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 "opto/addnode.hpp" 36 #include "opto/block.hpp" 37 #include "opto/c2compiler.hpp" 38 #include "opto/callGenerator.hpp" 39 #include "opto/callnode.hpp" 40 #include "opto/cfgnode.hpp" 41 #include "opto/chaitin.hpp" 42 #include "opto/compile.hpp" 43 #include "opto/connode.hpp" 44 #include "opto/divnode.hpp" 45 #include "opto/escape.hpp" 46 #include "opto/idealGraphPrinter.hpp" 47 #include "opto/loopnode.hpp" 48 #include "opto/machnode.hpp" 49 #include "opto/macro.hpp" 50 #include "opto/matcher.hpp" 51 #include "opto/mathexactnode.hpp" 52 #include "opto/memnode.hpp" 53 #include "opto/mulnode.hpp" 54 #include "opto/node.hpp" 55 #include "opto/opcodes.hpp" 56 #include "opto/output.hpp" 57 #include "opto/parse.hpp" 58 #include "opto/phaseX.hpp" 59 #include "opto/rootnode.hpp" 60 #include "opto/runtime.hpp" 61 #include "opto/stringopts.hpp" 62 #include "opto/type.hpp" 63 #include "opto/vectornode.hpp" 64 #include "runtime/arguments.hpp" 65 #include "runtime/signature.hpp" 66 #include "runtime/stubRoutines.hpp" 67 #include "runtime/timer.hpp" 68 #include "trace/tracing.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 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == NULL) { 89 _mach_constant_base_node = new (C) MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 101 #ifdef ASSERT 102 for (int i = 1; i < _intrinsics->length(); i++) { 103 CallGenerator* cg1 = _intrinsics->at(i-1); 104 CallGenerator* cg2 = _intrinsics->at(i); 105 assert(cg1->method() != cg2->method() 106 ? cg1->method() < cg2->method() 107 : cg1->is_virtual() < cg2->is_virtual(), 108 "compiler intrinsics list must stay sorted"); 109 } 110 #endif 111 // Binary search sorted list, in decreasing intervals [lo, hi]. 112 int lo = 0, hi = _intrinsics->length()-1; 113 while (lo <= hi) { 114 int mid = (uint)(hi + lo) / 2; 115 ciMethod* mid_m = _intrinsics->at(mid)->method(); 116 if (m < mid_m) { 117 hi = mid-1; 118 } else if (m > mid_m) { 119 lo = mid+1; 120 } else { 121 // look at minor sort key 122 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 123 if (is_virtual < mid_virt) { 124 hi = mid-1; 125 } else if (is_virtual > mid_virt) { 126 lo = mid+1; 127 } else { 128 return mid; // exact match 129 } 130 } 131 } 132 return lo; // inexact match 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 if (_intrinsics == NULL) { 137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 138 } 139 // This code is stolen from ciObjectFactory::insert. 140 // Really, GrowableArray should have methods for 141 // insert_at, remove_at, and binary_search. 142 int len = _intrinsics->length(); 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 144 if (index == len) { 145 _intrinsics->append(cg); 146 } else { 147 #ifdef ASSERT 148 CallGenerator* oldcg = _intrinsics->at(index); 149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 150 #endif 151 _intrinsics->append(_intrinsics->at(len-1)); 152 int pos; 153 for (pos = len-2; pos >= index; pos--) { 154 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 155 } 156 _intrinsics->at_put(index, cg); 157 } 158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 159 } 160 161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 162 assert(m->is_loaded(), "don't try this on unloaded methods"); 163 if (_intrinsics != NULL) { 164 int index = intrinsic_insertion_index(m, is_virtual); 165 if (index < _intrinsics->length() 166 && _intrinsics->at(index)->method() == m 167 && _intrinsics->at(index)->is_virtual() == is_virtual) { 168 return _intrinsics->at(index); 169 } 170 } 171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 172 if (m->intrinsic_id() != vmIntrinsics::_none && 173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 175 if (cg != NULL) { 176 // Save it for next time: 177 register_intrinsic(cg); 178 return cg; 179 } else { 180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 181 } 182 } 183 return NULL; 184 } 185 186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 187 // in library_call.cpp. 188 189 190 #ifndef PRODUCT 191 // statistics gathering... 192 193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 195 196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 198 int oflags = _intrinsic_hist_flags[id]; 199 assert(flags != 0, "what happened?"); 200 if (is_virtual) { 201 flags |= _intrinsic_virtual; 202 } 203 bool changed = (flags != oflags); 204 if ((flags & _intrinsic_worked) != 0) { 205 juint count = (_intrinsic_hist_count[id] += 1); 206 if (count == 1) { 207 changed = true; // first time 208 } 209 // increment the overall count also: 210 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 211 } 212 if (changed) { 213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 214 // Something changed about the intrinsic's virtuality. 215 if ((flags & _intrinsic_virtual) != 0) { 216 // This is the first use of this intrinsic as a virtual call. 217 if (oflags != 0) { 218 // We already saw it as a non-virtual, so note both cases. 219 flags |= _intrinsic_both; 220 } 221 } else if ((oflags & _intrinsic_both) == 0) { 222 // This is the first use of this intrinsic as a non-virtual 223 flags |= _intrinsic_both; 224 } 225 } 226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 227 } 228 // update the overall flags also: 229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 230 return changed; 231 } 232 233 static char* format_flags(int flags, char* buf) { 234 buf[0] = 0; 235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 240 if (buf[0] == 0) strcat(buf, ","); 241 assert(buf[0] == ',', "must be"); 242 return &buf[1]; 243 } 244 245 void Compile::print_intrinsic_statistics() { 246 char flagsbuf[100]; 247 ttyLocker ttyl; 248 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 249 tty->print_cr("Compiler intrinsic usage:"); 250 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 251 if (total == 0) total = 1; // avoid div0 in case of no successes 252 #define PRINT_STAT_LINE(name, c, f) \ 253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 255 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 256 int flags = _intrinsic_hist_flags[id]; 257 juint count = _intrinsic_hist_count[id]; 258 if ((flags | count) != 0) { 259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 260 } 261 } 262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 263 if (xtty != NULL) xtty->tail("statistics"); 264 } 265 266 void Compile::print_statistics() { 267 { ttyLocker ttyl; 268 if (xtty != NULL) xtty->head("statistics type='opto'"); 269 Parse::print_statistics(); 270 PhaseCCP::print_statistics(); 271 PhaseRegAlloc::print_statistics(); 272 Scheduling::print_statistics(); 273 PhasePeephole::print_statistics(); 274 PhaseIdealLoop::print_statistics(); 275 if (xtty != NULL) xtty->tail("statistics"); 276 } 277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 278 // put this under its own <statistics> element. 279 print_intrinsic_statistics(); 280 } 281 } 282 #endif //PRODUCT 283 284 // Support for bundling info 285 Bundle* Compile::node_bundling(const Node *n) { 286 assert(valid_bundle_info(n), "oob"); 287 return &_node_bundling_base[n->_idx]; 288 } 289 290 bool Compile::valid_bundle_info(const Node *n) { 291 return (_node_bundling_limit > n->_idx); 292 } 293 294 295 void Compile::gvn_replace_by(Node* n, Node* nn) { 296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 297 Node* use = n->last_out(i); 298 bool is_in_table = initial_gvn()->hash_delete(use); 299 uint uses_found = 0; 300 for (uint j = 0; j < use->len(); j++) { 301 if (use->in(j) == n) { 302 if (j < use->req()) 303 use->set_req(j, nn); 304 else 305 use->set_prec(j, nn); 306 uses_found++; 307 } 308 } 309 if (is_in_table) { 310 // reinsert into table 311 initial_gvn()->hash_find_insert(use); 312 } 313 record_for_igvn(use); 314 i -= uses_found; // we deleted 1 or more copies of this edge 315 } 316 } 317 318 319 static inline bool not_a_node(const Node* n) { 320 if (n == NULL) return true; 321 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 322 if (*(address*)n == badAddress) return true; // kill by Node::destruct 323 return false; 324 } 325 326 // Identify all nodes that are reachable from below, useful. 327 // Use breadth-first pass that records state in a Unique_Node_List, 328 // recursive traversal is slower. 329 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 330 int estimated_worklist_size = live_nodes(); 331 useful.map( estimated_worklist_size, NULL ); // preallocate space 332 333 // Initialize worklist 334 if (root() != NULL) { useful.push(root()); } 335 // If 'top' is cached, declare it useful to preserve cached node 336 if( cached_top_node() ) { useful.push(cached_top_node()); } 337 338 // Push all useful nodes onto the list, breadthfirst 339 for( uint next = 0; next < useful.size(); ++next ) { 340 assert( next < unique(), "Unique useful nodes < total nodes"); 341 Node *n = useful.at(next); 342 uint max = n->len(); 343 for( uint i = 0; i < max; ++i ) { 344 Node *m = n->in(i); 345 if (not_a_node(m)) continue; 346 useful.push(m); 347 } 348 } 349 } 350 351 // Update dead_node_list with any missing dead nodes using useful 352 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 353 void Compile::update_dead_node_list(Unique_Node_List &useful) { 354 uint max_idx = unique(); 355 VectorSet& useful_node_set = useful.member_set(); 356 357 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 358 // If node with index node_idx is not in useful set, 359 // mark it as dead in dead node list. 360 if (! useful_node_set.test(node_idx) ) { 361 record_dead_node(node_idx); 362 } 363 } 364 } 365 366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 367 int shift = 0; 368 for (int i = 0; i < inlines->length(); i++) { 369 CallGenerator* cg = inlines->at(i); 370 CallNode* call = cg->call_node(); 371 if (shift > 0) { 372 inlines->at_put(i-shift, cg); 373 } 374 if (!useful.member(call)) { 375 shift++; 376 } 377 } 378 inlines->trunc_to(inlines->length()-shift); 379 } 380 381 // Disconnect all useless nodes by disconnecting those at the boundary. 382 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 383 uint next = 0; 384 while (next < useful.size()) { 385 Node *n = useful.at(next++); 386 if (n->is_SafePoint()) { 387 // We're done with a parsing phase. Replaced nodes are not valid 388 // beyond that point. 389 n->as_SafePoint()->delete_replaced_nodes(); 390 } 391 // Use raw traversal of out edges since this code removes out edges 392 int max = n->outcnt(); 393 for (int j = 0; j < max; ++j) { 394 Node* child = n->raw_out(j); 395 if (! useful.member(child)) { 396 assert(!child->is_top() || child != top(), 397 "If top is cached in Compile object it is in useful list"); 398 // Only need to remove this out-edge to the useless node 399 n->raw_del_out(j); 400 --j; 401 --max; 402 } 403 } 404 if (n->outcnt() == 1 && n->has_special_unique_user()) { 405 record_for_igvn(n->unique_out()); 406 } 407 } 408 // Remove useless macro and predicate opaq nodes 409 for (int i = C->macro_count()-1; i >= 0; i--) { 410 Node* n = C->macro_node(i); 411 if (!useful.member(n)) { 412 remove_macro_node(n); 413 } 414 } 415 // Remove useless CastII nodes with range check dependency 416 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 417 Node* cast = range_check_cast_node(i); 418 if (!useful.member(cast)) { 419 remove_range_check_cast(cast); 420 } 421 } 422 // Remove useless expensive node 423 for (int i = C->expensive_count()-1; i >= 0; i--) { 424 Node* n = C->expensive_node(i); 425 if (!useful.member(n)) { 426 remove_expensive_node(n); 427 } 428 } 429 // clean up the late inline lists 430 remove_useless_late_inlines(&_string_late_inlines, useful); 431 remove_useless_late_inlines(&_boxing_late_inlines, useful); 432 remove_useless_late_inlines(&_late_inlines, useful); 433 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 434 } 435 436 //------------------------------frame_size_in_words----------------------------- 437 // frame_slots in units of words 438 int Compile::frame_size_in_words() const { 439 // shift is 0 in LP32 and 1 in LP64 440 const int shift = (LogBytesPerWord - LogBytesPerInt); 441 int words = _frame_slots >> shift; 442 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 443 return words; 444 } 445 446 // To bang the stack of this compiled method we use the stack size 447 // that the interpreter would need in case of a deoptimization. This 448 // removes the need to bang the stack in the deoptimization blob which 449 // in turn simplifies stack overflow handling. 450 int Compile::bang_size_in_bytes() const { 451 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 452 } 453 454 // ============================================================================ 455 //------------------------------CompileWrapper--------------------------------- 456 class CompileWrapper : public StackObj { 457 Compile *const _compile; 458 public: 459 CompileWrapper(Compile* compile); 460 461 ~CompileWrapper(); 462 }; 463 464 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 465 // the Compile* pointer is stored in the current ciEnv: 466 ciEnv* env = compile->env(); 467 assert(env == ciEnv::current(), "must already be a ciEnv active"); 468 assert(env->compiler_data() == NULL, "compile already active?"); 469 env->set_compiler_data(compile); 470 assert(compile == Compile::current(), "sanity"); 471 472 compile->set_type_dict(NULL); 473 compile->set_type_hwm(NULL); 474 compile->set_type_last_size(0); 475 compile->set_last_tf(NULL, NULL); 476 compile->set_indexSet_arena(NULL); 477 compile->set_indexSet_free_block_list(NULL); 478 compile->init_type_arena(); 479 Type::Initialize(compile); 480 _compile->set_scratch_buffer_blob(NULL); 481 _compile->begin_method(); 482 } 483 CompileWrapper::~CompileWrapper() { 484 _compile->end_method(); 485 if (_compile->scratch_buffer_blob() != NULL) 486 BufferBlob::free(_compile->scratch_buffer_blob()); 487 _compile->env()->set_compiler_data(NULL); 488 } 489 490 491 //----------------------------print_compile_messages--------------------------- 492 void Compile::print_compile_messages() { 493 #ifndef PRODUCT 494 // Check if recompiling 495 if (_subsume_loads == false && PrintOpto) { 496 // Recompiling without allowing machine instructions to subsume loads 497 tty->print_cr("*********************************************************"); 498 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 499 tty->print_cr("*********************************************************"); 500 } 501 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 502 // Recompiling without escape analysis 503 tty->print_cr("*********************************************************"); 504 tty->print_cr("** Bailout: Recompile without escape analysis **"); 505 tty->print_cr("*********************************************************"); 506 } 507 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 508 // Recompiling without boxing elimination 509 tty->print_cr("*********************************************************"); 510 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 511 tty->print_cr("*********************************************************"); 512 } 513 if (env()->break_at_compile()) { 514 // Open the debugger when compiling this method. 515 tty->print("### Breaking when compiling: "); 516 method()->print_short_name(); 517 tty->cr(); 518 BREAKPOINT; 519 } 520 521 if( PrintOpto ) { 522 if (is_osr_compilation()) { 523 tty->print("[OSR]%3d", _compile_id); 524 } else { 525 tty->print("%3d", _compile_id); 526 } 527 } 528 #endif 529 } 530 531 532 //-----------------------init_scratch_buffer_blob------------------------------ 533 // Construct a temporary BufferBlob and cache it for this compile. 534 void Compile::init_scratch_buffer_blob(int const_size) { 535 // If there is already a scratch buffer blob allocated and the 536 // constant section is big enough, use it. Otherwise free the 537 // current and allocate a new one. 538 BufferBlob* blob = scratch_buffer_blob(); 539 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 540 // Use the current blob. 541 } else { 542 if (blob != NULL) { 543 BufferBlob::free(blob); 544 } 545 546 ResourceMark rm; 547 _scratch_const_size = const_size; 548 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 549 blob = BufferBlob::create("Compile::scratch_buffer", size); 550 // Record the buffer blob for next time. 551 set_scratch_buffer_blob(blob); 552 // Have we run out of code space? 553 if (scratch_buffer_blob() == NULL) { 554 // Let CompilerBroker disable further compilations. 555 record_failure("Not enough space for scratch buffer in CodeCache"); 556 return; 557 } 558 } 559 560 // Initialize the relocation buffers 561 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 562 set_scratch_locs_memory(locs_buf); 563 } 564 565 566 //-----------------------scratch_emit_size------------------------------------- 567 // Helper function that computes size by emitting code 568 uint Compile::scratch_emit_size(const Node* n) { 569 // Start scratch_emit_size section. 570 set_in_scratch_emit_size(true); 571 572 // Emit into a trash buffer and count bytes emitted. 573 // This is a pretty expensive way to compute a size, 574 // but it works well enough if seldom used. 575 // All common fixed-size instructions are given a size 576 // method by the AD file. 577 // Note that the scratch buffer blob and locs memory are 578 // allocated at the beginning of the compile task, and 579 // may be shared by several calls to scratch_emit_size. 580 // The allocation of the scratch buffer blob is particularly 581 // expensive, since it has to grab the code cache lock. 582 BufferBlob* blob = this->scratch_buffer_blob(); 583 assert(blob != NULL, "Initialize BufferBlob at start"); 584 assert(blob->size() > MAX_inst_size, "sanity"); 585 relocInfo* locs_buf = scratch_locs_memory(); 586 address blob_begin = blob->content_begin(); 587 address blob_end = (address)locs_buf; 588 assert(blob->content_contains(blob_end), "sanity"); 589 CodeBuffer buf(blob_begin, blob_end - blob_begin); 590 buf.initialize_consts_size(_scratch_const_size); 591 buf.initialize_stubs_size(MAX_stubs_size); 592 assert(locs_buf != NULL, "sanity"); 593 int lsize = MAX_locs_size / 3; 594 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 595 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 596 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 597 598 // Do the emission. 599 600 Label fakeL; // Fake label for branch instructions. 601 Label* saveL = NULL; 602 uint save_bnum = 0; 603 bool is_branch = n->is_MachBranch(); 604 if (is_branch) { 605 MacroAssembler masm(&buf); 606 masm.bind(fakeL); 607 n->as_MachBranch()->save_label(&saveL, &save_bnum); 608 n->as_MachBranch()->label_set(&fakeL, 0); 609 } 610 n->emit(buf, this->regalloc()); 611 612 // Emitting into the scratch buffer should not fail 613 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 614 615 if (is_branch) // Restore label. 616 n->as_MachBranch()->label_set(saveL, save_bnum); 617 618 // End scratch_emit_size section. 619 set_in_scratch_emit_size(false); 620 621 return buf.insts_size(); 622 } 623 624 625 // ============================================================================ 626 //------------------------------Compile standard------------------------------- 627 debug_only( int Compile::_debug_idx = 100000; ) 628 629 // Compile a method. entry_bci is -1 for normal compilations and indicates 630 // the continuation bci for on stack replacement. 631 632 633 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 634 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 635 : Phase(Compiler), 636 _env(ci_env), 637 _log(ci_env->log()), 638 _compile_id(ci_env->compile_id()), 639 _save_argument_registers(false), 640 _stub_name(NULL), 641 _stub_function(NULL), 642 _stub_entry_point(NULL), 643 _method(target), 644 _entry_bci(osr_bci), 645 _initial_gvn(NULL), 646 _for_igvn(NULL), 647 _warm_calls(NULL), 648 _subsume_loads(subsume_loads), 649 _do_escape_analysis(do_escape_analysis), 650 _eliminate_boxing(eliminate_boxing), 651 _failure_reason(NULL), 652 _code_buffer("Compile::Fill_buffer"), 653 _orig_pc_slot(0), 654 _orig_pc_slot_offset_in_bytes(0), 655 _has_method_handle_invokes(false), 656 _mach_constant_base_node(NULL), 657 _node_bundling_limit(0), 658 _node_bundling_base(NULL), 659 _java_calls(0), 660 _inner_loops(0), 661 _scratch_const_size(-1), 662 _in_scratch_emit_size(false), 663 _dead_node_list(comp_arena()), 664 _dead_node_count(0), 665 #ifndef PRODUCT 666 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 667 _in_dump_cnt(0), 668 _printer(IdealGraphPrinter::printer()), 669 #endif 670 _congraph(NULL), 671 _comp_arena(mtCompiler), 672 _node_arena(mtCompiler), 673 _old_arena(mtCompiler), 674 _Compile_types(mtCompiler), 675 _replay_inline_data(NULL), 676 _late_inlines(comp_arena(), 2, 0, NULL), 677 _string_late_inlines(comp_arena(), 2, 0, NULL), 678 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 679 _late_inlines_pos(0), 680 _number_of_mh_late_inlines(0), 681 _inlining_progress(false), 682 _inlining_incrementally(false), 683 _print_inlining_list(NULL), 684 _print_inlining_idx(0), 685 _interpreter_frame_size(0), 686 _max_node_limit(MaxNodeLimit) { 687 C = this; 688 689 CompileWrapper cw(this); 690 #ifndef PRODUCT 691 if (TimeCompiler2) { 692 tty->print(" "); 693 target->holder()->name()->print(); 694 tty->print("."); 695 target->print_short_name(); 696 tty->print(" "); 697 } 698 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 699 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 700 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 701 if (!print_opto_assembly) { 702 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 703 if (print_assembly && !Disassembler::can_decode()) { 704 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 705 print_opto_assembly = true; 706 } 707 } 708 set_print_assembly(print_opto_assembly); 709 set_parsed_irreducible_loop(false); 710 711 if (method()->has_option("ReplayInline")) { 712 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 713 } 714 #endif 715 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 716 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 717 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 718 719 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 720 // Make sure the method being compiled gets its own MDO, 721 // so we can at least track the decompile_count(). 722 // Need MDO to record RTM code generation state. 723 method()->ensure_method_data(); 724 } 725 726 Init(::AliasLevel); 727 728 729 print_compile_messages(); 730 731 _ilt = InlineTree::build_inline_tree_root(); 732 733 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 734 assert(num_alias_types() >= AliasIdxRaw, ""); 735 736 #define MINIMUM_NODE_HASH 1023 737 // Node list that Iterative GVN will start with 738 Unique_Node_List for_igvn(comp_arena()); 739 set_for_igvn(&for_igvn); 740 741 // GVN that will be run immediately on new nodes 742 uint estimated_size = method()->code_size()*4+64; 743 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 744 PhaseGVN gvn(node_arena(), estimated_size); 745 set_initial_gvn(&gvn); 746 747 if (print_inlining() || print_intrinsics()) { 748 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 749 } 750 { // Scope for timing the parser 751 TracePhase t3("parse", &_t_parser, true); 752 753 // Put top into the hash table ASAP. 754 initial_gvn()->transform_no_reclaim(top()); 755 756 // Set up tf(), start(), and find a CallGenerator. 757 CallGenerator* cg = NULL; 758 if (is_osr_compilation()) { 759 const TypeTuple *domain = StartOSRNode::osr_domain(); 760 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 761 init_tf(TypeFunc::make(domain, range)); 762 StartNode* s = new (this) StartOSRNode(root(), domain); 763 initial_gvn()->set_type_bottom(s); 764 init_start(s); 765 cg = CallGenerator::for_osr(method(), entry_bci()); 766 } else { 767 // Normal case. 768 init_tf(TypeFunc::make(method())); 769 StartNode* s = new (this) StartNode(root(), tf()->domain()); 770 initial_gvn()->set_type_bottom(s); 771 init_start(s); 772 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 773 // With java.lang.ref.reference.get() we must go through the 774 // intrinsic when G1 is enabled - even when get() is the root 775 // method of the compile - so that, if necessary, the value in 776 // the referent field of the reference object gets recorded by 777 // the pre-barrier code. 778 // Specifically, if G1 is enabled, the value in the referent 779 // field is recorded by the G1 SATB pre barrier. This will 780 // result in the referent being marked live and the reference 781 // object removed from the list of discovered references during 782 // reference processing. 783 cg = find_intrinsic(method(), false); 784 } 785 if (cg == NULL) { 786 float past_uses = method()->interpreter_invocation_count(); 787 float expected_uses = past_uses; 788 cg = CallGenerator::for_inline(method(), expected_uses); 789 } 790 } 791 if (failing()) return; 792 if (cg == NULL) { 793 record_method_not_compilable_all_tiers("cannot parse method"); 794 return; 795 } 796 JVMState* jvms = build_start_state(start(), tf()); 797 if ((jvms = cg->generate(jvms)) == NULL) { 798 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 799 record_method_not_compilable("method parse failed"); 800 } 801 return; 802 } 803 GraphKit kit(jvms); 804 805 if (!kit.stopped()) { 806 // Accept return values, and transfer control we know not where. 807 // This is done by a special, unique ReturnNode bound to root. 808 return_values(kit.jvms()); 809 } 810 811 if (kit.has_exceptions()) { 812 // Any exceptions that escape from this call must be rethrown 813 // to whatever caller is dynamically above us on the stack. 814 // This is done by a special, unique RethrowNode bound to root. 815 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 816 } 817 818 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 819 820 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 821 inline_string_calls(true); 822 } 823 824 if (failing()) return; 825 826 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 827 828 // Remove clutter produced by parsing. 829 if (!failing()) { 830 ResourceMark rm; 831 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 832 } 833 } 834 835 // Note: Large methods are capped off in do_one_bytecode(). 836 if (failing()) return; 837 838 // After parsing, node notes are no longer automagic. 839 // They must be propagated by register_new_node_with_optimizer(), 840 // clone(), or the like. 841 set_default_node_notes(NULL); 842 843 for (;;) { 844 int successes = Inline_Warm(); 845 if (failing()) return; 846 if (successes == 0) break; 847 } 848 849 // Drain the list. 850 Finish_Warm(); 851 #ifndef PRODUCT 852 if (_printer) { 853 _printer->print_inlining(this); 854 } 855 #endif 856 857 if (failing()) return; 858 NOT_PRODUCT( verify_graph_edges(); ) 859 860 // Now optimize 861 Optimize(); 862 if (failing()) return; 863 NOT_PRODUCT( verify_graph_edges(); ) 864 865 #ifndef PRODUCT 866 if (PrintIdeal) { 867 ttyLocker ttyl; // keep the following output all in one block 868 // This output goes directly to the tty, not the compiler log. 869 // To enable tools to match it up with the compilation activity, 870 // be sure to tag this tty output with the compile ID. 871 if (xtty != NULL) { 872 xtty->head("ideal compile_id='%d'%s", compile_id(), 873 is_osr_compilation() ? " compile_kind='osr'" : 874 ""); 875 } 876 root()->dump(9999); 877 if (xtty != NULL) { 878 xtty->tail("ideal"); 879 } 880 } 881 #endif 882 883 NOT_PRODUCT( verify_barriers(); ) 884 885 // Dump compilation data to replay it. 886 if (method()->has_option("DumpReplay")) { 887 env()->dump_replay_data(_compile_id); 888 } 889 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 890 env()->dump_inline_data(_compile_id); 891 } 892 893 // Now that we know the size of all the monitors we can add a fixed slot 894 // for the original deopt pc. 895 896 _orig_pc_slot = fixed_slots(); 897 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 898 set_fixed_slots(next_slot); 899 900 // Compute when to use implicit null checks. Used by matching trap based 901 // nodes and NullCheck optimization. 902 set_allowed_deopt_reasons(); 903 904 // Now generate code 905 Code_Gen(); 906 if (failing()) return; 907 908 // Check if we want to skip execution of all compiled code. 909 { 910 #ifndef PRODUCT 911 if (OptoNoExecute) { 912 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 913 return; 914 } 915 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 916 #endif 917 918 if (is_osr_compilation()) { 919 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 920 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 921 } else { 922 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 923 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 924 } 925 926 env()->register_method(_method, _entry_bci, 927 &_code_offsets, 928 _orig_pc_slot_offset_in_bytes, 929 code_buffer(), 930 frame_size_in_words(), _oop_map_set, 931 &_handler_table, &_inc_table, 932 compiler, 933 env()->comp_level(), 934 has_unsafe_access(), 935 SharedRuntime::is_wide_vector(max_vector_size()), 936 rtm_state() 937 ); 938 939 if (log() != NULL) // Print code cache state into compiler log 940 log()->code_cache_state(); 941 } 942 } 943 944 //------------------------------Compile---------------------------------------- 945 // Compile a runtime stub 946 Compile::Compile( ciEnv* ci_env, 947 TypeFunc_generator generator, 948 address stub_function, 949 const char *stub_name, 950 int is_fancy_jump, 951 bool pass_tls, 952 bool save_arg_registers, 953 bool return_pc ) 954 : Phase(Compiler), 955 _env(ci_env), 956 _log(ci_env->log()), 957 _compile_id(0), 958 _save_argument_registers(save_arg_registers), 959 _method(NULL), 960 _stub_name(stub_name), 961 _stub_function(stub_function), 962 _stub_entry_point(NULL), 963 _entry_bci(InvocationEntryBci), 964 _initial_gvn(NULL), 965 _for_igvn(NULL), 966 _warm_calls(NULL), 967 _orig_pc_slot(0), 968 _orig_pc_slot_offset_in_bytes(0), 969 _subsume_loads(true), 970 _do_escape_analysis(false), 971 _eliminate_boxing(false), 972 _failure_reason(NULL), 973 _code_buffer("Compile::Fill_buffer"), 974 _has_method_handle_invokes(false), 975 _mach_constant_base_node(NULL), 976 _node_bundling_limit(0), 977 _node_bundling_base(NULL), 978 _java_calls(0), 979 _inner_loops(0), 980 #ifndef PRODUCT 981 _trace_opto_output(TraceOptoOutput), 982 _in_dump_cnt(0), 983 _printer(NULL), 984 #endif 985 _comp_arena(mtCompiler), 986 _node_arena(mtCompiler), 987 _old_arena(mtCompiler), 988 _Compile_types(mtCompiler), 989 _dead_node_list(comp_arena()), 990 _dead_node_count(0), 991 _congraph(NULL), 992 _replay_inline_data(NULL), 993 _number_of_mh_late_inlines(0), 994 _inlining_progress(false), 995 _inlining_incrementally(false), 996 _print_inlining_list(NULL), 997 _print_inlining_idx(0), 998 _allowed_reasons(0), 999 _interpreter_frame_size(0), 1000 _max_node_limit(MaxNodeLimit) { 1001 C = this; 1002 1003 #ifndef PRODUCT 1004 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 1005 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 1006 set_print_assembly(PrintFrameConverterAssembly); 1007 set_parsed_irreducible_loop(false); 1008 #endif 1009 set_has_irreducible_loop(false); // no loops 1010 1011 CompileWrapper cw(this); 1012 Init(/*AliasLevel=*/ 0); 1013 init_tf((*generator)()); 1014 1015 { 1016 // The following is a dummy for the sake of GraphKit::gen_stub 1017 Unique_Node_List for_igvn(comp_arena()); 1018 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1019 PhaseGVN gvn(Thread::current()->resource_area(),255); 1020 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1021 gvn.transform_no_reclaim(top()); 1022 1023 GraphKit kit; 1024 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1025 } 1026 1027 NOT_PRODUCT( verify_graph_edges(); ) 1028 Code_Gen(); 1029 if (failing()) return; 1030 1031 1032 // Entry point will be accessed using compile->stub_entry_point(); 1033 if (code_buffer() == NULL) { 1034 Matcher::soft_match_failure(); 1035 } else { 1036 if (PrintAssembly && (WizardMode || Verbose)) 1037 tty->print_cr("### Stub::%s", stub_name); 1038 1039 if (!failing()) { 1040 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1041 1042 // Make the NMethod 1043 // For now we mark the frame as never safe for profile stackwalking 1044 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1045 code_buffer(), 1046 CodeOffsets::frame_never_safe, 1047 // _code_offsets.value(CodeOffsets::Frame_Complete), 1048 frame_size_in_words(), 1049 _oop_map_set, 1050 save_arg_registers); 1051 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1052 1053 _stub_entry_point = rs->entry_point(); 1054 } 1055 } 1056 } 1057 1058 //------------------------------Init------------------------------------------- 1059 // Prepare for a single compilation 1060 void Compile::Init(int aliaslevel) { 1061 _unique = 0; 1062 _regalloc = NULL; 1063 1064 _tf = NULL; // filled in later 1065 _top = NULL; // cached later 1066 _matcher = NULL; // filled in later 1067 _cfg = NULL; // filled in later 1068 1069 set_24_bit_selection_and_mode(Use24BitFP, false); 1070 1071 _node_note_array = NULL; 1072 _default_node_notes = NULL; 1073 1074 _immutable_memory = NULL; // filled in at first inquiry 1075 1076 // Globally visible Nodes 1077 // First set TOP to NULL to give safe behavior during creation of RootNode 1078 set_cached_top_node(NULL); 1079 set_root(new (this) RootNode()); 1080 // Now that you have a Root to point to, create the real TOP 1081 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1082 set_recent_alloc(NULL, NULL); 1083 1084 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1085 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1086 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1087 env()->set_dependencies(new Dependencies(env())); 1088 1089 _fixed_slots = 0; 1090 set_has_split_ifs(false); 1091 set_has_loops(has_method() && method()->has_loops()); // first approximation 1092 set_has_stringbuilder(false); 1093 set_has_boxed_value(false); 1094 _trap_can_recompile = false; // no traps emitted yet 1095 _major_progress = true; // start out assuming good things will happen 1096 set_has_unsafe_access(false); 1097 set_max_vector_size(0); 1098 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1099 set_decompile_count(0); 1100 1101 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1102 set_num_loop_opts(LoopOptsCount); 1103 set_do_inlining(Inline); 1104 set_max_inline_size(MaxInlineSize); 1105 set_freq_inline_size(FreqInlineSize); 1106 set_do_scheduling(OptoScheduling); 1107 set_do_count_invocations(false); 1108 set_do_method_data_update(false); 1109 set_rtm_state(NoRTM); // No RTM lock eliding by default 1110 method_has_option_value("MaxNodeLimit", _max_node_limit); 1111 #if INCLUDE_RTM_OPT 1112 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1113 int rtm_state = method()->method_data()->rtm_state(); 1114 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1115 // Don't generate RTM lock eliding code. 1116 set_rtm_state(NoRTM); 1117 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1118 // Generate RTM lock eliding code without abort ratio calculation code. 1119 set_rtm_state(UseRTM); 1120 } else if (UseRTMDeopt) { 1121 // Generate RTM lock eliding code and include abort ratio calculation 1122 // code if UseRTMDeopt is on. 1123 set_rtm_state(ProfileRTM); 1124 } 1125 } 1126 #endif 1127 if (debug_info()->recording_non_safepoints()) { 1128 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1129 (comp_arena(), 8, 0, NULL)); 1130 set_default_node_notes(Node_Notes::make(this)); 1131 } 1132 1133 // // -- Initialize types before each compile -- 1134 // // Update cached type information 1135 // if( _method && _method->constants() ) 1136 // Type::update_loaded_types(_method, _method->constants()); 1137 1138 // Init alias_type map. 1139 if (!_do_escape_analysis && aliaslevel == 3) 1140 aliaslevel = 2; // No unique types without escape analysis 1141 _AliasLevel = aliaslevel; 1142 const int grow_ats = 16; 1143 _max_alias_types = grow_ats; 1144 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1145 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1146 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1147 { 1148 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1149 } 1150 // Initialize the first few types. 1151 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1152 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1153 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1154 _num_alias_types = AliasIdxRaw+1; 1155 // Zero out the alias type cache. 1156 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1157 // A NULL adr_type hits in the cache right away. Preload the right answer. 1158 probe_alias_cache(NULL)->_index = AliasIdxTop; 1159 1160 _intrinsics = NULL; 1161 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1162 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1163 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1164 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1165 register_library_intrinsics(); 1166 } 1167 1168 //---------------------------init_start---------------------------------------- 1169 // Install the StartNode on this compile object. 1170 void Compile::init_start(StartNode* s) { 1171 if (failing()) 1172 return; // already failing 1173 assert(s == start(), ""); 1174 } 1175 1176 StartNode* Compile::start() const { 1177 assert(!failing(), ""); 1178 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1179 Node* start = root()->fast_out(i); 1180 if( start->is_Start() ) 1181 return start->as_Start(); 1182 } 1183 fatal("Did not find Start node!"); 1184 return NULL; 1185 } 1186 1187 //-------------------------------immutable_memory------------------------------------- 1188 // Access immutable memory 1189 Node* Compile::immutable_memory() { 1190 if (_immutable_memory != NULL) { 1191 return _immutable_memory; 1192 } 1193 StartNode* s = start(); 1194 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1195 Node *p = s->fast_out(i); 1196 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1197 _immutable_memory = p; 1198 return _immutable_memory; 1199 } 1200 } 1201 ShouldNotReachHere(); 1202 return NULL; 1203 } 1204 1205 //----------------------set_cached_top_node------------------------------------ 1206 // Install the cached top node, and make sure Node::is_top works correctly. 1207 void Compile::set_cached_top_node(Node* tn) { 1208 if (tn != NULL) verify_top(tn); 1209 Node* old_top = _top; 1210 _top = tn; 1211 // Calling Node::setup_is_top allows the nodes the chance to adjust 1212 // their _out arrays. 1213 if (_top != NULL) _top->setup_is_top(); 1214 if (old_top != NULL) old_top->setup_is_top(); 1215 assert(_top == NULL || top()->is_top(), ""); 1216 } 1217 1218 #ifdef ASSERT 1219 uint Compile::count_live_nodes_by_graph_walk() { 1220 Unique_Node_List useful(comp_arena()); 1221 // Get useful node list by walking the graph. 1222 identify_useful_nodes(useful); 1223 return useful.size(); 1224 } 1225 1226 void Compile::print_missing_nodes() { 1227 1228 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1229 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1230 return; 1231 } 1232 1233 // This is an expensive function. It is executed only when the user 1234 // specifies VerifyIdealNodeCount option or otherwise knows the 1235 // additional work that needs to be done to identify reachable nodes 1236 // by walking the flow graph and find the missing ones using 1237 // _dead_node_list. 1238 1239 Unique_Node_List useful(comp_arena()); 1240 // Get useful node list by walking the graph. 1241 identify_useful_nodes(useful); 1242 1243 uint l_nodes = C->live_nodes(); 1244 uint l_nodes_by_walk = useful.size(); 1245 1246 if (l_nodes != l_nodes_by_walk) { 1247 if (_log != NULL) { 1248 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1249 _log->stamp(); 1250 _log->end_head(); 1251 } 1252 VectorSet& useful_member_set = useful.member_set(); 1253 int last_idx = l_nodes_by_walk; 1254 for (int i = 0; i < last_idx; i++) { 1255 if (useful_member_set.test(i)) { 1256 if (_dead_node_list.test(i)) { 1257 if (_log != NULL) { 1258 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1259 } 1260 if (PrintIdealNodeCount) { 1261 // Print the log message to tty 1262 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1263 useful.at(i)->dump(); 1264 } 1265 } 1266 } 1267 else if (! _dead_node_list.test(i)) { 1268 if (_log != NULL) { 1269 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1270 } 1271 if (PrintIdealNodeCount) { 1272 // Print the log message to tty 1273 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1274 } 1275 } 1276 } 1277 if (_log != NULL) { 1278 _log->tail("mismatched_nodes"); 1279 } 1280 } 1281 } 1282 #endif 1283 1284 #ifndef PRODUCT 1285 void Compile::verify_top(Node* tn) const { 1286 if (tn != NULL) { 1287 assert(tn->is_Con(), "top node must be a constant"); 1288 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1289 assert(tn->in(0) != NULL, "must have live top node"); 1290 } 1291 } 1292 #endif 1293 1294 1295 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1296 1297 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1298 guarantee(arr != NULL, ""); 1299 int num_blocks = arr->length(); 1300 if (grow_by < num_blocks) grow_by = num_blocks; 1301 int num_notes = grow_by * _node_notes_block_size; 1302 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1303 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1304 while (num_notes > 0) { 1305 arr->append(notes); 1306 notes += _node_notes_block_size; 1307 num_notes -= _node_notes_block_size; 1308 } 1309 assert(num_notes == 0, "exact multiple, please"); 1310 } 1311 1312 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1313 if (source == NULL || dest == NULL) return false; 1314 1315 if (dest->is_Con()) 1316 return false; // Do not push debug info onto constants. 1317 1318 #ifdef ASSERT 1319 // Leave a bread crumb trail pointing to the original node: 1320 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1321 dest->set_debug_orig(source); 1322 } 1323 #endif 1324 1325 if (node_note_array() == NULL) 1326 return false; // Not collecting any notes now. 1327 1328 // This is a copy onto a pre-existing node, which may already have notes. 1329 // If both nodes have notes, do not overwrite any pre-existing notes. 1330 Node_Notes* source_notes = node_notes_at(source->_idx); 1331 if (source_notes == NULL || source_notes->is_clear()) return false; 1332 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1333 if (dest_notes == NULL || dest_notes->is_clear()) { 1334 return set_node_notes_at(dest->_idx, source_notes); 1335 } 1336 1337 Node_Notes merged_notes = (*source_notes); 1338 // The order of operations here ensures that dest notes will win... 1339 merged_notes.update_from(dest_notes); 1340 return set_node_notes_at(dest->_idx, &merged_notes); 1341 } 1342 1343 1344 //--------------------------allow_range_check_smearing------------------------- 1345 // Gating condition for coalescing similar range checks. 1346 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1347 // single covering check that is at least as strong as any of them. 1348 // If the optimization succeeds, the simplified (strengthened) range check 1349 // will always succeed. If it fails, we will deopt, and then give up 1350 // on the optimization. 1351 bool Compile::allow_range_check_smearing() const { 1352 // If this method has already thrown a range-check, 1353 // assume it was because we already tried range smearing 1354 // and it failed. 1355 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1356 return !already_trapped; 1357 } 1358 1359 1360 //------------------------------flatten_alias_type----------------------------- 1361 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1362 int offset = tj->offset(); 1363 TypePtr::PTR ptr = tj->ptr(); 1364 1365 // Known instance (scalarizable allocation) alias only with itself. 1366 bool is_known_inst = tj->isa_oopptr() != NULL && 1367 tj->is_oopptr()->is_known_instance(); 1368 1369 // Process weird unsafe references. 1370 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1371 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1372 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1373 tj = TypeOopPtr::BOTTOM; 1374 ptr = tj->ptr(); 1375 offset = tj->offset(); 1376 } 1377 1378 // Array pointers need some flattening 1379 const TypeAryPtr *ta = tj->isa_aryptr(); 1380 if (ta && ta->is_stable()) { 1381 // Erase stability property for alias analysis. 1382 tj = ta = ta->cast_to_stable(false); 1383 } 1384 if( ta && is_known_inst ) { 1385 if ( offset != Type::OffsetBot && 1386 offset > arrayOopDesc::length_offset_in_bytes() ) { 1387 offset = Type::OffsetBot; // Flatten constant access into array body only 1388 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1389 } 1390 } else if( ta && _AliasLevel >= 2 ) { 1391 // For arrays indexed by constant indices, we flatten the alias 1392 // space to include all of the array body. Only the header, klass 1393 // and array length can be accessed un-aliased. 1394 if( offset != Type::OffsetBot ) { 1395 if( ta->const_oop() ) { // MethodData* or Method* 1396 offset = Type::OffsetBot; // Flatten constant access into array body 1397 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1398 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1399 // range is OK as-is. 1400 tj = ta = TypeAryPtr::RANGE; 1401 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1402 tj = TypeInstPtr::KLASS; // all klass loads look alike 1403 ta = TypeAryPtr::RANGE; // generic ignored junk 1404 ptr = TypePtr::BotPTR; 1405 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1406 tj = TypeInstPtr::MARK; 1407 ta = TypeAryPtr::RANGE; // generic ignored junk 1408 ptr = TypePtr::BotPTR; 1409 } else { // Random constant offset into array body 1410 offset = Type::OffsetBot; // Flatten constant access into array body 1411 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1412 } 1413 } 1414 // Arrays of fixed size alias with arrays of unknown size. 1415 if (ta->size() != TypeInt::POS) { 1416 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1418 } 1419 // Arrays of known objects become arrays of unknown objects. 1420 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1421 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1422 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1423 } 1424 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1425 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1426 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1427 } 1428 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1429 // cannot be distinguished by bytecode alone. 1430 if (ta->elem() == TypeInt::BOOL) { 1431 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1432 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1433 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1434 } 1435 // During the 2nd round of IterGVN, NotNull castings are removed. 1436 // Make sure the Bottom and NotNull variants alias the same. 1437 // Also, make sure exact and non-exact variants alias the same. 1438 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1439 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1440 } 1441 } 1442 1443 // Oop pointers need some flattening 1444 const TypeInstPtr *to = tj->isa_instptr(); 1445 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1446 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1447 if( ptr == TypePtr::Constant ) { 1448 if (to->klass() != ciEnv::current()->Class_klass() || 1449 offset < k->size_helper() * wordSize) { 1450 // No constant oop pointers (such as Strings); they alias with 1451 // unknown strings. 1452 assert(!is_known_inst, "not scalarizable allocation"); 1453 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1454 } 1455 } else if( is_known_inst ) { 1456 tj = to; // Keep NotNull and klass_is_exact for instance type 1457 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1458 // During the 2nd round of IterGVN, NotNull castings are removed. 1459 // Make sure the Bottom and NotNull variants alias the same. 1460 // Also, make sure exact and non-exact variants alias the same. 1461 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1462 } 1463 if (to->speculative() != NULL) { 1464 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1465 } 1466 // Canonicalize the holder of this field 1467 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1468 // First handle header references such as a LoadKlassNode, even if the 1469 // object's klass is unloaded at compile time (4965979). 1470 if (!is_known_inst) { // Do it only for non-instance types 1471 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1472 } 1473 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1474 // Static fields are in the space above the normal instance 1475 // fields in the java.lang.Class instance. 1476 if (to->klass() != ciEnv::current()->Class_klass()) { 1477 to = NULL; 1478 tj = TypeOopPtr::BOTTOM; 1479 offset = tj->offset(); 1480 } 1481 } else { 1482 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1483 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1484 if( is_known_inst ) { 1485 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1486 } else { 1487 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1488 } 1489 } 1490 } 1491 } 1492 1493 // Klass pointers to object array klasses need some flattening 1494 const TypeKlassPtr *tk = tj->isa_klassptr(); 1495 if( tk ) { 1496 // If we are referencing a field within a Klass, we need 1497 // to assume the worst case of an Object. Both exact and 1498 // inexact types must flatten to the same alias class so 1499 // use NotNull as the PTR. 1500 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1501 1502 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1503 TypeKlassPtr::OBJECT->klass(), 1504 offset); 1505 } 1506 1507 ciKlass* klass = tk->klass(); 1508 if( klass->is_obj_array_klass() ) { 1509 ciKlass* k = TypeAryPtr::OOPS->klass(); 1510 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1511 k = TypeInstPtr::BOTTOM->klass(); 1512 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1513 } 1514 1515 // Check for precise loads from the primary supertype array and force them 1516 // to the supertype cache alias index. Check for generic array loads from 1517 // the primary supertype array and also force them to the supertype cache 1518 // alias index. Since the same load can reach both, we need to merge 1519 // these 2 disparate memories into the same alias class. Since the 1520 // primary supertype array is read-only, there's no chance of confusion 1521 // where we bypass an array load and an array store. 1522 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1523 if (offset == Type::OffsetBot || 1524 (offset >= primary_supers_offset && 1525 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1526 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1527 offset = in_bytes(Klass::secondary_super_cache_offset()); 1528 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1529 } 1530 } 1531 1532 // Flatten all Raw pointers together. 1533 if (tj->base() == Type::RawPtr) 1534 tj = TypeRawPtr::BOTTOM; 1535 1536 if (tj->base() == Type::AnyPtr) 1537 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1538 1539 // Flatten all to bottom for now 1540 switch( _AliasLevel ) { 1541 case 0: 1542 tj = TypePtr::BOTTOM; 1543 break; 1544 case 1: // Flatten to: oop, static, field or array 1545 switch (tj->base()) { 1546 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1547 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1548 case Type::AryPtr: // do not distinguish arrays at all 1549 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1550 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1551 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1552 default: ShouldNotReachHere(); 1553 } 1554 break; 1555 case 2: // No collapsing at level 2; keep all splits 1556 case 3: // No collapsing at level 3; keep all splits 1557 break; 1558 default: 1559 Unimplemented(); 1560 } 1561 1562 offset = tj->offset(); 1563 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1564 1565 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1566 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1567 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1568 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1569 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1570 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1571 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1572 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1573 assert( tj->ptr() != TypePtr::TopPTR && 1574 tj->ptr() != TypePtr::AnyNull && 1575 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1576 // assert( tj->ptr() != TypePtr::Constant || 1577 // tj->base() == Type::RawPtr || 1578 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1579 1580 return tj; 1581 } 1582 1583 void Compile::AliasType::Init(int i, const TypePtr* at) { 1584 _index = i; 1585 _adr_type = at; 1586 _field = NULL; 1587 _element = NULL; 1588 _is_rewritable = true; // default 1589 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1590 if (atoop != NULL && atoop->is_known_instance()) { 1591 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1592 _general_index = Compile::current()->get_alias_index(gt); 1593 } else { 1594 _general_index = 0; 1595 } 1596 } 1597 1598 BasicType Compile::AliasType::basic_type() const { 1599 if (element() != NULL) { 1600 const Type* element = adr_type()->is_aryptr()->elem(); 1601 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1602 } if (field() != NULL) { 1603 return field()->layout_type(); 1604 } else { 1605 return T_ILLEGAL; // unknown 1606 } 1607 } 1608 1609 //---------------------------------print_on------------------------------------ 1610 #ifndef PRODUCT 1611 void Compile::AliasType::print_on(outputStream* st) { 1612 if (index() < 10) 1613 st->print("@ <%d> ", index()); 1614 else st->print("@ <%d>", index()); 1615 st->print(is_rewritable() ? " " : " RO"); 1616 int offset = adr_type()->offset(); 1617 if (offset == Type::OffsetBot) 1618 st->print(" +any"); 1619 else st->print(" +%-3d", offset); 1620 st->print(" in "); 1621 adr_type()->dump_on(st); 1622 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1623 if (field() != NULL && tjp) { 1624 if (tjp->klass() != field()->holder() || 1625 tjp->offset() != field()->offset_in_bytes()) { 1626 st->print(" != "); 1627 field()->print(); 1628 st->print(" ***"); 1629 } 1630 } 1631 } 1632 1633 void print_alias_types() { 1634 Compile* C = Compile::current(); 1635 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1636 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1637 C->alias_type(idx)->print_on(tty); 1638 tty->cr(); 1639 } 1640 } 1641 #endif 1642 1643 1644 //----------------------------probe_alias_cache-------------------------------- 1645 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1646 intptr_t key = (intptr_t) adr_type; 1647 key ^= key >> logAliasCacheSize; 1648 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1649 } 1650 1651 1652 //-----------------------------grow_alias_types-------------------------------- 1653 void Compile::grow_alias_types() { 1654 const int old_ats = _max_alias_types; // how many before? 1655 const int new_ats = old_ats; // how many more? 1656 const int grow_ats = old_ats+new_ats; // how many now? 1657 _max_alias_types = grow_ats; 1658 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1659 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1660 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1661 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1662 } 1663 1664 1665 //--------------------------------find_alias_type------------------------------ 1666 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1667 if (_AliasLevel == 0) 1668 return alias_type(AliasIdxBot); 1669 1670 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1671 if (ace->_adr_type == adr_type) { 1672 return alias_type(ace->_index); 1673 } 1674 1675 // Handle special cases. 1676 if (adr_type == NULL) return alias_type(AliasIdxTop); 1677 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1678 1679 // Do it the slow way. 1680 const TypePtr* flat = flatten_alias_type(adr_type); 1681 1682 #ifdef ASSERT 1683 assert(flat == flatten_alias_type(flat), "idempotent"); 1684 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1685 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1686 const TypeOopPtr* foop = flat->is_oopptr(); 1687 // Scalarizable allocations have exact klass always. 1688 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1689 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1690 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1691 } 1692 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1693 #endif 1694 1695 int idx = AliasIdxTop; 1696 for (int i = 0; i < num_alias_types(); i++) { 1697 if (alias_type(i)->adr_type() == flat) { 1698 idx = i; 1699 break; 1700 } 1701 } 1702 1703 if (idx == AliasIdxTop) { 1704 if (no_create) return NULL; 1705 // Grow the array if necessary. 1706 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1707 // Add a new alias type. 1708 idx = _num_alias_types++; 1709 _alias_types[idx]->Init(idx, flat); 1710 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1711 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1712 if (flat->isa_instptr()) { 1713 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1714 && flat->is_instptr()->klass() == env()->Class_klass()) 1715 alias_type(idx)->set_rewritable(false); 1716 } 1717 if (flat->isa_aryptr()) { 1718 #ifdef ASSERT 1719 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1720 // (T_BYTE has the weakest alignment and size restrictions...) 1721 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1722 #endif 1723 if (flat->offset() == TypePtr::OffsetBot) { 1724 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1725 } 1726 } 1727 if (flat->isa_klassptr()) { 1728 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1729 alias_type(idx)->set_rewritable(false); 1730 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1731 alias_type(idx)->set_rewritable(false); 1732 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1733 alias_type(idx)->set_rewritable(false); 1734 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1735 alias_type(idx)->set_rewritable(false); 1736 } 1737 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1738 // but the base pointer type is not distinctive enough to identify 1739 // references into JavaThread.) 1740 1741 // Check for final fields. 1742 const TypeInstPtr* tinst = flat->isa_instptr(); 1743 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1744 ciField* field; 1745 if (tinst->const_oop() != NULL && 1746 tinst->klass() == ciEnv::current()->Class_klass() && 1747 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1748 // static field 1749 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1750 field = k->get_field_by_offset(tinst->offset(), true); 1751 } else { 1752 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1753 field = k->get_field_by_offset(tinst->offset(), false); 1754 } 1755 assert(field == NULL || 1756 original_field == NULL || 1757 (field->holder() == original_field->holder() && 1758 field->offset() == original_field->offset() && 1759 field->is_static() == original_field->is_static()), "wrong field?"); 1760 // Set field() and is_rewritable() attributes. 1761 if (field != NULL) alias_type(idx)->set_field(field); 1762 } 1763 } 1764 1765 // Fill the cache for next time. 1766 ace->_adr_type = adr_type; 1767 ace->_index = idx; 1768 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1769 1770 // Might as well try to fill the cache for the flattened version, too. 1771 AliasCacheEntry* face = probe_alias_cache(flat); 1772 if (face->_adr_type == NULL) { 1773 face->_adr_type = flat; 1774 face->_index = idx; 1775 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1776 } 1777 1778 return alias_type(idx); 1779 } 1780 1781 1782 Compile::AliasType* Compile::alias_type(ciField* field) { 1783 const TypeOopPtr* t; 1784 if (field->is_static()) 1785 t = TypeInstPtr::make(field->holder()->java_mirror()); 1786 else 1787 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1788 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1789 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1790 return atp; 1791 } 1792 1793 1794 //------------------------------have_alias_type-------------------------------- 1795 bool Compile::have_alias_type(const TypePtr* adr_type) { 1796 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1797 if (ace->_adr_type == adr_type) { 1798 return true; 1799 } 1800 1801 // Handle special cases. 1802 if (adr_type == NULL) return true; 1803 if (adr_type == TypePtr::BOTTOM) return true; 1804 1805 return find_alias_type(adr_type, true, NULL) != NULL; 1806 } 1807 1808 //-----------------------------must_alias-------------------------------------- 1809 // True if all values of the given address type are in the given alias category. 1810 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1811 if (alias_idx == AliasIdxBot) return true; // the universal category 1812 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1813 if (alias_idx == AliasIdxTop) return false; // the empty category 1814 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1815 1816 // the only remaining possible overlap is identity 1817 int adr_idx = get_alias_index(adr_type); 1818 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1819 assert(adr_idx == alias_idx || 1820 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1821 && adr_type != TypeOopPtr::BOTTOM), 1822 "should not be testing for overlap with an unsafe pointer"); 1823 return adr_idx == alias_idx; 1824 } 1825 1826 //------------------------------can_alias-------------------------------------- 1827 // True if any values of the given address type are in the given alias category. 1828 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1829 if (alias_idx == AliasIdxTop) return false; // the empty category 1830 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1831 if (alias_idx == AliasIdxBot) return true; // the universal category 1832 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1833 1834 // the only remaining possible overlap is identity 1835 int adr_idx = get_alias_index(adr_type); 1836 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1837 return adr_idx == alias_idx; 1838 } 1839 1840 1841 1842 //---------------------------pop_warm_call------------------------------------- 1843 WarmCallInfo* Compile::pop_warm_call() { 1844 WarmCallInfo* wci = _warm_calls; 1845 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1846 return wci; 1847 } 1848 1849 //----------------------------Inline_Warm-------------------------------------- 1850 int Compile::Inline_Warm() { 1851 // If there is room, try to inline some more warm call sites. 1852 // %%% Do a graph index compaction pass when we think we're out of space? 1853 if (!InlineWarmCalls) return 0; 1854 1855 int calls_made_hot = 0; 1856 int room_to_grow = NodeCountInliningCutoff - unique(); 1857 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1858 int amount_grown = 0; 1859 WarmCallInfo* call; 1860 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1861 int est_size = (int)call->size(); 1862 if (est_size > (room_to_grow - amount_grown)) { 1863 // This one won't fit anyway. Get rid of it. 1864 call->make_cold(); 1865 continue; 1866 } 1867 call->make_hot(); 1868 calls_made_hot++; 1869 amount_grown += est_size; 1870 amount_to_grow -= est_size; 1871 } 1872 1873 if (calls_made_hot > 0) set_major_progress(); 1874 return calls_made_hot; 1875 } 1876 1877 1878 //----------------------------Finish_Warm-------------------------------------- 1879 void Compile::Finish_Warm() { 1880 if (!InlineWarmCalls) return; 1881 if (failing()) return; 1882 if (warm_calls() == NULL) return; 1883 1884 // Clean up loose ends, if we are out of space for inlining. 1885 WarmCallInfo* call; 1886 while ((call = pop_warm_call()) != NULL) { 1887 call->make_cold(); 1888 } 1889 } 1890 1891 //---------------------cleanup_loop_predicates----------------------- 1892 // Remove the opaque nodes that protect the predicates so that all unused 1893 // checks and uncommon_traps will be eliminated from the ideal graph 1894 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1895 if (predicate_count()==0) return; 1896 for (int i = predicate_count(); i > 0; i--) { 1897 Node * n = predicate_opaque1_node(i-1); 1898 assert(n->Opcode() == Op_Opaque1, "must be"); 1899 igvn.replace_node(n, n->in(1)); 1900 } 1901 assert(predicate_count()==0, "should be clean!"); 1902 } 1903 1904 void Compile::add_range_check_cast(Node* n) { 1905 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1906 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1907 _range_check_casts->append(n); 1908 } 1909 1910 // Remove all range check dependent CastIINodes. 1911 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1912 for (int i = range_check_cast_count(); i > 0; i--) { 1913 Node* cast = range_check_cast_node(i-1); 1914 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1915 igvn.replace_node(cast, cast->in(1)); 1916 } 1917 assert(range_check_cast_count() == 0, "should be empty"); 1918 } 1919 1920 // StringOpts and late inlining of string methods 1921 void Compile::inline_string_calls(bool parse_time) { 1922 { 1923 // remove useless nodes to make the usage analysis simpler 1924 ResourceMark rm; 1925 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1926 } 1927 1928 { 1929 ResourceMark rm; 1930 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1931 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1932 print_method(PHASE_AFTER_STRINGOPTS, 3); 1933 } 1934 1935 // now inline anything that we skipped the first time around 1936 if (!parse_time) { 1937 _late_inlines_pos = _late_inlines.length(); 1938 } 1939 1940 while (_string_late_inlines.length() > 0) { 1941 CallGenerator* cg = _string_late_inlines.pop(); 1942 cg->do_late_inline(); 1943 if (failing()) return; 1944 } 1945 _string_late_inlines.trunc_to(0); 1946 } 1947 1948 // Late inlining of boxing methods 1949 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1950 if (_boxing_late_inlines.length() > 0) { 1951 assert(has_boxed_value(), "inconsistent"); 1952 1953 PhaseGVN* gvn = initial_gvn(); 1954 set_inlining_incrementally(true); 1955 1956 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1957 for_igvn()->clear(); 1958 gvn->replace_with(&igvn); 1959 1960 _late_inlines_pos = _late_inlines.length(); 1961 1962 while (_boxing_late_inlines.length() > 0) { 1963 CallGenerator* cg = _boxing_late_inlines.pop(); 1964 cg->do_late_inline(); 1965 if (failing()) return; 1966 } 1967 _boxing_late_inlines.trunc_to(0); 1968 1969 { 1970 ResourceMark rm; 1971 PhaseRemoveUseless pru(gvn, for_igvn()); 1972 } 1973 1974 igvn = PhaseIterGVN(gvn); 1975 igvn.optimize(); 1976 1977 set_inlining_progress(false); 1978 set_inlining_incrementally(false); 1979 } 1980 } 1981 1982 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1983 assert(IncrementalInline, "incremental inlining should be on"); 1984 PhaseGVN* gvn = initial_gvn(); 1985 1986 set_inlining_progress(false); 1987 for_igvn()->clear(); 1988 gvn->replace_with(&igvn); 1989 1990 int i = 0; 1991 1992 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1993 CallGenerator* cg = _late_inlines.at(i); 1994 _late_inlines_pos = i+1; 1995 cg->do_late_inline(); 1996 if (failing()) return; 1997 } 1998 int j = 0; 1999 for (; i < _late_inlines.length(); i++, j++) { 2000 _late_inlines.at_put(j, _late_inlines.at(i)); 2001 } 2002 _late_inlines.trunc_to(j); 2003 2004 { 2005 ResourceMark rm; 2006 PhaseRemoveUseless pru(gvn, for_igvn()); 2007 } 2008 2009 igvn = PhaseIterGVN(gvn); 2010 } 2011 2012 // Perform incremental inlining until bound on number of live nodes is reached 2013 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2014 PhaseGVN* gvn = initial_gvn(); 2015 2016 set_inlining_incrementally(true); 2017 set_inlining_progress(true); 2018 uint low_live_nodes = 0; 2019 2020 while(inlining_progress() && _late_inlines.length() > 0) { 2021 2022 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2023 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2024 // PhaseIdealLoop is expensive so we only try it once we are 2025 // out of live nodes and we only try it again if the previous 2026 // helped got the number of nodes down significantly 2027 PhaseIdealLoop ideal_loop( igvn, false, true ); 2028 if (failing()) return; 2029 low_live_nodes = live_nodes(); 2030 _major_progress = true; 2031 } 2032 2033 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2034 break; 2035 } 2036 } 2037 2038 inline_incrementally_one(igvn); 2039 2040 if (failing()) return; 2041 2042 igvn.optimize(); 2043 2044 if (failing()) return; 2045 } 2046 2047 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2048 2049 if (_string_late_inlines.length() > 0) { 2050 assert(has_stringbuilder(), "inconsistent"); 2051 for_igvn()->clear(); 2052 initial_gvn()->replace_with(&igvn); 2053 2054 inline_string_calls(false); 2055 2056 if (failing()) return; 2057 2058 { 2059 ResourceMark rm; 2060 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2061 } 2062 2063 igvn = PhaseIterGVN(gvn); 2064 2065 igvn.optimize(); 2066 } 2067 2068 set_inlining_incrementally(false); 2069 } 2070 2071 2072 //------------------------------Optimize--------------------------------------- 2073 // Given a graph, optimize it. 2074 void Compile::Optimize() { 2075 TracePhase t1("optimizer", &_t_optimizer, true); 2076 2077 #ifndef PRODUCT 2078 if (env()->break_at_compile()) { 2079 BREAKPOINT; 2080 } 2081 2082 #endif 2083 2084 ResourceMark rm; 2085 int loop_opts_cnt; 2086 2087 NOT_PRODUCT( verify_graph_edges(); ) 2088 2089 print_method(PHASE_AFTER_PARSING); 2090 2091 { 2092 // Iterative Global Value Numbering, including ideal transforms 2093 // Initialize IterGVN with types and values from parse-time GVN 2094 PhaseIterGVN igvn(initial_gvn()); 2095 { 2096 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2097 igvn.optimize(); 2098 } 2099 2100 print_method(PHASE_ITER_GVN1, 2); 2101 2102 if (failing()) return; 2103 2104 { 2105 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2106 inline_incrementally(igvn); 2107 } 2108 2109 print_method(PHASE_INCREMENTAL_INLINE, 2); 2110 2111 if (failing()) return; 2112 2113 if (eliminate_boxing()) { 2114 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2115 // Inline valueOf() methods now. 2116 inline_boxing_calls(igvn); 2117 2118 if (AlwaysIncrementalInline) { 2119 inline_incrementally(igvn); 2120 } 2121 2122 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2123 2124 if (failing()) return; 2125 } 2126 2127 // Remove the speculative part of types and clean up the graph from 2128 // the extra CastPP nodes whose only purpose is to carry them. Do 2129 // that early so that optimizations are not disrupted by the extra 2130 // CastPP nodes. 2131 remove_speculative_types(igvn); 2132 2133 // No more new expensive nodes will be added to the list from here 2134 // so keep only the actual candidates for optimizations. 2135 cleanup_expensive_nodes(igvn); 2136 2137 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2138 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);) 2139 initial_gvn()->replace_with(&igvn); 2140 for_igvn()->clear(); 2141 Unique_Node_List new_worklist(C->comp_arena()); 2142 { 2143 ResourceMark rm; 2144 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2145 } 2146 set_for_igvn(&new_worklist); 2147 igvn = PhaseIterGVN(initial_gvn()); 2148 igvn.optimize(); 2149 } 2150 2151 // Perform escape analysis 2152 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2153 if (has_loops()) { 2154 // Cleanup graph (remove dead nodes). 2155 TracePhase t2("idealLoop", &_t_idealLoop, true); 2156 PhaseIdealLoop ideal_loop( igvn, false, true ); 2157 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2158 if (failing()) return; 2159 } 2160 ConnectionGraph::do_analysis(this, &igvn); 2161 2162 if (failing()) return; 2163 2164 // Optimize out fields loads from scalar replaceable allocations. 2165 igvn.optimize(); 2166 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2167 2168 if (failing()) return; 2169 2170 if (congraph() != NULL && macro_count() > 0) { 2171 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2172 PhaseMacroExpand mexp(igvn); 2173 mexp.eliminate_macro_nodes(); 2174 igvn.set_delay_transform(false); 2175 2176 igvn.optimize(); 2177 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2178 2179 if (failing()) return; 2180 } 2181 } 2182 2183 // Loop transforms on the ideal graph. Range Check Elimination, 2184 // peeling, unrolling, etc. 2185 2186 // Set loop opts counter 2187 loop_opts_cnt = num_loop_opts(); 2188 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2189 { 2190 TracePhase t2("idealLoop", &_t_idealLoop, true); 2191 PhaseIdealLoop ideal_loop( igvn, true ); 2192 loop_opts_cnt--; 2193 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2194 if (failing()) return; 2195 } 2196 // Loop opts pass if partial peeling occurred in previous pass 2197 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2198 TracePhase t3("idealLoop", &_t_idealLoop, true); 2199 PhaseIdealLoop ideal_loop( igvn, false ); 2200 loop_opts_cnt--; 2201 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2202 if (failing()) return; 2203 } 2204 // Loop opts pass for loop-unrolling before CCP 2205 if(major_progress() && (loop_opts_cnt > 0)) { 2206 TracePhase t4("idealLoop", &_t_idealLoop, true); 2207 PhaseIdealLoop ideal_loop( igvn, false ); 2208 loop_opts_cnt--; 2209 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2210 } 2211 if (!failing()) { 2212 // Verify that last round of loop opts produced a valid graph 2213 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2214 PhaseIdealLoop::verify(igvn); 2215 } 2216 } 2217 if (failing()) return; 2218 2219 // Conditional Constant Propagation; 2220 PhaseCCP ccp( &igvn ); 2221 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2222 { 2223 TracePhase t2("ccp", &_t_ccp, true); 2224 ccp.do_transform(); 2225 } 2226 print_method(PHASE_CPP1, 2); 2227 2228 assert( true, "Break here to ccp.dump_old2new_map()"); 2229 2230 // Iterative Global Value Numbering, including ideal transforms 2231 { 2232 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2233 igvn = ccp; 2234 igvn.optimize(); 2235 } 2236 2237 print_method(PHASE_ITER_GVN2, 2); 2238 2239 if (failing()) return; 2240 2241 // Loop transforms on the ideal graph. Range Check Elimination, 2242 // peeling, unrolling, etc. 2243 if(loop_opts_cnt > 0) { 2244 debug_only( int cnt = 0; ); 2245 while(major_progress() && (loop_opts_cnt > 0)) { 2246 TracePhase t2("idealLoop", &_t_idealLoop, true); 2247 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2248 PhaseIdealLoop ideal_loop( igvn, true); 2249 loop_opts_cnt--; 2250 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2251 if (failing()) return; 2252 } 2253 } 2254 2255 { 2256 // Verify that all previous optimizations produced a valid graph 2257 // at least to this point, even if no loop optimizations were done. 2258 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2259 PhaseIdealLoop::verify(igvn); 2260 } 2261 2262 if (range_check_cast_count() > 0) { 2263 // No more loop optimizations. Remove all range check dependent CastIINodes. 2264 C->remove_range_check_casts(igvn); 2265 igvn.optimize(); 2266 } 2267 2268 { 2269 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2270 PhaseMacroExpand mex(igvn); 2271 if (mex.expand_macro_nodes()) { 2272 assert(failing(), "must bail out w/ explicit message"); 2273 return; 2274 } 2275 } 2276 2277 } // (End scope of igvn; run destructor if necessary for asserts.) 2278 2279 dump_inlining(); 2280 // A method with only infinite loops has no edges entering loops from root 2281 { 2282 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2283 if (final_graph_reshaping()) { 2284 assert(failing(), "must bail out w/ explicit message"); 2285 return; 2286 } 2287 } 2288 2289 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2290 } 2291 2292 2293 //------------------------------Code_Gen--------------------------------------- 2294 // Given a graph, generate code for it 2295 void Compile::Code_Gen() { 2296 if (failing()) { 2297 return; 2298 } 2299 2300 // Perform instruction selection. You might think we could reclaim Matcher 2301 // memory PDQ, but actually the Matcher is used in generating spill code. 2302 // Internals of the Matcher (including some VectorSets) must remain live 2303 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2304 // set a bit in reclaimed memory. 2305 2306 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2307 // nodes. Mapping is only valid at the root of each matched subtree. 2308 NOT_PRODUCT( verify_graph_edges(); ) 2309 2310 Matcher matcher; 2311 _matcher = &matcher; 2312 { 2313 TracePhase t2("matcher", &_t_matcher, true); 2314 matcher.match(); 2315 } 2316 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2317 // nodes. Mapping is only valid at the root of each matched subtree. 2318 NOT_PRODUCT( verify_graph_edges(); ) 2319 2320 // If you have too many nodes, or if matching has failed, bail out 2321 check_node_count(0, "out of nodes matching instructions"); 2322 if (failing()) { 2323 return; 2324 } 2325 2326 // Build a proper-looking CFG 2327 PhaseCFG cfg(node_arena(), root(), matcher); 2328 _cfg = &cfg; 2329 { 2330 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2331 bool success = cfg.do_global_code_motion(); 2332 if (!success) { 2333 return; 2334 } 2335 2336 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2337 NOT_PRODUCT( verify_graph_edges(); ) 2338 debug_only( cfg.verify(); ) 2339 } 2340 2341 PhaseChaitin regalloc(unique(), cfg, matcher); 2342 _regalloc = ®alloc; 2343 { 2344 TracePhase t2("regalloc", &_t_registerAllocation, true); 2345 // Perform register allocation. After Chaitin, use-def chains are 2346 // no longer accurate (at spill code) and so must be ignored. 2347 // Node->LRG->reg mappings are still accurate. 2348 _regalloc->Register_Allocate(); 2349 2350 // Bail out if the allocator builds too many nodes 2351 if (failing()) { 2352 return; 2353 } 2354 } 2355 2356 // Prior to register allocation we kept empty basic blocks in case the 2357 // the allocator needed a place to spill. After register allocation we 2358 // are not adding any new instructions. If any basic block is empty, we 2359 // can now safely remove it. 2360 { 2361 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2362 cfg.remove_empty_blocks(); 2363 if (do_freq_based_layout()) { 2364 PhaseBlockLayout layout(cfg); 2365 } else { 2366 cfg.set_loop_alignment(); 2367 } 2368 cfg.fixup_flow(); 2369 } 2370 2371 // Apply peephole optimizations 2372 if( OptoPeephole ) { 2373 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2374 PhasePeephole peep( _regalloc, cfg); 2375 peep.do_transform(); 2376 } 2377 2378 // Do late expand if CPU requires this. 2379 if (Matcher::require_postalloc_expand) { 2380 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2381 cfg.postalloc_expand(_regalloc); 2382 } 2383 2384 // Convert Nodes to instruction bits in a buffer 2385 { 2386 // %%%% workspace merge brought two timers together for one job 2387 TracePhase t2a("output", &_t_output, true); 2388 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2389 Output(); 2390 } 2391 2392 print_method(PHASE_FINAL_CODE); 2393 2394 // He's dead, Jim. 2395 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2396 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2397 } 2398 2399 2400 //------------------------------dump_asm--------------------------------------- 2401 // Dump formatted assembly 2402 #ifndef PRODUCT 2403 void Compile::dump_asm(int *pcs, uint pc_limit) { 2404 bool cut_short = false; 2405 tty->print_cr("#"); 2406 tty->print("# "); _tf->dump(); tty->cr(); 2407 tty->print_cr("#"); 2408 2409 // For all blocks 2410 int pc = 0x0; // Program counter 2411 char starts_bundle = ' '; 2412 _regalloc->dump_frame(); 2413 2414 Node *n = NULL; 2415 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2416 if (VMThread::should_terminate()) { 2417 cut_short = true; 2418 break; 2419 } 2420 Block* block = _cfg->get_block(i); 2421 if (block->is_connector() && !Verbose) { 2422 continue; 2423 } 2424 n = block->head(); 2425 if (pcs && n->_idx < pc_limit) { 2426 tty->print("%3.3x ", pcs[n->_idx]); 2427 } else { 2428 tty->print(" "); 2429 } 2430 block->dump_head(_cfg); 2431 if (block->is_connector()) { 2432 tty->print_cr(" # Empty connector block"); 2433 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2434 tty->print_cr(" # Block is sole successor of call"); 2435 } 2436 2437 // For all instructions 2438 Node *delay = NULL; 2439 for (uint j = 0; j < block->number_of_nodes(); j++) { 2440 if (VMThread::should_terminate()) { 2441 cut_short = true; 2442 break; 2443 } 2444 n = block->get_node(j); 2445 if (valid_bundle_info(n)) { 2446 Bundle* bundle = node_bundling(n); 2447 if (bundle->used_in_unconditional_delay()) { 2448 delay = n; 2449 continue; 2450 } 2451 if (bundle->starts_bundle()) { 2452 starts_bundle = '+'; 2453 } 2454 } 2455 2456 if (WizardMode) { 2457 n->dump(); 2458 } 2459 2460 if( !n->is_Region() && // Dont print in the Assembly 2461 !n->is_Phi() && // a few noisely useless nodes 2462 !n->is_Proj() && 2463 !n->is_MachTemp() && 2464 !n->is_SafePointScalarObject() && 2465 !n->is_Catch() && // Would be nice to print exception table targets 2466 !n->is_MergeMem() && // Not very interesting 2467 !n->is_top() && // Debug info table constants 2468 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2469 ) { 2470 if (pcs && n->_idx < pc_limit) 2471 tty->print("%3.3x", pcs[n->_idx]); 2472 else 2473 tty->print(" "); 2474 tty->print(" %c ", starts_bundle); 2475 starts_bundle = ' '; 2476 tty->print("\t"); 2477 n->format(_regalloc, tty); 2478 tty->cr(); 2479 } 2480 2481 // If we have an instruction with a delay slot, and have seen a delay, 2482 // then back up and print it 2483 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2484 assert(delay != NULL, "no unconditional delay instruction"); 2485 if (WizardMode) delay->dump(); 2486 2487 if (node_bundling(delay)->starts_bundle()) 2488 starts_bundle = '+'; 2489 if (pcs && n->_idx < pc_limit) 2490 tty->print("%3.3x", pcs[n->_idx]); 2491 else 2492 tty->print(" "); 2493 tty->print(" %c ", starts_bundle); 2494 starts_bundle = ' '; 2495 tty->print("\t"); 2496 delay->format(_regalloc, tty); 2497 tty->cr(); 2498 delay = NULL; 2499 } 2500 2501 // Dump the exception table as well 2502 if( n->is_Catch() && (Verbose || WizardMode) ) { 2503 // Print the exception table for this offset 2504 _handler_table.print_subtable_for(pc); 2505 } 2506 } 2507 2508 if (pcs && n->_idx < pc_limit) 2509 tty->print_cr("%3.3x", pcs[n->_idx]); 2510 else 2511 tty->cr(); 2512 2513 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2514 2515 } // End of per-block dump 2516 tty->cr(); 2517 2518 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2519 } 2520 #endif 2521 2522 //------------------------------Final_Reshape_Counts--------------------------- 2523 // This class defines counters to help identify when a method 2524 // may/must be executed using hardware with only 24-bit precision. 2525 struct Final_Reshape_Counts : public StackObj { 2526 int _call_count; // count non-inlined 'common' calls 2527 int _float_count; // count float ops requiring 24-bit precision 2528 int _double_count; // count double ops requiring more precision 2529 int _java_call_count; // count non-inlined 'java' calls 2530 int _inner_loop_count; // count loops which need alignment 2531 VectorSet _visited; // Visitation flags 2532 Node_List _tests; // Set of IfNodes & PCTableNodes 2533 2534 Final_Reshape_Counts() : 2535 _call_count(0), _float_count(0), _double_count(0), 2536 _java_call_count(0), _inner_loop_count(0), 2537 _visited( Thread::current()->resource_area() ) { } 2538 2539 void inc_call_count () { _call_count ++; } 2540 void inc_float_count () { _float_count ++; } 2541 void inc_double_count() { _double_count++; } 2542 void inc_java_call_count() { _java_call_count++; } 2543 void inc_inner_loop_count() { _inner_loop_count++; } 2544 2545 int get_call_count () const { return _call_count ; } 2546 int get_float_count () const { return _float_count ; } 2547 int get_double_count() const { return _double_count; } 2548 int get_java_call_count() const { return _java_call_count; } 2549 int get_inner_loop_count() const { return _inner_loop_count; } 2550 }; 2551 2552 #ifdef ASSERT 2553 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2554 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2555 // Make sure the offset goes inside the instance layout. 2556 return k->contains_field_offset(tp->offset()); 2557 // Note that OffsetBot and OffsetTop are very negative. 2558 } 2559 #endif 2560 2561 // Eliminate trivially redundant StoreCMs and accumulate their 2562 // precedence edges. 2563 void Compile::eliminate_redundant_card_marks(Node* n) { 2564 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2565 if (n->in(MemNode::Address)->outcnt() > 1) { 2566 // There are multiple users of the same address so it might be 2567 // possible to eliminate some of the StoreCMs 2568 Node* mem = n->in(MemNode::Memory); 2569 Node* adr = n->in(MemNode::Address); 2570 Node* val = n->in(MemNode::ValueIn); 2571 Node* prev = n; 2572 bool done = false; 2573 // Walk the chain of StoreCMs eliminating ones that match. As 2574 // long as it's a chain of single users then the optimization is 2575 // safe. Eliminating partially redundant StoreCMs would require 2576 // cloning copies down the other paths. 2577 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2578 if (adr == mem->in(MemNode::Address) && 2579 val == mem->in(MemNode::ValueIn)) { 2580 // redundant StoreCM 2581 if (mem->req() > MemNode::OopStore) { 2582 // Hasn't been processed by this code yet. 2583 n->add_prec(mem->in(MemNode::OopStore)); 2584 } else { 2585 // Already converted to precedence edge 2586 for (uint i = mem->req(); i < mem->len(); i++) { 2587 // Accumulate any precedence edges 2588 if (mem->in(i) != NULL) { 2589 n->add_prec(mem->in(i)); 2590 } 2591 } 2592 // Everything above this point has been processed. 2593 done = true; 2594 } 2595 // Eliminate the previous StoreCM 2596 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2597 assert(mem->outcnt() == 0, "should be dead"); 2598 mem->disconnect_inputs(NULL, this); 2599 } else { 2600 prev = mem; 2601 } 2602 mem = prev->in(MemNode::Memory); 2603 } 2604 } 2605 } 2606 2607 //------------------------------final_graph_reshaping_impl---------------------- 2608 // Implement items 1-5 from final_graph_reshaping below. 2609 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2610 2611 if ( n->outcnt() == 0 ) return; // dead node 2612 uint nop = n->Opcode(); 2613 2614 // Check for 2-input instruction with "last use" on right input. 2615 // Swap to left input. Implements item (2). 2616 if( n->req() == 3 && // two-input instruction 2617 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2618 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2619 n->in(2)->outcnt() == 1 &&// right use IS a last use 2620 !n->in(2)->is_Con() ) { // right use is not a constant 2621 // Check for commutative opcode 2622 switch( nop ) { 2623 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2624 case Op_MaxI: case Op_MinI: 2625 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2626 case Op_AndL: case Op_XorL: case Op_OrL: 2627 case Op_AndI: case Op_XorI: case Op_OrI: { 2628 // Move "last use" input to left by swapping inputs 2629 n->swap_edges(1, 2); 2630 break; 2631 } 2632 default: 2633 break; 2634 } 2635 } 2636 2637 #ifdef ASSERT 2638 if( n->is_Mem() ) { 2639 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2640 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2641 // oop will be recorded in oop map if load crosses safepoint 2642 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2643 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2644 "raw memory operations should have control edge"); 2645 } 2646 #endif 2647 // Count FPU ops and common calls, implements item (3) 2648 switch( nop ) { 2649 // Count all float operations that may use FPU 2650 case Op_AddF: 2651 case Op_SubF: 2652 case Op_MulF: 2653 case Op_DivF: 2654 case Op_NegF: 2655 case Op_ModF: 2656 case Op_ConvI2F: 2657 case Op_ConF: 2658 case Op_CmpF: 2659 case Op_CmpF3: 2660 // case Op_ConvL2F: // longs are split into 32-bit halves 2661 frc.inc_float_count(); 2662 break; 2663 2664 case Op_ConvF2D: 2665 case Op_ConvD2F: 2666 frc.inc_float_count(); 2667 frc.inc_double_count(); 2668 break; 2669 2670 // Count all double operations that may use FPU 2671 case Op_AddD: 2672 case Op_SubD: 2673 case Op_MulD: 2674 case Op_DivD: 2675 case Op_NegD: 2676 case Op_ModD: 2677 case Op_ConvI2D: 2678 case Op_ConvD2I: 2679 // case Op_ConvL2D: // handled by leaf call 2680 // case Op_ConvD2L: // handled by leaf call 2681 case Op_ConD: 2682 case Op_CmpD: 2683 case Op_CmpD3: 2684 frc.inc_double_count(); 2685 break; 2686 case Op_Opaque1: // Remove Opaque Nodes before matching 2687 case Op_Opaque2: // Remove Opaque Nodes before matching 2688 case Op_Opaque3: 2689 n->subsume_by(n->in(1), this); 2690 break; 2691 case Op_CallStaticJava: 2692 case Op_CallJava: 2693 case Op_CallDynamicJava: 2694 frc.inc_java_call_count(); // Count java call site; 2695 case Op_CallRuntime: 2696 case Op_CallLeaf: 2697 case Op_CallLeafNoFP: { 2698 assert( n->is_Call(), "" ); 2699 CallNode *call = n->as_Call(); 2700 // Count call sites where the FP mode bit would have to be flipped. 2701 // Do not count uncommon runtime calls: 2702 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2703 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2704 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2705 frc.inc_call_count(); // Count the call site 2706 } else { // See if uncommon argument is shared 2707 Node *n = call->in(TypeFunc::Parms); 2708 int nop = n->Opcode(); 2709 // Clone shared simple arguments to uncommon calls, item (1). 2710 if( n->outcnt() > 1 && 2711 !n->is_Proj() && 2712 nop != Op_CreateEx && 2713 nop != Op_CheckCastPP && 2714 nop != Op_DecodeN && 2715 nop != Op_DecodeNKlass && 2716 !n->is_Mem() ) { 2717 Node *x = n->clone(); 2718 call->set_req( TypeFunc::Parms, x ); 2719 } 2720 } 2721 break; 2722 } 2723 2724 case Op_StoreD: 2725 case Op_LoadD: 2726 case Op_LoadD_unaligned: 2727 frc.inc_double_count(); 2728 goto handle_mem; 2729 case Op_StoreF: 2730 case Op_LoadF: 2731 frc.inc_float_count(); 2732 goto handle_mem; 2733 2734 case Op_StoreCM: 2735 { 2736 // Convert OopStore dependence into precedence edge 2737 Node* prec = n->in(MemNode::OopStore); 2738 n->del_req(MemNode::OopStore); 2739 n->add_prec(prec); 2740 eliminate_redundant_card_marks(n); 2741 } 2742 2743 // fall through 2744 2745 case Op_StoreB: 2746 case Op_StoreC: 2747 case Op_StorePConditional: 2748 case Op_StoreI: 2749 case Op_StoreL: 2750 case Op_StoreIConditional: 2751 case Op_StoreLConditional: 2752 case Op_CompareAndSwapI: 2753 case Op_CompareAndSwapL: 2754 case Op_CompareAndSwapP: 2755 case Op_CompareAndSwapN: 2756 case Op_GetAndAddI: 2757 case Op_GetAndAddL: 2758 case Op_GetAndSetI: 2759 case Op_GetAndSetL: 2760 case Op_GetAndSetP: 2761 case Op_GetAndSetN: 2762 case Op_StoreP: 2763 case Op_StoreN: 2764 case Op_StoreNKlass: 2765 case Op_LoadB: 2766 case Op_LoadUB: 2767 case Op_LoadUS: 2768 case Op_LoadI: 2769 case Op_LoadKlass: 2770 case Op_LoadNKlass: 2771 case Op_LoadL: 2772 case Op_LoadL_unaligned: 2773 case Op_LoadPLocked: 2774 case Op_LoadP: 2775 case Op_LoadN: 2776 case Op_LoadRange: 2777 case Op_LoadS: { 2778 handle_mem: 2779 #ifdef ASSERT 2780 if( VerifyOptoOopOffsets ) { 2781 assert( n->is_Mem(), "" ); 2782 MemNode *mem = (MemNode*)n; 2783 // Check to see if address types have grounded out somehow. 2784 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2785 assert( !tp || oop_offset_is_sane(tp), "" ); 2786 } 2787 #endif 2788 break; 2789 } 2790 2791 case Op_AddP: { // Assert sane base pointers 2792 Node *addp = n->in(AddPNode::Address); 2793 assert( !addp->is_AddP() || 2794 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2795 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2796 "Base pointers must match" ); 2797 #ifdef _LP64 2798 if ((UseCompressedOops || UseCompressedClassPointers) && 2799 addp->Opcode() == Op_ConP && 2800 addp == n->in(AddPNode::Base) && 2801 n->in(AddPNode::Offset)->is_Con()) { 2802 // Use addressing with narrow klass to load with offset on x86. 2803 // On sparc loading 32-bits constant and decoding it have less 2804 // instructions (4) then load 64-bits constant (7). 2805 // Do this transformation here since IGVN will convert ConN back to ConP. 2806 const Type* t = addp->bottom_type(); 2807 if (t->isa_oopptr() || t->isa_klassptr()) { 2808 Node* nn = NULL; 2809 2810 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2811 2812 // Look for existing ConN node of the same exact type. 2813 Node* r = root(); 2814 uint cnt = r->outcnt(); 2815 for (uint i = 0; i < cnt; i++) { 2816 Node* m = r->raw_out(i); 2817 if (m!= NULL && m->Opcode() == op && 2818 m->bottom_type()->make_ptr() == t) { 2819 nn = m; 2820 break; 2821 } 2822 } 2823 if (nn != NULL) { 2824 // Decode a narrow oop to match address 2825 // [R12 + narrow_oop_reg<<3 + offset] 2826 if (t->isa_oopptr()) { 2827 nn = new (this) DecodeNNode(nn, t); 2828 } else { 2829 nn = new (this) DecodeNKlassNode(nn, t); 2830 } 2831 n->set_req(AddPNode::Base, nn); 2832 n->set_req(AddPNode::Address, nn); 2833 if (addp->outcnt() == 0) { 2834 addp->disconnect_inputs(NULL, this); 2835 } 2836 } 2837 } 2838 } 2839 #endif 2840 break; 2841 } 2842 2843 #ifdef _LP64 2844 case Op_CastPP: 2845 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2846 Node* in1 = n->in(1); 2847 const Type* t = n->bottom_type(); 2848 Node* new_in1 = in1->clone(); 2849 new_in1->as_DecodeN()->set_type(t); 2850 2851 if (!Matcher::narrow_oop_use_complex_address()) { 2852 // 2853 // x86, ARM and friends can handle 2 adds in addressing mode 2854 // and Matcher can fold a DecodeN node into address by using 2855 // a narrow oop directly and do implicit NULL check in address: 2856 // 2857 // [R12 + narrow_oop_reg<<3 + offset] 2858 // NullCheck narrow_oop_reg 2859 // 2860 // On other platforms (Sparc) we have to keep new DecodeN node and 2861 // use it to do implicit NULL check in address: 2862 // 2863 // decode_not_null narrow_oop_reg, base_reg 2864 // [base_reg + offset] 2865 // NullCheck base_reg 2866 // 2867 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2868 // to keep the information to which NULL check the new DecodeN node 2869 // corresponds to use it as value in implicit_null_check(). 2870 // 2871 new_in1->set_req(0, n->in(0)); 2872 } 2873 2874 n->subsume_by(new_in1, this); 2875 if (in1->outcnt() == 0) { 2876 in1->disconnect_inputs(NULL, this); 2877 } 2878 } 2879 break; 2880 2881 case Op_CmpP: 2882 // Do this transformation here to preserve CmpPNode::sub() and 2883 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2884 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2885 Node* in1 = n->in(1); 2886 Node* in2 = n->in(2); 2887 if (!in1->is_DecodeNarrowPtr()) { 2888 in2 = in1; 2889 in1 = n->in(2); 2890 } 2891 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2892 2893 Node* new_in2 = NULL; 2894 if (in2->is_DecodeNarrowPtr()) { 2895 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2896 new_in2 = in2->in(1); 2897 } else if (in2->Opcode() == Op_ConP) { 2898 const Type* t = in2->bottom_type(); 2899 if (t == TypePtr::NULL_PTR) { 2900 assert(in1->is_DecodeN(), "compare klass to null?"); 2901 // Don't convert CmpP null check into CmpN if compressed 2902 // oops implicit null check is not generated. 2903 // This will allow to generate normal oop implicit null check. 2904 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2905 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2906 // 2907 // This transformation together with CastPP transformation above 2908 // will generated code for implicit NULL checks for compressed oops. 2909 // 2910 // The original code after Optimize() 2911 // 2912 // LoadN memory, narrow_oop_reg 2913 // decode narrow_oop_reg, base_reg 2914 // CmpP base_reg, NULL 2915 // CastPP base_reg // NotNull 2916 // Load [base_reg + offset], val_reg 2917 // 2918 // after these transformations will be 2919 // 2920 // LoadN memory, narrow_oop_reg 2921 // CmpN narrow_oop_reg, NULL 2922 // decode_not_null narrow_oop_reg, base_reg 2923 // Load [base_reg + offset], val_reg 2924 // 2925 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2926 // since narrow oops can be used in debug info now (see the code in 2927 // final_graph_reshaping_walk()). 2928 // 2929 // At the end the code will be matched to 2930 // on x86: 2931 // 2932 // Load_narrow_oop memory, narrow_oop_reg 2933 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2934 // NullCheck narrow_oop_reg 2935 // 2936 // and on sparc: 2937 // 2938 // Load_narrow_oop memory, narrow_oop_reg 2939 // decode_not_null narrow_oop_reg, base_reg 2940 // Load [base_reg + offset], val_reg 2941 // NullCheck base_reg 2942 // 2943 } else if (t->isa_oopptr()) { 2944 new_in2 = ConNode::make(this, t->make_narrowoop()); 2945 } else if (t->isa_klassptr()) { 2946 new_in2 = ConNode::make(this, t->make_narrowklass()); 2947 } 2948 } 2949 if (new_in2 != NULL) { 2950 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2951 n->subsume_by(cmpN, this); 2952 if (in1->outcnt() == 0) { 2953 in1->disconnect_inputs(NULL, this); 2954 } 2955 if (in2->outcnt() == 0) { 2956 in2->disconnect_inputs(NULL, this); 2957 } 2958 } 2959 } 2960 break; 2961 2962 case Op_DecodeN: 2963 case Op_DecodeNKlass: 2964 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2965 // DecodeN could be pinned when it can't be fold into 2966 // an address expression, see the code for Op_CastPP above. 2967 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2968 break; 2969 2970 case Op_EncodeP: 2971 case Op_EncodePKlass: { 2972 Node* in1 = n->in(1); 2973 if (in1->is_DecodeNarrowPtr()) { 2974 n->subsume_by(in1->in(1), this); 2975 } else if (in1->Opcode() == Op_ConP) { 2976 const Type* t = in1->bottom_type(); 2977 if (t == TypePtr::NULL_PTR) { 2978 assert(t->isa_oopptr(), "null klass?"); 2979 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2980 } else if (t->isa_oopptr()) { 2981 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2982 } else if (t->isa_klassptr()) { 2983 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2984 } 2985 } 2986 if (in1->outcnt() == 0) { 2987 in1->disconnect_inputs(NULL, this); 2988 } 2989 break; 2990 } 2991 2992 case Op_Proj: { 2993 if (OptimizeStringConcat) { 2994 ProjNode* p = n->as_Proj(); 2995 if (p->_is_io_use) { 2996 // Separate projections were used for the exception path which 2997 // are normally removed by a late inline. If it wasn't inlined 2998 // then they will hang around and should just be replaced with 2999 // the original one. 3000 Node* proj = NULL; 3001 // Replace with just one 3002 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3003 Node *use = i.get(); 3004 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3005 proj = use; 3006 break; 3007 } 3008 } 3009 assert(proj != NULL, "must be found"); 3010 p->subsume_by(proj, this); 3011 } 3012 } 3013 break; 3014 } 3015 3016 case Op_Phi: 3017 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3018 // The EncodeP optimization may create Phi with the same edges 3019 // for all paths. It is not handled well by Register Allocator. 3020 Node* unique_in = n->in(1); 3021 assert(unique_in != NULL, ""); 3022 uint cnt = n->req(); 3023 for (uint i = 2; i < cnt; i++) { 3024 Node* m = n->in(i); 3025 assert(m != NULL, ""); 3026 if (unique_in != m) 3027 unique_in = NULL; 3028 } 3029 if (unique_in != NULL) { 3030 n->subsume_by(unique_in, this); 3031 } 3032 } 3033 break; 3034 3035 #endif 3036 3037 #ifdef ASSERT 3038 case Op_CastII: 3039 // Verify that all range check dependent CastII nodes were removed. 3040 if (n->isa_CastII()->has_range_check()) { 3041 n->dump(3); 3042 assert(false, "Range check dependent CastII node was not removed"); 3043 } 3044 break; 3045 #endif 3046 3047 case Op_ModI: 3048 if (UseDivMod) { 3049 // Check if a%b and a/b both exist 3050 Node* d = n->find_similar(Op_DivI); 3051 if (d) { 3052 // Replace them with a fused divmod if supported 3053 if (Matcher::has_match_rule(Op_DivModI)) { 3054 DivModINode* divmod = DivModINode::make(this, n); 3055 d->subsume_by(divmod->div_proj(), this); 3056 n->subsume_by(divmod->mod_proj(), this); 3057 } else { 3058 // replace a%b with a-((a/b)*b) 3059 Node* mult = new (this) MulINode(d, d->in(2)); 3060 Node* sub = new (this) SubINode(d->in(1), mult); 3061 n->subsume_by(sub, this); 3062 } 3063 } 3064 } 3065 break; 3066 3067 case Op_ModL: 3068 if (UseDivMod) { 3069 // Check if a%b and a/b both exist 3070 Node* d = n->find_similar(Op_DivL); 3071 if (d) { 3072 // Replace them with a fused divmod if supported 3073 if (Matcher::has_match_rule(Op_DivModL)) { 3074 DivModLNode* divmod = DivModLNode::make(this, n); 3075 d->subsume_by(divmod->div_proj(), this); 3076 n->subsume_by(divmod->mod_proj(), this); 3077 } else { 3078 // replace a%b with a-((a/b)*b) 3079 Node* mult = new (this) MulLNode(d, d->in(2)); 3080 Node* sub = new (this) SubLNode(d->in(1), mult); 3081 n->subsume_by(sub, this); 3082 } 3083 } 3084 } 3085 break; 3086 3087 case Op_LoadVector: 3088 case Op_StoreVector: 3089 break; 3090 3091 case Op_PackB: 3092 case Op_PackS: 3093 case Op_PackI: 3094 case Op_PackF: 3095 case Op_PackL: 3096 case Op_PackD: 3097 if (n->req()-1 > 2) { 3098 // Replace many operand PackNodes with a binary tree for matching 3099 PackNode* p = (PackNode*) n; 3100 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3101 n->subsume_by(btp, this); 3102 } 3103 break; 3104 case Op_Loop: 3105 case Op_CountedLoop: 3106 if (n->as_Loop()->is_inner_loop()) { 3107 frc.inc_inner_loop_count(); 3108 } 3109 break; 3110 case Op_LShiftI: 3111 case Op_RShiftI: 3112 case Op_URShiftI: 3113 case Op_LShiftL: 3114 case Op_RShiftL: 3115 case Op_URShiftL: 3116 if (Matcher::need_masked_shift_count) { 3117 // The cpu's shift instructions don't restrict the count to the 3118 // lower 5/6 bits. We need to do the masking ourselves. 3119 Node* in2 = n->in(2); 3120 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3121 const TypeInt* t = in2->find_int_type(); 3122 if (t != NULL && t->is_con()) { 3123 juint shift = t->get_con(); 3124 if (shift > mask) { // Unsigned cmp 3125 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3126 } 3127 } else { 3128 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3129 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3130 n->set_req(2, shift); 3131 } 3132 } 3133 if (in2->outcnt() == 0) { // Remove dead node 3134 in2->disconnect_inputs(NULL, this); 3135 } 3136 } 3137 break; 3138 case Op_MemBarStoreStore: 3139 case Op_MemBarRelease: 3140 // Break the link with AllocateNode: it is no longer useful and 3141 // confuses register allocation. 3142 if (n->req() > MemBarNode::Precedent) { 3143 n->set_req(MemBarNode::Precedent, top()); 3144 } 3145 break; 3146 default: 3147 assert( !n->is_Call(), "" ); 3148 assert( !n->is_Mem(), "" ); 3149 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3150 break; 3151 } 3152 3153 // Collect CFG split points 3154 if (n->is_MultiBranch()) 3155 frc._tests.push(n); 3156 } 3157 3158 //------------------------------final_graph_reshaping_walk--------------------- 3159 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3160 // requires that the walk visits a node's inputs before visiting the node. 3161 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3162 ResourceArea *area = Thread::current()->resource_area(); 3163 Unique_Node_List sfpt(area); 3164 3165 frc._visited.set(root->_idx); // first, mark node as visited 3166 uint cnt = root->req(); 3167 Node *n = root; 3168 uint i = 0; 3169 while (true) { 3170 if (i < cnt) { 3171 // Place all non-visited non-null inputs onto stack 3172 Node* m = n->in(i); 3173 ++i; 3174 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3175 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3176 // compute worst case interpreter size in case of a deoptimization 3177 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3178 3179 sfpt.push(m); 3180 } 3181 cnt = m->req(); 3182 nstack.push(n, i); // put on stack parent and next input's index 3183 n = m; 3184 i = 0; 3185 } 3186 } else { 3187 // Now do post-visit work 3188 final_graph_reshaping_impl( n, frc ); 3189 if (nstack.is_empty()) 3190 break; // finished 3191 n = nstack.node(); // Get node from stack 3192 cnt = n->req(); 3193 i = nstack.index(); 3194 nstack.pop(); // Shift to the next node on stack 3195 } 3196 } 3197 3198 // Skip next transformation if compressed oops are not used. 3199 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3200 (!UseCompressedOops && !UseCompressedClassPointers)) 3201 return; 3202 3203 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3204 // It could be done for an uncommon traps or any safepoints/calls 3205 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3206 while (sfpt.size() > 0) { 3207 n = sfpt.pop(); 3208 JVMState *jvms = n->as_SafePoint()->jvms(); 3209 assert(jvms != NULL, "sanity"); 3210 int start = jvms->debug_start(); 3211 int end = n->req(); 3212 bool is_uncommon = (n->is_CallStaticJava() && 3213 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3214 for (int j = start; j < end; j++) { 3215 Node* in = n->in(j); 3216 if (in->is_DecodeNarrowPtr()) { 3217 bool safe_to_skip = true; 3218 if (!is_uncommon ) { 3219 // Is it safe to skip? 3220 for (uint i = 0; i < in->outcnt(); i++) { 3221 Node* u = in->raw_out(i); 3222 if (!u->is_SafePoint() || 3223 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3224 safe_to_skip = false; 3225 } 3226 } 3227 } 3228 if (safe_to_skip) { 3229 n->set_req(j, in->in(1)); 3230 } 3231 if (in->outcnt() == 0) { 3232 in->disconnect_inputs(NULL, this); 3233 } 3234 } 3235 } 3236 } 3237 } 3238 3239 //------------------------------final_graph_reshaping-------------------------- 3240 // Final Graph Reshaping. 3241 // 3242 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3243 // and not commoned up and forced early. Must come after regular 3244 // optimizations to avoid GVN undoing the cloning. Clone constant 3245 // inputs to Loop Phis; these will be split by the allocator anyways. 3246 // Remove Opaque nodes. 3247 // (2) Move last-uses by commutative operations to the left input to encourage 3248 // Intel update-in-place two-address operations and better register usage 3249 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3250 // calls canonicalizing them back. 3251 // (3) Count the number of double-precision FP ops, single-precision FP ops 3252 // and call sites. On Intel, we can get correct rounding either by 3253 // forcing singles to memory (requires extra stores and loads after each 3254 // FP bytecode) or we can set a rounding mode bit (requires setting and 3255 // clearing the mode bit around call sites). The mode bit is only used 3256 // if the relative frequency of single FP ops to calls is low enough. 3257 // This is a key transform for SPEC mpeg_audio. 3258 // (4) Detect infinite loops; blobs of code reachable from above but not 3259 // below. Several of the Code_Gen algorithms fail on such code shapes, 3260 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3261 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3262 // Detection is by looking for IfNodes where only 1 projection is 3263 // reachable from below or CatchNodes missing some targets. 3264 // (5) Assert for insane oop offsets in debug mode. 3265 3266 bool Compile::final_graph_reshaping() { 3267 // an infinite loop may have been eliminated by the optimizer, 3268 // in which case the graph will be empty. 3269 if (root()->req() == 1) { 3270 record_method_not_compilable("trivial infinite loop"); 3271 return true; 3272 } 3273 3274 // Expensive nodes have their control input set to prevent the GVN 3275 // from freely commoning them. There's no GVN beyond this point so 3276 // no need to keep the control input. We want the expensive nodes to 3277 // be freely moved to the least frequent code path by gcm. 3278 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3279 for (int i = 0; i < expensive_count(); i++) { 3280 _expensive_nodes->at(i)->set_req(0, NULL); 3281 } 3282 3283 Final_Reshape_Counts frc; 3284 3285 // Visit everybody reachable! 3286 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3287 Node_Stack nstack(live_nodes() >> 1); 3288 final_graph_reshaping_walk(nstack, root(), frc); 3289 3290 // Check for unreachable (from below) code (i.e., infinite loops). 3291 for( uint i = 0; i < frc._tests.size(); i++ ) { 3292 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3293 // Get number of CFG targets. 3294 // Note that PCTables include exception targets after calls. 3295 uint required_outcnt = n->required_outcnt(); 3296 if (n->outcnt() != required_outcnt) { 3297 // Check for a few special cases. Rethrow Nodes never take the 3298 // 'fall-thru' path, so expected kids is 1 less. 3299 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3300 if (n->in(0)->in(0)->is_Call()) { 3301 CallNode *call = n->in(0)->in(0)->as_Call(); 3302 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3303 required_outcnt--; // Rethrow always has 1 less kid 3304 } else if (call->req() > TypeFunc::Parms && 3305 call->is_CallDynamicJava()) { 3306 // Check for null receiver. In such case, the optimizer has 3307 // detected that the virtual call will always result in a null 3308 // pointer exception. The fall-through projection of this CatchNode 3309 // will not be populated. 3310 Node *arg0 = call->in(TypeFunc::Parms); 3311 if (arg0->is_Type() && 3312 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3313 required_outcnt--; 3314 } 3315 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3316 call->req() > TypeFunc::Parms+1 && 3317 call->is_CallStaticJava()) { 3318 // Check for negative array length. In such case, the optimizer has 3319 // detected that the allocation attempt will always result in an 3320 // exception. There is no fall-through projection of this CatchNode . 3321 Node *arg1 = call->in(TypeFunc::Parms+1); 3322 if (arg1->is_Type() && 3323 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3324 required_outcnt--; 3325 } 3326 } 3327 } 3328 } 3329 // Recheck with a better notion of 'required_outcnt' 3330 if (n->outcnt() != required_outcnt) { 3331 record_method_not_compilable("malformed control flow"); 3332 return true; // Not all targets reachable! 3333 } 3334 } 3335 // Check that I actually visited all kids. Unreached kids 3336 // must be infinite loops. 3337 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3338 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3339 record_method_not_compilable("infinite loop"); 3340 return true; // Found unvisited kid; must be unreach 3341 } 3342 } 3343 3344 // If original bytecodes contained a mixture of floats and doubles 3345 // check if the optimizer has made it homogenous, item (3). 3346 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3347 frc.get_float_count() > 32 && 3348 frc.get_double_count() == 0 && 3349 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3350 set_24_bit_selection_and_mode( false, true ); 3351 } 3352 3353 set_java_calls(frc.get_java_call_count()); 3354 set_inner_loops(frc.get_inner_loop_count()); 3355 3356 // No infinite loops, no reason to bail out. 3357 return false; 3358 } 3359 3360 //-----------------------------too_many_traps---------------------------------- 3361 // Report if there are too many traps at the current method and bci. 3362 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3363 bool Compile::too_many_traps(ciMethod* method, 3364 int bci, 3365 Deoptimization::DeoptReason reason) { 3366 ciMethodData* md = method->method_data(); 3367 if (md->is_empty()) { 3368 // Assume the trap has not occurred, or that it occurred only 3369 // because of a transient condition during start-up in the interpreter. 3370 return false; 3371 } 3372 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3373 if (md->has_trap_at(bci, m, reason) != 0) { 3374 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3375 // Also, if there are multiple reasons, or if there is no per-BCI record, 3376 // assume the worst. 3377 if (log()) 3378 log()->elem("observe trap='%s' count='%d'", 3379 Deoptimization::trap_reason_name(reason), 3380 md->trap_count(reason)); 3381 return true; 3382 } else { 3383 // Ignore method/bci and see if there have been too many globally. 3384 return too_many_traps(reason, md); 3385 } 3386 } 3387 3388 // Less-accurate variant which does not require a method and bci. 3389 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3390 ciMethodData* logmd) { 3391 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3392 // Too many traps globally. 3393 // Note that we use cumulative trap_count, not just md->trap_count. 3394 if (log()) { 3395 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3396 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3397 Deoptimization::trap_reason_name(reason), 3398 mcount, trap_count(reason)); 3399 } 3400 return true; 3401 } else { 3402 // The coast is clear. 3403 return false; 3404 } 3405 } 3406 3407 //--------------------------too_many_recompiles-------------------------------- 3408 // Report if there are too many recompiles at the current method and bci. 3409 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3410 // Is not eager to return true, since this will cause the compiler to use 3411 // Action_none for a trap point, to avoid too many recompilations. 3412 bool Compile::too_many_recompiles(ciMethod* method, 3413 int bci, 3414 Deoptimization::DeoptReason reason) { 3415 ciMethodData* md = method->method_data(); 3416 if (md->is_empty()) { 3417 // Assume the trap has not occurred, or that it occurred only 3418 // because of a transient condition during start-up in the interpreter. 3419 return false; 3420 } 3421 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3422 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3423 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3424 Deoptimization::DeoptReason per_bc_reason 3425 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3426 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3427 if ((per_bc_reason == Deoptimization::Reason_none 3428 || md->has_trap_at(bci, m, reason) != 0) 3429 // The trap frequency measure we care about is the recompile count: 3430 && md->trap_recompiled_at(bci, m) 3431 && md->overflow_recompile_count() >= bc_cutoff) { 3432 // Do not emit a trap here if it has already caused recompilations. 3433 // Also, if there are multiple reasons, or if there is no per-BCI record, 3434 // assume the worst. 3435 if (log()) 3436 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3437 Deoptimization::trap_reason_name(reason), 3438 md->trap_count(reason), 3439 md->overflow_recompile_count()); 3440 return true; 3441 } else if (trap_count(reason) != 0 3442 && decompile_count() >= m_cutoff) { 3443 // Too many recompiles globally, and we have seen this sort of trap. 3444 // Use cumulative decompile_count, not just md->decompile_count. 3445 if (log()) 3446 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3447 Deoptimization::trap_reason_name(reason), 3448 md->trap_count(reason), trap_count(reason), 3449 md->decompile_count(), decompile_count()); 3450 return true; 3451 } else { 3452 // The coast is clear. 3453 return false; 3454 } 3455 } 3456 3457 // Compute when not to trap. Used by matching trap based nodes and 3458 // NullCheck optimization. 3459 void Compile::set_allowed_deopt_reasons() { 3460 _allowed_reasons = 0; 3461 if (is_method_compilation()) { 3462 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3463 assert(rs < BitsPerInt, "recode bit map"); 3464 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3465 _allowed_reasons |= nth_bit(rs); 3466 } 3467 } 3468 } 3469 } 3470 3471 #ifndef PRODUCT 3472 //------------------------------verify_graph_edges--------------------------- 3473 // Walk the Graph and verify that there is a one-to-one correspondence 3474 // between Use-Def edges and Def-Use edges in the graph. 3475 void Compile::verify_graph_edges(bool no_dead_code) { 3476 if (VerifyGraphEdges) { 3477 ResourceArea *area = Thread::current()->resource_area(); 3478 Unique_Node_List visited(area); 3479 // Call recursive graph walk to check edges 3480 _root->verify_edges(visited); 3481 if (no_dead_code) { 3482 // Now make sure that no visited node is used by an unvisited node. 3483 bool dead_nodes = false; 3484 Unique_Node_List checked(area); 3485 while (visited.size() > 0) { 3486 Node* n = visited.pop(); 3487 checked.push(n); 3488 for (uint i = 0; i < n->outcnt(); i++) { 3489 Node* use = n->raw_out(i); 3490 if (checked.member(use)) continue; // already checked 3491 if (visited.member(use)) continue; // already in the graph 3492 if (use->is_Con()) continue; // a dead ConNode is OK 3493 // At this point, we have found a dead node which is DU-reachable. 3494 if (!dead_nodes) { 3495 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3496 dead_nodes = true; 3497 } 3498 use->dump(2); 3499 tty->print_cr("---"); 3500 checked.push(use); // No repeats; pretend it is now checked. 3501 } 3502 } 3503 assert(!dead_nodes, "using nodes must be reachable from root"); 3504 } 3505 } 3506 } 3507 3508 // Verify GC barriers consistency 3509 // Currently supported: 3510 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3511 void Compile::verify_barriers() { 3512 if (UseG1GC) { 3513 // Verify G1 pre-barriers 3514 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3515 3516 ResourceArea *area = Thread::current()->resource_area(); 3517 Unique_Node_List visited(area); 3518 Node_List worklist(area); 3519 // We're going to walk control flow backwards starting from the Root 3520 worklist.push(_root); 3521 while (worklist.size() > 0) { 3522 Node* x = worklist.pop(); 3523 if (x == NULL || x == top()) continue; 3524 if (visited.member(x)) { 3525 continue; 3526 } else { 3527 visited.push(x); 3528 } 3529 3530 if (x->is_Region()) { 3531 for (uint i = 1; i < x->req(); i++) { 3532 worklist.push(x->in(i)); 3533 } 3534 } else { 3535 worklist.push(x->in(0)); 3536 // We are looking for the pattern: 3537 // /->ThreadLocal 3538 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3539 // \->ConI(0) 3540 // We want to verify that the If and the LoadB have the same control 3541 // See GraphKit::g1_write_barrier_pre() 3542 if (x->is_If()) { 3543 IfNode *iff = x->as_If(); 3544 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3545 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3546 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3547 && cmp->in(1)->is_Load()) { 3548 LoadNode* load = cmp->in(1)->as_Load(); 3549 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3550 && load->in(2)->in(3)->is_Con() 3551 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3552 3553 Node* if_ctrl = iff->in(0); 3554 Node* load_ctrl = load->in(0); 3555 3556 if (if_ctrl != load_ctrl) { 3557 // Skip possible CProj->NeverBranch in infinite loops 3558 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3559 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3560 if_ctrl = if_ctrl->in(0)->in(0); 3561 } 3562 } 3563 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3564 } 3565 } 3566 } 3567 } 3568 } 3569 } 3570 } 3571 } 3572 3573 #endif 3574 3575 // The Compile object keeps track of failure reasons separately from the ciEnv. 3576 // This is required because there is not quite a 1-1 relation between the 3577 // ciEnv and its compilation task and the Compile object. Note that one 3578 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3579 // to backtrack and retry without subsuming loads. Other than this backtracking 3580 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3581 // by the logic in C2Compiler. 3582 void Compile::record_failure(const char* reason) { 3583 if (log() != NULL) { 3584 log()->elem("failure reason='%s' phase='compile'", reason); 3585 } 3586 if (_failure_reason == NULL) { 3587 // Record the first failure reason. 3588 _failure_reason = reason; 3589 } 3590 3591 EventCompilerFailure event; 3592 if (event.should_commit()) { 3593 event.set_compileID(Compile::compile_id()); 3594 event.set_failure(reason); 3595 event.commit(); 3596 } 3597 3598 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3599 C->print_method(PHASE_FAILURE); 3600 } 3601 _root = NULL; // flush the graph, too 3602 } 3603 3604 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3605 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3606 _phase_name(name), _dolog(dolog) 3607 { 3608 if (dolog) { 3609 C = Compile::current(); 3610 _log = C->log(); 3611 } else { 3612 C = NULL; 3613 _log = NULL; 3614 } 3615 if (_log != NULL) { 3616 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3617 _log->stamp(); 3618 _log->end_head(); 3619 } 3620 } 3621 3622 Compile::TracePhase::~TracePhase() { 3623 3624 C = Compile::current(); 3625 if (_dolog) { 3626 _log = C->log(); 3627 } else { 3628 _log = NULL; 3629 } 3630 3631 #ifdef ASSERT 3632 if (PrintIdealNodeCount) { 3633 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3634 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3635 } 3636 3637 if (VerifyIdealNodeCount) { 3638 Compile::current()->print_missing_nodes(); 3639 } 3640 #endif 3641 3642 if (_log != NULL) { 3643 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3644 } 3645 } 3646 3647 //============================================================================= 3648 // Two Constant's are equal when the type and the value are equal. 3649 bool Compile::Constant::operator==(const Constant& other) { 3650 if (type() != other.type() ) return false; 3651 if (can_be_reused() != other.can_be_reused()) return false; 3652 // For floating point values we compare the bit pattern. 3653 switch (type()) { 3654 case T_FLOAT: return (_v._value.i == other._v._value.i); 3655 case T_LONG: 3656 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3657 case T_OBJECT: 3658 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3659 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3660 case T_METADATA: return (_v._metadata == other._v._metadata); 3661 default: ShouldNotReachHere(); 3662 } 3663 return false; 3664 } 3665 3666 static int type_to_size_in_bytes(BasicType t) { 3667 switch (t) { 3668 case T_LONG: return sizeof(jlong ); 3669 case T_FLOAT: return sizeof(jfloat ); 3670 case T_DOUBLE: return sizeof(jdouble); 3671 case T_METADATA: return sizeof(Metadata*); 3672 // We use T_VOID as marker for jump-table entries (labels) which 3673 // need an internal word relocation. 3674 case T_VOID: 3675 case T_ADDRESS: 3676 case T_OBJECT: return sizeof(jobject); 3677 } 3678 3679 ShouldNotReachHere(); 3680 return -1; 3681 } 3682 3683 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3684 // sort descending 3685 if (a->freq() > b->freq()) return -1; 3686 if (a->freq() < b->freq()) return 1; 3687 return 0; 3688 } 3689 3690 void Compile::ConstantTable::calculate_offsets_and_size() { 3691 // First, sort the array by frequencies. 3692 _constants.sort(qsort_comparator); 3693 3694 #ifdef ASSERT 3695 // Make sure all jump-table entries were sorted to the end of the 3696 // array (they have a negative frequency). 3697 bool found_void = false; 3698 for (int i = 0; i < _constants.length(); i++) { 3699 Constant con = _constants.at(i); 3700 if (con.type() == T_VOID) 3701 found_void = true; // jump-tables 3702 else 3703 assert(!found_void, "wrong sorting"); 3704 } 3705 #endif 3706 3707 int offset = 0; 3708 for (int i = 0; i < _constants.length(); i++) { 3709 Constant* con = _constants.adr_at(i); 3710 3711 // Align offset for type. 3712 int typesize = type_to_size_in_bytes(con->type()); 3713 offset = align_size_up(offset, typesize); 3714 con->set_offset(offset); // set constant's offset 3715 3716 if (con->type() == T_VOID) { 3717 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3718 offset = offset + typesize * n->outcnt(); // expand jump-table 3719 } else { 3720 offset = offset + typesize; 3721 } 3722 } 3723 3724 // Align size up to the next section start (which is insts; see 3725 // CodeBuffer::align_at_start). 3726 assert(_size == -1, "already set?"); 3727 _size = align_size_up(offset, CodeEntryAlignment); 3728 } 3729 3730 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3731 MacroAssembler _masm(&cb); 3732 for (int i = 0; i < _constants.length(); i++) { 3733 Constant con = _constants.at(i); 3734 address constant_addr = NULL; 3735 switch (con.type()) { 3736 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3737 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3738 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3739 case T_OBJECT: { 3740 jobject obj = con.get_jobject(); 3741 int oop_index = _masm.oop_recorder()->find_index(obj); 3742 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3743 break; 3744 } 3745 case T_ADDRESS: { 3746 address addr = (address) con.get_jobject(); 3747 constant_addr = _masm.address_constant(addr); 3748 break; 3749 } 3750 // We use T_VOID as marker for jump-table entries (labels) which 3751 // need an internal word relocation. 3752 case T_VOID: { 3753 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3754 // Fill the jump-table with a dummy word. The real value is 3755 // filled in later in fill_jump_table. 3756 address dummy = (address) n; 3757 constant_addr = _masm.address_constant(dummy); 3758 // Expand jump-table 3759 for (uint i = 1; i < n->outcnt(); i++) { 3760 address temp_addr = _masm.address_constant(dummy + i); 3761 assert(temp_addr, "consts section too small"); 3762 } 3763 break; 3764 } 3765 case T_METADATA: { 3766 Metadata* obj = con.get_metadata(); 3767 int metadata_index = _masm.oop_recorder()->find_index(obj); 3768 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3769 break; 3770 } 3771 default: ShouldNotReachHere(); 3772 } 3773 assert(constant_addr, "consts section too small"); 3774 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3775 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3776 } 3777 } 3778 3779 int Compile::ConstantTable::find_offset(Constant& con) const { 3780 int idx = _constants.find(con); 3781 assert(idx != -1, "constant must be in constant table"); 3782 int offset = _constants.at(idx).offset(); 3783 assert(offset != -1, "constant table not emitted yet?"); 3784 return offset; 3785 } 3786 3787 void Compile::ConstantTable::add(Constant& con) { 3788 if (con.can_be_reused()) { 3789 int idx = _constants.find(con); 3790 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3791 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3792 return; 3793 } 3794 } 3795 (void) _constants.append(con); 3796 } 3797 3798 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3799 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3800 Constant con(type, value, b->_freq); 3801 add(con); 3802 return con; 3803 } 3804 3805 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3806 Constant con(metadata); 3807 add(con); 3808 return con; 3809 } 3810 3811 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3812 jvalue value; 3813 BasicType type = oper->type()->basic_type(); 3814 switch (type) { 3815 case T_LONG: value.j = oper->constantL(); break; 3816 case T_FLOAT: value.f = oper->constantF(); break; 3817 case T_DOUBLE: value.d = oper->constantD(); break; 3818 case T_OBJECT: 3819 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3820 case T_METADATA: return add((Metadata*)oper->constant()); break; 3821 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3822 } 3823 return add(n, type, value); 3824 } 3825 3826 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3827 jvalue value; 3828 // We can use the node pointer here to identify the right jump-table 3829 // as this method is called from Compile::Fill_buffer right before 3830 // the MachNodes are emitted and the jump-table is filled (means the 3831 // MachNode pointers do not change anymore). 3832 value.l = (jobject) n; 3833 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3834 add(con); 3835 return con; 3836 } 3837 3838 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3839 // If called from Compile::scratch_emit_size do nothing. 3840 if (Compile::current()->in_scratch_emit_size()) return; 3841 3842 assert(labels.is_nonempty(), "must be"); 3843 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3844 3845 // Since MachConstantNode::constant_offset() also contains 3846 // table_base_offset() we need to subtract the table_base_offset() 3847 // to get the plain offset into the constant table. 3848 int offset = n->constant_offset() - table_base_offset(); 3849 3850 MacroAssembler _masm(&cb); 3851 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3852 3853 for (uint i = 0; i < n->outcnt(); i++) { 3854 address* constant_addr = &jump_table_base[i]; 3855 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))); 3856 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3857 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3858 } 3859 } 3860 3861 void Compile::dump_inlining() { 3862 if (print_inlining() || print_intrinsics()) { 3863 // Print inlining message for candidates that we couldn't inline 3864 // for lack of space or non constant receiver 3865 for (int i = 0; i < _late_inlines.length(); i++) { 3866 CallGenerator* cg = _late_inlines.at(i); 3867 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3868 } 3869 Unique_Node_List useful; 3870 useful.push(root()); 3871 for (uint next = 0; next < useful.size(); ++next) { 3872 Node* n = useful.at(next); 3873 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3874 CallNode* call = n->as_Call(); 3875 CallGenerator* cg = call->generator(); 3876 cg->print_inlining_late("receiver not constant"); 3877 } 3878 uint max = n->len(); 3879 for ( uint i = 0; i < max; ++i ) { 3880 Node *m = n->in(i); 3881 if ( m == NULL ) continue; 3882 useful.push(m); 3883 } 3884 } 3885 for (int i = 0; i < _print_inlining_list->length(); i++) { 3886 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3887 } 3888 } 3889 } 3890 3891 // Dump inlining replay data to the stream. 3892 // Don't change thread state and acquire any locks. 3893 void Compile::dump_inline_data(outputStream* out) { 3894 InlineTree* inl_tree = ilt(); 3895 if (inl_tree != NULL) { 3896 out->print(" inline %d", inl_tree->count()); 3897 inl_tree->dump_replay_data(out); 3898 } 3899 } 3900 3901 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3902 if (n1->Opcode() < n2->Opcode()) return -1; 3903 else if (n1->Opcode() > n2->Opcode()) return 1; 3904 3905 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())); 3906 for (uint i = 1; i < n1->req(); i++) { 3907 if (n1->in(i) < n2->in(i)) return -1; 3908 else if (n1->in(i) > n2->in(i)) return 1; 3909 } 3910 3911 return 0; 3912 } 3913 3914 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3915 Node* n1 = *n1p; 3916 Node* n2 = *n2p; 3917 3918 return cmp_expensive_nodes(n1, n2); 3919 } 3920 3921 void Compile::sort_expensive_nodes() { 3922 if (!expensive_nodes_sorted()) { 3923 _expensive_nodes->sort(cmp_expensive_nodes); 3924 } 3925 } 3926 3927 bool Compile::expensive_nodes_sorted() const { 3928 for (int i = 1; i < _expensive_nodes->length(); i++) { 3929 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3930 return false; 3931 } 3932 } 3933 return true; 3934 } 3935 3936 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3937 if (_expensive_nodes->length() == 0) { 3938 return false; 3939 } 3940 3941 assert(OptimizeExpensiveOps, "optimization off?"); 3942 3943 // Take this opportunity to remove dead nodes from the list 3944 int j = 0; 3945 for (int i = 0; i < _expensive_nodes->length(); i++) { 3946 Node* n = _expensive_nodes->at(i); 3947 if (!n->is_unreachable(igvn)) { 3948 assert(n->is_expensive(), "should be expensive"); 3949 _expensive_nodes->at_put(j, n); 3950 j++; 3951 } 3952 } 3953 _expensive_nodes->trunc_to(j); 3954 3955 // Then sort the list so that similar nodes are next to each other 3956 // and check for at least two nodes of identical kind with same data 3957 // inputs. 3958 sort_expensive_nodes(); 3959 3960 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3961 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3962 return true; 3963 } 3964 } 3965 3966 return false; 3967 } 3968 3969 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3970 if (_expensive_nodes->length() == 0) { 3971 return; 3972 } 3973 3974 assert(OptimizeExpensiveOps, "optimization off?"); 3975 3976 // Sort to bring similar nodes next to each other and clear the 3977 // control input of nodes for which there's only a single copy. 3978 sort_expensive_nodes(); 3979 3980 int j = 0; 3981 int identical = 0; 3982 int i = 0; 3983 for (; i < _expensive_nodes->length()-1; i++) { 3984 assert(j <= i, "can't write beyond current index"); 3985 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3986 identical++; 3987 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3988 continue; 3989 } 3990 if (identical > 0) { 3991 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3992 identical = 0; 3993 } else { 3994 Node* n = _expensive_nodes->at(i); 3995 igvn.hash_delete(n); 3996 n->set_req(0, NULL); 3997 igvn.hash_insert(n); 3998 } 3999 } 4000 if (identical > 0) { 4001 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4002 } else if (_expensive_nodes->length() >= 1) { 4003 Node* n = _expensive_nodes->at(i); 4004 igvn.hash_delete(n); 4005 n->set_req(0, NULL); 4006 igvn.hash_insert(n); 4007 } 4008 _expensive_nodes->trunc_to(j); 4009 } 4010 4011 void Compile::add_expensive_node(Node * n) { 4012 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4013 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4014 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4015 if (OptimizeExpensiveOps) { 4016 _expensive_nodes->append(n); 4017 } else { 4018 // Clear control input and let IGVN optimize expensive nodes if 4019 // OptimizeExpensiveOps is off. 4020 n->set_req(0, NULL); 4021 } 4022 } 4023 4024 /** 4025 * Remove the speculative part of types and clean up the graph 4026 */ 4027 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4028 if (UseTypeSpeculation) { 4029 Unique_Node_List worklist; 4030 worklist.push(root()); 4031 int modified = 0; 4032 // Go over all type nodes that carry a speculative type, drop the 4033 // speculative part of the type and enqueue the node for an igvn 4034 // which may optimize it out. 4035 for (uint next = 0; next < worklist.size(); ++next) { 4036 Node *n = worklist.at(next); 4037 if (n->is_Type()) { 4038 TypeNode* tn = n->as_Type(); 4039 const Type* t = tn->type(); 4040 const Type* t_no_spec = t->remove_speculative(); 4041 if (t_no_spec != t) { 4042 bool in_hash = igvn.hash_delete(n); 4043 assert(in_hash, "node should be in igvn hash table"); 4044 tn->set_type(t_no_spec); 4045 igvn.hash_insert(n); 4046 igvn._worklist.push(n); // give it a chance to go away 4047 modified++; 4048 } 4049 } 4050 uint max = n->len(); 4051 for( uint i = 0; i < max; ++i ) { 4052 Node *m = n->in(i); 4053 if (not_a_node(m)) continue; 4054 worklist.push(m); 4055 } 4056 } 4057 // Drop the speculative part of all types in the igvn's type table 4058 igvn.remove_speculative_types(); 4059 if (modified > 0) { 4060 igvn.optimize(); 4061 } 4062 #ifdef ASSERT 4063 // Verify that after the IGVN is over no speculative type has resurfaced 4064 worklist.clear(); 4065 worklist.push(root()); 4066 for (uint next = 0; next < worklist.size(); ++next) { 4067 Node *n = worklist.at(next); 4068 const Type* t = igvn.type_or_null(n); 4069 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4070 if (n->is_Type()) { 4071 t = n->as_Type()->type(); 4072 assert(t == t->remove_speculative(), "no more speculative types"); 4073 } 4074 uint max = n->len(); 4075 for( uint i = 0; i < max; ++i ) { 4076 Node *m = n->in(i); 4077 if (not_a_node(m)) continue; 4078 worklist.push(m); 4079 } 4080 } 4081 igvn.check_no_speculative_types(); 4082 #endif 4083 } 4084 } 4085 4086 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4087 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4088 if (ctrl != NULL) { 4089 // Express control dependency by a CastII node with a narrow type. 4090 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */); 4091 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4092 // node from floating above the range check during loop optimizations. Otherwise, the 4093 // ConvI2L node may be eliminated independently of the range check, causing the data path 4094 // to become TOP while the control path is still there (although it's unreachable). 4095 value->set_req(0, ctrl); 4096 // Save CastII node to remove it after loop optimizations. 4097 phase->C->add_range_check_cast(value); 4098 value = phase->transform(value); 4099 } 4100 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4101 return phase->transform(new (phase->C) ConvI2LNode(value, ltype)); 4102 } 4103 4104 // Auxiliary method to support randomized stressing/fuzzing. 4105 // 4106 // This method can be called the arbitrary number of times, with current count 4107 // as the argument. The logic allows selecting a single candidate from the 4108 // running list of candidates as follows: 4109 // int count = 0; 4110 // Cand* selected = null; 4111 // while(cand = cand->next()) { 4112 // if (randomized_select(++count)) { 4113 // selected = cand; 4114 // } 4115 // } 4116 // 4117 // Including count equalizes the chances any candidate is "selected". 4118 // This is useful when we don't have the complete list of candidates to choose 4119 // from uniformly. In this case, we need to adjust the randomicity of the 4120 // selection, or else we will end up biasing the selection towards the latter 4121 // candidates. 4122 // 4123 // Quick back-envelope calculation shows that for the list of n candidates 4124 // the equal probability for the candidate to persist as "best" can be 4125 // achieved by replacing it with "next" k-th candidate with the probability 4126 // of 1/k. It can be easily shown that by the end of the run, the 4127 // probability for any candidate is converged to 1/n, thus giving the 4128 // uniform distribution among all the candidates. 4129 // 4130 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4131 #define RANDOMIZED_DOMAIN_POW 29 4132 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4133 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4134 bool Compile::randomized_select(int count) { 4135 assert(count > 0, "only positive"); 4136 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4137 }