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