1 /* 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "jfr/jfrEvents.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/divnode.hpp" 46 #include "opto/escape.hpp" 47 #include "opto/idealGraphPrinter.hpp" 48 #include "opto/loopnode.hpp" 49 #include "opto/machnode.hpp" 50 #include "opto/macro.hpp" 51 #include "opto/matcher.hpp" 52 #include "opto/mathexactnode.hpp" 53 #include "opto/memnode.hpp" 54 #include "opto/mulnode.hpp" 55 #include "opto/node.hpp" 56 #include "opto/opcodes.hpp" 57 #include "opto/output.hpp" 58 #include "opto/parse.hpp" 59 #include "opto/phaseX.hpp" 60 #include "opto/rootnode.hpp" 61 #include "opto/runtime.hpp" 62 #include "opto/stringopts.hpp" 63 #include "opto/type.hpp" 64 #include "opto/vectornode.hpp" 65 #include "runtime/arguments.hpp" 66 #include "runtime/signature.hpp" 67 #include "runtime/stubRoutines.hpp" 68 #include "runtime/timer.hpp" 69 #include "utilities/copy.hpp" 70 #if defined AD_MD_HPP 71 # include AD_MD_HPP 72 #elif defined TARGET_ARCH_MODEL_x86_32 73 # include "adfiles/ad_x86_32.hpp" 74 #elif defined TARGET_ARCH_MODEL_x86_64 75 # include "adfiles/ad_x86_64.hpp" 76 #elif defined TARGET_ARCH_MODEL_aarch64 77 # include "adfiles/ad_aarch64.hpp" 78 #elif defined TARGET_ARCH_MODEL_sparc 79 # include "adfiles/ad_sparc.hpp" 80 #elif defined TARGET_ARCH_MODEL_zero 81 # include "adfiles/ad_zero.hpp" 82 #elif defined TARGET_ARCH_MODEL_ppc_64 83 # include "adfiles/ad_ppc_64.hpp" 84 #endif 85 86 // -------------------- Compile::mach_constant_base_node ----------------------- 87 // Constant table base node singleton. 88 MachConstantBaseNode* Compile::mach_constant_base_node() { 89 if (_mach_constant_base_node == NULL) { 90 _mach_constant_base_node = new (C) MachConstantBaseNode(); 91 _mach_constant_base_node->add_req(C->root()); 92 } 93 return _mach_constant_base_node; 94 } 95 96 97 /// Support for intrinsics. 98 99 // Return the index at which m must be inserted (or already exists). 100 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 101 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 102 #ifdef ASSERT 103 for (int i = 1; i < _intrinsics->length(); i++) { 104 CallGenerator* cg1 = _intrinsics->at(i-1); 105 CallGenerator* cg2 = _intrinsics->at(i); 106 assert(cg1->method() != cg2->method() 107 ? cg1->method() < cg2->method() 108 : cg1->is_virtual() < cg2->is_virtual(), 109 "compiler intrinsics list must stay sorted"); 110 } 111 #endif 112 // Binary search sorted list, in decreasing intervals [lo, hi]. 113 int lo = 0, hi = _intrinsics->length()-1; 114 while (lo <= hi) { 115 int mid = (uint)(hi + lo) / 2; 116 ciMethod* mid_m = _intrinsics->at(mid)->method(); 117 if (m < mid_m) { 118 hi = mid-1; 119 } else if (m > mid_m) { 120 lo = mid+1; 121 } else { 122 // look at minor sort key 123 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 124 if (is_virtual < mid_virt) { 125 hi = mid-1; 126 } else if (is_virtual > mid_virt) { 127 lo = mid+1; 128 } else { 129 return mid; // exact match 130 } 131 } 132 } 133 return lo; // inexact match 134 } 135 136 void Compile::register_intrinsic(CallGenerator* cg) { 137 if (_intrinsics == NULL) { 138 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 139 } 140 // This code is stolen from ciObjectFactory::insert. 141 // Really, GrowableArray should have methods for 142 // insert_at, remove_at, and binary_search. 143 int len = _intrinsics->length(); 144 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 145 if (index == len) { 146 _intrinsics->append(cg); 147 } else { 148 #ifdef ASSERT 149 CallGenerator* oldcg = _intrinsics->at(index); 150 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 151 #endif 152 _intrinsics->append(_intrinsics->at(len-1)); 153 int pos; 154 for (pos = len-2; pos >= index; pos--) { 155 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 156 } 157 _intrinsics->at_put(index, cg); 158 } 159 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 160 } 161 162 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 163 assert(m->is_loaded(), "don't try this on unloaded methods"); 164 if (_intrinsics != NULL) { 165 int index = intrinsic_insertion_index(m, is_virtual); 166 if (index < _intrinsics->length() 167 && _intrinsics->at(index)->method() == m 168 && _intrinsics->at(index)->is_virtual() == is_virtual) { 169 return _intrinsics->at(index); 170 } 171 } 172 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 173 if (m->intrinsic_id() != vmIntrinsics::_none && 174 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 175 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 176 if (cg != NULL) { 177 // Save it for next time: 178 register_intrinsic(cg); 179 return cg; 180 } else { 181 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 182 } 183 } 184 return NULL; 185 } 186 187 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 188 // in library_call.cpp. 189 190 191 #ifndef PRODUCT 192 // statistics gathering... 193 194 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 195 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 196 197 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 198 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 199 int oflags = _intrinsic_hist_flags[id]; 200 assert(flags != 0, "what happened?"); 201 if (is_virtual) { 202 flags |= _intrinsic_virtual; 203 } 204 bool changed = (flags != oflags); 205 if ((flags & _intrinsic_worked) != 0) { 206 juint count = (_intrinsic_hist_count[id] += 1); 207 if (count == 1) { 208 changed = true; // first time 209 } 210 // increment the overall count also: 211 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 212 } 213 if (changed) { 214 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 215 // Something changed about the intrinsic's virtuality. 216 if ((flags & _intrinsic_virtual) != 0) { 217 // This is the first use of this intrinsic as a virtual call. 218 if (oflags != 0) { 219 // We already saw it as a non-virtual, so note both cases. 220 flags |= _intrinsic_both; 221 } 222 } else if ((oflags & _intrinsic_both) == 0) { 223 // This is the first use of this intrinsic as a non-virtual 224 flags |= _intrinsic_both; 225 } 226 } 227 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 228 } 229 // update the overall flags also: 230 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 231 return changed; 232 } 233 234 static char* format_flags(int flags, char* buf) { 235 buf[0] = 0; 236 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 237 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 238 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 239 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 240 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 241 if (buf[0] == 0) strcat(buf, ","); 242 assert(buf[0] == ',', "must be"); 243 return &buf[1]; 244 } 245 246 void Compile::print_intrinsic_statistics() { 247 char flagsbuf[100]; 248 ttyLocker ttyl; 249 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 250 tty->print_cr("Compiler intrinsic usage:"); 251 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 252 if (total == 0) total = 1; // avoid div0 in case of no successes 253 #define PRINT_STAT_LINE(name, c, f) \ 254 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 255 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 256 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 257 int flags = _intrinsic_hist_flags[id]; 258 juint count = _intrinsic_hist_count[id]; 259 if ((flags | count) != 0) { 260 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 261 } 262 } 263 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 264 if (xtty != NULL) xtty->tail("statistics"); 265 } 266 267 void Compile::print_statistics() { 268 { ttyLocker ttyl; 269 if (xtty != NULL) xtty->head("statistics type='opto'"); 270 Parse::print_statistics(); 271 PhaseCCP::print_statistics(); 272 PhaseRegAlloc::print_statistics(); 273 Scheduling::print_statistics(); 274 PhasePeephole::print_statistics(); 275 PhaseIdealLoop::print_statistics(); 276 if (xtty != NULL) xtty->tail("statistics"); 277 } 278 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 279 // put this under its own <statistics> element. 280 print_intrinsic_statistics(); 281 } 282 } 283 #endif //PRODUCT 284 285 // Support for bundling info 286 Bundle* Compile::node_bundling(const Node *n) { 287 assert(valid_bundle_info(n), "oob"); 288 return &_node_bundling_base[n->_idx]; 289 } 290 291 bool Compile::valid_bundle_info(const Node *n) { 292 return (_node_bundling_limit > n->_idx); 293 } 294 295 296 void Compile::gvn_replace_by(Node* n, Node* nn) { 297 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 298 Node* use = n->last_out(i); 299 bool is_in_table = initial_gvn()->hash_delete(use); 300 uint uses_found = 0; 301 for (uint j = 0; j < use->len(); j++) { 302 if (use->in(j) == n) { 303 if (j < use->req()) 304 use->set_req(j, nn); 305 else 306 use->set_prec(j, nn); 307 uses_found++; 308 } 309 } 310 if (is_in_table) { 311 // reinsert into table 312 initial_gvn()->hash_find_insert(use); 313 } 314 record_for_igvn(use); 315 i -= uses_found; // we deleted 1 or more copies of this edge 316 } 317 } 318 319 320 static inline bool not_a_node(const Node* n) { 321 if (n == NULL) return true; 322 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 323 if (*(address*)n == badAddress) return true; // kill by Node::destruct 324 return false; 325 } 326 327 // Identify all nodes that are reachable from below, useful. 328 // Use breadth-first pass that records state in a Unique_Node_List, 329 // recursive traversal is slower. 330 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 331 int estimated_worklist_size = live_nodes(); 332 useful.map( estimated_worklist_size, NULL ); // preallocate space 333 334 // Initialize worklist 335 if (root() != NULL) { useful.push(root()); } 336 // If 'top' is cached, declare it useful to preserve cached node 337 if( cached_top_node() ) { useful.push(cached_top_node()); } 338 339 // Push all useful nodes onto the list, breadthfirst 340 for( uint next = 0; next < useful.size(); ++next ) { 341 assert( next < unique(), "Unique useful nodes < total nodes"); 342 Node *n = useful.at(next); 343 uint max = n->len(); 344 for( uint i = 0; i < max; ++i ) { 345 Node *m = n->in(i); 346 if (not_a_node(m)) continue; 347 useful.push(m); 348 } 349 } 350 } 351 352 // Update dead_node_list with any missing dead nodes using useful 353 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 354 void Compile::update_dead_node_list(Unique_Node_List &useful) { 355 uint max_idx = unique(); 356 VectorSet& useful_node_set = useful.member_set(); 357 358 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 359 // If node with index node_idx is not in useful set, 360 // mark it as dead in dead node list. 361 if (! useful_node_set.test(node_idx) ) { 362 record_dead_node(node_idx); 363 } 364 } 365 } 366 367 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 368 int shift = 0; 369 for (int i = 0; i < inlines->length(); i++) { 370 CallGenerator* cg = inlines->at(i); 371 CallNode* call = cg->call_node(); 372 if (shift > 0) { 373 inlines->at_put(i-shift, cg); 374 } 375 if (!useful.member(call)) { 376 shift++; 377 } 378 } 379 inlines->trunc_to(inlines->length()-shift); 380 } 381 382 // Disconnect all useless nodes by disconnecting those at the boundary. 383 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 384 uint next = 0; 385 while (next < useful.size()) { 386 Node *n = useful.at(next++); 387 if (n->is_SafePoint()) { 388 // We're done with a parsing phase. Replaced nodes are not valid 389 // beyond that point. 390 n->as_SafePoint()->delete_replaced_nodes(); 391 } 392 // Use raw traversal of out edges since this code removes out edges 393 int max = n->outcnt(); 394 for (int j = 0; j < max; ++j) { 395 Node* child = n->raw_out(j); 396 if (! useful.member(child)) { 397 assert(!child->is_top() || child != top(), 398 "If top is cached in Compile object it is in useful list"); 399 // Only need to remove this out-edge to the useless node 400 n->raw_del_out(j); 401 --j; 402 --max; 403 } 404 } 405 if (n->outcnt() == 1 && n->has_special_unique_user()) { 406 record_for_igvn(n->unique_out()); 407 } 408 } 409 // Remove useless macro and predicate opaq nodes 410 for (int i = C->macro_count()-1; i >= 0; i--) { 411 Node* n = C->macro_node(i); 412 if (!useful.member(n)) { 413 remove_macro_node(n); 414 } 415 } 416 // Remove useless CastII nodes with range check dependency 417 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 418 Node* cast = range_check_cast_node(i); 419 if (!useful.member(cast)) { 420 remove_range_check_cast(cast); 421 } 422 } 423 // Remove useless expensive node 424 for (int i = C->expensive_count()-1; i >= 0; i--) { 425 Node* n = C->expensive_node(i); 426 if (!useful.member(n)) { 427 remove_expensive_node(n); 428 } 429 } 430 // clean up the late inline lists 431 remove_useless_late_inlines(&_string_late_inlines, useful); 432 remove_useless_late_inlines(&_boxing_late_inlines, useful); 433 remove_useless_late_inlines(&_late_inlines, useful); 434 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 435 } 436 437 //------------------------------frame_size_in_words----------------------------- 438 // frame_slots in units of words 439 int Compile::frame_size_in_words() const { 440 // shift is 0 in LP32 and 1 in LP64 441 const int shift = (LogBytesPerWord - LogBytesPerInt); 442 int words = _frame_slots >> shift; 443 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 444 return words; 445 } 446 447 // To bang the stack of this compiled method we use the stack size 448 // that the interpreter would need in case of a deoptimization. This 449 // removes the need to bang the stack in the deoptimization blob which 450 // in turn simplifies stack overflow handling. 451 int Compile::bang_size_in_bytes() const { 452 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 453 } 454 455 // ============================================================================ 456 //------------------------------CompileWrapper--------------------------------- 457 class CompileWrapper : public StackObj { 458 Compile *const _compile; 459 public: 460 CompileWrapper(Compile* compile); 461 462 ~CompileWrapper(); 463 }; 464 465 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 466 // the Compile* pointer is stored in the current ciEnv: 467 ciEnv* env = compile->env(); 468 assert(env == ciEnv::current(), "must already be a ciEnv active"); 469 assert(env->compiler_data() == NULL, "compile already active?"); 470 env->set_compiler_data(compile); 471 assert(compile == Compile::current(), "sanity"); 472 473 compile->set_type_dict(NULL); 474 compile->set_type_hwm(NULL); 475 compile->set_type_last_size(0); 476 compile->set_last_tf(NULL, NULL); 477 compile->set_indexSet_arena(NULL); 478 compile->set_indexSet_free_block_list(NULL); 479 compile->init_type_arena(); 480 Type::Initialize(compile); 481 _compile->set_scratch_buffer_blob(NULL); 482 _compile->begin_method(); 483 } 484 CompileWrapper::~CompileWrapper() { 485 _compile->end_method(); 486 if (_compile->scratch_buffer_blob() != NULL) 487 BufferBlob::free(_compile->scratch_buffer_blob()); 488 _compile->env()->set_compiler_data(NULL); 489 } 490 491 492 //----------------------------print_compile_messages--------------------------- 493 void Compile::print_compile_messages() { 494 #ifndef PRODUCT 495 // Check if recompiling 496 if (_subsume_loads == false && PrintOpto) { 497 // Recompiling without allowing machine instructions to subsume loads 498 tty->print_cr("*********************************************************"); 499 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 500 tty->print_cr("*********************************************************"); 501 } 502 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 503 // Recompiling without escape analysis 504 tty->print_cr("*********************************************************"); 505 tty->print_cr("** Bailout: Recompile without escape analysis **"); 506 tty->print_cr("*********************************************************"); 507 } 508 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 509 // Recompiling without boxing elimination 510 tty->print_cr("*********************************************************"); 511 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 512 tty->print_cr("*********************************************************"); 513 } 514 if (env()->break_at_compile()) { 515 // Open the debugger when compiling this method. 516 tty->print("### Breaking when compiling: "); 517 method()->print_short_name(); 518 tty->cr(); 519 BREAKPOINT; 520 } 521 522 if( PrintOpto ) { 523 if (is_osr_compilation()) { 524 tty->print("[OSR]%3d", _compile_id); 525 } else { 526 tty->print("%3d", _compile_id); 527 } 528 } 529 #endif 530 } 531 532 533 //-----------------------init_scratch_buffer_blob------------------------------ 534 // Construct a temporary BufferBlob and cache it for this compile. 535 void Compile::init_scratch_buffer_blob(int const_size) { 536 // If there is already a scratch buffer blob allocated and the 537 // constant section is big enough, use it. Otherwise free the 538 // current and allocate a new one. 539 BufferBlob* blob = scratch_buffer_blob(); 540 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 541 // Use the current blob. 542 } else { 543 if (blob != NULL) { 544 BufferBlob::free(blob); 545 } 546 547 ResourceMark rm; 548 _scratch_const_size = const_size; 549 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 550 blob = BufferBlob::create("Compile::scratch_buffer", size); 551 // Record the buffer blob for next time. 552 set_scratch_buffer_blob(blob); 553 // Have we run out of code space? 554 if (scratch_buffer_blob() == NULL) { 555 // Let CompilerBroker disable further compilations. 556 record_failure("Not enough space for scratch buffer in CodeCache"); 557 return; 558 } 559 } 560 561 // Initialize the relocation buffers 562 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 563 set_scratch_locs_memory(locs_buf); 564 } 565 566 567 //-----------------------scratch_emit_size------------------------------------- 568 // Helper function that computes size by emitting code 569 uint Compile::scratch_emit_size(const Node* n) { 570 // Start scratch_emit_size section. 571 set_in_scratch_emit_size(true); 572 573 // Emit into a trash buffer and count bytes emitted. 574 // This is a pretty expensive way to compute a size, 575 // but it works well enough if seldom used. 576 // All common fixed-size instructions are given a size 577 // method by the AD file. 578 // Note that the scratch buffer blob and locs memory are 579 // allocated at the beginning of the compile task, and 580 // may be shared by several calls to scratch_emit_size. 581 // The allocation of the scratch buffer blob is particularly 582 // expensive, since it has to grab the code cache lock. 583 BufferBlob* blob = this->scratch_buffer_blob(); 584 assert(blob != NULL, "Initialize BufferBlob at start"); 585 assert(blob->size() > MAX_inst_size, "sanity"); 586 relocInfo* locs_buf = scratch_locs_memory(); 587 address blob_begin = blob->content_begin(); 588 address blob_end = (address)locs_buf; 589 assert(blob->content_contains(blob_end), "sanity"); 590 CodeBuffer buf(blob_begin, blob_end - blob_begin); 591 buf.initialize_consts_size(_scratch_const_size); 592 buf.initialize_stubs_size(MAX_stubs_size); 593 assert(locs_buf != NULL, "sanity"); 594 int lsize = MAX_locs_size / 3; 595 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 596 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 597 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 598 599 // Do the emission. 600 601 Label fakeL; // Fake label for branch instructions. 602 Label* saveL = NULL; 603 uint save_bnum = 0; 604 bool is_branch = n->is_MachBranch(); 605 if (is_branch) { 606 MacroAssembler masm(&buf); 607 masm.bind(fakeL); 608 n->as_MachBranch()->save_label(&saveL, &save_bnum); 609 n->as_MachBranch()->label_set(&fakeL, 0); 610 } 611 n->emit(buf, this->regalloc()); 612 613 // Emitting into the scratch buffer should not fail 614 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 615 616 if (is_branch) // Restore label. 617 n->as_MachBranch()->label_set(saveL, save_bnum); 618 619 // End scratch_emit_size section. 620 set_in_scratch_emit_size(false); 621 622 return buf.insts_size(); 623 } 624 625 626 // ============================================================================ 627 //------------------------------Compile standard------------------------------- 628 debug_only( int Compile::_debug_idx = 100000; ) 629 630 // Compile a method. entry_bci is -1 for normal compilations and indicates 631 // the continuation bci for on stack replacement. 632 633 634 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 635 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 636 : Phase(Compiler), 637 _env(ci_env), 638 _log(ci_env->log()), 639 _compile_id(ci_env->compile_id()), 640 _save_argument_registers(false), 641 _stub_name(NULL), 642 _stub_function(NULL), 643 _stub_entry_point(NULL), 644 _method(target), 645 _entry_bci(osr_bci), 646 _initial_gvn(NULL), 647 _for_igvn(NULL), 648 _warm_calls(NULL), 649 _subsume_loads(subsume_loads), 650 _do_escape_analysis(do_escape_analysis), 651 _eliminate_boxing(eliminate_boxing), 652 _failure_reason(NULL), 653 _code_buffer("Compile::Fill_buffer"), 654 _orig_pc_slot(0), 655 _orig_pc_slot_offset_in_bytes(0), 656 _has_method_handle_invokes(false), 657 _mach_constant_base_node(NULL), 658 _node_bundling_limit(0), 659 _node_bundling_base(NULL), 660 _java_calls(0), 661 _inner_loops(0), 662 _scratch_const_size(-1), 663 _in_scratch_emit_size(false), 664 _dead_node_list(comp_arena()), 665 _dead_node_count(0), 666 #ifndef PRODUCT 667 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 668 _in_dump_cnt(0), 669 _printer(IdealGraphPrinter::printer()), 670 #endif 671 _congraph(NULL), 672 _comp_arena(mtCompiler), 673 _node_arena(mtCompiler), 674 _old_arena(mtCompiler), 675 _Compile_types(mtCompiler), 676 _replay_inline_data(NULL), 677 _late_inlines(comp_arena(), 2, 0, NULL), 678 _string_late_inlines(comp_arena(), 2, 0, NULL), 679 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 680 _late_inlines_pos(0), 681 _number_of_mh_late_inlines(0), 682 _inlining_progress(false), 683 _inlining_incrementally(false), 684 _print_inlining_list(NULL), 685 _print_inlining_idx(0), 686 _interpreter_frame_size(0), 687 _max_node_limit(MaxNodeLimit) { 688 C = this; 689 690 CompileWrapper cw(this); 691 #ifndef PRODUCT 692 if (TimeCompiler2) { 693 tty->print(" "); 694 target->holder()->name()->print(); 695 tty->print("."); 696 target->print_short_name(); 697 tty->print(" "); 698 } 699 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 700 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 701 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 702 if (!print_opto_assembly) { 703 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 704 if (print_assembly && !Disassembler::can_decode()) { 705 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 706 print_opto_assembly = true; 707 } 708 } 709 set_print_assembly(print_opto_assembly); 710 set_parsed_irreducible_loop(false); 711 712 if (method()->has_option("ReplayInline")) { 713 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 714 } 715 #endif 716 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 717 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 718 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 719 720 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 721 // Make sure the method being compiled gets its own MDO, 722 // so we can at least track the decompile_count(). 723 // Need MDO to record RTM code generation state. 724 method()->ensure_method_data(); 725 } 726 727 Init(::AliasLevel); 728 729 730 print_compile_messages(); 731 732 _ilt = InlineTree::build_inline_tree_root(); 733 734 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 735 assert(num_alias_types() >= AliasIdxRaw, ""); 736 737 #define MINIMUM_NODE_HASH 1023 738 // Node list that Iterative GVN will start with 739 Unique_Node_List for_igvn(comp_arena()); 740 set_for_igvn(&for_igvn); 741 742 // GVN that will be run immediately on new nodes 743 uint estimated_size = method()->code_size()*4+64; 744 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 745 PhaseGVN gvn(node_arena(), estimated_size); 746 set_initial_gvn(&gvn); 747 748 if (print_inlining() || print_intrinsics()) { 749 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 750 } 751 { // Scope for timing the parser 752 TracePhase t3("parse", &_t_parser, true); 753 754 // Put top into the hash table ASAP. 755 initial_gvn()->transform_no_reclaim(top()); 756 757 // Set up tf(), start(), and find a CallGenerator. 758 CallGenerator* cg = NULL; 759 if (is_osr_compilation()) { 760 const TypeTuple *domain = StartOSRNode::osr_domain(); 761 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 762 init_tf(TypeFunc::make(domain, range)); 763 StartNode* s = new (this) StartOSRNode(root(), domain); 764 initial_gvn()->set_type_bottom(s); 765 init_start(s); 766 cg = CallGenerator::for_osr(method(), entry_bci()); 767 } else { 768 // Normal case. 769 init_tf(TypeFunc::make(method())); 770 StartNode* s = new (this) StartNode(root(), tf()->domain()); 771 initial_gvn()->set_type_bottom(s); 772 init_start(s); 773 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 774 // With java.lang.ref.reference.get() we must go through the 775 // intrinsic when G1 is enabled - even when get() is the root 776 // method of the compile - so that, if necessary, the value in 777 // the referent field of the reference object gets recorded by 778 // the pre-barrier code. 779 // Specifically, if G1 is enabled, the value in the referent 780 // field is recorded by the G1 SATB pre barrier. This will 781 // result in the referent being marked live and the reference 782 // object removed from the list of discovered references during 783 // reference processing. 784 cg = find_intrinsic(method(), false); 785 } 786 if (cg == NULL) { 787 float past_uses = method()->interpreter_invocation_count(); 788 float expected_uses = past_uses; 789 cg = CallGenerator::for_inline(method(), expected_uses); 790 } 791 } 792 if (failing()) return; 793 if (cg == NULL) { 794 record_method_not_compilable_all_tiers("cannot parse method"); 795 return; 796 } 797 JVMState* jvms = build_start_state(start(), tf()); 798 if ((jvms = cg->generate(jvms)) == NULL) { 799 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 800 record_method_not_compilable("method parse failed"); 801 } 802 return; 803 } 804 GraphKit kit(jvms); 805 806 if (!kit.stopped()) { 807 // Accept return values, and transfer control we know not where. 808 // This is done by a special, unique ReturnNode bound to root. 809 return_values(kit.jvms()); 810 } 811 812 if (kit.has_exceptions()) { 813 // Any exceptions that escape from this call must be rethrown 814 // to whatever caller is dynamically above us on the stack. 815 // This is done by a special, unique RethrowNode bound to root. 816 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 817 } 818 819 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 820 821 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 822 inline_string_calls(true); 823 } 824 825 if (failing()) return; 826 827 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 828 829 // Remove clutter produced by parsing. 830 if (!failing()) { 831 ResourceMark rm; 832 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 833 } 834 } 835 836 // Note: Large methods are capped off in do_one_bytecode(). 837 if (failing()) return; 838 839 // After parsing, node notes are no longer automagic. 840 // They must be propagated by register_new_node_with_optimizer(), 841 // clone(), or the like. 842 set_default_node_notes(NULL); 843 844 for (;;) { 845 int successes = Inline_Warm(); 846 if (failing()) return; 847 if (successes == 0) break; 848 } 849 850 // Drain the list. 851 Finish_Warm(); 852 #ifndef PRODUCT 853 if (_printer) { 854 _printer->print_inlining(this); 855 } 856 #endif 857 858 if (failing()) return; 859 NOT_PRODUCT( verify_graph_edges(); ) 860 861 // Now optimize 862 Optimize(); 863 if (failing()) return; 864 NOT_PRODUCT( verify_graph_edges(); ) 865 866 #ifndef PRODUCT 867 if (PrintIdeal) { 868 ttyLocker ttyl; // keep the following output all in one block 869 // This output goes directly to the tty, not the compiler log. 870 // To enable tools to match it up with the compilation activity, 871 // be sure to tag this tty output with the compile ID. 872 if (xtty != NULL) { 873 xtty->head("ideal compile_id='%d'%s", compile_id(), 874 is_osr_compilation() ? " compile_kind='osr'" : 875 ""); 876 } 877 root()->dump(9999); 878 if (xtty != NULL) { 879 xtty->tail("ideal"); 880 } 881 } 882 #endif 883 884 NOT_PRODUCT( verify_barriers(); ) 885 886 // Dump compilation data to replay it. 887 if (method()->has_option("DumpReplay")) { 888 env()->dump_replay_data(_compile_id); 889 } 890 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 891 env()->dump_inline_data(_compile_id); 892 } 893 894 // Now that we know the size of all the monitors we can add a fixed slot 895 // for the original deopt pc. 896 897 _orig_pc_slot = fixed_slots(); 898 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 899 set_fixed_slots(next_slot); 900 901 // Compute when to use implicit null checks. Used by matching trap based 902 // nodes and NullCheck optimization. 903 set_allowed_deopt_reasons(); 904 905 // Now generate code 906 Code_Gen(); 907 if (failing()) return; 908 909 // Check if we want to skip execution of all compiled code. 910 { 911 #ifndef PRODUCT 912 if (OptoNoExecute) { 913 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 914 return; 915 } 916 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 917 #endif 918 919 if (is_osr_compilation()) { 920 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 921 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 922 } else { 923 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 924 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 925 } 926 927 env()->register_method(_method, _entry_bci, 928 &_code_offsets, 929 _orig_pc_slot_offset_in_bytes, 930 code_buffer(), 931 frame_size_in_words(), _oop_map_set, 932 &_handler_table, &_inc_table, 933 compiler, 934 env()->comp_level(), 935 has_unsafe_access(), 936 SharedRuntime::is_wide_vector(max_vector_size()), 937 rtm_state() 938 ); 939 940 if (log() != NULL) // Print code cache state into compiler log 941 log()->code_cache_state(); 942 } 943 } 944 945 //------------------------------Compile---------------------------------------- 946 // Compile a runtime stub 947 Compile::Compile( ciEnv* ci_env, 948 TypeFunc_generator generator, 949 address stub_function, 950 const char *stub_name, 951 int is_fancy_jump, 952 bool pass_tls, 953 bool save_arg_registers, 954 bool return_pc ) 955 : Phase(Compiler), 956 _env(ci_env), 957 _log(ci_env->log()), 958 _compile_id(0), 959 _save_argument_registers(save_arg_registers), 960 _method(NULL), 961 _stub_name(stub_name), 962 _stub_function(stub_function), 963 _stub_entry_point(NULL), 964 _entry_bci(InvocationEntryBci), 965 _initial_gvn(NULL), 966 _for_igvn(NULL), 967 _warm_calls(NULL), 968 _orig_pc_slot(0), 969 _orig_pc_slot_offset_in_bytes(0), 970 _subsume_loads(true), 971 _do_escape_analysis(false), 972 _eliminate_boxing(false), 973 _failure_reason(NULL), 974 _code_buffer("Compile::Fill_buffer"), 975 _has_method_handle_invokes(false), 976 _mach_constant_base_node(NULL), 977 _node_bundling_limit(0), 978 _node_bundling_base(NULL), 979 _java_calls(0), 980 _inner_loops(0), 981 #ifndef PRODUCT 982 _trace_opto_output(TraceOptoOutput), 983 _in_dump_cnt(0), 984 _printer(NULL), 985 #endif 986 _comp_arena(mtCompiler), 987 _node_arena(mtCompiler), 988 _old_arena(mtCompiler), 989 _Compile_types(mtCompiler), 990 _dead_node_list(comp_arena()), 991 _dead_node_count(0), 992 _congraph(NULL), 993 _replay_inline_data(NULL), 994 _number_of_mh_late_inlines(0), 995 _inlining_progress(false), 996 _inlining_incrementally(false), 997 _print_inlining_list(NULL), 998 _print_inlining_idx(0), 999 _allowed_reasons(0), 1000 _interpreter_frame_size(0), 1001 _max_node_limit(MaxNodeLimit) { 1002 C = this; 1003 1004 #ifndef PRODUCT 1005 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 1006 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 1007 set_print_assembly(PrintFrameConverterAssembly); 1008 set_parsed_irreducible_loop(false); 1009 #endif 1010 set_has_irreducible_loop(false); // no loops 1011 1012 CompileWrapper cw(this); 1013 Init(/*AliasLevel=*/ 0); 1014 init_tf((*generator)()); 1015 1016 { 1017 // The following is a dummy for the sake of GraphKit::gen_stub 1018 Unique_Node_List for_igvn(comp_arena()); 1019 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1020 PhaseGVN gvn(Thread::current()->resource_area(),255); 1021 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1022 gvn.transform_no_reclaim(top()); 1023 1024 GraphKit kit; 1025 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1026 } 1027 1028 NOT_PRODUCT( verify_graph_edges(); ) 1029 Code_Gen(); 1030 if (failing()) return; 1031 1032 1033 // Entry point will be accessed using compile->stub_entry_point(); 1034 if (code_buffer() == NULL) { 1035 Matcher::soft_match_failure(); 1036 } else { 1037 if (PrintAssembly && (WizardMode || Verbose)) 1038 tty->print_cr("### Stub::%s", stub_name); 1039 1040 if (!failing()) { 1041 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1042 1043 // Make the NMethod 1044 // For now we mark the frame as never safe for profile stackwalking 1045 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1046 code_buffer(), 1047 CodeOffsets::frame_never_safe, 1048 // _code_offsets.value(CodeOffsets::Frame_Complete), 1049 frame_size_in_words(), 1050 _oop_map_set, 1051 save_arg_registers); 1052 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1053 1054 _stub_entry_point = rs->entry_point(); 1055 } 1056 } 1057 } 1058 1059 //------------------------------Init------------------------------------------- 1060 // Prepare for a single compilation 1061 void Compile::Init(int aliaslevel) { 1062 _unique = 0; 1063 _regalloc = NULL; 1064 1065 _tf = NULL; // filled in later 1066 _top = NULL; // cached later 1067 _matcher = NULL; // filled in later 1068 _cfg = NULL; // filled in later 1069 1070 set_24_bit_selection_and_mode(Use24BitFP, false); 1071 1072 _node_note_array = NULL; 1073 _default_node_notes = NULL; 1074 1075 _immutable_memory = NULL; // filled in at first inquiry 1076 1077 // Globally visible Nodes 1078 // First set TOP to NULL to give safe behavior during creation of RootNode 1079 set_cached_top_node(NULL); 1080 set_root(new (this) RootNode()); 1081 // Now that you have a Root to point to, create the real TOP 1082 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1083 set_recent_alloc(NULL, NULL); 1084 1085 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1086 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1087 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1088 env()->set_dependencies(new Dependencies(env())); 1089 1090 _fixed_slots = 0; 1091 set_has_split_ifs(false); 1092 set_has_loops(has_method() && method()->has_loops()); // first approximation 1093 set_has_stringbuilder(false); 1094 set_has_boxed_value(false); 1095 _trap_can_recompile = false; // no traps emitted yet 1096 _major_progress = true; // start out assuming good things will happen 1097 set_has_unsafe_access(false); 1098 set_max_vector_size(0); 1099 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1100 set_decompile_count(0); 1101 1102 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1103 set_num_loop_opts(LoopOptsCount); 1104 set_do_inlining(Inline); 1105 set_max_inline_size(MaxInlineSize); 1106 set_freq_inline_size(FreqInlineSize); 1107 set_do_scheduling(OptoScheduling); 1108 set_do_count_invocations(false); 1109 set_do_method_data_update(false); 1110 set_rtm_state(NoRTM); // No RTM lock eliding by default 1111 method_has_option_value("MaxNodeLimit", _max_node_limit); 1112 #if INCLUDE_RTM_OPT 1113 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1114 int rtm_state = method()->method_data()->rtm_state(); 1115 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1116 // Don't generate RTM lock eliding code. 1117 set_rtm_state(NoRTM); 1118 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1119 // Generate RTM lock eliding code without abort ratio calculation code. 1120 set_rtm_state(UseRTM); 1121 } else if (UseRTMDeopt) { 1122 // Generate RTM lock eliding code and include abort ratio calculation 1123 // code if UseRTMDeopt is on. 1124 set_rtm_state(ProfileRTM); 1125 } 1126 } 1127 #endif 1128 if (debug_info()->recording_non_safepoints()) { 1129 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1130 (comp_arena(), 8, 0, NULL)); 1131 set_default_node_notes(Node_Notes::make(this)); 1132 } 1133 1134 // // -- Initialize types before each compile -- 1135 // // Update cached type information 1136 // if( _method && _method->constants() ) 1137 // Type::update_loaded_types(_method, _method->constants()); 1138 1139 // Init alias_type map. 1140 if (!_do_escape_analysis && aliaslevel == 3) 1141 aliaslevel = 2; // No unique types without escape analysis 1142 _AliasLevel = aliaslevel; 1143 const int grow_ats = 16; 1144 _max_alias_types = grow_ats; 1145 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1146 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1147 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1148 { 1149 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1150 } 1151 // Initialize the first few types. 1152 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1153 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1154 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1155 _num_alias_types = AliasIdxRaw+1; 1156 // Zero out the alias type cache. 1157 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1158 // A NULL adr_type hits in the cache right away. Preload the right answer. 1159 probe_alias_cache(NULL)->_index = AliasIdxTop; 1160 1161 _intrinsics = NULL; 1162 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1163 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1164 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1165 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1166 register_library_intrinsics(); 1167 #ifdef ASSERT 1168 _type_verify_symmetry = true; 1169 #endif 1170 } 1171 1172 //---------------------------init_start---------------------------------------- 1173 // Install the StartNode on this compile object. 1174 void Compile::init_start(StartNode* s) { 1175 if (failing()) 1176 return; // already failing 1177 assert(s == start(), ""); 1178 } 1179 1180 StartNode* Compile::start() const { 1181 assert(!failing(), ""); 1182 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1183 Node* start = root()->fast_out(i); 1184 if( start->is_Start() ) 1185 return start->as_Start(); 1186 } 1187 fatal("Did not find Start node!"); 1188 return NULL; 1189 } 1190 1191 //-------------------------------immutable_memory------------------------------------- 1192 // Access immutable memory 1193 Node* Compile::immutable_memory() { 1194 if (_immutable_memory != NULL) { 1195 return _immutable_memory; 1196 } 1197 StartNode* s = start(); 1198 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1199 Node *p = s->fast_out(i); 1200 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1201 _immutable_memory = p; 1202 return _immutable_memory; 1203 } 1204 } 1205 ShouldNotReachHere(); 1206 return NULL; 1207 } 1208 1209 //----------------------set_cached_top_node------------------------------------ 1210 // Install the cached top node, and make sure Node::is_top works correctly. 1211 void Compile::set_cached_top_node(Node* tn) { 1212 if (tn != NULL) verify_top(tn); 1213 Node* old_top = _top; 1214 _top = tn; 1215 // Calling Node::setup_is_top allows the nodes the chance to adjust 1216 // their _out arrays. 1217 if (_top != NULL) _top->setup_is_top(); 1218 if (old_top != NULL) old_top->setup_is_top(); 1219 assert(_top == NULL || top()->is_top(), ""); 1220 } 1221 1222 #ifdef ASSERT 1223 uint Compile::count_live_nodes_by_graph_walk() { 1224 Unique_Node_List useful(comp_arena()); 1225 // Get useful node list by walking the graph. 1226 identify_useful_nodes(useful); 1227 return useful.size(); 1228 } 1229 1230 void Compile::print_missing_nodes() { 1231 1232 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1233 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1234 return; 1235 } 1236 1237 // This is an expensive function. It is executed only when the user 1238 // specifies VerifyIdealNodeCount option or otherwise knows the 1239 // additional work that needs to be done to identify reachable nodes 1240 // by walking the flow graph and find the missing ones using 1241 // _dead_node_list. 1242 1243 Unique_Node_List useful(comp_arena()); 1244 // Get useful node list by walking the graph. 1245 identify_useful_nodes(useful); 1246 1247 uint l_nodes = C->live_nodes(); 1248 uint l_nodes_by_walk = useful.size(); 1249 1250 if (l_nodes != l_nodes_by_walk) { 1251 if (_log != NULL) { 1252 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1253 _log->stamp(); 1254 _log->end_head(); 1255 } 1256 VectorSet& useful_member_set = useful.member_set(); 1257 int last_idx = l_nodes_by_walk; 1258 for (int i = 0; i < last_idx; i++) { 1259 if (useful_member_set.test(i)) { 1260 if (_dead_node_list.test(i)) { 1261 if (_log != NULL) { 1262 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1263 } 1264 if (PrintIdealNodeCount) { 1265 // Print the log message to tty 1266 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1267 useful.at(i)->dump(); 1268 } 1269 } 1270 } 1271 else if (! _dead_node_list.test(i)) { 1272 if (_log != NULL) { 1273 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1274 } 1275 if (PrintIdealNodeCount) { 1276 // Print the log message to tty 1277 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1278 } 1279 } 1280 } 1281 if (_log != NULL) { 1282 _log->tail("mismatched_nodes"); 1283 } 1284 } 1285 } 1286 #endif 1287 1288 #ifndef PRODUCT 1289 void Compile::verify_top(Node* tn) const { 1290 if (tn != NULL) { 1291 assert(tn->is_Con(), "top node must be a constant"); 1292 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1293 assert(tn->in(0) != NULL, "must have live top node"); 1294 } 1295 } 1296 #endif 1297 1298 1299 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1300 1301 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1302 guarantee(arr != NULL, ""); 1303 int num_blocks = arr->length(); 1304 if (grow_by < num_blocks) grow_by = num_blocks; 1305 int num_notes = grow_by * _node_notes_block_size; 1306 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1307 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1308 while (num_notes > 0) { 1309 arr->append(notes); 1310 notes += _node_notes_block_size; 1311 num_notes -= _node_notes_block_size; 1312 } 1313 assert(num_notes == 0, "exact multiple, please"); 1314 } 1315 1316 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1317 if (source == NULL || dest == NULL) return false; 1318 1319 if (dest->is_Con()) 1320 return false; // Do not push debug info onto constants. 1321 1322 #ifdef ASSERT 1323 // Leave a bread crumb trail pointing to the original node: 1324 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1325 dest->set_debug_orig(source); 1326 } 1327 #endif 1328 1329 if (node_note_array() == NULL) 1330 return false; // Not collecting any notes now. 1331 1332 // This is a copy onto a pre-existing node, which may already have notes. 1333 // If both nodes have notes, do not overwrite any pre-existing notes. 1334 Node_Notes* source_notes = node_notes_at(source->_idx); 1335 if (source_notes == NULL || source_notes->is_clear()) return false; 1336 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1337 if (dest_notes == NULL || dest_notes->is_clear()) { 1338 return set_node_notes_at(dest->_idx, source_notes); 1339 } 1340 1341 Node_Notes merged_notes = (*source_notes); 1342 // The order of operations here ensures that dest notes will win... 1343 merged_notes.update_from(dest_notes); 1344 return set_node_notes_at(dest->_idx, &merged_notes); 1345 } 1346 1347 1348 //--------------------------allow_range_check_smearing------------------------- 1349 // Gating condition for coalescing similar range checks. 1350 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1351 // single covering check that is at least as strong as any of them. 1352 // If the optimization succeeds, the simplified (strengthened) range check 1353 // will always succeed. If it fails, we will deopt, and then give up 1354 // on the optimization. 1355 bool Compile::allow_range_check_smearing() const { 1356 // If this method has already thrown a range-check, 1357 // assume it was because we already tried range smearing 1358 // and it failed. 1359 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1360 return !already_trapped; 1361 } 1362 1363 1364 //------------------------------flatten_alias_type----------------------------- 1365 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1366 int offset = tj->offset(); 1367 TypePtr::PTR ptr = tj->ptr(); 1368 1369 // Known instance (scalarizable allocation) alias only with itself. 1370 bool is_known_inst = tj->isa_oopptr() != NULL && 1371 tj->is_oopptr()->is_known_instance(); 1372 1373 // Process weird unsafe references. 1374 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1375 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1376 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1377 tj = TypeOopPtr::BOTTOM; 1378 ptr = tj->ptr(); 1379 offset = tj->offset(); 1380 } 1381 1382 // Array pointers need some flattening 1383 const TypeAryPtr *ta = tj->isa_aryptr(); 1384 if (ta && ta->is_stable()) { 1385 // Erase stability property for alias analysis. 1386 tj = ta = ta->cast_to_stable(false); 1387 } 1388 if( ta && is_known_inst ) { 1389 if ( offset != Type::OffsetBot && 1390 offset > arrayOopDesc::length_offset_in_bytes() ) { 1391 offset = Type::OffsetBot; // Flatten constant access into array body only 1392 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1393 } 1394 } else if( ta && _AliasLevel >= 2 ) { 1395 // For arrays indexed by constant indices, we flatten the alias 1396 // space to include all of the array body. Only the header, klass 1397 // and array length can be accessed un-aliased. 1398 if( offset != Type::OffsetBot ) { 1399 if( ta->const_oop() ) { // MethodData* or Method* 1400 offset = Type::OffsetBot; // Flatten constant access into array body 1401 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1402 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1403 // range is OK as-is. 1404 tj = ta = TypeAryPtr::RANGE; 1405 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1406 tj = TypeInstPtr::KLASS; // all klass loads look alike 1407 ta = TypeAryPtr::RANGE; // generic ignored junk 1408 ptr = TypePtr::BotPTR; 1409 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1410 tj = TypeInstPtr::MARK; 1411 ta = TypeAryPtr::RANGE; // generic ignored junk 1412 ptr = TypePtr::BotPTR; 1413 } else { // Random constant offset into array body 1414 offset = Type::OffsetBot; // Flatten constant access into array body 1415 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1416 } 1417 } 1418 // Arrays of fixed size alias with arrays of unknown size. 1419 if (ta->size() != TypeInt::POS) { 1420 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1421 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1422 } 1423 // Arrays of known objects become arrays of unknown objects. 1424 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1425 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1426 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1427 } 1428 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1429 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1430 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1431 } 1432 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1433 // cannot be distinguished by bytecode alone. 1434 if (ta->elem() == TypeInt::BOOL) { 1435 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1436 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1437 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1438 } 1439 // During the 2nd round of IterGVN, NotNull castings are removed. 1440 // Make sure the Bottom and NotNull variants alias the same. 1441 // Also, make sure exact and non-exact variants alias the same. 1442 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1443 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1444 } 1445 } 1446 1447 // Oop pointers need some flattening 1448 const TypeInstPtr *to = tj->isa_instptr(); 1449 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1450 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1451 if( ptr == TypePtr::Constant ) { 1452 if (to->klass() != ciEnv::current()->Class_klass() || 1453 offset < k->size_helper() * wordSize) { 1454 // No constant oop pointers (such as Strings); they alias with 1455 // unknown strings. 1456 assert(!is_known_inst, "not scalarizable allocation"); 1457 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1458 } 1459 } else if( is_known_inst ) { 1460 tj = to; // Keep NotNull and klass_is_exact for instance type 1461 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1462 // During the 2nd round of IterGVN, NotNull castings are removed. 1463 // Make sure the Bottom and NotNull variants alias the same. 1464 // Also, make sure exact and non-exact variants alias the same. 1465 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1466 } 1467 if (to->speculative() != NULL) { 1468 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1469 } 1470 // Canonicalize the holder of this field 1471 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1472 // First handle header references such as a LoadKlassNode, even if the 1473 // object's klass is unloaded at compile time (4965979). 1474 if (!is_known_inst) { // Do it only for non-instance types 1475 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1476 } 1477 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1478 // Static fields are in the space above the normal instance 1479 // fields in the java.lang.Class instance. 1480 if (to->klass() != ciEnv::current()->Class_klass()) { 1481 to = NULL; 1482 tj = TypeOopPtr::BOTTOM; 1483 offset = tj->offset(); 1484 } 1485 } else { 1486 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1487 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1488 if( is_known_inst ) { 1489 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1490 } else { 1491 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1492 } 1493 } 1494 } 1495 } 1496 1497 // Klass pointers to object array klasses need some flattening 1498 const TypeKlassPtr *tk = tj->isa_klassptr(); 1499 if( tk ) { 1500 // If we are referencing a field within a Klass, we need 1501 // to assume the worst case of an Object. Both exact and 1502 // inexact types must flatten to the same alias class so 1503 // use NotNull as the PTR. 1504 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1505 1506 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1507 TypeKlassPtr::OBJECT->klass(), 1508 offset); 1509 } 1510 1511 ciKlass* klass = tk->klass(); 1512 if( klass->is_obj_array_klass() ) { 1513 ciKlass* k = TypeAryPtr::OOPS->klass(); 1514 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1515 k = TypeInstPtr::BOTTOM->klass(); 1516 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1517 } 1518 1519 // Check for precise loads from the primary supertype array and force them 1520 // to the supertype cache alias index. Check for generic array loads from 1521 // the primary supertype array and also force them to the supertype cache 1522 // alias index. Since the same load can reach both, we need to merge 1523 // these 2 disparate memories into the same alias class. Since the 1524 // primary supertype array is read-only, there's no chance of confusion 1525 // where we bypass an array load and an array store. 1526 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1527 if (offset == Type::OffsetBot || 1528 (offset >= primary_supers_offset && 1529 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1530 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1531 offset = in_bytes(Klass::secondary_super_cache_offset()); 1532 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1533 } 1534 } 1535 1536 // Flatten all Raw pointers together. 1537 if (tj->base() == Type::RawPtr) 1538 tj = TypeRawPtr::BOTTOM; 1539 1540 if (tj->base() == Type::AnyPtr) 1541 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1542 1543 // Flatten all to bottom for now 1544 switch( _AliasLevel ) { 1545 case 0: 1546 tj = TypePtr::BOTTOM; 1547 break; 1548 case 1: // Flatten to: oop, static, field or array 1549 switch (tj->base()) { 1550 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1551 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1552 case Type::AryPtr: // do not distinguish arrays at all 1553 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1554 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1555 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1556 default: ShouldNotReachHere(); 1557 } 1558 break; 1559 case 2: // No collapsing at level 2; keep all splits 1560 case 3: // No collapsing at level 3; keep all splits 1561 break; 1562 default: 1563 Unimplemented(); 1564 } 1565 1566 offset = tj->offset(); 1567 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1568 1569 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1570 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1571 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1572 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1573 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1574 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1575 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1576 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1577 assert( tj->ptr() != TypePtr::TopPTR && 1578 tj->ptr() != TypePtr::AnyNull && 1579 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1580 // assert( tj->ptr() != TypePtr::Constant || 1581 // tj->base() == Type::RawPtr || 1582 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1583 1584 return tj; 1585 } 1586 1587 void Compile::AliasType::Init(int i, const TypePtr* at) { 1588 _index = i; 1589 _adr_type = at; 1590 _field = NULL; 1591 _element = NULL; 1592 _is_rewritable = true; // default 1593 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1594 if (atoop != NULL && atoop->is_known_instance()) { 1595 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1596 _general_index = Compile::current()->get_alias_index(gt); 1597 } else { 1598 _general_index = 0; 1599 } 1600 } 1601 1602 BasicType Compile::AliasType::basic_type() const { 1603 if (element() != NULL) { 1604 const Type* element = adr_type()->is_aryptr()->elem(); 1605 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1606 } if (field() != NULL) { 1607 return field()->layout_type(); 1608 } else { 1609 return T_ILLEGAL; // unknown 1610 } 1611 } 1612 1613 //---------------------------------print_on------------------------------------ 1614 #ifndef PRODUCT 1615 void Compile::AliasType::print_on(outputStream* st) { 1616 if (index() < 10) 1617 st->print("@ <%d> ", index()); 1618 else st->print("@ <%d>", index()); 1619 st->print(is_rewritable() ? " " : " RO"); 1620 int offset = adr_type()->offset(); 1621 if (offset == Type::OffsetBot) 1622 st->print(" +any"); 1623 else st->print(" +%-3d", offset); 1624 st->print(" in "); 1625 adr_type()->dump_on(st); 1626 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1627 if (field() != NULL && tjp) { 1628 if (tjp->klass() != field()->holder() || 1629 tjp->offset() != field()->offset_in_bytes()) { 1630 st->print(" != "); 1631 field()->print(); 1632 st->print(" ***"); 1633 } 1634 } 1635 } 1636 1637 void print_alias_types() { 1638 Compile* C = Compile::current(); 1639 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1640 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1641 C->alias_type(idx)->print_on(tty); 1642 tty->cr(); 1643 } 1644 } 1645 #endif 1646 1647 1648 //----------------------------probe_alias_cache-------------------------------- 1649 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1650 intptr_t key = (intptr_t) adr_type; 1651 key ^= key >> logAliasCacheSize; 1652 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1653 } 1654 1655 1656 //-----------------------------grow_alias_types-------------------------------- 1657 void Compile::grow_alias_types() { 1658 const int old_ats = _max_alias_types; // how many before? 1659 const int new_ats = old_ats; // how many more? 1660 const int grow_ats = old_ats+new_ats; // how many now? 1661 _max_alias_types = grow_ats; 1662 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1663 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1664 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1665 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1666 } 1667 1668 1669 //--------------------------------find_alias_type------------------------------ 1670 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1671 if (_AliasLevel == 0) 1672 return alias_type(AliasIdxBot); 1673 1674 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1675 if (ace->_adr_type == adr_type) { 1676 return alias_type(ace->_index); 1677 } 1678 1679 // Handle special cases. 1680 if (adr_type == NULL) return alias_type(AliasIdxTop); 1681 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1682 1683 // Do it the slow way. 1684 const TypePtr* flat = flatten_alias_type(adr_type); 1685 1686 #ifdef ASSERT 1687 { 1688 ResourceMark rm; 1689 assert(flat == flatten_alias_type(flat), 1690 err_msg("not idempotent: adr_type = %s; flat = %s => %s", Type::str(adr_type), 1691 Type::str(flat), Type::str(flatten_alias_type(flat)))); 1692 assert(flat != TypePtr::BOTTOM, 1693 err_msg("cannot alias-analyze an untyped ptr: adr_type = %s", Type::str(adr_type))); 1694 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1695 const TypeOopPtr* foop = flat->is_oopptr(); 1696 // Scalarizable allocations have exact klass always. 1697 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1698 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1699 assert(foop == flatten_alias_type(xoop), 1700 err_msg("exactness must not affect alias type: foop = %s; xoop = %s", 1701 Type::str(foop), Type::str(xoop))); 1702 } 1703 } 1704 #endif 1705 1706 int idx = AliasIdxTop; 1707 for (int i = 0; i < num_alias_types(); i++) { 1708 if (alias_type(i)->adr_type() == flat) { 1709 idx = i; 1710 break; 1711 } 1712 } 1713 1714 if (idx == AliasIdxTop) { 1715 if (no_create) return NULL; 1716 // Grow the array if necessary. 1717 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1718 // Add a new alias type. 1719 idx = _num_alias_types++; 1720 _alias_types[idx]->Init(idx, flat); 1721 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1722 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1723 if (flat->isa_instptr()) { 1724 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1725 && flat->is_instptr()->klass() == env()->Class_klass()) 1726 alias_type(idx)->set_rewritable(false); 1727 } 1728 if (flat->isa_aryptr()) { 1729 #ifdef ASSERT 1730 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1731 // (T_BYTE has the weakest alignment and size restrictions...) 1732 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1733 #endif 1734 if (flat->offset() == TypePtr::OffsetBot) { 1735 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1736 } 1737 } 1738 if (flat->isa_klassptr()) { 1739 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1740 alias_type(idx)->set_rewritable(false); 1741 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1742 alias_type(idx)->set_rewritable(false); 1743 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1744 alias_type(idx)->set_rewritable(false); 1745 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1746 alias_type(idx)->set_rewritable(false); 1747 } 1748 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1749 // but the base pointer type is not distinctive enough to identify 1750 // references into JavaThread.) 1751 1752 // Check for final fields. 1753 const TypeInstPtr* tinst = flat->isa_instptr(); 1754 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1755 ciField* field; 1756 if (tinst->const_oop() != NULL && 1757 tinst->klass() == ciEnv::current()->Class_klass() && 1758 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1759 // static field 1760 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1761 field = k->get_field_by_offset(tinst->offset(), true); 1762 } else { 1763 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1764 field = k->get_field_by_offset(tinst->offset(), false); 1765 } 1766 assert(field == NULL || 1767 original_field == NULL || 1768 (field->holder() == original_field->holder() && 1769 field->offset() == original_field->offset() && 1770 field->is_static() == original_field->is_static()), "wrong field?"); 1771 // Set field() and is_rewritable() attributes. 1772 if (field != NULL) alias_type(idx)->set_field(field); 1773 } 1774 } 1775 1776 // Fill the cache for next time. 1777 ace->_adr_type = adr_type; 1778 ace->_index = idx; 1779 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1780 1781 // Might as well try to fill the cache for the flattened version, too. 1782 AliasCacheEntry* face = probe_alias_cache(flat); 1783 if (face->_adr_type == NULL) { 1784 face->_adr_type = flat; 1785 face->_index = idx; 1786 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1787 } 1788 1789 return alias_type(idx); 1790 } 1791 1792 1793 Compile::AliasType* Compile::alias_type(ciField* field) { 1794 const TypeOopPtr* t; 1795 if (field->is_static()) 1796 t = TypeInstPtr::make(field->holder()->java_mirror()); 1797 else 1798 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1799 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1800 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1801 return atp; 1802 } 1803 1804 1805 //------------------------------have_alias_type-------------------------------- 1806 bool Compile::have_alias_type(const TypePtr* adr_type) { 1807 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1808 if (ace->_adr_type == adr_type) { 1809 return true; 1810 } 1811 1812 // Handle special cases. 1813 if (adr_type == NULL) return true; 1814 if (adr_type == TypePtr::BOTTOM) return true; 1815 1816 return find_alias_type(adr_type, true, NULL) != NULL; 1817 } 1818 1819 //-----------------------------must_alias-------------------------------------- 1820 // True if all values of the given address type are in the given alias category. 1821 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1822 if (alias_idx == AliasIdxBot) return true; // the universal category 1823 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1824 if (alias_idx == AliasIdxTop) return false; // the empty category 1825 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1826 1827 // the only remaining possible overlap is identity 1828 int adr_idx = get_alias_index(adr_type); 1829 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1830 assert(adr_idx == alias_idx || 1831 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1832 && adr_type != TypeOopPtr::BOTTOM), 1833 "should not be testing for overlap with an unsafe pointer"); 1834 return adr_idx == alias_idx; 1835 } 1836 1837 //------------------------------can_alias-------------------------------------- 1838 // True if any values of the given address type are in the given alias category. 1839 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1840 if (alias_idx == AliasIdxTop) return false; // the empty category 1841 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1842 if (alias_idx == AliasIdxBot) return true; // the universal category 1843 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1844 1845 // the only remaining possible overlap is identity 1846 int adr_idx = get_alias_index(adr_type); 1847 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1848 return adr_idx == alias_idx; 1849 } 1850 1851 1852 1853 //---------------------------pop_warm_call------------------------------------- 1854 WarmCallInfo* Compile::pop_warm_call() { 1855 WarmCallInfo* wci = _warm_calls; 1856 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1857 return wci; 1858 } 1859 1860 //----------------------------Inline_Warm-------------------------------------- 1861 int Compile::Inline_Warm() { 1862 // If there is room, try to inline some more warm call sites. 1863 // %%% Do a graph index compaction pass when we think we're out of space? 1864 if (!InlineWarmCalls) return 0; 1865 1866 int calls_made_hot = 0; 1867 int room_to_grow = NodeCountInliningCutoff - unique(); 1868 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1869 int amount_grown = 0; 1870 WarmCallInfo* call; 1871 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1872 int est_size = (int)call->size(); 1873 if (est_size > (room_to_grow - amount_grown)) { 1874 // This one won't fit anyway. Get rid of it. 1875 call->make_cold(); 1876 continue; 1877 } 1878 call->make_hot(); 1879 calls_made_hot++; 1880 amount_grown += est_size; 1881 amount_to_grow -= est_size; 1882 } 1883 1884 if (calls_made_hot > 0) set_major_progress(); 1885 return calls_made_hot; 1886 } 1887 1888 1889 //----------------------------Finish_Warm-------------------------------------- 1890 void Compile::Finish_Warm() { 1891 if (!InlineWarmCalls) return; 1892 if (failing()) return; 1893 if (warm_calls() == NULL) return; 1894 1895 // Clean up loose ends, if we are out of space for inlining. 1896 WarmCallInfo* call; 1897 while ((call = pop_warm_call()) != NULL) { 1898 call->make_cold(); 1899 } 1900 } 1901 1902 //---------------------cleanup_loop_predicates----------------------- 1903 // Remove the opaque nodes that protect the predicates so that all unused 1904 // checks and uncommon_traps will be eliminated from the ideal graph 1905 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1906 if (predicate_count()==0) return; 1907 for (int i = predicate_count(); i > 0; i--) { 1908 Node * n = predicate_opaque1_node(i-1); 1909 assert(n->Opcode() == Op_Opaque1, "must be"); 1910 igvn.replace_node(n, n->in(1)); 1911 } 1912 assert(predicate_count()==0, "should be clean!"); 1913 } 1914 1915 void Compile::add_range_check_cast(Node* n) { 1916 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1917 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1918 _range_check_casts->append(n); 1919 } 1920 1921 // Remove all range check dependent CastIINodes. 1922 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1923 for (int i = range_check_cast_count(); i > 0; i--) { 1924 Node* cast = range_check_cast_node(i-1); 1925 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1926 igvn.replace_node(cast, cast->in(1)); 1927 } 1928 assert(range_check_cast_count() == 0, "should be empty"); 1929 } 1930 1931 // StringOpts and late inlining of string methods 1932 void Compile::inline_string_calls(bool parse_time) { 1933 { 1934 // remove useless nodes to make the usage analysis simpler 1935 ResourceMark rm; 1936 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1937 } 1938 1939 { 1940 ResourceMark rm; 1941 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1942 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1943 print_method(PHASE_AFTER_STRINGOPTS, 3); 1944 } 1945 1946 // now inline anything that we skipped the first time around 1947 if (!parse_time) { 1948 _late_inlines_pos = _late_inlines.length(); 1949 } 1950 1951 while (_string_late_inlines.length() > 0) { 1952 CallGenerator* cg = _string_late_inlines.pop(); 1953 cg->do_late_inline(); 1954 if (failing()) return; 1955 } 1956 _string_late_inlines.trunc_to(0); 1957 } 1958 1959 // Late inlining of boxing methods 1960 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1961 if (_boxing_late_inlines.length() > 0) { 1962 assert(has_boxed_value(), "inconsistent"); 1963 1964 PhaseGVN* gvn = initial_gvn(); 1965 set_inlining_incrementally(true); 1966 1967 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1968 for_igvn()->clear(); 1969 gvn->replace_with(&igvn); 1970 1971 _late_inlines_pos = _late_inlines.length(); 1972 1973 while (_boxing_late_inlines.length() > 0) { 1974 CallGenerator* cg = _boxing_late_inlines.pop(); 1975 cg->do_late_inline(); 1976 if (failing()) return; 1977 } 1978 _boxing_late_inlines.trunc_to(0); 1979 1980 { 1981 ResourceMark rm; 1982 PhaseRemoveUseless pru(gvn, for_igvn()); 1983 } 1984 1985 igvn = PhaseIterGVN(gvn); 1986 igvn.optimize(); 1987 1988 set_inlining_progress(false); 1989 set_inlining_incrementally(false); 1990 } 1991 } 1992 1993 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1994 assert(IncrementalInline, "incremental inlining should be on"); 1995 PhaseGVN* gvn = initial_gvn(); 1996 1997 set_inlining_progress(false); 1998 for_igvn()->clear(); 1999 gvn->replace_with(&igvn); 2000 2001 int i = 0; 2002 2003 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2004 CallGenerator* cg = _late_inlines.at(i); 2005 _late_inlines_pos = i+1; 2006 cg->do_late_inline(); 2007 if (failing()) return; 2008 } 2009 int j = 0; 2010 for (; i < _late_inlines.length(); i++, j++) { 2011 _late_inlines.at_put(j, _late_inlines.at(i)); 2012 } 2013 _late_inlines.trunc_to(j); 2014 2015 { 2016 ResourceMark rm; 2017 PhaseRemoveUseless pru(gvn, for_igvn()); 2018 } 2019 2020 igvn = PhaseIterGVN(gvn); 2021 } 2022 2023 // Perform incremental inlining until bound on number of live nodes is reached 2024 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2025 PhaseGVN* gvn = initial_gvn(); 2026 2027 set_inlining_incrementally(true); 2028 set_inlining_progress(true); 2029 uint low_live_nodes = 0; 2030 2031 while(inlining_progress() && _late_inlines.length() > 0) { 2032 2033 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2034 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2035 // PhaseIdealLoop is expensive so we only try it once we are 2036 // out of live nodes and we only try it again if the previous 2037 // helped got the number of nodes down significantly 2038 PhaseIdealLoop ideal_loop( igvn, false, true ); 2039 if (failing()) return; 2040 low_live_nodes = live_nodes(); 2041 _major_progress = true; 2042 } 2043 2044 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2045 break; 2046 } 2047 } 2048 2049 inline_incrementally_one(igvn); 2050 2051 if (failing()) return; 2052 2053 igvn.optimize(); 2054 2055 if (failing()) return; 2056 } 2057 2058 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2059 2060 if (_string_late_inlines.length() > 0) { 2061 assert(has_stringbuilder(), "inconsistent"); 2062 for_igvn()->clear(); 2063 initial_gvn()->replace_with(&igvn); 2064 2065 inline_string_calls(false); 2066 2067 if (failing()) return; 2068 2069 { 2070 ResourceMark rm; 2071 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2072 } 2073 2074 igvn = PhaseIterGVN(gvn); 2075 2076 igvn.optimize(); 2077 } 2078 2079 set_inlining_incrementally(false); 2080 } 2081 2082 2083 // Remove edges from "root" to each SafePoint at a backward branch. 2084 // They were inserted during parsing (see add_safepoint()) to make 2085 // infinite loops without calls or exceptions visible to root, i.e., 2086 // useful. 2087 void Compile::remove_root_to_sfpts_edges() { 2088 Node *r = root(); 2089 if (r != NULL) { 2090 for (uint i = r->req(); i < r->len(); ++i) { 2091 Node *n = r->in(i); 2092 if (n != NULL && n->is_SafePoint()) { 2093 r->rm_prec(i); 2094 --i; 2095 } 2096 } 2097 } 2098 } 2099 2100 //------------------------------Optimize--------------------------------------- 2101 // Given a graph, optimize it. 2102 void Compile::Optimize() { 2103 TracePhase t1("optimizer", &_t_optimizer, true); 2104 2105 #ifndef PRODUCT 2106 if (env()->break_at_compile()) { 2107 BREAKPOINT; 2108 } 2109 2110 #endif 2111 2112 ResourceMark rm; 2113 int loop_opts_cnt; 2114 2115 NOT_PRODUCT( verify_graph_edges(); ) 2116 2117 print_method(PHASE_AFTER_PARSING); 2118 2119 { 2120 // Iterative Global Value Numbering, including ideal transforms 2121 // Initialize IterGVN with types and values from parse-time GVN 2122 PhaseIterGVN igvn(initial_gvn()); 2123 { 2124 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2125 igvn.optimize(); 2126 } 2127 2128 print_method(PHASE_ITER_GVN1, 2); 2129 2130 if (failing()) return; 2131 2132 { 2133 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2134 inline_incrementally(igvn); 2135 } 2136 2137 print_method(PHASE_INCREMENTAL_INLINE, 2); 2138 2139 if (failing()) return; 2140 2141 if (eliminate_boxing()) { 2142 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2143 // Inline valueOf() methods now. 2144 inline_boxing_calls(igvn); 2145 2146 if (AlwaysIncrementalInline) { 2147 inline_incrementally(igvn); 2148 } 2149 2150 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2151 2152 if (failing()) return; 2153 } 2154 2155 // Now that all inlining is over, cut edge from root to loop 2156 // safepoints 2157 remove_root_to_sfpts_edges(); 2158 2159 // Remove the speculative part of types and clean up the graph from 2160 // the extra CastPP nodes whose only purpose is to carry them. Do 2161 // that early so that optimizations are not disrupted by the extra 2162 // CastPP nodes. 2163 remove_speculative_types(igvn); 2164 2165 // No more new expensive nodes will be added to the list from here 2166 // so keep only the actual candidates for optimizations. 2167 cleanup_expensive_nodes(igvn); 2168 2169 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2170 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);) 2171 initial_gvn()->replace_with(&igvn); 2172 for_igvn()->clear(); 2173 Unique_Node_List new_worklist(C->comp_arena()); 2174 { 2175 ResourceMark rm; 2176 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2177 } 2178 set_for_igvn(&new_worklist); 2179 igvn = PhaseIterGVN(initial_gvn()); 2180 igvn.optimize(); 2181 } 2182 2183 // Perform escape analysis 2184 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2185 if (has_loops()) { 2186 // Cleanup graph (remove dead nodes). 2187 TracePhase t2("idealLoop", &_t_idealLoop, true); 2188 PhaseIdealLoop ideal_loop( igvn, false, true ); 2189 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2190 if (failing()) return; 2191 } 2192 ConnectionGraph::do_analysis(this, &igvn); 2193 2194 if (failing()) return; 2195 2196 // Optimize out fields loads from scalar replaceable allocations. 2197 igvn.optimize(); 2198 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2199 2200 if (failing()) return; 2201 2202 if (congraph() != NULL && macro_count() > 0) { 2203 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2204 PhaseMacroExpand mexp(igvn); 2205 mexp.eliminate_macro_nodes(); 2206 igvn.set_delay_transform(false); 2207 2208 igvn.optimize(); 2209 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2210 2211 if (failing()) return; 2212 } 2213 } 2214 2215 // Loop transforms on the ideal graph. Range Check Elimination, 2216 // peeling, unrolling, etc. 2217 2218 // Set loop opts counter 2219 loop_opts_cnt = num_loop_opts(); 2220 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2221 { 2222 TracePhase t2("idealLoop", &_t_idealLoop, true); 2223 PhaseIdealLoop ideal_loop( igvn, true ); 2224 loop_opts_cnt--; 2225 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2226 if (failing()) return; 2227 } 2228 // Loop opts pass if partial peeling occurred in previous pass 2229 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2230 TracePhase t3("idealLoop", &_t_idealLoop, true); 2231 PhaseIdealLoop ideal_loop( igvn, false ); 2232 loop_opts_cnt--; 2233 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2234 if (failing()) return; 2235 } 2236 // Loop opts pass for loop-unrolling before CCP 2237 if(major_progress() && (loop_opts_cnt > 0)) { 2238 TracePhase t4("idealLoop", &_t_idealLoop, true); 2239 PhaseIdealLoop ideal_loop( igvn, false ); 2240 loop_opts_cnt--; 2241 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2242 } 2243 if (!failing()) { 2244 // Verify that last round of loop opts produced a valid graph 2245 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2246 PhaseIdealLoop::verify(igvn); 2247 } 2248 } 2249 if (failing()) return; 2250 2251 // Conditional Constant Propagation; 2252 PhaseCCP ccp( &igvn ); 2253 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2254 { 2255 TracePhase t2("ccp", &_t_ccp, true); 2256 ccp.do_transform(); 2257 } 2258 print_method(PHASE_CPP1, 2); 2259 2260 assert( true, "Break here to ccp.dump_old2new_map()"); 2261 2262 // Iterative Global Value Numbering, including ideal transforms 2263 { 2264 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2265 igvn = ccp; 2266 igvn.optimize(); 2267 } 2268 2269 print_method(PHASE_ITER_GVN2, 2); 2270 2271 if (failing()) return; 2272 2273 // Loop transforms on the ideal graph. Range Check Elimination, 2274 // peeling, unrolling, etc. 2275 if(loop_opts_cnt > 0) { 2276 debug_only( int cnt = 0; ); 2277 while(major_progress() && (loop_opts_cnt > 0)) { 2278 TracePhase t2("idealLoop", &_t_idealLoop, true); 2279 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2280 PhaseIdealLoop ideal_loop( igvn, true); 2281 loop_opts_cnt--; 2282 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2283 if (failing()) return; 2284 } 2285 } 2286 2287 { 2288 // Verify that all previous optimizations produced a valid graph 2289 // at least to this point, even if no loop optimizations were done. 2290 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2291 PhaseIdealLoop::verify(igvn); 2292 } 2293 2294 if (range_check_cast_count() > 0) { 2295 // No more loop optimizations. Remove all range check dependent CastIINodes. 2296 C->remove_range_check_casts(igvn); 2297 igvn.optimize(); 2298 } 2299 2300 { 2301 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2302 PhaseMacroExpand mex(igvn); 2303 if (mex.expand_macro_nodes()) { 2304 assert(failing(), "must bail out w/ explicit message"); 2305 return; 2306 } 2307 } 2308 2309 } // (End scope of igvn; run destructor if necessary for asserts.) 2310 2311 dump_inlining(); 2312 // A method with only infinite loops has no edges entering loops from root 2313 { 2314 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2315 if (final_graph_reshaping()) { 2316 assert(failing(), "must bail out w/ explicit message"); 2317 return; 2318 } 2319 } 2320 2321 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2322 } 2323 2324 2325 //------------------------------Code_Gen--------------------------------------- 2326 // Given a graph, generate code for it 2327 void Compile::Code_Gen() { 2328 if (failing()) { 2329 return; 2330 } 2331 2332 // Perform instruction selection. You might think we could reclaim Matcher 2333 // memory PDQ, but actually the Matcher is used in generating spill code. 2334 // Internals of the Matcher (including some VectorSets) must remain live 2335 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2336 // set a bit in reclaimed memory. 2337 2338 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2339 // nodes. Mapping is only valid at the root of each matched subtree. 2340 NOT_PRODUCT( verify_graph_edges(); ) 2341 2342 Matcher matcher; 2343 _matcher = &matcher; 2344 { 2345 TracePhase t2("matcher", &_t_matcher, true); 2346 matcher.match(); 2347 } 2348 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2349 // nodes. Mapping is only valid at the root of each matched subtree. 2350 NOT_PRODUCT( verify_graph_edges(); ) 2351 2352 // If you have too many nodes, or if matching has failed, bail out 2353 check_node_count(0, "out of nodes matching instructions"); 2354 if (failing()) { 2355 return; 2356 } 2357 2358 // Build a proper-looking CFG 2359 PhaseCFG cfg(node_arena(), root(), matcher); 2360 _cfg = &cfg; 2361 { 2362 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2363 bool success = cfg.do_global_code_motion(); 2364 if (!success) { 2365 return; 2366 } 2367 2368 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2369 NOT_PRODUCT( verify_graph_edges(); ) 2370 debug_only( cfg.verify(); ) 2371 } 2372 2373 PhaseChaitin regalloc(unique(), cfg, matcher); 2374 _regalloc = ®alloc; 2375 { 2376 TracePhase t2("regalloc", &_t_registerAllocation, true); 2377 // Perform register allocation. After Chaitin, use-def chains are 2378 // no longer accurate (at spill code) and so must be ignored. 2379 // Node->LRG->reg mappings are still accurate. 2380 _regalloc->Register_Allocate(); 2381 2382 // Bail out if the allocator builds too many nodes 2383 if (failing()) { 2384 return; 2385 } 2386 } 2387 2388 // Prior to register allocation we kept empty basic blocks in case the 2389 // the allocator needed a place to spill. After register allocation we 2390 // are not adding any new instructions. If any basic block is empty, we 2391 // can now safely remove it. 2392 { 2393 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2394 cfg.remove_empty_blocks(); 2395 if (do_freq_based_layout()) { 2396 PhaseBlockLayout layout(cfg); 2397 } else { 2398 cfg.set_loop_alignment(); 2399 } 2400 cfg.fixup_flow(); 2401 } 2402 2403 // Apply peephole optimizations 2404 if( OptoPeephole ) { 2405 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2406 PhasePeephole peep( _regalloc, cfg); 2407 peep.do_transform(); 2408 } 2409 2410 // Do late expand if CPU requires this. 2411 if (Matcher::require_postalloc_expand) { 2412 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2413 cfg.postalloc_expand(_regalloc); 2414 } 2415 2416 // Convert Nodes to instruction bits in a buffer 2417 { 2418 // %%%% workspace merge brought two timers together for one job 2419 TracePhase t2a("output", &_t_output, true); 2420 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2421 Output(); 2422 } 2423 2424 print_method(PHASE_FINAL_CODE); 2425 2426 // He's dead, Jim. 2427 _cfg = (PhaseCFG*)((intptr_t)0xdeadbeef); 2428 _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef); 2429 } 2430 2431 2432 //------------------------------dump_asm--------------------------------------- 2433 // Dump formatted assembly 2434 #ifndef PRODUCT 2435 void Compile::dump_asm(int *pcs, uint pc_limit) { 2436 bool cut_short = false; 2437 tty->print_cr("#"); 2438 tty->print("# "); _tf->dump(); tty->cr(); 2439 tty->print_cr("#"); 2440 2441 // For all blocks 2442 int pc = 0x0; // Program counter 2443 char starts_bundle = ' '; 2444 _regalloc->dump_frame(); 2445 2446 Node *n = NULL; 2447 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2448 if (VMThread::should_terminate()) { 2449 cut_short = true; 2450 break; 2451 } 2452 Block* block = _cfg->get_block(i); 2453 if (block->is_connector() && !Verbose) { 2454 continue; 2455 } 2456 n = block->head(); 2457 if (pcs && n->_idx < pc_limit) { 2458 tty->print("%3.3x ", pcs[n->_idx]); 2459 } else { 2460 tty->print(" "); 2461 } 2462 block->dump_head(_cfg); 2463 if (block->is_connector()) { 2464 tty->print_cr(" # Empty connector block"); 2465 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2466 tty->print_cr(" # Block is sole successor of call"); 2467 } 2468 2469 // For all instructions 2470 Node *delay = NULL; 2471 for (uint j = 0; j < block->number_of_nodes(); j++) { 2472 if (VMThread::should_terminate()) { 2473 cut_short = true; 2474 break; 2475 } 2476 n = block->get_node(j); 2477 if (valid_bundle_info(n)) { 2478 Bundle* bundle = node_bundling(n); 2479 if (bundle->used_in_unconditional_delay()) { 2480 delay = n; 2481 continue; 2482 } 2483 if (bundle->starts_bundle()) { 2484 starts_bundle = '+'; 2485 } 2486 } 2487 2488 if (WizardMode) { 2489 n->dump(); 2490 } 2491 2492 if( !n->is_Region() && // Dont print in the Assembly 2493 !n->is_Phi() && // a few noisely useless nodes 2494 !n->is_Proj() && 2495 !n->is_MachTemp() && 2496 !n->is_SafePointScalarObject() && 2497 !n->is_Catch() && // Would be nice to print exception table targets 2498 !n->is_MergeMem() && // Not very interesting 2499 !n->is_top() && // Debug info table constants 2500 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2501 ) { 2502 if (pcs && n->_idx < pc_limit) 2503 tty->print("%3.3x", pcs[n->_idx]); 2504 else 2505 tty->print(" "); 2506 tty->print(" %c ", starts_bundle); 2507 starts_bundle = ' '; 2508 tty->print("\t"); 2509 n->format(_regalloc, tty); 2510 tty->cr(); 2511 } 2512 2513 // If we have an instruction with a delay slot, and have seen a delay, 2514 // then back up and print it 2515 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2516 assert(delay != NULL, "no unconditional delay instruction"); 2517 if (WizardMode) delay->dump(); 2518 2519 if (node_bundling(delay)->starts_bundle()) 2520 starts_bundle = '+'; 2521 if (pcs && n->_idx < pc_limit) 2522 tty->print("%3.3x", pcs[n->_idx]); 2523 else 2524 tty->print(" "); 2525 tty->print(" %c ", starts_bundle); 2526 starts_bundle = ' '; 2527 tty->print("\t"); 2528 delay->format(_regalloc, tty); 2529 tty->cr(); 2530 delay = NULL; 2531 } 2532 2533 // Dump the exception table as well 2534 if( n->is_Catch() && (Verbose || WizardMode) ) { 2535 // Print the exception table for this offset 2536 _handler_table.print_subtable_for(pc); 2537 } 2538 } 2539 2540 if (pcs && n->_idx < pc_limit) 2541 tty->print_cr("%3.3x", pcs[n->_idx]); 2542 else 2543 tty->cr(); 2544 2545 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2546 2547 } // End of per-block dump 2548 tty->cr(); 2549 2550 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2551 } 2552 #endif 2553 2554 //------------------------------Final_Reshape_Counts--------------------------- 2555 // This class defines counters to help identify when a method 2556 // may/must be executed using hardware with only 24-bit precision. 2557 struct Final_Reshape_Counts : public StackObj { 2558 int _call_count; // count non-inlined 'common' calls 2559 int _float_count; // count float ops requiring 24-bit precision 2560 int _double_count; // count double ops requiring more precision 2561 int _java_call_count; // count non-inlined 'java' calls 2562 int _inner_loop_count; // count loops which need alignment 2563 VectorSet _visited; // Visitation flags 2564 Node_List _tests; // Set of IfNodes & PCTableNodes 2565 2566 Final_Reshape_Counts() : 2567 _call_count(0), _float_count(0), _double_count(0), 2568 _java_call_count(0), _inner_loop_count(0), 2569 _visited( Thread::current()->resource_area() ) { } 2570 2571 void inc_call_count () { _call_count ++; } 2572 void inc_float_count () { _float_count ++; } 2573 void inc_double_count() { _double_count++; } 2574 void inc_java_call_count() { _java_call_count++; } 2575 void inc_inner_loop_count() { _inner_loop_count++; } 2576 2577 int get_call_count () const { return _call_count ; } 2578 int get_float_count () const { return _float_count ; } 2579 int get_double_count() const { return _double_count; } 2580 int get_java_call_count() const { return _java_call_count; } 2581 int get_inner_loop_count() const { return _inner_loop_count; } 2582 }; 2583 2584 #ifdef ASSERT 2585 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2586 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2587 // Make sure the offset goes inside the instance layout. 2588 return k->contains_field_offset(tp->offset()); 2589 // Note that OffsetBot and OffsetTop are very negative. 2590 } 2591 #endif 2592 2593 // Eliminate trivially redundant StoreCMs and accumulate their 2594 // precedence edges. 2595 void Compile::eliminate_redundant_card_marks(Node* n) { 2596 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2597 if (n->in(MemNode::Address)->outcnt() > 1) { 2598 // There are multiple users of the same address so it might be 2599 // possible to eliminate some of the StoreCMs 2600 Node* mem = n->in(MemNode::Memory); 2601 Node* adr = n->in(MemNode::Address); 2602 Node* val = n->in(MemNode::ValueIn); 2603 Node* prev = n; 2604 bool done = false; 2605 // Walk the chain of StoreCMs eliminating ones that match. As 2606 // long as it's a chain of single users then the optimization is 2607 // safe. Eliminating partially redundant StoreCMs would require 2608 // cloning copies down the other paths. 2609 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2610 if (adr == mem->in(MemNode::Address) && 2611 val == mem->in(MemNode::ValueIn)) { 2612 // redundant StoreCM 2613 if (mem->req() > MemNode::OopStore) { 2614 // Hasn't been processed by this code yet. 2615 n->add_prec(mem->in(MemNode::OopStore)); 2616 } else { 2617 // Already converted to precedence edge 2618 for (uint i = mem->req(); i < mem->len(); i++) { 2619 // Accumulate any precedence edges 2620 if (mem->in(i) != NULL) { 2621 n->add_prec(mem->in(i)); 2622 } 2623 } 2624 // Everything above this point has been processed. 2625 done = true; 2626 } 2627 // Eliminate the previous StoreCM 2628 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2629 assert(mem->outcnt() == 0, "should be dead"); 2630 mem->disconnect_inputs(NULL, this); 2631 } else { 2632 prev = mem; 2633 } 2634 mem = prev->in(MemNode::Memory); 2635 } 2636 } 2637 } 2638 2639 //------------------------------final_graph_reshaping_impl---------------------- 2640 // Implement items 1-5 from final_graph_reshaping below. 2641 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2642 2643 if ( n->outcnt() == 0 ) return; // dead node 2644 uint nop = n->Opcode(); 2645 2646 // Check for 2-input instruction with "last use" on right input. 2647 // Swap to left input. Implements item (2). 2648 if( n->req() == 3 && // two-input instruction 2649 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2650 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2651 n->in(2)->outcnt() == 1 &&// right use IS a last use 2652 !n->in(2)->is_Con() ) { // right use is not a constant 2653 // Check for commutative opcode 2654 switch( nop ) { 2655 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2656 case Op_MaxI: case Op_MinI: 2657 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2658 case Op_AndL: case Op_XorL: case Op_OrL: 2659 case Op_AndI: case Op_XorI: case Op_OrI: { 2660 // Move "last use" input to left by swapping inputs 2661 n->swap_edges(1, 2); 2662 break; 2663 } 2664 default: 2665 break; 2666 } 2667 } 2668 2669 #ifdef ASSERT 2670 if( n->is_Mem() ) { 2671 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2672 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2673 // oop will be recorded in oop map if load crosses safepoint 2674 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2675 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2676 "raw memory operations should have control edge"); 2677 } 2678 if (n->is_MemBar()) { 2679 MemBarNode* mb = n->as_MemBar(); 2680 if (mb->trailing_store() || mb->trailing_load_store()) { 2681 assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair"); 2682 Node* mem = mb->in(MemBarNode::Precedent); 2683 assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) || 2684 (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op"); 2685 } else if (mb->leading()) { 2686 assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair"); 2687 } 2688 } 2689 #endif 2690 // Count FPU ops and common calls, implements item (3) 2691 switch( nop ) { 2692 // Count all float operations that may use FPU 2693 case Op_AddF: 2694 case Op_SubF: 2695 case Op_MulF: 2696 case Op_DivF: 2697 case Op_NegF: 2698 case Op_ModF: 2699 case Op_ConvI2F: 2700 case Op_ConF: 2701 case Op_CmpF: 2702 case Op_CmpF3: 2703 // case Op_ConvL2F: // longs are split into 32-bit halves 2704 frc.inc_float_count(); 2705 break; 2706 2707 case Op_ConvF2D: 2708 case Op_ConvD2F: 2709 frc.inc_float_count(); 2710 frc.inc_double_count(); 2711 break; 2712 2713 // Count all double operations that may use FPU 2714 case Op_AddD: 2715 case Op_SubD: 2716 case Op_MulD: 2717 case Op_DivD: 2718 case Op_NegD: 2719 case Op_ModD: 2720 case Op_ConvI2D: 2721 case Op_ConvD2I: 2722 // case Op_ConvL2D: // handled by leaf call 2723 // case Op_ConvD2L: // handled by leaf call 2724 case Op_ConD: 2725 case Op_CmpD: 2726 case Op_CmpD3: 2727 frc.inc_double_count(); 2728 break; 2729 case Op_Opaque1: // Remove Opaque Nodes before matching 2730 case Op_Opaque2: // Remove Opaque Nodes before matching 2731 case Op_Opaque3: 2732 n->subsume_by(n->in(1), this); 2733 break; 2734 case Op_CallStaticJava: 2735 case Op_CallJava: 2736 case Op_CallDynamicJava: 2737 frc.inc_java_call_count(); // Count java call site; 2738 case Op_CallRuntime: 2739 case Op_CallLeaf: 2740 case Op_CallLeafNoFP: { 2741 assert( n->is_Call(), "" ); 2742 CallNode *call = n->as_Call(); 2743 // Count call sites where the FP mode bit would have to be flipped. 2744 // Do not count uncommon runtime calls: 2745 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2746 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2747 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2748 frc.inc_call_count(); // Count the call site 2749 } else { // See if uncommon argument is shared 2750 Node *n = call->in(TypeFunc::Parms); 2751 int nop = n->Opcode(); 2752 // Clone shared simple arguments to uncommon calls, item (1). 2753 if( n->outcnt() > 1 && 2754 !n->is_Proj() && 2755 nop != Op_CreateEx && 2756 nop != Op_CheckCastPP && 2757 nop != Op_DecodeN && 2758 nop != Op_DecodeNKlass && 2759 !n->is_Mem() ) { 2760 Node *x = n->clone(); 2761 call->set_req( TypeFunc::Parms, x ); 2762 } 2763 } 2764 break; 2765 } 2766 2767 case Op_StoreD: 2768 case Op_LoadD: 2769 case Op_LoadD_unaligned: 2770 frc.inc_double_count(); 2771 goto handle_mem; 2772 case Op_StoreF: 2773 case Op_LoadF: 2774 frc.inc_float_count(); 2775 goto handle_mem; 2776 2777 case Op_StoreCM: 2778 { 2779 // Convert OopStore dependence into precedence edge 2780 Node* prec = n->in(MemNode::OopStore); 2781 n->del_req(MemNode::OopStore); 2782 n->add_prec(prec); 2783 eliminate_redundant_card_marks(n); 2784 } 2785 2786 // fall through 2787 2788 case Op_StoreB: 2789 case Op_StoreC: 2790 case Op_StorePConditional: 2791 case Op_StoreI: 2792 case Op_StoreL: 2793 case Op_StoreIConditional: 2794 case Op_StoreLConditional: 2795 case Op_CompareAndSwapI: 2796 case Op_CompareAndSwapL: 2797 case Op_CompareAndSwapP: 2798 case Op_CompareAndSwapN: 2799 case Op_GetAndAddI: 2800 case Op_GetAndAddL: 2801 case Op_GetAndSetI: 2802 case Op_GetAndSetL: 2803 case Op_GetAndSetP: 2804 case Op_GetAndSetN: 2805 case Op_StoreP: 2806 case Op_StoreN: 2807 case Op_StoreNKlass: 2808 case Op_LoadB: 2809 case Op_LoadUB: 2810 case Op_LoadUS: 2811 case Op_LoadI: 2812 case Op_LoadKlass: 2813 case Op_LoadNKlass: 2814 case Op_LoadL: 2815 case Op_LoadL_unaligned: 2816 case Op_LoadPLocked: 2817 case Op_LoadP: 2818 case Op_LoadN: 2819 case Op_LoadRange: 2820 case Op_LoadS: { 2821 handle_mem: 2822 #ifdef ASSERT 2823 if( VerifyOptoOopOffsets ) { 2824 assert( n->is_Mem(), "" ); 2825 MemNode *mem = (MemNode*)n; 2826 // Check to see if address types have grounded out somehow. 2827 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2828 assert( !tp || oop_offset_is_sane(tp), "" ); 2829 } 2830 #endif 2831 break; 2832 } 2833 2834 case Op_AddP: { // Assert sane base pointers 2835 Node *addp = n->in(AddPNode::Address); 2836 assert( !addp->is_AddP() || 2837 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2838 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2839 "Base pointers must match" ); 2840 #ifdef _LP64 2841 if ((UseCompressedOops || UseCompressedClassPointers) && 2842 addp->Opcode() == Op_ConP && 2843 addp == n->in(AddPNode::Base) && 2844 n->in(AddPNode::Offset)->is_Con()) { 2845 // Use addressing with narrow klass to load with offset on x86. 2846 // On sparc loading 32-bits constant and decoding it have less 2847 // instructions (4) then load 64-bits constant (7). 2848 // Do this transformation here since IGVN will convert ConN back to ConP. 2849 const Type* t = addp->bottom_type(); 2850 if (t->isa_oopptr() || t->isa_klassptr()) { 2851 Node* nn = NULL; 2852 2853 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2854 2855 // Look for existing ConN node of the same exact type. 2856 Node* r = root(); 2857 uint cnt = r->outcnt(); 2858 for (uint i = 0; i < cnt; i++) { 2859 Node* m = r->raw_out(i); 2860 if (m!= NULL && m->Opcode() == op && 2861 m->bottom_type()->make_ptr() == t) { 2862 nn = m; 2863 break; 2864 } 2865 } 2866 if (nn != NULL) { 2867 // Decode a narrow oop to match address 2868 // [R12 + narrow_oop_reg<<3 + offset] 2869 if (t->isa_oopptr()) { 2870 nn = new (this) DecodeNNode(nn, t); 2871 } else { 2872 nn = new (this) DecodeNKlassNode(nn, t); 2873 } 2874 n->set_req(AddPNode::Base, nn); 2875 n->set_req(AddPNode::Address, nn); 2876 if (addp->outcnt() == 0) { 2877 addp->disconnect_inputs(NULL, this); 2878 } 2879 } 2880 } 2881 } 2882 #endif 2883 break; 2884 } 2885 2886 case Op_CastPP: 2887 { 2888 // Remove CastPP nodes to gain more freedom during scheduling but 2889 // keep the dependency they encode as control or precedence edges 2890 // (if control is set already) on memory operations. Some CastPP 2891 // nodes don't have a control (don't carry a dependency): skip 2892 // those. 2893 if (n->in(0) != NULL) { 2894 ResourceMark rm; 2895 Unique_Node_List wq; 2896 wq.push(n); 2897 for (uint next = 0; next < wq.size(); ++next) { 2898 Node *m = wq.at(next); 2899 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2900 Node* use = m->fast_out(i); 2901 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2902 use->ensure_control_or_add_prec(n->in(0)); 2903 } else if (use->in(0) == NULL) { 2904 switch(use->Opcode()) { 2905 case Op_AddP: 2906 case Op_DecodeN: 2907 case Op_DecodeNKlass: 2908 case Op_CheckCastPP: 2909 case Op_CastPP: 2910 wq.push(use); 2911 break; 2912 } 2913 } 2914 } 2915 } 2916 } 2917 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2918 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2919 Node* in1 = n->in(1); 2920 const Type* t = n->bottom_type(); 2921 Node* new_in1 = in1->clone(); 2922 new_in1->as_DecodeN()->set_type(t); 2923 2924 if (!Matcher::narrow_oop_use_complex_address()) { 2925 // 2926 // x86, ARM and friends can handle 2 adds in addressing mode 2927 // and Matcher can fold a DecodeN node into address by using 2928 // a narrow oop directly and do implicit NULL check in address: 2929 // 2930 // [R12 + narrow_oop_reg<<3 + offset] 2931 // NullCheck narrow_oop_reg 2932 // 2933 // On other platforms (Sparc) we have to keep new DecodeN node and 2934 // use it to do implicit NULL check in address: 2935 // 2936 // decode_not_null narrow_oop_reg, base_reg 2937 // [base_reg + offset] 2938 // NullCheck base_reg 2939 // 2940 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2941 // to keep the information to which NULL check the new DecodeN node 2942 // corresponds to use it as value in implicit_null_check(). 2943 // 2944 new_in1->set_req(0, n->in(0)); 2945 } 2946 2947 n->subsume_by(new_in1, this); 2948 if (in1->outcnt() == 0) { 2949 in1->disconnect_inputs(NULL, this); 2950 } 2951 } else { 2952 n->subsume_by(n->in(1), this); 2953 if (n->outcnt() == 0) { 2954 n->disconnect_inputs(NULL, this); 2955 } 2956 } 2957 break; 2958 } 2959 #ifdef _LP64 2960 case Op_CmpP: 2961 // Do this transformation here to preserve CmpPNode::sub() and 2962 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2963 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2964 Node* in1 = n->in(1); 2965 Node* in2 = n->in(2); 2966 if (!in1->is_DecodeNarrowPtr()) { 2967 in2 = in1; 2968 in1 = n->in(2); 2969 } 2970 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2971 2972 Node* new_in2 = NULL; 2973 if (in2->is_DecodeNarrowPtr()) { 2974 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2975 new_in2 = in2->in(1); 2976 } else if (in2->Opcode() == Op_ConP) { 2977 const Type* t = in2->bottom_type(); 2978 if (t == TypePtr::NULL_PTR) { 2979 assert(in1->is_DecodeN(), "compare klass to null?"); 2980 // Don't convert CmpP null check into CmpN if compressed 2981 // oops implicit null check is not generated. 2982 // This will allow to generate normal oop implicit null check. 2983 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2984 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2985 // 2986 // This transformation together with CastPP transformation above 2987 // will generated code for implicit NULL checks for compressed oops. 2988 // 2989 // The original code after Optimize() 2990 // 2991 // LoadN memory, narrow_oop_reg 2992 // decode narrow_oop_reg, base_reg 2993 // CmpP base_reg, NULL 2994 // CastPP base_reg // NotNull 2995 // Load [base_reg + offset], val_reg 2996 // 2997 // after these transformations will be 2998 // 2999 // LoadN memory, narrow_oop_reg 3000 // CmpN narrow_oop_reg, NULL 3001 // decode_not_null narrow_oop_reg, base_reg 3002 // Load [base_reg + offset], val_reg 3003 // 3004 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3005 // since narrow oops can be used in debug info now (see the code in 3006 // final_graph_reshaping_walk()). 3007 // 3008 // At the end the code will be matched to 3009 // on x86: 3010 // 3011 // Load_narrow_oop memory, narrow_oop_reg 3012 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3013 // NullCheck narrow_oop_reg 3014 // 3015 // and on sparc: 3016 // 3017 // Load_narrow_oop memory, narrow_oop_reg 3018 // decode_not_null narrow_oop_reg, base_reg 3019 // Load [base_reg + offset], val_reg 3020 // NullCheck base_reg 3021 // 3022 } else if (t->isa_oopptr()) { 3023 new_in2 = ConNode::make(this, t->make_narrowoop()); 3024 } else if (t->isa_klassptr()) { 3025 new_in2 = ConNode::make(this, t->make_narrowklass()); 3026 } 3027 } 3028 if (new_in2 != NULL) { 3029 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 3030 n->subsume_by(cmpN, this); 3031 if (in1->outcnt() == 0) { 3032 in1->disconnect_inputs(NULL, this); 3033 } 3034 if (in2->outcnt() == 0) { 3035 in2->disconnect_inputs(NULL, this); 3036 } 3037 } 3038 } 3039 break; 3040 3041 case Op_DecodeN: 3042 case Op_DecodeNKlass: 3043 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3044 // DecodeN could be pinned when it can't be fold into 3045 // an address expression, see the code for Op_CastPP above. 3046 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3047 break; 3048 3049 case Op_EncodeP: 3050 case Op_EncodePKlass: { 3051 Node* in1 = n->in(1); 3052 if (in1->is_DecodeNarrowPtr()) { 3053 n->subsume_by(in1->in(1), this); 3054 } else if (in1->Opcode() == Op_ConP) { 3055 const Type* t = in1->bottom_type(); 3056 if (t == TypePtr::NULL_PTR) { 3057 assert(t->isa_oopptr(), "null klass?"); 3058 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 3059 } else if (t->isa_oopptr()) { 3060 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 3061 } else if (t->isa_klassptr()) { 3062 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 3063 } 3064 } 3065 if (in1->outcnt() == 0) { 3066 in1->disconnect_inputs(NULL, this); 3067 } 3068 break; 3069 } 3070 3071 case Op_Proj: { 3072 if (OptimizeStringConcat) { 3073 ProjNode* p = n->as_Proj(); 3074 if (p->_is_io_use) { 3075 // Separate projections were used for the exception path which 3076 // are normally removed by a late inline. If it wasn't inlined 3077 // then they will hang around and should just be replaced with 3078 // the original one. 3079 Node* proj = NULL; 3080 // Replace with just one 3081 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3082 Node *use = i.get(); 3083 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3084 proj = use; 3085 break; 3086 } 3087 } 3088 assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop"); 3089 if (proj != NULL) { 3090 p->subsume_by(proj, this); 3091 } 3092 } 3093 } 3094 break; 3095 } 3096 3097 case Op_Phi: 3098 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3099 // The EncodeP optimization may create Phi with the same edges 3100 // for all paths. It is not handled well by Register Allocator. 3101 Node* unique_in = n->in(1); 3102 assert(unique_in != NULL, ""); 3103 uint cnt = n->req(); 3104 for (uint i = 2; i < cnt; i++) { 3105 Node* m = n->in(i); 3106 assert(m != NULL, ""); 3107 if (unique_in != m) 3108 unique_in = NULL; 3109 } 3110 if (unique_in != NULL) { 3111 n->subsume_by(unique_in, this); 3112 } 3113 } 3114 break; 3115 3116 #endif 3117 3118 #ifdef ASSERT 3119 case Op_CastII: 3120 // Verify that all range check dependent CastII nodes were removed. 3121 if (n->isa_CastII()->has_range_check()) { 3122 n->dump(3); 3123 assert(false, "Range check dependent CastII node was not removed"); 3124 } 3125 break; 3126 #endif 3127 3128 case Op_ModI: 3129 if (UseDivMod) { 3130 // Check if a%b and a/b both exist 3131 Node* d = n->find_similar(Op_DivI); 3132 if (d) { 3133 // Replace them with a fused divmod if supported 3134 if (Matcher::has_match_rule(Op_DivModI)) { 3135 DivModINode* divmod = DivModINode::make(this, n); 3136 d->subsume_by(divmod->div_proj(), this); 3137 n->subsume_by(divmod->mod_proj(), this); 3138 } else { 3139 // replace a%b with a-((a/b)*b) 3140 Node* mult = new (this) MulINode(d, d->in(2)); 3141 Node* sub = new (this) SubINode(d->in(1), mult); 3142 n->subsume_by(sub, this); 3143 } 3144 } 3145 } 3146 break; 3147 3148 case Op_ModL: 3149 if (UseDivMod) { 3150 // Check if a%b and a/b both exist 3151 Node* d = n->find_similar(Op_DivL); 3152 if (d) { 3153 // Replace them with a fused divmod if supported 3154 if (Matcher::has_match_rule(Op_DivModL)) { 3155 DivModLNode* divmod = DivModLNode::make(this, n); 3156 d->subsume_by(divmod->div_proj(), this); 3157 n->subsume_by(divmod->mod_proj(), this); 3158 } else { 3159 // replace a%b with a-((a/b)*b) 3160 Node* mult = new (this) MulLNode(d, d->in(2)); 3161 Node* sub = new (this) SubLNode(d->in(1), mult); 3162 n->subsume_by(sub, this); 3163 } 3164 } 3165 } 3166 break; 3167 3168 case Op_LoadVector: 3169 case Op_StoreVector: 3170 break; 3171 3172 case Op_PackB: 3173 case Op_PackS: 3174 case Op_PackI: 3175 case Op_PackF: 3176 case Op_PackL: 3177 case Op_PackD: 3178 if (n->req()-1 > 2) { 3179 // Replace many operand PackNodes with a binary tree for matching 3180 PackNode* p = (PackNode*) n; 3181 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3182 n->subsume_by(btp, this); 3183 } 3184 break; 3185 case Op_Loop: 3186 case Op_CountedLoop: 3187 if (n->as_Loop()->is_inner_loop()) { 3188 frc.inc_inner_loop_count(); 3189 } 3190 break; 3191 case Op_LShiftI: 3192 case Op_RShiftI: 3193 case Op_URShiftI: 3194 case Op_LShiftL: 3195 case Op_RShiftL: 3196 case Op_URShiftL: 3197 if (Matcher::need_masked_shift_count) { 3198 // The cpu's shift instructions don't restrict the count to the 3199 // lower 5/6 bits. We need to do the masking ourselves. 3200 Node* in2 = n->in(2); 3201 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3202 const TypeInt* t = in2->find_int_type(); 3203 if (t != NULL && t->is_con()) { 3204 juint shift = t->get_con(); 3205 if (shift > mask) { // Unsigned cmp 3206 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3207 } 3208 } else { 3209 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3210 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3211 n->set_req(2, shift); 3212 } 3213 } 3214 if (in2->outcnt() == 0) { // Remove dead node 3215 in2->disconnect_inputs(NULL, this); 3216 } 3217 } 3218 break; 3219 case Op_MemBarStoreStore: 3220 case Op_MemBarRelease: 3221 // Break the link with AllocateNode: it is no longer useful and 3222 // confuses register allocation. 3223 if (n->req() > MemBarNode::Precedent) { 3224 n->set_req(MemBarNode::Precedent, top()); 3225 } 3226 break; 3227 default: 3228 assert( !n->is_Call(), "" ); 3229 assert( !n->is_Mem(), "" ); 3230 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3231 break; 3232 } 3233 3234 // Collect CFG split points 3235 if (n->is_MultiBranch()) 3236 frc._tests.push(n); 3237 } 3238 3239 //------------------------------final_graph_reshaping_walk--------------------- 3240 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3241 // requires that the walk visits a node's inputs before visiting the node. 3242 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3243 ResourceArea *area = Thread::current()->resource_area(); 3244 Unique_Node_List sfpt(area); 3245 3246 frc._visited.set(root->_idx); // first, mark node as visited 3247 uint cnt = root->req(); 3248 Node *n = root; 3249 uint i = 0; 3250 while (true) { 3251 if (i < cnt) { 3252 // Place all non-visited non-null inputs onto stack 3253 Node* m = n->in(i); 3254 ++i; 3255 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3256 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3257 // compute worst case interpreter size in case of a deoptimization 3258 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3259 3260 sfpt.push(m); 3261 } 3262 cnt = m->req(); 3263 nstack.push(n, i); // put on stack parent and next input's index 3264 n = m; 3265 i = 0; 3266 } 3267 } else { 3268 // Now do post-visit work 3269 final_graph_reshaping_impl( n, frc ); 3270 if (nstack.is_empty()) 3271 break; // finished 3272 n = nstack.node(); // Get node from stack 3273 cnt = n->req(); 3274 i = nstack.index(); 3275 nstack.pop(); // Shift to the next node on stack 3276 } 3277 } 3278 3279 // Skip next transformation if compressed oops are not used. 3280 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3281 (!UseCompressedOops && !UseCompressedClassPointers)) 3282 return; 3283 3284 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3285 // It could be done for an uncommon traps or any safepoints/calls 3286 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3287 while (sfpt.size() > 0) { 3288 n = sfpt.pop(); 3289 JVMState *jvms = n->as_SafePoint()->jvms(); 3290 assert(jvms != NULL, "sanity"); 3291 int start = jvms->debug_start(); 3292 int end = n->req(); 3293 bool is_uncommon = (n->is_CallStaticJava() && 3294 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3295 for (int j = start; j < end; j++) { 3296 Node* in = n->in(j); 3297 if (in->is_DecodeNarrowPtr()) { 3298 bool safe_to_skip = true; 3299 if (!is_uncommon ) { 3300 // Is it safe to skip? 3301 for (uint i = 0; i < in->outcnt(); i++) { 3302 Node* u = in->raw_out(i); 3303 if (!u->is_SafePoint() || 3304 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3305 safe_to_skip = false; 3306 } 3307 } 3308 } 3309 if (safe_to_skip) { 3310 n->set_req(j, in->in(1)); 3311 } 3312 if (in->outcnt() == 0) { 3313 in->disconnect_inputs(NULL, this); 3314 } 3315 } 3316 } 3317 } 3318 } 3319 3320 //------------------------------final_graph_reshaping-------------------------- 3321 // Final Graph Reshaping. 3322 // 3323 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3324 // and not commoned up and forced early. Must come after regular 3325 // optimizations to avoid GVN undoing the cloning. Clone constant 3326 // inputs to Loop Phis; these will be split by the allocator anyways. 3327 // Remove Opaque nodes. 3328 // (2) Move last-uses by commutative operations to the left input to encourage 3329 // Intel update-in-place two-address operations and better register usage 3330 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3331 // calls canonicalizing them back. 3332 // (3) Count the number of double-precision FP ops, single-precision FP ops 3333 // and call sites. On Intel, we can get correct rounding either by 3334 // forcing singles to memory (requires extra stores and loads after each 3335 // FP bytecode) or we can set a rounding mode bit (requires setting and 3336 // clearing the mode bit around call sites). The mode bit is only used 3337 // if the relative frequency of single FP ops to calls is low enough. 3338 // This is a key transform for SPEC mpeg_audio. 3339 // (4) Detect infinite loops; blobs of code reachable from above but not 3340 // below. Several of the Code_Gen algorithms fail on such code shapes, 3341 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3342 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3343 // Detection is by looking for IfNodes where only 1 projection is 3344 // reachable from below or CatchNodes missing some targets. 3345 // (5) Assert for insane oop offsets in debug mode. 3346 3347 bool Compile::final_graph_reshaping() { 3348 // an infinite loop may have been eliminated by the optimizer, 3349 // in which case the graph will be empty. 3350 if (root()->req() == 1) { 3351 record_method_not_compilable("trivial infinite loop"); 3352 return true; 3353 } 3354 3355 // Expensive nodes have their control input set to prevent the GVN 3356 // from freely commoning them. There's no GVN beyond this point so 3357 // no need to keep the control input. We want the expensive nodes to 3358 // be freely moved to the least frequent code path by gcm. 3359 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3360 for (int i = 0; i < expensive_count(); i++) { 3361 _expensive_nodes->at(i)->set_req(0, NULL); 3362 } 3363 3364 Final_Reshape_Counts frc; 3365 3366 // Visit everybody reachable! 3367 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3368 Node_Stack nstack(live_nodes() >> 1); 3369 final_graph_reshaping_walk(nstack, root(), frc); 3370 3371 // Check for unreachable (from below) code (i.e., infinite loops). 3372 for( uint i = 0; i < frc._tests.size(); i++ ) { 3373 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3374 // Get number of CFG targets. 3375 // Note that PCTables include exception targets after calls. 3376 uint required_outcnt = n->required_outcnt(); 3377 if (n->outcnt() != required_outcnt) { 3378 // Check for a few special cases. Rethrow Nodes never take the 3379 // 'fall-thru' path, so expected kids is 1 less. 3380 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3381 if (n->in(0)->in(0)->is_Call()) { 3382 CallNode *call = n->in(0)->in(0)->as_Call(); 3383 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3384 required_outcnt--; // Rethrow always has 1 less kid 3385 } else if (call->req() > TypeFunc::Parms && 3386 call->is_CallDynamicJava()) { 3387 // Check for null receiver. In such case, the optimizer has 3388 // detected that the virtual call will always result in a null 3389 // pointer exception. The fall-through projection of this CatchNode 3390 // will not be populated. 3391 Node *arg0 = call->in(TypeFunc::Parms); 3392 if (arg0->is_Type() && 3393 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3394 required_outcnt--; 3395 } 3396 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3397 call->req() > TypeFunc::Parms+1 && 3398 call->is_CallStaticJava()) { 3399 // Check for negative array length. In such case, the optimizer has 3400 // detected that the allocation attempt will always result in an 3401 // exception. There is no fall-through projection of this CatchNode . 3402 Node *arg1 = call->in(TypeFunc::Parms+1); 3403 if (arg1->is_Type() && 3404 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3405 required_outcnt--; 3406 } 3407 } 3408 } 3409 } 3410 // Recheck with a better notion of 'required_outcnt' 3411 if (n->outcnt() != required_outcnt) { 3412 record_method_not_compilable("malformed control flow"); 3413 return true; // Not all targets reachable! 3414 } 3415 } 3416 // Check that I actually visited all kids. Unreached kids 3417 // must be infinite loops. 3418 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3419 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3420 record_method_not_compilable("infinite loop"); 3421 return true; // Found unvisited kid; must be unreach 3422 } 3423 } 3424 3425 // If original bytecodes contained a mixture of floats and doubles 3426 // check if the optimizer has made it homogenous, item (3). 3427 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3428 frc.get_float_count() > 32 && 3429 frc.get_double_count() == 0 && 3430 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3431 set_24_bit_selection_and_mode( false, true ); 3432 } 3433 3434 set_java_calls(frc.get_java_call_count()); 3435 set_inner_loops(frc.get_inner_loop_count()); 3436 3437 // No infinite loops, no reason to bail out. 3438 return false; 3439 } 3440 3441 //-----------------------------too_many_traps---------------------------------- 3442 // Report if there are too many traps at the current method and bci. 3443 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3444 bool Compile::too_many_traps(ciMethod* method, 3445 int bci, 3446 Deoptimization::DeoptReason reason) { 3447 ciMethodData* md = method->method_data(); 3448 if (md->is_empty()) { 3449 // Assume the trap has not occurred, or that it occurred only 3450 // because of a transient condition during start-up in the interpreter. 3451 return false; 3452 } 3453 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3454 if (md->has_trap_at(bci, m, reason) != 0) { 3455 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3456 // Also, if there are multiple reasons, or if there is no per-BCI record, 3457 // assume the worst. 3458 if (log()) 3459 log()->elem("observe trap='%s' count='%d'", 3460 Deoptimization::trap_reason_name(reason), 3461 md->trap_count(reason)); 3462 return true; 3463 } else { 3464 // Ignore method/bci and see if there have been too many globally. 3465 return too_many_traps(reason, md); 3466 } 3467 } 3468 3469 // Less-accurate variant which does not require a method and bci. 3470 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3471 ciMethodData* logmd) { 3472 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3473 // Too many traps globally. 3474 // Note that we use cumulative trap_count, not just md->trap_count. 3475 if (log()) { 3476 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3477 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3478 Deoptimization::trap_reason_name(reason), 3479 mcount, trap_count(reason)); 3480 } 3481 return true; 3482 } else { 3483 // The coast is clear. 3484 return false; 3485 } 3486 } 3487 3488 //--------------------------too_many_recompiles-------------------------------- 3489 // Report if there are too many recompiles at the current method and bci. 3490 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3491 // Is not eager to return true, since this will cause the compiler to use 3492 // Action_none for a trap point, to avoid too many recompilations. 3493 bool Compile::too_many_recompiles(ciMethod* method, 3494 int bci, 3495 Deoptimization::DeoptReason reason) { 3496 ciMethodData* md = method->method_data(); 3497 if (md->is_empty()) { 3498 // Assume the trap has not occurred, or that it occurred only 3499 // because of a transient condition during start-up in the interpreter. 3500 return false; 3501 } 3502 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3503 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3504 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3505 Deoptimization::DeoptReason per_bc_reason 3506 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3507 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3508 if ((per_bc_reason == Deoptimization::Reason_none 3509 || md->has_trap_at(bci, m, reason) != 0) 3510 // The trap frequency measure we care about is the recompile count: 3511 && md->trap_recompiled_at(bci, m) 3512 && md->overflow_recompile_count() >= bc_cutoff) { 3513 // Do not emit a trap here if it has already caused recompilations. 3514 // Also, if there are multiple reasons, or if there is no per-BCI record, 3515 // assume the worst. 3516 if (log()) 3517 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3518 Deoptimization::trap_reason_name(reason), 3519 md->trap_count(reason), 3520 md->overflow_recompile_count()); 3521 return true; 3522 } else if (trap_count(reason) != 0 3523 && decompile_count() >= m_cutoff) { 3524 // Too many recompiles globally, and we have seen this sort of trap. 3525 // Use cumulative decompile_count, not just md->decompile_count. 3526 if (log()) 3527 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3528 Deoptimization::trap_reason_name(reason), 3529 md->trap_count(reason), trap_count(reason), 3530 md->decompile_count(), decompile_count()); 3531 return true; 3532 } else { 3533 // The coast is clear. 3534 return false; 3535 } 3536 } 3537 3538 // Compute when not to trap. Used by matching trap based nodes and 3539 // NullCheck optimization. 3540 void Compile::set_allowed_deopt_reasons() { 3541 _allowed_reasons = 0; 3542 if (is_method_compilation()) { 3543 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3544 assert(rs < BitsPerInt, "recode bit map"); 3545 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3546 _allowed_reasons |= nth_bit(rs); 3547 } 3548 } 3549 } 3550 } 3551 3552 #ifndef PRODUCT 3553 //------------------------------verify_graph_edges--------------------------- 3554 // Walk the Graph and verify that there is a one-to-one correspondence 3555 // between Use-Def edges and Def-Use edges in the graph. 3556 void Compile::verify_graph_edges(bool no_dead_code) { 3557 if (VerifyGraphEdges) { 3558 ResourceArea *area = Thread::current()->resource_area(); 3559 Unique_Node_List visited(area); 3560 // Call recursive graph walk to check edges 3561 _root->verify_edges(visited); 3562 if (no_dead_code) { 3563 // Now make sure that no visited node is used by an unvisited node. 3564 bool dead_nodes = false; 3565 Unique_Node_List checked(area); 3566 while (visited.size() > 0) { 3567 Node* n = visited.pop(); 3568 checked.push(n); 3569 for (uint i = 0; i < n->outcnt(); i++) { 3570 Node* use = n->raw_out(i); 3571 if (checked.member(use)) continue; // already checked 3572 if (visited.member(use)) continue; // already in the graph 3573 if (use->is_Con()) continue; // a dead ConNode is OK 3574 // At this point, we have found a dead node which is DU-reachable. 3575 if (!dead_nodes) { 3576 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3577 dead_nodes = true; 3578 } 3579 use->dump(2); 3580 tty->print_cr("---"); 3581 checked.push(use); // No repeats; pretend it is now checked. 3582 } 3583 } 3584 assert(!dead_nodes, "using nodes must be reachable from root"); 3585 } 3586 } 3587 } 3588 3589 // Verify GC barriers consistency 3590 // Currently supported: 3591 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3592 void Compile::verify_barriers() { 3593 if (UseG1GC) { 3594 // Verify G1 pre-barriers 3595 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3596 3597 ResourceArea *area = Thread::current()->resource_area(); 3598 Unique_Node_List visited(area); 3599 Node_List worklist(area); 3600 // We're going to walk control flow backwards starting from the Root 3601 worklist.push(_root); 3602 while (worklist.size() > 0) { 3603 Node* x = worklist.pop(); 3604 if (x == NULL || x == top()) continue; 3605 if (visited.member(x)) { 3606 continue; 3607 } else { 3608 visited.push(x); 3609 } 3610 3611 if (x->is_Region()) { 3612 for (uint i = 1; i < x->req(); i++) { 3613 worklist.push(x->in(i)); 3614 } 3615 } else { 3616 worklist.push(x->in(0)); 3617 // We are looking for the pattern: 3618 // /->ThreadLocal 3619 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3620 // \->ConI(0) 3621 // We want to verify that the If and the LoadB have the same control 3622 // See GraphKit::g1_write_barrier_pre() 3623 if (x->is_If()) { 3624 IfNode *iff = x->as_If(); 3625 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3626 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3627 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3628 && cmp->in(1)->is_Load()) { 3629 LoadNode* load = cmp->in(1)->as_Load(); 3630 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3631 && load->in(2)->in(3)->is_Con() 3632 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3633 3634 Node* if_ctrl = iff->in(0); 3635 Node* load_ctrl = load->in(0); 3636 3637 if (if_ctrl != load_ctrl) { 3638 // Skip possible CProj->NeverBranch in infinite loops 3639 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3640 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3641 if_ctrl = if_ctrl->in(0)->in(0); 3642 } 3643 } 3644 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3645 } 3646 } 3647 } 3648 } 3649 } 3650 } 3651 } 3652 } 3653 3654 #endif 3655 3656 // The Compile object keeps track of failure reasons separately from the ciEnv. 3657 // This is required because there is not quite a 1-1 relation between the 3658 // ciEnv and its compilation task and the Compile object. Note that one 3659 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3660 // to backtrack and retry without subsuming loads. Other than this backtracking 3661 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3662 // by the logic in C2Compiler. 3663 void Compile::record_failure(const char* reason) { 3664 if (log() != NULL) { 3665 log()->elem("failure reason='%s' phase='compile'", reason); 3666 } 3667 if (_failure_reason == NULL) { 3668 // Record the first failure reason. 3669 _failure_reason = reason; 3670 } 3671 3672 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3673 C->print_method(PHASE_FAILURE); 3674 } 3675 _root = NULL; // flush the graph, too 3676 } 3677 3678 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3679 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3680 _phase_name(name), _dolog(dolog) 3681 { 3682 if (dolog) { 3683 C = Compile::current(); 3684 _log = C->log(); 3685 } else { 3686 C = NULL; 3687 _log = NULL; 3688 } 3689 if (_log != NULL) { 3690 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3691 _log->stamp(); 3692 _log->end_head(); 3693 } 3694 } 3695 3696 Compile::TracePhase::~TracePhase() { 3697 3698 C = Compile::current(); 3699 if (_dolog) { 3700 _log = C->log(); 3701 } else { 3702 _log = NULL; 3703 } 3704 3705 #ifdef ASSERT 3706 if (PrintIdealNodeCount) { 3707 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3708 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3709 } 3710 3711 if (VerifyIdealNodeCount) { 3712 Compile::current()->print_missing_nodes(); 3713 } 3714 #endif 3715 3716 if (_log != NULL) { 3717 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3718 } 3719 } 3720 3721 //============================================================================= 3722 // Two Constant's are equal when the type and the value are equal. 3723 bool Compile::Constant::operator==(const Constant& other) { 3724 if (type() != other.type() ) return false; 3725 if (can_be_reused() != other.can_be_reused()) return false; 3726 // For floating point values we compare the bit pattern. 3727 switch (type()) { 3728 case T_FLOAT: return (_v._value.i == other._v._value.i); 3729 case T_LONG: 3730 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3731 case T_OBJECT: 3732 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3733 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3734 case T_METADATA: return (_v._metadata == other._v._metadata); 3735 default: ShouldNotReachHere(); 3736 } 3737 return false; 3738 } 3739 3740 static int type_to_size_in_bytes(BasicType t) { 3741 switch (t) { 3742 case T_LONG: return sizeof(jlong ); 3743 case T_FLOAT: return sizeof(jfloat ); 3744 case T_DOUBLE: return sizeof(jdouble); 3745 case T_METADATA: return sizeof(Metadata*); 3746 // We use T_VOID as marker for jump-table entries (labels) which 3747 // need an internal word relocation. 3748 case T_VOID: 3749 case T_ADDRESS: 3750 case T_OBJECT: return sizeof(jobject); 3751 } 3752 3753 ShouldNotReachHere(); 3754 return -1; 3755 } 3756 3757 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3758 // sort descending 3759 if (a->freq() > b->freq()) return -1; 3760 if (a->freq() < b->freq()) return 1; 3761 return 0; 3762 } 3763 3764 void Compile::ConstantTable::calculate_offsets_and_size() { 3765 // First, sort the array by frequencies. 3766 _constants.sort(qsort_comparator); 3767 3768 #ifdef ASSERT 3769 // Make sure all jump-table entries were sorted to the end of the 3770 // array (they have a negative frequency). 3771 bool found_void = false; 3772 for (int i = 0; i < _constants.length(); i++) { 3773 Constant con = _constants.at(i); 3774 if (con.type() == T_VOID) 3775 found_void = true; // jump-tables 3776 else 3777 assert(!found_void, "wrong sorting"); 3778 } 3779 #endif 3780 3781 int offset = 0; 3782 for (int i = 0; i < _constants.length(); i++) { 3783 Constant* con = _constants.adr_at(i); 3784 3785 // Align offset for type. 3786 int typesize = type_to_size_in_bytes(con->type()); 3787 offset = align_size_up(offset, typesize); 3788 con->set_offset(offset); // set constant's offset 3789 3790 if (con->type() == T_VOID) { 3791 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3792 offset = offset + typesize * n->outcnt(); // expand jump-table 3793 } else { 3794 offset = offset + typesize; 3795 } 3796 } 3797 3798 // Align size up to the next section start (which is insts; see 3799 // CodeBuffer::align_at_start). 3800 assert(_size == -1, "already set?"); 3801 _size = align_size_up(offset, CodeEntryAlignment); 3802 } 3803 3804 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3805 MacroAssembler _masm(&cb); 3806 for (int i = 0; i < _constants.length(); i++) { 3807 Constant con = _constants.at(i); 3808 address constant_addr = NULL; 3809 switch (con.type()) { 3810 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3811 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3812 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3813 case T_OBJECT: { 3814 jobject obj = con.get_jobject(); 3815 int oop_index = _masm.oop_recorder()->find_index(obj); 3816 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3817 break; 3818 } 3819 case T_ADDRESS: { 3820 address addr = (address) con.get_jobject(); 3821 constant_addr = _masm.address_constant(addr); 3822 break; 3823 } 3824 // We use T_VOID as marker for jump-table entries (labels) which 3825 // need an internal word relocation. 3826 case T_VOID: { 3827 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3828 // Fill the jump-table with a dummy word. The real value is 3829 // filled in later in fill_jump_table. 3830 address dummy = (address) n; 3831 constant_addr = _masm.address_constant(dummy); 3832 // Expand jump-table 3833 for (uint i = 1; i < n->outcnt(); i++) { 3834 address temp_addr = _masm.address_constant(dummy + i); 3835 assert(temp_addr, "consts section too small"); 3836 } 3837 break; 3838 } 3839 case T_METADATA: { 3840 Metadata* obj = con.get_metadata(); 3841 int metadata_index = _masm.oop_recorder()->find_index(obj); 3842 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3843 break; 3844 } 3845 default: ShouldNotReachHere(); 3846 } 3847 assert(constant_addr, "consts section too small"); 3848 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3849 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3850 } 3851 } 3852 3853 int Compile::ConstantTable::find_offset(Constant& con) const { 3854 int idx = _constants.find(con); 3855 assert(idx != -1, "constant must be in constant table"); 3856 int offset = _constants.at(idx).offset(); 3857 assert(offset != -1, "constant table not emitted yet?"); 3858 return offset; 3859 } 3860 3861 void Compile::ConstantTable::add(Constant& con) { 3862 if (con.can_be_reused()) { 3863 int idx = _constants.find(con); 3864 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3865 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3866 return; 3867 } 3868 } 3869 (void) _constants.append(con); 3870 } 3871 3872 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3873 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3874 Constant con(type, value, b->_freq); 3875 add(con); 3876 return con; 3877 } 3878 3879 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3880 Constant con(metadata); 3881 add(con); 3882 return con; 3883 } 3884 3885 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3886 jvalue value; 3887 BasicType type = oper->type()->basic_type(); 3888 switch (type) { 3889 case T_LONG: value.j = oper->constantL(); break; 3890 case T_FLOAT: value.f = oper->constantF(); break; 3891 case T_DOUBLE: value.d = oper->constantD(); break; 3892 case T_OBJECT: 3893 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3894 case T_METADATA: return add((Metadata*)oper->constant()); break; 3895 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3896 } 3897 return add(n, type, value); 3898 } 3899 3900 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3901 jvalue value; 3902 // We can use the node pointer here to identify the right jump-table 3903 // as this method is called from Compile::Fill_buffer right before 3904 // the MachNodes are emitted and the jump-table is filled (means the 3905 // MachNode pointers do not change anymore). 3906 value.l = (jobject) n; 3907 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3908 add(con); 3909 return con; 3910 } 3911 3912 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3913 // If called from Compile::scratch_emit_size do nothing. 3914 if (Compile::current()->in_scratch_emit_size()) return; 3915 3916 assert(labels.is_nonempty(), "must be"); 3917 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3918 3919 // Since MachConstantNode::constant_offset() also contains 3920 // table_base_offset() we need to subtract the table_base_offset() 3921 // to get the plain offset into the constant table. 3922 int offset = n->constant_offset() - table_base_offset(); 3923 3924 MacroAssembler _masm(&cb); 3925 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3926 3927 for (uint i = 0; i < n->outcnt(); i++) { 3928 address* constant_addr = &jump_table_base[i]; 3929 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))); 3930 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3931 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3932 } 3933 } 3934 3935 void Compile::dump_inlining() { 3936 if (print_inlining() || print_intrinsics()) { 3937 // Print inlining message for candidates that we couldn't inline 3938 // for lack of space or non constant receiver 3939 for (int i = 0; i < _late_inlines.length(); i++) { 3940 CallGenerator* cg = _late_inlines.at(i); 3941 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3942 } 3943 Unique_Node_List useful; 3944 useful.push(root()); 3945 for (uint next = 0; next < useful.size(); ++next) { 3946 Node* n = useful.at(next); 3947 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3948 CallNode* call = n->as_Call(); 3949 CallGenerator* cg = call->generator(); 3950 cg->print_inlining_late("receiver not constant"); 3951 } 3952 uint max = n->len(); 3953 for ( uint i = 0; i < max; ++i ) { 3954 Node *m = n->in(i); 3955 if ( m == NULL ) continue; 3956 useful.push(m); 3957 } 3958 } 3959 for (int i = 0; i < _print_inlining_list->length(); i++) { 3960 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3961 } 3962 } 3963 } 3964 3965 // Dump inlining replay data to the stream. 3966 // Don't change thread state and acquire any locks. 3967 void Compile::dump_inline_data(outputStream* out) { 3968 InlineTree* inl_tree = ilt(); 3969 if (inl_tree != NULL) { 3970 out->print(" inline %d", inl_tree->count()); 3971 inl_tree->dump_replay_data(out); 3972 } 3973 } 3974 3975 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3976 if (n1->Opcode() < n2->Opcode()) return -1; 3977 else if (n1->Opcode() > n2->Opcode()) return 1; 3978 3979 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())); 3980 for (uint i = 1; i < n1->req(); i++) { 3981 if (n1->in(i) < n2->in(i)) return -1; 3982 else if (n1->in(i) > n2->in(i)) return 1; 3983 } 3984 3985 return 0; 3986 } 3987 3988 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3989 Node* n1 = *n1p; 3990 Node* n2 = *n2p; 3991 3992 return cmp_expensive_nodes(n1, n2); 3993 } 3994 3995 void Compile::sort_expensive_nodes() { 3996 if (!expensive_nodes_sorted()) { 3997 _expensive_nodes->sort(cmp_expensive_nodes); 3998 } 3999 } 4000 4001 bool Compile::expensive_nodes_sorted() const { 4002 for (int i = 1; i < _expensive_nodes->length(); i++) { 4003 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4004 return false; 4005 } 4006 } 4007 return true; 4008 } 4009 4010 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4011 if (_expensive_nodes->length() == 0) { 4012 return false; 4013 } 4014 4015 assert(OptimizeExpensiveOps, "optimization off?"); 4016 4017 // Take this opportunity to remove dead nodes from the list 4018 int j = 0; 4019 for (int i = 0; i < _expensive_nodes->length(); i++) { 4020 Node* n = _expensive_nodes->at(i); 4021 if (!n->is_unreachable(igvn)) { 4022 assert(n->is_expensive(), "should be expensive"); 4023 _expensive_nodes->at_put(j, n); 4024 j++; 4025 } 4026 } 4027 _expensive_nodes->trunc_to(j); 4028 4029 // Then sort the list so that similar nodes are next to each other 4030 // and check for at least two nodes of identical kind with same data 4031 // inputs. 4032 sort_expensive_nodes(); 4033 4034 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4035 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4036 return true; 4037 } 4038 } 4039 4040 return false; 4041 } 4042 4043 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4044 if (_expensive_nodes->length() == 0) { 4045 return; 4046 } 4047 4048 assert(OptimizeExpensiveOps, "optimization off?"); 4049 4050 // Sort to bring similar nodes next to each other and clear the 4051 // control input of nodes for which there's only a single copy. 4052 sort_expensive_nodes(); 4053 4054 int j = 0; 4055 int identical = 0; 4056 int i = 0; 4057 for (; i < _expensive_nodes->length()-1; i++) { 4058 assert(j <= i, "can't write beyond current index"); 4059 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4060 identical++; 4061 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4062 continue; 4063 } 4064 if (identical > 0) { 4065 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4066 identical = 0; 4067 } else { 4068 Node* n = _expensive_nodes->at(i); 4069 igvn.hash_delete(n); 4070 n->set_req(0, NULL); 4071 igvn.hash_insert(n); 4072 } 4073 } 4074 if (identical > 0) { 4075 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4076 } else if (_expensive_nodes->length() >= 1) { 4077 Node* n = _expensive_nodes->at(i); 4078 igvn.hash_delete(n); 4079 n->set_req(0, NULL); 4080 igvn.hash_insert(n); 4081 } 4082 _expensive_nodes->trunc_to(j); 4083 } 4084 4085 void Compile::add_expensive_node(Node * n) { 4086 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4087 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4088 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4089 if (OptimizeExpensiveOps) { 4090 _expensive_nodes->append(n); 4091 } else { 4092 // Clear control input and let IGVN optimize expensive nodes if 4093 // OptimizeExpensiveOps is off. 4094 n->set_req(0, NULL); 4095 } 4096 } 4097 4098 /** 4099 * Remove the speculative part of types and clean up the graph 4100 */ 4101 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4102 if (UseTypeSpeculation) { 4103 Unique_Node_List worklist; 4104 worklist.push(root()); 4105 int modified = 0; 4106 // Go over all type nodes that carry a speculative type, drop the 4107 // speculative part of the type and enqueue the node for an igvn 4108 // which may optimize it out. 4109 for (uint next = 0; next < worklist.size(); ++next) { 4110 Node *n = worklist.at(next); 4111 if (n->is_Type()) { 4112 TypeNode* tn = n->as_Type(); 4113 const Type* t = tn->type(); 4114 const Type* t_no_spec = t->remove_speculative(); 4115 if (t_no_spec != t) { 4116 bool in_hash = igvn.hash_delete(n); 4117 assert(in_hash, "node should be in igvn hash table"); 4118 tn->set_type(t_no_spec); 4119 igvn.hash_insert(n); 4120 igvn._worklist.push(n); // give it a chance to go away 4121 modified++; 4122 } 4123 } 4124 uint max = n->len(); 4125 for( uint i = 0; i < max; ++i ) { 4126 Node *m = n->in(i); 4127 if (not_a_node(m)) continue; 4128 worklist.push(m); 4129 } 4130 } 4131 // Drop the speculative part of all types in the igvn's type table 4132 igvn.remove_speculative_types(); 4133 if (modified > 0) { 4134 igvn.optimize(); 4135 } 4136 #ifdef ASSERT 4137 // Verify that after the IGVN is over no speculative type has resurfaced 4138 worklist.clear(); 4139 worklist.push(root()); 4140 for (uint next = 0; next < worklist.size(); ++next) { 4141 Node *n = worklist.at(next); 4142 const Type* t = igvn.type_or_null(n); 4143 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4144 if (n->is_Type()) { 4145 t = n->as_Type()->type(); 4146 assert(t == t->remove_speculative(), "no more speculative types"); 4147 } 4148 uint max = n->len(); 4149 for( uint i = 0; i < max; ++i ) { 4150 Node *m = n->in(i); 4151 if (not_a_node(m)) continue; 4152 worklist.push(m); 4153 } 4154 } 4155 igvn.check_no_speculative_types(); 4156 #endif 4157 } 4158 } 4159 4160 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4161 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4162 if (ctrl != NULL) { 4163 // Express control dependency by a CastII node with a narrow type. 4164 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */); 4165 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4166 // node from floating above the range check during loop optimizations. Otherwise, the 4167 // ConvI2L node may be eliminated independently of the range check, causing the data path 4168 // to become TOP while the control path is still there (although it's unreachable). 4169 value->set_req(0, ctrl); 4170 // Save CastII node to remove it after loop optimizations. 4171 phase->C->add_range_check_cast(value); 4172 value = phase->transform(value); 4173 } 4174 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4175 return phase->transform(new (phase->C) ConvI2LNode(value, ltype)); 4176 } 4177 4178 // Auxiliary method to support randomized stressing/fuzzing. 4179 // 4180 // This method can be called the arbitrary number of times, with current count 4181 // as the argument. The logic allows selecting a single candidate from the 4182 // running list of candidates as follows: 4183 // int count = 0; 4184 // Cand* selected = null; 4185 // while(cand = cand->next()) { 4186 // if (randomized_select(++count)) { 4187 // selected = cand; 4188 // } 4189 // } 4190 // 4191 // Including count equalizes the chances any candidate is "selected". 4192 // This is useful when we don't have the complete list of candidates to choose 4193 // from uniformly. In this case, we need to adjust the randomicity of the 4194 // selection, or else we will end up biasing the selection towards the latter 4195 // candidates. 4196 // 4197 // Quick back-envelope calculation shows that for the list of n candidates 4198 // the equal probability for the candidate to persist as "best" can be 4199 // achieved by replacing it with "next" k-th candidate with the probability 4200 // of 1/k. It can be easily shown that by the end of the run, the 4201 // probability for any candidate is converged to 1/n, thus giving the 4202 // uniform distribution among all the candidates. 4203 // 4204 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4205 #define RANDOMIZED_DOMAIN_POW 29 4206 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4207 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4208 bool Compile::randomized_select(int count) { 4209 assert(count > 0, "only positive"); 4210 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4211 }