1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "gc/shared/barrierSet.hpp" 37 #include "gc/shared/c2/barrierSetC2.hpp" 38 #include "memory/resourceArea.hpp" 39 #include "opto/addnode.hpp" 40 #include "opto/block.hpp" 41 #include "opto/c2compiler.hpp" 42 #include "opto/callGenerator.hpp" 43 #include "opto/callnode.hpp" 44 #include "opto/castnode.hpp" 45 #include "opto/cfgnode.hpp" 46 #include "opto/chaitin.hpp" 47 #include "opto/compile.hpp" 48 #include "opto/connode.hpp" 49 #include "opto/convertnode.hpp" 50 #include "opto/divnode.hpp" 51 #include "opto/escape.hpp" 52 #include "opto/idealGraphPrinter.hpp" 53 #include "opto/loopnode.hpp" 54 #include "opto/machnode.hpp" 55 #include "opto/macro.hpp" 56 #include "opto/matcher.hpp" 57 #include "opto/mathexactnode.hpp" 58 #include "opto/memnode.hpp" 59 #include "opto/mulnode.hpp" 60 #include "opto/narrowptrnode.hpp" 61 #include "opto/node.hpp" 62 #include "opto/opcodes.hpp" 63 #include "opto/output.hpp" 64 #include "opto/parse.hpp" 65 #include "opto/phaseX.hpp" 66 #include "opto/rootnode.hpp" 67 #include "opto/runtime.hpp" 68 #include "opto/stringopts.hpp" 69 #include "opto/type.hpp" 70 #include "opto/valuetypenode.hpp" 71 #include "opto/vectornode.hpp" 72 #include "runtime/arguments.hpp" 73 #include "runtime/sharedRuntime.hpp" 74 #include "runtime/signature.hpp" 75 #include "runtime/stubRoutines.hpp" 76 #include "runtime/timer.hpp" 77 #include "utilities/align.hpp" 78 #include "utilities/copy.hpp" 79 #include "utilities/macros.hpp" 80 #if INCLUDE_G1GC 81 #include "gc/g1/g1ThreadLocalData.hpp" 82 #endif // INCLUDE_G1GC 83 #if INCLUDE_ZGC 84 #include "gc/z/c2/zBarrierSetC2.hpp" 85 #endif 86 87 88 // -------------------- Compile::mach_constant_base_node ----------------------- 89 // Constant table base node singleton. 90 MachConstantBaseNode* Compile::mach_constant_base_node() { 91 if (_mach_constant_base_node == NULL) { 92 _mach_constant_base_node = new MachConstantBaseNode(); 93 _mach_constant_base_node->add_req(C->root()); 94 } 95 return _mach_constant_base_node; 96 } 97 98 99 /// Support for intrinsics. 100 101 // Return the index at which m must be inserted (or already exists). 102 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 103 class IntrinsicDescPair { 104 private: 105 ciMethod* _m; 106 bool _is_virtual; 107 public: 108 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 109 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 110 ciMethod* m= elt->method(); 111 ciMethod* key_m = key->_m; 112 if (key_m < m) return -1; 113 else if (key_m > m) return 1; 114 else { 115 bool is_virtual = elt->is_virtual(); 116 bool key_virtual = key->_is_virtual; 117 if (key_virtual < is_virtual) return -1; 118 else if (key_virtual > is_virtual) return 1; 119 else return 0; 120 } 121 } 122 }; 123 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 124 #ifdef ASSERT 125 for (int i = 1; i < _intrinsics->length(); i++) { 126 CallGenerator* cg1 = _intrinsics->at(i-1); 127 CallGenerator* cg2 = _intrinsics->at(i); 128 assert(cg1->method() != cg2->method() 129 ? cg1->method() < cg2->method() 130 : cg1->is_virtual() < cg2->is_virtual(), 131 "compiler intrinsics list must stay sorted"); 132 } 133 #endif 134 IntrinsicDescPair pair(m, is_virtual); 135 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 136 } 137 138 void Compile::register_intrinsic(CallGenerator* cg) { 139 if (_intrinsics == NULL) { 140 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 141 } 142 int len = _intrinsics->length(); 143 bool found = false; 144 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 145 assert(!found, "registering twice"); 146 _intrinsics->insert_before(index, cg); 147 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 148 } 149 150 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 151 assert(m->is_loaded(), "don't try this on unloaded methods"); 152 if (_intrinsics != NULL) { 153 bool found = false; 154 int index = intrinsic_insertion_index(m, is_virtual, found); 155 if (found) { 156 return _intrinsics->at(index); 157 } 158 } 159 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 160 if (m->intrinsic_id() != vmIntrinsics::_none && 161 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 162 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 163 if (cg != NULL) { 164 // Save it for next time: 165 register_intrinsic(cg); 166 return cg; 167 } else { 168 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 169 } 170 } 171 return NULL; 172 } 173 174 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 175 // in library_call.cpp. 176 177 178 #ifndef PRODUCT 179 // statistics gathering... 180 181 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 182 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 183 184 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 185 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 186 int oflags = _intrinsic_hist_flags[id]; 187 assert(flags != 0, "what happened?"); 188 if (is_virtual) { 189 flags |= _intrinsic_virtual; 190 } 191 bool changed = (flags != oflags); 192 if ((flags & _intrinsic_worked) != 0) { 193 juint count = (_intrinsic_hist_count[id] += 1); 194 if (count == 1) { 195 changed = true; // first time 196 } 197 // increment the overall count also: 198 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 199 } 200 if (changed) { 201 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 202 // Something changed about the intrinsic's virtuality. 203 if ((flags & _intrinsic_virtual) != 0) { 204 // This is the first use of this intrinsic as a virtual call. 205 if (oflags != 0) { 206 // We already saw it as a non-virtual, so note both cases. 207 flags |= _intrinsic_both; 208 } 209 } else if ((oflags & _intrinsic_both) == 0) { 210 // This is the first use of this intrinsic as a non-virtual 211 flags |= _intrinsic_both; 212 } 213 } 214 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 215 } 216 // update the overall flags also: 217 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 218 return changed; 219 } 220 221 static char* format_flags(int flags, char* buf) { 222 buf[0] = 0; 223 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 224 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 225 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 226 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 227 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 228 if (buf[0] == 0) strcat(buf, ","); 229 assert(buf[0] == ',', "must be"); 230 return &buf[1]; 231 } 232 233 void Compile::print_intrinsic_statistics() { 234 char flagsbuf[100]; 235 ttyLocker ttyl; 236 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 237 tty->print_cr("Compiler intrinsic usage:"); 238 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 239 if (total == 0) total = 1; // avoid div0 in case of no successes 240 #define PRINT_STAT_LINE(name, c, f) \ 241 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 242 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 243 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 244 int flags = _intrinsic_hist_flags[id]; 245 juint count = _intrinsic_hist_count[id]; 246 if ((flags | count) != 0) { 247 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 248 } 249 } 250 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 251 if (xtty != NULL) xtty->tail("statistics"); 252 } 253 254 void Compile::print_statistics() { 255 { ttyLocker ttyl; 256 if (xtty != NULL) xtty->head("statistics type='opto'"); 257 Parse::print_statistics(); 258 PhaseCCP::print_statistics(); 259 PhaseRegAlloc::print_statistics(); 260 Scheduling::print_statistics(); 261 PhasePeephole::print_statistics(); 262 PhaseIdealLoop::print_statistics(); 263 if (xtty != NULL) xtty->tail("statistics"); 264 } 265 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 266 // put this under its own <statistics> element. 267 print_intrinsic_statistics(); 268 } 269 } 270 #endif //PRODUCT 271 272 // Support for bundling info 273 Bundle* Compile::node_bundling(const Node *n) { 274 assert(valid_bundle_info(n), "oob"); 275 return &_node_bundling_base[n->_idx]; 276 } 277 278 bool Compile::valid_bundle_info(const Node *n) { 279 return (_node_bundling_limit > n->_idx); 280 } 281 282 283 void Compile::gvn_replace_by(Node* n, Node* nn) { 284 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 285 Node* use = n->last_out(i); 286 bool is_in_table = initial_gvn()->hash_delete(use); 287 uint uses_found = 0; 288 for (uint j = 0; j < use->len(); j++) { 289 if (use->in(j) == n) { 290 if (j < use->req()) 291 use->set_req(j, nn); 292 else 293 use->set_prec(j, nn); 294 uses_found++; 295 } 296 } 297 if (is_in_table) { 298 // reinsert into table 299 initial_gvn()->hash_find_insert(use); 300 } 301 record_for_igvn(use); 302 i -= uses_found; // we deleted 1 or more copies of this edge 303 } 304 } 305 306 307 static inline bool not_a_node(const Node* n) { 308 if (n == NULL) return true; 309 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 310 if (*(address*)n == badAddress) return true; // kill by Node::destruct 311 return false; 312 } 313 314 // Identify all nodes that are reachable from below, useful. 315 // Use breadth-first pass that records state in a Unique_Node_List, 316 // recursive traversal is slower. 317 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 318 int estimated_worklist_size = live_nodes(); 319 useful.map( estimated_worklist_size, NULL ); // preallocate space 320 321 // Initialize worklist 322 if (root() != NULL) { useful.push(root()); } 323 // If 'top' is cached, declare it useful to preserve cached node 324 if( cached_top_node() ) { useful.push(cached_top_node()); } 325 326 // Push all useful nodes onto the list, breadthfirst 327 for( uint next = 0; next < useful.size(); ++next ) { 328 assert( next < unique(), "Unique useful nodes < total nodes"); 329 Node *n = useful.at(next); 330 uint max = n->len(); 331 for( uint i = 0; i < max; ++i ) { 332 Node *m = n->in(i); 333 if (not_a_node(m)) continue; 334 useful.push(m); 335 } 336 } 337 } 338 339 // Update dead_node_list with any missing dead nodes using useful 340 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 341 void Compile::update_dead_node_list(Unique_Node_List &useful) { 342 uint max_idx = unique(); 343 VectorSet& useful_node_set = useful.member_set(); 344 345 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 346 // If node with index node_idx is not in useful set, 347 // mark it as dead in dead node list. 348 if (! useful_node_set.test(node_idx) ) { 349 record_dead_node(node_idx); 350 } 351 } 352 } 353 354 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 355 int shift = 0; 356 for (int i = 0; i < inlines->length(); i++) { 357 CallGenerator* cg = inlines->at(i); 358 CallNode* call = cg->call_node(); 359 if (shift > 0) { 360 inlines->at_put(i-shift, cg); 361 } 362 if (!useful.member(call)) { 363 shift++; 364 } 365 } 366 inlines->trunc_to(inlines->length()-shift); 367 } 368 369 // Disconnect all useless nodes by disconnecting those at the boundary. 370 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 371 uint next = 0; 372 while (next < useful.size()) { 373 Node *n = useful.at(next++); 374 if (n->is_SafePoint()) { 375 // We're done with a parsing phase. Replaced nodes are not valid 376 // beyond that point. 377 n->as_SafePoint()->delete_replaced_nodes(); 378 } 379 // Use raw traversal of out edges since this code removes out edges 380 int max = n->outcnt(); 381 for (int j = 0; j < max; ++j) { 382 Node* child = n->raw_out(j); 383 if (! useful.member(child)) { 384 assert(!child->is_top() || child != top(), 385 "If top is cached in Compile object it is in useful list"); 386 // Only need to remove this out-edge to the useless node 387 n->raw_del_out(j); 388 --j; 389 --max; 390 } 391 } 392 if (n->outcnt() == 1 && n->has_special_unique_user()) { 393 record_for_igvn(n->unique_out()); 394 } 395 } 396 // Remove useless macro and predicate opaq nodes 397 for (int i = C->macro_count()-1; i >= 0; i--) { 398 Node* n = C->macro_node(i); 399 if (!useful.member(n)) { 400 remove_macro_node(n); 401 } 402 } 403 // Remove useless CastII nodes with range check dependency 404 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 405 Node* cast = range_check_cast_node(i); 406 if (!useful.member(cast)) { 407 remove_range_check_cast(cast); 408 } 409 } 410 // Remove useless expensive nodes 411 for (int i = C->expensive_count()-1; i >= 0; i--) { 412 Node* n = C->expensive_node(i); 413 if (!useful.member(n)) { 414 remove_expensive_node(n); 415 } 416 } 417 // Remove useless Opaque4 nodes 418 for (int i = opaque4_count() - 1; i >= 0; i--) { 419 Node* opaq = opaque4_node(i); 420 if (!useful.member(opaq)) { 421 remove_opaque4_node(opaq); 422 } 423 } 424 // Remove useless value type nodes 425 if (_value_type_nodes != NULL) { 426 _value_type_nodes->remove_useless_nodes(useful.member_set()); 427 } 428 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 429 bs->eliminate_useless_gc_barriers(useful); 430 // clean up the late inline lists 431 remove_useless_late_inlines(&_string_late_inlines, useful); 432 remove_useless_late_inlines(&_boxing_late_inlines, useful); 433 remove_useless_late_inlines(&_late_inlines, useful); 434 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 435 } 436 437 //------------------------------frame_size_in_words----------------------------- 438 // frame_slots in units of words 439 int Compile::frame_size_in_words() const { 440 // shift is 0 in LP32 and 1 in LP64 441 const int shift = (LogBytesPerWord - LogBytesPerInt); 442 int words = _frame_slots >> shift; 443 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 444 return words; 445 } 446 447 // To bang the stack of this compiled method we use the stack size 448 // that the interpreter would need in case of a deoptimization. This 449 // removes the need to bang the stack in the deoptimization blob which 450 // in turn simplifies stack overflow handling. 451 int Compile::bang_size_in_bytes() const { 452 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 453 } 454 455 // ============================================================================ 456 //------------------------------CompileWrapper--------------------------------- 457 class CompileWrapper : public StackObj { 458 Compile *const _compile; 459 public: 460 CompileWrapper(Compile* compile); 461 462 ~CompileWrapper(); 463 }; 464 465 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 466 // the Compile* pointer is stored in the current ciEnv: 467 ciEnv* env = compile->env(); 468 assert(env == ciEnv::current(), "must already be a ciEnv active"); 469 assert(env->compiler_data() == NULL, "compile already active?"); 470 env->set_compiler_data(compile); 471 assert(compile == Compile::current(), "sanity"); 472 473 compile->set_type_dict(NULL); 474 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 475 compile->clone_map().set_clone_idx(0); 476 compile->set_type_hwm(NULL); 477 compile->set_type_last_size(0); 478 compile->set_last_tf(NULL, NULL); 479 compile->set_indexSet_arena(NULL); 480 compile->set_indexSet_free_block_list(NULL); 481 compile->init_type_arena(); 482 Type::Initialize(compile); 483 _compile->set_scratch_buffer_blob(NULL); 484 _compile->begin_method(); 485 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 486 } 487 CompileWrapper::~CompileWrapper() { 488 _compile->end_method(); 489 if (_compile->scratch_buffer_blob() != NULL) 490 BufferBlob::free(_compile->scratch_buffer_blob()); 491 _compile->env()->set_compiler_data(NULL); 492 } 493 494 495 //----------------------------print_compile_messages--------------------------- 496 void Compile::print_compile_messages() { 497 #ifndef PRODUCT 498 // Check if recompiling 499 if (_subsume_loads == false && PrintOpto) { 500 // Recompiling without allowing machine instructions to subsume loads 501 tty->print_cr("*********************************************************"); 502 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 503 tty->print_cr("*********************************************************"); 504 } 505 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 506 // Recompiling without escape analysis 507 tty->print_cr("*********************************************************"); 508 tty->print_cr("** Bailout: Recompile without escape analysis **"); 509 tty->print_cr("*********************************************************"); 510 } 511 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 512 // Recompiling without boxing elimination 513 tty->print_cr("*********************************************************"); 514 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 515 tty->print_cr("*********************************************************"); 516 } 517 if (C->directive()->BreakAtCompileOption) { 518 // Open the debugger when compiling this method. 519 tty->print("### Breaking when compiling: "); 520 method()->print_short_name(); 521 tty->cr(); 522 BREAKPOINT; 523 } 524 525 if( PrintOpto ) { 526 if (is_osr_compilation()) { 527 tty->print("[OSR]%3d", _compile_id); 528 } else { 529 tty->print("%3d", _compile_id); 530 } 531 } 532 #endif 533 } 534 535 536 //-----------------------init_scratch_buffer_blob------------------------------ 537 // Construct a temporary BufferBlob and cache it for this compile. 538 void Compile::init_scratch_buffer_blob(int const_size) { 539 // If there is already a scratch buffer blob allocated and the 540 // constant section is big enough, use it. Otherwise free the 541 // current and allocate a new one. 542 BufferBlob* blob = scratch_buffer_blob(); 543 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 544 // Use the current blob. 545 } else { 546 if (blob != NULL) { 547 BufferBlob::free(blob); 548 } 549 550 ResourceMark rm; 551 _scratch_const_size = const_size; 552 int locs_size = sizeof(relocInfo) * MAX_locs_size; 553 int slop = 2 * CodeSection::end_slop(); // space between sections 554 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size + slop + locs_size); 555 #ifdef ASSERT 556 if (C->has_scalarized_args()) { 557 // Oop verification for loading object fields from scalarized value types in the new entry point requires lots of space 558 size += 5120; 559 } 560 #endif 561 blob = BufferBlob::create("Compile::scratch_buffer", size); 562 // Record the buffer blob for next time. 563 set_scratch_buffer_blob(blob); 564 // Have we run out of code space? 565 if (scratch_buffer_blob() == NULL) { 566 // Let CompilerBroker disable further compilations. 567 record_failure("Not enough space for scratch buffer in CodeCache"); 568 return; 569 } 570 } 571 572 // Initialize the relocation buffers 573 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 574 set_scratch_locs_memory(locs_buf); 575 } 576 577 578 //-----------------------scratch_emit_size------------------------------------- 579 // Helper function that computes size by emitting code 580 uint Compile::scratch_emit_size(const Node* n) { 581 // Start scratch_emit_size section. 582 set_in_scratch_emit_size(true); 583 584 // Emit into a trash buffer and count bytes emitted. 585 // This is a pretty expensive way to compute a size, 586 // but it works well enough if seldom used. 587 // All common fixed-size instructions are given a size 588 // method by the AD file. 589 // Note that the scratch buffer blob and locs memory are 590 // allocated at the beginning of the compile task, and 591 // may be shared by several calls to scratch_emit_size. 592 // The allocation of the scratch buffer blob is particularly 593 // expensive, since it has to grab the code cache lock. 594 BufferBlob* blob = this->scratch_buffer_blob(); 595 assert(blob != NULL, "Initialize BufferBlob at start"); 596 assert(blob->size() > MAX_inst_size, "sanity"); 597 relocInfo* locs_buf = scratch_locs_memory(); 598 address blob_begin = blob->content_begin(); 599 address blob_end = (address)locs_buf; 600 assert(blob->contains(blob_end), "sanity"); 601 CodeBuffer buf(blob_begin, blob_end - blob_begin); 602 buf.initialize_consts_size(_scratch_const_size); 603 buf.initialize_stubs_size(MAX_stubs_size); 604 assert(locs_buf != NULL, "sanity"); 605 int lsize = MAX_locs_size / 3; 606 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 607 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 608 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 609 // Mark as scratch buffer. 610 buf.consts()->set_scratch_emit(); 611 buf.insts()->set_scratch_emit(); 612 buf.stubs()->set_scratch_emit(); 613 614 // Do the emission. 615 616 Label fakeL; // Fake label for branch instructions. 617 Label* saveL = NULL; 618 uint save_bnum = 0; 619 bool is_branch = n->is_MachBranch(); 620 if (is_branch) { 621 MacroAssembler masm(&buf); 622 masm.bind(fakeL); 623 n->as_MachBranch()->save_label(&saveL, &save_bnum); 624 n->as_MachBranch()->label_set(&fakeL, 0); 625 } else if (n->is_MachProlog()) { 626 saveL = ((MachPrologNode*)n)->_verified_entry; 627 ((MachPrologNode*)n)->_verified_entry = &fakeL; 628 } 629 n->emit(buf, this->regalloc()); 630 631 // Emitting into the scratch buffer should not fail 632 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 633 634 // Restore label. 635 if (is_branch) { 636 n->as_MachBranch()->label_set(saveL, save_bnum); 637 } else if (n->is_MachProlog()) { 638 ((MachPrologNode*)n)->_verified_entry = saveL; 639 } 640 641 // End scratch_emit_size section. 642 set_in_scratch_emit_size(false); 643 644 return buf.insts_size(); 645 } 646 647 648 // ============================================================================ 649 //------------------------------Compile standard------------------------------- 650 debug_only( int Compile::_debug_idx = 100000; ) 651 652 // Compile a method. entry_bci is -1 for normal compilations and indicates 653 // the continuation bci for on stack replacement. 654 655 656 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 657 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 658 : Phase(Compiler), 659 _compile_id(ci_env->compile_id()), 660 _save_argument_registers(false), 661 _subsume_loads(subsume_loads), 662 _do_escape_analysis(do_escape_analysis), 663 _eliminate_boxing(eliminate_boxing), 664 _method(target), 665 _entry_bci(osr_bci), 666 _stub_function(NULL), 667 _stub_name(NULL), 668 _stub_entry_point(NULL), 669 _max_node_limit(MaxNodeLimit), 670 _orig_pc_slot(0), 671 _orig_pc_slot_offset_in_bytes(0), 672 _sp_inc_slot(0), 673 _sp_inc_slot_offset_in_bytes(0), 674 _inlining_progress(false), 675 _inlining_incrementally(false), 676 _has_reserved_stack_access(target->has_reserved_stack_access()), 677 #ifndef PRODUCT 678 _trace_opto_output(directive->TraceOptoOutputOption), 679 #endif 680 _has_method_handle_invokes(false), 681 _comp_arena(mtCompiler), 682 _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())), 683 _env(ci_env), 684 _directive(directive), 685 _log(ci_env->log()), 686 _failure_reason(NULL), 687 _congraph(NULL), 688 #ifndef PRODUCT 689 _printer(IdealGraphPrinter::printer()), 690 #endif 691 _dead_node_list(comp_arena()), 692 _dead_node_count(0), 693 _node_arena(mtCompiler), 694 _old_arena(mtCompiler), 695 _mach_constant_base_node(NULL), 696 _Compile_types(mtCompiler), 697 _initial_gvn(NULL), 698 _for_igvn(NULL), 699 _warm_calls(NULL), 700 _late_inlines(comp_arena(), 2, 0, NULL), 701 _string_late_inlines(comp_arena(), 2, 0, NULL), 702 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 703 _late_inlines_pos(0), 704 _number_of_mh_late_inlines(0), 705 _print_inlining_stream(NULL), 706 _print_inlining_list(NULL), 707 _print_inlining_idx(0), 708 _print_inlining_output(NULL), 709 _replay_inline_data(NULL), 710 _java_calls(0), 711 _inner_loops(0), 712 _interpreter_frame_size(0), 713 _node_bundling_limit(0), 714 _node_bundling_base(NULL), 715 _code_buffer("Compile::Fill_buffer"), 716 _scratch_const_size(-1), 717 _in_scratch_emit_size(false) 718 #ifndef PRODUCT 719 , _in_dump_cnt(0) 720 #endif 721 { 722 C = this; 723 #ifndef PRODUCT 724 if (_printer != NULL) { 725 _printer->set_compile(this); 726 } 727 #endif 728 CompileWrapper cw(this); 729 730 if (CITimeVerbose) { 731 tty->print(" "); 732 target->holder()->name()->print(); 733 tty->print("."); 734 target->print_short_name(); 735 tty->print(" "); 736 } 737 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 738 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 739 740 #ifndef PRODUCT 741 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 742 if (!print_opto_assembly) { 743 bool print_assembly = directive->PrintAssemblyOption; 744 if (print_assembly && !Disassembler::can_decode()) { 745 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 746 print_opto_assembly = true; 747 } 748 } 749 set_print_assembly(print_opto_assembly); 750 set_parsed_irreducible_loop(false); 751 752 if (directive->ReplayInlineOption) { 753 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 754 } 755 #endif 756 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 757 set_print_intrinsics(directive->PrintIntrinsicsOption); 758 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 759 760 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 761 // Make sure the method being compiled gets its own MDO, 762 // so we can at least track the decompile_count(). 763 // Need MDO to record RTM code generation state. 764 method()->ensure_method_data(); 765 } 766 767 Init(::AliasLevel); 768 769 770 print_compile_messages(); 771 772 _ilt = InlineTree::build_inline_tree_root(); 773 774 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 775 assert(num_alias_types() >= AliasIdxRaw, ""); 776 777 #define MINIMUM_NODE_HASH 1023 778 // Node list that Iterative GVN will start with 779 Unique_Node_List for_igvn(comp_arena()); 780 set_for_igvn(&for_igvn); 781 782 // GVN that will be run immediately on new nodes 783 uint estimated_size = method()->code_size()*4+64; 784 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 785 PhaseGVN gvn(node_arena(), estimated_size); 786 set_initial_gvn(&gvn); 787 788 print_inlining_init(); 789 { // Scope for timing the parser 790 TracePhase tp("parse", &timers[_t_parser]); 791 792 // Put top into the hash table ASAP. 793 initial_gvn()->transform_no_reclaim(top()); 794 795 // Set up tf(), start(), and find a CallGenerator. 796 CallGenerator* cg = NULL; 797 if (is_osr_compilation()) { 798 const TypeTuple *domain = StartOSRNode::osr_domain(); 799 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 800 init_tf(TypeFunc::make(domain, range)); 801 StartNode* s = new StartOSRNode(root(), domain); 802 initial_gvn()->set_type_bottom(s); 803 init_start(s); 804 cg = CallGenerator::for_osr(method(), entry_bci()); 805 } else { 806 // Normal case. 807 init_tf(TypeFunc::make(method())); 808 StartNode* s = new StartNode(root(), tf()->domain_cc()); 809 initial_gvn()->set_type_bottom(s); 810 init_start(s); 811 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 812 // With java.lang.ref.reference.get() we must go through the 813 // intrinsic - even when get() is the root 814 // method of the compile - so that, if necessary, the value in 815 // the referent field of the reference object gets recorded by 816 // the pre-barrier code. 817 cg = find_intrinsic(method(), false); 818 } 819 if (cg == NULL) { 820 float past_uses = method()->interpreter_invocation_count(); 821 float expected_uses = past_uses; 822 cg = CallGenerator::for_inline(method(), expected_uses); 823 } 824 } 825 if (failing()) return; 826 if (cg == NULL) { 827 record_method_not_compilable("cannot parse method"); 828 return; 829 } 830 JVMState* jvms = build_start_state(start(), tf()); 831 if ((jvms = cg->generate(jvms)) == NULL) { 832 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 833 record_method_not_compilable("method parse failed"); 834 } 835 return; 836 } 837 GraphKit kit(jvms); 838 839 if (!kit.stopped()) { 840 // Accept return values, and transfer control we know not where. 841 // This is done by a special, unique ReturnNode bound to root. 842 return_values(kit.jvms()); 843 } 844 845 if (kit.has_exceptions()) { 846 // Any exceptions that escape from this call must be rethrown 847 // to whatever caller is dynamically above us on the stack. 848 // This is done by a special, unique RethrowNode bound to root. 849 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 850 } 851 852 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 853 854 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 855 inline_string_calls(true); 856 } 857 858 if (failing()) return; 859 860 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 861 862 // Remove clutter produced by parsing. 863 if (!failing()) { 864 ResourceMark rm; 865 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 866 } 867 } 868 869 // Note: Large methods are capped off in do_one_bytecode(). 870 if (failing()) return; 871 872 // After parsing, node notes are no longer automagic. 873 // They must be propagated by register_new_node_with_optimizer(), 874 // clone(), or the like. 875 set_default_node_notes(NULL); 876 877 for (;;) { 878 int successes = Inline_Warm(); 879 if (failing()) return; 880 if (successes == 0) break; 881 } 882 883 // Drain the list. 884 Finish_Warm(); 885 #ifndef PRODUCT 886 if (_printer && _printer->should_print(1)) { 887 _printer->print_inlining(); 888 } 889 #endif 890 891 if (failing()) return; 892 NOT_PRODUCT( verify_graph_edges(); ) 893 894 // Now optimize 895 Optimize(); 896 if (failing()) return; 897 NOT_PRODUCT( verify_graph_edges(); ) 898 899 #ifndef PRODUCT 900 if (PrintIdeal) { 901 ttyLocker ttyl; // keep the following output all in one block 902 // This output goes directly to the tty, not the compiler log. 903 // To enable tools to match it up with the compilation activity, 904 // be sure to tag this tty output with the compile ID. 905 if (xtty != NULL) { 906 xtty->head("ideal compile_id='%d'%s", compile_id(), 907 is_osr_compilation() ? " compile_kind='osr'" : 908 ""); 909 } 910 root()->dump(9999); 911 if (xtty != NULL) { 912 xtty->tail("ideal"); 913 } 914 } 915 #endif 916 917 NOT_PRODUCT( verify_barriers(); ) 918 919 // Dump compilation data to replay it. 920 if (directive->DumpReplayOption) { 921 env()->dump_replay_data(_compile_id); 922 } 923 if (directive->DumpInlineOption && (ilt() != NULL)) { 924 env()->dump_inline_data(_compile_id); 925 } 926 927 // Now that we know the size of all the monitors we can add a fixed slot 928 // for the original deopt pc. 929 930 _orig_pc_slot = fixed_slots(); 931 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 932 933 if (method()->get_Method()->needs_stack_repair()) { 934 // One extra slot for the special stack increment value 935 _sp_inc_slot = next_slot; 936 next_slot += 2; 937 } 938 939 set_fixed_slots(next_slot); 940 941 // Compute when to use implicit null checks. Used by matching trap based 942 // nodes and NullCheck optimization. 943 set_allowed_deopt_reasons(); 944 945 // Now generate code 946 Code_Gen(); 947 if (failing()) return; 948 949 // Check if we want to skip execution of all compiled code. 950 { 951 #ifndef PRODUCT 952 if (OptoNoExecute) { 953 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 954 return; 955 } 956 #endif 957 TracePhase tp("install_code", &timers[_t_registerMethod]); 958 959 if (is_osr_compilation()) { 960 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 961 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 962 } else { 963 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 964 if (_code_offsets.value(CodeOffsets::Verified_Value_Entry) == -1) { 965 _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, _first_block_size); 966 } 967 if (_code_offsets.value(CodeOffsets::Entry) == -1) { 968 // We emitted a value type entry point, adjust normal entry 969 _code_offsets.set_value(CodeOffsets::Entry, _first_block_size); 970 } 971 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 972 } 973 974 env()->register_method(_method, _entry_bci, 975 &_code_offsets, 976 _orig_pc_slot_offset_in_bytes, 977 code_buffer(), 978 frame_size_in_words(), _oop_map_set, 979 &_handler_table, &_inc_table, 980 compiler, 981 has_unsafe_access(), 982 SharedRuntime::is_wide_vector(max_vector_size()), 983 rtm_state() 984 ); 985 986 if (log() != NULL) // Print code cache state into compiler log 987 log()->code_cache_state(); 988 } 989 } 990 991 //------------------------------Compile---------------------------------------- 992 // Compile a runtime stub 993 Compile::Compile( ciEnv* ci_env, 994 TypeFunc_generator generator, 995 address stub_function, 996 const char *stub_name, 997 int is_fancy_jump, 998 bool pass_tls, 999 bool save_arg_registers, 1000 bool return_pc, 1001 DirectiveSet* directive) 1002 : Phase(Compiler), 1003 _compile_id(0), 1004 _save_argument_registers(save_arg_registers), 1005 _subsume_loads(true), 1006 _do_escape_analysis(false), 1007 _eliminate_boxing(false), 1008 _method(NULL), 1009 _entry_bci(InvocationEntryBci), 1010 _stub_function(stub_function), 1011 _stub_name(stub_name), 1012 _stub_entry_point(NULL), 1013 _max_node_limit(MaxNodeLimit), 1014 _orig_pc_slot(0), 1015 _orig_pc_slot_offset_in_bytes(0), 1016 _sp_inc_slot(0), 1017 _sp_inc_slot_offset_in_bytes(0), 1018 _inlining_progress(false), 1019 _inlining_incrementally(false), 1020 _has_reserved_stack_access(false), 1021 #ifndef PRODUCT 1022 _trace_opto_output(directive->TraceOptoOutputOption), 1023 #endif 1024 _has_method_handle_invokes(false), 1025 _comp_arena(mtCompiler), 1026 _env(ci_env), 1027 _directive(directive), 1028 _log(ci_env->log()), 1029 _failure_reason(NULL), 1030 _congraph(NULL), 1031 #ifndef PRODUCT 1032 _printer(NULL), 1033 #endif 1034 _dead_node_list(comp_arena()), 1035 _dead_node_count(0), 1036 _node_arena(mtCompiler), 1037 _old_arena(mtCompiler), 1038 _mach_constant_base_node(NULL), 1039 _Compile_types(mtCompiler), 1040 _initial_gvn(NULL), 1041 _for_igvn(NULL), 1042 _warm_calls(NULL), 1043 _number_of_mh_late_inlines(0), 1044 _print_inlining_stream(NULL), 1045 _print_inlining_list(NULL), 1046 _print_inlining_idx(0), 1047 _print_inlining_output(NULL), 1048 _replay_inline_data(NULL), 1049 _java_calls(0), 1050 _inner_loops(0), 1051 _interpreter_frame_size(0), 1052 _node_bundling_limit(0), 1053 _node_bundling_base(NULL), 1054 _code_buffer("Compile::Fill_buffer"), 1055 #ifndef PRODUCT 1056 _in_dump_cnt(0), 1057 #endif 1058 _allowed_reasons(0) { 1059 C = this; 1060 1061 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1062 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1063 1064 #ifndef PRODUCT 1065 set_print_assembly(PrintFrameConverterAssembly); 1066 set_parsed_irreducible_loop(false); 1067 #endif 1068 set_has_irreducible_loop(false); // no loops 1069 1070 CompileWrapper cw(this); 1071 Init(/*AliasLevel=*/ 0); 1072 init_tf((*generator)()); 1073 1074 { 1075 // The following is a dummy for the sake of GraphKit::gen_stub 1076 Unique_Node_List for_igvn(comp_arena()); 1077 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1078 PhaseGVN gvn(Thread::current()->resource_area(),255); 1079 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1080 gvn.transform_no_reclaim(top()); 1081 1082 GraphKit kit; 1083 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1084 } 1085 1086 NOT_PRODUCT( verify_graph_edges(); ) 1087 Code_Gen(); 1088 if (failing()) return; 1089 1090 1091 // Entry point will be accessed using compile->stub_entry_point(); 1092 if (code_buffer() == NULL) { 1093 Matcher::soft_match_failure(); 1094 } else { 1095 if (PrintAssembly && (WizardMode || Verbose)) 1096 tty->print_cr("### Stub::%s", stub_name); 1097 1098 if (!failing()) { 1099 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1100 1101 // Make the NMethod 1102 // For now we mark the frame as never safe for profile stackwalking 1103 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1104 code_buffer(), 1105 CodeOffsets::frame_never_safe, 1106 // _code_offsets.value(CodeOffsets::Frame_Complete), 1107 frame_size_in_words(), 1108 _oop_map_set, 1109 save_arg_registers); 1110 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1111 1112 _stub_entry_point = rs->entry_point(); 1113 } 1114 } 1115 } 1116 1117 //------------------------------Init------------------------------------------- 1118 // Prepare for a single compilation 1119 void Compile::Init(int aliaslevel) { 1120 _unique = 0; 1121 _regalloc = NULL; 1122 1123 _tf = NULL; // filled in later 1124 _top = NULL; // cached later 1125 _matcher = NULL; // filled in later 1126 _cfg = NULL; // filled in later 1127 1128 set_24_bit_selection_and_mode(Use24BitFP, false); 1129 1130 _node_note_array = NULL; 1131 _default_node_notes = NULL; 1132 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1133 1134 _immutable_memory = NULL; // filled in at first inquiry 1135 1136 // Globally visible Nodes 1137 // First set TOP to NULL to give safe behavior during creation of RootNode 1138 set_cached_top_node(NULL); 1139 set_root(new RootNode()); 1140 // Now that you have a Root to point to, create the real TOP 1141 set_cached_top_node( new ConNode(Type::TOP) ); 1142 set_recent_alloc(NULL, NULL); 1143 1144 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1145 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1146 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1147 env()->set_dependencies(new Dependencies(env())); 1148 1149 _fixed_slots = 0; 1150 set_has_split_ifs(false); 1151 set_has_loops(has_method() && method()->has_loops()); // first approximation 1152 set_has_stringbuilder(false); 1153 set_has_boxed_value(false); 1154 _trap_can_recompile = false; // no traps emitted yet 1155 _major_progress = true; // start out assuming good things will happen 1156 set_has_unsafe_access(false); 1157 set_max_vector_size(0); 1158 set_clear_upper_avx(false); //false as default for clear upper bits of ymm registers 1159 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1160 set_decompile_count(0); 1161 1162 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1163 set_num_loop_opts(LoopOptsCount); 1164 set_do_inlining(Inline); 1165 set_max_inline_size(MaxInlineSize); 1166 set_freq_inline_size(FreqInlineSize); 1167 set_do_scheduling(OptoScheduling); 1168 set_do_count_invocations(false); 1169 set_do_method_data_update(false); 1170 1171 set_do_vector_loop(false); 1172 1173 if (AllowVectorizeOnDemand) { 1174 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1175 set_do_vector_loop(true); 1176 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());}) 1177 } else if (has_method() && method()->name() != 0 && 1178 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1179 set_do_vector_loop(true); 1180 } 1181 } 1182 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1183 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());}) 1184 1185 set_age_code(has_method() && method()->profile_aging()); 1186 set_rtm_state(NoRTM); // No RTM lock eliding by default 1187 _max_node_limit = _directive->MaxNodeLimitOption; 1188 1189 #if INCLUDE_RTM_OPT 1190 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1191 int rtm_state = method()->method_data()->rtm_state(); 1192 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1193 // Don't generate RTM lock eliding code. 1194 set_rtm_state(NoRTM); 1195 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1196 // Generate RTM lock eliding code without abort ratio calculation code. 1197 set_rtm_state(UseRTM); 1198 } else if (UseRTMDeopt) { 1199 // Generate RTM lock eliding code and include abort ratio calculation 1200 // code if UseRTMDeopt is on. 1201 set_rtm_state(ProfileRTM); 1202 } 1203 } 1204 #endif 1205 if (debug_info()->recording_non_safepoints()) { 1206 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1207 (comp_arena(), 8, 0, NULL)); 1208 set_default_node_notes(Node_Notes::make(this)); 1209 } 1210 1211 // // -- Initialize types before each compile -- 1212 // // Update cached type information 1213 // if( _method && _method->constants() ) 1214 // Type::update_loaded_types(_method, _method->constants()); 1215 1216 // Init alias_type map. 1217 if (!_do_escape_analysis && aliaslevel == 3) 1218 aliaslevel = 2; // No unique types without escape analysis 1219 _AliasLevel = aliaslevel; 1220 const int grow_ats = 16; 1221 _max_alias_types = grow_ats; 1222 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1223 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1224 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1225 { 1226 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1227 } 1228 // Initialize the first few types. 1229 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1230 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1231 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1232 _num_alias_types = AliasIdxRaw+1; 1233 // Zero out the alias type cache. 1234 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1235 // A NULL adr_type hits in the cache right away. Preload the right answer. 1236 probe_alias_cache(NULL)->_index = AliasIdxTop; 1237 1238 _intrinsics = NULL; 1239 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1240 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1241 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1242 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1243 _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1244 _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 1245 register_library_intrinsics(); 1246 } 1247 1248 //---------------------------init_start---------------------------------------- 1249 // Install the StartNode on this compile object. 1250 void Compile::init_start(StartNode* s) { 1251 if (failing()) 1252 return; // already failing 1253 assert(s == start(), ""); 1254 } 1255 1256 /** 1257 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1258 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1259 * the ideal graph. 1260 */ 1261 StartNode* Compile::start() const { 1262 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1263 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1264 Node* start = root()->fast_out(i); 1265 if (start->is_Start()) { 1266 return start->as_Start(); 1267 } 1268 } 1269 fatal("Did not find Start node!"); 1270 return NULL; 1271 } 1272 1273 //-------------------------------immutable_memory------------------------------------- 1274 // Access immutable memory 1275 Node* Compile::immutable_memory() { 1276 if (_immutable_memory != NULL) { 1277 return _immutable_memory; 1278 } 1279 StartNode* s = start(); 1280 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1281 Node *p = s->fast_out(i); 1282 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1283 _immutable_memory = p; 1284 return _immutable_memory; 1285 } 1286 } 1287 ShouldNotReachHere(); 1288 return NULL; 1289 } 1290 1291 //----------------------set_cached_top_node------------------------------------ 1292 // Install the cached top node, and make sure Node::is_top works correctly. 1293 void Compile::set_cached_top_node(Node* tn) { 1294 if (tn != NULL) verify_top(tn); 1295 Node* old_top = _top; 1296 _top = tn; 1297 // Calling Node::setup_is_top allows the nodes the chance to adjust 1298 // their _out arrays. 1299 if (_top != NULL) _top->setup_is_top(); 1300 if (old_top != NULL) old_top->setup_is_top(); 1301 assert(_top == NULL || top()->is_top(), ""); 1302 } 1303 1304 #ifdef ASSERT 1305 uint Compile::count_live_nodes_by_graph_walk() { 1306 Unique_Node_List useful(comp_arena()); 1307 // Get useful node list by walking the graph. 1308 identify_useful_nodes(useful); 1309 return useful.size(); 1310 } 1311 1312 void Compile::print_missing_nodes() { 1313 1314 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1315 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1316 return; 1317 } 1318 1319 // This is an expensive function. It is executed only when the user 1320 // specifies VerifyIdealNodeCount option or otherwise knows the 1321 // additional work that needs to be done to identify reachable nodes 1322 // by walking the flow graph and find the missing ones using 1323 // _dead_node_list. 1324 1325 Unique_Node_List useful(comp_arena()); 1326 // Get useful node list by walking the graph. 1327 identify_useful_nodes(useful); 1328 1329 uint l_nodes = C->live_nodes(); 1330 uint l_nodes_by_walk = useful.size(); 1331 1332 if (l_nodes != l_nodes_by_walk) { 1333 if (_log != NULL) { 1334 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1335 _log->stamp(); 1336 _log->end_head(); 1337 } 1338 VectorSet& useful_member_set = useful.member_set(); 1339 int last_idx = l_nodes_by_walk; 1340 for (int i = 0; i < last_idx; i++) { 1341 if (useful_member_set.test(i)) { 1342 if (_dead_node_list.test(i)) { 1343 if (_log != NULL) { 1344 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1345 } 1346 if (PrintIdealNodeCount) { 1347 // Print the log message to tty 1348 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1349 useful.at(i)->dump(); 1350 } 1351 } 1352 } 1353 else if (! _dead_node_list.test(i)) { 1354 if (_log != NULL) { 1355 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1356 } 1357 if (PrintIdealNodeCount) { 1358 // Print the log message to tty 1359 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1360 } 1361 } 1362 } 1363 if (_log != NULL) { 1364 _log->tail("mismatched_nodes"); 1365 } 1366 } 1367 } 1368 void Compile::record_modified_node(Node* n) { 1369 if (_modified_nodes != NULL && !_inlining_incrementally && 1370 n->outcnt() != 0 && !n->is_Con()) { 1371 _modified_nodes->push(n); 1372 } 1373 } 1374 1375 void Compile::remove_modified_node(Node* n) { 1376 if (_modified_nodes != NULL) { 1377 _modified_nodes->remove(n); 1378 } 1379 } 1380 #endif 1381 1382 #ifndef PRODUCT 1383 void Compile::verify_top(Node* tn) const { 1384 if (tn != NULL) { 1385 assert(tn->is_Con(), "top node must be a constant"); 1386 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1387 assert(tn->in(0) != NULL, "must have live top node"); 1388 } 1389 } 1390 #endif 1391 1392 1393 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1394 1395 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1396 guarantee(arr != NULL, ""); 1397 int num_blocks = arr->length(); 1398 if (grow_by < num_blocks) grow_by = num_blocks; 1399 int num_notes = grow_by * _node_notes_block_size; 1400 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1401 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1402 while (num_notes > 0) { 1403 arr->append(notes); 1404 notes += _node_notes_block_size; 1405 num_notes -= _node_notes_block_size; 1406 } 1407 assert(num_notes == 0, "exact multiple, please"); 1408 } 1409 1410 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1411 if (source == NULL || dest == NULL) return false; 1412 1413 if (dest->is_Con()) 1414 return false; // Do not push debug info onto constants. 1415 1416 #ifdef ASSERT 1417 // Leave a bread crumb trail pointing to the original node: 1418 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1419 dest->set_debug_orig(source); 1420 } 1421 #endif 1422 1423 if (node_note_array() == NULL) 1424 return false; // Not collecting any notes now. 1425 1426 // This is a copy onto a pre-existing node, which may already have notes. 1427 // If both nodes have notes, do not overwrite any pre-existing notes. 1428 Node_Notes* source_notes = node_notes_at(source->_idx); 1429 if (source_notes == NULL || source_notes->is_clear()) return false; 1430 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1431 if (dest_notes == NULL || dest_notes->is_clear()) { 1432 return set_node_notes_at(dest->_idx, source_notes); 1433 } 1434 1435 Node_Notes merged_notes = (*source_notes); 1436 // The order of operations here ensures that dest notes will win... 1437 merged_notes.update_from(dest_notes); 1438 return set_node_notes_at(dest->_idx, &merged_notes); 1439 } 1440 1441 1442 //--------------------------allow_range_check_smearing------------------------- 1443 // Gating condition for coalescing similar range checks. 1444 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1445 // single covering check that is at least as strong as any of them. 1446 // If the optimization succeeds, the simplified (strengthened) range check 1447 // will always succeed. If it fails, we will deopt, and then give up 1448 // on the optimization. 1449 bool Compile::allow_range_check_smearing() const { 1450 // If this method has already thrown a range-check, 1451 // assume it was because we already tried range smearing 1452 // and it failed. 1453 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1454 return !already_trapped; 1455 } 1456 1457 1458 //------------------------------flatten_alias_type----------------------------- 1459 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1460 int offset = tj->offset(); 1461 TypePtr::PTR ptr = tj->ptr(); 1462 1463 // Known instance (scalarizable allocation) alias only with itself. 1464 bool is_known_inst = tj->isa_oopptr() != NULL && 1465 tj->is_oopptr()->is_known_instance(); 1466 1467 // Process weird unsafe references. 1468 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1469 bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass(); 1470 assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops"); 1471 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1472 tj = TypeOopPtr::BOTTOM; 1473 ptr = tj->ptr(); 1474 offset = tj->offset(); 1475 } 1476 1477 // Array pointers need some flattening 1478 const TypeAryPtr *ta = tj->isa_aryptr(); 1479 if (ta && ta->is_stable()) { 1480 // Erase stability property for alias analysis. 1481 tj = ta = ta->cast_to_stable(false); 1482 } 1483 if( ta && is_known_inst ) { 1484 if ( offset != Type::OffsetBot && 1485 offset > arrayOopDesc::length_offset_in_bytes() ) { 1486 offset = Type::OffsetBot; // Flatten constant access into array body only 1487 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1488 } 1489 } else if( ta && _AliasLevel >= 2 ) { 1490 // For arrays indexed by constant indices, we flatten the alias 1491 // space to include all of the array body. Only the header, klass 1492 // and array length can be accessed un-aliased. 1493 // For flattened value type array, each field has its own slice so 1494 // we must include the field offset. 1495 if( offset != Type::OffsetBot ) { 1496 if( ta->const_oop() ) { // MethodData* or Method* 1497 offset = Type::OffsetBot; // Flatten constant access into array body 1498 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1499 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1500 // range is OK as-is. 1501 tj = ta = TypeAryPtr::RANGE; 1502 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1503 tj = TypeInstPtr::KLASS; // all klass loads look alike 1504 ta = TypeAryPtr::RANGE; // generic ignored junk 1505 ptr = TypePtr::BotPTR; 1506 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1507 tj = TypeInstPtr::MARK; 1508 ta = TypeAryPtr::RANGE; // generic ignored junk 1509 ptr = TypePtr::BotPTR; 1510 } else { // Random constant offset into array body 1511 offset = Type::OffsetBot; // Flatten constant access into array body 1512 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1513 } 1514 } 1515 // Arrays of fixed size alias with arrays of unknown size. 1516 if (ta->size() != TypeInt::POS) { 1517 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1518 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1519 } 1520 // Arrays of known objects become arrays of unknown objects. 1521 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1522 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1523 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1524 } 1525 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1526 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1527 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1528 } 1529 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1530 // cannot be distinguished by bytecode alone. 1531 if (ta->elem() == TypeInt::BOOL) { 1532 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1533 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1534 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1535 } 1536 // During the 2nd round of IterGVN, NotNull castings are removed. 1537 // Make sure the Bottom and NotNull variants alias the same. 1538 // Also, make sure exact and non-exact variants alias the same. 1539 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1540 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1541 } 1542 } 1543 1544 // Oop pointers need some flattening 1545 const TypeInstPtr *to = tj->isa_instptr(); 1546 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1547 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1548 if( ptr == TypePtr::Constant ) { 1549 if (to->klass() != ciEnv::current()->Class_klass() || 1550 offset < k->size_helper() * wordSize) { 1551 // No constant oop pointers (such as Strings); they alias with 1552 // unknown strings. 1553 assert(!is_known_inst, "not scalarizable allocation"); 1554 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1555 } 1556 } else if( is_known_inst ) { 1557 tj = to; // Keep NotNull and klass_is_exact for instance type 1558 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1559 // During the 2nd round of IterGVN, NotNull castings are removed. 1560 // Make sure the Bottom and NotNull variants alias the same. 1561 // Also, make sure exact and non-exact variants alias the same. 1562 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1563 } 1564 if (to->speculative() != NULL) { 1565 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1566 } 1567 // Canonicalize the holder of this field 1568 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1569 // First handle header references such as a LoadKlassNode, even if the 1570 // object's klass is unloaded at compile time (4965979). 1571 if (!is_known_inst) { // Do it only for non-instance types 1572 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1573 } 1574 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1575 // Static fields are in the space above the normal instance 1576 // fields in the java.lang.Class instance. 1577 if (to->klass() != ciEnv::current()->Class_klass()) { 1578 to = NULL; 1579 tj = TypeOopPtr::BOTTOM; 1580 offset = tj->offset(); 1581 } 1582 } else { 1583 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1584 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1585 if( is_known_inst ) { 1586 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1587 } else { 1588 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1589 } 1590 } 1591 } 1592 } 1593 1594 // Klass pointers to object array klasses need some flattening 1595 const TypeKlassPtr *tk = tj->isa_klassptr(); 1596 if( tk ) { 1597 // If we are referencing a field within a Klass, we need 1598 // to assume the worst case of an Object. Both exact and 1599 // inexact types must flatten to the same alias class so 1600 // use NotNull as the PTR. 1601 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1602 1603 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1604 TypeKlassPtr::OBJECT->klass(), 1605 Type::Offset(offset)); 1606 } 1607 1608 ciKlass* klass = tk->klass(); 1609 if (klass != NULL && klass->is_obj_array_klass()) { 1610 ciKlass* k = TypeAryPtr::OOPS->klass(); 1611 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1612 k = TypeInstPtr::BOTTOM->klass(); 1613 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1614 } 1615 1616 // Check for precise loads from the primary supertype array and force them 1617 // to the supertype cache alias index. Check for generic array loads from 1618 // the primary supertype array and also force them to the supertype cache 1619 // alias index. Since the same load can reach both, we need to merge 1620 // these 2 disparate memories into the same alias class. Since the 1621 // primary supertype array is read-only, there's no chance of confusion 1622 // where we bypass an array load and an array store. 1623 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1624 if (offset == Type::OffsetBot || 1625 (offset >= primary_supers_offset && 1626 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1627 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1628 offset = in_bytes(Klass::secondary_super_cache_offset()); 1629 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1630 } 1631 } 1632 1633 // Flatten all Raw pointers together. 1634 if (tj->base() == Type::RawPtr) 1635 tj = TypeRawPtr::BOTTOM; 1636 1637 if (tj->base() == Type::AnyPtr) 1638 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1639 1640 // Flatten all to bottom for now 1641 switch( _AliasLevel ) { 1642 case 0: 1643 tj = TypePtr::BOTTOM; 1644 break; 1645 case 1: // Flatten to: oop, static, field or array 1646 switch (tj->base()) { 1647 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1648 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1649 case Type::AryPtr: // do not distinguish arrays at all 1650 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1651 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1652 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1653 default: ShouldNotReachHere(); 1654 } 1655 break; 1656 case 2: // No collapsing at level 2; keep all splits 1657 case 3: // No collapsing at level 3; keep all splits 1658 break; 1659 default: 1660 Unimplemented(); 1661 } 1662 1663 offset = tj->offset(); 1664 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1665 1666 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1667 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1668 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1669 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1670 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1671 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1672 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1673 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1674 assert( tj->ptr() != TypePtr::TopPTR && 1675 tj->ptr() != TypePtr::AnyNull && 1676 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1677 // assert( tj->ptr() != TypePtr::Constant || 1678 // tj->base() == Type::RawPtr || 1679 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1680 1681 return tj; 1682 } 1683 1684 void Compile::AliasType::Init(int i, const TypePtr* at) { 1685 _index = i; 1686 _adr_type = at; 1687 _field = NULL; 1688 _element = NULL; 1689 _is_rewritable = true; // default 1690 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1691 if (atoop != NULL && atoop->is_known_instance()) { 1692 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1693 _general_index = Compile::current()->get_alias_index(gt); 1694 } else { 1695 _general_index = 0; 1696 } 1697 } 1698 1699 BasicType Compile::AliasType::basic_type() const { 1700 if (element() != NULL) { 1701 const Type* element = adr_type()->is_aryptr()->elem(); 1702 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1703 } if (field() != NULL) { 1704 return field()->layout_type(); 1705 } else { 1706 return T_ILLEGAL; // unknown 1707 } 1708 } 1709 1710 //---------------------------------print_on------------------------------------ 1711 #ifndef PRODUCT 1712 void Compile::AliasType::print_on(outputStream* st) { 1713 if (index() < 10) 1714 st->print("@ <%d> ", index()); 1715 else st->print("@ <%d>", index()); 1716 st->print(is_rewritable() ? " " : " RO"); 1717 int offset = adr_type()->offset(); 1718 if (offset == Type::OffsetBot) 1719 st->print(" +any"); 1720 else st->print(" +%-3d", offset); 1721 st->print(" in "); 1722 adr_type()->dump_on(st); 1723 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1724 if (field() != NULL && tjp) { 1725 if (tjp->klass() != field()->holder() || 1726 tjp->offset() != field()->offset_in_bytes()) { 1727 st->print(" != "); 1728 field()->print(); 1729 st->print(" ***"); 1730 } 1731 } 1732 } 1733 1734 void print_alias_types() { 1735 Compile* C = Compile::current(); 1736 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1737 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1738 C->alias_type(idx)->print_on(tty); 1739 tty->cr(); 1740 } 1741 } 1742 #endif 1743 1744 1745 //----------------------------probe_alias_cache-------------------------------- 1746 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1747 intptr_t key = (intptr_t) adr_type; 1748 key ^= key >> logAliasCacheSize; 1749 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1750 } 1751 1752 1753 //-----------------------------grow_alias_types-------------------------------- 1754 void Compile::grow_alias_types() { 1755 const int old_ats = _max_alias_types; // how many before? 1756 const int new_ats = old_ats; // how many more? 1757 const int grow_ats = old_ats+new_ats; // how many now? 1758 _max_alias_types = grow_ats; 1759 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1760 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1761 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1762 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1763 } 1764 1765 1766 //--------------------------------find_alias_type------------------------------ 1767 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1768 if (_AliasLevel == 0) 1769 return alias_type(AliasIdxBot); 1770 1771 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1772 if (ace->_adr_type == adr_type) { 1773 return alias_type(ace->_index); 1774 } 1775 1776 // Handle special cases. 1777 if (adr_type == NULL) return alias_type(AliasIdxTop); 1778 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1779 1780 // Do it the slow way. 1781 const TypePtr* flat = flatten_alias_type(adr_type); 1782 1783 #ifdef ASSERT 1784 { 1785 ResourceMark rm; 1786 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1787 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1788 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1789 Type::str(adr_type)); 1790 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1791 const TypeOopPtr* foop = flat->is_oopptr(); 1792 // Scalarizable allocations have exact klass always. 1793 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1794 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1795 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1796 Type::str(foop), Type::str(xoop)); 1797 } 1798 } 1799 #endif 1800 1801 int idx = AliasIdxTop; 1802 for (int i = 0; i < num_alias_types(); i++) { 1803 if (alias_type(i)->adr_type() == flat) { 1804 idx = i; 1805 break; 1806 } 1807 } 1808 1809 if (idx == AliasIdxTop) { 1810 if (no_create) return NULL; 1811 // Grow the array if necessary. 1812 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1813 // Add a new alias type. 1814 idx = _num_alias_types++; 1815 _alias_types[idx]->Init(idx, flat); 1816 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1817 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1818 if (flat->isa_instptr()) { 1819 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1820 && flat->is_instptr()->klass() == env()->Class_klass()) 1821 alias_type(idx)->set_rewritable(false); 1822 } 1823 ciField* field = NULL; 1824 if (flat->isa_aryptr()) { 1825 #ifdef ASSERT 1826 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1827 // (T_BYTE has the weakest alignment and size restrictions...) 1828 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1829 #endif 1830 const Type* elemtype = flat->is_aryptr()->elem(); 1831 if (flat->offset() == TypePtr::OffsetBot) { 1832 alias_type(idx)->set_element(elemtype); 1833 } 1834 int field_offset = flat->is_aryptr()->field_offset().get(); 1835 if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) { 1836 ciValueKlass* vk = elemtype->is_valuetype()->value_klass(); 1837 field_offset += vk->first_field_offset(); 1838 field = vk->get_field_by_offset(field_offset, false); 1839 } 1840 } 1841 if (flat->isa_klassptr()) { 1842 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1843 alias_type(idx)->set_rewritable(false); 1844 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1845 alias_type(idx)->set_rewritable(false); 1846 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1847 alias_type(idx)->set_rewritable(false); 1848 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1849 alias_type(idx)->set_rewritable(false); 1850 } 1851 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1852 // but the base pointer type is not distinctive enough to identify 1853 // references into JavaThread.) 1854 1855 // Check for final fields. 1856 const TypeInstPtr* tinst = flat->isa_instptr(); 1857 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1858 if (tinst->const_oop() != NULL && 1859 tinst->klass() == ciEnv::current()->Class_klass() && 1860 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1861 // static field 1862 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1863 field = k->get_field_by_offset(tinst->offset(), true); 1864 } else if (tinst->klass()->is_valuetype()) { 1865 // Value type field 1866 ciValueKlass* vk = tinst->value_klass(); 1867 field = vk->get_field_by_offset(tinst->offset(), false); 1868 } else { 1869 ciInstanceKlass* k = tinst->klass()->as_instance_klass(); 1870 field = k->get_field_by_offset(tinst->offset(), false); 1871 } 1872 } 1873 assert(field == NULL || 1874 original_field == NULL || 1875 (field->holder() == original_field->holder() && 1876 field->offset() == original_field->offset() && 1877 field->is_static() == original_field->is_static()), "wrong field?"); 1878 // Set field() and is_rewritable() attributes. 1879 if (field != NULL) alias_type(idx)->set_field(field); 1880 } 1881 1882 // Fill the cache for next time. 1883 ace->_adr_type = adr_type; 1884 ace->_index = idx; 1885 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1886 1887 // Might as well try to fill the cache for the flattened version, too. 1888 AliasCacheEntry* face = probe_alias_cache(flat); 1889 if (face->_adr_type == NULL) { 1890 face->_adr_type = flat; 1891 face->_index = idx; 1892 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1893 } 1894 1895 return alias_type(idx); 1896 } 1897 1898 1899 Compile::AliasType* Compile::alias_type(ciField* field) { 1900 const TypeOopPtr* t; 1901 if (field->is_static()) 1902 t = TypeInstPtr::make(field->holder()->java_mirror()); 1903 else 1904 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1905 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1906 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1907 return atp; 1908 } 1909 1910 1911 //------------------------------have_alias_type-------------------------------- 1912 bool Compile::have_alias_type(const TypePtr* adr_type) { 1913 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1914 if (ace->_adr_type == adr_type) { 1915 return true; 1916 } 1917 1918 // Handle special cases. 1919 if (adr_type == NULL) return true; 1920 if (adr_type == TypePtr::BOTTOM) return true; 1921 1922 return find_alias_type(adr_type, true, NULL) != NULL; 1923 } 1924 1925 //-----------------------------must_alias-------------------------------------- 1926 // True if all values of the given address type are in the given alias category. 1927 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1928 if (alias_idx == AliasIdxBot) return true; // the universal category 1929 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1930 if (alias_idx == AliasIdxTop) return false; // the empty category 1931 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1932 1933 // the only remaining possible overlap is identity 1934 int adr_idx = get_alias_index(adr_type); 1935 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1936 assert(adr_idx == alias_idx || 1937 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1938 && adr_type != TypeOopPtr::BOTTOM), 1939 "should not be testing for overlap with an unsafe pointer"); 1940 return adr_idx == alias_idx; 1941 } 1942 1943 //------------------------------can_alias-------------------------------------- 1944 // True if any values of the given address type are in the given alias category. 1945 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1946 if (alias_idx == AliasIdxTop) return false; // the empty category 1947 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1948 if (alias_idx == AliasIdxBot) return true; // the universal category 1949 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1950 1951 // the only remaining possible overlap is identity 1952 int adr_idx = get_alias_index(adr_type); 1953 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1954 return adr_idx == alias_idx; 1955 } 1956 1957 1958 1959 //---------------------------pop_warm_call------------------------------------- 1960 WarmCallInfo* Compile::pop_warm_call() { 1961 WarmCallInfo* wci = _warm_calls; 1962 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1963 return wci; 1964 } 1965 1966 //----------------------------Inline_Warm-------------------------------------- 1967 int Compile::Inline_Warm() { 1968 // If there is room, try to inline some more warm call sites. 1969 // %%% Do a graph index compaction pass when we think we're out of space? 1970 if (!InlineWarmCalls) return 0; 1971 1972 int calls_made_hot = 0; 1973 int room_to_grow = NodeCountInliningCutoff - unique(); 1974 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1975 int amount_grown = 0; 1976 WarmCallInfo* call; 1977 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1978 int est_size = (int)call->size(); 1979 if (est_size > (room_to_grow - amount_grown)) { 1980 // This one won't fit anyway. Get rid of it. 1981 call->make_cold(); 1982 continue; 1983 } 1984 call->make_hot(); 1985 calls_made_hot++; 1986 amount_grown += est_size; 1987 amount_to_grow -= est_size; 1988 } 1989 1990 if (calls_made_hot > 0) set_major_progress(); 1991 return calls_made_hot; 1992 } 1993 1994 1995 //----------------------------Finish_Warm-------------------------------------- 1996 void Compile::Finish_Warm() { 1997 if (!InlineWarmCalls) return; 1998 if (failing()) return; 1999 if (warm_calls() == NULL) return; 2000 2001 // Clean up loose ends, if we are out of space for inlining. 2002 WarmCallInfo* call; 2003 while ((call = pop_warm_call()) != NULL) { 2004 call->make_cold(); 2005 } 2006 } 2007 2008 //---------------------cleanup_loop_predicates----------------------- 2009 // Remove the opaque nodes that protect the predicates so that all unused 2010 // checks and uncommon_traps will be eliminated from the ideal graph 2011 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 2012 if (predicate_count()==0) return; 2013 for (int i = predicate_count(); i > 0; i--) { 2014 Node * n = predicate_opaque1_node(i-1); 2015 assert(n->Opcode() == Op_Opaque1, "must be"); 2016 igvn.replace_node(n, n->in(1)); 2017 } 2018 assert(predicate_count()==0, "should be clean!"); 2019 } 2020 2021 void Compile::add_range_check_cast(Node* n) { 2022 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2023 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 2024 _range_check_casts->append(n); 2025 } 2026 2027 // Remove all range check dependent CastIINodes. 2028 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 2029 for (int i = range_check_cast_count(); i > 0; i--) { 2030 Node* cast = range_check_cast_node(i-1); 2031 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 2032 igvn.replace_node(cast, cast->in(1)); 2033 } 2034 assert(range_check_cast_count() == 0, "should be empty"); 2035 } 2036 2037 void Compile::add_opaque4_node(Node* n) { 2038 assert(n->Opcode() == Op_Opaque4, "Opaque4 only"); 2039 assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list"); 2040 _opaque4_nodes->append(n); 2041 } 2042 2043 // Remove all Opaque4 nodes. 2044 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) { 2045 for (int i = opaque4_count(); i > 0; i--) { 2046 Node* opaq = opaque4_node(i-1); 2047 assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only"); 2048 igvn.replace_node(opaq, opaq->in(2)); 2049 } 2050 assert(opaque4_count() == 0, "should be empty"); 2051 } 2052 2053 void Compile::add_value_type(Node* n) { 2054 assert(n->is_ValueTypeBase(), "unexpected node"); 2055 if (_value_type_nodes != NULL) { 2056 _value_type_nodes->push(n); 2057 } 2058 } 2059 2060 void Compile::remove_value_type(Node* n) { 2061 assert(n->is_ValueTypeBase(), "unexpected node"); 2062 if (_value_type_nodes != NULL) { 2063 _value_type_nodes->remove(n); 2064 } 2065 } 2066 2067 void Compile::process_value_types(PhaseIterGVN &igvn) { 2068 // Make value types scalar in safepoints 2069 while (_value_type_nodes->size() != 0) { 2070 ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase(); 2071 vt->make_scalar_in_safepoints(igvn.C->root(), &igvn); 2072 if (vt->is_ValueTypePtr()) { 2073 igvn.replace_node(vt, vt->get_oop()); 2074 } 2075 } 2076 _value_type_nodes = NULL; 2077 igvn.optimize(); 2078 } 2079 2080 // StringOpts and late inlining of string methods 2081 void Compile::inline_string_calls(bool parse_time) { 2082 { 2083 // remove useless nodes to make the usage analysis simpler 2084 ResourceMark rm; 2085 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2086 } 2087 2088 { 2089 ResourceMark rm; 2090 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2091 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2092 print_method(PHASE_AFTER_STRINGOPTS, 3); 2093 } 2094 2095 // now inline anything that we skipped the first time around 2096 if (!parse_time) { 2097 _late_inlines_pos = _late_inlines.length(); 2098 } 2099 2100 while (_string_late_inlines.length() > 0) { 2101 CallGenerator* cg = _string_late_inlines.pop(); 2102 cg->do_late_inline(); 2103 if (failing()) return; 2104 } 2105 _string_late_inlines.trunc_to(0); 2106 } 2107 2108 // Late inlining of boxing methods 2109 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2110 if (_boxing_late_inlines.length() > 0) { 2111 assert(has_boxed_value(), "inconsistent"); 2112 2113 PhaseGVN* gvn = initial_gvn(); 2114 set_inlining_incrementally(true); 2115 2116 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2117 for_igvn()->clear(); 2118 gvn->replace_with(&igvn); 2119 2120 _late_inlines_pos = _late_inlines.length(); 2121 2122 while (_boxing_late_inlines.length() > 0) { 2123 CallGenerator* cg = _boxing_late_inlines.pop(); 2124 cg->do_late_inline(); 2125 if (failing()) return; 2126 } 2127 _boxing_late_inlines.trunc_to(0); 2128 2129 { 2130 ResourceMark rm; 2131 PhaseRemoveUseless pru(gvn, for_igvn()); 2132 } 2133 2134 igvn = PhaseIterGVN(gvn); 2135 igvn.optimize(); 2136 2137 set_inlining_progress(false); 2138 set_inlining_incrementally(false); 2139 } 2140 } 2141 2142 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2143 assert(IncrementalInline, "incremental inlining should be on"); 2144 PhaseGVN* gvn = initial_gvn(); 2145 2146 set_inlining_progress(false); 2147 for_igvn()->clear(); 2148 gvn->replace_with(&igvn); 2149 2150 { 2151 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2152 int i = 0; 2153 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2154 CallGenerator* cg = _late_inlines.at(i); 2155 _late_inlines_pos = i+1; 2156 cg->do_late_inline(); 2157 if (failing()) return; 2158 } 2159 int j = 0; 2160 for (; i < _late_inlines.length(); i++, j++) { 2161 _late_inlines.at_put(j, _late_inlines.at(i)); 2162 } 2163 _late_inlines.trunc_to(j); 2164 } 2165 2166 { 2167 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2168 ResourceMark rm; 2169 PhaseRemoveUseless pru(gvn, for_igvn()); 2170 } 2171 2172 { 2173 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2174 igvn = PhaseIterGVN(gvn); 2175 } 2176 } 2177 2178 // Perform incremental inlining until bound on number of live nodes is reached 2179 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2180 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2181 2182 PhaseGVN* gvn = initial_gvn(); 2183 2184 set_inlining_incrementally(true); 2185 set_inlining_progress(true); 2186 uint low_live_nodes = 0; 2187 2188 while(inlining_progress() && _late_inlines.length() > 0) { 2189 2190 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2191 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2192 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2193 // PhaseIdealLoop is expensive so we only try it once we are 2194 // out of live nodes and we only try it again if the previous 2195 // helped got the number of nodes down significantly 2196 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2197 if (failing()) return; 2198 low_live_nodes = live_nodes(); 2199 _major_progress = true; 2200 } 2201 2202 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2203 break; 2204 } 2205 } 2206 2207 inline_incrementally_one(igvn); 2208 2209 if (failing()) return; 2210 2211 { 2212 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2213 igvn.optimize(); 2214 } 2215 2216 if (failing()) return; 2217 } 2218 2219 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2220 2221 if (_string_late_inlines.length() > 0) { 2222 assert(has_stringbuilder(), "inconsistent"); 2223 for_igvn()->clear(); 2224 initial_gvn()->replace_with(&igvn); 2225 2226 inline_string_calls(false); 2227 2228 if (failing()) return; 2229 2230 { 2231 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2232 ResourceMark rm; 2233 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2234 } 2235 2236 { 2237 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2238 igvn = PhaseIterGVN(gvn); 2239 igvn.optimize(); 2240 } 2241 } 2242 2243 set_inlining_incrementally(false); 2244 } 2245 2246 2247 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) { 2248 if(loop_opts_cnt > 0) { 2249 debug_only( int cnt = 0; ); 2250 while(major_progress() && (loop_opts_cnt > 0)) { 2251 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2252 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2253 PhaseIdealLoop ideal_loop(igvn, mode); 2254 loop_opts_cnt--; 2255 if (failing()) return false; 2256 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2257 } 2258 } 2259 return true; 2260 } 2261 2262 //------------------------------Optimize--------------------------------------- 2263 // Given a graph, optimize it. 2264 void Compile::Optimize() { 2265 TracePhase tp("optimizer", &timers[_t_optimizer]); 2266 2267 #ifndef PRODUCT 2268 if (_directive->BreakAtCompileOption) { 2269 BREAKPOINT; 2270 } 2271 2272 #endif 2273 2274 #ifdef ASSERT 2275 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2276 bs->verify_gc_barriers(true); 2277 #endif 2278 2279 ResourceMark rm; 2280 int loop_opts_cnt; 2281 2282 print_inlining_reinit(); 2283 2284 NOT_PRODUCT( verify_graph_edges(); ) 2285 2286 print_method(PHASE_AFTER_PARSING); 2287 2288 { 2289 // Iterative Global Value Numbering, including ideal transforms 2290 // Initialize IterGVN with types and values from parse-time GVN 2291 PhaseIterGVN igvn(initial_gvn()); 2292 #ifdef ASSERT 2293 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2294 #endif 2295 { 2296 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2297 igvn.optimize(); 2298 } 2299 2300 if (failing()) return; 2301 2302 print_method(PHASE_ITER_GVN1, 2); 2303 2304 inline_incrementally(igvn); 2305 2306 print_method(PHASE_INCREMENTAL_INLINE, 2); 2307 2308 if (failing()) return; 2309 2310 if (eliminate_boxing()) { 2311 // Inline valueOf() methods now. 2312 inline_boxing_calls(igvn); 2313 2314 if (AlwaysIncrementalInline) { 2315 inline_incrementally(igvn); 2316 } 2317 2318 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2319 2320 if (failing()) return; 2321 } 2322 2323 // Remove the speculative part of types and clean up the graph from 2324 // the extra CastPP nodes whose only purpose is to carry them. Do 2325 // that early so that optimizations are not disrupted by the extra 2326 // CastPP nodes. 2327 remove_speculative_types(igvn); 2328 2329 // No more new expensive nodes will be added to the list from here 2330 // so keep only the actual candidates for optimizations. 2331 cleanup_expensive_nodes(igvn); 2332 2333 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2334 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2335 initial_gvn()->replace_with(&igvn); 2336 for_igvn()->clear(); 2337 Unique_Node_List new_worklist(C->comp_arena()); 2338 { 2339 ResourceMark rm; 2340 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2341 } 2342 set_for_igvn(&new_worklist); 2343 igvn = PhaseIterGVN(initial_gvn()); 2344 igvn.optimize(); 2345 } 2346 2347 if (_value_type_nodes->size() > 0) { 2348 // Do this once all inlining is over to avoid getting inconsistent debug info 2349 process_value_types(igvn); 2350 } 2351 2352 // Perform escape analysis 2353 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2354 if (has_loops()) { 2355 // Cleanup graph (remove dead nodes). 2356 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2357 PhaseIdealLoop ideal_loop(igvn, LoopOptsNone); 2358 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2359 if (failing()) return; 2360 } 2361 ConnectionGraph::do_analysis(this, &igvn); 2362 2363 if (failing()) return; 2364 2365 // Optimize out fields loads from scalar replaceable allocations. 2366 igvn.optimize(); 2367 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2368 2369 if (failing()) return; 2370 2371 if (congraph() != NULL && macro_count() > 0) { 2372 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2373 PhaseMacroExpand mexp(igvn); 2374 mexp.eliminate_macro_nodes(); 2375 igvn.set_delay_transform(false); 2376 2377 igvn.optimize(); 2378 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2379 2380 if (failing()) return; 2381 } 2382 } 2383 2384 // Loop transforms on the ideal graph. Range Check Elimination, 2385 // peeling, unrolling, etc. 2386 2387 // Set loop opts counter 2388 loop_opts_cnt = num_loop_opts(); 2389 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2390 { 2391 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2392 PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault); 2393 loop_opts_cnt--; 2394 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2395 if (failing()) return; 2396 } 2397 // Loop opts pass if partial peeling occurred in previous pass 2398 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2399 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2400 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2401 loop_opts_cnt--; 2402 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2403 if (failing()) return; 2404 } 2405 // Loop opts pass for loop-unrolling before CCP 2406 if(major_progress() && (loop_opts_cnt > 0)) { 2407 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2408 PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf); 2409 loop_opts_cnt--; 2410 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2411 } 2412 if (!failing()) { 2413 // Verify that last round of loop opts produced a valid graph 2414 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2415 PhaseIdealLoop::verify(igvn); 2416 } 2417 } 2418 if (failing()) return; 2419 2420 // Conditional Constant Propagation; 2421 PhaseCCP ccp( &igvn ); 2422 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2423 { 2424 TracePhase tp("ccp", &timers[_t_ccp]); 2425 ccp.do_transform(); 2426 } 2427 print_method(PHASE_CPP1, 2); 2428 2429 assert( true, "Break here to ccp.dump_old2new_map()"); 2430 2431 // Iterative Global Value Numbering, including ideal transforms 2432 { 2433 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2434 igvn = ccp; 2435 igvn.optimize(); 2436 } 2437 2438 print_method(PHASE_ITER_GVN2, 2); 2439 2440 if (failing()) return; 2441 2442 // Loop transforms on the ideal graph. Range Check Elimination, 2443 // peeling, unrolling, etc. 2444 if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) { 2445 return; 2446 } 2447 2448 #if INCLUDE_ZGC 2449 if (UseZGC) { 2450 ZBarrierSetC2::find_dominating_barriers(igvn); 2451 } 2452 #endif 2453 2454 if (failing()) return; 2455 2456 // Ensure that major progress is now clear 2457 C->clear_major_progress(); 2458 2459 { 2460 // Verify that all previous optimizations produced a valid graph 2461 // at least to this point, even if no loop optimizations were done. 2462 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2463 PhaseIdealLoop::verify(igvn); 2464 } 2465 2466 if (range_check_cast_count() > 0) { 2467 // No more loop optimizations. Remove all range check dependent CastIINodes. 2468 C->remove_range_check_casts(igvn); 2469 igvn.optimize(); 2470 } 2471 2472 #ifdef ASSERT 2473 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2474 bs->verify_gc_barriers(false); 2475 #endif 2476 2477 { 2478 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2479 PhaseMacroExpand mex(igvn); 2480 print_method(PHASE_BEFORE_MACRO_EXPANSION, 2); 2481 if (mex.expand_macro_nodes()) { 2482 assert(failing(), "must bail out w/ explicit message"); 2483 return; 2484 } 2485 } 2486 2487 if (opaque4_count() > 0) { 2488 C->remove_opaque4_nodes(igvn); 2489 igvn.optimize(); 2490 } 2491 2492 DEBUG_ONLY( _modified_nodes = NULL; ) 2493 } // (End scope of igvn; run destructor if necessary for asserts.) 2494 2495 process_print_inlining(); 2496 // A method with only infinite loops has no edges entering loops from root 2497 { 2498 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2499 if (final_graph_reshaping()) { 2500 assert(failing(), "must bail out w/ explicit message"); 2501 return; 2502 } 2503 } 2504 2505 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2506 } 2507 2508 //------------------------------Code_Gen--------------------------------------- 2509 // Given a graph, generate code for it 2510 void Compile::Code_Gen() { 2511 if (failing()) { 2512 return; 2513 } 2514 2515 // Perform instruction selection. You might think we could reclaim Matcher 2516 // memory PDQ, but actually the Matcher is used in generating spill code. 2517 // Internals of the Matcher (including some VectorSets) must remain live 2518 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2519 // set a bit in reclaimed memory. 2520 2521 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2522 // nodes. Mapping is only valid at the root of each matched subtree. 2523 NOT_PRODUCT( verify_graph_edges(); ) 2524 2525 Matcher matcher; 2526 _matcher = &matcher; 2527 { 2528 TracePhase tp("matcher", &timers[_t_matcher]); 2529 matcher.match(); 2530 } 2531 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2532 // nodes. Mapping is only valid at the root of each matched subtree. 2533 NOT_PRODUCT( verify_graph_edges(); ) 2534 2535 // If you have too many nodes, or if matching has failed, bail out 2536 check_node_count(0, "out of nodes matching instructions"); 2537 if (failing()) { 2538 return; 2539 } 2540 2541 print_method(PHASE_MATCHING, 2); 2542 2543 // Build a proper-looking CFG 2544 PhaseCFG cfg(node_arena(), root(), matcher); 2545 _cfg = &cfg; 2546 { 2547 TracePhase tp("scheduler", &timers[_t_scheduler]); 2548 bool success = cfg.do_global_code_motion(); 2549 if (!success) { 2550 return; 2551 } 2552 2553 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2554 NOT_PRODUCT( verify_graph_edges(); ) 2555 debug_only( cfg.verify(); ) 2556 } 2557 2558 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2559 _regalloc = ®alloc; 2560 { 2561 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2562 // Perform register allocation. After Chaitin, use-def chains are 2563 // no longer accurate (at spill code) and so must be ignored. 2564 // Node->LRG->reg mappings are still accurate. 2565 _regalloc->Register_Allocate(); 2566 2567 // Bail out if the allocator builds too many nodes 2568 if (failing()) { 2569 return; 2570 } 2571 } 2572 2573 // Prior to register allocation we kept empty basic blocks in case the 2574 // the allocator needed a place to spill. After register allocation we 2575 // are not adding any new instructions. If any basic block is empty, we 2576 // can now safely remove it. 2577 { 2578 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2579 cfg.remove_empty_blocks(); 2580 if (do_freq_based_layout()) { 2581 PhaseBlockLayout layout(cfg); 2582 } else { 2583 cfg.set_loop_alignment(); 2584 } 2585 cfg.fixup_flow(); 2586 } 2587 2588 // Apply peephole optimizations 2589 if( OptoPeephole ) { 2590 TracePhase tp("peephole", &timers[_t_peephole]); 2591 PhasePeephole peep( _regalloc, cfg); 2592 peep.do_transform(); 2593 } 2594 2595 // Do late expand if CPU requires this. 2596 if (Matcher::require_postalloc_expand) { 2597 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2598 cfg.postalloc_expand(_regalloc); 2599 } 2600 2601 // Convert Nodes to instruction bits in a buffer 2602 { 2603 TraceTime tp("output", &timers[_t_output], CITime); 2604 Output(); 2605 } 2606 2607 print_method(PHASE_FINAL_CODE); 2608 2609 // He's dead, Jim. 2610 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2611 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2612 } 2613 2614 2615 //------------------------------dump_asm--------------------------------------- 2616 // Dump formatted assembly 2617 #ifndef PRODUCT 2618 void Compile::dump_asm(int *pcs, uint pc_limit) { 2619 bool cut_short = false; 2620 tty->print_cr("#"); 2621 tty->print("# "); _tf->dump(); tty->cr(); 2622 tty->print_cr("#"); 2623 2624 // For all blocks 2625 int pc = 0x0; // Program counter 2626 char starts_bundle = ' '; 2627 _regalloc->dump_frame(); 2628 2629 Node *n = NULL; 2630 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2631 if (VMThread::should_terminate()) { 2632 cut_short = true; 2633 break; 2634 } 2635 Block* block = _cfg->get_block(i); 2636 if (block->is_connector() && !Verbose) { 2637 continue; 2638 } 2639 n = block->head(); 2640 if (pcs && n->_idx < pc_limit) { 2641 tty->print("%3.3x ", pcs[n->_idx]); 2642 } else { 2643 tty->print(" "); 2644 } 2645 block->dump_head(_cfg); 2646 if (block->is_connector()) { 2647 tty->print_cr(" # Empty connector block"); 2648 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2649 tty->print_cr(" # Block is sole successor of call"); 2650 } 2651 2652 // For all instructions 2653 Node *delay = NULL; 2654 for (uint j = 0; j < block->number_of_nodes(); j++) { 2655 if (VMThread::should_terminate()) { 2656 cut_short = true; 2657 break; 2658 } 2659 n = block->get_node(j); 2660 if (valid_bundle_info(n)) { 2661 Bundle* bundle = node_bundling(n); 2662 if (bundle->used_in_unconditional_delay()) { 2663 delay = n; 2664 continue; 2665 } 2666 if (bundle->starts_bundle()) { 2667 starts_bundle = '+'; 2668 } 2669 } 2670 2671 if (WizardMode) { 2672 n->dump(); 2673 } 2674 2675 if( !n->is_Region() && // Dont print in the Assembly 2676 !n->is_Phi() && // a few noisely useless nodes 2677 !n->is_Proj() && 2678 !n->is_MachTemp() && 2679 !n->is_SafePointScalarObject() && 2680 !n->is_Catch() && // Would be nice to print exception table targets 2681 !n->is_MergeMem() && // Not very interesting 2682 !n->is_top() && // Debug info table constants 2683 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2684 ) { 2685 if (pcs && n->_idx < pc_limit) 2686 tty->print("%3.3x", pcs[n->_idx]); 2687 else 2688 tty->print(" "); 2689 tty->print(" %c ", starts_bundle); 2690 starts_bundle = ' '; 2691 tty->print("\t"); 2692 n->format(_regalloc, tty); 2693 tty->cr(); 2694 } 2695 2696 // If we have an instruction with a delay slot, and have seen a delay, 2697 // then back up and print it 2698 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2699 assert(delay != NULL, "no unconditional delay instruction"); 2700 if (WizardMode) delay->dump(); 2701 2702 if (node_bundling(delay)->starts_bundle()) 2703 starts_bundle = '+'; 2704 if (pcs && n->_idx < pc_limit) 2705 tty->print("%3.3x", pcs[n->_idx]); 2706 else 2707 tty->print(" "); 2708 tty->print(" %c ", starts_bundle); 2709 starts_bundle = ' '; 2710 tty->print("\t"); 2711 delay->format(_regalloc, tty); 2712 tty->cr(); 2713 delay = NULL; 2714 } 2715 2716 // Dump the exception table as well 2717 if( n->is_Catch() && (Verbose || WizardMode) ) { 2718 // Print the exception table for this offset 2719 _handler_table.print_subtable_for(pc); 2720 } 2721 } 2722 2723 if (pcs && n->_idx < pc_limit) 2724 tty->print_cr("%3.3x", pcs[n->_idx]); 2725 else 2726 tty->cr(); 2727 2728 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2729 2730 } // End of per-block dump 2731 tty->cr(); 2732 2733 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2734 } 2735 #endif 2736 2737 //------------------------------Final_Reshape_Counts--------------------------- 2738 // This class defines counters to help identify when a method 2739 // may/must be executed using hardware with only 24-bit precision. 2740 struct Final_Reshape_Counts : public StackObj { 2741 int _call_count; // count non-inlined 'common' calls 2742 int _float_count; // count float ops requiring 24-bit precision 2743 int _double_count; // count double ops requiring more precision 2744 int _java_call_count; // count non-inlined 'java' calls 2745 int _inner_loop_count; // count loops which need alignment 2746 VectorSet _visited; // Visitation flags 2747 Node_List _tests; // Set of IfNodes & PCTableNodes 2748 2749 Final_Reshape_Counts() : 2750 _call_count(0), _float_count(0), _double_count(0), 2751 _java_call_count(0), _inner_loop_count(0), 2752 _visited( Thread::current()->resource_area() ) { } 2753 2754 void inc_call_count () { _call_count ++; } 2755 void inc_float_count () { _float_count ++; } 2756 void inc_double_count() { _double_count++; } 2757 void inc_java_call_count() { _java_call_count++; } 2758 void inc_inner_loop_count() { _inner_loop_count++; } 2759 2760 int get_call_count () const { return _call_count ; } 2761 int get_float_count () const { return _float_count ; } 2762 int get_double_count() const { return _double_count; } 2763 int get_java_call_count() const { return _java_call_count; } 2764 int get_inner_loop_count() const { return _inner_loop_count; } 2765 }; 2766 2767 #ifdef ASSERT 2768 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2769 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2770 // Make sure the offset goes inside the instance layout. 2771 return k->contains_field_offset(tp->offset()); 2772 // Note that OffsetBot and OffsetTop are very negative. 2773 } 2774 #endif 2775 2776 // Eliminate trivially redundant StoreCMs and accumulate their 2777 // precedence edges. 2778 void Compile::eliminate_redundant_card_marks(Node* n) { 2779 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2780 if (n->in(MemNode::Address)->outcnt() > 1) { 2781 // There are multiple users of the same address so it might be 2782 // possible to eliminate some of the StoreCMs 2783 Node* mem = n->in(MemNode::Memory); 2784 Node* adr = n->in(MemNode::Address); 2785 Node* val = n->in(MemNode::ValueIn); 2786 Node* prev = n; 2787 bool done = false; 2788 // Walk the chain of StoreCMs eliminating ones that match. As 2789 // long as it's a chain of single users then the optimization is 2790 // safe. Eliminating partially redundant StoreCMs would require 2791 // cloning copies down the other paths. 2792 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2793 if (adr == mem->in(MemNode::Address) && 2794 val == mem->in(MemNode::ValueIn)) { 2795 // redundant StoreCM 2796 if (mem->req() > MemNode::OopStore) { 2797 // Hasn't been processed by this code yet. 2798 n->add_prec(mem->in(MemNode::OopStore)); 2799 } else { 2800 // Already converted to precedence edge 2801 for (uint i = mem->req(); i < mem->len(); i++) { 2802 // Accumulate any precedence edges 2803 if (mem->in(i) != NULL) { 2804 n->add_prec(mem->in(i)); 2805 } 2806 } 2807 // Everything above this point has been processed. 2808 done = true; 2809 } 2810 // Eliminate the previous StoreCM 2811 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2812 assert(mem->outcnt() == 0, "should be dead"); 2813 mem->disconnect_inputs(NULL, this); 2814 } else { 2815 prev = mem; 2816 } 2817 mem = prev->in(MemNode::Memory); 2818 } 2819 } 2820 } 2821 2822 2823 //------------------------------final_graph_reshaping_impl---------------------- 2824 // Implement items 1-5 from final_graph_reshaping below. 2825 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2826 2827 if ( n->outcnt() == 0 ) return; // dead node 2828 uint nop = n->Opcode(); 2829 2830 // Check for 2-input instruction with "last use" on right input. 2831 // Swap to left input. Implements item (2). 2832 if( n->req() == 3 && // two-input instruction 2833 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2834 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2835 n->in(2)->outcnt() == 1 &&// right use IS a last use 2836 !n->in(2)->is_Con() ) { // right use is not a constant 2837 // Check for commutative opcode 2838 switch( nop ) { 2839 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2840 case Op_MaxI: case Op_MinI: 2841 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2842 case Op_AndL: case Op_XorL: case Op_OrL: 2843 case Op_AndI: case Op_XorI: case Op_OrI: { 2844 // Move "last use" input to left by swapping inputs 2845 n->swap_edges(1, 2); 2846 break; 2847 } 2848 default: 2849 break; 2850 } 2851 } 2852 2853 #ifdef ASSERT 2854 if( n->is_Mem() ) { 2855 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2856 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2857 // oop will be recorded in oop map if load crosses safepoint 2858 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2859 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2860 "raw memory operations should have control edge"); 2861 } 2862 if (n->is_MemBar()) { 2863 MemBarNode* mb = n->as_MemBar(); 2864 if (mb->trailing_store() || mb->trailing_load_store()) { 2865 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2866 Node* mem = mb->in(MemBarNode::Precedent); 2867 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2868 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2869 } else if (mb->leading()) { 2870 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2871 } 2872 } 2873 #endif 2874 // Count FPU ops and common calls, implements item (3) 2875 switch( nop ) { 2876 // Count all float operations that may use FPU 2877 case Op_AddF: 2878 case Op_SubF: 2879 case Op_MulF: 2880 case Op_DivF: 2881 case Op_NegF: 2882 case Op_ModF: 2883 case Op_ConvI2F: 2884 case Op_ConF: 2885 case Op_CmpF: 2886 case Op_CmpF3: 2887 // case Op_ConvL2F: // longs are split into 32-bit halves 2888 frc.inc_float_count(); 2889 break; 2890 2891 case Op_ConvF2D: 2892 case Op_ConvD2F: 2893 frc.inc_float_count(); 2894 frc.inc_double_count(); 2895 break; 2896 2897 // Count all double operations that may use FPU 2898 case Op_AddD: 2899 case Op_SubD: 2900 case Op_MulD: 2901 case Op_DivD: 2902 case Op_NegD: 2903 case Op_ModD: 2904 case Op_ConvI2D: 2905 case Op_ConvD2I: 2906 // case Op_ConvL2D: // handled by leaf call 2907 // case Op_ConvD2L: // handled by leaf call 2908 case Op_ConD: 2909 case Op_CmpD: 2910 case Op_CmpD3: 2911 frc.inc_double_count(); 2912 break; 2913 case Op_Opaque1: // Remove Opaque Nodes before matching 2914 case Op_Opaque2: // Remove Opaque Nodes before matching 2915 case Op_Opaque3: 2916 n->subsume_by(n->in(1), this); 2917 break; 2918 case Op_CallStaticJava: 2919 case Op_CallJava: 2920 case Op_CallDynamicJava: 2921 frc.inc_java_call_count(); // Count java call site; 2922 case Op_CallRuntime: 2923 case Op_CallLeaf: 2924 case Op_CallLeafNoFP: { 2925 assert (n->is_Call(), ""); 2926 CallNode *call = n->as_Call(); 2927 // Count call sites where the FP mode bit would have to be flipped. 2928 // Do not count uncommon runtime calls: 2929 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2930 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2931 if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) { 2932 frc.inc_call_count(); // Count the call site 2933 } else { // See if uncommon argument is shared 2934 Node *n = call->in(TypeFunc::Parms); 2935 int nop = n->Opcode(); 2936 // Clone shared simple arguments to uncommon calls, item (1). 2937 if (n->outcnt() > 1 && 2938 !n->is_Proj() && 2939 nop != Op_CreateEx && 2940 nop != Op_CheckCastPP && 2941 nop != Op_DecodeN && 2942 nop != Op_DecodeNKlass && 2943 !n->is_Mem() && 2944 !n->is_Phi()) { 2945 Node *x = n->clone(); 2946 call->set_req(TypeFunc::Parms, x); 2947 } 2948 } 2949 break; 2950 } 2951 2952 case Op_StoreD: 2953 case Op_LoadD: 2954 case Op_LoadD_unaligned: 2955 frc.inc_double_count(); 2956 goto handle_mem; 2957 case Op_StoreF: 2958 case Op_LoadF: 2959 frc.inc_float_count(); 2960 goto handle_mem; 2961 2962 case Op_StoreCM: 2963 { 2964 // Convert OopStore dependence into precedence edge 2965 Node* prec = n->in(MemNode::OopStore); 2966 n->del_req(MemNode::OopStore); 2967 n->add_prec(prec); 2968 eliminate_redundant_card_marks(n); 2969 } 2970 2971 // fall through 2972 2973 case Op_StoreB: 2974 case Op_StoreC: 2975 case Op_StorePConditional: 2976 case Op_StoreI: 2977 case Op_StoreL: 2978 case Op_StoreIConditional: 2979 case Op_StoreLConditional: 2980 case Op_CompareAndSwapB: 2981 case Op_CompareAndSwapS: 2982 case Op_CompareAndSwapI: 2983 case Op_CompareAndSwapL: 2984 case Op_CompareAndSwapP: 2985 case Op_CompareAndSwapN: 2986 case Op_WeakCompareAndSwapB: 2987 case Op_WeakCompareAndSwapS: 2988 case Op_WeakCompareAndSwapI: 2989 case Op_WeakCompareAndSwapL: 2990 case Op_WeakCompareAndSwapP: 2991 case Op_WeakCompareAndSwapN: 2992 case Op_CompareAndExchangeB: 2993 case Op_CompareAndExchangeS: 2994 case Op_CompareAndExchangeI: 2995 case Op_CompareAndExchangeL: 2996 case Op_CompareAndExchangeP: 2997 case Op_CompareAndExchangeN: 2998 case Op_GetAndAddS: 2999 case Op_GetAndAddB: 3000 case Op_GetAndAddI: 3001 case Op_GetAndAddL: 3002 case Op_GetAndSetS: 3003 case Op_GetAndSetB: 3004 case Op_GetAndSetI: 3005 case Op_GetAndSetL: 3006 case Op_GetAndSetP: 3007 case Op_GetAndSetN: 3008 case Op_StoreP: 3009 case Op_StoreN: 3010 case Op_StoreNKlass: 3011 case Op_LoadB: 3012 case Op_LoadUB: 3013 case Op_LoadUS: 3014 case Op_LoadI: 3015 case Op_LoadKlass: 3016 case Op_LoadNKlass: 3017 case Op_LoadL: 3018 case Op_LoadL_unaligned: 3019 case Op_LoadPLocked: 3020 case Op_LoadP: 3021 #if INCLUDE_ZGC 3022 case Op_LoadBarrierSlowReg: 3023 case Op_LoadBarrierWeakSlowReg: 3024 #endif 3025 case Op_LoadN: 3026 case Op_LoadRange: 3027 case Op_LoadS: { 3028 handle_mem: 3029 #ifdef ASSERT 3030 if( VerifyOptoOopOffsets ) { 3031 assert( n->is_Mem(), "" ); 3032 MemNode *mem = (MemNode*)n; 3033 // Check to see if address types have grounded out somehow. 3034 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 3035 assert( !tp || oop_offset_is_sane(tp), "" ); 3036 } 3037 #endif 3038 break; 3039 } 3040 3041 case Op_AddP: { // Assert sane base pointers 3042 Node *addp = n->in(AddPNode::Address); 3043 assert( !addp->is_AddP() || 3044 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3045 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3046 "Base pointers must match (addp %u)", addp->_idx ); 3047 #ifdef _LP64 3048 if ((UseCompressedOops || UseCompressedClassPointers) && 3049 addp->Opcode() == Op_ConP && 3050 addp == n->in(AddPNode::Base) && 3051 n->in(AddPNode::Offset)->is_Con()) { 3052 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3053 // on the platform and on the compressed oops mode. 3054 // Use addressing with narrow klass to load with offset on x86. 3055 // Some platforms can use the constant pool to load ConP. 3056 // Do this transformation here since IGVN will convert ConN back to ConP. 3057 const Type* t = addp->bottom_type(); 3058 bool is_oop = t->isa_oopptr() != NULL; 3059 bool is_klass = t->isa_klassptr() != NULL; 3060 3061 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3062 (is_klass && Matcher::const_klass_prefer_decode())) { 3063 Node* nn = NULL; 3064 3065 int op = is_oop ? Op_ConN : Op_ConNKlass; 3066 3067 // Look for existing ConN node of the same exact type. 3068 Node* r = root(); 3069 uint cnt = r->outcnt(); 3070 for (uint i = 0; i < cnt; i++) { 3071 Node* m = r->raw_out(i); 3072 if (m!= NULL && m->Opcode() == op && 3073 m->bottom_type()->make_ptr() == t) { 3074 nn = m; 3075 break; 3076 } 3077 } 3078 if (nn != NULL) { 3079 // Decode a narrow oop to match address 3080 // [R12 + narrow_oop_reg<<3 + offset] 3081 if (is_oop) { 3082 nn = new DecodeNNode(nn, t); 3083 } else { 3084 nn = new DecodeNKlassNode(nn, t); 3085 } 3086 // Check for succeeding AddP which uses the same Base. 3087 // Otherwise we will run into the assertion above when visiting that guy. 3088 for (uint i = 0; i < n->outcnt(); ++i) { 3089 Node *out_i = n->raw_out(i); 3090 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3091 out_i->set_req(AddPNode::Base, nn); 3092 #ifdef ASSERT 3093 for (uint j = 0; j < out_i->outcnt(); ++j) { 3094 Node *out_j = out_i->raw_out(j); 3095 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3096 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3097 } 3098 #endif 3099 } 3100 } 3101 n->set_req(AddPNode::Base, nn); 3102 n->set_req(AddPNode::Address, nn); 3103 if (addp->outcnt() == 0) { 3104 addp->disconnect_inputs(NULL, this); 3105 } 3106 } 3107 } 3108 } 3109 #endif 3110 // platform dependent reshaping of the address expression 3111 reshape_address(n->as_AddP()); 3112 break; 3113 } 3114 3115 case Op_CastPP: { 3116 // Remove CastPP nodes to gain more freedom during scheduling but 3117 // keep the dependency they encode as control or precedence edges 3118 // (if control is set already) on memory operations. Some CastPP 3119 // nodes don't have a control (don't carry a dependency): skip 3120 // those. 3121 if (n->in(0) != NULL) { 3122 ResourceMark rm; 3123 Unique_Node_List wq; 3124 wq.push(n); 3125 for (uint next = 0; next < wq.size(); ++next) { 3126 Node *m = wq.at(next); 3127 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3128 Node* use = m->fast_out(i); 3129 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3130 use->ensure_control_or_add_prec(n->in(0)); 3131 } else { 3132 switch(use->Opcode()) { 3133 case Op_AddP: 3134 case Op_DecodeN: 3135 case Op_DecodeNKlass: 3136 case Op_CheckCastPP: 3137 case Op_CastPP: 3138 wq.push(use); 3139 break; 3140 } 3141 } 3142 } 3143 } 3144 } 3145 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3146 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3147 Node* in1 = n->in(1); 3148 const Type* t = n->bottom_type(); 3149 Node* new_in1 = in1->clone(); 3150 new_in1->as_DecodeN()->set_type(t); 3151 3152 if (!Matcher::narrow_oop_use_complex_address()) { 3153 // 3154 // x86, ARM and friends can handle 2 adds in addressing mode 3155 // and Matcher can fold a DecodeN node into address by using 3156 // a narrow oop directly and do implicit NULL check in address: 3157 // 3158 // [R12 + narrow_oop_reg<<3 + offset] 3159 // NullCheck narrow_oop_reg 3160 // 3161 // On other platforms (Sparc) we have to keep new DecodeN node and 3162 // use it to do implicit NULL check in address: 3163 // 3164 // decode_not_null narrow_oop_reg, base_reg 3165 // [base_reg + offset] 3166 // NullCheck base_reg 3167 // 3168 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3169 // to keep the information to which NULL check the new DecodeN node 3170 // corresponds to use it as value in implicit_null_check(). 3171 // 3172 new_in1->set_req(0, n->in(0)); 3173 } 3174 3175 n->subsume_by(new_in1, this); 3176 if (in1->outcnt() == 0) { 3177 in1->disconnect_inputs(NULL, this); 3178 } 3179 } else { 3180 n->subsume_by(n->in(1), this); 3181 if (n->outcnt() == 0) { 3182 n->disconnect_inputs(NULL, this); 3183 } 3184 } 3185 break; 3186 } 3187 #ifdef _LP64 3188 case Op_CmpP: 3189 // Do this transformation here to preserve CmpPNode::sub() and 3190 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3191 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3192 Node* in1 = n->in(1); 3193 Node* in2 = n->in(2); 3194 if (!in1->is_DecodeNarrowPtr()) { 3195 in2 = in1; 3196 in1 = n->in(2); 3197 } 3198 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3199 3200 Node* new_in2 = NULL; 3201 if (in2->is_DecodeNarrowPtr()) { 3202 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3203 new_in2 = in2->in(1); 3204 } else if (in2->Opcode() == Op_ConP) { 3205 const Type* t = in2->bottom_type(); 3206 if (t == TypePtr::NULL_PTR) { 3207 assert(in1->is_DecodeN(), "compare klass to null?"); 3208 // Don't convert CmpP null check into CmpN if compressed 3209 // oops implicit null check is not generated. 3210 // This will allow to generate normal oop implicit null check. 3211 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3212 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3213 // 3214 // This transformation together with CastPP transformation above 3215 // will generated code for implicit NULL checks for compressed oops. 3216 // 3217 // The original code after Optimize() 3218 // 3219 // LoadN memory, narrow_oop_reg 3220 // decode narrow_oop_reg, base_reg 3221 // CmpP base_reg, NULL 3222 // CastPP base_reg // NotNull 3223 // Load [base_reg + offset], val_reg 3224 // 3225 // after these transformations will be 3226 // 3227 // LoadN memory, narrow_oop_reg 3228 // CmpN narrow_oop_reg, NULL 3229 // decode_not_null narrow_oop_reg, base_reg 3230 // Load [base_reg + offset], val_reg 3231 // 3232 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3233 // since narrow oops can be used in debug info now (see the code in 3234 // final_graph_reshaping_walk()). 3235 // 3236 // At the end the code will be matched to 3237 // on x86: 3238 // 3239 // Load_narrow_oop memory, narrow_oop_reg 3240 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3241 // NullCheck narrow_oop_reg 3242 // 3243 // and on sparc: 3244 // 3245 // Load_narrow_oop memory, narrow_oop_reg 3246 // decode_not_null narrow_oop_reg, base_reg 3247 // Load [base_reg + offset], val_reg 3248 // NullCheck base_reg 3249 // 3250 } else if (t->isa_oopptr()) { 3251 new_in2 = ConNode::make(t->make_narrowoop()); 3252 } else if (t->isa_klassptr()) { 3253 new_in2 = ConNode::make(t->make_narrowklass()); 3254 } 3255 } 3256 if (new_in2 != NULL) { 3257 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3258 n->subsume_by(cmpN, this); 3259 if (in1->outcnt() == 0) { 3260 in1->disconnect_inputs(NULL, this); 3261 } 3262 if (in2->outcnt() == 0) { 3263 in2->disconnect_inputs(NULL, this); 3264 } 3265 } 3266 } 3267 break; 3268 3269 case Op_DecodeN: 3270 case Op_DecodeNKlass: 3271 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3272 // DecodeN could be pinned when it can't be fold into 3273 // an address expression, see the code for Op_CastPP above. 3274 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3275 break; 3276 3277 case Op_EncodeP: 3278 case Op_EncodePKlass: { 3279 Node* in1 = n->in(1); 3280 if (in1->is_DecodeNarrowPtr()) { 3281 n->subsume_by(in1->in(1), this); 3282 } else if (in1->Opcode() == Op_ConP) { 3283 const Type* t = in1->bottom_type(); 3284 if (t == TypePtr::NULL_PTR) { 3285 assert(t->isa_oopptr(), "null klass?"); 3286 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3287 } else if (t->isa_oopptr()) { 3288 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3289 } else if (t->isa_klassptr()) { 3290 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3291 } 3292 } 3293 if (in1->outcnt() == 0) { 3294 in1->disconnect_inputs(NULL, this); 3295 } 3296 break; 3297 } 3298 3299 case Op_Proj: { 3300 if (OptimizeStringConcat) { 3301 ProjNode* p = n->as_Proj(); 3302 if (p->_is_io_use) { 3303 // Separate projections were used for the exception path which 3304 // are normally removed by a late inline. If it wasn't inlined 3305 // then they will hang around and should just be replaced with 3306 // the original one. 3307 Node* proj = NULL; 3308 // Replace with just one 3309 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3310 Node *use = i.get(); 3311 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3312 proj = use; 3313 break; 3314 } 3315 } 3316 assert(proj != NULL, "must be found"); 3317 p->subsume_by(proj, this); 3318 } 3319 } 3320 break; 3321 } 3322 3323 case Op_Phi: 3324 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3325 // The EncodeP optimization may create Phi with the same edges 3326 // for all paths. It is not handled well by Register Allocator. 3327 Node* unique_in = n->in(1); 3328 assert(unique_in != NULL, ""); 3329 uint cnt = n->req(); 3330 for (uint i = 2; i < cnt; i++) { 3331 Node* m = n->in(i); 3332 assert(m != NULL, ""); 3333 if (unique_in != m) 3334 unique_in = NULL; 3335 } 3336 if (unique_in != NULL) { 3337 n->subsume_by(unique_in, this); 3338 } 3339 } 3340 break; 3341 3342 #endif 3343 3344 #ifdef ASSERT 3345 case Op_CastII: 3346 // Verify that all range check dependent CastII nodes were removed. 3347 if (n->isa_CastII()->has_range_check()) { 3348 n->dump(3); 3349 assert(false, "Range check dependent CastII node was not removed"); 3350 } 3351 break; 3352 #endif 3353 3354 case Op_ModI: 3355 if (UseDivMod) { 3356 // Check if a%b and a/b both exist 3357 Node* d = n->find_similar(Op_DivI); 3358 if (d) { 3359 // Replace them with a fused divmod if supported 3360 if (Matcher::has_match_rule(Op_DivModI)) { 3361 DivModINode* divmod = DivModINode::make(n); 3362 d->subsume_by(divmod->div_proj(), this); 3363 n->subsume_by(divmod->mod_proj(), this); 3364 } else { 3365 // replace a%b with a-((a/b)*b) 3366 Node* mult = new MulINode(d, d->in(2)); 3367 Node* sub = new SubINode(d->in(1), mult); 3368 n->subsume_by(sub, this); 3369 } 3370 } 3371 } 3372 break; 3373 3374 case Op_ModL: 3375 if (UseDivMod) { 3376 // Check if a%b and a/b both exist 3377 Node* d = n->find_similar(Op_DivL); 3378 if (d) { 3379 // Replace them with a fused divmod if supported 3380 if (Matcher::has_match_rule(Op_DivModL)) { 3381 DivModLNode* divmod = DivModLNode::make(n); 3382 d->subsume_by(divmod->div_proj(), this); 3383 n->subsume_by(divmod->mod_proj(), this); 3384 } else { 3385 // replace a%b with a-((a/b)*b) 3386 Node* mult = new MulLNode(d, d->in(2)); 3387 Node* sub = new SubLNode(d->in(1), mult); 3388 n->subsume_by(sub, this); 3389 } 3390 } 3391 } 3392 break; 3393 3394 case Op_LoadVector: 3395 case Op_StoreVector: 3396 break; 3397 3398 case Op_AddReductionVI: 3399 case Op_AddReductionVL: 3400 case Op_AddReductionVF: 3401 case Op_AddReductionVD: 3402 case Op_MulReductionVI: 3403 case Op_MulReductionVL: 3404 case Op_MulReductionVF: 3405 case Op_MulReductionVD: 3406 break; 3407 3408 case Op_PackB: 3409 case Op_PackS: 3410 case Op_PackI: 3411 case Op_PackF: 3412 case Op_PackL: 3413 case Op_PackD: 3414 if (n->req()-1 > 2) { 3415 // Replace many operand PackNodes with a binary tree for matching 3416 PackNode* p = (PackNode*) n; 3417 Node* btp = p->binary_tree_pack(1, n->req()); 3418 n->subsume_by(btp, this); 3419 } 3420 break; 3421 case Op_Loop: 3422 case Op_CountedLoop: 3423 case Op_OuterStripMinedLoop: 3424 if (n->as_Loop()->is_inner_loop()) { 3425 frc.inc_inner_loop_count(); 3426 } 3427 n->as_Loop()->verify_strip_mined(0); 3428 break; 3429 case Op_LShiftI: 3430 case Op_RShiftI: 3431 case Op_URShiftI: 3432 case Op_LShiftL: 3433 case Op_RShiftL: 3434 case Op_URShiftL: 3435 if (Matcher::need_masked_shift_count) { 3436 // The cpu's shift instructions don't restrict the count to the 3437 // lower 5/6 bits. We need to do the masking ourselves. 3438 Node* in2 = n->in(2); 3439 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3440 const TypeInt* t = in2->find_int_type(); 3441 if (t != NULL && t->is_con()) { 3442 juint shift = t->get_con(); 3443 if (shift > mask) { // Unsigned cmp 3444 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3445 } 3446 } else { 3447 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3448 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3449 n->set_req(2, shift); 3450 } 3451 } 3452 if (in2->outcnt() == 0) { // Remove dead node 3453 in2->disconnect_inputs(NULL, this); 3454 } 3455 } 3456 break; 3457 case Op_MemBarStoreStore: 3458 case Op_MemBarRelease: 3459 // Break the link with AllocateNode: it is no longer useful and 3460 // confuses register allocation. 3461 if (n->req() > MemBarNode::Precedent) { 3462 n->set_req(MemBarNode::Precedent, top()); 3463 } 3464 break; 3465 case Op_RangeCheck: { 3466 RangeCheckNode* rc = n->as_RangeCheck(); 3467 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3468 n->subsume_by(iff, this); 3469 frc._tests.push(iff); 3470 break; 3471 } 3472 case Op_ConvI2L: { 3473 if (!Matcher::convi2l_type_required) { 3474 // Code generation on some platforms doesn't need accurate 3475 // ConvI2L types. Widening the type can help remove redundant 3476 // address computations. 3477 n->as_Type()->set_type(TypeLong::INT); 3478 ResourceMark rm; 3479 Node_List wq; 3480 wq.push(n); 3481 for (uint next = 0; next < wq.size(); next++) { 3482 Node *m = wq.at(next); 3483 3484 for(;;) { 3485 // Loop over all nodes with identical inputs edges as m 3486 Node* k = m->find_similar(m->Opcode()); 3487 if (k == NULL) { 3488 break; 3489 } 3490 // Push their uses so we get a chance to remove node made 3491 // redundant 3492 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3493 Node* u = k->fast_out(i); 3494 assert(!wq.contains(u), "shouldn't process one node several times"); 3495 if (u->Opcode() == Op_LShiftL || 3496 u->Opcode() == Op_AddL || 3497 u->Opcode() == Op_SubL || 3498 u->Opcode() == Op_AddP) { 3499 wq.push(u); 3500 } 3501 } 3502 // Replace all nodes with identical edges as m with m 3503 k->subsume_by(m, this); 3504 } 3505 } 3506 } 3507 break; 3508 } 3509 case Op_CmpUL: { 3510 if (!Matcher::has_match_rule(Op_CmpUL)) { 3511 // We don't support unsigned long comparisons. Set 'max_idx_expr' 3512 // to max_julong if < 0 to make the signed comparison fail. 3513 ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1)); 3514 Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos); 3515 Node* orl = new OrLNode(n->in(1), sign_bit_mask); 3516 ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong)); 3517 Node* andl = new AndLNode(orl, remove_sign_mask); 3518 Node* cmp = new CmpLNode(andl, n->in(2)); 3519 n->subsume_by(cmp, this); 3520 } 3521 break; 3522 } 3523 #ifdef ASSERT 3524 case Op_ValueTypePtr: 3525 case Op_ValueType: { 3526 n->dump(-1); 3527 assert(false, "value type node was not removed"); 3528 break; 3529 } 3530 #endif 3531 default: 3532 assert( !n->is_Call(), "" ); 3533 assert( !n->is_Mem(), "" ); 3534 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3535 break; 3536 } 3537 3538 // Collect CFG split points 3539 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3540 frc._tests.push(n); 3541 } 3542 } 3543 3544 //------------------------------final_graph_reshaping_walk--------------------- 3545 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3546 // requires that the walk visits a node's inputs before visiting the node. 3547 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3548 ResourceArea *area = Thread::current()->resource_area(); 3549 Unique_Node_List sfpt(area); 3550 3551 frc._visited.set(root->_idx); // first, mark node as visited 3552 uint cnt = root->req(); 3553 Node *n = root; 3554 uint i = 0; 3555 while (true) { 3556 if (i < cnt) { 3557 // Place all non-visited non-null inputs onto stack 3558 Node* m = n->in(i); 3559 ++i; 3560 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3561 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3562 // compute worst case interpreter size in case of a deoptimization 3563 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3564 3565 sfpt.push(m); 3566 } 3567 cnt = m->req(); 3568 nstack.push(n, i); // put on stack parent and next input's index 3569 n = m; 3570 i = 0; 3571 } 3572 } else { 3573 // Now do post-visit work 3574 final_graph_reshaping_impl( n, frc ); 3575 if (nstack.is_empty()) 3576 break; // finished 3577 n = nstack.node(); // Get node from stack 3578 cnt = n->req(); 3579 i = nstack.index(); 3580 nstack.pop(); // Shift to the next node on stack 3581 } 3582 } 3583 3584 // Skip next transformation if compressed oops are not used. 3585 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3586 (!UseCompressedOops && !UseCompressedClassPointers)) 3587 return; 3588 3589 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3590 // It could be done for an uncommon traps or any safepoints/calls 3591 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3592 while (sfpt.size() > 0) { 3593 n = sfpt.pop(); 3594 JVMState *jvms = n->as_SafePoint()->jvms(); 3595 assert(jvms != NULL, "sanity"); 3596 int start = jvms->debug_start(); 3597 int end = n->req(); 3598 bool is_uncommon = (n->is_CallStaticJava() && 3599 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3600 for (int j = start; j < end; j++) { 3601 Node* in = n->in(j); 3602 if (in->is_DecodeNarrowPtr()) { 3603 bool safe_to_skip = true; 3604 if (!is_uncommon ) { 3605 // Is it safe to skip? 3606 for (uint i = 0; i < in->outcnt(); i++) { 3607 Node* u = in->raw_out(i); 3608 if (!u->is_SafePoint() || 3609 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) { 3610 safe_to_skip = false; 3611 } 3612 } 3613 } 3614 if (safe_to_skip) { 3615 n->set_req(j, in->in(1)); 3616 } 3617 if (in->outcnt() == 0) { 3618 in->disconnect_inputs(NULL, this); 3619 } 3620 } 3621 } 3622 } 3623 } 3624 3625 //------------------------------final_graph_reshaping-------------------------- 3626 // Final Graph Reshaping. 3627 // 3628 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3629 // and not commoned up and forced early. Must come after regular 3630 // optimizations to avoid GVN undoing the cloning. Clone constant 3631 // inputs to Loop Phis; these will be split by the allocator anyways. 3632 // Remove Opaque nodes. 3633 // (2) Move last-uses by commutative operations to the left input to encourage 3634 // Intel update-in-place two-address operations and better register usage 3635 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3636 // calls canonicalizing them back. 3637 // (3) Count the number of double-precision FP ops, single-precision FP ops 3638 // and call sites. On Intel, we can get correct rounding either by 3639 // forcing singles to memory (requires extra stores and loads after each 3640 // FP bytecode) or we can set a rounding mode bit (requires setting and 3641 // clearing the mode bit around call sites). The mode bit is only used 3642 // if the relative frequency of single FP ops to calls is low enough. 3643 // This is a key transform for SPEC mpeg_audio. 3644 // (4) Detect infinite loops; blobs of code reachable from above but not 3645 // below. Several of the Code_Gen algorithms fail on such code shapes, 3646 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3647 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3648 // Detection is by looking for IfNodes where only 1 projection is 3649 // reachable from below or CatchNodes missing some targets. 3650 // (5) Assert for insane oop offsets in debug mode. 3651 3652 bool Compile::final_graph_reshaping() { 3653 // an infinite loop may have been eliminated by the optimizer, 3654 // in which case the graph will be empty. 3655 if (root()->req() == 1) { 3656 record_method_not_compilable("trivial infinite loop"); 3657 return true; 3658 } 3659 3660 // Expensive nodes have their control input set to prevent the GVN 3661 // from freely commoning them. There's no GVN beyond this point so 3662 // no need to keep the control input. We want the expensive nodes to 3663 // be freely moved to the least frequent code path by gcm. 3664 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3665 for (int i = 0; i < expensive_count(); i++) { 3666 _expensive_nodes->at(i)->set_req(0, NULL); 3667 } 3668 3669 Final_Reshape_Counts frc; 3670 3671 // Visit everybody reachable! 3672 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3673 Node_Stack nstack(live_nodes() >> 1); 3674 final_graph_reshaping_walk(nstack, root(), frc); 3675 3676 // Check for unreachable (from below) code (i.e., infinite loops). 3677 for( uint i = 0; i < frc._tests.size(); i++ ) { 3678 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3679 // Get number of CFG targets. 3680 // Note that PCTables include exception targets after calls. 3681 uint required_outcnt = n->required_outcnt(); 3682 if (n->outcnt() != required_outcnt) { 3683 // Check for a few special cases. Rethrow Nodes never take the 3684 // 'fall-thru' path, so expected kids is 1 less. 3685 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3686 if (n->in(0)->in(0)->is_Call()) { 3687 CallNode *call = n->in(0)->in(0)->as_Call(); 3688 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3689 required_outcnt--; // Rethrow always has 1 less kid 3690 } else if (call->req() > TypeFunc::Parms && 3691 call->is_CallDynamicJava()) { 3692 // Check for null receiver. In such case, the optimizer has 3693 // detected that the virtual call will always result in a null 3694 // pointer exception. The fall-through projection of this CatchNode 3695 // will not be populated. 3696 Node *arg0 = call->in(TypeFunc::Parms); 3697 if (arg0->is_Type() && 3698 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3699 required_outcnt--; 3700 } 3701 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3702 call->req() > TypeFunc::Parms+1 && 3703 call->is_CallStaticJava()) { 3704 // Check for negative array length. In such case, the optimizer has 3705 // detected that the allocation attempt will always result in an 3706 // exception. There is no fall-through projection of this CatchNode . 3707 Node *arg1 = call->in(TypeFunc::Parms+1); 3708 if (arg1->is_Type() && 3709 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3710 required_outcnt--; 3711 } 3712 } 3713 } 3714 } 3715 // Recheck with a better notion of 'required_outcnt' 3716 if (n->outcnt() != required_outcnt) { 3717 record_method_not_compilable("malformed control flow"); 3718 return true; // Not all targets reachable! 3719 } 3720 } 3721 // Check that I actually visited all kids. Unreached kids 3722 // must be infinite loops. 3723 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3724 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3725 record_method_not_compilable("infinite loop"); 3726 return true; // Found unvisited kid; must be unreach 3727 } 3728 3729 // Here so verification code in final_graph_reshaping_walk() 3730 // always see an OuterStripMinedLoopEnd 3731 if (n->is_OuterStripMinedLoopEnd()) { 3732 IfNode* init_iff = n->as_If(); 3733 Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt); 3734 n->subsume_by(iff, this); 3735 } 3736 } 3737 3738 // If original bytecodes contained a mixture of floats and doubles 3739 // check if the optimizer has made it homogenous, item (3). 3740 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3741 frc.get_float_count() > 32 && 3742 frc.get_double_count() == 0 && 3743 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3744 set_24_bit_selection_and_mode( false, true ); 3745 } 3746 3747 set_java_calls(frc.get_java_call_count()); 3748 set_inner_loops(frc.get_inner_loop_count()); 3749 3750 // No infinite loops, no reason to bail out. 3751 return false; 3752 } 3753 3754 //-----------------------------too_many_traps---------------------------------- 3755 // Report if there are too many traps at the current method and bci. 3756 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3757 bool Compile::too_many_traps(ciMethod* method, 3758 int bci, 3759 Deoptimization::DeoptReason reason) { 3760 ciMethodData* md = method->method_data(); 3761 if (md->is_empty()) { 3762 // Assume the trap has not occurred, or that it occurred only 3763 // because of a transient condition during start-up in the interpreter. 3764 return false; 3765 } 3766 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3767 if (md->has_trap_at(bci, m, reason) != 0) { 3768 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3769 // Also, if there are multiple reasons, or if there is no per-BCI record, 3770 // assume the worst. 3771 if (log()) 3772 log()->elem("observe trap='%s' count='%d'", 3773 Deoptimization::trap_reason_name(reason), 3774 md->trap_count(reason)); 3775 return true; 3776 } else { 3777 // Ignore method/bci and see if there have been too many globally. 3778 return too_many_traps(reason, md); 3779 } 3780 } 3781 3782 // Less-accurate variant which does not require a method and bci. 3783 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3784 ciMethodData* logmd) { 3785 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3786 // Too many traps globally. 3787 // Note that we use cumulative trap_count, not just md->trap_count. 3788 if (log()) { 3789 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3790 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3791 Deoptimization::trap_reason_name(reason), 3792 mcount, trap_count(reason)); 3793 } 3794 return true; 3795 } else { 3796 // The coast is clear. 3797 return false; 3798 } 3799 } 3800 3801 //--------------------------too_many_recompiles-------------------------------- 3802 // Report if there are too many recompiles at the current method and bci. 3803 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3804 // Is not eager to return true, since this will cause the compiler to use 3805 // Action_none for a trap point, to avoid too many recompilations. 3806 bool Compile::too_many_recompiles(ciMethod* method, 3807 int bci, 3808 Deoptimization::DeoptReason reason) { 3809 ciMethodData* md = method->method_data(); 3810 if (md->is_empty()) { 3811 // Assume the trap has not occurred, or that it occurred only 3812 // because of a transient condition during start-up in the interpreter. 3813 return false; 3814 } 3815 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3816 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3817 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3818 Deoptimization::DeoptReason per_bc_reason 3819 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3820 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3821 if ((per_bc_reason == Deoptimization::Reason_none 3822 || md->has_trap_at(bci, m, reason) != 0) 3823 // The trap frequency measure we care about is the recompile count: 3824 && md->trap_recompiled_at(bci, m) 3825 && md->overflow_recompile_count() >= bc_cutoff) { 3826 // Do not emit a trap here if it has already caused recompilations. 3827 // Also, if there are multiple reasons, or if there is no per-BCI record, 3828 // assume the worst. 3829 if (log()) 3830 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3831 Deoptimization::trap_reason_name(reason), 3832 md->trap_count(reason), 3833 md->overflow_recompile_count()); 3834 return true; 3835 } else if (trap_count(reason) != 0 3836 && decompile_count() >= m_cutoff) { 3837 // Too many recompiles globally, and we have seen this sort of trap. 3838 // Use cumulative decompile_count, not just md->decompile_count. 3839 if (log()) 3840 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3841 Deoptimization::trap_reason_name(reason), 3842 md->trap_count(reason), trap_count(reason), 3843 md->decompile_count(), decompile_count()); 3844 return true; 3845 } else { 3846 // The coast is clear. 3847 return false; 3848 } 3849 } 3850 3851 // Compute when not to trap. Used by matching trap based nodes and 3852 // NullCheck optimization. 3853 void Compile::set_allowed_deopt_reasons() { 3854 _allowed_reasons = 0; 3855 if (is_method_compilation()) { 3856 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3857 assert(rs < BitsPerInt, "recode bit map"); 3858 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3859 _allowed_reasons |= nth_bit(rs); 3860 } 3861 } 3862 } 3863 } 3864 3865 #ifndef PRODUCT 3866 //------------------------------verify_graph_edges--------------------------- 3867 // Walk the Graph and verify that there is a one-to-one correspondence 3868 // between Use-Def edges and Def-Use edges in the graph. 3869 void Compile::verify_graph_edges(bool no_dead_code) { 3870 if (VerifyGraphEdges) { 3871 ResourceArea *area = Thread::current()->resource_area(); 3872 Unique_Node_List visited(area); 3873 // Call recursive graph walk to check edges 3874 _root->verify_edges(visited); 3875 if (no_dead_code) { 3876 // Now make sure that no visited node is used by an unvisited node. 3877 bool dead_nodes = false; 3878 Unique_Node_List checked(area); 3879 while (visited.size() > 0) { 3880 Node* n = visited.pop(); 3881 checked.push(n); 3882 for (uint i = 0; i < n->outcnt(); i++) { 3883 Node* use = n->raw_out(i); 3884 if (checked.member(use)) continue; // already checked 3885 if (visited.member(use)) continue; // already in the graph 3886 if (use->is_Con()) continue; // a dead ConNode is OK 3887 // At this point, we have found a dead node which is DU-reachable. 3888 if (!dead_nodes) { 3889 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3890 dead_nodes = true; 3891 } 3892 use->dump(2); 3893 tty->print_cr("---"); 3894 checked.push(use); // No repeats; pretend it is now checked. 3895 } 3896 } 3897 assert(!dead_nodes, "using nodes must be reachable from root"); 3898 } 3899 } 3900 } 3901 3902 // Verify GC barriers consistency 3903 // Currently supported: 3904 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3905 void Compile::verify_barriers() { 3906 #if INCLUDE_G1GC 3907 if (UseG1GC) { 3908 // Verify G1 pre-barriers 3909 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()); 3910 3911 ResourceArea *area = Thread::current()->resource_area(); 3912 Unique_Node_List visited(area); 3913 Node_List worklist(area); 3914 // We're going to walk control flow backwards starting from the Root 3915 worklist.push(_root); 3916 while (worklist.size() > 0) { 3917 Node* x = worklist.pop(); 3918 if (x == NULL || x == top()) continue; 3919 if (visited.member(x)) { 3920 continue; 3921 } else { 3922 visited.push(x); 3923 } 3924 3925 if (x->is_Region()) { 3926 for (uint i = 1; i < x->req(); i++) { 3927 worklist.push(x->in(i)); 3928 } 3929 } else { 3930 worklist.push(x->in(0)); 3931 // We are looking for the pattern: 3932 // /->ThreadLocal 3933 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3934 // \->ConI(0) 3935 // We want to verify that the If and the LoadB have the same control 3936 // See GraphKit::g1_write_barrier_pre() 3937 if (x->is_If()) { 3938 IfNode *iff = x->as_If(); 3939 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3940 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3941 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3942 && cmp->in(1)->is_Load()) { 3943 LoadNode* load = cmp->in(1)->as_Load(); 3944 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3945 && load->in(2)->in(3)->is_Con() 3946 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3947 3948 Node* if_ctrl = iff->in(0); 3949 Node* load_ctrl = load->in(0); 3950 3951 if (if_ctrl != load_ctrl) { 3952 // Skip possible CProj->NeverBranch in infinite loops 3953 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3954 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3955 if_ctrl = if_ctrl->in(0)->in(0); 3956 } 3957 } 3958 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3959 } 3960 } 3961 } 3962 } 3963 } 3964 } 3965 } 3966 #endif 3967 } 3968 3969 #endif 3970 3971 // The Compile object keeps track of failure reasons separately from the ciEnv. 3972 // This is required because there is not quite a 1-1 relation between the 3973 // ciEnv and its compilation task and the Compile object. Note that one 3974 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3975 // to backtrack and retry without subsuming loads. Other than this backtracking 3976 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3977 // by the logic in C2Compiler. 3978 void Compile::record_failure(const char* reason) { 3979 if (log() != NULL) { 3980 log()->elem("failure reason='%s' phase='compile'", reason); 3981 } 3982 if (_failure_reason == NULL) { 3983 // Record the first failure reason. 3984 _failure_reason = reason; 3985 } 3986 3987 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3988 C->print_method(PHASE_FAILURE); 3989 } 3990 _root = NULL; // flush the graph, too 3991 } 3992 3993 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3994 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3995 _phase_name(name), _dolog(CITimeVerbose) 3996 { 3997 if (_dolog) { 3998 C = Compile::current(); 3999 _log = C->log(); 4000 } else { 4001 C = NULL; 4002 _log = NULL; 4003 } 4004 if (_log != NULL) { 4005 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4006 _log->stamp(); 4007 _log->end_head(); 4008 } 4009 } 4010 4011 Compile::TracePhase::~TracePhase() { 4012 4013 C = Compile::current(); 4014 if (_dolog) { 4015 _log = C->log(); 4016 } else { 4017 _log = NULL; 4018 } 4019 4020 #ifdef ASSERT 4021 if (PrintIdealNodeCount) { 4022 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 4023 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 4024 } 4025 4026 if (VerifyIdealNodeCount) { 4027 Compile::current()->print_missing_nodes(); 4028 } 4029 #endif 4030 4031 if (_log != NULL) { 4032 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 4033 } 4034 } 4035 4036 //============================================================================= 4037 // Two Constant's are equal when the type and the value are equal. 4038 bool Compile::Constant::operator==(const Constant& other) { 4039 if (type() != other.type() ) return false; 4040 if (can_be_reused() != other.can_be_reused()) return false; 4041 // For floating point values we compare the bit pattern. 4042 switch (type()) { 4043 case T_INT: 4044 case T_FLOAT: return (_v._value.i == other._v._value.i); 4045 case T_LONG: 4046 case T_DOUBLE: return (_v._value.j == other._v._value.j); 4047 case T_OBJECT: 4048 case T_ADDRESS: return (_v._value.l == other._v._value.l); 4049 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 4050 case T_METADATA: return (_v._metadata == other._v._metadata); 4051 default: ShouldNotReachHere(); return false; 4052 } 4053 } 4054 4055 static int type_to_size_in_bytes(BasicType t) { 4056 switch (t) { 4057 case T_INT: return sizeof(jint ); 4058 case T_LONG: return sizeof(jlong ); 4059 case T_FLOAT: return sizeof(jfloat ); 4060 case T_DOUBLE: return sizeof(jdouble); 4061 case T_METADATA: return sizeof(Metadata*); 4062 // We use T_VOID as marker for jump-table entries (labels) which 4063 // need an internal word relocation. 4064 case T_VOID: 4065 case T_ADDRESS: 4066 case T_OBJECT: return sizeof(jobject); 4067 default: 4068 ShouldNotReachHere(); 4069 return -1; 4070 } 4071 } 4072 4073 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4074 // sort descending 4075 if (a->freq() > b->freq()) return -1; 4076 if (a->freq() < b->freq()) return 1; 4077 return 0; 4078 } 4079 4080 void Compile::ConstantTable::calculate_offsets_and_size() { 4081 // First, sort the array by frequencies. 4082 _constants.sort(qsort_comparator); 4083 4084 #ifdef ASSERT 4085 // Make sure all jump-table entries were sorted to the end of the 4086 // array (they have a negative frequency). 4087 bool found_void = false; 4088 for (int i = 0; i < _constants.length(); i++) { 4089 Constant con = _constants.at(i); 4090 if (con.type() == T_VOID) 4091 found_void = true; // jump-tables 4092 else 4093 assert(!found_void, "wrong sorting"); 4094 } 4095 #endif 4096 4097 int offset = 0; 4098 for (int i = 0; i < _constants.length(); i++) { 4099 Constant* con = _constants.adr_at(i); 4100 4101 // Align offset for type. 4102 int typesize = type_to_size_in_bytes(con->type()); 4103 offset = align_up(offset, typesize); 4104 con->set_offset(offset); // set constant's offset 4105 4106 if (con->type() == T_VOID) { 4107 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4108 offset = offset + typesize * n->outcnt(); // expand jump-table 4109 } else { 4110 offset = offset + typesize; 4111 } 4112 } 4113 4114 // Align size up to the next section start (which is insts; see 4115 // CodeBuffer::align_at_start). 4116 assert(_size == -1, "already set?"); 4117 _size = align_up(offset, (int)CodeEntryAlignment); 4118 } 4119 4120 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4121 MacroAssembler _masm(&cb); 4122 for (int i = 0; i < _constants.length(); i++) { 4123 Constant con = _constants.at(i); 4124 address constant_addr = NULL; 4125 switch (con.type()) { 4126 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4127 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4128 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4129 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4130 case T_OBJECT: { 4131 jobject obj = con.get_jobject(); 4132 int oop_index = _masm.oop_recorder()->find_index(obj); 4133 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4134 break; 4135 } 4136 case T_ADDRESS: { 4137 address addr = (address) con.get_jobject(); 4138 constant_addr = _masm.address_constant(addr); 4139 break; 4140 } 4141 // We use T_VOID as marker for jump-table entries (labels) which 4142 // need an internal word relocation. 4143 case T_VOID: { 4144 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4145 // Fill the jump-table with a dummy word. The real value is 4146 // filled in later in fill_jump_table. 4147 address dummy = (address) n; 4148 constant_addr = _masm.address_constant(dummy); 4149 // Expand jump-table 4150 for (uint i = 1; i < n->outcnt(); i++) { 4151 address temp_addr = _masm.address_constant(dummy + i); 4152 assert(temp_addr, "consts section too small"); 4153 } 4154 break; 4155 } 4156 case T_METADATA: { 4157 Metadata* obj = con.get_metadata(); 4158 int metadata_index = _masm.oop_recorder()->find_index(obj); 4159 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4160 break; 4161 } 4162 default: ShouldNotReachHere(); 4163 } 4164 assert(constant_addr, "consts section too small"); 4165 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4166 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4167 } 4168 } 4169 4170 int Compile::ConstantTable::find_offset(Constant& con) const { 4171 int idx = _constants.find(con); 4172 guarantee(idx != -1, "constant must be in constant table"); 4173 int offset = _constants.at(idx).offset(); 4174 guarantee(offset != -1, "constant table not emitted yet?"); 4175 return offset; 4176 } 4177 4178 void Compile::ConstantTable::add(Constant& con) { 4179 if (con.can_be_reused()) { 4180 int idx = _constants.find(con); 4181 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4182 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4183 return; 4184 } 4185 } 4186 (void) _constants.append(con); 4187 } 4188 4189 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4190 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4191 Constant con(type, value, b->_freq); 4192 add(con); 4193 return con; 4194 } 4195 4196 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4197 Constant con(metadata); 4198 add(con); 4199 return con; 4200 } 4201 4202 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4203 jvalue value; 4204 BasicType type = oper->type()->basic_type(); 4205 switch (type) { 4206 case T_LONG: value.j = oper->constantL(); break; 4207 case T_FLOAT: value.f = oper->constantF(); break; 4208 case T_DOUBLE: value.d = oper->constantD(); break; 4209 case T_OBJECT: 4210 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4211 case T_METADATA: return add((Metadata*)oper->constant()); break; 4212 default: guarantee(false, "unhandled type: %s", type2name(type)); 4213 } 4214 return add(n, type, value); 4215 } 4216 4217 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4218 jvalue value; 4219 // We can use the node pointer here to identify the right jump-table 4220 // as this method is called from Compile::Fill_buffer right before 4221 // the MachNodes are emitted and the jump-table is filled (means the 4222 // MachNode pointers do not change anymore). 4223 value.l = (jobject) n; 4224 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4225 add(con); 4226 return con; 4227 } 4228 4229 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4230 // If called from Compile::scratch_emit_size do nothing. 4231 if (Compile::current()->in_scratch_emit_size()) return; 4232 4233 assert(labels.is_nonempty(), "must be"); 4234 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4235 4236 // Since MachConstantNode::constant_offset() also contains 4237 // table_base_offset() we need to subtract the table_base_offset() 4238 // to get the plain offset into the constant table. 4239 int offset = n->constant_offset() - table_base_offset(); 4240 4241 MacroAssembler _masm(&cb); 4242 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4243 4244 for (uint i = 0; i < n->outcnt(); i++) { 4245 address* constant_addr = &jump_table_base[i]; 4246 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)); 4247 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4248 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4249 } 4250 } 4251 4252 //----------------------------static_subtype_check----------------------------- 4253 // Shortcut important common cases when superklass is exact: 4254 // (0) superklass is java.lang.Object (can occur in reflective code) 4255 // (1) subklass is already limited to a subtype of superklass => always ok 4256 // (2) subklass does not overlap with superklass => always fail 4257 // (3) superklass has NO subtypes and we can check with a simple compare. 4258 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4259 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4260 return SSC_full_test; // Let caller generate the general case. 4261 } 4262 4263 if (!EnableValhalla && superk == env()->Object_klass()) { 4264 return SSC_always_true; // (0) this test cannot fail 4265 } 4266 4267 ciType* superelem = superk; 4268 if (superelem->is_array_klass()) 4269 superelem = superelem->as_array_klass()->base_element_type(); 4270 4271 if (!subk->is_interface()) { // cannot trust static interface types yet 4272 if (subk->is_subtype_of(superk)) { 4273 return SSC_always_true; // (1) false path dead; no dynamic test needed 4274 } 4275 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4276 !superk->is_subtype_of(subk)) { 4277 return SSC_always_false; 4278 } 4279 } 4280 4281 // If casting to an instance klass, it must have no subtypes 4282 if (superk->is_interface()) { 4283 // Cannot trust interfaces yet. 4284 // %%% S.B. superk->nof_implementors() == 1 4285 } else if (superelem->is_instance_klass()) { 4286 ciInstanceKlass* ik = superelem->as_instance_klass(); 4287 if (!ik->has_subklass() && !ik->is_interface()) { 4288 if (!ik->is_final()) { 4289 // Add a dependency if there is a chance of a later subclass. 4290 dependencies()->assert_leaf_type(ik); 4291 } 4292 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4293 } 4294 } else { 4295 // A primitive array type has no subtypes. 4296 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4297 } 4298 4299 return SSC_full_test; 4300 } 4301 4302 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4303 #ifdef _LP64 4304 // The scaled index operand to AddP must be a clean 64-bit value. 4305 // Java allows a 32-bit int to be incremented to a negative 4306 // value, which appears in a 64-bit register as a large 4307 // positive number. Using that large positive number as an 4308 // operand in pointer arithmetic has bad consequences. 4309 // On the other hand, 32-bit overflow is rare, and the possibility 4310 // can often be excluded, if we annotate the ConvI2L node with 4311 // a type assertion that its value is known to be a small positive 4312 // number. (The prior range check has ensured this.) 4313 // This assertion is used by ConvI2LNode::Ideal. 4314 int index_max = max_jint - 1; // array size is max_jint, index is one less 4315 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4316 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4317 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4318 #endif 4319 return idx; 4320 } 4321 4322 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4323 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4324 if (ctrl != NULL) { 4325 // Express control dependency by a CastII node with a narrow type. 4326 value = new CastIINode(value, itype, false, true /* range check dependency */); 4327 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4328 // node from floating above the range check during loop optimizations. Otherwise, the 4329 // ConvI2L node may be eliminated independently of the range check, causing the data path 4330 // to become TOP while the control path is still there (although it's unreachable). 4331 value->set_req(0, ctrl); 4332 // Save CastII node to remove it after loop optimizations. 4333 phase->C->add_range_check_cast(value); 4334 value = phase->transform(value); 4335 } 4336 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4337 return phase->transform(new ConvI2LNode(value, ltype)); 4338 } 4339 4340 // The message about the current inlining is accumulated in 4341 // _print_inlining_stream and transfered into the _print_inlining_list 4342 // once we know whether inlining succeeds or not. For regular 4343 // inlining, messages are appended to the buffer pointed by 4344 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4345 // a new buffer is added after _print_inlining_idx in the list. This 4346 // way we can update the inlining message for late inlining call site 4347 // when the inlining is attempted again. 4348 void Compile::print_inlining_init() { 4349 if (print_inlining() || print_intrinsics()) { 4350 _print_inlining_stream = new stringStream(); 4351 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4352 } 4353 } 4354 4355 void Compile::print_inlining_reinit() { 4356 if (print_inlining() || print_intrinsics()) { 4357 // Re allocate buffer when we change ResourceMark 4358 _print_inlining_stream = new stringStream(); 4359 } 4360 } 4361 4362 void Compile::print_inlining_reset() { 4363 _print_inlining_stream->reset(); 4364 } 4365 4366 void Compile::print_inlining_commit() { 4367 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4368 // Transfer the message from _print_inlining_stream to the current 4369 // _print_inlining_list buffer and clear _print_inlining_stream. 4370 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4371 print_inlining_reset(); 4372 } 4373 4374 void Compile::print_inlining_push() { 4375 // Add new buffer to the _print_inlining_list at current position 4376 _print_inlining_idx++; 4377 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4378 } 4379 4380 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4381 return _print_inlining_list->at(_print_inlining_idx); 4382 } 4383 4384 void Compile::print_inlining_update(CallGenerator* cg) { 4385 if (print_inlining() || print_intrinsics()) { 4386 if (!cg->is_late_inline()) { 4387 if (print_inlining_current().cg() != NULL) { 4388 print_inlining_push(); 4389 } 4390 print_inlining_commit(); 4391 } else { 4392 if (print_inlining_current().cg() != cg && 4393 (print_inlining_current().cg() != NULL || 4394 print_inlining_current().ss()->size() != 0)) { 4395 print_inlining_push(); 4396 } 4397 print_inlining_commit(); 4398 print_inlining_current().set_cg(cg); 4399 } 4400 } 4401 } 4402 4403 void Compile::print_inlining_move_to(CallGenerator* cg) { 4404 // We resume inlining at a late inlining call site. Locate the 4405 // corresponding inlining buffer so that we can update it. 4406 if (print_inlining()) { 4407 for (int i = 0; i < _print_inlining_list->length(); i++) { 4408 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4409 _print_inlining_idx = i; 4410 return; 4411 } 4412 } 4413 ShouldNotReachHere(); 4414 } 4415 } 4416 4417 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4418 if (print_inlining()) { 4419 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4420 assert(print_inlining_current().cg() == cg, "wrong entry"); 4421 // replace message with new message 4422 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4423 print_inlining_commit(); 4424 print_inlining_current().set_cg(cg); 4425 } 4426 } 4427 4428 void Compile::print_inlining_assert_ready() { 4429 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4430 } 4431 4432 void Compile::process_print_inlining() { 4433 bool do_print_inlining = print_inlining() || print_intrinsics(); 4434 if (do_print_inlining || log() != NULL) { 4435 // Print inlining message for candidates that we couldn't inline 4436 // for lack of space 4437 for (int i = 0; i < _late_inlines.length(); i++) { 4438 CallGenerator* cg = _late_inlines.at(i); 4439 if (!cg->is_mh_late_inline()) { 4440 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4441 if (do_print_inlining) { 4442 cg->print_inlining_late(msg); 4443 } 4444 log_late_inline_failure(cg, msg); 4445 } 4446 } 4447 } 4448 if (do_print_inlining) { 4449 ResourceMark rm; 4450 stringStream ss; 4451 for (int i = 0; i < _print_inlining_list->length(); i++) { 4452 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4453 } 4454 size_t end = ss.size(); 4455 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4456 strncpy(_print_inlining_output, ss.base(), end+1); 4457 _print_inlining_output[end] = 0; 4458 } 4459 } 4460 4461 void Compile::dump_print_inlining() { 4462 if (_print_inlining_output != NULL) { 4463 tty->print_raw(_print_inlining_output); 4464 } 4465 } 4466 4467 void Compile::log_late_inline(CallGenerator* cg) { 4468 if (log() != NULL) { 4469 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4470 cg->unique_id()); 4471 JVMState* p = cg->call_node()->jvms(); 4472 while (p != NULL) { 4473 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4474 p = p->caller(); 4475 } 4476 log()->tail("late_inline"); 4477 } 4478 } 4479 4480 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4481 log_late_inline(cg); 4482 if (log() != NULL) { 4483 log()->inline_fail(msg); 4484 } 4485 } 4486 4487 void Compile::log_inline_id(CallGenerator* cg) { 4488 if (log() != NULL) { 4489 // The LogCompilation tool needs a unique way to identify late 4490 // inline call sites. This id must be unique for this call site in 4491 // this compilation. Try to have it unique across compilations as 4492 // well because it can be convenient when grepping through the log 4493 // file. 4494 // Distinguish OSR compilations from others in case CICountOSR is 4495 // on. 4496 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4497 cg->set_unique_id(id); 4498 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4499 } 4500 } 4501 4502 void Compile::log_inline_failure(const char* msg) { 4503 if (C->log() != NULL) { 4504 C->log()->inline_fail(msg); 4505 } 4506 } 4507 4508 4509 // Dump inlining replay data to the stream. 4510 // Don't change thread state and acquire any locks. 4511 void Compile::dump_inline_data(outputStream* out) { 4512 InlineTree* inl_tree = ilt(); 4513 if (inl_tree != NULL) { 4514 out->print(" inline %d", inl_tree->count()); 4515 inl_tree->dump_replay_data(out); 4516 } 4517 } 4518 4519 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4520 if (n1->Opcode() < n2->Opcode()) return -1; 4521 else if (n1->Opcode() > n2->Opcode()) return 1; 4522 4523 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4524 for (uint i = 1; i < n1->req(); i++) { 4525 if (n1->in(i) < n2->in(i)) return -1; 4526 else if (n1->in(i) > n2->in(i)) return 1; 4527 } 4528 4529 return 0; 4530 } 4531 4532 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4533 Node* n1 = *n1p; 4534 Node* n2 = *n2p; 4535 4536 return cmp_expensive_nodes(n1, n2); 4537 } 4538 4539 void Compile::sort_expensive_nodes() { 4540 if (!expensive_nodes_sorted()) { 4541 _expensive_nodes->sort(cmp_expensive_nodes); 4542 } 4543 } 4544 4545 bool Compile::expensive_nodes_sorted() const { 4546 for (int i = 1; i < _expensive_nodes->length(); i++) { 4547 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4548 return false; 4549 } 4550 } 4551 return true; 4552 } 4553 4554 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4555 if (_expensive_nodes->length() == 0) { 4556 return false; 4557 } 4558 4559 assert(OptimizeExpensiveOps, "optimization off?"); 4560 4561 // Take this opportunity to remove dead nodes from the list 4562 int j = 0; 4563 for (int i = 0; i < _expensive_nodes->length(); i++) { 4564 Node* n = _expensive_nodes->at(i); 4565 if (!n->is_unreachable(igvn)) { 4566 assert(n->is_expensive(), "should be expensive"); 4567 _expensive_nodes->at_put(j, n); 4568 j++; 4569 } 4570 } 4571 _expensive_nodes->trunc_to(j); 4572 4573 // Then sort the list so that similar nodes are next to each other 4574 // and check for at least two nodes of identical kind with same data 4575 // inputs. 4576 sort_expensive_nodes(); 4577 4578 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4579 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4580 return true; 4581 } 4582 } 4583 4584 return false; 4585 } 4586 4587 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4588 if (_expensive_nodes->length() == 0) { 4589 return; 4590 } 4591 4592 assert(OptimizeExpensiveOps, "optimization off?"); 4593 4594 // Sort to bring similar nodes next to each other and clear the 4595 // control input of nodes for which there's only a single copy. 4596 sort_expensive_nodes(); 4597 4598 int j = 0; 4599 int identical = 0; 4600 int i = 0; 4601 bool modified = false; 4602 for (; i < _expensive_nodes->length()-1; i++) { 4603 assert(j <= i, "can't write beyond current index"); 4604 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4605 identical++; 4606 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4607 continue; 4608 } 4609 if (identical > 0) { 4610 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4611 identical = 0; 4612 } else { 4613 Node* n = _expensive_nodes->at(i); 4614 igvn.replace_input_of(n, 0, NULL); 4615 igvn.hash_insert(n); 4616 modified = true; 4617 } 4618 } 4619 if (identical > 0) { 4620 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4621 } else if (_expensive_nodes->length() >= 1) { 4622 Node* n = _expensive_nodes->at(i); 4623 igvn.replace_input_of(n, 0, NULL); 4624 igvn.hash_insert(n); 4625 modified = true; 4626 } 4627 _expensive_nodes->trunc_to(j); 4628 if (modified) { 4629 igvn.optimize(); 4630 } 4631 } 4632 4633 void Compile::add_expensive_node(Node * n) { 4634 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4635 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4636 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4637 if (OptimizeExpensiveOps) { 4638 _expensive_nodes->append(n); 4639 } else { 4640 // Clear control input and let IGVN optimize expensive nodes if 4641 // OptimizeExpensiveOps is off. 4642 n->set_req(0, NULL); 4643 } 4644 } 4645 4646 /** 4647 * Remove the speculative part of types and clean up the graph 4648 */ 4649 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4650 if (UseTypeSpeculation) { 4651 Unique_Node_List worklist; 4652 worklist.push(root()); 4653 int modified = 0; 4654 // Go over all type nodes that carry a speculative type, drop the 4655 // speculative part of the type and enqueue the node for an igvn 4656 // which may optimize it out. 4657 for (uint next = 0; next < worklist.size(); ++next) { 4658 Node *n = worklist.at(next); 4659 if (n->is_Type()) { 4660 TypeNode* tn = n->as_Type(); 4661 const Type* t = tn->type(); 4662 const Type* t_no_spec = t->remove_speculative(); 4663 if (t_no_spec != t) { 4664 bool in_hash = igvn.hash_delete(n); 4665 assert(in_hash, "node should be in igvn hash table"); 4666 tn->set_type(t_no_spec); 4667 igvn.hash_insert(n); 4668 igvn._worklist.push(n); // give it a chance to go away 4669 modified++; 4670 } 4671 } 4672 uint max = n->len(); 4673 for( uint i = 0; i < max; ++i ) { 4674 Node *m = n->in(i); 4675 if (not_a_node(m)) continue; 4676 worklist.push(m); 4677 } 4678 } 4679 // Drop the speculative part of all types in the igvn's type table 4680 igvn.remove_speculative_types(); 4681 if (modified > 0) { 4682 igvn.optimize(); 4683 } 4684 #ifdef ASSERT 4685 // Verify that after the IGVN is over no speculative type has resurfaced 4686 worklist.clear(); 4687 worklist.push(root()); 4688 for (uint next = 0; next < worklist.size(); ++next) { 4689 Node *n = worklist.at(next); 4690 const Type* t = igvn.type_or_null(n); 4691 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4692 if (n->is_Type()) { 4693 t = n->as_Type()->type(); 4694 assert(t == t->remove_speculative(), "no more speculative types"); 4695 } 4696 uint max = n->len(); 4697 for( uint i = 0; i < max; ++i ) { 4698 Node *m = n->in(i); 4699 if (not_a_node(m)) continue; 4700 worklist.push(m); 4701 } 4702 } 4703 igvn.check_no_speculative_types(); 4704 #endif 4705 } 4706 } 4707 4708 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) { 4709 const TypeInstPtr* ta = phase->type(a)->isa_instptr(); 4710 const TypeInstPtr* tb = phase->type(b)->isa_instptr(); 4711 if (!EnableValhalla || ta == NULL || tb == NULL || 4712 ta->is_zero_type() || tb->is_zero_type() || 4713 !ta->can_be_value_type() || !tb->can_be_value_type()) { 4714 // Use old acmp if one operand is null or not a value type 4715 return new CmpPNode(a, b); 4716 } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) { 4717 // We know that one operand is a value type. Therefore, 4718 // new acmp will only return true if both operands are NULL. 4719 // Check if both operands are null by or'ing the oops. 4720 a = phase->transform(new CastP2XNode(NULL, a)); 4721 b = phase->transform(new CastP2XNode(NULL, b)); 4722 a = phase->transform(new OrXNode(a, b)); 4723 return new CmpXNode(a, phase->MakeConX(0)); 4724 } 4725 // Use new acmp 4726 return NULL; 4727 } 4728 4729 // Auxiliary method to support randomized stressing/fuzzing. 4730 // 4731 // This method can be called the arbitrary number of times, with current count 4732 // as the argument. The logic allows selecting a single candidate from the 4733 // running list of candidates as follows: 4734 // int count = 0; 4735 // Cand* selected = null; 4736 // while(cand = cand->next()) { 4737 // if (randomized_select(++count)) { 4738 // selected = cand; 4739 // } 4740 // } 4741 // 4742 // Including count equalizes the chances any candidate is "selected". 4743 // This is useful when we don't have the complete list of candidates to choose 4744 // from uniformly. In this case, we need to adjust the randomicity of the 4745 // selection, or else we will end up biasing the selection towards the latter 4746 // candidates. 4747 // 4748 // Quick back-envelope calculation shows that for the list of n candidates 4749 // the equal probability for the candidate to persist as "best" can be 4750 // achieved by replacing it with "next" k-th candidate with the probability 4751 // of 1/k. It can be easily shown that by the end of the run, the 4752 // probability for any candidate is converged to 1/n, thus giving the 4753 // uniform distribution among all the candidates. 4754 // 4755 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4756 #define RANDOMIZED_DOMAIN_POW 29 4757 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4758 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4759 bool Compile::randomized_select(int count) { 4760 assert(count > 0, "only positive"); 4761 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4762 } 4763 4764 CloneMap& Compile::clone_map() { return _clone_map; } 4765 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4766 4767 void NodeCloneInfo::dump() const { 4768 tty->print(" {%d:%d} ", idx(), gen()); 4769 } 4770 4771 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4772 uint64_t val = value(old->_idx); 4773 NodeCloneInfo cio(val); 4774 assert(val != 0, "old node should be in the map"); 4775 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4776 insert(nnn->_idx, cin.get()); 4777 #ifndef PRODUCT 4778 if (is_debug()) { 4779 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4780 } 4781 #endif 4782 } 4783 4784 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4785 NodeCloneInfo cio(value(old->_idx)); 4786 if (cio.get() == 0) { 4787 cio.set(old->_idx, 0); 4788 insert(old->_idx, cio.get()); 4789 #ifndef PRODUCT 4790 if (is_debug()) { 4791 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4792 } 4793 #endif 4794 } 4795 clone(old, nnn, gen); 4796 } 4797 4798 int CloneMap::max_gen() const { 4799 int g = 0; 4800 DictI di(_dict); 4801 for(; di.test(); ++di) { 4802 int t = gen(di._key); 4803 if (g < t) { 4804 g = t; 4805 #ifndef PRODUCT 4806 if (is_debug()) { 4807 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4808 } 4809 #endif 4810 } 4811 } 4812 return g; 4813 } 4814 4815 void CloneMap::dump(node_idx_t key) const { 4816 uint64_t val = value(key); 4817 if (val != 0) { 4818 NodeCloneInfo ni(val); 4819 ni.dump(); 4820 } 4821 }