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