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