1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #if INCLUDE_JFR 36 #include "jfr/jfrEvents.hpp" 37 #endif 38 #include "opto/addnode.hpp" 39 #include "opto/block.hpp" 40 #include "opto/c2compiler.hpp" 41 #include "opto/callGenerator.hpp" 42 #include "opto/callnode.hpp" 43 #include "opto/cfgnode.hpp" 44 #include "opto/chaitin.hpp" 45 #include "opto/compile.hpp" 46 #include "opto/connode.hpp" 47 #include "opto/divnode.hpp" 48 #include "opto/escape.hpp" 49 #include "opto/idealGraphPrinter.hpp" 50 #include "opto/loopnode.hpp" 51 #include "opto/machnode.hpp" 52 #include "opto/macro.hpp" 53 #include "opto/matcher.hpp" 54 #include "opto/mathexactnode.hpp" 55 #include "opto/memnode.hpp" 56 #include "opto/mulnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/signature.hpp" 69 #include "runtime/stubRoutines.hpp" 70 #include "runtime/timer.hpp" 71 #include "utilities/copy.hpp" 72 #if defined AD_MD_HPP 73 # include AD_MD_HPP 74 #elif defined TARGET_ARCH_MODEL_x86_32 75 # include "adfiles/ad_x86_32.hpp" 76 #elif defined TARGET_ARCH_MODEL_x86_64 77 # include "adfiles/ad_x86_64.hpp" 78 #elif defined TARGET_ARCH_MODEL_sparc 79 # include "adfiles/ad_sparc.hpp" 80 #elif defined TARGET_ARCH_MODEL_zero 81 # include "adfiles/ad_zero.hpp" 82 #elif defined TARGET_ARCH_MODEL_ppc_64 83 # include "adfiles/ad_ppc_64.hpp" 84 #endif 85 86 // -------------------- Compile::mach_constant_base_node ----------------------- 87 // Constant table base node singleton. 88 MachConstantBaseNode* Compile::mach_constant_base_node() { 89 if (_mach_constant_base_node == NULL) { 90 _mach_constant_base_node = new (C) MachConstantBaseNode(); 91 _mach_constant_base_node->add_req(C->root()); 92 } 93 return _mach_constant_base_node; 94 } 95 96 97 /// Support for intrinsics. 98 99 // Return the index at which m must be inserted (or already exists). 100 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 101 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 102 #ifdef ASSERT 103 for (int i = 1; i < _intrinsics->length(); i++) { 104 CallGenerator* cg1 = _intrinsics->at(i-1); 105 CallGenerator* cg2 = _intrinsics->at(i); 106 assert(cg1->method() != cg2->method() 107 ? cg1->method() < cg2->method() 108 : cg1->is_virtual() < cg2->is_virtual(), 109 "compiler intrinsics list must stay sorted"); 110 } 111 #endif 112 // Binary search sorted list, in decreasing intervals [lo, hi]. 113 int lo = 0, hi = _intrinsics->length()-1; 114 while (lo <= hi) { 115 int mid = (uint)(hi + lo) / 2; 116 ciMethod* mid_m = _intrinsics->at(mid)->method(); 117 if (m < mid_m) { 118 hi = mid-1; 119 } else if (m > mid_m) { 120 lo = mid+1; 121 } else { 122 // look at minor sort key 123 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 124 if (is_virtual < mid_virt) { 125 hi = mid-1; 126 } else if (is_virtual > mid_virt) { 127 lo = mid+1; 128 } else { 129 return mid; // exact match 130 } 131 } 132 } 133 return lo; // inexact match 134 } 135 136 void Compile::register_intrinsic(CallGenerator* cg) { 137 if (_intrinsics == NULL) { 138 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 139 } 140 // This code is stolen from ciObjectFactory::insert. 141 // Really, GrowableArray should have methods for 142 // insert_at, remove_at, and binary_search. 143 int len = _intrinsics->length(); 144 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 145 if (index == len) { 146 _intrinsics->append(cg); 147 } else { 148 #ifdef ASSERT 149 CallGenerator* oldcg = _intrinsics->at(index); 150 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 151 #endif 152 _intrinsics->append(_intrinsics->at(len-1)); 153 int pos; 154 for (pos = len-2; pos >= index; pos--) { 155 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 156 } 157 _intrinsics->at_put(index, cg); 158 } 159 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 160 } 161 162 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 163 assert(m->is_loaded(), "don't try this on unloaded methods"); 164 if (_intrinsics != NULL) { 165 int index = intrinsic_insertion_index(m, is_virtual); 166 if (index < _intrinsics->length() 167 && _intrinsics->at(index)->method() == m 168 && _intrinsics->at(index)->is_virtual() == is_virtual) { 169 return _intrinsics->at(index); 170 } 171 } 172 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 173 if (m->intrinsic_id() != vmIntrinsics::_none && 174 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 175 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 176 if (cg != NULL) { 177 // Save it for next time: 178 register_intrinsic(cg); 179 return cg; 180 } else { 181 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 182 } 183 } 184 return NULL; 185 } 186 187 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 188 // in library_call.cpp. 189 190 191 #ifndef PRODUCT 192 // statistics gathering... 193 194 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 195 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 196 197 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 198 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 199 int oflags = _intrinsic_hist_flags[id]; 200 assert(flags != 0, "what happened?"); 201 if (is_virtual) { 202 flags |= _intrinsic_virtual; 203 } 204 bool changed = (flags != oflags); 205 if ((flags & _intrinsic_worked) != 0) { 206 juint count = (_intrinsic_hist_count[id] += 1); 207 if (count == 1) { 208 changed = true; // first time 209 } 210 // increment the overall count also: 211 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 212 } 213 if (changed) { 214 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 215 // Something changed about the intrinsic's virtuality. 216 if ((flags & _intrinsic_virtual) != 0) { 217 // This is the first use of this intrinsic as a virtual call. 218 if (oflags != 0) { 219 // We already saw it as a non-virtual, so note both cases. 220 flags |= _intrinsic_both; 221 } 222 } else if ((oflags & _intrinsic_both) == 0) { 223 // This is the first use of this intrinsic as a non-virtual 224 flags |= _intrinsic_both; 225 } 226 } 227 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 228 } 229 // update the overall flags also: 230 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 231 return changed; 232 } 233 234 static char* format_flags(int flags, char* buf) { 235 buf[0] = 0; 236 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 237 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 238 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 239 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 240 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 241 if (buf[0] == 0) strcat(buf, ","); 242 assert(buf[0] == ',', "must be"); 243 return &buf[1]; 244 } 245 246 void Compile::print_intrinsic_statistics() { 247 char flagsbuf[100]; 248 ttyLocker ttyl; 249 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 250 tty->print_cr("Compiler intrinsic usage:"); 251 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 252 if (total == 0) total = 1; // avoid div0 in case of no successes 253 #define PRINT_STAT_LINE(name, c, f) \ 254 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 255 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 256 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 257 int flags = _intrinsic_hist_flags[id]; 258 juint count = _intrinsic_hist_count[id]; 259 if ((flags | count) != 0) { 260 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 261 } 262 } 263 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 264 if (xtty != NULL) xtty->tail("statistics"); 265 } 266 267 void Compile::print_statistics() { 268 { ttyLocker ttyl; 269 if (xtty != NULL) xtty->head("statistics type='opto'"); 270 Parse::print_statistics(); 271 PhaseCCP::print_statistics(); 272 PhaseRegAlloc::print_statistics(); 273 Scheduling::print_statistics(); 274 PhasePeephole::print_statistics(); 275 PhaseIdealLoop::print_statistics(); 276 if (xtty != NULL) xtty->tail("statistics"); 277 } 278 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 279 // put this under its own <statistics> element. 280 print_intrinsic_statistics(); 281 } 282 } 283 #endif //PRODUCT 284 285 // Support for bundling info 286 Bundle* Compile::node_bundling(const Node *n) { 287 assert(valid_bundle_info(n), "oob"); 288 return &_node_bundling_base[n->_idx]; 289 } 290 291 bool Compile::valid_bundle_info(const Node *n) { 292 return (_node_bundling_limit > n->_idx); 293 } 294 295 296 void Compile::gvn_replace_by(Node* n, Node* nn) { 297 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 298 Node* use = n->last_out(i); 299 bool is_in_table = initial_gvn()->hash_delete(use); 300 uint uses_found = 0; 301 for (uint j = 0; j < use->len(); j++) { 302 if (use->in(j) == n) { 303 if (j < use->req()) 304 use->set_req(j, nn); 305 else 306 use->set_prec(j, nn); 307 uses_found++; 308 } 309 } 310 if (is_in_table) { 311 // reinsert into table 312 initial_gvn()->hash_find_insert(use); 313 } 314 record_for_igvn(use); 315 i -= uses_found; // we deleted 1 or more copies of this edge 316 } 317 } 318 319 320 static inline bool not_a_node(const Node* n) { 321 if (n == NULL) return true; 322 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 323 if (*(address*)n == badAddress) return true; // kill by Node::destruct 324 return false; 325 } 326 327 // Identify all nodes that are reachable from below, useful. 328 // Use breadth-first pass that records state in a Unique_Node_List, 329 // recursive traversal is slower. 330 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 331 int estimated_worklist_size = live_nodes(); 332 useful.map( estimated_worklist_size, NULL ); // preallocate space 333 334 // Initialize worklist 335 if (root() != NULL) { useful.push(root()); } 336 // If 'top' is cached, declare it useful to preserve cached node 337 if( cached_top_node() ) { useful.push(cached_top_node()); } 338 339 // Push all useful nodes onto the list, breadthfirst 340 for( uint next = 0; next < useful.size(); ++next ) { 341 assert( next < unique(), "Unique useful nodes < total nodes"); 342 Node *n = useful.at(next); 343 uint max = n->len(); 344 for( uint i = 0; i < max; ++i ) { 345 Node *m = n->in(i); 346 if (not_a_node(m)) continue; 347 useful.push(m); 348 } 349 } 350 } 351 352 // Update dead_node_list with any missing dead nodes using useful 353 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 354 void Compile::update_dead_node_list(Unique_Node_List &useful) { 355 uint max_idx = unique(); 356 VectorSet& useful_node_set = useful.member_set(); 357 358 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 359 // If node with index node_idx is not in useful set, 360 // mark it as dead in dead node list. 361 if (! useful_node_set.test(node_idx) ) { 362 record_dead_node(node_idx); 363 } 364 } 365 } 366 367 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 368 int shift = 0; 369 for (int i = 0; i < inlines->length(); i++) { 370 CallGenerator* cg = inlines->at(i); 371 CallNode* call = cg->call_node(); 372 if (shift > 0) { 373 inlines->at_put(i-shift, cg); 374 } 375 if (!useful.member(call)) { 376 shift++; 377 } 378 } 379 inlines->trunc_to(inlines->length()-shift); 380 } 381 382 // Disconnect all useless nodes by disconnecting those at the boundary. 383 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 384 uint next = 0; 385 while (next < useful.size()) { 386 Node *n = useful.at(next++); 387 if (n->is_SafePoint()) { 388 // We're done with a parsing phase. Replaced nodes are not valid 389 // beyond that point. 390 n->as_SafePoint()->delete_replaced_nodes(); 391 } 392 // Use raw traversal of out edges since this code removes out edges 393 int max = n->outcnt(); 394 for (int j = 0; j < max; ++j) { 395 Node* child = n->raw_out(j); 396 if (! useful.member(child)) { 397 assert(!child->is_top() || child != top(), 398 "If top is cached in Compile object it is in useful list"); 399 // Only need to remove this out-edge to the useless node 400 n->raw_del_out(j); 401 --j; 402 --max; 403 } 404 } 405 if (n->outcnt() == 1 && n->has_special_unique_user()) { 406 record_for_igvn(n->unique_out()); 407 } 408 } 409 // Remove useless macro and predicate opaq nodes 410 for (int i = C->macro_count()-1; i >= 0; i--) { 411 Node* n = C->macro_node(i); 412 if (!useful.member(n)) { 413 remove_macro_node(n); 414 } 415 } 416 // Remove useless CastII nodes with range check dependency 417 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 418 Node* cast = range_check_cast_node(i); 419 if (!useful.member(cast)) { 420 remove_range_check_cast(cast); 421 } 422 } 423 // Remove useless expensive node 424 for (int i = C->expensive_count()-1; i >= 0; i--) { 425 Node* n = C->expensive_node(i); 426 if (!useful.member(n)) { 427 remove_expensive_node(n); 428 } 429 } 430 // clean up the late inline lists 431 remove_useless_late_inlines(&_string_late_inlines, useful); 432 remove_useless_late_inlines(&_boxing_late_inlines, useful); 433 remove_useless_late_inlines(&_late_inlines, useful); 434 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 435 } 436 437 //------------------------------frame_size_in_words----------------------------- 438 // frame_slots in units of words 439 int Compile::frame_size_in_words() const { 440 // shift is 0 in LP32 and 1 in LP64 441 const int shift = (LogBytesPerWord - LogBytesPerInt); 442 int words = _frame_slots >> shift; 443 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 444 return words; 445 } 446 447 // To bang the stack of this compiled method we use the stack size 448 // that the interpreter would need in case of a deoptimization. This 449 // removes the need to bang the stack in the deoptimization blob which 450 // in turn simplifies stack overflow handling. 451 int Compile::bang_size_in_bytes() const { 452 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 453 } 454 455 // ============================================================================ 456 //------------------------------CompileWrapper--------------------------------- 457 class CompileWrapper : public StackObj { 458 Compile *const _compile; 459 public: 460 CompileWrapper(Compile* compile); 461 462 ~CompileWrapper(); 463 }; 464 465 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 466 // the Compile* pointer is stored in the current ciEnv: 467 ciEnv* env = compile->env(); 468 assert(env == ciEnv::current(), "must already be a ciEnv active"); 469 assert(env->compiler_data() == NULL, "compile already active?"); 470 env->set_compiler_data(compile); 471 assert(compile == Compile::current(), "sanity"); 472 473 compile->set_type_dict(NULL); 474 compile->set_type_hwm(NULL); 475 compile->set_type_last_size(0); 476 compile->set_last_tf(NULL, NULL); 477 compile->set_indexSet_arena(NULL); 478 compile->set_indexSet_free_block_list(NULL); 479 compile->init_type_arena(); 480 Type::Initialize(compile); 481 _compile->set_scratch_buffer_blob(NULL); 482 _compile->begin_method(); 483 } 484 CompileWrapper::~CompileWrapper() { 485 _compile->end_method(); 486 if (_compile->scratch_buffer_blob() != NULL) 487 BufferBlob::free(_compile->scratch_buffer_blob()); 488 _compile->env()->set_compiler_data(NULL); 489 } 490 491 492 //----------------------------print_compile_messages--------------------------- 493 void Compile::print_compile_messages() { 494 #ifndef PRODUCT 495 // Check if recompiling 496 if (_subsume_loads == false && PrintOpto) { 497 // Recompiling without allowing machine instructions to subsume loads 498 tty->print_cr("*********************************************************"); 499 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 500 tty->print_cr("*********************************************************"); 501 } 502 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 503 // Recompiling without escape analysis 504 tty->print_cr("*********************************************************"); 505 tty->print_cr("** Bailout: Recompile without escape analysis **"); 506 tty->print_cr("*********************************************************"); 507 } 508 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 509 // Recompiling without boxing elimination 510 tty->print_cr("*********************************************************"); 511 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 512 tty->print_cr("*********************************************************"); 513 } 514 if (env()->break_at_compile()) { 515 // Open the debugger when compiling this method. 516 tty->print("### Breaking when compiling: "); 517 method()->print_short_name(); 518 tty->cr(); 519 BREAKPOINT; 520 } 521 522 if( PrintOpto ) { 523 if (is_osr_compilation()) { 524 tty->print("[OSR]%3d", _compile_id); 525 } else { 526 tty->print("%3d", _compile_id); 527 } 528 } 529 #endif 530 } 531 532 533 //-----------------------init_scratch_buffer_blob------------------------------ 534 // Construct a temporary BufferBlob and cache it for this compile. 535 void Compile::init_scratch_buffer_blob(int const_size) { 536 // If there is already a scratch buffer blob allocated and the 537 // constant section is big enough, use it. Otherwise free the 538 // current and allocate a new one. 539 BufferBlob* blob = scratch_buffer_blob(); 540 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 541 // Use the current blob. 542 } else { 543 if (blob != NULL) { 544 BufferBlob::free(blob); 545 } 546 547 ResourceMark rm; 548 _scratch_const_size = const_size; 549 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 550 blob = BufferBlob::create("Compile::scratch_buffer", size); 551 // Record the buffer blob for next time. 552 set_scratch_buffer_blob(blob); 553 // Have we run out of code space? 554 if (scratch_buffer_blob() == NULL) { 555 // Let CompilerBroker disable further compilations. 556 record_failure("Not enough space for scratch buffer in CodeCache"); 557 return; 558 } 559 } 560 561 // Initialize the relocation buffers 562 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 563 set_scratch_locs_memory(locs_buf); 564 } 565 566 567 //-----------------------scratch_emit_size------------------------------------- 568 // Helper function that computes size by emitting code 569 uint Compile::scratch_emit_size(const Node* n) { 570 // Start scratch_emit_size section. 571 set_in_scratch_emit_size(true); 572 573 // Emit into a trash buffer and count bytes emitted. 574 // This is a pretty expensive way to compute a size, 575 // but it works well enough if seldom used. 576 // All common fixed-size instructions are given a size 577 // method by the AD file. 578 // Note that the scratch buffer blob and locs memory are 579 // allocated at the beginning of the compile task, and 580 // may be shared by several calls to scratch_emit_size. 581 // The allocation of the scratch buffer blob is particularly 582 // expensive, since it has to grab the code cache lock. 583 BufferBlob* blob = this->scratch_buffer_blob(); 584 assert(blob != NULL, "Initialize BufferBlob at start"); 585 assert(blob->size() > MAX_inst_size, "sanity"); 586 relocInfo* locs_buf = scratch_locs_memory(); 587 address blob_begin = blob->content_begin(); 588 address blob_end = (address)locs_buf; 589 assert(blob->content_contains(blob_end), "sanity"); 590 CodeBuffer buf(blob_begin, blob_end - blob_begin); 591 buf.initialize_consts_size(_scratch_const_size); 592 buf.initialize_stubs_size(MAX_stubs_size); 593 assert(locs_buf != NULL, "sanity"); 594 int lsize = MAX_locs_size / 3; 595 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 596 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 597 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 598 599 // Do the emission. 600 601 Label fakeL; // Fake label for branch instructions. 602 Label* saveL = NULL; 603 uint save_bnum = 0; 604 bool is_branch = n->is_MachBranch(); 605 if (is_branch) { 606 MacroAssembler masm(&buf); 607 masm.bind(fakeL); 608 n->as_MachBranch()->save_label(&saveL, &save_bnum); 609 n->as_MachBranch()->label_set(&fakeL, 0); 610 } 611 n->emit(buf, this->regalloc()); 612 613 // Emitting into the scratch buffer should not fail 614 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 615 616 if (is_branch) // Restore label. 617 n->as_MachBranch()->label_set(saveL, save_bnum); 618 619 // End scratch_emit_size section. 620 set_in_scratch_emit_size(false); 621 622 return buf.insts_size(); 623 } 624 625 626 // ============================================================================ 627 //------------------------------Compile standard------------------------------- 628 debug_only( int Compile::_debug_idx = 100000; ) 629 630 // Compile a method. entry_bci is -1 for normal compilations and indicates 631 // the continuation bci for on stack replacement. 632 633 634 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 635 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 636 : Phase(Compiler), 637 _env(ci_env), 638 _log(ci_env->log()), 639 _compile_id(ci_env->compile_id()), 640 _save_argument_registers(false), 641 _stub_name(NULL), 642 _stub_function(NULL), 643 _stub_entry_point(NULL), 644 _method(target), 645 _entry_bci(osr_bci), 646 _initial_gvn(NULL), 647 _for_igvn(NULL), 648 _warm_calls(NULL), 649 _subsume_loads(subsume_loads), 650 _do_escape_analysis(do_escape_analysis), 651 _eliminate_boxing(eliminate_boxing), 652 _failure_reason(NULL), 653 _code_buffer("Compile::Fill_buffer"), 654 _orig_pc_slot(0), 655 _orig_pc_slot_offset_in_bytes(0), 656 _has_method_handle_invokes(false), 657 _mach_constant_base_node(NULL), 658 _node_bundling_limit(0), 659 _node_bundling_base(NULL), 660 _java_calls(0), 661 _inner_loops(0), 662 _scratch_const_size(-1), 663 _in_scratch_emit_size(false), 664 _dead_node_list(comp_arena()), 665 _dead_node_count(0), 666 #ifndef PRODUCT 667 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 668 _in_dump_cnt(0), 669 _printer(IdealGraphPrinter::printer()), 670 #endif 671 _congraph(NULL), 672 _comp_arena(mtCompiler), 673 _node_arena(mtCompiler), 674 _old_arena(mtCompiler), 675 _Compile_types(mtCompiler), 676 _replay_inline_data(NULL), 677 _late_inlines(comp_arena(), 2, 0, NULL), 678 _string_late_inlines(comp_arena(), 2, 0, NULL), 679 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 680 _late_inlines_pos(0), 681 _number_of_mh_late_inlines(0), 682 _inlining_progress(false), 683 _inlining_incrementally(false), 684 _print_inlining_list(NULL), 685 _print_inlining_idx(0), 686 _interpreter_frame_size(0), 687 _max_node_limit(MaxNodeLimit) { 688 C = this; 689 690 CompileWrapper cw(this); 691 #ifndef PRODUCT 692 if (TimeCompiler2) { 693 tty->print(" "); 694 target->holder()->name()->print(); 695 tty->print("."); 696 target->print_short_name(); 697 tty->print(" "); 698 } 699 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 700 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 701 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 702 if (!print_opto_assembly) { 703 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 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 (method()->has_option("ReplayInline")) { 713 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 714 } 715 #endif 716 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 717 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 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 if (print_inlining() || print_intrinsics()) { 749 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 750 } 751 { // Scope for timing the parser 752 TracePhase t3("parse", &_t_parser, true); 753 754 // Put top into the hash table ASAP. 755 initial_gvn()->transform_no_reclaim(top()); 756 757 // Set up tf(), start(), and find a CallGenerator. 758 CallGenerator* cg = NULL; 759 if (is_osr_compilation()) { 760 const TypeTuple *domain = StartOSRNode::osr_domain(); 761 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 762 init_tf(TypeFunc::make(domain, range)); 763 StartNode* s = new (this) StartOSRNode(root(), domain); 764 initial_gvn()->set_type_bottom(s); 765 init_start(s); 766 cg = CallGenerator::for_osr(method(), entry_bci()); 767 } else { 768 // Normal case. 769 init_tf(TypeFunc::make(method())); 770 StartNode* s = new (this) StartNode(root(), tf()->domain()); 771 initial_gvn()->set_type_bottom(s); 772 init_start(s); 773 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 774 // With java.lang.ref.reference.get() we must go through the 775 // intrinsic when G1 is enabled - even when get() is the root 776 // method of the compile - so that, if necessary, the value in 777 // the referent field of the reference object gets recorded by 778 // the pre-barrier code. 779 // Specifically, if G1 is enabled, the value in the referent 780 // field is recorded by the G1 SATB pre barrier. This will 781 // result in the referent being marked live and the reference 782 // object removed from the list of discovered references during 783 // reference processing. 784 cg = find_intrinsic(method(), false); 785 } 786 if (cg == NULL) { 787 float past_uses = method()->interpreter_invocation_count(); 788 float expected_uses = past_uses; 789 cg = CallGenerator::for_inline(method(), expected_uses); 790 } 791 } 792 if (failing()) return; 793 if (cg == NULL) { 794 record_method_not_compilable_all_tiers("cannot parse method"); 795 return; 796 } 797 JVMState* jvms = build_start_state(start(), tf()); 798 if ((jvms = cg->generate(jvms)) == NULL) { 799 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 800 record_method_not_compilable("method parse failed"); 801 } 802 return; 803 } 804 GraphKit kit(jvms); 805 806 if (!kit.stopped()) { 807 // Accept return values, and transfer control we know not where. 808 // This is done by a special, unique ReturnNode bound to root. 809 return_values(kit.jvms()); 810 } 811 812 if (kit.has_exceptions()) { 813 // Any exceptions that escape from this call must be rethrown 814 // to whatever caller is dynamically above us on the stack. 815 // This is done by a special, unique RethrowNode bound to root. 816 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 817 } 818 819 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 820 821 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 822 inline_string_calls(true); 823 } 824 825 if (failing()) return; 826 827 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 828 829 // Remove clutter produced by parsing. 830 if (!failing()) { 831 ResourceMark rm; 832 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 833 } 834 } 835 836 // Note: Large methods are capped off in do_one_bytecode(). 837 if (failing()) return; 838 839 // After parsing, node notes are no longer automagic. 840 // They must be propagated by register_new_node_with_optimizer(), 841 // clone(), or the like. 842 set_default_node_notes(NULL); 843 844 for (;;) { 845 int successes = Inline_Warm(); 846 if (failing()) return; 847 if (successes == 0) break; 848 } 849 850 // Drain the list. 851 Finish_Warm(); 852 #ifndef PRODUCT 853 if (_printer) { 854 _printer->print_inlining(this); 855 } 856 #endif 857 858 if (failing()) return; 859 NOT_PRODUCT( verify_graph_edges(); ) 860 861 // Now optimize 862 Optimize(); 863 if (failing()) return; 864 NOT_PRODUCT( verify_graph_edges(); ) 865 866 #ifndef PRODUCT 867 if (PrintIdeal) { 868 ttyLocker ttyl; // keep the following output all in one block 869 // This output goes directly to the tty, not the compiler log. 870 // To enable tools to match it up with the compilation activity, 871 // be sure to tag this tty output with the compile ID. 872 if (xtty != NULL) { 873 xtty->head("ideal compile_id='%d'%s", compile_id(), 874 is_osr_compilation() ? " compile_kind='osr'" : 875 ""); 876 } 877 root()->dump(9999); 878 if (xtty != NULL) { 879 xtty->tail("ideal"); 880 } 881 } 882 #endif 883 884 NOT_PRODUCT( verify_barriers(); ) 885 886 // Dump compilation data to replay it. 887 if (method()->has_option("DumpReplay")) { 888 env()->dump_replay_data(_compile_id); 889 } 890 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 891 env()->dump_inline_data(_compile_id); 892 } 893 894 // Now that we know the size of all the monitors we can add a fixed slot 895 // for the original deopt pc. 896 897 _orig_pc_slot = fixed_slots(); 898 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 899 set_fixed_slots(next_slot); 900 901 // Compute when to use implicit null checks. Used by matching trap based 902 // nodes and NullCheck optimization. 903 set_allowed_deopt_reasons(); 904 905 // Now generate code 906 Code_Gen(); 907 if (failing()) return; 908 909 // Check if we want to skip execution of all compiled code. 910 { 911 #ifndef PRODUCT 912 if (OptoNoExecute) { 913 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 914 return; 915 } 916 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 917 #endif 918 919 if (is_osr_compilation()) { 920 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 921 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 922 } else { 923 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 924 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 925 } 926 927 env()->register_method(_method, _entry_bci, 928 &_code_offsets, 929 _orig_pc_slot_offset_in_bytes, 930 code_buffer(), 931 frame_size_in_words(), _oop_map_set, 932 &_handler_table, &_inc_table, 933 compiler, 934 env()->comp_level(), 935 has_unsafe_access(), 936 SharedRuntime::is_wide_vector(max_vector_size()), 937 rtm_state() 938 ); 939 940 if (log() != NULL) // Print code cache state into compiler log 941 log()->code_cache_state(); 942 } 943 } 944 945 //------------------------------Compile---------------------------------------- 946 // Compile a runtime stub 947 Compile::Compile( ciEnv* ci_env, 948 TypeFunc_generator generator, 949 address stub_function, 950 const char *stub_name, 951 int is_fancy_jump, 952 bool pass_tls, 953 bool save_arg_registers, 954 bool return_pc ) 955 : Phase(Compiler), 956 _env(ci_env), 957 _log(ci_env->log()), 958 _compile_id(0), 959 _save_argument_registers(save_arg_registers), 960 _method(NULL), 961 _stub_name(stub_name), 962 _stub_function(stub_function), 963 _stub_entry_point(NULL), 964 _entry_bci(InvocationEntryBci), 965 _initial_gvn(NULL), 966 _for_igvn(NULL), 967 _warm_calls(NULL), 968 _orig_pc_slot(0), 969 _orig_pc_slot_offset_in_bytes(0), 970 _subsume_loads(true), 971 _do_escape_analysis(false), 972 _eliminate_boxing(false), 973 _failure_reason(NULL), 974 _code_buffer("Compile::Fill_buffer"), 975 _has_method_handle_invokes(false), 976 _mach_constant_base_node(NULL), 977 _node_bundling_limit(0), 978 _node_bundling_base(NULL), 979 _java_calls(0), 980 _inner_loops(0), 981 #ifndef PRODUCT 982 _trace_opto_output(TraceOptoOutput), 983 _in_dump_cnt(0), 984 _printer(NULL), 985 #endif 986 _comp_arena(mtCompiler), 987 _node_arena(mtCompiler), 988 _old_arena(mtCompiler), 989 _Compile_types(mtCompiler), 990 _dead_node_list(comp_arena()), 991 _dead_node_count(0), 992 _congraph(NULL), 993 _replay_inline_data(NULL), 994 _number_of_mh_late_inlines(0), 995 _inlining_progress(false), 996 _inlining_incrementally(false), 997 _print_inlining_list(NULL), 998 _print_inlining_idx(0), 999 _allowed_reasons(0), 1000 _interpreter_frame_size(0), 1001 _max_node_limit(MaxNodeLimit) { 1002 C = this; 1003 1004 #ifndef PRODUCT 1005 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 1006 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 1007 set_print_assembly(PrintFrameConverterAssembly); 1008 set_parsed_irreducible_loop(false); 1009 #endif 1010 set_has_irreducible_loop(false); // no loops 1011 1012 CompileWrapper cw(this); 1013 Init(/*AliasLevel=*/ 0); 1014 init_tf((*generator)()); 1015 1016 { 1017 // The following is a dummy for the sake of GraphKit::gen_stub 1018 Unique_Node_List for_igvn(comp_arena()); 1019 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1020 PhaseGVN gvn(Thread::current()->resource_area(),255); 1021 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1022 gvn.transform_no_reclaim(top()); 1023 1024 GraphKit kit; 1025 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1026 } 1027 1028 NOT_PRODUCT( verify_graph_edges(); ) 1029 Code_Gen(); 1030 if (failing()) return; 1031 1032 1033 // Entry point will be accessed using compile->stub_entry_point(); 1034 if (code_buffer() == NULL) { 1035 Matcher::soft_match_failure(); 1036 } else { 1037 if (PrintAssembly && (WizardMode || Verbose)) 1038 tty->print_cr("### Stub::%s", stub_name); 1039 1040 if (!failing()) { 1041 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1042 1043 // Make the NMethod 1044 // For now we mark the frame as never safe for profile stackwalking 1045 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1046 code_buffer(), 1047 CodeOffsets::frame_never_safe, 1048 // _code_offsets.value(CodeOffsets::Frame_Complete), 1049 frame_size_in_words(), 1050 _oop_map_set, 1051 save_arg_registers); 1052 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1053 1054 _stub_entry_point = rs->entry_point(); 1055 } 1056 } 1057 } 1058 1059 //------------------------------Init------------------------------------------- 1060 // Prepare for a single compilation 1061 void Compile::Init(int aliaslevel) { 1062 _unique = 0; 1063 _regalloc = NULL; 1064 1065 _tf = NULL; // filled in later 1066 _top = NULL; // cached later 1067 _matcher = NULL; // filled in later 1068 _cfg = NULL; // filled in later 1069 1070 set_24_bit_selection_and_mode(Use24BitFP, false); 1071 1072 _node_note_array = NULL; 1073 _default_node_notes = NULL; 1074 1075 _immutable_memory = NULL; // filled in at first inquiry 1076 1077 // Globally visible Nodes 1078 // First set TOP to NULL to give safe behavior during creation of RootNode 1079 set_cached_top_node(NULL); 1080 set_root(new (this) RootNode()); 1081 // Now that you have a Root to point to, create the real TOP 1082 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1083 set_recent_alloc(NULL, NULL); 1084 1085 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1086 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1087 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1088 env()->set_dependencies(new Dependencies(env())); 1089 1090 _fixed_slots = 0; 1091 set_has_split_ifs(false); 1092 set_has_loops(has_method() && method()->has_loops()); // first approximation 1093 set_has_stringbuilder(false); 1094 set_has_boxed_value(false); 1095 _trap_can_recompile = false; // no traps emitted yet 1096 _major_progress = true; // start out assuming good things will happen 1097 set_has_unsafe_access(false); 1098 set_max_vector_size(0); 1099 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1100 set_decompile_count(0); 1101 1102 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1103 set_num_loop_opts(LoopOptsCount); 1104 set_do_inlining(Inline); 1105 set_max_inline_size(MaxInlineSize); 1106 set_freq_inline_size(FreqInlineSize); 1107 set_do_scheduling(OptoScheduling); 1108 set_do_count_invocations(false); 1109 set_do_method_data_update(false); 1110 set_rtm_state(NoRTM); // No RTM lock eliding by default 1111 method_has_option_value("MaxNodeLimit", _max_node_limit); 1112 #if INCLUDE_RTM_OPT 1113 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1114 int rtm_state = method()->method_data()->rtm_state(); 1115 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1116 // Don't generate RTM lock eliding code. 1117 set_rtm_state(NoRTM); 1118 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1119 // Generate RTM lock eliding code without abort ratio calculation code. 1120 set_rtm_state(UseRTM); 1121 } else if (UseRTMDeopt) { 1122 // Generate RTM lock eliding code and include abort ratio calculation 1123 // code if UseRTMDeopt is on. 1124 set_rtm_state(ProfileRTM); 1125 } 1126 } 1127 #endif 1128 if (debug_info()->recording_non_safepoints()) { 1129 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1130 (comp_arena(), 8, 0, NULL)); 1131 set_default_node_notes(Node_Notes::make(this)); 1132 } 1133 1134 // // -- Initialize types before each compile -- 1135 // // Update cached type information 1136 // if( _method && _method->constants() ) 1137 // Type::update_loaded_types(_method, _method->constants()); 1138 1139 // Init alias_type map. 1140 if (!_do_escape_analysis && aliaslevel == 3) 1141 aliaslevel = 2; // No unique types without escape analysis 1142 _AliasLevel = aliaslevel; 1143 const int grow_ats = 16; 1144 _max_alias_types = grow_ats; 1145 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1146 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1147 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1148 { 1149 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1150 } 1151 // Initialize the first few types. 1152 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1153 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1154 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1155 _num_alias_types = AliasIdxRaw+1; 1156 // Zero out the alias type cache. 1157 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1158 // A NULL adr_type hits in the cache right away. Preload the right answer. 1159 probe_alias_cache(NULL)->_index = AliasIdxTop; 1160 1161 _intrinsics = NULL; 1162 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1163 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1164 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1165 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1166 register_library_intrinsics(); 1167 } 1168 1169 //---------------------------init_start---------------------------------------- 1170 // Install the StartNode on this compile object. 1171 void Compile::init_start(StartNode* s) { 1172 if (failing()) 1173 return; // already failing 1174 assert(s == start(), ""); 1175 } 1176 1177 StartNode* Compile::start() const { 1178 assert(!failing(), ""); 1179 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1180 Node* start = root()->fast_out(i); 1181 if( start->is_Start() ) 1182 return start->as_Start(); 1183 } 1184 fatal("Did not find Start node!"); 1185 return NULL; 1186 } 1187 1188 //-------------------------------immutable_memory------------------------------------- 1189 // Access immutable memory 1190 Node* Compile::immutable_memory() { 1191 if (_immutable_memory != NULL) { 1192 return _immutable_memory; 1193 } 1194 StartNode* s = start(); 1195 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1196 Node *p = s->fast_out(i); 1197 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1198 _immutable_memory = p; 1199 return _immutable_memory; 1200 } 1201 } 1202 ShouldNotReachHere(); 1203 return NULL; 1204 } 1205 1206 //----------------------set_cached_top_node------------------------------------ 1207 // Install the cached top node, and make sure Node::is_top works correctly. 1208 void Compile::set_cached_top_node(Node* tn) { 1209 if (tn != NULL) verify_top(tn); 1210 Node* old_top = _top; 1211 _top = tn; 1212 // Calling Node::setup_is_top allows the nodes the chance to adjust 1213 // their _out arrays. 1214 if (_top != NULL) _top->setup_is_top(); 1215 if (old_top != NULL) old_top->setup_is_top(); 1216 assert(_top == NULL || top()->is_top(), ""); 1217 } 1218 1219 #ifdef ASSERT 1220 uint Compile::count_live_nodes_by_graph_walk() { 1221 Unique_Node_List useful(comp_arena()); 1222 // Get useful node list by walking the graph. 1223 identify_useful_nodes(useful); 1224 return useful.size(); 1225 } 1226 1227 void Compile::print_missing_nodes() { 1228 1229 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1230 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1231 return; 1232 } 1233 1234 // This is an expensive function. It is executed only when the user 1235 // specifies VerifyIdealNodeCount option or otherwise knows the 1236 // additional work that needs to be done to identify reachable nodes 1237 // by walking the flow graph and find the missing ones using 1238 // _dead_node_list. 1239 1240 Unique_Node_List useful(comp_arena()); 1241 // Get useful node list by walking the graph. 1242 identify_useful_nodes(useful); 1243 1244 uint l_nodes = C->live_nodes(); 1245 uint l_nodes_by_walk = useful.size(); 1246 1247 if (l_nodes != l_nodes_by_walk) { 1248 if (_log != NULL) { 1249 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1250 _log->stamp(); 1251 _log->end_head(); 1252 } 1253 VectorSet& useful_member_set = useful.member_set(); 1254 int last_idx = l_nodes_by_walk; 1255 for (int i = 0; i < last_idx; i++) { 1256 if (useful_member_set.test(i)) { 1257 if (_dead_node_list.test(i)) { 1258 if (_log != NULL) { 1259 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1260 } 1261 if (PrintIdealNodeCount) { 1262 // Print the log message to tty 1263 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1264 useful.at(i)->dump(); 1265 } 1266 } 1267 } 1268 else if (! _dead_node_list.test(i)) { 1269 if (_log != NULL) { 1270 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1271 } 1272 if (PrintIdealNodeCount) { 1273 // Print the log message to tty 1274 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1275 } 1276 } 1277 } 1278 if (_log != NULL) { 1279 _log->tail("mismatched_nodes"); 1280 } 1281 } 1282 } 1283 #endif 1284 1285 #ifndef PRODUCT 1286 void Compile::verify_top(Node* tn) const { 1287 if (tn != NULL) { 1288 assert(tn->is_Con(), "top node must be a constant"); 1289 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1290 assert(tn->in(0) != NULL, "must have live top node"); 1291 } 1292 } 1293 #endif 1294 1295 1296 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1297 1298 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1299 guarantee(arr != NULL, ""); 1300 int num_blocks = arr->length(); 1301 if (grow_by < num_blocks) grow_by = num_blocks; 1302 int num_notes = grow_by * _node_notes_block_size; 1303 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1304 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1305 while (num_notes > 0) { 1306 arr->append(notes); 1307 notes += _node_notes_block_size; 1308 num_notes -= _node_notes_block_size; 1309 } 1310 assert(num_notes == 0, "exact multiple, please"); 1311 } 1312 1313 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1314 if (source == NULL || dest == NULL) return false; 1315 1316 if (dest->is_Con()) 1317 return false; // Do not push debug info onto constants. 1318 1319 #ifdef ASSERT 1320 // Leave a bread crumb trail pointing to the original node: 1321 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1322 dest->set_debug_orig(source); 1323 } 1324 #endif 1325 1326 if (node_note_array() == NULL) 1327 return false; // Not collecting any notes now. 1328 1329 // This is a copy onto a pre-existing node, which may already have notes. 1330 // If both nodes have notes, do not overwrite any pre-existing notes. 1331 Node_Notes* source_notes = node_notes_at(source->_idx); 1332 if (source_notes == NULL || source_notes->is_clear()) return false; 1333 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1334 if (dest_notes == NULL || dest_notes->is_clear()) { 1335 return set_node_notes_at(dest->_idx, source_notes); 1336 } 1337 1338 Node_Notes merged_notes = (*source_notes); 1339 // The order of operations here ensures that dest notes will win... 1340 merged_notes.update_from(dest_notes); 1341 return set_node_notes_at(dest->_idx, &merged_notes); 1342 } 1343 1344 1345 //--------------------------allow_range_check_smearing------------------------- 1346 // Gating condition for coalescing similar range checks. 1347 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1348 // single covering check that is at least as strong as any of them. 1349 // If the optimization succeeds, the simplified (strengthened) range check 1350 // will always succeed. If it fails, we will deopt, and then give up 1351 // on the optimization. 1352 bool Compile::allow_range_check_smearing() const { 1353 // If this method has already thrown a range-check, 1354 // assume it was because we already tried range smearing 1355 // and it failed. 1356 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1357 return !already_trapped; 1358 } 1359 1360 1361 //------------------------------flatten_alias_type----------------------------- 1362 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1363 int offset = tj->offset(); 1364 TypePtr::PTR ptr = tj->ptr(); 1365 1366 // Known instance (scalarizable allocation) alias only with itself. 1367 bool is_known_inst = tj->isa_oopptr() != NULL && 1368 tj->is_oopptr()->is_known_instance(); 1369 1370 // Process weird unsafe references. 1371 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1372 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1373 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1374 tj = TypeOopPtr::BOTTOM; 1375 ptr = tj->ptr(); 1376 offset = tj->offset(); 1377 } 1378 1379 // Array pointers need some flattening 1380 const TypeAryPtr *ta = tj->isa_aryptr(); 1381 if (ta && ta->is_stable()) { 1382 // Erase stability property for alias analysis. 1383 tj = ta = ta->cast_to_stable(false); 1384 } 1385 if( ta && is_known_inst ) { 1386 if ( offset != Type::OffsetBot && 1387 offset > arrayOopDesc::length_offset_in_bytes() ) { 1388 offset = Type::OffsetBot; // Flatten constant access into array body only 1389 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1390 } 1391 } else if( ta && _AliasLevel >= 2 ) { 1392 // For arrays indexed by constant indices, we flatten the alias 1393 // space to include all of the array body. Only the header, klass 1394 // and array length can be accessed un-aliased. 1395 if( offset != Type::OffsetBot ) { 1396 if( ta->const_oop() ) { // MethodData* or Method* 1397 offset = Type::OffsetBot; // Flatten constant access into array body 1398 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1399 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1400 // range is OK as-is. 1401 tj = ta = TypeAryPtr::RANGE; 1402 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1403 tj = TypeInstPtr::KLASS; // all klass loads look alike 1404 ta = TypeAryPtr::RANGE; // generic ignored junk 1405 ptr = TypePtr::BotPTR; 1406 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1407 tj = TypeInstPtr::MARK; 1408 ta = TypeAryPtr::RANGE; // generic ignored junk 1409 ptr = TypePtr::BotPTR; 1410 } else { // Random constant offset into array body 1411 offset = Type::OffsetBot; // Flatten constant access into array body 1412 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1413 } 1414 } 1415 // Arrays of fixed size alias with arrays of unknown size. 1416 if (ta->size() != TypeInt::POS) { 1417 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1418 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1419 } 1420 // Arrays of known objects become arrays of unknown objects. 1421 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1422 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1423 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1424 } 1425 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1426 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1427 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1428 } 1429 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1430 // cannot be distinguished by bytecode alone. 1431 if (ta->elem() == TypeInt::BOOL) { 1432 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1433 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1434 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1435 } 1436 // During the 2nd round of IterGVN, NotNull castings are removed. 1437 // Make sure the Bottom and NotNull variants alias the same. 1438 // Also, make sure exact and non-exact variants alias the same. 1439 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1440 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1441 } 1442 } 1443 1444 // Oop pointers need some flattening 1445 const TypeInstPtr *to = tj->isa_instptr(); 1446 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1447 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1448 if( ptr == TypePtr::Constant ) { 1449 if (to->klass() != ciEnv::current()->Class_klass() || 1450 offset < k->size_helper() * wordSize) { 1451 // No constant oop pointers (such as Strings); they alias with 1452 // unknown strings. 1453 assert(!is_known_inst, "not scalarizable allocation"); 1454 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1455 } 1456 } else if( is_known_inst ) { 1457 tj = to; // Keep NotNull and klass_is_exact for instance type 1458 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1459 // During the 2nd round of IterGVN, NotNull castings are removed. 1460 // Make sure the Bottom and NotNull variants alias the same. 1461 // Also, make sure exact and non-exact variants alias the same. 1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1463 } 1464 if (to->speculative() != NULL) { 1465 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1466 } 1467 // Canonicalize the holder of this field 1468 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1469 // First handle header references such as a LoadKlassNode, even if the 1470 // object's klass is unloaded at compile time (4965979). 1471 if (!is_known_inst) { // Do it only for non-instance types 1472 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1473 } 1474 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1475 // Static fields are in the space above the normal instance 1476 // fields in the java.lang.Class instance. 1477 if (to->klass() != ciEnv::current()->Class_klass()) { 1478 to = NULL; 1479 tj = TypeOopPtr::BOTTOM; 1480 offset = tj->offset(); 1481 } 1482 } else { 1483 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1484 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1485 if( is_known_inst ) { 1486 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1487 } else { 1488 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1489 } 1490 } 1491 } 1492 } 1493 1494 // Klass pointers to object array klasses need some flattening 1495 const TypeKlassPtr *tk = tj->isa_klassptr(); 1496 if( tk ) { 1497 // If we are referencing a field within a Klass, we need 1498 // to assume the worst case of an Object. Both exact and 1499 // inexact types must flatten to the same alias class so 1500 // use NotNull as the PTR. 1501 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1502 1503 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1504 TypeKlassPtr::OBJECT->klass(), 1505 offset); 1506 } 1507 1508 ciKlass* klass = tk->klass(); 1509 if( klass->is_obj_array_klass() ) { 1510 ciKlass* k = TypeAryPtr::OOPS->klass(); 1511 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1512 k = TypeInstPtr::BOTTOM->klass(); 1513 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1514 } 1515 1516 // Check for precise loads from the primary supertype array and force them 1517 // to the supertype cache alias index. Check for generic array loads from 1518 // the primary supertype array and also force them to the supertype cache 1519 // alias index. Since the same load can reach both, we need to merge 1520 // these 2 disparate memories into the same alias class. Since the 1521 // primary supertype array is read-only, there's no chance of confusion 1522 // where we bypass an array load and an array store. 1523 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1524 if (offset == Type::OffsetBot || 1525 (offset >= primary_supers_offset && 1526 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1527 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1528 offset = in_bytes(Klass::secondary_super_cache_offset()); 1529 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1530 } 1531 } 1532 1533 // Flatten all Raw pointers together. 1534 if (tj->base() == Type::RawPtr) 1535 tj = TypeRawPtr::BOTTOM; 1536 1537 if (tj->base() == Type::AnyPtr) 1538 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1539 1540 // Flatten all to bottom for now 1541 switch( _AliasLevel ) { 1542 case 0: 1543 tj = TypePtr::BOTTOM; 1544 break; 1545 case 1: // Flatten to: oop, static, field or array 1546 switch (tj->base()) { 1547 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1548 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1549 case Type::AryPtr: // do not distinguish arrays at all 1550 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1551 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1552 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1553 default: ShouldNotReachHere(); 1554 } 1555 break; 1556 case 2: // No collapsing at level 2; keep all splits 1557 case 3: // No collapsing at level 3; keep all splits 1558 break; 1559 default: 1560 Unimplemented(); 1561 } 1562 1563 offset = tj->offset(); 1564 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1565 1566 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1567 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1568 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1569 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1570 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1571 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1572 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1573 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1574 assert( tj->ptr() != TypePtr::TopPTR && 1575 tj->ptr() != TypePtr::AnyNull && 1576 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1577 // assert( tj->ptr() != TypePtr::Constant || 1578 // tj->base() == Type::RawPtr || 1579 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1580 1581 return tj; 1582 } 1583 1584 void Compile::AliasType::Init(int i, const TypePtr* at) { 1585 _index = i; 1586 _adr_type = at; 1587 _field = NULL; 1588 _element = NULL; 1589 _is_rewritable = true; // default 1590 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1591 if (atoop != NULL && atoop->is_known_instance()) { 1592 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1593 _general_index = Compile::current()->get_alias_index(gt); 1594 } else { 1595 _general_index = 0; 1596 } 1597 } 1598 1599 BasicType Compile::AliasType::basic_type() const { 1600 if (element() != NULL) { 1601 const Type* element = adr_type()->is_aryptr()->elem(); 1602 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1603 } if (field() != NULL) { 1604 return field()->layout_type(); 1605 } else { 1606 return T_ILLEGAL; // unknown 1607 } 1608 } 1609 1610 //---------------------------------print_on------------------------------------ 1611 #ifndef PRODUCT 1612 void Compile::AliasType::print_on(outputStream* st) { 1613 if (index() < 10) 1614 st->print("@ <%d> ", index()); 1615 else st->print("@ <%d>", index()); 1616 st->print(is_rewritable() ? " " : " RO"); 1617 int offset = adr_type()->offset(); 1618 if (offset == Type::OffsetBot) 1619 st->print(" +any"); 1620 else st->print(" +%-3d", offset); 1621 st->print(" in "); 1622 adr_type()->dump_on(st); 1623 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1624 if (field() != NULL && tjp) { 1625 if (tjp->klass() != field()->holder() || 1626 tjp->offset() != field()->offset_in_bytes()) { 1627 st->print(" != "); 1628 field()->print(); 1629 st->print(" ***"); 1630 } 1631 } 1632 } 1633 1634 void print_alias_types() { 1635 Compile* C = Compile::current(); 1636 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1637 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1638 C->alias_type(idx)->print_on(tty); 1639 tty->cr(); 1640 } 1641 } 1642 #endif 1643 1644 1645 //----------------------------probe_alias_cache-------------------------------- 1646 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1647 intptr_t key = (intptr_t) adr_type; 1648 key ^= key >> logAliasCacheSize; 1649 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1650 } 1651 1652 1653 //-----------------------------grow_alias_types-------------------------------- 1654 void Compile::grow_alias_types() { 1655 const int old_ats = _max_alias_types; // how many before? 1656 const int new_ats = old_ats; // how many more? 1657 const int grow_ats = old_ats+new_ats; // how many now? 1658 _max_alias_types = grow_ats; 1659 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1660 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1661 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1662 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1663 } 1664 1665 1666 //--------------------------------find_alias_type------------------------------ 1667 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1668 if (_AliasLevel == 0) 1669 return alias_type(AliasIdxBot); 1670 1671 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1672 if (ace->_adr_type == adr_type) { 1673 return alias_type(ace->_index); 1674 } 1675 1676 // Handle special cases. 1677 if (adr_type == NULL) return alias_type(AliasIdxTop); 1678 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1679 1680 // Do it the slow way. 1681 const TypePtr* flat = flatten_alias_type(adr_type); 1682 1683 #ifdef ASSERT 1684 { 1685 ResourceMark rm; 1686 assert(flat == flatten_alias_type(flat), 1687 err_msg("not idempotent: adr_type = %s; flat = %s => %s", Type::str(adr_type), 1688 Type::str(flat), Type::str(flatten_alias_type(flat)))); 1689 assert(flat != TypePtr::BOTTOM, 1690 err_msg("cannot alias-analyze an untyped ptr: adr_type = %s", Type::str(adr_type))); 1691 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1692 const TypeOopPtr* foop = flat->is_oopptr(); 1693 // Scalarizable allocations have exact klass always. 1694 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1695 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1696 assert(foop == flatten_alias_type(xoop), 1697 err_msg("exactness must not affect alias type: foop = %s; xoop = %s", 1698 Type::str(foop), Type::str(xoop))); 1699 } 1700 } 1701 #endif 1702 1703 int idx = AliasIdxTop; 1704 for (int i = 0; i < num_alias_types(); i++) { 1705 if (alias_type(i)->adr_type() == flat) { 1706 idx = i; 1707 break; 1708 } 1709 } 1710 1711 if (idx == AliasIdxTop) { 1712 if (no_create) return NULL; 1713 // Grow the array if necessary. 1714 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1715 // Add a new alias type. 1716 idx = _num_alias_types++; 1717 _alias_types[idx]->Init(idx, flat); 1718 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1719 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1720 if (flat->isa_instptr()) { 1721 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1722 && flat->is_instptr()->klass() == env()->Class_klass()) 1723 alias_type(idx)->set_rewritable(false); 1724 } 1725 if (flat->isa_aryptr()) { 1726 #ifdef ASSERT 1727 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1728 // (T_BYTE has the weakest alignment and size restrictions...) 1729 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1730 #endif 1731 if (flat->offset() == TypePtr::OffsetBot) { 1732 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1733 } 1734 } 1735 if (flat->isa_klassptr()) { 1736 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1737 alias_type(idx)->set_rewritable(false); 1738 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1739 alias_type(idx)->set_rewritable(false); 1740 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1741 alias_type(idx)->set_rewritable(false); 1742 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1743 alias_type(idx)->set_rewritable(false); 1744 } 1745 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1746 // but the base pointer type is not distinctive enough to identify 1747 // references into JavaThread.) 1748 1749 // Check for final fields. 1750 const TypeInstPtr* tinst = flat->isa_instptr(); 1751 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1752 ciField* field; 1753 if (tinst->const_oop() != NULL && 1754 tinst->klass() == ciEnv::current()->Class_klass() && 1755 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1756 // static field 1757 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1758 field = k->get_field_by_offset(tinst->offset(), true); 1759 } else { 1760 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1761 field = k->get_field_by_offset(tinst->offset(), false); 1762 } 1763 assert(field == NULL || 1764 original_field == NULL || 1765 (field->holder() == original_field->holder() && 1766 field->offset() == original_field->offset() && 1767 field->is_static() == original_field->is_static()), "wrong field?"); 1768 // Set field() and is_rewritable() attributes. 1769 if (field != NULL) alias_type(idx)->set_field(field); 1770 } 1771 } 1772 1773 // Fill the cache for next time. 1774 ace->_adr_type = adr_type; 1775 ace->_index = idx; 1776 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1777 1778 // Might as well try to fill the cache for the flattened version, too. 1779 AliasCacheEntry* face = probe_alias_cache(flat); 1780 if (face->_adr_type == NULL) { 1781 face->_adr_type = flat; 1782 face->_index = idx; 1783 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1784 } 1785 1786 return alias_type(idx); 1787 } 1788 1789 1790 Compile::AliasType* Compile::alias_type(ciField* field) { 1791 const TypeOopPtr* t; 1792 if (field->is_static()) 1793 t = TypeInstPtr::make(field->holder()->java_mirror()); 1794 else 1795 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1796 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1797 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1798 return atp; 1799 } 1800 1801 1802 //------------------------------have_alias_type-------------------------------- 1803 bool Compile::have_alias_type(const TypePtr* adr_type) { 1804 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1805 if (ace->_adr_type == adr_type) { 1806 return true; 1807 } 1808 1809 // Handle special cases. 1810 if (adr_type == NULL) return true; 1811 if (adr_type == TypePtr::BOTTOM) return true; 1812 1813 return find_alias_type(adr_type, true, NULL) != NULL; 1814 } 1815 1816 //-----------------------------must_alias-------------------------------------- 1817 // True if all values of the given address type are in the given alias category. 1818 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1819 if (alias_idx == AliasIdxBot) return true; // the universal category 1820 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1821 if (alias_idx == AliasIdxTop) return false; // the empty category 1822 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1823 1824 // the only remaining possible overlap is identity 1825 int adr_idx = get_alias_index(adr_type); 1826 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1827 assert(adr_idx == alias_idx || 1828 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1829 && adr_type != TypeOopPtr::BOTTOM), 1830 "should not be testing for overlap with an unsafe pointer"); 1831 return adr_idx == alias_idx; 1832 } 1833 1834 //------------------------------can_alias-------------------------------------- 1835 // True if any values of the given address type are in the given alias category. 1836 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1837 if (alias_idx == AliasIdxTop) return false; // the empty category 1838 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1839 if (alias_idx == AliasIdxBot) return true; // the universal category 1840 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1841 1842 // the only remaining possible overlap is identity 1843 int adr_idx = get_alias_index(adr_type); 1844 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1845 return adr_idx == alias_idx; 1846 } 1847 1848 1849 1850 //---------------------------pop_warm_call------------------------------------- 1851 WarmCallInfo* Compile::pop_warm_call() { 1852 WarmCallInfo* wci = _warm_calls; 1853 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1854 return wci; 1855 } 1856 1857 //----------------------------Inline_Warm-------------------------------------- 1858 int Compile::Inline_Warm() { 1859 // If there is room, try to inline some more warm call sites. 1860 // %%% Do a graph index compaction pass when we think we're out of space? 1861 if (!InlineWarmCalls) return 0; 1862 1863 int calls_made_hot = 0; 1864 int room_to_grow = NodeCountInliningCutoff - unique(); 1865 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1866 int amount_grown = 0; 1867 WarmCallInfo* call; 1868 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1869 int est_size = (int)call->size(); 1870 if (est_size > (room_to_grow - amount_grown)) { 1871 // This one won't fit anyway. Get rid of it. 1872 call->make_cold(); 1873 continue; 1874 } 1875 call->make_hot(); 1876 calls_made_hot++; 1877 amount_grown += est_size; 1878 amount_to_grow -= est_size; 1879 } 1880 1881 if (calls_made_hot > 0) set_major_progress(); 1882 return calls_made_hot; 1883 } 1884 1885 1886 //----------------------------Finish_Warm-------------------------------------- 1887 void Compile::Finish_Warm() { 1888 if (!InlineWarmCalls) return; 1889 if (failing()) return; 1890 if (warm_calls() == NULL) return; 1891 1892 // Clean up loose ends, if we are out of space for inlining. 1893 WarmCallInfo* call; 1894 while ((call = pop_warm_call()) != NULL) { 1895 call->make_cold(); 1896 } 1897 } 1898 1899 //---------------------cleanup_loop_predicates----------------------- 1900 // Remove the opaque nodes that protect the predicates so that all unused 1901 // checks and uncommon_traps will be eliminated from the ideal graph 1902 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1903 if (predicate_count()==0) return; 1904 for (int i = predicate_count(); i > 0; i--) { 1905 Node * n = predicate_opaque1_node(i-1); 1906 assert(n->Opcode() == Op_Opaque1, "must be"); 1907 igvn.replace_node(n, n->in(1)); 1908 } 1909 assert(predicate_count()==0, "should be clean!"); 1910 } 1911 1912 void Compile::add_range_check_cast(Node* n) { 1913 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1914 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1915 _range_check_casts->append(n); 1916 } 1917 1918 // Remove all range check dependent CastIINodes. 1919 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1920 for (int i = range_check_cast_count(); i > 0; i--) { 1921 Node* cast = range_check_cast_node(i-1); 1922 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1923 igvn.replace_node(cast, cast->in(1)); 1924 } 1925 assert(range_check_cast_count() == 0, "should be empty"); 1926 } 1927 1928 // StringOpts and late inlining of string methods 1929 void Compile::inline_string_calls(bool parse_time) { 1930 { 1931 // remove useless nodes to make the usage analysis simpler 1932 ResourceMark rm; 1933 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1934 } 1935 1936 { 1937 ResourceMark rm; 1938 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1939 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1940 print_method(PHASE_AFTER_STRINGOPTS, 3); 1941 } 1942 1943 // now inline anything that we skipped the first time around 1944 if (!parse_time) { 1945 _late_inlines_pos = _late_inlines.length(); 1946 } 1947 1948 while (_string_late_inlines.length() > 0) { 1949 CallGenerator* cg = _string_late_inlines.pop(); 1950 cg->do_late_inline(); 1951 if (failing()) return; 1952 } 1953 _string_late_inlines.trunc_to(0); 1954 } 1955 1956 // Late inlining of boxing methods 1957 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1958 if (_boxing_late_inlines.length() > 0) { 1959 assert(has_boxed_value(), "inconsistent"); 1960 1961 PhaseGVN* gvn = initial_gvn(); 1962 set_inlining_incrementally(true); 1963 1964 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1965 for_igvn()->clear(); 1966 gvn->replace_with(&igvn); 1967 1968 _late_inlines_pos = _late_inlines.length(); 1969 1970 while (_boxing_late_inlines.length() > 0) { 1971 CallGenerator* cg = _boxing_late_inlines.pop(); 1972 cg->do_late_inline(); 1973 if (failing()) return; 1974 } 1975 _boxing_late_inlines.trunc_to(0); 1976 1977 { 1978 ResourceMark rm; 1979 PhaseRemoveUseless pru(gvn, for_igvn()); 1980 } 1981 1982 igvn = PhaseIterGVN(gvn); 1983 igvn.optimize(); 1984 1985 set_inlining_progress(false); 1986 set_inlining_incrementally(false); 1987 } 1988 } 1989 1990 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1991 assert(IncrementalInline, "incremental inlining should be on"); 1992 PhaseGVN* gvn = initial_gvn(); 1993 1994 set_inlining_progress(false); 1995 for_igvn()->clear(); 1996 gvn->replace_with(&igvn); 1997 1998 int i = 0; 1999 2000 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2001 CallGenerator* cg = _late_inlines.at(i); 2002 _late_inlines_pos = i+1; 2003 cg->do_late_inline(); 2004 if (failing()) return; 2005 } 2006 int j = 0; 2007 for (; i < _late_inlines.length(); i++, j++) { 2008 _late_inlines.at_put(j, _late_inlines.at(i)); 2009 } 2010 _late_inlines.trunc_to(j); 2011 2012 { 2013 ResourceMark rm; 2014 PhaseRemoveUseless pru(gvn, for_igvn()); 2015 } 2016 2017 igvn = PhaseIterGVN(gvn); 2018 } 2019 2020 // Perform incremental inlining until bound on number of live nodes is reached 2021 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2022 PhaseGVN* gvn = initial_gvn(); 2023 2024 set_inlining_incrementally(true); 2025 set_inlining_progress(true); 2026 uint low_live_nodes = 0; 2027 2028 while(inlining_progress() && _late_inlines.length() > 0) { 2029 2030 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2031 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2032 // PhaseIdealLoop is expensive so we only try it once we are 2033 // out of live nodes and we only try it again if the previous 2034 // helped got the number of nodes down significantly 2035 PhaseIdealLoop ideal_loop( igvn, false, true ); 2036 if (failing()) return; 2037 low_live_nodes = live_nodes(); 2038 _major_progress = true; 2039 } 2040 2041 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2042 break; 2043 } 2044 } 2045 2046 inline_incrementally_one(igvn); 2047 2048 if (failing()) return; 2049 2050 igvn.optimize(); 2051 2052 if (failing()) return; 2053 } 2054 2055 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2056 2057 if (_string_late_inlines.length() > 0) { 2058 assert(has_stringbuilder(), "inconsistent"); 2059 for_igvn()->clear(); 2060 initial_gvn()->replace_with(&igvn); 2061 2062 inline_string_calls(false); 2063 2064 if (failing()) return; 2065 2066 { 2067 ResourceMark rm; 2068 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2069 } 2070 2071 igvn = PhaseIterGVN(gvn); 2072 2073 igvn.optimize(); 2074 } 2075 2076 set_inlining_incrementally(false); 2077 } 2078 2079 2080 //------------------------------Optimize--------------------------------------- 2081 // Given a graph, optimize it. 2082 void Compile::Optimize() { 2083 TracePhase t1("optimizer", &_t_optimizer, true); 2084 2085 #ifndef PRODUCT 2086 if (env()->break_at_compile()) { 2087 BREAKPOINT; 2088 } 2089 2090 #endif 2091 2092 ResourceMark rm; 2093 int loop_opts_cnt; 2094 2095 NOT_PRODUCT( verify_graph_edges(); ) 2096 2097 print_method(PHASE_AFTER_PARSING); 2098 2099 { 2100 // Iterative Global Value Numbering, including ideal transforms 2101 // Initialize IterGVN with types and values from parse-time GVN 2102 PhaseIterGVN igvn(initial_gvn()); 2103 { 2104 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2105 igvn.optimize(); 2106 } 2107 2108 print_method(PHASE_ITER_GVN1, 2); 2109 2110 if (failing()) return; 2111 2112 { 2113 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2114 inline_incrementally(igvn); 2115 } 2116 2117 print_method(PHASE_INCREMENTAL_INLINE, 2); 2118 2119 if (failing()) return; 2120 2121 if (eliminate_boxing()) { 2122 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2123 // Inline valueOf() methods now. 2124 inline_boxing_calls(igvn); 2125 2126 if (AlwaysIncrementalInline) { 2127 inline_incrementally(igvn); 2128 } 2129 2130 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2131 2132 if (failing()) return; 2133 } 2134 2135 // Remove the speculative part of types and clean up the graph from 2136 // the extra CastPP nodes whose only purpose is to carry them. Do 2137 // that early so that optimizations are not disrupted by the extra 2138 // CastPP nodes. 2139 remove_speculative_types(igvn); 2140 2141 // No more new expensive nodes will be added to the list from here 2142 // so keep only the actual candidates for optimizations. 2143 cleanup_expensive_nodes(igvn); 2144 2145 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2146 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);) 2147 initial_gvn()->replace_with(&igvn); 2148 for_igvn()->clear(); 2149 Unique_Node_List new_worklist(C->comp_arena()); 2150 { 2151 ResourceMark rm; 2152 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2153 } 2154 set_for_igvn(&new_worklist); 2155 igvn = PhaseIterGVN(initial_gvn()); 2156 igvn.optimize(); 2157 } 2158 2159 // Perform escape analysis 2160 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2161 if (has_loops()) { 2162 // Cleanup graph (remove dead nodes). 2163 TracePhase t2("idealLoop", &_t_idealLoop, true); 2164 PhaseIdealLoop ideal_loop( igvn, false, true ); 2165 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2166 if (failing()) return; 2167 } 2168 ConnectionGraph::do_analysis(this, &igvn); 2169 2170 if (failing()) return; 2171 2172 // Optimize out fields loads from scalar replaceable allocations. 2173 igvn.optimize(); 2174 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2175 2176 if (failing()) return; 2177 2178 if (congraph() != NULL && macro_count() > 0) { 2179 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2180 PhaseMacroExpand mexp(igvn); 2181 mexp.eliminate_macro_nodes(); 2182 igvn.set_delay_transform(false); 2183 2184 igvn.optimize(); 2185 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2186 2187 if (failing()) return; 2188 } 2189 } 2190 2191 // Loop transforms on the ideal graph. Range Check Elimination, 2192 // peeling, unrolling, etc. 2193 2194 // Set loop opts counter 2195 loop_opts_cnt = num_loop_opts(); 2196 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2197 { 2198 TracePhase t2("idealLoop", &_t_idealLoop, true); 2199 PhaseIdealLoop ideal_loop( igvn, true ); 2200 loop_opts_cnt--; 2201 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2202 if (failing()) return; 2203 } 2204 // Loop opts pass if partial peeling occurred in previous pass 2205 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2206 TracePhase t3("idealLoop", &_t_idealLoop, true); 2207 PhaseIdealLoop ideal_loop( igvn, false ); 2208 loop_opts_cnt--; 2209 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2210 if (failing()) return; 2211 } 2212 // Loop opts pass for loop-unrolling before CCP 2213 if(major_progress() && (loop_opts_cnt > 0)) { 2214 TracePhase t4("idealLoop", &_t_idealLoop, true); 2215 PhaseIdealLoop ideal_loop( igvn, false ); 2216 loop_opts_cnt--; 2217 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2218 } 2219 if (!failing()) { 2220 // Verify that last round of loop opts produced a valid graph 2221 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2222 PhaseIdealLoop::verify(igvn); 2223 } 2224 } 2225 if (failing()) return; 2226 2227 // Conditional Constant Propagation; 2228 PhaseCCP ccp( &igvn ); 2229 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2230 { 2231 TracePhase t2("ccp", &_t_ccp, true); 2232 ccp.do_transform(); 2233 } 2234 print_method(PHASE_CPP1, 2); 2235 2236 assert( true, "Break here to ccp.dump_old2new_map()"); 2237 2238 // Iterative Global Value Numbering, including ideal transforms 2239 { 2240 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2241 igvn = ccp; 2242 igvn.optimize(); 2243 } 2244 2245 print_method(PHASE_ITER_GVN2, 2); 2246 2247 if (failing()) return; 2248 2249 // Loop transforms on the ideal graph. Range Check Elimination, 2250 // peeling, unrolling, etc. 2251 if(loop_opts_cnt > 0) { 2252 debug_only( int cnt = 0; ); 2253 while(major_progress() && (loop_opts_cnt > 0)) { 2254 TracePhase t2("idealLoop", &_t_idealLoop, true); 2255 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2256 PhaseIdealLoop ideal_loop( igvn, true); 2257 loop_opts_cnt--; 2258 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2259 if (failing()) return; 2260 } 2261 } 2262 2263 { 2264 // Verify that all previous optimizations produced a valid graph 2265 // at least to this point, even if no loop optimizations were done. 2266 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2267 PhaseIdealLoop::verify(igvn); 2268 } 2269 2270 if (range_check_cast_count() > 0) { 2271 // No more loop optimizations. Remove all range check dependent CastIINodes. 2272 C->remove_range_check_casts(igvn); 2273 igvn.optimize(); 2274 } 2275 2276 { 2277 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2278 PhaseMacroExpand mex(igvn); 2279 if (mex.expand_macro_nodes()) { 2280 assert(failing(), "must bail out w/ explicit message"); 2281 return; 2282 } 2283 } 2284 2285 } // (End scope of igvn; run destructor if necessary for asserts.) 2286 2287 dump_inlining(); 2288 // A method with only infinite loops has no edges entering loops from root 2289 { 2290 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2291 if (final_graph_reshaping()) { 2292 assert(failing(), "must bail out w/ explicit message"); 2293 return; 2294 } 2295 } 2296 2297 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2298 } 2299 2300 2301 //------------------------------Code_Gen--------------------------------------- 2302 // Given a graph, generate code for it 2303 void Compile::Code_Gen() { 2304 if (failing()) { 2305 return; 2306 } 2307 2308 // Perform instruction selection. You might think we could reclaim Matcher 2309 // memory PDQ, but actually the Matcher is used in generating spill code. 2310 // Internals of the Matcher (including some VectorSets) must remain live 2311 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2312 // set a bit in reclaimed memory. 2313 2314 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2315 // nodes. Mapping is only valid at the root of each matched subtree. 2316 NOT_PRODUCT( verify_graph_edges(); ) 2317 2318 Matcher matcher; 2319 _matcher = &matcher; 2320 { 2321 TracePhase t2("matcher", &_t_matcher, true); 2322 matcher.match(); 2323 } 2324 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2325 // nodes. Mapping is only valid at the root of each matched subtree. 2326 NOT_PRODUCT( verify_graph_edges(); ) 2327 2328 // If you have too many nodes, or if matching has failed, bail out 2329 check_node_count(0, "out of nodes matching instructions"); 2330 if (failing()) { 2331 return; 2332 } 2333 2334 // Build a proper-looking CFG 2335 PhaseCFG cfg(node_arena(), root(), matcher); 2336 _cfg = &cfg; 2337 { 2338 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2339 bool success = cfg.do_global_code_motion(); 2340 if (!success) { 2341 return; 2342 } 2343 2344 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2345 NOT_PRODUCT( verify_graph_edges(); ) 2346 debug_only( cfg.verify(); ) 2347 } 2348 2349 PhaseChaitin regalloc(unique(), cfg, matcher); 2350 _regalloc = ®alloc; 2351 { 2352 TracePhase t2("regalloc", &_t_registerAllocation, true); 2353 // Perform register allocation. After Chaitin, use-def chains are 2354 // no longer accurate (at spill code) and so must be ignored. 2355 // Node->LRG->reg mappings are still accurate. 2356 _regalloc->Register_Allocate(); 2357 2358 // Bail out if the allocator builds too many nodes 2359 if (failing()) { 2360 return; 2361 } 2362 } 2363 2364 // Prior to register allocation we kept empty basic blocks in case the 2365 // the allocator needed a place to spill. After register allocation we 2366 // are not adding any new instructions. If any basic block is empty, we 2367 // can now safely remove it. 2368 { 2369 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2370 cfg.remove_empty_blocks(); 2371 if (do_freq_based_layout()) { 2372 PhaseBlockLayout layout(cfg); 2373 } else { 2374 cfg.set_loop_alignment(); 2375 } 2376 cfg.fixup_flow(); 2377 } 2378 2379 // Apply peephole optimizations 2380 if( OptoPeephole ) { 2381 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2382 PhasePeephole peep( _regalloc, cfg); 2383 peep.do_transform(); 2384 } 2385 2386 // Do late expand if CPU requires this. 2387 if (Matcher::require_postalloc_expand) { 2388 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2389 cfg.postalloc_expand(_regalloc); 2390 } 2391 2392 // Convert Nodes to instruction bits in a buffer 2393 { 2394 // %%%% workspace merge brought two timers together for one job 2395 TracePhase t2a("output", &_t_output, true); 2396 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2397 Output(); 2398 } 2399 2400 print_method(PHASE_FINAL_CODE); 2401 2402 // He's dead, Jim. 2403 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2404 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2405 } 2406 2407 2408 //------------------------------dump_asm--------------------------------------- 2409 // Dump formatted assembly 2410 #ifndef PRODUCT 2411 void Compile::dump_asm(int *pcs, uint pc_limit) { 2412 bool cut_short = false; 2413 tty->print_cr("#"); 2414 tty->print("# "); _tf->dump(); tty->cr(); 2415 tty->print_cr("#"); 2416 2417 // For all blocks 2418 int pc = 0x0; // Program counter 2419 char starts_bundle = ' '; 2420 _regalloc->dump_frame(); 2421 2422 Node *n = NULL; 2423 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2424 if (VMThread::should_terminate()) { 2425 cut_short = true; 2426 break; 2427 } 2428 Block* block = _cfg->get_block(i); 2429 if (block->is_connector() && !Verbose) { 2430 continue; 2431 } 2432 n = block->head(); 2433 if (pcs && n->_idx < pc_limit) { 2434 tty->print("%3.3x ", pcs[n->_idx]); 2435 } else { 2436 tty->print(" "); 2437 } 2438 block->dump_head(_cfg); 2439 if (block->is_connector()) { 2440 tty->print_cr(" # Empty connector block"); 2441 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2442 tty->print_cr(" # Block is sole successor of call"); 2443 } 2444 2445 // For all instructions 2446 Node *delay = NULL; 2447 for (uint j = 0; j < block->number_of_nodes(); j++) { 2448 if (VMThread::should_terminate()) { 2449 cut_short = true; 2450 break; 2451 } 2452 n = block->get_node(j); 2453 if (valid_bundle_info(n)) { 2454 Bundle* bundle = node_bundling(n); 2455 if (bundle->used_in_unconditional_delay()) { 2456 delay = n; 2457 continue; 2458 } 2459 if (bundle->starts_bundle()) { 2460 starts_bundle = '+'; 2461 } 2462 } 2463 2464 if (WizardMode) { 2465 n->dump(); 2466 } 2467 2468 if( !n->is_Region() && // Dont print in the Assembly 2469 !n->is_Phi() && // a few noisely useless nodes 2470 !n->is_Proj() && 2471 !n->is_MachTemp() && 2472 !n->is_SafePointScalarObject() && 2473 !n->is_Catch() && // Would be nice to print exception table targets 2474 !n->is_MergeMem() && // Not very interesting 2475 !n->is_top() && // Debug info table constants 2476 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2477 ) { 2478 if (pcs && n->_idx < pc_limit) 2479 tty->print("%3.3x", pcs[n->_idx]); 2480 else 2481 tty->print(" "); 2482 tty->print(" %c ", starts_bundle); 2483 starts_bundle = ' '; 2484 tty->print("\t"); 2485 n->format(_regalloc, tty); 2486 tty->cr(); 2487 } 2488 2489 // If we have an instruction with a delay slot, and have seen a delay, 2490 // then back up and print it 2491 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2492 assert(delay != NULL, "no unconditional delay instruction"); 2493 if (WizardMode) delay->dump(); 2494 2495 if (node_bundling(delay)->starts_bundle()) 2496 starts_bundle = '+'; 2497 if (pcs && n->_idx < pc_limit) 2498 tty->print("%3.3x", pcs[n->_idx]); 2499 else 2500 tty->print(" "); 2501 tty->print(" %c ", starts_bundle); 2502 starts_bundle = ' '; 2503 tty->print("\t"); 2504 delay->format(_regalloc, tty); 2505 tty->cr(); 2506 delay = NULL; 2507 } 2508 2509 // Dump the exception table as well 2510 if( n->is_Catch() && (Verbose || WizardMode) ) { 2511 // Print the exception table for this offset 2512 _handler_table.print_subtable_for(pc); 2513 } 2514 } 2515 2516 if (pcs && n->_idx < pc_limit) 2517 tty->print_cr("%3.3x", pcs[n->_idx]); 2518 else 2519 tty->cr(); 2520 2521 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2522 2523 } // End of per-block dump 2524 tty->cr(); 2525 2526 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2527 } 2528 #endif 2529 2530 //------------------------------Final_Reshape_Counts--------------------------- 2531 // This class defines counters to help identify when a method 2532 // may/must be executed using hardware with only 24-bit precision. 2533 struct Final_Reshape_Counts : public StackObj { 2534 int _call_count; // count non-inlined 'common' calls 2535 int _float_count; // count float ops requiring 24-bit precision 2536 int _double_count; // count double ops requiring more precision 2537 int _java_call_count; // count non-inlined 'java' calls 2538 int _inner_loop_count; // count loops which need alignment 2539 VectorSet _visited; // Visitation flags 2540 Node_List _tests; // Set of IfNodes & PCTableNodes 2541 2542 Final_Reshape_Counts() : 2543 _call_count(0), _float_count(0), _double_count(0), 2544 _java_call_count(0), _inner_loop_count(0), 2545 _visited( Thread::current()->resource_area() ) { } 2546 2547 void inc_call_count () { _call_count ++; } 2548 void inc_float_count () { _float_count ++; } 2549 void inc_double_count() { _double_count++; } 2550 void inc_java_call_count() { _java_call_count++; } 2551 void inc_inner_loop_count() { _inner_loop_count++; } 2552 2553 int get_call_count () const { return _call_count ; } 2554 int get_float_count () const { return _float_count ; } 2555 int get_double_count() const { return _double_count; } 2556 int get_java_call_count() const { return _java_call_count; } 2557 int get_inner_loop_count() const { return _inner_loop_count; } 2558 }; 2559 2560 #ifdef ASSERT 2561 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2562 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2563 // Make sure the offset goes inside the instance layout. 2564 return k->contains_field_offset(tp->offset()); 2565 // Note that OffsetBot and OffsetTop are very negative. 2566 } 2567 #endif 2568 2569 // Eliminate trivially redundant StoreCMs and accumulate their 2570 // precedence edges. 2571 void Compile::eliminate_redundant_card_marks(Node* n) { 2572 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2573 if (n->in(MemNode::Address)->outcnt() > 1) { 2574 // There are multiple users of the same address so it might be 2575 // possible to eliminate some of the StoreCMs 2576 Node* mem = n->in(MemNode::Memory); 2577 Node* adr = n->in(MemNode::Address); 2578 Node* val = n->in(MemNode::ValueIn); 2579 Node* prev = n; 2580 bool done = false; 2581 // Walk the chain of StoreCMs eliminating ones that match. As 2582 // long as it's a chain of single users then the optimization is 2583 // safe. Eliminating partially redundant StoreCMs would require 2584 // cloning copies down the other paths. 2585 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2586 if (adr == mem->in(MemNode::Address) && 2587 val == mem->in(MemNode::ValueIn)) { 2588 // redundant StoreCM 2589 if (mem->req() > MemNode::OopStore) { 2590 // Hasn't been processed by this code yet. 2591 n->add_prec(mem->in(MemNode::OopStore)); 2592 } else { 2593 // Already converted to precedence edge 2594 for (uint i = mem->req(); i < mem->len(); i++) { 2595 // Accumulate any precedence edges 2596 if (mem->in(i) != NULL) { 2597 n->add_prec(mem->in(i)); 2598 } 2599 } 2600 // Everything above this point has been processed. 2601 done = true; 2602 } 2603 // Eliminate the previous StoreCM 2604 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2605 assert(mem->outcnt() == 0, "should be dead"); 2606 mem->disconnect_inputs(NULL, this); 2607 } else { 2608 prev = mem; 2609 } 2610 mem = prev->in(MemNode::Memory); 2611 } 2612 } 2613 } 2614 2615 //------------------------------final_graph_reshaping_impl---------------------- 2616 // Implement items 1-5 from final_graph_reshaping below. 2617 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2618 2619 if ( n->outcnt() == 0 ) return; // dead node 2620 uint nop = n->Opcode(); 2621 2622 // Check for 2-input instruction with "last use" on right input. 2623 // Swap to left input. Implements item (2). 2624 if( n->req() == 3 && // two-input instruction 2625 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2626 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2627 n->in(2)->outcnt() == 1 &&// right use IS a last use 2628 !n->in(2)->is_Con() ) { // right use is not a constant 2629 // Check for commutative opcode 2630 switch( nop ) { 2631 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2632 case Op_MaxI: case Op_MinI: 2633 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2634 case Op_AndL: case Op_XorL: case Op_OrL: 2635 case Op_AndI: case Op_XorI: case Op_OrI: { 2636 // Move "last use" input to left by swapping inputs 2637 n->swap_edges(1, 2); 2638 break; 2639 } 2640 default: 2641 break; 2642 } 2643 } 2644 2645 #ifdef ASSERT 2646 if( n->is_Mem() ) { 2647 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2648 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2649 // oop will be recorded in oop map if load crosses safepoint 2650 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2651 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2652 "raw memory operations should have control edge"); 2653 } 2654 #endif 2655 // Count FPU ops and common calls, implements item (3) 2656 switch( nop ) { 2657 // Count all float operations that may use FPU 2658 case Op_AddF: 2659 case Op_SubF: 2660 case Op_MulF: 2661 case Op_DivF: 2662 case Op_NegF: 2663 case Op_ModF: 2664 case Op_ConvI2F: 2665 case Op_ConF: 2666 case Op_CmpF: 2667 case Op_CmpF3: 2668 // case Op_ConvL2F: // longs are split into 32-bit halves 2669 frc.inc_float_count(); 2670 break; 2671 2672 case Op_ConvF2D: 2673 case Op_ConvD2F: 2674 frc.inc_float_count(); 2675 frc.inc_double_count(); 2676 break; 2677 2678 // Count all double operations that may use FPU 2679 case Op_AddD: 2680 case Op_SubD: 2681 case Op_MulD: 2682 case Op_DivD: 2683 case Op_NegD: 2684 case Op_ModD: 2685 case Op_ConvI2D: 2686 case Op_ConvD2I: 2687 // case Op_ConvL2D: // handled by leaf call 2688 // case Op_ConvD2L: // handled by leaf call 2689 case Op_ConD: 2690 case Op_CmpD: 2691 case Op_CmpD3: 2692 frc.inc_double_count(); 2693 break; 2694 case Op_Opaque1: // Remove Opaque Nodes before matching 2695 case Op_Opaque2: // Remove Opaque Nodes before matching 2696 case Op_Opaque3: 2697 n->subsume_by(n->in(1), this); 2698 break; 2699 case Op_CallStaticJava: 2700 case Op_CallJava: 2701 case Op_CallDynamicJava: 2702 frc.inc_java_call_count(); // Count java call site; 2703 case Op_CallRuntime: 2704 case Op_CallLeaf: 2705 case Op_CallLeafNoFP: { 2706 assert( n->is_Call(), "" ); 2707 CallNode *call = n->as_Call(); 2708 // Count call sites where the FP mode bit would have to be flipped. 2709 // Do not count uncommon runtime calls: 2710 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2711 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2712 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2713 frc.inc_call_count(); // Count the call site 2714 } else { // See if uncommon argument is shared 2715 Node *n = call->in(TypeFunc::Parms); 2716 int nop = n->Opcode(); 2717 // Clone shared simple arguments to uncommon calls, item (1). 2718 if( n->outcnt() > 1 && 2719 !n->is_Proj() && 2720 nop != Op_CreateEx && 2721 nop != Op_CheckCastPP && 2722 nop != Op_DecodeN && 2723 nop != Op_DecodeNKlass && 2724 !n->is_Mem() ) { 2725 Node *x = n->clone(); 2726 call->set_req( TypeFunc::Parms, x ); 2727 } 2728 } 2729 break; 2730 } 2731 2732 case Op_StoreD: 2733 case Op_LoadD: 2734 case Op_LoadD_unaligned: 2735 frc.inc_double_count(); 2736 goto handle_mem; 2737 case Op_StoreF: 2738 case Op_LoadF: 2739 frc.inc_float_count(); 2740 goto handle_mem; 2741 2742 case Op_StoreCM: 2743 { 2744 // Convert OopStore dependence into precedence edge 2745 Node* prec = n->in(MemNode::OopStore); 2746 n->del_req(MemNode::OopStore); 2747 n->add_prec(prec); 2748 eliminate_redundant_card_marks(n); 2749 } 2750 2751 // fall through 2752 2753 case Op_StoreB: 2754 case Op_StoreC: 2755 case Op_StorePConditional: 2756 case Op_StoreI: 2757 case Op_StoreL: 2758 case Op_StoreIConditional: 2759 case Op_StoreLConditional: 2760 case Op_CompareAndSwapI: 2761 case Op_CompareAndSwapL: 2762 case Op_CompareAndSwapP: 2763 case Op_CompareAndSwapN: 2764 case Op_GetAndAddI: 2765 case Op_GetAndAddL: 2766 case Op_GetAndSetI: 2767 case Op_GetAndSetL: 2768 case Op_GetAndSetP: 2769 case Op_GetAndSetN: 2770 case Op_StoreP: 2771 case Op_StoreN: 2772 case Op_StoreNKlass: 2773 case Op_LoadB: 2774 case Op_LoadUB: 2775 case Op_LoadUS: 2776 case Op_LoadI: 2777 case Op_LoadKlass: 2778 case Op_LoadNKlass: 2779 case Op_LoadL: 2780 case Op_LoadL_unaligned: 2781 case Op_LoadPLocked: 2782 case Op_LoadP: 2783 case Op_LoadN: 2784 case Op_LoadRange: 2785 case Op_LoadS: { 2786 handle_mem: 2787 #ifdef ASSERT 2788 if( VerifyOptoOopOffsets ) { 2789 assert( n->is_Mem(), "" ); 2790 MemNode *mem = (MemNode*)n; 2791 // Check to see if address types have grounded out somehow. 2792 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2793 assert( !tp || oop_offset_is_sane(tp), "" ); 2794 } 2795 #endif 2796 break; 2797 } 2798 2799 case Op_AddP: { // Assert sane base pointers 2800 Node *addp = n->in(AddPNode::Address); 2801 assert( !addp->is_AddP() || 2802 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2803 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2804 "Base pointers must match" ); 2805 #ifdef _LP64 2806 if ((UseCompressedOops || UseCompressedClassPointers) && 2807 addp->Opcode() == Op_ConP && 2808 addp == n->in(AddPNode::Base) && 2809 n->in(AddPNode::Offset)->is_Con()) { 2810 // Use addressing with narrow klass to load with offset on x86. 2811 // On sparc loading 32-bits constant and decoding it have less 2812 // instructions (4) then load 64-bits constant (7). 2813 // Do this transformation here since IGVN will convert ConN back to ConP. 2814 const Type* t = addp->bottom_type(); 2815 if (t->isa_oopptr() || t->isa_klassptr()) { 2816 Node* nn = NULL; 2817 2818 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2819 2820 // Look for existing ConN node of the same exact type. 2821 Node* r = root(); 2822 uint cnt = r->outcnt(); 2823 for (uint i = 0; i < cnt; i++) { 2824 Node* m = r->raw_out(i); 2825 if (m!= NULL && m->Opcode() == op && 2826 m->bottom_type()->make_ptr() == t) { 2827 nn = m; 2828 break; 2829 } 2830 } 2831 if (nn != NULL) { 2832 // Decode a narrow oop to match address 2833 // [R12 + narrow_oop_reg<<3 + offset] 2834 if (t->isa_oopptr()) { 2835 nn = new (this) DecodeNNode(nn, t); 2836 } else { 2837 nn = new (this) DecodeNKlassNode(nn, t); 2838 } 2839 n->set_req(AddPNode::Base, nn); 2840 n->set_req(AddPNode::Address, nn); 2841 if (addp->outcnt() == 0) { 2842 addp->disconnect_inputs(NULL, this); 2843 } 2844 } 2845 } 2846 } 2847 #endif 2848 break; 2849 } 2850 2851 #ifdef _LP64 2852 case Op_CastPP: 2853 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2854 Node* in1 = n->in(1); 2855 const Type* t = n->bottom_type(); 2856 Node* new_in1 = in1->clone(); 2857 new_in1->as_DecodeN()->set_type(t); 2858 2859 if (!Matcher::narrow_oop_use_complex_address()) { 2860 // 2861 // x86, ARM and friends can handle 2 adds in addressing mode 2862 // and Matcher can fold a DecodeN node into address by using 2863 // a narrow oop directly and do implicit NULL check in address: 2864 // 2865 // [R12 + narrow_oop_reg<<3 + offset] 2866 // NullCheck narrow_oop_reg 2867 // 2868 // On other platforms (Sparc) we have to keep new DecodeN node and 2869 // use it to do implicit NULL check in address: 2870 // 2871 // decode_not_null narrow_oop_reg, base_reg 2872 // [base_reg + offset] 2873 // NullCheck base_reg 2874 // 2875 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2876 // to keep the information to which NULL check the new DecodeN node 2877 // corresponds to use it as value in implicit_null_check(). 2878 // 2879 new_in1->set_req(0, n->in(0)); 2880 } 2881 2882 n->subsume_by(new_in1, this); 2883 if (in1->outcnt() == 0) { 2884 in1->disconnect_inputs(NULL, this); 2885 } 2886 } 2887 break; 2888 2889 case Op_CmpP: 2890 // Do this transformation here to preserve CmpPNode::sub() and 2891 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2892 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2893 Node* in1 = n->in(1); 2894 Node* in2 = n->in(2); 2895 if (!in1->is_DecodeNarrowPtr()) { 2896 in2 = in1; 2897 in1 = n->in(2); 2898 } 2899 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2900 2901 Node* new_in2 = NULL; 2902 if (in2->is_DecodeNarrowPtr()) { 2903 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2904 new_in2 = in2->in(1); 2905 } else if (in2->Opcode() == Op_ConP) { 2906 const Type* t = in2->bottom_type(); 2907 if (t == TypePtr::NULL_PTR) { 2908 assert(in1->is_DecodeN(), "compare klass to null?"); 2909 // Don't convert CmpP null check into CmpN if compressed 2910 // oops implicit null check is not generated. 2911 // This will allow to generate normal oop implicit null check. 2912 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2913 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2914 // 2915 // This transformation together with CastPP transformation above 2916 // will generated code for implicit NULL checks for compressed oops. 2917 // 2918 // The original code after Optimize() 2919 // 2920 // LoadN memory, narrow_oop_reg 2921 // decode narrow_oop_reg, base_reg 2922 // CmpP base_reg, NULL 2923 // CastPP base_reg // NotNull 2924 // Load [base_reg + offset], val_reg 2925 // 2926 // after these transformations will be 2927 // 2928 // LoadN memory, narrow_oop_reg 2929 // CmpN narrow_oop_reg, NULL 2930 // decode_not_null narrow_oop_reg, base_reg 2931 // Load [base_reg + offset], val_reg 2932 // 2933 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2934 // since narrow oops can be used in debug info now (see the code in 2935 // final_graph_reshaping_walk()). 2936 // 2937 // At the end the code will be matched to 2938 // on x86: 2939 // 2940 // Load_narrow_oop memory, narrow_oop_reg 2941 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2942 // NullCheck narrow_oop_reg 2943 // 2944 // and on sparc: 2945 // 2946 // Load_narrow_oop memory, narrow_oop_reg 2947 // decode_not_null narrow_oop_reg, base_reg 2948 // Load [base_reg + offset], val_reg 2949 // NullCheck base_reg 2950 // 2951 } else if (t->isa_oopptr()) { 2952 new_in2 = ConNode::make(this, t->make_narrowoop()); 2953 } else if (t->isa_klassptr()) { 2954 new_in2 = ConNode::make(this, t->make_narrowklass()); 2955 } 2956 } 2957 if (new_in2 != NULL) { 2958 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2959 n->subsume_by(cmpN, this); 2960 if (in1->outcnt() == 0) { 2961 in1->disconnect_inputs(NULL, this); 2962 } 2963 if (in2->outcnt() == 0) { 2964 in2->disconnect_inputs(NULL, this); 2965 } 2966 } 2967 } 2968 break; 2969 2970 case Op_DecodeN: 2971 case Op_DecodeNKlass: 2972 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2973 // DecodeN could be pinned when it can't be fold into 2974 // an address expression, see the code for Op_CastPP above. 2975 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2976 break; 2977 2978 case Op_EncodeP: 2979 case Op_EncodePKlass: { 2980 Node* in1 = n->in(1); 2981 if (in1->is_DecodeNarrowPtr()) { 2982 n->subsume_by(in1->in(1), this); 2983 } else if (in1->Opcode() == Op_ConP) { 2984 const Type* t = in1->bottom_type(); 2985 if (t == TypePtr::NULL_PTR) { 2986 assert(t->isa_oopptr(), "null klass?"); 2987 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2988 } else if (t->isa_oopptr()) { 2989 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2990 } else if (t->isa_klassptr()) { 2991 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2992 } 2993 } 2994 if (in1->outcnt() == 0) { 2995 in1->disconnect_inputs(NULL, this); 2996 } 2997 break; 2998 } 2999 3000 case Op_Proj: { 3001 if (OptimizeStringConcat) { 3002 ProjNode* p = n->as_Proj(); 3003 if (p->_is_io_use) { 3004 // Separate projections were used for the exception path which 3005 // are normally removed by a late inline. If it wasn't inlined 3006 // then they will hang around and should just be replaced with 3007 // the original one. 3008 Node* proj = NULL; 3009 // Replace with just one 3010 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3011 Node *use = i.get(); 3012 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3013 proj = use; 3014 break; 3015 } 3016 } 3017 assert(proj != NULL, "must be found"); 3018 p->subsume_by(proj, this); 3019 } 3020 } 3021 break; 3022 } 3023 3024 case Op_Phi: 3025 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3026 // The EncodeP optimization may create Phi with the same edges 3027 // for all paths. It is not handled well by Register Allocator. 3028 Node* unique_in = n->in(1); 3029 assert(unique_in != NULL, ""); 3030 uint cnt = n->req(); 3031 for (uint i = 2; i < cnt; i++) { 3032 Node* m = n->in(i); 3033 assert(m != NULL, ""); 3034 if (unique_in != m) 3035 unique_in = NULL; 3036 } 3037 if (unique_in != NULL) { 3038 n->subsume_by(unique_in, this); 3039 } 3040 } 3041 break; 3042 3043 #endif 3044 3045 #ifdef ASSERT 3046 case Op_CastII: 3047 // Verify that all range check dependent CastII nodes were removed. 3048 if (n->isa_CastII()->has_range_check()) { 3049 n->dump(3); 3050 assert(false, "Range check dependent CastII node was not removed"); 3051 } 3052 break; 3053 #endif 3054 3055 case Op_ModI: 3056 if (UseDivMod) { 3057 // Check if a%b and a/b both exist 3058 Node* d = n->find_similar(Op_DivI); 3059 if (d) { 3060 // Replace them with a fused divmod if supported 3061 if (Matcher::has_match_rule(Op_DivModI)) { 3062 DivModINode* divmod = DivModINode::make(this, n); 3063 d->subsume_by(divmod->div_proj(), this); 3064 n->subsume_by(divmod->mod_proj(), this); 3065 } else { 3066 // replace a%b with a-((a/b)*b) 3067 Node* mult = new (this) MulINode(d, d->in(2)); 3068 Node* sub = new (this) SubINode(d->in(1), mult); 3069 n->subsume_by(sub, this); 3070 } 3071 } 3072 } 3073 break; 3074 3075 case Op_ModL: 3076 if (UseDivMod) { 3077 // Check if a%b and a/b both exist 3078 Node* d = n->find_similar(Op_DivL); 3079 if (d) { 3080 // Replace them with a fused divmod if supported 3081 if (Matcher::has_match_rule(Op_DivModL)) { 3082 DivModLNode* divmod = DivModLNode::make(this, n); 3083 d->subsume_by(divmod->div_proj(), this); 3084 n->subsume_by(divmod->mod_proj(), this); 3085 } else { 3086 // replace a%b with a-((a/b)*b) 3087 Node* mult = new (this) MulLNode(d, d->in(2)); 3088 Node* sub = new (this) SubLNode(d->in(1), mult); 3089 n->subsume_by(sub, this); 3090 } 3091 } 3092 } 3093 break; 3094 3095 case Op_LoadVector: 3096 case Op_StoreVector: 3097 break; 3098 3099 case Op_PackB: 3100 case Op_PackS: 3101 case Op_PackI: 3102 case Op_PackF: 3103 case Op_PackL: 3104 case Op_PackD: 3105 if (n->req()-1 > 2) { 3106 // Replace many operand PackNodes with a binary tree for matching 3107 PackNode* p = (PackNode*) n; 3108 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3109 n->subsume_by(btp, this); 3110 } 3111 break; 3112 case Op_Loop: 3113 case Op_CountedLoop: 3114 if (n->as_Loop()->is_inner_loop()) { 3115 frc.inc_inner_loop_count(); 3116 } 3117 break; 3118 case Op_LShiftI: 3119 case Op_RShiftI: 3120 case Op_URShiftI: 3121 case Op_LShiftL: 3122 case Op_RShiftL: 3123 case Op_URShiftL: 3124 if (Matcher::need_masked_shift_count) { 3125 // The cpu's shift instructions don't restrict the count to the 3126 // lower 5/6 bits. We need to do the masking ourselves. 3127 Node* in2 = n->in(2); 3128 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3129 const TypeInt* t = in2->find_int_type(); 3130 if (t != NULL && t->is_con()) { 3131 juint shift = t->get_con(); 3132 if (shift > mask) { // Unsigned cmp 3133 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3134 } 3135 } else { 3136 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3137 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3138 n->set_req(2, shift); 3139 } 3140 } 3141 if (in2->outcnt() == 0) { // Remove dead node 3142 in2->disconnect_inputs(NULL, this); 3143 } 3144 } 3145 break; 3146 case Op_MemBarStoreStore: 3147 case Op_MemBarRelease: 3148 // Break the link with AllocateNode: it is no longer useful and 3149 // confuses register allocation. 3150 if (n->req() > MemBarNode::Precedent) { 3151 n->set_req(MemBarNode::Precedent, top()); 3152 } 3153 break; 3154 default: 3155 assert( !n->is_Call(), "" ); 3156 assert( !n->is_Mem(), "" ); 3157 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3158 break; 3159 } 3160 3161 // Collect CFG split points 3162 if (n->is_MultiBranch()) 3163 frc._tests.push(n); 3164 } 3165 3166 //------------------------------final_graph_reshaping_walk--------------------- 3167 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3168 // requires that the walk visits a node's inputs before visiting the node. 3169 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3170 ResourceArea *area = Thread::current()->resource_area(); 3171 Unique_Node_List sfpt(area); 3172 3173 frc._visited.set(root->_idx); // first, mark node as visited 3174 uint cnt = root->req(); 3175 Node *n = root; 3176 uint i = 0; 3177 while (true) { 3178 if (i < cnt) { 3179 // Place all non-visited non-null inputs onto stack 3180 Node* m = n->in(i); 3181 ++i; 3182 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3183 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3184 // compute worst case interpreter size in case of a deoptimization 3185 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3186 3187 sfpt.push(m); 3188 } 3189 cnt = m->req(); 3190 nstack.push(n, i); // put on stack parent and next input's index 3191 n = m; 3192 i = 0; 3193 } 3194 } else { 3195 // Now do post-visit work 3196 final_graph_reshaping_impl( n, frc ); 3197 if (nstack.is_empty()) 3198 break; // finished 3199 n = nstack.node(); // Get node from stack 3200 cnt = n->req(); 3201 i = nstack.index(); 3202 nstack.pop(); // Shift to the next node on stack 3203 } 3204 } 3205 3206 // Skip next transformation if compressed oops are not used. 3207 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3208 (!UseCompressedOops && !UseCompressedClassPointers)) 3209 return; 3210 3211 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3212 // It could be done for an uncommon traps or any safepoints/calls 3213 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3214 while (sfpt.size() > 0) { 3215 n = sfpt.pop(); 3216 JVMState *jvms = n->as_SafePoint()->jvms(); 3217 assert(jvms != NULL, "sanity"); 3218 int start = jvms->debug_start(); 3219 int end = n->req(); 3220 bool is_uncommon = (n->is_CallStaticJava() && 3221 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3222 for (int j = start; j < end; j++) { 3223 Node* in = n->in(j); 3224 if (in->is_DecodeNarrowPtr()) { 3225 bool safe_to_skip = true; 3226 if (!is_uncommon ) { 3227 // Is it safe to skip? 3228 for (uint i = 0; i < in->outcnt(); i++) { 3229 Node* u = in->raw_out(i); 3230 if (!u->is_SafePoint() || 3231 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3232 safe_to_skip = false; 3233 } 3234 } 3235 } 3236 if (safe_to_skip) { 3237 n->set_req(j, in->in(1)); 3238 } 3239 if (in->outcnt() == 0) { 3240 in->disconnect_inputs(NULL, this); 3241 } 3242 } 3243 } 3244 } 3245 } 3246 3247 //------------------------------final_graph_reshaping-------------------------- 3248 // Final Graph Reshaping. 3249 // 3250 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3251 // and not commoned up and forced early. Must come after regular 3252 // optimizations to avoid GVN undoing the cloning. Clone constant 3253 // inputs to Loop Phis; these will be split by the allocator anyways. 3254 // Remove Opaque nodes. 3255 // (2) Move last-uses by commutative operations to the left input to encourage 3256 // Intel update-in-place two-address operations and better register usage 3257 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3258 // calls canonicalizing them back. 3259 // (3) Count the number of double-precision FP ops, single-precision FP ops 3260 // and call sites. On Intel, we can get correct rounding either by 3261 // forcing singles to memory (requires extra stores and loads after each 3262 // FP bytecode) or we can set a rounding mode bit (requires setting and 3263 // clearing the mode bit around call sites). The mode bit is only used 3264 // if the relative frequency of single FP ops to calls is low enough. 3265 // This is a key transform for SPEC mpeg_audio. 3266 // (4) Detect infinite loops; blobs of code reachable from above but not 3267 // below. Several of the Code_Gen algorithms fail on such code shapes, 3268 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3269 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3270 // Detection is by looking for IfNodes where only 1 projection is 3271 // reachable from below or CatchNodes missing some targets. 3272 // (5) Assert for insane oop offsets in debug mode. 3273 3274 bool Compile::final_graph_reshaping() { 3275 // an infinite loop may have been eliminated by the optimizer, 3276 // in which case the graph will be empty. 3277 if (root()->req() == 1) { 3278 record_method_not_compilable("trivial infinite loop"); 3279 return true; 3280 } 3281 3282 // Expensive nodes have their control input set to prevent the GVN 3283 // from freely commoning them. There's no GVN beyond this point so 3284 // no need to keep the control input. We want the expensive nodes to 3285 // be freely moved to the least frequent code path by gcm. 3286 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3287 for (int i = 0; i < expensive_count(); i++) { 3288 _expensive_nodes->at(i)->set_req(0, NULL); 3289 } 3290 3291 Final_Reshape_Counts frc; 3292 3293 // Visit everybody reachable! 3294 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3295 Node_Stack nstack(live_nodes() >> 1); 3296 final_graph_reshaping_walk(nstack, root(), frc); 3297 3298 // Check for unreachable (from below) code (i.e., infinite loops). 3299 for( uint i = 0; i < frc._tests.size(); i++ ) { 3300 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3301 // Get number of CFG targets. 3302 // Note that PCTables include exception targets after calls. 3303 uint required_outcnt = n->required_outcnt(); 3304 if (n->outcnt() != required_outcnt) { 3305 // Check for a few special cases. Rethrow Nodes never take the 3306 // 'fall-thru' path, so expected kids is 1 less. 3307 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3308 if (n->in(0)->in(0)->is_Call()) { 3309 CallNode *call = n->in(0)->in(0)->as_Call(); 3310 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3311 required_outcnt--; // Rethrow always has 1 less kid 3312 } else if (call->req() > TypeFunc::Parms && 3313 call->is_CallDynamicJava()) { 3314 // Check for null receiver. In such case, the optimizer has 3315 // detected that the virtual call will always result in a null 3316 // pointer exception. The fall-through projection of this CatchNode 3317 // will not be populated. 3318 Node *arg0 = call->in(TypeFunc::Parms); 3319 if (arg0->is_Type() && 3320 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3321 required_outcnt--; 3322 } 3323 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3324 call->req() > TypeFunc::Parms+1 && 3325 call->is_CallStaticJava()) { 3326 // Check for negative array length. In such case, the optimizer has 3327 // detected that the allocation attempt will always result in an 3328 // exception. There is no fall-through projection of this CatchNode . 3329 Node *arg1 = call->in(TypeFunc::Parms+1); 3330 if (arg1->is_Type() && 3331 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3332 required_outcnt--; 3333 } 3334 } 3335 } 3336 } 3337 // Recheck with a better notion of 'required_outcnt' 3338 if (n->outcnt() != required_outcnt) { 3339 record_method_not_compilable("malformed control flow"); 3340 return true; // Not all targets reachable! 3341 } 3342 } 3343 // Check that I actually visited all kids. Unreached kids 3344 // must be infinite loops. 3345 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3346 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3347 record_method_not_compilable("infinite loop"); 3348 return true; // Found unvisited kid; must be unreach 3349 } 3350 } 3351 3352 // If original bytecodes contained a mixture of floats and doubles 3353 // check if the optimizer has made it homogenous, item (3). 3354 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3355 frc.get_float_count() > 32 && 3356 frc.get_double_count() == 0 && 3357 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3358 set_24_bit_selection_and_mode( false, true ); 3359 } 3360 3361 set_java_calls(frc.get_java_call_count()); 3362 set_inner_loops(frc.get_inner_loop_count()); 3363 3364 // No infinite loops, no reason to bail out. 3365 return false; 3366 } 3367 3368 //-----------------------------too_many_traps---------------------------------- 3369 // Report if there are too many traps at the current method and bci. 3370 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3371 bool Compile::too_many_traps(ciMethod* method, 3372 int bci, 3373 Deoptimization::DeoptReason reason) { 3374 ciMethodData* md = method->method_data(); 3375 if (md->is_empty()) { 3376 // Assume the trap has not occurred, or that it occurred only 3377 // because of a transient condition during start-up in the interpreter. 3378 return false; 3379 } 3380 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3381 if (md->has_trap_at(bci, m, reason) != 0) { 3382 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3383 // Also, if there are multiple reasons, or if there is no per-BCI record, 3384 // assume the worst. 3385 if (log()) 3386 log()->elem("observe trap='%s' count='%d'", 3387 Deoptimization::trap_reason_name(reason), 3388 md->trap_count(reason)); 3389 return true; 3390 } else { 3391 // Ignore method/bci and see if there have been too many globally. 3392 return too_many_traps(reason, md); 3393 } 3394 } 3395 3396 // Less-accurate variant which does not require a method and bci. 3397 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3398 ciMethodData* logmd) { 3399 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3400 // Too many traps globally. 3401 // Note that we use cumulative trap_count, not just md->trap_count. 3402 if (log()) { 3403 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3404 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3405 Deoptimization::trap_reason_name(reason), 3406 mcount, trap_count(reason)); 3407 } 3408 return true; 3409 } else { 3410 // The coast is clear. 3411 return false; 3412 } 3413 } 3414 3415 //--------------------------too_many_recompiles-------------------------------- 3416 // Report if there are too many recompiles at the current method and bci. 3417 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3418 // Is not eager to return true, since this will cause the compiler to use 3419 // Action_none for a trap point, to avoid too many recompilations. 3420 bool Compile::too_many_recompiles(ciMethod* method, 3421 int bci, 3422 Deoptimization::DeoptReason reason) { 3423 ciMethodData* md = method->method_data(); 3424 if (md->is_empty()) { 3425 // Assume the trap has not occurred, or that it occurred only 3426 // because of a transient condition during start-up in the interpreter. 3427 return false; 3428 } 3429 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3430 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3431 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3432 Deoptimization::DeoptReason per_bc_reason 3433 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3434 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3435 if ((per_bc_reason == Deoptimization::Reason_none 3436 || md->has_trap_at(bci, m, reason) != 0) 3437 // The trap frequency measure we care about is the recompile count: 3438 && md->trap_recompiled_at(bci, m) 3439 && md->overflow_recompile_count() >= bc_cutoff) { 3440 // Do not emit a trap here if it has already caused recompilations. 3441 // Also, if there are multiple reasons, or if there is no per-BCI record, 3442 // assume the worst. 3443 if (log()) 3444 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3445 Deoptimization::trap_reason_name(reason), 3446 md->trap_count(reason), 3447 md->overflow_recompile_count()); 3448 return true; 3449 } else if (trap_count(reason) != 0 3450 && decompile_count() >= m_cutoff) { 3451 // Too many recompiles globally, and we have seen this sort of trap. 3452 // Use cumulative decompile_count, not just md->decompile_count. 3453 if (log()) 3454 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3455 Deoptimization::trap_reason_name(reason), 3456 md->trap_count(reason), trap_count(reason), 3457 md->decompile_count(), decompile_count()); 3458 return true; 3459 } else { 3460 // The coast is clear. 3461 return false; 3462 } 3463 } 3464 3465 // Compute when not to trap. Used by matching trap based nodes and 3466 // NullCheck optimization. 3467 void Compile::set_allowed_deopt_reasons() { 3468 _allowed_reasons = 0; 3469 if (is_method_compilation()) { 3470 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3471 assert(rs < BitsPerInt, "recode bit map"); 3472 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3473 _allowed_reasons |= nth_bit(rs); 3474 } 3475 } 3476 } 3477 } 3478 3479 #ifndef PRODUCT 3480 //------------------------------verify_graph_edges--------------------------- 3481 // Walk the Graph and verify that there is a one-to-one correspondence 3482 // between Use-Def edges and Def-Use edges in the graph. 3483 void Compile::verify_graph_edges(bool no_dead_code) { 3484 if (VerifyGraphEdges) { 3485 ResourceArea *area = Thread::current()->resource_area(); 3486 Unique_Node_List visited(area); 3487 // Call recursive graph walk to check edges 3488 _root->verify_edges(visited); 3489 if (no_dead_code) { 3490 // Now make sure that no visited node is used by an unvisited node. 3491 bool dead_nodes = false; 3492 Unique_Node_List checked(area); 3493 while (visited.size() > 0) { 3494 Node* n = visited.pop(); 3495 checked.push(n); 3496 for (uint i = 0; i < n->outcnt(); i++) { 3497 Node* use = n->raw_out(i); 3498 if (checked.member(use)) continue; // already checked 3499 if (visited.member(use)) continue; // already in the graph 3500 if (use->is_Con()) continue; // a dead ConNode is OK 3501 // At this point, we have found a dead node which is DU-reachable. 3502 if (!dead_nodes) { 3503 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3504 dead_nodes = true; 3505 } 3506 use->dump(2); 3507 tty->print_cr("---"); 3508 checked.push(use); // No repeats; pretend it is now checked. 3509 } 3510 } 3511 assert(!dead_nodes, "using nodes must be reachable from root"); 3512 } 3513 } 3514 } 3515 3516 // Verify GC barriers consistency 3517 // Currently supported: 3518 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3519 void Compile::verify_barriers() { 3520 if (UseG1GC) { 3521 // Verify G1 pre-barriers 3522 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3523 3524 ResourceArea *area = Thread::current()->resource_area(); 3525 Unique_Node_List visited(area); 3526 Node_List worklist(area); 3527 // We're going to walk control flow backwards starting from the Root 3528 worklist.push(_root); 3529 while (worklist.size() > 0) { 3530 Node* x = worklist.pop(); 3531 if (x == NULL || x == top()) continue; 3532 if (visited.member(x)) { 3533 continue; 3534 } else { 3535 visited.push(x); 3536 } 3537 3538 if (x->is_Region()) { 3539 for (uint i = 1; i < x->req(); i++) { 3540 worklist.push(x->in(i)); 3541 } 3542 } else { 3543 worklist.push(x->in(0)); 3544 // We are looking for the pattern: 3545 // /->ThreadLocal 3546 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3547 // \->ConI(0) 3548 // We want to verify that the If and the LoadB have the same control 3549 // See GraphKit::g1_write_barrier_pre() 3550 if (x->is_If()) { 3551 IfNode *iff = x->as_If(); 3552 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3553 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3554 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3555 && cmp->in(1)->is_Load()) { 3556 LoadNode* load = cmp->in(1)->as_Load(); 3557 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3558 && load->in(2)->in(3)->is_Con() 3559 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3560 3561 Node* if_ctrl = iff->in(0); 3562 Node* load_ctrl = load->in(0); 3563 3564 if (if_ctrl != load_ctrl) { 3565 // Skip possible CProj->NeverBranch in infinite loops 3566 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3567 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3568 if_ctrl = if_ctrl->in(0)->in(0); 3569 } 3570 } 3571 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3572 } 3573 } 3574 } 3575 } 3576 } 3577 } 3578 } 3579 } 3580 3581 #endif 3582 3583 // The Compile object keeps track of failure reasons separately from the ciEnv. 3584 // This is required because there is not quite a 1-1 relation between the 3585 // ciEnv and its compilation task and the Compile object. Note that one 3586 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3587 // to backtrack and retry without subsuming loads. Other than this backtracking 3588 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3589 // by the logic in C2Compiler. 3590 void Compile::record_failure(const char* reason) { 3591 if (log() != NULL) { 3592 log()->elem("failure reason='%s' phase='compile'", reason); 3593 } 3594 if (_failure_reason == NULL) { 3595 // Record the first failure reason. 3596 _failure_reason = reason; 3597 } 3598 3599 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3600 C->print_method(PHASE_FAILURE); 3601 } 3602 _root = NULL; // flush the graph, too 3603 } 3604 3605 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3606 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3607 _phase_name(name), _dolog(dolog) 3608 { 3609 if (dolog) { 3610 C = Compile::current(); 3611 _log = C->log(); 3612 } else { 3613 C = NULL; 3614 _log = NULL; 3615 } 3616 if (_log != NULL) { 3617 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3618 _log->stamp(); 3619 _log->end_head(); 3620 } 3621 } 3622 3623 Compile::TracePhase::~TracePhase() { 3624 3625 C = Compile::current(); 3626 if (_dolog) { 3627 _log = C->log(); 3628 } else { 3629 _log = NULL; 3630 } 3631 3632 #ifdef ASSERT 3633 if (PrintIdealNodeCount) { 3634 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3635 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3636 } 3637 3638 if (VerifyIdealNodeCount) { 3639 Compile::current()->print_missing_nodes(); 3640 } 3641 #endif 3642 3643 if (_log != NULL) { 3644 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3645 } 3646 } 3647 3648 //============================================================================= 3649 // Two Constant's are equal when the type and the value are equal. 3650 bool Compile::Constant::operator==(const Constant& other) { 3651 if (type() != other.type() ) return false; 3652 if (can_be_reused() != other.can_be_reused()) return false; 3653 // For floating point values we compare the bit pattern. 3654 switch (type()) { 3655 case T_FLOAT: return (_v._value.i == other._v._value.i); 3656 case T_LONG: 3657 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3658 case T_OBJECT: 3659 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3660 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3661 case T_METADATA: return (_v._metadata == other._v._metadata); 3662 default: ShouldNotReachHere(); 3663 } 3664 return false; 3665 } 3666 3667 static int type_to_size_in_bytes(BasicType t) { 3668 switch (t) { 3669 case T_LONG: return sizeof(jlong ); 3670 case T_FLOAT: return sizeof(jfloat ); 3671 case T_DOUBLE: return sizeof(jdouble); 3672 case T_METADATA: return sizeof(Metadata*); 3673 // We use T_VOID as marker for jump-table entries (labels) which 3674 // need an internal word relocation. 3675 case T_VOID: 3676 case T_ADDRESS: 3677 case T_OBJECT: return sizeof(jobject); 3678 } 3679 3680 ShouldNotReachHere(); 3681 return -1; 3682 } 3683 3684 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3685 // sort descending 3686 if (a->freq() > b->freq()) return -1; 3687 if (a->freq() < b->freq()) return 1; 3688 return 0; 3689 } 3690 3691 void Compile::ConstantTable::calculate_offsets_and_size() { 3692 // First, sort the array by frequencies. 3693 _constants.sort(qsort_comparator); 3694 3695 #ifdef ASSERT 3696 // Make sure all jump-table entries were sorted to the end of the 3697 // array (they have a negative frequency). 3698 bool found_void = false; 3699 for (int i = 0; i < _constants.length(); i++) { 3700 Constant con = _constants.at(i); 3701 if (con.type() == T_VOID) 3702 found_void = true; // jump-tables 3703 else 3704 assert(!found_void, "wrong sorting"); 3705 } 3706 #endif 3707 3708 int offset = 0; 3709 for (int i = 0; i < _constants.length(); i++) { 3710 Constant* con = _constants.adr_at(i); 3711 3712 // Align offset for type. 3713 int typesize = type_to_size_in_bytes(con->type()); 3714 offset = align_size_up(offset, typesize); 3715 con->set_offset(offset); // set constant's offset 3716 3717 if (con->type() == T_VOID) { 3718 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3719 offset = offset + typesize * n->outcnt(); // expand jump-table 3720 } else { 3721 offset = offset + typesize; 3722 } 3723 } 3724 3725 // Align size up to the next section start (which is insts; see 3726 // CodeBuffer::align_at_start). 3727 assert(_size == -1, "already set?"); 3728 _size = align_size_up(offset, CodeEntryAlignment); 3729 } 3730 3731 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3732 MacroAssembler _masm(&cb); 3733 for (int i = 0; i < _constants.length(); i++) { 3734 Constant con = _constants.at(i); 3735 address constant_addr = NULL; 3736 switch (con.type()) { 3737 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3738 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3739 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3740 case T_OBJECT: { 3741 jobject obj = con.get_jobject(); 3742 int oop_index = _masm.oop_recorder()->find_index(obj); 3743 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3744 break; 3745 } 3746 case T_ADDRESS: { 3747 address addr = (address) con.get_jobject(); 3748 constant_addr = _masm.address_constant(addr); 3749 break; 3750 } 3751 // We use T_VOID as marker for jump-table entries (labels) which 3752 // need an internal word relocation. 3753 case T_VOID: { 3754 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3755 // Fill the jump-table with a dummy word. The real value is 3756 // filled in later in fill_jump_table. 3757 address dummy = (address) n; 3758 constant_addr = _masm.address_constant(dummy); 3759 // Expand jump-table 3760 for (uint i = 1; i < n->outcnt(); i++) { 3761 address temp_addr = _masm.address_constant(dummy + i); 3762 assert(temp_addr, "consts section too small"); 3763 } 3764 break; 3765 } 3766 case T_METADATA: { 3767 Metadata* obj = con.get_metadata(); 3768 int metadata_index = _masm.oop_recorder()->find_index(obj); 3769 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3770 break; 3771 } 3772 default: ShouldNotReachHere(); 3773 } 3774 assert(constant_addr, "consts section too small"); 3775 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3776 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3777 } 3778 } 3779 3780 int Compile::ConstantTable::find_offset(Constant& con) const { 3781 int idx = _constants.find(con); 3782 assert(idx != -1, "constant must be in constant table"); 3783 int offset = _constants.at(idx).offset(); 3784 assert(offset != -1, "constant table not emitted yet?"); 3785 return offset; 3786 } 3787 3788 void Compile::ConstantTable::add(Constant& con) { 3789 if (con.can_be_reused()) { 3790 int idx = _constants.find(con); 3791 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3792 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3793 return; 3794 } 3795 } 3796 (void) _constants.append(con); 3797 } 3798 3799 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3800 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3801 Constant con(type, value, b->_freq); 3802 add(con); 3803 return con; 3804 } 3805 3806 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3807 Constant con(metadata); 3808 add(con); 3809 return con; 3810 } 3811 3812 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3813 jvalue value; 3814 BasicType type = oper->type()->basic_type(); 3815 switch (type) { 3816 case T_LONG: value.j = oper->constantL(); break; 3817 case T_FLOAT: value.f = oper->constantF(); break; 3818 case T_DOUBLE: value.d = oper->constantD(); break; 3819 case T_OBJECT: 3820 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3821 case T_METADATA: return add((Metadata*)oper->constant()); break; 3822 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3823 } 3824 return add(n, type, value); 3825 } 3826 3827 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3828 jvalue value; 3829 // We can use the node pointer here to identify the right jump-table 3830 // as this method is called from Compile::Fill_buffer right before 3831 // the MachNodes are emitted and the jump-table is filled (means the 3832 // MachNode pointers do not change anymore). 3833 value.l = (jobject) n; 3834 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3835 add(con); 3836 return con; 3837 } 3838 3839 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3840 // If called from Compile::scratch_emit_size do nothing. 3841 if (Compile::current()->in_scratch_emit_size()) return; 3842 3843 assert(labels.is_nonempty(), "must be"); 3844 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3845 3846 // Since MachConstantNode::constant_offset() also contains 3847 // table_base_offset() we need to subtract the table_base_offset() 3848 // to get the plain offset into the constant table. 3849 int offset = n->constant_offset() - table_base_offset(); 3850 3851 MacroAssembler _masm(&cb); 3852 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3853 3854 for (uint i = 0; i < n->outcnt(); i++) { 3855 address* constant_addr = &jump_table_base[i]; 3856 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i))); 3857 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3858 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3859 } 3860 } 3861 3862 void Compile::dump_inlining() { 3863 if (print_inlining() || print_intrinsics()) { 3864 // Print inlining message for candidates that we couldn't inline 3865 // for lack of space or non constant receiver 3866 for (int i = 0; i < _late_inlines.length(); i++) { 3867 CallGenerator* cg = _late_inlines.at(i); 3868 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3869 } 3870 Unique_Node_List useful; 3871 useful.push(root()); 3872 for (uint next = 0; next < useful.size(); ++next) { 3873 Node* n = useful.at(next); 3874 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3875 CallNode* call = n->as_Call(); 3876 CallGenerator* cg = call->generator(); 3877 cg->print_inlining_late("receiver not constant"); 3878 } 3879 uint max = n->len(); 3880 for ( uint i = 0; i < max; ++i ) { 3881 Node *m = n->in(i); 3882 if ( m == NULL ) continue; 3883 useful.push(m); 3884 } 3885 } 3886 for (int i = 0; i < _print_inlining_list->length(); i++) { 3887 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3888 } 3889 } 3890 } 3891 3892 // Dump inlining replay data to the stream. 3893 // Don't change thread state and acquire any locks. 3894 void Compile::dump_inline_data(outputStream* out) { 3895 InlineTree* inl_tree = ilt(); 3896 if (inl_tree != NULL) { 3897 out->print(" inline %d", inl_tree->count()); 3898 inl_tree->dump_replay_data(out); 3899 } 3900 } 3901 3902 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3903 if (n1->Opcode() < n2->Opcode()) return -1; 3904 else if (n1->Opcode() > n2->Opcode()) return 1; 3905 3906 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req())); 3907 for (uint i = 1; i < n1->req(); i++) { 3908 if (n1->in(i) < n2->in(i)) return -1; 3909 else if (n1->in(i) > n2->in(i)) return 1; 3910 } 3911 3912 return 0; 3913 } 3914 3915 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3916 Node* n1 = *n1p; 3917 Node* n2 = *n2p; 3918 3919 return cmp_expensive_nodes(n1, n2); 3920 } 3921 3922 void Compile::sort_expensive_nodes() { 3923 if (!expensive_nodes_sorted()) { 3924 _expensive_nodes->sort(cmp_expensive_nodes); 3925 } 3926 } 3927 3928 bool Compile::expensive_nodes_sorted() const { 3929 for (int i = 1; i < _expensive_nodes->length(); i++) { 3930 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3931 return false; 3932 } 3933 } 3934 return true; 3935 } 3936 3937 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3938 if (_expensive_nodes->length() == 0) { 3939 return false; 3940 } 3941 3942 assert(OptimizeExpensiveOps, "optimization off?"); 3943 3944 // Take this opportunity to remove dead nodes from the list 3945 int j = 0; 3946 for (int i = 0; i < _expensive_nodes->length(); i++) { 3947 Node* n = _expensive_nodes->at(i); 3948 if (!n->is_unreachable(igvn)) { 3949 assert(n->is_expensive(), "should be expensive"); 3950 _expensive_nodes->at_put(j, n); 3951 j++; 3952 } 3953 } 3954 _expensive_nodes->trunc_to(j); 3955 3956 // Then sort the list so that similar nodes are next to each other 3957 // and check for at least two nodes of identical kind with same data 3958 // inputs. 3959 sort_expensive_nodes(); 3960 3961 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3962 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3963 return true; 3964 } 3965 } 3966 3967 return false; 3968 } 3969 3970 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3971 if (_expensive_nodes->length() == 0) { 3972 return; 3973 } 3974 3975 assert(OptimizeExpensiveOps, "optimization off?"); 3976 3977 // Sort to bring similar nodes next to each other and clear the 3978 // control input of nodes for which there's only a single copy. 3979 sort_expensive_nodes(); 3980 3981 int j = 0; 3982 int identical = 0; 3983 int i = 0; 3984 for (; i < _expensive_nodes->length()-1; i++) { 3985 assert(j <= i, "can't write beyond current index"); 3986 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3987 identical++; 3988 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3989 continue; 3990 } 3991 if (identical > 0) { 3992 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3993 identical = 0; 3994 } else { 3995 Node* n = _expensive_nodes->at(i); 3996 igvn.hash_delete(n); 3997 n->set_req(0, NULL); 3998 igvn.hash_insert(n); 3999 } 4000 } 4001 if (identical > 0) { 4002 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4003 } else if (_expensive_nodes->length() >= 1) { 4004 Node* n = _expensive_nodes->at(i); 4005 igvn.hash_delete(n); 4006 n->set_req(0, NULL); 4007 igvn.hash_insert(n); 4008 } 4009 _expensive_nodes->trunc_to(j); 4010 } 4011 4012 void Compile::add_expensive_node(Node * n) { 4013 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4014 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4015 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4016 if (OptimizeExpensiveOps) { 4017 _expensive_nodes->append(n); 4018 } else { 4019 // Clear control input and let IGVN optimize expensive nodes if 4020 // OptimizeExpensiveOps is off. 4021 n->set_req(0, NULL); 4022 } 4023 } 4024 4025 /** 4026 * Remove the speculative part of types and clean up the graph 4027 */ 4028 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4029 if (UseTypeSpeculation) { 4030 Unique_Node_List worklist; 4031 worklist.push(root()); 4032 int modified = 0; 4033 // Go over all type nodes that carry a speculative type, drop the 4034 // speculative part of the type and enqueue the node for an igvn 4035 // which may optimize it out. 4036 for (uint next = 0; next < worklist.size(); ++next) { 4037 Node *n = worklist.at(next); 4038 if (n->is_Type()) { 4039 TypeNode* tn = n->as_Type(); 4040 const Type* t = tn->type(); 4041 const Type* t_no_spec = t->remove_speculative(); 4042 if (t_no_spec != t) { 4043 bool in_hash = igvn.hash_delete(n); 4044 assert(in_hash, "node should be in igvn hash table"); 4045 tn->set_type(t_no_spec); 4046 igvn.hash_insert(n); 4047 igvn._worklist.push(n); // give it a chance to go away 4048 modified++; 4049 } 4050 } 4051 uint max = n->len(); 4052 for( uint i = 0; i < max; ++i ) { 4053 Node *m = n->in(i); 4054 if (not_a_node(m)) continue; 4055 worklist.push(m); 4056 } 4057 } 4058 // Drop the speculative part of all types in the igvn's type table 4059 igvn.remove_speculative_types(); 4060 if (modified > 0) { 4061 igvn.optimize(); 4062 } 4063 #ifdef ASSERT 4064 // Verify that after the IGVN is over no speculative type has resurfaced 4065 worklist.clear(); 4066 worklist.push(root()); 4067 for (uint next = 0; next < worklist.size(); ++next) { 4068 Node *n = worklist.at(next); 4069 const Type* t = igvn.type_or_null(n); 4070 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4071 if (n->is_Type()) { 4072 t = n->as_Type()->type(); 4073 assert(t == t->remove_speculative(), "no more speculative types"); 4074 } 4075 uint max = n->len(); 4076 for( uint i = 0; i < max; ++i ) { 4077 Node *m = n->in(i); 4078 if (not_a_node(m)) continue; 4079 worklist.push(m); 4080 } 4081 } 4082 igvn.check_no_speculative_types(); 4083 #endif 4084 } 4085 } 4086 4087 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4088 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4089 if (ctrl != NULL) { 4090 // Express control dependency by a CastII node with a narrow type. 4091 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */); 4092 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4093 // node from floating above the range check during loop optimizations. Otherwise, the 4094 // ConvI2L node may be eliminated independently of the range check, causing the data path 4095 // to become TOP while the control path is still there (although it's unreachable). 4096 value->set_req(0, ctrl); 4097 // Save CastII node to remove it after loop optimizations. 4098 phase->C->add_range_check_cast(value); 4099 value = phase->transform(value); 4100 } 4101 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4102 return phase->transform(new (phase->C) ConvI2LNode(value, ltype)); 4103 } 4104 4105 // Auxiliary method to support randomized stressing/fuzzing. 4106 // 4107 // This method can be called the arbitrary number of times, with current count 4108 // as the argument. The logic allows selecting a single candidate from the 4109 // running list of candidates as follows: 4110 // int count = 0; 4111 // Cand* selected = null; 4112 // while(cand = cand->next()) { 4113 // if (randomized_select(++count)) { 4114 // selected = cand; 4115 // } 4116 // } 4117 // 4118 // Including count equalizes the chances any candidate is "selected". 4119 // This is useful when we don't have the complete list of candidates to choose 4120 // from uniformly. In this case, we need to adjust the randomicity of the 4121 // selection, or else we will end up biasing the selection towards the latter 4122 // candidates. 4123 // 4124 // Quick back-envelope calculation shows that for the list of n candidates 4125 // the equal probability for the candidate to persist as "best" can be 4126 // achieved by replacing it with "next" k-th candidate with the probability 4127 // of 1/k. It can be easily shown that by the end of the run, the 4128 // probability for any candidate is converged to 1/n, thus giving the 4129 // uniform distribution among all the candidates. 4130 // 4131 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4132 #define RANDOMIZED_DOMAIN_POW 29 4133 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4134 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4135 bool Compile::randomized_select(int count) { 4136 assert(count > 0, "only positive"); 4137 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4138 }