1 /* 2 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "opto/addnode.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/callGenerator.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/castnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/chaitin.hpp" 47 #include "opto/compile.hpp" 48 #include "opto/connode.hpp" 49 #include "opto/convertnode.hpp" 50 #include "opto/divnode.hpp" 51 #include "opto/escape.hpp" 52 #include "opto/idealGraphPrinter.hpp" 53 #include "opto/loopnode.hpp" 54 #include "opto/machnode.hpp" 55 #include "opto/macro.hpp" 56 #include "opto/matcher.hpp" 57 #include "opto/mathexactnode.hpp" 58 #include "opto/memnode.hpp" 59 #include "opto/mulnode.hpp" 60 #include "opto/narrowptrnode.hpp" 61 #include "opto/node.hpp" 62 #include "opto/opcodes.hpp" 63 #include "opto/output.hpp" 64 #include "opto/parse.hpp" 65 #include "opto/phaseX.hpp" 66 #include "opto/rootnode.hpp" 67 #include "opto/runtime.hpp" 68 #include "opto/stringopts.hpp" 69 #include "opto/type.hpp" 70 #include "opto/vectornode.hpp" 71 #include "runtime/arguments.hpp" 72 #include "runtime/sharedRuntime.hpp" 73 #include "runtime/signature.hpp" 74 #include "runtime/stubRoutines.hpp" 75 #include "runtime/timer.hpp" 76 #include "utilities/align.hpp" 77 #include "utilities/copy.hpp" 78 #include "utilities/macros.hpp" 79 #if INCLUDE_ZGC 80 #include "gc/z/c2/zBarrierSetC2.hpp" 81 #endif 82 83 84 // -------------------- Compile::mach_constant_base_node ----------------------- 85 // Constant table base node singleton. 86 MachConstantBaseNode* Compile::mach_constant_base_node() { 87 if (_mach_constant_base_node == NULL) { 88 _mach_constant_base_node = new MachConstantBaseNode(); 89 _mach_constant_base_node->add_req(C->root()); 90 } 91 return _mach_constant_base_node; 92 } 93 94 95 /// Support for intrinsics. 96 97 // Return the index at which m must be inserted (or already exists). 98 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 99 class IntrinsicDescPair { 100 private: 101 ciMethod* _m; 102 bool _is_virtual; 103 public: 104 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 105 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 106 ciMethod* m= elt->method(); 107 ciMethod* key_m = key->_m; 108 if (key_m < m) return -1; 109 else if (key_m > m) return 1; 110 else { 111 bool is_virtual = elt->is_virtual(); 112 bool key_virtual = key->_is_virtual; 113 if (key_virtual < is_virtual) return -1; 114 else if (key_virtual > is_virtual) return 1; 115 else return 0; 116 } 117 } 118 }; 119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 120 #ifdef ASSERT 121 for (int i = 1; i < _intrinsics->length(); i++) { 122 CallGenerator* cg1 = _intrinsics->at(i-1); 123 CallGenerator* cg2 = _intrinsics->at(i); 124 assert(cg1->method() != cg2->method() 125 ? cg1->method() < cg2->method() 126 : cg1->is_virtual() < cg2->is_virtual(), 127 "compiler intrinsics list must stay sorted"); 128 } 129 #endif 130 IntrinsicDescPair pair(m, is_virtual); 131 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 132 } 133 134 void Compile::register_intrinsic(CallGenerator* cg) { 135 if (_intrinsics == NULL) { 136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 137 } 138 int len = _intrinsics->length(); 139 bool found = false; 140 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 141 assert(!found, "registering twice"); 142 _intrinsics->insert_before(index, cg); 143 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 144 } 145 146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 147 assert(m->is_loaded(), "don't try this on unloaded methods"); 148 if (_intrinsics != NULL) { 149 bool found = false; 150 int index = intrinsic_insertion_index(m, is_virtual, found); 151 if (found) { 152 return _intrinsics->at(index); 153 } 154 } 155 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 156 if (m->intrinsic_id() != vmIntrinsics::_none && 157 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 158 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 159 if (cg != NULL) { 160 // Save it for next time: 161 register_intrinsic(cg); 162 return cg; 163 } else { 164 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 165 } 166 } 167 return NULL; 168 } 169 170 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 171 // in library_call.cpp. 172 173 174 #ifndef PRODUCT 175 // statistics gathering... 176 177 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 179 180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 181 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 182 int oflags = _intrinsic_hist_flags[id]; 183 assert(flags != 0, "what happened?"); 184 if (is_virtual) { 185 flags |= _intrinsic_virtual; 186 } 187 bool changed = (flags != oflags); 188 if ((flags & _intrinsic_worked) != 0) { 189 juint count = (_intrinsic_hist_count[id] += 1); 190 if (count == 1) { 191 changed = true; // first time 192 } 193 // increment the overall count also: 194 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 195 } 196 if (changed) { 197 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 198 // Something changed about the intrinsic's virtuality. 199 if ((flags & _intrinsic_virtual) != 0) { 200 // This is the first use of this intrinsic as a virtual call. 201 if (oflags != 0) { 202 // We already saw it as a non-virtual, so note both cases. 203 flags |= _intrinsic_both; 204 } 205 } else if ((oflags & _intrinsic_both) == 0) { 206 // This is the first use of this intrinsic as a non-virtual 207 flags |= _intrinsic_both; 208 } 209 } 210 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 211 } 212 // update the overall flags also: 213 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 214 return changed; 215 } 216 217 static char* format_flags(int flags, char* buf) { 218 buf[0] = 0; 219 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 220 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 221 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 222 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 223 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 224 if (buf[0] == 0) strcat(buf, ","); 225 assert(buf[0] == ',', "must be"); 226 return &buf[1]; 227 } 228 229 void Compile::print_intrinsic_statistics() { 230 char flagsbuf[100]; 231 ttyLocker ttyl; 232 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 233 tty->print_cr("Compiler intrinsic usage:"); 234 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 235 if (total == 0) total = 1; // avoid div0 in case of no successes 236 #define PRINT_STAT_LINE(name, c, f) \ 237 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 238 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 239 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 240 int flags = _intrinsic_hist_flags[id]; 241 juint count = _intrinsic_hist_count[id]; 242 if ((flags | count) != 0) { 243 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 244 } 245 } 246 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 247 if (xtty != NULL) xtty->tail("statistics"); 248 } 249 250 void Compile::print_statistics() { 251 { ttyLocker ttyl; 252 if (xtty != NULL) xtty->head("statistics type='opto'"); 253 Parse::print_statistics(); 254 PhaseCCP::print_statistics(); 255 PhaseRegAlloc::print_statistics(); 256 Scheduling::print_statistics(); 257 PhasePeephole::print_statistics(); 258 PhaseIdealLoop::print_statistics(); 259 if (xtty != NULL) xtty->tail("statistics"); 260 } 261 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 262 // put this under its own <statistics> element. 263 print_intrinsic_statistics(); 264 } 265 } 266 #endif //PRODUCT 267 268 // Support for bundling info 269 Bundle* Compile::node_bundling(const Node *n) { 270 assert(valid_bundle_info(n), "oob"); 271 return &_node_bundling_base[n->_idx]; 272 } 273 274 bool Compile::valid_bundle_info(const Node *n) { 275 return (_node_bundling_limit > n->_idx); 276 } 277 278 279 void Compile::gvn_replace_by(Node* n, Node* nn) { 280 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 281 Node* use = n->last_out(i); 282 bool is_in_table = initial_gvn()->hash_delete(use); 283 uint uses_found = 0; 284 for (uint j = 0; j < use->len(); j++) { 285 if (use->in(j) == n) { 286 if (j < use->req()) 287 use->set_req(j, nn); 288 else 289 use->set_prec(j, nn); 290 uses_found++; 291 } 292 } 293 if (is_in_table) { 294 // reinsert into table 295 initial_gvn()->hash_find_insert(use); 296 } 297 record_for_igvn(use); 298 i -= uses_found; // we deleted 1 or more copies of this edge 299 } 300 } 301 302 303 static inline bool not_a_node(const Node* n) { 304 if (n == NULL) return true; 305 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 306 if (*(address*)n == badAddress) return true; // kill by Node::destruct 307 return false; 308 } 309 310 // Identify all nodes that are reachable from below, useful. 311 // Use breadth-first pass that records state in a Unique_Node_List, 312 // recursive traversal is slower. 313 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 314 int estimated_worklist_size = live_nodes(); 315 useful.map( estimated_worklist_size, NULL ); // preallocate space 316 317 // Initialize worklist 318 if (root() != NULL) { useful.push(root()); } 319 // If 'top' is cached, declare it useful to preserve cached node 320 if( cached_top_node() ) { useful.push(cached_top_node()); } 321 322 // Push all useful nodes onto the list, breadthfirst 323 for( uint next = 0; next < useful.size(); ++next ) { 324 assert( next < unique(), "Unique useful nodes < total nodes"); 325 Node *n = useful.at(next); 326 uint max = n->len(); 327 for( uint i = 0; i < max; ++i ) { 328 Node *m = n->in(i); 329 if (not_a_node(m)) continue; 330 useful.push(m); 331 } 332 } 333 } 334 335 // Update dead_node_list with any missing dead nodes using useful 336 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 337 void Compile::update_dead_node_list(Unique_Node_List &useful) { 338 uint max_idx = unique(); 339 VectorSet& useful_node_set = useful.member_set(); 340 341 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 342 // If node with index node_idx is not in useful set, 343 // mark it as dead in dead node list. 344 if (! useful_node_set.test(node_idx) ) { 345 record_dead_node(node_idx); 346 } 347 } 348 } 349 350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 351 int shift = 0; 352 for (int i = 0; i < inlines->length(); i++) { 353 CallGenerator* cg = inlines->at(i); 354 CallNode* call = cg->call_node(); 355 if (shift > 0) { 356 inlines->at_put(i-shift, cg); 357 } 358 if (!useful.member(call)) { 359 shift++; 360 } 361 } 362 inlines->trunc_to(inlines->length()-shift); 363 } 364 365 // Disconnect all useless nodes by disconnecting those at the boundary. 366 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 367 uint next = 0; 368 while (next < useful.size()) { 369 Node *n = useful.at(next++); 370 if (n->is_SafePoint()) { 371 // We're done with a parsing phase. Replaced nodes are not valid 372 // beyond that point. 373 n->as_SafePoint()->delete_replaced_nodes(); 374 } 375 // Use raw traversal of out edges since this code removes out edges 376 int max = n->outcnt(); 377 for (int j = 0; j < max; ++j) { 378 Node* child = n->raw_out(j); 379 if (! useful.member(child)) { 380 assert(!child->is_top() || child != top(), 381 "If top is cached in Compile object it is in useful list"); 382 // Only need to remove this out-edge to the useless node 383 n->raw_del_out(j); 384 --j; 385 --max; 386 } 387 } 388 if (n->outcnt() == 1 && n->has_special_unique_user()) { 389 record_for_igvn(n->unique_out()); 390 } 391 } 392 // Remove useless macro and predicate opaq nodes 393 for (int i = C->macro_count()-1; i >= 0; i--) { 394 Node* n = C->macro_node(i); 395 if (!useful.member(n)) { 396 remove_macro_node(n); 397 } 398 } 399 // Remove useless CastII nodes with range check dependency 400 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 401 Node* cast = range_check_cast_node(i); 402 if (!useful.member(cast)) { 403 remove_range_check_cast(cast); 404 } 405 } 406 // Remove useless expensive nodes 407 for (int i = C->expensive_count()-1; i >= 0; i--) { 408 Node* n = C->expensive_node(i); 409 if (!useful.member(n)) { 410 remove_expensive_node(n); 411 } 412 } 413 // Remove useless Opaque4 nodes 414 for (int i = opaque4_count() - 1; i >= 0; i--) { 415 Node* opaq = opaque4_node(i); 416 if (!useful.member(opaq)) { 417 remove_opaque4_node(opaq); 418 } 419 } 420 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 421 bs->eliminate_useless_gc_barriers(useful, this); 422 // clean up the late inline lists 423 remove_useless_late_inlines(&_string_late_inlines, useful); 424 remove_useless_late_inlines(&_boxing_late_inlines, useful); 425 remove_useless_late_inlines(&_late_inlines, useful); 426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 427 } 428 429 //------------------------------frame_size_in_words----------------------------- 430 // frame_slots in units of words 431 int Compile::frame_size_in_words() const { 432 // shift is 0 in LP32 and 1 in LP64 433 const int shift = (LogBytesPerWord - LogBytesPerInt); 434 int words = _frame_slots >> shift; 435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 436 return words; 437 } 438 439 // To bang the stack of this compiled method we use the stack size 440 // that the interpreter would need in case of a deoptimization. This 441 // removes the need to bang the stack in the deoptimization blob which 442 // in turn simplifies stack overflow handling. 443 int Compile::bang_size_in_bytes() const { 444 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 445 } 446 447 // ============================================================================ 448 //------------------------------CompileWrapper--------------------------------- 449 class CompileWrapper : public StackObj { 450 Compile *const _compile; 451 public: 452 CompileWrapper(Compile* compile); 453 454 ~CompileWrapper(); 455 }; 456 457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 458 // the Compile* pointer is stored in the current ciEnv: 459 ciEnv* env = compile->env(); 460 assert(env == ciEnv::current(), "must already be a ciEnv active"); 461 assert(env->compiler_data() == NULL, "compile already active?"); 462 env->set_compiler_data(compile); 463 assert(compile == Compile::current(), "sanity"); 464 465 compile->set_type_dict(NULL); 466 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 467 compile->clone_map().set_clone_idx(0); 468 compile->set_type_hwm(NULL); 469 compile->set_type_last_size(0); 470 compile->set_last_tf(NULL, NULL); 471 compile->set_indexSet_arena(NULL); 472 compile->set_indexSet_free_block_list(NULL); 473 compile->init_type_arena(); 474 Type::Initialize(compile); 475 _compile->set_scratch_buffer_blob(NULL); 476 _compile->begin_method(); 477 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 478 } 479 CompileWrapper::~CompileWrapper() { 480 _compile->end_method(); 481 if (_compile->scratch_buffer_blob() != NULL) 482 BufferBlob::free(_compile->scratch_buffer_blob()); 483 _compile->env()->set_compiler_data(NULL); 484 } 485 486 487 //----------------------------print_compile_messages--------------------------- 488 void Compile::print_compile_messages() { 489 #ifndef PRODUCT 490 // Check if recompiling 491 if (_subsume_loads == false && PrintOpto) { 492 // Recompiling without allowing machine instructions to subsume loads 493 tty->print_cr("*********************************************************"); 494 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 495 tty->print_cr("*********************************************************"); 496 } 497 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 498 // Recompiling without escape analysis 499 tty->print_cr("*********************************************************"); 500 tty->print_cr("** Bailout: Recompile without escape analysis **"); 501 tty->print_cr("*********************************************************"); 502 } 503 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 504 // Recompiling without boxing elimination 505 tty->print_cr("*********************************************************"); 506 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 507 tty->print_cr("*********************************************************"); 508 } 509 if (C->directive()->BreakAtCompileOption) { 510 // Open the debugger when compiling this method. 511 tty->print("### Breaking when compiling: "); 512 method()->print_short_name(); 513 tty->cr(); 514 BREAKPOINT; 515 } 516 517 if( PrintOpto ) { 518 if (is_osr_compilation()) { 519 tty->print("[OSR]%3d", _compile_id); 520 } else { 521 tty->print("%3d", _compile_id); 522 } 523 } 524 #endif 525 } 526 527 528 //-----------------------init_scratch_buffer_blob------------------------------ 529 // Construct a temporary BufferBlob and cache it for this compile. 530 void Compile::init_scratch_buffer_blob(int const_size) { 531 // If there is already a scratch buffer blob allocated and the 532 // constant section is big enough, use it. Otherwise free the 533 // current and allocate a new one. 534 BufferBlob* blob = scratch_buffer_blob(); 535 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 536 // Use the current blob. 537 } else { 538 if (blob != NULL) { 539 BufferBlob::free(blob); 540 } 541 542 ResourceMark rm; 543 _scratch_const_size = const_size; 544 int size = C2Compiler::initial_code_buffer_size(const_size); 545 blob = BufferBlob::create("Compile::scratch_buffer", size); 546 // Record the buffer blob for next time. 547 set_scratch_buffer_blob(blob); 548 // Have we run out of code space? 549 if (scratch_buffer_blob() == NULL) { 550 // Let CompilerBroker disable further compilations. 551 record_failure("Not enough space for scratch buffer in CodeCache"); 552 return; 553 } 554 } 555 556 // Initialize the relocation buffers 557 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 558 set_scratch_locs_memory(locs_buf); 559 } 560 561 562 //-----------------------scratch_emit_size------------------------------------- 563 // Helper function that computes size by emitting code 564 uint Compile::scratch_emit_size(const Node* n) { 565 // Start scratch_emit_size section. 566 set_in_scratch_emit_size(true); 567 568 // Emit into a trash buffer and count bytes emitted. 569 // This is a pretty expensive way to compute a size, 570 // but it works well enough if seldom used. 571 // All common fixed-size instructions are given a size 572 // method by the AD file. 573 // Note that the scratch buffer blob and locs memory are 574 // allocated at the beginning of the compile task, and 575 // may be shared by several calls to scratch_emit_size. 576 // The allocation of the scratch buffer blob is particularly 577 // expensive, since it has to grab the code cache lock. 578 BufferBlob* blob = this->scratch_buffer_blob(); 579 assert(blob != NULL, "Initialize BufferBlob at start"); 580 assert(blob->size() > MAX_inst_size, "sanity"); 581 relocInfo* locs_buf = scratch_locs_memory(); 582 address blob_begin = blob->content_begin(); 583 address blob_end = (address)locs_buf; 584 assert(blob->contains(blob_end), "sanity"); 585 CodeBuffer buf(blob_begin, blob_end - blob_begin); 586 buf.initialize_consts_size(_scratch_const_size); 587 buf.initialize_stubs_size(MAX_stubs_size); 588 assert(locs_buf != NULL, "sanity"); 589 int lsize = MAX_locs_size / 3; 590 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 591 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 592 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 593 // Mark as scratch buffer. 594 buf.consts()->set_scratch_emit(); 595 buf.insts()->set_scratch_emit(); 596 buf.stubs()->set_scratch_emit(); 597 598 // Do the emission. 599 600 Label fakeL; // Fake label for branch instructions. 601 Label* saveL = NULL; 602 uint save_bnum = 0; 603 bool is_branch = n->is_MachBranch(); 604 if (is_branch) { 605 MacroAssembler masm(&buf); 606 masm.bind(fakeL); 607 n->as_MachBranch()->save_label(&saveL, &save_bnum); 608 n->as_MachBranch()->label_set(&fakeL, 0); 609 } 610 n->emit(buf, this->regalloc()); 611 612 // Emitting into the scratch buffer should not fail 613 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 614 615 if (is_branch) // Restore label. 616 n->as_MachBranch()->label_set(saveL, save_bnum); 617 618 // End scratch_emit_size section. 619 set_in_scratch_emit_size(false); 620 621 return buf.insts_size(); 622 } 623 624 625 // ============================================================================ 626 //------------------------------Compile standard------------------------------- 627 debug_only( int Compile::_debug_idx = 100000; ) 628 629 // Compile a method. entry_bci is -1 for normal compilations and indicates 630 // the continuation bci for on stack replacement. 631 632 633 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 634 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 635 : Phase(Compiler), 636 _compile_id(ci_env->compile_id()), 637 _save_argument_registers(false), 638 _subsume_loads(subsume_loads), 639 _do_escape_analysis(do_escape_analysis), 640 _eliminate_boxing(eliminate_boxing), 641 _method(target), 642 _entry_bci(osr_bci), 643 _stub_function(NULL), 644 _stub_name(NULL), 645 _stub_entry_point(NULL), 646 _max_node_limit(MaxNodeLimit), 647 _orig_pc_slot(0), 648 _orig_pc_slot_offset_in_bytes(0), 649 _inlining_progress(false), 650 _inlining_incrementally(false), 651 _do_cleanup(false), 652 _has_reserved_stack_access(target->has_reserved_stack_access()), 653 #ifndef PRODUCT 654 _trace_opto_output(directive->TraceOptoOutputOption), 655 #endif 656 _has_method_handle_invokes(false), 657 _comp_arena(mtCompiler), 658 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 659 _env(ci_env), 660 _directive(directive), 661 _log(ci_env->log()), 662 _failure_reason(NULL), 663 _congraph(NULL), 664 #ifndef PRODUCT 665 _printer(IdealGraphPrinter::printer()), 666 #endif 667 _dead_node_list(comp_arena()), 668 _dead_node_count(0), 669 _node_arena(mtCompiler), 670 _old_arena(mtCompiler), 671 _mach_constant_base_node(NULL), 672 _Compile_types(mtCompiler), 673 _initial_gvn(NULL), 674 _for_igvn(NULL), 675 _warm_calls(NULL), 676 _late_inlines(comp_arena(), 2, 0, NULL), 677 _string_late_inlines(comp_arena(), 2, 0, NULL), 678 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 679 _late_inlines_pos(0), 680 _number_of_mh_late_inlines(0), 681 _print_inlining_stream(NULL), 682 _print_inlining_list(NULL), 683 _print_inlining_idx(0), 684 _print_inlining_output(NULL), 685 _replay_inline_data(NULL), 686 _java_calls(0), 687 _inner_loops(0), 688 _interpreter_frame_size(0), 689 _node_bundling_limit(0), 690 _node_bundling_base(NULL), 691 _code_buffer("Compile::Fill_buffer"), 692 _scratch_const_size(-1), 693 _in_scratch_emit_size(false) 694 #ifndef PRODUCT 695 , _in_dump_cnt(0) 696 #endif 697 { 698 C = this; 699 #ifndef PRODUCT 700 if (_printer != NULL) { 701 _printer->set_compile(this); 702 } 703 #endif 704 CompileWrapper cw(this); 705 706 if (CITimeVerbose) { 707 tty->print(" "); 708 target->holder()->name()->print(); 709 tty->print("."); 710 target->print_short_name(); 711 tty->print(" "); 712 } 713 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 714 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 715 716 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY) 717 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 718 // We can always print a disassembly, either abstract (hex dump) or 719 // with the help of a suitable hsdis library. Thus, we should not 720 // couple print_assembly and print_opto_assembly controls. 721 // But: always print opto and regular assembly on compile command 'print'. 722 bool print_assembly = directive->PrintAssemblyOption; 723 set_print_assembly(print_opto_assembly || print_assembly); 724 #else 725 set_print_assembly(false); // must initialize. 726 #endif 727 728 #ifndef PRODUCT 729 set_parsed_irreducible_loop(false); 730 731 if (directive->ReplayInlineOption) { 732 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 733 } 734 #endif 735 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 736 set_print_intrinsics(directive->PrintIntrinsicsOption); 737 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 738 739 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 740 // Make sure the method being compiled gets its own MDO, 741 // so we can at least track the decompile_count(). 742 // Need MDO to record RTM code generation state. 743 method()->ensure_method_data(); 744 } 745 746 Init(::AliasLevel); 747 748 749 print_compile_messages(); 750 751 _ilt = InlineTree::build_inline_tree_root(); 752 753 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 754 assert(num_alias_types() >= AliasIdxRaw, ""); 755 756 #define MINIMUM_NODE_HASH 1023 757 // Node list that Iterative GVN will start with 758 Unique_Node_List for_igvn(comp_arena()); 759 set_for_igvn(&for_igvn); 760 761 // GVN that will be run immediately on new nodes 762 uint estimated_size = method()->code_size()*4+64; 763 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 764 PhaseGVN gvn(node_arena(), estimated_size); 765 set_initial_gvn(&gvn); 766 767 print_inlining_init(); 768 { // Scope for timing the parser 769 TracePhase tp("parse", &timers[_t_parser]); 770 771 // Put top into the hash table ASAP. 772 initial_gvn()->transform_no_reclaim(top()); 773 774 // Set up tf(), start(), and find a CallGenerator. 775 CallGenerator* cg = NULL; 776 if (is_osr_compilation()) { 777 const TypeTuple *domain = StartOSRNode::osr_domain(); 778 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 779 init_tf(TypeFunc::make(domain, range)); 780 StartNode* s = new StartOSRNode(root(), domain); 781 initial_gvn()->set_type_bottom(s); 782 init_start(s); 783 cg = CallGenerator::for_osr(method(), entry_bci()); 784 } else { 785 // Normal case. 786 init_tf(TypeFunc::make(method())); 787 StartNode* s = new StartNode(root(), tf()->domain()); 788 initial_gvn()->set_type_bottom(s); 789 init_start(s); 790 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 791 // With java.lang.ref.reference.get() we must go through the 792 // intrinsic - even when get() is the root 793 // method of the compile - so that, if necessary, the value in 794 // the referent field of the reference object gets recorded by 795 // the pre-barrier code. 796 cg = find_intrinsic(method(), false); 797 } 798 if (cg == NULL) { 799 float past_uses = method()->interpreter_invocation_count(); 800 float expected_uses = past_uses; 801 cg = CallGenerator::for_inline(method(), expected_uses); 802 } 803 } 804 if (failing()) return; 805 if (cg == NULL) { 806 record_method_not_compilable("cannot parse method"); 807 return; 808 } 809 JVMState* jvms = build_start_state(start(), tf()); 810 if ((jvms = cg->generate(jvms)) == NULL) { 811 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 812 record_method_not_compilable("method parse failed"); 813 } 814 return; 815 } 816 GraphKit kit(jvms); 817 818 if (!kit.stopped()) { 819 // Accept return values, and transfer control we know not where. 820 // This is done by a special, unique ReturnNode bound to root. 821 return_values(kit.jvms()); 822 } 823 824 if (kit.has_exceptions()) { 825 // Any exceptions that escape from this call must be rethrown 826 // to whatever caller is dynamically above us on the stack. 827 // This is done by a special, unique RethrowNode bound to root. 828 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 829 } 830 831 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 832 833 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 834 inline_string_calls(true); 835 } 836 837 if (failing()) return; 838 839 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 840 841 // Remove clutter produced by parsing. 842 if (!failing()) { 843 ResourceMark rm; 844 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 845 } 846 } 847 848 // Note: Large methods are capped off in do_one_bytecode(). 849 if (failing()) return; 850 851 // After parsing, node notes are no longer automagic. 852 // They must be propagated by register_new_node_with_optimizer(), 853 // clone(), or the like. 854 set_default_node_notes(NULL); 855 856 for (;;) { 857 int successes = Inline_Warm(); 858 if (failing()) return; 859 if (successes == 0) break; 860 } 861 862 // Drain the list. 863 Finish_Warm(); 864 #ifndef PRODUCT 865 if (_printer && _printer->should_print(1)) { 866 _printer->print_inlining(); 867 } 868 #endif 869 870 if (failing()) return; 871 NOT_PRODUCT( verify_graph_edges(); ) 872 873 // Now optimize 874 Optimize(); 875 if (failing()) return; 876 NOT_PRODUCT( verify_graph_edges(); ) 877 878 #ifndef PRODUCT 879 if (PrintIdeal) { 880 ttyLocker ttyl; // keep the following output all in one block 881 // This output goes directly to the tty, not the compiler log. 882 // To enable tools to match it up with the compilation activity, 883 // be sure to tag this tty output with the compile ID. 884 if (xtty != NULL) { 885 xtty->head("ideal compile_id='%d'%s", compile_id(), 886 is_osr_compilation() ? " compile_kind='osr'" : 887 ""); 888 } 889 root()->dump(9999); 890 if (xtty != NULL) { 891 xtty->tail("ideal"); 892 } 893 } 894 #endif 895 896 #ifdef ASSERT 897 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 898 bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen); 899 #endif 900 901 // Dump compilation data to replay it. 902 if (directive->DumpReplayOption) { 903 env()->dump_replay_data(_compile_id); 904 } 905 if (directive->DumpInlineOption && (ilt() != NULL)) { 906 env()->dump_inline_data(_compile_id); 907 } 908 909 // Now that we know the size of all the monitors we can add a fixed slot 910 // for the original deopt pc. 911 912 _orig_pc_slot = fixed_slots(); 913 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 914 set_fixed_slots(next_slot); 915 916 // Compute when to use implicit null checks. Used by matching trap based 917 // nodes and NullCheck optimization. 918 set_allowed_deopt_reasons(); 919 920 // Now generate code 921 Code_Gen(); 922 if (failing()) return; 923 924 // Check if we want to skip execution of all compiled code. 925 { 926 #ifndef PRODUCT 927 if (OptoNoExecute) { 928 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 929 return; 930 } 931 #endif 932 TracePhase tp("install_code", &timers[_t_registerMethod]); 933 934 if (is_osr_compilation()) { 935 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 936 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 937 } else { 938 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 939 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 940 } 941 942 env()->register_method(_method, _entry_bci, 943 &_code_offsets, 944 _orig_pc_slot_offset_in_bytes, 945 code_buffer(), 946 frame_size_in_words(), _oop_map_set, 947 &_handler_table, &_inc_table, 948 compiler, 949 has_unsafe_access(), 950 SharedRuntime::is_wide_vector(max_vector_size()), 951 rtm_state() 952 ); 953 954 if (log() != NULL) // Print code cache state into compiler log 955 log()->code_cache_state(); 956 } 957 } 958 959 //------------------------------Compile---------------------------------------- 960 // Compile a runtime stub 961 Compile::Compile( ciEnv* ci_env, 962 TypeFunc_generator generator, 963 address stub_function, 964 const char *stub_name, 965 int is_fancy_jump, 966 bool pass_tls, 967 bool save_arg_registers, 968 bool return_pc, 969 DirectiveSet* directive) 970 : Phase(Compiler), 971 _compile_id(0), 972 _save_argument_registers(save_arg_registers), 973 _subsume_loads(true), 974 _do_escape_analysis(false), 975 _eliminate_boxing(false), 976 _method(NULL), 977 _entry_bci(InvocationEntryBci), 978 _stub_function(stub_function), 979 _stub_name(stub_name), 980 _stub_entry_point(NULL), 981 _max_node_limit(MaxNodeLimit), 982 _orig_pc_slot(0), 983 _orig_pc_slot_offset_in_bytes(0), 984 _inlining_progress(false), 985 _inlining_incrementally(false), 986 _has_reserved_stack_access(false), 987 #ifndef PRODUCT 988 _trace_opto_output(directive->TraceOptoOutputOption), 989 #endif 990 _has_method_handle_invokes(false), 991 _comp_arena(mtCompiler), 992 _env(ci_env), 993 _directive(directive), 994 _log(ci_env->log()), 995 _failure_reason(NULL), 996 _congraph(NULL), 997 #ifndef PRODUCT 998 _printer(NULL), 999 #endif 1000 _dead_node_list(comp_arena()), 1001 _dead_node_count(0), 1002 _node_arena(mtCompiler), 1003 _old_arena(mtCompiler), 1004 _mach_constant_base_node(NULL), 1005 _Compile_types(mtCompiler), 1006 _initial_gvn(NULL), 1007 _for_igvn(NULL), 1008 _warm_calls(NULL), 1009 _number_of_mh_late_inlines(0), 1010 _print_inlining_stream(NULL), 1011 _print_inlining_list(NULL), 1012 _print_inlining_idx(0), 1013 _print_inlining_output(NULL), 1014 _replay_inline_data(NULL), 1015 _java_calls(0), 1016 _inner_loops(0), 1017 _interpreter_frame_size(0), 1018 _node_bundling_limit(0), 1019 _node_bundling_base(NULL), 1020 _code_buffer("Compile::Fill_buffer"), 1021 #ifndef PRODUCT 1022 _in_dump_cnt(0), 1023 #endif 1024 _allowed_reasons(0) { 1025 C = this; 1026 1027 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1028 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1029 1030 #ifndef PRODUCT 1031 set_print_assembly(PrintFrameConverterAssembly); 1032 set_parsed_irreducible_loop(false); 1033 #else 1034 set_print_assembly(false); // Must initialize. 1035 #endif 1036 set_has_irreducible_loop(false); // no loops 1037 1038 CompileWrapper cw(this); 1039 Init(/*AliasLevel=*/ 0); 1040 init_tf((*generator)()); 1041 1042 { 1043 // The following is a dummy for the sake of GraphKit::gen_stub 1044 Unique_Node_List for_igvn(comp_arena()); 1045 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1046 PhaseGVN gvn(Thread::current()->resource_area(),255); 1047 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1048 gvn.transform_no_reclaim(top()); 1049 1050 GraphKit kit; 1051 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1052 } 1053 1054 NOT_PRODUCT( verify_graph_edges(); ) 1055 Code_Gen(); 1056 if (failing()) return; 1057 1058 1059 // Entry point will be accessed using compile->stub_entry_point(); 1060 if (code_buffer() == NULL) { 1061 Matcher::soft_match_failure(); 1062 } else { 1063 if (PrintAssembly && (WizardMode || Verbose)) 1064 tty->print_cr("### Stub::%s", stub_name); 1065 1066 if (!failing()) { 1067 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1068 1069 // Make the NMethod 1070 // For now we mark the frame as never safe for profile stackwalking 1071 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1072 code_buffer(), 1073 CodeOffsets::frame_never_safe, 1074 // _code_offsets.value(CodeOffsets::Frame_Complete), 1075 frame_size_in_words(), 1076 _oop_map_set, 1077 save_arg_registers); 1078 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1079 1080 _stub_entry_point = rs->entry_point(); 1081 } 1082 } 1083 } 1084 1085 //------------------------------Init------------------------------------------- 1086 // Prepare for a single compilation 1087 void Compile::Init(int aliaslevel) { 1088 _unique = 0; 1089 _regalloc = NULL; 1090 1091 _tf = NULL; // filled in later 1092 _top = NULL; // cached later 1093 _matcher = NULL; // filled in later 1094 _cfg = NULL; // filled in later 1095 1096 set_24_bit_selection_and_mode(Use24BitFP, false); 1097 1098 _node_note_array = NULL; 1099 _default_node_notes = NULL; 1100 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1101 1102 _immutable_memory = NULL; // filled in at first inquiry 1103 1104 // Globally visible Nodes 1105 // First set TOP to NULL to give safe behavior during creation of RootNode 1106 set_cached_top_node(NULL); 1107 set_root(new RootNode()); 1108 // Now that you have a Root to point to, create the real TOP 1109 set_cached_top_node( new ConNode(Type::TOP) ); 1110 set_recent_alloc(NULL, NULL); 1111 1112 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1113 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1114 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1115 env()->set_dependencies(new Dependencies(env())); 1116 1117 _fixed_slots = 0; 1118 set_has_split_ifs(false); 1119 set_has_loops(has_method() && method()->has_loops()); // first approximation 1120 set_has_stringbuilder(false); 1121 set_has_boxed_value(false); 1122 _trap_can_recompile = false; // no traps emitted yet 1123 _major_progress = true; // start out assuming good things will happen 1124 set_has_unsafe_access(false); 1125 set_max_vector_size(0); 1126 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1127 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1128 set_decompile_count(0); 1129 1130 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1131 _loop_opts_cnt = LoopOptsCount; 1132 set_do_inlining(Inline); 1133 set_max_inline_size(MaxInlineSize); 1134 set_freq_inline_size(FreqInlineSize); 1135 set_do_scheduling(OptoScheduling); 1136 set_do_count_invocations(false); 1137 set_do_method_data_update(false); 1138 1139 set_do_vector_loop(false); 1140 1141 if (AllowVectorizeOnDemand) { 1142 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1143 set_do_vector_loop(true); 1144 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());}) 1145 } else if (has_method() && method()->name() != 0 && 1146 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1147 set_do_vector_loop(true); 1148 } 1149 } 1150 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1151 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());}) 1152 1153 set_age_code(has_method() && method()->profile_aging()); 1154 set_rtm_state(NoRTM); // No RTM lock eliding by default 1155 _max_node_limit = _directive->MaxNodeLimitOption; 1156 1157 #if INCLUDE_RTM_OPT 1158 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1159 int rtm_state = method()->method_data()->rtm_state(); 1160 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1161 // Don't generate RTM lock eliding code. 1162 set_rtm_state(NoRTM); 1163 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1164 // Generate RTM lock eliding code without abort ratio calculation code. 1165 set_rtm_state(UseRTM); 1166 } else if (UseRTMDeopt) { 1167 // Generate RTM lock eliding code and include abort ratio calculation 1168 // code if UseRTMDeopt is on. 1169 set_rtm_state(ProfileRTM); 1170 } 1171 } 1172 #endif 1173 if (debug_info()->recording_non_safepoints()) { 1174 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1175 (comp_arena(), 8, 0, NULL)); 1176 set_default_node_notes(Node_Notes::make(this)); 1177 } 1178 1179 // // -- Initialize types before each compile -- 1180 // // Update cached type information 1181 // if( _method && _method->constants() ) 1182 // Type::update_loaded_types(_method, _method->constants()); 1183 1184 // Init alias_type map. 1185 if (!_do_escape_analysis && aliaslevel == 3) 1186 aliaslevel = 2; // No unique types without escape analysis 1187 _AliasLevel = aliaslevel; 1188 const int grow_ats = 16; 1189 _max_alias_types = grow_ats; 1190 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1191 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1192 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1193 { 1194 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1195 } 1196 // Initialize the first few types. 1197 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1198 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1199 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1200 _num_alias_types = AliasIdxRaw+1; 1201 // Zero out the alias type cache. 1202 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1203 // A NULL adr_type hits in the cache right away. Preload the right answer. 1204 probe_alias_cache(NULL)->_index = AliasIdxTop; 1205 1206 _intrinsics = NULL; 1207 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1208 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1209 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1210 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1211 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1212 register_library_intrinsics(); 1213 } 1214 1215 //---------------------------init_start---------------------------------------- 1216 // Install the StartNode on this compile object. 1217 void Compile::init_start(StartNode* s) { 1218 if (failing()) 1219 return; // already failing 1220 assert(s == start(), ""); 1221 } 1222 1223 /** 1224 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1225 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1226 * the ideal graph. 1227 */ 1228 StartNode* Compile::start() const { 1229 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1230 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1231 Node* start = root()->fast_out(i); 1232 if (start->is_Start()) { 1233 return start->as_Start(); 1234 } 1235 } 1236 fatal("Did not find Start node!"); 1237 return NULL; 1238 } 1239 1240 //-------------------------------immutable_memory------------------------------------- 1241 // Access immutable memory 1242 Node* Compile::immutable_memory() { 1243 if (_immutable_memory != NULL) { 1244 return _immutable_memory; 1245 } 1246 StartNode* s = start(); 1247 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1248 Node *p = s->fast_out(i); 1249 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1250 _immutable_memory = p; 1251 return _immutable_memory; 1252 } 1253 } 1254 ShouldNotReachHere(); 1255 return NULL; 1256 } 1257 1258 //----------------------set_cached_top_node------------------------------------ 1259 // Install the cached top node, and make sure Node::is_top works correctly. 1260 void Compile::set_cached_top_node(Node* tn) { 1261 if (tn != NULL) verify_top(tn); 1262 Node* old_top = _top; 1263 _top = tn; 1264 // Calling Node::setup_is_top allows the nodes the chance to adjust 1265 // their _out arrays. 1266 if (_top != NULL) _top->setup_is_top(); 1267 if (old_top != NULL) old_top->setup_is_top(); 1268 assert(_top == NULL || top()->is_top(), ""); 1269 } 1270 1271 #ifdef ASSERT 1272 uint Compile::count_live_nodes_by_graph_walk() { 1273 Unique_Node_List useful(comp_arena()); 1274 // Get useful node list by walking the graph. 1275 identify_useful_nodes(useful); 1276 return useful.size(); 1277 } 1278 1279 void Compile::print_missing_nodes() { 1280 1281 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1282 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1283 return; 1284 } 1285 1286 // This is an expensive function. It is executed only when the user 1287 // specifies VerifyIdealNodeCount option or otherwise knows the 1288 // additional work that needs to be done to identify reachable nodes 1289 // by walking the flow graph and find the missing ones using 1290 // _dead_node_list. 1291 1292 Unique_Node_List useful(comp_arena()); 1293 // Get useful node list by walking the graph. 1294 identify_useful_nodes(useful); 1295 1296 uint l_nodes = C->live_nodes(); 1297 uint l_nodes_by_walk = useful.size(); 1298 1299 if (l_nodes != l_nodes_by_walk) { 1300 if (_log != NULL) { 1301 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1302 _log->stamp(); 1303 _log->end_head(); 1304 } 1305 VectorSet& useful_member_set = useful.member_set(); 1306 int last_idx = l_nodes_by_walk; 1307 for (int i = 0; i < last_idx; i++) { 1308 if (useful_member_set.test(i)) { 1309 if (_dead_node_list.test(i)) { 1310 if (_log != NULL) { 1311 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1312 } 1313 if (PrintIdealNodeCount) { 1314 // Print the log message to tty 1315 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1316 useful.at(i)->dump(); 1317 } 1318 } 1319 } 1320 else if (! _dead_node_list.test(i)) { 1321 if (_log != NULL) { 1322 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1323 } 1324 if (PrintIdealNodeCount) { 1325 // Print the log message to tty 1326 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1327 } 1328 } 1329 } 1330 if (_log != NULL) { 1331 _log->tail("mismatched_nodes"); 1332 } 1333 } 1334 } 1335 void Compile::record_modified_node(Node* n) { 1336 if (_modified_nodes != NULL && !_inlining_incrementally && 1337 n->outcnt() != 0 && !n->is_Con()) { 1338 _modified_nodes->push(n); 1339 } 1340 } 1341 1342 void Compile::remove_modified_node(Node* n) { 1343 if (_modified_nodes != NULL) { 1344 _modified_nodes->remove(n); 1345 } 1346 } 1347 #endif 1348 1349 #ifndef PRODUCT 1350 void Compile::verify_top(Node* tn) const { 1351 if (tn != NULL) { 1352 assert(tn->is_Con(), "top node must be a constant"); 1353 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1354 assert(tn->in(0) != NULL, "must have live top node"); 1355 } 1356 } 1357 #endif 1358 1359 1360 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1361 1362 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1363 guarantee(arr != NULL, ""); 1364 int num_blocks = arr->length(); 1365 if (grow_by < num_blocks) grow_by = num_blocks; 1366 int num_notes = grow_by * _node_notes_block_size; 1367 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1368 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1369 while (num_notes > 0) { 1370 arr->append(notes); 1371 notes += _node_notes_block_size; 1372 num_notes -= _node_notes_block_size; 1373 } 1374 assert(num_notes == 0, "exact multiple, please"); 1375 } 1376 1377 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1378 if (source == NULL || dest == NULL) return false; 1379 1380 if (dest->is_Con()) 1381 return false; // Do not push debug info onto constants. 1382 1383 #ifdef ASSERT 1384 // Leave a bread crumb trail pointing to the original node: 1385 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1386 dest->set_debug_orig(source); 1387 } 1388 #endif 1389 1390 if (node_note_array() == NULL) 1391 return false; // Not collecting any notes now. 1392 1393 // This is a copy onto a pre-existing node, which may already have notes. 1394 // If both nodes have notes, do not overwrite any pre-existing notes. 1395 Node_Notes* source_notes = node_notes_at(source->_idx); 1396 if (source_notes == NULL || source_notes->is_clear()) return false; 1397 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1398 if (dest_notes == NULL || dest_notes->is_clear()) { 1399 return set_node_notes_at(dest->_idx, source_notes); 1400 } 1401 1402 Node_Notes merged_notes = (*source_notes); 1403 // The order of operations here ensures that dest notes will win... 1404 merged_notes.update_from(dest_notes); 1405 return set_node_notes_at(dest->_idx, &merged_notes); 1406 } 1407 1408 1409 //--------------------------allow_range_check_smearing------------------------- 1410 // Gating condition for coalescing similar range checks. 1411 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1412 // single covering check that is at least as strong as any of them. 1413 // If the optimization succeeds, the simplified (strengthened) range check 1414 // will always succeed. If it fails, we will deopt, and then give up 1415 // on the optimization. 1416 bool Compile::allow_range_check_smearing() const { 1417 // If this method has already thrown a range-check, 1418 // assume it was because we already tried range smearing 1419 // and it failed. 1420 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1421 return !already_trapped; 1422 } 1423 1424 1425 //------------------------------flatten_alias_type----------------------------- 1426 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1427 int offset = tj->offset(); 1428 TypePtr::PTR ptr = tj->ptr(); 1429 1430 // Known instance (scalarizable allocation) alias only with itself. 1431 bool is_known_inst = tj->isa_oopptr() != NULL && 1432 tj->is_oopptr()->is_known_instance(); 1433 1434 // Process weird unsafe references. 1435 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1436 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1437 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1438 tj = TypeOopPtr::BOTTOM; 1439 ptr = tj->ptr(); 1440 offset = tj->offset(); 1441 } 1442 1443 // Array pointers need some flattening 1444 const TypeAryPtr *ta = tj->isa_aryptr(); 1445 if (ta && ta->is_stable()) { 1446 // Erase stability property for alias analysis. 1447 tj = ta = ta->cast_to_stable(false); 1448 } 1449 if( ta && is_known_inst ) { 1450 if ( offset != Type::OffsetBot && 1451 offset > arrayOopDesc::length_offset_in_bytes() ) { 1452 offset = Type::OffsetBot; // Flatten constant access into array body only 1453 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1454 } 1455 } else if( ta && _AliasLevel >= 2 ) { 1456 // For arrays indexed by constant indices, we flatten the alias 1457 // space to include all of the array body. Only the header, klass 1458 // and array length can be accessed un-aliased. 1459 if( offset != Type::OffsetBot ) { 1460 if( ta->const_oop() ) { // MethodData* or Method* 1461 offset = Type::OffsetBot; // Flatten constant access into array body 1462 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1463 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1464 // range is OK as-is. 1465 tj = ta = TypeAryPtr::RANGE; 1466 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1467 tj = TypeInstPtr::KLASS; // all klass loads look alike 1468 ta = TypeAryPtr::RANGE; // generic ignored junk 1469 ptr = TypePtr::BotPTR; 1470 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1471 tj = TypeInstPtr::MARK; 1472 ta = TypeAryPtr::RANGE; // generic ignored junk 1473 ptr = TypePtr::BotPTR; 1474 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1475 ta = tj->isa_aryptr(); 1476 } else { // Random constant offset into array body 1477 offset = Type::OffsetBot; // Flatten constant access into array body 1478 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1479 } 1480 } 1481 // Arrays of fixed size alias with arrays of unknown size. 1482 if (ta->size() != TypeInt::POS) { 1483 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1484 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1485 } 1486 // Arrays of known objects become arrays of unknown objects. 1487 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1488 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1489 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1490 } 1491 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1492 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1493 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1494 } 1495 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1496 // cannot be distinguished by bytecode alone. 1497 if (ta->elem() == TypeInt::BOOL) { 1498 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1499 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1500 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1501 } 1502 // During the 2nd round of IterGVN, NotNull castings are removed. 1503 // Make sure the Bottom and NotNull variants alias the same. 1504 // Also, make sure exact and non-exact variants alias the same. 1505 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1506 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1507 } 1508 } 1509 1510 // Oop pointers need some flattening 1511 const TypeInstPtr *to = tj->isa_instptr(); 1512 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1513 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1514 if( ptr == TypePtr::Constant ) { 1515 if (to->klass() != ciEnv::current()->Class_klass() || 1516 offset < k->size_helper() * wordSize) { 1517 // No constant oop pointers (such as Strings); they alias with 1518 // unknown strings. 1519 assert(!is_known_inst, "not scalarizable allocation"); 1520 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1521 } 1522 } else if( is_known_inst ) { 1523 tj = to; // Keep NotNull and klass_is_exact for instance type 1524 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1525 // During the 2nd round of IterGVN, NotNull castings are removed. 1526 // Make sure the Bottom and NotNull variants alias the same. 1527 // Also, make sure exact and non-exact variants alias the same. 1528 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1529 } 1530 if (to->speculative() != NULL) { 1531 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1532 } 1533 // Canonicalize the holder of this field 1534 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1535 // First handle header references such as a LoadKlassNode, even if the 1536 // object's klass is unloaded at compile time (4965979). 1537 if (!is_known_inst) { // Do it only for non-instance types 1538 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1539 } 1540 } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) { 1541 to = tj->is_instptr(); 1542 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1543 // Static fields are in the space above the normal instance 1544 // fields in the java.lang.Class instance. 1545 if (to->klass() != ciEnv::current()->Class_klass()) { 1546 to = NULL; 1547 tj = TypeOopPtr::BOTTOM; 1548 offset = tj->offset(); 1549 } 1550 } else { 1551 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1552 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1553 if( is_known_inst ) { 1554 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1555 } else { 1556 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1557 } 1558 } 1559 } 1560 } 1561 1562 // Klass pointers to object array klasses need some flattening 1563 const TypeKlassPtr *tk = tj->isa_klassptr(); 1564 if( tk ) { 1565 // If we are referencing a field within a Klass, we need 1566 // to assume the worst case of an Object. Both exact and 1567 // inexact types must flatten to the same alias class so 1568 // use NotNull as the PTR. 1569 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1570 1571 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1572 TypeKlassPtr::OBJECT->klass(), 1573 offset); 1574 } 1575 1576 ciKlass* klass = tk->klass(); 1577 if( klass->is_obj_array_klass() ) { 1578 ciKlass* k = TypeAryPtr::OOPS->klass(); 1579 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1580 k = TypeInstPtr::BOTTOM->klass(); 1581 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1582 } 1583 1584 // Check for precise loads from the primary supertype array and force them 1585 // to the supertype cache alias index. Check for generic array loads from 1586 // the primary supertype array and also force them to the supertype cache 1587 // alias index. Since the same load can reach both, we need to merge 1588 // these 2 disparate memories into the same alias class. Since the 1589 // primary supertype array is read-only, there's no chance of confusion 1590 // where we bypass an array load and an array store. 1591 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1592 if (offset == Type::OffsetBot || 1593 (offset >= primary_supers_offset && 1594 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1595 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1596 offset = in_bytes(Klass::secondary_super_cache_offset()); 1597 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1598 } 1599 } 1600 1601 // Flatten all Raw pointers together. 1602 if (tj->base() == Type::RawPtr) 1603 tj = TypeRawPtr::BOTTOM; 1604 1605 if (tj->base() == Type::AnyPtr) 1606 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1607 1608 // Flatten all to bottom for now 1609 switch( _AliasLevel ) { 1610 case 0: 1611 tj = TypePtr::BOTTOM; 1612 break; 1613 case 1: // Flatten to: oop, static, field or array 1614 switch (tj->base()) { 1615 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1616 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1617 case Type::AryPtr: // do not distinguish arrays at all 1618 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1619 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1620 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1621 default: ShouldNotReachHere(); 1622 } 1623 break; 1624 case 2: // No collapsing at level 2; keep all splits 1625 case 3: // No collapsing at level 3; keep all splits 1626 break; 1627 default: 1628 Unimplemented(); 1629 } 1630 1631 offset = tj->offset(); 1632 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1633 1634 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1635 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1636 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1637 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1638 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1639 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1640 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1641 (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)), 1642 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1643 assert( tj->ptr() != TypePtr::TopPTR && 1644 tj->ptr() != TypePtr::AnyNull && 1645 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1646 // assert( tj->ptr() != TypePtr::Constant || 1647 // tj->base() == Type::RawPtr || 1648 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1649 1650 return tj; 1651 } 1652 1653 void Compile::AliasType::Init(int i, const TypePtr* at) { 1654 _index = i; 1655 _adr_type = at; 1656 _field = NULL; 1657 _element = NULL; 1658 _is_rewritable = true; // default 1659 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1660 if (atoop != NULL && atoop->is_known_instance()) { 1661 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1662 _general_index = Compile::current()->get_alias_index(gt); 1663 } else { 1664 _general_index = 0; 1665 } 1666 } 1667 1668 BasicType Compile::AliasType::basic_type() const { 1669 if (element() != NULL) { 1670 const Type* element = adr_type()->is_aryptr()->elem(); 1671 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1672 } if (field() != NULL) { 1673 return field()->layout_type(); 1674 } else { 1675 return T_ILLEGAL; // unknown 1676 } 1677 } 1678 1679 //---------------------------------print_on------------------------------------ 1680 #ifndef PRODUCT 1681 void Compile::AliasType::print_on(outputStream* st) { 1682 if (index() < 10) 1683 st->print("@ <%d> ", index()); 1684 else st->print("@ <%d>", index()); 1685 st->print(is_rewritable() ? " " : " RO"); 1686 int offset = adr_type()->offset(); 1687 if (offset == Type::OffsetBot) 1688 st->print(" +any"); 1689 else st->print(" +%-3d", offset); 1690 st->print(" in "); 1691 adr_type()->dump_on(st); 1692 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1693 if (field() != NULL && tjp) { 1694 if (tjp->klass() != field()->holder() || 1695 tjp->offset() != field()->offset_in_bytes()) { 1696 st->print(" != "); 1697 field()->print(); 1698 st->print(" ***"); 1699 } 1700 } 1701 } 1702 1703 void print_alias_types() { 1704 Compile* C = Compile::current(); 1705 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1706 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1707 C->alias_type(idx)->print_on(tty); 1708 tty->cr(); 1709 } 1710 } 1711 #endif 1712 1713 1714 //----------------------------probe_alias_cache-------------------------------- 1715 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1716 intptr_t key = (intptr_t) adr_type; 1717 key ^= key >> logAliasCacheSize; 1718 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1719 } 1720 1721 1722 //-----------------------------grow_alias_types-------------------------------- 1723 void Compile::grow_alias_types() { 1724 const int old_ats = _max_alias_types; // how many before? 1725 const int new_ats = old_ats; // how many more? 1726 const int grow_ats = old_ats+new_ats; // how many now? 1727 _max_alias_types = grow_ats; 1728 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1729 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1730 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1731 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1732 } 1733 1734 1735 //--------------------------------find_alias_type------------------------------ 1736 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1737 if (_AliasLevel == 0) 1738 return alias_type(AliasIdxBot); 1739 1740 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1741 if (ace->_adr_type == adr_type) { 1742 return alias_type(ace->_index); 1743 } 1744 1745 // Handle special cases. 1746 if (adr_type == NULL) return alias_type(AliasIdxTop); 1747 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1748 1749 // Do it the slow way. 1750 const TypePtr* flat = flatten_alias_type(adr_type); 1751 1752 #ifdef ASSERT 1753 { 1754 ResourceMark rm; 1755 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1756 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1757 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1758 Type::str(adr_type)); 1759 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1760 const TypeOopPtr* foop = flat->is_oopptr(); 1761 // Scalarizable allocations have exact klass always. 1762 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1763 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1764 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1765 Type::str(foop), Type::str(xoop)); 1766 } 1767 } 1768 #endif 1769 1770 int idx = AliasIdxTop; 1771 for (int i = 0; i < num_alias_types(); i++) { 1772 if (alias_type(i)->adr_type() == flat) { 1773 idx = i; 1774 break; 1775 } 1776 } 1777 1778 if (idx == AliasIdxTop) { 1779 if (no_create) return NULL; 1780 // Grow the array if necessary. 1781 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1782 // Add a new alias type. 1783 idx = _num_alias_types++; 1784 _alias_types[idx]->Init(idx, flat); 1785 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1786 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1787 if (flat->isa_instptr()) { 1788 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1789 && flat->is_instptr()->klass() == env()->Class_klass()) 1790 alias_type(idx)->set_rewritable(false); 1791 } 1792 if (flat->isa_aryptr()) { 1793 #ifdef ASSERT 1794 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1795 // (T_BYTE has the weakest alignment and size restrictions...) 1796 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1797 #endif 1798 if (flat->offset() == TypePtr::OffsetBot) { 1799 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1800 } 1801 } 1802 if (flat->isa_klassptr()) { 1803 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1804 alias_type(idx)->set_rewritable(false); 1805 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1806 alias_type(idx)->set_rewritable(false); 1807 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1808 alias_type(idx)->set_rewritable(false); 1809 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1810 alias_type(idx)->set_rewritable(false); 1811 } 1812 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1813 // but the base pointer type is not distinctive enough to identify 1814 // references into JavaThread.) 1815 1816 // Check for final fields. 1817 const TypeInstPtr* tinst = flat->isa_instptr(); 1818 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1819 ciField* field; 1820 if (tinst->const_oop() != NULL && 1821 tinst->klass() == ciEnv::current()->Class_klass() && 1822 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1823 // static field 1824 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1825 field = k->get_field_by_offset(tinst->offset(), true); 1826 } else { 1827 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1828 field = k->get_field_by_offset(tinst->offset(), false); 1829 } 1830 assert(field == NULL || 1831 original_field == NULL || 1832 (field->holder() == original_field->holder() && 1833 field->offset() == original_field->offset() && 1834 field->is_static() == original_field->is_static()), "wrong field?"); 1835 // Set field() and is_rewritable() attributes. 1836 if (field != NULL) alias_type(idx)->set_field(field); 1837 } 1838 } 1839 1840 // Fill the cache for next time. 1841 ace->_adr_type = adr_type; 1842 ace->_index = idx; 1843 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1844 1845 // Might as well try to fill the cache for the flattened version, too. 1846 AliasCacheEntry* face = probe_alias_cache(flat); 1847 if (face->_adr_type == NULL) { 1848 face->_adr_type = flat; 1849 face->_index = idx; 1850 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1851 } 1852 1853 return alias_type(idx); 1854 } 1855 1856 1857 Compile::AliasType* Compile::alias_type(ciField* field) { 1858 const TypeOopPtr* t; 1859 if (field->is_static()) 1860 t = TypeInstPtr::make(field->holder()->java_mirror()); 1861 else 1862 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1863 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1864 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1865 return atp; 1866 } 1867 1868 1869 //------------------------------have_alias_type-------------------------------- 1870 bool Compile::have_alias_type(const TypePtr* adr_type) { 1871 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1872 if (ace->_adr_type == adr_type) { 1873 return true; 1874 } 1875 1876 // Handle special cases. 1877 if (adr_type == NULL) return true; 1878 if (adr_type == TypePtr::BOTTOM) return true; 1879 1880 return find_alias_type(adr_type, true, NULL) != NULL; 1881 } 1882 1883 //-----------------------------must_alias-------------------------------------- 1884 // True if all values of the given address type are in the given alias category. 1885 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1886 if (alias_idx == AliasIdxBot) return true; // the universal category 1887 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1888 if (alias_idx == AliasIdxTop) return false; // the empty category 1889 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1890 1891 // the only remaining possible overlap is identity 1892 int adr_idx = get_alias_index(adr_type); 1893 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1894 assert(adr_idx == alias_idx || 1895 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1896 && adr_type != TypeOopPtr::BOTTOM), 1897 "should not be testing for overlap with an unsafe pointer"); 1898 return adr_idx == alias_idx; 1899 } 1900 1901 //------------------------------can_alias-------------------------------------- 1902 // True if any values of the given address type are in the given alias category. 1903 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1904 if (alias_idx == AliasIdxTop) return false; // the empty category 1905 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1906 if (alias_idx == AliasIdxBot) return true; // the universal category 1907 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1908 1909 // the only remaining possible overlap is identity 1910 int adr_idx = get_alias_index(adr_type); 1911 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1912 return adr_idx == alias_idx; 1913 } 1914 1915 1916 1917 //---------------------------pop_warm_call------------------------------------- 1918 WarmCallInfo* Compile::pop_warm_call() { 1919 WarmCallInfo* wci = _warm_calls; 1920 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1921 return wci; 1922 } 1923 1924 //----------------------------Inline_Warm-------------------------------------- 1925 int Compile::Inline_Warm() { 1926 // If there is room, try to inline some more warm call sites. 1927 // %%% Do a graph index compaction pass when we think we're out of space? 1928 if (!InlineWarmCalls) return 0; 1929 1930 int calls_made_hot = 0; 1931 int room_to_grow = NodeCountInliningCutoff - unique(); 1932 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1933 int amount_grown = 0; 1934 WarmCallInfo* call; 1935 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1936 int est_size = (int)call->size(); 1937 if (est_size > (room_to_grow - amount_grown)) { 1938 // This one won't fit anyway. Get rid of it. 1939 call->make_cold(); 1940 continue; 1941 } 1942 call->make_hot(); 1943 calls_made_hot++; 1944 amount_grown += est_size; 1945 amount_to_grow -= est_size; 1946 } 1947 1948 if (calls_made_hot > 0) set_major_progress(); 1949 return calls_made_hot; 1950 } 1951 1952 1953 //----------------------------Finish_Warm-------------------------------------- 1954 void Compile::Finish_Warm() { 1955 if (!InlineWarmCalls) return; 1956 if (failing()) return; 1957 if (warm_calls() == NULL) return; 1958 1959 // Clean up loose ends, if we are out of space for inlining. 1960 WarmCallInfo* call; 1961 while ((call = pop_warm_call()) != NULL) { 1962 call->make_cold(); 1963 } 1964 } 1965 1966 //---------------------cleanup_loop_predicates----------------------- 1967 // Remove the opaque nodes that protect the predicates so that all unused 1968 // checks and uncommon_traps will be eliminated from the ideal graph 1969 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1970 if (predicate_count()==0) return; 1971 for (int i = predicate_count(); i > 0; i--) { 1972 Node * n = predicate_opaque1_node(i-1); 1973 assert(n->Opcode() == Op_Opaque1, "must be"); 1974 igvn.replace_node(n, n->in(1)); 1975 } 1976 assert(predicate_count()==0, "should be clean!"); 1977 } 1978 1979 void Compile::add_range_check_cast(Node* n) { 1980 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1981 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1982 _range_check_casts->append(n); 1983 } 1984 1985 // Remove all range check dependent CastIINodes. 1986 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1987 for (int i = range_check_cast_count(); i > 0; i--) { 1988 Node* cast = range_check_cast_node(i-1); 1989 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1990 igvn.replace_node(cast, cast->in(1)); 1991 } 1992 assert(range_check_cast_count() == 0, "should be empty"); 1993 } 1994 1995 void Compile::add_opaque4_node(Node* n) { 1996 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 1997 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 1998 _opaque4_nodes->append(n); 1999 } 2000 2001 // Remove all Opaque4 nodes. 2002 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 2003 for (int i = opaque4_count(); i > 0; i--) { 2004 Node* opaq = opaque4_node(i-1); 2005 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2006 igvn.replace_node(opaq, opaq->in(2)); 2007 } 2008 assert(opaque4_count() == 0, "should be empty"); 2009 } 2010 2011 // StringOpts and late inlining of string methods 2012 void Compile::inline_string_calls(bool parse_time) { 2013 { 2014 // remove useless nodes to make the usage analysis simpler 2015 ResourceMark rm; 2016 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2017 } 2018 2019 { 2020 ResourceMark rm; 2021 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2022 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2023 print_method(PHASE_AFTER_STRINGOPTS, 3); 2024 } 2025 2026 // now inline anything that we skipped the first time around 2027 if (!parse_time) { 2028 _late_inlines_pos = _late_inlines.length(); 2029 } 2030 2031 while (_string_late_inlines.length() > 0) { 2032 CallGenerator* cg = _string_late_inlines.pop(); 2033 cg->do_late_inline(); 2034 if (failing()) return; 2035 } 2036 _string_late_inlines.trunc_to(0); 2037 } 2038 2039 // Late inlining of boxing methods 2040 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2041 if (_boxing_late_inlines.length() > 0) { 2042 assert(has_boxed_value(), "inconsistent"); 2043 2044 PhaseGVN* gvn = initial_gvn(); 2045 set_inlining_incrementally(true); 2046 2047 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2048 for_igvn()->clear(); 2049 gvn->replace_with(&igvn); 2050 2051 _late_inlines_pos = _late_inlines.length(); 2052 2053 while (_boxing_late_inlines.length() > 0) { 2054 CallGenerator* cg = _boxing_late_inlines.pop(); 2055 cg->do_late_inline(); 2056 if (failing()) return; 2057 } 2058 _boxing_late_inlines.trunc_to(0); 2059 2060 inline_incrementally_cleanup(igvn); 2061 2062 set_inlining_incrementally(false); 2063 } 2064 } 2065 2066 bool Compile::inline_incrementally_one() { 2067 assert(IncrementalInline, "incremental inlining should be on"); 2068 2069 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2070 set_inlining_progress(false); 2071 set_do_cleanup(false); 2072 int i = 0; 2073 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2074 CallGenerator* cg = _late_inlines.at(i); 2075 _late_inlines_pos = i+1; 2076 cg->do_late_inline(); 2077 if (failing()) return false; 2078 } 2079 int j = 0; 2080 for (; i < _late_inlines.length(); i++, j++) { 2081 _late_inlines.at_put(j, _late_inlines.at(i)); 2082 } 2083 _late_inlines.trunc_to(j); 2084 assert(inlining_progress() || _late_inlines.length() == 0, ""); 2085 2086 bool needs_cleanup = do_cleanup() || over_inlining_cutoff(); 2087 2088 set_inlining_progress(false); 2089 set_do_cleanup(false); 2090 return (_late_inlines.length() > 0) && !needs_cleanup; 2091 } 2092 2093 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) { 2094 { 2095 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2096 ResourceMark rm; 2097 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2098 } 2099 { 2100 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2101 igvn = PhaseIterGVN(initial_gvn()); 2102 igvn.optimize(); 2103 } 2104 } 2105 2106 // Perform incremental inlining until bound on number of live nodes is reached 2107 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2108 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2109 2110 set_inlining_incrementally(true); 2111 uint low_live_nodes = 0; 2112 2113 while (_late_inlines.length() > 0) { 2114 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2115 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2116 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2117 // PhaseIdealLoop is expensive so we only try it once we are 2118 // out of live nodes and we only try it again if the previous 2119 // helped got the number of nodes down significantly 2120 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2121 if (failing()) return; 2122 low_live_nodes = live_nodes(); 2123 _major_progress = true; 2124 } 2125 2126 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2127 break; // finish 2128 } 2129 } 2130 2131 for_igvn()->clear(); 2132 initial_gvn()->replace_with(&igvn); 2133 2134 while (inline_incrementally_one()) { 2135 assert(!failing(), "inconsistent"); 2136 } 2137 2138 if (failing()) return; 2139 2140 inline_incrementally_cleanup(igvn); 2141 2142 if (failing()) return; 2143 } 2144 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2145 2146 if (_string_late_inlines.length() > 0) { 2147 assert(has_stringbuilder(), "inconsistent"); 2148 for_igvn()->clear(); 2149 initial_gvn()->replace_with(&igvn); 2150 2151 inline_string_calls(false); 2152 2153 if (failing()) return; 2154 2155 inline_incrementally_cleanup(igvn); 2156 } 2157 2158 set_inlining_incrementally(false); 2159 } 2160 2161 2162 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) { 2163 if(_loop_opts_cnt > 0) { 2164 debug_only( int cnt = 0; ); 2165 while(major_progress() && (_loop_opts_cnt > 0)) { 2166 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2167 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2168 PhaseIdealLoop::optimize(igvn, mode); 2169 _loop_opts_cnt--; 2170 if (failing()) return false; 2171 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2172 } 2173 } 2174 return true; 2175 } 2176 2177 // Remove edges from "root" to each SafePoint at a backward branch. 2178 // They were inserted during parsing (see add_safepoint()) to make 2179 // infinite loops without calls or exceptions visible to root, i.e., 2180 // useful. 2181 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) { 2182 Node *r = root(); 2183 if (r != NULL) { 2184 for (uint i = r->req(); i < r->len(); ++i) { 2185 Node *n = r->in(i); 2186 if (n != NULL && n->is_SafePoint()) { 2187 r->rm_prec(i); 2188 if (n->outcnt() == 0) { 2189 igvn.remove_dead_node(n); 2190 } 2191 --i; 2192 } 2193 } 2194 } 2195 } 2196 2197 //------------------------------Optimize--------------------------------------- 2198 // Given a graph, optimize it. 2199 void Compile::Optimize() { 2200 TracePhase tp("optimizer", &timers[_t_optimizer]); 2201 2202 #ifndef PRODUCT 2203 if (_directive->BreakAtCompileOption) { 2204 BREAKPOINT; 2205 } 2206 2207 #endif 2208 2209 #ifdef ASSERT 2210 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2211 bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize); 2212 #endif 2213 2214 ResourceMark rm; 2215 2216 print_inlining_reinit(); 2217 2218 NOT_PRODUCT( verify_graph_edges(); ) 2219 2220 print_method(PHASE_AFTER_PARSING); 2221 2222 { 2223 // Iterative Global Value Numbering, including ideal transforms 2224 // Initialize IterGVN with types and values from parse-time GVN 2225 PhaseIterGVN igvn(initial_gvn()); 2226 #ifdef ASSERT 2227 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2228 #endif 2229 { 2230 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2231 igvn.optimize(); 2232 } 2233 2234 if (failing()) return; 2235 2236 print_method(PHASE_ITER_GVN1, 2); 2237 2238 inline_incrementally(igvn); 2239 2240 print_method(PHASE_INCREMENTAL_INLINE, 2); 2241 2242 if (failing()) return; 2243 2244 if (eliminate_boxing()) { 2245 // Inline valueOf() methods now. 2246 inline_boxing_calls(igvn); 2247 2248 if (AlwaysIncrementalInline) { 2249 inline_incrementally(igvn); 2250 } 2251 2252 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2253 2254 if (failing()) return; 2255 } 2256 2257 // Now that all inlining is over, cut edge from root to loop 2258 // safepoints 2259 remove_root_to_sfpts_edges(igvn); 2260 2261 // Remove the speculative part of types and clean up the graph from 2262 // the extra CastPP nodes whose only purpose is to carry them. Do 2263 // that early so that optimizations are not disrupted by the extra 2264 // CastPP nodes. 2265 remove_speculative_types(igvn); 2266 2267 // No more new expensive nodes will be added to the list from here 2268 // so keep only the actual candidates for optimizations. 2269 cleanup_expensive_nodes(igvn); 2270 2271 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2272 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2273 initial_gvn()->replace_with(&igvn); 2274 for_igvn()->clear(); 2275 Unique_Node_List new_worklist(C->comp_arena()); 2276 { 2277 ResourceMark rm; 2278 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2279 } 2280 set_for_igvn(&new_worklist); 2281 igvn = PhaseIterGVN(initial_gvn()); 2282 igvn.optimize(); 2283 } 2284 2285 // Perform escape analysis 2286 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2287 if (has_loops()) { 2288 // Cleanup graph (remove dead nodes). 2289 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2290 PhaseIdealLoop::optimize(igvn, LoopOptsNone); 2291 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2292 if (failing()) return; 2293 } 2294 ConnectionGraph::do_analysis(this, &igvn); 2295 2296 if (failing()) return; 2297 2298 // Optimize out fields loads from scalar replaceable allocations. 2299 igvn.optimize(); 2300 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2301 2302 if (failing()) return; 2303 2304 if (congraph() != NULL && macro_count() > 0) { 2305 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2306 PhaseMacroExpand mexp(igvn); 2307 mexp.eliminate_macro_nodes(); 2308 igvn.set_delay_transform(false); 2309 2310 igvn.optimize(); 2311 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2312 2313 if (failing()) return; 2314 } 2315 } 2316 2317 // Loop transforms on the ideal graph. Range Check Elimination, 2318 // peeling, unrolling, etc. 2319 2320 // Set loop opts counter 2321 if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2322 { 2323 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2324 PhaseIdealLoop::optimize(igvn, LoopOptsDefault); 2325 _loop_opts_cnt--; 2326 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2327 if (failing()) return; 2328 } 2329 // Loop opts pass if partial peeling occurred in previous pass 2330 if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) { 2331 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2332 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2333 _loop_opts_cnt--; 2334 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2335 if (failing()) return; 2336 } 2337 // Loop opts pass for loop-unrolling before CCP 2338 if(major_progress() && (_loop_opts_cnt > 0)) { 2339 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2340 PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf); 2341 _loop_opts_cnt--; 2342 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2343 } 2344 if (!failing()) { 2345 // Verify that last round of loop opts produced a valid graph 2346 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2347 PhaseIdealLoop::verify(igvn); 2348 } 2349 } 2350 if (failing()) return; 2351 2352 // Conditional Constant Propagation; 2353 PhaseCCP ccp( &igvn ); 2354 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2355 { 2356 TracePhase tp("ccp", &timers[_t_ccp]); 2357 ccp.do_transform(); 2358 } 2359 print_method(PHASE_CPP1, 2); 2360 2361 assert( true, "Break here to ccp.dump_old2new_map()"); 2362 2363 // Iterative Global Value Numbering, including ideal transforms 2364 { 2365 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2366 igvn = ccp; 2367 igvn.optimize(); 2368 } 2369 2370 print_method(PHASE_ITER_GVN2, 2); 2371 2372 if (failing()) return; 2373 2374 // Loop transforms on the ideal graph. Range Check Elimination, 2375 // peeling, unrolling, etc. 2376 if (!optimize_loops(igvn, LoopOptsDefault)) { 2377 return; 2378 } 2379 2380 #if INCLUDE_ZGC 2381 if (UseZGC) { 2382 ZBarrierSetC2::find_dominating_barriers(igvn); 2383 } 2384 #endif 2385 2386 if (failing()) return; 2387 2388 // Ensure that major progress is now clear 2389 C->clear_major_progress(); 2390 2391 { 2392 // Verify that all previous optimizations produced a valid graph 2393 // at least to this point, even if no loop optimizations were done. 2394 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2395 PhaseIdealLoop::verify(igvn); 2396 } 2397 2398 if (range_check_cast_count() > 0) { 2399 // No more loop optimizations. Remove all range check dependent CastIINodes. 2400 C->remove_range_check_casts(igvn); 2401 igvn.optimize(); 2402 } 2403 2404 #ifdef ASSERT 2405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2406 bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand); 2407 #endif 2408 2409 { 2410 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2411 PhaseMacroExpand mex(igvn); 2412 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2413 if (mex.expand_macro_nodes()) { 2414 assert(failing(), "must bail out w/ explicit message"); 2415 return; 2416 } 2417 } 2418 2419 { 2420 TracePhase tp("barrierExpand", &timers[_t_barrierExpand]); 2421 print_method(PHASE_BEFORE_BARRIER_EXPAND, 2); 2422 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2423 if (bs->expand_barriers(this, igvn)) { 2424 assert(failing(), "must bail out w/ explicit message"); 2425 return; 2426 } 2427 } 2428 2429 if (opaque4_count() > 0) { 2430 C->remove_opaque4_nodes(igvn); 2431 igvn.optimize(); 2432 } 2433 2434 DEBUG_ONLY( _modified_nodes = NULL; ) 2435 } // (End scope of igvn; run destructor if necessary for asserts.) 2436 2437 process_print_inlining(); 2438 // A method with only infinite loops has no edges entering loops from root 2439 { 2440 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2441 if (final_graph_reshaping()) { 2442 assert(failing(), "must bail out w/ explicit message"); 2443 return; 2444 } 2445 } 2446 2447 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2448 } 2449 2450 2451 //------------------------------Code_Gen--------------------------------------- 2452 // Given a graph, generate code for it 2453 void Compile::Code_Gen() { 2454 if (failing()) { 2455 return; 2456 } 2457 2458 // Perform instruction selection. You might think we could reclaim Matcher 2459 // memory PDQ, but actually the Matcher is used in generating spill code. 2460 // Internals of the Matcher (including some VectorSets) must remain live 2461 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2462 // set a bit in reclaimed memory. 2463 2464 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2465 // nodes. Mapping is only valid at the root of each matched subtree. 2466 NOT_PRODUCT( verify_graph_edges(); ) 2467 2468 Matcher matcher; 2469 _matcher = &matcher; 2470 { 2471 TracePhase tp("matcher", &timers[_t_matcher]); 2472 matcher.match(); 2473 } 2474 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2475 // nodes. Mapping is only valid at the root of each matched subtree. 2476 NOT_PRODUCT( verify_graph_edges(); ) 2477 2478 // If you have too many nodes, or if matching has failed, bail out 2479 check_node_count(0, "out of nodes matching instructions"); 2480 if (failing()) { 2481 return; 2482 } 2483 2484 print_method(PHASE_MATCHING, 2); 2485 2486 // Build a proper-looking CFG 2487 PhaseCFG cfg(node_arena(), root(), matcher); 2488 _cfg = &cfg; 2489 { 2490 TracePhase tp("scheduler", &timers[_t_scheduler]); 2491 bool success = cfg.do_global_code_motion(); 2492 if (!success) { 2493 return; 2494 } 2495 2496 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2497 NOT_PRODUCT( verify_graph_edges(); ) 2498 debug_only( cfg.verify(); ) 2499 } 2500 2501 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2502 _regalloc = ®alloc; 2503 { 2504 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2505 // Perform register allocation. After Chaitin, use-def chains are 2506 // no longer accurate (at spill code) and so must be ignored. 2507 // Node->LRG->reg mappings are still accurate. 2508 _regalloc->Register_Allocate(); 2509 2510 // Bail out if the allocator builds too many nodes 2511 if (failing()) { 2512 return; 2513 } 2514 } 2515 2516 // Prior to register allocation we kept empty basic blocks in case the 2517 // the allocator needed a place to spill. After register allocation we 2518 // are not adding any new instructions. If any basic block is empty, we 2519 // can now safely remove it. 2520 { 2521 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2522 cfg.remove_empty_blocks(); 2523 if (do_freq_based_layout()) { 2524 PhaseBlockLayout layout(cfg); 2525 } else { 2526 cfg.set_loop_alignment(); 2527 } 2528 cfg.fixup_flow(); 2529 } 2530 2531 // Apply peephole optimizations 2532 if( OptoPeephole ) { 2533 TracePhase tp("peephole", &timers[_t_peephole]); 2534 PhasePeephole peep( _regalloc, cfg); 2535 peep.do_transform(); 2536 } 2537 2538 // Do late expand if CPU requires this. 2539 if (Matcher::require_postalloc_expand) { 2540 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2541 cfg.postalloc_expand(_regalloc); 2542 } 2543 2544 // Convert Nodes to instruction bits in a buffer 2545 { 2546 TraceTime tp("output", &timers[_t_output], CITime); 2547 Output(); 2548 } 2549 2550 print_method(PHASE_FINAL_CODE); 2551 2552 // He's dead, Jim. 2553 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2554 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2555 } 2556 2557 2558 //------------------------------dump_asm--------------------------------------- 2559 // Dump formatted assembly 2560 #if defined(SUPPORT_OPTO_ASSEMBLY) 2561 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) { 2562 2563 int pc_digits = 3; // #chars required for pc 2564 int sb_chars = 3; // #chars for "start bundle" indicator 2565 int tab_size = 8; 2566 if (pcs != NULL) { 2567 int max_pc = 0; 2568 for (uint i = 0; i < pc_limit; i++) { 2569 max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc; 2570 } 2571 pc_digits = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc 2572 } 2573 int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size; 2574 2575 bool cut_short = false; 2576 st->print_cr("#"); 2577 st->print("# "); _tf->dump_on(st); st->cr(); 2578 st->print_cr("#"); 2579 2580 // For all blocks 2581 int pc = 0x0; // Program counter 2582 char starts_bundle = ' '; 2583 _regalloc->dump_frame(); 2584 2585 Node *n = NULL; 2586 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2587 if (VMThread::should_terminate()) { 2588 cut_short = true; 2589 break; 2590 } 2591 Block* block = _cfg->get_block(i); 2592 if (block->is_connector() && !Verbose) { 2593 continue; 2594 } 2595 n = block->head(); 2596 if ((pcs != NULL) && (n->_idx < pc_limit)) { 2597 pc = pcs[n->_idx]; 2598 st->print("%*.*x", pc_digits, pc_digits, pc); 2599 } 2600 st->fill_to(prefix_len); 2601 block->dump_head(_cfg, st); 2602 if (block->is_connector()) { 2603 st->fill_to(prefix_len); 2604 st->print_cr("# Empty connector block"); 2605 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2606 st->fill_to(prefix_len); 2607 st->print_cr("# Block is sole successor of call"); 2608 } 2609 2610 // For all instructions 2611 Node *delay = NULL; 2612 for (uint j = 0; j < block->number_of_nodes(); j++) { 2613 if (VMThread::should_terminate()) { 2614 cut_short = true; 2615 break; 2616 } 2617 n = block->get_node(j); 2618 if (valid_bundle_info(n)) { 2619 Bundle* bundle = node_bundling(n); 2620 if (bundle->used_in_unconditional_delay()) { 2621 delay = n; 2622 continue; 2623 } 2624 if (bundle->starts_bundle()) { 2625 starts_bundle = '+'; 2626 } 2627 } 2628 2629 if (WizardMode) { 2630 n->dump(); 2631 } 2632 2633 if( !n->is_Region() && // Dont print in the Assembly 2634 !n->is_Phi() && // a few noisely useless nodes 2635 !n->is_Proj() && 2636 !n->is_MachTemp() && 2637 !n->is_SafePointScalarObject() && 2638 !n->is_Catch() && // Would be nice to print exception table targets 2639 !n->is_MergeMem() && // Not very interesting 2640 !n->is_top() && // Debug info table constants 2641 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2642 ) { 2643 if ((pcs != NULL) && (n->_idx < pc_limit)) { 2644 pc = pcs[n->_idx]; 2645 st->print("%*.*x", pc_digits, pc_digits, pc); 2646 } else { 2647 st->fill_to(pc_digits); 2648 } 2649 st->print(" %c ", starts_bundle); 2650 starts_bundle = ' '; 2651 st->fill_to(prefix_len); 2652 n->format(_regalloc, st); 2653 st->cr(); 2654 } 2655 2656 // If we have an instruction with a delay slot, and have seen a delay, 2657 // then back up and print it 2658 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2659 // Coverity finding - Explicit null dereferenced. 2660 guarantee(delay != NULL, "no unconditional delay instruction"); 2661 if (WizardMode) delay->dump(); 2662 2663 if (node_bundling(delay)->starts_bundle()) 2664 starts_bundle = '+'; 2665 if ((pcs != NULL) && (n->_idx < pc_limit)) { 2666 pc = pcs[n->_idx]; 2667 st->print("%*.*x", pc_digits, pc_digits, pc); 2668 } else { 2669 st->fill_to(pc_digits); 2670 } 2671 st->print(" %c ", starts_bundle); 2672 starts_bundle = ' '; 2673 st->fill_to(prefix_len); 2674 delay->format(_regalloc, st); 2675 st->cr(); 2676 delay = NULL; 2677 } 2678 2679 // Dump the exception table as well 2680 if( n->is_Catch() && (Verbose || WizardMode) ) { 2681 // Print the exception table for this offset 2682 _handler_table.print_subtable_for(pc); 2683 } 2684 st->bol(); // Make sure we start on a new line 2685 } 2686 st->cr(); // one empty line between blocks 2687 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2688 } // End of per-block dump 2689 2690 if (cut_short) st->print_cr("*** disassembly is cut short ***"); 2691 } 2692 #endif 2693 2694 //------------------------------Final_Reshape_Counts--------------------------- 2695 // This class defines counters to help identify when a method 2696 // may/must be executed using hardware with only 24-bit precision. 2697 struct Final_Reshape_Counts : public StackObj { 2698 int _call_count; // count non-inlined 'common' calls 2699 int _float_count; // count float ops requiring 24-bit precision 2700 int _double_count; // count double ops requiring more precision 2701 int _java_call_count; // count non-inlined 'java' calls 2702 int _inner_loop_count; // count loops which need alignment 2703 VectorSet _visited; // Visitation flags 2704 Node_List _tests; // Set of IfNodes & PCTableNodes 2705 2706 Final_Reshape_Counts() : 2707 _call_count(0), _float_count(0), _double_count(0), 2708 _java_call_count(0), _inner_loop_count(0), 2709 _visited( Thread::current()->resource_area() ) { } 2710 2711 void inc_call_count () { _call_count ++; } 2712 void inc_float_count () { _float_count ++; } 2713 void inc_double_count() { _double_count++; } 2714 void inc_java_call_count() { _java_call_count++; } 2715 void inc_inner_loop_count() { _inner_loop_count++; } 2716 2717 int get_call_count () const { return _call_count ; } 2718 int get_float_count () const { return _float_count ; } 2719 int get_double_count() const { return _double_count; } 2720 int get_java_call_count() const { return _java_call_count; } 2721 int get_inner_loop_count() const { return _inner_loop_count; } 2722 }; 2723 2724 #ifdef ASSERT 2725 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2726 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2727 // Make sure the offset goes inside the instance layout. 2728 return k->contains_field_offset(tp->offset()); 2729 // Note that OffsetBot and OffsetTop are very negative. 2730 } 2731 #endif 2732 2733 // Eliminate trivially redundant StoreCMs and accumulate their 2734 // precedence edges. 2735 void Compile::eliminate_redundant_card_marks(Node* n) { 2736 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2737 if (n->in(MemNode::Address)->outcnt() > 1) { 2738 // There are multiple users of the same address so it might be 2739 // possible to eliminate some of the StoreCMs 2740 Node* mem = n->in(MemNode::Memory); 2741 Node* adr = n->in(MemNode::Address); 2742 Node* val = n->in(MemNode::ValueIn); 2743 Node* prev = n; 2744 bool done = false; 2745 // Walk the chain of StoreCMs eliminating ones that match. As 2746 // long as it's a chain of single users then the optimization is 2747 // safe. Eliminating partially redundant StoreCMs would require 2748 // cloning copies down the other paths. 2749 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2750 if (adr == mem->in(MemNode::Address) && 2751 val == mem->in(MemNode::ValueIn)) { 2752 // redundant StoreCM 2753 if (mem->req() > MemNode::OopStore) { 2754 // Hasn't been processed by this code yet. 2755 n->add_prec(mem->in(MemNode::OopStore)); 2756 } else { 2757 // Already converted to precedence edge 2758 for (uint i = mem->req(); i < mem->len(); i++) { 2759 // Accumulate any precedence edges 2760 if (mem->in(i) != NULL) { 2761 n->add_prec(mem->in(i)); 2762 } 2763 } 2764 // Everything above this point has been processed. 2765 done = true; 2766 } 2767 // Eliminate the previous StoreCM 2768 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2769 assert(mem->outcnt() == 0, "should be dead"); 2770 mem->disconnect_inputs(NULL, this); 2771 } else { 2772 prev = mem; 2773 } 2774 mem = prev->in(MemNode::Memory); 2775 } 2776 } 2777 } 2778 2779 //------------------------------final_graph_reshaping_impl---------------------- 2780 // Implement items 1-5 from final_graph_reshaping below. 2781 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2782 2783 if ( n->outcnt() == 0 ) return; // dead node 2784 uint nop = n->Opcode(); 2785 2786 // Check for 2-input instruction with "last use" on right input. 2787 // Swap to left input. Implements item (2). 2788 if( n->req() == 3 && // two-input instruction 2789 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2790 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2791 n->in(2)->outcnt() == 1 &&// right use IS a last use 2792 !n->in(2)->is_Con() ) { // right use is not a constant 2793 // Check for commutative opcode 2794 switch( nop ) { 2795 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2796 case Op_MaxI: case Op_MinI: 2797 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2798 case Op_AndL: case Op_XorL: case Op_OrL: 2799 case Op_AndI: case Op_XorI: case Op_OrI: { 2800 // Move "last use" input to left by swapping inputs 2801 n->swap_edges(1, 2); 2802 break; 2803 } 2804 default: 2805 break; 2806 } 2807 } 2808 2809 #ifdef ASSERT 2810 if( n->is_Mem() ) { 2811 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2812 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2813 // oop will be recorded in oop map if load crosses safepoint 2814 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2815 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2816 "raw memory operations should have control edge"); 2817 } 2818 if (n->is_MemBar()) { 2819 MemBarNode* mb = n->as_MemBar(); 2820 if (mb->trailing_store() || mb->trailing_load_store()) { 2821 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2822 Node* mem = mb->in(MemBarNode::Precedent); 2823 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2824 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2825 } else if (mb->leading()) { 2826 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2827 } 2828 } 2829 #endif 2830 // Count FPU ops and common calls, implements item (3) 2831 bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop); 2832 if (!gc_handled) { 2833 final_graph_reshaping_main_switch(n, frc, nop); 2834 } 2835 2836 // Collect CFG split points 2837 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 2838 frc._tests.push(n); 2839 } 2840 } 2841 2842 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) { 2843 switch( nop ) { 2844 // Count all float operations that may use FPU 2845 case Op_AddF: 2846 case Op_SubF: 2847 case Op_MulF: 2848 case Op_DivF: 2849 case Op_NegF: 2850 case Op_ModF: 2851 case Op_ConvI2F: 2852 case Op_ConF: 2853 case Op_CmpF: 2854 case Op_CmpF3: 2855 // case Op_ConvL2F: // longs are split into 32-bit halves 2856 frc.inc_float_count(); 2857 break; 2858 2859 case Op_ConvF2D: 2860 case Op_ConvD2F: 2861 frc.inc_float_count(); 2862 frc.inc_double_count(); 2863 break; 2864 2865 // Count all double operations that may use FPU 2866 case Op_AddD: 2867 case Op_SubD: 2868 case Op_MulD: 2869 case Op_DivD: 2870 case Op_NegD: 2871 case Op_ModD: 2872 case Op_ConvI2D: 2873 case Op_ConvD2I: 2874 // case Op_ConvL2D: // handled by leaf call 2875 // case Op_ConvD2L: // handled by leaf call 2876 case Op_ConD: 2877 case Op_CmpD: 2878 case Op_CmpD3: 2879 frc.inc_double_count(); 2880 break; 2881 case Op_Opaque1: // Remove Opaque Nodes before matching 2882 case Op_Opaque2: // Remove Opaque Nodes before matching 2883 case Op_Opaque3: 2884 n->subsume_by(n->in(1), this); 2885 break; 2886 case Op_CallStaticJava: 2887 case Op_CallJava: 2888 case Op_CallDynamicJava: 2889 frc.inc_java_call_count(); // Count java call site; 2890 case Op_CallRuntime: 2891 case Op_CallLeaf: 2892 case Op_CallLeafNoFP: { 2893 assert (n->is_Call(), ""); 2894 CallNode *call = n->as_Call(); 2895 // Count call sites where the FP mode bit would have to be flipped. 2896 // Do not count uncommon runtime calls: 2897 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2898 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2899 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2900 frc.inc_call_count(); // Count the call site 2901 } else { // See if uncommon argument is shared 2902 Node *n = call->in(TypeFunc::Parms); 2903 int nop = n->Opcode(); 2904 // Clone shared simple arguments to uncommon calls, item (1). 2905 if (n->outcnt() > 1 && 2906 !n->is_Proj() && 2907 nop != Op_CreateEx && 2908 nop != Op_CheckCastPP && 2909 nop != Op_DecodeN && 2910 nop != Op_DecodeNKlass && 2911 !n->is_Mem() && 2912 !n->is_Phi()) { 2913 Node *x = n->clone(); 2914 call->set_req(TypeFunc::Parms, x); 2915 } 2916 } 2917 break; 2918 } 2919 2920 case Op_StoreD: 2921 case Op_LoadD: 2922 case Op_LoadD_unaligned: 2923 frc.inc_double_count(); 2924 goto handle_mem; 2925 case Op_StoreF: 2926 case Op_LoadF: 2927 frc.inc_float_count(); 2928 goto handle_mem; 2929 2930 case Op_StoreCM: 2931 { 2932 // Convert OopStore dependence into precedence edge 2933 Node* prec = n->in(MemNode::OopStore); 2934 n->del_req(MemNode::OopStore); 2935 n->add_prec(prec); 2936 eliminate_redundant_card_marks(n); 2937 } 2938 2939 // fall through 2940 2941 case Op_StoreB: 2942 case Op_StoreC: 2943 case Op_StorePConditional: 2944 case Op_StoreI: 2945 case Op_StoreL: 2946 case Op_StoreIConditional: 2947 case Op_StoreLConditional: 2948 case Op_CompareAndSwapB: 2949 case Op_CompareAndSwapS: 2950 case Op_CompareAndSwapI: 2951 case Op_CompareAndSwapL: 2952 case Op_CompareAndSwapP: 2953 case Op_CompareAndSwapN: 2954 case Op_WeakCompareAndSwapB: 2955 case Op_WeakCompareAndSwapS: 2956 case Op_WeakCompareAndSwapI: 2957 case Op_WeakCompareAndSwapL: 2958 case Op_WeakCompareAndSwapP: 2959 case Op_WeakCompareAndSwapN: 2960 case Op_CompareAndExchangeB: 2961 case Op_CompareAndExchangeS: 2962 case Op_CompareAndExchangeI: 2963 case Op_CompareAndExchangeL: 2964 case Op_CompareAndExchangeP: 2965 case Op_CompareAndExchangeN: 2966 case Op_GetAndAddS: 2967 case Op_GetAndAddB: 2968 case Op_GetAndAddI: 2969 case Op_GetAndAddL: 2970 case Op_GetAndSetS: 2971 case Op_GetAndSetB: 2972 case Op_GetAndSetI: 2973 case Op_GetAndSetL: 2974 case Op_GetAndSetP: 2975 case Op_GetAndSetN: 2976 case Op_StoreP: 2977 case Op_StoreN: 2978 case Op_StoreNKlass: 2979 case Op_LoadB: 2980 case Op_LoadUB: 2981 case Op_LoadUS: 2982 case Op_LoadI: 2983 case Op_LoadKlass: 2984 case Op_LoadNKlass: 2985 case Op_LoadL: 2986 case Op_LoadL_unaligned: 2987 case Op_LoadPLocked: 2988 case Op_LoadP: 2989 case Op_LoadN: 2990 case Op_LoadRange: 2991 case Op_LoadS: { 2992 handle_mem: 2993 #ifdef ASSERT 2994 if( VerifyOptoOopOffsets ) { 2995 MemNode* mem = n->as_Mem(); 2996 // Check to see if address types have grounded out somehow. 2997 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2998 assert( !tp || oop_offset_is_sane(tp), "" ); 2999 } 3000 #endif 3001 break; 3002 } 3003 3004 case Op_AddP: { // Assert sane base pointers 3005 Node *addp = n->in(AddPNode::Address); 3006 assert( !addp->is_AddP() || 3007 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3008 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3009 "Base pointers must match (addp %u)", addp->_idx ); 3010 #ifdef _LP64 3011 if ((UseCompressedOops || UseCompressedClassPointers) && 3012 addp->Opcode() == Op_ConP && 3013 addp == n->in(AddPNode::Base) && 3014 n->in(AddPNode::Offset)->is_Con()) { 3015 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3016 // on the platform and on the compressed oops mode. 3017 // Use addressing with narrow klass to load with offset on x86. 3018 // Some platforms can use the constant pool to load ConP. 3019 // Do this transformation here since IGVN will convert ConN back to ConP. 3020 const Type* t = addp->bottom_type(); 3021 bool is_oop = t->isa_oopptr() != NULL; 3022 bool is_klass = t->isa_klassptr() != NULL; 3023 3024 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3025 (is_klass && Matcher::const_klass_prefer_decode())) { 3026 Node* nn = NULL; 3027 3028 int op = is_oop ? Op_ConN : Op_ConNKlass; 3029 3030 // Look for existing ConN node of the same exact type. 3031 Node* r = root(); 3032 uint cnt = r->outcnt(); 3033 for (uint i = 0; i < cnt; i++) { 3034 Node* m = r->raw_out(i); 3035 if (m!= NULL && m->Opcode() == op && 3036 m->bottom_type()->make_ptr() == t) { 3037 nn = m; 3038 break; 3039 } 3040 } 3041 if (nn != NULL) { 3042 // Decode a narrow oop to match address 3043 // [R12 + narrow_oop_reg<<3 + offset] 3044 if (is_oop) { 3045 nn = new DecodeNNode(nn, t); 3046 } else { 3047 nn = new DecodeNKlassNode(nn, t); 3048 } 3049 // Check for succeeding AddP which uses the same Base. 3050 // Otherwise we will run into the assertion above when visiting that guy. 3051 for (uint i = 0; i < n->outcnt(); ++i) { 3052 Node *out_i = n->raw_out(i); 3053 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3054 out_i->set_req(AddPNode::Base, nn); 3055 #ifdef ASSERT 3056 for (uint j = 0; j < out_i->outcnt(); ++j) { 3057 Node *out_j = out_i->raw_out(j); 3058 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3059 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3060 } 3061 #endif 3062 } 3063 } 3064 n->set_req(AddPNode::Base, nn); 3065 n->set_req(AddPNode::Address, nn); 3066 if (addp->outcnt() == 0) { 3067 addp->disconnect_inputs(NULL, this); 3068 } 3069 } 3070 } 3071 } 3072 #endif 3073 // platform dependent reshaping of the address expression 3074 reshape_address(n->as_AddP()); 3075 break; 3076 } 3077 3078 case Op_CastPP: { 3079 // Remove CastPP nodes to gain more freedom during scheduling but 3080 // keep the dependency they encode as control or precedence edges 3081 // (if control is set already) on memory operations. Some CastPP 3082 // nodes don't have a control (don't carry a dependency): skip 3083 // those. 3084 if (n->in(0) != NULL) { 3085 ResourceMark rm; 3086 Unique_Node_List wq; 3087 wq.push(n); 3088 for (uint next = 0; next < wq.size(); ++next) { 3089 Node *m = wq.at(next); 3090 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3091 Node* use = m->fast_out(i); 3092 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3093 use->ensure_control_or_add_prec(n->in(0)); 3094 } else { 3095 switch(use->Opcode()) { 3096 case Op_AddP: 3097 case Op_DecodeN: 3098 case Op_DecodeNKlass: 3099 case Op_CheckCastPP: 3100 case Op_CastPP: 3101 wq.push(use); 3102 break; 3103 } 3104 } 3105 } 3106 } 3107 } 3108 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3109 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3110 Node* in1 = n->in(1); 3111 const Type* t = n->bottom_type(); 3112 Node* new_in1 = in1->clone(); 3113 new_in1->as_DecodeN()->set_type(t); 3114 3115 if (!Matcher::narrow_oop_use_complex_address()) { 3116 // 3117 // x86, ARM and friends can handle 2 adds in addressing mode 3118 // and Matcher can fold a DecodeN node into address by using 3119 // a narrow oop directly and do implicit NULL check in address: 3120 // 3121 // [R12 + narrow_oop_reg<<3 + offset] 3122 // NullCheck narrow_oop_reg 3123 // 3124 // On other platforms (Sparc) we have to keep new DecodeN node and 3125 // use it to do implicit NULL check in address: 3126 // 3127 // decode_not_null narrow_oop_reg, base_reg 3128 // [base_reg + offset] 3129 // NullCheck base_reg 3130 // 3131 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3132 // to keep the information to which NULL check the new DecodeN node 3133 // corresponds to use it as value in implicit_null_check(). 3134 // 3135 new_in1->set_req(0, n->in(0)); 3136 } 3137 3138 n->subsume_by(new_in1, this); 3139 if (in1->outcnt() == 0) { 3140 in1->disconnect_inputs(NULL, this); 3141 } 3142 } else { 3143 n->subsume_by(n->in(1), this); 3144 if (n->outcnt() == 0) { 3145 n->disconnect_inputs(NULL, this); 3146 } 3147 } 3148 break; 3149 } 3150 #ifdef _LP64 3151 case Op_CmpP: 3152 // Do this transformation here to preserve CmpPNode::sub() and 3153 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3154 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3155 Node* in1 = n->in(1); 3156 Node* in2 = n->in(2); 3157 if (!in1->is_DecodeNarrowPtr()) { 3158 in2 = in1; 3159 in1 = n->in(2); 3160 } 3161 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3162 3163 Node* new_in2 = NULL; 3164 if (in2->is_DecodeNarrowPtr()) { 3165 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3166 new_in2 = in2->in(1); 3167 } else if (in2->Opcode() == Op_ConP) { 3168 const Type* t = in2->bottom_type(); 3169 if (t == TypePtr::NULL_PTR) { 3170 assert(in1->is_DecodeN(), "compare klass to null?"); 3171 // Don't convert CmpP null check into CmpN if compressed 3172 // oops implicit null check is not generated. 3173 // This will allow to generate normal oop implicit null check. 3174 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3175 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3176 // 3177 // This transformation together with CastPP transformation above 3178 // will generated code for implicit NULL checks for compressed oops. 3179 // 3180 // The original code after Optimize() 3181 // 3182 // LoadN memory, narrow_oop_reg 3183 // decode narrow_oop_reg, base_reg 3184 // CmpP base_reg, NULL 3185 // CastPP base_reg // NotNull 3186 // Load [base_reg + offset], val_reg 3187 // 3188 // after these transformations will be 3189 // 3190 // LoadN memory, narrow_oop_reg 3191 // CmpN narrow_oop_reg, NULL 3192 // decode_not_null narrow_oop_reg, base_reg 3193 // Load [base_reg + offset], val_reg 3194 // 3195 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3196 // since narrow oops can be used in debug info now (see the code in 3197 // final_graph_reshaping_walk()). 3198 // 3199 // At the end the code will be matched to 3200 // on x86: 3201 // 3202 // Load_narrow_oop memory, narrow_oop_reg 3203 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3204 // NullCheck narrow_oop_reg 3205 // 3206 // and on sparc: 3207 // 3208 // Load_narrow_oop memory, narrow_oop_reg 3209 // decode_not_null narrow_oop_reg, base_reg 3210 // Load [base_reg + offset], val_reg 3211 // NullCheck base_reg 3212 // 3213 } else if (t->isa_oopptr()) { 3214 new_in2 = ConNode::make(t->make_narrowoop()); 3215 } else if (t->isa_klassptr()) { 3216 new_in2 = ConNode::make(t->make_narrowklass()); 3217 } 3218 } 3219 if (new_in2 != NULL) { 3220 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3221 n->subsume_by(cmpN, this); 3222 if (in1->outcnt() == 0) { 3223 in1->disconnect_inputs(NULL, this); 3224 } 3225 if (in2->outcnt() == 0) { 3226 in2->disconnect_inputs(NULL, this); 3227 } 3228 } 3229 } 3230 break; 3231 3232 case Op_DecodeN: 3233 case Op_DecodeNKlass: 3234 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3235 // DecodeN could be pinned when it can't be fold into 3236 // an address expression, see the code for Op_CastPP above. 3237 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3238 break; 3239 3240 case Op_EncodeP: 3241 case Op_EncodePKlass: { 3242 Node* in1 = n->in(1); 3243 if (in1->is_DecodeNarrowPtr()) { 3244 n->subsume_by(in1->in(1), this); 3245 } else if (in1->Opcode() == Op_ConP) { 3246 const Type* t = in1->bottom_type(); 3247 if (t == TypePtr::NULL_PTR) { 3248 assert(t->isa_oopptr(), "null klass?"); 3249 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3250 } else if (t->isa_oopptr()) { 3251 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3252 } else if (t->isa_klassptr()) { 3253 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3254 } 3255 } 3256 if (in1->outcnt() == 0) { 3257 in1->disconnect_inputs(NULL, this); 3258 } 3259 break; 3260 } 3261 3262 case Op_Proj: { 3263 if (OptimizeStringConcat) { 3264 ProjNode* p = n->as_Proj(); 3265 if (p->_is_io_use) { 3266 // Separate projections were used for the exception path which 3267 // are normally removed by a late inline. If it wasn't inlined 3268 // then they will hang around and should just be replaced with 3269 // the original one. 3270 Node* proj = NULL; 3271 // Replace with just one 3272 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3273 Node *use = i.get(); 3274 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3275 proj = use; 3276 break; 3277 } 3278 } 3279 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3280 if (proj != NULL) { 3281 p->subsume_by(proj, this); 3282 } 3283 } 3284 } 3285 break; 3286 } 3287 3288 case Op_Phi: 3289 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3290 // The EncodeP optimization may create Phi with the same edges 3291 // for all paths. It is not handled well by Register Allocator. 3292 Node* unique_in = n->in(1); 3293 assert(unique_in != NULL, ""); 3294 uint cnt = n->req(); 3295 for (uint i = 2; i < cnt; i++) { 3296 Node* m = n->in(i); 3297 assert(m != NULL, ""); 3298 if (unique_in != m) 3299 unique_in = NULL; 3300 } 3301 if (unique_in != NULL) { 3302 n->subsume_by(unique_in, this); 3303 } 3304 } 3305 break; 3306 3307 #endif 3308 3309 #ifdef ASSERT 3310 case Op_CastII: 3311 // Verify that all range check dependent CastII nodes were removed. 3312 if (n->isa_CastII()->has_range_check()) { 3313 n->dump(3); 3314 assert(false, "Range check dependent CastII node was not removed"); 3315 } 3316 break; 3317 #endif 3318 3319 case Op_ModI: 3320 if (UseDivMod) { 3321 // Check if a%b and a/b both exist 3322 Node* d = n->find_similar(Op_DivI); 3323 if (d) { 3324 // Replace them with a fused divmod if supported 3325 if (Matcher::has_match_rule(Op_DivModI)) { 3326 DivModINode* divmod = DivModINode::make(n); 3327 d->subsume_by(divmod->div_proj(), this); 3328 n->subsume_by(divmod->mod_proj(), this); 3329 } else { 3330 // replace a%b with a-((a/b)*b) 3331 Node* mult = new MulINode(d, d->in(2)); 3332 Node* sub = new SubINode(d->in(1), mult); 3333 n->subsume_by(sub, this); 3334 } 3335 } 3336 } 3337 break; 3338 3339 case Op_ModL: 3340 if (UseDivMod) { 3341 // Check if a%b and a/b both exist 3342 Node* d = n->find_similar(Op_DivL); 3343 if (d) { 3344 // Replace them with a fused divmod if supported 3345 if (Matcher::has_match_rule(Op_DivModL)) { 3346 DivModLNode* divmod = DivModLNode::make(n); 3347 d->subsume_by(divmod->div_proj(), this); 3348 n->subsume_by(divmod->mod_proj(), this); 3349 } else { 3350 // replace a%b with a-((a/b)*b) 3351 Node* mult = new MulLNode(d, d->in(2)); 3352 Node* sub = new SubLNode(d->in(1), mult); 3353 n->subsume_by(sub, this); 3354 } 3355 } 3356 } 3357 break; 3358 3359 case Op_LoadVector: 3360 case Op_StoreVector: 3361 break; 3362 3363 case Op_AddReductionVI: 3364 case Op_AddReductionVL: 3365 case Op_AddReductionVF: 3366 case Op_AddReductionVD: 3367 case Op_MulReductionVI: 3368 case Op_MulReductionVL: 3369 case Op_MulReductionVF: 3370 case Op_MulReductionVD: 3371 case Op_MinReductionV: 3372 case Op_MaxReductionV: 3373 break; 3374 3375 case Op_PackB: 3376 case Op_PackS: 3377 case Op_PackI: 3378 case Op_PackF: 3379 case Op_PackL: 3380 case Op_PackD: 3381 if (n->req()-1 > 2) { 3382 // Replace many operand PackNodes with a binary tree for matching 3383 PackNode* p = (PackNode*) n; 3384 Node* btp = p->binary_tree_pack(1, n->req()); 3385 n->subsume_by(btp, this); 3386 } 3387 break; 3388 case Op_Loop: 3389 case Op_CountedLoop: 3390 case Op_OuterStripMinedLoop: 3391 if (n->as_Loop()->is_inner_loop()) { 3392 frc.inc_inner_loop_count(); 3393 } 3394 n->as_Loop()->verify_strip_mined(0); 3395 break; 3396 case Op_LShiftI: 3397 case Op_RShiftI: 3398 case Op_URShiftI: 3399 case Op_LShiftL: 3400 case Op_RShiftL: 3401 case Op_URShiftL: 3402 if (Matcher::need_masked_shift_count) { 3403 // The cpu's shift instructions don't restrict the count to the 3404 // lower 5/6 bits. We need to do the masking ourselves. 3405 Node* in2 = n->in(2); 3406 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3407 const TypeInt* t = in2->find_int_type(); 3408 if (t != NULL && t->is_con()) { 3409 juint shift = t->get_con(); 3410 if (shift > mask) { // Unsigned cmp 3411 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3412 } 3413 } else { 3414 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3415 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3416 n->set_req(2, shift); 3417 } 3418 } 3419 if (in2->outcnt() == 0) { // Remove dead node 3420 in2->disconnect_inputs(NULL, this); 3421 } 3422 } 3423 break; 3424 case Op_MemBarStoreStore: 3425 case Op_MemBarRelease: 3426 // Break the link with AllocateNode: it is no longer useful and 3427 // confuses register allocation. 3428 if (n->req() > MemBarNode::Precedent) { 3429 n->set_req(MemBarNode::Precedent, top()); 3430 } 3431 break; 3432 case Op_MemBarAcquire: { 3433 if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) { 3434 // At parse time, the trailing MemBarAcquire for a volatile load 3435 // is created with an edge to the load. After optimizations, 3436 // that input may be a chain of Phis. If those phis have no 3437 // other use, then the MemBarAcquire keeps them alive and 3438 // register allocation can be confused. 3439 ResourceMark rm; 3440 Unique_Node_List wq; 3441 wq.push(n->in(MemBarNode::Precedent)); 3442 n->set_req(MemBarNode::Precedent, top()); 3443 while (wq.size() > 0) { 3444 Node* m = wq.pop(); 3445 if (m->outcnt() == 0) { 3446 for (uint j = 0; j < m->req(); j++) { 3447 Node* in = m->in(j); 3448 if (in != NULL) { 3449 wq.push(in); 3450 } 3451 } 3452 m->disconnect_inputs(NULL, this); 3453 } 3454 } 3455 } 3456 break; 3457 } 3458 case Op_RangeCheck: { 3459 RangeCheckNode* rc = n->as_RangeCheck(); 3460 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3461 n->subsume_by(iff, this); 3462 frc._tests.push(iff); 3463 break; 3464 } 3465 case Op_ConvI2L: { 3466 if (!Matcher::convi2l_type_required) { 3467 // Code generation on some platforms doesn't need accurate 3468 // ConvI2L types. Widening the type can help remove redundant 3469 // address computations. 3470 n->as_Type()->set_type(TypeLong::INT); 3471 ResourceMark rm; 3472 Node_List wq; 3473 wq.push(n); 3474 for (uint next = 0; next < wq.size(); next++) { 3475 Node *m = wq.at(next); 3476 3477 for(;;) { 3478 // Loop over all nodes with identical inputs edges as m 3479 Node* k = m->find_similar(m->Opcode()); 3480 if (k == NULL) { 3481 break; 3482 } 3483 // Push their uses so we get a chance to remove node made 3484 // redundant 3485 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3486 Node* u = k->fast_out(i); 3487 assert(!wq.contains(u), "shouldn't process one node several times"); 3488 if (u->Opcode() == Op_LShiftL || 3489 u->Opcode() == Op_AddL || 3490 u->Opcode() == Op_SubL || 3491 u->Opcode() == Op_AddP) { 3492 wq.push(u); 3493 } 3494 } 3495 // Replace all nodes with identical edges as m with m 3496 k->subsume_by(m, this); 3497 } 3498 } 3499 } 3500 break; 3501 } 3502 case Op_CmpUL: { 3503 if (!Matcher::has_match_rule(Op_CmpUL)) { 3504 // No support for unsigned long comparisons 3505 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3506 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3507 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3508 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3509 Node* andl = new AndLNode(orl, remove_sign_mask); 3510 Node* cmp = new CmpLNode(andl, n->in(2)); 3511 n->subsume_by(cmp, this); 3512 } 3513 break; 3514 } 3515 default: 3516 assert(!n->is_Call(), ""); 3517 assert(!n->is_Mem(), ""); 3518 assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3519 break; 3520 } 3521 } 3522 3523 //------------------------------final_graph_reshaping_walk--------------------- 3524 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3525 // requires that the walk visits a node's inputs before visiting the node. 3526 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3527 ResourceArea *area = Thread::current()->resource_area(); 3528 Unique_Node_List sfpt(area); 3529 3530 frc._visited.set(root->_idx); // first, mark node as visited 3531 uint cnt = root->req(); 3532 Node *n = root; 3533 uint i = 0; 3534 while (true) { 3535 if (i < cnt) { 3536 // Place all non-visited non-null inputs onto stack 3537 Node* m = n->in(i); 3538 ++i; 3539 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3540 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3541 // compute worst case interpreter size in case of a deoptimization 3542 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3543 3544 sfpt.push(m); 3545 } 3546 cnt = m->req(); 3547 nstack.push(n, i); // put on stack parent and next input's index 3548 n = m; 3549 i = 0; 3550 } 3551 } else { 3552 // Now do post-visit work 3553 final_graph_reshaping_impl( n, frc ); 3554 if (nstack.is_empty()) 3555 break; // finished 3556 n = nstack.node(); // Get node from stack 3557 cnt = n->req(); 3558 i = nstack.index(); 3559 nstack.pop(); // Shift to the next node on stack 3560 } 3561 } 3562 3563 // Skip next transformation if compressed oops are not used. 3564 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3565 (!UseCompressedOops && !UseCompressedClassPointers)) 3566 return; 3567 3568 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3569 // It could be done for an uncommon traps or any safepoints/calls 3570 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3571 while (sfpt.size() > 0) { 3572 n = sfpt.pop(); 3573 JVMState *jvms = n->as_SafePoint()->jvms(); 3574 assert(jvms != NULL, "sanity"); 3575 int start = jvms->debug_start(); 3576 int end = n->req(); 3577 bool is_uncommon = (n->is_CallStaticJava() && 3578 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3579 for (int j = start; j < end; j++) { 3580 Node* in = n->in(j); 3581 if (in->is_DecodeNarrowPtr()) { 3582 bool safe_to_skip = true; 3583 if (!is_uncommon ) { 3584 // Is it safe to skip? 3585 for (uint i = 0; i < in->outcnt(); i++) { 3586 Node* u = in->raw_out(i); 3587 if (!u->is_SafePoint() || 3588 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3589 safe_to_skip = false; 3590 } 3591 } 3592 } 3593 if (safe_to_skip) { 3594 n->set_req(j, in->in(1)); 3595 } 3596 if (in->outcnt() == 0) { 3597 in->disconnect_inputs(NULL, this); 3598 } 3599 } 3600 } 3601 } 3602 } 3603 3604 //------------------------------final_graph_reshaping-------------------------- 3605 // Final Graph Reshaping. 3606 // 3607 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3608 // and not commoned up and forced early. Must come after regular 3609 // optimizations to avoid GVN undoing the cloning. Clone constant 3610 // inputs to Loop Phis; these will be split by the allocator anyways. 3611 // Remove Opaque nodes. 3612 // (2) Move last-uses by commutative operations to the left input to encourage 3613 // Intel update-in-place two-address operations and better register usage 3614 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3615 // calls canonicalizing them back. 3616 // (3) Count the number of double-precision FP ops, single-precision FP ops 3617 // and call sites. On Intel, we can get correct rounding either by 3618 // forcing singles to memory (requires extra stores and loads after each 3619 // FP bytecode) or we can set a rounding mode bit (requires setting and 3620 // clearing the mode bit around call sites). The mode bit is only used 3621 // if the relative frequency of single FP ops to calls is low enough. 3622 // This is a key transform for SPEC mpeg_audio. 3623 // (4) Detect infinite loops; blobs of code reachable from above but not 3624 // below. Several of the Code_Gen algorithms fail on such code shapes, 3625 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3626 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3627 // Detection is by looking for IfNodes where only 1 projection is 3628 // reachable from below or CatchNodes missing some targets. 3629 // (5) Assert for insane oop offsets in debug mode. 3630 3631 bool Compile::final_graph_reshaping() { 3632 // an infinite loop may have been eliminated by the optimizer, 3633 // in which case the graph will be empty. 3634 if (root()->req() == 1) { 3635 record_method_not_compilable("trivial infinite loop"); 3636 return true; 3637 } 3638 3639 // Expensive nodes have their control input set to prevent the GVN 3640 // from freely commoning them. There's no GVN beyond this point so 3641 // no need to keep the control input. We want the expensive nodes to 3642 // be freely moved to the least frequent code path by gcm. 3643 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3644 for (int i = 0; i < expensive_count(); i++) { 3645 _expensive_nodes->at(i)->set_req(0, NULL); 3646 } 3647 3648 Final_Reshape_Counts frc; 3649 3650 // Visit everybody reachable! 3651 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3652 Node_Stack nstack(live_nodes() >> 1); 3653 final_graph_reshaping_walk(nstack, root(), frc); 3654 3655 // Check for unreachable (from below) code (i.e., infinite loops). 3656 for( uint i = 0; i < frc._tests.size(); i++ ) { 3657 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3658 // Get number of CFG targets. 3659 // Note that PCTables include exception targets after calls. 3660 uint required_outcnt = n->required_outcnt(); 3661 if (n->outcnt() != required_outcnt) { 3662 // Check for a few special cases. Rethrow Nodes never take the 3663 // 'fall-thru' path, so expected kids is 1 less. 3664 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3665 if (n->in(0)->in(0)->is_Call()) { 3666 CallNode *call = n->in(0)->in(0)->as_Call(); 3667 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3668 required_outcnt--; // Rethrow always has 1 less kid 3669 } else if (call->req() > TypeFunc::Parms && 3670 call->is_CallDynamicJava()) { 3671 // Check for null receiver. In such case, the optimizer has 3672 // detected that the virtual call will always result in a null 3673 // pointer exception. The fall-through projection of this CatchNode 3674 // will not be populated. 3675 Node *arg0 = call->in(TypeFunc::Parms); 3676 if (arg0->is_Type() && 3677 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3678 required_outcnt--; 3679 } 3680 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3681 call->req() > TypeFunc::Parms+1 && 3682 call->is_CallStaticJava()) { 3683 // Check for negative array length. In such case, the optimizer has 3684 // detected that the allocation attempt will always result in an 3685 // exception. There is no fall-through projection of this CatchNode . 3686 Node *arg1 = call->in(TypeFunc::Parms+1); 3687 if (arg1->is_Type() && 3688 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3689 required_outcnt--; 3690 } 3691 } 3692 } 3693 } 3694 // Recheck with a better notion of 'required_outcnt' 3695 if (n->outcnt() != required_outcnt) { 3696 record_method_not_compilable("malformed control flow"); 3697 return true; // Not all targets reachable! 3698 } 3699 } 3700 // Check that I actually visited all kids. Unreached kids 3701 // must be infinite loops. 3702 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3703 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3704 record_method_not_compilable("infinite loop"); 3705 return true; // Found unvisited kid; must be unreach 3706 } 3707 3708 // Here so verification code in final_graph_reshaping_walk() 3709 // always see an OuterStripMinedLoopEnd 3710 if (n->is_OuterStripMinedLoopEnd()) { 3711 IfNode* init_iff = n->as_If(); 3712 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3713 n->subsume_by(iff, this); 3714 } 3715 } 3716 3717 // If original bytecodes contained a mixture of floats and doubles 3718 // check if the optimizer has made it homogenous, item (3). 3719 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3720 frc.get_float_count() > 32 && 3721 frc.get_double_count() == 0 && 3722 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3723 set_24_bit_selection_and_mode( false, true ); 3724 } 3725 3726 set_java_calls(frc.get_java_call_count()); 3727 set_inner_loops(frc.get_inner_loop_count()); 3728 3729 // No infinite loops, no reason to bail out. 3730 return false; 3731 } 3732 3733 //-----------------------------too_many_traps---------------------------------- 3734 // Report if there are too many traps at the current method and bci. 3735 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3736 bool Compile::too_many_traps(ciMethod* method, 3737 int bci, 3738 Deoptimization::DeoptReason reason) { 3739 ciMethodData* md = method->method_data(); 3740 if (md->is_empty()) { 3741 // Assume the trap has not occurred, or that it occurred only 3742 // because of a transient condition during start-up in the interpreter. 3743 return false; 3744 } 3745 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3746 if (md->has_trap_at(bci, m, reason) != 0) { 3747 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3748 // Also, if there are multiple reasons, or if there is no per-BCI record, 3749 // assume the worst. 3750 if (log()) 3751 log()->elem("observe trap='%s' count='%d'", 3752 Deoptimization::trap_reason_name(reason), 3753 md->trap_count(reason)); 3754 return true; 3755 } else { 3756 // Ignore method/bci and see if there have been too many globally. 3757 return too_many_traps(reason, md); 3758 } 3759 } 3760 3761 // Less-accurate variant which does not require a method and bci. 3762 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3763 ciMethodData* logmd) { 3764 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3765 // Too many traps globally. 3766 // Note that we use cumulative trap_count, not just md->trap_count. 3767 if (log()) { 3768 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3769 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3770 Deoptimization::trap_reason_name(reason), 3771 mcount, trap_count(reason)); 3772 } 3773 return true; 3774 } else { 3775 // The coast is clear. 3776 return false; 3777 } 3778 } 3779 3780 //--------------------------too_many_recompiles-------------------------------- 3781 // Report if there are too many recompiles at the current method and bci. 3782 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3783 // Is not eager to return true, since this will cause the compiler to use 3784 // Action_none for a trap point, to avoid too many recompilations. 3785 bool Compile::too_many_recompiles(ciMethod* method, 3786 int bci, 3787 Deoptimization::DeoptReason reason) { 3788 ciMethodData* md = method->method_data(); 3789 if (md->is_empty()) { 3790 // Assume the trap has not occurred, or that it occurred only 3791 // because of a transient condition during start-up in the interpreter. 3792 return false; 3793 } 3794 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3795 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3796 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3797 Deoptimization::DeoptReason per_bc_reason 3798 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3799 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3800 if ((per_bc_reason == Deoptimization::Reason_none 3801 || md->has_trap_at(bci, m, reason) != 0) 3802 // The trap frequency measure we care about is the recompile count: 3803 && md->trap_recompiled_at(bci, m) 3804 && md->overflow_recompile_count() >= bc_cutoff) { 3805 // Do not emit a trap here if it has already caused recompilations. 3806 // Also, if there are multiple reasons, or if there is no per-BCI record, 3807 // assume the worst. 3808 if (log()) 3809 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3810 Deoptimization::trap_reason_name(reason), 3811 md->trap_count(reason), 3812 md->overflow_recompile_count()); 3813 return true; 3814 } else if (trap_count(reason) != 0 3815 && decompile_count() >= m_cutoff) { 3816 // Too many recompiles globally, and we have seen this sort of trap. 3817 // Use cumulative decompile_count, not just md->decompile_count. 3818 if (log()) 3819 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3820 Deoptimization::trap_reason_name(reason), 3821 md->trap_count(reason), trap_count(reason), 3822 md->decompile_count(), decompile_count()); 3823 return true; 3824 } else { 3825 // The coast is clear. 3826 return false; 3827 } 3828 } 3829 3830 // Compute when not to trap. Used by matching trap based nodes and 3831 // NullCheck optimization. 3832 void Compile::set_allowed_deopt_reasons() { 3833 _allowed_reasons = 0; 3834 if (is_method_compilation()) { 3835 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3836 assert(rs < BitsPerInt, "recode bit map"); 3837 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3838 _allowed_reasons |= nth_bit(rs); 3839 } 3840 } 3841 } 3842 } 3843 3844 bool Compile::is_compiling_clinit_for(ciKlass* k) { 3845 ciMethod* root = method(); // the root method of compilation 3846 return root->is_static_initializer() && root->holder() == k; // access in the context of clinit 3847 } 3848 3849 #ifndef PRODUCT 3850 //------------------------------verify_graph_edges--------------------------- 3851 // Walk the Graph and verify that there is a one-to-one correspondence 3852 // between Use-Def edges and Def-Use edges in the graph. 3853 void Compile::verify_graph_edges(bool no_dead_code) { 3854 if (VerifyGraphEdges) { 3855 ResourceArea *area = Thread::current()->resource_area(); 3856 Unique_Node_List visited(area); 3857 // Call recursive graph walk to check edges 3858 _root->verify_edges(visited); 3859 if (no_dead_code) { 3860 // Now make sure that no visited node is used by an unvisited node. 3861 bool dead_nodes = false; 3862 Unique_Node_List checked(area); 3863 while (visited.size() > 0) { 3864 Node* n = visited.pop(); 3865 checked.push(n); 3866 for (uint i = 0; i < n->outcnt(); i++) { 3867 Node* use = n->raw_out(i); 3868 if (checked.member(use)) continue; // already checked 3869 if (visited.member(use)) continue; // already in the graph 3870 if (use->is_Con()) continue; // a dead ConNode is OK 3871 // At this point, we have found a dead node which is DU-reachable. 3872 if (!dead_nodes) { 3873 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3874 dead_nodes = true; 3875 } 3876 use->dump(2); 3877 tty->print_cr("---"); 3878 checked.push(use); // No repeats; pretend it is now checked. 3879 } 3880 } 3881 assert(!dead_nodes, "using nodes must be reachable from root"); 3882 } 3883 } 3884 } 3885 #endif 3886 3887 // The Compile object keeps track of failure reasons separately from the ciEnv. 3888 // This is required because there is not quite a 1-1 relation between the 3889 // ciEnv and its compilation task and the Compile object. Note that one 3890 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3891 // to backtrack and retry without subsuming loads. Other than this backtracking 3892 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3893 // by the logic in C2Compiler. 3894 void Compile::record_failure(const char* reason) { 3895 if (log() != NULL) { 3896 log()->elem("failure reason='%s' phase='compile'", reason); 3897 } 3898 if (_failure_reason == NULL) { 3899 // Record the first failure reason. 3900 _failure_reason = reason; 3901 } 3902 3903 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3904 C->print_method(PHASE_FAILURE); 3905 } 3906 _root = NULL; // flush the graph, too 3907 } 3908 3909 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3910 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3911 _phase_name(name), _dolog(CITimeVerbose) 3912 { 3913 if (_dolog) { 3914 C = Compile::current(); 3915 _log = C->log(); 3916 } else { 3917 C = NULL; 3918 _log = NULL; 3919 } 3920 if (_log != NULL) { 3921 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3922 _log->stamp(); 3923 _log->end_head(); 3924 } 3925 } 3926 3927 Compile::TracePhase::~TracePhase() { 3928 3929 C = Compile::current(); 3930 if (_dolog) { 3931 _log = C->log(); 3932 } else { 3933 _log = NULL; 3934 } 3935 3936 #ifdef ASSERT 3937 if (PrintIdealNodeCount) { 3938 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3939 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3940 } 3941 3942 if (VerifyIdealNodeCount) { 3943 Compile::current()->print_missing_nodes(); 3944 } 3945 #endif 3946 3947 if (_log != NULL) { 3948 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3949 } 3950 } 3951 3952 //============================================================================= 3953 // Two Constant's are equal when the type and the value are equal. 3954 bool Compile::Constant::operator==(const Constant& other) { 3955 if (type() != other.type() ) return false; 3956 if (can_be_reused() != other.can_be_reused()) return false; 3957 // For floating point values we compare the bit pattern. 3958 switch (type()) { 3959 case T_INT: 3960 case T_FLOAT: return (_v._value.i == other._v._value.i); 3961 case T_LONG: 3962 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3963 case T_OBJECT: 3964 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3965 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3966 case T_METADATA: return (_v._metadata == other._v._metadata); 3967 default: ShouldNotReachHere(); return false; 3968 } 3969 } 3970 3971 static int type_to_size_in_bytes(BasicType t) { 3972 switch (t) { 3973 case T_INT: return sizeof(jint ); 3974 case T_LONG: return sizeof(jlong ); 3975 case T_FLOAT: return sizeof(jfloat ); 3976 case T_DOUBLE: return sizeof(jdouble); 3977 case T_METADATA: return sizeof(Metadata*); 3978 // We use T_VOID as marker for jump-table entries (labels) which 3979 // need an internal word relocation. 3980 case T_VOID: 3981 case T_ADDRESS: 3982 case T_OBJECT: return sizeof(jobject); 3983 default: 3984 ShouldNotReachHere(); 3985 return -1; 3986 } 3987 } 3988 3989 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3990 // sort descending 3991 if (a->freq() > b->freq()) return -1; 3992 if (a->freq() < b->freq()) return 1; 3993 return 0; 3994 } 3995 3996 void Compile::ConstantTable::calculate_offsets_and_size() { 3997 // First, sort the array by frequencies. 3998 _constants.sort(qsort_comparator); 3999 4000 #ifdef ASSERT 4001 // Make sure all jump-table entries were sorted to the end of the 4002 // array (they have a negative frequency). 4003 bool found_void = false; 4004 for (int i = 0; i < _constants.length(); i++) { 4005 Constant con = _constants.at(i); 4006 if (con.type() == T_VOID) 4007 found_void = true; // jump-tables 4008 else 4009 assert(!found_void, "wrong sorting"); 4010 } 4011 #endif 4012 4013 int offset = 0; 4014 for (int i = 0; i < _constants.length(); i++) { 4015 Constant* con = _constants.adr_at(i); 4016 4017 // Align offset for type. 4018 int typesize = type_to_size_in_bytes(con->type()); 4019 offset = align_up(offset, typesize); 4020 con->set_offset(offset); // set constant's offset 4021 4022 if (con->type() == T_VOID) { 4023 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4024 offset = offset + typesize * n->outcnt(); // expand jump-table 4025 } else { 4026 offset = offset + typesize; 4027 } 4028 } 4029 4030 // Align size up to the next section start (which is insts; see 4031 // CodeBuffer::align_at_start). 4032 assert(_size == -1, "already set?"); 4033 _size = align_up(offset, (int)CodeEntryAlignment); 4034 } 4035 4036 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4037 MacroAssembler _masm(&cb); 4038 for (int i = 0; i < _constants.length(); i++) { 4039 Constant con = _constants.at(i); 4040 address constant_addr = NULL; 4041 switch (con.type()) { 4042 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4043 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4044 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4045 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4046 case T_OBJECT: { 4047 jobject obj = con.get_jobject(); 4048 int oop_index = _masm.oop_recorder()->find_index(obj); 4049 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4050 break; 4051 } 4052 case T_ADDRESS: { 4053 address addr = (address) con.get_jobject(); 4054 constant_addr = _masm.address_constant(addr); 4055 break; 4056 } 4057 // We use T_VOID as marker for jump-table entries (labels) which 4058 // need an internal word relocation. 4059 case T_VOID: { 4060 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4061 // Fill the jump-table with a dummy word. The real value is 4062 // filled in later in fill_jump_table. 4063 address dummy = (address) n; 4064 constant_addr = _masm.address_constant(dummy); 4065 // Expand jump-table 4066 for (uint i = 1; i < n->outcnt(); i++) { 4067 address temp_addr = _masm.address_constant(dummy + i); 4068 assert(temp_addr, "consts section too small"); 4069 } 4070 break; 4071 } 4072 case T_METADATA: { 4073 Metadata* obj = con.get_metadata(); 4074 int metadata_index = _masm.oop_recorder()->find_index(obj); 4075 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4076 break; 4077 } 4078 default: ShouldNotReachHere(); 4079 } 4080 assert(constant_addr, "consts section too small"); 4081 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4082 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4083 } 4084 } 4085 4086 int Compile::ConstantTable::find_offset(Constant& con) const { 4087 int idx = _constants.find(con); 4088 guarantee(idx != -1, "constant must be in constant table"); 4089 int offset = _constants.at(idx).offset(); 4090 guarantee(offset != -1, "constant table not emitted yet?"); 4091 return offset; 4092 } 4093 4094 void Compile::ConstantTable::add(Constant& con) { 4095 if (con.can_be_reused()) { 4096 int idx = _constants.find(con); 4097 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4098 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4099 return; 4100 } 4101 } 4102 (void) _constants.append(con); 4103 } 4104 4105 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4106 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4107 Constant con(type, value, b->_freq); 4108 add(con); 4109 return con; 4110 } 4111 4112 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4113 Constant con(metadata); 4114 add(con); 4115 return con; 4116 } 4117 4118 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4119 jvalue value; 4120 BasicType type = oper->type()->basic_type(); 4121 switch (type) { 4122 case T_LONG: value.j = oper->constantL(); break; 4123 case T_FLOAT: value.f = oper->constantF(); break; 4124 case T_DOUBLE: value.d = oper->constantD(); break; 4125 case T_OBJECT: 4126 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4127 case T_METADATA: return add((Metadata*)oper->constant()); break; 4128 default: guarantee(false, "unhandled type: %s", type2name(type)); 4129 } 4130 return add(n, type, value); 4131 } 4132 4133 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4134 jvalue value; 4135 // We can use the node pointer here to identify the right jump-table 4136 // as this method is called from Compile::Fill_buffer right before 4137 // the MachNodes are emitted and the jump-table is filled (means the 4138 // MachNode pointers do not change anymore). 4139 value.l = (jobject) n; 4140 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4141 add(con); 4142 return con; 4143 } 4144 4145 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4146 // If called from Compile::scratch_emit_size do nothing. 4147 if (Compile::current()->in_scratch_emit_size()) return; 4148 4149 assert(labels.is_nonempty(), "must be"); 4150 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4151 4152 // Since MachConstantNode::constant_offset() also contains 4153 // table_base_offset() we need to subtract the table_base_offset() 4154 // to get the plain offset into the constant table. 4155 int offset = n->constant_offset() - table_base_offset(); 4156 4157 MacroAssembler _masm(&cb); 4158 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4159 4160 for (uint i = 0; i < n->outcnt(); i++) { 4161 address* constant_addr = &jump_table_base[i]; 4162 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)); 4163 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4164 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4165 } 4166 } 4167 4168 //----------------------------static_subtype_check----------------------------- 4169 // Shortcut important common cases when superklass is exact: 4170 // (0) superklass is java.lang.Object (can occur in reflective code) 4171 // (1) subklass is already limited to a subtype of superklass => always ok 4172 // (2) subklass does not overlap with superklass => always fail 4173 // (3) superklass has NO subtypes and we can check with a simple compare. 4174 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4175 if (StressReflectiveCode) { 4176 return SSC_full_test; // Let caller generate the general case. 4177 } 4178 4179 if (superk == env()->Object_klass()) { 4180 return SSC_always_true; // (0) this test cannot fail 4181 } 4182 4183 ciType* superelem = superk; 4184 if (superelem->is_array_klass()) 4185 superelem = superelem->as_array_klass()->base_element_type(); 4186 4187 if (!subk->is_interface()) { // cannot trust static interface types yet 4188 if (subk->is_subtype_of(superk)) { 4189 return SSC_always_true; // (1) false path dead; no dynamic test needed 4190 } 4191 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4192 !superk->is_subtype_of(subk)) { 4193 return SSC_always_false; 4194 } 4195 } 4196 4197 // If casting to an instance klass, it must have no subtypes 4198 if (superk->is_interface()) { 4199 // Cannot trust interfaces yet. 4200 // %%% S.B. superk->nof_implementors() == 1 4201 } else if (superelem->is_instance_klass()) { 4202 ciInstanceKlass* ik = superelem->as_instance_klass(); 4203 if (!ik->has_subklass() && !ik->is_interface()) { 4204 if (!ik->is_final()) { 4205 // Add a dependency if there is a chance of a later subclass. 4206 dependencies()->assert_leaf_type(ik); 4207 } 4208 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4209 } 4210 } else { 4211 // A primitive array type has no subtypes. 4212 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4213 } 4214 4215 return SSC_full_test; 4216 } 4217 4218 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4219 #ifdef _LP64 4220 // The scaled index operand to AddP must be a clean 64-bit value. 4221 // Java allows a 32-bit int to be incremented to a negative 4222 // value, which appears in a 64-bit register as a large 4223 // positive number. Using that large positive number as an 4224 // operand in pointer arithmetic has bad consequences. 4225 // On the other hand, 32-bit overflow is rare, and the possibility 4226 // can often be excluded, if we annotate the ConvI2L node with 4227 // a type assertion that its value is known to be a small positive 4228 // number. (The prior range check has ensured this.) 4229 // This assertion is used by ConvI2LNode::Ideal. 4230 int index_max = max_jint - 1; // array size is max_jint, index is one less 4231 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4232 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4233 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4234 #endif 4235 return idx; 4236 } 4237 4238 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4239 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4240 if (ctrl != NULL) { 4241 // Express control dependency by a CastII node with a narrow type. 4242 value = new CastIINode(value, itype, false, true /* range check dependency */); 4243 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4244 // node from floating above the range check during loop optimizations. Otherwise, the 4245 // ConvI2L node may be eliminated independently of the range check, causing the data path 4246 // to become TOP while the control path is still there (although it's unreachable). 4247 value->set_req(0, ctrl); 4248 // Save CastII node to remove it after loop optimizations. 4249 phase->C->add_range_check_cast(value); 4250 value = phase->transform(value); 4251 } 4252 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4253 return phase->transform(new ConvI2LNode(value, ltype)); 4254 } 4255 4256 // The message about the current inlining is accumulated in 4257 // _print_inlining_stream and transfered into the _print_inlining_list 4258 // once we know whether inlining succeeds or not. For regular 4259 // inlining, messages are appended to the buffer pointed by 4260 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4261 // a new buffer is added after _print_inlining_idx in the list. This 4262 // way we can update the inlining message for late inlining call site 4263 // when the inlining is attempted again. 4264 void Compile::print_inlining_init() { 4265 if (print_inlining() || print_intrinsics()) { 4266 _print_inlining_stream = new stringStream(); 4267 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4268 } 4269 } 4270 4271 void Compile::print_inlining_reinit() { 4272 if (print_inlining() || print_intrinsics()) { 4273 // Re allocate buffer when we change ResourceMark 4274 _print_inlining_stream = new stringStream(); 4275 } 4276 } 4277 4278 void Compile::print_inlining_reset() { 4279 _print_inlining_stream->reset(); 4280 } 4281 4282 void Compile::print_inlining_commit() { 4283 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4284 // Transfer the message from _print_inlining_stream to the current 4285 // _print_inlining_list buffer and clear _print_inlining_stream. 4286 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4287 print_inlining_reset(); 4288 } 4289 4290 void Compile::print_inlining_push() { 4291 // Add new buffer to the _print_inlining_list at current position 4292 _print_inlining_idx++; 4293 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4294 } 4295 4296 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4297 return _print_inlining_list->at(_print_inlining_idx); 4298 } 4299 4300 void Compile::print_inlining_update(CallGenerator* cg) { 4301 if (print_inlining() || print_intrinsics()) { 4302 if (!cg->is_late_inline()) { 4303 if (print_inlining_current().cg() != NULL) { 4304 print_inlining_push(); 4305 } 4306 print_inlining_commit(); 4307 } else { 4308 if (print_inlining_current().cg() != cg && 4309 (print_inlining_current().cg() != NULL || 4310 print_inlining_current().ss()->size() != 0)) { 4311 print_inlining_push(); 4312 } 4313 print_inlining_commit(); 4314 print_inlining_current().set_cg(cg); 4315 } 4316 } 4317 } 4318 4319 void Compile::print_inlining_move_to(CallGenerator* cg) { 4320 // We resume inlining at a late inlining call site. Locate the 4321 // corresponding inlining buffer so that we can update it. 4322 if (print_inlining()) { 4323 for (int i = 0; i < _print_inlining_list->length(); i++) { 4324 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4325 _print_inlining_idx = i; 4326 return; 4327 } 4328 } 4329 ShouldNotReachHere(); 4330 } 4331 } 4332 4333 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4334 if (print_inlining()) { 4335 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4336 assert(print_inlining_current().cg() == cg, "wrong entry"); 4337 // replace message with new message 4338 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4339 print_inlining_commit(); 4340 print_inlining_current().set_cg(cg); 4341 } 4342 } 4343 4344 void Compile::print_inlining_assert_ready() { 4345 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4346 } 4347 4348 void Compile::process_print_inlining() { 4349 bool do_print_inlining = print_inlining() || print_intrinsics(); 4350 if (do_print_inlining || log() != NULL) { 4351 // Print inlining message for candidates that we couldn't inline 4352 // for lack of space 4353 for (int i = 0; i < _late_inlines.length(); i++) { 4354 CallGenerator* cg = _late_inlines.at(i); 4355 if (!cg->is_mh_late_inline()) { 4356 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4357 if (do_print_inlining) { 4358 cg->print_inlining_late(msg); 4359 } 4360 log_late_inline_failure(cg, msg); 4361 } 4362 } 4363 } 4364 if (do_print_inlining) { 4365 ResourceMark rm; 4366 stringStream ss; 4367 for (int i = 0; i < _print_inlining_list->length(); i++) { 4368 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4369 } 4370 size_t end = ss.size(); 4371 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4372 strncpy(_print_inlining_output, ss.base(), end+1); 4373 _print_inlining_output[end] = 0; 4374 } 4375 } 4376 4377 void Compile::dump_print_inlining() { 4378 if (_print_inlining_output != NULL) { 4379 tty->print_raw(_print_inlining_output); 4380 } 4381 } 4382 4383 void Compile::log_late_inline(CallGenerator* cg) { 4384 if (log() != NULL) { 4385 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4386 cg->unique_id()); 4387 JVMState* p = cg->call_node()->jvms(); 4388 while (p != NULL) { 4389 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4390 p = p->caller(); 4391 } 4392 log()->tail("late_inline"); 4393 } 4394 } 4395 4396 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4397 log_late_inline(cg); 4398 if (log() != NULL) { 4399 log()->inline_fail(msg); 4400 } 4401 } 4402 4403 void Compile::log_inline_id(CallGenerator* cg) { 4404 if (log() != NULL) { 4405 // The LogCompilation tool needs a unique way to identify late 4406 // inline call sites. This id must be unique for this call site in 4407 // this compilation. Try to have it unique across compilations as 4408 // well because it can be convenient when grepping through the log 4409 // file. 4410 // Distinguish OSR compilations from others in case CICountOSR is 4411 // on. 4412 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4413 cg->set_unique_id(id); 4414 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4415 } 4416 } 4417 4418 void Compile::log_inline_failure(const char* msg) { 4419 if (C->log() != NULL) { 4420 C->log()->inline_fail(msg); 4421 } 4422 } 4423 4424 4425 // Dump inlining replay data to the stream. 4426 // Don't change thread state and acquire any locks. 4427 void Compile::dump_inline_data(outputStream* out) { 4428 InlineTree* inl_tree = ilt(); 4429 if (inl_tree != NULL) { 4430 out->print(" inline %d", inl_tree->count()); 4431 inl_tree->dump_replay_data(out); 4432 } 4433 } 4434 4435 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4436 if (n1->Opcode() < n2->Opcode()) return -1; 4437 else if (n1->Opcode() > n2->Opcode()) return 1; 4438 4439 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4440 for (uint i = 1; i < n1->req(); i++) { 4441 if (n1->in(i) < n2->in(i)) return -1; 4442 else if (n1->in(i) > n2->in(i)) return 1; 4443 } 4444 4445 return 0; 4446 } 4447 4448 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4449 Node* n1 = *n1p; 4450 Node* n2 = *n2p; 4451 4452 return cmp_expensive_nodes(n1, n2); 4453 } 4454 4455 void Compile::sort_expensive_nodes() { 4456 if (!expensive_nodes_sorted()) { 4457 _expensive_nodes->sort(cmp_expensive_nodes); 4458 } 4459 } 4460 4461 bool Compile::expensive_nodes_sorted() const { 4462 for (int i = 1; i < _expensive_nodes->length(); i++) { 4463 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4464 return false; 4465 } 4466 } 4467 return true; 4468 } 4469 4470 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4471 if (_expensive_nodes->length() == 0) { 4472 return false; 4473 } 4474 4475 assert(OptimizeExpensiveOps, "optimization off?"); 4476 4477 // Take this opportunity to remove dead nodes from the list 4478 int j = 0; 4479 for (int i = 0; i < _expensive_nodes->length(); i++) { 4480 Node* n = _expensive_nodes->at(i); 4481 if (!n->is_unreachable(igvn)) { 4482 assert(n->is_expensive(), "should be expensive"); 4483 _expensive_nodes->at_put(j, n); 4484 j++; 4485 } 4486 } 4487 _expensive_nodes->trunc_to(j); 4488 4489 // Then sort the list so that similar nodes are next to each other 4490 // and check for at least two nodes of identical kind with same data 4491 // inputs. 4492 sort_expensive_nodes(); 4493 4494 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4495 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4496 return true; 4497 } 4498 } 4499 4500 return false; 4501 } 4502 4503 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4504 if (_expensive_nodes->length() == 0) { 4505 return; 4506 } 4507 4508 assert(OptimizeExpensiveOps, "optimization off?"); 4509 4510 // Sort to bring similar nodes next to each other and clear the 4511 // control input of nodes for which there's only a single copy. 4512 sort_expensive_nodes(); 4513 4514 int j = 0; 4515 int identical = 0; 4516 int i = 0; 4517 bool modified = false; 4518 for (; i < _expensive_nodes->length()-1; i++) { 4519 assert(j <= i, "can't write beyond current index"); 4520 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4521 identical++; 4522 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4523 continue; 4524 } 4525 if (identical > 0) { 4526 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4527 identical = 0; 4528 } else { 4529 Node* n = _expensive_nodes->at(i); 4530 igvn.replace_input_of(n, 0, NULL); 4531 igvn.hash_insert(n); 4532 modified = true; 4533 } 4534 } 4535 if (identical > 0) { 4536 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4537 } else if (_expensive_nodes->length() >= 1) { 4538 Node* n = _expensive_nodes->at(i); 4539 igvn.replace_input_of(n, 0, NULL); 4540 igvn.hash_insert(n); 4541 modified = true; 4542 } 4543 _expensive_nodes->trunc_to(j); 4544 if (modified) { 4545 igvn.optimize(); 4546 } 4547 } 4548 4549 void Compile::add_expensive_node(Node * n) { 4550 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4551 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4552 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4553 if (OptimizeExpensiveOps) { 4554 _expensive_nodes->append(n); 4555 } else { 4556 // Clear control input and let IGVN optimize expensive nodes if 4557 // OptimizeExpensiveOps is off. 4558 n->set_req(0, NULL); 4559 } 4560 } 4561 4562 /** 4563 * Remove the speculative part of types and clean up the graph 4564 */ 4565 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4566 if (UseTypeSpeculation) { 4567 Unique_Node_List worklist; 4568 worklist.push(root()); 4569 int modified = 0; 4570 // Go over all type nodes that carry a speculative type, drop the 4571 // speculative part of the type and enqueue the node for an igvn 4572 // which may optimize it out. 4573 for (uint next = 0; next < worklist.size(); ++next) { 4574 Node *n = worklist.at(next); 4575 if (n->is_Type()) { 4576 TypeNode* tn = n->as_Type(); 4577 const Type* t = tn->type(); 4578 const Type* t_no_spec = t->remove_speculative(); 4579 if (t_no_spec != t) { 4580 bool in_hash = igvn.hash_delete(n); 4581 assert(in_hash, "node should be in igvn hash table"); 4582 tn->set_type(t_no_spec); 4583 igvn.hash_insert(n); 4584 igvn._worklist.push(n); // give it a chance to go away 4585 modified++; 4586 } 4587 } 4588 uint max = n->len(); 4589 for( uint i = 0; i < max; ++i ) { 4590 Node *m = n->in(i); 4591 if (not_a_node(m)) continue; 4592 worklist.push(m); 4593 } 4594 } 4595 // Drop the speculative part of all types in the igvn's type table 4596 igvn.remove_speculative_types(); 4597 if (modified > 0) { 4598 igvn.optimize(); 4599 } 4600 #ifdef ASSERT 4601 // Verify that after the IGVN is over no speculative type has resurfaced 4602 worklist.clear(); 4603 worklist.push(root()); 4604 for (uint next = 0; next < worklist.size(); ++next) { 4605 Node *n = worklist.at(next); 4606 const Type* t = igvn.type_or_null(n); 4607 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4608 if (n->is_Type()) { 4609 t = n->as_Type()->type(); 4610 assert(t == t->remove_speculative(), "no more speculative types"); 4611 } 4612 uint max = n->len(); 4613 for( uint i = 0; i < max; ++i ) { 4614 Node *m = n->in(i); 4615 if (not_a_node(m)) continue; 4616 worklist.push(m); 4617 } 4618 } 4619 igvn.check_no_speculative_types(); 4620 #endif 4621 } 4622 } 4623 4624 // Auxiliary method to support randomized stressing/fuzzing. 4625 // 4626 // This method can be called the arbitrary number of times, with current count 4627 // as the argument. The logic allows selecting a single candidate from the 4628 // running list of candidates as follows: 4629 // int count = 0; 4630 // Cand* selected = null; 4631 // while(cand = cand->next()) { 4632 // if (randomized_select(++count)) { 4633 // selected = cand; 4634 // } 4635 // } 4636 // 4637 // Including count equalizes the chances any candidate is "selected". 4638 // This is useful when we don't have the complete list of candidates to choose 4639 // from uniformly. In this case, we need to adjust the randomicity of the 4640 // selection, or else we will end up biasing the selection towards the latter 4641 // candidates. 4642 // 4643 // Quick back-envelope calculation shows that for the list of n candidates 4644 // the equal probability for the candidate to persist as "best" can be 4645 // achieved by replacing it with "next" k-th candidate with the probability 4646 // of 1/k. It can be easily shown that by the end of the run, the 4647 // probability for any candidate is converged to 1/n, thus giving the 4648 // uniform distribution among all the candidates. 4649 // 4650 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4651 #define RANDOMIZED_DOMAIN_POW 29 4652 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4653 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4654 bool Compile::randomized_select(int count) { 4655 assert(count > 0, "only positive"); 4656 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4657 } 4658 4659 CloneMap& Compile::clone_map() { return _clone_map; } 4660 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4661 4662 void NodeCloneInfo::dump() const { 4663 tty->print(" {%d:%d} ", idx(), gen()); 4664 } 4665 4666 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4667 uint64_t val = value(old->_idx); 4668 NodeCloneInfo cio(val); 4669 assert(val != 0, "old node should be in the map"); 4670 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4671 insert(nnn->_idx, cin.get()); 4672 #ifndef PRODUCT 4673 if (is_debug()) { 4674 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4675 } 4676 #endif 4677 } 4678 4679 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4680 NodeCloneInfo cio(value(old->_idx)); 4681 if (cio.get() == 0) { 4682 cio.set(old->_idx, 0); 4683 insert(old->_idx, cio.get()); 4684 #ifndef PRODUCT 4685 if (is_debug()) { 4686 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4687 } 4688 #endif 4689 } 4690 clone(old, nnn, gen); 4691 } 4692 4693 int CloneMap::max_gen() const { 4694 int g = 0; 4695 DictI di(_dict); 4696 for(; di.test(); ++di) { 4697 int t = gen(di._key); 4698 if (g < t) { 4699 g = t; 4700 #ifndef PRODUCT 4701 if (is_debug()) { 4702 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4703 } 4704 #endif 4705 } 4706 } 4707 return g; 4708 } 4709 4710 void CloneMap::dump(node_idx_t key) const { 4711 uint64_t val = value(key); 4712 if (val != 0) { 4713 NodeCloneInfo ni(val); 4714 ni.dump(); 4715 } 4716 }