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