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