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 #include "jfr/jfrEvents.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/divnode.hpp" 46 #include "opto/escape.hpp" 47 #include "opto/idealGraphPrinter.hpp" 48 #include "opto/loopnode.hpp" 49 #include "opto/machnode.hpp" 50 #include "opto/macro.hpp" 51 #include "opto/matcher.hpp" 52 #include "opto/mathexactnode.hpp" 53 #include "opto/memnode.hpp" 54 #include "opto/mulnode.hpp" 55 #include "opto/node.hpp" 56 #include "opto/opcodes.hpp" 57 #include "opto/output.hpp" 58 #include "opto/parse.hpp" 59 #include "opto/phaseX.hpp" 60 #include "opto/rootnode.hpp" 61 #include "opto/runtime.hpp" 62 #include "opto/stringopts.hpp" 63 #include "opto/type.hpp" 64 #include "opto/vectornode.hpp" 65 #include "runtime/arguments.hpp" 66 #include "runtime/signature.hpp" 67 #include "runtime/stubRoutines.hpp" 68 #include "runtime/timer.hpp" 69 #include "utilities/copy.hpp" 70 #if defined AD_MD_HPP 71 # include AD_MD_HPP 72 #elif defined TARGET_ARCH_MODEL_x86_32 73 # include "adfiles/ad_x86_32.hpp" 74 #elif defined TARGET_ARCH_MODEL_x86_64 75 # include "adfiles/ad_x86_64.hpp" 76 #elif defined TARGET_ARCH_MODEL_sparc 77 # include "adfiles/ad_sparc.hpp" 78 #elif defined TARGET_ARCH_MODEL_zero 79 # include "adfiles/ad_zero.hpp" 80 #elif defined TARGET_ARCH_MODEL_ppc_64 81 # include "adfiles/ad_ppc_64.hpp" 82 #endif 83 84 // -------------------- Compile::mach_constant_base_node ----------------------- 85 // Constant table base node singleton. 86 MachConstantBaseNode* Compile::mach_constant_base_node() { 87 if (_mach_constant_base_node == NULL) { 88 _mach_constant_base_node = new (C) MachConstantBaseNode(); 89 _mach_constant_base_node->add_req(C->root()); 90 } 91 return _mach_constant_base_node; 92 } 93 94 95 /// Support for intrinsics. 96 97 // Return the index at which m must be inserted (or already exists). 98 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 99 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 100 #ifdef ASSERT 101 for (int i = 1; i < _intrinsics->length(); i++) { 102 CallGenerator* cg1 = _intrinsics->at(i-1); 103 CallGenerator* cg2 = _intrinsics->at(i); 104 assert(cg1->method() != cg2->method() 105 ? cg1->method() < cg2->method() 106 : cg1->is_virtual() < cg2->is_virtual(), 107 "compiler intrinsics list must stay sorted"); 108 } 109 #endif 110 // Binary search sorted list, in decreasing intervals [lo, hi]. 111 int lo = 0, hi = _intrinsics->length()-1; 112 while (lo <= hi) { 113 int mid = (uint)(hi + lo) / 2; 114 ciMethod* mid_m = _intrinsics->at(mid)->method(); 115 if (m < mid_m) { 116 hi = mid-1; 117 } else if (m > mid_m) { 118 lo = mid+1; 119 } else { 120 // look at minor sort key 121 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 122 if (is_virtual < mid_virt) { 123 hi = mid-1; 124 } else if (is_virtual > mid_virt) { 125 lo = mid+1; 126 } else { 127 return mid; // exact match 128 } 129 } 130 } 131 return lo; // inexact match 132 } 133 134 void Compile::register_intrinsic(CallGenerator* cg) { 135 if (_intrinsics == NULL) { 136 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 137 } 138 // This code is stolen from ciObjectFactory::insert. 139 // Really, GrowableArray should have methods for 140 // insert_at, remove_at, and binary_search. 141 int len = _intrinsics->length(); 142 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 143 if (index == len) { 144 _intrinsics->append(cg); 145 } else { 146 #ifdef ASSERT 147 CallGenerator* oldcg = _intrinsics->at(index); 148 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 149 #endif 150 _intrinsics->append(_intrinsics->at(len-1)); 151 int pos; 152 for (pos = len-2; pos >= index; pos--) { 153 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 154 } 155 _intrinsics->at_put(index, cg); 156 } 157 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 158 } 159 160 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 161 assert(m->is_loaded(), "don't try this on unloaded methods"); 162 if (_intrinsics != NULL) { 163 int index = intrinsic_insertion_index(m, is_virtual); 164 if (index < _intrinsics->length() 165 && _intrinsics->at(index)->method() == m 166 && _intrinsics->at(index)->is_virtual() == is_virtual) { 167 return _intrinsics->at(index); 168 } 169 } 170 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 171 if (m->intrinsic_id() != vmIntrinsics::_none && 172 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 173 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 174 if (cg != NULL) { 175 // Save it for next time: 176 register_intrinsic(cg); 177 return cg; 178 } else { 179 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 180 } 181 } 182 return NULL; 183 } 184 185 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 186 // in library_call.cpp. 187 188 189 #ifndef PRODUCT 190 // statistics gathering... 191 192 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 193 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 194 195 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 196 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 197 int oflags = _intrinsic_hist_flags[id]; 198 assert(flags != 0, "what happened?"); 199 if (is_virtual) { 200 flags |= _intrinsic_virtual; 201 } 202 bool changed = (flags != oflags); 203 if ((flags & _intrinsic_worked) != 0) { 204 juint count = (_intrinsic_hist_count[id] += 1); 205 if (count == 1) { 206 changed = true; // first time 207 } 208 // increment the overall count also: 209 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 210 } 211 if (changed) { 212 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 213 // Something changed about the intrinsic's virtuality. 214 if ((flags & _intrinsic_virtual) != 0) { 215 // This is the first use of this intrinsic as a virtual call. 216 if (oflags != 0) { 217 // We already saw it as a non-virtual, so note both cases. 218 flags |= _intrinsic_both; 219 } 220 } else if ((oflags & _intrinsic_both) == 0) { 221 // This is the first use of this intrinsic as a non-virtual 222 flags |= _intrinsic_both; 223 } 224 } 225 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 226 } 227 // update the overall flags also: 228 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 229 return changed; 230 } 231 232 static char* format_flags(int flags, char* buf) { 233 buf[0] = 0; 234 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 235 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 236 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 237 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 238 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 239 if (buf[0] == 0) strcat(buf, ","); 240 assert(buf[0] == ',', "must be"); 241 return &buf[1]; 242 } 243 244 void Compile::print_intrinsic_statistics() { 245 char flagsbuf[100]; 246 ttyLocker ttyl; 247 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 248 tty->print_cr("Compiler intrinsic usage:"); 249 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 250 if (total == 0) total = 1; // avoid div0 in case of no successes 251 #define PRINT_STAT_LINE(name, c, f) \ 252 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 253 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 254 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 255 int flags = _intrinsic_hist_flags[id]; 256 juint count = _intrinsic_hist_count[id]; 257 if ((flags | count) != 0) { 258 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 259 } 260 } 261 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 262 if (xtty != NULL) xtty->tail("statistics"); 263 } 264 265 void Compile::print_statistics() { 266 { ttyLocker ttyl; 267 if (xtty != NULL) xtty->head("statistics type='opto'"); 268 Parse::print_statistics(); 269 PhaseCCP::print_statistics(); 270 PhaseRegAlloc::print_statistics(); 271 Scheduling::print_statistics(); 272 PhasePeephole::print_statistics(); 273 PhaseIdealLoop::print_statistics(); 274 if (xtty != NULL) xtty->tail("statistics"); 275 } 276 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 277 // put this under its own <statistics> element. 278 print_intrinsic_statistics(); 279 } 280 } 281 #endif //PRODUCT 282 283 // Support for bundling info 284 Bundle* Compile::node_bundling(const Node *n) { 285 assert(valid_bundle_info(n), "oob"); 286 return &_node_bundling_base[n->_idx]; 287 } 288 289 bool Compile::valid_bundle_info(const Node *n) { 290 return (_node_bundling_limit > n->_idx); 291 } 292 293 294 void Compile::gvn_replace_by(Node* n, Node* nn) { 295 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 296 Node* use = n->last_out(i); 297 bool is_in_table = initial_gvn()->hash_delete(use); 298 uint uses_found = 0; 299 for (uint j = 0; j < use->len(); j++) { 300 if (use->in(j) == n) { 301 if (j < use->req()) 302 use->set_req(j, nn); 303 else 304 use->set_prec(j, nn); 305 uses_found++; 306 } 307 } 308 if (is_in_table) { 309 // reinsert into table 310 initial_gvn()->hash_find_insert(use); 311 } 312 record_for_igvn(use); 313 i -= uses_found; // we deleted 1 or more copies of this edge 314 } 315 } 316 317 318 static inline bool not_a_node(const Node* n) { 319 if (n == NULL) return true; 320 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 321 if (*(address*)n == badAddress) return true; // kill by Node::destruct 322 return false; 323 } 324 325 // Identify all nodes that are reachable from below, useful. 326 // Use breadth-first pass that records state in a Unique_Node_List, 327 // recursive traversal is slower. 328 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 329 int estimated_worklist_size = live_nodes(); 330 useful.map( estimated_worklist_size, NULL ); // preallocate space 331 332 // Initialize worklist 333 if (root() != NULL) { useful.push(root()); } 334 // If 'top' is cached, declare it useful to preserve cached node 335 if( cached_top_node() ) { useful.push(cached_top_node()); } 336 337 // Push all useful nodes onto the list, breadthfirst 338 for( uint next = 0; next < useful.size(); ++next ) { 339 assert( next < unique(), "Unique useful nodes < total nodes"); 340 Node *n = useful.at(next); 341 uint max = n->len(); 342 for( uint i = 0; i < max; ++i ) { 343 Node *m = n->in(i); 344 if (not_a_node(m)) continue; 345 useful.push(m); 346 } 347 } 348 } 349 350 // Update dead_node_list with any missing dead nodes using useful 351 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 352 void Compile::update_dead_node_list(Unique_Node_List &useful) { 353 uint max_idx = unique(); 354 VectorSet& useful_node_set = useful.member_set(); 355 356 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 357 // If node with index node_idx is not in useful set, 358 // mark it as dead in dead node list. 359 if (! useful_node_set.test(node_idx) ) { 360 record_dead_node(node_idx); 361 } 362 } 363 } 364 365 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 366 int shift = 0; 367 for (int i = 0; i < inlines->length(); i++) { 368 CallGenerator* cg = inlines->at(i); 369 CallNode* call = cg->call_node(); 370 if (shift > 0) { 371 inlines->at_put(i-shift, cg); 372 } 373 if (!useful.member(call)) { 374 shift++; 375 } 376 } 377 inlines->trunc_to(inlines->length()-shift); 378 } 379 380 // Disconnect all useless nodes by disconnecting those at the boundary. 381 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 382 uint next = 0; 383 while (next < useful.size()) { 384 Node *n = useful.at(next++); 385 if (n->is_SafePoint()) { 386 // We're done with a parsing phase. Replaced nodes are not valid 387 // beyond that point. 388 n->as_SafePoint()->delete_replaced_nodes(); 389 } 390 // Use raw traversal of out edges since this code removes out edges 391 int max = n->outcnt(); 392 for (int j = 0; j < max; ++j) { 393 Node* child = n->raw_out(j); 394 if (! useful.member(child)) { 395 assert(!child->is_top() || child != top(), 396 "If top is cached in Compile object it is in useful list"); 397 // Only need to remove this out-edge to the useless node 398 n->raw_del_out(j); 399 --j; 400 --max; 401 } 402 } 403 if (n->outcnt() == 1 && n->has_special_unique_user()) { 404 record_for_igvn(n->unique_out()); 405 } 406 } 407 // Remove useless macro and predicate opaq nodes 408 for (int i = C->macro_count()-1; i >= 0; i--) { 409 Node* n = C->macro_node(i); 410 if (!useful.member(n)) { 411 remove_macro_node(n); 412 } 413 } 414 // Remove useless CastII nodes with range check dependency 415 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 416 Node* cast = range_check_cast_node(i); 417 if (!useful.member(cast)) { 418 remove_range_check_cast(cast); 419 } 420 } 421 // Remove useless expensive node 422 for (int i = C->expensive_count()-1; i >= 0; i--) { 423 Node* n = C->expensive_node(i); 424 if (!useful.member(n)) { 425 remove_expensive_node(n); 426 } 427 } 428 // clean up the late inline lists 429 remove_useless_late_inlines(&_string_late_inlines, useful); 430 remove_useless_late_inlines(&_boxing_late_inlines, useful); 431 remove_useless_late_inlines(&_late_inlines, useful); 432 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 433 } 434 435 //------------------------------frame_size_in_words----------------------------- 436 // frame_slots in units of words 437 int Compile::frame_size_in_words() const { 438 // shift is 0 in LP32 and 1 in LP64 439 const int shift = (LogBytesPerWord - LogBytesPerInt); 440 int words = _frame_slots >> shift; 441 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 442 return words; 443 } 444 445 // To bang the stack of this compiled method we use the stack size 446 // that the interpreter would need in case of a deoptimization. This 447 // removes the need to bang the stack in the deoptimization blob which 448 // in turn simplifies stack overflow handling. 449 int Compile::bang_size_in_bytes() const { 450 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 451 } 452 453 // ============================================================================ 454 //------------------------------CompileWrapper--------------------------------- 455 class CompileWrapper : public StackObj { 456 Compile *const _compile; 457 public: 458 CompileWrapper(Compile* compile); 459 460 ~CompileWrapper(); 461 }; 462 463 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 464 // the Compile* pointer is stored in the current ciEnv: 465 ciEnv* env = compile->env(); 466 assert(env == ciEnv::current(), "must already be a ciEnv active"); 467 assert(env->compiler_data() == NULL, "compile already active?"); 468 env->set_compiler_data(compile); 469 assert(compile == Compile::current(), "sanity"); 470 471 compile->set_type_dict(NULL); 472 compile->set_type_hwm(NULL); 473 compile->set_type_last_size(0); 474 compile->set_last_tf(NULL, NULL); 475 compile->set_indexSet_arena(NULL); 476 compile->set_indexSet_free_block_list(NULL); 477 compile->init_type_arena(); 478 Type::Initialize(compile); 479 _compile->set_scratch_buffer_blob(NULL); 480 _compile->begin_method(); 481 } 482 CompileWrapper::~CompileWrapper() { 483 _compile->end_method(); 484 if (_compile->scratch_buffer_blob() != NULL) 485 BufferBlob::free(_compile->scratch_buffer_blob()); 486 _compile->env()->set_compiler_data(NULL); 487 } 488 489 490 //----------------------------print_compile_messages--------------------------- 491 void Compile::print_compile_messages() { 492 #ifndef PRODUCT 493 // Check if recompiling 494 if (_subsume_loads == false && PrintOpto) { 495 // Recompiling without allowing machine instructions to subsume loads 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 501 // Recompiling without escape analysis 502 tty->print_cr("*********************************************************"); 503 tty->print_cr("** Bailout: Recompile without escape analysis **"); 504 tty->print_cr("*********************************************************"); 505 } 506 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 507 // Recompiling without boxing elimination 508 tty->print_cr("*********************************************************"); 509 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 510 tty->print_cr("*********************************************************"); 511 } 512 if (env()->break_at_compile()) { 513 // Open the debugger when compiling this method. 514 tty->print("### Breaking when compiling: "); 515 method()->print_short_name(); 516 tty->cr(); 517 BREAKPOINT; 518 } 519 520 if( PrintOpto ) { 521 if (is_osr_compilation()) { 522 tty->print("[OSR]%3d", _compile_id); 523 } else { 524 tty->print("%3d", _compile_id); 525 } 526 } 527 #endif 528 } 529 530 531 //-----------------------init_scratch_buffer_blob------------------------------ 532 // Construct a temporary BufferBlob and cache it for this compile. 533 void Compile::init_scratch_buffer_blob(int const_size) { 534 // If there is already a scratch buffer blob allocated and the 535 // constant section is big enough, use it. Otherwise free the 536 // current and allocate a new one. 537 BufferBlob* blob = scratch_buffer_blob(); 538 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 539 // Use the current blob. 540 } else { 541 if (blob != NULL) { 542 BufferBlob::free(blob); 543 } 544 545 ResourceMark rm; 546 _scratch_const_size = const_size; 547 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 548 blob = BufferBlob::create("Compile::scratch_buffer", size); 549 // Record the buffer blob for next time. 550 set_scratch_buffer_blob(blob); 551 // Have we run out of code space? 552 if (scratch_buffer_blob() == NULL) { 553 // Let CompilerBroker disable further compilations. 554 record_failure("Not enough space for scratch buffer in CodeCache"); 555 return; 556 } 557 } 558 559 // Initialize the relocation buffers 560 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 561 set_scratch_locs_memory(locs_buf); 562 } 563 564 565 //-----------------------scratch_emit_size------------------------------------- 566 // Helper function that computes size by emitting code 567 uint Compile::scratch_emit_size(const Node* n) { 568 // Start scratch_emit_size section. 569 set_in_scratch_emit_size(true); 570 571 // Emit into a trash buffer and count bytes emitted. 572 // This is a pretty expensive way to compute a size, 573 // but it works well enough if seldom used. 574 // All common fixed-size instructions are given a size 575 // method by the AD file. 576 // Note that the scratch buffer blob and locs memory are 577 // allocated at the beginning of the compile task, and 578 // may be shared by several calls to scratch_emit_size. 579 // The allocation of the scratch buffer blob is particularly 580 // expensive, since it has to grab the code cache lock. 581 BufferBlob* blob = this->scratch_buffer_blob(); 582 assert(blob != NULL, "Initialize BufferBlob at start"); 583 assert(blob->size() > MAX_inst_size, "sanity"); 584 relocInfo* locs_buf = scratch_locs_memory(); 585 address blob_begin = blob->content_begin(); 586 address blob_end = (address)locs_buf; 587 assert(blob->content_contains(blob_end), "sanity"); 588 CodeBuffer buf(blob_begin, blob_end - blob_begin); 589 buf.initialize_consts_size(_scratch_const_size); 590 buf.initialize_stubs_size(MAX_stubs_size); 591 assert(locs_buf != NULL, "sanity"); 592 int lsize = MAX_locs_size / 3; 593 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 594 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 595 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 596 597 // Do the emission. 598 599 Label fakeL; // Fake label for branch instructions. 600 Label* saveL = NULL; 601 uint save_bnum = 0; 602 bool is_branch = n->is_MachBranch(); 603 if (is_branch) { 604 MacroAssembler masm(&buf); 605 masm.bind(fakeL); 606 n->as_MachBranch()->save_label(&saveL, &save_bnum); 607 n->as_MachBranch()->label_set(&fakeL, 0); 608 } 609 n->emit(buf, this->regalloc()); 610 611 // Emitting into the scratch buffer should not fail 612 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 613 614 if (is_branch) // Restore label. 615 n->as_MachBranch()->label_set(saveL, save_bnum); 616 617 // End scratch_emit_size section. 618 set_in_scratch_emit_size(false); 619 620 return buf.insts_size(); 621 } 622 623 624 // ============================================================================ 625 //------------------------------Compile standard------------------------------- 626 debug_only( int Compile::_debug_idx = 100000; ) 627 628 // Compile a method. entry_bci is -1 for normal compilations and indicates 629 // the continuation bci for on stack replacement. 630 631 632 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 633 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 634 : Phase(Compiler), 635 _env(ci_env), 636 _log(ci_env->log()), 637 _compile_id(ci_env->compile_id()), 638 _save_argument_registers(false), 639 _stub_name(NULL), 640 _stub_function(NULL), 641 _stub_entry_point(NULL), 642 _method(target), 643 _entry_bci(osr_bci), 644 _initial_gvn(NULL), 645 _for_igvn(NULL), 646 _warm_calls(NULL), 647 _subsume_loads(subsume_loads), 648 _do_escape_analysis(do_escape_analysis), 649 _eliminate_boxing(eliminate_boxing), 650 _failure_reason(NULL), 651 _code_buffer("Compile::Fill_buffer"), 652 _orig_pc_slot(0), 653 _orig_pc_slot_offset_in_bytes(0), 654 _has_method_handle_invokes(false), 655 _mach_constant_base_node(NULL), 656 _node_bundling_limit(0), 657 _node_bundling_base(NULL), 658 _java_calls(0), 659 _inner_loops(0), 660 _scratch_const_size(-1), 661 _in_scratch_emit_size(false), 662 _dead_node_list(comp_arena()), 663 _dead_node_count(0), 664 #ifndef PRODUCT 665 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 666 _in_dump_cnt(0), 667 _printer(IdealGraphPrinter::printer()), 668 #endif 669 _congraph(NULL), 670 _comp_arena(mtCompiler), 671 _node_arena(mtCompiler), 672 _old_arena(mtCompiler), 673 _Compile_types(mtCompiler), 674 _replay_inline_data(NULL), 675 _late_inlines(comp_arena(), 2, 0, NULL), 676 _string_late_inlines(comp_arena(), 2, 0, NULL), 677 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 678 _late_inlines_pos(0), 679 _number_of_mh_late_inlines(0), 680 _inlining_progress(false), 681 _inlining_incrementally(false), 682 _print_inlining_list(NULL), 683 _print_inlining_idx(0), 684 _interpreter_frame_size(0), 685 _max_node_limit(MaxNodeLimit) { 686 C = this; 687 688 CompileWrapper cw(this); 689 #ifndef PRODUCT 690 if (TimeCompiler2) { 691 tty->print(" "); 692 target->holder()->name()->print(); 693 tty->print("."); 694 target->print_short_name(); 695 tty->print(" "); 696 } 697 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 698 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 699 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 700 if (!print_opto_assembly) { 701 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 702 if (print_assembly && !Disassembler::can_decode()) { 703 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 704 print_opto_assembly = true; 705 } 706 } 707 set_print_assembly(print_opto_assembly); 708 set_parsed_irreducible_loop(false); 709 710 if (method()->has_option("ReplayInline")) { 711 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 712 } 713 #endif 714 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 715 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 716 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 717 718 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 719 // Make sure the method being compiled gets its own MDO, 720 // so we can at least track the decompile_count(). 721 // Need MDO to record RTM code generation state. 722 method()->ensure_method_data(); 723 } 724 725 Init(::AliasLevel); 726 727 728 print_compile_messages(); 729 730 _ilt = InlineTree::build_inline_tree_root(); 731 732 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 733 assert(num_alias_types() >= AliasIdxRaw, ""); 734 735 #define MINIMUM_NODE_HASH 1023 736 // Node list that Iterative GVN will start with 737 Unique_Node_List for_igvn(comp_arena()); 738 set_for_igvn(&for_igvn); 739 740 // GVN that will be run immediately on new nodes 741 uint estimated_size = method()->code_size()*4+64; 742 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 743 PhaseGVN gvn(node_arena(), estimated_size); 744 set_initial_gvn(&gvn); 745 746 if (print_inlining() || print_intrinsics()) { 747 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 748 } 749 { // Scope for timing the parser 750 TracePhase t3("parse", &_t_parser, true); 751 752 // Put top into the hash table ASAP. 753 initial_gvn()->transform_no_reclaim(top()); 754 755 // Set up tf(), start(), and find a CallGenerator. 756 CallGenerator* cg = NULL; 757 if (is_osr_compilation()) { 758 const TypeTuple *domain = StartOSRNode::osr_domain(); 759 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 760 init_tf(TypeFunc::make(domain, range)); 761 StartNode* s = new (this) StartOSRNode(root(), domain); 762 initial_gvn()->set_type_bottom(s); 763 init_start(s); 764 cg = CallGenerator::for_osr(method(), entry_bci()); 765 } else { 766 // Normal case. 767 init_tf(TypeFunc::make(method())); 768 StartNode* s = new (this) StartNode(root(), tf()->domain()); 769 initial_gvn()->set_type_bottom(s); 770 init_start(s); 771 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 772 // With java.lang.ref.reference.get() we must go through the 773 // intrinsic when G1 is enabled - even when get() is the root 774 // method of the compile - so that, if necessary, the value in 775 // the referent field of the reference object gets recorded by 776 // the pre-barrier code. 777 // Specifically, if G1 is enabled, the value in the referent 778 // field is recorded by the G1 SATB pre barrier. This will 779 // result in the referent being marked live and the reference 780 // object removed from the list of discovered references during 781 // reference processing. 782 cg = find_intrinsic(method(), false); 783 } 784 if (cg == NULL) { 785 float past_uses = method()->interpreter_invocation_count(); 786 float expected_uses = past_uses; 787 cg = CallGenerator::for_inline(method(), expected_uses); 788 } 789 } 790 if (failing()) return; 791 if (cg == NULL) { 792 record_method_not_compilable_all_tiers("cannot parse method"); 793 return; 794 } 795 JVMState* jvms = build_start_state(start(), tf()); 796 if ((jvms = cg->generate(jvms)) == NULL) { 797 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 798 record_method_not_compilable("method parse failed"); 799 } 800 return; 801 } 802 GraphKit kit(jvms); 803 804 if (!kit.stopped()) { 805 // Accept return values, and transfer control we know not where. 806 // This is done by a special, unique ReturnNode bound to root. 807 return_values(kit.jvms()); 808 } 809 810 if (kit.has_exceptions()) { 811 // Any exceptions that escape from this call must be rethrown 812 // to whatever caller is dynamically above us on the stack. 813 // This is done by a special, unique RethrowNode bound to root. 814 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 815 } 816 817 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 818 819 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 820 inline_string_calls(true); 821 } 822 823 if (failing()) return; 824 825 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 826 827 // Remove clutter produced by parsing. 828 if (!failing()) { 829 ResourceMark rm; 830 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 831 } 832 } 833 834 // Note: Large methods are capped off in do_one_bytecode(). 835 if (failing()) return; 836 837 // After parsing, node notes are no longer automagic. 838 // They must be propagated by register_new_node_with_optimizer(), 839 // clone(), or the like. 840 set_default_node_notes(NULL); 841 842 for (;;) { 843 int successes = Inline_Warm(); 844 if (failing()) return; 845 if (successes == 0) break; 846 } 847 848 // Drain the list. 849 Finish_Warm(); 850 #ifndef PRODUCT 851 if (_printer) { 852 _printer->print_inlining(this); 853 } 854 #endif 855 856 if (failing()) return; 857 NOT_PRODUCT( verify_graph_edges(); ) 858 859 // Now optimize 860 Optimize(); 861 if (failing()) return; 862 NOT_PRODUCT( verify_graph_edges(); ) 863 864 #ifndef PRODUCT 865 if (PrintIdeal) { 866 ttyLocker ttyl; // keep the following output all in one block 867 // This output goes directly to the tty, not the compiler log. 868 // To enable tools to match it up with the compilation activity, 869 // be sure to tag this tty output with the compile ID. 870 if (xtty != NULL) { 871 xtty->head("ideal compile_id='%d'%s", compile_id(), 872 is_osr_compilation() ? " compile_kind='osr'" : 873 ""); 874 } 875 root()->dump(9999); 876 if (xtty != NULL) { 877 xtty->tail("ideal"); 878 } 879 } 880 #endif 881 882 NOT_PRODUCT( verify_barriers(); ) 883 884 // Dump compilation data to replay it. 885 if (method()->has_option("DumpReplay")) { 886 env()->dump_replay_data(_compile_id); 887 } 888 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 889 env()->dump_inline_data(_compile_id); 890 } 891 892 // Now that we know the size of all the monitors we can add a fixed slot 893 // for the original deopt pc. 894 895 _orig_pc_slot = fixed_slots(); 896 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 897 set_fixed_slots(next_slot); 898 899 // Compute when to use implicit null checks. Used by matching trap based 900 // nodes and NullCheck optimization. 901 set_allowed_deopt_reasons(); 902 903 // Now generate code 904 Code_Gen(); 905 if (failing()) return; 906 907 // Check if we want to skip execution of all compiled code. 908 { 909 #ifndef PRODUCT 910 if (OptoNoExecute) { 911 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 912 return; 913 } 914 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 915 #endif 916 917 if (is_osr_compilation()) { 918 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 919 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 920 } else { 921 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 922 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 923 } 924 925 env()->register_method(_method, _entry_bci, 926 &_code_offsets, 927 _orig_pc_slot_offset_in_bytes, 928 code_buffer(), 929 frame_size_in_words(), _oop_map_set, 930 &_handler_table, &_inc_table, 931 compiler, 932 env()->comp_level(), 933 has_unsafe_access(), 934 SharedRuntime::is_wide_vector(max_vector_size()), 935 rtm_state() 936 ); 937 938 if (log() != NULL) // Print code cache state into compiler log 939 log()->code_cache_state(); 940 } 941 } 942 943 //------------------------------Compile---------------------------------------- 944 // Compile a runtime stub 945 Compile::Compile( ciEnv* ci_env, 946 TypeFunc_generator generator, 947 address stub_function, 948 const char *stub_name, 949 int is_fancy_jump, 950 bool pass_tls, 951 bool save_arg_registers, 952 bool return_pc ) 953 : Phase(Compiler), 954 _env(ci_env), 955 _log(ci_env->log()), 956 _compile_id(0), 957 _save_argument_registers(save_arg_registers), 958 _method(NULL), 959 _stub_name(stub_name), 960 _stub_function(stub_function), 961 _stub_entry_point(NULL), 962 _entry_bci(InvocationEntryBci), 963 _initial_gvn(NULL), 964 _for_igvn(NULL), 965 _warm_calls(NULL), 966 _orig_pc_slot(0), 967 _orig_pc_slot_offset_in_bytes(0), 968 _subsume_loads(true), 969 _do_escape_analysis(false), 970 _eliminate_boxing(false), 971 _failure_reason(NULL), 972 _code_buffer("Compile::Fill_buffer"), 973 _has_method_handle_invokes(false), 974 _mach_constant_base_node(NULL), 975 _node_bundling_limit(0), 976 _node_bundling_base(NULL), 977 _java_calls(0), 978 _inner_loops(0), 979 #ifndef PRODUCT 980 _trace_opto_output(TraceOptoOutput), 981 _in_dump_cnt(0), 982 _printer(NULL), 983 #endif 984 _comp_arena(mtCompiler), 985 _node_arena(mtCompiler), 986 _old_arena(mtCompiler), 987 _Compile_types(mtCompiler), 988 _dead_node_list(comp_arena()), 989 _dead_node_count(0), 990 _congraph(NULL), 991 _replay_inline_data(NULL), 992 _number_of_mh_late_inlines(0), 993 _inlining_progress(false), 994 _inlining_incrementally(false), 995 _print_inlining_list(NULL), 996 _print_inlining_idx(0), 997 _allowed_reasons(0), 998 _interpreter_frame_size(0), 999 _max_node_limit(MaxNodeLimit) { 1000 C = this; 1001 1002 #ifndef PRODUCT 1003 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 1004 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 1005 set_print_assembly(PrintFrameConverterAssembly); 1006 set_parsed_irreducible_loop(false); 1007 #endif 1008 set_has_irreducible_loop(false); // no loops 1009 1010 CompileWrapper cw(this); 1011 Init(/*AliasLevel=*/ 0); 1012 init_tf((*generator)()); 1013 1014 { 1015 // The following is a dummy for the sake of GraphKit::gen_stub 1016 Unique_Node_List for_igvn(comp_arena()); 1017 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1018 PhaseGVN gvn(Thread::current()->resource_area(),255); 1019 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1020 gvn.transform_no_reclaim(top()); 1021 1022 GraphKit kit; 1023 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1024 } 1025 1026 NOT_PRODUCT( verify_graph_edges(); ) 1027 Code_Gen(); 1028 if (failing()) return; 1029 1030 1031 // Entry point will be accessed using compile->stub_entry_point(); 1032 if (code_buffer() == NULL) { 1033 Matcher::soft_match_failure(); 1034 } else { 1035 if (PrintAssembly && (WizardMode || Verbose)) 1036 tty->print_cr("### Stub::%s", stub_name); 1037 1038 if (!failing()) { 1039 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1040 1041 // Make the NMethod 1042 // For now we mark the frame as never safe for profile stackwalking 1043 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1044 code_buffer(), 1045 CodeOffsets::frame_never_safe, 1046 // _code_offsets.value(CodeOffsets::Frame_Complete), 1047 frame_size_in_words(), 1048 _oop_map_set, 1049 save_arg_registers); 1050 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1051 1052 _stub_entry_point = rs->entry_point(); 1053 } 1054 } 1055 } 1056 1057 //------------------------------Init------------------------------------------- 1058 // Prepare for a single compilation 1059 void Compile::Init(int aliaslevel) { 1060 _unique = 0; 1061 _regalloc = NULL; 1062 1063 _tf = NULL; // filled in later 1064 _top = NULL; // cached later 1065 _matcher = NULL; // filled in later 1066 _cfg = NULL; // filled in later 1067 1068 set_24_bit_selection_and_mode(Use24BitFP, false); 1069 1070 _node_note_array = NULL; 1071 _default_node_notes = NULL; 1072 1073 _immutable_memory = NULL; // filled in at first inquiry 1074 1075 // Globally visible Nodes 1076 // First set TOP to NULL to give safe behavior during creation of RootNode 1077 set_cached_top_node(NULL); 1078 set_root(new (this) RootNode()); 1079 // Now that you have a Root to point to, create the real TOP 1080 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1081 set_recent_alloc(NULL, NULL); 1082 1083 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1084 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1085 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1086 env()->set_dependencies(new Dependencies(env())); 1087 1088 _fixed_slots = 0; 1089 set_has_split_ifs(false); 1090 set_has_loops(has_method() && method()->has_loops()); // first approximation 1091 set_has_stringbuilder(false); 1092 set_has_boxed_value(false); 1093 _trap_can_recompile = false; // no traps emitted yet 1094 _major_progress = true; // start out assuming good things will happen 1095 set_has_unsafe_access(false); 1096 set_max_vector_size(0); 1097 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1098 set_decompile_count(0); 1099 1100 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1101 set_num_loop_opts(LoopOptsCount); 1102 set_do_inlining(Inline); 1103 set_max_inline_size(MaxInlineSize); 1104 set_freq_inline_size(FreqInlineSize); 1105 set_do_scheduling(OptoScheduling); 1106 set_do_count_invocations(false); 1107 set_do_method_data_update(false); 1108 set_rtm_state(NoRTM); // No RTM lock eliding by default 1109 method_has_option_value("MaxNodeLimit", _max_node_limit); 1110 #if INCLUDE_RTM_OPT 1111 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1112 int rtm_state = method()->method_data()->rtm_state(); 1113 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1114 // Don't generate RTM lock eliding code. 1115 set_rtm_state(NoRTM); 1116 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1117 // Generate RTM lock eliding code without abort ratio calculation code. 1118 set_rtm_state(UseRTM); 1119 } else if (UseRTMDeopt) { 1120 // Generate RTM lock eliding code and include abort ratio calculation 1121 // code if UseRTMDeopt is on. 1122 set_rtm_state(ProfileRTM); 1123 } 1124 } 1125 #endif 1126 if (debug_info()->recording_non_safepoints()) { 1127 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1128 (comp_arena(), 8, 0, NULL)); 1129 set_default_node_notes(Node_Notes::make(this)); 1130 } 1131 1132 // // -- Initialize types before each compile -- 1133 // // Update cached type information 1134 // if( _method && _method->constants() ) 1135 // Type::update_loaded_types(_method, _method->constants()); 1136 1137 // Init alias_type map. 1138 if (!_do_escape_analysis && aliaslevel == 3) 1139 aliaslevel = 2; // No unique types without escape analysis 1140 _AliasLevel = aliaslevel; 1141 const int grow_ats = 16; 1142 _max_alias_types = grow_ats; 1143 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1144 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1145 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1146 { 1147 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1148 } 1149 // Initialize the first few types. 1150 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1151 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1152 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1153 _num_alias_types = AliasIdxRaw+1; 1154 // Zero out the alias type cache. 1155 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1156 // A NULL adr_type hits in the cache right away. Preload the right answer. 1157 probe_alias_cache(NULL)->_index = AliasIdxTop; 1158 1159 _intrinsics = NULL; 1160 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1161 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1162 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1163 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1164 register_library_intrinsics(); 1165 #ifdef ASSERT 1166 _type_verify_symmetry = true; 1167 #endif 1168 } 1169 1170 //---------------------------init_start---------------------------------------- 1171 // Install the StartNode on this compile object. 1172 void Compile::init_start(StartNode* s) { 1173 if (failing()) 1174 return; // already failing 1175 assert(s == start(), ""); 1176 } 1177 1178 StartNode* Compile::start() const { 1179 assert(!failing(), ""); 1180 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1181 Node* start = root()->fast_out(i); 1182 if( start->is_Start() ) 1183 return start->as_Start(); 1184 } 1185 fatal("Did not find Start node!"); 1186 return NULL; 1187 } 1188 1189 //-------------------------------immutable_memory------------------------------------- 1190 // Access immutable memory 1191 Node* Compile::immutable_memory() { 1192 if (_immutable_memory != NULL) { 1193 return _immutable_memory; 1194 } 1195 StartNode* s = start(); 1196 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1197 Node *p = s->fast_out(i); 1198 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1199 _immutable_memory = p; 1200 return _immutable_memory; 1201 } 1202 } 1203 ShouldNotReachHere(); 1204 return NULL; 1205 } 1206 1207 //----------------------set_cached_top_node------------------------------------ 1208 // Install the cached top node, and make sure Node::is_top works correctly. 1209 void Compile::set_cached_top_node(Node* tn) { 1210 if (tn != NULL) verify_top(tn); 1211 Node* old_top = _top; 1212 _top = tn; 1213 // Calling Node::setup_is_top allows the nodes the chance to adjust 1214 // their _out arrays. 1215 if (_top != NULL) _top->setup_is_top(); 1216 if (old_top != NULL) old_top->setup_is_top(); 1217 assert(_top == NULL || top()->is_top(), ""); 1218 } 1219 1220 #ifdef ASSERT 1221 uint Compile::count_live_nodes_by_graph_walk() { 1222 Unique_Node_List useful(comp_arena()); 1223 // Get useful node list by walking the graph. 1224 identify_useful_nodes(useful); 1225 return useful.size(); 1226 } 1227 1228 void Compile::print_missing_nodes() { 1229 1230 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1231 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1232 return; 1233 } 1234 1235 // This is an expensive function. It is executed only when the user 1236 // specifies VerifyIdealNodeCount option or otherwise knows the 1237 // additional work that needs to be done to identify reachable nodes 1238 // by walking the flow graph and find the missing ones using 1239 // _dead_node_list. 1240 1241 Unique_Node_List useful(comp_arena()); 1242 // Get useful node list by walking the graph. 1243 identify_useful_nodes(useful); 1244 1245 uint l_nodes = C->live_nodes(); 1246 uint l_nodes_by_walk = useful.size(); 1247 1248 if (l_nodes != l_nodes_by_walk) { 1249 if (_log != NULL) { 1250 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1251 _log->stamp(); 1252 _log->end_head(); 1253 } 1254 VectorSet& useful_member_set = useful.member_set(); 1255 int last_idx = l_nodes_by_walk; 1256 for (int i = 0; i < last_idx; i++) { 1257 if (useful_member_set.test(i)) { 1258 if (_dead_node_list.test(i)) { 1259 if (_log != NULL) { 1260 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1261 } 1262 if (PrintIdealNodeCount) { 1263 // Print the log message to tty 1264 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1265 useful.at(i)->dump(); 1266 } 1267 } 1268 } 1269 else if (! _dead_node_list.test(i)) { 1270 if (_log != NULL) { 1271 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1272 } 1273 if (PrintIdealNodeCount) { 1274 // Print the log message to tty 1275 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1276 } 1277 } 1278 } 1279 if (_log != NULL) { 1280 _log->tail("mismatched_nodes"); 1281 } 1282 } 1283 } 1284 #endif 1285 1286 #ifndef PRODUCT 1287 void Compile::verify_top(Node* tn) const { 1288 if (tn != NULL) { 1289 assert(tn->is_Con(), "top node must be a constant"); 1290 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1291 assert(tn->in(0) != NULL, "must have live top node"); 1292 } 1293 } 1294 #endif 1295 1296 1297 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1298 1299 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1300 guarantee(arr != NULL, ""); 1301 int num_blocks = arr->length(); 1302 if (grow_by < num_blocks) grow_by = num_blocks; 1303 int num_notes = grow_by * _node_notes_block_size; 1304 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1305 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1306 while (num_notes > 0) { 1307 arr->append(notes); 1308 notes += _node_notes_block_size; 1309 num_notes -= _node_notes_block_size; 1310 } 1311 assert(num_notes == 0, "exact multiple, please"); 1312 } 1313 1314 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1315 if (source == NULL || dest == NULL) return false; 1316 1317 if (dest->is_Con()) 1318 return false; // Do not push debug info onto constants. 1319 1320 #ifdef ASSERT 1321 // Leave a bread crumb trail pointing to the original node: 1322 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1323 dest->set_debug_orig(source); 1324 } 1325 #endif 1326 1327 if (node_note_array() == NULL) 1328 return false; // Not collecting any notes now. 1329 1330 // This is a copy onto a pre-existing node, which may already have notes. 1331 // If both nodes have notes, do not overwrite any pre-existing notes. 1332 Node_Notes* source_notes = node_notes_at(source->_idx); 1333 if (source_notes == NULL || source_notes->is_clear()) return false; 1334 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1335 if (dest_notes == NULL || dest_notes->is_clear()) { 1336 return set_node_notes_at(dest->_idx, source_notes); 1337 } 1338 1339 Node_Notes merged_notes = (*source_notes); 1340 // The order of operations here ensures that dest notes will win... 1341 merged_notes.update_from(dest_notes); 1342 return set_node_notes_at(dest->_idx, &merged_notes); 1343 } 1344 1345 1346 //--------------------------allow_range_check_smearing------------------------- 1347 // Gating condition for coalescing similar range checks. 1348 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1349 // single covering check that is at least as strong as any of them. 1350 // If the optimization succeeds, the simplified (strengthened) range check 1351 // will always succeed. If it fails, we will deopt, and then give up 1352 // on the optimization. 1353 bool Compile::allow_range_check_smearing() const { 1354 // If this method has already thrown a range-check, 1355 // assume it was because we already tried range smearing 1356 // and it failed. 1357 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1358 return !already_trapped; 1359 } 1360 1361 1362 //------------------------------flatten_alias_type----------------------------- 1363 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1364 int offset = tj->offset(); 1365 TypePtr::PTR ptr = tj->ptr(); 1366 1367 // Known instance (scalarizable allocation) alias only with itself. 1368 bool is_known_inst = tj->isa_oopptr() != NULL && 1369 tj->is_oopptr()->is_known_instance(); 1370 1371 // Process weird unsafe references. 1372 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1373 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1374 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1375 tj = TypeOopPtr::BOTTOM; 1376 ptr = tj->ptr(); 1377 offset = tj->offset(); 1378 } 1379 1380 // Array pointers need some flattening 1381 const TypeAryPtr *ta = tj->isa_aryptr(); 1382 if (ta && ta->is_stable()) { 1383 // Erase stability property for alias analysis. 1384 tj = ta = ta->cast_to_stable(false); 1385 } 1386 if( ta && is_known_inst ) { 1387 if ( offset != Type::OffsetBot && 1388 offset > arrayOopDesc::length_offset_in_bytes() ) { 1389 offset = Type::OffsetBot; // Flatten constant access into array body only 1390 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1391 } 1392 } else if( ta && _AliasLevel >= 2 ) { 1393 // For arrays indexed by constant indices, we flatten the alias 1394 // space to include all of the array body. Only the header, klass 1395 // and array length can be accessed un-aliased. 1396 if( offset != Type::OffsetBot ) { 1397 if( ta->const_oop() ) { // MethodData* or Method* 1398 offset = Type::OffsetBot; // Flatten constant access into array body 1399 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1400 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1401 // range is OK as-is. 1402 tj = ta = TypeAryPtr::RANGE; 1403 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1404 tj = TypeInstPtr::KLASS; // all klass loads look alike 1405 ta = TypeAryPtr::RANGE; // generic ignored junk 1406 ptr = TypePtr::BotPTR; 1407 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1408 tj = TypeInstPtr::MARK; 1409 ta = TypeAryPtr::RANGE; // generic ignored junk 1410 ptr = TypePtr::BotPTR; 1411 } else { // Random constant offset into array body 1412 offset = Type::OffsetBot; // Flatten constant access into array body 1413 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1414 } 1415 } 1416 // Arrays of fixed size alias with arrays of unknown size. 1417 if (ta->size() != TypeInt::POS) { 1418 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1419 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1420 } 1421 // Arrays of known objects become arrays of unknown objects. 1422 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1423 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1425 } 1426 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1427 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1428 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1429 } 1430 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1431 // cannot be distinguished by bytecode alone. 1432 if (ta->elem() == TypeInt::BOOL) { 1433 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1434 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1435 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1436 } 1437 // During the 2nd round of IterGVN, NotNull castings are removed. 1438 // Make sure the Bottom and NotNull variants alias the same. 1439 // Also, make sure exact and non-exact variants alias the same. 1440 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1441 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1442 } 1443 } 1444 1445 // Oop pointers need some flattening 1446 const TypeInstPtr *to = tj->isa_instptr(); 1447 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1448 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1449 if( ptr == TypePtr::Constant ) { 1450 if (to->klass() != ciEnv::current()->Class_klass() || 1451 offset < k->size_helper() * wordSize) { 1452 // No constant oop pointers (such as Strings); they alias with 1453 // unknown strings. 1454 assert(!is_known_inst, "not scalarizable allocation"); 1455 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1456 } 1457 } else if( is_known_inst ) { 1458 tj = to; // Keep NotNull and klass_is_exact for instance type 1459 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1460 // During the 2nd round of IterGVN, NotNull castings are removed. 1461 // Make sure the Bottom and NotNull variants alias the same. 1462 // Also, make sure exact and non-exact variants alias the same. 1463 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1464 } 1465 if (to->speculative() != NULL) { 1466 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1467 } 1468 // Canonicalize the holder of this field 1469 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1470 // First handle header references such as a LoadKlassNode, even if the 1471 // object's klass is unloaded at compile time (4965979). 1472 if (!is_known_inst) { // Do it only for non-instance types 1473 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1474 } 1475 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1476 // Static fields are in the space above the normal instance 1477 // fields in the java.lang.Class instance. 1478 if (to->klass() != ciEnv::current()->Class_klass()) { 1479 to = NULL; 1480 tj = TypeOopPtr::BOTTOM; 1481 offset = tj->offset(); 1482 } 1483 } else { 1484 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1485 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1486 if( is_known_inst ) { 1487 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1488 } else { 1489 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1490 } 1491 } 1492 } 1493 } 1494 1495 // Klass pointers to object array klasses need some flattening 1496 const TypeKlassPtr *tk = tj->isa_klassptr(); 1497 if( tk ) { 1498 // If we are referencing a field within a Klass, we need 1499 // to assume the worst case of an Object. Both exact and 1500 // inexact types must flatten to the same alias class so 1501 // use NotNull as the PTR. 1502 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1503 1504 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1505 TypeKlassPtr::OBJECT->klass(), 1506 offset); 1507 } 1508 1509 ciKlass* klass = tk->klass(); 1510 if( klass->is_obj_array_klass() ) { 1511 ciKlass* k = TypeAryPtr::OOPS->klass(); 1512 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1513 k = TypeInstPtr::BOTTOM->klass(); 1514 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1515 } 1516 1517 // Check for precise loads from the primary supertype array and force them 1518 // to the supertype cache alias index. Check for generic array loads from 1519 // the primary supertype array and also force them to the supertype cache 1520 // alias index. Since the same load can reach both, we need to merge 1521 // these 2 disparate memories into the same alias class. Since the 1522 // primary supertype array is read-only, there's no chance of confusion 1523 // where we bypass an array load and an array store. 1524 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1525 if (offset == Type::OffsetBot || 1526 (offset >= primary_supers_offset && 1527 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1528 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1529 offset = in_bytes(Klass::secondary_super_cache_offset()); 1530 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1531 } 1532 } 1533 1534 // Flatten all Raw pointers together. 1535 if (tj->base() == Type::RawPtr) 1536 tj = TypeRawPtr::BOTTOM; 1537 1538 if (tj->base() == Type::AnyPtr) 1539 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1540 1541 // Flatten all to bottom for now 1542 switch( _AliasLevel ) { 1543 case 0: 1544 tj = TypePtr::BOTTOM; 1545 break; 1546 case 1: // Flatten to: oop, static, field or array 1547 switch (tj->base()) { 1548 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1549 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1550 case Type::AryPtr: // do not distinguish arrays at all 1551 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1552 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1553 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1554 default: ShouldNotReachHere(); 1555 } 1556 break; 1557 case 2: // No collapsing at level 2; keep all splits 1558 case 3: // No collapsing at level 3; keep all splits 1559 break; 1560 default: 1561 Unimplemented(); 1562 } 1563 1564 offset = tj->offset(); 1565 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1566 1567 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1568 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1569 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1570 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1571 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1572 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1573 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1574 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1575 assert( tj->ptr() != TypePtr::TopPTR && 1576 tj->ptr() != TypePtr::AnyNull && 1577 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1578 // assert( tj->ptr() != TypePtr::Constant || 1579 // tj->base() == Type::RawPtr || 1580 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1581 1582 return tj; 1583 } 1584 1585 void Compile::AliasType::Init(int i, const TypePtr* at) { 1586 _index = i; 1587 _adr_type = at; 1588 _field = NULL; 1589 _element = NULL; 1590 _is_rewritable = true; // default 1591 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1592 if (atoop != NULL && atoop->is_known_instance()) { 1593 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1594 _general_index = Compile::current()->get_alias_index(gt); 1595 } else { 1596 _general_index = 0; 1597 } 1598 } 1599 1600 BasicType Compile::AliasType::basic_type() const { 1601 if (element() != NULL) { 1602 const Type* element = adr_type()->is_aryptr()->elem(); 1603 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1604 } if (field() != NULL) { 1605 return field()->layout_type(); 1606 } else { 1607 return T_ILLEGAL; // unknown 1608 } 1609 } 1610 1611 //---------------------------------print_on------------------------------------ 1612 #ifndef PRODUCT 1613 void Compile::AliasType::print_on(outputStream* st) { 1614 if (index() < 10) 1615 st->print("@ <%d> ", index()); 1616 else st->print("@ <%d>", index()); 1617 st->print(is_rewritable() ? " " : " RO"); 1618 int offset = adr_type()->offset(); 1619 if (offset == Type::OffsetBot) 1620 st->print(" +any"); 1621 else st->print(" +%-3d", offset); 1622 st->print(" in "); 1623 adr_type()->dump_on(st); 1624 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1625 if (field() != NULL && tjp) { 1626 if (tjp->klass() != field()->holder() || 1627 tjp->offset() != field()->offset_in_bytes()) { 1628 st->print(" != "); 1629 field()->print(); 1630 st->print(" ***"); 1631 } 1632 } 1633 } 1634 1635 void print_alias_types() { 1636 Compile* C = Compile::current(); 1637 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1638 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1639 C->alias_type(idx)->print_on(tty); 1640 tty->cr(); 1641 } 1642 } 1643 #endif 1644 1645 1646 //----------------------------probe_alias_cache-------------------------------- 1647 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1648 intptr_t key = (intptr_t) adr_type; 1649 key ^= key >> logAliasCacheSize; 1650 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1651 } 1652 1653 1654 //-----------------------------grow_alias_types-------------------------------- 1655 void Compile::grow_alias_types() { 1656 const int old_ats = _max_alias_types; // how many before? 1657 const int new_ats = old_ats; // how many more? 1658 const int grow_ats = old_ats+new_ats; // how many now? 1659 _max_alias_types = grow_ats; 1660 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1661 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1662 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1663 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1664 } 1665 1666 1667 //--------------------------------find_alias_type------------------------------ 1668 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1669 if (_AliasLevel == 0) 1670 return alias_type(AliasIdxBot); 1671 1672 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1673 if (ace->_adr_type == adr_type) { 1674 return alias_type(ace->_index); 1675 } 1676 1677 // Handle special cases. 1678 if (adr_type == NULL) return alias_type(AliasIdxTop); 1679 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1680 1681 // Do it the slow way. 1682 const TypePtr* flat = flatten_alias_type(adr_type); 1683 1684 #ifdef ASSERT 1685 { 1686 ResourceMark rm; 1687 assert(flat == flatten_alias_type(flat), 1688 err_msg("not idempotent: adr_type = %s; flat = %s => %s", Type::str(adr_type), 1689 Type::str(flat), Type::str(flatten_alias_type(flat)))); 1690 assert(flat != TypePtr::BOTTOM, 1691 err_msg("cannot alias-analyze an untyped ptr: adr_type = %s", Type::str(adr_type))); 1692 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1693 const TypeOopPtr* foop = flat->is_oopptr(); 1694 // Scalarizable allocations have exact klass always. 1695 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1696 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1697 assert(foop == flatten_alias_type(xoop), 1698 err_msg("exactness must not affect alias type: foop = %s; xoop = %s", 1699 Type::str(foop), Type::str(xoop))); 1700 } 1701 } 1702 #endif 1703 1704 int idx = AliasIdxTop; 1705 for (int i = 0; i < num_alias_types(); i++) { 1706 if (alias_type(i)->adr_type() == flat) { 1707 idx = i; 1708 break; 1709 } 1710 } 1711 1712 if (idx == AliasIdxTop) { 1713 if (no_create) return NULL; 1714 // Grow the array if necessary. 1715 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1716 // Add a new alias type. 1717 idx = _num_alias_types++; 1718 _alias_types[idx]->Init(idx, flat); 1719 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1720 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1721 if (flat->isa_instptr()) { 1722 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1723 && flat->is_instptr()->klass() == env()->Class_klass()) 1724 alias_type(idx)->set_rewritable(false); 1725 } 1726 if (flat->isa_aryptr()) { 1727 #ifdef ASSERT 1728 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1729 // (T_BYTE has the weakest alignment and size restrictions...) 1730 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1731 #endif 1732 if (flat->offset() == TypePtr::OffsetBot) { 1733 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1734 } 1735 } 1736 if (flat->isa_klassptr()) { 1737 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1738 alias_type(idx)->set_rewritable(false); 1739 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1740 alias_type(idx)->set_rewritable(false); 1741 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1742 alias_type(idx)->set_rewritable(false); 1743 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1744 alias_type(idx)->set_rewritable(false); 1745 } 1746 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1747 // but the base pointer type is not distinctive enough to identify 1748 // references into JavaThread.) 1749 1750 // Check for final fields. 1751 const TypeInstPtr* tinst = flat->isa_instptr(); 1752 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1753 ciField* field; 1754 if (tinst->const_oop() != NULL && 1755 tinst->klass() == ciEnv::current()->Class_klass() && 1756 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1757 // static field 1758 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1759 field = k->get_field_by_offset(tinst->offset(), true); 1760 } else { 1761 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1762 field = k->get_field_by_offset(tinst->offset(), false); 1763 } 1764 assert(field == NULL || 1765 original_field == NULL || 1766 (field->holder() == original_field->holder() && 1767 field->offset() == original_field->offset() && 1768 field->is_static() == original_field->is_static()), "wrong field?"); 1769 // Set field() and is_rewritable() attributes. 1770 if (field != NULL) alias_type(idx)->set_field(field); 1771 } 1772 } 1773 1774 // Fill the cache for next time. 1775 ace->_adr_type = adr_type; 1776 ace->_index = idx; 1777 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1778 1779 // Might as well try to fill the cache for the flattened version, too. 1780 AliasCacheEntry* face = probe_alias_cache(flat); 1781 if (face->_adr_type == NULL) { 1782 face->_adr_type = flat; 1783 face->_index = idx; 1784 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1785 } 1786 1787 return alias_type(idx); 1788 } 1789 1790 1791 Compile::AliasType* Compile::alias_type(ciField* field) { 1792 const TypeOopPtr* t; 1793 if (field->is_static()) 1794 t = TypeInstPtr::make(field->holder()->java_mirror()); 1795 else 1796 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1797 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1798 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1799 return atp; 1800 } 1801 1802 1803 //------------------------------have_alias_type-------------------------------- 1804 bool Compile::have_alias_type(const TypePtr* adr_type) { 1805 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1806 if (ace->_adr_type == adr_type) { 1807 return true; 1808 } 1809 1810 // Handle special cases. 1811 if (adr_type == NULL) return true; 1812 if (adr_type == TypePtr::BOTTOM) return true; 1813 1814 return find_alias_type(adr_type, true, NULL) != NULL; 1815 } 1816 1817 //-----------------------------must_alias-------------------------------------- 1818 // True if all values of the given address type are in the given alias category. 1819 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1820 if (alias_idx == AliasIdxBot) return true; // the universal category 1821 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1822 if (alias_idx == AliasIdxTop) return false; // the empty category 1823 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1824 1825 // the only remaining possible overlap is identity 1826 int adr_idx = get_alias_index(adr_type); 1827 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1828 assert(adr_idx == alias_idx || 1829 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1830 && adr_type != TypeOopPtr::BOTTOM), 1831 "should not be testing for overlap with an unsafe pointer"); 1832 return adr_idx == alias_idx; 1833 } 1834 1835 //------------------------------can_alias-------------------------------------- 1836 // True if any values of the given address type are in the given alias category. 1837 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1838 if (alias_idx == AliasIdxTop) return false; // the empty category 1839 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1840 if (alias_idx == AliasIdxBot) return true; // the universal category 1841 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1842 1843 // the only remaining possible overlap is identity 1844 int adr_idx = get_alias_index(adr_type); 1845 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1846 return adr_idx == alias_idx; 1847 } 1848 1849 1850 1851 //---------------------------pop_warm_call------------------------------------- 1852 WarmCallInfo* Compile::pop_warm_call() { 1853 WarmCallInfo* wci = _warm_calls; 1854 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1855 return wci; 1856 } 1857 1858 //----------------------------Inline_Warm-------------------------------------- 1859 int Compile::Inline_Warm() { 1860 // If there is room, try to inline some more warm call sites. 1861 // %%% Do a graph index compaction pass when we think we're out of space? 1862 if (!InlineWarmCalls) return 0; 1863 1864 int calls_made_hot = 0; 1865 int room_to_grow = NodeCountInliningCutoff - unique(); 1866 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1867 int amount_grown = 0; 1868 WarmCallInfo* call; 1869 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1870 int est_size = (int)call->size(); 1871 if (est_size > (room_to_grow - amount_grown)) { 1872 // This one won't fit anyway. Get rid of it. 1873 call->make_cold(); 1874 continue; 1875 } 1876 call->make_hot(); 1877 calls_made_hot++; 1878 amount_grown += est_size; 1879 amount_to_grow -= est_size; 1880 } 1881 1882 if (calls_made_hot > 0) set_major_progress(); 1883 return calls_made_hot; 1884 } 1885 1886 1887 //----------------------------Finish_Warm-------------------------------------- 1888 void Compile::Finish_Warm() { 1889 if (!InlineWarmCalls) return; 1890 if (failing()) return; 1891 if (warm_calls() == NULL) return; 1892 1893 // Clean up loose ends, if we are out of space for inlining. 1894 WarmCallInfo* call; 1895 while ((call = pop_warm_call()) != NULL) { 1896 call->make_cold(); 1897 } 1898 } 1899 1900 //---------------------cleanup_loop_predicates----------------------- 1901 // Remove the opaque nodes that protect the predicates so that all unused 1902 // checks and uncommon_traps will be eliminated from the ideal graph 1903 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1904 if (predicate_count()==0) return; 1905 for (int i = predicate_count(); i > 0; i--) { 1906 Node * n = predicate_opaque1_node(i-1); 1907 assert(n->Opcode() == Op_Opaque1, "must be"); 1908 igvn.replace_node(n, n->in(1)); 1909 } 1910 assert(predicate_count()==0, "should be clean!"); 1911 } 1912 1913 void Compile::add_range_check_cast(Node* n) { 1914 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1915 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1916 _range_check_casts->append(n); 1917 } 1918 1919 // Remove all range check dependent CastIINodes. 1920 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1921 for (int i = range_check_cast_count(); i > 0; i--) { 1922 Node* cast = range_check_cast_node(i-1); 1923 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1924 igvn.replace_node(cast, cast->in(1)); 1925 } 1926 assert(range_check_cast_count() == 0, "should be empty"); 1927 } 1928 1929 // StringOpts and late inlining of string methods 1930 void Compile::inline_string_calls(bool parse_time) { 1931 { 1932 // remove useless nodes to make the usage analysis simpler 1933 ResourceMark rm; 1934 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1935 } 1936 1937 { 1938 ResourceMark rm; 1939 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1940 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1941 print_method(PHASE_AFTER_STRINGOPTS, 3); 1942 } 1943 1944 // now inline anything that we skipped the first time around 1945 if (!parse_time) { 1946 _late_inlines_pos = _late_inlines.length(); 1947 } 1948 1949 while (_string_late_inlines.length() > 0) { 1950 CallGenerator* cg = _string_late_inlines.pop(); 1951 cg->do_late_inline(); 1952 if (failing()) return; 1953 } 1954 _string_late_inlines.trunc_to(0); 1955 } 1956 1957 // Late inlining of boxing methods 1958 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1959 if (_boxing_late_inlines.length() > 0) { 1960 assert(has_boxed_value(), "inconsistent"); 1961 1962 PhaseGVN* gvn = initial_gvn(); 1963 set_inlining_incrementally(true); 1964 1965 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1966 for_igvn()->clear(); 1967 gvn->replace_with(&igvn); 1968 1969 _late_inlines_pos = _late_inlines.length(); 1970 1971 while (_boxing_late_inlines.length() > 0) { 1972 CallGenerator* cg = _boxing_late_inlines.pop(); 1973 cg->do_late_inline(); 1974 if (failing()) return; 1975 } 1976 _boxing_late_inlines.trunc_to(0); 1977 1978 { 1979 ResourceMark rm; 1980 PhaseRemoveUseless pru(gvn, for_igvn()); 1981 } 1982 1983 igvn = PhaseIterGVN(gvn); 1984 igvn.optimize(); 1985 1986 set_inlining_progress(false); 1987 set_inlining_incrementally(false); 1988 } 1989 } 1990 1991 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1992 assert(IncrementalInline, "incremental inlining should be on"); 1993 PhaseGVN* gvn = initial_gvn(); 1994 1995 set_inlining_progress(false); 1996 for_igvn()->clear(); 1997 gvn->replace_with(&igvn); 1998 1999 int i = 0; 2000 2001 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2002 CallGenerator* cg = _late_inlines.at(i); 2003 _late_inlines_pos = i+1; 2004 cg->do_late_inline(); 2005 if (failing()) return; 2006 } 2007 int j = 0; 2008 for (; i < _late_inlines.length(); i++, j++) { 2009 _late_inlines.at_put(j, _late_inlines.at(i)); 2010 } 2011 _late_inlines.trunc_to(j); 2012 2013 { 2014 ResourceMark rm; 2015 PhaseRemoveUseless pru(gvn, for_igvn()); 2016 } 2017 2018 igvn = PhaseIterGVN(gvn); 2019 } 2020 2021 // Perform incremental inlining until bound on number of live nodes is reached 2022 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2023 PhaseGVN* gvn = initial_gvn(); 2024 2025 set_inlining_incrementally(true); 2026 set_inlining_progress(true); 2027 uint low_live_nodes = 0; 2028 2029 while(inlining_progress() && _late_inlines.length() > 0) { 2030 2031 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2032 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2033 // PhaseIdealLoop is expensive so we only try it once we are 2034 // out of live nodes and we only try it again if the previous 2035 // helped got the number of nodes down significantly 2036 PhaseIdealLoop ideal_loop( igvn, false, true ); 2037 if (failing()) return; 2038 low_live_nodes = live_nodes(); 2039 _major_progress = true; 2040 } 2041 2042 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2043 break; 2044 } 2045 } 2046 2047 inline_incrementally_one(igvn); 2048 2049 if (failing()) return; 2050 2051 igvn.optimize(); 2052 2053 if (failing()) return; 2054 } 2055 2056 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2057 2058 if (_string_late_inlines.length() > 0) { 2059 assert(has_stringbuilder(), "inconsistent"); 2060 for_igvn()->clear(); 2061 initial_gvn()->replace_with(&igvn); 2062 2063 inline_string_calls(false); 2064 2065 if (failing()) return; 2066 2067 { 2068 ResourceMark rm; 2069 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2070 } 2071 2072 igvn = PhaseIterGVN(gvn); 2073 2074 igvn.optimize(); 2075 } 2076 2077 set_inlining_incrementally(false); 2078 } 2079 2080 2081 // Remove edges from "root" to each SafePoint at a backward branch. 2082 // They were inserted during parsing (see add_safepoint()) to make 2083 // infinite loops without calls or exceptions visible to root, i.e., 2084 // useful. 2085 void Compile::remove_root_to_sfpts_edges() { 2086 Node *r = root(); 2087 if (r != NULL) { 2088 for (uint i = r->req(); i < r->len(); ++i) { 2089 Node *n = r->in(i); 2090 if (n != NULL && n->is_SafePoint()) { 2091 r->rm_prec(i); 2092 --i; 2093 } 2094 } 2095 } 2096 } 2097 2098 //------------------------------Optimize--------------------------------------- 2099 // Given a graph, optimize it. 2100 void Compile::Optimize() { 2101 TracePhase t1("optimizer", &_t_optimizer, true); 2102 2103 #ifndef PRODUCT 2104 if (env()->break_at_compile()) { 2105 BREAKPOINT; 2106 } 2107 2108 #endif 2109 2110 ResourceMark rm; 2111 int loop_opts_cnt; 2112 2113 NOT_PRODUCT( verify_graph_edges(); ) 2114 2115 print_method(PHASE_AFTER_PARSING); 2116 2117 { 2118 // Iterative Global Value Numbering, including ideal transforms 2119 // Initialize IterGVN with types and values from parse-time GVN 2120 PhaseIterGVN igvn(initial_gvn()); 2121 { 2122 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2123 igvn.optimize(); 2124 } 2125 2126 print_method(PHASE_ITER_GVN1, 2); 2127 2128 if (failing()) return; 2129 2130 { 2131 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2132 inline_incrementally(igvn); 2133 } 2134 2135 print_method(PHASE_INCREMENTAL_INLINE, 2); 2136 2137 if (failing()) return; 2138 2139 if (eliminate_boxing()) { 2140 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2141 // Inline valueOf() methods now. 2142 inline_boxing_calls(igvn); 2143 2144 if (AlwaysIncrementalInline) { 2145 inline_incrementally(igvn); 2146 } 2147 2148 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2149 2150 if (failing()) return; 2151 } 2152 2153 // Now that all inlining is over, cut edge from root to loop 2154 // safepoints 2155 remove_root_to_sfpts_edges(); 2156 2157 // Remove the speculative part of types and clean up the graph from 2158 // the extra CastPP nodes whose only purpose is to carry them. Do 2159 // that early so that optimizations are not disrupted by the extra 2160 // CastPP nodes. 2161 remove_speculative_types(igvn); 2162 2163 // No more new expensive nodes will be added to the list from here 2164 // so keep only the actual candidates for optimizations. 2165 cleanup_expensive_nodes(igvn); 2166 2167 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2168 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);) 2169 initial_gvn()->replace_with(&igvn); 2170 for_igvn()->clear(); 2171 Unique_Node_List new_worklist(C->comp_arena()); 2172 { 2173 ResourceMark rm; 2174 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2175 } 2176 set_for_igvn(&new_worklist); 2177 igvn = PhaseIterGVN(initial_gvn()); 2178 igvn.optimize(); 2179 } 2180 2181 // Perform escape analysis 2182 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2183 if (has_loops()) { 2184 // Cleanup graph (remove dead nodes). 2185 TracePhase t2("idealLoop", &_t_idealLoop, true); 2186 PhaseIdealLoop ideal_loop( igvn, false, true ); 2187 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2188 if (failing()) return; 2189 } 2190 ConnectionGraph::do_analysis(this, &igvn); 2191 2192 if (failing()) return; 2193 2194 // Optimize out fields loads from scalar replaceable allocations. 2195 igvn.optimize(); 2196 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2197 2198 if (failing()) return; 2199 2200 if (congraph() != NULL && macro_count() > 0) { 2201 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2202 PhaseMacroExpand mexp(igvn); 2203 mexp.eliminate_macro_nodes(); 2204 igvn.set_delay_transform(false); 2205 2206 igvn.optimize(); 2207 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2208 2209 if (failing()) return; 2210 } 2211 } 2212 2213 // Loop transforms on the ideal graph. Range Check Elimination, 2214 // peeling, unrolling, etc. 2215 2216 // Set loop opts counter 2217 loop_opts_cnt = num_loop_opts(); 2218 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2219 { 2220 TracePhase t2("idealLoop", &_t_idealLoop, true); 2221 PhaseIdealLoop ideal_loop( igvn, true ); 2222 loop_opts_cnt--; 2223 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2224 if (failing()) return; 2225 } 2226 // Loop opts pass if partial peeling occurred in previous pass 2227 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2228 TracePhase t3("idealLoop", &_t_idealLoop, true); 2229 PhaseIdealLoop ideal_loop( igvn, false ); 2230 loop_opts_cnt--; 2231 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2232 if (failing()) return; 2233 } 2234 // Loop opts pass for loop-unrolling before CCP 2235 if(major_progress() && (loop_opts_cnt > 0)) { 2236 TracePhase t4("idealLoop", &_t_idealLoop, true); 2237 PhaseIdealLoop ideal_loop( igvn, false ); 2238 loop_opts_cnt--; 2239 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2240 } 2241 if (!failing()) { 2242 // Verify that last round of loop opts produced a valid graph 2243 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2244 PhaseIdealLoop::verify(igvn); 2245 } 2246 } 2247 if (failing()) return; 2248 2249 // Conditional Constant Propagation; 2250 PhaseCCP ccp( &igvn ); 2251 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2252 { 2253 TracePhase t2("ccp", &_t_ccp, true); 2254 ccp.do_transform(); 2255 } 2256 print_method(PHASE_CPP1, 2); 2257 2258 assert( true, "Break here to ccp.dump_old2new_map()"); 2259 2260 // Iterative Global Value Numbering, including ideal transforms 2261 { 2262 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2263 igvn = ccp; 2264 igvn.optimize(); 2265 } 2266 2267 print_method(PHASE_ITER_GVN2, 2); 2268 2269 if (failing()) return; 2270 2271 // Loop transforms on the ideal graph. Range Check Elimination, 2272 // peeling, unrolling, etc. 2273 if(loop_opts_cnt > 0) { 2274 debug_only( int cnt = 0; ); 2275 while(major_progress() && (loop_opts_cnt > 0)) { 2276 TracePhase t2("idealLoop", &_t_idealLoop, true); 2277 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2278 PhaseIdealLoop ideal_loop( igvn, true); 2279 loop_opts_cnt--; 2280 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2281 if (failing()) return; 2282 } 2283 } 2284 2285 { 2286 // Verify that all previous optimizations produced a valid graph 2287 // at least to this point, even if no loop optimizations were done. 2288 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2289 PhaseIdealLoop::verify(igvn); 2290 } 2291 2292 if (range_check_cast_count() > 0) { 2293 // No more loop optimizations. Remove all range check dependent CastIINodes. 2294 C->remove_range_check_casts(igvn); 2295 igvn.optimize(); 2296 } 2297 2298 { 2299 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2300 PhaseMacroExpand mex(igvn); 2301 if (mex.expand_macro_nodes()) { 2302 assert(failing(), "must bail out w/ explicit message"); 2303 return; 2304 } 2305 } 2306 2307 } // (End scope of igvn; run destructor if necessary for asserts.) 2308 2309 dump_inlining(); 2310 // A method with only infinite loops has no edges entering loops from root 2311 { 2312 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2313 if (final_graph_reshaping()) { 2314 assert(failing(), "must bail out w/ explicit message"); 2315 return; 2316 } 2317 } 2318 2319 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2320 } 2321 2322 2323 //------------------------------Code_Gen--------------------------------------- 2324 // Given a graph, generate code for it 2325 void Compile::Code_Gen() { 2326 if (failing()) { 2327 return; 2328 } 2329 2330 // Perform instruction selection. You might think we could reclaim Matcher 2331 // memory PDQ, but actually the Matcher is used in generating spill code. 2332 // Internals of the Matcher (including some VectorSets) must remain live 2333 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2334 // set a bit in reclaimed memory. 2335 2336 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2337 // nodes. Mapping is only valid at the root of each matched subtree. 2338 NOT_PRODUCT( verify_graph_edges(); ) 2339 2340 Matcher matcher; 2341 _matcher = &matcher; 2342 { 2343 TracePhase t2("matcher", &_t_matcher, true); 2344 matcher.match(); 2345 } 2346 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2347 // nodes. Mapping is only valid at the root of each matched subtree. 2348 NOT_PRODUCT( verify_graph_edges(); ) 2349 2350 // If you have too many nodes, or if matching has failed, bail out 2351 check_node_count(0, "out of nodes matching instructions"); 2352 if (failing()) { 2353 return; 2354 } 2355 2356 // Build a proper-looking CFG 2357 PhaseCFG cfg(node_arena(), root(), matcher); 2358 _cfg = &cfg; 2359 { 2360 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2361 bool success = cfg.do_global_code_motion(); 2362 if (!success) { 2363 return; 2364 } 2365 2366 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2367 NOT_PRODUCT( verify_graph_edges(); ) 2368 debug_only( cfg.verify(); ) 2369 } 2370 2371 PhaseChaitin regalloc(unique(), cfg, matcher); 2372 _regalloc = ®alloc; 2373 { 2374 TracePhase t2("regalloc", &_t_registerAllocation, true); 2375 // Perform register allocation. After Chaitin, use-def chains are 2376 // no longer accurate (at spill code) and so must be ignored. 2377 // Node->LRG->reg mappings are still accurate. 2378 _regalloc->Register_Allocate(); 2379 2380 // Bail out if the allocator builds too many nodes 2381 if (failing()) { 2382 return; 2383 } 2384 } 2385 2386 // Prior to register allocation we kept empty basic blocks in case the 2387 // the allocator needed a place to spill. After register allocation we 2388 // are not adding any new instructions. If any basic block is empty, we 2389 // can now safely remove it. 2390 { 2391 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2392 cfg.remove_empty_blocks(); 2393 if (do_freq_based_layout()) { 2394 PhaseBlockLayout layout(cfg); 2395 } else { 2396 cfg.set_loop_alignment(); 2397 } 2398 cfg.fixup_flow(); 2399 } 2400 2401 // Apply peephole optimizations 2402 if( OptoPeephole ) { 2403 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2404 PhasePeephole peep( _regalloc, cfg); 2405 peep.do_transform(); 2406 } 2407 2408 // Do late expand if CPU requires this. 2409 if (Matcher::require_postalloc_expand) { 2410 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2411 cfg.postalloc_expand(_regalloc); 2412 } 2413 2414 // Convert Nodes to instruction bits in a buffer 2415 { 2416 // %%%% workspace merge brought two timers together for one job 2417 TracePhase t2a("output", &_t_output, true); 2418 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2419 Output(); 2420 } 2421 2422 print_method(PHASE_FINAL_CODE); 2423 2424 // He's dead, Jim. 2425 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2426 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2427 } 2428 2429 2430 //------------------------------dump_asm--------------------------------------- 2431 // Dump formatted assembly 2432 #ifndef PRODUCT 2433 void Compile::dump_asm(int *pcs, uint pc_limit) { 2434 bool cut_short = false; 2435 tty->print_cr("#"); 2436 tty->print("# "); _tf->dump(); tty->cr(); 2437 tty->print_cr("#"); 2438 2439 // For all blocks 2440 int pc = 0x0; // Program counter 2441 char starts_bundle = ' '; 2442 _regalloc->dump_frame(); 2443 2444 Node *n = NULL; 2445 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2446 if (VMThread::should_terminate()) { 2447 cut_short = true; 2448 break; 2449 } 2450 Block* block = _cfg->get_block(i); 2451 if (block->is_connector() && !Verbose) { 2452 continue; 2453 } 2454 n = block->head(); 2455 if (pcs && n->_idx < pc_limit) { 2456 tty->print("%3.3x ", pcs[n->_idx]); 2457 } else { 2458 tty->print(" "); 2459 } 2460 block->dump_head(_cfg); 2461 if (block->is_connector()) { 2462 tty->print_cr(" # Empty connector block"); 2463 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2464 tty->print_cr(" # Block is sole successor of call"); 2465 } 2466 2467 // For all instructions 2468 Node *delay = NULL; 2469 for (uint j = 0; j < block->number_of_nodes(); j++) { 2470 if (VMThread::should_terminate()) { 2471 cut_short = true; 2472 break; 2473 } 2474 n = block->get_node(j); 2475 if (valid_bundle_info(n)) { 2476 Bundle* bundle = node_bundling(n); 2477 if (bundle->used_in_unconditional_delay()) { 2478 delay = n; 2479 continue; 2480 } 2481 if (bundle->starts_bundle()) { 2482 starts_bundle = '+'; 2483 } 2484 } 2485 2486 if (WizardMode) { 2487 n->dump(); 2488 } 2489 2490 if( !n->is_Region() && // Dont print in the Assembly 2491 !n->is_Phi() && // a few noisely useless nodes 2492 !n->is_Proj() && 2493 !n->is_MachTemp() && 2494 !n->is_SafePointScalarObject() && 2495 !n->is_Catch() && // Would be nice to print exception table targets 2496 !n->is_MergeMem() && // Not very interesting 2497 !n->is_top() && // Debug info table constants 2498 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2499 ) { 2500 if (pcs && n->_idx < pc_limit) 2501 tty->print("%3.3x", pcs[n->_idx]); 2502 else 2503 tty->print(" "); 2504 tty->print(" %c ", starts_bundle); 2505 starts_bundle = ' '; 2506 tty->print("\t"); 2507 n->format(_regalloc, tty); 2508 tty->cr(); 2509 } 2510 2511 // If we have an instruction with a delay slot, and have seen a delay, 2512 // then back up and print it 2513 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2514 assert(delay != NULL, "no unconditional delay instruction"); 2515 if (WizardMode) delay->dump(); 2516 2517 if (node_bundling(delay)->starts_bundle()) 2518 starts_bundle = '+'; 2519 if (pcs && n->_idx < pc_limit) 2520 tty->print("%3.3x", pcs[n->_idx]); 2521 else 2522 tty->print(" "); 2523 tty->print(" %c ", starts_bundle); 2524 starts_bundle = ' '; 2525 tty->print("\t"); 2526 delay->format(_regalloc, tty); 2527 tty->cr(); 2528 delay = NULL; 2529 } 2530 2531 // Dump the exception table as well 2532 if( n->is_Catch() && (Verbose || WizardMode) ) { 2533 // Print the exception table for this offset 2534 _handler_table.print_subtable_for(pc); 2535 } 2536 } 2537 2538 if (pcs && n->_idx < pc_limit) 2539 tty->print_cr("%3.3x", pcs[n->_idx]); 2540 else 2541 tty->cr(); 2542 2543 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2544 2545 } // End of per-block dump 2546 tty->cr(); 2547 2548 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2549 } 2550 #endif 2551 2552 //------------------------------Final_Reshape_Counts--------------------------- 2553 // This class defines counters to help identify when a method 2554 // may/must be executed using hardware with only 24-bit precision. 2555 struct Final_Reshape_Counts : public StackObj { 2556 int _call_count; // count non-inlined 'common' calls 2557 int _float_count; // count float ops requiring 24-bit precision 2558 int _double_count; // count double ops requiring more precision 2559 int _java_call_count; // count non-inlined 'java' calls 2560 int _inner_loop_count; // count loops which need alignment 2561 VectorSet _visited; // Visitation flags 2562 Node_List _tests; // Set of IfNodes & PCTableNodes 2563 2564 Final_Reshape_Counts() : 2565 _call_count(0), _float_count(0), _double_count(0), 2566 _java_call_count(0), _inner_loop_count(0), 2567 _visited( Thread::current()->resource_area() ) { } 2568 2569 void inc_call_count () { _call_count ++; } 2570 void inc_float_count () { _float_count ++; } 2571 void inc_double_count() { _double_count++; } 2572 void inc_java_call_count() { _java_call_count++; } 2573 void inc_inner_loop_count() { _inner_loop_count++; } 2574 2575 int get_call_count () const { return _call_count ; } 2576 int get_float_count () const { return _float_count ; } 2577 int get_double_count() const { return _double_count; } 2578 int get_java_call_count() const { return _java_call_count; } 2579 int get_inner_loop_count() const { return _inner_loop_count; } 2580 }; 2581 2582 #ifdef ASSERT 2583 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2584 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2585 // Make sure the offset goes inside the instance layout. 2586 return k->contains_field_offset(tp->offset()); 2587 // Note that OffsetBot and OffsetTop are very negative. 2588 } 2589 #endif 2590 2591 // Eliminate trivially redundant StoreCMs and accumulate their 2592 // precedence edges. 2593 void Compile::eliminate_redundant_card_marks(Node* n) { 2594 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2595 if (n->in(MemNode::Address)->outcnt() > 1) { 2596 // There are multiple users of the same address so it might be 2597 // possible to eliminate some of the StoreCMs 2598 Node* mem = n->in(MemNode::Memory); 2599 Node* adr = n->in(MemNode::Address); 2600 Node* val = n->in(MemNode::ValueIn); 2601 Node* prev = n; 2602 bool done = false; 2603 // Walk the chain of StoreCMs eliminating ones that match. As 2604 // long as it's a chain of single users then the optimization is 2605 // safe. Eliminating partially redundant StoreCMs would require 2606 // cloning copies down the other paths. 2607 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2608 if (adr == mem->in(MemNode::Address) && 2609 val == mem->in(MemNode::ValueIn)) { 2610 // redundant StoreCM 2611 if (mem->req() > MemNode::OopStore) { 2612 // Hasn't been processed by this code yet. 2613 n->add_prec(mem->in(MemNode::OopStore)); 2614 } else { 2615 // Already converted to precedence edge 2616 for (uint i = mem->req(); i < mem->len(); i++) { 2617 // Accumulate any precedence edges 2618 if (mem->in(i) != NULL) { 2619 n->add_prec(mem->in(i)); 2620 } 2621 } 2622 // Everything above this point has been processed. 2623 done = true; 2624 } 2625 // Eliminate the previous StoreCM 2626 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2627 assert(mem->outcnt() == 0, "should be dead"); 2628 mem->disconnect_inputs(NULL, this); 2629 } else { 2630 prev = mem; 2631 } 2632 mem = prev->in(MemNode::Memory); 2633 } 2634 } 2635 } 2636 2637 //------------------------------final_graph_reshaping_impl---------------------- 2638 // Implement items 1-5 from final_graph_reshaping below. 2639 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2640 2641 if ( n->outcnt() == 0 ) return; // dead node 2642 uint nop = n->Opcode(); 2643 2644 // Check for 2-input instruction with "last use" on right input. 2645 // Swap to left input. Implements item (2). 2646 if( n->req() == 3 && // two-input instruction 2647 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2648 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2649 n->in(2)->outcnt() == 1 &&// right use IS a last use 2650 !n->in(2)->is_Con() ) { // right use is not a constant 2651 // Check for commutative opcode 2652 switch( nop ) { 2653 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2654 case Op_MaxI: case Op_MinI: 2655 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2656 case Op_AndL: case Op_XorL: case Op_OrL: 2657 case Op_AndI: case Op_XorI: case Op_OrI: { 2658 // Move "last use" input to left by swapping inputs 2659 n->swap_edges(1, 2); 2660 break; 2661 } 2662 default: 2663 break; 2664 } 2665 } 2666 2667 #ifdef ASSERT 2668 if( n->is_Mem() ) { 2669 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2670 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2671 // oop will be recorded in oop map if load crosses safepoint 2672 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2673 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2674 "raw memory operations should have control edge"); 2675 } 2676 #endif 2677 // Count FPU ops and common calls, implements item (3) 2678 switch( nop ) { 2679 // Count all float operations that may use FPU 2680 case Op_AddF: 2681 case Op_SubF: 2682 case Op_MulF: 2683 case Op_DivF: 2684 case Op_NegF: 2685 case Op_ModF: 2686 case Op_ConvI2F: 2687 case Op_ConF: 2688 case Op_CmpF: 2689 case Op_CmpF3: 2690 // case Op_ConvL2F: // longs are split into 32-bit halves 2691 frc.inc_float_count(); 2692 break; 2693 2694 case Op_ConvF2D: 2695 case Op_ConvD2F: 2696 frc.inc_float_count(); 2697 frc.inc_double_count(); 2698 break; 2699 2700 // Count all double operations that may use FPU 2701 case Op_AddD: 2702 case Op_SubD: 2703 case Op_MulD: 2704 case Op_DivD: 2705 case Op_NegD: 2706 case Op_ModD: 2707 case Op_ConvI2D: 2708 case Op_ConvD2I: 2709 // case Op_ConvL2D: // handled by leaf call 2710 // case Op_ConvD2L: // handled by leaf call 2711 case Op_ConD: 2712 case Op_CmpD: 2713 case Op_CmpD3: 2714 frc.inc_double_count(); 2715 break; 2716 case Op_Opaque1: // Remove Opaque Nodes before matching 2717 case Op_Opaque2: // Remove Opaque Nodes before matching 2718 case Op_Opaque3: 2719 n->subsume_by(n->in(1), this); 2720 break; 2721 case Op_CallStaticJava: 2722 case Op_CallJava: 2723 case Op_CallDynamicJava: 2724 frc.inc_java_call_count(); // Count java call site; 2725 case Op_CallRuntime: 2726 case Op_CallLeaf: 2727 case Op_CallLeafNoFP: { 2728 assert( n->is_Call(), "" ); 2729 CallNode *call = n->as_Call(); 2730 // Count call sites where the FP mode bit would have to be flipped. 2731 // Do not count uncommon runtime calls: 2732 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2733 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2734 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2735 frc.inc_call_count(); // Count the call site 2736 } else { // See if uncommon argument is shared 2737 Node *n = call->in(TypeFunc::Parms); 2738 int nop = n->Opcode(); 2739 // Clone shared simple arguments to uncommon calls, item (1). 2740 if( n->outcnt() > 1 && 2741 !n->is_Proj() && 2742 nop != Op_CreateEx && 2743 nop != Op_CheckCastPP && 2744 nop != Op_DecodeN && 2745 nop != Op_DecodeNKlass && 2746 !n->is_Mem() ) { 2747 Node *x = n->clone(); 2748 call->set_req( TypeFunc::Parms, x ); 2749 } 2750 } 2751 break; 2752 } 2753 2754 case Op_StoreD: 2755 case Op_LoadD: 2756 case Op_LoadD_unaligned: 2757 frc.inc_double_count(); 2758 goto handle_mem; 2759 case Op_StoreF: 2760 case Op_LoadF: 2761 frc.inc_float_count(); 2762 goto handle_mem; 2763 2764 case Op_StoreCM: 2765 { 2766 // Convert OopStore dependence into precedence edge 2767 Node* prec = n->in(MemNode::OopStore); 2768 n->del_req(MemNode::OopStore); 2769 n->add_prec(prec); 2770 eliminate_redundant_card_marks(n); 2771 } 2772 2773 // fall through 2774 2775 case Op_StoreB: 2776 case Op_StoreC: 2777 case Op_StorePConditional: 2778 case Op_StoreI: 2779 case Op_StoreL: 2780 case Op_StoreIConditional: 2781 case Op_StoreLConditional: 2782 case Op_CompareAndSwapI: 2783 case Op_CompareAndSwapL: 2784 case Op_CompareAndSwapP: 2785 case Op_CompareAndSwapN: 2786 case Op_GetAndAddI: 2787 case Op_GetAndAddL: 2788 case Op_GetAndSetI: 2789 case Op_GetAndSetL: 2790 case Op_GetAndSetP: 2791 case Op_GetAndSetN: 2792 case Op_StoreP: 2793 case Op_StoreN: 2794 case Op_StoreNKlass: 2795 case Op_LoadB: 2796 case Op_LoadUB: 2797 case Op_LoadUS: 2798 case Op_LoadI: 2799 case Op_LoadKlass: 2800 case Op_LoadNKlass: 2801 case Op_LoadL: 2802 case Op_LoadL_unaligned: 2803 case Op_LoadPLocked: 2804 case Op_LoadP: 2805 case Op_LoadN: 2806 case Op_LoadRange: 2807 case Op_LoadS: { 2808 handle_mem: 2809 #ifdef ASSERT 2810 if( VerifyOptoOopOffsets ) { 2811 assert( n->is_Mem(), "" ); 2812 MemNode *mem = (MemNode*)n; 2813 // Check to see if address types have grounded out somehow. 2814 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2815 assert( !tp || oop_offset_is_sane(tp), "" ); 2816 } 2817 #endif 2818 break; 2819 } 2820 2821 case Op_AddP: { // Assert sane base pointers 2822 Node *addp = n->in(AddPNode::Address); 2823 assert( !addp->is_AddP() || 2824 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2825 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2826 "Base pointers must match" ); 2827 #ifdef _LP64 2828 if ((UseCompressedOops || UseCompressedClassPointers) && 2829 addp->Opcode() == Op_ConP && 2830 addp == n->in(AddPNode::Base) && 2831 n->in(AddPNode::Offset)->is_Con()) { 2832 // Use addressing with narrow klass to load with offset on x86. 2833 // On sparc loading 32-bits constant and decoding it have less 2834 // instructions (4) then load 64-bits constant (7). 2835 // Do this transformation here since IGVN will convert ConN back to ConP. 2836 const Type* t = addp->bottom_type(); 2837 if (t->isa_oopptr() || t->isa_klassptr()) { 2838 Node* nn = NULL; 2839 2840 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2841 2842 // Look for existing ConN node of the same exact type. 2843 Node* r = root(); 2844 uint cnt = r->outcnt(); 2845 for (uint i = 0; i < cnt; i++) { 2846 Node* m = r->raw_out(i); 2847 if (m!= NULL && m->Opcode() == op && 2848 m->bottom_type()->make_ptr() == t) { 2849 nn = m; 2850 break; 2851 } 2852 } 2853 if (nn != NULL) { 2854 // Decode a narrow oop to match address 2855 // [R12 + narrow_oop_reg<<3 + offset] 2856 if (t->isa_oopptr()) { 2857 nn = new (this) DecodeNNode(nn, t); 2858 } else { 2859 nn = new (this) DecodeNKlassNode(nn, t); 2860 } 2861 n->set_req(AddPNode::Base, nn); 2862 n->set_req(AddPNode::Address, nn); 2863 if (addp->outcnt() == 0) { 2864 addp->disconnect_inputs(NULL, this); 2865 } 2866 } 2867 } 2868 } 2869 #endif 2870 break; 2871 } 2872 2873 #ifdef _LP64 2874 case Op_CastPP: 2875 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2876 Node* in1 = n->in(1); 2877 const Type* t = n->bottom_type(); 2878 Node* new_in1 = in1->clone(); 2879 new_in1->as_DecodeN()->set_type(t); 2880 2881 if (!Matcher::narrow_oop_use_complex_address()) { 2882 // 2883 // x86, ARM and friends can handle 2 adds in addressing mode 2884 // and Matcher can fold a DecodeN node into address by using 2885 // a narrow oop directly and do implicit NULL check in address: 2886 // 2887 // [R12 + narrow_oop_reg<<3 + offset] 2888 // NullCheck narrow_oop_reg 2889 // 2890 // On other platforms (Sparc) we have to keep new DecodeN node and 2891 // use it to do implicit NULL check in address: 2892 // 2893 // decode_not_null narrow_oop_reg, base_reg 2894 // [base_reg + offset] 2895 // NullCheck base_reg 2896 // 2897 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2898 // to keep the information to which NULL check the new DecodeN node 2899 // corresponds to use it as value in implicit_null_check(). 2900 // 2901 new_in1->set_req(0, n->in(0)); 2902 } 2903 2904 n->subsume_by(new_in1, this); 2905 if (in1->outcnt() == 0) { 2906 in1->disconnect_inputs(NULL, this); 2907 } 2908 } 2909 break; 2910 2911 case Op_CmpP: 2912 // Do this transformation here to preserve CmpPNode::sub() and 2913 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2914 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2915 Node* in1 = n->in(1); 2916 Node* in2 = n->in(2); 2917 if (!in1->is_DecodeNarrowPtr()) { 2918 in2 = in1; 2919 in1 = n->in(2); 2920 } 2921 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2922 2923 Node* new_in2 = NULL; 2924 if (in2->is_DecodeNarrowPtr()) { 2925 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2926 new_in2 = in2->in(1); 2927 } else if (in2->Opcode() == Op_ConP) { 2928 const Type* t = in2->bottom_type(); 2929 if (t == TypePtr::NULL_PTR) { 2930 assert(in1->is_DecodeN(), "compare klass to null?"); 2931 // Don't convert CmpP null check into CmpN if compressed 2932 // oops implicit null check is not generated. 2933 // This will allow to generate normal oop implicit null check. 2934 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2935 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2936 // 2937 // This transformation together with CastPP transformation above 2938 // will generated code for implicit NULL checks for compressed oops. 2939 // 2940 // The original code after Optimize() 2941 // 2942 // LoadN memory, narrow_oop_reg 2943 // decode narrow_oop_reg, base_reg 2944 // CmpP base_reg, NULL 2945 // CastPP base_reg // NotNull 2946 // Load [base_reg + offset], val_reg 2947 // 2948 // after these transformations will be 2949 // 2950 // LoadN memory, narrow_oop_reg 2951 // CmpN narrow_oop_reg, NULL 2952 // decode_not_null narrow_oop_reg, base_reg 2953 // Load [base_reg + offset], val_reg 2954 // 2955 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2956 // since narrow oops can be used in debug info now (see the code in 2957 // final_graph_reshaping_walk()). 2958 // 2959 // At the end the code will be matched to 2960 // on x86: 2961 // 2962 // Load_narrow_oop memory, narrow_oop_reg 2963 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2964 // NullCheck narrow_oop_reg 2965 // 2966 // and on sparc: 2967 // 2968 // Load_narrow_oop memory, narrow_oop_reg 2969 // decode_not_null narrow_oop_reg, base_reg 2970 // Load [base_reg + offset], val_reg 2971 // NullCheck base_reg 2972 // 2973 } else if (t->isa_oopptr()) { 2974 new_in2 = ConNode::make(this, t->make_narrowoop()); 2975 } else if (t->isa_klassptr()) { 2976 new_in2 = ConNode::make(this, t->make_narrowklass()); 2977 } 2978 } 2979 if (new_in2 != NULL) { 2980 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2981 n->subsume_by(cmpN, this); 2982 if (in1->outcnt() == 0) { 2983 in1->disconnect_inputs(NULL, this); 2984 } 2985 if (in2->outcnt() == 0) { 2986 in2->disconnect_inputs(NULL, this); 2987 } 2988 } 2989 } 2990 break; 2991 2992 case Op_DecodeN: 2993 case Op_DecodeNKlass: 2994 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2995 // DecodeN could be pinned when it can't be fold into 2996 // an address expression, see the code for Op_CastPP above. 2997 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2998 break; 2999 3000 case Op_EncodeP: 3001 case Op_EncodePKlass: { 3002 Node* in1 = n->in(1); 3003 if (in1->is_DecodeNarrowPtr()) { 3004 n->subsume_by(in1->in(1), this); 3005 } else if (in1->Opcode() == Op_ConP) { 3006 const Type* t = in1->bottom_type(); 3007 if (t == TypePtr::NULL_PTR) { 3008 assert(t->isa_oopptr(), "null klass?"); 3009 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 3010 } else if (t->isa_oopptr()) { 3011 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 3012 } else if (t->isa_klassptr()) { 3013 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 3014 } 3015 } 3016 if (in1->outcnt() == 0) { 3017 in1->disconnect_inputs(NULL, this); 3018 } 3019 break; 3020 } 3021 3022 case Op_Proj: { 3023 if (OptimizeStringConcat) { 3024 ProjNode* p = n->as_Proj(); 3025 if (p->_is_io_use) { 3026 // Separate projections were used for the exception path which 3027 // are normally removed by a late inline. If it wasn't inlined 3028 // then they will hang around and should just be replaced with 3029 // the original one. 3030 Node* proj = NULL; 3031 // Replace with just one 3032 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3033 Node *use = i.get(); 3034 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3035 proj = use; 3036 break; 3037 } 3038 } 3039 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3040 if (proj != NULL) { 3041 p->subsume_by(proj, this); 3042 } 3043 } 3044 } 3045 break; 3046 } 3047 3048 case Op_Phi: 3049 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3050 // The EncodeP optimization may create Phi with the same edges 3051 // for all paths. It is not handled well by Register Allocator. 3052 Node* unique_in = n->in(1); 3053 assert(unique_in != NULL, ""); 3054 uint cnt = n->req(); 3055 for (uint i = 2; i < cnt; i++) { 3056 Node* m = n->in(i); 3057 assert(m != NULL, ""); 3058 if (unique_in != m) 3059 unique_in = NULL; 3060 } 3061 if (unique_in != NULL) { 3062 n->subsume_by(unique_in, this); 3063 } 3064 } 3065 break; 3066 3067 #endif 3068 3069 #ifdef ASSERT 3070 case Op_CastII: 3071 // Verify that all range check dependent CastII nodes were removed. 3072 if (n->isa_CastII()->has_range_check()) { 3073 n->dump(3); 3074 assert(false, "Range check dependent CastII node was not removed"); 3075 } 3076 break; 3077 #endif 3078 3079 case Op_ModI: 3080 if (UseDivMod) { 3081 // Check if a%b and a/b both exist 3082 Node* d = n->find_similar(Op_DivI); 3083 if (d) { 3084 // Replace them with a fused divmod if supported 3085 if (Matcher::has_match_rule(Op_DivModI)) { 3086 DivModINode* divmod = DivModINode::make(this, n); 3087 d->subsume_by(divmod->div_proj(), this); 3088 n->subsume_by(divmod->mod_proj(), this); 3089 } else { 3090 // replace a%b with a-((a/b)*b) 3091 Node* mult = new (this) MulINode(d, d->in(2)); 3092 Node* sub = new (this) SubINode(d->in(1), mult); 3093 n->subsume_by(sub, this); 3094 } 3095 } 3096 } 3097 break; 3098 3099 case Op_ModL: 3100 if (UseDivMod) { 3101 // Check if a%b and a/b both exist 3102 Node* d = n->find_similar(Op_DivL); 3103 if (d) { 3104 // Replace them with a fused divmod if supported 3105 if (Matcher::has_match_rule(Op_DivModL)) { 3106 DivModLNode* divmod = DivModLNode::make(this, n); 3107 d->subsume_by(divmod->div_proj(), this); 3108 n->subsume_by(divmod->mod_proj(), this); 3109 } else { 3110 // replace a%b with a-((a/b)*b) 3111 Node* mult = new (this) MulLNode(d, d->in(2)); 3112 Node* sub = new (this) SubLNode(d->in(1), mult); 3113 n->subsume_by(sub, this); 3114 } 3115 } 3116 } 3117 break; 3118 3119 case Op_LoadVector: 3120 case Op_StoreVector: 3121 break; 3122 3123 case Op_PackB: 3124 case Op_PackS: 3125 case Op_PackI: 3126 case Op_PackF: 3127 case Op_PackL: 3128 case Op_PackD: 3129 if (n->req()-1 > 2) { 3130 // Replace many operand PackNodes with a binary tree for matching 3131 PackNode* p = (PackNode*) n; 3132 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3133 n->subsume_by(btp, this); 3134 } 3135 break; 3136 case Op_Loop: 3137 case Op_CountedLoop: 3138 if (n->as_Loop()->is_inner_loop()) { 3139 frc.inc_inner_loop_count(); 3140 } 3141 break; 3142 case Op_LShiftI: 3143 case Op_RShiftI: 3144 case Op_URShiftI: 3145 case Op_LShiftL: 3146 case Op_RShiftL: 3147 case Op_URShiftL: 3148 if (Matcher::need_masked_shift_count) { 3149 // The cpu's shift instructions don't restrict the count to the 3150 // lower 5/6 bits. We need to do the masking ourselves. 3151 Node* in2 = n->in(2); 3152 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3153 const TypeInt* t = in2->find_int_type(); 3154 if (t != NULL && t->is_con()) { 3155 juint shift = t->get_con(); 3156 if (shift > mask) { // Unsigned cmp 3157 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3158 } 3159 } else { 3160 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3161 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3162 n->set_req(2, shift); 3163 } 3164 } 3165 if (in2->outcnt() == 0) { // Remove dead node 3166 in2->disconnect_inputs(NULL, this); 3167 } 3168 } 3169 break; 3170 case Op_MemBarStoreStore: 3171 case Op_MemBarRelease: 3172 // Break the link with AllocateNode: it is no longer useful and 3173 // confuses register allocation. 3174 if (n->req() > MemBarNode::Precedent) { 3175 n->set_req(MemBarNode::Precedent, top()); 3176 } 3177 break; 3178 default: 3179 assert( !n->is_Call(), "" ); 3180 assert( !n->is_Mem(), "" ); 3181 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3182 break; 3183 } 3184 3185 // Collect CFG split points 3186 if (n->is_MultiBranch()) 3187 frc._tests.push(n); 3188 } 3189 3190 //------------------------------final_graph_reshaping_walk--------------------- 3191 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3192 // requires that the walk visits a node's inputs before visiting the node. 3193 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3194 ResourceArea *area = Thread::current()->resource_area(); 3195 Unique_Node_List sfpt(area); 3196 3197 frc._visited.set(root->_idx); // first, mark node as visited 3198 uint cnt = root->req(); 3199 Node *n = root; 3200 uint i = 0; 3201 while (true) { 3202 if (i < cnt) { 3203 // Place all non-visited non-null inputs onto stack 3204 Node* m = n->in(i); 3205 ++i; 3206 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3207 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3208 // compute worst case interpreter size in case of a deoptimization 3209 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3210 3211 sfpt.push(m); 3212 } 3213 cnt = m->req(); 3214 nstack.push(n, i); // put on stack parent and next input's index 3215 n = m; 3216 i = 0; 3217 } 3218 } else { 3219 // Now do post-visit work 3220 final_graph_reshaping_impl( n, frc ); 3221 if (nstack.is_empty()) 3222 break; // finished 3223 n = nstack.node(); // Get node from stack 3224 cnt = n->req(); 3225 i = nstack.index(); 3226 nstack.pop(); // Shift to the next node on stack 3227 } 3228 } 3229 3230 // Skip next transformation if compressed oops are not used. 3231 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3232 (!UseCompressedOops && !UseCompressedClassPointers)) 3233 return; 3234 3235 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3236 // It could be done for an uncommon traps or any safepoints/calls 3237 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3238 while (sfpt.size() > 0) { 3239 n = sfpt.pop(); 3240 JVMState *jvms = n->as_SafePoint()->jvms(); 3241 assert(jvms != NULL, "sanity"); 3242 int start = jvms->debug_start(); 3243 int end = n->req(); 3244 bool is_uncommon = (n->is_CallStaticJava() && 3245 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3246 for (int j = start; j < end; j++) { 3247 Node* in = n->in(j); 3248 if (in->is_DecodeNarrowPtr()) { 3249 bool safe_to_skip = true; 3250 if (!is_uncommon ) { 3251 // Is it safe to skip? 3252 for (uint i = 0; i < in->outcnt(); i++) { 3253 Node* u = in->raw_out(i); 3254 if (!u->is_SafePoint() || 3255 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3256 safe_to_skip = false; 3257 } 3258 } 3259 } 3260 if (safe_to_skip) { 3261 n->set_req(j, in->in(1)); 3262 } 3263 if (in->outcnt() == 0) { 3264 in->disconnect_inputs(NULL, this); 3265 } 3266 } 3267 } 3268 } 3269 } 3270 3271 //------------------------------final_graph_reshaping-------------------------- 3272 // Final Graph Reshaping. 3273 // 3274 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3275 // and not commoned up and forced early. Must come after regular 3276 // optimizations to avoid GVN undoing the cloning. Clone constant 3277 // inputs to Loop Phis; these will be split by the allocator anyways. 3278 // Remove Opaque nodes. 3279 // (2) Move last-uses by commutative operations to the left input to encourage 3280 // Intel update-in-place two-address operations and better register usage 3281 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3282 // calls canonicalizing them back. 3283 // (3) Count the number of double-precision FP ops, single-precision FP ops 3284 // and call sites. On Intel, we can get correct rounding either by 3285 // forcing singles to memory (requires extra stores and loads after each 3286 // FP bytecode) or we can set a rounding mode bit (requires setting and 3287 // clearing the mode bit around call sites). The mode bit is only used 3288 // if the relative frequency of single FP ops to calls is low enough. 3289 // This is a key transform for SPEC mpeg_audio. 3290 // (4) Detect infinite loops; blobs of code reachable from above but not 3291 // below. Several of the Code_Gen algorithms fail on such code shapes, 3292 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3293 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3294 // Detection is by looking for IfNodes where only 1 projection is 3295 // reachable from below or CatchNodes missing some targets. 3296 // (5) Assert for insane oop offsets in debug mode. 3297 3298 bool Compile::final_graph_reshaping() { 3299 // an infinite loop may have been eliminated by the optimizer, 3300 // in which case the graph will be empty. 3301 if (root()->req() == 1) { 3302 record_method_not_compilable("trivial infinite loop"); 3303 return true; 3304 } 3305 3306 // Expensive nodes have their control input set to prevent the GVN 3307 // from freely commoning them. There's no GVN beyond this point so 3308 // no need to keep the control input. We want the expensive nodes to 3309 // be freely moved to the least frequent code path by gcm. 3310 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3311 for (int i = 0; i < expensive_count(); i++) { 3312 _expensive_nodes->at(i)->set_req(0, NULL); 3313 } 3314 3315 Final_Reshape_Counts frc; 3316 3317 // Visit everybody reachable! 3318 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3319 Node_Stack nstack(live_nodes() >> 1); 3320 final_graph_reshaping_walk(nstack, root(), frc); 3321 3322 // Check for unreachable (from below) code (i.e., infinite loops). 3323 for( uint i = 0; i < frc._tests.size(); i++ ) { 3324 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3325 // Get number of CFG targets. 3326 // Note that PCTables include exception targets after calls. 3327 uint required_outcnt = n->required_outcnt(); 3328 if (n->outcnt() != required_outcnt) { 3329 // Check for a few special cases. Rethrow Nodes never take the 3330 // 'fall-thru' path, so expected kids is 1 less. 3331 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3332 if (n->in(0)->in(0)->is_Call()) { 3333 CallNode *call = n->in(0)->in(0)->as_Call(); 3334 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3335 required_outcnt--; // Rethrow always has 1 less kid 3336 } else if (call->req() > TypeFunc::Parms && 3337 call->is_CallDynamicJava()) { 3338 // Check for null receiver. In such case, the optimizer has 3339 // detected that the virtual call will always result in a null 3340 // pointer exception. The fall-through projection of this CatchNode 3341 // will not be populated. 3342 Node *arg0 = call->in(TypeFunc::Parms); 3343 if (arg0->is_Type() && 3344 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3345 required_outcnt--; 3346 } 3347 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3348 call->req() > TypeFunc::Parms+1 && 3349 call->is_CallStaticJava()) { 3350 // Check for negative array length. In such case, the optimizer has 3351 // detected that the allocation attempt will always result in an 3352 // exception. There is no fall-through projection of this CatchNode . 3353 Node *arg1 = call->in(TypeFunc::Parms+1); 3354 if (arg1->is_Type() && 3355 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3356 required_outcnt--; 3357 } 3358 } 3359 } 3360 } 3361 // Recheck with a better notion of 'required_outcnt' 3362 if (n->outcnt() != required_outcnt) { 3363 record_method_not_compilable("malformed control flow"); 3364 return true; // Not all targets reachable! 3365 } 3366 } 3367 // Check that I actually visited all kids. Unreached kids 3368 // must be infinite loops. 3369 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3370 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3371 record_method_not_compilable("infinite loop"); 3372 return true; // Found unvisited kid; must be unreach 3373 } 3374 } 3375 3376 // If original bytecodes contained a mixture of floats and doubles 3377 // check if the optimizer has made it homogenous, item (3). 3378 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3379 frc.get_float_count() > 32 && 3380 frc.get_double_count() == 0 && 3381 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3382 set_24_bit_selection_and_mode( false, true ); 3383 } 3384 3385 set_java_calls(frc.get_java_call_count()); 3386 set_inner_loops(frc.get_inner_loop_count()); 3387 3388 // No infinite loops, no reason to bail out. 3389 return false; 3390 } 3391 3392 //-----------------------------too_many_traps---------------------------------- 3393 // Report if there are too many traps at the current method and bci. 3394 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3395 bool Compile::too_many_traps(ciMethod* method, 3396 int bci, 3397 Deoptimization::DeoptReason reason) { 3398 ciMethodData* md = method->method_data(); 3399 if (md->is_empty()) { 3400 // Assume the trap has not occurred, or that it occurred only 3401 // because of a transient condition during start-up in the interpreter. 3402 return false; 3403 } 3404 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3405 if (md->has_trap_at(bci, m, reason) != 0) { 3406 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3407 // Also, if there are multiple reasons, or if there is no per-BCI record, 3408 // assume the worst. 3409 if (log()) 3410 log()->elem("observe trap='%s' count='%d'", 3411 Deoptimization::trap_reason_name(reason), 3412 md->trap_count(reason)); 3413 return true; 3414 } else { 3415 // Ignore method/bci and see if there have been too many globally. 3416 return too_many_traps(reason, md); 3417 } 3418 } 3419 3420 // Less-accurate variant which does not require a method and bci. 3421 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3422 ciMethodData* logmd) { 3423 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3424 // Too many traps globally. 3425 // Note that we use cumulative trap_count, not just md->trap_count. 3426 if (log()) { 3427 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3428 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3429 Deoptimization::trap_reason_name(reason), 3430 mcount, trap_count(reason)); 3431 } 3432 return true; 3433 } else { 3434 // The coast is clear. 3435 return false; 3436 } 3437 } 3438 3439 //--------------------------too_many_recompiles-------------------------------- 3440 // Report if there are too many recompiles at the current method and bci. 3441 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3442 // Is not eager to return true, since this will cause the compiler to use 3443 // Action_none for a trap point, to avoid too many recompilations. 3444 bool Compile::too_many_recompiles(ciMethod* method, 3445 int bci, 3446 Deoptimization::DeoptReason reason) { 3447 ciMethodData* md = method->method_data(); 3448 if (md->is_empty()) { 3449 // Assume the trap has not occurred, or that it occurred only 3450 // because of a transient condition during start-up in the interpreter. 3451 return false; 3452 } 3453 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3454 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3455 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3456 Deoptimization::DeoptReason per_bc_reason 3457 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3458 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3459 if ((per_bc_reason == Deoptimization::Reason_none 3460 || md->has_trap_at(bci, m, reason) != 0) 3461 // The trap frequency measure we care about is the recompile count: 3462 && md->trap_recompiled_at(bci, m) 3463 && md->overflow_recompile_count() >= bc_cutoff) { 3464 // Do not emit a trap here if it has already caused recompilations. 3465 // Also, if there are multiple reasons, or if there is no per-BCI record, 3466 // assume the worst. 3467 if (log()) 3468 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3469 Deoptimization::trap_reason_name(reason), 3470 md->trap_count(reason), 3471 md->overflow_recompile_count()); 3472 return true; 3473 } else if (trap_count(reason) != 0 3474 && decompile_count() >= m_cutoff) { 3475 // Too many recompiles globally, and we have seen this sort of trap. 3476 // Use cumulative decompile_count, not just md->decompile_count. 3477 if (log()) 3478 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3479 Deoptimization::trap_reason_name(reason), 3480 md->trap_count(reason), trap_count(reason), 3481 md->decompile_count(), decompile_count()); 3482 return true; 3483 } else { 3484 // The coast is clear. 3485 return false; 3486 } 3487 } 3488 3489 // Compute when not to trap. Used by matching trap based nodes and 3490 // NullCheck optimization. 3491 void Compile::set_allowed_deopt_reasons() { 3492 _allowed_reasons = 0; 3493 if (is_method_compilation()) { 3494 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3495 assert(rs < BitsPerInt, "recode bit map"); 3496 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3497 _allowed_reasons |= nth_bit(rs); 3498 } 3499 } 3500 } 3501 } 3502 3503 #ifndef PRODUCT 3504 //------------------------------verify_graph_edges--------------------------- 3505 // Walk the Graph and verify that there is a one-to-one correspondence 3506 // between Use-Def edges and Def-Use edges in the graph. 3507 void Compile::verify_graph_edges(bool no_dead_code) { 3508 if (VerifyGraphEdges) { 3509 ResourceArea *area = Thread::current()->resource_area(); 3510 Unique_Node_List visited(area); 3511 // Call recursive graph walk to check edges 3512 _root->verify_edges(visited); 3513 if (no_dead_code) { 3514 // Now make sure that no visited node is used by an unvisited node. 3515 bool dead_nodes = false; 3516 Unique_Node_List checked(area); 3517 while (visited.size() > 0) { 3518 Node* n = visited.pop(); 3519 checked.push(n); 3520 for (uint i = 0; i < n->outcnt(); i++) { 3521 Node* use = n->raw_out(i); 3522 if (checked.member(use)) continue; // already checked 3523 if (visited.member(use)) continue; // already in the graph 3524 if (use->is_Con()) continue; // a dead ConNode is OK 3525 // At this point, we have found a dead node which is DU-reachable. 3526 if (!dead_nodes) { 3527 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3528 dead_nodes = true; 3529 } 3530 use->dump(2); 3531 tty->print_cr("---"); 3532 checked.push(use); // No repeats; pretend it is now checked. 3533 } 3534 } 3535 assert(!dead_nodes, "using nodes must be reachable from root"); 3536 } 3537 } 3538 } 3539 3540 // Verify GC barriers consistency 3541 // Currently supported: 3542 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3543 void Compile::verify_barriers() { 3544 if (UseG1GC) { 3545 // Verify G1 pre-barriers 3546 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3547 3548 ResourceArea *area = Thread::current()->resource_area(); 3549 Unique_Node_List visited(area); 3550 Node_List worklist(area); 3551 // We're going to walk control flow backwards starting from the Root 3552 worklist.push(_root); 3553 while (worklist.size() > 0) { 3554 Node* x = worklist.pop(); 3555 if (x == NULL || x == top()) continue; 3556 if (visited.member(x)) { 3557 continue; 3558 } else { 3559 visited.push(x); 3560 } 3561 3562 if (x->is_Region()) { 3563 for (uint i = 1; i < x->req(); i++) { 3564 worklist.push(x->in(i)); 3565 } 3566 } else { 3567 worklist.push(x->in(0)); 3568 // We are looking for the pattern: 3569 // /->ThreadLocal 3570 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3571 // \->ConI(0) 3572 // We want to verify that the If and the LoadB have the same control 3573 // See GraphKit::g1_write_barrier_pre() 3574 if (x->is_If()) { 3575 IfNode *iff = x->as_If(); 3576 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3577 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3578 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3579 && cmp->in(1)->is_Load()) { 3580 LoadNode* load = cmp->in(1)->as_Load(); 3581 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3582 && load->in(2)->in(3)->is_Con() 3583 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3584 3585 Node* if_ctrl = iff->in(0); 3586 Node* load_ctrl = load->in(0); 3587 3588 if (if_ctrl != load_ctrl) { 3589 // Skip possible CProj->NeverBranch in infinite loops 3590 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3591 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3592 if_ctrl = if_ctrl->in(0)->in(0); 3593 } 3594 } 3595 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3596 } 3597 } 3598 } 3599 } 3600 } 3601 } 3602 } 3603 } 3604 3605 #endif 3606 3607 // The Compile object keeps track of failure reasons separately from the ciEnv. 3608 // This is required because there is not quite a 1-1 relation between the 3609 // ciEnv and its compilation task and the Compile object. Note that one 3610 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3611 // to backtrack and retry without subsuming loads. Other than this backtracking 3612 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3613 // by the logic in C2Compiler. 3614 void Compile::record_failure(const char* reason) { 3615 if (log() != NULL) { 3616 log()->elem("failure reason='%s' phase='compile'", reason); 3617 } 3618 if (_failure_reason == NULL) { 3619 // Record the first failure reason. 3620 _failure_reason = reason; 3621 } 3622 3623 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3624 C->print_method(PHASE_FAILURE); 3625 } 3626 _root = NULL; // flush the graph, too 3627 } 3628 3629 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3630 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3631 _phase_name(name), _dolog(dolog) 3632 { 3633 if (dolog) { 3634 C = Compile::current(); 3635 _log = C->log(); 3636 } else { 3637 C = NULL; 3638 _log = NULL; 3639 } 3640 if (_log != NULL) { 3641 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3642 _log->stamp(); 3643 _log->end_head(); 3644 } 3645 } 3646 3647 Compile::TracePhase::~TracePhase() { 3648 3649 C = Compile::current(); 3650 if (_dolog) { 3651 _log = C->log(); 3652 } else { 3653 _log = NULL; 3654 } 3655 3656 #ifdef ASSERT 3657 if (PrintIdealNodeCount) { 3658 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3659 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3660 } 3661 3662 if (VerifyIdealNodeCount) { 3663 Compile::current()->print_missing_nodes(); 3664 } 3665 #endif 3666 3667 if (_log != NULL) { 3668 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3669 } 3670 } 3671 3672 //============================================================================= 3673 // Two Constant's are equal when the type and the value are equal. 3674 bool Compile::Constant::operator==(const Constant& other) { 3675 if (type() != other.type() ) return false; 3676 if (can_be_reused() != other.can_be_reused()) return false; 3677 // For floating point values we compare the bit pattern. 3678 switch (type()) { 3679 case T_FLOAT: return (_v._value.i == other._v._value.i); 3680 case T_LONG: 3681 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3682 case T_OBJECT: 3683 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3684 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3685 case T_METADATA: return (_v._metadata == other._v._metadata); 3686 default: ShouldNotReachHere(); 3687 } 3688 return false; 3689 } 3690 3691 static int type_to_size_in_bytes(BasicType t) { 3692 switch (t) { 3693 case T_LONG: return sizeof(jlong ); 3694 case T_FLOAT: return sizeof(jfloat ); 3695 case T_DOUBLE: return sizeof(jdouble); 3696 case T_METADATA: return sizeof(Metadata*); 3697 // We use T_VOID as marker for jump-table entries (labels) which 3698 // need an internal word relocation. 3699 case T_VOID: 3700 case T_ADDRESS: 3701 case T_OBJECT: return sizeof(jobject); 3702 } 3703 3704 ShouldNotReachHere(); 3705 return -1; 3706 } 3707 3708 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3709 // sort descending 3710 if (a->freq() > b->freq()) return -1; 3711 if (a->freq() < b->freq()) return 1; 3712 return 0; 3713 } 3714 3715 void Compile::ConstantTable::calculate_offsets_and_size() { 3716 // First, sort the array by frequencies. 3717 _constants.sort(qsort_comparator); 3718 3719 #ifdef ASSERT 3720 // Make sure all jump-table entries were sorted to the end of the 3721 // array (they have a negative frequency). 3722 bool found_void = false; 3723 for (int i = 0; i < _constants.length(); i++) { 3724 Constant con = _constants.at(i); 3725 if (con.type() == T_VOID) 3726 found_void = true; // jump-tables 3727 else 3728 assert(!found_void, "wrong sorting"); 3729 } 3730 #endif 3731 3732 int offset = 0; 3733 for (int i = 0; i < _constants.length(); i++) { 3734 Constant* con = _constants.adr_at(i); 3735 3736 // Align offset for type. 3737 int typesize = type_to_size_in_bytes(con->type()); 3738 offset = align_size_up(offset, typesize); 3739 con->set_offset(offset); // set constant's offset 3740 3741 if (con->type() == T_VOID) { 3742 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3743 offset = offset + typesize * n->outcnt(); // expand jump-table 3744 } else { 3745 offset = offset + typesize; 3746 } 3747 } 3748 3749 // Align size up to the next section start (which is insts; see 3750 // CodeBuffer::align_at_start). 3751 assert(_size == -1, "already set?"); 3752 _size = align_size_up(offset, CodeEntryAlignment); 3753 } 3754 3755 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3756 MacroAssembler _masm(&cb); 3757 for (int i = 0; i < _constants.length(); i++) { 3758 Constant con = _constants.at(i); 3759 address constant_addr = NULL; 3760 switch (con.type()) { 3761 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3762 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3763 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3764 case T_OBJECT: { 3765 jobject obj = con.get_jobject(); 3766 int oop_index = _masm.oop_recorder()->find_index(obj); 3767 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3768 break; 3769 } 3770 case T_ADDRESS: { 3771 address addr = (address) con.get_jobject(); 3772 constant_addr = _masm.address_constant(addr); 3773 break; 3774 } 3775 // We use T_VOID as marker for jump-table entries (labels) which 3776 // need an internal word relocation. 3777 case T_VOID: { 3778 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3779 // Fill the jump-table with a dummy word. The real value is 3780 // filled in later in fill_jump_table. 3781 address dummy = (address) n; 3782 constant_addr = _masm.address_constant(dummy); 3783 // Expand jump-table 3784 for (uint i = 1; i < n->outcnt(); i++) { 3785 address temp_addr = _masm.address_constant(dummy + i); 3786 assert(temp_addr, "consts section too small"); 3787 } 3788 break; 3789 } 3790 case T_METADATA: { 3791 Metadata* obj = con.get_metadata(); 3792 int metadata_index = _masm.oop_recorder()->find_index(obj); 3793 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3794 break; 3795 } 3796 default: ShouldNotReachHere(); 3797 } 3798 assert(constant_addr, "consts section too small"); 3799 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3800 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3801 } 3802 } 3803 3804 int Compile::ConstantTable::find_offset(Constant& con) const { 3805 int idx = _constants.find(con); 3806 assert(idx != -1, "constant must be in constant table"); 3807 int offset = _constants.at(idx).offset(); 3808 assert(offset != -1, "constant table not emitted yet?"); 3809 return offset; 3810 } 3811 3812 void Compile::ConstantTable::add(Constant& con) { 3813 if (con.can_be_reused()) { 3814 int idx = _constants.find(con); 3815 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3816 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3817 return; 3818 } 3819 } 3820 (void) _constants.append(con); 3821 } 3822 3823 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3824 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3825 Constant con(type, value, b->_freq); 3826 add(con); 3827 return con; 3828 } 3829 3830 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3831 Constant con(metadata); 3832 add(con); 3833 return con; 3834 } 3835 3836 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3837 jvalue value; 3838 BasicType type = oper->type()->basic_type(); 3839 switch (type) { 3840 case T_LONG: value.j = oper->constantL(); break; 3841 case T_FLOAT: value.f = oper->constantF(); break; 3842 case T_DOUBLE: value.d = oper->constantD(); break; 3843 case T_OBJECT: 3844 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3845 case T_METADATA: return add((Metadata*)oper->constant()); break; 3846 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3847 } 3848 return add(n, type, value); 3849 } 3850 3851 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3852 jvalue value; 3853 // We can use the node pointer here to identify the right jump-table 3854 // as this method is called from Compile::Fill_buffer right before 3855 // the MachNodes are emitted and the jump-table is filled (means the 3856 // MachNode pointers do not change anymore). 3857 value.l = (jobject) n; 3858 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3859 add(con); 3860 return con; 3861 } 3862 3863 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3864 // If called from Compile::scratch_emit_size do nothing. 3865 if (Compile::current()->in_scratch_emit_size()) return; 3866 3867 assert(labels.is_nonempty(), "must be"); 3868 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3869 3870 // Since MachConstantNode::constant_offset() also contains 3871 // table_base_offset() we need to subtract the table_base_offset() 3872 // to get the plain offset into the constant table. 3873 int offset = n->constant_offset() - table_base_offset(); 3874 3875 MacroAssembler _masm(&cb); 3876 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3877 3878 for (uint i = 0; i < n->outcnt(); i++) { 3879 address* constant_addr = &jump_table_base[i]; 3880 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))); 3881 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3882 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3883 } 3884 } 3885 3886 void Compile::dump_inlining() { 3887 if (print_inlining() || print_intrinsics()) { 3888 // Print inlining message for candidates that we couldn't inline 3889 // for lack of space or non constant receiver 3890 for (int i = 0; i < _late_inlines.length(); i++) { 3891 CallGenerator* cg = _late_inlines.at(i); 3892 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3893 } 3894 Unique_Node_List useful; 3895 useful.push(root()); 3896 for (uint next = 0; next < useful.size(); ++next) { 3897 Node* n = useful.at(next); 3898 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3899 CallNode* call = n->as_Call(); 3900 CallGenerator* cg = call->generator(); 3901 cg->print_inlining_late("receiver not constant"); 3902 } 3903 uint max = n->len(); 3904 for ( uint i = 0; i < max; ++i ) { 3905 Node *m = n->in(i); 3906 if ( m == NULL ) continue; 3907 useful.push(m); 3908 } 3909 } 3910 for (int i = 0; i < _print_inlining_list->length(); i++) { 3911 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3912 } 3913 } 3914 } 3915 3916 // Dump inlining replay data to the stream. 3917 // Don't change thread state and acquire any locks. 3918 void Compile::dump_inline_data(outputStream* out) { 3919 InlineTree* inl_tree = ilt(); 3920 if (inl_tree != NULL) { 3921 out->print(" inline %d", inl_tree->count()); 3922 inl_tree->dump_replay_data(out); 3923 } 3924 } 3925 3926 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3927 if (n1->Opcode() < n2->Opcode()) return -1; 3928 else if (n1->Opcode() > n2->Opcode()) return 1; 3929 3930 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())); 3931 for (uint i = 1; i < n1->req(); i++) { 3932 if (n1->in(i) < n2->in(i)) return -1; 3933 else if (n1->in(i) > n2->in(i)) return 1; 3934 } 3935 3936 return 0; 3937 } 3938 3939 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3940 Node* n1 = *n1p; 3941 Node* n2 = *n2p; 3942 3943 return cmp_expensive_nodes(n1, n2); 3944 } 3945 3946 void Compile::sort_expensive_nodes() { 3947 if (!expensive_nodes_sorted()) { 3948 _expensive_nodes->sort(cmp_expensive_nodes); 3949 } 3950 } 3951 3952 bool Compile::expensive_nodes_sorted() const { 3953 for (int i = 1; i < _expensive_nodes->length(); i++) { 3954 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3955 return false; 3956 } 3957 } 3958 return true; 3959 } 3960 3961 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3962 if (_expensive_nodes->length() == 0) { 3963 return false; 3964 } 3965 3966 assert(OptimizeExpensiveOps, "optimization off?"); 3967 3968 // Take this opportunity to remove dead nodes from the list 3969 int j = 0; 3970 for (int i = 0; i < _expensive_nodes->length(); i++) { 3971 Node* n = _expensive_nodes->at(i); 3972 if (!n->is_unreachable(igvn)) { 3973 assert(n->is_expensive(), "should be expensive"); 3974 _expensive_nodes->at_put(j, n); 3975 j++; 3976 } 3977 } 3978 _expensive_nodes->trunc_to(j); 3979 3980 // Then sort the list so that similar nodes are next to each other 3981 // and check for at least two nodes of identical kind with same data 3982 // inputs. 3983 sort_expensive_nodes(); 3984 3985 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3986 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3987 return true; 3988 } 3989 } 3990 3991 return false; 3992 } 3993 3994 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3995 if (_expensive_nodes->length() == 0) { 3996 return; 3997 } 3998 3999 assert(OptimizeExpensiveOps, "optimization off?"); 4000 4001 // Sort to bring similar nodes next to each other and clear the 4002 // control input of nodes for which there's only a single copy. 4003 sort_expensive_nodes(); 4004 4005 int j = 0; 4006 int identical = 0; 4007 int i = 0; 4008 for (; i < _expensive_nodes->length()-1; i++) { 4009 assert(j <= i, "can't write beyond current index"); 4010 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4011 identical++; 4012 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4013 continue; 4014 } 4015 if (identical > 0) { 4016 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4017 identical = 0; 4018 } else { 4019 Node* n = _expensive_nodes->at(i); 4020 igvn.hash_delete(n); 4021 n->set_req(0, NULL); 4022 igvn.hash_insert(n); 4023 } 4024 } 4025 if (identical > 0) { 4026 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4027 } else if (_expensive_nodes->length() >= 1) { 4028 Node* n = _expensive_nodes->at(i); 4029 igvn.hash_delete(n); 4030 n->set_req(0, NULL); 4031 igvn.hash_insert(n); 4032 } 4033 _expensive_nodes->trunc_to(j); 4034 } 4035 4036 void Compile::add_expensive_node(Node * n) { 4037 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4038 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4039 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4040 if (OptimizeExpensiveOps) { 4041 _expensive_nodes->append(n); 4042 } else { 4043 // Clear control input and let IGVN optimize expensive nodes if 4044 // OptimizeExpensiveOps is off. 4045 n->set_req(0, NULL); 4046 } 4047 } 4048 4049 /** 4050 * Remove the speculative part of types and clean up the graph 4051 */ 4052 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4053 if (UseTypeSpeculation) { 4054 Unique_Node_List worklist; 4055 worklist.push(root()); 4056 int modified = 0; 4057 // Go over all type nodes that carry a speculative type, drop the 4058 // speculative part of the type and enqueue the node for an igvn 4059 // which may optimize it out. 4060 for (uint next = 0; next < worklist.size(); ++next) { 4061 Node *n = worklist.at(next); 4062 if (n->is_Type()) { 4063 TypeNode* tn = n->as_Type(); 4064 const Type* t = tn->type(); 4065 const Type* t_no_spec = t->remove_speculative(); 4066 if (t_no_spec != t) { 4067 bool in_hash = igvn.hash_delete(n); 4068 assert(in_hash, "node should be in igvn hash table"); 4069 tn->set_type(t_no_spec); 4070 igvn.hash_insert(n); 4071 igvn._worklist.push(n); // give it a chance to go away 4072 modified++; 4073 } 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 // Drop the speculative part of all types in the igvn's type table 4083 igvn.remove_speculative_types(); 4084 if (modified > 0) { 4085 igvn.optimize(); 4086 } 4087 #ifdef ASSERT 4088 // Verify that after the IGVN is over no speculative type has resurfaced 4089 worklist.clear(); 4090 worklist.push(root()); 4091 for (uint next = 0; next < worklist.size(); ++next) { 4092 Node *n = worklist.at(next); 4093 const Type* t = igvn.type_or_null(n); 4094 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4095 if (n->is_Type()) { 4096 t = n->as_Type()->type(); 4097 assert(t == t->remove_speculative(), "no more speculative types"); 4098 } 4099 uint max = n->len(); 4100 for( uint i = 0; i < max; ++i ) { 4101 Node *m = n->in(i); 4102 if (not_a_node(m)) continue; 4103 worklist.push(m); 4104 } 4105 } 4106 igvn.check_no_speculative_types(); 4107 #endif 4108 } 4109 } 4110 4111 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4112 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4113 if (ctrl != NULL) { 4114 // Express control dependency by a CastII node with a narrow type. 4115 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */); 4116 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4117 // node from floating above the range check during loop optimizations. Otherwise, the 4118 // ConvI2L node may be eliminated independently of the range check, causing the data path 4119 // to become TOP while the control path is still there (although it's unreachable). 4120 value->set_req(0, ctrl); 4121 // Save CastII node to remove it after loop optimizations. 4122 phase->C->add_range_check_cast(value); 4123 value = phase->transform(value); 4124 } 4125 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4126 return phase->transform(new (phase->C) ConvI2LNode(value, ltype)); 4127 } 4128 4129 // Auxiliary method to support randomized stressing/fuzzing. 4130 // 4131 // This method can be called the arbitrary number of times, with current count 4132 // as the argument. The logic allows selecting a single candidate from the 4133 // running list of candidates as follows: 4134 // int count = 0; 4135 // Cand* selected = null; 4136 // while(cand = cand->next()) { 4137 // if (randomized_select(++count)) { 4138 // selected = cand; 4139 // } 4140 // } 4141 // 4142 // Including count equalizes the chances any candidate is "selected". 4143 // This is useful when we don't have the complete list of candidates to choose 4144 // from uniformly. In this case, we need to adjust the randomicity of the 4145 // selection, or else we will end up biasing the selection towards the latter 4146 // candidates. 4147 // 4148 // Quick back-envelope calculation shows that for the list of n candidates 4149 // the equal probability for the candidate to persist as "best" can be 4150 // achieved by replacing it with "next" k-th candidate with the probability 4151 // of 1/k. It can be easily shown that by the end of the run, the 4152 // probability for any candidate is converged to 1/n, thus giving the 4153 // uniform distribution among all the candidates. 4154 // 4155 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4156 #define RANDOMIZED_DOMAIN_POW 29 4157 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4158 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4159 bool Compile::randomized_select(int count) { 4160 assert(count > 0, "only positive"); 4161 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4162 }