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