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/valuetypenode.hpp" 69 #include "opto/vectornode.hpp" 70 #include "runtime/arguments.hpp" 71 #include "runtime/sharedRuntime.hpp" 72 #include "runtime/signature.hpp" 73 #include "runtime/stubRoutines.hpp" 74 #include "runtime/timer.hpp" 75 #include "utilities/copy.hpp" 76 77 78 // -------------------- Compile::mach_constant_base_node ----------------------- 79 // Constant table base node singleton. 80 MachConstantBaseNode* Compile::mach_constant_base_node() { 81 if (_mach_constant_base_node == NULL) { 82 _mach_constant_base_node = new MachConstantBaseNode(); 83 _mach_constant_base_node->add_req(C->root()); 84 } 85 return _mach_constant_base_node; 86 } 87 88 89 /// Support for intrinsics. 90 91 // Return the index at which m must be inserted (or already exists). 92 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 93 class IntrinsicDescPair { 94 private: 95 ciMethod* _m; 96 bool _is_virtual; 97 public: 98 IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {} 99 static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) { 100 ciMethod* m= elt->method(); 101 ciMethod* key_m = key->_m; 102 if (key_m < m) return -1; 103 else if (key_m > m) return 1; 104 else { 105 bool is_virtual = elt->is_virtual(); 106 bool key_virtual = key->_is_virtual; 107 if (key_virtual < is_virtual) return -1; 108 else if (key_virtual > is_virtual) return 1; 109 else return 0; 110 } 111 } 112 }; 113 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) { 114 #ifdef ASSERT 115 for (int i = 1; i < _intrinsics->length(); i++) { 116 CallGenerator* cg1 = _intrinsics->at(i-1); 117 CallGenerator* cg2 = _intrinsics->at(i); 118 assert(cg1->method() != cg2->method() 119 ? cg1->method() < cg2->method() 120 : cg1->is_virtual() < cg2->is_virtual(), 121 "compiler intrinsics list must stay sorted"); 122 } 123 #endif 124 IntrinsicDescPair pair(m, is_virtual); 125 return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found); 126 } 127 128 void Compile::register_intrinsic(CallGenerator* cg) { 129 if (_intrinsics == NULL) { 130 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 131 } 132 int len = _intrinsics->length(); 133 bool found = false; 134 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found); 135 assert(!found, "registering twice"); 136 _intrinsics->insert_before(index, cg); 137 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 138 } 139 140 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 141 assert(m->is_loaded(), "don't try this on unloaded methods"); 142 if (_intrinsics != NULL) { 143 bool found = false; 144 int index = intrinsic_insertion_index(m, is_virtual, found); 145 if (found) { 146 return _intrinsics->at(index); 147 } 148 } 149 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 150 if (m->intrinsic_id() != vmIntrinsics::_none && 151 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 152 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 153 if (cg != NULL) { 154 // Save it for next time: 155 register_intrinsic(cg); 156 return cg; 157 } else { 158 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 159 } 160 } 161 return NULL; 162 } 163 164 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 165 // in library_call.cpp. 166 167 168 #ifndef PRODUCT 169 // statistics gathering... 170 171 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 172 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 173 174 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 175 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 176 int oflags = _intrinsic_hist_flags[id]; 177 assert(flags != 0, "what happened?"); 178 if (is_virtual) { 179 flags |= _intrinsic_virtual; 180 } 181 bool changed = (flags != oflags); 182 if ((flags & _intrinsic_worked) != 0) { 183 juint count = (_intrinsic_hist_count[id] += 1); 184 if (count == 1) { 185 changed = true; // first time 186 } 187 // increment the overall count also: 188 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 189 } 190 if (changed) { 191 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 192 // Something changed about the intrinsic's virtuality. 193 if ((flags & _intrinsic_virtual) != 0) { 194 // This is the first use of this intrinsic as a virtual call. 195 if (oflags != 0) { 196 // We already saw it as a non-virtual, so note both cases. 197 flags |= _intrinsic_both; 198 } 199 } else if ((oflags & _intrinsic_both) == 0) { 200 // This is the first use of this intrinsic as a non-virtual 201 flags |= _intrinsic_both; 202 } 203 } 204 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 205 } 206 // update the overall flags also: 207 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 208 return changed; 209 } 210 211 static char* format_flags(int flags, char* buf) { 212 buf[0] = 0; 213 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 214 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 215 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 216 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 217 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 218 if (buf[0] == 0) strcat(buf, ","); 219 assert(buf[0] == ',', "must be"); 220 return &buf[1]; 221 } 222 223 void Compile::print_intrinsic_statistics() { 224 char flagsbuf[100]; 225 ttyLocker ttyl; 226 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 227 tty->print_cr("Compiler intrinsic usage:"); 228 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 229 if (total == 0) total = 1; // avoid div0 in case of no successes 230 #define PRINT_STAT_LINE(name, c, f) \ 231 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 232 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 233 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 234 int flags = _intrinsic_hist_flags[id]; 235 juint count = _intrinsic_hist_count[id]; 236 if ((flags | count) != 0) { 237 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 238 } 239 } 240 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 241 if (xtty != NULL) xtty->tail("statistics"); 242 } 243 244 void Compile::print_statistics() { 245 { ttyLocker ttyl; 246 if (xtty != NULL) xtty->head("statistics type='opto'"); 247 Parse::print_statistics(); 248 PhaseCCP::print_statistics(); 249 PhaseRegAlloc::print_statistics(); 250 Scheduling::print_statistics(); 251 PhasePeephole::print_statistics(); 252 PhaseIdealLoop::print_statistics(); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 256 // put this under its own <statistics> element. 257 print_intrinsic_statistics(); 258 } 259 } 260 #endif //PRODUCT 261 262 // Support for bundling info 263 Bundle* Compile::node_bundling(const Node *n) { 264 assert(valid_bundle_info(n), "oob"); 265 return &_node_bundling_base[n->_idx]; 266 } 267 268 bool Compile::valid_bundle_info(const Node *n) { 269 return (_node_bundling_limit > n->_idx); 270 } 271 272 273 void Compile::gvn_replace_by(Node* n, Node* nn) { 274 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 275 Node* use = n->last_out(i); 276 bool is_in_table = initial_gvn()->hash_delete(use); 277 uint uses_found = 0; 278 for (uint j = 0; j < use->len(); j++) { 279 if (use->in(j) == n) { 280 if (j < use->req()) 281 use->set_req(j, nn); 282 else 283 use->set_prec(j, nn); 284 uses_found++; 285 } 286 } 287 if (is_in_table) { 288 // reinsert into table 289 initial_gvn()->hash_find_insert(use); 290 } 291 record_for_igvn(use); 292 i -= uses_found; // we deleted 1 or more copies of this edge 293 } 294 } 295 296 297 static inline bool not_a_node(const Node* n) { 298 if (n == NULL) return true; 299 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 300 if (*(address*)n == badAddress) return true; // kill by Node::destruct 301 return false; 302 } 303 304 // Identify all nodes that are reachable from below, useful. 305 // Use breadth-first pass that records state in a Unique_Node_List, 306 // recursive traversal is slower. 307 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 308 int estimated_worklist_size = live_nodes(); 309 useful.map( estimated_worklist_size, NULL ); // preallocate space 310 311 // Initialize worklist 312 if (root() != NULL) { useful.push(root()); } 313 // If 'top' is cached, declare it useful to preserve cached node 314 if( cached_top_node() ) { useful.push(cached_top_node()); } 315 316 // Push all useful nodes onto the list, breadthfirst 317 for( uint next = 0; next < useful.size(); ++next ) { 318 assert( next < unique(), "Unique useful nodes < total nodes"); 319 Node *n = useful.at(next); 320 uint max = n->len(); 321 for( uint i = 0; i < max; ++i ) { 322 Node *m = n->in(i); 323 if (not_a_node(m)) continue; 324 useful.push(m); 325 } 326 } 327 } 328 329 // Update dead_node_list with any missing dead nodes using useful 330 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 331 void Compile::update_dead_node_list(Unique_Node_List &useful) { 332 uint max_idx = unique(); 333 VectorSet& useful_node_set = useful.member_set(); 334 335 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 336 // If node with index node_idx is not in useful set, 337 // mark it as dead in dead node list. 338 if (! useful_node_set.test(node_idx) ) { 339 record_dead_node(node_idx); 340 } 341 } 342 } 343 344 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 345 int shift = 0; 346 for (int i = 0; i < inlines->length(); i++) { 347 CallGenerator* cg = inlines->at(i); 348 CallNode* call = cg->call_node(); 349 if (shift > 0) { 350 inlines->at_put(i-shift, cg); 351 } 352 if (!useful.member(call)) { 353 shift++; 354 } 355 } 356 inlines->trunc_to(inlines->length()-shift); 357 } 358 359 // Disconnect all useless nodes by disconnecting those at the boundary. 360 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 361 uint next = 0; 362 while (next < useful.size()) { 363 Node *n = useful.at(next++); 364 if (n->is_SafePoint()) { 365 // We're done with a parsing phase. Replaced nodes are not valid 366 // beyond that point. 367 n->as_SafePoint()->delete_replaced_nodes(); 368 } 369 // Use raw traversal of out edges since this code removes out edges 370 int max = n->outcnt(); 371 for (int j = 0; j < max; ++j) { 372 Node* child = n->raw_out(j); 373 if (! useful.member(child)) { 374 assert(!child->is_top() || child != top(), 375 "If top is cached in Compile object it is in useful list"); 376 // Only need to remove this out-edge to the useless node 377 n->raw_del_out(j); 378 --j; 379 --max; 380 } 381 } 382 if (n->outcnt() == 1 && n->has_special_unique_user()) { 383 record_for_igvn(n->unique_out()); 384 } 385 } 386 // Remove useless macro and predicate opaq nodes 387 for (int i = C->macro_count()-1; i >= 0; i--) { 388 Node* n = C->macro_node(i); 389 if (!useful.member(n)) { 390 remove_macro_node(n); 391 } 392 } 393 // Remove useless CastII nodes with range check dependency 394 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 395 Node* cast = range_check_cast_node(i); 396 if (!useful.member(cast)) { 397 remove_range_check_cast(cast); 398 } 399 } 400 // Remove useless expensive node 401 for (int i = C->expensive_count()-1; i >= 0; i--) { 402 Node* n = C->expensive_node(i); 403 if (!useful.member(n)) { 404 remove_expensive_node(n); 405 } 406 } 407 for (int i = value_type_ptr_count() - 1; i >= 0; i--) { 408 ValueTypePtrNode* vtptr = value_type_ptr(i); 409 if (!useful.member(vtptr)) { 410 remove_value_type_ptr(vtptr); 411 } 412 } 413 // clean up the late inline lists 414 remove_useless_late_inlines(&_string_late_inlines, useful); 415 remove_useless_late_inlines(&_boxing_late_inlines, useful); 416 remove_useless_late_inlines(&_late_inlines, useful); 417 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 418 } 419 420 //------------------------------frame_size_in_words----------------------------- 421 // frame_slots in units of words 422 int Compile::frame_size_in_words() const { 423 // shift is 0 in LP32 and 1 in LP64 424 const int shift = (LogBytesPerWord - LogBytesPerInt); 425 int words = _frame_slots >> shift; 426 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 427 return words; 428 } 429 430 // To bang the stack of this compiled method we use the stack size 431 // that the interpreter would need in case of a deoptimization. This 432 // removes the need to bang the stack in the deoptimization blob which 433 // in turn simplifies stack overflow handling. 434 int Compile::bang_size_in_bytes() const { 435 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 436 } 437 438 // ============================================================================ 439 //------------------------------CompileWrapper--------------------------------- 440 class CompileWrapper : public StackObj { 441 Compile *const _compile; 442 public: 443 CompileWrapper(Compile* compile); 444 445 ~CompileWrapper(); 446 }; 447 448 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 449 // the Compile* pointer is stored in the current ciEnv: 450 ciEnv* env = compile->env(); 451 assert(env == ciEnv::current(), "must already be a ciEnv active"); 452 assert(env->compiler_data() == NULL, "compile already active?"); 453 env->set_compiler_data(compile); 454 assert(compile == Compile::current(), "sanity"); 455 456 compile->set_type_dict(NULL); 457 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 458 compile->clone_map().set_clone_idx(0); 459 compile->set_type_hwm(NULL); 460 compile->set_type_last_size(0); 461 compile->set_last_tf(NULL, NULL); 462 compile->set_indexSet_arena(NULL); 463 compile->set_indexSet_free_block_list(NULL); 464 compile->init_type_arena(); 465 Type::Initialize(compile); 466 _compile->set_scratch_buffer_blob(NULL); 467 _compile->begin_method(); 468 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 469 } 470 CompileWrapper::~CompileWrapper() { 471 _compile->end_method(); 472 if (_compile->scratch_buffer_blob() != NULL) 473 BufferBlob::free(_compile->scratch_buffer_blob()); 474 _compile->env()->set_compiler_data(NULL); 475 } 476 477 478 //----------------------------print_compile_messages--------------------------- 479 void Compile::print_compile_messages() { 480 #ifndef PRODUCT 481 // Check if recompiling 482 if (_subsume_loads == false && PrintOpto) { 483 // Recompiling without allowing machine instructions to subsume loads 484 tty->print_cr("*********************************************************"); 485 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 486 tty->print_cr("*********************************************************"); 487 } 488 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 489 // Recompiling without escape analysis 490 tty->print_cr("*********************************************************"); 491 tty->print_cr("** Bailout: Recompile without escape analysis **"); 492 tty->print_cr("*********************************************************"); 493 } 494 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 495 // Recompiling without boxing elimination 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (C->directive()->BreakAtCompileOption) { 501 // Open the debugger when compiling this method. 502 tty->print("### Breaking when compiling: "); 503 method()->print_short_name(); 504 tty->cr(); 505 BREAKPOINT; 506 } 507 508 if( PrintOpto ) { 509 if (is_osr_compilation()) { 510 tty->print("[OSR]%3d", _compile_id); 511 } else { 512 tty->print("%3d", _compile_id); 513 } 514 } 515 #endif 516 } 517 518 519 //-----------------------init_scratch_buffer_blob------------------------------ 520 // Construct a temporary BufferBlob and cache it for this compile. 521 void Compile::init_scratch_buffer_blob(int const_size) { 522 // If there is already a scratch buffer blob allocated and the 523 // constant section is big enough, use it. Otherwise free the 524 // current and allocate a new one. 525 BufferBlob* blob = scratch_buffer_blob(); 526 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 527 // Use the current blob. 528 } else { 529 if (blob != NULL) { 530 BufferBlob::free(blob); 531 } 532 533 ResourceMark rm; 534 _scratch_const_size = const_size; 535 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 536 blob = BufferBlob::create("Compile::scratch_buffer", size); 537 // Record the buffer blob for next time. 538 set_scratch_buffer_blob(blob); 539 // Have we run out of code space? 540 if (scratch_buffer_blob() == NULL) { 541 // Let CompilerBroker disable further compilations. 542 record_failure("Not enough space for scratch buffer in CodeCache"); 543 return; 544 } 545 } 546 547 // Initialize the relocation buffers 548 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 549 set_scratch_locs_memory(locs_buf); 550 } 551 552 553 //-----------------------scratch_emit_size------------------------------------- 554 // Helper function that computes size by emitting code 555 uint Compile::scratch_emit_size(const Node* n) { 556 // Start scratch_emit_size section. 557 set_in_scratch_emit_size(true); 558 559 // Emit into a trash buffer and count bytes emitted. 560 // This is a pretty expensive way to compute a size, 561 // but it works well enough if seldom used. 562 // All common fixed-size instructions are given a size 563 // method by the AD file. 564 // Note that the scratch buffer blob and locs memory are 565 // allocated at the beginning of the compile task, and 566 // may be shared by several calls to scratch_emit_size. 567 // The allocation of the scratch buffer blob is particularly 568 // expensive, since it has to grab the code cache lock. 569 BufferBlob* blob = this->scratch_buffer_blob(); 570 assert(blob != NULL, "Initialize BufferBlob at start"); 571 assert(blob->size() > MAX_inst_size, "sanity"); 572 relocInfo* locs_buf = scratch_locs_memory(); 573 address blob_begin = blob->content_begin(); 574 address blob_end = (address)locs_buf; 575 assert(blob->contains(blob_end), "sanity"); 576 CodeBuffer buf(blob_begin, blob_end - blob_begin); 577 buf.initialize_consts_size(_scratch_const_size); 578 buf.initialize_stubs_size(MAX_stubs_size); 579 assert(locs_buf != NULL, "sanity"); 580 int lsize = MAX_locs_size / 3; 581 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 582 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 583 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 584 // Mark as scratch buffer. 585 buf.consts()->set_scratch_emit(); 586 buf.insts()->set_scratch_emit(); 587 buf.stubs()->set_scratch_emit(); 588 589 // Do the emission. 590 591 Label fakeL; // Fake label for branch instructions. 592 Label* saveL = NULL; 593 uint save_bnum = 0; 594 bool is_branch = n->is_MachBranch(); 595 if (is_branch) { 596 MacroAssembler masm(&buf); 597 masm.bind(fakeL); 598 n->as_MachBranch()->save_label(&saveL, &save_bnum); 599 n->as_MachBranch()->label_set(&fakeL, 0); 600 } 601 n->emit(buf, this->regalloc()); 602 603 // Emitting into the scratch buffer should not fail 604 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 605 606 if (is_branch) // Restore label. 607 n->as_MachBranch()->label_set(saveL, save_bnum); 608 609 // End scratch_emit_size section. 610 set_in_scratch_emit_size(false); 611 612 return buf.insts_size(); 613 } 614 615 616 // ============================================================================ 617 //------------------------------Compile standard------------------------------- 618 debug_only( int Compile::_debug_idx = 100000; ) 619 620 // Compile a method. entry_bci is -1 for normal compilations and indicates 621 // the continuation bci for on stack replacement. 622 623 624 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 625 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 626 : Phase(Compiler), 627 _env(ci_env), 628 _directive(directive), 629 _log(ci_env->log()), 630 _compile_id(ci_env->compile_id()), 631 _save_argument_registers(false), 632 _stub_name(NULL), 633 _stub_function(NULL), 634 _stub_entry_point(NULL), 635 _method(target), 636 _entry_bci(osr_bci), 637 _initial_gvn(NULL), 638 _for_igvn(NULL), 639 _warm_calls(NULL), 640 _subsume_loads(subsume_loads), 641 _do_escape_analysis(do_escape_analysis), 642 _eliminate_boxing(eliminate_boxing), 643 _failure_reason(NULL), 644 _code_buffer("Compile::Fill_buffer"), 645 _orig_pc_slot(0), 646 _orig_pc_slot_offset_in_bytes(0), 647 _has_method_handle_invokes(false), 648 _mach_constant_base_node(NULL), 649 _node_bundling_limit(0), 650 _node_bundling_base(NULL), 651 _java_calls(0), 652 _inner_loops(0), 653 _scratch_const_size(-1), 654 _in_scratch_emit_size(false), 655 _dead_node_list(comp_arena()), 656 _dead_node_count(0), 657 #ifndef PRODUCT 658 _trace_opto_output(directive->TraceOptoOutputOption), 659 _in_dump_cnt(0), 660 _printer(IdealGraphPrinter::printer()), 661 #endif 662 _congraph(NULL), 663 _comp_arena(mtCompiler), 664 _node_arena(mtCompiler), 665 _old_arena(mtCompiler), 666 _Compile_types(mtCompiler), 667 _replay_inline_data(NULL), 668 _late_inlines(comp_arena(), 2, 0, NULL), 669 _string_late_inlines(comp_arena(), 2, 0, NULL), 670 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 671 _late_inlines_pos(0), 672 _number_of_mh_late_inlines(0), 673 _inlining_progress(false), 674 _inlining_incrementally(false), 675 _print_inlining_list(NULL), 676 _print_inlining_stream(NULL), 677 _print_inlining_idx(0), 678 _print_inlining_output(NULL), 679 _interpreter_frame_size(0), 680 _max_node_limit(MaxNodeLimit), 681 _has_reserved_stack_access(target->has_reserved_stack_access()) { 682 C = this; 683 #ifndef PRODUCT 684 if (_printer != NULL) { 685 _printer->set_compile(this); 686 } 687 #endif 688 CompileWrapper cw(this); 689 690 if (CITimeVerbose) { 691 tty->print(" "); 692 target->holder()->name()->print(); 693 tty->print("."); 694 target->print_short_name(); 695 tty->print(" "); 696 } 697 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 698 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 699 700 #ifndef PRODUCT 701 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 702 if (!print_opto_assembly) { 703 bool print_assembly = directive->PrintAssemblyOption; 704 if (print_assembly && !Disassembler::can_decode()) { 705 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 706 print_opto_assembly = true; 707 } 708 } 709 set_print_assembly(print_opto_assembly); 710 set_parsed_irreducible_loop(false); 711 712 if (directive->ReplayInlineOption) { 713 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 714 } 715 #endif 716 set_print_inlining(directive->PrintInliningOption || PrintOptoInlining); 717 set_print_intrinsics(directive->PrintIntrinsicsOption); 718 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 719 720 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 721 // Make sure the method being compiled gets its own MDO, 722 // so we can at least track the decompile_count(). 723 // Need MDO to record RTM code generation state. 724 method()->ensure_method_data(); 725 } 726 727 Init(::AliasLevel); 728 729 730 print_compile_messages(); 731 732 _ilt = InlineTree::build_inline_tree_root(); 733 734 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 735 assert(num_alias_types() >= AliasIdxRaw, ""); 736 737 #define MINIMUM_NODE_HASH 1023 738 // Node list that Iterative GVN will start with 739 Unique_Node_List for_igvn(comp_arena()); 740 set_for_igvn(&for_igvn); 741 742 // GVN that will be run immediately on new nodes 743 uint estimated_size = method()->code_size()*4+64; 744 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 745 PhaseGVN gvn(node_arena(), estimated_size); 746 set_initial_gvn(&gvn); 747 748 print_inlining_init(); 749 { // Scope for timing the parser 750 TracePhase tp("parse", &timers[_t_parser]); 751 752 // Put top into the hash table ASAP. 753 initial_gvn()->transform_no_reclaim(top()); 754 755 // Set up tf(), start(), and find a CallGenerator. 756 CallGenerator* cg = NULL; 757 if (is_osr_compilation()) { 758 const TypeTuple *domain = StartOSRNode::osr_domain(); 759 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 760 init_tf(TypeFunc::make(domain, range)); 761 StartNode* s = new StartOSRNode(root(), domain); 762 initial_gvn()->set_type_bottom(s); 763 init_start(s); 764 cg = CallGenerator::for_osr(method(), entry_bci()); 765 } else { 766 // Normal case. 767 init_tf(TypeFunc::make(method())); 768 StartNode* s = new StartNode(root(), tf()->domain_cc()); 769 initial_gvn()->set_type_bottom(s); 770 init_start(s); 771 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 772 // With java.lang.ref.reference.get() we must go through the 773 // intrinsic when G1 is enabled - even when get() is the root 774 // method of the compile - so that, if necessary, the value in 775 // the referent field of the reference object gets recorded by 776 // the pre-barrier code. 777 // Specifically, if G1 is enabled, the value in the referent 778 // field is recorded by the G1 SATB pre barrier. This will 779 // result in the referent being marked live and the reference 780 // object removed from the list of discovered references during 781 // reference processing. 782 cg = find_intrinsic(method(), false); 783 } 784 if (cg == NULL) { 785 float past_uses = method()->interpreter_invocation_count(); 786 float expected_uses = past_uses; 787 cg = CallGenerator::for_inline(method(), expected_uses); 788 } 789 } 790 if (failing()) return; 791 if (cg == NULL) { 792 record_method_not_compilable("cannot parse method"); 793 return; 794 } 795 JVMState* jvms = build_start_state(start(), tf()); 796 if ((jvms = cg->generate(jvms)) == NULL) { 797 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 798 record_method_not_compilable("method parse failed"); 799 } 800 return; 801 } 802 GraphKit kit(jvms); 803 804 if (!kit.stopped()) { 805 // Accept return values, and transfer control we know not where. 806 // This is done by a special, unique ReturnNode bound to root. 807 return_values(kit.jvms()); 808 } 809 810 if (kit.has_exceptions()) { 811 // Any exceptions that escape from this call must be rethrown 812 // to whatever caller is dynamically above us on the stack. 813 // This is done by a special, unique RethrowNode bound to root. 814 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 815 } 816 817 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 818 819 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 820 inline_string_calls(true); 821 } 822 823 if (failing()) return; 824 825 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 826 827 // Remove clutter produced by parsing. 828 if (!failing()) { 829 ResourceMark rm; 830 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 831 } 832 } 833 834 // Note: Large methods are capped off in do_one_bytecode(). 835 if (failing()) return; 836 837 // After parsing, node notes are no longer automagic. 838 // They must be propagated by register_new_node_with_optimizer(), 839 // clone(), or the like. 840 set_default_node_notes(NULL); 841 842 for (;;) { 843 int successes = Inline_Warm(); 844 if (failing()) return; 845 if (successes == 0) break; 846 } 847 848 // Drain the list. 849 Finish_Warm(); 850 #ifndef PRODUCT 851 if (_printer && _printer->should_print(1)) { 852 _printer->print_inlining(); 853 } 854 #endif 855 856 if (failing()) return; 857 NOT_PRODUCT( verify_graph_edges(); ) 858 859 // Now optimize 860 Optimize(); 861 if (failing()) return; 862 NOT_PRODUCT( verify_graph_edges(); ) 863 864 #ifndef PRODUCT 865 if (PrintIdeal) { 866 ttyLocker ttyl; // keep the following output all in one block 867 // This output goes directly to the tty, not the compiler log. 868 // To enable tools to match it up with the compilation activity, 869 // be sure to tag this tty output with the compile ID. 870 if (xtty != NULL) { 871 xtty->head("ideal compile_id='%d'%s", compile_id(), 872 is_osr_compilation() ? " compile_kind='osr'" : 873 ""); 874 } 875 root()->dump(9999); 876 if (xtty != NULL) { 877 xtty->tail("ideal"); 878 } 879 } 880 #endif 881 882 NOT_PRODUCT( verify_barriers(); ) 883 884 // Dump compilation data to replay it. 885 if (directive->DumpReplayOption) { 886 env()->dump_replay_data(_compile_id); 887 } 888 if (directive->DumpInlineOption && (ilt() != NULL)) { 889 env()->dump_inline_data(_compile_id); 890 } 891 892 // Now that we know the size of all the monitors we can add a fixed slot 893 // for the original deopt pc. 894 895 _orig_pc_slot = fixed_slots(); 896 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 897 set_fixed_slots(next_slot); 898 899 // Compute when to use implicit null checks. Used by matching trap based 900 // nodes and NullCheck optimization. 901 set_allowed_deopt_reasons(); 902 903 // Now generate code 904 Code_Gen(); 905 if (failing()) return; 906 907 // Check if we want to skip execution of all compiled code. 908 { 909 #ifndef PRODUCT 910 if (OptoNoExecute) { 911 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 912 return; 913 } 914 #endif 915 TracePhase tp("install_code", &timers[_t_registerMethod]); 916 917 if (is_osr_compilation()) { 918 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 919 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 920 } else { 921 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 922 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 923 } 924 925 env()->register_method(_method, _entry_bci, 926 &_code_offsets, 927 _orig_pc_slot_offset_in_bytes, 928 code_buffer(), 929 frame_size_in_words(), _oop_map_set, 930 &_handler_table, &_inc_table, 931 compiler, 932 has_unsafe_access(), 933 SharedRuntime::is_wide_vector(max_vector_size()), 934 rtm_state() 935 ); 936 937 if (log() != NULL) // Print code cache state into compiler log 938 log()->code_cache_state(); 939 } 940 } 941 942 //------------------------------Compile---------------------------------------- 943 // Compile a runtime stub 944 Compile::Compile( ciEnv* ci_env, 945 TypeFunc_generator generator, 946 address stub_function, 947 const char *stub_name, 948 int is_fancy_jump, 949 bool pass_tls, 950 bool save_arg_registers, 951 bool return_pc, 952 DirectiveSet* directive) 953 : Phase(Compiler), 954 _env(ci_env), 955 _directive(directive), 956 _log(ci_env->log()), 957 _compile_id(0), 958 _save_argument_registers(save_arg_registers), 959 _method(NULL), 960 _stub_name(stub_name), 961 _stub_function(stub_function), 962 _stub_entry_point(NULL), 963 _entry_bci(InvocationEntryBci), 964 _initial_gvn(NULL), 965 _for_igvn(NULL), 966 _warm_calls(NULL), 967 _orig_pc_slot(0), 968 _orig_pc_slot_offset_in_bytes(0), 969 _subsume_loads(true), 970 _do_escape_analysis(false), 971 _eliminate_boxing(false), 972 _failure_reason(NULL), 973 _code_buffer("Compile::Fill_buffer"), 974 _has_method_handle_invokes(false), 975 _mach_constant_base_node(NULL), 976 _node_bundling_limit(0), 977 _node_bundling_base(NULL), 978 _java_calls(0), 979 _inner_loops(0), 980 #ifndef PRODUCT 981 _trace_opto_output(directive->TraceOptoOutputOption), 982 _in_dump_cnt(0), 983 _printer(NULL), 984 #endif 985 _comp_arena(mtCompiler), 986 _node_arena(mtCompiler), 987 _old_arena(mtCompiler), 988 _Compile_types(mtCompiler), 989 _dead_node_list(comp_arena()), 990 _dead_node_count(0), 991 _congraph(NULL), 992 _replay_inline_data(NULL), 993 _number_of_mh_late_inlines(0), 994 _inlining_progress(false), 995 _inlining_incrementally(false), 996 _print_inlining_list(NULL), 997 _print_inlining_stream(NULL), 998 _print_inlining_idx(0), 999 _print_inlining_output(NULL), 1000 _allowed_reasons(0), 1001 _interpreter_frame_size(0), 1002 _max_node_limit(MaxNodeLimit) { 1003 C = this; 1004 1005 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1006 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1007 1008 #ifndef PRODUCT 1009 set_print_assembly(PrintFrameConverterAssembly); 1010 set_parsed_irreducible_loop(false); 1011 #endif 1012 set_has_irreducible_loop(false); // no loops 1013 1014 CompileWrapper cw(this); 1015 Init(/*AliasLevel=*/ 0); 1016 init_tf((*generator)()); 1017 1018 { 1019 // The following is a dummy for the sake of GraphKit::gen_stub 1020 Unique_Node_List for_igvn(comp_arena()); 1021 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1022 PhaseGVN gvn(Thread::current()->resource_area(),255); 1023 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1024 gvn.transform_no_reclaim(top()); 1025 1026 GraphKit kit; 1027 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1028 } 1029 1030 NOT_PRODUCT( verify_graph_edges(); ) 1031 Code_Gen(); 1032 if (failing()) return; 1033 1034 1035 // Entry point will be accessed using compile->stub_entry_point(); 1036 if (code_buffer() == NULL) { 1037 Matcher::soft_match_failure(); 1038 } else { 1039 if (PrintAssembly && (WizardMode || Verbose)) 1040 tty->print_cr("### Stub::%s", stub_name); 1041 1042 if (!failing()) { 1043 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1044 1045 // Make the NMethod 1046 // For now we mark the frame as never safe for profile stackwalking 1047 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1048 code_buffer(), 1049 CodeOffsets::frame_never_safe, 1050 // _code_offsets.value(CodeOffsets::Frame_Complete), 1051 frame_size_in_words(), 1052 _oop_map_set, 1053 save_arg_registers); 1054 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1055 1056 _stub_entry_point = rs->entry_point(); 1057 } 1058 } 1059 } 1060 1061 //------------------------------Init------------------------------------------- 1062 // Prepare for a single compilation 1063 void Compile::Init(int aliaslevel) { 1064 _unique = 0; 1065 _regalloc = NULL; 1066 1067 _tf = NULL; // filled in later 1068 _top = NULL; // cached later 1069 _matcher = NULL; // filled in later 1070 _cfg = NULL; // filled in later 1071 1072 set_24_bit_selection_and_mode(Use24BitFP, false); 1073 1074 _node_note_array = NULL; 1075 _default_node_notes = NULL; 1076 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1077 1078 _immutable_memory = NULL; // filled in at first inquiry 1079 1080 // Globally visible Nodes 1081 // First set TOP to NULL to give safe behavior during creation of RootNode 1082 set_cached_top_node(NULL); 1083 set_root(new RootNode()); 1084 // Now that you have a Root to point to, create the real TOP 1085 set_cached_top_node( new ConNode(Type::TOP) ); 1086 set_recent_alloc(NULL, NULL); 1087 1088 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1089 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1090 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1091 env()->set_dependencies(new Dependencies(env())); 1092 1093 _fixed_slots = 0; 1094 set_has_split_ifs(false); 1095 set_has_loops(has_method() && method()->has_loops()); // first approximation 1096 set_has_stringbuilder(false); 1097 set_has_boxed_value(false); 1098 _trap_can_recompile = false; // no traps emitted yet 1099 _major_progress = true; // start out assuming good things will happen 1100 set_has_unsafe_access(false); 1101 set_max_vector_size(0); 1102 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1103 set_decompile_count(0); 1104 1105 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1106 set_num_loop_opts(LoopOptsCount); 1107 set_do_inlining(Inline); 1108 set_max_inline_size(MaxInlineSize); 1109 set_freq_inline_size(FreqInlineSize); 1110 set_do_scheduling(OptoScheduling); 1111 set_do_count_invocations(false); 1112 set_do_method_data_update(false); 1113 1114 set_do_vector_loop(false); 1115 1116 if (AllowVectorizeOnDemand) { 1117 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1118 set_do_vector_loop(true); 1119 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());}) 1120 } else if (has_method() && method()->name() != 0 && 1121 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1122 set_do_vector_loop(true); 1123 } 1124 } 1125 set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally 1126 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());}) 1127 1128 set_age_code(has_method() && method()->profile_aging()); 1129 set_rtm_state(NoRTM); // No RTM lock eliding by default 1130 _max_node_limit = _directive->MaxNodeLimitOption; 1131 1132 #if INCLUDE_RTM_OPT 1133 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1134 int rtm_state = method()->method_data()->rtm_state(); 1135 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1136 // Don't generate RTM lock eliding code. 1137 set_rtm_state(NoRTM); 1138 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1139 // Generate RTM lock eliding code without abort ratio calculation code. 1140 set_rtm_state(UseRTM); 1141 } else if (UseRTMDeopt) { 1142 // Generate RTM lock eliding code and include abort ratio calculation 1143 // code if UseRTMDeopt is on. 1144 set_rtm_state(ProfileRTM); 1145 } 1146 } 1147 #endif 1148 if (debug_info()->recording_non_safepoints()) { 1149 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1150 (comp_arena(), 8, 0, NULL)); 1151 set_default_node_notes(Node_Notes::make(this)); 1152 } 1153 1154 // // -- Initialize types before each compile -- 1155 // // Update cached type information 1156 // if( _method && _method->constants() ) 1157 // Type::update_loaded_types(_method, _method->constants()); 1158 1159 // Init alias_type map. 1160 if (!_do_escape_analysis && aliaslevel == 3) 1161 aliaslevel = 2; // No unique types without escape analysis 1162 _AliasLevel = aliaslevel; 1163 const int grow_ats = 16; 1164 _max_alias_types = grow_ats; 1165 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1166 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1167 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1168 { 1169 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1170 } 1171 // Initialize the first few types. 1172 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1173 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1174 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1175 _num_alias_types = AliasIdxRaw+1; 1176 // Zero out the alias type cache. 1177 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1178 // A NULL adr_type hits in the cache right away. Preload the right answer. 1179 probe_alias_cache(NULL)->_index = AliasIdxTop; 1180 1181 _intrinsics = NULL; 1182 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1183 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1184 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1185 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1186 _value_type_ptr_nodes = new(comp_arena()) GrowableArray<ValueTypePtrNode*>(comp_arena(), 8, 0, NULL); 1187 register_library_intrinsics(); 1188 } 1189 1190 //---------------------------init_start---------------------------------------- 1191 // Install the StartNode on this compile object. 1192 void Compile::init_start(StartNode* s) { 1193 if (failing()) 1194 return; // already failing 1195 assert(s == start(), ""); 1196 } 1197 1198 /** 1199 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1200 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1201 * the ideal graph. 1202 */ 1203 StartNode* Compile::start() const { 1204 assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason()); 1205 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1206 Node* start = root()->fast_out(i); 1207 if (start->is_Start()) { 1208 return start->as_Start(); 1209 } 1210 } 1211 fatal("Did not find Start node!"); 1212 return NULL; 1213 } 1214 1215 //-------------------------------immutable_memory------------------------------------- 1216 // Access immutable memory 1217 Node* Compile::immutable_memory() { 1218 if (_immutable_memory != NULL) { 1219 return _immutable_memory; 1220 } 1221 StartNode* s = start(); 1222 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1223 Node *p = s->fast_out(i); 1224 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1225 _immutable_memory = p; 1226 return _immutable_memory; 1227 } 1228 } 1229 ShouldNotReachHere(); 1230 return NULL; 1231 } 1232 1233 //----------------------set_cached_top_node------------------------------------ 1234 // Install the cached top node, and make sure Node::is_top works correctly. 1235 void Compile::set_cached_top_node(Node* tn) { 1236 if (tn != NULL) verify_top(tn); 1237 Node* old_top = _top; 1238 _top = tn; 1239 // Calling Node::setup_is_top allows the nodes the chance to adjust 1240 // their _out arrays. 1241 if (_top != NULL) _top->setup_is_top(); 1242 if (old_top != NULL) old_top->setup_is_top(); 1243 assert(_top == NULL || top()->is_top(), ""); 1244 } 1245 1246 #ifdef ASSERT 1247 uint Compile::count_live_nodes_by_graph_walk() { 1248 Unique_Node_List useful(comp_arena()); 1249 // Get useful node list by walking the graph. 1250 identify_useful_nodes(useful); 1251 return useful.size(); 1252 } 1253 1254 void Compile::print_missing_nodes() { 1255 1256 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1257 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1258 return; 1259 } 1260 1261 // This is an expensive function. It is executed only when the user 1262 // specifies VerifyIdealNodeCount option or otherwise knows the 1263 // additional work that needs to be done to identify reachable nodes 1264 // by walking the flow graph and find the missing ones using 1265 // _dead_node_list. 1266 1267 Unique_Node_List useful(comp_arena()); 1268 // Get useful node list by walking the graph. 1269 identify_useful_nodes(useful); 1270 1271 uint l_nodes = C->live_nodes(); 1272 uint l_nodes_by_walk = useful.size(); 1273 1274 if (l_nodes != l_nodes_by_walk) { 1275 if (_log != NULL) { 1276 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1277 _log->stamp(); 1278 _log->end_head(); 1279 } 1280 VectorSet& useful_member_set = useful.member_set(); 1281 int last_idx = l_nodes_by_walk; 1282 for (int i = 0; i < last_idx; i++) { 1283 if (useful_member_set.test(i)) { 1284 if (_dead_node_list.test(i)) { 1285 if (_log != NULL) { 1286 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1287 } 1288 if (PrintIdealNodeCount) { 1289 // Print the log message to tty 1290 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1291 useful.at(i)->dump(); 1292 } 1293 } 1294 } 1295 else if (! _dead_node_list.test(i)) { 1296 if (_log != NULL) { 1297 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1298 } 1299 if (PrintIdealNodeCount) { 1300 // Print the log message to tty 1301 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1302 } 1303 } 1304 } 1305 if (_log != NULL) { 1306 _log->tail("mismatched_nodes"); 1307 } 1308 } 1309 } 1310 void Compile::record_modified_node(Node* n) { 1311 if (_modified_nodes != NULL && !_inlining_incrementally && 1312 n->outcnt() != 0 && !n->is_Con()) { 1313 _modified_nodes->push(n); 1314 } 1315 } 1316 1317 void Compile::remove_modified_node(Node* n) { 1318 if (_modified_nodes != NULL) { 1319 _modified_nodes->remove(n); 1320 } 1321 } 1322 #endif 1323 1324 #ifndef PRODUCT 1325 void Compile::verify_top(Node* tn) const { 1326 if (tn != NULL) { 1327 assert(tn->is_Con(), "top node must be a constant"); 1328 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1329 assert(tn->in(0) != NULL, "must have live top node"); 1330 } 1331 } 1332 #endif 1333 1334 1335 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1336 1337 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1338 guarantee(arr != NULL, ""); 1339 int num_blocks = arr->length(); 1340 if (grow_by < num_blocks) grow_by = num_blocks; 1341 int num_notes = grow_by * _node_notes_block_size; 1342 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1343 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1344 while (num_notes > 0) { 1345 arr->append(notes); 1346 notes += _node_notes_block_size; 1347 num_notes -= _node_notes_block_size; 1348 } 1349 assert(num_notes == 0, "exact multiple, please"); 1350 } 1351 1352 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1353 if (source == NULL || dest == NULL) return false; 1354 1355 if (dest->is_Con()) 1356 return false; // Do not push debug info onto constants. 1357 1358 #ifdef ASSERT 1359 // Leave a bread crumb trail pointing to the original node: 1360 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1361 dest->set_debug_orig(source); 1362 } 1363 #endif 1364 1365 if (node_note_array() == NULL) 1366 return false; // Not collecting any notes now. 1367 1368 // This is a copy onto a pre-existing node, which may already have notes. 1369 // If both nodes have notes, do not overwrite any pre-existing notes. 1370 Node_Notes* source_notes = node_notes_at(source->_idx); 1371 if (source_notes == NULL || source_notes->is_clear()) return false; 1372 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1373 if (dest_notes == NULL || dest_notes->is_clear()) { 1374 return set_node_notes_at(dest->_idx, source_notes); 1375 } 1376 1377 Node_Notes merged_notes = (*source_notes); 1378 // The order of operations here ensures that dest notes will win... 1379 merged_notes.update_from(dest_notes); 1380 return set_node_notes_at(dest->_idx, &merged_notes); 1381 } 1382 1383 1384 //--------------------------allow_range_check_smearing------------------------- 1385 // Gating condition for coalescing similar range checks. 1386 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1387 // single covering check that is at least as strong as any of them. 1388 // If the optimization succeeds, the simplified (strengthened) range check 1389 // will always succeed. If it fails, we will deopt, and then give up 1390 // on the optimization. 1391 bool Compile::allow_range_check_smearing() const { 1392 // If this method has already thrown a range-check, 1393 // assume it was because we already tried range smearing 1394 // and it failed. 1395 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1396 return !already_trapped; 1397 } 1398 1399 1400 //------------------------------flatten_alias_type----------------------------- 1401 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1402 int offset = tj->offset(); 1403 TypePtr::PTR ptr = tj->ptr(); 1404 1405 // Known instance (scalarizable allocation) alias only with itself. 1406 bool is_known_inst = tj->isa_oopptr() != NULL && 1407 tj->is_oopptr()->is_known_instance(); 1408 1409 // Process weird unsafe references. 1410 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1411 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1412 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1413 tj = TypeOopPtr::BOTTOM; 1414 ptr = tj->ptr(); 1415 offset = tj->offset(); 1416 } 1417 1418 // Array pointers need some flattening 1419 const TypeAryPtr *ta = tj->isa_aryptr(); 1420 if (ta && ta->is_stable()) { 1421 // Erase stability property for alias analysis. 1422 tj = ta = ta->cast_to_stable(false); 1423 } 1424 if( ta && is_known_inst ) { 1425 if ( offset != Type::OffsetBot && 1426 offset > arrayOopDesc::length_offset_in_bytes() ) { 1427 offset = Type::OffsetBot; // Flatten constant access into array body only 1428 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id()); 1429 } 1430 } else if( ta && _AliasLevel >= 2 ) { 1431 // For arrays indexed by constant indices, we flatten the alias 1432 // space to include all of the array body. Only the header, klass 1433 // and array length can be accessed un-aliased. 1434 // For flattened value type array, each field has its own slice so 1435 // we must include the field offset. 1436 if( offset != Type::OffsetBot ) { 1437 if( ta->const_oop() ) { // MethodData* or Method* 1438 offset = Type::OffsetBot; // Flatten constant access into array body 1439 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1440 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1441 // range is OK as-is. 1442 tj = ta = TypeAryPtr::RANGE; 1443 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1444 tj = TypeInstPtr::KLASS; // all klass loads look alike 1445 ta = TypeAryPtr::RANGE; // generic ignored junk 1446 ptr = TypePtr::BotPTR; 1447 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1448 tj = TypeInstPtr::MARK; 1449 ta = TypeAryPtr::RANGE; // generic ignored junk 1450 ptr = TypePtr::BotPTR; 1451 } else { // Random constant offset into array body 1452 offset = Type::OffsetBot; // Flatten constant access into array body 1453 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1454 } 1455 } 1456 // Arrays of fixed size alias with arrays of unknown size. 1457 if (ta->size() != TypeInt::POS) { 1458 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1459 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1460 } 1461 // Arrays of known objects become arrays of unknown objects. 1462 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1463 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1464 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1465 } 1466 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1467 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1468 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset()); 1469 } 1470 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1471 // cannot be distinguished by bytecode alone. 1472 if (ta->elem() == TypeInt::BOOL) { 1473 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1474 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1475 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset()); 1476 } 1477 // During the 2nd round of IterGVN, NotNull castings are removed. 1478 // Make sure the Bottom and NotNull variants alias the same. 1479 // Also, make sure exact and non-exact variants alias the same. 1480 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1481 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset()); 1482 } 1483 } 1484 1485 // Oop pointers need some flattening 1486 const TypeInstPtr *to = tj->isa_instptr(); 1487 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1488 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1489 if( ptr == TypePtr::Constant ) { 1490 if (to->klass() != ciEnv::current()->Class_klass() || 1491 offset < k->size_helper() * wordSize) { 1492 // No constant oop pointers (such as Strings); they alias with 1493 // unknown strings. 1494 assert(!is_known_inst, "not scalarizable allocation"); 1495 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1496 } 1497 } else if( is_known_inst ) { 1498 tj = to; // Keep NotNull and klass_is_exact for instance type 1499 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1500 // During the 2nd round of IterGVN, NotNull castings are removed. 1501 // Make sure the Bottom and NotNull variants alias the same. 1502 // Also, make sure exact and non-exact variants alias the same. 1503 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset)); 1504 } 1505 if (to->speculative() != NULL) { 1506 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id()); 1507 } 1508 // Canonicalize the holder of this field 1509 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1510 // First handle header references such as a LoadKlassNode, even if the 1511 // object's klass is unloaded at compile time (4965979). 1512 if (!is_known_inst) { // Do it only for non-instance types 1513 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset)); 1514 } 1515 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1516 // Static fields are in the space above the normal instance 1517 // fields in the java.lang.Class instance. 1518 if (to->klass() != ciEnv::current()->Class_klass()) { 1519 to = NULL; 1520 tj = TypeOopPtr::BOTTOM; 1521 offset = tj->offset(); 1522 } 1523 } else { 1524 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1525 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1526 if( is_known_inst ) { 1527 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id()); 1528 } else { 1529 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset)); 1530 } 1531 } 1532 } 1533 } 1534 1535 // Klass pointers to object array klasses need some flattening 1536 const TypeKlassPtr *tk = tj->isa_klassptr(); 1537 if( tk ) { 1538 // If we are referencing a field within a Klass, we need 1539 // to assume the worst case of an Object. Both exact and 1540 // inexact types must flatten to the same alias class so 1541 // use NotNull as the PTR. 1542 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1543 1544 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1545 TypeKlassPtr::OBJECT->klass(), 1546 Type::Offset(offset)); 1547 } 1548 1549 ciKlass* klass = tk->klass(); 1550 if (klass != NULL && klass->is_obj_array_klass()) { 1551 ciKlass* k = TypeAryPtr::OOPS->klass(); 1552 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1553 k = TypeInstPtr::BOTTOM->klass(); 1554 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset)); 1555 } 1556 1557 // Check for precise loads from the primary supertype array and force them 1558 // to the supertype cache alias index. Check for generic array loads from 1559 // the primary supertype array and also force them to the supertype cache 1560 // alias index. Since the same load can reach both, we need to merge 1561 // these 2 disparate memories into the same alias class. Since the 1562 // primary supertype array is read-only, there's no chance of confusion 1563 // where we bypass an array load and an array store. 1564 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1565 if (offset == Type::OffsetBot || 1566 (offset >= primary_supers_offset && 1567 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1568 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1569 offset = in_bytes(Klass::secondary_super_cache_offset()); 1570 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset)); 1571 } 1572 } 1573 1574 // Flatten all Raw pointers together. 1575 if (tj->base() == Type::RawPtr) 1576 tj = TypeRawPtr::BOTTOM; 1577 1578 if (tj->base() == Type::AnyPtr) 1579 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1580 1581 // Flatten all to bottom for now 1582 switch( _AliasLevel ) { 1583 case 0: 1584 tj = TypePtr::BOTTOM; 1585 break; 1586 case 1: // Flatten to: oop, static, field or array 1587 switch (tj->base()) { 1588 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1589 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1590 case Type::AryPtr: // do not distinguish arrays at all 1591 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1592 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1593 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1594 default: ShouldNotReachHere(); 1595 } 1596 break; 1597 case 2: // No collapsing at level 2; keep all splits 1598 case 3: // No collapsing at level 3; keep all splits 1599 break; 1600 default: 1601 Unimplemented(); 1602 } 1603 1604 offset = tj->offset(); 1605 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1606 1607 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1608 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1609 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1610 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1611 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1612 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1613 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1614 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1615 assert( tj->ptr() != TypePtr::TopPTR && 1616 tj->ptr() != TypePtr::AnyNull && 1617 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1618 // assert( tj->ptr() != TypePtr::Constant || 1619 // tj->base() == Type::RawPtr || 1620 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1621 1622 return tj; 1623 } 1624 1625 void Compile::AliasType::Init(int i, const TypePtr* at) { 1626 _index = i; 1627 _adr_type = at; 1628 _field = NULL; 1629 _element = NULL; 1630 _is_rewritable = true; // default 1631 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1632 if (atoop != NULL && atoop->is_known_instance()) { 1633 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1634 _general_index = Compile::current()->get_alias_index(gt); 1635 } else { 1636 _general_index = 0; 1637 } 1638 } 1639 1640 BasicType Compile::AliasType::basic_type() const { 1641 if (element() != NULL) { 1642 const Type* element = adr_type()->is_aryptr()->elem(); 1643 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1644 } if (field() != NULL) { 1645 return field()->layout_type(); 1646 } else { 1647 return T_ILLEGAL; // unknown 1648 } 1649 } 1650 1651 //---------------------------------print_on------------------------------------ 1652 #ifndef PRODUCT 1653 void Compile::AliasType::print_on(outputStream* st) { 1654 if (index() < 10) 1655 st->print("@ <%d> ", index()); 1656 else st->print("@ <%d>", index()); 1657 st->print(is_rewritable() ? " " : " RO"); 1658 int offset = adr_type()->offset(); 1659 if (offset == Type::OffsetBot) 1660 st->print(" +any"); 1661 else st->print(" +%-3d", offset); 1662 st->print(" in "); 1663 adr_type()->dump_on(st); 1664 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1665 if (field() != NULL && tjp) { 1666 if (tjp->klass() != field()->holder() || 1667 tjp->offset() != field()->offset_in_bytes()) { 1668 st->print(" != "); 1669 field()->print(); 1670 st->print(" ***"); 1671 } 1672 } 1673 } 1674 1675 void print_alias_types() { 1676 Compile* C = Compile::current(); 1677 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1678 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1679 C->alias_type(idx)->print_on(tty); 1680 tty->cr(); 1681 } 1682 } 1683 #endif 1684 1685 1686 //----------------------------probe_alias_cache-------------------------------- 1687 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1688 intptr_t key = (intptr_t) adr_type; 1689 key ^= key >> logAliasCacheSize; 1690 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1691 } 1692 1693 1694 //-----------------------------grow_alias_types-------------------------------- 1695 void Compile::grow_alias_types() { 1696 const int old_ats = _max_alias_types; // how many before? 1697 const int new_ats = old_ats; // how many more? 1698 const int grow_ats = old_ats+new_ats; // how many now? 1699 _max_alias_types = grow_ats; 1700 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1701 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1702 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1703 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1704 } 1705 1706 1707 //--------------------------------find_alias_type------------------------------ 1708 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1709 if (_AliasLevel == 0) 1710 return alias_type(AliasIdxBot); 1711 1712 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1713 if (ace->_adr_type == adr_type) { 1714 return alias_type(ace->_index); 1715 } 1716 1717 // Handle special cases. 1718 if (adr_type == NULL) return alias_type(AliasIdxTop); 1719 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1720 1721 // Do it the slow way. 1722 const TypePtr* flat = flatten_alias_type(adr_type); 1723 1724 #ifdef ASSERT 1725 { 1726 ResourceMark rm; 1727 assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s", 1728 Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat))); 1729 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s", 1730 Type::str(adr_type)); 1731 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1732 const TypeOopPtr* foop = flat->is_oopptr(); 1733 // Scalarizable allocations have exact klass always. 1734 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1735 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1736 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s", 1737 Type::str(foop), Type::str(xoop)); 1738 } 1739 } 1740 #endif 1741 1742 int idx = AliasIdxTop; 1743 for (int i = 0; i < num_alias_types(); i++) { 1744 if (alias_type(i)->adr_type() == flat) { 1745 idx = i; 1746 break; 1747 } 1748 } 1749 1750 if (idx == AliasIdxTop) { 1751 if (no_create) return NULL; 1752 // Grow the array if necessary. 1753 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1754 // Add a new alias type. 1755 idx = _num_alias_types++; 1756 _alias_types[idx]->Init(idx, flat); 1757 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1758 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1759 if (flat->isa_instptr()) { 1760 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1761 && flat->is_instptr()->klass() == env()->Class_klass()) 1762 alias_type(idx)->set_rewritable(false); 1763 } 1764 if (flat->isa_aryptr()) { 1765 #ifdef ASSERT 1766 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1767 // (T_BYTE has the weakest alignment and size restrictions...) 1768 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1769 #endif 1770 if (flat->offset() == TypePtr::OffsetBot) { 1771 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1772 } 1773 } 1774 if (flat->isa_klassptr()) { 1775 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1776 alias_type(idx)->set_rewritable(false); 1777 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1778 alias_type(idx)->set_rewritable(false); 1779 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1780 alias_type(idx)->set_rewritable(false); 1781 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1782 alias_type(idx)->set_rewritable(false); 1783 } 1784 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1785 // but the base pointer type is not distinctive enough to identify 1786 // references into JavaThread.) 1787 1788 // Check for final fields. 1789 const TypeInstPtr* tinst = flat->isa_instptr(); 1790 const TypeValueTypePtr* vtptr = flat->isa_valuetypeptr(); 1791 ciField* field = NULL; 1792 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1793 if (tinst->const_oop() != NULL && 1794 tinst->klass() == ciEnv::current()->Class_klass() && 1795 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1796 // static field 1797 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1798 field = k->get_field_by_offset(tinst->offset(), true); 1799 } else { 1800 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1801 field = k->get_field_by_offset(tinst->offset(), false); 1802 } 1803 } else if (vtptr) { 1804 // Value type field 1805 ciValueKlass* vk = vtptr->klass()->as_value_klass(); 1806 field = vk->get_field_by_offset(vtptr->offset(), false); 1807 } 1808 assert(field == NULL || 1809 original_field == NULL || 1810 (field->holder() == original_field->holder() && 1811 field->offset() == original_field->offset() && 1812 field->is_static() == original_field->is_static()), "wrong field?"); 1813 // Set field() and is_rewritable() attributes. 1814 if (field != NULL) alias_type(idx)->set_field(field); 1815 } 1816 1817 // Fill the cache for next time. 1818 ace->_adr_type = adr_type; 1819 ace->_index = idx; 1820 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1821 1822 // Might as well try to fill the cache for the flattened version, too. 1823 AliasCacheEntry* face = probe_alias_cache(flat); 1824 if (face->_adr_type == NULL) { 1825 face->_adr_type = flat; 1826 face->_index = idx; 1827 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1828 } 1829 1830 return alias_type(idx); 1831 } 1832 1833 1834 Compile::AliasType* Compile::alias_type(ciField* field) { 1835 const TypeOopPtr* t; 1836 if (field->is_static()) 1837 t = TypeInstPtr::make(field->holder()->java_mirror()); 1838 else 1839 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1840 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1841 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1842 return atp; 1843 } 1844 1845 1846 //------------------------------have_alias_type-------------------------------- 1847 bool Compile::have_alias_type(const TypePtr* adr_type) { 1848 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1849 if (ace->_adr_type == adr_type) { 1850 return true; 1851 } 1852 1853 // Handle special cases. 1854 if (adr_type == NULL) return true; 1855 if (adr_type == TypePtr::BOTTOM) return true; 1856 1857 return find_alias_type(adr_type, true, NULL) != NULL; 1858 } 1859 1860 //-----------------------------must_alias-------------------------------------- 1861 // True if all values of the given address type are in the given alias category. 1862 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1863 if (alias_idx == AliasIdxBot) return true; // the universal category 1864 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1865 if (alias_idx == AliasIdxTop) return false; // the empty category 1866 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1867 1868 // the only remaining possible overlap is identity 1869 int adr_idx = get_alias_index(adr_type); 1870 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1871 assert(adr_idx == alias_idx || 1872 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1873 && adr_type != TypeOopPtr::BOTTOM), 1874 "should not be testing for overlap with an unsafe pointer"); 1875 return adr_idx == alias_idx; 1876 } 1877 1878 //------------------------------can_alias-------------------------------------- 1879 // True if any values of the given address type are in the given alias category. 1880 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1881 if (alias_idx == AliasIdxTop) return false; // the empty category 1882 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1883 if (alias_idx == AliasIdxBot) return true; // the universal category 1884 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1885 1886 // the only remaining possible overlap is identity 1887 int adr_idx = get_alias_index(adr_type); 1888 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1889 return adr_idx == alias_idx; 1890 } 1891 1892 1893 1894 //---------------------------pop_warm_call------------------------------------- 1895 WarmCallInfo* Compile::pop_warm_call() { 1896 WarmCallInfo* wci = _warm_calls; 1897 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1898 return wci; 1899 } 1900 1901 //----------------------------Inline_Warm-------------------------------------- 1902 int Compile::Inline_Warm() { 1903 // If there is room, try to inline some more warm call sites. 1904 // %%% Do a graph index compaction pass when we think we're out of space? 1905 if (!InlineWarmCalls) return 0; 1906 1907 int calls_made_hot = 0; 1908 int room_to_grow = NodeCountInliningCutoff - unique(); 1909 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1910 int amount_grown = 0; 1911 WarmCallInfo* call; 1912 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1913 int est_size = (int)call->size(); 1914 if (est_size > (room_to_grow - amount_grown)) { 1915 // This one won't fit anyway. Get rid of it. 1916 call->make_cold(); 1917 continue; 1918 } 1919 call->make_hot(); 1920 calls_made_hot++; 1921 amount_grown += est_size; 1922 amount_to_grow -= est_size; 1923 } 1924 1925 if (calls_made_hot > 0) set_major_progress(); 1926 return calls_made_hot; 1927 } 1928 1929 1930 //----------------------------Finish_Warm-------------------------------------- 1931 void Compile::Finish_Warm() { 1932 if (!InlineWarmCalls) return; 1933 if (failing()) return; 1934 if (warm_calls() == NULL) return; 1935 1936 // Clean up loose ends, if we are out of space for inlining. 1937 WarmCallInfo* call; 1938 while ((call = pop_warm_call()) != NULL) { 1939 call->make_cold(); 1940 } 1941 } 1942 1943 //---------------------cleanup_loop_predicates----------------------- 1944 // Remove the opaque nodes that protect the predicates so that all unused 1945 // checks and uncommon_traps will be eliminated from the ideal graph 1946 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1947 if (predicate_count()==0) return; 1948 for (int i = predicate_count(); i > 0; i--) { 1949 Node * n = predicate_opaque1_node(i-1); 1950 assert(n->Opcode() == Op_Opaque1, "must be"); 1951 igvn.replace_node(n, n->in(1)); 1952 } 1953 assert(predicate_count()==0, "should be clean!"); 1954 } 1955 1956 void Compile::add_range_check_cast(Node* n) { 1957 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1958 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1959 _range_check_casts->append(n); 1960 } 1961 1962 // Remove all range check dependent CastIINodes. 1963 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1964 for (int i = range_check_cast_count(); i > 0; i--) { 1965 Node* cast = range_check_cast_node(i-1); 1966 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1967 igvn.replace_node(cast, cast->in(1)); 1968 } 1969 assert(range_check_cast_count() == 0, "should be empty"); 1970 } 1971 1972 void Compile::add_value_type_ptr(ValueTypePtrNode* n) { 1973 assert(!_value_type_ptr_nodes->contains(n), "duplicate entry"); 1974 _value_type_ptr_nodes->append(n); 1975 } 1976 1977 void Compile::process_value_type_ptr_nodes(PhaseIterGVN &igvn) { 1978 for (int i = value_type_ptr_count(); i > 0; i--) { 1979 ValueTypePtrNode* vtptr = value_type_ptr(i-1); 1980 // once all inlining is over otherwise debug info can get 1981 // inconsistent 1982 vtptr->make_scalar_in_safepoints(igvn.C->root(), &igvn); 1983 igvn.replace_node(vtptr, vtptr->get_oop()); 1984 } 1985 assert(value_type_ptr_count() == 0, "should be empty"); 1986 igvn.optimize(); 1987 } 1988 1989 // StringOpts and late inlining of string methods 1990 void Compile::inline_string_calls(bool parse_time) { 1991 { 1992 // remove useless nodes to make the usage analysis simpler 1993 ResourceMark rm; 1994 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1995 } 1996 1997 { 1998 ResourceMark rm; 1999 print_method(PHASE_BEFORE_STRINGOPTS, 3); 2000 PhaseStringOpts pso(initial_gvn(), for_igvn()); 2001 print_method(PHASE_AFTER_STRINGOPTS, 3); 2002 } 2003 2004 // now inline anything that we skipped the first time around 2005 if (!parse_time) { 2006 _late_inlines_pos = _late_inlines.length(); 2007 } 2008 2009 while (_string_late_inlines.length() > 0) { 2010 CallGenerator* cg = _string_late_inlines.pop(); 2011 cg->do_late_inline(); 2012 if (failing()) return; 2013 } 2014 _string_late_inlines.trunc_to(0); 2015 } 2016 2017 // Late inlining of boxing methods 2018 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2019 if (_boxing_late_inlines.length() > 0) { 2020 assert(has_boxed_value(), "inconsistent"); 2021 2022 PhaseGVN* gvn = initial_gvn(); 2023 set_inlining_incrementally(true); 2024 2025 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2026 for_igvn()->clear(); 2027 gvn->replace_with(&igvn); 2028 2029 _late_inlines_pos = _late_inlines.length(); 2030 2031 while (_boxing_late_inlines.length() > 0) { 2032 CallGenerator* cg = _boxing_late_inlines.pop(); 2033 cg->do_late_inline(); 2034 if (failing()) return; 2035 } 2036 _boxing_late_inlines.trunc_to(0); 2037 2038 { 2039 ResourceMark rm; 2040 PhaseRemoveUseless pru(gvn, for_igvn()); 2041 } 2042 2043 igvn = PhaseIterGVN(gvn); 2044 igvn.optimize(); 2045 2046 set_inlining_progress(false); 2047 set_inlining_incrementally(false); 2048 } 2049 } 2050 2051 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2052 assert(IncrementalInline, "incremental inlining should be on"); 2053 PhaseGVN* gvn = initial_gvn(); 2054 2055 set_inlining_progress(false); 2056 for_igvn()->clear(); 2057 gvn->replace_with(&igvn); 2058 2059 { 2060 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2061 int i = 0; 2062 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2063 CallGenerator* cg = _late_inlines.at(i); 2064 _late_inlines_pos = i+1; 2065 cg->do_late_inline(); 2066 if (failing()) return; 2067 } 2068 int j = 0; 2069 for (; i < _late_inlines.length(); i++, j++) { 2070 _late_inlines.at_put(j, _late_inlines.at(i)); 2071 } 2072 _late_inlines.trunc_to(j); 2073 } 2074 2075 { 2076 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2077 ResourceMark rm; 2078 PhaseRemoveUseless pru(gvn, for_igvn()); 2079 } 2080 2081 { 2082 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2083 igvn = PhaseIterGVN(gvn); 2084 } 2085 } 2086 2087 // Perform incremental inlining until bound on number of live nodes is reached 2088 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2089 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2090 2091 PhaseGVN* gvn = initial_gvn(); 2092 2093 set_inlining_incrementally(true); 2094 set_inlining_progress(true); 2095 uint low_live_nodes = 0; 2096 2097 while(inlining_progress() && _late_inlines.length() > 0) { 2098 2099 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2100 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2101 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2102 // PhaseIdealLoop is expensive so we only try it once we are 2103 // out of live nodes and we only try it again if the previous 2104 // helped got the number of nodes down significantly 2105 PhaseIdealLoop ideal_loop( igvn, false, true ); 2106 if (failing()) return; 2107 low_live_nodes = live_nodes(); 2108 _major_progress = true; 2109 } 2110 2111 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2112 break; 2113 } 2114 } 2115 2116 inline_incrementally_one(igvn); 2117 2118 if (failing()) return; 2119 2120 { 2121 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2122 igvn.optimize(); 2123 } 2124 2125 if (failing()) return; 2126 } 2127 2128 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2129 2130 if (_string_late_inlines.length() > 0) { 2131 assert(has_stringbuilder(), "inconsistent"); 2132 for_igvn()->clear(); 2133 initial_gvn()->replace_with(&igvn); 2134 2135 inline_string_calls(false); 2136 2137 if (failing()) return; 2138 2139 { 2140 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2141 ResourceMark rm; 2142 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2143 } 2144 2145 { 2146 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2147 igvn = PhaseIterGVN(gvn); 2148 igvn.optimize(); 2149 } 2150 } 2151 2152 set_inlining_incrementally(false); 2153 } 2154 2155 2156 //------------------------------Optimize--------------------------------------- 2157 // Given a graph, optimize it. 2158 void Compile::Optimize() { 2159 TracePhase tp("optimizer", &timers[_t_optimizer]); 2160 2161 #ifndef PRODUCT 2162 if (_directive->BreakAtCompileOption) { 2163 BREAKPOINT; 2164 } 2165 2166 #endif 2167 2168 ResourceMark rm; 2169 int loop_opts_cnt; 2170 2171 print_inlining_reinit(); 2172 2173 NOT_PRODUCT( verify_graph_edges(); ) 2174 2175 print_method(PHASE_AFTER_PARSING); 2176 2177 { 2178 // Iterative Global Value Numbering, including ideal transforms 2179 // Initialize IterGVN with types and values from parse-time GVN 2180 PhaseIterGVN igvn(initial_gvn()); 2181 #ifdef ASSERT 2182 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2183 #endif 2184 { 2185 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2186 igvn.optimize(); 2187 } 2188 2189 print_method(PHASE_ITER_GVN1, 2); 2190 2191 if (failing()) return; 2192 2193 inline_incrementally(igvn); 2194 2195 print_method(PHASE_INCREMENTAL_INLINE, 2); 2196 2197 if (failing()) return; 2198 2199 if (eliminate_boxing()) { 2200 // Inline valueOf() methods now. 2201 inline_boxing_calls(igvn); 2202 2203 if (AlwaysIncrementalInline) { 2204 inline_incrementally(igvn); 2205 } 2206 2207 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2208 2209 if (failing()) return; 2210 } 2211 2212 // Remove the speculative part of types and clean up the graph from 2213 // the extra CastPP nodes whose only purpose is to carry them. Do 2214 // that early so that optimizations are not disrupted by the extra 2215 // CastPP nodes. 2216 remove_speculative_types(igvn); 2217 2218 // No more new expensive nodes will be added to the list from here 2219 // so keep only the actual candidates for optimizations. 2220 cleanup_expensive_nodes(igvn); 2221 2222 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2223 Compile::TracePhase tp("", &timers[_t_renumberLive]); 2224 initial_gvn()->replace_with(&igvn); 2225 for_igvn()->clear(); 2226 Unique_Node_List new_worklist(C->comp_arena()); 2227 { 2228 ResourceMark rm; 2229 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2230 } 2231 set_for_igvn(&new_worklist); 2232 igvn = PhaseIterGVN(initial_gvn()); 2233 igvn.optimize(); 2234 } 2235 2236 if (value_type_ptr_count() > 0) { 2237 process_value_type_ptr_nodes(igvn); 2238 } 2239 2240 // Perform escape analysis 2241 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2242 if (has_loops()) { 2243 // Cleanup graph (remove dead nodes). 2244 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2245 PhaseIdealLoop ideal_loop( igvn, false, true ); 2246 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2247 if (failing()) return; 2248 } 2249 ConnectionGraph::do_analysis(this, &igvn); 2250 2251 if (failing()) return; 2252 2253 // Optimize out fields loads from scalar replaceable allocations. 2254 igvn.optimize(); 2255 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2256 2257 if (failing()) return; 2258 2259 if (congraph() != NULL && macro_count() > 0) { 2260 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2261 PhaseMacroExpand mexp(igvn); 2262 mexp.eliminate_macro_nodes(); 2263 igvn.set_delay_transform(false); 2264 2265 igvn.optimize(); 2266 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2267 2268 if (failing()) return; 2269 } 2270 } 2271 2272 // Loop transforms on the ideal graph. Range Check Elimination, 2273 // peeling, unrolling, etc. 2274 2275 // Set loop opts counter 2276 loop_opts_cnt = num_loop_opts(); 2277 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2278 { 2279 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2280 PhaseIdealLoop ideal_loop( igvn, true ); 2281 loop_opts_cnt--; 2282 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2283 if (failing()) return; 2284 } 2285 // Loop opts pass if partial peeling occurred in previous pass 2286 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2287 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2288 PhaseIdealLoop ideal_loop( igvn, false ); 2289 loop_opts_cnt--; 2290 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2291 if (failing()) return; 2292 } 2293 // Loop opts pass for loop-unrolling before CCP 2294 if(major_progress() && (loop_opts_cnt > 0)) { 2295 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2296 PhaseIdealLoop ideal_loop( igvn, false ); 2297 loop_opts_cnt--; 2298 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2299 } 2300 if (!failing()) { 2301 // Verify that last round of loop opts produced a valid graph 2302 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2303 PhaseIdealLoop::verify(igvn); 2304 } 2305 } 2306 if (failing()) return; 2307 2308 // Conditional Constant Propagation; 2309 PhaseCCP ccp( &igvn ); 2310 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2311 { 2312 TracePhase tp("ccp", &timers[_t_ccp]); 2313 ccp.do_transform(); 2314 } 2315 print_method(PHASE_CPP1, 2); 2316 2317 assert( true, "Break here to ccp.dump_old2new_map()"); 2318 2319 // Iterative Global Value Numbering, including ideal transforms 2320 { 2321 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2322 igvn = ccp; 2323 igvn.optimize(); 2324 } 2325 2326 print_method(PHASE_ITER_GVN2, 2); 2327 2328 if (failing()) return; 2329 2330 // Loop transforms on the ideal graph. Range Check Elimination, 2331 // peeling, unrolling, etc. 2332 if(loop_opts_cnt > 0) { 2333 debug_only( int cnt = 0; ); 2334 while(major_progress() && (loop_opts_cnt > 0)) { 2335 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2336 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2337 PhaseIdealLoop ideal_loop( igvn, true); 2338 loop_opts_cnt--; 2339 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2340 if (failing()) return; 2341 } 2342 } 2343 // Ensure that major progress is now clear 2344 C->clear_major_progress(); 2345 2346 { 2347 // Verify that all previous optimizations produced a valid graph 2348 // at least to this point, even if no loop optimizations were done. 2349 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2350 PhaseIdealLoop::verify(igvn); 2351 } 2352 2353 if (range_check_cast_count() > 0) { 2354 // No more loop optimizations. Remove all range check dependent CastIINodes. 2355 C->remove_range_check_casts(igvn); 2356 igvn.optimize(); 2357 } 2358 2359 { 2360 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2361 PhaseMacroExpand mex(igvn); 2362 if (mex.expand_macro_nodes()) { 2363 assert(failing(), "must bail out w/ explicit message"); 2364 return; 2365 } 2366 } 2367 2368 DEBUG_ONLY( _modified_nodes = NULL; ) 2369 } // (End scope of igvn; run destructor if necessary for asserts.) 2370 2371 process_print_inlining(); 2372 // A method with only infinite loops has no edges entering loops from root 2373 { 2374 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2375 if (final_graph_reshaping()) { 2376 assert(failing(), "must bail out w/ explicit message"); 2377 return; 2378 } 2379 } 2380 2381 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2382 } 2383 2384 // Fixme remove 2385 static void check_for_value_node(Node &n, void* C) { 2386 if (n.is_ValueType()) { 2387 #ifdef ASSERT 2388 ((Compile*)C)->method()->print_short_name(); 2389 tty->print_cr(""); 2390 n.dump(-1); 2391 assert(false, "Unable to match ValueTypeNode"); 2392 #endif 2393 ((Compile*)C)->record_failure("Unable to match ValueTypeNode"); 2394 } 2395 } 2396 2397 //------------------------------Code_Gen--------------------------------------- 2398 // Given a graph, generate code for it 2399 void Compile::Code_Gen() { 2400 // FIXME remove 2401 root()->walk(Node::nop, check_for_value_node, this); 2402 2403 if (failing()) { 2404 return; 2405 } 2406 2407 // Perform instruction selection. You might think we could reclaim Matcher 2408 // memory PDQ, but actually the Matcher is used in generating spill code. 2409 // Internals of the Matcher (including some VectorSets) must remain live 2410 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2411 // set a bit in reclaimed memory. 2412 2413 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2414 // nodes. Mapping is only valid at the root of each matched subtree. 2415 NOT_PRODUCT( verify_graph_edges(); ) 2416 2417 Matcher matcher; 2418 _matcher = &matcher; 2419 { 2420 TracePhase tp("matcher", &timers[_t_matcher]); 2421 matcher.match(); 2422 } 2423 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2424 // nodes. Mapping is only valid at the root of each matched subtree. 2425 NOT_PRODUCT( verify_graph_edges(); ) 2426 2427 // If you have too many nodes, or if matching has failed, bail out 2428 check_node_count(0, "out of nodes matching instructions"); 2429 if (failing()) { 2430 return; 2431 } 2432 2433 // Build a proper-looking CFG 2434 PhaseCFG cfg(node_arena(), root(), matcher); 2435 _cfg = &cfg; 2436 { 2437 TracePhase tp("scheduler", &timers[_t_scheduler]); 2438 bool success = cfg.do_global_code_motion(); 2439 if (!success) { 2440 return; 2441 } 2442 2443 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2444 NOT_PRODUCT( verify_graph_edges(); ) 2445 debug_only( cfg.verify(); ) 2446 } 2447 2448 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2449 _regalloc = ®alloc; 2450 { 2451 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2452 // Perform register allocation. After Chaitin, use-def chains are 2453 // no longer accurate (at spill code) and so must be ignored. 2454 // Node->LRG->reg mappings are still accurate. 2455 _regalloc->Register_Allocate(); 2456 2457 // Bail out if the allocator builds too many nodes 2458 if (failing()) { 2459 return; 2460 } 2461 } 2462 2463 // Prior to register allocation we kept empty basic blocks in case the 2464 // the allocator needed a place to spill. After register allocation we 2465 // are not adding any new instructions. If any basic block is empty, we 2466 // can now safely remove it. 2467 { 2468 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2469 cfg.remove_empty_blocks(); 2470 if (do_freq_based_layout()) { 2471 PhaseBlockLayout layout(cfg); 2472 } else { 2473 cfg.set_loop_alignment(); 2474 } 2475 cfg.fixup_flow(); 2476 } 2477 2478 // Apply peephole optimizations 2479 if( OptoPeephole ) { 2480 TracePhase tp("peephole", &timers[_t_peephole]); 2481 PhasePeephole peep( _regalloc, cfg); 2482 peep.do_transform(); 2483 } 2484 2485 // Do late expand if CPU requires this. 2486 if (Matcher::require_postalloc_expand) { 2487 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2488 cfg.postalloc_expand(_regalloc); 2489 } 2490 2491 // Convert Nodes to instruction bits in a buffer 2492 { 2493 TraceTime tp("output", &timers[_t_output], CITime); 2494 Output(); 2495 } 2496 2497 print_method(PHASE_FINAL_CODE); 2498 2499 // He's dead, Jim. 2500 _cfg = (PhaseCFG*)0xdeadbeef; 2501 _regalloc = (PhaseChaitin*)0xdeadbeef; 2502 } 2503 2504 2505 //------------------------------dump_asm--------------------------------------- 2506 // Dump formatted assembly 2507 #ifndef PRODUCT 2508 void Compile::dump_asm(int *pcs, uint pc_limit) { 2509 bool cut_short = false; 2510 tty->print_cr("#"); 2511 tty->print("# "); _tf->dump(); tty->cr(); 2512 tty->print_cr("#"); 2513 2514 // For all blocks 2515 int pc = 0x0; // Program counter 2516 char starts_bundle = ' '; 2517 _regalloc->dump_frame(); 2518 2519 Node *n = NULL; 2520 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2521 if (VMThread::should_terminate()) { 2522 cut_short = true; 2523 break; 2524 } 2525 Block* block = _cfg->get_block(i); 2526 if (block->is_connector() && !Verbose) { 2527 continue; 2528 } 2529 n = block->head(); 2530 if (pcs && n->_idx < pc_limit) { 2531 tty->print("%3.3x ", pcs[n->_idx]); 2532 } else { 2533 tty->print(" "); 2534 } 2535 block->dump_head(_cfg); 2536 if (block->is_connector()) { 2537 tty->print_cr(" # Empty connector block"); 2538 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2539 tty->print_cr(" # Block is sole successor of call"); 2540 } 2541 2542 // For all instructions 2543 Node *delay = NULL; 2544 for (uint j = 0; j < block->number_of_nodes(); j++) { 2545 if (VMThread::should_terminate()) { 2546 cut_short = true; 2547 break; 2548 } 2549 n = block->get_node(j); 2550 if (valid_bundle_info(n)) { 2551 Bundle* bundle = node_bundling(n); 2552 if (bundle->used_in_unconditional_delay()) { 2553 delay = n; 2554 continue; 2555 } 2556 if (bundle->starts_bundle()) { 2557 starts_bundle = '+'; 2558 } 2559 } 2560 2561 if (WizardMode) { 2562 n->dump(); 2563 } 2564 2565 if( !n->is_Region() && // Dont print in the Assembly 2566 !n->is_Phi() && // a few noisely useless nodes 2567 !n->is_Proj() && 2568 !n->is_MachTemp() && 2569 !n->is_SafePointScalarObject() && 2570 !n->is_Catch() && // Would be nice to print exception table targets 2571 !n->is_MergeMem() && // Not very interesting 2572 !n->is_top() && // Debug info table constants 2573 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2574 ) { 2575 if (pcs && n->_idx < pc_limit) 2576 tty->print("%3.3x", pcs[n->_idx]); 2577 else 2578 tty->print(" "); 2579 tty->print(" %c ", starts_bundle); 2580 starts_bundle = ' '; 2581 tty->print("\t"); 2582 n->format(_regalloc, tty); 2583 tty->cr(); 2584 } 2585 2586 // If we have an instruction with a delay slot, and have seen a delay, 2587 // then back up and print it 2588 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2589 assert(delay != NULL, "no unconditional delay instruction"); 2590 if (WizardMode) delay->dump(); 2591 2592 if (node_bundling(delay)->starts_bundle()) 2593 starts_bundle = '+'; 2594 if (pcs && n->_idx < pc_limit) 2595 tty->print("%3.3x", pcs[n->_idx]); 2596 else 2597 tty->print(" "); 2598 tty->print(" %c ", starts_bundle); 2599 starts_bundle = ' '; 2600 tty->print("\t"); 2601 delay->format(_regalloc, tty); 2602 tty->cr(); 2603 delay = NULL; 2604 } 2605 2606 // Dump the exception table as well 2607 if( n->is_Catch() && (Verbose || WizardMode) ) { 2608 // Print the exception table for this offset 2609 _handler_table.print_subtable_for(pc); 2610 } 2611 } 2612 2613 if (pcs && n->_idx < pc_limit) 2614 tty->print_cr("%3.3x", pcs[n->_idx]); 2615 else 2616 tty->cr(); 2617 2618 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2619 2620 } // End of per-block dump 2621 tty->cr(); 2622 2623 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2624 } 2625 #endif 2626 2627 //------------------------------Final_Reshape_Counts--------------------------- 2628 // This class defines counters to help identify when a method 2629 // may/must be executed using hardware with only 24-bit precision. 2630 struct Final_Reshape_Counts : public StackObj { 2631 int _call_count; // count non-inlined 'common' calls 2632 int _float_count; // count float ops requiring 24-bit precision 2633 int _double_count; // count double ops requiring more precision 2634 int _java_call_count; // count non-inlined 'java' calls 2635 int _inner_loop_count; // count loops which need alignment 2636 VectorSet _visited; // Visitation flags 2637 Node_List _tests; // Set of IfNodes & PCTableNodes 2638 2639 Final_Reshape_Counts() : 2640 _call_count(0), _float_count(0), _double_count(0), 2641 _java_call_count(0), _inner_loop_count(0), 2642 _visited( Thread::current()->resource_area() ) { } 2643 2644 void inc_call_count () { _call_count ++; } 2645 void inc_float_count () { _float_count ++; } 2646 void inc_double_count() { _double_count++; } 2647 void inc_java_call_count() { _java_call_count++; } 2648 void inc_inner_loop_count() { _inner_loop_count++; } 2649 2650 int get_call_count () const { return _call_count ; } 2651 int get_float_count () const { return _float_count ; } 2652 int get_double_count() const { return _double_count; } 2653 int get_java_call_count() const { return _java_call_count; } 2654 int get_inner_loop_count() const { return _inner_loop_count; } 2655 }; 2656 2657 #ifdef ASSERT 2658 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2659 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2660 // Make sure the offset goes inside the instance layout. 2661 return k->contains_field_offset(tp->offset()); 2662 // Note that OffsetBot and OffsetTop are very negative. 2663 } 2664 #endif 2665 2666 // Eliminate trivially redundant StoreCMs and accumulate their 2667 // precedence edges. 2668 void Compile::eliminate_redundant_card_marks(Node* n) { 2669 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2670 if (n->in(MemNode::Address)->outcnt() > 1) { 2671 // There are multiple users of the same address so it might be 2672 // possible to eliminate some of the StoreCMs 2673 Node* mem = n->in(MemNode::Memory); 2674 Node* adr = n->in(MemNode::Address); 2675 Node* val = n->in(MemNode::ValueIn); 2676 Node* prev = n; 2677 bool done = false; 2678 // Walk the chain of StoreCMs eliminating ones that match. As 2679 // long as it's a chain of single users then the optimization is 2680 // safe. Eliminating partially redundant StoreCMs would require 2681 // cloning copies down the other paths. 2682 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2683 if (adr == mem->in(MemNode::Address) && 2684 val == mem->in(MemNode::ValueIn)) { 2685 // redundant StoreCM 2686 if (mem->req() > MemNode::OopStore) { 2687 // Hasn't been processed by this code yet. 2688 n->add_prec(mem->in(MemNode::OopStore)); 2689 } else { 2690 // Already converted to precedence edge 2691 for (uint i = mem->req(); i < mem->len(); i++) { 2692 // Accumulate any precedence edges 2693 if (mem->in(i) != NULL) { 2694 n->add_prec(mem->in(i)); 2695 } 2696 } 2697 // Everything above this point has been processed. 2698 done = true; 2699 } 2700 // Eliminate the previous StoreCM 2701 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2702 assert(mem->outcnt() == 0, "should be dead"); 2703 mem->disconnect_inputs(NULL, this); 2704 } else { 2705 prev = mem; 2706 } 2707 mem = prev->in(MemNode::Memory); 2708 } 2709 } 2710 } 2711 2712 void Compile::value_type_return_from_mh_intrinsic(CallNode *call, Final_Reshape_Counts &frc) { 2713 if (ValueTypeReturnedAsFields && 2714 call->is_CallStaticJava() && 2715 call->as_CallStaticJava()->method() != NULL && 2716 call->as_CallStaticJava()->method()->is_method_handle_intrinsic() && 2717 call->proj_out(TypeFunc::Parms) != NULL && 2718 call->proj_out(TypeFunc::Parms)->bottom_type()->isa_valuetypeptr()) { 2719 // A value type is returned from the call but we don't know its 2720 // type. One of the values being returned is the klass of the 2721 // value type. We need to allocate a value type instance of that 2722 // type and initialize it with other values being returned. This 2723 // is done with the stub call below that we add right after this 2724 // call. 2725 Node* ret = call->proj_out(TypeFunc::Parms); 2726 assert(ret->bottom_type()->is_valuetypeptr()->klass() == env()->___Value_klass(), "unexpected return type from MH intrinsic"); 2727 const TypeFunc* tf = call->_tf; 2728 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc(); 2729 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain); 2730 call->_tf = new_tf; 2731 2732 CallProjections projs; 2733 call->extract_projections(&projs, true, true); 2734 Node* ctl = projs.fallthrough_catchproj; 2735 Node* mem = projs.fallthrough_memproj; 2736 Node* io = projs.fallthrough_ioproj; 2737 Node* ex_ctl = projs.catchall_catchproj; 2738 Node* ex_mem = projs.catchall_memproj; 2739 Node* ex_io = projs.catchall_ioproj; 2740 CallStaticJavaNode* rt_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(), 2741 StubRoutines::store_value_type_fields_to_buf(), 2742 "store_value_type_fields", 2743 call->jvms()->bci(), 2744 TypePtr::BOTTOM); 2745 Node* out_ctl = new ProjNode(rt_call, TypeFunc::Control); 2746 Node* out_mem = new ProjNode(rt_call, TypeFunc::Memory); 2747 Node* out_io = new ProjNode(rt_call, TypeFunc::I_O); 2748 Node* res = new ProjNode(rt_call, TypeFunc::Parms); 2749 2750 Node* catc = new CatchNode(out_ctl, out_io, 2); 2751 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci); 2752 Node* excp = new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci); 2753 Node* r = new RegionNode(3); 2754 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM); 2755 Node* io_phi = new PhiNode(r, Type::ABIO); 2756 r->init_req(1, excp); 2757 mem_phi->init_req(1, out_mem); 2758 io_phi->init_req(1, out_io); 2759 2760 frc._visited.set(norm->_idx); 2761 frc._visited.set(excp->_idx); 2762 2763 ctl->replace_by(norm); 2764 mem->replace_by(out_mem); 2765 io->replace_by(out_io); 2766 ret->replace_by(res); 2767 ex_ctl->replace_by(r); 2768 ex_mem->replace_by(mem_phi); 2769 ex_io->replace_by(io_phi); 2770 2771 r->init_req(2, ex_ctl); 2772 mem_phi->init_req(2, ex_mem); 2773 io_phi->init_req(2, ex_io); 2774 2775 rt_call->init_req(TypeFunc::Control, ctl); 2776 rt_call->init_req(TypeFunc::Memory, mem); 2777 rt_call->init_req(TypeFunc::I_O, io); 2778 rt_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr)); 2779 rt_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr)); 2780 2781 rt_call->init_req(TypeFunc::Parms, ret); 2782 // We don't know how many values are returned. This assumes the 2783 // worst case, that all available registers are used. 2784 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) { 2785 if (domain->field_at(i) == Type::HALF) { 2786 rt_call->init_req(i, top()); 2787 continue; 2788 } 2789 Node* proj = new ProjNode(call, i); 2790 rt_call->init_req(i, proj); 2791 } 2792 2793 // We can safepoint at that new call 2794 add_safepoint_edges(rt_call, call->jvms()); 2795 } 2796 } 2797 2798 //------------------------------final_graph_reshaping_impl---------------------- 2799 // Implement items 1-5 from final_graph_reshaping below. 2800 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2801 2802 if ( n->outcnt() == 0 ) return; // dead node 2803 uint nop = n->Opcode(); 2804 2805 // Check for 2-input instruction with "last use" on right input. 2806 // Swap to left input. Implements item (2). 2807 if( n->req() == 3 && // two-input instruction 2808 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2809 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2810 n->in(2)->outcnt() == 1 &&// right use IS a last use 2811 !n->in(2)->is_Con() ) { // right use is not a constant 2812 // Check for commutative opcode 2813 switch( nop ) { 2814 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2815 case Op_MaxI: case Op_MinI: 2816 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2817 case Op_AndL: case Op_XorL: case Op_OrL: 2818 case Op_AndI: case Op_XorI: case Op_OrI: { 2819 // Move "last use" input to left by swapping inputs 2820 n->swap_edges(1, 2); 2821 break; 2822 } 2823 default: 2824 break; 2825 } 2826 } 2827 2828 #ifdef ASSERT 2829 if( n->is_Mem() ) { 2830 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2831 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2832 // oop will be recorded in oop map if load crosses safepoint 2833 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2834 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2835 "raw memory operations should have control edge"); 2836 } 2837 #endif 2838 // Count FPU ops and common calls, implements item (3) 2839 switch( nop ) { 2840 // Count all float operations that may use FPU 2841 case Op_AddF: 2842 case Op_SubF: 2843 case Op_MulF: 2844 case Op_DivF: 2845 case Op_NegF: 2846 case Op_ModF: 2847 case Op_ConvI2F: 2848 case Op_ConF: 2849 case Op_CmpF: 2850 case Op_CmpF3: 2851 // case Op_ConvL2F: // longs are split into 32-bit halves 2852 frc.inc_float_count(); 2853 break; 2854 2855 case Op_ConvF2D: 2856 case Op_ConvD2F: 2857 frc.inc_float_count(); 2858 frc.inc_double_count(); 2859 break; 2860 2861 // Count all double operations that may use FPU 2862 case Op_AddD: 2863 case Op_SubD: 2864 case Op_MulD: 2865 case Op_DivD: 2866 case Op_NegD: 2867 case Op_ModD: 2868 case Op_ConvI2D: 2869 case Op_ConvD2I: 2870 // case Op_ConvL2D: // handled by leaf call 2871 // case Op_ConvD2L: // handled by leaf call 2872 case Op_ConD: 2873 case Op_CmpD: 2874 case Op_CmpD3: 2875 frc.inc_double_count(); 2876 break; 2877 case Op_Opaque1: // Remove Opaque Nodes before matching 2878 case Op_Opaque2: // Remove Opaque Nodes before matching 2879 case Op_Opaque3: 2880 n->subsume_by(n->in(1), this); 2881 break; 2882 case Op_CallStaticJava: 2883 case Op_CallJava: 2884 case Op_CallDynamicJava: 2885 frc.inc_java_call_count(); // Count java call site; 2886 case Op_CallRuntime: 2887 case Op_CallLeaf: 2888 case Op_CallLeafNoFP: { 2889 assert( n->is_Call(), "" ); 2890 CallNode *call = n->as_Call(); 2891 // Count call sites where the FP mode bit would have to be flipped. 2892 // Do not count uncommon runtime calls: 2893 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2894 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2895 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2896 frc.inc_call_count(); // Count the call site 2897 } else { // See if uncommon argument is shared 2898 Node *n = call->in(TypeFunc::Parms); 2899 int nop = n->Opcode(); 2900 // Clone shared simple arguments to uncommon calls, item (1). 2901 if( n->outcnt() > 1 && 2902 !n->is_Proj() && 2903 nop != Op_CreateEx && 2904 nop != Op_CheckCastPP && 2905 nop != Op_DecodeN && 2906 nop != Op_DecodeNKlass && 2907 !n->is_Mem() ) { 2908 Node *x = n->clone(); 2909 call->set_req( TypeFunc::Parms, x ); 2910 } 2911 } 2912 value_type_return_from_mh_intrinsic(call, frc); 2913 break; 2914 } 2915 2916 case Op_StoreD: 2917 case Op_LoadD: 2918 case Op_LoadD_unaligned: 2919 frc.inc_double_count(); 2920 goto handle_mem; 2921 case Op_StoreF: 2922 case Op_LoadF: 2923 frc.inc_float_count(); 2924 goto handle_mem; 2925 2926 case Op_StoreCM: 2927 { 2928 // Convert OopStore dependence into precedence edge 2929 Node* prec = n->in(MemNode::OopStore); 2930 n->del_req(MemNode::OopStore); 2931 n->add_prec(prec); 2932 eliminate_redundant_card_marks(n); 2933 } 2934 2935 // fall through 2936 2937 case Op_StoreB: 2938 case Op_StoreC: 2939 case Op_StorePConditional: 2940 case Op_StoreI: 2941 case Op_StoreL: 2942 case Op_StoreIConditional: 2943 case Op_StoreLConditional: 2944 case Op_CompareAndSwapB: 2945 case Op_CompareAndSwapS: 2946 case Op_CompareAndSwapI: 2947 case Op_CompareAndSwapL: 2948 case Op_CompareAndSwapP: 2949 case Op_CompareAndSwapN: 2950 case Op_WeakCompareAndSwapB: 2951 case Op_WeakCompareAndSwapS: 2952 case Op_WeakCompareAndSwapI: 2953 case Op_WeakCompareAndSwapL: 2954 case Op_WeakCompareAndSwapP: 2955 case Op_WeakCompareAndSwapN: 2956 case Op_CompareAndExchangeB: 2957 case Op_CompareAndExchangeS: 2958 case Op_CompareAndExchangeI: 2959 case Op_CompareAndExchangeL: 2960 case Op_CompareAndExchangeP: 2961 case Op_CompareAndExchangeN: 2962 case Op_GetAndAddS: 2963 case Op_GetAndAddB: 2964 case Op_GetAndAddI: 2965 case Op_GetAndAddL: 2966 case Op_GetAndSetS: 2967 case Op_GetAndSetB: 2968 case Op_GetAndSetI: 2969 case Op_GetAndSetL: 2970 case Op_GetAndSetP: 2971 case Op_GetAndSetN: 2972 case Op_StoreP: 2973 case Op_StoreN: 2974 case Op_StoreNKlass: 2975 case Op_LoadB: 2976 case Op_LoadUB: 2977 case Op_LoadUS: 2978 case Op_LoadI: 2979 case Op_LoadKlass: 2980 case Op_LoadNKlass: 2981 case Op_LoadL: 2982 case Op_LoadL_unaligned: 2983 case Op_LoadPLocked: 2984 case Op_LoadP: 2985 case Op_LoadN: 2986 case Op_LoadRange: 2987 case Op_LoadS: { 2988 handle_mem: 2989 #ifdef ASSERT 2990 if( VerifyOptoOopOffsets ) { 2991 assert( n->is_Mem(), "" ); 2992 MemNode *mem = (MemNode*)n; 2993 // Check to see if address types have grounded out somehow. 2994 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2995 assert( !tp || oop_offset_is_sane(tp), "" ); 2996 } 2997 #endif 2998 break; 2999 } 3000 3001 case Op_AddP: { // Assert sane base pointers 3002 Node *addp = n->in(AddPNode::Address); 3003 assert( !addp->is_AddP() || 3004 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 3005 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 3006 "Base pointers must match (addp %u)", addp->_idx ); 3007 #ifdef _LP64 3008 if ((UseCompressedOops || UseCompressedClassPointers) && 3009 addp->Opcode() == Op_ConP && 3010 addp == n->in(AddPNode::Base) && 3011 n->in(AddPNode::Offset)->is_Con()) { 3012 // If the transformation of ConP to ConN+DecodeN is beneficial depends 3013 // on the platform and on the compressed oops mode. 3014 // Use addressing with narrow klass to load with offset on x86. 3015 // Some platforms can use the constant pool to load ConP. 3016 // Do this transformation here since IGVN will convert ConN back to ConP. 3017 const Type* t = addp->bottom_type(); 3018 bool is_oop = t->isa_oopptr() != NULL; 3019 bool is_klass = t->isa_klassptr() != NULL; 3020 3021 if ((is_oop && Matcher::const_oop_prefer_decode() ) || 3022 (is_klass && Matcher::const_klass_prefer_decode())) { 3023 Node* nn = NULL; 3024 3025 int op = is_oop ? Op_ConN : Op_ConNKlass; 3026 3027 // Look for existing ConN node of the same exact type. 3028 Node* r = root(); 3029 uint cnt = r->outcnt(); 3030 for (uint i = 0; i < cnt; i++) { 3031 Node* m = r->raw_out(i); 3032 if (m!= NULL && m->Opcode() == op && 3033 m->bottom_type()->make_ptr() == t) { 3034 nn = m; 3035 break; 3036 } 3037 } 3038 if (nn != NULL) { 3039 // Decode a narrow oop to match address 3040 // [R12 + narrow_oop_reg<<3 + offset] 3041 if (is_oop) { 3042 nn = new DecodeNNode(nn, t); 3043 } else { 3044 nn = new DecodeNKlassNode(nn, t); 3045 } 3046 // Check for succeeding AddP which uses the same Base. 3047 // Otherwise we will run into the assertion above when visiting that guy. 3048 for (uint i = 0; i < n->outcnt(); ++i) { 3049 Node *out_i = n->raw_out(i); 3050 if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) { 3051 out_i->set_req(AddPNode::Base, nn); 3052 #ifdef ASSERT 3053 for (uint j = 0; j < out_i->outcnt(); ++j) { 3054 Node *out_j = out_i->raw_out(j); 3055 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp, 3056 "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx); 3057 } 3058 #endif 3059 } 3060 } 3061 n->set_req(AddPNode::Base, nn); 3062 n->set_req(AddPNode::Address, nn); 3063 if (addp->outcnt() == 0) { 3064 addp->disconnect_inputs(NULL, this); 3065 } 3066 } 3067 } 3068 } 3069 #endif 3070 // platform dependent reshaping of the address expression 3071 reshape_address(n->as_AddP()); 3072 break; 3073 } 3074 3075 case Op_CastPP: { 3076 // Remove CastPP nodes to gain more freedom during scheduling but 3077 // keep the dependency they encode as control or precedence edges 3078 // (if control is set already) on memory operations. Some CastPP 3079 // nodes don't have a control (don't carry a dependency): skip 3080 // those. 3081 if (n->in(0) != NULL) { 3082 ResourceMark rm; 3083 Unique_Node_List wq; 3084 wq.push(n); 3085 for (uint next = 0; next < wq.size(); ++next) { 3086 Node *m = wq.at(next); 3087 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 3088 Node* use = m->fast_out(i); 3089 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 3090 use->ensure_control_or_add_prec(n->in(0)); 3091 } else { 3092 switch(use->Opcode()) { 3093 case Op_AddP: 3094 case Op_DecodeN: 3095 case Op_DecodeNKlass: 3096 case Op_CheckCastPP: 3097 case Op_CastPP: 3098 wq.push(use); 3099 break; 3100 } 3101 } 3102 } 3103 } 3104 } 3105 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 3106 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 3107 Node* in1 = n->in(1); 3108 const Type* t = n->bottom_type(); 3109 Node* new_in1 = in1->clone(); 3110 new_in1->as_DecodeN()->set_type(t); 3111 3112 if (!Matcher::narrow_oop_use_complex_address()) { 3113 // 3114 // x86, ARM and friends can handle 2 adds in addressing mode 3115 // and Matcher can fold a DecodeN node into address by using 3116 // a narrow oop directly and do implicit NULL check in address: 3117 // 3118 // [R12 + narrow_oop_reg<<3 + offset] 3119 // NullCheck narrow_oop_reg 3120 // 3121 // On other platforms (Sparc) we have to keep new DecodeN node and 3122 // use it to do implicit NULL check in address: 3123 // 3124 // decode_not_null narrow_oop_reg, base_reg 3125 // [base_reg + offset] 3126 // NullCheck base_reg 3127 // 3128 // Pin the new DecodeN node to non-null path on these platform (Sparc) 3129 // to keep the information to which NULL check the new DecodeN node 3130 // corresponds to use it as value in implicit_null_check(). 3131 // 3132 new_in1->set_req(0, n->in(0)); 3133 } 3134 3135 n->subsume_by(new_in1, this); 3136 if (in1->outcnt() == 0) { 3137 in1->disconnect_inputs(NULL, this); 3138 } 3139 } else { 3140 n->subsume_by(n->in(1), this); 3141 if (n->outcnt() == 0) { 3142 n->disconnect_inputs(NULL, this); 3143 } 3144 } 3145 break; 3146 } 3147 #ifdef _LP64 3148 case Op_CmpP: 3149 // Do this transformation here to preserve CmpPNode::sub() and 3150 // other TypePtr related Ideal optimizations (for example, ptr nullness). 3151 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 3152 Node* in1 = n->in(1); 3153 Node* in2 = n->in(2); 3154 if (!in1->is_DecodeNarrowPtr()) { 3155 in2 = in1; 3156 in1 = n->in(2); 3157 } 3158 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3159 3160 Node* new_in2 = NULL; 3161 if (in2->is_DecodeNarrowPtr()) { 3162 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3163 new_in2 = in2->in(1); 3164 } else if (in2->Opcode() == Op_ConP) { 3165 const Type* t = in2->bottom_type(); 3166 if (t == TypePtr::NULL_PTR) { 3167 assert(in1->is_DecodeN(), "compare klass to null?"); 3168 // Don't convert CmpP null check into CmpN if compressed 3169 // oops implicit null check is not generated. 3170 // This will allow to generate normal oop implicit null check. 3171 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3172 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 3173 // 3174 // This transformation together with CastPP transformation above 3175 // will generated code for implicit NULL checks for compressed oops. 3176 // 3177 // The original code after Optimize() 3178 // 3179 // LoadN memory, narrow_oop_reg 3180 // decode narrow_oop_reg, base_reg 3181 // CmpP base_reg, NULL 3182 // CastPP base_reg // NotNull 3183 // Load [base_reg + offset], val_reg 3184 // 3185 // after these transformations will be 3186 // 3187 // LoadN memory, narrow_oop_reg 3188 // CmpN narrow_oop_reg, NULL 3189 // decode_not_null narrow_oop_reg, base_reg 3190 // Load [base_reg + offset], val_reg 3191 // 3192 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3193 // since narrow oops can be used in debug info now (see the code in 3194 // final_graph_reshaping_walk()). 3195 // 3196 // At the end the code will be matched to 3197 // on x86: 3198 // 3199 // Load_narrow_oop memory, narrow_oop_reg 3200 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3201 // NullCheck narrow_oop_reg 3202 // 3203 // and on sparc: 3204 // 3205 // Load_narrow_oop memory, narrow_oop_reg 3206 // decode_not_null narrow_oop_reg, base_reg 3207 // Load [base_reg + offset], val_reg 3208 // NullCheck base_reg 3209 // 3210 } else if (t->isa_oopptr()) { 3211 new_in2 = ConNode::make(t->make_narrowoop()); 3212 } else if (t->isa_klassptr()) { 3213 new_in2 = ConNode::make(t->make_narrowklass()); 3214 } 3215 } 3216 if (new_in2 != NULL) { 3217 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 3218 n->subsume_by(cmpN, this); 3219 if (in1->outcnt() == 0) { 3220 in1->disconnect_inputs(NULL, this); 3221 } 3222 if (in2->outcnt() == 0) { 3223 in2->disconnect_inputs(NULL, this); 3224 } 3225 } 3226 } 3227 break; 3228 3229 case Op_DecodeN: 3230 case Op_DecodeNKlass: 3231 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3232 // DecodeN could be pinned when it can't be fold into 3233 // an address expression, see the code for Op_CastPP above. 3234 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3235 break; 3236 3237 case Op_EncodeP: 3238 case Op_EncodePKlass: { 3239 Node* in1 = n->in(1); 3240 if (in1->is_DecodeNarrowPtr()) { 3241 n->subsume_by(in1->in(1), this); 3242 } else if (in1->Opcode() == Op_ConP) { 3243 const Type* t = in1->bottom_type(); 3244 if (t == TypePtr::NULL_PTR) { 3245 assert(t->isa_oopptr(), "null klass?"); 3246 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3247 } else if (t->isa_oopptr()) { 3248 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3249 } else if (t->isa_klassptr()) { 3250 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3251 } 3252 } 3253 if (in1->outcnt() == 0) { 3254 in1->disconnect_inputs(NULL, this); 3255 } 3256 break; 3257 } 3258 3259 case Op_Proj: { 3260 if (OptimizeStringConcat) { 3261 ProjNode* p = n->as_Proj(); 3262 if (p->_is_io_use) { 3263 // Separate projections were used for the exception path which 3264 // are normally removed by a late inline. If it wasn't inlined 3265 // then they will hang around and should just be replaced with 3266 // the original one. 3267 Node* proj = NULL; 3268 // Replace with just one 3269 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3270 Node *use = i.get(); 3271 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3272 proj = use; 3273 break; 3274 } 3275 } 3276 assert(proj != NULL, "must be found"); 3277 p->subsume_by(proj, this); 3278 } 3279 } 3280 break; 3281 } 3282 3283 case Op_Phi: 3284 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3285 // The EncodeP optimization may create Phi with the same edges 3286 // for all paths. It is not handled well by Register Allocator. 3287 Node* unique_in = n->in(1); 3288 assert(unique_in != NULL, ""); 3289 uint cnt = n->req(); 3290 for (uint i = 2; i < cnt; i++) { 3291 Node* m = n->in(i); 3292 assert(m != NULL, ""); 3293 if (unique_in != m) 3294 unique_in = NULL; 3295 } 3296 if (unique_in != NULL) { 3297 n->subsume_by(unique_in, this); 3298 } 3299 } 3300 break; 3301 3302 #endif 3303 3304 #ifdef ASSERT 3305 case Op_CastII: 3306 // Verify that all range check dependent CastII nodes were removed. 3307 if (n->isa_CastII()->has_range_check()) { 3308 n->dump(3); 3309 assert(false, "Range check dependent CastII node was not removed"); 3310 } 3311 break; 3312 #endif 3313 3314 case Op_ModI: 3315 if (UseDivMod) { 3316 // Check if a%b and a/b both exist 3317 Node* d = n->find_similar(Op_DivI); 3318 if (d) { 3319 // Replace them with a fused divmod if supported 3320 if (Matcher::has_match_rule(Op_DivModI)) { 3321 DivModINode* divmod = DivModINode::make(n); 3322 d->subsume_by(divmod->div_proj(), this); 3323 n->subsume_by(divmod->mod_proj(), this); 3324 } else { 3325 // replace a%b with a-((a/b)*b) 3326 Node* mult = new MulINode(d, d->in(2)); 3327 Node* sub = new SubINode(d->in(1), mult); 3328 n->subsume_by(sub, this); 3329 } 3330 } 3331 } 3332 break; 3333 3334 case Op_ModL: 3335 if (UseDivMod) { 3336 // Check if a%b and a/b both exist 3337 Node* d = n->find_similar(Op_DivL); 3338 if (d) { 3339 // Replace them with a fused divmod if supported 3340 if (Matcher::has_match_rule(Op_DivModL)) { 3341 DivModLNode* divmod = DivModLNode::make(n); 3342 d->subsume_by(divmod->div_proj(), this); 3343 n->subsume_by(divmod->mod_proj(), this); 3344 } else { 3345 // replace a%b with a-((a/b)*b) 3346 Node* mult = new MulLNode(d, d->in(2)); 3347 Node* sub = new SubLNode(d->in(1), mult); 3348 n->subsume_by(sub, this); 3349 } 3350 } 3351 } 3352 break; 3353 3354 case Op_LoadVector: 3355 case Op_StoreVector: 3356 break; 3357 3358 case Op_AddReductionVI: 3359 case Op_AddReductionVL: 3360 case Op_AddReductionVF: 3361 case Op_AddReductionVD: 3362 case Op_MulReductionVI: 3363 case Op_MulReductionVL: 3364 case Op_MulReductionVF: 3365 case Op_MulReductionVD: 3366 break; 3367 3368 case Op_PackB: 3369 case Op_PackS: 3370 case Op_PackI: 3371 case Op_PackF: 3372 case Op_PackL: 3373 case Op_PackD: 3374 if (n->req()-1 > 2) { 3375 // Replace many operand PackNodes with a binary tree for matching 3376 PackNode* p = (PackNode*) n; 3377 Node* btp = p->binary_tree_pack(1, n->req()); 3378 n->subsume_by(btp, this); 3379 } 3380 break; 3381 case Op_Loop: 3382 case Op_CountedLoop: 3383 if (n->as_Loop()->is_inner_loop()) { 3384 frc.inc_inner_loop_count(); 3385 } 3386 break; 3387 case Op_LShiftI: 3388 case Op_RShiftI: 3389 case Op_URShiftI: 3390 case Op_LShiftL: 3391 case Op_RShiftL: 3392 case Op_URShiftL: 3393 if (Matcher::need_masked_shift_count) { 3394 // The cpu's shift instructions don't restrict the count to the 3395 // lower 5/6 bits. We need to do the masking ourselves. 3396 Node* in2 = n->in(2); 3397 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3398 const TypeInt* t = in2->find_int_type(); 3399 if (t != NULL && t->is_con()) { 3400 juint shift = t->get_con(); 3401 if (shift > mask) { // Unsigned cmp 3402 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3403 } 3404 } else { 3405 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3406 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3407 n->set_req(2, shift); 3408 } 3409 } 3410 if (in2->outcnt() == 0) { // Remove dead node 3411 in2->disconnect_inputs(NULL, this); 3412 } 3413 } 3414 break; 3415 case Op_MemBarStoreStore: 3416 case Op_MemBarRelease: 3417 // Break the link with AllocateNode: it is no longer useful and 3418 // confuses register allocation. 3419 if (n->req() > MemBarNode::Precedent) { 3420 n->set_req(MemBarNode::Precedent, top()); 3421 } 3422 break; 3423 case Op_RangeCheck: { 3424 RangeCheckNode* rc = n->as_RangeCheck(); 3425 Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt); 3426 n->subsume_by(iff, this); 3427 frc._tests.push(iff); 3428 break; 3429 } 3430 case Op_ConvI2L: { 3431 if (!Matcher::convi2l_type_required) { 3432 // Code generation on some platforms doesn't need accurate 3433 // ConvI2L types. Widening the type can help remove redundant 3434 // address computations. 3435 n->as_Type()->set_type(TypeLong::INT); 3436 ResourceMark rm; 3437 Node_List wq; 3438 wq.push(n); 3439 for (uint next = 0; next < wq.size(); next++) { 3440 Node *m = wq.at(next); 3441 3442 for(;;) { 3443 // Loop over all nodes with identical inputs edges as m 3444 Node* k = m->find_similar(m->Opcode()); 3445 if (k == NULL) { 3446 break; 3447 } 3448 // Push their uses so we get a chance to remove node made 3449 // redundant 3450 for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) { 3451 Node* u = k->fast_out(i); 3452 assert(!wq.contains(u), "shouldn't process one node several times"); 3453 if (u->Opcode() == Op_LShiftL || 3454 u->Opcode() == Op_AddL || 3455 u->Opcode() == Op_SubL || 3456 u->Opcode() == Op_AddP) { 3457 wq.push(u); 3458 } 3459 } 3460 // Replace all nodes with identical edges as m with m 3461 k->subsume_by(m, this); 3462 } 3463 } 3464 } 3465 break; 3466 } 3467 case Op_ValueType: { 3468 ValueTypeNode* vt = n->as_ValueType(); 3469 vt->make_scalar_in_safepoints(root(), NULL); 3470 if (vt->outcnt() == 0) { 3471 vt->disconnect_inputs(NULL, this); 3472 } 3473 break; 3474 } 3475 case Op_ValueTypePtr: { 3476 ShouldNotReachHere(); 3477 break; 3478 } 3479 default: 3480 assert( !n->is_Call(), "" ); 3481 assert( !n->is_Mem(), "" ); 3482 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3483 break; 3484 } 3485 3486 // Collect CFG split points 3487 if (n->is_MultiBranch() && !n->is_RangeCheck()) { 3488 frc._tests.push(n); 3489 } 3490 } 3491 3492 //------------------------------final_graph_reshaping_walk--------------------- 3493 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3494 // requires that the walk visits a node's inputs before visiting the node. 3495 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3496 ResourceArea *area = Thread::current()->resource_area(); 3497 Unique_Node_List sfpt(area); 3498 3499 frc._visited.set(root->_idx); // first, mark node as visited 3500 uint cnt = root->req(); 3501 Node *n = root; 3502 uint i = 0; 3503 while (true) { 3504 if (i < cnt) { 3505 // Place all non-visited non-null inputs onto stack 3506 Node* m = n->in(i); 3507 ++i; 3508 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3509 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3510 // compute worst case interpreter size in case of a deoptimization 3511 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3512 3513 sfpt.push(m); 3514 } 3515 cnt = m->req(); 3516 nstack.push(n, i); // put on stack parent and next input's index 3517 n = m; 3518 i = 0; 3519 } 3520 } else { 3521 // Now do post-visit work 3522 final_graph_reshaping_impl( n, frc ); 3523 if (nstack.is_empty()) 3524 break; // finished 3525 n = nstack.node(); // Get node from stack 3526 cnt = n->req(); 3527 i = nstack.index(); 3528 nstack.pop(); // Shift to the next node on stack 3529 } 3530 } 3531 3532 // Skip next transformation if compressed oops are not used. 3533 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3534 (!UseCompressedOops && !UseCompressedClassPointers)) 3535 return; 3536 3537 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3538 // It could be done for an uncommon traps or any safepoints/calls 3539 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3540 while (sfpt.size() > 0) { 3541 n = sfpt.pop(); 3542 JVMState *jvms = n->as_SafePoint()->jvms(); 3543 assert(jvms != NULL, "sanity"); 3544 int start = jvms->debug_start(); 3545 int end = n->req(); 3546 bool is_uncommon = (n->is_CallStaticJava() && 3547 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3548 for (int j = start; j < end; j++) { 3549 Node* in = n->in(j); 3550 if (in->is_DecodeNarrowPtr()) { 3551 bool safe_to_skip = true; 3552 if (!is_uncommon ) { 3553 // Is it safe to skip? 3554 for (uint i = 0; i < in->outcnt(); i++) { 3555 Node* u = in->raw_out(i); 3556 if (!u->is_SafePoint() || 3557 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3558 safe_to_skip = false; 3559 } 3560 } 3561 } 3562 if (safe_to_skip) { 3563 n->set_req(j, in->in(1)); 3564 } 3565 if (in->outcnt() == 0) { 3566 in->disconnect_inputs(NULL, this); 3567 } 3568 } 3569 } 3570 } 3571 } 3572 3573 //------------------------------final_graph_reshaping-------------------------- 3574 // Final Graph Reshaping. 3575 // 3576 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3577 // and not commoned up and forced early. Must come after regular 3578 // optimizations to avoid GVN undoing the cloning. Clone constant 3579 // inputs to Loop Phis; these will be split by the allocator anyways. 3580 // Remove Opaque nodes. 3581 // (2) Move last-uses by commutative operations to the left input to encourage 3582 // Intel update-in-place two-address operations and better register usage 3583 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3584 // calls canonicalizing them back. 3585 // (3) Count the number of double-precision FP ops, single-precision FP ops 3586 // and call sites. On Intel, we can get correct rounding either by 3587 // forcing singles to memory (requires extra stores and loads after each 3588 // FP bytecode) or we can set a rounding mode bit (requires setting and 3589 // clearing the mode bit around call sites). The mode bit is only used 3590 // if the relative frequency of single FP ops to calls is low enough. 3591 // This is a key transform for SPEC mpeg_audio. 3592 // (4) Detect infinite loops; blobs of code reachable from above but not 3593 // below. Several of the Code_Gen algorithms fail on such code shapes, 3594 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3595 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3596 // Detection is by looking for IfNodes where only 1 projection is 3597 // reachable from below or CatchNodes missing some targets. 3598 // (5) Assert for insane oop offsets in debug mode. 3599 3600 bool Compile::final_graph_reshaping() { 3601 // an infinite loop may have been eliminated by the optimizer, 3602 // in which case the graph will be empty. 3603 if (root()->req() == 1) { 3604 record_method_not_compilable("trivial infinite loop"); 3605 return true; 3606 } 3607 3608 // Expensive nodes have their control input set to prevent the GVN 3609 // from freely commoning them. There's no GVN beyond this point so 3610 // no need to keep the control input. We want the expensive nodes to 3611 // be freely moved to the least frequent code path by gcm. 3612 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3613 for (int i = 0; i < expensive_count(); i++) { 3614 _expensive_nodes->at(i)->set_req(0, NULL); 3615 } 3616 3617 Final_Reshape_Counts frc; 3618 3619 // Visit everybody reachable! 3620 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3621 Node_Stack nstack(live_nodes() >> 1); 3622 final_graph_reshaping_walk(nstack, root(), frc); 3623 3624 // Check for unreachable (from below) code (i.e., infinite loops). 3625 for( uint i = 0; i < frc._tests.size(); i++ ) { 3626 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3627 // Get number of CFG targets. 3628 // Note that PCTables include exception targets after calls. 3629 uint required_outcnt = n->required_outcnt(); 3630 if (n->outcnt() != required_outcnt) { 3631 // Check for a few special cases. Rethrow Nodes never take the 3632 // 'fall-thru' path, so expected kids is 1 less. 3633 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3634 if (n->in(0)->in(0)->is_Call()) { 3635 CallNode *call = n->in(0)->in(0)->as_Call(); 3636 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3637 required_outcnt--; // Rethrow always has 1 less kid 3638 } else if (call->req() > TypeFunc::Parms && 3639 call->is_CallDynamicJava()) { 3640 // Check for null receiver. In such case, the optimizer has 3641 // detected that the virtual call will always result in a null 3642 // pointer exception. The fall-through projection of this CatchNode 3643 // will not be populated. 3644 Node *arg0 = call->in(TypeFunc::Parms); 3645 if (arg0->is_Type() && 3646 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3647 required_outcnt--; 3648 } 3649 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3650 call->req() > TypeFunc::Parms+1 && 3651 call->is_CallStaticJava()) { 3652 // Check for negative array length. In such case, the optimizer has 3653 // detected that the allocation attempt will always result in an 3654 // exception. There is no fall-through projection of this CatchNode . 3655 Node *arg1 = call->in(TypeFunc::Parms+1); 3656 if (arg1->is_Type() && 3657 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3658 required_outcnt--; 3659 } 3660 } 3661 } 3662 } 3663 // Recheck with a better notion of 'required_outcnt' 3664 if (n->outcnt() != required_outcnt) { 3665 record_method_not_compilable("malformed control flow"); 3666 return true; // Not all targets reachable! 3667 } 3668 } 3669 // Check that I actually visited all kids. Unreached kids 3670 // must be infinite loops. 3671 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3672 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3673 record_method_not_compilable("infinite loop"); 3674 return true; // Found unvisited kid; must be unreach 3675 } 3676 } 3677 3678 // If original bytecodes contained a mixture of floats and doubles 3679 // check if the optimizer has made it homogenous, item (3). 3680 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3681 frc.get_float_count() > 32 && 3682 frc.get_double_count() == 0 && 3683 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3684 set_24_bit_selection_and_mode( false, true ); 3685 } 3686 3687 set_java_calls(frc.get_java_call_count()); 3688 set_inner_loops(frc.get_inner_loop_count()); 3689 3690 // No infinite loops, no reason to bail out. 3691 return false; 3692 } 3693 3694 //-----------------------------too_many_traps---------------------------------- 3695 // Report if there are too many traps at the current method and bci. 3696 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3697 bool Compile::too_many_traps(ciMethod* method, 3698 int bci, 3699 Deoptimization::DeoptReason reason) { 3700 ciMethodData* md = method->method_data(); 3701 if (md->is_empty()) { 3702 // Assume the trap has not occurred, or that it occurred only 3703 // because of a transient condition during start-up in the interpreter. 3704 return false; 3705 } 3706 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3707 if (md->has_trap_at(bci, m, reason) != 0) { 3708 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3709 // Also, if there are multiple reasons, or if there is no per-BCI record, 3710 // assume the worst. 3711 if (log()) 3712 log()->elem("observe trap='%s' count='%d'", 3713 Deoptimization::trap_reason_name(reason), 3714 md->trap_count(reason)); 3715 return true; 3716 } else { 3717 // Ignore method/bci and see if there have been too many globally. 3718 return too_many_traps(reason, md); 3719 } 3720 } 3721 3722 // Less-accurate variant which does not require a method and bci. 3723 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3724 ciMethodData* logmd) { 3725 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3726 // Too many traps globally. 3727 // Note that we use cumulative trap_count, not just md->trap_count. 3728 if (log()) { 3729 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3730 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3731 Deoptimization::trap_reason_name(reason), 3732 mcount, trap_count(reason)); 3733 } 3734 return true; 3735 } else { 3736 // The coast is clear. 3737 return false; 3738 } 3739 } 3740 3741 //--------------------------too_many_recompiles-------------------------------- 3742 // Report if there are too many recompiles at the current method and bci. 3743 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3744 // Is not eager to return true, since this will cause the compiler to use 3745 // Action_none for a trap point, to avoid too many recompilations. 3746 bool Compile::too_many_recompiles(ciMethod* method, 3747 int bci, 3748 Deoptimization::DeoptReason reason) { 3749 ciMethodData* md = method->method_data(); 3750 if (md->is_empty()) { 3751 // Assume the trap has not occurred, or that it occurred only 3752 // because of a transient condition during start-up in the interpreter. 3753 return false; 3754 } 3755 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3756 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3757 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3758 Deoptimization::DeoptReason per_bc_reason 3759 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3760 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3761 if ((per_bc_reason == Deoptimization::Reason_none 3762 || md->has_trap_at(bci, m, reason) != 0) 3763 // The trap frequency measure we care about is the recompile count: 3764 && md->trap_recompiled_at(bci, m) 3765 && md->overflow_recompile_count() >= bc_cutoff) { 3766 // Do not emit a trap here if it has already caused recompilations. 3767 // Also, if there are multiple reasons, or if there is no per-BCI record, 3768 // assume the worst. 3769 if (log()) 3770 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3771 Deoptimization::trap_reason_name(reason), 3772 md->trap_count(reason), 3773 md->overflow_recompile_count()); 3774 return true; 3775 } else if (trap_count(reason) != 0 3776 && decompile_count() >= m_cutoff) { 3777 // Too many recompiles globally, and we have seen this sort of trap. 3778 // Use cumulative decompile_count, not just md->decompile_count. 3779 if (log()) 3780 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3781 Deoptimization::trap_reason_name(reason), 3782 md->trap_count(reason), trap_count(reason), 3783 md->decompile_count(), decompile_count()); 3784 return true; 3785 } else { 3786 // The coast is clear. 3787 return false; 3788 } 3789 } 3790 3791 // Compute when not to trap. Used by matching trap based nodes and 3792 // NullCheck optimization. 3793 void Compile::set_allowed_deopt_reasons() { 3794 _allowed_reasons = 0; 3795 if (is_method_compilation()) { 3796 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3797 assert(rs < BitsPerInt, "recode bit map"); 3798 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3799 _allowed_reasons |= nth_bit(rs); 3800 } 3801 } 3802 } 3803 } 3804 3805 #ifndef PRODUCT 3806 //------------------------------verify_graph_edges--------------------------- 3807 // Walk the Graph and verify that there is a one-to-one correspondence 3808 // between Use-Def edges and Def-Use edges in the graph. 3809 void Compile::verify_graph_edges(bool no_dead_code) { 3810 if (VerifyGraphEdges) { 3811 ResourceArea *area = Thread::current()->resource_area(); 3812 Unique_Node_List visited(area); 3813 // Call recursive graph walk to check edges 3814 _root->verify_edges(visited); 3815 if (no_dead_code) { 3816 // Now make sure that no visited node is used by an unvisited node. 3817 bool dead_nodes = false; 3818 Unique_Node_List checked(area); 3819 while (visited.size() > 0) { 3820 Node* n = visited.pop(); 3821 checked.push(n); 3822 for (uint i = 0; i < n->outcnt(); i++) { 3823 Node* use = n->raw_out(i); 3824 if (checked.member(use)) continue; // already checked 3825 if (visited.member(use)) continue; // already in the graph 3826 if (use->is_Con()) continue; // a dead ConNode is OK 3827 // At this point, we have found a dead node which is DU-reachable. 3828 if (!dead_nodes) { 3829 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3830 dead_nodes = true; 3831 } 3832 use->dump(2); 3833 tty->print_cr("---"); 3834 checked.push(use); // No repeats; pretend it is now checked. 3835 } 3836 } 3837 assert(!dead_nodes, "using nodes must be reachable from root"); 3838 } 3839 } 3840 } 3841 3842 // Verify GC barriers consistency 3843 // Currently supported: 3844 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3845 void Compile::verify_barriers() { 3846 if (UseG1GC) { 3847 // Verify G1 pre-barriers 3848 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active()); 3849 3850 ResourceArea *area = Thread::current()->resource_area(); 3851 Unique_Node_List visited(area); 3852 Node_List worklist(area); 3853 // We're going to walk control flow backwards starting from the Root 3854 worklist.push(_root); 3855 while (worklist.size() > 0) { 3856 Node* x = worklist.pop(); 3857 if (x == NULL || x == top()) continue; 3858 if (visited.member(x)) { 3859 continue; 3860 } else { 3861 visited.push(x); 3862 } 3863 3864 if (x->is_Region()) { 3865 for (uint i = 1; i < x->req(); i++) { 3866 worklist.push(x->in(i)); 3867 } 3868 } else { 3869 worklist.push(x->in(0)); 3870 // We are looking for the pattern: 3871 // /->ThreadLocal 3872 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3873 // \->ConI(0) 3874 // We want to verify that the If and the LoadB have the same control 3875 // See GraphKit::g1_write_barrier_pre() 3876 if (x->is_If()) { 3877 IfNode *iff = x->as_If(); 3878 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3879 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3880 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3881 && cmp->in(1)->is_Load()) { 3882 LoadNode* load = cmp->in(1)->as_Load(); 3883 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3884 && load->in(2)->in(3)->is_Con() 3885 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3886 3887 Node* if_ctrl = iff->in(0); 3888 Node* load_ctrl = load->in(0); 3889 3890 if (if_ctrl != load_ctrl) { 3891 // Skip possible CProj->NeverBranch in infinite loops 3892 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3893 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3894 if_ctrl = if_ctrl->in(0)->in(0); 3895 } 3896 } 3897 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3898 } 3899 } 3900 } 3901 } 3902 } 3903 } 3904 } 3905 } 3906 3907 #endif 3908 3909 // The Compile object keeps track of failure reasons separately from the ciEnv. 3910 // This is required because there is not quite a 1-1 relation between the 3911 // ciEnv and its compilation task and the Compile object. Note that one 3912 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3913 // to backtrack and retry without subsuming loads. Other than this backtracking 3914 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3915 // by the logic in C2Compiler. 3916 void Compile::record_failure(const char* reason) { 3917 if (log() != NULL) { 3918 log()->elem("failure reason='%s' phase='compile'", reason); 3919 } 3920 if (_failure_reason == NULL) { 3921 // Record the first failure reason. 3922 _failure_reason = reason; 3923 } 3924 3925 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3926 C->print_method(PHASE_FAILURE); 3927 } 3928 _root = NULL; // flush the graph, too 3929 } 3930 3931 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3932 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3933 _phase_name(name), _dolog(CITimeVerbose) 3934 { 3935 if (_dolog) { 3936 C = Compile::current(); 3937 _log = C->log(); 3938 } else { 3939 C = NULL; 3940 _log = NULL; 3941 } 3942 if (_log != NULL) { 3943 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3944 _log->stamp(); 3945 _log->end_head(); 3946 } 3947 } 3948 3949 Compile::TracePhase::~TracePhase() { 3950 3951 C = Compile::current(); 3952 if (_dolog) { 3953 _log = C->log(); 3954 } else { 3955 _log = NULL; 3956 } 3957 3958 #ifdef ASSERT 3959 if (PrintIdealNodeCount) { 3960 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3961 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3962 } 3963 3964 if (VerifyIdealNodeCount) { 3965 Compile::current()->print_missing_nodes(); 3966 } 3967 #endif 3968 3969 if (_log != NULL) { 3970 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3971 } 3972 } 3973 3974 //============================================================================= 3975 // Two Constant's are equal when the type and the value are equal. 3976 bool Compile::Constant::operator==(const Constant& other) { 3977 if (type() != other.type() ) return false; 3978 if (can_be_reused() != other.can_be_reused()) return false; 3979 // For floating point values we compare the bit pattern. 3980 switch (type()) { 3981 case T_INT: 3982 case T_FLOAT: return (_v._value.i == other._v._value.i); 3983 case T_LONG: 3984 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3985 case T_OBJECT: 3986 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3987 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3988 case T_METADATA: return (_v._metadata == other._v._metadata); 3989 default: ShouldNotReachHere(); 3990 } 3991 return false; 3992 } 3993 3994 static int type_to_size_in_bytes(BasicType t) { 3995 switch (t) { 3996 case T_INT: return sizeof(jint ); 3997 case T_LONG: return sizeof(jlong ); 3998 case T_FLOAT: return sizeof(jfloat ); 3999 case T_DOUBLE: return sizeof(jdouble); 4000 case T_METADATA: return sizeof(Metadata*); 4001 // We use T_VOID as marker for jump-table entries (labels) which 4002 // need an internal word relocation. 4003 case T_VOID: 4004 case T_ADDRESS: 4005 case T_OBJECT: return sizeof(jobject); 4006 } 4007 4008 ShouldNotReachHere(); 4009 return -1; 4010 } 4011 4012 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 4013 // sort descending 4014 if (a->freq() > b->freq()) return -1; 4015 if (a->freq() < b->freq()) return 1; 4016 return 0; 4017 } 4018 4019 void Compile::ConstantTable::calculate_offsets_and_size() { 4020 // First, sort the array by frequencies. 4021 _constants.sort(qsort_comparator); 4022 4023 #ifdef ASSERT 4024 // Make sure all jump-table entries were sorted to the end of the 4025 // array (they have a negative frequency). 4026 bool found_void = false; 4027 for (int i = 0; i < _constants.length(); i++) { 4028 Constant con = _constants.at(i); 4029 if (con.type() == T_VOID) 4030 found_void = true; // jump-tables 4031 else 4032 assert(!found_void, "wrong sorting"); 4033 } 4034 #endif 4035 4036 int offset = 0; 4037 for (int i = 0; i < _constants.length(); i++) { 4038 Constant* con = _constants.adr_at(i); 4039 4040 // Align offset for type. 4041 int typesize = type_to_size_in_bytes(con->type()); 4042 offset = align_size_up(offset, typesize); 4043 con->set_offset(offset); // set constant's offset 4044 4045 if (con->type() == T_VOID) { 4046 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 4047 offset = offset + typesize * n->outcnt(); // expand jump-table 4048 } else { 4049 offset = offset + typesize; 4050 } 4051 } 4052 4053 // Align size up to the next section start (which is insts; see 4054 // CodeBuffer::align_at_start). 4055 assert(_size == -1, "already set?"); 4056 _size = align_size_up(offset, CodeEntryAlignment); 4057 } 4058 4059 void Compile::ConstantTable::emit(CodeBuffer& cb) { 4060 MacroAssembler _masm(&cb); 4061 for (int i = 0; i < _constants.length(); i++) { 4062 Constant con = _constants.at(i); 4063 address constant_addr = NULL; 4064 switch (con.type()) { 4065 case T_INT: constant_addr = _masm.int_constant( con.get_jint() ); break; 4066 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 4067 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 4068 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 4069 case T_OBJECT: { 4070 jobject obj = con.get_jobject(); 4071 int oop_index = _masm.oop_recorder()->find_index(obj); 4072 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 4073 break; 4074 } 4075 case T_ADDRESS: { 4076 address addr = (address) con.get_jobject(); 4077 constant_addr = _masm.address_constant(addr); 4078 break; 4079 } 4080 // We use T_VOID as marker for jump-table entries (labels) which 4081 // need an internal word relocation. 4082 case T_VOID: { 4083 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 4084 // Fill the jump-table with a dummy word. The real value is 4085 // filled in later in fill_jump_table. 4086 address dummy = (address) n; 4087 constant_addr = _masm.address_constant(dummy); 4088 // Expand jump-table 4089 for (uint i = 1; i < n->outcnt(); i++) { 4090 address temp_addr = _masm.address_constant(dummy + i); 4091 assert(temp_addr, "consts section too small"); 4092 } 4093 break; 4094 } 4095 case T_METADATA: { 4096 Metadata* obj = con.get_metadata(); 4097 int metadata_index = _masm.oop_recorder()->find_index(obj); 4098 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 4099 break; 4100 } 4101 default: ShouldNotReachHere(); 4102 } 4103 assert(constant_addr, "consts section too small"); 4104 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 4105 "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())); 4106 } 4107 } 4108 4109 int Compile::ConstantTable::find_offset(Constant& con) const { 4110 int idx = _constants.find(con); 4111 assert(idx != -1, "constant must be in constant table"); 4112 int offset = _constants.at(idx).offset(); 4113 assert(offset != -1, "constant table not emitted yet?"); 4114 return offset; 4115 } 4116 4117 void Compile::ConstantTable::add(Constant& con) { 4118 if (con.can_be_reused()) { 4119 int idx = _constants.find(con); 4120 if (idx != -1 && _constants.at(idx).can_be_reused()) { 4121 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 4122 return; 4123 } 4124 } 4125 (void) _constants.append(con); 4126 } 4127 4128 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 4129 Block* b = Compile::current()->cfg()->get_block_for_node(n); 4130 Constant con(type, value, b->_freq); 4131 add(con); 4132 return con; 4133 } 4134 4135 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 4136 Constant con(metadata); 4137 add(con); 4138 return con; 4139 } 4140 4141 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 4142 jvalue value; 4143 BasicType type = oper->type()->basic_type(); 4144 switch (type) { 4145 case T_LONG: value.j = oper->constantL(); break; 4146 case T_FLOAT: value.f = oper->constantF(); break; 4147 case T_DOUBLE: value.d = oper->constantD(); break; 4148 case T_OBJECT: 4149 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 4150 case T_METADATA: return add((Metadata*)oper->constant()); break; 4151 default: guarantee(false, "unhandled type: %s", type2name(type)); 4152 } 4153 return add(n, type, value); 4154 } 4155 4156 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 4157 jvalue value; 4158 // We can use the node pointer here to identify the right jump-table 4159 // as this method is called from Compile::Fill_buffer right before 4160 // the MachNodes are emitted and the jump-table is filled (means the 4161 // MachNode pointers do not change anymore). 4162 value.l = (jobject) n; 4163 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 4164 add(con); 4165 return con; 4166 } 4167 4168 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 4169 // If called from Compile::scratch_emit_size do nothing. 4170 if (Compile::current()->in_scratch_emit_size()) return; 4171 4172 assert(labels.is_nonempty(), "must be"); 4173 assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt()); 4174 4175 // Since MachConstantNode::constant_offset() also contains 4176 // table_base_offset() we need to subtract the table_base_offset() 4177 // to get the plain offset into the constant table. 4178 int offset = n->constant_offset() - table_base_offset(); 4179 4180 MacroAssembler _masm(&cb); 4181 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 4182 4183 for (uint i = 0; i < n->outcnt(); i++) { 4184 address* constant_addr = &jump_table_base[i]; 4185 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)); 4186 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 4187 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 4188 } 4189 } 4190 4191 //----------------------------static_subtype_check----------------------------- 4192 // Shortcut important common cases when superklass is exact: 4193 // (0) superklass is java.lang.Object (can occur in reflective code) 4194 // (1) subklass is already limited to a subtype of superklass => always ok 4195 // (2) subklass does not overlap with superklass => always fail 4196 // (3) superklass has NO subtypes and we can check with a simple compare. 4197 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 4198 if (StressReflectiveCode || superk == NULL || subk == NULL) { 4199 return SSC_full_test; // Let caller generate the general case. 4200 } 4201 4202 if (!EnableMVT && !EnableValhalla && superk == env()->Object_klass()) { 4203 return SSC_always_true; // (0) this test cannot fail 4204 } 4205 4206 ciType* superelem = superk; 4207 if (superelem->is_array_klass()) 4208 superelem = superelem->as_array_klass()->base_element_type(); 4209 4210 if (!subk->is_interface()) { // cannot trust static interface types yet 4211 if (subk->is_subtype_of(superk)) { 4212 return SSC_always_true; // (1) false path dead; no dynamic test needed 4213 } 4214 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 4215 !superk->is_subtype_of(subk)) { 4216 return SSC_always_false; 4217 } 4218 } 4219 4220 // If casting to an instance klass, it must have no subtypes 4221 if (superk->is_interface()) { 4222 // Cannot trust interfaces yet. 4223 // %%% S.B. superk->nof_implementors() == 1 4224 } else if (superelem->is_instance_klass()) { 4225 ciInstanceKlass* ik = superelem->as_instance_klass(); 4226 if (!ik->has_subklass() && !ik->is_interface()) { 4227 if (!ik->is_final()) { 4228 // Add a dependency if there is a chance of a later subclass. 4229 dependencies()->assert_leaf_type(ik); 4230 } 4231 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4232 } 4233 } else { 4234 // A primitive array type has no subtypes. 4235 return SSC_easy_test; // (3) caller can do a simple ptr comparison 4236 } 4237 4238 return SSC_full_test; 4239 } 4240 4241 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) { 4242 #ifdef _LP64 4243 // The scaled index operand to AddP must be a clean 64-bit value. 4244 // Java allows a 32-bit int to be incremented to a negative 4245 // value, which appears in a 64-bit register as a large 4246 // positive number. Using that large positive number as an 4247 // operand in pointer arithmetic has bad consequences. 4248 // On the other hand, 32-bit overflow is rare, and the possibility 4249 // can often be excluded, if we annotate the ConvI2L node with 4250 // a type assertion that its value is known to be a small positive 4251 // number. (The prior range check has ensured this.) 4252 // This assertion is used by ConvI2LNode::Ideal. 4253 int index_max = max_jint - 1; // array size is max_jint, index is one less 4254 if (sizetype != NULL) index_max = sizetype->_hi - 1; 4255 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax); 4256 idx = constrained_convI2L(phase, idx, iidxtype, ctrl); 4257 #endif 4258 return idx; 4259 } 4260 4261 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4262 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4263 if (ctrl != NULL) { 4264 // Express control dependency by a CastII node with a narrow type. 4265 value = new CastIINode(value, itype, false, true /* range check dependency */); 4266 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4267 // node from floating above the range check during loop optimizations. Otherwise, the 4268 // ConvI2L node may be eliminated independently of the range check, causing the data path 4269 // to become TOP while the control path is still there (although it's unreachable). 4270 value->set_req(0, ctrl); 4271 // Save CastII node to remove it after loop optimizations. 4272 phase->C->add_range_check_cast(value); 4273 value = phase->transform(value); 4274 } 4275 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4276 return phase->transform(new ConvI2LNode(value, ltype)); 4277 } 4278 4279 // The message about the current inlining is accumulated in 4280 // _print_inlining_stream and transfered into the _print_inlining_list 4281 // once we know whether inlining succeeds or not. For regular 4282 // inlining, messages are appended to the buffer pointed by 4283 // _print_inlining_idx in the _print_inlining_list. For late inlining, 4284 // a new buffer is added after _print_inlining_idx in the list. This 4285 // way we can update the inlining message for late inlining call site 4286 // when the inlining is attempted again. 4287 void Compile::print_inlining_init() { 4288 if (print_inlining() || print_intrinsics()) { 4289 _print_inlining_stream = new stringStream(); 4290 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 4291 } 4292 } 4293 4294 void Compile::print_inlining_reinit() { 4295 if (print_inlining() || print_intrinsics()) { 4296 // Re allocate buffer when we change ResourceMark 4297 _print_inlining_stream = new stringStream(); 4298 } 4299 } 4300 4301 void Compile::print_inlining_reset() { 4302 _print_inlining_stream->reset(); 4303 } 4304 4305 void Compile::print_inlining_commit() { 4306 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 4307 // Transfer the message from _print_inlining_stream to the current 4308 // _print_inlining_list buffer and clear _print_inlining_stream. 4309 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 4310 print_inlining_reset(); 4311 } 4312 4313 void Compile::print_inlining_push() { 4314 // Add new buffer to the _print_inlining_list at current position 4315 _print_inlining_idx++; 4316 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4317 } 4318 4319 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4320 return _print_inlining_list->at(_print_inlining_idx); 4321 } 4322 4323 void Compile::print_inlining_update(CallGenerator* cg) { 4324 if (print_inlining() || print_intrinsics()) { 4325 if (!cg->is_late_inline()) { 4326 if (print_inlining_current().cg() != NULL) { 4327 print_inlining_push(); 4328 } 4329 print_inlining_commit(); 4330 } else { 4331 if (print_inlining_current().cg() != cg && 4332 (print_inlining_current().cg() != NULL || 4333 print_inlining_current().ss()->size() != 0)) { 4334 print_inlining_push(); 4335 } 4336 print_inlining_commit(); 4337 print_inlining_current().set_cg(cg); 4338 } 4339 } 4340 } 4341 4342 void Compile::print_inlining_move_to(CallGenerator* cg) { 4343 // We resume inlining at a late inlining call site. Locate the 4344 // corresponding inlining buffer so that we can update it. 4345 if (print_inlining()) { 4346 for (int i = 0; i < _print_inlining_list->length(); i++) { 4347 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4348 _print_inlining_idx = i; 4349 return; 4350 } 4351 } 4352 ShouldNotReachHere(); 4353 } 4354 } 4355 4356 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4357 if (print_inlining()) { 4358 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4359 assert(print_inlining_current().cg() == cg, "wrong entry"); 4360 // replace message with new message 4361 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4362 print_inlining_commit(); 4363 print_inlining_current().set_cg(cg); 4364 } 4365 } 4366 4367 void Compile::print_inlining_assert_ready() { 4368 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4369 } 4370 4371 void Compile::process_print_inlining() { 4372 bool do_print_inlining = print_inlining() || print_intrinsics(); 4373 if (do_print_inlining || log() != NULL) { 4374 // Print inlining message for candidates that we couldn't inline 4375 // for lack of space 4376 for (int i = 0; i < _late_inlines.length(); i++) { 4377 CallGenerator* cg = _late_inlines.at(i); 4378 if (!cg->is_mh_late_inline()) { 4379 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4380 if (do_print_inlining) { 4381 cg->print_inlining_late(msg); 4382 } 4383 log_late_inline_failure(cg, msg); 4384 } 4385 } 4386 } 4387 if (do_print_inlining) { 4388 ResourceMark rm; 4389 stringStream ss; 4390 for (int i = 0; i < _print_inlining_list->length(); i++) { 4391 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4392 } 4393 size_t end = ss.size(); 4394 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4395 strncpy(_print_inlining_output, ss.base(), end+1); 4396 _print_inlining_output[end] = 0; 4397 } 4398 } 4399 4400 void Compile::dump_print_inlining() { 4401 if (_print_inlining_output != NULL) { 4402 tty->print_raw(_print_inlining_output); 4403 } 4404 } 4405 4406 void Compile::log_late_inline(CallGenerator* cg) { 4407 if (log() != NULL) { 4408 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4409 cg->unique_id()); 4410 JVMState* p = cg->call_node()->jvms(); 4411 while (p != NULL) { 4412 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4413 p = p->caller(); 4414 } 4415 log()->tail("late_inline"); 4416 } 4417 } 4418 4419 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4420 log_late_inline(cg); 4421 if (log() != NULL) { 4422 log()->inline_fail(msg); 4423 } 4424 } 4425 4426 void Compile::log_inline_id(CallGenerator* cg) { 4427 if (log() != NULL) { 4428 // The LogCompilation tool needs a unique way to identify late 4429 // inline call sites. This id must be unique for this call site in 4430 // this compilation. Try to have it unique across compilations as 4431 // well because it can be convenient when grepping through the log 4432 // file. 4433 // Distinguish OSR compilations from others in case CICountOSR is 4434 // on. 4435 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4436 cg->set_unique_id(id); 4437 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4438 } 4439 } 4440 4441 void Compile::log_inline_failure(const char* msg) { 4442 if (C->log() != NULL) { 4443 C->log()->inline_fail(msg); 4444 } 4445 } 4446 4447 4448 // Dump inlining replay data to the stream. 4449 // Don't change thread state and acquire any locks. 4450 void Compile::dump_inline_data(outputStream* out) { 4451 InlineTree* inl_tree = ilt(); 4452 if (inl_tree != NULL) { 4453 out->print(" inline %d", inl_tree->count()); 4454 inl_tree->dump_replay_data(out); 4455 } 4456 } 4457 4458 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4459 if (n1->Opcode() < n2->Opcode()) return -1; 4460 else if (n1->Opcode() > n2->Opcode()) return 1; 4461 4462 assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()); 4463 for (uint i = 1; i < n1->req(); i++) { 4464 if (n1->in(i) < n2->in(i)) return -1; 4465 else if (n1->in(i) > n2->in(i)) return 1; 4466 } 4467 4468 return 0; 4469 } 4470 4471 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4472 Node* n1 = *n1p; 4473 Node* n2 = *n2p; 4474 4475 return cmp_expensive_nodes(n1, n2); 4476 } 4477 4478 void Compile::sort_expensive_nodes() { 4479 if (!expensive_nodes_sorted()) { 4480 _expensive_nodes->sort(cmp_expensive_nodes); 4481 } 4482 } 4483 4484 bool Compile::expensive_nodes_sorted() const { 4485 for (int i = 1; i < _expensive_nodes->length(); i++) { 4486 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4487 return false; 4488 } 4489 } 4490 return true; 4491 } 4492 4493 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4494 if (_expensive_nodes->length() == 0) { 4495 return false; 4496 } 4497 4498 assert(OptimizeExpensiveOps, "optimization off?"); 4499 4500 // Take this opportunity to remove dead nodes from the list 4501 int j = 0; 4502 for (int i = 0; i < _expensive_nodes->length(); i++) { 4503 Node* n = _expensive_nodes->at(i); 4504 if (!n->is_unreachable(igvn)) { 4505 assert(n->is_expensive(), "should be expensive"); 4506 _expensive_nodes->at_put(j, n); 4507 j++; 4508 } 4509 } 4510 _expensive_nodes->trunc_to(j); 4511 4512 // Then sort the list so that similar nodes are next to each other 4513 // and check for at least two nodes of identical kind with same data 4514 // inputs. 4515 sort_expensive_nodes(); 4516 4517 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4518 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4519 return true; 4520 } 4521 } 4522 4523 return false; 4524 } 4525 4526 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4527 if (_expensive_nodes->length() == 0) { 4528 return; 4529 } 4530 4531 assert(OptimizeExpensiveOps, "optimization off?"); 4532 4533 // Sort to bring similar nodes next to each other and clear the 4534 // control input of nodes for which there's only a single copy. 4535 sort_expensive_nodes(); 4536 4537 int j = 0; 4538 int identical = 0; 4539 int i = 0; 4540 bool modified = false; 4541 for (; i < _expensive_nodes->length()-1; i++) { 4542 assert(j <= i, "can't write beyond current index"); 4543 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4544 identical++; 4545 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4546 continue; 4547 } 4548 if (identical > 0) { 4549 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4550 identical = 0; 4551 } else { 4552 Node* n = _expensive_nodes->at(i); 4553 igvn.replace_input_of(n, 0, NULL); 4554 igvn.hash_insert(n); 4555 modified = true; 4556 } 4557 } 4558 if (identical > 0) { 4559 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4560 } else if (_expensive_nodes->length() >= 1) { 4561 Node* n = _expensive_nodes->at(i); 4562 igvn.replace_input_of(n, 0, NULL); 4563 igvn.hash_insert(n); 4564 modified = true; 4565 } 4566 _expensive_nodes->trunc_to(j); 4567 if (modified) { 4568 igvn.optimize(); 4569 } 4570 } 4571 4572 void Compile::add_expensive_node(Node * n) { 4573 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4574 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4575 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4576 if (OptimizeExpensiveOps) { 4577 _expensive_nodes->append(n); 4578 } else { 4579 // Clear control input and let IGVN optimize expensive nodes if 4580 // OptimizeExpensiveOps is off. 4581 n->set_req(0, NULL); 4582 } 4583 } 4584 4585 /** 4586 * Remove the speculative part of types and clean up the graph 4587 */ 4588 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4589 if (UseTypeSpeculation) { 4590 Unique_Node_List worklist; 4591 worklist.push(root()); 4592 int modified = 0; 4593 // Go over all type nodes that carry a speculative type, drop the 4594 // speculative part of the type and enqueue the node for an igvn 4595 // which may optimize it out. 4596 for (uint next = 0; next < worklist.size(); ++next) { 4597 Node *n = worklist.at(next); 4598 if (n->is_Type()) { 4599 TypeNode* tn = n->as_Type(); 4600 const Type* t = tn->type(); 4601 const Type* t_no_spec = t->remove_speculative(); 4602 if (t_no_spec != t) { 4603 bool in_hash = igvn.hash_delete(n); 4604 assert(in_hash, "node should be in igvn hash table"); 4605 tn->set_type(t_no_spec); 4606 igvn.hash_insert(n); 4607 igvn._worklist.push(n); // give it a chance to go away 4608 modified++; 4609 } 4610 } 4611 uint max = n->len(); 4612 for( uint i = 0; i < max; ++i ) { 4613 Node *m = n->in(i); 4614 if (not_a_node(m)) continue; 4615 worklist.push(m); 4616 } 4617 } 4618 // Drop the speculative part of all types in the igvn's type table 4619 igvn.remove_speculative_types(); 4620 if (modified > 0) { 4621 igvn.optimize(); 4622 } 4623 #ifdef ASSERT 4624 // Verify that after the IGVN is over no speculative type has resurfaced 4625 worklist.clear(); 4626 worklist.push(root()); 4627 for (uint next = 0; next < worklist.size(); ++next) { 4628 Node *n = worklist.at(next); 4629 const Type* t = igvn.type_or_null(n); 4630 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4631 if (n->is_Type()) { 4632 t = n->as_Type()->type(); 4633 assert(t == t->remove_speculative(), "no more speculative types"); 4634 } 4635 uint max = n->len(); 4636 for( uint i = 0; i < max; ++i ) { 4637 Node *m = n->in(i); 4638 if (not_a_node(m)) continue; 4639 worklist.push(m); 4640 } 4641 } 4642 igvn.check_no_speculative_types(); 4643 #endif 4644 } 4645 } 4646 4647 // Auxiliary method to support randomized stressing/fuzzing. 4648 // 4649 // This method can be called the arbitrary number of times, with current count 4650 // as the argument. The logic allows selecting a single candidate from the 4651 // running list of candidates as follows: 4652 // int count = 0; 4653 // Cand* selected = null; 4654 // while(cand = cand->next()) { 4655 // if (randomized_select(++count)) { 4656 // selected = cand; 4657 // } 4658 // } 4659 // 4660 // Including count equalizes the chances any candidate is "selected". 4661 // This is useful when we don't have the complete list of candidates to choose 4662 // from uniformly. In this case, we need to adjust the randomicity of the 4663 // selection, or else we will end up biasing the selection towards the latter 4664 // candidates. 4665 // 4666 // Quick back-envelope calculation shows that for the list of n candidates 4667 // the equal probability for the candidate to persist as "best" can be 4668 // achieved by replacing it with "next" k-th candidate with the probability 4669 // of 1/k. It can be easily shown that by the end of the run, the 4670 // probability for any candidate is converged to 1/n, thus giving the 4671 // uniform distribution among all the candidates. 4672 // 4673 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4674 #define RANDOMIZED_DOMAIN_POW 29 4675 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4676 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4677 bool Compile::randomized_select(int count) { 4678 assert(count > 0, "only positive"); 4679 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4680 } 4681 4682 CloneMap& Compile::clone_map() { return _clone_map; } 4683 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4684 4685 void NodeCloneInfo::dump() const { 4686 tty->print(" {%d:%d} ", idx(), gen()); 4687 } 4688 4689 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4690 uint64_t val = value(old->_idx); 4691 NodeCloneInfo cio(val); 4692 assert(val != 0, "old node should be in the map"); 4693 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4694 insert(nnn->_idx, cin.get()); 4695 #ifndef PRODUCT 4696 if (is_debug()) { 4697 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4698 } 4699 #endif 4700 } 4701 4702 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4703 NodeCloneInfo cio(value(old->_idx)); 4704 if (cio.get() == 0) { 4705 cio.set(old->_idx, 0); 4706 insert(old->_idx, cio.get()); 4707 #ifndef PRODUCT 4708 if (is_debug()) { 4709 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4710 } 4711 #endif 4712 } 4713 clone(old, nnn, gen); 4714 } 4715 4716 int CloneMap::max_gen() const { 4717 int g = 0; 4718 DictI di(_dict); 4719 for(; di.test(); ++di) { 4720 int t = gen(di._key); 4721 if (g < t) { 4722 g = t; 4723 #ifndef PRODUCT 4724 if (is_debug()) { 4725 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4726 } 4727 #endif 4728 } 4729 } 4730 return g; 4731 } 4732 4733 void CloneMap::dump(node_idx_t key) const { 4734 uint64_t val = value(key); 4735 if (val != 0) { 4736 NodeCloneInfo ni(val); 4737 ni.dump(); 4738 } 4739 } 4740 4741 // Helper function for enforcing certain bytecodes to reexecute if 4742 // deoptimization happens 4743 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) { 4744 ciMethod* cur_method = jvms->method(); 4745 int cur_bci = jvms->bci(); 4746 if (cur_method != NULL && cur_bci != InvocationEntryBci) { 4747 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci); 4748 return Interpreter::bytecode_should_reexecute(code) || 4749 is_anewarray && code == Bytecodes::_multianewarray; 4750 // Reexecute _multianewarray bytecode which was replaced with 4751 // sequence of [a]newarray. See Parse::do_multianewarray(). 4752 // 4753 // Note: interpreter should not have it set since this optimization 4754 // is limited by dimensions and guarded by flag so in some cases 4755 // multianewarray() runtime calls will be generated and 4756 // the bytecode should not be reexecutes (stack will not be reset). 4757 } else 4758 return false; 4759 } 4760 4761 void Compile::add_safepoint_edges(SafePointNode* call, JVMState* youngest_jvms, bool can_prune_locals, uint stack_slots_not_pruned) { 4762 // do not scribble on the input jvms 4763 JVMState* out_jvms = youngest_jvms->clone_deep(C); 4764 call->set_jvms(out_jvms); // Start jvms list for call node 4765 4766 // For a known set of bytecodes, the interpreter should reexecute them if 4767 // deoptimization happens. We set the reexecute state for them here 4768 if (out_jvms->is_reexecute_undefined() && //don't change if already specified 4769 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) { 4770 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed 4771 } 4772 4773 // Presize the call: 4774 DEBUG_ONLY(uint non_debug_edges = call->req()); 4775 call->add_req_batch(top(), youngest_jvms->debug_depth()); 4776 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), ""); 4777 4778 // Set up edges so that the call looks like this: 4779 // Call [state:] ctl io mem fptr retadr 4780 // [parms:] parm0 ... parmN 4781 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 4782 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...] 4783 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN 4784 // Note that caller debug info precedes callee debug info. 4785 4786 // Fill pointer walks backwards from "young:" to "root:" in the diagram above: 4787 uint debug_ptr = call->req(); 4788 4789 // Loop over the map input edges associated with jvms, add them 4790 // to the call node, & reset all offsets to match call node array. 4791 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) { 4792 uint debug_end = debug_ptr; 4793 uint debug_start = debug_ptr - in_jvms->debug_size(); 4794 debug_ptr = debug_start; // back up the ptr 4795 4796 uint p = debug_start; // walks forward in [debug_start, debug_end) 4797 uint j, k, l; 4798 SafePointNode* in_map = in_jvms->map(); 4799 out_jvms->set_map(call); 4800 4801 if (can_prune_locals) { 4802 assert(in_jvms->method() == out_jvms->method(), "sanity"); 4803 // If the current throw can reach an exception handler in this JVMS, 4804 // then we must keep everything live that can reach that handler. 4805 // As a quick and dirty approximation, we look for any handlers at all. 4806 if (in_jvms->method()->has_exception_handlers()) { 4807 can_prune_locals = false; 4808 } 4809 } 4810 4811 // Add the Locals 4812 k = in_jvms->locoff(); 4813 l = in_jvms->loc_size(); 4814 out_jvms->set_locoff(p); 4815 if (!can_prune_locals) { 4816 for (j = 0; j < l; j++) 4817 call->set_req(p++, in_map->in(k+j)); 4818 } else { 4819 p += l; // already set to top above by add_req_batch 4820 } 4821 4822 // Add the Expression Stack 4823 k = in_jvms->stkoff(); 4824 l = in_jvms->sp(); 4825 out_jvms->set_stkoff(p); 4826 if (!can_prune_locals) { 4827 for (j = 0; j < l; j++) 4828 call->set_req(p++, in_map->in(k+j)); 4829 } else if (can_prune_locals && stack_slots_not_pruned != 0) { 4830 // Divide stack into {S0,...,S1}, where S0 is set to top. 4831 uint s1 = stack_slots_not_pruned; 4832 stack_slots_not_pruned = 0; // for next iteration 4833 if (s1 > l) s1 = l; 4834 uint s0 = l - s1; 4835 p += s0; // skip the tops preinstalled by add_req_batch 4836 for (j = s0; j < l; j++) 4837 call->set_req(p++, in_map->in(k+j)); 4838 } else { 4839 p += l; // already set to top above by add_req_batch 4840 } 4841 4842 // Add the Monitors 4843 k = in_jvms->monoff(); 4844 l = in_jvms->mon_size(); 4845 out_jvms->set_monoff(p); 4846 for (j = 0; j < l; j++) 4847 call->set_req(p++, in_map->in(k+j)); 4848 4849 // Copy any scalar object fields. 4850 k = in_jvms->scloff(); 4851 l = in_jvms->scl_size(); 4852 out_jvms->set_scloff(p); 4853 for (j = 0; j < l; j++) 4854 call->set_req(p++, in_map->in(k+j)); 4855 4856 // Finish the new jvms. 4857 out_jvms->set_endoff(p); 4858 4859 assert(out_jvms->endoff() == debug_end, "fill ptr must match"); 4860 assert(out_jvms->depth() == in_jvms->depth(), "depth must match"); 4861 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match"); 4862 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match"); 4863 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match"); 4864 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match"); 4865 4866 // Update the two tail pointers in parallel. 4867 out_jvms = out_jvms->caller(); 4868 in_jvms = in_jvms->caller(); 4869 } 4870 4871 assert(debug_ptr == non_debug_edges, "debug info must fit exactly"); 4872 4873 // Test the correctness of JVMState::debug_xxx accessors: 4874 assert(call->jvms()->debug_start() == non_debug_edges, ""); 4875 assert(call->jvms()->debug_end() == call->req(), ""); 4876 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, ""); 4877 }