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