->outcnt(); 378 for (int j = 0; j < max; ++j) { 379 Node* child = n->raw_out(j); 380 if (! useful.member(child)) { 381 assert(!child->is_top() || child != top(), 382 "If top is cached in Compile object it is in useful list"); 383 // Only need to remove this out-edge to the useless node 384 n->raw_del_out(j); 385 --j; 386 --max; 387 } 388 } 389 if (n->outcnt() == 1 && n->has_special_unique_user()) { 390 record_for_igvn(n->unique_out()); 391 } 392 } 393 // Remove useless macro and predicate opaq nodes 394 for (int i = C->macro_count()-1; i >= 0; i--) { 395 Node* n = C->macro_node(i); 396 if (!useful.member(n)) { 397 remove_macro_node(n); 398 } 399 } 400 // Remove useless CastII nodes with range check dependency 401 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 402 Node* cast = range_check_cast_node(i); 403 if (!useful.member(cast)) { 404 remove_range_check_cast(cast); 405 } 406 } 407 // Remove useless expensive nodes 408 for (int i = C->expensive_count()-1; i >= 0; i--) { 409 Node* n = C->expensive_node(i); 410 if (!useful.member(n)) { 411 remove_expensive_node(n); 412 } 413 } 414 // Remove useless Opaque4 nodes 415 for (int i = opaque4_count() - 1; i >= 0; i--) { 416 Node* opaq = opaque4_node(i); 417 if (!useful.member(opaq)) { 418 remove_opaque4_node(opaq); 419 } 420 } 421 // Remove useless value type nodes 422 if (_value_type_nodes != NULL) { 423 _value_type_nodes->remove_useless_nodes(useful.member_set()); 424 } 425 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 426 bs->eliminate_useless_gc_barriers(useful, this); 427 // clean up the late inline lists 428 remove_useless_late_inlines(&_string_late_inlines, useful); 429 remove_useless_late_inlines(&_boxing_late_inlines, useful); 430 remove_useless_late_inlines(&_late_inlines, useful); 431 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 432 } 433 434 //------------------------------frame_size_in_words----------------------------- 435 // frame_slots in units of words 436 int Compile::frame_size_in_words() const { 437 // shift is 0 in LP32 and 1 in LP64 438 const int shift = (LogBytesPerWord - LogBytesPerInt); 439 int words = _frame_slots >> shift; 440 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 441 return words; 442 } 443 444 // To bang the stack of this compiled method we use the stack size 445 // that the interpreter would need in case of a deoptimization. This 446 // removes the need to bang the stack in the deoptimization blob which 447 // in turn simplifies stack overflow handling. 448 int Compile::bang_size_in_bytes() const { 449 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 450 } 451 452 // ============================================================================ 453 //------------------------------CompileWrapper--------------------------------- 454 class CompileWrapper : public StackObj { 455 Compile *const _compile; 456 public: 457 CompileWrapper(Compile* compile); 458 459 ~CompileWrapper(); 460 }; 461 462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 463 // the Compile* pointer is stored in the current ciEnv: 464 ciEnv* env = compile->env(); 465 assert(env == ciEnv::current(), "must already be a ciEnv active"); 466 assert(env->compiler_data() == NULL, "compile already active?"); 467 env->set_compiler_data(compile); 468 assert(compile == Compile::current(), "sanity"); 469 470 compile->set_type_dict(NULL); 471 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 472 compile->clone_map().set_clone_idx(0); 473 compile->set_type_hwm(NULL); 474 compile->set_type_last_size(0); 475 compile->set_last_tf(NULL, NULL); 476 compile->set_indexSet_arena(NULL); 477 compile->set_indexSet_free_block_list(NULL); 478 compile->init_type_arena(); 479 Type::Initialize(compile); 480 _compile->set_scratch_buffer_blob(NULL); 481 _compile->begin_method(); 482 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 483 } 484 CompileWrapper::~CompileWrapper() { 485 _compile->end_method(); 486 if (_compile->scratch_buffer_blob() != NULL) 487 BufferBlob::free(_compile->scratch_buffer_blob()); 488 _compile->env()->set_compiler_data(NULL); 489 } 490 491 492 //----------------------------print_compile_messages--------------------------- 493 void Compile::print_compile_messages() { 494 #ifndef PRODUCT 495 // Check if recompiling 496 if (_subsume_loads == false && PrintOpto) { 497 // Recompiling without allowing machine instructions to subsume loads 498 tty->print_cr("*********************************************************"); 499 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 500 tty->print_cr("*********************************************************"); 501 } 502 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 503 // Recompiling without escape analysis 504 tty->print_cr("*********************************************************"); 505 tty->print_cr("** Bailout: Recompile without escape analysis **"); 506 tty->print_cr("*********************************************************"); 507 } 508 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 509 // Recompiling without boxing elimination 510 tty->print_cr("*********************************************************"); 511 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 512 tty->print_cr("*********************************************************"); 513 } 514 if (C->directive()->BreakAtCompileOption) { 515 // Open the debugger when compiling this method. 516 tty->print("### Breaking when compiling: "); 517 method()->print_short_name(); 518 tty->cr(); 519 BREAKPOINT; 520 } 521 522 if( PrintOpto ) { 523 if (is_osr_compilation()) { 524 tty->print("[OSR]%3d", _compile_id); 525 } else { 526 tty->print("%3d", _compile_id); 527 } 528 } 529 #endif 530 } 531 532 533 //-----------------------init_scratch_buffer_blob------------------------------ 534 // Construct a temporary BufferBlob and cache it for this compile. 535 void Compile::init_scratch_buffer_blob(int const_size) { 536 // If there is already a scratch buffer blob allocated and the 537 // constant section is big enough, use it. Otherwise free the 538 // current and allocate a new one. 539 BufferBlob* blob = scratch_buffer_blob(); 540 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 541 // Use the current blob. 542 } else { 543 if (blob != NULL) { 544 BufferBlob::free(blob); 545 } 546 547 ResourceMark rm; 548 _scratch_const_size = const_size; 549 int size = C2Compiler::initial_code_buffer_size(const_size); 550 #ifdef ASSERT 551 if (C->has_scalarized_args()) { 552 // Oop verification for loading object fields from scalarized value types in the new entry point requires lots of space 553 size += 5120; 554 } 555 #endif 556 blob = BufferBlob::create("Compile::scratch_buffer", size); 557 // Record the buffer blob for next time. 558 set_scratch_buffer_blob(blob); 559 // Have we run out of code space? 560 if (scratch_buffer_blob() == NULL) { 561 // Let CompilerBroker disable further compilations. 562 record_failure("Not enough space for scratch buffer in CodeCache"); 563 return; 564 } 565 } 566 567 // Initialize the relocation buffers 568 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 569 set_scratch_locs_memory(locs_buf); 570 } 571 572 573 //-----------------------scratch_emit_size------------------------------------- 574 // Helper function that computes size by emitting code 575 uint Compile::scratch_emit_size(const Node* n) { 576 // Start scratch_emit_size section. 577 set_in_scratch_emit_size(true); 578 579 // Emit into a trash buffer and count bytes emitted. 580 // This is a pretty expensive way to compute a size, 581 // but it works well enough if seldom used. 582 // All common fixed-size instructions are given a size 583 // method by the AD file. 584 // Note that the scratch buffer blob and locs memory are 585 // allocated at the beginning of the compile task, and 586 // may be shared by several calls to scratch_emit_size. 587 // The allocation of the scratch buffer blob is particularly 588 // expensive, since it has to grab the code cache lock. 589 BufferBlob* blob = this->scratch_buffer_blob(); 590 assert(blob != NULL, "Initialize BufferBlob at start"); 591 assert(blob->size() > MAX_inst_size, "sanity"); 592 relocInfo* locs_buf = scratch_locs_memory(); 593 address blob_begin = blob->content_begin(); 594 address blob_end = (address)locs_buf; 595 assert(blob->contains(blob_end), "sanity"); 596 CodeBuffer buf(blob_begin, blob_end - blob_begin); 597 buf.initialize_consts_size(_scratch_const_size); 598 buf.initialize_stubs_size(MAX_stubs_size); 599 assert(locs_buf != NULL, "sanity"); 600 int lsize = MAX_locs_size / 3; 601 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 602 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 603 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 604 // Mark as scratch buffer. 605 buf.consts()->set_scratch_emit(); 606 buf.insts()->set_scratch_emit(); 607 buf.stubs()->set_scratch_emit(); 608 609 // Do the emission. 610 611 Label fakeL; // Fake label for branch instructions. 612 Label* saveL = NULL; 613 uint save_bnum = 0; 614 bool is_branch = n->is_MachBranch(); 615 if (is_branch) { 616 MacroAssembler masm(&buf); 617 masm.bind(fakeL); 618 n->as_MachBranch()->save_label(&saveL, &save_bnum); 619 n->as_MachBranch()->label_set(&fakeL, 0); 620 } else if (n->is_MachProlog()) { 621 saveL = ((MachPrologNode*)n)->_verified_entry; 622 ((MachPrologNode*)n)->_verified_entry = &fakeL; 623 } 624 n->emit(buf, this->regalloc()); 625 626 // Emitting into the scratch buffer should not fail 627 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 628 629 // Restore label. 630 if (is_branch) { 631 n->as_MachBranch()->label_set(saveL, save_bnum); 632 } else if (n->is_MachProlog()) { 633 ((MachPrologNode*)n)->_verified_entry = saveL; 634 } 635 636 // End scratch_emit_size section. 637 set_in_scratch_emit_size(false); 638 639 return buf.insts_size(); 640 } 641 642 643 // ============================================================================ 644 //------------------------------Compile standard------------------------------- 645 debug_only( int Compile::_debug_idx = 100000; ) 646 647 // Compile a method. entry_bci is -1 for normal compilations and indicates 648 // the continuation bci for on stack replacement. 649 650 651 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 652 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 653 : Phase(Compiler), 654 _compile_id(ci_env->compile_id()), 655 _save_argument_registers(false), 656 _subsume_loads(subsume_loads), 657 _do_escape_analysis(do_escape_analysis), 658 _eliminate_boxing(eliminate_boxing), 659 _method(target), 660 _entry_bci(osr_bci), 661 _stub_function(NULL), 662 _stub_name(NULL), 663 _stub_entry_point(NULL), 664 _max_node_limit(MaxNodeLimit), 665 _orig_pc_slot(0), 666 _orig_pc_slot_offset_in_bytes(0), 667 _sp_inc_slot(0), 668 _sp_inc_slot_offset_in_bytes(0), 669 _inlining_progress(false), 670 _inlining_incrementally(false), 671 _do_cleanup(false), 672 _has_reserved_stack_access(target->has_reserved_stack_access()), 673 #ifndef PRODUCT 674 _trace_opto_output(directive->TraceOptoOutputOption), 675 #endif 676 _has_method_handle_invokes(false), 677 _comp_arena(mtCompiler), 678 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 679 _env(ci_env), 680 _directive(directive), 681 _log(ci_env->log()), 682 _failure_reason(NULL), 683 _congraph(NULL), 684 #ifndef PRODUCT 685 _printer(IdealGraphPrinter::printer()), 686 #endif 687 _dead_node_list(comp_arena()), 688 _dead_node_count(0), 689 _node_arena(mtCompiler), 690 _old_arena(mtCompiler), 691 _mach_constant_base_node(NULL), 692 _Compile_types(mtCompiler), 693 _initial_gvn(NULL), 694 _for_igvn(NULL), 695 _warm_calls(NULL), 696 _late_inlines(comp_arena(), 2, 0, NULL), 697 _string_late_inlines(comp_arena(), 2, 0, NULL), 698 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 699 _late_inlines_pos(0), 700 _number_of_mh_late_inlines(0), 701 _print_inlining_stream(NULL), 702 _print_inlining_list(NULL), 703 _print_inlining_idx(0), 704 _print_inlining_output(NULL), 705 _replay_inline_data(NULL), 706 _java_calls(0), 707 _inner_loops(0), 708 _interpreter_frame_size(0), 709 _node_bundling_limit(0), 710 _node_bundling_base(NULL), 711 _code_buffer("Compile::Fill_buffer"), 712 _scratch_const_size(-1), 713 _in_scratch_emit_size(false) 714 #ifndef PRODUCT 715 , _in_dump_cnt(0) 716 #endif 717 { 718 C = this; 719 #ifndef PRODUCT 720 if (_printer != NULL) { 721 _printer->set_compile(this); 722 } 723 #endif 724 CompileWrapper cw(this); 725 726 if (CITimeVerbose) { 727 tty->print(" "); 728 target->holder()->name()->print(); 729 tty->print("."); 730 target->print_short_name(); 731 tty->print(" "); 732 } 733 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 734 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 735 736 #ifndef PRODUCT 737 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 738 if (!print_opto_assembly) { 739 bool print_assembly = directive->PrintAssemblyOption; 740 if (print_assembly && !Disassembler::can_decode()) { 741 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 742 print_opto_assembly = true; 743 } 744 } 745 set_print_assembly(print_opto_assembly); 746 set_parsed_irreducible_loop(false); 747 748 if (directive->ReplayInlineOption) { 749 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 750 } 751 #endif 752 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 753 set_print_intrinsics(directive->PrintIntrinsicsOption); 754 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 755 756 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 757 // Make sure the method being compiled gets its own MDO, 758 // so we can at least track the decompile_count(). 759 // Need MDO to record RTM code generation state. 760 method()->ensure_method_data(); 761 } 762 763 Init(::AliasLevel); 764 765 766 print_compile_messages(); 767 768 _ilt = InlineTree::build_inline_tree_root(); 769 770 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 771 assert(num_alias_types() >= AliasIdxRaw, ""); 772 773 #define MINIMUM_NODE_HASH 1023 774 // Node list that Iterative GVN will start with 775 Unique_Node_List for_igvn(comp_arena()); 776 set_for_igvn(&for_igvn); 777 778 // GVN that will be run immediately on new nodes 779 uint estimated_size = method()->code_size()*4+64; 780 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 781 PhaseGVN gvn(node_arena(), estimated_size); 782 set_initial_gvn(&gvn); 783 784 print_inlining_init(); 785 { // Scope for timing the parser 786 TracePhase tp("parse", &timers[_t_parser]); 787 788 // Put top into the hash table ASAP. 789 initial_gvn()->transform_no_reclaim(top()); 790 791 // Set up tf(), start(), and find a CallGenerator. 792 CallGenerator* cg = NULL; 793 if (is_osr_compilation()) { 794 const TypeTuple *domain = StartOSRNode::osr_domain(); 795 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 796 init_tf(TypeFunc::make(domain, range)); 797 StartNode* s = new StartOSRNode(root(), domain); 798 initial_gvn()->set_type_bottom(s); 799 init_start(s); 800 cg = CallGenerator::for_osr(method(), entry_bci()); 801 } else { 802 // Normal case. 803 init_tf(TypeFunc::make(method())); 804 StartNode* s = new StartNode(root(), tf()->domain_cc()); 805 initial_gvn()->set_type_bottom(s); 806 init_start(s); 807 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 808 // With java.lang.ref.reference.get() we must go through the 809 // intrinsic - even when get() is the root 810 // method of the compile - so that, if necessary, the value in 811 // the referent field of the reference object gets recorded by 812 // the pre-barrier code. 813 cg = find_intrinsic(method(), false); 814 } 815 if (cg == NULL) { 816 float past_uses = method()->interpreter_invocation_count(); 817 float expected_uses = past_uses; 818 cg = CallGenerator::for_inline(method(), expected_uses); 819 } 820 } 821 if (failing()) return; 822 if (cg == NULL) { 823 record_method_not_compilable("cannot parse method"); 824 return; 825 } 826 JVMState* jvms = build_start_state(start(), tf()); 827 if ((jvms = cg->generate(jvms)) == NULL) { 828 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 829 record_method_not_compilable("method parse failed"); 830 } 831 return; 832 } 833 GraphKit kit(jvms); 834 835 if (!kit.stopped()) { 836 // Accept return values, and transfer control we know not where. 837 // This is done by a special, unique ReturnNode bound to root. 838 return_values(kit.jvms()); 839 } 840 841 if (kit.has_exceptions()) { 842 // Any exceptions that escape from this call must be rethrown 843 // to whatever caller is dynamically above us on the stack. 844 // This is done by a special, unique RethrowNode bound to root. 845 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 846 } 847 848 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 849 850 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 851 inline_string_calls(true); 852 } 853 854 if (failing()) return; 855 856 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 857 858 // Remove clutter produced by parsing. 859 if (!failing()) { 860 ResourceMark rm; 861 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 862 } 863 } 864 865 // Note: Large methods are capped off in do_one_bytecode(). 866 if (failing()) return; 867 868 // After parsing, node notes are no longer automagic. 869 // They must be propagated by register_new_node_with_optimizer(), 870 // clone(), or the like. 871 set_default_node_notes(NULL); 872 873 for (;;) { 874 int successes = Inline_Warm(); 875 if (failing()) return; 876 if (successes == 0) break; 877 } 878 879 // Drain the list. 880 Finish_Warm(); 881 #ifndef PRODUCT 882 if (_printer && _printer->should_print(1)) { 883 _printer->print_inlining(); 884 } 885 #endif 886 887 if (failing()) return; 888 NOT_PRODUCT( verify_graph_edges(); ) 889 890 // Now optimize 891 Optimize(); 892 if (failing()) return; 893 NOT_PRODUCT( verify_graph_edges(); ) 894 895 #ifndef PRODUCT 896 if (PrintIdeal) { 897 ttyLocker ttyl; // keep the following output all in one block 898 // This output goes directly to the tty, not the compiler log. 899 // To enable tools to match it up with the compilation activity, 900 // be sure to tag this tty output with the compile ID. 901 if (xtty != NULL) { 902 xtty->head("ideal compile_id='%d'%s", compile_id(), 903 is_osr_compilation() ? " compile_kind='osr'" : 904 ""); 905 } 906 root()->dump(9999); 907 if (xtty != NULL) { 908 xtty->tail("ideal"); 909 } 910 } 911 #endif 912 913 #ifdef ASSERT 914 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 915 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 916 #endif 917 918 // Dump compilation data to replay it. 919 if (directive->DumpReplayOption) { 920 env()->dump_replay_data(_compile_id); 921 } 922 if (directive->DumpInlineOption && (ilt() != NULL)) { 923 env()->dump_inline_data(_compile_id); 924 } 925 926 // Now that we know the size of all the monitors we can add a fixed slot 927 // for the original deopt pc. 928 929 _orig_pc_slot = fixed_slots(); 930 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 931 932 if (needs_stack_repair()) { 933 // One extra slot for the special stack increment value 934 _sp_inc_slot = next_slot; 935 next_slot += 2; 936 } 937 938 set_fixed_slots(next_slot); 939 940 // Compute when to use implicit null checks. Used by matching trap based 941 // nodes and NullCheck optimization. 942 set_allowed_deopt_reasons(); 943 944 // Now generate code 945 Code_Gen(); 946 if (failing()) return; 947 948 // Check if we want to skip execution of all compiled code. 949 { 950 #ifndef PRODUCT 951 if (OptoNoExecute) { 952 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 953 return; 954 } 955 #endif 956 TracePhase tp("install_code", &timers[_t_registerMethod]); 957 958 if (is_osr_compilation()) { 959 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 960 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 961 } else { 962 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 963 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry) == -1) { 964 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, _first_block_size); 965 } 966 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry_RO) == -1) { 967 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, _first_block_size); 968 } 969 if (_code_offsets.value(CodeOffsets::Entry) == -1) { 970 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size); 971 } 972 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 973 } 974 975 env()->register_method(_method, _entry_bci, 976 &_code_offsets, 977 _orig_pc_slot_offset_in_bytes, 978 code_buffer(), 979 frame_size_in_words(), _oop_map_set, 980 &_handler_table, &_inc_table, 981 compiler, 982 has_unsafe_access(), 983 SharedRuntime::is_wide_vector(max_vector_size()), 984 rtm_state() 985 ); 986 987 if (log() != NULL) // Print code cache state into compiler log 988 log()->code_cache_state(); 989 } 990 } 991 992 //------------------------------Compile---------------------------------------- 993 // Compile a runtime stub 994 Compile::Compile( ciEnv* ci_env, 995 TypeFunc_generator generator, 996 address stub_function, 997 const char *stub_name, 998 int is_fancy_jump, 999 bool pass_tls, 1000 bool save_arg_registers, 1001 bool return_pc, 1002 DirectiveSet* directive) 1003 : Phase(Compiler), 1004 _compile_id(0), 1005 _save_argument_registers(save_arg_registers), 1006 _subsume_loads(true), 1007 _do_escape_analysis(false), 1008 _eliminate_boxing(false), 1009 _method(NULL), 1010 _entry_bci(InvocationEntryBci), 1011 _stub_function(stub_function), 1012 _stub_name(stub_name), 1013 _stub_entry_point(NULL), 1014 _max_node_limit(MaxNodeLimit), 1015 _orig_pc_slot(0), 1016 _orig_pc_slot_offset_in_bytes(0), 1017 _sp_inc_slot(0), 1018 _sp_inc_slot_offset_in_bytes(0), 1019 _inlining_progress(false), 1020 _inlining_incrementally(false), 1021 _has_reserved_stack_access(false), 1022 #ifndef PRODUCT 1023 _trace_opto_output(directive->TraceOptoOutputOption), 1024 #endif 1025 _has_method_handle_invokes(false), 1026 _comp_arena(mtCompiler), 1027 _env(ci_env), 1028 _directive(directive), 1029 _log(ci_env->log()), 1030 _failure_reason(NULL), 1031 _congraph(NULL), 1032 #ifndef PRODUCT 1033 _printer(NULL), 1034 #endif 1035 _dead_node_list(comp_arena()), 1036 _dead_node_count(0), 1037 _node_arena(mtCompiler), 1038 _old_arena(mtCompiler), 1039 _mach_constant_base_node(NULL), 1040 _Compile_types(mtCompiler), 1041 _initial_gvn(NULL), 1042 _for_igvn(NULL), 1043 _warm_calls(NULL), 1044 _number_of_mh_late_inlines(0), 1045 _print_inlining_stream(NULL), 1046 _print_inlining_list(NULL), 1047 _print_inlining_idx(0), 1048 _print_inlining_output(NULL), 1049 _replay_inline_data(NULL), 1050 _java_calls(0), 1051 _inner_loops(0), 1052 _interpreter_frame_size(0), 1053 _node_bundling_limit(0), 1054 _node_bundling_base(NULL), 1055 _code_buffer("Compile::Fill_buffer"), 1056 #ifndef PRODUCT 1057 _in_dump_cnt(0), 1058 #endif 1059 _allowed_reasons(0) { 1060 C = this; 1061 1062 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1063 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1064 1065 #ifndef PRODUCT 1066 set_print_assembly(PrintFrameConverterAssembly); 1067 set_parsed_irreducible_loop(false); 1068 #endif 1069 set_has_irreducible_loop(false); // no loops 1070 1071 CompileWrapper cw(this); 1072 Init(/*AliasLevel=*/ 0); 1073 init_tf((*generator)()); 1074 1075 { 1076 // The following is a dummy for the sake of GraphKit::gen_stub 1077 Unique_Node_List for_igvn(comp_arena()); 1078 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1079 PhaseGVN gvn(Thread::current()->resource_area(),255); 1080 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1081 gvn.transform_no_reclaim(top()); 1082 1083 GraphKit kit; 1084 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1085 } 1086 1087 NOT_PRODUCT( verify_graph_edges(); ) 1088 Code_Gen(); 1089 if (failing()) return; 1090 1091 1092 // Entry point will be accessed using compile->stub_entry_point(); 1093 if (code_buffer() == NULL) { 1094 Matcher::soft_match_failure(); 1095 } else { 1096 if (PrintAssembly && (WizardMode || Verbose)) 1097 tty->print_cr("### Stub::%s", stub_name); 1098 1099 if (!failing()) { 1100 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1101 1102 // Make the NMethod 1103 // For now we mark the frame as never safe for profile stackwalking 1104 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1105 code_buffer(), 1106 CodeOffsets::frame_never_safe, 1107 // _code_offsets.value(CodeOffsets::Frame_Complete), 1108 frame_size_in_words(), 1109 _oop_map_set, 1110 save_arg_registers); 1111 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1112 1113 _stub_entry_point = rs->entry_point(); 1114 } 1115 } 1116 } 1117 1118 //------------------------------Init------------------------------------------- 1119 // Prepare for a single compilation 1120 void Compile::Init(int aliaslevel) { 1121 _unique = 0; 1122 _regalloc = NULL; 1123 1124 _tf = NULL; // filled in later 1125 _top = NULL; // cached later 1126 _matcher = NULL; // filled in later 1127 _cfg = NULL; // filled in later 1128 1129 set_24_bit_selection_and_mode(Use24BitFP, false); 1130 1131 _node_note_array = NULL; 1132 _default_node_notes = NULL; 1133 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1134 1135 _immutable_memory = NULL; // filled in at first inquiry 1136 1137 // Globally visible Nodes 1138 // First set TOP to NULL to give safe behavior during creation of RootNode 1139 set_cached_top_node(NULL); 1140 set_root(new RootNode()); 1141 // Now that you have a Root to point to, create the real TOP 1142 set_cached_top_node( new ConNode(Type::TOP) ); 1143 set_recent_alloc(NULL, NULL); 1144 1145 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1146 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1147 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1148 env()->set_dependencies(new Dependencies(env())); 1149 1150 _fixed_slots = 0; 1151 set_has_split_ifs(false); 1152 set_has_loops(has_method() && method()->has_loops()); // first approximation 1153 set_has_stringbuilder(false); 1154 set_has_boxed_value(false); 1155 _trap_can_recompile = false; // no traps emitted yet 1156 _major_progress = true; // start out assuming good things will happen 1157 set_has_unsafe_access(false); 1158 set_max_vector_size(0); 1159 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1160 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1161 set_decompile_count(0); 1162 1163 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1164 _loop_opts_cnt = LoopOptsCount; 1165 set_do_inlining(Inline); 1166 set_max_inline_size(MaxInlineSize); 1167 set_freq_inline_size(FreqInlineSize); 1168 set_do_scheduling(OptoScheduling); 1169 set_do_count_invocations(false); 1170 set_do_method_data_update(false); 1171 1172 set_do_vector_loop(false); 1173 1174 if (AllowVectorizeOnDemand) { 1175 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1176 set_do_vector_loop(true); 1177 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());}) 1178 } else if (has_method() && method()->name() != 0 && 1179 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1180 set_do_vector_loop(true); 1181 } 1182 } 1183 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1184 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());}) 1185 1186 set_age_code(has_method() && method()->profile_aging()); 1187 set_rtm_state(NoRTM); // No RTM lock eliding by default 1188 _max_node_limit = _directive->MaxNodeLimitOption; 1189 1190 #if INCLUDE_RTM_OPT 1191 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1192 int rtm_state = method()->method_data()->rtm_state(); 1193 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1194 // Don't generate RTM lock eliding code. 1195 set_rtm_state(NoRTM); 1196 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1197 // Generate RTM lock eliding code without abort ratio calculation code. 1198 set_rtm_state(UseRTM); 1199 } else if (UseRTMDeopt) { 1200 // Generate RTM lock eliding code and include abort ratio calculation 1201 // code if UseRTMDeopt is on. 1202 set_rtm_state(ProfileRTM); 1203 } 1204 } 1205 #endif 1206 if (debug_info()->recording_non_safepoints()) { 1207 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1208 (comp_arena(), 8, 0, NULL)); 1209 set_default_node_notes(Node_Notes::make(this)); 1210 } 1211 1212 // // -- Initialize types before each compile -- 1213 // // Update cached type information 1214 // if( _method && _method->constants() ) 1215 // Type::update_loaded_types(_method, _method->constants()); 1216 1217 // Init alias_type map. 1218 if (!_do_escape_analysis && aliaslevel == 3) 1219 aliaslevel = 2; // No unique types without escape analysis 1220 _AliasLevel = aliaslevel; 1221 const int grow_ats = 16; 1222 _max_alias_types = grow_ats; 1223 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1224 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1225 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1226 { 1227 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1228 } 1229 // Initialize the first few types. 1230 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1231 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1232 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1233 _num_alias_types = AliasIdxRaw+1; 1234 // Zero out the alias type cache. 1235 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1236 // A NULL adr_type hits in the cache right away. Preload the right answer. 1237 probe_alias_cache(NULL)->_index = AliasIdxTop; 1238 1239 _intrinsics = NULL; 1240 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1241 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1242 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1243 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1244 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1245 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1246 register_library_intrinsics(); 1247 } 1248 1249 //---------------------------init_start---------------------------------------- 1250 // Install the StartNode on this compile object. 1251 void Compile::init_start(StartNode* s) { 1252 if (failing()) 1253 return; // already failing 1254 assert(s == start(), ""); 1255 } 1256 1257 /** 1258 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1259 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1260 * the ideal graph. 1261 */ 1262 StartNode* Compile::start() const { 1263 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1264 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1265 Node* start = root()->fast_out(i); 1266 if (start->is_Start()) { 1267 return start->as_Start(); 1268 } 1269 } 1270 fatal("Did not find Start node!"); 1271 return NULL; 1272 } 1273 1274 //-------------------------------immutable_memory------------------------------------- 1275 // Access immutable memory 1276 Node* Compile::immutable_memory() { 1277 if (_immutable_memory != NULL) { 1278 return _immutable_memory; 1279 } 1280 StartNode* s = start(); 1281 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1282 Node *p = s->fast_out(i); 1283 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1284 _immutable_memory = p; 1285 return _immutable_memory; 1286 } 1287 } 1288 ShouldNotReachHere(); 1289 return NULL; 1290 } 1291 1292 //----------------------set_cached_top_node------------------------------------ 1293 // Install the cached top node, and make sure Node::is_top works correctly. 1294 void Compile::set_cached_top_node(Node* tn) { 1295 if (tn != NULL) verify_top(tn); 1296 Node* old_top = _top; 1297 _top = tn; 1298 // Calling Node::setup_is_top allows the nodes the chance to adjust 1299 // their _out arrays. 1300 if (_top != NULL) _top->setup_is_top(); 1301 if (old_top != NULL) old_top->setup_is_top(); 1302 assert(_top == NULL || top()->is_top(), ""); 1303 } 1304 1305 #ifdef ASSERT 1306 uint Compile::count_live_nodes_by_graph_walk() { 1307 Unique_Node_List useful(comp_arena()); 1308 // Get useful node list by walking the graph. 1309 identify_useful_nodes(useful); 1310 return useful.size(); 1311 } 1312 1313 void Compile::print_missing_nodes() { 1314 1315 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1316 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1317 return; 1318 } 1319 1320 // This is an expensive function. It is executed only when the user 1321 // specifies VerifyIdealNodeCount option or otherwise knows the 1322 // additional work that needs to be done to identify reachable nodes 1323 // by walking the flow graph and find the missing ones using 1324 // _dead_node_list. 1325 1326 Unique_Node_List useful(comp_arena()); 1327 // Get useful node list by walking the graph. 1328 identify_useful_nodes(useful); 1329 1330 uint l_nodes = C->live_nodes(); 1331 uint l_nodes_by_walk = useful.size(); 1332 1333 if (l_nodes != l_nodes_by_walk) { 1334 if (_log != NULL) { 1335 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1336 _log->stamp(); 1337 _log->end_head(); 1338 } 1339 VectorSet& useful_member_set = useful.member_set(); 1340 int last_idx = l_nodes_by_walk; 1341 for (int i = 0; i < last_idx; i++) { 1342 if (useful_member_set.test(i)) { 1343 if (_dead_node_list.test(i)) { 1344 if (_log != NULL) { 1345 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1346 } 1347 if (PrintIdealNodeCount) { 1348 // Print the log message to tty 1349 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1350 useful.at(i)->dump(); 1351 } 1352 } 1353 } 1354 else if (! _dead_node_list.test(i)) { 1355 if (_log != NULL) { 1356 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1357 } 1358 if (PrintIdealNodeCount) { 1359 // Print the log message to tty 1360 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1361 } 1362 } 1363 } 1364 if (_log != NULL) { 1365 _log->tail("mismatched_nodes"); 1366 } 1367 } 1368 } 1369 void Compile::record_modified_node(Node* n) { 1370 if (_modified_nodes != NULL && !_inlining_incrementally && 1371 n->outcnt() != 0 && !n->is_Con()) { 1372 _modified_nodes->push(n); 1373 } 1374 } 1375 1376 void Compile::remove_modified_node(Node* n) { 1377 if (_modified_nodes != NULL) { 1378 _modified_nodes->remove(n); 1379 } 1380 } 1381 #endif 1382 1383 #ifndef PRODUCT 1384 void Compile::verify_top(Node* tn) const { 1385 if (tn != NULL) { 1386 assert(tn->is_Con(), "top node must be a constant"); 1387 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1388 assert(tn->in(0) != NULL, "must have live top node"); 1389 } 1390 } 1391 #endif 1392 1393 1394 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1395 1396 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1397 guarantee(arr != NULL, ""); 1398 int num_blocks = arr->length(); 1399 if (grow_by < num_blocks) grow_by = num_blocks; 1400 int num_notes = grow_by * _node_notes_block_size; 1401 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1402 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1403 while (num_notes > 0) { 1404 arr->append(notes); 1405 notes += _node_notes_block_size; 1406 num_notes -= _node_notes_block_size; 1407 } 1408 assert(num_notes == 0, "exact multiple, please"); 1409 } 1410 1411 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1412 if (source == NULL || dest == NULL) return false; 1413 1414 if (dest->is_Con()) 1415 return false; // Do not push debug info onto constants. 1416 1417 #ifdef ASSERT 1418 // Leave a bread crumb trail pointing to the original node: 1419 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1420 dest->set_debug_orig(source); 1421 } 1422 #endif 1423 1424 if (node_note_array() == NULL) 1425 return false; // Not collecting any notes now. 1426 1427 // This is a copy onto a pre-existing node, which may already have notes. 1428 // If both nodes have notes, do not overwrite any pre-existing notes. 1429 Node_Notes* source_notes = node_notes_at(source->_idx); 1430 if (source_notes == NULL || source_notes->is_clear()) return false; 1431 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1432 if (dest_notes == NULL || dest_notes->is_clear()) { 1433 return set_node_notes_at(dest->_idx, source_notes); 1434 } 1435 1436 Node_Notes merged_notes = (*source_notes); 1437 // The order of operations here ensures that dest notes will win... 1438 merged_notes.update_from(dest_notes); 1439 return set_node_notes_at(dest->_idx, &merged_notes); 1440 } 1441 1442 1443 //--------------------------allow_range_check_smearing------------------------- 1444 // Gating condition for coalescing similar range checks. 1445 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1446 // single covering check that is at least as strong as any of them. 1447 // If the optimization succeeds, the simplified (strengthened) range check 1448 // will always succeed. If it fails, we will deopt, and then give up 1449 // on the optimization. 1450 bool Compile::allow_range_check_smearing() const { 1451 // If this method has already thrown a range-check, 1452 // assume it was because we already tried range smearing 1453 // and it failed. 1454 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1455 return !already_trapped; 1456 } 1457 1458 1459 //------------------------------flatten_alias_type----------------------------- 1460 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1461 int offset = tj->offset(); 1462 TypePtr::PTR ptr = tj->ptr(); 1463 1464 // Known instance (scalarizable allocation) alias only with itself. 1465 bool is_known_inst = tj->isa_oopptr() != NULL && 1466 tj->is_oopptr()->is_known_instance(); 1467 1468 // Process weird unsafe references. 1469 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1470 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass(); 1471 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops"); 1472 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1473 tj = TypeOopPtr::BOTTOM; 1474 ptr = tj->ptr(); 1475 offset = tj->offset(); 1476 } 1477 1478 // Array pointers need some flattening 1479 const TypeAryPtr *ta = tj->isa_aryptr(); 1480 if (ta && ta->is_stable()) { 1481 // Erase stability property for alias analysis. 1482 tj = ta = ta->cast_to_stable(false); 1483 } 1484 if( ta && is_known_inst ) { 1485 if ( offset != Type::OffsetBot && 1486 offset > arrayOopDesc::length_offset_in_bytes() ) { 1487 offset = Type::OffsetBot; // Flatten constant access into array body only 1488 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1489 } 1490 } else if( ta && _AliasLevel >= 2 ) { 1491 // For arrays indexed by constant indices, we flatten the alias 1492 // space to include all of the array body. Only the header, klass 1493 // and array length can be accessed un-aliased. 1494 // For flattened value type array, each field has its own slice so 1495 // we must include the field offset. 1496 if( offset != Type::OffsetBot ) { 1497 if( ta->const_oop() ) { // MethodData* or Method* 1498 offset = Type::OffsetBot; // Flatten constant access into array body 1499 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1500 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1501 // range is OK as-is. 1502 tj = ta = TypeAryPtr::RANGE; 1503 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1504 tj = TypeInstPtr::KLASS; // all klass loads look alike 1505 ta = TypeAryPtr::RANGE; // generic ignored junk 1506 ptr = TypePtr::BotPTR; 1507 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1508 tj = TypeInstPtr::MARK; 1509 ta = TypeAryPtr::RANGE; // generic ignored junk 1510 ptr = TypePtr::BotPTR; 1511 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1512 ta = tj->isa_aryptr(); 1513 } else { // Random constant offset into array body 1514 offset = Type::OffsetBot; // Flatten constant access into array body 1515 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1516 } 1517 } 1518 // Arrays of fixed size alias with arrays of unknown size. 1519 if (ta->size() != TypeInt::POS) { 1520 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1521 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1522 } 1523 // Arrays of known objects become arrays of unknown objects. 1524 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1525 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1526 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1527 } 1528 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1529 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1530 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1531 } 1532 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1533 // cannot be distinguished by bytecode alone. 1534 if (ta->elem() == TypeInt::BOOL) { 1535 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1536 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1537 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1538 } 1539 // During the 2nd round of IterGVN, NotNull castings are removed. 1540 // Make sure the Bottom and NotNull variants alias the same. 1541 // Also, make sure exact and non-exact variants alias the same. 1542 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1543 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1544 } 1545 } 1546 1547 // Oop pointers need some flattening 1548 const TypeInstPtr *to = tj->isa_instptr(); 1549 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1550 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1551 if( ptr == TypePtr::Constant ) { 1552 if (to->klass() != ciEnv::current()->Class_klass() || 1553 offset < k->size_helper() * wordSize) { 1554 // No constant oop pointers (such as Strings); they alias with 1555 // unknown strings. 1556 assert(!is_known_inst, "not scalarizable allocation"); 1557 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1558 } 1559 } else if( is_known_inst ) { 1560 tj = to; // Keep NotNull and klass_is_exact for instance type 1561 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1562 // During the 2nd round of IterGVN, NotNull castings are removed. 1563 // Make sure the Bottom and NotNull variants alias the same. 1564 // Also, make sure exact and non-exact variants alias the same. 1565 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1566 } 1567 if (to->speculative() != NULL) { 1568 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1569 } 1570 // Canonicalize the holder of this field 1571 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1572 // First handle header references such as a LoadKlassNode, even if the 1573 // object's klass is unloaded at compile time (4965979). 1574 if (!is_known_inst) { // Do it only for non-instance types 1575 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1576 } 1577 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1578 to = tj->is_instptr(); 1579 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1580 // Static fields are in the space above the normal instance 1581 // fields in the java.lang.Class instance. 1582 if (to->klass() != ciEnv::current()->Class_klass()) { 1583 to = NULL; 1584 tj = TypeOopPtr::BOTTOM; 1585 offset = tj->offset(); 1586 } 1587 } else { 1588 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1589 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1590 if( is_known_inst ) { 1591 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1592 } else { 1593 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1594 } 1595 } 1596 } 1597 } 1598 1599 // Klass pointers to object array klasses need some flattening 1600 const TypeKlassPtr *tk = tj->isa_klassptr(); 1601 if( tk ) { 1602 // If we are referencing a field within a Klass, we need 1603 // to assume the worst case of an Object. Both exact and 1604 // inexact types must flatten to the same alias class so 1605 // use NotNull as the PTR. 1606 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1607 1608 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1609 TypeKlassPtr::OBJECT->klass(), 1610 Type::Offset(offset)); 1611 } 1612 1613 ciKlass* klass = tk->klass(); 1614 if (klass != NULL && klass->is_obj_array_klass()) { 1615 ciKlass* k = TypeAryPtr::OOPS->klass(); 1616 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1617 k = TypeInstPtr::BOTTOM->klass(); 1618 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1619 } 1620 1621 // Check for precise loads from the primary supertype array and force them 1622 // to the supertype cache alias index. Check for generic array loads from 1623 // the primary supertype array and also force them to the supertype cache 1624 // alias index. Since the same load can reach both, we need to merge 1625 // these 2 disparate memories into the same alias class. Since the 1626 // primary supertype array is read-only, there's no chance of confusion 1627 // where we bypass an array load and an array store. 1628 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1629 if (offset == Type::OffsetBot || 1630 (offset >= primary_supers_offset && 1631 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1632 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1633 offset = in_bytes(Klass::secondary_super_cache_offset()); 1634 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1635 } 1636 } 1637 1638 // Flatten all Raw pointers together. 1639 if (tj->base() == Type::RawPtr) 1640 tj = TypeRawPtr::BOTTOM; 1641 1642 if (tj->base() == Type::AnyPtr) 1643 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1644 1645 // Flatten all to bottom for now 1646 switch( _AliasLevel ) { 1647 case 0: 1648 tj = TypePtr::BOTTOM; 1649 break; 1650 case 1: // Flatten to: oop, static, field or array 1651 switch (tj->base()) { 1652 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1653 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1654 case Type::AryPtr: // do not distinguish arrays at all 1655 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1656 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1657 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1658 default: ShouldNotReachHere(); 1659 } 1660 break; 1661 case 2: // No collapsing at level 2; keep all splits 1662 case 3: // No collapsing at level 3; keep all splits 1663 break; 1664 default: 1665 Unimplemented(); 1666 } 1667 1668 offset = tj->offset(); 1669 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1670 1671 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1672 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1673 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1674 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1675 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1676 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1677 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1678 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), 1679 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1680 assert( tj->ptr() != TypePtr::TopPTR && 1681 tj->ptr() != TypePtr::AnyNull && 1682 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1683 // assert( tj->ptr() != TypePtr::Constant || 1684 // tj->base() == Type::RawPtr || 1685 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1686 1687 return tj; 1688 } 1689 1690 void Compile::AliasType::Init(int i, const TypePtr* at) { 1691 _index = i; 1692 _adr_type = at; 1693 _field = NULL; 1694 _element = NULL; 1695 _is_rewritable = true; // default 1696 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1697 if (atoop != NULL && atoop->is_known_instance()) { 1698 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1699 _general_index = Compile::current()->get_alias_index(gt); 1700 } else { 1701 _general_index = 0; 1702 } 1703 } 1704 1705 BasicType Compile::AliasType::basic_type() const { 1706 if (element() != NULL) { 1707 const Type* element = adr_type()->is_aryptr()->elem(); 1708 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1709 } if (field() != NULL) { 1710 return field()->layout_type(); 1711 } else { 1712 return T_ILLEGAL; // unknown 1713 } 1714 } 1715 1716 //---------------------------------print_on------------------------------------ 1717 #ifndef PRODUCT 1718 void Compile::AliasType::print_on(outputStream* st) { 1719 if (index() < 10) 1720 st->print("@ <%d> ", index()); 1721 else st->print("@ <%d>", index()); 1722 st->print(is_rewritable() ? " " : " RO"); 1723 int offset = adr_type()->offset(); 1724 if (offset == Type::OffsetBot) 1725 st->print(" +any"); 1726 else st->print(" +%-3d", offset); 1727 st->print(" in "); 1728 adr_type()->dump_on(st); 1729 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1730 if (field() != NULL && tjp) { 1731 if (tjp->klass() != field()->holder() || 1732 tjp->offset() != field()->offset_in_bytes()) { 1733 st->print(" != "); 1734 field()->print(); 1735 st->print(" ***"); 1736 } 1737 } 1738 } 1739 1740 void print_alias_types() { 1741 Compile* C = Compile::current(); 1742 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1743 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1744 C->alias_type(idx)->print_on(tty); 1745 tty->cr(); 1746 } 1747 } 1748 #endif 1749 1750 1751 //----------------------------probe_alias_cache-------------------------------- 1752 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1753 intptr_t key = (intptr_t) adr_type; 1754 key ^= key >> logAliasCacheSize; 1755 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1756 } 1757 1758 1759 //-----------------------------grow_alias_types-------------------------------- 1760 void Compile::grow_alias_types() { 1761 const int old_ats = _max_alias_types; // how many before? 1762 const int new_ats = old_ats; // how many more? 1763 const int grow_ats = old_ats+new_ats; // how many now? 1764 _max_alias_types = grow_ats; 1765 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1766 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1767 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1768 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1769 } 1770 1771 1772 //--------------------------------find_alias_type------------------------------ 1773 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1774 if (_AliasLevel == 0) 1775 return alias_type(AliasIdxBot); 1776 1777 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1778 if (ace->_adr_type == adr_type) { 1779 return alias_type(ace->_index); 1780 } 1781 1782 // Handle special cases. 1783 if (adr_type == NULL) return alias_type(AliasIdxTop); 1784 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1785 1786 // Do it the slow way. 1787 const TypePtr* flat = flatten_alias_type(adr_type); 1788 1789 #ifdef ASSERT 1790 { 1791 ResourceMark rm; 1792 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1793 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1794 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1795 Type::str(adr_type)); 1796 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1797 const TypeOopPtr* foop = flat->is_oopptr(); 1798 // Scalarizable allocations have exact klass always. 1799 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1800 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1801 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1802 Type::str(foop), Type::str(xoop)); 1803 } 1804 } 1805 #endif 1806 1807 int idx = AliasIdxTop; 1808 for (int i = 0; i < num_alias_types(); i++) { 1809 if (alias_type(i)->adr_type() == flat) { 1810 idx = i; 1811 break; 1812 } 1813 } 1814 1815 if (idx == AliasIdxTop) { 1816 if (no_create) return NULL; 1817 // Grow the array if necessary. 1818 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1819 // Add a new alias type. 1820 idx = _num_alias_types++; 1821 _alias_types[idx]->Init(idx, flat); 1822 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1823 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1824 if (flat->isa_instptr()) { 1825 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1826 && flat->is_instptr()->klass() == env()->Class_klass()) 1827 alias_type(idx)->set_rewritable(false); 1828 } 1829 ciField* field = NULL; 1830 if (flat->isa_aryptr()) { 1831 #ifdef ASSERT 1832 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1833 // (T_BYTE has the weakest alignment and size restrictions...) 1834 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1835 #endif 1836 const Type* elemtype = flat->is_aryptr()->elem(); 1837 if (flat->offset() == TypePtr::OffsetBot) { 1838 alias_type(idx)->set_element(elemtype); 1839 } 1840 int field_offset = flat->is_aryptr()->field_offset().get(); 1841 if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) { 1842 ciValueKlass* vk = elemtype->is_valuetype()->value_klass(); 1843 field_offset += vk->first_field_offset(); 1844 field = vk->get_field_by_offset(field_offset, false); 1845 } 1846 } 1847 if (flat->isa_klassptr()) { 1848 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1849 alias_type(idx)->set_rewritable(false); 1850 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1851 alias_type(idx)->set_rewritable(false); 1852 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1853 alias_type(idx)->set_rewritable(false); 1854 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1855 alias_type(idx)->set_rewritable(false); 1856 } 1857 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1858 // but the base pointer type is not distinctive enough to identify 1859 // references into JavaThread.) 1860 1861 // Check for final fields. 1862 const TypeInstPtr* tinst = flat->isa_instptr(); 1863 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1864 if (tinst->const_oop() != NULL && 1865 tinst->klass() == ciEnv::current()->Class_klass() && 1866 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1867 // static field 1868 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1869 field = k->get_field_by_offset(tinst->offset(), true); 1870 } else if (tinst->klass()->is_valuetype()) { 1871 // Value type field 1872 ciValueKlass* vk = tinst->value_klass(); 1873 field = vk->get_field_by_offset(tinst->offset(), false); 1874 } else { 1875 ciInstanceKlass* k = tinst->klass()->as_instance_klass(); 1876 field = k->get_field_by_offset(tinst->offset(), false); 1877 } 1878 } 1879 assert(field == NULL || 1880 original_field == NULL || 1881 (field->holder() == original_field->holder() && 1882 field->offset() == original_field->offset() && 1883 field->is_static() == original_field->is_static()), "wrong field?"); 1884 // Set field() and is_rewritable() attributes. 1885 if (field != NULL) alias_type(idx)->set_field(field); 1886 } 1887 1888 // Fill the cache for next time. 1889 ace->_adr_type = adr_type; 1890 ace->_index = idx; 1891 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1892 1893 // Might as well try to fill the cache for the flattened version, too. 1894 AliasCacheEntry* face = probe_alias_cache(flat); 1895 if (face->_adr_type == NULL) { 1896 face->_adr_type = flat; 1897 face->_index = idx; 1898 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1899 } 1900 1901 return alias_type(idx); 1902 } 1903 1904 1905 Compile::AliasType* Compile::alias_type(ciField* field) { 1906 const TypeOopPtr* t; 1907 if (field->is_static()) 1908 t = TypeInstPtr::make(field->holder()->java_mirror()); 1909 else 1910 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1911 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1912 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1913 return atp; 1914 } 1915 1916 1917 //------------------------------have_alias_type-------------------------------- 1918 bool Compile::have_alias_type(const TypePtr* adr_type) { 1919 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1920 if (ace->_adr_type == adr_type) { 1921 return true; 1922 } 1923 1924 // Handle special cases. 1925 if (adr_type == NULL) return true; 1926 if (adr_type == TypePtr::BOTTOM) return true; 1927 1928 return find_alias_type(adr_type, true, NULL) != NULL; 1929 } 1930 1931 //-----------------------------must_alias-------------------------------------- 1932 // True if all values of the given address type are in the given alias category. 1933 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1934 if (alias_idx == AliasIdxBot) return true; // the universal category 1935 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1936 if (alias_idx == AliasIdxTop) return false; // the empty category 1937 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1938 1939 // the only remaining possible overlap is identity 1940 int adr_idx = get_alias_index(adr_type); 1941 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1942 assert(adr_idx == alias_idx || 1943 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1944 && adr_type != TypeOopPtr::BOTTOM), 1945 "should not be testing for overlap with an unsafe pointer"); 1946 return adr_idx == alias_idx; 1947 } 1948 1949 //------------------------------can_alias-------------------------------------- 1950 // True if any values of the given address type are in the given alias category. 1951 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1952 if (alias_idx == AliasIdxTop) return false; // the empty category 1953 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1954 if (alias_idx == AliasIdxBot) return true; // the universal category 1955 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1956 1957 // the only remaining possible overlap is identity 1958 int adr_idx = get_alias_index(adr_type); 1959 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1960 return adr_idx == alias_idx; 1961 } 1962 1963 1964 1965 //---------------------------pop_warm_call------------------------------------- 1966 WarmCallInfo* Compile::pop_warm_call() { 1967 WarmCallInfo* wci = _warm_calls; 1968 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1969 return wci; 1970 } 1971 1972 //----------------------------Inline_Warm-------------------------------------- 1973 int Compile::Inline_Warm() { 1974 // If there is room, try to inline some more warm call sites. 1975 // %%% Do a graph index compaction pass when we think we're out of space? 1976 if (!InlineWarmCalls) return 0; 1977 1978 int calls_made_hot = 0; 1979 int room_to_grow = NodeCountInliningCutoff - unique(); 1980 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1981 int amount_grown = 0; 1982 WarmCallInfo* call; 1983 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1984 int est_size = (int)call->size(); 1985 if (est_size > (room_to_grow - amount_grown)) { 1986 // This one won't fit anyway. Get rid of it. 1987 call->make_cold(); 1988 continue; 1989 } 1990 call->make_hot(); 1991 calls_made_hot++; 1992 amount_grown += est_size; 1993 amount_to_grow -= est_size; 1994 } 1995 1996 if (calls_made_hot > 0) set_major_progress(); 1997 return calls_made_hot; 1998 } 1999 2000 2001 //----------------------------Finish_Warm-------------------------------------- 2002 void Compile::Finish_Warm() { 2003 if (!InlineWarmCalls) return; 2004 if (failing()) return; 2005 if (warm_calls() == NULL) return; 2006 2007 // Clean up loose ends, if we are out of space for inlining. 2008 WarmCallInfo* call; 2009 while ((call = pop_warm_call()) != NULL) { 2010 call->make_cold(); 2011 } 2012 } 2013 2014 //---------------------cleanup_loop_predicates----------------------- 2015 // Remove the opaque nodes that protect the predicates so that all unused 2016 // checks and uncommon_traps will be eliminated from the ideal graph 2017 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 2018 if (predicate_count()==0) return; 2019 for (int i = predicate_count(); i > 0; i--) { 2020 Node * n = predicate_opaque1_node(i-1); 2021 assert(n->Opcode() == Op_Opaque1, "must be"); 2022 igvn.replace_node(n, n->in(1)); 2023 } 2024 assert(predicate_count()==0, "should be clean!"); 2025 } 2026 2027 void Compile::add_range_check_cast(Node* n) { 2028 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2029 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 2030 _range_check_casts->append(n); 2031 } 2032 2033 // Remove all range check dependent CastIINodes. 2034 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 2035 for (int i = range_check_cast_count(); i > 0; i--) { 2036 Node* cast = range_check_cast_node(i-1); 2037 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2038 igvn.replace_node(cast, cast->in(1)); 2039 } 2040 assert(range_check_cast_count() == 0, "should be empty"); 2041 } 2042 2043 void Compile::add_opaque4_node(Node* n) { 2044 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 2045 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 2046 _opaque4_nodes->append(n); 2047 } 2048 2049 // Remove all Opaque4 nodes. 2050 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 2051 for (int i = opaque4_count(); i > 0; i--) { 2052 Node* opaq = opaque4_node(i-1); 2053 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2054 igvn.replace_node(opaq, opaq->in(2)); 2055 } 2056 assert(opaque4_count() == 0, "should be empty"); 2057 } 2058 2059 void Compile::add_value_type(Node* n) { 2060 assert(n->is_ValueTypeBase(), "unexpected node"); 2061 if (_value_type_nodes != NULL) { 2062 _value_type_nodes->push(n); 2063 } 2064 } 2065 2066 void Compile::remove_value_type(Node* n) { 2067 assert(n->is_ValueTypeBase(), "unexpected node"); 2068 if (_value_type_nodes != NULL) { 2069 _value_type_nodes->remove(n); 2070 } 2071 } 2072 2073 void Compile::process_value_types(PhaseIterGVN &igvn) { 2074 // Make value types scalar in safepoints 2075 while (_value_type_nodes->size() != 0) { 2076 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 2077 vt->make_scalar_in_safepoints(&igvn); 2078 if (vt->is_ValueTypePtr()) { 2079 igvn.replace_node(vt, vt->get_oop()); 2080 } else { 2081 if (vt->outcnt() == 0) { 2082 igvn.remove_dead_node(vt); 2083 } 2084 } 2085 } 2086 _value_type_nodes = NULL; 2087 igvn.optimize(); 2088 } 2089 2090 // StringOpts and late inlining of string methods 2091 void Compile::inline_string_calls(bool parse_time) { 2092 { 2093 // remove useless nodes to make the usage analysis simpler 2094 ResourceMark rm; 2095 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2096 } 2097 2098 { 2099 ResourceMark rm; 2100 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2101 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2102 print_method(PHASE_AFTER_STRINGOPTS, 3); 2103 } 2104 2105 // now inline anything that we skipped the first time around 2106 if (!parse_time) { 2107 _late_inlines_pos = _late_inlines.length(); 2108 } 2109 2110 while (_string_late_inlines.length() > 0) { 2111 CallGenerator* cg = _string_late_inlines.pop(); 2112 cg->do_late_inline(); 2113 if (failing()) return; 2114 } 2115 _string_late_inlines.trunc_to(0); 2116 } 2117 2118 // Late inlining of boxing methods 2119 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2120 if (_boxing_late_inlines.length() > 0) { 2121 assert(has_boxed_value(), "inconsistent"); 2122 2123 PhaseGVN* gvn = initial_gvn(); 2124 set_inlining_incrementally(true); 2125 2126 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2127 for_igvn()->clear(); 2128 gvn->replace_with(&igvn); 2129 2130 _late_inlines_pos = _late_inlines.length(); 2131 2132 while (_boxing_late_inlines.length() > 0) { 2133 CallGenerator* cg = _boxing_late_inlines.pop(); 2134 cg->do_late_inline(); 2135 if (failing()) return; 2136 } 2137 _boxing_late_inlines.trunc_to(0); 2138 2139 inline_incrementally_cleanup(igvn); 2140 2141 set_inlining_incrementally(false); 2142 } 2143 } 2144 2145 bool Compile::inline_incrementally_one() { 2146 assert(IncrementalInline, "incremental inlining should be on"); 2147 2148 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2149 set_inlining_progress(false); 2150 set_do_cleanup(false); 2151 int i = 0; 2152 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2153 CallGenerator* cg = _late_inlines.at(i); 2154 _late_inlines_pos = i+1; 2155 cg->do_late_inline(); 2156 if (failing()) return false; 2157 } 2158 int j = 0; 2159 for (; i < _late_inlines.length(); i++, j++) { 2160 _late_inlines.at_put(j, _late_inlines.at(i)); 2161 } 2162 _late_inlines.trunc_to(j); 2163 assert(inlining_progress() || _late_inlines.length() == 0, ""); 2164 2165 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2166 2167 set_inlining_progress(false); 2168 set_do_cleanup(false); 2169 return (_late_inlines.length() > 0) && !needs_cleanup; 2170 } 2171 2172 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2173 { 2174 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2175 ResourceMark rm; 2176 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2177 } 2178 { 2179 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2180 igvn = PhaseIterGVN(initial_gvn()); 2181 igvn.optimize(); 2182 } 2183 } 2184 2185 // Perform incremental inlining until bound on number of live nodes is reached 2186 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2187 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2188 2189 set_inlining_incrementally(true); 2190 uint low_live_nodes = 0; 2191 2192 while (_late_inlines.length() > 0) { 2193 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2194 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2195 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2196 // PhaseIdealLoop is expensive so we only try it once we are 2197 // out of live nodes and we only try it again if the previous 2198 // helped got the number of nodes down significantly 2199 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2200 if (failing()) return; 2201 low_live_nodes = live_nodes(); 2202 _major_progress = true; 2203 } 2204 2205 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2206 break; // finish 2207 } 2208 } 2209 2210 for_igvn()->clear(); 2211 initial_gvn()->replace_with(&igvn); 2212 2213 while (inline_incrementally_one()) { 2214 assert(!failing(), "inconsistent"); 2215 } 2216 2217 if (failing()) return; 2218 2219 inline_incrementally_cleanup(igvn); 2220 2221 if (failing()) return; 2222 } 2223 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2224 2225 if (_string_late_inlines.length() > 0) { 2226 assert(has_stringbuilder(), "inconsistent"); 2227 for_igvn()->clear(); 2228 initial_gvn()->replace_with(&igvn); 2229 2230 inline_string_calls(false); 2231 2232 if (failing()) return; 2233 2234 inline_incrementally_cleanup(igvn); 2235 } 2236 2237 set_inlining_incrementally(false); 2238 } 2239 2240 2241 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2242 if(_loop_opts_cnt > 0) { 2243 debug_only( int cnt = 0; ); 2244 while(major_progress() && (_loop_opts_cnt > 0)) { 2245 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2246 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2247 PhaseIdealLoop::optimize(igvn, mode); 2248 _loop_opts_cnt--; 2249 if (failing()) return false; 2250 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2251 } 2252 } 2253 return true; 2254 } 2255 2256 // Remove edges from "root" to each SafePoint at a backward branch. 2257 // They were inserted during parsing (see add_safepoint()) to make 2258 // infinite loops without calls or exceptions visible to root, i.e., 2259 // useful. 2260 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2261 Node *r = root(); 2262 if (r != NULL) { 2263 for (uint i = r->req(); i < r->len(); ++i) { 2264 Node *n = r->in(i); 2265 if (n != NULL && n->is_SafePoint()) { 2266 r->rm_prec(i); 2267 if (n->outcnt() == 0) { 2268 igvn.remove_dead_node(n); 2269 } 2270 --i; 2271 } 2272 } 2273 } 2274 } 2275 2276 //------------------------------Optimize--------------------------------------- 2277 // Given a graph, optimize it. 2278 void Compile::Optimize() { 2279 TracePhase tp("optimizer", &timers[_t_optimizer]); 2280 2281 #ifndef PRODUCT 2282 if (_directive->BreakAtCompileOption) { 2283 BREAKPOINT; 2284 } 2285 2286 #endif 2287 2288 #ifdef ASSERT 2289 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2290 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2291 #endif 2292 2293 ResourceMark rm; 2294 2295 print_inlining_reinit(); 2296 2297 NOT_PRODUCT( verify_graph_edges(); ) 2298 2299 print_method(PHASE_AFTER_PARSING); 2300 2301 { 2302 // Iterative Global Value Numbering, including ideal transforms 2303 // Initialize IterGVN with types and values from parse-time GVN 2304 PhaseIterGVN igvn(initial_gvn()); 2305 #ifdef ASSERT 2306 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2307 #endif 2308 { 2309 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2310 igvn.optimize(); 2311 } 2312 2313 if (failing()) return; 2314 2315 print_method(PHASE_ITER_GVN1, 2); 2316 2317 inline_incrementally(igvn); 2318 2319 print_method(PHASE_INCREMENTAL_INLINE, 2); 2320 2321 if (failing()) return; 2322 2323 if (eliminate_boxing()) { 2324 // Inline valueOf() methods now. 2325 inline_boxing_calls(igvn); 2326 2327 if (AlwaysIncrementalInline) { 2328 inline_incrementally(igvn); 2329 } 2330 2331 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2332 2333 if (failing()) return; 2334 } 2335 2336 // Now that all inlining is over, cut edge from root to loop 2337 // safepoints 2338 remove_root_to_sfpts_edges(igvn); 2339 2340 // Remove the speculative part of types and clean up the graph from 2341 // the extra CastPP nodes whose only purpose is to carry them. Do 2342 // that early so that optimizations are not disrupted by the extra 2343 // CastPP nodes. 2344 remove_speculative_types(igvn); 2345 2346 // No more new expensive nodes will be added to the list from here 2347 // so keep only the actual candidates for optimizations. 2348 cleanup_expensive_nodes(igvn); 2349 2350 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2351 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2352 initial_gvn()->replace_with(&igvn); 2353 for_igvn()->clear(); 2354 Unique_Node_List new_worklist(C->comp_arena()); 2355 { 2356 ResourceMark rm; 2357 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2358 } 2359 set_for_igvn(&new_worklist); 2360 igvn = PhaseIterGVN(initial_gvn()); 2361 igvn.optimize(); 2362 } 2363 2364 if (_value_type_nodes->size() > 0) { 2365 // Do this once all inlining is over to avoid getting inconsistent debug info 2366 process_value_types(igvn); 2367 } 2368 2369 // Perform escape analysis 2370 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2371 if (has_loops()) { 2372 // Cleanup graph (remove dead nodes). 2373 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2374 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2375 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2376 if (failing()) return; 2377 } 2378 ConnectionGraph::do_analysis(this, &igvn); 2379 2380 if (failing()) return; 2381 2382 // Optimize out fields loads from scalar replaceable allocations. 2383 igvn.optimize(); 2384 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2385 2386 if (failing()) return; 2387 2388 if (congraph() != NULL && macro_count() > 0) { 2389 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2390 PhaseMacroExpand mexp(igvn); 2391 mexp.eliminate_macro_nodes(); 2392 igvn.set_delay_transform(false); 2393 2394 igvn.optimize(); 2395 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2396 2397 if (failing()) return; 2398 } 2399 } 2400 2401 // Loop transforms on the ideal graph. Range Check Elimination, 2402 // peeling, unrolling, etc. 2403 2404 // Set loop opts counter 2405 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2406 { 2407 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2408 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2409 _loop_opts_cnt--; 2410 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2411 if (failing()) return; 2412 } 2413 // Loop opts pass if partial peeling occurred in previous pass 2414 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2415 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2416 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2417 _loop_opts_cnt--; 2418 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2419 if (failing()) return; 2420 } 2421 // Loop opts pass for loop-unrolling before CCP 2422 if(major_progress() && (_loop_opts_cnt > 0)) { 2423 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2424 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2425 _loop_opts_cnt--; 2426 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2427 } 2428 if (!failing()) { 2429 // Verify that last round of loop opts produced a valid graph 2430 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2431 PhaseIdealLoop::verify(igvn); 2432 } 2433 } 2434 if (failing()) return; 2435 2436 // Conditional Constant Propagation; 2437 PhaseCCP ccp( &igvn ); 2438 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2439 { 2440 TracePhase tp("ccp", &timers[_t_ccp]); 2441 ccp.do_transform(); 2442 } 2443 print_method(PHASE_CPP1, 2); 2444 2445 assert( true, "Break here to ccp.dump_old2new_map()"); 2446 2447 // Iterative Global Value Numbering, including ideal transforms 2448 { 2449 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2450 igvn = ccp; 2451 igvn.optimize(); 2452 } 2453 2454 print_method(PHASE_ITER_GVN2, 2); 2455 2456 if (failing()) return; 2457 2458 // Loop transforms on the ideal graph. Range Check Elimination, 2459 // peeling, unrolling, etc. 2460 if (!optimize_loops(igvn, LoopOptsDefault)) { 2461 return; 2462 } 2463 2464 #if INCLUDE_ZGC 2465 if (UseZGC) { 2466 ZBarrierSetC2::find_dominating_barriers(igvn); 2467 } 2468 #endif 2469 2470 if (failing()) return; 2471 2472 // Ensure that major progress is now clear 2473 C->clear_major_progress(); 2474 2475 { 2476 // Verify that all previous optimizations produced a valid graph 2477 // at least to this point, even if no loop optimizations were done. 2478 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2479 PhaseIdealLoop::verify(igvn); 2480 } 2481 2482 if (range_check_cast_count() > 0) { 2483 // No more loop optimizations. Remove all range check dependent CastIINodes. 2484 C->remove_range_check_casts(igvn); 2485 igvn.optimize(); 2486 } 2487 2488 #ifdef ASSERT 2489 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2490 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand); 2491 #endif 2492 2493 { 2494 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2495 PhaseMacroExpand mex(igvn); 2496 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2497 if (mex.expand_macro_nodes()) { 2498 assert(failing(), "must bail out w/ explicit message"); 2499 return; 2500 } 2501 } 2502 2503 { 2504 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2505 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2); 2506 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2507 if (bs->expand_barriers(this, igvn)) { 2508 assert(failing(), "must bail out w/ explicit message"); 2509 return; 2510 } 2511 } 2512 2513 if (opaque4_count() > 0) { 2514 C->remove_opaque4_nodes(igvn); 2515 igvn.optimize(); 2516 } 2517 2518 DEBUG_ONLY( _modified_nodes = NULL; ) 2519 } // (End scope of igvn; run destructor if necessary for asserts.) 2520 2521 process_print_inlining(); 2522 // A method with only infinite loops has no edges entering loops from root 2523 { 2524 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2525 if (final_graph_reshaping()) { 2526 assert(failing(), "must bail out w/ explicit message"); 2527 return; 2528 } 2529 } 2530 2531 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2532 } 2533 2534 //------------------------------Code_Gen--------------------------------------- 2535 // Given a graph, generate code for it 2536 void Compile::Code_Gen() { 2537 if (failing()) { 2538 return; 2539 } 2540 2541 // Perform instruction selection. You might think we could reclaim Matcher 2542 // memory PDQ, but actually the Matcher is used in generating spill code. 2543 // Internals of the Matcher (including some VectorSets) must remain live 2544 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2545 // set a bit in reclaimed memory. 2546 2547 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2548 // nodes. Mapping is only valid at the root of each matched subtree. 2549 NOT_PRODUCT( verify_graph_edges(); ) 2550 2551 Matcher matcher; 2552 _matcher = &matcher; 2553 { 2554 TracePhase tp("matcher", &timers[_t_matcher]); 2555 matcher.match(); 2556 } 2557 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2558 // nodes. Mapping is only valid at the root of each matched subtree. 2559 NOT_PRODUCT( verify_graph_edges(); ) 2560 2561 // If you have too many nodes, or if matching has failed, bail out 2562 check_node_count(0, "out of nodes matching instructions"); 2563 if (failing()) { 2564 return; 2565 } 2566 2567 print_method(PHASE_MATCHING, 2); 2568 2569 // Build a proper-looking CFG 2570 PhaseCFG cfg(node_arena(), root(), matcher); 2571 _cfg = &cfg; 2572 { 2573 TracePhase tp("scheduler", &timers[_t_scheduler]); 2574 bool success = cfg.do_global_code_motion(); 2575 if (!success) { 2576 return; 2577 } 2578 2579 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2580 NOT_PRODUCT( verify_graph_edges(); ) 2581 debug_only( cfg.verify(); ) 2582 } 2583 2584 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2585 _regalloc = ®alloc; 2586 { 2587 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2588 // Perform register allocation. After Chaitin, use-def chains are 2589 // no longer accurate (at spill code) and so must be ignored. 2590 // Node->LRG->reg mappings are still accurate. 2591 _regalloc->Register_Allocate(); 2592 2593 // Bail out if the allocator builds too many nodes 2594 if (failing()) { 2595 return; 2596 } 2597 } 2598 2599 // Prior to register allocation we kept empty basic blocks in case the 2600 // the allocator needed a place to spill. After register allocation we 2601 // are not adding any new instructions. If any basic block is empty, we 2602 // can now safely remove it. 2603 { 2604 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2605 cfg.remove_empty_blocks(); 2606 if (do_freq_based_layout()) { 2607 PhaseBlockLayout layout(cfg); 2608 } else { 2609 cfg.set_loop_alignment(); 2610 } 2611 cfg.fixup_flow(); 2612 } 2613 2614 // Apply peephole optimizations 2615 if( OptoPeephole ) { 2616 TracePhase tp("peephole", &timers[_t_peephole]); 2617 PhasePeephole peep( _regalloc, cfg); 2618 peep.do_transform(); 2619 } 2620 2621 // Do late expand if CPU requires this. 2622 if (Matcher::require_postalloc_expand) { 2623 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2624 cfg.postalloc_expand(_regalloc); 2625 } 2626 2627 // Convert Nodes to instruction bits in a buffer 2628 { 2629 TraceTime tp("output", &timers[_t_output], CITime); 2630 Output(); 2631 } 2632 2633 print_method(PHASE_FINAL_CODE); 2634 2635 // He's dead, Jim. 2636 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2637 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2638 } 2639 2640 2641 //------------------------------dump_asm--------------------------------------- 2642 // Dump formatted assembly 2643 #ifndef PRODUCT 2644 void Compile::dump_asm(int *pcs, uint pc_limit) { 2645 bool cut_short = false; 2646 tty->print_cr("#"); 2647 tty->print("# "); _tf->dump(); tty->cr(); 2648 tty->print_cr("#"); 2649 2650 // For all blocks 2651 int pc = 0x0; // Program counter 2652 char starts_bundle = ' '; 2653 _regalloc->dump_frame(); 2654 2655 Node *n = NULL; 2656 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2657 if (VMThread::should_terminate()) { 2658 cut_short = true; 2659 break; 2660 } 2661 Block* block = _cfg->get_block(i); 2662 if (block->is_connector() && !Verbose) { 2663 continue; 2664 } 2665 n = block->head(); 2666 if (pcs && n->_idx < pc_limit) { 2667 tty->print("%3.3x ", pcs[n->_idx]); 2668 } else { 2669 tty->print(" "); 2670 } 2671 block->dump_head(_cfg); 2672 if (block->is_connector()) { 2673 tty->print_cr(" # Empty connector block"); 2674 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2675 tty->print_cr(" # Block is sole successor of call"); 2676 } 2677 2678 // For all instructions 2679 Node *delay = NULL; 2680 for (uint j = 0; j < block->number_of_nodes(); j++) { 2681 if (VMThread::should_terminate()) { 2682 cut_short = true; 2683 break; 2684 } 2685 n = block->get_node(j); 2686 if (valid_bundle_info(n)) { 2687 Bundle* bundle = node_bundling(n); 2688 if (bundle->used_in_unconditional_delay()) { 2689 delay = n; 2690 continue; 2691 } 2692 if (bundle->starts_bundle()) { 2693 starts_bundle = '+'; 2694 } 2695 } 2696 2697 if (WizardMode) { 2698 n->dump(); 2699 } 2700 2701 if( !n->is_Region() && // Dont print in the Assembly 2702 !n->is_Phi() && // a few noisely useless nodes 2703 !n->is_Proj() && 2704 !n->is_MachTemp() && 2705 !n->is_SafePointScalarObject() && 2706 !n->is_Catch() && // Would be nice to print exception table targets 2707 !n->is_MergeMem() && // Not very interesting 2708 !n->is_top() && // Debug info table constants 2709 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2710 ) { 2711 if (pcs && n->_idx < pc_limit) 2712 tty->print("%3.3x", pcs[n->_idx]); 2713 else 2714 tty->print(" "); 2715 tty->print(" %c ", starts_bundle); 2716 starts_bundle = ' '; 2717 tty->print("\t"); 2718 n->format(_regalloc, tty); 2719 tty->cr(); 2720 } 2721 2722 // If we have an instruction with a delay slot, and have seen a delay, 2723 // then back up and print it 2724 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2725 assert(delay != NULL, "no unconditional delay instruction"); 2726 if (WizardMode) delay->dump(); 2727 2728 if (node_bundling(delay)->starts_bundle()) 2729 starts_bundle = '+'; 2730 if (pcs && n->_idx < pc_limit) 2731 tty->print("%3.3x", pcs[n->_idx]); 2732 else 2733 tty->print(" "); 2734 tty->print(" %c ", starts_bundle); 2735 starts_bundle = ' '; 2736 tty->print("\t"); 2737 delay->format(_regalloc, tty); 2738 tty->cr(); 2739 delay = NULL; 2740 } 2741 2742 // Dump the exception table as well 2743 if( n->is_Catch() && (Verbose || WizardMode) ) { 2744 // Print the exception table for this offset 2745 _handler_table.print_subtable_for(pc); 2746 } 2747 } 2748 2749 if (pcs && n->_idx < pc_limit) 2750 tty->print_cr("%3.3x", pcs[n->_idx]); 2751 else 2752 tty->cr(); 2753 2754 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2755 2756 } // End of per-block dump 2757 tty->cr(); 2758 2759 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2760 } 2761 #endif 2762 2763 //------------------------------Final_Reshape_Counts--------------------------- 2764 // This class defines counters to help identify when a method 2765 // may/must be executed using hardware with only 24-bit precision. 2766 struct Final_Reshape_Counts : public StackObj { 2767 int _call_count; // count non-inlined 'common' calls 2768 int _float_count; // count float ops requiring 24-bit precision 2769 int _double_count; // count double ops requiring more precision 2770 int _java_call_count; // count non-inlined 'java' calls 2771 int _inner_loop_count; // count loops which need alignment 2772 VectorSet _visited; // Visitation flags 2773 Node_List _tests; // Set of IfNodes & PCTableNodes 2774 2775 Final_Reshape_Counts() : 2776 _call_count(0), _float_count(0), _double_count(0), 2777 _java_call_count(0), _inner_loop_count(0), 2778 _visited( Thread::current()->resource_area() ) { } 2779 2780 void inc_call_count () { _call_count ++; } 2781 void inc_float_count () { _float_count ++; } 2782 void inc_double_count() { _double_count++; } 2783 void inc_java_call_count() { _java_call_count++; } 2784 void inc_inner_loop_count() { _inner_loop_count++; } 2785 2786 int get_call_count () const { return _call_count ; } 2787 int get_float_count () const { return _float_count ; } 2788 int get_double_count() const { return _double_count; } 2789 int get_java_call_count() const { return _java_call_count; } 2790 int get_inner_loop_count() const { return _inner_loop_count; } 2791 }; 2792 2793 #ifdef ASSERT 2794 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2795 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2796 // Make sure the offset goes inside the instance layout. 2797 return k->contains_field_offset(tp->offset()); 2798 // Note that OffsetBot and OffsetTop are very negative. 2799 } 2800 #endif 2801 2802 // Eliminate trivially redundant StoreCMs and accumulate their 2803 // precedence edges. 2804 void Compile::eliminate_redundant_card_marks(Node* n) { 2805 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2806 if (n->in(MemNode::Address)->outcnt() > 1) { 2807 // There are multiple users of the same address so it might be 2808 // possible to eliminate some of the StoreCMs 2809 Node* mem = n->in(MemNode::Memory); 2810 Node* adr = n->in(MemNode::Address); 2811 Node* val = n->in(MemNode::ValueIn); 2812 Node* prev = n; 2813 bool done = false; 2814 // Walk the chain of StoreCMs eliminating ones that match. As 2815 // long as it's a chain of single users then the optimization is 2816 // safe. Eliminating partially redundant StoreCMs would require 2817 // cloning copies down the other paths. 2818 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2819 if (adr == mem->in(MemNode::Address) && 2820 val == mem->in(MemNode::ValueIn)) { 2821 // redundant StoreCM 2822 if (mem->req() > MemNode::OopStore) { 2823 // Hasn't been processed by this code yet. 2824 n->add_prec(mem->in(MemNode::OopStore)); 2825 } else { 2826 // Already converted to precedence edge 2827 for (uint i = mem->req(); i < mem->len(); i++) { 2828 // Accumulate any precedence edges 2829 if (mem->in(i) != NULL) { 2830 n->add_prec(mem->in(i)); 2831 } 2832 } 2833 // Everything above this point has been processed. 2834 done = true; 2835 } 2836 // Eliminate the previous StoreCM 2837 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2838 assert(mem->outcnt() == 0, "should be dead"); 2839 mem->disconnect_inputs(NULL, this); 2840 } else { 2841 prev = mem; 2842 } 2843 mem = prev->in(MemNode::Memory); 2844 } 2845 } 2846 } 2847 2848 2849 //------------------------------final_graph_reshaping_impl---------------------- 2850 // Implement items 1-5 from final_graph_reshaping below. 2851 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2852 2853 if ( n->outcnt() == 0 ) return; // dead node 2854 uint nop = n->Opcode(); 2855 2856 // Check for 2-input instruction with "last use" on right input. 2857 // Swap to left input. Implements item (2). 2858 if( n->req() == 3 && // two-input instruction 2859 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2860 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2861 n->in(2)->outcnt() == 1 &&// right use IS a last use 2862 !n->in(2)->is_Con() ) { // right use is not a constant 2863 // Check for commutative opcode 2864 switch( nop ) { 2865 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2866 case Op_MaxI: case Op_MinI: 2867 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2868 case Op_AndL: case Op_XorL: case Op_OrL: 2869 case Op_AndI: case Op_XorI: case Op_OrI: { 2870 // Move "last use" input to left by swapping inputs 2871 n->swap_edges(1, 2); 2872 break; 2873 } 2874 default: 2875 break; 2876 } 2877 } 2878 2879 #ifdef ASSERT 2880 if( n->is_Mem() ) { 2881 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2882 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2883 // oop will be recorded in oop map if load crosses safepoint 2884 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2885 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2886 "raw memory operations should have control edge"); 2887 } 2888 if (n->is_MemBar()) { 2889 MemBarNode* mb = n->as_MemBar(); 2890 if (mb->trailing_store() || mb->trailing_load_store()) { 2891 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2892 Node* mem = mb->in(MemBarNode::Precedent); 2893 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2894 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2895 } else if (mb->leading()) { 2896 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2897 } 2898 } 2899 #endif 2900 // Count FPU ops and common calls, implements item (3) 2901 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2902 if (!gc_handled) { 2903 final_graph_reshaping_main_switch(n, frc, nop); 2904 } 2905 2906 // Collect CFG split points 2907 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2908 frc._tests.push(n); 2909 } 2910 } 2911 2912 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2913 switch( nop ) { 2914 // Count all float operations that may use FPU 2915 case Op_AddF: 2916 case Op_SubF: 2917 case Op_MulF: 2918 case Op_DivF: 2919 case Op_NegF: 2920 case Op_ModF: 2921 case Op_ConvI2F: 2922 case Op_ConF: 2923 case Op_CmpF: 2924 case Op_CmpF3: 2925 // case Op_ConvL2F: // longs are split into 32-bit halves 2926 frc.inc_float_count(); 2927 break; 2928 2929 case Op_ConvF2D: 2930 case Op_ConvD2F: 2931 frc.inc_float_count(); 2932 frc.inc_double_count(); 2933 break; 2934 2935 // Count all double operations that may use FPU 2936 case Op_AddD: 2937 case Op_SubD: 2938 case Op_MulD: 2939 case Op_DivD: 2940 case Op_NegD: 2941 case Op_ModD: 2942 case Op_ConvI2D: 2943 case Op_ConvD2I: 2944 // case Op_ConvL2D: // handled by leaf call 2945 // case Op_ConvD2L: // handled by leaf call 2946 case Op_ConD: 2947 case Op_CmpD: 2948 case Op_CmpD3: 2949 frc.inc_double_count(); 2950 break; 2951 case Op_Opaque1: // Remove Opaque Nodes before matching 2952 case Op_Opaque2: // Remove Opaque Nodes before matching 2953 case Op_Opaque3: 2954 n->subsume_by(n->in(1), this); 2955 break; 2956 case Op_CallStaticJava: 2957 case Op_CallJava: 2958 case Op_CallDynamicJava: 2959 frc.inc_java_call_count(); // Count java call site; 2960 case Op_CallRuntime: 2961 case Op_CallLeaf: 2962 case Op_CallLeafNoFP: { 2963 assert (n->is_Call(), ""); 2964 CallNode *call = n->as_Call(); 2965 // Count call sites where the FP mode bit would have to be flipped. 2966 // Do not count uncommon runtime calls: 2967 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2968 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2969 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2970 frc.inc_call_count(); // Count the call site 2971 } else { // See if uncommon argument is shared 2972 Node *n = call->in(TypeFunc::Parms); 2973 int nop = n->Opcode(); 2974 // Clone shared simple arguments to uncommon calls, item (1). 2975 if (n->outcnt() > 1 && 2976 !n->is_Proj() && 2977 nop != Op_CreateEx && 2978 nop != Op_CheckCastPP && 2979 nop != Op_DecodeN && 2980 nop != Op_DecodeNKlass && 2981 !n->is_Mem() && 2982 !n->is_Phi()) { 2983 Node *x = n->clone(); 2984 call->set_req(TypeFunc::Parms, x); 2985 } 2986 } 2987 break; 2988 } 2989 2990 case Op_StoreD: 2991 case Op_LoadD: 2992 case Op_LoadD_unaligned: 2993 frc.inc_double_count(); 2994 goto handle_mem; 2995 case Op_StoreF: 2996 case Op_LoadF: 2997 frc.inc_float_count(); 2998 goto handle_mem; 2999 3000 case Op_StoreCM: 3001 { 3002 // Convert OopStore dependence into precedence edge 3003 Node* prec = n->in(MemNode::OopStore); 3004 n->del_req(MemNode::OopStore); 3005 n->add_prec(prec); 3006 eliminate_redundant_card_marks(n); 3007 } 3008 3009 // fall through 3010 3011 case Op_StoreB: 3012 case Op_StoreC: 3013 case Op_StorePConditional: 3014 case Op_StoreI: 3015 case Op_StoreL: 3016 case Op_StoreIConditional: 3017 case Op_StoreLConditional: 3018 case Op_CompareAndSwapB: 3019 case Op_CompareAndSwapS: 3020 case Op_CompareAndSwapI: 3021 case Op_CompareAndSwapL: 3022 case Op_CompareAndSwapP: 3023 case Op_CompareAndSwapN: 3024 case Op_WeakCompareAndSwapB: 3025 case Op_WeakCompareAndSwapS: 3026 case Op_WeakCompareAndSwapI: 3027 case Op_WeakCompareAndSwapL: 3028 case Op_WeakCompareAndSwapP: 3029 case Op_WeakCompareAndSwapN: 3030 case Op_CompareAndExchangeB: 3031 case Op_CompareAndExchangeS: 3032 case Op_CompareAndExchangeI: 3033 case Op_CompareAndExchangeL: 3034 case Op_CompareAndExchangeP: 3035 case Op_CompareAndExchangeN: 3036 case Op_GetAndAddS: 3037 case Op_GetAndAddB: 3038 case Op_GetAndAddI: 3039 case Op_GetAndAddL: 3040 case Op_GetAndSetS: 3041 case Op_GetAndSetB: 3042 case Op_GetAndSetI: 3043 case Op_GetAndSetL: 3044 case Op_GetAndSetP: 3045 case Op_GetAndSetN: 3046 case Op_StoreP: 3047 case Op_StoreN: 3048 case Op_StoreNKlass: 3049 case Op_LoadB: 3050 case Op_LoadUB: 3051 case Op_LoadUS: 3052 case Op_LoadI: 3053 case Op_LoadKlass: 3054 case Op_LoadNKlass: 3055 case Op_LoadL: 3056 case Op_LoadL_unaligned: 3057 case Op_LoadPLocked: 3058 case Op_LoadP: 3059 case Op_LoadN: 3060 case Op_LoadRange: 3061 case Op_LoadS: { 3062 handle_mem: 3063 #ifdef ASSERT 3064 if( VerifyOptoOopOffsets ) { 3065 MemNode* mem = n->as_Mem(); 3066 // Check to see if address types have grounded out somehow. 3067 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3068 assert( !tp || oop_offset_is_sane(tp), "" ); 3069 } 3070 #endif 3071 break; 3072 } 3073 3074 case Op_AddP: { // Assert sane base pointers 3075 Node *addp = n->in(AddPNode::Address); 3076 assert( !addp->is_AddP() || 3077 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3078 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3079 "Base pointers must match (addp %u)", addp->_idx ); 3080 #ifdef _LP64 3081 if ((UseCompressedOops || UseCompressedClassPointers) && 3082 addp->Opcode() == Op_ConP && 3083 addp == n->in(AddPNode::Base) && 3084 n->in(AddPNode::Offset)->is_Con()) { 3085 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3086 // on the platform and on the compressed oops mode. 3087 // Use addressing with narrow klass to load with offset on x86. 3088 // Some platforms can use the constant pool to load ConP. 3089 // Do this transformation here since IGVN will convert ConN back to ConP. 3090 const Type* t = addp->bottom_type(); 3091 bool is_oop = t->isa_oopptr() != NULL; 3092 bool is_klass = t->isa_klassptr() != NULL; 3093 3094 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3095 (is_klass && Matcher::const_klass_prefer_decode())) { 3096 Node* nn = NULL; 3097 3098 int op = is_oop ? Op_ConN : Op_ConNKlass; 3099 3100 // Look for existing ConN node of the same exact type. 3101 Node* r = root(); 3102 uint cnt = r->outcnt(); 3103 for (uint i = 0; i < cnt; i++) { 3104 Node* m = r->raw_out(i); 3105 if (m!= NULL && m->Opcode() == op && 3106 m->bottom_type()->make_ptr() == t) { 3107 nn = m; 3108 break; 3109 } 3110 } 3111 if (nn != NULL) { 3112 // Decode a narrow oop to match address 3113 // [R12 + narrow_oop_reg<<3 + offset] 3114 if (is_oop) { 3115 nn = new DecodeNNode(nn, t); 3116 } else { 3117 nn = new DecodeNKlassNode(nn, t); 3118 } 3119 // Check for succeeding AddP which uses the same Base. 3120 // Otherwise we will run into the assertion above when visiting that guy. 3121 for (uint i = 0; i < n->outcnt(); ++i) { 3122 Node *out_i = n->raw_out(i); 3123 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3124 out_i->set_req(AddPNode::Base, nn); 3125 #ifdef ASSERT 3126 for (uint j = 0; j < out_i->outcnt(); ++j) { 3127 Node *out_j = out_i->raw_out(j); 3128 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3129 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3130 } 3131 #endif 3132 } 3133 } 3134 n->set_req(AddPNode::Base, nn); 3135 n->set_req(AddPNode::Address, nn); 3136 if (addp->outcnt() == 0) { 3137 addp->disconnect_inputs(NULL, this); 3138 } 3139 } 3140 } 3141 } 3142 #endif 3143 // platform dependent reshaping of the address expression 3144 reshape_address(n->as_AddP()); 3145 break; 3146 } 3147 3148 case Op_CastPP: { 3149 // Remove CastPP nodes to gain more freedom during scheduling but 3150 // keep the dependency they encode as control or precedence edges 3151 // (if control is set already) on memory operations. Some CastPP 3152 // nodes don't have a control (don't carry a dependency): skip 3153 // those. 3154 if (n->in(0) != NULL) { 3155 ResourceMark rm; 3156 Unique_Node_List wq; 3157 wq.push(n); 3158 for (uint next = 0; next < wq.size(); ++next) { 3159 Node *m = wq.at(next); 3160 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3161 Node* use = m->fast_out(i); 3162 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) { 3163 use->ensure_control_or_add_prec(n->in(0)); 3164 } else { 3165 switch(use->Opcode()) { 3166 case Op_AddP: 3167 case Op_DecodeN: 3168 case Op_DecodeNKlass: 3169 case Op_CheckCastPP: 3170 case Op_CastPP: 3171 wq.push(use); 3172 break; 3173 } 3174 } 3175 } 3176 } 3177 } 3178 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3179 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3180 Node* in1 = n->in(1); 3181 const Type* t = n->bottom_type(); 3182 Node* new_in1 = in1->clone(); 3183 new_in1->as_DecodeN()->set_type(t); 3184 3185 if (!Matcher::narrow_oop_use_complex_address()) { 3186 // 3187 // x86, ARM and friends can handle 2 adds in addressing mode 3188 // and Matcher can fold a DecodeN node into address by using 3189 // a narrow oop directly and do implicit NULL check in address: 3190 // 3191 // [R12 + narrow_oop_reg<<3 + offset] 3192 // NullCheck narrow_oop_reg 3193 // 3194 // On other platforms (Sparc) we have to keep new DecodeN node and 3195 // use it to do implicit NULL check in address: 3196 // 3197 // decode_not_null narrow_oop_reg, base_reg 3198 // [base_reg + offset] 3199 // NullCheck base_reg 3200 // 3201 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3202 // to keep the information to which NULL check the new DecodeN node 3203 // corresponds to use it as value in implicit_null_check(). 3204 // 3205 new_in1->set_req(0, n->in(0)); 3206 } 3207 3208 n->subsume_by(new_in1, this); 3209 if (in1->outcnt() == 0) { 3210 in1->disconnect_inputs(NULL, this); 3211 } 3212 } else { 3213 n->subsume_by(n->in(1), this); 3214 if (n->outcnt() == 0) { 3215 n->disconnect_inputs(NULL, this); 3216 } 3217 } 3218 break; 3219 } 3220 #ifdef _LP64 3221 case Op_CmpP: 3222 // Do this transformation here to preserve CmpPNode::sub() and 3223 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3224 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3225 Node* in1 = n->in(1); 3226 Node* in2 = n->in(2); 3227 if (!in1->is_DecodeNarrowPtr()) { 3228 in2 = in1; 3229 in1 = n->in(2); 3230 } 3231 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3232 3233 Node* new_in2 = NULL; 3234 if (in2->is_DecodeNarrowPtr()) { 3235 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3236 new_in2 = in2->in(1); 3237 } else if (in2->Opcode() == Op_ConP) { 3238 const Type* t = in2->bottom_type(); 3239 if (t == TypePtr::NULL_PTR) { 3240 assert(in1->is_DecodeN(), "compare klass to null?"); 3241 // Don't convert CmpP null check into CmpN if compressed 3242 // oops implicit null check is not generated. 3243 // This will allow to generate normal oop implicit null check. 3244 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3245 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3246 // 3247 // This transformation together with CastPP transformation above 3248 // will generated code for implicit NULL checks for compressed oops. 3249 // 3250 // The original code after Optimize() 3251 // 3252 // LoadN memory, narrow_oop_reg 3253 // decode narrow_oop_reg, base_reg 3254 // CmpP base_reg, NULL 3255 // CastPP base_reg // NotNull 3256 // Load [base_reg + offset], val_reg 3257 // 3258 // after these transformations will be 3259 // 3260 // LoadN memory, narrow_oop_reg 3261 // CmpN narrow_oop_reg, NULL 3262 // decode_not_null narrow_oop_reg, base_reg 3263 // Load [base_reg + offset], val_reg 3264 // 3265 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3266 // since narrow oops can be used in debug info now (see the code in 3267 // final_graph_reshaping_walk()). 3268 // 3269 // At the end the code will be matched to 3270 // on x86: 3271 // 3272 // Load_narrow_oop memory, narrow_oop_reg 3273 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3274 // NullCheck narrow_oop_reg 3275 // 3276 // and on sparc: 3277 // 3278 // Load_narrow_oop memory, narrow_oop_reg 3279 // decode_not_null narrow_oop_reg, base_reg 3280 // Load [base_reg + offset], val_reg 3281 // NullCheck base_reg 3282 // 3283 } else if (t->isa_oopptr()) { 3284 new_in2 = ConNode::make(t->make_narrowoop()); 3285 } else if (t->isa_klassptr()) { 3286 new_in2 = ConNode::make(t->make_narrowklass()); 3287 } 3288 } 3289 if (new_in2 != NULL) { 3290 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3291 n->subsume_by(cmpN, this); 3292 if (in1->outcnt() == 0) { 3293 in1->disconnect_inputs(NULL, this); 3294 } 3295 if (in2->outcnt() == 0) { 3296 in2->disconnect_inputs(NULL, this); 3297 } 3298 } 3299 } 3300 break; 3301 3302 case Op_DecodeN: 3303 case Op_DecodeNKlass: 3304 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3305 // DecodeN could be pinned when it can't be fold into 3306 // an address expression, see the code for Op_CastPP above. 3307 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3308 break; 3309 3310 case Op_EncodeP: 3311 case Op_EncodePKlass: { 3312 Node* in1 = n->in(1); 3313 if (in1->is_DecodeNarrowPtr()) { 3314 n->subsume_by(in1->in(1), this); 3315 } else if (in1->Opcode() == Op_ConP) { 3316 const Type* t = in1->bottom_type(); 3317 if (t == TypePtr::NULL_PTR) { 3318 assert(t->isa_oopptr(), "null klass?"); 3319 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3320 } else if (t->isa_oopptr()) { 3321 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3322 } else if (t->isa_klassptr()) { 3323 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3324 } 3325 } 3326 if (in1->outcnt() == 0) { 3327 in1->disconnect_inputs(NULL, this); 3328 } 3329 break; 3330 } 3331 3332 case Op_Proj: { 3333 if (OptimizeStringConcat) { 3334 ProjNode* p = n->as_Proj(); 3335 if (p->_is_io_use) { 3336 // Separate projections were used for the exception path which 3337 // are normally removed by a late inline. If it wasn't inlined 3338 // then they will hang around and should just be replaced with 3339 // the original one. 3340 Node* proj = NULL; 3341 // Replace with just one 3342 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3343 Node *use = i.get(); 3344 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3345 proj = use; 3346 break; 3347 } 3348 } 3349 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3350 if (proj != NULL) { 3351 p->subsume_by(proj, this); 3352 } 3353 } 3354 } 3355 break; 3356 } 3357 3358 case Op_Phi: 3359 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3360 // The EncodeP optimization may create Phi with the same edges 3361 // for all paths. It is not handled well by Register Allocator. 3362 Node* unique_in = n->in(1); 3363 assert(unique_in != NULL, ""); 3364 uint cnt = n->req(); 3365 for (uint i = 2; i < cnt; i++) { 3366 Node* m = n->in(i); 3367 assert(m != NULL, ""); 3368 if (unique_in != m) 3369 unique_in = NULL; 3370 } 3371 if (unique_in != NULL) { 3372 n->subsume_by(unique_in, this); 3373 } 3374 } 3375 break; 3376 3377 #endif 3378 3379 #ifdef ASSERT 3380 case Op_CastII: 3381 // Verify that all range check dependent CastII nodes were removed. 3382 if (n->isa_CastII()->has_range_check()) { 3383 n->dump(3); 3384 assert(false, "Range check dependent CastII node was not removed"); 3385 } 3386 break; 3387 #endif 3388 3389 case Op_ModI: 3390 if (UseDivMod) { 3391 // Check if a%b and a/b both exist 3392 Node* d = n->find_similar(Op_DivI); 3393 if (d) { 3394 // Replace them with a fused divmod if supported 3395 if (Matcher::has_match_rule(Op_DivModI)) { 3396 DivModINode* divmod = DivModINode::make(n); 3397 d->subsume_by(divmod->div_proj(), this); 3398 n->subsume_by(divmod->mod_proj(), this); 3399 } else { 3400 // replace a%b with a-((a/b)*b) 3401 Node* mult = new MulINode(d, d->in(2)); 3402 Node* sub = new SubINode(d->in(1), mult); 3403 n->subsume_by(sub, this); 3404 } 3405 } 3406 } 3407 break; 3408 3409 case Op_ModL: 3410 if (UseDivMod) { 3411 // Check if a%b and a/b both exist 3412 Node* d = n->find_similar(Op_DivL); 3413 if (d) { 3414 // Replace them with a fused divmod if supported 3415 if (Matcher::has_match_rule(Op_DivModL)) { 3416 DivModLNode* divmod = DivModLNode::make(n); 3417 d->subsume_by(divmod->div_proj(), this); 3418 n->subsume_by(divmod->mod_proj(), this); 3419 } else { 3420 // replace a%b with a-((a/b)*b) 3421 Node* mult = new MulLNode(d, d->in(2)); 3422 Node* sub = new SubLNode(d->in(1), mult); 3423 n->subsume_by(sub, this); 3424 } 3425 } 3426 } 3427 break; 3428 3429 case Op_LoadVector: 3430 case Op_StoreVector: 3431 break; 3432 3433 case Op_AddReductionVI: 3434 case Op_AddReductionVL: 3435 case Op_AddReductionVF: 3436 case Op_AddReductionVD: 3437 case Op_MulReductionVI: 3438 case Op_MulReductionVL: 3439 case Op_MulReductionVF: 3440 case Op_MulReductionVD: 3441 break; 3442 3443 case Op_PackB: 3444 case Op_PackS: 3445 case Op_PackI: 3446 case Op_PackF: 3447 case Op_PackL: 3448 case Op_PackD: 3449 if (n->req()-1 > 2) { 3450 // Replace many operand PackNodes with a binary tree for matching 3451 PackNode* p = (PackNode*) n; 3452 Node* btp = p->binary_tree_pack(1, n->req()); 3453 n->subsume_by(btp, this); 3454 } 3455 break; 3456 case Op_Loop: 3457 case Op_CountedLoop: 3458 case Op_OuterStripMinedLoop: 3459 if (n->as_Loop()->is_inner_loop()) { 3460 frc.inc_inner_loop_count(); 3461 } 3462 n->as_Loop()->verify_strip_mined(0); 3463 break; 3464 case Op_LShiftI: 3465 case Op_RShiftI: 3466 case Op_URShiftI: 3467 case Op_LShiftL: 3468 case Op_RShiftL: 3469 case Op_URShiftL: 3470 if (Matcher::need_masked_shift_count) { 3471 // The cpu's shift instructions don't restrict the count to the 3472 // lower 5/6 bits. We need to do the masking ourselves. 3473 Node* in2 = n->in(2); 3474 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3475 const TypeInt* t = in2->find_int_type(); 3476 if (t != NULL && t->is_con()) { 3477 juint shift = t->get_con(); 3478 if (shift > mask) { // Unsigned cmp 3479 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3480 } 3481 } else { 3482 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3483 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3484 n->set_req(2, shift); 3485 } 3486 } 3487 if (in2->outcnt() == 0) { // Remove dead node 3488 in2->disconnect_inputs(NULL, this); 3489 } 3490 } 3491 break; 3492 case Op_MemBarStoreStore: 3493 case Op_MemBarRelease: 3494 // Break the link with AllocateNode: it is no longer useful and 3495 // confuses register allocation. 3496 if (n->req() > MemBarNode::Precedent) { 3497 n->set_req(MemBarNode::Precedent, top()); 3498 } 3499 break; 3500 case Op_MemBarAcquire: { 3501 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3502 // At parse time, the trailing MemBarAcquire for a volatile load 3503 // is created with an edge to the load. After optimizations, 3504 // that input may be a chain of Phis. If those phis have no 3505 // other use, then the MemBarAcquire keeps them alive and 3506 // register allocation can be confused. 3507 ResourceMark rm; 3508 Unique_Node_List wq; 3509 wq.push(n->in(MemBarNode::Precedent)); 3510 n->set_req(MemBarNode::Precedent, top()); 3511 while (wq.size() > 0) { 3512 Node* m = wq.pop(); 3513 if (m->outcnt() == 0) { 3514 for (uint j = 0; j < m->req(); j++) { 3515 Node* in = m->in(j); 3516 if (in != NULL) { 3517 wq.push(in); 3518 } 3519 } 3520 m->disconnect_inputs(NULL, this); 3521 } 3522 } 3523 } 3524 break; 3525 } 3526 case Op_RangeCheck: { 3527 RangeCheckNode* rc = n->as_RangeCheck(); 3528 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3529 n->subsume_by(iff, this); 3530 frc._tests.push(iff); 3531 break; 3532 } 3533 case Op_ConvI2L: { 3534 if (!Matcher::convi2l_type_required) { 3535 // Code generation on some platforms doesn't need accurate 3536 // ConvI2L types. Widening the type can help remove redundant 3537 // address computations. 3538 n->as_Type()->set_type(TypeLong::INT); 3539 ResourceMark rm; 3540 Node_List wq; 3541 wq.push(n); 3542 for (uint next = 0; next < wq.size(); next++) { 3543 Node *m = wq.at(next); 3544 3545 for(;;) { 3546 // Loop over all nodes with identical inputs edges as m 3547 Node* k = m->find_similar(m->Opcode()); 3548 if (k == NULL) { 3549 break; 3550 } 3551 // Push their uses so we get a chance to remove node made 3552 // redundant 3553 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3554 Node* u = k->fast_out(i); 3555 assert(!wq.contains(u), "shouldn't process one node several times"); 3556 if (u->Opcode() == Op_LShiftL || 3557 u->Opcode() == Op_AddL || 3558 u->Opcode() == Op_SubL || 3559 u->Opcode() == Op_AddP) { 3560 wq.push(u); 3561 } 3562 } 3563 // Replace all nodes with identical edges as m with m 3564 k->subsume_by(m, this); 3565 } 3566 } 3567 } 3568 break; 3569 } 3570 case Op_CmpUL: { 3571 if (!Matcher::has_match_rule(Op_CmpUL)) { 3572 // No support for unsigned long comparisons 3573 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3574 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3575 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3576 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3577 Node* andl = new AndLNode(orl, remove_sign_mask); 3578 Node* cmp = new CmpLNode(andl, n->in(2)); 3579 n->subsume_by(cmp, this); 3580 } 3581 break; 3582 } 3583 #ifdef ASSERT 3584 case Op_ValueTypePtr: 3585 case Op_ValueType: { 3586 n->dump(-1); 3587 assert(false, "value type node was not removed"); 3588 break; 3589 } 3590 #endif 3591 default: 3592 assert(!n->is_Call(), ""); 3593 assert(!n->is_Mem(), ""); 3594 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3595 break; 3596 } 3597 } 3598 3599 //------------------------------final_graph_reshaping_walk--------------------- 3600 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3601 // requires that the walk visits a node's inputs before visiting the node. 3602 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3603 ResourceArea *area = Thread::current()->resource_area(); 3604 Unique_Node_List sfpt(area); 3605 3606 frc._visited.set(root->_idx); // first, mark node as visited 3607 uint cnt = root->req(); 3608 Node *n = root; 3609 uint i = 0; 3610 while (true) { 3611 if (i < cnt) { 3612 // Place all non-visited non-null inputs onto stack 3613 Node* m = n->in(i); 3614 ++i; 3615 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3616 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3617 // compute worst case interpreter size in case of a deoptimization 3618 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3619 3620 sfpt.push(m); 3621 } 3622 cnt = m->req(); 3623 nstack.push(n, i); // put on stack parent and next input's index 3624 n = m; 3625 i = 0; 3626 } 3627 } else { 3628 // Now do post-visit work 3629 final_graph_reshaping_impl( n, frc ); 3630 if (nstack.is_empty()) 3631 break; // finished 3632 n = nstack.node(); // Get node from stack 3633 cnt = n->req(); 3634 i = nstack.index(); 3635 nstack.pop(); // Shift to the next node on stack 3636 } 3637 } 3638 3639 // Skip next transformation if compressed oops are not used. 3640 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3641 (!UseCompressedOops && !UseCompressedClassPointers)) 3642 return; 3643 3644 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3645 // It could be done for an uncommon traps or any safepoints/calls 3646 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3647 while (sfpt.size() > 0) { 3648 n = sfpt.pop(); 3649 JVMState *jvms = n->as_SafePoint()->jvms(); 3650 assert(jvms != NULL, "sanity"); 3651 int start = jvms->debug_start(); 3652 int end = n->req(); 3653 bool is_uncommon = (n->is_CallStaticJava() && 3654 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3655 for (int j = start; j < end; j++) { 3656 Node* in = n->in(j); 3657 if (in->is_DecodeNarrowPtr()) { 3658 bool safe_to_skip = true; 3659 if (!is_uncommon ) { 3660 // Is it safe to skip? 3661 for (uint i = 0; i < in->outcnt(); i++) { 3662 Node* u = in->raw_out(i); 3663 if (!u->is_SafePoint() || 3664 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3665 safe_to_skip = false; 3666 } 3667 } 3668 } 3669 if (safe_to_skip) { 3670 n->set_req(j, in->in(1)); 3671 } 3672 if (in->outcnt() == 0) { 3673 in->disconnect_inputs(NULL, this); 3674 } 3675 } 3676 } 3677 } 3678 } 3679 3680 //------------------------------final_graph_reshaping-------------------------- 3681 // Final Graph Reshaping. 3682 // 3683 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3684 // and not commoned up and forced early. Must come after regular 3685 // optimizations to avoid GVN undoing the cloning. Clone constant 3686 // inputs to Loop Phis; these will be split by the allocator anyways. 3687 // Remove Opaque nodes. 3688 // (2) Move last-uses by commutative operations to the left input to encourage 3689 // Intel update-in-place two-address operations and better register usage 3690 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3691 // calls canonicalizing them back. 3692 // (3) Count the number of double-precision FP ops, single-precision FP ops 3693 // and call sites. On Intel, we can get correct rounding either by 3694 // forcing singles to memory (requires extra stores and loads after each 3695 // FP bytecode) or we can set a rounding mode bit (requires setting and 3696 // clearing the mode bit around call sites). The mode bit is only used 3697 // if the relative frequency of single FP ops to calls is low enough. 3698 // This is a key transform for SPEC mpeg_audio. 3699 // (4) Detect infinite loops; blobs of code reachable from above but not 3700 // below. Several of the Code_Gen algorithms fail on such code shapes, 3701 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3702 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3703 // Detection is by looking for IfNodes where only 1 projection is 3704 // reachable from below or CatchNodes missing some targets. 3705 // (5) Assert for insane oop offsets in debug mode. 3706 3707 bool Compile::final_graph_reshaping() { 3708 // an infinite loop may have been eliminated by the optimizer, 3709 // in which case the graph will be empty. 3710 if (root()->req() == 1) { 3711 record_method_not_compilable("trivial infinite loop"); 3712 return true; 3713 } 3714 3715 // Expensive nodes have their control input set to prevent the GVN 3716 // from freely commoning them. There's no GVN beyond this point so 3717 // no need to keep the control input. We want the expensive nodes to 3718 // be freely moved to the least frequent code path by gcm. 3719 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3720 for (int i = 0; i < expensive_count(); i++) { 3721 _expensive_nodes->at(i)->set_req(0, NULL); 3722 } 3723 3724 Final_Reshape_Counts frc; 3725 3726 // Visit everybody reachable! 3727 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3728 Node_Stack nstack(live_nodes() >> 1); 3729 final_graph_reshaping_walk(nstack, root(), frc); 3730 3731 // Check for unreachable (from below) code (i.e., infinite loops). 3732 for( uint i = 0; i < frc._tests.size(); i++ ) { 3733 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3734 // Get number of CFG targets. 3735 // Note that PCTables include exception targets after calls. 3736 uint required_outcnt = n->required_outcnt(); 3737 if (n->outcnt() != required_outcnt) { 3738 // Check for a few special cases. Rethrow Nodes never take the 3739 // 'fall-thru' path, so expected kids is 1 less. 3740 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3741 if (n->in(0)->in(0)->is_Call()) { 3742 CallNode *call = n->in(0)->in(0)->as_Call(); 3743 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3744 required_outcnt--; // Rethrow always has 1 less kid 3745 } else if (call->req() > TypeFunc::Parms && 3746 call->is_CallDynamicJava()) { 3747 // Check for null receiver. In such case, the optimizer has 3748 // detected that the virtual call will always result in a null 3749 // pointer exception. The fall-through projection of this CatchNode 3750 // will not be populated. 3751 Node *arg0 = call->in(TypeFunc::Parms); 3752 if (arg0->is_Type() && 3753 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3754 required_outcnt--; 3755 } 3756 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3757 call->req() > TypeFunc::Parms+1 && 3758 call->is_CallStaticJava()) { 3759 // Check for negative array length. In such case, the optimizer has 3760 // detected that the allocation attempt will always result in an 3761 // exception. There is no fall-through projection of this CatchNode . 3762 Node *arg1 = call->in(TypeFunc::Parms+1); 3763 if (arg1->is_Type() && 3764 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3765 required_outcnt--; 3766 } 3767 } 3768 } 3769 } 3770 // Recheck with a better notion of 'required_outcnt' 3771 if (n->outcnt() != required_outcnt) { 3772 record_method_not_compilable("malformed control flow"); 3773 return true; // Not all targets reachable! 3774 } 3775 } 3776 // Check that I actually visited all kids. Unreached kids 3777 // must be infinite loops. 3778 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3779 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3780 record_method_not_compilable("infinite loop"); 3781 return true; // Found unvisited kid; must be unreach 3782 } 3783 3784 // Here so verification code in final_graph_reshaping_walk() 3785 // always see an OuterStripMinedLoopEnd 3786 if (n->is_OuterStripMinedLoopEnd()) { 3787 IfNode* init_iff = n->as_If(); 3788 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3789 n->subsume_by(iff, this); 3790 } 3791 } 3792 3793 // If original bytecodes contained a mixture of floats and doubles 3794 // check if the optimizer has made it homogenous, item (3). 3795 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3796 frc.get_float_count() > 32 && 3797 frc.get_double_count() == 0 && 3798 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3799 set_24_bit_selection_and_mode( false, true ); 3800 } 3801 3802 set_java_calls(frc.get_java_call_count()); 3803 set_inner_loops(frc.get_inner_loop_count()); 3804 3805 // No infinite loops, no reason to bail out. 3806 return false; 3807 } 3808 3809 //-----------------------------too_many_traps---------------------------------- 3810 // Report if there are too many traps at the current method and bci. 3811 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3812 bool Compile::too_many_traps(ciMethod* method, 3813 int bci, 3814 Deoptimization::DeoptReason reason) { 3815 ciMethodData* md = method->method_data(); 3816 if (md->is_empty()) { 3817 // Assume the trap has not occurred, or that it occurred only 3818 // because of a transient condition during start-up in the interpreter. 3819 return false; 3820 } 3821 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3822 if (md->has_trap_at(bci, m, reason) != 0) { 3823 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3824 // Also, if there are multiple reasons, or if there is no per-BCI record, 3825 // assume the worst. 3826 if (log()) 3827 log()->elem("observe trap='%s' count='%d'", 3828 Deoptimization::trap_reason_name(reason), 3829 md->trap_count(reason)); 3830 return true; 3831 } else { 3832 // Ignore method/bci and see if there have been too many globally. 3833 return too_many_traps(reason, md); 3834 } 3835 } 3836 3837 // Less-accurate variant which does not require a method and bci. 3838 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3839 ciMethodData* logmd) { 3840 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3841 // Too many traps globally. 3842 // Note that we use cumulative trap_count, not just md->trap_count. 3843 if (log()) { 3844 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3845 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3846 Deoptimization::trap_reason_name(reason), 3847 mcount, trap_count(reason)); 3848 } 3849 return true; 3850 } else { 3851 // The coast is clear. 3852 return false; 3853 } 3854 } 3855 3856 //--------------------------too_many_recompiles-------------------------------- 3857 // Report if there are too many recompiles at the current method and bci. 3858 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3859 // Is not eager to return true, since this will cause the compiler to use 3860 // Action_none for a trap point, to avoid too many recompilations. 3861 bool Compile::too_many_recompiles(ciMethod* method, 3862 int bci, 3863 Deoptimization::DeoptReason reason) { 3864 ciMethodData* md = method->method_data(); 3865 if (md->is_empty()) { 3866 // Assume the trap has not occurred, or that it occurred only 3867 // because of a transient condition during start-up in the interpreter. 3868 return false; 3869 } 3870 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3871 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3872 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3873 Deoptimization::DeoptReason per_bc_reason 3874 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3875 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3876 if ((per_bc_reason == Deoptimization::Reason_none 3877 || md->has_trap_at(bci, m, reason) != 0) 3878 // The trap frequency measure we care about is the recompile count: 3879 && md->trap_recompiled_at(bci, m) 3880 && md->overflow_recompile_count() >= bc_cutoff) { 3881 // Do not emit a trap here if it has already caused recompilations. 3882 // Also, if there are multiple reasons, or if there is no per-BCI record, 3883 // assume the worst. 3884 if (log()) 3885 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3886 Deoptimization::trap_reason_name(reason), 3887 md->trap_count(reason), 3888 md->overflow_recompile_count()); 3889 return true; 3890 } else if (trap_count(reason) != 0 3891 && decompile_count() >= m_cutoff) { 3892 // Too many recompiles globally, and we have seen this sort of trap. 3893 // Use cumulative decompile_count, not just md->decompile_count. 3894 if (log()) 3895 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3896 Deoptimization::trap_reason_name(reason), 3897 md->trap_count(reason), trap_count(reason), 3898 md->decompile_count(), decompile_count()); 3899 return true; 3900 } else { 3901 // The coast is clear. 3902 return false; 3903 } 3904 } 3905 3906 // Compute when not to trap. Used by matching trap based nodes and 3907 // NullCheck optimization. 3908 void Compile::set_allowed_deopt_reasons() { 3909 _allowed_reasons = 0; 3910 if (is_method_compilation()) { 3911 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3912 assert(rs < BitsPerInt, "recode bit map"); 3913 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3914 _allowed_reasons |= nth_bit(rs); 3915 } 3916 } 3917 } 3918 } 3919 3920 bool Compile::is_compiling_clinit_for(ciKlass* k) { 3921 ciMethod* root = method(); // the root method of compilation 3922 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit 3923 } 3924 3925 #ifndef PRODUCT 3926 //------------------------------verify_graph_edges--------------------------- 3927 // Walk the Graph and verify that there is a one-to-one correspondence 3928 // between Use-Def edges and Def-Use edges in the graph. 3929 void Compile::verify_graph_edges(bool no_dead_code) { 3930 if (VerifyGraphEdges) { 3931 ResourceArea *area = Thread::current()->resource_area(); 3932 Unique_Node_List visited(area); 3933 // Call recursive graph walk to check edges 3934 _root->verify_edges(visited); 3935 if (no_dead_code) { 3936 // Now make sure that no visited node is used by an unvisited node. 3937 bool dead_nodes = false; 3938 Unique_Node_List checked(area); 3939 while (visited.size() > 0) { 3940 Node* n = visited.pop(); 3941 checked.push(n); 3942 for (uint i = 0; i < n->outcnt(); i++) { 3943 Node* use = n->raw_out(i); 3944 if (checked.member(use)) continue; // already checked 3945 if (visited.member(use)) continue; // already in the graph 3946 if (use->is_Con()) continue; // a dead ConNode is OK 3947 // At this point, we have found a dead node which is DU-reachable. 3948 if (!dead_nodes) { 3949 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3950 dead_nodes = true; 3951 } 3952 use->dump(2); 3953 tty->print_cr("---"); 3954 checked.push(use); // No repeats; pretend it is now checked. 3955 } 3956 } 3957 assert(!dead_nodes, "using nodes must be reachable from root"); 3958 } 3959 } 3960 } 3961 #endif 3962 3963 // The Compile object keeps track of failure reasons separately from the ciEnv. 3964 // This is required because there is not quite a 1-1 relation between the 3965 // ciEnv and its compilation task and the Compile object. Note that one 3966 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3967 // to backtrack and retry without subsuming loads. Other than this backtracking 3968 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3969 // by the logic in C2Compiler. 3970 void Compile::record_failure(const char* reason) { 3971 if (log() != NULL) { 3972 log()->elem("failure reason='%s' phase='compile'", reason); 3973 } 3974 if (_failure_reason == NULL) { 3975 // Record the first failure reason. 3976 _failure_reason = reason; 3977 } 3978 3979 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3980 C->print_method(PHASE_FAILURE); 3981 } 3982 _root = NULL; // flush the graph, too 3983 } 3984 3985 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3986 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3987 _phase_name(name), _dolog(CITimeVerbose) 3988 { 3989 if (_dolog) { 3990 C = Compile::current(); 3991 _log = C->log(); 3992 } else { 3993 C = NULL; 3994 _log = NULL; 3995 } 3996 if (_log != NULL) { 3997 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3998 _log->stamp(); 3999 _log->end_head(); 4000 } 4001 } 4002 4003 Compile::TracePhase::~TracePhase() { 4004 4005 C = Compile::current(); 4006 if (_dolog) { 4007 _log = C->log(); 4008 } else { 4009 _log = NULL; 4010 } 4011 4012 #ifdef ASSERT 4013 if (PrintIdealNodeCount) { 4014 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4015 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4016 } 4017 4018 if (VerifyIdealNodeCount) { 4019 Compile::current()->print_missing_nodes(); 4020 } 4021 #endif 4022 4023 if (_log != NULL) { 4024 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4025 } 4026 } 4027 4028 //============================================================================= 4029 // Two Constant's are equal when the type and the value are equal. 4030 bool Compile::Constant::operator==(const Constant& other) { 4031 if (type() != other.type() ) return false; 4032 if (can_be_reused() != other.can_be_reused()) return false; 4033 // For floating point values we compare the bit pattern. 4034 switch (type()) { 4035 case T_INT: 4036 case T_FLOAT: return (_v._value.i == other._v._value.i); 4037 case T_LONG: 4038 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4039 case T_OBJECT: 4040 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4041 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4042 case T_METADATA: return (_v._metadata == other._v._metadata); 4043 default: ShouldNotReachHere(); return false; 4044 } 4045 } 4046 4047 static int type_to_size_in_bytes(BasicType t) { 4048 switch (t) { 4049 case T_INT: return sizeof(jint ); 4050 case T_LONG: return sizeof(jlong ); 4051 case T_FLOAT: return sizeof(jfloat ); 4052 case T_DOUBLE: return sizeof(jdouble); 4053 case T_METADATA: return sizeof(Metadata*); 4054 // We use T_VOID as marker for jump-table entries (labels) which 4055 // need an internal word relocation. 4056 case T_VOID: 4057 case T_ADDRESS: 4058 case T_OBJECT: return sizeof(jobject); 4059 default: 4060 ShouldNotReachHere(); 4061 return -1; 4062 } 4063 } 4064 4065 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4066 // sort descending 4067 if (a->freq() > b->freq()) return -1; 4068 if (a->freq() < b->freq()) return 1; 4069 return 0; 4070 } 4071 4072 void Compile::ConstantTable::calculate_offsets_and_size() { 4073 // First, sort the array by frequencies. 4074 _constants.sort(qsort_comparator); 4075 4076 #ifdef ASSERT 4077 // Make sure all jump-table entries were sorted to the end of the 4078 // array (they have a negative frequency). 4079 bool found_void = false; 4080 for (int i = 0; i < _constants.length(); i++) { 4081 Constant con = _constants.at(i); 4082 if (con.type() == T_VOID) 4083 found_void = true; // jump-tables 4084 else 4085 assert(!found_void, "wrong sorting"); 4086 } 4087 #endif 4088 4089 int offset = 0; 4090 for (int i = 0; i < _constants.length(); i++) { 4091 Constant* con = _constants.adr_at(i); 4092 4093 // Align offset for type. 4094 int typesize = type_to_size_in_bytes(con->type()); 4095 offset = align_up(offset, typesize); 4096 con->set_offset(offset); // set constant's offset 4097 4098 if (con->type() == T_VOID) { 4099 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4100 offset = offset + typesize * n->outcnt(); // expand jump-table 4101 } else { 4102 offset = offset + typesize; 4103 } 4104 } 4105 4106 // Align size up to the next section start (which is insts; see 4107 // CodeBuffer::align_at_start). 4108 assert(_size == -1, "already set?"); 4109 _size = align_up(offset, (int)CodeEntryAlignment); 4110 } 4111 4112 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4113 MacroAssembler _masm(&cb); 4114 for (int i = 0; i < _constants.length(); i++) { 4115 Constant con = _constants.at(i); 4116 address constant_addr = NULL; 4117 switch (con.type()) { 4118 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4119 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4120 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4121 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4122 case T_OBJECT: { 4123 jobject obj = con.get_jobject(); 4124 int oop_index = _masm.oop_recorder()->find_index(obj); 4125 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4126 break; 4127 } 4128 case T_ADDRESS: { 4129 address addr = (address) con.get_jobject(); 4130 constant_addr = _masm.address_constant(addr); 4131 break; 4132 } 4133 // We use T_VOID as marker for jump-table entries (labels) which 4134 // need an internal word relocation. 4135 case T_VOID: { 4136 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4137 // Fill the jump-table with a dummy word. The real value is 4138 // filled in later in fill_jump_table. 4139 address dummy = (address) n; 4140 constant_addr = _masm.address_constant(dummy); 4141 // Expand jump-table 4142 for (uint i = 1; i < n->outcnt(); i++) { 4143 address temp_addr = _masm.address_constant(dummy + i); 4144 assert(temp_addr, "consts section too small"); 4145 } 4146 break; 4147 } 4148 case T_METADATA: { 4149 Metadata* obj = con.get_metadata(); 4150 int metadata_index = _masm.oop_recorder()->find_index(obj); 4151 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4152 break; 4153 } 4154 default: ShouldNotReachHere(); 4155 } 4156 assert(constant_addr, "consts section too small"); 4157 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4158 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4159 } 4160 } 4161 4162 int Compile::ConstantTable::find_offset(Constant& con) const { 4163 int idx = _constants.find(con); 4164 guarantee(idx != -1, "constant must be in constant table"); 4165 int offset = _constants.at(idx).offset(); 4166 guarantee(offset != -1, "constant table not emitted yet?"); 4167 return offset; 4168 } 4169 4170 void Compile::ConstantTable::add(Constant& con) { 4171 if (con.can_be_reused()) { 4172 int idx = _constants.find(con); 4173 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4174 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4175 return; 4176 } 4177 } 4178 (void) _constants.append(con); 4179 } 4180 4181 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4182 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4183 Constant con(type, value, b->_freq); 4184 add(con); 4185 return con; 4186 } 4187 4188 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4189 Constant con(metadata); 4190 add(con); 4191 return con; 4192 } 4193 4194 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4195 jvalue value; 4196 BasicType type = oper->type()->basic_type(); 4197 switch (type) { 4198 case T_LONG: value.j = oper->constantL(); break; 4199 case T_FLOAT: value.f = oper->constantF(); break; 4200 case T_DOUBLE: value.d = oper->constantD(); break; 4201 case T_OBJECT: 4202 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4203 case T_METADATA: return add((Metadata*)oper->constant()); break; 4204 default: guarantee(false, "unhandled type: %s", type2name(type)); 4205 } 4206 return add(n, type, value); 4207 } 4208 4209 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4210 jvalue value; 4211 // We can use the node pointer here to identify the right jump-table 4212 // as this method is called from Compile::Fill_buffer right before 4213 // the MachNodes are emitted and the jump-table is filled (means the 4214 // MachNode pointers do not change anymore). 4215 value.l = (jobject) n; 4216 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4217 add(con); 4218 return con; 4219 } 4220 4221 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4222 // If called from Compile::scratch_emit_size do nothing. 4223 if (Compile::current()->in_scratch_emit_size()) return; 4224 4225 assert(labels.is_nonempty(), "must be"); 4226 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4227 4228 // Since MachConstantNode::constant_offset() also contains 4229 // table_base_offset() we need to subtract the table_base_offset() 4230 // to get the plain offset into the constant table. 4231 int offset = n->constant_offset() - table_base_offset(); 4232 4233 MacroAssembler _masm(&cb); 4234 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4235 4236 for (uint i = 0; i < n->outcnt(); i++) { 4237 address* constant_addr = &jump_table_base[i]; 4238 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)); 4239 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4240 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4241 } 4242 } 4243 4244 //----------------------------static_subtype_check----------------------------- 4245 // Shortcut important common cases when superklass is exact: 4246 // (0) superklass is java.lang.Object (can occur in reflective code) 4247 // (1) subklass is already limited to a subtype of superklass => always ok 4248 // (2) subklass does not overlap with superklass => always fail 4249 // (3) superklass has NO subtypes and we can check with a simple compare. 4250 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4251 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4252 return SSC_full_test; // Let caller generate the general case. 4253 } 4254 4255 if (superk == env()->Object_klass()) { 4256 return SSC_always_true; // (0) this test cannot fail 4257 } 4258 4259 ciType* superelem = superk; 4260 if (superelem->is_array_klass()) 4261 superelem = superelem->as_array_klass()->base_element_type(); 4262 4263 if (!subk->is_interface()) { // cannot trust static interface types yet 4264 if (subk->is_subtype_of(superk)) { 4265 return SSC_always_true; // (1) false path dead; no dynamic test needed 4266 } 4267 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4268 !superk->is_subtype_of(subk)) { 4269 return SSC_always_false; 4270 } 4271 } 4272 4273 // If casting to an instance klass, it must have no subtypes 4274 if (superk->is_interface()) { 4275 // Cannot trust interfaces yet. 4276 // %%% S.B. superk->nof_implementors() == 1 4277 } else if (superelem->is_instance_klass()) { 4278 ciInstanceKlass* ik = superelem->as_instance_klass(); 4279 if (!ik->has_subklass() && !ik->is_interface()) { 4280 if (!ik->is_final()) { 4281 // Add a dependency if there is a chance of a later subclass. 4282 dependencies()->assert_leaf_type(ik); 4283 } 4284 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4285 } 4286 } else { 4287 // A primitive array type has no subtypes. 4288 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4289 } 4290 4291 return SSC_full_test; 4292 } 4293 4294 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4295 #ifdef _LP64 4296 // The scaled index operand to AddP must be a clean 64-bit value. 4297 // Java allows a 32-bit int to be incremented to a negative 4298 // value, which appears in a 64-bit register as a large 4299 // positive number. Using that large positive number as an 4300 // operand in pointer arithmetic has bad consequences. 4301 // On the other hand, 32-bit overflow is rare, and the possibility 4302 // can often be excluded, if we annotate the ConvI2L node with 4303 // a type assertion that its value is known to be a small positive 4304 // number. (The prior range check has ensured this.) 4305 // This assertion is used by ConvI2LNode::Ideal. 4306 int index_max = max_jint - 1; // array size is max_jint, index is one less 4307 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4308 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4309 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4310 #endif 4311 return idx; 4312 } 4313 4314 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4315 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4316 if (ctrl != NULL) { 4317 // Express control dependency by a CastII node with a narrow type. 4318 value = new CastIINode(value, itype, false, true /* range check dependency */); 4319 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4320 // node from floating above the range check during loop optimizations. Otherwise, the 4321 // ConvI2L node may be eliminated independently of the range check, causing the data path 4322 // to become TOP while the control path is still there (although it's unreachable). 4323 value->set_req(0, ctrl); 4324 // Save CastII node to remove it after loop optimizations. 4325 phase->C->add_range_check_cast(value); 4326 value = phase->transform(value); 4327 } 4328 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4329 return phase->transform(new ConvI2LNode(value, ltype)); 4330 } 4331 4332 // The message about the current inlining is accumulated in 4333 // _print_inlining_stream and transfered into the _print_inlining_list 4334 // once we know whether inlining succeeds or not. For regular 4335 // inlining, messages are appended to the buffer pointed by 4336 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4337 // a new buffer is added after _print_inlining_idx in the list. This 4338 // way we can update the inlining message for late inlining call site 4339 // when the inlining is attempted again. 4340 void Compile::print_inlining_init() { 4341 if (print_inlining() || print_intrinsics()) { 4342 _print_inlining_stream = new stringStream(); 4343 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4344 } 4345 } 4346 4347 void Compile::print_inlining_reinit() { 4348 if (print_inlining() || print_intrinsics()) { 4349 // Re allocate buffer when we change ResourceMark 4350 _print_inlining_stream = new stringStream(); 4351 } 4352 } 4353 4354 void Compile::print_inlining_reset() { 4355 _print_inlining_stream->reset(); 4356 } 4357 4358 void Compile::print_inlining_commit() { 4359 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4360 // Transfer the message from _print_inlining_stream to the current 4361 // _print_inlining_list buffer and clear _print_inlining_stream. 4362 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4363 print_inlining_reset(); 4364 } 4365 4366 void Compile::print_inlining_push() { 4367 // Add new buffer to the _print_inlining_list at current position 4368 _print_inlining_idx++; 4369 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4370 } 4371 4372 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4373 return _print_inlining_list->at(_print_inlining_idx); 4374 } 4375 4376 void Compile::print_inlining_update(CallGenerator* cg) { 4377 if (print_inlining() || print_intrinsics()) { 4378 if (!cg->is_late_inline()) { 4379 if (print_inlining_current().cg() != NULL) { 4380 print_inlining_push(); 4381 } 4382 print_inlining_commit(); 4383 } else { 4384 if (print_inlining_current().cg() != cg && 4385 (print_inlining_current().cg() != NULL || 4386 print_inlining_current().ss()->size() != 0)) { 4387 print_inlining_push(); 4388 } 4389 print_inlining_commit(); 4390 print_inlining_current().set_cg(cg); 4391 } 4392 } 4393 } 4394 4395 void Compile::print_inlining_move_to(CallGenerator* cg) { 4396 // We resume inlining at a late inlining call site. Locate the 4397 // corresponding inlining buffer so that we can update it. 4398 if (print_inlining()) { 4399 for (int i = 0; i < _print_inlining_list->length(); i++) { 4400 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4401 _print_inlining_idx = i; 4402 return; 4403 } 4404 } 4405 ShouldNotReachHere(); 4406 } 4407 } 4408 4409 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4410 if (print_inlining()) { 4411 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4412 assert(print_inlining_current().cg() == cg, "wrong entry"); 4413 // replace message with new message 4414 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4415 print_inlining_commit(); 4416 print_inlining_current().set_cg(cg); 4417 } 4418 } 4419 4420 void Compile::print_inlining_assert_ready() { 4421 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4422 } 4423 4424 void Compile::process_print_inlining() { 4425 bool do_print_inlining = print_inlining() || print_intrinsics(); 4426 if (do_print_inlining || log() != NULL) { 4427 // Print inlining message for candidates that we couldn't inline 4428 // for lack of space 4429 for (int i = 0; i < _late_inlines.length(); i++) { 4430 CallGenerator* cg = _late_inlines.at(i); 4431 if (!cg->is_mh_late_inline()) { 4432 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4433 if (do_print_inlining) { 4434 cg->print_inlining_late(msg); 4435 } 4436 log_late_inline_failure(cg, msg); 4437 } 4438 } 4439 } 4440 if (do_print_inlining) { 4441 ResourceMark rm; 4442 stringStream ss; 4443 for (int i = 0; i < _print_inlining_list->length(); i++) { 4444 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4445 } 4446 size_t end = ss.size(); 4447 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4448 strncpy(_print_inlining_output, ss.base(), end+1); 4449 _print_inlining_output[end] = 0; 4450 } 4451 } 4452 4453 void Compile::dump_print_inlining() { 4454 if (_print_inlining_output != NULL) { 4455 tty->print_raw(_print_inlining_output); 4456 } 4457 } 4458 4459 void Compile::log_late_inline(CallGenerator* cg) { 4460 if (log() != NULL) { 4461 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4462 cg->unique_id()); 4463 JVMState* p = cg->call_node()->jvms(); 4464 while (p != NULL) { 4465 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4466 p = p->caller(); 4467 } 4468 log()->tail("late_inline"); 4469 } 4470 } 4471 4472 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4473 log_late_inline(cg); 4474 if (log() != NULL) { 4475 log()->inline_fail(msg); 4476 } 4477 } 4478 4479 void Compile::log_inline_id(CallGenerator* cg) { 4480 if (log() != NULL) { 4481 // The LogCompilation tool needs a unique way to identify late 4482 // inline call sites. This id must be unique for this call site in 4483 // this compilation. Try to have it unique across compilations as 4484 // well because it can be convenient when grepping through the log 4485 // file. 4486 // Distinguish OSR compilations from others in case CICountOSR is 4487 // on. 4488 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4489 cg->set_unique_id(id); 4490 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4491 } 4492 } 4493 4494 void Compile::log_inline_failure(const char* msg) { 4495 if (C->log() != NULL) { 4496 C->log()->inline_fail(msg); 4497 } 4498 } 4499 4500 4501 // Dump inlining replay data to the stream. 4502 // Don't change thread state and acquire any locks. 4503 void Compile::dump_inline_data(outputStream* out) { 4504 InlineTree* inl_tree = ilt(); 4505 if (inl_tree != NULL) { 4506 out->print(" inline %d", inl_tree->count()); 4507 inl_tree->dump_replay_data(out); 4508 } 4509 } 4510 4511 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4512 if (n1->Opcode() < n2->Opcode()) return -1; 4513 else if (n1->Opcode() > n2->Opcode()) return 1; 4514 4515 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4516 for (uint i = 1; i < n1->req(); i++) { 4517 if (n1->in(i) < n2->in(i)) return -1; 4518 else if (n1->in(i) > n2->in(i)) return 1; 4519 } 4520 4521 return 0; 4522 } 4523 4524 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4525 Node* n1 = *n1p; 4526 Node* n2 = *n2p; 4527 4528 return cmp_expensive_nodes(n1, n2); 4529 } 4530 4531 void Compile::sort_expensive_nodes() { 4532 if (!expensive_nodes_sorted()) { 4533 _expensive_nodes->sort(cmp_expensive_nodes); 4534 } 4535 } 4536 4537 bool Compile::expensive_nodes_sorted() const { 4538 for (int i = 1; i < _expensive_nodes->length(); i++) { 4539 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4540 return false; 4541 } 4542 } 4543 return true; 4544 } 4545 4546 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4547 if (_expensive_nodes->length() == 0) { 4548 return false; 4549 } 4550 4551 assert(OptimizeExpensiveOps, "optimization off?"); 4552 4553 // Take this opportunity to remove dead nodes from the list 4554 int j = 0; 4555 for (int i = 0; i < _expensive_nodes->length(); i++) { 4556 Node* n = _expensive_nodes->at(i); 4557 if (!n->is_unreachable(igvn)) { 4558 assert(n->is_expensive(), "should be expensive"); 4559 _expensive_nodes->at_put(j, n); 4560 j++; 4561 } 4562 } 4563 _expensive_nodes->trunc_to(j); 4564 4565 // Then sort the list so that similar nodes are next to each other 4566 // and check for at least two nodes of identical kind with same data 4567 // inputs. 4568 sort_expensive_nodes(); 4569 4570 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4571 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4572 return true; 4573 } 4574 } 4575 4576 return false; 4577 } 4578 4579 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4580 if (_expensive_nodes->length() == 0) { 4581 return; 4582 } 4583 4584 assert(OptimizeExpensiveOps, "optimization off?"); 4585 4586 // Sort to bring similar nodes next to each other and clear the 4587 // control input of nodes for which there's only a single copy. 4588 sort_expensive_nodes(); 4589 4590 int j = 0; 4591 int identical = 0; 4592 int i = 0; 4593 bool modified = false; 4594 for (; i < _expensive_nodes->length()-1; i++) { 4595 assert(j <= i, "can't write beyond current index"); 4596 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4597 identical++; 4598 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4599 continue; 4600 } 4601 if (identical > 0) { 4602 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4603 identical = 0; 4604 } else { 4605 Node* n = _expensive_nodes->at(i); 4606 igvn.replace_input_of(n, 0, NULL); 4607 igvn.hash_insert(n); 4608 modified = true; 4609 } 4610 } 4611 if (identical > 0) { 4612 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4613 } else if (_expensive_nodes->length() >= 1) { 4614 Node* n = _expensive_nodes->at(i); 4615 igvn.replace_input_of(n, 0, NULL); 4616 igvn.hash_insert(n); 4617 modified = true; 4618 } 4619 _expensive_nodes->trunc_to(j); 4620 if (modified) { 4621 igvn.optimize(); 4622 } 4623 } 4624 4625 void Compile::add_expensive_node(Node * n) { 4626 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4627 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4628 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4629 if (OptimizeExpensiveOps) { 4630 _expensive_nodes->append(n); 4631 } else { 4632 // Clear control input and let IGVN optimize expensive nodes if 4633 // OptimizeExpensiveOps is off. 4634 n->set_req(0, NULL); 4635 } 4636 } 4637 4638 /** 4639 * Remove the speculative part of types and clean up the graph 4640 */ 4641 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4642 if (UseTypeSpeculation) { 4643 Unique_Node_List worklist; 4644 worklist.push(root()); 4645 int modified = 0; 4646 // Go over all type nodes that carry a speculative type, drop the 4647 // speculative part of the type and enqueue the node for an igvn 4648 // which may optimize it out. 4649 for (uint next = 0; next < worklist.size(); ++next) { 4650 Node *n = worklist.at(next); 4651 if (n->is_Type()) { 4652 TypeNode* tn = n->as_Type(); 4653 const Type* t = tn->type(); 4654 const Type* t_no_spec = t->remove_speculative(); 4655 if (t_no_spec != t) { 4656 bool in_hash = igvn.hash_delete(n); 4657 assert(in_hash, "node should be in igvn hash table"); 4658 tn->set_type(t_no_spec); 4659 igvn.hash_insert(n); 4660 igvn._worklist.push(n); // give it a chance to go away 4661 modified++; 4662 } 4663 } 4664 uint max = n->len(); 4665 for( uint i = 0; i < max; ++i ) { 4666 Node *m = n->in(i); 4667 if (not_a_node(m)) continue; 4668 worklist.push(m); 4669 } 4670 } 4671 // Drop the speculative part of all types in the igvn's type table 4672 igvn.remove_speculative_types(); 4673 if (modified > 0) { 4674 igvn.optimize(); 4675 } 4676 #ifdef ASSERT 4677 // Verify that after the IGVN is over no speculative type has resurfaced 4678 worklist.clear(); 4679 worklist.push(root()); 4680 for (uint next = 0; next < worklist.size(); ++next) { 4681 Node *n = worklist.at(next); 4682 const Type* t = igvn.type_or_null(n); 4683 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4684 if (n->is_Type()) { 4685 t = n->as_Type()->type(); 4686 assert(t == t->remove_speculative(), "no more speculative types"); 4687 } 4688 uint max = n->len(); 4689 for( uint i = 0; i < max; ++i ) { 4690 Node *m = n->in(i); 4691 if (not_a_node(m)) continue; 4692 worklist.push(m); 4693 } 4694 } 4695 igvn.check_no_speculative_types(); 4696 #endif 4697 } 4698 } 4699 4700 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 4701 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 4702 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 4703 if (!EnableValhalla || ta == NULL || tb == NULL || 4704 ta->is_zero_type() || tb->is_zero_type() || 4705 !ta->can_be_value_type() || !tb->can_be_value_type()) { 4706 // Use old acmp if one operand is null or not a value type 4707 return new CmpPNode(a, b); 4708 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) { 4709 // We know that one operand is a value type. Therefore, 4710 // new acmp will only return true if both operands are NULL. 4711 // Check if both operands are null by or'ing the oops. 4712 a = phase->transform(new CastP2XNode(NULL, a)); 4713 b = phase->transform(new CastP2XNode(NULL, b)); 4714 a = phase->transform(new OrXNode(a, b)); 4715 return new CmpXNode(a, phase->MakeConX(0)); 4716 } 4717 // Use new acmp 4718 return NULL; 4719 } 4720 4721 // Auxiliary method to support randomized stressing/fuzzing. 4722 // 4723 // This method can be called the arbitrary number of times, with current count 4724 // as the argument. The logic allows selecting a single candidate from the 4725 // running list of candidates as follows: 4726 // int count = 0; 4727 // Cand* selected = null; 4728 // while(cand = cand->next()) { 4729 // if (randomized_select(++count)) { 4730 // selected = cand; 4731 // } 4732 // } 4733 // 4734 // Including count equalizes the chances any candidate is "selected". 4735 // This is useful when we don't have the complete list of candidates to choose 4736 // from uniformly. In this case, we need to adjust the randomicity of the 4737 // selection, or else we will end up biasing the selection towards the latter 4738 // candidates. 4739 // 4740 // Quick back-envelope calculation shows that for the list of n candidates 4741 // the equal probability for the candidate to persist as "best" can be 4742 // achieved by replacing it with "next" k-th candidate with the probability 4743 // of 1/k. It can be easily shown that by the end of the run, the 4744 // probability for any candidate is converged to 1/n, thus giving the 4745 // uniform distribution among all the candidates. 4746 // 4747 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4748 #define RANDOMIZED_DOMAIN_POW 29 4749 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4750 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4751 bool Compile::randomized_select(int count) { 4752 assert(count > 0, "only positive"); 4753 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4754 } 4755 4756 CloneMap& Compile::clone_map() { return _clone_map; } 4757 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4758 4759 void NodeCloneInfo::dump() const { 4760 tty->print(" {%d:%d} ", idx(), gen()); 4761 } 4762 4763 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4764 uint64_t val = value(old->_idx); 4765 NodeCloneInfo cio(val); 4766 assert(val != 0, "old node should be in the map"); 4767 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4768 insert(nnn->_idx, cin.get()); 4769 #ifndef PRODUCT 4770 if (is_debug()) { 4771 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4772 } 4773 #endif 4774 } 4775 4776 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4777 NodeCloneInfo cio(value(old->_idx)); 4778 if (cio.get() == 0) { 4779 cio.set(old->_idx, 0); 4780 insert(old->_idx, cio.get()); 4781 #ifndef PRODUCT 4782 if (is_debug()) { 4783 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4784 } 4785 #endif 4786 } 4787 clone(old, nnn, gen); 4788 } 4789 4790 int CloneMap::max_gen() const { 4791 int g = 0; 4792 DictI di(_dict); 4793 for(; di.test(); ++di) { 4794 int t = gen(di._key); 4795 if (g < t) { 4796 g = t; 4797 #ifndef PRODUCT 4798 if (is_debug()) { 4799 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4800 } 4801 #endif 4802 } 4803 } 4804 return g; 4805 } 4806 4807 void CloneMap::dump(node_idx_t key) const { 4808 uint64_t val = value(key); 4809 if (val != 0) { 4810 NodeCloneInfo ni(val); 4811 ni.dump(); 4812 } 4813 }