1 /* 2 * Copyright (c) 1997, 2016, 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/copy.hpp" 75 76 77 // -------------------- Compile::mach_constant_base_node ----------------------- 78 // Constant table base node singleton. 79 MachConstantBaseNode* Compile::mach_constant_base_node() { 80 if (_mach_constant_base_node == NULL) { 81 _mach_constant_base_node = new MachConstantBaseNode(); 82 _mach_constant_base_node->add_req(C->root()); 83 } 84 return _mach_constant_base_node; 85 } 86 87 88 /// Support for intrinsics. 89 90 // Return the index at which m must be inserted (or already exists). 91 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 92 class IntrinsicDescPair { 93 private: 94 ciMethod* _m; 95 bool _is_virtual; 96 public: 97 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 98 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 99 ciMethod* m= elt->method(); 100 ciMethod* key_m = key->_m; 101 if (key_m < m) return -1; 102 else if (key_m > m) return 1; 103 else { 104 bool is_virtual = elt->is_virtual(); 105 bool key_virtual = key->_is_virtual; 106 if (key_virtual < is_virtual) return -1; 107 else if (key_virtual > is_virtual) return 1; 108 else return 0; 109 } 110 } 111 }; 112 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 113 #ifdef ASSERT 114 for (int i = 1; i < _intrinsics->length(); i++) { 115 CallGenerator* cg1 = _intrinsics->at(i-1); 116 CallGenerator* cg2 = _intrinsics->at(i); 117 assert(cg1->method() != cg2->method() 118 ? cg1->method() < cg2->method() 119 : cg1->is_virtual() < cg2->is_virtual(), 120 "compiler intrinsics list must stay sorted"); 121 } 122 #endif 123 IntrinsicDescPair pair(m, is_virtual); 124 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 125 } 126 127 void Compile::register_intrinsic(CallGenerator* cg) { 128 if (_intrinsics == NULL) { 129 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 130 } 131 int len = _intrinsics->length(); 132 bool found = false; 133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 134 assert(!found, "registering twice"); 135 _intrinsics->insert_before(index, cg); 136 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 137 } 138 139 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 140 assert(m->is_loaded(), "don't try this on unloaded methods"); 141 if (_intrinsics != NULL) { 142 bool found = false; 143 int index = intrinsic_insertion_index(m, is_virtual, found); 144 if (found) { 145 return _intrinsics->at(index); 146 } 147 } 148 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 149 if (m->intrinsic_id() != vmIntrinsics::_none && 150 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 151 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 152 if (cg != NULL) { 153 // Save it for next time: 154 register_intrinsic(cg); 155 return cg; 156 } else { 157 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 158 } 159 } 160 return NULL; 161 } 162 163 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 164 // in library_call.cpp. 165 166 167 #ifndef PRODUCT 168 // statistics gathering... 169 170 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 171 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 172 173 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 174 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 175 int oflags = _intrinsic_hist_flags[id]; 176 assert(flags != 0, "what happened?"); 177 if (is_virtual) { 178 flags |= _intrinsic_virtual; 179 } 180 bool changed = (flags != oflags); 181 if ((flags & _intrinsic_worked) != 0) { 182 juint count = (_intrinsic_hist_count[id] += 1); 183 if (count == 1) { 184 changed = true; // first time 185 } 186 // increment the overall count also: 187 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 188 } 189 if (changed) { 190 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 191 // Something changed about the intrinsic's virtuality. 192 if ((flags & _intrinsic_virtual) != 0) { 193 // This is the first use of this intrinsic as a virtual call. 194 if (oflags != 0) { 195 // We already saw it as a non-virtual, so note both cases. 196 flags |= _intrinsic_both; 197 } 198 } else if ((oflags & _intrinsic_both) == 0) { 199 // This is the first use of this intrinsic as a non-virtual 200 flags |= _intrinsic_both; 201 } 202 } 203 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 204 } 205 // update the overall flags also: 206 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 207 return changed; 208 } 209 210 static char* format_flags(int flags, char* buf) { 211 buf[0] = 0; 212 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 213 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 214 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 215 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 216 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 217 if (buf[0] == 0) strcat(buf, ","); 218 assert(buf[0] == ',', "must be"); 219 return &buf[1]; 220 } 221 222 void Compile::print_intrinsic_statistics() { 223 char flagsbuf[100]; 224 ttyLocker ttyl; 225 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 226 tty->print_cr("Compiler intrinsic usage:"); 227 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 228 if (total == 0) total = 1; // avoid div0 in case of no successes 229 #define PRINT_STAT_LINE(name, c, f) \ 230 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 231 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 232 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 233 int flags = _intrinsic_hist_flags[id]; 234 juint count = _intrinsic_hist_count[id]; 235 if ((flags | count) != 0) { 236 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 237 } 238 } 239 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 240 if (xtty != NULL) xtty->tail("statistics"); 241 } 242 243 void Compile::print_statistics() { 244 { ttyLocker ttyl; 245 if (xtty != NULL) xtty->head("statistics type='opto'"); 246 Parse::print_statistics(); 247 PhaseCCP::print_statistics(); 248 PhaseRegAlloc::print_statistics(); 249 Scheduling::print_statistics(); 250 PhasePeephole::print_statistics(); 251 PhaseIdealLoop::print_statistics(); 252 if (xtty != NULL) xtty->tail("statistics"); 253 } 254 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 255 // put this under its own <statistics> element. 256 print_intrinsic_statistics(); 257 } 258 } 259 #endif //PRODUCT 260 261 // Support for bundling info 262 Bundle* Compile::node_bundling(const Node *n) { 263 assert(valid_bundle_info(n), "oob"); 264 return &_node_bundling_base[n->_idx]; 265 } 266 267 bool Compile::valid_bundle_info(const Node *n) { 268 return (_node_bundling_limit > n->_idx); 269 } 270 271 272 void Compile::gvn_replace_by(Node* n, Node* nn) { 273 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 274 Node* use = n->last_out(i); 275 bool is_in_table = initial_gvn()->hash_delete(use); 276 uint uses_found = 0; 277 for (uint j = 0; j < use->len(); j++) { 278 if (use->in(j) == n) { 279 if (j < use->req()) 280 use->set_req(j, nn); 281 else 282 use->set_prec(j, nn); 283 uses_found++; 284 } 285 } 286 if (is_in_table) { 287 // reinsert into table 288 initial_gvn()->hash_find_insert(use); 289 } 290 record_for_igvn(use); 291 i -= uses_found; // we deleted 1 or more copies of this edge 292 } 293 } 294 295 296 static inline bool not_a_node(const Node* n) { 297 if (n == NULL) return true; 298 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 299 if (*(address*)n == badAddress) return true; // kill by Node::destruct 300 return false; 301 } 302 303 // Identify all nodes that are reachable from below, useful. 304 // Use breadth-first pass that records state in a Unique_Node_List, 305 // recursive traversal is slower. 306 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 307 int estimated_worklist_size = live_nodes(); 308 useful.map( estimated_worklist_size, NULL ); // preallocate space 309 310 // Initialize worklist 311 if (root() != NULL) { useful.push(root()); } 312 // If 'top' is cached, declare it useful to preserve cached node 313 if( cached_top_node() ) { useful.push(cached_top_node()); } 314 315 // Push all useful nodes onto the list, breadthfirst 316 for( uint next = 0; next < useful.size(); ++next ) { 317 assert( next < unique(), "Unique useful nodes < total nodes"); 318 Node *n = useful.at(next); 319 uint max = n->len(); 320 for( uint i = 0; i < max; ++i ) { 321 Node *m = n->in(i); 322 if (not_a_node(m)) continue; 323 useful.push(m); 324 } 325 } 326 } 327 328 // Update dead_node_list with any missing dead nodes using useful 329 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 330 void Compile::update_dead_node_list(Unique_Node_List &useful) { 331 uint max_idx = unique(); 332 VectorSet& useful_node_set = useful.member_set(); 333 334 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 335 // If node with index node_idx is not in useful set, 336 // mark it as dead in dead node list. 337 if (! useful_node_set.test(node_idx) ) { 338 record_dead_node(node_idx); 339 } 340 } 341 } 342 343 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 344 int shift = 0; 345 for (int i = 0; i < inlines->length(); i++) { 346 CallGenerator* cg = inlines->at(i); 347 CallNode* call = cg->call_node(); 348 if (shift > 0) { 349 inlines->at_put(i-shift, cg); 350 } 351 if (!useful.member(call)) { 352 shift++; 353 } 354 } 355 inlines->trunc_to(inlines->length()-shift); 356 } 357 358 // Disconnect all useless nodes by disconnecting those at the boundary. 359 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 360 uint next = 0; 361 while (next < useful.size()) { 362 Node *n = useful.at(next++); 363 if (n->is_SafePoint()) { 364 // We're done with a parsing phase. Replaced nodes are not valid 365 // beyond that point. 366 n->as_SafePoint()->delete_replaced_nodes(); 367 } 368 // Use raw traversal of out edges since this code removes out edges 369 int max = n->outcnt(); 370 for (int j = 0; j < max; ++j) { 371 Node* child = n->raw_out(j); 372 if (! useful.member(child)) { 373 assert(!child->is_top() || child != top(), 374 "If top is cached in Compile object it is in useful list"); 375 // Only need to remove this out-edge to the useless node 376 n->raw_del_out(j); 377 --j; 378 --max; 379 } 380 } 381 if (n->outcnt() == 1 && n->has_special_unique_user()) { 382 record_for_igvn(n->unique_out()); 383 } 384 } 385 // Remove useless macro and predicate opaq nodes 386 for (int i = C->macro_count()-1; i >= 0; i--) { 387 Node* n = C->macro_node(i); 388 if (!useful.member(n)) { 389 remove_macro_node(n); 390 } 391 } 392 // Remove useless CastII nodes with range check dependency 393 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 394 Node* cast = range_check_cast_node(i); 395 if (!useful.member(cast)) { 396 remove_range_check_cast(cast); 397 } 398 } 399 // Remove useless expensive node 400 for (int i = C->expensive_count()-1; i >= 0; i--) { 401 Node* n = C->expensive_node(i); 402 if (!useful.member(n)) { 403 remove_expensive_node(n); 404 } 405 } 406 // clean up the late inline lists 407 remove_useless_late_inlines(&_string_late_inlines, useful); 408 remove_useless_late_inlines(&_boxing_late_inlines, useful); 409 remove_useless_late_inlines(&_late_inlines, useful); 410 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 411 } 412 413 //------------------------------frame_size_in_words----------------------------- 414 // frame_slots in units of words 415 int Compile::frame_size_in_words() const { 416 // shift is 0 in LP32 and 1 in LP64 417 const int shift = (LogBytesPerWord - LogBytesPerInt); 418 int words = _frame_slots >> shift; 419 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 420 return words; 421 } 422 423 // To bang the stack of this compiled method we use the stack size 424 // that the interpreter would need in case of a deoptimization. This 425 // removes the need to bang the stack in the deoptimization blob which 426 // in turn simplifies stack overflow handling. 427 int Compile::bang_size_in_bytes() const { 428 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 429 } 430 431 // ============================================================================ 432 //------------------------------CompileWrapper--------------------------------- 433 class CompileWrapper : public StackObj { 434 Compile *const _compile; 435 public: 436 CompileWrapper(Compile* compile); 437 438 ~CompileWrapper(); 439 }; 440 441 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 442 // the Compile* pointer is stored in the current ciEnv: 443 ciEnv* env = compile->env(); 444 assert(env == ciEnv::current(), "must already be a ciEnv active"); 445 assert(env->compiler_data() == NULL, "compile already active?"); 446 env->set_compiler_data(compile); 447 assert(compile == Compile::current(), "sanity"); 448 449 compile->set_type_dict(NULL); 450 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 451 compile->clone_map().set_clone_idx(0); 452 compile->set_type_hwm(NULL); 453 compile->set_type_last_size(0); 454 compile->set_last_tf(NULL, NULL); 455 compile->set_indexSet_arena(NULL); 456 compile->set_indexSet_free_block_list(NULL); 457 compile->init_type_arena(); 458 Type::Initialize(compile); 459 _compile->set_scratch_buffer_blob(NULL); 460 _compile->begin_method(); 461 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 462 } 463 CompileWrapper::~CompileWrapper() { 464 _compile->end_method(); 465 if (_compile->scratch_buffer_blob() != NULL) 466 BufferBlob::free(_compile->scratch_buffer_blob()); 467 _compile->env()->set_compiler_data(NULL); 468 } 469 470 471 //----------------------------print_compile_messages--------------------------- 472 void Compile::print_compile_messages() { 473 #ifndef PRODUCT 474 // Check if recompiling 475 if (_subsume_loads == false && PrintOpto) { 476 // Recompiling without allowing machine instructions to subsume loads 477 tty->print_cr("*********************************************************"); 478 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 479 tty->print_cr("*********************************************************"); 480 } 481 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 482 // Recompiling without escape analysis 483 tty->print_cr("*********************************************************"); 484 tty->print_cr("** Bailout: Recompile without escape analysis **"); 485 tty->print_cr("*********************************************************"); 486 } 487 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 488 // Recompiling without boxing elimination 489 tty->print_cr("*********************************************************"); 490 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 491 tty->print_cr("*********************************************************"); 492 } 493 if (C->directive()->BreakAtCompileOption) { 494 // Open the debugger when compiling this method. 495 tty->print("### Breaking when compiling: "); 496 method()->print_short_name(); 497 tty->cr(); 498 BREAKPOINT; 499 } 500 501 if( PrintOpto ) { 502 if (is_osr_compilation()) { 503 tty->print("[OSR]%3d", _compile_id); 504 } else { 505 tty->print("%3d", _compile_id); 506 } 507 } 508 #endif 509 } 510 511 512 //-----------------------init_scratch_buffer_blob------------------------------ 513 // Construct a temporary BufferBlob and cache it for this compile. 514 void Compile::init_scratch_buffer_blob(int const_size) { 515 // If there is already a scratch buffer blob allocated and the 516 // constant section is big enough, use it. Otherwise free the 517 // current and allocate a new one. 518 BufferBlob* blob = scratch_buffer_blob(); 519 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 520 // Use the current blob. 521 } else { 522 if (blob != NULL) { 523 BufferBlob::free(blob); 524 } 525 526 ResourceMark rm; 527 _scratch_const_size = const_size; 528 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 529 blob = BufferBlob::create("Compile::scratch_buffer", size); 530 // Record the buffer blob for next time. 531 set_scratch_buffer_blob(blob); 532 // Have we run out of code space? 533 if (scratch_buffer_blob() == NULL) { 534 // Let CompilerBroker disable further compilations. 535 record_failure("Not enough space for scratch buffer in CodeCache"); 536 return; 537 } 538 } 539 540 // Initialize the relocation buffers 541 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 542 set_scratch_locs_memory(locs_buf); 543 } 544 545 546 //-----------------------scratch_emit_size------------------------------------- 547 // Helper function that computes size by emitting code 548 uint Compile::scratch_emit_size(const Node* n) { 549 // Start scratch_emit_size section. 550 set_in_scratch_emit_size(true); 551 552 // Emit into a trash buffer and count bytes emitted. 553 // This is a pretty expensive way to compute a size, 554 // but it works well enough if seldom used. 555 // All common fixed-size instructions are given a size 556 // method by the AD file. 557 // Note that the scratch buffer blob and locs memory are 558 // allocated at the beginning of the compile task, and 559 // may be shared by several calls to scratch_emit_size. 560 // The allocation of the scratch buffer blob is particularly 561 // expensive, since it has to grab the code cache lock. 562 BufferBlob* blob = this->scratch_buffer_blob(); 563 assert(blob != NULL, "Initialize BufferBlob at start"); 564 assert(blob->size() > MAX_inst_size, "sanity"); 565 relocInfo* locs_buf = scratch_locs_memory(); 566 address blob_begin = blob->content_begin(); 567 address blob_end = (address)locs_buf; 568 assert(blob->contains(blob_end), "sanity"); 569 CodeBuffer buf(blob_begin, blob_end - blob_begin); 570 buf.initialize_consts_size(_scratch_const_size); 571 buf.initialize_stubs_size(MAX_stubs_size); 572 assert(locs_buf != NULL, "sanity"); 573 int lsize = MAX_locs_size / 3; 574 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 575 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 576 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 577 // Mark as scratch buffer. 578 buf.consts()->set_scratch_emit(); 579 buf.insts()->set_scratch_emit(); 580 buf.stubs()->set_scratch_emit(); 581 582 // Do the emission. 583 584 Label fakeL; // Fake label for branch instructions. 585 Label* saveL = NULL; 586 uint save_bnum = 0; 587 bool is_branch = n->is_MachBranch(); 588 if (is_branch) { 589 MacroAssembler masm(&buf); 590 masm.bind(fakeL); 591 n->as_MachBranch()->save_label(&saveL, &save_bnum); 592 n->as_MachBranch()->label_set(&fakeL, 0); 593 } 594 n->emit(buf, this->regalloc()); 595 596 // Emitting into the scratch buffer should not fail 597 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 598 599 if (is_branch) // Restore label. 600 n->as_MachBranch()->label_set(saveL, save_bnum); 601 602 // End scratch_emit_size section. 603 set_in_scratch_emit_size(false); 604 605 return buf.insts_size(); 606 } 607 608 609 // ============================================================================ 610 //------------------------------Compile standard------------------------------- 611 debug_only( int Compile::_debug_idx = 100000; ) 612 613 // Compile a method. entry_bci is -1 for normal compilations and indicates 614 // the continuation bci for on stack replacement. 615 616 617 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 618 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 619 : Phase(Compiler), 620 _env(ci_env), 621 _directive(directive), 622 _log(ci_env->log()), 623 _compile_id(ci_env->compile_id()), 624 _save_argument_registers(false), 625 _stub_name(NULL), 626 _stub_function(NULL), 627 _stub_entry_point(NULL), 628 _method(target), 629 _entry_bci(osr_bci), 630 _initial_gvn(NULL), 631 _for_igvn(NULL), 632 _warm_calls(NULL), 633 _subsume_loads(subsume_loads), 634 _do_escape_analysis(do_escape_analysis), 635 _eliminate_boxing(eliminate_boxing), 636 _failure_reason(NULL), 637 _code_buffer("Compile::Fill_buffer"), 638 _orig_pc_slot(0), 639 _orig_pc_slot_offset_in_bytes(0), 640 _has_method_handle_invokes(false), 641 _mach_constant_base_node(NULL), 642 _node_bundling_limit(0), 643 _node_bundling_base(NULL), 644 _java_calls(0), 645 _inner_loops(0), 646 _scratch_const_size(-1), 647 _in_scratch_emit_size(false), 648 _dead_node_list(comp_arena()), 649 _dead_node_count(0), 650 #ifndef PRODUCT 651 _trace_opto_output(directive->TraceOptoOutputOption), 652 _in_dump_cnt(0), 653 _printer(IdealGraphPrinter::printer()), 654 #endif 655 _congraph(NULL), 656 _comp_arena(mtCompiler), 657 _node_arena(mtCompiler), 658 _old_arena(mtCompiler), 659 _Compile_types(mtCompiler), 660 _replay_inline_data(NULL), 661 _late_inlines(comp_arena(), 2, 0, NULL), 662 _string_late_inlines(comp_arena(), 2, 0, NULL), 663 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 664 _late_inlines_pos(0), 665 _number_of_mh_late_inlines(0), 666 _inlining_progress(false), 667 _inlining_incrementally(false), 668 _print_inlining_list(NULL), 669 _print_inlining_stream(NULL), 670 _print_inlining_idx(0), 671 _print_inlining_output(NULL), 672 _interpreter_frame_size(0), 673 _max_node_limit(MaxNodeLimit), 674 _has_reserved_stack_access(target->has_reserved_stack_access()) { 675 C = this; 676 #ifndef PRODUCT 677 if (_printer != NULL) { 678 _printer->set_compile(this); 679 } 680 #endif 681 CompileWrapper cw(this); 682 683 if (CITimeVerbose) { 684 tty->print(" "); 685 target->holder()->name()->print(); 686 tty->print("."); 687 target->print_short_name(); 688 tty->print(" "); 689 } 690 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 691 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 692 693 #ifndef PRODUCT 694 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 695 if (!print_opto_assembly) { 696 bool print_assembly = directive->PrintAssemblyOption; 697 if (print_assembly && !Disassembler::can_decode()) { 698 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 699 print_opto_assembly = true; 700 } 701 } 702 set_print_assembly(print_opto_assembly); 703 set_parsed_irreducible_loop(false); 704 705 if (directive->ReplayInlineOption) { 706 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 707 } 708 #endif 709 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 710 set_print_intrinsics(directive->PrintIntrinsicsOption); 711 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 712 713 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 714 // Make sure the method being compiled gets its own MDO, 715 // so we can at least track the decompile_count(). 716 // Need MDO to record RTM code generation state. 717 method()->ensure_method_data(); 718 } 719 720 Init(::AliasLevel); 721 722 723 print_compile_messages(); 724 725 _ilt = InlineTree::build_inline_tree_root(); 726 727 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 728 assert(num_alias_types() >= AliasIdxRaw, ""); 729 730 #define MINIMUM_NODE_HASH 1023 731 // Node list that Iterative GVN will start with 732 Unique_Node_List for_igvn(comp_arena()); 733 set_for_igvn(&for_igvn); 734 735 // GVN that will be run immediately on new nodes 736 uint estimated_size = method()->code_size()*4+64; 737 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 738 PhaseGVN gvn(node_arena(), estimated_size); 739 set_initial_gvn(&gvn); 740 741 print_inlining_init(); 742 { // Scope for timing the parser 743 TracePhase tp("parse", &timers[_t_parser]); 744 745 // Put top into the hash table ASAP. 746 initial_gvn()->transform_no_reclaim(top()); 747 748 // Set up tf(), start(), and find a CallGenerator. 749 CallGenerator* cg = NULL; 750 if (is_osr_compilation()) { 751 const TypeTuple *domain = StartOSRNode::osr_domain(); 752 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 753 init_tf(TypeFunc::make(domain, range)); 754 StartNode* s = new StartOSRNode(root(), domain); 755 initial_gvn()->set_type_bottom(s); 756 init_start(s); 757 cg = CallGenerator::for_osr(method(), entry_bci()); 758 } else { 759 // Normal case. 760 init_tf(TypeFunc::make(method())); 761 StartNode* s = new StartNode(root(), tf()->domain()); 762 initial_gvn()->set_type_bottom(s); 763 init_start(s); 764 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) { 765 // With java.lang.ref.reference.get() we must go through the 766 // intrinsic - even when get() is the root 767 // method of the compile - so that, if necessary, the value in 768 // the referent field of the reference object gets recorded by 769 // the pre-barrier code. 770 cg = find_intrinsic(method(), false); 771 } 772 if (cg == NULL) { 773 float past_uses = method()->interpreter_invocation_count(); 774 float expected_uses = past_uses; 775 cg = CallGenerator::for_inline(method(), expected_uses); 776 } 777 } 778 if (failing()) return; 779 if (cg == NULL) { 780 record_method_not_compilable("cannot parse method"); 781 return; 782 } 783 JVMState* jvms = build_start_state(start(), tf()); 784 if ((jvms = cg->generate(jvms)) == NULL) { 785 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 786 record_method_not_compilable("method parse failed"); 787 } 788 return; 789 } 790 GraphKit kit(jvms); 791 792 if (!kit.stopped()) { 793 // Accept return values, and transfer control we know not where. 794 // This is done by a special, unique ReturnNode bound to root. 795 return_values(kit.jvms()); 796 } 797 798 if (kit.has_exceptions()) { 799 // Any exceptions that escape from this call must be rethrown 800 // to whatever caller is dynamically above us on the stack. 801 // This is done by a special, unique RethrowNode bound to root. 802 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 803 } 804 805 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 806 807 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 808 inline_string_calls(true); 809 } 810 811 if (failing()) return; 812 813 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 814 815 // Remove clutter produced by parsing. 816 if (!failing()) { 817 ResourceMark rm; 818 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 819 } 820 } 821 822 // Note: Large methods are capped off in do_one_bytecode(). 823 if (failing()) return; 824 825 // After parsing, node notes are no longer automagic. 826 // They must be propagated by register_new_node_with_optimizer(), 827 // clone(), or the like. 828 set_default_node_notes(NULL); 829 830 for (;;) { 831 int successes = Inline_Warm(); 832 if (failing()) return; 833 if (successes == 0) break; 834 } 835 836 // Drain the list. 837 Finish_Warm(); 838 #ifndef PRODUCT 839 if (_printer && _printer->should_print(1)) { 840 _printer->print_inlining(); 841 } 842 #endif 843 844 if (failing()) return; 845 NOT_PRODUCT( verify_graph_edges(); ) 846 847 // Now optimize 848 Optimize(); 849 if (failing()) return; 850 NOT_PRODUCT( verify_graph_edges(); ) 851 852 #ifndef PRODUCT 853 if (PrintIdeal) { 854 ttyLocker ttyl; // keep the following output all in one block 855 // This output goes directly to the tty, not the compiler log. 856 // To enable tools to match it up with the compilation activity, 857 // be sure to tag this tty output with the compile ID. 858 if (xtty != NULL) { 859 xtty->head("ideal compile_id='%d'%s", compile_id(), 860 is_osr_compilation() ? " compile_kind='osr'" : 861 ""); 862 } 863 root()->dump(9999); 864 if (xtty != NULL) { 865 xtty->tail("ideal"); 866 } 867 } 868 #endif 869 870 NOT_PRODUCT( verify_barriers(); ) 871 872 // Dump compilation data to replay it. 873 if (directive->DumpReplayOption) { 874 env()->dump_replay_data(_compile_id); 875 } 876 if (directive->DumpInlineOption && (ilt() != NULL)) { 877 env()->dump_inline_data(_compile_id); 878 } 879 880 // Now that we know the size of all the monitors we can add a fixed slot 881 // for the original deopt pc. 882 883 _orig_pc_slot = fixed_slots(); 884 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 885 set_fixed_slots(next_slot); 886 887 // Compute when to use implicit null checks. Used by matching trap based 888 // nodes and NullCheck optimization. 889 set_allowed_deopt_reasons(); 890 891 // Now generate code 892 Code_Gen(); 893 if (failing()) return; 894 895 // Check if we want to skip execution of all compiled code. 896 { 897 #ifndef PRODUCT 898 if (OptoNoExecute) { 899 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 900 return; 901 } 902 #endif 903 TracePhase tp("install_code", &timers[_t_registerMethod]); 904 905 if (is_osr_compilation()) { 906 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 907 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 908 } else { 909 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 910 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 911 } 912 913 env()->register_method(_method, _entry_bci, 914 &_code_offsets, 915 _orig_pc_slot_offset_in_bytes, 916 code_buffer(), 917 frame_size_in_words(), _oop_map_set, 918 &_handler_table, &_inc_table, 919 compiler, 920 has_unsafe_access(), 921 SharedRuntime::is_wide_vector(max_vector_size()), 922 rtm_state() 923 ); 924 925 if (log() != NULL) // Print code cache state into compiler log 926 log()->code_cache_state(); 927 } 928 } 929 930 //------------------------------Compile---------------------------------------- 931 // Compile a runtime stub 932 Compile::Compile( ciEnv* ci_env, 933 TypeFunc_generator generator, 934 address stub_function, 935 const char *stub_name, 936 int is_fancy_jump, 937 bool pass_tls, 938 bool save_arg_registers, 939 bool return_pc, 940 DirectiveSet* directive) 941 : Phase(Compiler), 942 _env(ci_env), 943 _directive(directive), 944 _log(ci_env->log()), 945 _compile_id(0), 946 _save_argument_registers(save_arg_registers), 947 _method(NULL), 948 _stub_name(stub_name), 949 _stub_function(stub_function), 950 _stub_entry_point(NULL), 951 _entry_bci(InvocationEntryBci), 952 _initial_gvn(NULL), 953 _for_igvn(NULL), 954 _warm_calls(NULL), 955 _orig_pc_slot(0), 956 _orig_pc_slot_offset_in_bytes(0), 957 _subsume_loads(true), 958 _do_escape_analysis(false), 959 _eliminate_boxing(false), 960 _failure_reason(NULL), 961 _code_buffer("Compile::Fill_buffer"), 962 _has_method_handle_invokes(false), 963 _mach_constant_base_node(NULL), 964 _node_bundling_limit(0), 965 _node_bundling_base(NULL), 966 _java_calls(0), 967 _inner_loops(0), 968 #ifndef PRODUCT 969 _trace_opto_output(directive->TraceOptoOutputOption), 970 _in_dump_cnt(0), 971 _printer(NULL), 972 #endif 973 _comp_arena(mtCompiler), 974 _node_arena(mtCompiler), 975 _old_arena(mtCompiler), 976 _Compile_types(mtCompiler), 977 _dead_node_list(comp_arena()), 978 _dead_node_count(0), 979 _congraph(NULL), 980 _replay_inline_data(NULL), 981 _number_of_mh_late_inlines(0), 982 _inlining_progress(false), 983 _inlining_incrementally(false), 984 _print_inlining_list(NULL), 985 _print_inlining_stream(NULL), 986 _print_inlining_idx(0), 987 _print_inlining_output(NULL), 988 _allowed_reasons(0), 989 _interpreter_frame_size(0), 990 _max_node_limit(MaxNodeLimit) { 991 C = this; 992 993 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 994 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 995 996 #ifndef PRODUCT 997 set_print_assembly(PrintFrameConverterAssembly); 998 set_parsed_irreducible_loop(false); 999 #endif 1000 set_has_irreducible_loop(false); // no loops 1001 1002 CompileWrapper cw(this); 1003 Init(/*AliasLevel=*/ 0); 1004 init_tf((*generator)()); 1005 1006 { 1007 // The following is a dummy for the sake of GraphKit::gen_stub 1008 Unique_Node_List for_igvn(comp_arena()); 1009 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1010 PhaseGVN gvn(Thread::current()->resource_area(),255); 1011 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1012 gvn.transform_no_reclaim(top()); 1013 1014 GraphKit kit; 1015 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1016 } 1017 1018 NOT_PRODUCT( verify_graph_edges(); ) 1019 Code_Gen(); 1020 if (failing()) return; 1021 1022 1023 // Entry point will be accessed using compile->stub_entry_point(); 1024 if (code_buffer() == NULL) { 1025 Matcher::soft_match_failure(); 1026 } else { 1027 if (PrintAssembly && (WizardMode || Verbose)) 1028 tty->print_cr("### Stub::%s", stub_name); 1029 1030 if (!failing()) { 1031 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1032 1033 // Make the NMethod 1034 // For now we mark the frame as never safe for profile stackwalking 1035 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1036 code_buffer(), 1037 CodeOffsets::frame_never_safe, 1038 // _code_offsets.value(CodeOffsets::Frame_Complete), 1039 frame_size_in_words(), 1040 _oop_map_set, 1041 save_arg_registers); 1042 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1043 1044 _stub_entry_point = rs->entry_point(); 1045 } 1046 } 1047 } 1048 1049 //------------------------------Init------------------------------------------- 1050 // Prepare for a single compilation 1051 void Compile::Init(int aliaslevel) { 1052 _unique = 0; 1053 _regalloc = NULL; 1054 1055 _tf = NULL; // filled in later 1056 _top = NULL; // cached later 1057 _matcher = NULL; // filled in later 1058 _cfg = NULL; // filled in later 1059 1060 set_24_bit_selection_and_mode(Use24BitFP, false); 1061 1062 _node_note_array = NULL; 1063 _default_node_notes = NULL; 1064 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1065 1066 _immutable_memory = NULL; // filled in at first inquiry 1067 1068 // Globally visible Nodes 1069 // First set TOP to NULL to give safe behavior during creation of RootNode 1070 set_cached_top_node(NULL); 1071 set_root(new RootNode()); 1072 // Now that you have a Root to point to, create the real TOP 1073 set_cached_top_node( new ConNode(Type::TOP) ); 1074 set_recent_alloc(NULL, NULL); 1075 1076 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1077 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1078 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1079 env()->set_dependencies(new Dependencies(env())); 1080 1081 _fixed_slots = 0; 1082 set_has_split_ifs(false); 1083 set_has_loops(has_method() && method()->has_loops()); // first approximation 1084 set_has_stringbuilder(false); 1085 set_has_boxed_value(false); 1086 _trap_can_recompile = false; // no traps emitted yet 1087 _major_progress = true; // start out assuming good things will happen 1088 set_has_unsafe_access(false); 1089 set_max_vector_size(0); 1090 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1091 set_decompile_count(0); 1092 1093 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1094 set_num_loop_opts(LoopOptsCount); 1095 set_do_inlining(Inline); 1096 set_max_inline_size(MaxInlineSize); 1097 set_freq_inline_size(FreqInlineSize); 1098 set_do_scheduling(OptoScheduling); 1099 set_do_count_invocations(false); 1100 set_do_method_data_update(false); 1101 1102 set_do_vector_loop(false); 1103 1104 if (AllowVectorizeOnDemand) { 1105 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1106 set_do_vector_loop(true); 1107 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());}) 1108 } else if (has_method() && method()->name() != 0 && 1109 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1110 set_do_vector_loop(true); 1111 } 1112 } 1113 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1114 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());}) 1115 1116 set_age_code(has_method() && method()->profile_aging()); 1117 set_rtm_state(NoRTM); // No RTM lock eliding by default 1118 _max_node_limit = _directive->MaxNodeLimitOption; 1119 1120 #if INCLUDE_RTM_OPT 1121 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1122 int rtm_state = method()->method_data()->rtm_state(); 1123 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1124 // Don't generate RTM lock eliding code. 1125 set_rtm_state(NoRTM); 1126 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1127 // Generate RTM lock eliding code without abort ratio calculation code. 1128 set_rtm_state(UseRTM); 1129 } else if (UseRTMDeopt) { 1130 // Generate RTM lock eliding code and include abort ratio calculation 1131 // code if UseRTMDeopt is on. 1132 set_rtm_state(ProfileRTM); 1133 } 1134 } 1135 #endif 1136 if (debug_info()->recording_non_safepoints()) { 1137 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1138 (comp_arena(), 8, 0, NULL)); 1139 set_default_node_notes(Node_Notes::make(this)); 1140 } 1141 1142 // // -- Initialize types before each compile -- 1143 // // Update cached type information 1144 // if( _method && _method->constants() ) 1145 // Type::update_loaded_types(_method, _method->constants()); 1146 1147 // Init alias_type map. 1148 if (!_do_escape_analysis && aliaslevel == 3) 1149 aliaslevel = 2; // No unique types without escape analysis 1150 _AliasLevel = aliaslevel; 1151 const int grow_ats = 16; 1152 _max_alias_types = grow_ats; 1153 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1154 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1155 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1156 { 1157 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1158 } 1159 // Initialize the first few types. 1160 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1161 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1162 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1163 _num_alias_types = AliasIdxRaw+1; 1164 // Zero out the alias type cache. 1165 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1166 // A NULL adr_type hits in the cache right away. Preload the right answer. 1167 probe_alias_cache(NULL)->_index = AliasIdxTop; 1168 1169 _intrinsics = NULL; 1170 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1171 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1172 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1173 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1174 register_library_intrinsics(); 1175 } 1176 1177 //---------------------------init_start---------------------------------------- 1178 // Install the StartNode on this compile object. 1179 void Compile::init_start(StartNode* s) { 1180 if (failing()) 1181 return; // already failing 1182 assert(s == start(), ""); 1183 } 1184 1185 /** 1186 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1187 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1188 * the ideal graph. 1189 */ 1190 StartNode* Compile::start() const { 1191 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1192 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1193 Node* start = root()->fast_out(i); 1194 if (start->is_Start()) { 1195 return start->as_Start(); 1196 } 1197 } 1198 fatal("Did not find Start node!"); 1199 return NULL; 1200 } 1201 1202 //-------------------------------immutable_memory------------------------------------- 1203 // Access immutable memory 1204 Node* Compile::immutable_memory() { 1205 if (_immutable_memory != NULL) { 1206 return _immutable_memory; 1207 } 1208 StartNode* s = start(); 1209 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1210 Node *p = s->fast_out(i); 1211 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1212 _immutable_memory = p; 1213 return _immutable_memory; 1214 } 1215 } 1216 ShouldNotReachHere(); 1217 return NULL; 1218 } 1219 1220 //----------------------set_cached_top_node------------------------------------ 1221 // Install the cached top node, and make sure Node::is_top works correctly. 1222 void Compile::set_cached_top_node(Node* tn) { 1223 if (tn != NULL) verify_top(tn); 1224 Node* old_top = _top; 1225 _top = tn; 1226 // Calling Node::setup_is_top allows the nodes the chance to adjust 1227 // their _out arrays. 1228 if (_top != NULL) _top->setup_is_top(); 1229 if (old_top != NULL) old_top->setup_is_top(); 1230 assert(_top == NULL || top()->is_top(), ""); 1231 } 1232 1233 #ifdef ASSERT 1234 uint Compile::count_live_nodes_by_graph_walk() { 1235 Unique_Node_List useful(comp_arena()); 1236 // Get useful node list by walking the graph. 1237 identify_useful_nodes(useful); 1238 return useful.size(); 1239 } 1240 1241 void Compile::print_missing_nodes() { 1242 1243 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1244 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1245 return; 1246 } 1247 1248 // This is an expensive function. It is executed only when the user 1249 // specifies VerifyIdealNodeCount option or otherwise knows the 1250 // additional work that needs to be done to identify reachable nodes 1251 // by walking the flow graph and find the missing ones using 1252 // _dead_node_list. 1253 1254 Unique_Node_List useful(comp_arena()); 1255 // Get useful node list by walking the graph. 1256 identify_useful_nodes(useful); 1257 1258 uint l_nodes = C->live_nodes(); 1259 uint l_nodes_by_walk = useful.size(); 1260 1261 if (l_nodes != l_nodes_by_walk) { 1262 if (_log != NULL) { 1263 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1264 _log->stamp(); 1265 _log->end_head(); 1266 } 1267 VectorSet& useful_member_set = useful.member_set(); 1268 int last_idx = l_nodes_by_walk; 1269 for (int i = 0; i < last_idx; i++) { 1270 if (useful_member_set.test(i)) { 1271 if (_dead_node_list.test(i)) { 1272 if (_log != NULL) { 1273 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1274 } 1275 if (PrintIdealNodeCount) { 1276 // Print the log message to tty 1277 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1278 useful.at(i)->dump(); 1279 } 1280 } 1281 } 1282 else if (! _dead_node_list.test(i)) { 1283 if (_log != NULL) { 1284 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1285 } 1286 if (PrintIdealNodeCount) { 1287 // Print the log message to tty 1288 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1289 } 1290 } 1291 } 1292 if (_log != NULL) { 1293 _log->tail("mismatched_nodes"); 1294 } 1295 } 1296 } 1297 void Compile::record_modified_node(Node* n) { 1298 if (_modified_nodes != NULL && !_inlining_incrementally && 1299 n->outcnt() != 0 && !n->is_Con()) { 1300 _modified_nodes->push(n); 1301 } 1302 } 1303 1304 void Compile::remove_modified_node(Node* n) { 1305 if (_modified_nodes != NULL) { 1306 _modified_nodes->remove(n); 1307 } 1308 } 1309 #endif 1310 1311 #ifndef PRODUCT 1312 void Compile::verify_top(Node* tn) const { 1313 if (tn != NULL) { 1314 assert(tn->is_Con(), "top node must be a constant"); 1315 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1316 assert(tn->in(0) != NULL, "must have live top node"); 1317 } 1318 } 1319 #endif 1320 1321 1322 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1323 1324 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1325 guarantee(arr != NULL, ""); 1326 int num_blocks = arr->length(); 1327 if (grow_by < num_blocks) grow_by = num_blocks; 1328 int num_notes = grow_by * _node_notes_block_size; 1329 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1330 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1331 while (num_notes > 0) { 1332 arr->append(notes); 1333 notes += _node_notes_block_size; 1334 num_notes -= _node_notes_block_size; 1335 } 1336 assert(num_notes == 0, "exact multiple, please"); 1337 } 1338 1339 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1340 if (source == NULL || dest == NULL) return false; 1341 1342 if (dest->is_Con()) 1343 return false; // Do not push debug info onto constants. 1344 1345 #ifdef ASSERT 1346 // Leave a bread crumb trail pointing to the original node: 1347 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1348 dest->set_debug_orig(source); 1349 } 1350 #endif 1351 1352 if (node_note_array() == NULL) 1353 return false; // Not collecting any notes now. 1354 1355 // This is a copy onto a pre-existing node, which may already have notes. 1356 // If both nodes have notes, do not overwrite any pre-existing notes. 1357 Node_Notes* source_notes = node_notes_at(source->_idx); 1358 if (source_notes == NULL || source_notes->is_clear()) return false; 1359 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1360 if (dest_notes == NULL || dest_notes->is_clear()) { 1361 return set_node_notes_at(dest->_idx, source_notes); 1362 } 1363 1364 Node_Notes merged_notes = (*source_notes); 1365 // The order of operations here ensures that dest notes will win... 1366 merged_notes.update_from(dest_notes); 1367 return set_node_notes_at(dest->_idx, &merged_notes); 1368 } 1369 1370 1371 //--------------------------allow_range_check_smearing------------------------- 1372 // Gating condition for coalescing similar range checks. 1373 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1374 // single covering check that is at least as strong as any of them. 1375 // If the optimization succeeds, the simplified (strengthened) range check 1376 // will always succeed. If it fails, we will deopt, and then give up 1377 // on the optimization. 1378 bool Compile::allow_range_check_smearing() const { 1379 // If this method has already thrown a range-check, 1380 // assume it was because we already tried range smearing 1381 // and it failed. 1382 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1383 return !already_trapped; 1384 } 1385 1386 1387 //------------------------------flatten_alias_type----------------------------- 1388 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1389 int offset = tj->offset(); 1390 TypePtr::PTR ptr = tj->ptr(); 1391 1392 // Known instance (scalarizable allocation) alias only with itself. 1393 bool is_known_inst = tj->isa_oopptr() != NULL && 1394 tj->is_oopptr()->is_known_instance(); 1395 1396 // Process weird unsafe references. 1397 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1398 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1399 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1400 tj = TypeOopPtr::BOTTOM; 1401 ptr = tj->ptr(); 1402 offset = tj->offset(); 1403 } 1404 1405 // Array pointers need some flattening 1406 const TypeAryPtr *ta = tj->isa_aryptr(); 1407 if (ta && ta->is_stable()) { 1408 // Erase stability property for alias analysis. 1409 tj = ta = ta->cast_to_stable(false); 1410 } 1411 if( ta && is_known_inst ) { 1412 if ( offset != Type::OffsetBot && 1413 offset > arrayOopDesc::length_offset_in_bytes() ) { 1414 offset = Type::OffsetBot; // Flatten constant access into array body only 1415 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1416 } 1417 } else if( ta && _AliasLevel >= 2 ) { 1418 // For arrays indexed by constant indices, we flatten the alias 1419 // space to include all of the array body. Only the header, klass 1420 // and array length can be accessed un-aliased. 1421 if( offset != Type::OffsetBot ) { 1422 if( ta->const_oop() ) { // MethodData* or Method* 1423 offset = Type::OffsetBot; // Flatten constant access into array body 1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1425 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1426 // range is OK as-is. 1427 tj = ta = TypeAryPtr::RANGE; 1428 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1429 tj = TypeInstPtr::KLASS; // all klass loads look alike 1430 ta = TypeAryPtr::RANGE; // generic ignored junk 1431 ptr = TypePtr::BotPTR; 1432 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1433 tj = TypeInstPtr::MARK; 1434 ta = TypeAryPtr::RANGE; // generic ignored junk 1435 ptr = TypePtr::BotPTR; 1436 } else { // Random constant offset into array body 1437 offset = Type::OffsetBot; // Flatten constant access into array body 1438 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1439 } 1440 } 1441 // Arrays of fixed size alias with arrays of unknown size. 1442 if (ta->size() != TypeInt::POS) { 1443 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1444 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1445 } 1446 // Arrays of known objects become arrays of unknown objects. 1447 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1448 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1449 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1450 } 1451 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1452 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1453 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1454 } 1455 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1456 // cannot be distinguished by bytecode alone. 1457 if (ta->elem() == TypeInt::BOOL) { 1458 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1459 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1460 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1461 } 1462 // During the 2nd round of IterGVN, NotNull castings are removed. 1463 // Make sure the Bottom and NotNull variants alias the same. 1464 // Also, make sure exact and non-exact variants alias the same. 1465 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1466 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1467 } 1468 } 1469 1470 // Oop pointers need some flattening 1471 const TypeInstPtr *to = tj->isa_instptr(); 1472 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1473 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1474 if( ptr == TypePtr::Constant ) { 1475 if (to->klass() != ciEnv::current()->Class_klass() || 1476 offset < k->size_helper() * wordSize) { 1477 // No constant oop pointers (such as Strings); they alias with 1478 // unknown strings. 1479 assert(!is_known_inst, "not scalarizable allocation"); 1480 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1481 } 1482 } else if( is_known_inst ) { 1483 tj = to; // Keep NotNull and klass_is_exact for instance type 1484 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1485 // During the 2nd round of IterGVN, NotNull castings are removed. 1486 // Make sure the Bottom and NotNull variants alias the same. 1487 // Also, make sure exact and non-exact variants alias the same. 1488 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1489 } 1490 if (to->speculative() != NULL) { 1491 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1492 } 1493 // Canonicalize the holder of this field 1494 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1495 // First handle header references such as a LoadKlassNode, even if the 1496 // object's klass is unloaded at compile time (4965979). 1497 if (!is_known_inst) { // Do it only for non-instance types 1498 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1499 } 1500 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1501 // Static fields are in the space above the normal instance 1502 // fields in the java.lang.Class instance. 1503 if (to->klass() != ciEnv::current()->Class_klass()) { 1504 to = NULL; 1505 tj = TypeOopPtr::BOTTOM; 1506 offset = tj->offset(); 1507 } 1508 } else { 1509 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1510 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1511 if( is_known_inst ) { 1512 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1513 } else { 1514 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1515 } 1516 } 1517 } 1518 } 1519 1520 // Klass pointers to object array klasses need some flattening 1521 const TypeKlassPtr *tk = tj->isa_klassptr(); 1522 if( tk ) { 1523 // If we are referencing a field within a Klass, we need 1524 // to assume the worst case of an Object. Both exact and 1525 // inexact types must flatten to the same alias class so 1526 // use NotNull as the PTR. 1527 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1528 1529 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1530 TypeKlassPtr::OBJECT->klass(), 1531 offset); 1532 } 1533 1534 ciKlass* klass = tk->klass(); 1535 if( klass->is_obj_array_klass() ) { 1536 ciKlass* k = TypeAryPtr::OOPS->klass(); 1537 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1538 k = TypeInstPtr::BOTTOM->klass(); 1539 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1540 } 1541 1542 // Check for precise loads from the primary supertype array and force them 1543 // to the supertype cache alias index. Check for generic array loads from 1544 // the primary supertype array and also force them to the supertype cache 1545 // alias index. Since the same load can reach both, we need to merge 1546 // these 2 disparate memories into the same alias class. Since the 1547 // primary supertype array is read-only, there's no chance of confusion 1548 // where we bypass an array load and an array store. 1549 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1550 if (offset == Type::OffsetBot || 1551 (offset >= primary_supers_offset && 1552 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1553 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1554 offset = in_bytes(Klass::secondary_super_cache_offset()); 1555 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1556 } 1557 } 1558 1559 // Flatten all Raw pointers together. 1560 if (tj->base() == Type::RawPtr) 1561 tj = TypeRawPtr::BOTTOM; 1562 1563 if (tj->base() == Type::AnyPtr) 1564 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1565 1566 // Flatten all to bottom for now 1567 switch( _AliasLevel ) { 1568 case 0: 1569 tj = TypePtr::BOTTOM; 1570 break; 1571 case 1: // Flatten to: oop, static, field or array 1572 switch (tj->base()) { 1573 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1574 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1575 case Type::AryPtr: // do not distinguish arrays at all 1576 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1577 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1578 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1579 default: ShouldNotReachHere(); 1580 } 1581 break; 1582 case 2: // No collapsing at level 2; keep all splits 1583 case 3: // No collapsing at level 3; keep all splits 1584 break; 1585 default: 1586 Unimplemented(); 1587 } 1588 1589 offset = tj->offset(); 1590 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1591 1592 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1593 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1594 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1595 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1596 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1597 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1598 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1599 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1600 assert( tj->ptr() != TypePtr::TopPTR && 1601 tj->ptr() != TypePtr::AnyNull && 1602 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1603 // assert( tj->ptr() != TypePtr::Constant || 1604 // tj->base() == Type::RawPtr || 1605 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1606 1607 return tj; 1608 } 1609 1610 void Compile::AliasType::Init(int i, const TypePtr* at) { 1611 _index = i; 1612 _adr_type = at; 1613 _field = NULL; 1614 _element = NULL; 1615 _is_rewritable = true; // default 1616 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1617 if (atoop != NULL && atoop->is_known_instance()) { 1618 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1619 _general_index = Compile::current()->get_alias_index(gt); 1620 } else { 1621 _general_index = 0; 1622 } 1623 } 1624 1625 BasicType Compile::AliasType::basic_type() const { 1626 if (element() != NULL) { 1627 const Type* element = adr_type()->is_aryptr()->elem(); 1628 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1629 } if (field() != NULL) { 1630 return field()->layout_type(); 1631 } else { 1632 return T_ILLEGAL; // unknown 1633 } 1634 } 1635 1636 //---------------------------------print_on------------------------------------ 1637 #ifndef PRODUCT 1638 void Compile::AliasType::print_on(outputStream* st) { 1639 if (index() < 10) 1640 st->print("@ <%d> ", index()); 1641 else st->print("@ <%d>", index()); 1642 st->print(is_rewritable() ? " " : " RO"); 1643 int offset = adr_type()->offset(); 1644 if (offset == Type::OffsetBot) 1645 st->print(" +any"); 1646 else st->print(" +%-3d", offset); 1647 st->print(" in "); 1648 adr_type()->dump_on(st); 1649 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1650 if (field() != NULL && tjp) { 1651 if (tjp->klass() != field()->holder() || 1652 tjp->offset() != field()->offset_in_bytes()) { 1653 st->print(" != "); 1654 field()->print(); 1655 st->print(" ***"); 1656 } 1657 } 1658 } 1659 1660 void print_alias_types() { 1661 Compile* C = Compile::current(); 1662 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1663 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1664 C->alias_type(idx)->print_on(tty); 1665 tty->cr(); 1666 } 1667 } 1668 #endif 1669 1670 1671 //----------------------------probe_alias_cache-------------------------------- 1672 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1673 intptr_t key = (intptr_t) adr_type; 1674 key ^= key >> logAliasCacheSize; 1675 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1676 } 1677 1678 1679 //-----------------------------grow_alias_types-------------------------------- 1680 void Compile::grow_alias_types() { 1681 const int old_ats = _max_alias_types; // how many before? 1682 const int new_ats = old_ats; // how many more? 1683 const int grow_ats = old_ats+new_ats; // how many now? 1684 _max_alias_types = grow_ats; 1685 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1686 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1687 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1688 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1689 } 1690 1691 1692 //--------------------------------find_alias_type------------------------------ 1693 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1694 if (_AliasLevel == 0) 1695 return alias_type(AliasIdxBot); 1696 1697 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1698 if (ace->_adr_type == adr_type) { 1699 return alias_type(ace->_index); 1700 } 1701 1702 // Handle special cases. 1703 if (adr_type == NULL) return alias_type(AliasIdxTop); 1704 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1705 1706 // Do it the slow way. 1707 const TypePtr* flat = flatten_alias_type(adr_type); 1708 1709 #ifdef ASSERT 1710 { 1711 ResourceMark rm; 1712 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1713 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1714 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1715 Type::str(adr_type)); 1716 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1717 const TypeOopPtr* foop = flat->is_oopptr(); 1718 // Scalarizable allocations have exact klass always. 1719 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1720 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1721 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1722 Type::str(foop), Type::str(xoop)); 1723 } 1724 } 1725 #endif 1726 1727 int idx = AliasIdxTop; 1728 for (int i = 0; i < num_alias_types(); i++) { 1729 if (alias_type(i)->adr_type() == flat) { 1730 idx = i; 1731 break; 1732 } 1733 } 1734 1735 if (idx == AliasIdxTop) { 1736 if (no_create) return NULL; 1737 // Grow the array if necessary. 1738 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1739 // Add a new alias type. 1740 idx = _num_alias_types++; 1741 _alias_types[idx]->Init(idx, flat); 1742 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1743 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1744 if (flat->isa_instptr()) { 1745 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1746 && flat->is_instptr()->klass() == env()->Class_klass()) 1747 alias_type(idx)->set_rewritable(false); 1748 } 1749 if (flat->isa_aryptr()) { 1750 #ifdef ASSERT 1751 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1752 // (T_BYTE has the weakest alignment and size restrictions...) 1753 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1754 #endif 1755 if (flat->offset() == TypePtr::OffsetBot) { 1756 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1757 } 1758 } 1759 if (flat->isa_klassptr()) { 1760 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1761 alias_type(idx)->set_rewritable(false); 1762 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1763 alias_type(idx)->set_rewritable(false); 1764 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1765 alias_type(idx)->set_rewritable(false); 1766 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1767 alias_type(idx)->set_rewritable(false); 1768 } 1769 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1770 // but the base pointer type is not distinctive enough to identify 1771 // references into JavaThread.) 1772 1773 // Check for final fields. 1774 const TypeInstPtr* tinst = flat->isa_instptr(); 1775 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1776 ciField* field; 1777 if (tinst->const_oop() != NULL && 1778 tinst->klass() == ciEnv::current()->Class_klass() && 1779 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1780 // static field 1781 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1782 field = k->get_field_by_offset(tinst->offset(), true); 1783 } else { 1784 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1785 field = k->get_field_by_offset(tinst->offset(), false); 1786 } 1787 assert(field == NULL || 1788 original_field == NULL || 1789 (field->holder() == original_field->holder() && 1790 field->offset() == original_field->offset() && 1791 field->is_static() == original_field->is_static()), "wrong field?"); 1792 // Set field() and is_rewritable() attributes. 1793 if (field != NULL) alias_type(idx)->set_field(field); 1794 } 1795 } 1796 1797 // Fill the cache for next time. 1798 ace->_adr_type = adr_type; 1799 ace->_index = idx; 1800 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1801 1802 // Might as well try to fill the cache for the flattened version, too. 1803 AliasCacheEntry* face = probe_alias_cache(flat); 1804 if (face->_adr_type == NULL) { 1805 face->_adr_type = flat; 1806 face->_index = idx; 1807 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1808 } 1809 1810 return alias_type(idx); 1811 } 1812 1813 1814 Compile::AliasType* Compile::alias_type(ciField* field) { 1815 const TypeOopPtr* t; 1816 if (field->is_static()) 1817 t = TypeInstPtr::make(field->holder()->java_mirror()); 1818 else 1819 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1820 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1821 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1822 return atp; 1823 } 1824 1825 1826 //------------------------------have_alias_type-------------------------------- 1827 bool Compile::have_alias_type(const TypePtr* adr_type) { 1828 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1829 if (ace->_adr_type == adr_type) { 1830 return true; 1831 } 1832 1833 // Handle special cases. 1834 if (adr_type == NULL) return true; 1835 if (adr_type == TypePtr::BOTTOM) return true; 1836 1837 return find_alias_type(adr_type, true, NULL) != NULL; 1838 } 1839 1840 //-----------------------------must_alias-------------------------------------- 1841 // True if all values of the given address type are in the given alias category. 1842 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1843 if (alias_idx == AliasIdxBot) return true; // the universal category 1844 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1845 if (alias_idx == AliasIdxTop) return false; // the empty category 1846 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1847 1848 // the only remaining possible overlap is identity 1849 int adr_idx = get_alias_index(adr_type); 1850 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1851 assert(adr_idx == alias_idx || 1852 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1853 && adr_type != TypeOopPtr::BOTTOM), 1854 "should not be testing for overlap with an unsafe pointer"); 1855 return adr_idx == alias_idx; 1856 } 1857 1858 //------------------------------can_alias-------------------------------------- 1859 // True if any values of the given address type are in the given alias category. 1860 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1861 if (alias_idx == AliasIdxTop) return false; // the empty category 1862 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1863 if (alias_idx == AliasIdxBot) return true; // the universal category 1864 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1865 1866 // the only remaining possible overlap is identity 1867 int adr_idx = get_alias_index(adr_type); 1868 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1869 return adr_idx == alias_idx; 1870 } 1871 1872 1873 1874 //---------------------------pop_warm_call------------------------------------- 1875 WarmCallInfo* Compile::pop_warm_call() { 1876 WarmCallInfo* wci = _warm_calls; 1877 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1878 return wci; 1879 } 1880 1881 //----------------------------Inline_Warm-------------------------------------- 1882 int Compile::Inline_Warm() { 1883 // If there is room, try to inline some more warm call sites. 1884 // %%% Do a graph index compaction pass when we think we're out of space? 1885 if (!InlineWarmCalls) return 0; 1886 1887 int calls_made_hot = 0; 1888 int room_to_grow = NodeCountInliningCutoff - unique(); 1889 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1890 int amount_grown = 0; 1891 WarmCallInfo* call; 1892 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1893 int est_size = (int)call->size(); 1894 if (est_size > (room_to_grow - amount_grown)) { 1895 // This one won't fit anyway. Get rid of it. 1896 call->make_cold(); 1897 continue; 1898 } 1899 call->make_hot(); 1900 calls_made_hot++; 1901 amount_grown += est_size; 1902 amount_to_grow -= est_size; 1903 } 1904 1905 if (calls_made_hot > 0) set_major_progress(); 1906 return calls_made_hot; 1907 } 1908 1909 1910 //----------------------------Finish_Warm-------------------------------------- 1911 void Compile::Finish_Warm() { 1912 if (!InlineWarmCalls) return; 1913 if (failing()) return; 1914 if (warm_calls() == NULL) return; 1915 1916 // Clean up loose ends, if we are out of space for inlining. 1917 WarmCallInfo* call; 1918 while ((call = pop_warm_call()) != NULL) { 1919 call->make_cold(); 1920 } 1921 } 1922 1923 //---------------------cleanup_loop_predicates----------------------- 1924 // Remove the opaque nodes that protect the predicates so that all unused 1925 // checks and uncommon_traps will be eliminated from the ideal graph 1926 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1927 if (predicate_count()==0) return; 1928 for (int i = predicate_count(); i > 0; i--) { 1929 Node * n = predicate_opaque1_node(i-1); 1930 assert(n->Opcode() == Op_Opaque1, "must be"); 1931 igvn.replace_node(n, n->in(1)); 1932 } 1933 assert(predicate_count()==0, "should be clean!"); 1934 } 1935 1936 void Compile::add_range_check_cast(Node* n) { 1937 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1938 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1939 _range_check_casts->append(n); 1940 } 1941 1942 // Remove all range check dependent CastIINodes. 1943 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1944 for (int i = range_check_cast_count(); i > 0; i--) { 1945 Node* cast = range_check_cast_node(i-1); 1946 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1947 igvn.replace_node(cast, cast->in(1)); 1948 } 1949 assert(range_check_cast_count() == 0, "should be empty"); 1950 } 1951 1952 // StringOpts and late inlining of string methods 1953 void Compile::inline_string_calls(bool parse_time) { 1954 { 1955 // remove useless nodes to make the usage analysis simpler 1956 ResourceMark rm; 1957 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1958 } 1959 1960 { 1961 ResourceMark rm; 1962 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1963 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1964 print_method(PHASE_AFTER_STRINGOPTS, 3); 1965 } 1966 1967 // now inline anything that we skipped the first time around 1968 if (!parse_time) { 1969 _late_inlines_pos = _late_inlines.length(); 1970 } 1971 1972 while (_string_late_inlines.length() > 0) { 1973 CallGenerator* cg = _string_late_inlines.pop(); 1974 cg->do_late_inline(); 1975 if (failing()) return; 1976 } 1977 _string_late_inlines.trunc_to(0); 1978 } 1979 1980 // Late inlining of boxing methods 1981 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1982 if (_boxing_late_inlines.length() > 0) { 1983 assert(has_boxed_value(), "inconsistent"); 1984 1985 PhaseGVN* gvn = initial_gvn(); 1986 set_inlining_incrementally(true); 1987 1988 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1989 for_igvn()->clear(); 1990 gvn->replace_with(&igvn); 1991 1992 _late_inlines_pos = _late_inlines.length(); 1993 1994 while (_boxing_late_inlines.length() > 0) { 1995 CallGenerator* cg = _boxing_late_inlines.pop(); 1996 cg->do_late_inline(); 1997 if (failing()) return; 1998 } 1999 _boxing_late_inlines.trunc_to(0); 2000 2001 { 2002 ResourceMark rm; 2003 PhaseRemoveUseless pru(gvn, for_igvn()); 2004 } 2005 2006 igvn = PhaseIterGVN(gvn); 2007 igvn.optimize(); 2008 2009 set_inlining_progress(false); 2010 set_inlining_incrementally(false); 2011 } 2012 } 2013 2014 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2015 assert(IncrementalInline, "incremental inlining should be on"); 2016 PhaseGVN* gvn = initial_gvn(); 2017 2018 set_inlining_progress(false); 2019 for_igvn()->clear(); 2020 gvn->replace_with(&igvn); 2021 2022 { 2023 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2024 int i = 0; 2025 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2026 CallGenerator* cg = _late_inlines.at(i); 2027 _late_inlines_pos = i+1; 2028 cg->do_late_inline(); 2029 if (failing()) return; 2030 } 2031 int j = 0; 2032 for (; i < _late_inlines.length(); i++, j++) { 2033 _late_inlines.at_put(j, _late_inlines.at(i)); 2034 } 2035 _late_inlines.trunc_to(j); 2036 } 2037 2038 { 2039 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2040 ResourceMark rm; 2041 PhaseRemoveUseless pru(gvn, for_igvn()); 2042 } 2043 2044 { 2045 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2046 igvn = PhaseIterGVN(gvn); 2047 } 2048 } 2049 2050 // Perform incremental inlining until bound on number of live nodes is reached 2051 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2052 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2053 2054 PhaseGVN* gvn = initial_gvn(); 2055 2056 set_inlining_incrementally(true); 2057 set_inlining_progress(true); 2058 uint low_live_nodes = 0; 2059 2060 while(inlining_progress() && _late_inlines.length() > 0) { 2061 2062 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2063 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2064 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2065 // PhaseIdealLoop is expensive so we only try it once we are 2066 // out of live nodes and we only try it again if the previous 2067 // helped got the number of nodes down significantly 2068 PhaseIdealLoop ideal_loop( igvn, false, true ); 2069 if (failing()) return; 2070 low_live_nodes = live_nodes(); 2071 _major_progress = true; 2072 } 2073 2074 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2075 break; 2076 } 2077 } 2078 2079 inline_incrementally_one(igvn); 2080 2081 if (failing()) return; 2082 2083 { 2084 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2085 igvn.optimize(); 2086 } 2087 2088 if (failing()) return; 2089 } 2090 2091 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2092 2093 if (_string_late_inlines.length() > 0) { 2094 assert(has_stringbuilder(), "inconsistent"); 2095 for_igvn()->clear(); 2096 initial_gvn()->replace_with(&igvn); 2097 2098 inline_string_calls(false); 2099 2100 if (failing()) return; 2101 2102 { 2103 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2104 ResourceMark rm; 2105 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2106 } 2107 2108 { 2109 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2110 igvn = PhaseIterGVN(gvn); 2111 igvn.optimize(); 2112 } 2113 } 2114 2115 set_inlining_incrementally(false); 2116 } 2117 2118 2119 //------------------------------Optimize--------------------------------------- 2120 // Given a graph, optimize it. 2121 void Compile::Optimize() { 2122 TracePhase tp("optimizer", &timers[_t_optimizer]); 2123 2124 #ifndef PRODUCT 2125 if (_directive->BreakAtCompileOption) { 2126 BREAKPOINT; 2127 } 2128 2129 #endif 2130 2131 ResourceMark rm; 2132 int loop_opts_cnt; 2133 2134 print_inlining_reinit(); 2135 2136 NOT_PRODUCT( verify_graph_edges(); ) 2137 2138 print_method(PHASE_AFTER_PARSING); 2139 2140 { 2141 // Iterative Global Value Numbering, including ideal transforms 2142 // Initialize IterGVN with types and values from parse-time GVN 2143 PhaseIterGVN igvn(initial_gvn()); 2144 #ifdef ASSERT 2145 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2146 #endif 2147 { 2148 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2149 igvn.optimize(); 2150 } 2151 2152 print_method(PHASE_ITER_GVN1, 2); 2153 2154 if (failing()) return; 2155 2156 inline_incrementally(igvn); 2157 2158 print_method(PHASE_INCREMENTAL_INLINE, 2); 2159 2160 if (failing()) return; 2161 2162 if (eliminate_boxing()) { 2163 // Inline valueOf() methods now. 2164 inline_boxing_calls(igvn); 2165 2166 if (AlwaysIncrementalInline) { 2167 inline_incrementally(igvn); 2168 } 2169 2170 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2171 2172 if (failing()) return; 2173 } 2174 2175 // Remove the speculative part of types and clean up the graph from 2176 // the extra CastPP nodes whose only purpose is to carry them. Do 2177 // that early so that optimizations are not disrupted by the extra 2178 // CastPP nodes. 2179 remove_speculative_types(igvn); 2180 2181 // No more new expensive nodes will be added to the list from here 2182 // so keep only the actual candidates for optimizations. 2183 cleanup_expensive_nodes(igvn); 2184 2185 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2186 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2187 initial_gvn()->replace_with(&igvn); 2188 for_igvn()->clear(); 2189 Unique_Node_List new_worklist(C->comp_arena()); 2190 { 2191 ResourceMark rm; 2192 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2193 } 2194 set_for_igvn(&new_worklist); 2195 igvn = PhaseIterGVN(initial_gvn()); 2196 igvn.optimize(); 2197 } 2198 2199 // Perform escape analysis 2200 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2201 if (has_loops()) { 2202 // Cleanup graph (remove dead nodes). 2203 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2204 PhaseIdealLoop ideal_loop( igvn, false, true ); 2205 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2206 if (failing()) return; 2207 } 2208 ConnectionGraph::do_analysis(this, &igvn); 2209 2210 if (failing()) return; 2211 2212 // Optimize out fields loads from scalar replaceable allocations. 2213 igvn.optimize(); 2214 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2215 2216 if (failing()) return; 2217 2218 if (congraph() != NULL && macro_count() > 0) { 2219 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2220 PhaseMacroExpand mexp(igvn); 2221 mexp.eliminate_macro_nodes(); 2222 igvn.set_delay_transform(false); 2223 2224 igvn.optimize(); 2225 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2226 2227 if (failing()) return; 2228 } 2229 } 2230 2231 // Loop transforms on the ideal graph. Range Check Elimination, 2232 // peeling, unrolling, etc. 2233 2234 // Set loop opts counter 2235 loop_opts_cnt = num_loop_opts(); 2236 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2237 { 2238 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2239 PhaseIdealLoop ideal_loop( igvn, true ); 2240 loop_opts_cnt--; 2241 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2242 if (failing()) return; 2243 } 2244 // Loop opts pass if partial peeling occurred in previous pass 2245 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2246 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2247 PhaseIdealLoop ideal_loop( igvn, false ); 2248 loop_opts_cnt--; 2249 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2250 if (failing()) return; 2251 } 2252 // Loop opts pass for loop-unrolling before CCP 2253 if(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_PHASEIDEALLOOP3, 2); 2258 } 2259 if (!failing()) { 2260 // Verify that last round of loop opts produced a valid graph 2261 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2262 PhaseIdealLoop::verify(igvn); 2263 } 2264 } 2265 if (failing()) return; 2266 2267 // Conditional Constant Propagation; 2268 PhaseCCP ccp( &igvn ); 2269 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2270 { 2271 TracePhase tp("ccp", &timers[_t_ccp]); 2272 ccp.do_transform(); 2273 } 2274 print_method(PHASE_CPP1, 2); 2275 2276 assert( true, "Break here to ccp.dump_old2new_map()"); 2277 2278 // Iterative Global Value Numbering, including ideal transforms 2279 { 2280 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2281 igvn = ccp; 2282 igvn.optimize(); 2283 } 2284 2285 print_method(PHASE_ITER_GVN2, 2); 2286 2287 if (failing()) return; 2288 2289 // Loop transforms on the ideal graph. Range Check Elimination, 2290 // peeling, unrolling, etc. 2291 if(loop_opts_cnt > 0) { 2292 debug_only( int cnt = 0; ); 2293 while(major_progress() && (loop_opts_cnt > 0)) { 2294 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2295 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2296 PhaseIdealLoop ideal_loop( igvn, true); 2297 loop_opts_cnt--; 2298 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2299 if (failing()) return; 2300 } 2301 } 2302 // Ensure that major progress is now clear 2303 C->clear_major_progress(); 2304 2305 { 2306 // Verify that all previous optimizations produced a valid graph 2307 // at least to this point, even if no loop optimizations were done. 2308 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2309 PhaseIdealLoop::verify(igvn); 2310 } 2311 2312 if (range_check_cast_count() > 0) { 2313 // No more loop optimizations. Remove all range check dependent CastIINodes. 2314 C->remove_range_check_casts(igvn); 2315 igvn.optimize(); 2316 } 2317 2318 { 2319 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2320 PhaseMacroExpand mex(igvn); 2321 if (mex.expand_macro_nodes()) { 2322 assert(failing(), "must bail out w/ explicit message"); 2323 return; 2324 } 2325 } 2326 2327 DEBUG_ONLY( _modified_nodes = NULL; ) 2328 } // (End scope of igvn; run destructor if necessary for asserts.) 2329 2330 process_print_inlining(); 2331 // A method with only infinite loops has no edges entering loops from root 2332 { 2333 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2334 if (final_graph_reshaping()) { 2335 assert(failing(), "must bail out w/ explicit message"); 2336 return; 2337 } 2338 } 2339 2340 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2341 } 2342 2343 2344 //------------------------------Code_Gen--------------------------------------- 2345 // Given a graph, generate code for it 2346 void Compile::Code_Gen() { 2347 if (failing()) { 2348 return; 2349 } 2350 2351 // Perform instruction selection. You might think we could reclaim Matcher 2352 // memory PDQ, but actually the Matcher is used in generating spill code. 2353 // Internals of the Matcher (including some VectorSets) must remain live 2354 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2355 // set a bit in reclaimed memory. 2356 2357 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2358 // nodes. Mapping is only valid at the root of each matched subtree. 2359 NOT_PRODUCT( verify_graph_edges(); ) 2360 2361 Matcher matcher; 2362 _matcher = &matcher; 2363 { 2364 TracePhase tp("matcher", &timers[_t_matcher]); 2365 matcher.match(); 2366 } 2367 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2368 // nodes. Mapping is only valid at the root of each matched subtree. 2369 NOT_PRODUCT( verify_graph_edges(); ) 2370 2371 // If you have too many nodes, or if matching has failed, bail out 2372 check_node_count(0, "out of nodes matching instructions"); 2373 if (failing()) { 2374 return; 2375 } 2376 2377 // Build a proper-looking CFG 2378 PhaseCFG cfg(node_arena(), root(), matcher); 2379 _cfg = &cfg; 2380 { 2381 TracePhase tp("scheduler", &timers[_t_scheduler]); 2382 bool success = cfg.do_global_code_motion(); 2383 if (!success) { 2384 return; 2385 } 2386 2387 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2388 NOT_PRODUCT( verify_graph_edges(); ) 2389 debug_only( cfg.verify(); ) 2390 } 2391 2392 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2393 _regalloc = ®alloc; 2394 { 2395 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2396 // Perform register allocation. After Chaitin, use-def chains are 2397 // no longer accurate (at spill code) and so must be ignored. 2398 // Node->LRG->reg mappings are still accurate. 2399 _regalloc->Register_Allocate(); 2400 2401 // Bail out if the allocator builds too many nodes 2402 if (failing()) { 2403 return; 2404 } 2405 } 2406 2407 // Prior to register allocation we kept empty basic blocks in case the 2408 // the allocator needed a place to spill. After register allocation we 2409 // are not adding any new instructions. If any basic block is empty, we 2410 // can now safely remove it. 2411 { 2412 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2413 cfg.remove_empty_blocks(); 2414 if (do_freq_based_layout()) { 2415 PhaseBlockLayout layout(cfg); 2416 } else { 2417 cfg.set_loop_alignment(); 2418 } 2419 cfg.fixup_flow(); 2420 } 2421 2422 // Apply peephole optimizations 2423 if( OptoPeephole ) { 2424 TracePhase tp("peephole", &timers[_t_peephole]); 2425 PhasePeephole peep( _regalloc, cfg); 2426 peep.do_transform(); 2427 } 2428 2429 // Do late expand if CPU requires this. 2430 if (Matcher::require_postalloc_expand) { 2431 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2432 cfg.postalloc_expand(_regalloc); 2433 } 2434 2435 // Convert Nodes to instruction bits in a buffer 2436 { 2437 TraceTime tp("output", &timers[_t_output], CITime); 2438 Output(); 2439 } 2440 2441 print_method(PHASE_FINAL_CODE); 2442 2443 // He's dead, Jim. 2444 _cfg = (PhaseCFG*)0xdeadbeef; 2445 _regalloc = (PhaseChaitin*)0xdeadbeef; 2446 } 2447 2448 2449 //------------------------------dump_asm--------------------------------------- 2450 // Dump formatted assembly 2451 #ifndef PRODUCT 2452 void Compile::dump_asm(int *pcs, uint pc_limit) { 2453 bool cut_short = false; 2454 tty->print_cr("#"); 2455 tty->print("# "); _tf->dump(); tty->cr(); 2456 tty->print_cr("#"); 2457 2458 // For all blocks 2459 int pc = 0x0; // Program counter 2460 char starts_bundle = ' '; 2461 _regalloc->dump_frame(); 2462 2463 Node *n = NULL; 2464 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2465 if (VMThread::should_terminate()) { 2466 cut_short = true; 2467 break; 2468 } 2469 Block* block = _cfg->get_block(i); 2470 if (block->is_connector() && !Verbose) { 2471 continue; 2472 } 2473 n = block->head(); 2474 if (pcs && n->_idx < pc_limit) { 2475 tty->print("%3.3x ", pcs[n->_idx]); 2476 } else { 2477 tty->print(" "); 2478 } 2479 block->dump_head(_cfg); 2480 if (block->is_connector()) { 2481 tty->print_cr(" # Empty connector block"); 2482 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2483 tty->print_cr(" # Block is sole successor of call"); 2484 } 2485 2486 // For all instructions 2487 Node *delay = NULL; 2488 for (uint j = 0; j < block->number_of_nodes(); j++) { 2489 if (VMThread::should_terminate()) { 2490 cut_short = true; 2491 break; 2492 } 2493 n = block->get_node(j); 2494 if (valid_bundle_info(n)) { 2495 Bundle* bundle = node_bundling(n); 2496 if (bundle->used_in_unconditional_delay()) { 2497 delay = n; 2498 continue; 2499 } 2500 if (bundle->starts_bundle()) { 2501 starts_bundle = '+'; 2502 } 2503 } 2504 2505 if (WizardMode) { 2506 n->dump(); 2507 } 2508 2509 if( !n->is_Region() && // Dont print in the Assembly 2510 !n->is_Phi() && // a few noisely useless nodes 2511 !n->is_Proj() && 2512 !n->is_MachTemp() && 2513 !n->is_SafePointScalarObject() && 2514 !n->is_Catch() && // Would be nice to print exception table targets 2515 !n->is_MergeMem() && // Not very interesting 2516 !n->is_top() && // Debug info table constants 2517 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2518 ) { 2519 if (pcs && n->_idx < pc_limit) 2520 tty->print("%3.3x", pcs[n->_idx]); 2521 else 2522 tty->print(" "); 2523 tty->print(" %c ", starts_bundle); 2524 starts_bundle = ' '; 2525 tty->print("\t"); 2526 n->format(_regalloc, tty); 2527 tty->cr(); 2528 } 2529 2530 // If we have an instruction with a delay slot, and have seen a delay, 2531 // then back up and print it 2532 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2533 assert(delay != NULL, "no unconditional delay instruction"); 2534 if (WizardMode) delay->dump(); 2535 2536 if (node_bundling(delay)->starts_bundle()) 2537 starts_bundle = '+'; 2538 if (pcs && n->_idx < pc_limit) 2539 tty->print("%3.3x", pcs[n->_idx]); 2540 else 2541 tty->print(" "); 2542 tty->print(" %c ", starts_bundle); 2543 starts_bundle = ' '; 2544 tty->print("\t"); 2545 delay->format(_regalloc, tty); 2546 tty->cr(); 2547 delay = NULL; 2548 } 2549 2550 // Dump the exception table as well 2551 if( n->is_Catch() && (Verbose || WizardMode) ) { 2552 // Print the exception table for this offset 2553 _handler_table.print_subtable_for(pc); 2554 } 2555 } 2556 2557 if (pcs && n->_idx < pc_limit) 2558 tty->print_cr("%3.3x", pcs[n->_idx]); 2559 else 2560 tty->cr(); 2561 2562 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2563 2564 } // End of per-block dump 2565 tty->cr(); 2566 2567 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2568 } 2569 #endif 2570 2571 //------------------------------Final_Reshape_Counts--------------------------- 2572 // This class defines counters to help identify when a method 2573 // may/must be executed using hardware with only 24-bit precision. 2574 struct Final_Reshape_Counts : public StackObj { 2575 int _call_count; // count non-inlined 'common' calls 2576 int _float_count; // count float ops requiring 24-bit precision 2577 int _double_count; // count double ops requiring more precision 2578 int _java_call_count; // count non-inlined 'java' calls 2579 int _inner_loop_count; // count loops which need alignment 2580 VectorSet _visited; // Visitation flags 2581 Node_List _tests; // Set of IfNodes & PCTableNodes 2582 2583 Final_Reshape_Counts() : 2584 _call_count(0), _float_count(0), _double_count(0), 2585 _java_call_count(0), _inner_loop_count(0), 2586 _visited( Thread::current()->resource_area() ) { } 2587 2588 void inc_call_count () { _call_count ++; } 2589 void inc_float_count () { _float_count ++; } 2590 void inc_double_count() { _double_count++; } 2591 void inc_java_call_count() { _java_call_count++; } 2592 void inc_inner_loop_count() { _inner_loop_count++; } 2593 2594 int get_call_count () const { return _call_count ; } 2595 int get_float_count () const { return _float_count ; } 2596 int get_double_count() const { return _double_count; } 2597 int get_java_call_count() const { return _java_call_count; } 2598 int get_inner_loop_count() const { return _inner_loop_count; } 2599 }; 2600 2601 #ifdef ASSERT 2602 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2603 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2604 // Make sure the offset goes inside the instance layout. 2605 return k->contains_field_offset(tp->offset()); 2606 // Note that OffsetBot and OffsetTop are very negative. 2607 } 2608 #endif 2609 2610 // Eliminate trivially redundant StoreCMs and accumulate their 2611 // precedence edges. 2612 void Compile::eliminate_redundant_card_marks(Node* n) { 2613 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2614 if (n->in(MemNode::Address)->outcnt() > 1) { 2615 // There are multiple users of the same address so it might be 2616 // possible to eliminate some of the StoreCMs 2617 Node* mem = n->in(MemNode::Memory); 2618 Node* adr = n->in(MemNode::Address); 2619 Node* val = n->in(MemNode::ValueIn); 2620 Node* prev = n; 2621 bool done = false; 2622 // Walk the chain of StoreCMs eliminating ones that match. As 2623 // long as it's a chain of single users then the optimization is 2624 // safe. Eliminating partially redundant StoreCMs would require 2625 // cloning copies down the other paths. 2626 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2627 if (adr == mem->in(MemNode::Address) && 2628 val == mem->in(MemNode::ValueIn)) { 2629 // redundant StoreCM 2630 if (mem->req() > MemNode::OopStore) { 2631 // Hasn't been processed by this code yet. 2632 n->add_prec(mem->in(MemNode::OopStore)); 2633 } else { 2634 // Already converted to precedence edge 2635 for (uint i = mem->req(); i < mem->len(); i++) { 2636 // Accumulate any precedence edges 2637 if (mem->in(i) != NULL) { 2638 n->add_prec(mem->in(i)); 2639 } 2640 } 2641 // Everything above this point has been processed. 2642 done = true; 2643 } 2644 // Eliminate the previous StoreCM 2645 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2646 assert(mem->outcnt() == 0, "should be dead"); 2647 mem->disconnect_inputs(NULL, this); 2648 } else { 2649 prev = mem; 2650 } 2651 mem = prev->in(MemNode::Memory); 2652 } 2653 } 2654 } 2655 2656 //------------------------------final_graph_reshaping_impl---------------------- 2657 // Implement items 1-5 from final_graph_reshaping below. 2658 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2659 2660 if ( n->outcnt() == 0 ) return; // dead node 2661 uint nop = n->Opcode(); 2662 2663 // Check for 2-input instruction with "last use" on right input. 2664 // Swap to left input. Implements item (2). 2665 if( n->req() == 3 && // two-input instruction 2666 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2667 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2668 n->in(2)->outcnt() == 1 &&// right use IS a last use 2669 !n->in(2)->is_Con() ) { // right use is not a constant 2670 // Check for commutative opcode 2671 switch( nop ) { 2672 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2673 case Op_MaxI: case Op_MinI: 2674 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2675 case Op_AndL: case Op_XorL: case Op_OrL: 2676 case Op_AndI: case Op_XorI: case Op_OrI: { 2677 // Move "last use" input to left by swapping inputs 2678 n->swap_edges(1, 2); 2679 break; 2680 } 2681 default: 2682 break; 2683 } 2684 } 2685 2686 #ifdef ASSERT 2687 if( n->is_Mem() ) { 2688 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2689 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2690 // oop will be recorded in oop map if load crosses safepoint 2691 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2692 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2693 "raw memory operations should have control edge"); 2694 } 2695 #endif 2696 // Count FPU ops and common calls, implements item (3) 2697 switch( nop ) { 2698 // Count all float operations that may use FPU 2699 case Op_AddF: 2700 case Op_SubF: 2701 case Op_MulF: 2702 case Op_DivF: 2703 case Op_NegF: 2704 case Op_ModF: 2705 case Op_ConvI2F: 2706 case Op_ConF: 2707 case Op_CmpF: 2708 case Op_CmpF3: 2709 // case Op_ConvL2F: // longs are split into 32-bit halves 2710 frc.inc_float_count(); 2711 break; 2712 2713 case Op_ConvF2D: 2714 case Op_ConvD2F: 2715 frc.inc_float_count(); 2716 frc.inc_double_count(); 2717 break; 2718 2719 // Count all double operations that may use FPU 2720 case Op_AddD: 2721 case Op_SubD: 2722 case Op_MulD: 2723 case Op_DivD: 2724 case Op_NegD: 2725 case Op_ModD: 2726 case Op_ConvI2D: 2727 case Op_ConvD2I: 2728 // case Op_ConvL2D: // handled by leaf call 2729 // case Op_ConvD2L: // handled by leaf call 2730 case Op_ConD: 2731 case Op_CmpD: 2732 case Op_CmpD3: 2733 frc.inc_double_count(); 2734 break; 2735 case Op_Opaque1: // Remove Opaque Nodes before matching 2736 case Op_Opaque2: // Remove Opaque Nodes before matching 2737 case Op_Opaque3: 2738 n->subsume_by(n->in(1), this); 2739 break; 2740 case Op_CallStaticJava: 2741 case Op_CallJava: 2742 case Op_CallDynamicJava: 2743 frc.inc_java_call_count(); // Count java call site; 2744 case Op_CallRuntime: 2745 case Op_CallLeaf: 2746 case Op_CallLeafNoFP: { 2747 assert( n->is_Call(), "" ); 2748 CallNode *call = n->as_Call(); 2749 // Count call sites where the FP mode bit would have to be flipped. 2750 // Do not count uncommon runtime calls: 2751 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2752 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2753 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2754 frc.inc_call_count(); // Count the call site 2755 } else { // See if uncommon argument is shared 2756 Node *n = call->in(TypeFunc::Parms); 2757 int nop = n->Opcode(); 2758 // Clone shared simple arguments to uncommon calls, item (1). 2759 if( n->outcnt() > 1 && 2760 !n->is_Proj() && 2761 nop != Op_CreateEx && 2762 nop != Op_CheckCastPP && 2763 nop != Op_DecodeN && 2764 nop != Op_DecodeNKlass && 2765 !n->is_Mem() ) { 2766 Node *x = n->clone(); 2767 call->set_req( TypeFunc::Parms, x ); 2768 } 2769 } 2770 break; 2771 } 2772 2773 case Op_StoreD: 2774 case Op_LoadD: 2775 case Op_LoadD_unaligned: 2776 frc.inc_double_count(); 2777 goto handle_mem; 2778 case Op_StoreF: 2779 case Op_LoadF: 2780 frc.inc_float_count(); 2781 goto handle_mem; 2782 2783 case Op_StoreCM: 2784 { 2785 // Convert OopStore dependence into precedence edge 2786 Node* prec = n->in(MemNode::OopStore); 2787 n->del_req(MemNode::OopStore); 2788 n->add_prec(prec); 2789 eliminate_redundant_card_marks(n); 2790 } 2791 2792 // fall through 2793 2794 case Op_StoreB: 2795 case Op_StoreC: 2796 case Op_StorePConditional: 2797 case Op_StoreI: 2798 case Op_StoreL: 2799 case Op_StoreIConditional: 2800 case Op_StoreLConditional: 2801 case Op_CompareAndSwapB: 2802 case Op_CompareAndSwapS: 2803 case Op_CompareAndSwapI: 2804 case Op_CompareAndSwapL: 2805 case Op_CompareAndSwapP: 2806 case Op_CompareAndSwapN: 2807 case Op_WeakCompareAndSwapB: 2808 case Op_WeakCompareAndSwapS: 2809 case Op_WeakCompareAndSwapI: 2810 case Op_WeakCompareAndSwapL: 2811 case Op_WeakCompareAndSwapP: 2812 case Op_WeakCompareAndSwapN: 2813 case Op_CompareAndExchangeB: 2814 case Op_CompareAndExchangeS: 2815 case Op_CompareAndExchangeI: 2816 case Op_CompareAndExchangeL: 2817 case Op_CompareAndExchangeP: 2818 case Op_CompareAndExchangeN: 2819 case Op_GetAndAddS: 2820 case Op_GetAndAddB: 2821 case Op_GetAndAddI: 2822 case Op_GetAndAddL: 2823 case Op_GetAndSetS: 2824 case Op_GetAndSetB: 2825 case Op_GetAndSetI: 2826 case Op_GetAndSetL: 2827 case Op_GetAndSetP: 2828 case Op_GetAndSetN: 2829 case Op_StoreP: 2830 case Op_StoreN: 2831 case Op_StoreNKlass: 2832 case Op_LoadB: 2833 case Op_LoadUB: 2834 case Op_LoadUS: 2835 case Op_LoadI: 2836 case Op_LoadKlass: 2837 case Op_LoadNKlass: 2838 case Op_LoadL: 2839 case Op_LoadL_unaligned: 2840 case Op_LoadPLocked: 2841 case Op_LoadP: 2842 case Op_LoadN: 2843 case Op_LoadRange: 2844 case Op_LoadS: { 2845 handle_mem: 2846 #ifdef ASSERT 2847 if( VerifyOptoOopOffsets ) { 2848 assert( n->is_Mem(), "" ); 2849 MemNode *mem = (MemNode*)n; 2850 // Check to see if address types have grounded out somehow. 2851 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2852 assert( !tp || oop_offset_is_sane(tp), "" ); 2853 } 2854 #endif 2855 break; 2856 } 2857 2858 case Op_AddP: { // Assert sane base pointers 2859 Node *addp = n->in(AddPNode::Address); 2860 assert( !addp->is_AddP() || 2861 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2862 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2863 "Base pointers must match (addp %u)", addp->_idx ); 2864 #ifdef _LP64 2865 if ((UseCompressedOops || UseCompressedClassPointers) && 2866 addp->Opcode() == Op_ConP && 2867 addp == n->in(AddPNode::Base) && 2868 n->in(AddPNode::Offset)->is_Con()) { 2869 // If the transformation of ConP to ConN+DecodeN is beneficial depends 2870 // on the platform and on the compressed oops mode. 2871 // Use addressing with narrow klass to load with offset on x86. 2872 // Some platforms can use the constant pool to load ConP. 2873 // Do this transformation here since IGVN will convert ConN back to ConP. 2874 const Type* t = addp->bottom_type(); 2875 bool is_oop = t->isa_oopptr() != NULL; 2876 bool is_klass = t->isa_klassptr() != NULL; 2877 2878 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 2879 (is_klass && Matcher::const_klass_prefer_decode())) { 2880 Node* nn = NULL; 2881 2882 int op = is_oop ? Op_ConN : Op_ConNKlass; 2883 2884 // Look for existing ConN node of the same exact type. 2885 Node* r = root(); 2886 uint cnt = r->outcnt(); 2887 for (uint i = 0; i < cnt; i++) { 2888 Node* m = r->raw_out(i); 2889 if (m!= NULL && m->Opcode() == op && 2890 m->bottom_type()->make_ptr() == t) { 2891 nn = m; 2892 break; 2893 } 2894 } 2895 if (nn != NULL) { 2896 // Decode a narrow oop to match address 2897 // [R12 + narrow_oop_reg<<3 + offset] 2898 if (is_oop) { 2899 nn = new DecodeNNode(nn, t); 2900 } else { 2901 nn = new DecodeNKlassNode(nn, t); 2902 } 2903 // Check for succeeding AddP which uses the same Base. 2904 // Otherwise we will run into the assertion above when visiting that guy. 2905 for (uint i = 0; i < n->outcnt(); ++i) { 2906 Node *out_i = n->raw_out(i); 2907 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 2908 out_i->set_req(AddPNode::Base, nn); 2909 #ifdef ASSERT 2910 for (uint j = 0; j < out_i->outcnt(); ++j) { 2911 Node *out_j = out_i->raw_out(j); 2912 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 2913 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 2914 } 2915 #endif 2916 } 2917 } 2918 n->set_req(AddPNode::Base, nn); 2919 n->set_req(AddPNode::Address, nn); 2920 if (addp->outcnt() == 0) { 2921 addp->disconnect_inputs(NULL, this); 2922 } 2923 } 2924 } 2925 } 2926 #endif 2927 // platform dependent reshaping of the address expression 2928 reshape_address(n->as_AddP()); 2929 break; 2930 } 2931 2932 case Op_CastPP: { 2933 // Remove CastPP nodes to gain more freedom during scheduling but 2934 // keep the dependency they encode as control or precedence edges 2935 // (if control is set already) on memory operations. Some CastPP 2936 // nodes don't have a control (don't carry a dependency): skip 2937 // those. 2938 if (n->in(0) != NULL) { 2939 ResourceMark rm; 2940 Unique_Node_List wq; 2941 wq.push(n); 2942 for (uint next = 0; next < wq.size(); ++next) { 2943 Node *m = wq.at(next); 2944 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2945 Node* use = m->fast_out(i); 2946 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2947 use->ensure_control_or_add_prec(n->in(0)); 2948 } else { 2949 switch(use->Opcode()) { 2950 case Op_AddP: 2951 case Op_DecodeN: 2952 case Op_DecodeNKlass: 2953 case Op_CheckCastPP: 2954 case Op_CastPP: 2955 wq.push(use); 2956 break; 2957 } 2958 } 2959 } 2960 } 2961 } 2962 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2963 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2964 Node* in1 = n->in(1); 2965 const Type* t = n->bottom_type(); 2966 Node* new_in1 = in1->clone(); 2967 new_in1->as_DecodeN()->set_type(t); 2968 2969 if (!Matcher::narrow_oop_use_complex_address()) { 2970 // 2971 // x86, ARM and friends can handle 2 adds in addressing mode 2972 // and Matcher can fold a DecodeN node into address by using 2973 // a narrow oop directly and do implicit NULL check in address: 2974 // 2975 // [R12 + narrow_oop_reg<<3 + offset] 2976 // NullCheck narrow_oop_reg 2977 // 2978 // On other platforms (Sparc) we have to keep new DecodeN node and 2979 // use it to do implicit NULL check in address: 2980 // 2981 // decode_not_null narrow_oop_reg, base_reg 2982 // [base_reg + offset] 2983 // NullCheck base_reg 2984 // 2985 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2986 // to keep the information to which NULL check the new DecodeN node 2987 // corresponds to use it as value in implicit_null_check(). 2988 // 2989 new_in1->set_req(0, n->in(0)); 2990 } 2991 2992 n->subsume_by(new_in1, this); 2993 if (in1->outcnt() == 0) { 2994 in1->disconnect_inputs(NULL, this); 2995 } 2996 } else { 2997 n->subsume_by(n->in(1), this); 2998 if (n->outcnt() == 0) { 2999 n->disconnect_inputs(NULL, this); 3000 } 3001 } 3002 break; 3003 } 3004 #ifdef _LP64 3005 case Op_CmpP: 3006 // Do this transformation here to preserve CmpPNode::sub() and 3007 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3008 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3009 Node* in1 = n->in(1); 3010 Node* in2 = n->in(2); 3011 if (!in1->is_DecodeNarrowPtr()) { 3012 in2 = in1; 3013 in1 = n->in(2); 3014 } 3015 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3016 3017 Node* new_in2 = NULL; 3018 if (in2->is_DecodeNarrowPtr()) { 3019 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3020 new_in2 = in2->in(1); 3021 } else if (in2->Opcode() == Op_ConP) { 3022 const Type* t = in2->bottom_type(); 3023 if (t == TypePtr::NULL_PTR) { 3024 assert(in1->is_DecodeN(), "compare klass to null?"); 3025 // Don't convert CmpP null check into CmpN if compressed 3026 // oops implicit null check is not generated. 3027 // This will allow to generate normal oop implicit null check. 3028 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3029 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3030 // 3031 // This transformation together with CastPP transformation above 3032 // will generated code for implicit NULL checks for compressed oops. 3033 // 3034 // The original code after Optimize() 3035 // 3036 // LoadN memory, narrow_oop_reg 3037 // decode narrow_oop_reg, base_reg 3038 // CmpP base_reg, NULL 3039 // CastPP base_reg // NotNull 3040 // Load [base_reg + offset], val_reg 3041 // 3042 // after these transformations will be 3043 // 3044 // LoadN memory, narrow_oop_reg 3045 // CmpN narrow_oop_reg, NULL 3046 // decode_not_null narrow_oop_reg, base_reg 3047 // Load [base_reg + offset], val_reg 3048 // 3049 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3050 // since narrow oops can be used in debug info now (see the code in 3051 // final_graph_reshaping_walk()). 3052 // 3053 // At the end the code will be matched to 3054 // on x86: 3055 // 3056 // Load_narrow_oop memory, narrow_oop_reg 3057 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3058 // NullCheck narrow_oop_reg 3059 // 3060 // and on sparc: 3061 // 3062 // Load_narrow_oop memory, narrow_oop_reg 3063 // decode_not_null narrow_oop_reg, base_reg 3064 // Load [base_reg + offset], val_reg 3065 // NullCheck base_reg 3066 // 3067 } else if (t->isa_oopptr()) { 3068 new_in2 = ConNode::make(t->make_narrowoop()); 3069 } else if (t->isa_klassptr()) { 3070 new_in2 = ConNode::make(t->make_narrowklass()); 3071 } 3072 } 3073 if (new_in2 != NULL) { 3074 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3075 n->subsume_by(cmpN, this); 3076 if (in1->outcnt() == 0) { 3077 in1->disconnect_inputs(NULL, this); 3078 } 3079 if (in2->outcnt() == 0) { 3080 in2->disconnect_inputs(NULL, this); 3081 } 3082 } 3083 } 3084 break; 3085 3086 case Op_DecodeN: 3087 case Op_DecodeNKlass: 3088 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3089 // DecodeN could be pinned when it can't be fold into 3090 // an address expression, see the code for Op_CastPP above. 3091 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3092 break; 3093 3094 case Op_EncodeP: 3095 case Op_EncodePKlass: { 3096 Node* in1 = n->in(1); 3097 if (in1->is_DecodeNarrowPtr()) { 3098 n->subsume_by(in1->in(1), this); 3099 } else if (in1->Opcode() == Op_ConP) { 3100 const Type* t = in1->bottom_type(); 3101 if (t == TypePtr::NULL_PTR) { 3102 assert(t->isa_oopptr(), "null klass?"); 3103 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3104 } else if (t->isa_oopptr()) { 3105 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3106 } else if (t->isa_klassptr()) { 3107 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3108 } 3109 } 3110 if (in1->outcnt() == 0) { 3111 in1->disconnect_inputs(NULL, this); 3112 } 3113 break; 3114 } 3115 3116 case Op_Proj: { 3117 if (OptimizeStringConcat) { 3118 ProjNode* p = n->as_Proj(); 3119 if (p->_is_io_use) { 3120 // Separate projections were used for the exception path which 3121 // are normally removed by a late inline. If it wasn't inlined 3122 // then they will hang around and should just be replaced with 3123 // the original one. 3124 Node* proj = NULL; 3125 // Replace with just one 3126 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3127 Node *use = i.get(); 3128 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3129 proj = use; 3130 break; 3131 } 3132 } 3133 assert(proj != NULL, "must be found"); 3134 p->subsume_by(proj, this); 3135 } 3136 } 3137 break; 3138 } 3139 3140 case Op_Phi: 3141 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3142 // The EncodeP optimization may create Phi with the same edges 3143 // for all paths. It is not handled well by Register Allocator. 3144 Node* unique_in = n->in(1); 3145 assert(unique_in != NULL, ""); 3146 uint cnt = n->req(); 3147 for (uint i = 2; i < cnt; i++) { 3148 Node* m = n->in(i); 3149 assert(m != NULL, ""); 3150 if (unique_in != m) 3151 unique_in = NULL; 3152 } 3153 if (unique_in != NULL) { 3154 n->subsume_by(unique_in, this); 3155 } 3156 } 3157 break; 3158 3159 #endif 3160 3161 #ifdef ASSERT 3162 case Op_CastII: 3163 // Verify that all range check dependent CastII nodes were removed. 3164 if (n->isa_CastII()->has_range_check()) { 3165 n->dump(3); 3166 assert(false, "Range check dependent CastII node was not removed"); 3167 } 3168 break; 3169 #endif 3170 3171 case Op_ModI: 3172 if (UseDivMod) { 3173 // Check if a%b and a/b both exist 3174 Node* d = n->find_similar(Op_DivI); 3175 if (d) { 3176 // Replace them with a fused divmod if supported 3177 if (Matcher::has_match_rule(Op_DivModI)) { 3178 DivModINode* divmod = DivModINode::make(n); 3179 d->subsume_by(divmod->div_proj(), this); 3180 n->subsume_by(divmod->mod_proj(), this); 3181 } else { 3182 // replace a%b with a-((a/b)*b) 3183 Node* mult = new MulINode(d, d->in(2)); 3184 Node* sub = new SubINode(d->in(1), mult); 3185 n->subsume_by(sub, this); 3186 } 3187 } 3188 } 3189 break; 3190 3191 case Op_ModL: 3192 if (UseDivMod) { 3193 // Check if a%b and a/b both exist 3194 Node* d = n->find_similar(Op_DivL); 3195 if (d) { 3196 // Replace them with a fused divmod if supported 3197 if (Matcher::has_match_rule(Op_DivModL)) { 3198 DivModLNode* divmod = DivModLNode::make(n); 3199 d->subsume_by(divmod->div_proj(), this); 3200 n->subsume_by(divmod->mod_proj(), this); 3201 } else { 3202 // replace a%b with a-((a/b)*b) 3203 Node* mult = new MulLNode(d, d->in(2)); 3204 Node* sub = new SubLNode(d->in(1), mult); 3205 n->subsume_by(sub, this); 3206 } 3207 } 3208 } 3209 break; 3210 3211 case Op_LoadVector: 3212 case Op_StoreVector: 3213 break; 3214 3215 case Op_AddReductionVI: 3216 case Op_AddReductionVL: 3217 case Op_AddReductionVF: 3218 case Op_AddReductionVD: 3219 case Op_MulReductionVI: 3220 case Op_MulReductionVL: 3221 case Op_MulReductionVF: 3222 case Op_MulReductionVD: 3223 break; 3224 3225 case Op_PackB: 3226 case Op_PackS: 3227 case Op_PackI: 3228 case Op_PackF: 3229 case Op_PackL: 3230 case Op_PackD: 3231 if (n->req()-1 > 2) { 3232 // Replace many operand PackNodes with a binary tree for matching 3233 PackNode* p = (PackNode*) n; 3234 Node* btp = p->binary_tree_pack(1, n->req()); 3235 n->subsume_by(btp, this); 3236 } 3237 break; 3238 case Op_Loop: 3239 case Op_CountedLoop: 3240 if (n->as_Loop()->is_inner_loop()) { 3241 frc.inc_inner_loop_count(); 3242 } 3243 break; 3244 case Op_LShiftI: 3245 case Op_RShiftI: 3246 case Op_URShiftI: 3247 case Op_LShiftL: 3248 case Op_RShiftL: 3249 case Op_URShiftL: 3250 if (Matcher::need_masked_shift_count) { 3251 // The cpu's shift instructions don't restrict the count to the 3252 // lower 5/6 bits. We need to do the masking ourselves. 3253 Node* in2 = n->in(2); 3254 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3255 const TypeInt* t = in2->find_int_type(); 3256 if (t != NULL && t->is_con()) { 3257 juint shift = t->get_con(); 3258 if (shift > mask) { // Unsigned cmp 3259 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3260 } 3261 } else { 3262 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3263 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3264 n->set_req(2, shift); 3265 } 3266 } 3267 if (in2->outcnt() == 0) { // Remove dead node 3268 in2->disconnect_inputs(NULL, this); 3269 } 3270 } 3271 break; 3272 case Op_MemBarStoreStore: 3273 case Op_MemBarRelease: 3274 // Break the link with AllocateNode: it is no longer useful and 3275 // confuses register allocation. 3276 if (n->req() > MemBarNode::Precedent) { 3277 n->set_req(MemBarNode::Precedent, top()); 3278 } 3279 break; 3280 case Op_RangeCheck: { 3281 RangeCheckNode* rc = n->as_RangeCheck(); 3282 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3283 n->subsume_by(iff, this); 3284 frc._tests.push(iff); 3285 break; 3286 } 3287 case Op_ConvI2L: { 3288 if (!Matcher::convi2l_type_required) { 3289 // Code generation on some platforms doesn't need accurate 3290 // ConvI2L types. Widening the type can help remove redundant 3291 // address computations. 3292 n->as_Type()->set_type(TypeLong::INT); 3293 ResourceMark rm; 3294 Node_List wq; 3295 wq.push(n); 3296 for (uint next = 0; next < wq.size(); next++) { 3297 Node *m = wq.at(next); 3298 3299 for(;;) { 3300 // Loop over all nodes with identical inputs edges as m 3301 Node* k = m->find_similar(m->Opcode()); 3302 if (k == NULL) { 3303 break; 3304 } 3305 // Push their uses so we get a chance to remove node made 3306 // redundant 3307 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3308 Node* u = k->fast_out(i); 3309 assert(!wq.contains(u), "shouldn't process one node several times"); 3310 if (u->Opcode() == Op_LShiftL || 3311 u->Opcode() == Op_AddL || 3312 u->Opcode() == Op_SubL || 3313 u->Opcode() == Op_AddP) { 3314 wq.push(u); 3315 } 3316 } 3317 // Replace all nodes with identical edges as m with m 3318 k->subsume_by(m, this); 3319 } 3320 } 3321 } 3322 break; 3323 } 3324 default: 3325 assert( !n->is_Call(), "" ); 3326 assert( !n->is_Mem(), "" ); 3327 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3328 break; 3329 } 3330 3331 // Collect CFG split points 3332 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3333 frc._tests.push(n); 3334 } 3335 } 3336 3337 //------------------------------final_graph_reshaping_walk--------------------- 3338 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3339 // requires that the walk visits a node's inputs before visiting the node. 3340 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3341 ResourceArea *area = Thread::current()->resource_area(); 3342 Unique_Node_List sfpt(area); 3343 3344 frc._visited.set(root->_idx); // first, mark node as visited 3345 uint cnt = root->req(); 3346 Node *n = root; 3347 uint i = 0; 3348 while (true) { 3349 if (i < cnt) { 3350 // Place all non-visited non-null inputs onto stack 3351 Node* m = n->in(i); 3352 ++i; 3353 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3354 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3355 // compute worst case interpreter size in case of a deoptimization 3356 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3357 3358 sfpt.push(m); 3359 } 3360 cnt = m->req(); 3361 nstack.push(n, i); // put on stack parent and next input's index 3362 n = m; 3363 i = 0; 3364 } 3365 } else { 3366 // Now do post-visit work 3367 final_graph_reshaping_impl( n, frc ); 3368 if (nstack.is_empty()) 3369 break; // finished 3370 n = nstack.node(); // Get node from stack 3371 cnt = n->req(); 3372 i = nstack.index(); 3373 nstack.pop(); // Shift to the next node on stack 3374 } 3375 } 3376 3377 // Skip next transformation if compressed oops are not used. 3378 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3379 (!UseCompressedOops && !UseCompressedClassPointers)) 3380 return; 3381 3382 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3383 // It could be done for an uncommon traps or any safepoints/calls 3384 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3385 while (sfpt.size() > 0) { 3386 n = sfpt.pop(); 3387 JVMState *jvms = n->as_SafePoint()->jvms(); 3388 assert(jvms != NULL, "sanity"); 3389 int start = jvms->debug_start(); 3390 int end = n->req(); 3391 bool is_uncommon = (n->is_CallStaticJava() && 3392 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3393 for (int j = start; j < end; j++) { 3394 Node* in = n->in(j); 3395 if (in->is_DecodeNarrowPtr()) { 3396 bool safe_to_skip = true; 3397 if (!is_uncommon ) { 3398 // Is it safe to skip? 3399 for (uint i = 0; i < in->outcnt(); i++) { 3400 Node* u = in->raw_out(i); 3401 if (!u->is_SafePoint() || 3402 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3403 safe_to_skip = false; 3404 } 3405 } 3406 } 3407 if (safe_to_skip) { 3408 n->set_req(j, in->in(1)); 3409 } 3410 if (in->outcnt() == 0) { 3411 in->disconnect_inputs(NULL, this); 3412 } 3413 } 3414 } 3415 } 3416 } 3417 3418 //------------------------------final_graph_reshaping-------------------------- 3419 // Final Graph Reshaping. 3420 // 3421 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3422 // and not commoned up and forced early. Must come after regular 3423 // optimizations to avoid GVN undoing the cloning. Clone constant 3424 // inputs to Loop Phis; these will be split by the allocator anyways. 3425 // Remove Opaque nodes. 3426 // (2) Move last-uses by commutative operations to the left input to encourage 3427 // Intel update-in-place two-address operations and better register usage 3428 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3429 // calls canonicalizing them back. 3430 // (3) Count the number of double-precision FP ops, single-precision FP ops 3431 // and call sites. On Intel, we can get correct rounding either by 3432 // forcing singles to memory (requires extra stores and loads after each 3433 // FP bytecode) or we can set a rounding mode bit (requires setting and 3434 // clearing the mode bit around call sites). The mode bit is only used 3435 // if the relative frequency of single FP ops to calls is low enough. 3436 // This is a key transform for SPEC mpeg_audio. 3437 // (4) Detect infinite loops; blobs of code reachable from above but not 3438 // below. Several of the Code_Gen algorithms fail on such code shapes, 3439 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3440 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3441 // Detection is by looking for IfNodes where only 1 projection is 3442 // reachable from below or CatchNodes missing some targets. 3443 // (5) Assert for insane oop offsets in debug mode. 3444 3445 bool Compile::final_graph_reshaping() { 3446 // an infinite loop may have been eliminated by the optimizer, 3447 // in which case the graph will be empty. 3448 if (root()->req() == 1) { 3449 record_method_not_compilable("trivial infinite loop"); 3450 return true; 3451 } 3452 3453 // Expensive nodes have their control input set to prevent the GVN 3454 // from freely commoning them. There's no GVN beyond this point so 3455 // no need to keep the control input. We want the expensive nodes to 3456 // be freely moved to the least frequent code path by gcm. 3457 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3458 for (int i = 0; i < expensive_count(); i++) { 3459 _expensive_nodes->at(i)->set_req(0, NULL); 3460 } 3461 3462 Final_Reshape_Counts frc; 3463 3464 // Visit everybody reachable! 3465 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3466 Node_Stack nstack(live_nodes() >> 1); 3467 final_graph_reshaping_walk(nstack, root(), frc); 3468 3469 // Check for unreachable (from below) code (i.e., infinite loops). 3470 for( uint i = 0; i < frc._tests.size(); i++ ) { 3471 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3472 // Get number of CFG targets. 3473 // Note that PCTables include exception targets after calls. 3474 uint required_outcnt = n->required_outcnt(); 3475 if (n->outcnt() != required_outcnt) { 3476 // Check for a few special cases. Rethrow Nodes never take the 3477 // 'fall-thru' path, so expected kids is 1 less. 3478 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3479 if (n->in(0)->in(0)->is_Call()) { 3480 CallNode *call = n->in(0)->in(0)->as_Call(); 3481 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3482 required_outcnt--; // Rethrow always has 1 less kid 3483 } else if (call->req() > TypeFunc::Parms && 3484 call->is_CallDynamicJava()) { 3485 // Check for null receiver. In such case, the optimizer has 3486 // detected that the virtual call will always result in a null 3487 // pointer exception. The fall-through projection of this CatchNode 3488 // will not be populated. 3489 Node *arg0 = call->in(TypeFunc::Parms); 3490 if (arg0->is_Type() && 3491 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3492 required_outcnt--; 3493 } 3494 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3495 call->req() > TypeFunc::Parms+1 && 3496 call->is_CallStaticJava()) { 3497 // Check for negative array length. In such case, the optimizer has 3498 // detected that the allocation attempt will always result in an 3499 // exception. There is no fall-through projection of this CatchNode . 3500 Node *arg1 = call->in(TypeFunc::Parms+1); 3501 if (arg1->is_Type() && 3502 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3503 required_outcnt--; 3504 } 3505 } 3506 } 3507 } 3508 // Recheck with a better notion of 'required_outcnt' 3509 if (n->outcnt() != required_outcnt) { 3510 record_method_not_compilable("malformed control flow"); 3511 return true; // Not all targets reachable! 3512 } 3513 } 3514 // Check that I actually visited all kids. Unreached kids 3515 // must be infinite loops. 3516 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3517 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3518 record_method_not_compilable("infinite loop"); 3519 return true; // Found unvisited kid; must be unreach 3520 } 3521 } 3522 3523 // If original bytecodes contained a mixture of floats and doubles 3524 // check if the optimizer has made it homogenous, item (3). 3525 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3526 frc.get_float_count() > 32 && 3527 frc.get_double_count() == 0 && 3528 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3529 set_24_bit_selection_and_mode( false, true ); 3530 } 3531 3532 set_java_calls(frc.get_java_call_count()); 3533 set_inner_loops(frc.get_inner_loop_count()); 3534 3535 // No infinite loops, no reason to bail out. 3536 return false; 3537 } 3538 3539 //-----------------------------too_many_traps---------------------------------- 3540 // Report if there are too many traps at the current method and bci. 3541 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3542 bool Compile::too_many_traps(ciMethod* method, 3543 int bci, 3544 Deoptimization::DeoptReason reason) { 3545 ciMethodData* md = method->method_data(); 3546 if (md->is_empty()) { 3547 // Assume the trap has not occurred, or that it occurred only 3548 // because of a transient condition during start-up in the interpreter. 3549 return false; 3550 } 3551 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3552 if (md->has_trap_at(bci, m, reason) != 0) { 3553 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3554 // Also, if there are multiple reasons, or if there is no per-BCI record, 3555 // assume the worst. 3556 if (log()) 3557 log()->elem("observe trap='%s' count='%d'", 3558 Deoptimization::trap_reason_name(reason), 3559 md->trap_count(reason)); 3560 return true; 3561 } else { 3562 // Ignore method/bci and see if there have been too many globally. 3563 return too_many_traps(reason, md); 3564 } 3565 } 3566 3567 // Less-accurate variant which does not require a method and bci. 3568 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3569 ciMethodData* logmd) { 3570 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3571 // Too many traps globally. 3572 // Note that we use cumulative trap_count, not just md->trap_count. 3573 if (log()) { 3574 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3575 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3576 Deoptimization::trap_reason_name(reason), 3577 mcount, trap_count(reason)); 3578 } 3579 return true; 3580 } else { 3581 // The coast is clear. 3582 return false; 3583 } 3584 } 3585 3586 //--------------------------too_many_recompiles-------------------------------- 3587 // Report if there are too many recompiles at the current method and bci. 3588 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3589 // Is not eager to return true, since this will cause the compiler to use 3590 // Action_none for a trap point, to avoid too many recompilations. 3591 bool Compile::too_many_recompiles(ciMethod* method, 3592 int bci, 3593 Deoptimization::DeoptReason reason) { 3594 ciMethodData* md = method->method_data(); 3595 if (md->is_empty()) { 3596 // Assume the trap has not occurred, or that it occurred only 3597 // because of a transient condition during start-up in the interpreter. 3598 return false; 3599 } 3600 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3601 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3602 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3603 Deoptimization::DeoptReason per_bc_reason 3604 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3605 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3606 if ((per_bc_reason == Deoptimization::Reason_none 3607 || md->has_trap_at(bci, m, reason) != 0) 3608 // The trap frequency measure we care about is the recompile count: 3609 && md->trap_recompiled_at(bci, m) 3610 && md->overflow_recompile_count() >= bc_cutoff) { 3611 // Do not emit a trap here if it has already caused recompilations. 3612 // Also, if there are multiple reasons, or if there is no per-BCI record, 3613 // assume the worst. 3614 if (log()) 3615 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3616 Deoptimization::trap_reason_name(reason), 3617 md->trap_count(reason), 3618 md->overflow_recompile_count()); 3619 return true; 3620 } else if (trap_count(reason) != 0 3621 && decompile_count() >= m_cutoff) { 3622 // Too many recompiles globally, and we have seen this sort of trap. 3623 // Use cumulative decompile_count, not just md->decompile_count. 3624 if (log()) 3625 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3626 Deoptimization::trap_reason_name(reason), 3627 md->trap_count(reason), trap_count(reason), 3628 md->decompile_count(), decompile_count()); 3629 return true; 3630 } else { 3631 // The coast is clear. 3632 return false; 3633 } 3634 } 3635 3636 // Compute when not to trap. Used by matching trap based nodes and 3637 // NullCheck optimization. 3638 void Compile::set_allowed_deopt_reasons() { 3639 _allowed_reasons = 0; 3640 if (is_method_compilation()) { 3641 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3642 assert(rs < BitsPerInt, "recode bit map"); 3643 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3644 _allowed_reasons |= nth_bit(rs); 3645 } 3646 } 3647 } 3648 } 3649 3650 #ifndef PRODUCT 3651 //------------------------------verify_graph_edges--------------------------- 3652 // Walk the Graph and verify that there is a one-to-one correspondence 3653 // between Use-Def edges and Def-Use edges in the graph. 3654 void Compile::verify_graph_edges(bool no_dead_code) { 3655 if (VerifyGraphEdges) { 3656 ResourceArea *area = Thread::current()->resource_area(); 3657 Unique_Node_List visited(area); 3658 // Call recursive graph walk to check edges 3659 _root->verify_edges(visited); 3660 if (no_dead_code) { 3661 // Now make sure that no visited node is used by an unvisited node. 3662 bool dead_nodes = false; 3663 Unique_Node_List checked(area); 3664 while (visited.size() > 0) { 3665 Node* n = visited.pop(); 3666 checked.push(n); 3667 for (uint i = 0; i < n->outcnt(); i++) { 3668 Node* use = n->raw_out(i); 3669 if (checked.member(use)) continue; // already checked 3670 if (visited.member(use)) continue; // already in the graph 3671 if (use->is_Con()) continue; // a dead ConNode is OK 3672 // At this point, we have found a dead node which is DU-reachable. 3673 if (!dead_nodes) { 3674 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3675 dead_nodes = true; 3676 } 3677 use->dump(2); 3678 tty->print_cr("---"); 3679 checked.push(use); // No repeats; pretend it is now checked. 3680 } 3681 } 3682 assert(!dead_nodes, "using nodes must be reachable from root"); 3683 } 3684 } 3685 } 3686 3687 // Verify GC barriers consistency 3688 // Currently supported: 3689 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3690 void Compile::verify_barriers() { 3691 if (UseG1GC) { 3692 // Verify G1 pre-barriers 3693 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active()); 3694 3695 ResourceArea *area = Thread::current()->resource_area(); 3696 Unique_Node_List visited(area); 3697 Node_List worklist(area); 3698 // We're going to walk control flow backwards starting from the Root 3699 worklist.push(_root); 3700 while (worklist.size() > 0) { 3701 Node* x = worklist.pop(); 3702 if (x == NULL || x == top()) continue; 3703 if (visited.member(x)) { 3704 continue; 3705 } else { 3706 visited.push(x); 3707 } 3708 3709 if (x->is_Region()) { 3710 for (uint i = 1; i < x->req(); i++) { 3711 worklist.push(x->in(i)); 3712 } 3713 } else { 3714 worklist.push(x->in(0)); 3715 // We are looking for the pattern: 3716 // /->ThreadLocal 3717 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3718 // \->ConI(0) 3719 // We want to verify that the If and the LoadB have the same control 3720 // See GraphKit::g1_write_barrier_pre() 3721 if (x->is_If()) { 3722 IfNode *iff = x->as_If(); 3723 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3724 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3725 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3726 && cmp->in(1)->is_Load()) { 3727 LoadNode* load = cmp->in(1)->as_Load(); 3728 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3729 && load->in(2)->in(3)->is_Con() 3730 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3731 3732 Node* if_ctrl = iff->in(0); 3733 Node* load_ctrl = load->in(0); 3734 3735 if (if_ctrl != load_ctrl) { 3736 // Skip possible CProj->NeverBranch in infinite loops 3737 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3738 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3739 if_ctrl = if_ctrl->in(0)->in(0); 3740 } 3741 } 3742 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3743 } 3744 } 3745 } 3746 } 3747 } 3748 } 3749 } 3750 } 3751 3752 #endif 3753 3754 // The Compile object keeps track of failure reasons separately from the ciEnv. 3755 // This is required because there is not quite a 1-1 relation between the 3756 // ciEnv and its compilation task and the Compile object. Note that one 3757 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3758 // to backtrack and retry without subsuming loads. Other than this backtracking 3759 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3760 // by the logic in C2Compiler. 3761 void Compile::record_failure(const char* reason) { 3762 if (log() != NULL) { 3763 log()->elem("failure reason='%s' phase='compile'", reason); 3764 } 3765 if (_failure_reason == NULL) { 3766 // Record the first failure reason. 3767 _failure_reason = reason; 3768 } 3769 3770 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3771 C->print_method(PHASE_FAILURE); 3772 } 3773 _root = NULL; // flush the graph, too 3774 } 3775 3776 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3777 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3778 _phase_name(name), _dolog(CITimeVerbose) 3779 { 3780 if (_dolog) { 3781 C = Compile::current(); 3782 _log = C->log(); 3783 } else { 3784 C = NULL; 3785 _log = NULL; 3786 } 3787 if (_log != NULL) { 3788 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3789 _log->stamp(); 3790 _log->end_head(); 3791 } 3792 } 3793 3794 Compile::TracePhase::~TracePhase() { 3795 3796 C = Compile::current(); 3797 if (_dolog) { 3798 _log = C->log(); 3799 } else { 3800 _log = NULL; 3801 } 3802 3803 #ifdef ASSERT 3804 if (PrintIdealNodeCount) { 3805 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3806 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3807 } 3808 3809 if (VerifyIdealNodeCount) { 3810 Compile::current()->print_missing_nodes(); 3811 } 3812 #endif 3813 3814 if (_log != NULL) { 3815 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3816 } 3817 } 3818 3819 //============================================================================= 3820 // Two Constant's are equal when the type and the value are equal. 3821 bool Compile::Constant::operator==(const Constant& other) { 3822 if (type() != other.type() ) return false; 3823 if (can_be_reused() != other.can_be_reused()) return false; 3824 // For floating point values we compare the bit pattern. 3825 switch (type()) { 3826 case T_INT: 3827 case T_FLOAT: return (_v._value.i == other._v._value.i); 3828 case T_LONG: 3829 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3830 case T_OBJECT: 3831 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3832 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3833 case T_METADATA: return (_v._metadata == other._v._metadata); 3834 default: ShouldNotReachHere(); 3835 } 3836 return false; 3837 } 3838 3839 static int type_to_size_in_bytes(BasicType t) { 3840 switch (t) { 3841 case T_INT: return sizeof(jint ); 3842 case T_LONG: return sizeof(jlong ); 3843 case T_FLOAT: return sizeof(jfloat ); 3844 case T_DOUBLE: return sizeof(jdouble); 3845 case T_METADATA: return sizeof(Metadata*); 3846 // We use T_VOID as marker for jump-table entries (labels) which 3847 // need an internal word relocation. 3848 case T_VOID: 3849 case T_ADDRESS: 3850 case T_OBJECT: return sizeof(jobject); 3851 } 3852 3853 ShouldNotReachHere(); 3854 return -1; 3855 } 3856 3857 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3858 // sort descending 3859 if (a->freq() > b->freq()) return -1; 3860 if (a->freq() < b->freq()) return 1; 3861 return 0; 3862 } 3863 3864 void Compile::ConstantTable::calculate_offsets_and_size() { 3865 // First, sort the array by frequencies. 3866 _constants.sort(qsort_comparator); 3867 3868 #ifdef ASSERT 3869 // Make sure all jump-table entries were sorted to the end of the 3870 // array (they have a negative frequency). 3871 bool found_void = false; 3872 for (int i = 0; i < _constants.length(); i++) { 3873 Constant con = _constants.at(i); 3874 if (con.type() == T_VOID) 3875 found_void = true; // jump-tables 3876 else 3877 assert(!found_void, "wrong sorting"); 3878 } 3879 #endif 3880 3881 int offset = 0; 3882 for (int i = 0; i < _constants.length(); i++) { 3883 Constant* con = _constants.adr_at(i); 3884 3885 // Align offset for type. 3886 int typesize = type_to_size_in_bytes(con->type()); 3887 offset = align_size_up(offset, typesize); 3888 con->set_offset(offset); // set constant's offset 3889 3890 if (con->type() == T_VOID) { 3891 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3892 offset = offset + typesize * n->outcnt(); // expand jump-table 3893 } else { 3894 offset = offset + typesize; 3895 } 3896 } 3897 3898 // Align size up to the next section start (which is insts; see 3899 // CodeBuffer::align_at_start). 3900 assert(_size == -1, "already set?"); 3901 _size = align_size_up(offset, CodeEntryAlignment); 3902 } 3903 3904 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3905 MacroAssembler _masm(&cb); 3906 for (int i = 0; i < _constants.length(); i++) { 3907 Constant con = _constants.at(i); 3908 address constant_addr = NULL; 3909 switch (con.type()) { 3910 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 3911 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3912 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3913 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3914 case T_OBJECT: { 3915 jobject obj = con.get_jobject(); 3916 int oop_index = _masm.oop_recorder()->find_index(obj); 3917 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3918 break; 3919 } 3920 case T_ADDRESS: { 3921 address addr = (address) con.get_jobject(); 3922 constant_addr = _masm.address_constant(addr); 3923 break; 3924 } 3925 // We use T_VOID as marker for jump-table entries (labels) which 3926 // need an internal word relocation. 3927 case T_VOID: { 3928 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3929 // Fill the jump-table with a dummy word. The real value is 3930 // filled in later in fill_jump_table. 3931 address dummy = (address) n; 3932 constant_addr = _masm.address_constant(dummy); 3933 // Expand jump-table 3934 for (uint i = 1; i < n->outcnt(); i++) { 3935 address temp_addr = _masm.address_constant(dummy + i); 3936 assert(temp_addr, "consts section too small"); 3937 } 3938 break; 3939 } 3940 case T_METADATA: { 3941 Metadata* obj = con.get_metadata(); 3942 int metadata_index = _masm.oop_recorder()->find_index(obj); 3943 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3944 break; 3945 } 3946 default: ShouldNotReachHere(); 3947 } 3948 assert(constant_addr, "consts section too small"); 3949 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3950 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 3951 } 3952 } 3953 3954 int Compile::ConstantTable::find_offset(Constant& con) const { 3955 int idx = _constants.find(con); 3956 assert(idx != -1, "constant must be in constant table"); 3957 int offset = _constants.at(idx).offset(); 3958 assert(offset != -1, "constant table not emitted yet?"); 3959 return offset; 3960 } 3961 3962 void Compile::ConstantTable::add(Constant& con) { 3963 if (con.can_be_reused()) { 3964 int idx = _constants.find(con); 3965 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3966 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3967 return; 3968 } 3969 } 3970 (void) _constants.append(con); 3971 } 3972 3973 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3974 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3975 Constant con(type, value, b->_freq); 3976 add(con); 3977 return con; 3978 } 3979 3980 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3981 Constant con(metadata); 3982 add(con); 3983 return con; 3984 } 3985 3986 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3987 jvalue value; 3988 BasicType type = oper->type()->basic_type(); 3989 switch (type) { 3990 case T_LONG: value.j = oper->constantL(); break; 3991 case T_FLOAT: value.f = oper->constantF(); break; 3992 case T_DOUBLE: value.d = oper->constantD(); break; 3993 case T_OBJECT: 3994 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3995 case T_METADATA: return add((Metadata*)oper->constant()); break; 3996 default: guarantee(false, "unhandled type: %s", type2name(type)); 3997 } 3998 return add(n, type, value); 3999 } 4000 4001 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4002 jvalue value; 4003 // We can use the node pointer here to identify the right jump-table 4004 // as this method is called from Compile::Fill_buffer right before 4005 // the MachNodes are emitted and the jump-table is filled (means the 4006 // MachNode pointers do not change anymore). 4007 value.l = (jobject) n; 4008 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4009 add(con); 4010 return con; 4011 } 4012 4013 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4014 // If called from Compile::scratch_emit_size do nothing. 4015 if (Compile::current()->in_scratch_emit_size()) return; 4016 4017 assert(labels.is_nonempty(), "must be"); 4018 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4019 4020 // Since MachConstantNode::constant_offset() also contains 4021 // table_base_offset() we need to subtract the table_base_offset() 4022 // to get the plain offset into the constant table. 4023 int offset = n->constant_offset() - table_base_offset(); 4024 4025 MacroAssembler _masm(&cb); 4026 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4027 4028 for (uint i = 0; i < n->outcnt(); i++) { 4029 address* constant_addr = &jump_table_base[i]; 4030 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)); 4031 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4032 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4033 } 4034 } 4035 4036 //----------------------------static_subtype_check----------------------------- 4037 // Shortcut important common cases when superklass is exact: 4038 // (0) superklass is java.lang.Object (can occur in reflective code) 4039 // (1) subklass is already limited to a subtype of superklass => always ok 4040 // (2) subklass does not overlap with superklass => always fail 4041 // (3) superklass has NO subtypes and we can check with a simple compare. 4042 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4043 if (StressReflectiveCode) { 4044 return SSC_full_test; // Let caller generate the general case. 4045 } 4046 4047 if (superk == env()->Object_klass()) { 4048 return SSC_always_true; // (0) this test cannot fail 4049 } 4050 4051 ciType* superelem = superk; 4052 if (superelem->is_array_klass()) 4053 superelem = superelem->as_array_klass()->base_element_type(); 4054 4055 if (!subk->is_interface()) { // cannot trust static interface types yet 4056 if (subk->is_subtype_of(superk)) { 4057 return SSC_always_true; // (1) false path dead; no dynamic test needed 4058 } 4059 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4060 !superk->is_subtype_of(subk)) { 4061 return SSC_always_false; 4062 } 4063 } 4064 4065 // If casting to an instance klass, it must have no subtypes 4066 if (superk->is_interface()) { 4067 // Cannot trust interfaces yet. 4068 // %%% S.B. superk->nof_implementors() == 1 4069 } else if (superelem->is_instance_klass()) { 4070 ciInstanceKlass* ik = superelem->as_instance_klass(); 4071 if (!ik->has_subklass() && !ik->is_interface()) { 4072 if (!ik->is_final()) { 4073 // Add a dependency if there is a chance of a later subclass. 4074 dependencies()->assert_leaf_type(ik); 4075 } 4076 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4077 } 4078 } else { 4079 // A primitive array type has no subtypes. 4080 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4081 } 4082 4083 return SSC_full_test; 4084 } 4085 4086 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4087 #ifdef _LP64 4088 // The scaled index operand to AddP must be a clean 64-bit value. 4089 // Java allows a 32-bit int to be incremented to a negative 4090 // value, which appears in a 64-bit register as a large 4091 // positive number. Using that large positive number as an 4092 // operand in pointer arithmetic has bad consequences. 4093 // On the other hand, 32-bit overflow is rare, and the possibility 4094 // can often be excluded, if we annotate the ConvI2L node with 4095 // a type assertion that its value is known to be a small positive 4096 // number. (The prior range check has ensured this.) 4097 // This assertion is used by ConvI2LNode::Ideal. 4098 int index_max = max_jint - 1; // array size is max_jint, index is one less 4099 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4100 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4101 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4102 #endif 4103 return idx; 4104 } 4105 4106 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4107 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4108 if (ctrl != NULL) { 4109 // Express control dependency by a CastII node with a narrow type. 4110 value = new CastIINode(value, itype, false, true /* range check dependency */); 4111 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4112 // node from floating above the range check during loop optimizations. Otherwise, the 4113 // ConvI2L node may be eliminated independently of the range check, causing the data path 4114 // to become TOP while the control path is still there (although it's unreachable). 4115 value->set_req(0, ctrl); 4116 // Save CastII node to remove it after loop optimizations. 4117 phase->C->add_range_check_cast(value); 4118 value = phase->transform(value); 4119 } 4120 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4121 return phase->transform(new ConvI2LNode(value, ltype)); 4122 } 4123 4124 // The message about the current inlining is accumulated in 4125 // _print_inlining_stream and transfered into the _print_inlining_list 4126 // once we know whether inlining succeeds or not. For regular 4127 // inlining, messages are appended to the buffer pointed by 4128 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4129 // a new buffer is added after _print_inlining_idx in the list. This 4130 // way we can update the inlining message for late inlining call site 4131 // when the inlining is attempted again. 4132 void Compile::print_inlining_init() { 4133 if (print_inlining() || print_intrinsics()) { 4134 _print_inlining_stream = new stringStream(); 4135 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4136 } 4137 } 4138 4139 void Compile::print_inlining_reinit() { 4140 if (print_inlining() || print_intrinsics()) { 4141 // Re allocate buffer when we change ResourceMark 4142 _print_inlining_stream = new stringStream(); 4143 } 4144 } 4145 4146 void Compile::print_inlining_reset() { 4147 _print_inlining_stream->reset(); 4148 } 4149 4150 void Compile::print_inlining_commit() { 4151 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4152 // Transfer the message from _print_inlining_stream to the current 4153 // _print_inlining_list buffer and clear _print_inlining_stream. 4154 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4155 print_inlining_reset(); 4156 } 4157 4158 void Compile::print_inlining_push() { 4159 // Add new buffer to the _print_inlining_list at current position 4160 _print_inlining_idx++; 4161 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4162 } 4163 4164 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4165 return _print_inlining_list->at(_print_inlining_idx); 4166 } 4167 4168 void Compile::print_inlining_update(CallGenerator* cg) { 4169 if (print_inlining() || print_intrinsics()) { 4170 if (!cg->is_late_inline()) { 4171 if (print_inlining_current().cg() != NULL) { 4172 print_inlining_push(); 4173 } 4174 print_inlining_commit(); 4175 } else { 4176 if (print_inlining_current().cg() != cg && 4177 (print_inlining_current().cg() != NULL || 4178 print_inlining_current().ss()->size() != 0)) { 4179 print_inlining_push(); 4180 } 4181 print_inlining_commit(); 4182 print_inlining_current().set_cg(cg); 4183 } 4184 } 4185 } 4186 4187 void Compile::print_inlining_move_to(CallGenerator* cg) { 4188 // We resume inlining at a late inlining call site. Locate the 4189 // corresponding inlining buffer so that we can update it. 4190 if (print_inlining()) { 4191 for (int i = 0; i < _print_inlining_list->length(); i++) { 4192 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4193 _print_inlining_idx = i; 4194 return; 4195 } 4196 } 4197 ShouldNotReachHere(); 4198 } 4199 } 4200 4201 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4202 if (print_inlining()) { 4203 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4204 assert(print_inlining_current().cg() == cg, "wrong entry"); 4205 // replace message with new message 4206 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4207 print_inlining_commit(); 4208 print_inlining_current().set_cg(cg); 4209 } 4210 } 4211 4212 void Compile::print_inlining_assert_ready() { 4213 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4214 } 4215 4216 void Compile::process_print_inlining() { 4217 bool do_print_inlining = print_inlining() || print_intrinsics(); 4218 if (do_print_inlining || log() != NULL) { 4219 // Print inlining message for candidates that we couldn't inline 4220 // for lack of space 4221 for (int i = 0; i < _late_inlines.length(); i++) { 4222 CallGenerator* cg = _late_inlines.at(i); 4223 if (!cg->is_mh_late_inline()) { 4224 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4225 if (do_print_inlining) { 4226 cg->print_inlining_late(msg); 4227 } 4228 log_late_inline_failure(cg, msg); 4229 } 4230 } 4231 } 4232 if (do_print_inlining) { 4233 ResourceMark rm; 4234 stringStream ss; 4235 for (int i = 0; i < _print_inlining_list->length(); i++) { 4236 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4237 } 4238 size_t end = ss.size(); 4239 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4240 strncpy(_print_inlining_output, ss.base(), end+1); 4241 _print_inlining_output[end] = 0; 4242 } 4243 } 4244 4245 void Compile::dump_print_inlining() { 4246 if (_print_inlining_output != NULL) { 4247 tty->print_raw(_print_inlining_output); 4248 } 4249 } 4250 4251 void Compile::log_late_inline(CallGenerator* cg) { 4252 if (log() != NULL) { 4253 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4254 cg->unique_id()); 4255 JVMState* p = cg->call_node()->jvms(); 4256 while (p != NULL) { 4257 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4258 p = p->caller(); 4259 } 4260 log()->tail("late_inline"); 4261 } 4262 } 4263 4264 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4265 log_late_inline(cg); 4266 if (log() != NULL) { 4267 log()->inline_fail(msg); 4268 } 4269 } 4270 4271 void Compile::log_inline_id(CallGenerator* cg) { 4272 if (log() != NULL) { 4273 // The LogCompilation tool needs a unique way to identify late 4274 // inline call sites. This id must be unique for this call site in 4275 // this compilation. Try to have it unique across compilations as 4276 // well because it can be convenient when grepping through the log 4277 // file. 4278 // Distinguish OSR compilations from others in case CICountOSR is 4279 // on. 4280 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4281 cg->set_unique_id(id); 4282 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4283 } 4284 } 4285 4286 void Compile::log_inline_failure(const char* msg) { 4287 if (C->log() != NULL) { 4288 C->log()->inline_fail(msg); 4289 } 4290 } 4291 4292 4293 // Dump inlining replay data to the stream. 4294 // Don't change thread state and acquire any locks. 4295 void Compile::dump_inline_data(outputStream* out) { 4296 InlineTree* inl_tree = ilt(); 4297 if (inl_tree != NULL) { 4298 out->print(" inline %d", inl_tree->count()); 4299 inl_tree->dump_replay_data(out); 4300 } 4301 } 4302 4303 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4304 if (n1->Opcode() < n2->Opcode()) return -1; 4305 else if (n1->Opcode() > n2->Opcode()) return 1; 4306 4307 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4308 for (uint i = 1; i < n1->req(); i++) { 4309 if (n1->in(i) < n2->in(i)) return -1; 4310 else if (n1->in(i) > n2->in(i)) return 1; 4311 } 4312 4313 return 0; 4314 } 4315 4316 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4317 Node* n1 = *n1p; 4318 Node* n2 = *n2p; 4319 4320 return cmp_expensive_nodes(n1, n2); 4321 } 4322 4323 void Compile::sort_expensive_nodes() { 4324 if (!expensive_nodes_sorted()) { 4325 _expensive_nodes->sort(cmp_expensive_nodes); 4326 } 4327 } 4328 4329 bool Compile::expensive_nodes_sorted() const { 4330 for (int i = 1; i < _expensive_nodes->length(); i++) { 4331 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4332 return false; 4333 } 4334 } 4335 return true; 4336 } 4337 4338 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4339 if (_expensive_nodes->length() == 0) { 4340 return false; 4341 } 4342 4343 assert(OptimizeExpensiveOps, "optimization off?"); 4344 4345 // Take this opportunity to remove dead nodes from the list 4346 int j = 0; 4347 for (int i = 0; i < _expensive_nodes->length(); i++) { 4348 Node* n = _expensive_nodes->at(i); 4349 if (!n->is_unreachable(igvn)) { 4350 assert(n->is_expensive(), "should be expensive"); 4351 _expensive_nodes->at_put(j, n); 4352 j++; 4353 } 4354 } 4355 _expensive_nodes->trunc_to(j); 4356 4357 // Then sort the list so that similar nodes are next to each other 4358 // and check for at least two nodes of identical kind with same data 4359 // inputs. 4360 sort_expensive_nodes(); 4361 4362 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4363 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4364 return true; 4365 } 4366 } 4367 4368 return false; 4369 } 4370 4371 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4372 if (_expensive_nodes->length() == 0) { 4373 return; 4374 } 4375 4376 assert(OptimizeExpensiveOps, "optimization off?"); 4377 4378 // Sort to bring similar nodes next to each other and clear the 4379 // control input of nodes for which there's only a single copy. 4380 sort_expensive_nodes(); 4381 4382 int j = 0; 4383 int identical = 0; 4384 int i = 0; 4385 bool modified = false; 4386 for (; i < _expensive_nodes->length()-1; i++) { 4387 assert(j <= i, "can't write beyond current index"); 4388 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4389 identical++; 4390 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4391 continue; 4392 } 4393 if (identical > 0) { 4394 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4395 identical = 0; 4396 } else { 4397 Node* n = _expensive_nodes->at(i); 4398 igvn.replace_input_of(n, 0, NULL); 4399 igvn.hash_insert(n); 4400 modified = true; 4401 } 4402 } 4403 if (identical > 0) { 4404 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4405 } else if (_expensive_nodes->length() >= 1) { 4406 Node* n = _expensive_nodes->at(i); 4407 igvn.replace_input_of(n, 0, NULL); 4408 igvn.hash_insert(n); 4409 modified = true; 4410 } 4411 _expensive_nodes->trunc_to(j); 4412 if (modified) { 4413 igvn.optimize(); 4414 } 4415 } 4416 4417 void Compile::add_expensive_node(Node * n) { 4418 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4419 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4420 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4421 if (OptimizeExpensiveOps) { 4422 _expensive_nodes->append(n); 4423 } else { 4424 // Clear control input and let IGVN optimize expensive nodes if 4425 // OptimizeExpensiveOps is off. 4426 n->set_req(0, NULL); 4427 } 4428 } 4429 4430 /** 4431 * Remove the speculative part of types and clean up the graph 4432 */ 4433 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4434 if (UseTypeSpeculation) { 4435 Unique_Node_List worklist; 4436 worklist.push(root()); 4437 int modified = 0; 4438 // Go over all type nodes that carry a speculative type, drop the 4439 // speculative part of the type and enqueue the node for an igvn 4440 // which may optimize it out. 4441 for (uint next = 0; next < worklist.size(); ++next) { 4442 Node *n = worklist.at(next); 4443 if (n->is_Type()) { 4444 TypeNode* tn = n->as_Type(); 4445 const Type* t = tn->type(); 4446 const Type* t_no_spec = t->remove_speculative(); 4447 if (t_no_spec != t) { 4448 bool in_hash = igvn.hash_delete(n); 4449 assert(in_hash, "node should be in igvn hash table"); 4450 tn->set_type(t_no_spec); 4451 igvn.hash_insert(n); 4452 igvn._worklist.push(n); // give it a chance to go away 4453 modified++; 4454 } 4455 } 4456 uint max = n->len(); 4457 for( uint i = 0; i < max; ++i ) { 4458 Node *m = n->in(i); 4459 if (not_a_node(m)) continue; 4460 worklist.push(m); 4461 } 4462 } 4463 // Drop the speculative part of all types in the igvn's type table 4464 igvn.remove_speculative_types(); 4465 if (modified > 0) { 4466 igvn.optimize(); 4467 } 4468 #ifdef ASSERT 4469 // Verify that after the IGVN is over no speculative type has resurfaced 4470 worklist.clear(); 4471 worklist.push(root()); 4472 for (uint next = 0; next < worklist.size(); ++next) { 4473 Node *n = worklist.at(next); 4474 const Type* t = igvn.type_or_null(n); 4475 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4476 if (n->is_Type()) { 4477 t = n->as_Type()->type(); 4478 assert(t == t->remove_speculative(), "no more speculative types"); 4479 } 4480 uint max = n->len(); 4481 for( uint i = 0; i < max; ++i ) { 4482 Node *m = n->in(i); 4483 if (not_a_node(m)) continue; 4484 worklist.push(m); 4485 } 4486 } 4487 igvn.check_no_speculative_types(); 4488 #endif 4489 } 4490 } 4491 4492 // Auxiliary method to support randomized stressing/fuzzing. 4493 // 4494 // This method can be called the arbitrary number of times, with current count 4495 // as the argument. The logic allows selecting a single candidate from the 4496 // running list of candidates as follows: 4497 // int count = 0; 4498 // Cand* selected = null; 4499 // while(cand = cand->next()) { 4500 // if (randomized_select(++count)) { 4501 // selected = cand; 4502 // } 4503 // } 4504 // 4505 // Including count equalizes the chances any candidate is "selected". 4506 // This is useful when we don't have the complete list of candidates to choose 4507 // from uniformly. In this case, we need to adjust the randomicity of the 4508 // selection, or else we will end up biasing the selection towards the latter 4509 // candidates. 4510 // 4511 // Quick back-envelope calculation shows that for the list of n candidates 4512 // the equal probability for the candidate to persist as "best" can be 4513 // achieved by replacing it with "next" k-th candidate with the probability 4514 // of 1/k. It can be easily shown that by the end of the run, the 4515 // probability for any candidate is converged to 1/n, thus giving the 4516 // uniform distribution among all the candidates. 4517 // 4518 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4519 #define RANDOMIZED_DOMAIN_POW 29 4520 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4521 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4522 bool Compile::randomized_select(int count) { 4523 assert(count > 0, "only positive"); 4524 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4525 } 4526 4527 CloneMap& Compile::clone_map() { return _clone_map; } 4528 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4529 4530 void NodeCloneInfo::dump() const { 4531 tty->print(" {%d:%d} ", idx(), gen()); 4532 } 4533 4534 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4535 uint64_t val = value(old->_idx); 4536 NodeCloneInfo cio(val); 4537 assert(val != 0, "old node should be in the map"); 4538 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4539 insert(nnn->_idx, cin.get()); 4540 #ifndef PRODUCT 4541 if (is_debug()) { 4542 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4543 } 4544 #endif 4545 } 4546 4547 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4548 NodeCloneInfo cio(value(old->_idx)); 4549 if (cio.get() == 0) { 4550 cio.set(old->_idx, 0); 4551 insert(old->_idx, cio.get()); 4552 #ifndef PRODUCT 4553 if (is_debug()) { 4554 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4555 } 4556 #endif 4557 } 4558 clone(old, nnn, gen); 4559 } 4560 4561 int CloneMap::max_gen() const { 4562 int g = 0; 4563 DictI di(_dict); 4564 for(; di.test(); ++di) { 4565 int t = gen(di._key); 4566 if (g < t) { 4567 g = t; 4568 #ifndef PRODUCT 4569 if (is_debug()) { 4570 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4571 } 4572 #endif 4573 } 4574 } 4575 return g; 4576 } 4577 4578 void CloneMap::dump(node_idx_t key) const { 4579 uint64_t val = value(key); 4580 if (val != 0) { 4581 NodeCloneInfo ni(val); 4582 ni.dump(); 4583 } 4584 }