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_32 85 # include "adfiles/ad_ppc_32.hpp" 86 #endif 87 #ifdef TARGET_ARCH_MODEL_ppc_64 88 # include "adfiles/ad_ppc_64.hpp" 89 #endif 90 91 92 // -------------------- Compile::mach_constant_base_node ----------------------- 93 // Constant table base node singleton. 94 MachConstantBaseNode* Compile::mach_constant_base_node() { 95 if (_mach_constant_base_node == NULL) { 96 _mach_constant_base_node = new (C) MachConstantBaseNode(); 97 _mach_constant_base_node->add_req(C->root()); 98 } 99 return _mach_constant_base_node; 100 } 101 102 103 /// Support for intrinsics. 104 105 // Return the index at which m must be inserted (or already exists). 106 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 107 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 108 #ifdef ASSERT 109 for (int i = 1; i < _intrinsics->length(); i++) { 110 CallGenerator* cg1 = _intrinsics->at(i-1); 111 CallGenerator* cg2 = _intrinsics->at(i); 112 assert(cg1->method() != cg2->method() 113 ? cg1->method() < cg2->method() 114 : cg1->is_virtual() < cg2->is_virtual(), 115 "compiler intrinsics list must stay sorted"); 116 } 117 #endif 118 // Binary search sorted list, in decreasing intervals [lo, hi]. 119 int lo = 0, hi = _intrinsics->length()-1; 120 while (lo <= hi) { 121 int mid = (uint)(hi + lo) / 2; 122 ciMethod* mid_m = _intrinsics->at(mid)->method(); 123 if (m < mid_m) { 124 hi = mid-1; 125 } else if (m > mid_m) { 126 lo = mid+1; 127 } else { 128 // look at minor sort key 129 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 130 if (is_virtual < mid_virt) { 131 hi = mid-1; 132 } else if (is_virtual > mid_virt) { 133 lo = mid+1; 134 } else { 135 return mid; // exact match 136 } 137 } 138 } 139 return lo; // inexact match 140 } 141 142 void Compile::register_intrinsic(CallGenerator* cg) { 143 if (_intrinsics == NULL) { 144 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 145 } 146 // This code is stolen from ciObjectFactory::insert. 147 // Really, GrowableArray should have methods for 148 // insert_at, remove_at, and binary_search. 149 int len = _intrinsics->length(); 150 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 151 if (index == len) { 152 _intrinsics->append(cg); 153 } else { 154 #ifdef ASSERT 155 CallGenerator* oldcg = _intrinsics->at(index); 156 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 157 #endif 158 _intrinsics->append(_intrinsics->at(len-1)); 159 int pos; 160 for (pos = len-2; pos >= index; pos--) { 161 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 162 } 163 _intrinsics->at_put(index, cg); 164 } 165 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 166 } 167 168 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 169 assert(m->is_loaded(), "don't try this on unloaded methods"); 170 if (_intrinsics != NULL) { 171 int index = intrinsic_insertion_index(m, is_virtual); 172 if (index < _intrinsics->length() 173 && _intrinsics->at(index)->method() == m 174 && _intrinsics->at(index)->is_virtual() == is_virtual) { 175 return _intrinsics->at(index); 176 } 177 } 178 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 179 if (m->intrinsic_id() != vmIntrinsics::_none && 180 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 181 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 182 if (cg != NULL) { 183 // Save it for next time: 184 register_intrinsic(cg); 185 return cg; 186 } else { 187 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 188 } 189 } 190 return NULL; 191 } 192 193 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 194 // in library_call.cpp. 195 196 197 #ifndef PRODUCT 198 // statistics gathering... 199 200 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 201 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 202 203 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 204 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 205 int oflags = _intrinsic_hist_flags[id]; 206 assert(flags != 0, "what happened?"); 207 if (is_virtual) { 208 flags |= _intrinsic_virtual; 209 } 210 bool changed = (flags != oflags); 211 if ((flags & _intrinsic_worked) != 0) { 212 juint count = (_intrinsic_hist_count[id] += 1); 213 if (count == 1) { 214 changed = true; // first time 215 } 216 // increment the overall count also: 217 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 218 } 219 if (changed) { 220 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 221 // Something changed about the intrinsic's virtuality. 222 if ((flags & _intrinsic_virtual) != 0) { 223 // This is the first use of this intrinsic as a virtual call. 224 if (oflags != 0) { 225 // We already saw it as a non-virtual, so note both cases. 226 flags |= _intrinsic_both; 227 } 228 } else if ((oflags & _intrinsic_both) == 0) { 229 // This is the first use of this intrinsic as a non-virtual 230 flags |= _intrinsic_both; 231 } 232 } 233 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 234 } 235 // update the overall flags also: 236 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 237 return changed; 238 } 239 240 static char* format_flags(int flags, char* buf) { 241 buf[0] = 0; 242 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 243 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 244 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 245 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 246 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 247 if (buf[0] == 0) strcat(buf, ","); 248 assert(buf[0] == ',', "must be"); 249 return &buf[1]; 250 } 251 252 void Compile::print_intrinsic_statistics() { 253 char flagsbuf[100]; 254 ttyLocker ttyl; 255 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 256 tty->print_cr("Compiler intrinsic usage:"); 257 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 258 if (total == 0) total = 1; // avoid div0 in case of no successes 259 #define PRINT_STAT_LINE(name, c, f) \ 260 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 261 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 262 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 263 int flags = _intrinsic_hist_flags[id]; 264 juint count = _intrinsic_hist_count[id]; 265 if ((flags | count) != 0) { 266 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 267 } 268 } 269 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 270 if (xtty != NULL) xtty->tail("statistics"); 271 } 272 273 void Compile::print_statistics() { 274 { ttyLocker ttyl; 275 if (xtty != NULL) xtty->head("statistics type='opto'"); 276 Parse::print_statistics(); 277 PhaseCCP::print_statistics(); 278 PhaseRegAlloc::print_statistics(); 279 Scheduling::print_statistics(); 280 PhasePeephole::print_statistics(); 281 PhaseIdealLoop::print_statistics(); 282 if (xtty != NULL) xtty->tail("statistics"); 283 } 284 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 285 // put this under its own <statistics> element. 286 print_intrinsic_statistics(); 287 } 288 } 289 #endif //PRODUCT 290 291 // Support for bundling info 292 Bundle* Compile::node_bundling(const Node *n) { 293 assert(valid_bundle_info(n), "oob"); 294 return &_node_bundling_base[n->_idx]; 295 } 296 297 bool Compile::valid_bundle_info(const Node *n) { 298 return (_node_bundling_limit > n->_idx); 299 } 300 301 302 void Compile::gvn_replace_by(Node* n, Node* nn) { 303 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 304 Node* use = n->last_out(i); 305 bool is_in_table = initial_gvn()->hash_delete(use); 306 uint uses_found = 0; 307 for (uint j = 0; j < use->len(); j++) { 308 if (use->in(j) == n) { 309 if (j < use->req()) 310 use->set_req(j, nn); 311 else 312 use->set_prec(j, nn); 313 uses_found++; 314 } 315 } 316 if (is_in_table) { 317 // reinsert into table 318 initial_gvn()->hash_find_insert(use); 319 } 320 record_for_igvn(use); 321 i -= uses_found; // we deleted 1 or more copies of this edge 322 } 323 } 324 325 326 static inline bool not_a_node(const Node* n) { 327 if (n == NULL) return true; 328 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 329 if (*(address*)n == badAddress) return true; // kill by Node::destruct 330 return false; 331 } 332 333 // Identify all nodes that are reachable from below, useful. 334 // Use breadth-first pass that records state in a Unique_Node_List, 335 // recursive traversal is slower. 336 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 337 int estimated_worklist_size = unique(); 338 useful.map( estimated_worklist_size, NULL ); // preallocate space 339 340 // Initialize worklist 341 if (root() != NULL) { useful.push(root()); } 342 // If 'top' is cached, declare it useful to preserve cached node 343 if( cached_top_node() ) { useful.push(cached_top_node()); } 344 345 // Push all useful nodes onto the list, breadthfirst 346 for( uint next = 0; next < useful.size(); ++next ) { 347 assert( next < unique(), "Unique useful nodes < total nodes"); 348 Node *n = useful.at(next); 349 uint max = n->len(); 350 for( uint i = 0; i < max; ++i ) { 351 Node *m = n->in(i); 352 if (not_a_node(m)) continue; 353 useful.push(m); 354 } 355 } 356 } 357 358 // Update dead_node_list with any missing dead nodes using useful 359 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 360 void Compile::update_dead_node_list(Unique_Node_List &useful) { 361 uint max_idx = unique(); 362 VectorSet& useful_node_set = useful.member_set(); 363 364 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 365 // If node with index node_idx is not in useful set, 366 // mark it as dead in dead node list. 367 if (! useful_node_set.test(node_idx) ) { 368 record_dead_node(node_idx); 369 } 370 } 371 } 372 373 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 374 int shift = 0; 375 for (int i = 0; i < inlines->length(); i++) { 376 CallGenerator* cg = inlines->at(i); 377 CallNode* call = cg->call_node(); 378 if (shift > 0) { 379 inlines->at_put(i-shift, cg); 380 } 381 if (!useful.member(call)) { 382 shift++; 383 } 384 } 385 inlines->trunc_to(inlines->length()-shift); 386 } 387 388 // Disconnect all useless nodes by disconnecting those at the boundary. 389 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 390 uint next = 0; 391 while (next < useful.size()) { 392 Node *n = useful.at(next++); 393 // Use raw traversal of out edges since this code removes out edges 394 int max = n->outcnt(); 395 for (int j = 0; j < max; ++j) { 396 Node* child = n->raw_out(j); 397 if (! useful.member(child)) { 398 assert(!child->is_top() || child != top(), 399 "If top is cached in Compile object it is in useful list"); 400 // Only need to remove this out-edge to the useless node 401 n->raw_del_out(j); 402 --j; 403 --max; 404 } 405 } 406 if (n->outcnt() == 1 && n->has_special_unique_user()) { 407 record_for_igvn(n->unique_out()); 408 } 409 } 410 // Remove useless macro and predicate opaq nodes 411 for (int i = C->macro_count()-1; i >= 0; i--) { 412 Node* n = C->macro_node(i); 413 if (!useful.member(n)) { 414 remove_macro_node(n); 415 } 416 } 417 // Remove useless expensive node 418 for (int i = C->expensive_count()-1; i >= 0; i--) { 419 Node* n = C->expensive_node(i); 420 if (!useful.member(n)) { 421 remove_expensive_node(n); 422 } 423 } 424 // clean up the late inline lists 425 remove_useless_late_inlines(&_string_late_inlines, useful); 426 remove_useless_late_inlines(&_boxing_late_inlines, useful); 427 remove_useless_late_inlines(&_late_inlines, useful); 428 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 429 } 430 431 //------------------------------frame_size_in_words----------------------------- 432 // frame_slots in units of words 433 int Compile::frame_size_in_words() const { 434 // shift is 0 in LP32 and 1 in LP64 435 const int shift = (LogBytesPerWord - LogBytesPerInt); 436 int words = _frame_slots >> shift; 437 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 438 return words; 439 } 440 441 // ============================================================================ 442 //------------------------------CompileWrapper--------------------------------- 443 class CompileWrapper : public StackObj { 444 Compile *const _compile; 445 public: 446 CompileWrapper(Compile* compile); 447 448 ~CompileWrapper(); 449 }; 450 451 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 452 // the Compile* pointer is stored in the current ciEnv: 453 ciEnv* env = compile->env(); 454 assert(env == ciEnv::current(), "must already be a ciEnv active"); 455 assert(env->compiler_data() == NULL, "compile already active?"); 456 env->set_compiler_data(compile); 457 assert(compile == Compile::current(), "sanity"); 458 459 compile->set_type_dict(NULL); 460 compile->set_type_hwm(NULL); 461 compile->set_type_last_size(0); 462 compile->set_last_tf(NULL, NULL); 463 compile->set_indexSet_arena(NULL); 464 compile->set_indexSet_free_block_list(NULL); 465 compile->init_type_arena(); 466 Type::Initialize(compile); 467 _compile->set_scratch_buffer_blob(NULL); 468 _compile->begin_method(); 469 } 470 CompileWrapper::~CompileWrapper() { 471 _compile->end_method(); 472 if (_compile->scratch_buffer_blob() != NULL) 473 BufferBlob::free(_compile->scratch_buffer_blob()); 474 _compile->env()->set_compiler_data(NULL); 475 } 476 477 478 //----------------------------print_compile_messages--------------------------- 479 void Compile::print_compile_messages() { 480 #ifndef PRODUCT 481 // Check if recompiling 482 if (_subsume_loads == false && PrintOpto) { 483 // Recompiling without allowing machine instructions to subsume loads 484 tty->print_cr("*********************************************************"); 485 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 486 tty->print_cr("*********************************************************"); 487 } 488 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 489 // Recompiling without escape analysis 490 tty->print_cr("*********************************************************"); 491 tty->print_cr("** Bailout: Recompile without escape analysis **"); 492 tty->print_cr("*********************************************************"); 493 } 494 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 495 // Recompiling without boxing elimination 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (env()->break_at_compile()) { 501 // Open the debugger when compiling this method. 502 tty->print("### Breaking when compiling: "); 503 method()->print_short_name(); 504 tty->cr(); 505 BREAKPOINT; 506 } 507 508 if( PrintOpto ) { 509 if (is_osr_compilation()) { 510 tty->print("[OSR]%3d", _compile_id); 511 } else { 512 tty->print("%3d", _compile_id); 513 } 514 } 515 #endif 516 } 517 518 519 //-----------------------init_scratch_buffer_blob------------------------------ 520 // Construct a temporary BufferBlob and cache it for this compile. 521 void Compile::init_scratch_buffer_blob(int const_size) { 522 // If there is already a scratch buffer blob allocated and the 523 // constant section is big enough, use it. Otherwise free the 524 // current and allocate a new one. 525 BufferBlob* blob = scratch_buffer_blob(); 526 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 527 // Use the current blob. 528 } else { 529 if (blob != NULL) { 530 BufferBlob::free(blob); 531 } 532 533 ResourceMark rm; 534 _scratch_const_size = const_size; 535 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 536 blob = BufferBlob::create("Compile::scratch_buffer", size); 537 // Record the buffer blob for next time. 538 set_scratch_buffer_blob(blob); 539 // Have we run out of code space? 540 if (scratch_buffer_blob() == NULL) { 541 // Let CompilerBroker disable further compilations. 542 record_failure("Not enough space for scratch buffer in CodeCache"); 543 return; 544 } 545 } 546 547 // Initialize the relocation buffers 548 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 549 set_scratch_locs_memory(locs_buf); 550 } 551 552 553 //-----------------------scratch_emit_size------------------------------------- 554 // Helper function that computes size by emitting code 555 uint Compile::scratch_emit_size(const Node* n) { 556 // Start scratch_emit_size section. 557 set_in_scratch_emit_size(true); 558 559 // Emit into a trash buffer and count bytes emitted. 560 // This is a pretty expensive way to compute a size, 561 // but it works well enough if seldom used. 562 // All common fixed-size instructions are given a size 563 // method by the AD file. 564 // Note that the scratch buffer blob and locs memory are 565 // allocated at the beginning of the compile task, and 566 // may be shared by several calls to scratch_emit_size. 567 // The allocation of the scratch buffer blob is particularly 568 // expensive, since it has to grab the code cache lock. 569 BufferBlob* blob = this->scratch_buffer_blob(); 570 assert(blob != NULL, "Initialize BufferBlob at start"); 571 assert(blob->size() > MAX_inst_size, "sanity"); 572 relocInfo* locs_buf = scratch_locs_memory(); 573 address blob_begin = blob->content_begin(); 574 address blob_end = (address)locs_buf; 575 assert(blob->content_contains(blob_end), "sanity"); 576 CodeBuffer buf(blob_begin, blob_end - blob_begin); 577 buf.initialize_consts_size(_scratch_const_size); 578 buf.initialize_stubs_size(MAX_stubs_size); 579 assert(locs_buf != NULL, "sanity"); 580 int lsize = MAX_locs_size / 3; 581 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 582 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 583 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 584 585 // Do the emission. 586 587 Label fakeL; // Fake label for branch instructions. 588 Label* saveL = NULL; 589 uint save_bnum = 0; 590 bool is_branch = n->is_MachBranch(); 591 if (is_branch) { 592 MacroAssembler masm(&buf); 593 masm.bind(fakeL); 594 n->as_MachBranch()->save_label(&saveL, &save_bnum); 595 n->as_MachBranch()->label_set(&fakeL, 0); 596 } 597 n->emit(buf, this->regalloc()); 598 if (is_branch) // Restore label. 599 n->as_MachBranch()->label_set(saveL, save_bnum); 600 601 // End scratch_emit_size section. 602 set_in_scratch_emit_size(false); 603 604 return buf.insts_size(); 605 } 606 607 608 // ============================================================================ 609 //------------------------------Compile standard------------------------------- 610 debug_only( int Compile::_debug_idx = 100000; ) 611 612 // Compile a method. entry_bci is -1 for normal compilations and indicates 613 // the continuation bci for on stack replacement. 614 615 616 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 617 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 618 : Phase(Compiler), 619 _env(ci_env), 620 _log(ci_env->log()), 621 _compile_id(ci_env->compile_id()), 622 _save_argument_registers(false), 623 _stub_name(NULL), 624 _stub_function(NULL), 625 _stub_entry_point(NULL), 626 _method(target), 627 _entry_bci(osr_bci), 628 _initial_gvn(NULL), 629 _for_igvn(NULL), 630 _warm_calls(NULL), 631 _subsume_loads(subsume_loads), 632 _do_escape_analysis(do_escape_analysis), 633 _eliminate_boxing(eliminate_boxing), 634 _failure_reason(NULL), 635 _code_buffer("Compile::Fill_buffer"), 636 _orig_pc_slot(0), 637 _orig_pc_slot_offset_in_bytes(0), 638 _has_method_handle_invokes(false), 639 _mach_constant_base_node(NULL), 640 _node_bundling_limit(0), 641 _node_bundling_base(NULL), 642 _java_calls(0), 643 _inner_loops(0), 644 _scratch_const_size(-1), 645 _in_scratch_emit_size(false), 646 _dead_node_list(comp_arena()), 647 _dead_node_count(0), 648 #ifndef PRODUCT 649 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 650 _in_dump_cnt(0), 651 _printer(IdealGraphPrinter::printer()), 652 #endif 653 _congraph(NULL), 654 _late_inlines(comp_arena(), 2, 0, NULL), 655 _string_late_inlines(comp_arena(), 2, 0, NULL), 656 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 657 _late_inlines_pos(0), 658 _number_of_mh_late_inlines(0), 659 _inlining_progress(false), 660 _inlining_incrementally(false), 661 _print_inlining_list(NULL), 662 _print_inlining_idx(0), 663 _preserve_jvm_state(0) { 664 C = this; 665 666 CompileWrapper cw(this); 667 #ifndef PRODUCT 668 if (TimeCompiler2) { 669 tty->print(" "); 670 target->holder()->name()->print(); 671 tty->print("."); 672 target->print_short_name(); 673 tty->print(" "); 674 } 675 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 676 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 677 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 678 if (!print_opto_assembly) { 679 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 680 if (print_assembly && !Disassembler::can_decode()) { 681 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 682 print_opto_assembly = true; 683 } 684 } 685 set_print_assembly(print_opto_assembly); 686 set_parsed_irreducible_loop(false); 687 #endif 688 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 689 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 690 691 if (ProfileTraps) { 692 // Make sure the method being compiled gets its own MDO, 693 // so we can at least track the decompile_count(). 694 method()->ensure_method_data(); 695 } 696 697 Init(::AliasLevel); 698 699 700 print_compile_messages(); 701 702 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 703 _ilt = InlineTree::build_inline_tree_root(); 704 else 705 _ilt = NULL; 706 707 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 708 assert(num_alias_types() >= AliasIdxRaw, ""); 709 710 #define MINIMUM_NODE_HASH 1023 711 // Node list that Iterative GVN will start with 712 Unique_Node_List for_igvn(comp_arena()); 713 set_for_igvn(&for_igvn); 714 715 // GVN that will be run immediately on new nodes 716 uint estimated_size = method()->code_size()*4+64; 717 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 718 PhaseGVN gvn(node_arena(), estimated_size); 719 set_initial_gvn(&gvn); 720 721 if (print_inlining() || print_intrinsics()) { 722 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 723 } 724 { // Scope for timing the parser 725 TracePhase t3("parse", &_t_parser, true); 726 727 // Put top into the hash table ASAP. 728 initial_gvn()->transform_no_reclaim(top()); 729 730 // Set up tf(), start(), and find a CallGenerator. 731 CallGenerator* cg = NULL; 732 if (is_osr_compilation()) { 733 const TypeTuple *domain = StartOSRNode::osr_domain(); 734 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 735 init_tf(TypeFunc::make(domain, range)); 736 StartNode* s = new (this) StartOSRNode(root(), domain); 737 initial_gvn()->set_type_bottom(s); 738 init_start(s); 739 cg = CallGenerator::for_osr(method(), entry_bci()); 740 } else { 741 // Normal case. 742 init_tf(TypeFunc::make(method())); 743 StartNode* s = new (this) StartNode(root(), tf()->domain()); 744 initial_gvn()->set_type_bottom(s); 745 init_start(s); 746 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 747 // With java.lang.ref.reference.get() we must go through the 748 // intrinsic when G1 is enabled - even when get() is the root 749 // method of the compile - so that, if necessary, the value in 750 // the referent field of the reference object gets recorded by 751 // the pre-barrier code. 752 // Specifically, if G1 is enabled, the value in the referent 753 // field is recorded by the G1 SATB pre barrier. This will 754 // result in the referent being marked live and the reference 755 // object removed from the list of discovered references during 756 // reference processing. 757 cg = find_intrinsic(method(), false); 758 } 759 if (cg == NULL) { 760 float past_uses = method()->interpreter_invocation_count(); 761 float expected_uses = past_uses; 762 cg = CallGenerator::for_inline(method(), expected_uses); 763 } 764 } 765 if (failing()) return; 766 if (cg == NULL) { 767 record_method_not_compilable_all_tiers("cannot parse method"); 768 return; 769 } 770 JVMState* jvms = build_start_state(start(), tf()); 771 if ((jvms = cg->generate(jvms, NULL)) == NULL) { 772 record_method_not_compilable("method parse failed"); 773 return; 774 } 775 GraphKit kit(jvms); 776 777 if (!kit.stopped()) { 778 // Accept return values, and transfer control we know not where. 779 // This is done by a special, unique ReturnNode bound to root. 780 return_values(kit.jvms()); 781 } 782 783 if (kit.has_exceptions()) { 784 // Any exceptions that escape from this call must be rethrown 785 // to whatever caller is dynamically above us on the stack. 786 // This is done by a special, unique RethrowNode bound to root. 787 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 788 } 789 790 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 791 792 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 793 inline_string_calls(true); 794 } 795 796 if (failing()) return; 797 798 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 799 800 // Remove clutter produced by parsing. 801 if (!failing()) { 802 ResourceMark rm; 803 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 804 } 805 } 806 807 // Note: Large methods are capped off in do_one_bytecode(). 808 if (failing()) return; 809 810 // After parsing, node notes are no longer automagic. 811 // They must be propagated by register_new_node_with_optimizer(), 812 // clone(), or the like. 813 set_default_node_notes(NULL); 814 815 for (;;) { 816 int successes = Inline_Warm(); 817 if (failing()) return; 818 if (successes == 0) break; 819 } 820 821 // Drain the list. 822 Finish_Warm(); 823 #ifndef PRODUCT 824 if (_printer) { 825 _printer->print_inlining(this); 826 } 827 #endif 828 829 if (failing()) return; 830 NOT_PRODUCT( verify_graph_edges(); ) 831 832 // Now optimize 833 Optimize(); 834 if (failing()) return; 835 NOT_PRODUCT( verify_graph_edges(); ) 836 837 #ifndef PRODUCT 838 if (PrintIdeal) { 839 ttyLocker ttyl; // keep the following output all in one block 840 // This output goes directly to the tty, not the compiler log. 841 // To enable tools to match it up with the compilation activity, 842 // be sure to tag this tty output with the compile ID. 843 if (xtty != NULL) { 844 xtty->head("ideal compile_id='%d'%s", compile_id(), 845 is_osr_compilation() ? " compile_kind='osr'" : 846 ""); 847 } 848 root()->dump(9999); 849 if (xtty != NULL) { 850 xtty->tail("ideal"); 851 } 852 } 853 #endif 854 855 NOT_PRODUCT( verify_barriers(); ) 856 // Now that we know the size of all the monitors we can add a fixed slot 857 // for the original deopt pc. 858 859 _orig_pc_slot = fixed_slots(); 860 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 861 set_fixed_slots(next_slot); 862 863 // Now generate code 864 Code_Gen(); 865 if (failing()) return; 866 867 // Check if we want to skip execution of all compiled code. 868 { 869 #ifndef PRODUCT 870 if (OptoNoExecute) { 871 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 872 return; 873 } 874 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 875 #endif 876 877 if (is_osr_compilation()) { 878 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 879 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 880 } else { 881 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 882 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 883 } 884 885 env()->register_method(_method, _entry_bci, 886 &_code_offsets, 887 _orig_pc_slot_offset_in_bytes, 888 code_buffer(), 889 frame_size_in_words(), _oop_map_set, 890 &_handler_table, &_inc_table, 891 compiler, 892 env()->comp_level(), 893 has_unsafe_access(), 894 SharedRuntime::is_wide_vector(max_vector_size()) 895 ); 896 897 if (log() != NULL) // Print code cache state into compiler log 898 log()->code_cache_state(); 899 } 900 } 901 902 //------------------------------Compile---------------------------------------- 903 // Compile a runtime stub 904 Compile::Compile( ciEnv* ci_env, 905 TypeFunc_generator generator, 906 address stub_function, 907 const char *stub_name, 908 int is_fancy_jump, 909 bool pass_tls, 910 bool save_arg_registers, 911 bool return_pc ) 912 : Phase(Compiler), 913 _env(ci_env), 914 _log(ci_env->log()), 915 _compile_id(0), 916 _save_argument_registers(save_arg_registers), 917 _method(NULL), 918 _stub_name(stub_name), 919 _stub_function(stub_function), 920 _stub_entry_point(NULL), 921 _entry_bci(InvocationEntryBci), 922 _initial_gvn(NULL), 923 _for_igvn(NULL), 924 _warm_calls(NULL), 925 _orig_pc_slot(0), 926 _orig_pc_slot_offset_in_bytes(0), 927 _subsume_loads(true), 928 _do_escape_analysis(false), 929 _eliminate_boxing(false), 930 _failure_reason(NULL), 931 _code_buffer("Compile::Fill_buffer"), 932 _has_method_handle_invokes(false), 933 _mach_constant_base_node(NULL), 934 _node_bundling_limit(0), 935 _node_bundling_base(NULL), 936 _java_calls(0), 937 _inner_loops(0), 938 #ifndef PRODUCT 939 _trace_opto_output(TraceOptoOutput), 940 _in_dump_cnt(0), 941 _printer(NULL), 942 #endif 943 _dead_node_list(comp_arena()), 944 _dead_node_count(0), 945 _congraph(NULL), 946 _number_of_mh_late_inlines(0), 947 _inlining_progress(false), 948 _inlining_incrementally(false), 949 _print_inlining_list(NULL), 950 _print_inlining_idx(0), 951 _preserve_jvm_state(0) { 952 C = this; 953 954 #ifndef PRODUCT 955 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 956 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 957 set_print_assembly(PrintFrameConverterAssembly); 958 set_parsed_irreducible_loop(false); 959 #endif 960 CompileWrapper cw(this); 961 Init(/*AliasLevel=*/ 0); 962 init_tf((*generator)()); 963 964 { 965 // The following is a dummy for the sake of GraphKit::gen_stub 966 Unique_Node_List for_igvn(comp_arena()); 967 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 968 PhaseGVN gvn(Thread::current()->resource_area(),255); 969 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 970 gvn.transform_no_reclaim(top()); 971 972 GraphKit kit; 973 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 974 } 975 976 NOT_PRODUCT( verify_graph_edges(); ) 977 Code_Gen(); 978 if (failing()) return; 979 980 981 // Entry point will be accessed using compile->stub_entry_point(); 982 if (code_buffer() == NULL) { 983 Matcher::soft_match_failure(); 984 } else { 985 if (PrintAssembly && (WizardMode || Verbose)) 986 tty->print_cr("### Stub::%s", stub_name); 987 988 if (!failing()) { 989 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 990 991 // Make the NMethod 992 // For now we mark the frame as never safe for profile stackwalking 993 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 994 code_buffer(), 995 CodeOffsets::frame_never_safe, 996 // _code_offsets.value(CodeOffsets::Frame_Complete), 997 frame_size_in_words(), 998 _oop_map_set, 999 save_arg_registers); 1000 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1001 1002 _stub_entry_point = rs->entry_point(); 1003 } 1004 } 1005 } 1006 1007 //------------------------------Init------------------------------------------- 1008 // Prepare for a single compilation 1009 void Compile::Init(int aliaslevel) { 1010 _unique = 0; 1011 _regalloc = NULL; 1012 1013 _tf = NULL; // filled in later 1014 _top = NULL; // cached later 1015 _matcher = NULL; // filled in later 1016 _cfg = NULL; // filled in later 1017 1018 set_24_bit_selection_and_mode(Use24BitFP, false); 1019 1020 _node_note_array = NULL; 1021 _default_node_notes = NULL; 1022 1023 _immutable_memory = NULL; // filled in at first inquiry 1024 1025 // Globally visible Nodes 1026 // First set TOP to NULL to give safe behavior during creation of RootNode 1027 set_cached_top_node(NULL); 1028 set_root(new (this) RootNode()); 1029 // Now that you have a Root to point to, create the real TOP 1030 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1031 set_recent_alloc(NULL, NULL); 1032 1033 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1034 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1035 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1036 env()->set_dependencies(new Dependencies(env())); 1037 1038 _fixed_slots = 0; 1039 set_has_split_ifs(false); 1040 set_has_loops(has_method() && method()->has_loops()); // first approximation 1041 set_has_stringbuilder(false); 1042 set_has_boxed_value(false); 1043 _trap_can_recompile = false; // no traps emitted yet 1044 _major_progress = true; // start out assuming good things will happen 1045 set_has_unsafe_access(false); 1046 set_max_vector_size(0); 1047 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1048 set_decompile_count(0); 1049 1050 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1051 set_num_loop_opts(LoopOptsCount); 1052 set_do_inlining(Inline); 1053 set_max_inline_size(MaxInlineSize); 1054 set_freq_inline_size(FreqInlineSize); 1055 set_do_scheduling(OptoScheduling); 1056 set_do_count_invocations(false); 1057 set_do_method_data_update(false); 1058 1059 if (debug_info()->recording_non_safepoints()) { 1060 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1061 (comp_arena(), 8, 0, NULL)); 1062 set_default_node_notes(Node_Notes::make(this)); 1063 } 1064 1065 // // -- Initialize types before each compile -- 1066 // // Update cached type information 1067 // if( _method && _method->constants() ) 1068 // Type::update_loaded_types(_method, _method->constants()); 1069 1070 // Init alias_type map. 1071 if (!_do_escape_analysis && aliaslevel == 3) 1072 aliaslevel = 2; // No unique types without escape analysis 1073 _AliasLevel = aliaslevel; 1074 const int grow_ats = 16; 1075 _max_alias_types = grow_ats; 1076 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1077 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1078 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1079 { 1080 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1081 } 1082 // Initialize the first few types. 1083 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1084 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1085 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1086 _num_alias_types = AliasIdxRaw+1; 1087 // Zero out the alias type cache. 1088 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1089 // A NULL adr_type hits in the cache right away. Preload the right answer. 1090 probe_alias_cache(NULL)->_index = AliasIdxTop; 1091 1092 _intrinsics = NULL; 1093 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1094 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1095 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1096 register_library_intrinsics(); 1097 } 1098 1099 //---------------------------init_start---------------------------------------- 1100 // Install the StartNode on this compile object. 1101 void Compile::init_start(StartNode* s) { 1102 if (failing()) 1103 return; // already failing 1104 assert(s == start(), ""); 1105 } 1106 1107 StartNode* Compile::start() const { 1108 assert(!failing(), ""); 1109 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1110 Node* start = root()->fast_out(i); 1111 if( start->is_Start() ) 1112 return start->as_Start(); 1113 } 1114 ShouldNotReachHere(); 1115 return NULL; 1116 } 1117 1118 //-------------------------------immutable_memory------------------------------------- 1119 // Access immutable memory 1120 Node* Compile::immutable_memory() { 1121 if (_immutable_memory != NULL) { 1122 return _immutable_memory; 1123 } 1124 StartNode* s = start(); 1125 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1126 Node *p = s->fast_out(i); 1127 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1128 _immutable_memory = p; 1129 return _immutable_memory; 1130 } 1131 } 1132 ShouldNotReachHere(); 1133 return NULL; 1134 } 1135 1136 //----------------------set_cached_top_node------------------------------------ 1137 // Install the cached top node, and make sure Node::is_top works correctly. 1138 void Compile::set_cached_top_node(Node* tn) { 1139 if (tn != NULL) verify_top(tn); 1140 Node* old_top = _top; 1141 _top = tn; 1142 // Calling Node::setup_is_top allows the nodes the chance to adjust 1143 // their _out arrays. 1144 if (_top != NULL) _top->setup_is_top(); 1145 if (old_top != NULL) old_top->setup_is_top(); 1146 assert(_top == NULL || top()->is_top(), ""); 1147 } 1148 1149 #ifdef ASSERT 1150 uint Compile::count_live_nodes_by_graph_walk() { 1151 Unique_Node_List useful(comp_arena()); 1152 // Get useful node list by walking the graph. 1153 identify_useful_nodes(useful); 1154 return useful.size(); 1155 } 1156 1157 void Compile::print_missing_nodes() { 1158 1159 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1160 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1161 return; 1162 } 1163 1164 // This is an expensive function. It is executed only when the user 1165 // specifies VerifyIdealNodeCount option or otherwise knows the 1166 // additional work that needs to be done to identify reachable nodes 1167 // by walking the flow graph and find the missing ones using 1168 // _dead_node_list. 1169 1170 Unique_Node_List useful(comp_arena()); 1171 // Get useful node list by walking the graph. 1172 identify_useful_nodes(useful); 1173 1174 uint l_nodes = C->live_nodes(); 1175 uint l_nodes_by_walk = useful.size(); 1176 1177 if (l_nodes != l_nodes_by_walk) { 1178 if (_log != NULL) { 1179 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1180 _log->stamp(); 1181 _log->end_head(); 1182 } 1183 VectorSet& useful_member_set = useful.member_set(); 1184 int last_idx = l_nodes_by_walk; 1185 for (int i = 0; i < last_idx; i++) { 1186 if (useful_member_set.test(i)) { 1187 if (_dead_node_list.test(i)) { 1188 if (_log != NULL) { 1189 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1190 } 1191 if (PrintIdealNodeCount) { 1192 // Print the log message to tty 1193 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1194 useful.at(i)->dump(); 1195 } 1196 } 1197 } 1198 else if (! _dead_node_list.test(i)) { 1199 if (_log != NULL) { 1200 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1201 } 1202 if (PrintIdealNodeCount) { 1203 // Print the log message to tty 1204 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1205 } 1206 } 1207 } 1208 if (_log != NULL) { 1209 _log->tail("mismatched_nodes"); 1210 } 1211 } 1212 } 1213 #endif 1214 1215 #ifndef PRODUCT 1216 void Compile::verify_top(Node* tn) const { 1217 if (tn != NULL) { 1218 assert(tn->is_Con(), "top node must be a constant"); 1219 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1220 assert(tn->in(0) != NULL, "must have live top node"); 1221 } 1222 } 1223 #endif 1224 1225 1226 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1227 1228 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1229 guarantee(arr != NULL, ""); 1230 int num_blocks = arr->length(); 1231 if (grow_by < num_blocks) grow_by = num_blocks; 1232 int num_notes = grow_by * _node_notes_block_size; 1233 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1234 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1235 while (num_notes > 0) { 1236 arr->append(notes); 1237 notes += _node_notes_block_size; 1238 num_notes -= _node_notes_block_size; 1239 } 1240 assert(num_notes == 0, "exact multiple, please"); 1241 } 1242 1243 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1244 if (source == NULL || dest == NULL) return false; 1245 1246 if (dest->is_Con()) 1247 return false; // Do not push debug info onto constants. 1248 1249 #ifdef ASSERT 1250 // Leave a bread crumb trail pointing to the original node: 1251 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1252 dest->set_debug_orig(source); 1253 } 1254 #endif 1255 1256 if (node_note_array() == NULL) 1257 return false; // Not collecting any notes now. 1258 1259 // This is a copy onto a pre-existing node, which may already have notes. 1260 // If both nodes have notes, do not overwrite any pre-existing notes. 1261 Node_Notes* source_notes = node_notes_at(source->_idx); 1262 if (source_notes == NULL || source_notes->is_clear()) return false; 1263 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1264 if (dest_notes == NULL || dest_notes->is_clear()) { 1265 return set_node_notes_at(dest->_idx, source_notes); 1266 } 1267 1268 Node_Notes merged_notes = (*source_notes); 1269 // The order of operations here ensures that dest notes will win... 1270 merged_notes.update_from(dest_notes); 1271 return set_node_notes_at(dest->_idx, &merged_notes); 1272 } 1273 1274 1275 //--------------------------allow_range_check_smearing------------------------- 1276 // Gating condition for coalescing similar range checks. 1277 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1278 // single covering check that is at least as strong as any of them. 1279 // If the optimization succeeds, the simplified (strengthened) range check 1280 // will always succeed. If it fails, we will deopt, and then give up 1281 // on the optimization. 1282 bool Compile::allow_range_check_smearing() const { 1283 // If this method has already thrown a range-check, 1284 // assume it was because we already tried range smearing 1285 // and it failed. 1286 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1287 return !already_trapped; 1288 } 1289 1290 1291 //------------------------------flatten_alias_type----------------------------- 1292 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1293 int offset = tj->offset(); 1294 TypePtr::PTR ptr = tj->ptr(); 1295 1296 // Known instance (scalarizable allocation) alias only with itself. 1297 bool is_known_inst = tj->isa_oopptr() != NULL && 1298 tj->is_oopptr()->is_known_instance(); 1299 1300 // Process weird unsafe references. 1301 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1302 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1303 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1304 tj = TypeOopPtr::BOTTOM; 1305 ptr = tj->ptr(); 1306 offset = tj->offset(); 1307 } 1308 1309 // Array pointers need some flattening 1310 const TypeAryPtr *ta = tj->isa_aryptr(); 1311 if (ta && ta->is_stable()) { 1312 // Erase stability property for alias analysis. 1313 tj = ta = ta->cast_to_stable(false); 1314 } 1315 if( ta && is_known_inst ) { 1316 if ( offset != Type::OffsetBot && 1317 offset > arrayOopDesc::length_offset_in_bytes() ) { 1318 offset = Type::OffsetBot; // Flatten constant access into array body only 1319 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1320 } 1321 } else if( ta && _AliasLevel >= 2 ) { 1322 // For arrays indexed by constant indices, we flatten the alias 1323 // space to include all of the array body. Only the header, klass 1324 // and array length can be accessed un-aliased. 1325 if( offset != Type::OffsetBot ) { 1326 if( ta->const_oop() ) { // MethodData* or Method* 1327 offset = Type::OffsetBot; // Flatten constant access into array body 1328 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1329 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1330 // range is OK as-is. 1331 tj = ta = TypeAryPtr::RANGE; 1332 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1333 tj = TypeInstPtr::KLASS; // all klass loads look alike 1334 ta = TypeAryPtr::RANGE; // generic ignored junk 1335 ptr = TypePtr::BotPTR; 1336 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1337 tj = TypeInstPtr::MARK; 1338 ta = TypeAryPtr::RANGE; // generic ignored junk 1339 ptr = TypePtr::BotPTR; 1340 } else { // Random constant offset into array body 1341 offset = Type::OffsetBot; // Flatten constant access into array body 1342 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1343 } 1344 } 1345 // Arrays of fixed size alias with arrays of unknown size. 1346 if (ta->size() != TypeInt::POS) { 1347 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1348 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1349 } 1350 // Arrays of known objects become arrays of unknown objects. 1351 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1352 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1353 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1354 } 1355 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1356 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1357 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1358 } 1359 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1360 // cannot be distinguished by bytecode alone. 1361 if (ta->elem() == TypeInt::BOOL) { 1362 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1363 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1364 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1365 } 1366 // During the 2nd round of IterGVN, NotNull castings are removed. 1367 // Make sure the Bottom and NotNull variants alias the same. 1368 // Also, make sure exact and non-exact variants alias the same. 1369 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1370 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1371 } 1372 } 1373 1374 // Oop pointers need some flattening 1375 const TypeInstPtr *to = tj->isa_instptr(); 1376 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1377 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1378 if( ptr == TypePtr::Constant ) { 1379 if (to->klass() != ciEnv::current()->Class_klass() || 1380 offset < k->size_helper() * wordSize) { 1381 // No constant oop pointers (such as Strings); they alias with 1382 // unknown strings. 1383 assert(!is_known_inst, "not scalarizable allocation"); 1384 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1385 } 1386 } else if( is_known_inst ) { 1387 tj = to; // Keep NotNull and klass_is_exact for instance type 1388 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1389 // During the 2nd round of IterGVN, NotNull castings are removed. 1390 // Make sure the Bottom and NotNull variants alias the same. 1391 // Also, make sure exact and non-exact variants alias the same. 1392 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1393 } 1394 if (to->speculative() != NULL) { 1395 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1396 } 1397 // Canonicalize the holder of this field 1398 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1399 // First handle header references such as a LoadKlassNode, even if the 1400 // object's klass is unloaded at compile time (4965979). 1401 if (!is_known_inst) { // Do it only for non-instance types 1402 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1403 } 1404 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1405 // Static fields are in the space above the normal instance 1406 // fields in the java.lang.Class instance. 1407 if (to->klass() != ciEnv::current()->Class_klass()) { 1408 to = NULL; 1409 tj = TypeOopPtr::BOTTOM; 1410 offset = tj->offset(); 1411 } 1412 } else { 1413 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1414 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1415 if( is_known_inst ) { 1416 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1417 } else { 1418 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1419 } 1420 } 1421 } 1422 } 1423 1424 // Klass pointers to object array klasses need some flattening 1425 const TypeKlassPtr *tk = tj->isa_klassptr(); 1426 if( tk ) { 1427 // If we are referencing a field within a Klass, we need 1428 // to assume the worst case of an Object. Both exact and 1429 // inexact types must flatten to the same alias class so 1430 // use NotNull as the PTR. 1431 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1432 1433 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1434 TypeKlassPtr::OBJECT->klass(), 1435 offset); 1436 } 1437 1438 ciKlass* klass = tk->klass(); 1439 if( klass->is_obj_array_klass() ) { 1440 ciKlass* k = TypeAryPtr::OOPS->klass(); 1441 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1442 k = TypeInstPtr::BOTTOM->klass(); 1443 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1444 } 1445 1446 // Check for precise loads from the primary supertype array and force them 1447 // to the supertype cache alias index. Check for generic array loads from 1448 // the primary supertype array and also force them to the supertype cache 1449 // alias index. Since the same load can reach both, we need to merge 1450 // these 2 disparate memories into the same alias class. Since the 1451 // primary supertype array is read-only, there's no chance of confusion 1452 // where we bypass an array load and an array store. 1453 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1454 if (offset == Type::OffsetBot || 1455 (offset >= primary_supers_offset && 1456 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1457 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1458 offset = in_bytes(Klass::secondary_super_cache_offset()); 1459 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1460 } 1461 } 1462 1463 // Flatten all Raw pointers together. 1464 if (tj->base() == Type::RawPtr) 1465 tj = TypeRawPtr::BOTTOM; 1466 1467 if (tj->base() == Type::AnyPtr) 1468 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1469 1470 // Flatten all to bottom for now 1471 switch( _AliasLevel ) { 1472 case 0: 1473 tj = TypePtr::BOTTOM; 1474 break; 1475 case 1: // Flatten to: oop, static, field or array 1476 switch (tj->base()) { 1477 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1478 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1479 case Type::AryPtr: // do not distinguish arrays at all 1480 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1481 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1482 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1483 default: ShouldNotReachHere(); 1484 } 1485 break; 1486 case 2: // No collapsing at level 2; keep all splits 1487 case 3: // No collapsing at level 3; keep all splits 1488 break; 1489 default: 1490 Unimplemented(); 1491 } 1492 1493 offset = tj->offset(); 1494 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1495 1496 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1497 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1498 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1499 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1500 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1501 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1502 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1503 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1504 assert( tj->ptr() != TypePtr::TopPTR && 1505 tj->ptr() != TypePtr::AnyNull && 1506 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1507 // assert( tj->ptr() != TypePtr::Constant || 1508 // tj->base() == Type::RawPtr || 1509 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1510 1511 return tj; 1512 } 1513 1514 void Compile::AliasType::Init(int i, const TypePtr* at) { 1515 _index = i; 1516 _adr_type = at; 1517 _field = NULL; 1518 _element = NULL; 1519 _is_rewritable = true; // default 1520 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1521 if (atoop != NULL && atoop->is_known_instance()) { 1522 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1523 _general_index = Compile::current()->get_alias_index(gt); 1524 } else { 1525 _general_index = 0; 1526 } 1527 } 1528 1529 //---------------------------------print_on------------------------------------ 1530 #ifndef PRODUCT 1531 void Compile::AliasType::print_on(outputStream* st) { 1532 if (index() < 10) 1533 st->print("@ <%d> ", index()); 1534 else st->print("@ <%d>", index()); 1535 st->print(is_rewritable() ? " " : " RO"); 1536 int offset = adr_type()->offset(); 1537 if (offset == Type::OffsetBot) 1538 st->print(" +any"); 1539 else st->print(" +%-3d", offset); 1540 st->print(" in "); 1541 adr_type()->dump_on(st); 1542 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1543 if (field() != NULL && tjp) { 1544 if (tjp->klass() != field()->holder() || 1545 tjp->offset() != field()->offset_in_bytes()) { 1546 st->print(" != "); 1547 field()->print(); 1548 st->print(" ***"); 1549 } 1550 } 1551 } 1552 1553 void print_alias_types() { 1554 Compile* C = Compile::current(); 1555 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1556 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1557 C->alias_type(idx)->print_on(tty); 1558 tty->cr(); 1559 } 1560 } 1561 #endif 1562 1563 1564 //----------------------------probe_alias_cache-------------------------------- 1565 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1566 intptr_t key = (intptr_t) adr_type; 1567 key ^= key >> logAliasCacheSize; 1568 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1569 } 1570 1571 1572 //-----------------------------grow_alias_types-------------------------------- 1573 void Compile::grow_alias_types() { 1574 const int old_ats = _max_alias_types; // how many before? 1575 const int new_ats = old_ats; // how many more? 1576 const int grow_ats = old_ats+new_ats; // how many now? 1577 _max_alias_types = grow_ats; 1578 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1579 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1580 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1581 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1582 } 1583 1584 1585 //--------------------------------find_alias_type------------------------------ 1586 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1587 if (_AliasLevel == 0) 1588 return alias_type(AliasIdxBot); 1589 1590 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1591 if (ace->_adr_type == adr_type) { 1592 return alias_type(ace->_index); 1593 } 1594 1595 // Handle special cases. 1596 if (adr_type == NULL) return alias_type(AliasIdxTop); 1597 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1598 1599 // Do it the slow way. 1600 const TypePtr* flat = flatten_alias_type(adr_type); 1601 1602 #ifdef ASSERT 1603 assert(flat == flatten_alias_type(flat), "idempotent"); 1604 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1605 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1606 const TypeOopPtr* foop = flat->is_oopptr(); 1607 // Scalarizable allocations have exact klass always. 1608 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1609 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1610 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1611 } 1612 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1613 #endif 1614 1615 int idx = AliasIdxTop; 1616 for (int i = 0; i < num_alias_types(); i++) { 1617 if (alias_type(i)->adr_type() == flat) { 1618 idx = i; 1619 break; 1620 } 1621 } 1622 1623 if (idx == AliasIdxTop) { 1624 if (no_create) return NULL; 1625 // Grow the array if necessary. 1626 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1627 // Add a new alias type. 1628 idx = _num_alias_types++; 1629 _alias_types[idx]->Init(idx, flat); 1630 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1631 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1632 if (flat->isa_instptr()) { 1633 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1634 && flat->is_instptr()->klass() == env()->Class_klass()) 1635 alias_type(idx)->set_rewritable(false); 1636 } 1637 if (flat->isa_aryptr()) { 1638 #ifdef ASSERT 1639 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1640 // (T_BYTE has the weakest alignment and size restrictions...) 1641 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1642 #endif 1643 if (flat->offset() == TypePtr::OffsetBot) { 1644 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1645 } 1646 } 1647 if (flat->isa_klassptr()) { 1648 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1649 alias_type(idx)->set_rewritable(false); 1650 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1651 alias_type(idx)->set_rewritable(false); 1652 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1653 alias_type(idx)->set_rewritable(false); 1654 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1655 alias_type(idx)->set_rewritable(false); 1656 } 1657 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1658 // but the base pointer type is not distinctive enough to identify 1659 // references into JavaThread.) 1660 1661 // Check for final fields. 1662 const TypeInstPtr* tinst = flat->isa_instptr(); 1663 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1664 ciField* field; 1665 if (tinst->const_oop() != NULL && 1666 tinst->klass() == ciEnv::current()->Class_klass() && 1667 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1668 // static field 1669 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1670 field = k->get_field_by_offset(tinst->offset(), true); 1671 } else { 1672 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1673 field = k->get_field_by_offset(tinst->offset(), false); 1674 } 1675 assert(field == NULL || 1676 original_field == NULL || 1677 (field->holder() == original_field->holder() && 1678 field->offset() == original_field->offset() && 1679 field->is_static() == original_field->is_static()), "wrong field?"); 1680 // Set field() and is_rewritable() attributes. 1681 if (field != NULL) alias_type(idx)->set_field(field); 1682 } 1683 } 1684 1685 // Fill the cache for next time. 1686 ace->_adr_type = adr_type; 1687 ace->_index = idx; 1688 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1689 1690 // Might as well try to fill the cache for the flattened version, too. 1691 AliasCacheEntry* face = probe_alias_cache(flat); 1692 if (face->_adr_type == NULL) { 1693 face->_adr_type = flat; 1694 face->_index = idx; 1695 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1696 } 1697 1698 return alias_type(idx); 1699 } 1700 1701 1702 Compile::AliasType* Compile::alias_type(ciField* field) { 1703 const TypeOopPtr* t; 1704 if (field->is_static()) 1705 t = TypeInstPtr::make(field->holder()->java_mirror()); 1706 else 1707 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1708 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1709 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1710 return atp; 1711 } 1712 1713 1714 //------------------------------have_alias_type-------------------------------- 1715 bool Compile::have_alias_type(const TypePtr* adr_type) { 1716 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1717 if (ace->_adr_type == adr_type) { 1718 return true; 1719 } 1720 1721 // Handle special cases. 1722 if (adr_type == NULL) return true; 1723 if (adr_type == TypePtr::BOTTOM) return true; 1724 1725 return find_alias_type(adr_type, true, NULL) != NULL; 1726 } 1727 1728 //-----------------------------must_alias-------------------------------------- 1729 // True if all values of the given address type are in the given alias category. 1730 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1731 if (alias_idx == AliasIdxBot) return true; // the universal category 1732 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1733 if (alias_idx == AliasIdxTop) return false; // the empty category 1734 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1735 1736 // the only remaining possible overlap is identity 1737 int adr_idx = get_alias_index(adr_type); 1738 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1739 assert(adr_idx == alias_idx || 1740 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1741 && adr_type != TypeOopPtr::BOTTOM), 1742 "should not be testing for overlap with an unsafe pointer"); 1743 return adr_idx == alias_idx; 1744 } 1745 1746 //------------------------------can_alias-------------------------------------- 1747 // True if any values of the given address type are in the given alias category. 1748 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1749 if (alias_idx == AliasIdxTop) return false; // the empty category 1750 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1751 if (alias_idx == AliasIdxBot) return true; // the universal category 1752 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1753 1754 // the only remaining possible overlap is identity 1755 int adr_idx = get_alias_index(adr_type); 1756 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1757 return adr_idx == alias_idx; 1758 } 1759 1760 1761 1762 //---------------------------pop_warm_call------------------------------------- 1763 WarmCallInfo* Compile::pop_warm_call() { 1764 WarmCallInfo* wci = _warm_calls; 1765 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1766 return wci; 1767 } 1768 1769 //----------------------------Inline_Warm-------------------------------------- 1770 int Compile::Inline_Warm() { 1771 // If there is room, try to inline some more warm call sites. 1772 // %%% Do a graph index compaction pass when we think we're out of space? 1773 if (!InlineWarmCalls) return 0; 1774 1775 int calls_made_hot = 0; 1776 int room_to_grow = NodeCountInliningCutoff - unique(); 1777 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1778 int amount_grown = 0; 1779 WarmCallInfo* call; 1780 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1781 int est_size = (int)call->size(); 1782 if (est_size > (room_to_grow - amount_grown)) { 1783 // This one won't fit anyway. Get rid of it. 1784 call->make_cold(); 1785 continue; 1786 } 1787 call->make_hot(); 1788 calls_made_hot++; 1789 amount_grown += est_size; 1790 amount_to_grow -= est_size; 1791 } 1792 1793 if (calls_made_hot > 0) set_major_progress(); 1794 return calls_made_hot; 1795 } 1796 1797 1798 //----------------------------Finish_Warm-------------------------------------- 1799 void Compile::Finish_Warm() { 1800 if (!InlineWarmCalls) return; 1801 if (failing()) return; 1802 if (warm_calls() == NULL) return; 1803 1804 // Clean up loose ends, if we are out of space for inlining. 1805 WarmCallInfo* call; 1806 while ((call = pop_warm_call()) != NULL) { 1807 call->make_cold(); 1808 } 1809 } 1810 1811 //---------------------cleanup_loop_predicates----------------------- 1812 // Remove the opaque nodes that protect the predicates so that all unused 1813 // checks and uncommon_traps will be eliminated from the ideal graph 1814 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1815 if (predicate_count()==0) return; 1816 for (int i = predicate_count(); i > 0; i--) { 1817 Node * n = predicate_opaque1_node(i-1); 1818 assert(n->Opcode() == Op_Opaque1, "must be"); 1819 igvn.replace_node(n, n->in(1)); 1820 } 1821 assert(predicate_count()==0, "should be clean!"); 1822 } 1823 1824 // StringOpts and late inlining of string methods 1825 void Compile::inline_string_calls(bool parse_time) { 1826 { 1827 // remove useless nodes to make the usage analysis simpler 1828 ResourceMark rm; 1829 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1830 } 1831 1832 { 1833 ResourceMark rm; 1834 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1835 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1836 print_method(PHASE_AFTER_STRINGOPTS, 3); 1837 } 1838 1839 // now inline anything that we skipped the first time around 1840 if (!parse_time) { 1841 _late_inlines_pos = _late_inlines.length(); 1842 } 1843 1844 while (_string_late_inlines.length() > 0) { 1845 CallGenerator* cg = _string_late_inlines.pop(); 1846 cg->do_late_inline(); 1847 if (failing()) return; 1848 } 1849 _string_late_inlines.trunc_to(0); 1850 } 1851 1852 // Late inlining of boxing methods 1853 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1854 if (_boxing_late_inlines.length() > 0) { 1855 assert(has_boxed_value(), "inconsistent"); 1856 1857 PhaseGVN* gvn = initial_gvn(); 1858 set_inlining_incrementally(true); 1859 1860 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1861 for_igvn()->clear(); 1862 gvn->replace_with(&igvn); 1863 1864 while (_boxing_late_inlines.length() > 0) { 1865 CallGenerator* cg = _boxing_late_inlines.pop(); 1866 cg->do_late_inline(); 1867 if (failing()) return; 1868 } 1869 _boxing_late_inlines.trunc_to(0); 1870 1871 { 1872 ResourceMark rm; 1873 PhaseRemoveUseless pru(gvn, for_igvn()); 1874 } 1875 1876 igvn = PhaseIterGVN(gvn); 1877 igvn.optimize(); 1878 1879 set_inlining_progress(false); 1880 set_inlining_incrementally(false); 1881 } 1882 } 1883 1884 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1885 assert(IncrementalInline, "incremental inlining should be on"); 1886 PhaseGVN* gvn = initial_gvn(); 1887 1888 set_inlining_progress(false); 1889 for_igvn()->clear(); 1890 gvn->replace_with(&igvn); 1891 1892 int i = 0; 1893 1894 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1895 CallGenerator* cg = _late_inlines.at(i); 1896 _late_inlines_pos = i+1; 1897 cg->do_late_inline(); 1898 if (failing()) return; 1899 } 1900 int j = 0; 1901 for (; i < _late_inlines.length(); i++, j++) { 1902 _late_inlines.at_put(j, _late_inlines.at(i)); 1903 } 1904 _late_inlines.trunc_to(j); 1905 1906 { 1907 ResourceMark rm; 1908 PhaseRemoveUseless pru(gvn, for_igvn()); 1909 } 1910 1911 igvn = PhaseIterGVN(gvn); 1912 } 1913 1914 // Perform incremental inlining until bound on number of live nodes is reached 1915 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1916 PhaseGVN* gvn = initial_gvn(); 1917 1918 set_inlining_incrementally(true); 1919 set_inlining_progress(true); 1920 uint low_live_nodes = 0; 1921 1922 while(inlining_progress() && _late_inlines.length() > 0) { 1923 1924 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1925 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1926 // PhaseIdealLoop is expensive so we only try it once we are 1927 // out of loop and we only try it again if the previous helped 1928 // got the number of nodes down significantly 1929 PhaseIdealLoop ideal_loop( igvn, false, true ); 1930 if (failing()) return; 1931 low_live_nodes = live_nodes(); 1932 _major_progress = true; 1933 } 1934 1935 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1936 break; 1937 } 1938 } 1939 1940 inline_incrementally_one(igvn); 1941 1942 if (failing()) return; 1943 1944 igvn.optimize(); 1945 1946 if (failing()) return; 1947 } 1948 1949 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1950 1951 if (_string_late_inlines.length() > 0) { 1952 assert(has_stringbuilder(), "inconsistent"); 1953 for_igvn()->clear(); 1954 initial_gvn()->replace_with(&igvn); 1955 1956 inline_string_calls(false); 1957 1958 if (failing()) return; 1959 1960 { 1961 ResourceMark rm; 1962 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1963 } 1964 1965 igvn = PhaseIterGVN(gvn); 1966 1967 igvn.optimize(); 1968 } 1969 1970 set_inlining_incrementally(false); 1971 } 1972 1973 1974 //------------------------------Optimize--------------------------------------- 1975 // Given a graph, optimize it. 1976 void Compile::Optimize() { 1977 TracePhase t1("optimizer", &_t_optimizer, true); 1978 1979 #ifndef PRODUCT 1980 if (env()->break_at_compile()) { 1981 BREAKPOINT; 1982 } 1983 1984 #endif 1985 1986 ResourceMark rm; 1987 int loop_opts_cnt; 1988 1989 NOT_PRODUCT( verify_graph_edges(); ) 1990 1991 print_method(PHASE_AFTER_PARSING); 1992 1993 { 1994 // Iterative Global Value Numbering, including ideal transforms 1995 // Initialize IterGVN with types and values from parse-time GVN 1996 PhaseIterGVN igvn(initial_gvn()); 1997 { 1998 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1999 igvn.optimize(); 2000 } 2001 2002 print_method(PHASE_ITER_GVN1, 2); 2003 2004 if (failing()) return; 2005 2006 { 2007 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2008 inline_incrementally(igvn); 2009 } 2010 2011 print_method(PHASE_INCREMENTAL_INLINE, 2); 2012 2013 if (failing()) return; 2014 2015 if (eliminate_boxing()) { 2016 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2017 // Inline valueOf() methods now. 2018 inline_boxing_calls(igvn); 2019 2020 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2021 2022 if (failing()) return; 2023 } 2024 2025 // Remove the speculative part of types and clean up the graph from 2026 // the extra CastPP nodes whose only purpose is to carry them. Do 2027 // that early so that optimizations are not disrupted by the extra 2028 // CastPP nodes. 2029 remove_speculative_types(igvn); 2030 2031 // No more new expensive nodes will be added to the list from here 2032 // so keep only the actual candidates for optimizations. 2033 cleanup_expensive_nodes(igvn); 2034 2035 // Perform escape analysis 2036 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2037 if (has_loops()) { 2038 // Cleanup graph (remove dead nodes). 2039 TracePhase t2("idealLoop", &_t_idealLoop, true); 2040 PhaseIdealLoop ideal_loop( igvn, false, true ); 2041 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2042 if (failing()) return; 2043 } 2044 ConnectionGraph::do_analysis(this, &igvn); 2045 2046 if (failing()) return; 2047 2048 // Optimize out fields loads from scalar replaceable allocations. 2049 igvn.optimize(); 2050 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2051 2052 if (failing()) return; 2053 2054 if (congraph() != NULL && macro_count() > 0) { 2055 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2056 PhaseMacroExpand mexp(igvn); 2057 mexp.eliminate_macro_nodes(); 2058 igvn.set_delay_transform(false); 2059 2060 igvn.optimize(); 2061 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2062 2063 if (failing()) return; 2064 } 2065 } 2066 2067 // Loop transforms on the ideal graph. Range Check Elimination, 2068 // peeling, unrolling, etc. 2069 2070 // Set loop opts counter 2071 loop_opts_cnt = num_loop_opts(); 2072 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2073 { 2074 TracePhase t2("idealLoop", &_t_idealLoop, true); 2075 PhaseIdealLoop ideal_loop( igvn, true ); 2076 loop_opts_cnt--; 2077 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2078 if (failing()) return; 2079 } 2080 // Loop opts pass if partial peeling occurred in previous pass 2081 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2082 TracePhase t3("idealLoop", &_t_idealLoop, true); 2083 PhaseIdealLoop ideal_loop( igvn, false ); 2084 loop_opts_cnt--; 2085 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2086 if (failing()) return; 2087 } 2088 // Loop opts pass for loop-unrolling before CCP 2089 if(major_progress() && (loop_opts_cnt > 0)) { 2090 TracePhase t4("idealLoop", &_t_idealLoop, true); 2091 PhaseIdealLoop ideal_loop( igvn, false ); 2092 loop_opts_cnt--; 2093 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2094 } 2095 if (!failing()) { 2096 // Verify that last round of loop opts produced a valid graph 2097 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2098 PhaseIdealLoop::verify(igvn); 2099 } 2100 } 2101 if (failing()) return; 2102 2103 // Conditional Constant Propagation; 2104 PhaseCCP ccp( &igvn ); 2105 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2106 { 2107 TracePhase t2("ccp", &_t_ccp, true); 2108 ccp.do_transform(); 2109 } 2110 print_method(PHASE_CPP1, 2); 2111 2112 assert( true, "Break here to ccp.dump_old2new_map()"); 2113 2114 // Iterative Global Value Numbering, including ideal transforms 2115 { 2116 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2117 igvn = ccp; 2118 igvn.optimize(); 2119 } 2120 2121 print_method(PHASE_ITER_GVN2, 2); 2122 2123 if (failing()) return; 2124 2125 // Loop transforms on the ideal graph. Range Check Elimination, 2126 // peeling, unrolling, etc. 2127 if(loop_opts_cnt > 0) { 2128 debug_only( int cnt = 0; ); 2129 while(major_progress() && (loop_opts_cnt > 0)) { 2130 TracePhase t2("idealLoop", &_t_idealLoop, true); 2131 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2132 PhaseIdealLoop ideal_loop( igvn, true); 2133 loop_opts_cnt--; 2134 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2135 if (failing()) return; 2136 } 2137 } 2138 2139 { 2140 // Verify that all previous optimizations produced a valid graph 2141 // at least to this point, even if no loop optimizations were done. 2142 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2143 PhaseIdealLoop::verify(igvn); 2144 } 2145 2146 { 2147 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2148 PhaseMacroExpand mex(igvn); 2149 if (mex.expand_macro_nodes()) { 2150 assert(failing(), "must bail out w/ explicit message"); 2151 return; 2152 } 2153 } 2154 2155 } // (End scope of igvn; run destructor if necessary for asserts.) 2156 2157 dump_inlining(); 2158 // A method with only infinite loops has no edges entering loops from root 2159 { 2160 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2161 if (final_graph_reshaping()) { 2162 assert(failing(), "must bail out w/ explicit message"); 2163 return; 2164 } 2165 } 2166 2167 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2168 } 2169 2170 2171 //------------------------------Code_Gen--------------------------------------- 2172 // Given a graph, generate code for it 2173 void Compile::Code_Gen() { 2174 if (failing()) { 2175 return; 2176 } 2177 2178 // Perform instruction selection. You might think we could reclaim Matcher 2179 // memory PDQ, but actually the Matcher is used in generating spill code. 2180 // Internals of the Matcher (including some VectorSets) must remain live 2181 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2182 // set a bit in reclaimed memory. 2183 2184 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2185 // nodes. Mapping is only valid at the root of each matched subtree. 2186 NOT_PRODUCT( verify_graph_edges(); ) 2187 2188 Matcher matcher; 2189 _matcher = &matcher; 2190 { 2191 TracePhase t2("matcher", &_t_matcher, true); 2192 matcher.match(); 2193 } 2194 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2195 // nodes. Mapping is only valid at the root of each matched subtree. 2196 NOT_PRODUCT( verify_graph_edges(); ) 2197 2198 // If you have too many nodes, or if matching has failed, bail out 2199 check_node_count(0, "out of nodes matching instructions"); 2200 if (failing()) { 2201 return; 2202 } 2203 2204 // Build a proper-looking CFG 2205 PhaseCFG cfg(node_arena(), root(), matcher); 2206 _cfg = &cfg; 2207 { 2208 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2209 bool success = cfg.do_global_code_motion(); 2210 if (!success) { 2211 return; 2212 } 2213 2214 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2215 NOT_PRODUCT( verify_graph_edges(); ) 2216 debug_only( cfg.verify(); ) 2217 } 2218 2219 PhaseChaitin regalloc(unique(), cfg, matcher); 2220 _regalloc = ®alloc; 2221 { 2222 TracePhase t2("regalloc", &_t_registerAllocation, true); 2223 // Perform register allocation. After Chaitin, use-def chains are 2224 // no longer accurate (at spill code) and so must be ignored. 2225 // Node->LRG->reg mappings are still accurate. 2226 _regalloc->Register_Allocate(); 2227 2228 // Bail out if the allocator builds too many nodes 2229 if (failing()) { 2230 return; 2231 } 2232 } 2233 2234 // Prior to register allocation we kept empty basic blocks in case the 2235 // the allocator needed a place to spill. After register allocation we 2236 // are not adding any new instructions. If any basic block is empty, we 2237 // can now safely remove it. 2238 { 2239 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2240 cfg.remove_empty_blocks(); 2241 if (do_freq_based_layout()) { 2242 PhaseBlockLayout layout(cfg); 2243 } else { 2244 cfg.set_loop_alignment(); 2245 } 2246 cfg.fixup_flow(); 2247 } 2248 2249 // Apply peephole optimizations 2250 if( OptoPeephole ) { 2251 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2252 PhasePeephole peep( _regalloc, cfg); 2253 peep.do_transform(); 2254 } 2255 2256 // Do late expand if CPU requires this. 2257 if (Matcher::require_postalloc_expand) { 2258 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2259 cfg.postalloc_expand(_regalloc); 2260 } 2261 2262 // Convert Nodes to instruction bits in a buffer 2263 { 2264 // %%%% workspace merge brought two timers together for one job 2265 TracePhase t2a("output", &_t_output, true); 2266 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2267 Output(); 2268 } 2269 2270 print_method(PHASE_FINAL_CODE); 2271 2272 // He's dead, Jim. 2273 _cfg = (PhaseCFG*)0xdeadbeef; 2274 _regalloc = (PhaseChaitin*)0xdeadbeef; 2275 } 2276 2277 2278 //------------------------------dump_asm--------------------------------------- 2279 // Dump formatted assembly 2280 #ifndef PRODUCT 2281 void Compile::dump_asm(int *pcs, uint pc_limit) { 2282 bool cut_short = false; 2283 tty->print_cr("#"); 2284 tty->print("# "); _tf->dump(); tty->cr(); 2285 tty->print_cr("#"); 2286 2287 // For all blocks 2288 int pc = 0x0; // Program counter 2289 char starts_bundle = ' '; 2290 _regalloc->dump_frame(); 2291 2292 Node *n = NULL; 2293 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2294 if (VMThread::should_terminate()) { 2295 cut_short = true; 2296 break; 2297 } 2298 Block* block = _cfg->get_block(i); 2299 if (block->is_connector() && !Verbose) { 2300 continue; 2301 } 2302 n = block->head(); 2303 if (pcs && n->_idx < pc_limit) { 2304 tty->print("%3.3x ", pcs[n->_idx]); 2305 } else { 2306 tty->print(" "); 2307 } 2308 block->dump_head(_cfg); 2309 if (block->is_connector()) { 2310 tty->print_cr(" # Empty connector block"); 2311 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2312 tty->print_cr(" # Block is sole successor of call"); 2313 } 2314 2315 // For all instructions 2316 Node *delay = NULL; 2317 for (uint j = 0; j < block->number_of_nodes(); j++) { 2318 if (VMThread::should_terminate()) { 2319 cut_short = true; 2320 break; 2321 } 2322 n = block->get_node(j); 2323 if (valid_bundle_info(n)) { 2324 Bundle* bundle = node_bundling(n); 2325 if (bundle->used_in_unconditional_delay()) { 2326 delay = n; 2327 continue; 2328 } 2329 if (bundle->starts_bundle()) { 2330 starts_bundle = '+'; 2331 } 2332 } 2333 2334 if (WizardMode) { 2335 n->dump(); 2336 } 2337 2338 if( !n->is_Region() && // Dont print in the Assembly 2339 !n->is_Phi() && // a few noisely useless nodes 2340 !n->is_Proj() && 2341 !n->is_MachTemp() && 2342 !n->is_SafePointScalarObject() && 2343 !n->is_Catch() && // Would be nice to print exception table targets 2344 !n->is_MergeMem() && // Not very interesting 2345 !n->is_top() && // Debug info table constants 2346 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2347 ) { 2348 if (pcs && n->_idx < pc_limit) 2349 tty->print("%3.3x", pcs[n->_idx]); 2350 else 2351 tty->print(" "); 2352 tty->print(" %c ", starts_bundle); 2353 starts_bundle = ' '; 2354 tty->print("\t"); 2355 n->format(_regalloc, tty); 2356 tty->cr(); 2357 } 2358 2359 // If we have an instruction with a delay slot, and have seen a delay, 2360 // then back up and print it 2361 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2362 assert(delay != NULL, "no unconditional delay instruction"); 2363 if (WizardMode) delay->dump(); 2364 2365 if (node_bundling(delay)->starts_bundle()) 2366 starts_bundle = '+'; 2367 if (pcs && n->_idx < pc_limit) 2368 tty->print("%3.3x", pcs[n->_idx]); 2369 else 2370 tty->print(" "); 2371 tty->print(" %c ", starts_bundle); 2372 starts_bundle = ' '; 2373 tty->print("\t"); 2374 delay->format(_regalloc, tty); 2375 tty->print_cr(""); 2376 delay = NULL; 2377 } 2378 2379 // Dump the exception table as well 2380 if( n->is_Catch() && (Verbose || WizardMode) ) { 2381 // Print the exception table for this offset 2382 _handler_table.print_subtable_for(pc); 2383 } 2384 } 2385 2386 if (pcs && n->_idx < pc_limit) 2387 tty->print_cr("%3.3x", pcs[n->_idx]); 2388 else 2389 tty->print_cr(""); 2390 2391 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2392 2393 } // End of per-block dump 2394 tty->print_cr(""); 2395 2396 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2397 } 2398 #endif 2399 2400 //------------------------------Final_Reshape_Counts--------------------------- 2401 // This class defines counters to help identify when a method 2402 // may/must be executed using hardware with only 24-bit precision. 2403 struct Final_Reshape_Counts : public StackObj { 2404 int _call_count; // count non-inlined 'common' calls 2405 int _float_count; // count float ops requiring 24-bit precision 2406 int _double_count; // count double ops requiring more precision 2407 int _java_call_count; // count non-inlined 'java' calls 2408 int _inner_loop_count; // count loops which need alignment 2409 VectorSet _visited; // Visitation flags 2410 Node_List _tests; // Set of IfNodes & PCTableNodes 2411 2412 Final_Reshape_Counts() : 2413 _call_count(0), _float_count(0), _double_count(0), 2414 _java_call_count(0), _inner_loop_count(0), 2415 _visited( Thread::current()->resource_area() ) { } 2416 2417 void inc_call_count () { _call_count ++; } 2418 void inc_float_count () { _float_count ++; } 2419 void inc_double_count() { _double_count++; } 2420 void inc_java_call_count() { _java_call_count++; } 2421 void inc_inner_loop_count() { _inner_loop_count++; } 2422 2423 int get_call_count () const { return _call_count ; } 2424 int get_float_count () const { return _float_count ; } 2425 int get_double_count() const { return _double_count; } 2426 int get_java_call_count() const { return _java_call_count; } 2427 int get_inner_loop_count() const { return _inner_loop_count; } 2428 }; 2429 2430 #ifdef ASSERT 2431 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2432 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2433 // Make sure the offset goes inside the instance layout. 2434 return k->contains_field_offset(tp->offset()); 2435 // Note that OffsetBot and OffsetTop are very negative. 2436 } 2437 #endif 2438 2439 // Eliminate trivially redundant StoreCMs and accumulate their 2440 // precedence edges. 2441 void Compile::eliminate_redundant_card_marks(Node* n) { 2442 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2443 if (n->in(MemNode::Address)->outcnt() > 1) { 2444 // There are multiple users of the same address so it might be 2445 // possible to eliminate some of the StoreCMs 2446 Node* mem = n->in(MemNode::Memory); 2447 Node* adr = n->in(MemNode::Address); 2448 Node* val = n->in(MemNode::ValueIn); 2449 Node* prev = n; 2450 bool done = false; 2451 // Walk the chain of StoreCMs eliminating ones that match. As 2452 // long as it's a chain of single users then the optimization is 2453 // safe. Eliminating partially redundant StoreCMs would require 2454 // cloning copies down the other paths. 2455 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2456 if (adr == mem->in(MemNode::Address) && 2457 val == mem->in(MemNode::ValueIn)) { 2458 // redundant StoreCM 2459 if (mem->req() > MemNode::OopStore) { 2460 // Hasn't been processed by this code yet. 2461 n->add_prec(mem->in(MemNode::OopStore)); 2462 } else { 2463 // Already converted to precedence edge 2464 for (uint i = mem->req(); i < mem->len(); i++) { 2465 // Accumulate any precedence edges 2466 if (mem->in(i) != NULL) { 2467 n->add_prec(mem->in(i)); 2468 } 2469 } 2470 // Everything above this point has been processed. 2471 done = true; 2472 } 2473 // Eliminate the previous StoreCM 2474 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2475 assert(mem->outcnt() == 0, "should be dead"); 2476 mem->disconnect_inputs(NULL, this); 2477 } else { 2478 prev = mem; 2479 } 2480 mem = prev->in(MemNode::Memory); 2481 } 2482 } 2483 } 2484 2485 //------------------------------final_graph_reshaping_impl---------------------- 2486 // Implement items 1-5 from final_graph_reshaping below. 2487 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2488 2489 if ( n->outcnt() == 0 ) return; // dead node 2490 uint nop = n->Opcode(); 2491 2492 // Check for 2-input instruction with "last use" on right input. 2493 // Swap to left input. Implements item (2). 2494 if( n->req() == 3 && // two-input instruction 2495 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2496 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2497 n->in(2)->outcnt() == 1 &&// right use IS a last use 2498 !n->in(2)->is_Con() ) { // right use is not a constant 2499 // Check for commutative opcode 2500 switch( nop ) { 2501 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2502 case Op_MaxI: case Op_MinI: 2503 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2504 case Op_AndL: case Op_XorL: case Op_OrL: 2505 case Op_AndI: case Op_XorI: case Op_OrI: { 2506 // Move "last use" input to left by swapping inputs 2507 n->swap_edges(1, 2); 2508 break; 2509 } 2510 default: 2511 break; 2512 } 2513 } 2514 2515 #ifdef ASSERT 2516 if( n->is_Mem() ) { 2517 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2518 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2519 // oop will be recorded in oop map if load crosses safepoint 2520 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2521 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2522 "raw memory operations should have control edge"); 2523 } 2524 #endif 2525 // Count FPU ops and common calls, implements item (3) 2526 switch( nop ) { 2527 // Count all float operations that may use FPU 2528 case Op_AddF: 2529 case Op_SubF: 2530 case Op_MulF: 2531 case Op_DivF: 2532 case Op_NegF: 2533 case Op_ModF: 2534 case Op_ConvI2F: 2535 case Op_ConF: 2536 case Op_CmpF: 2537 case Op_CmpF3: 2538 // case Op_ConvL2F: // longs are split into 32-bit halves 2539 frc.inc_float_count(); 2540 break; 2541 2542 case Op_ConvF2D: 2543 case Op_ConvD2F: 2544 frc.inc_float_count(); 2545 frc.inc_double_count(); 2546 break; 2547 2548 // Count all double operations that may use FPU 2549 case Op_AddD: 2550 case Op_SubD: 2551 case Op_MulD: 2552 case Op_DivD: 2553 case Op_NegD: 2554 case Op_ModD: 2555 case Op_ConvI2D: 2556 case Op_ConvD2I: 2557 // case Op_ConvL2D: // handled by leaf call 2558 // case Op_ConvD2L: // handled by leaf call 2559 case Op_ConD: 2560 case Op_CmpD: 2561 case Op_CmpD3: 2562 frc.inc_double_count(); 2563 break; 2564 case Op_Opaque1: // Remove Opaque Nodes before matching 2565 case Op_Opaque2: // Remove Opaque Nodes before matching 2566 n->subsume_by(n->in(1), this); 2567 break; 2568 case Op_CallStaticJava: 2569 case Op_CallJava: 2570 case Op_CallDynamicJava: 2571 frc.inc_java_call_count(); // Count java call site; 2572 case Op_CallRuntime: 2573 case Op_CallLeaf: 2574 case Op_CallLeafNoFP: { 2575 assert( n->is_Call(), "" ); 2576 CallNode *call = n->as_Call(); 2577 // Count call sites where the FP mode bit would have to be flipped. 2578 // Do not count uncommon runtime calls: 2579 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2580 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2581 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2582 frc.inc_call_count(); // Count the call site 2583 } else { // See if uncommon argument is shared 2584 Node *n = call->in(TypeFunc::Parms); 2585 int nop = n->Opcode(); 2586 // Clone shared simple arguments to uncommon calls, item (1). 2587 if( n->outcnt() > 1 && 2588 !n->is_Proj() && 2589 nop != Op_CreateEx && 2590 nop != Op_CheckCastPP && 2591 nop != Op_DecodeN && 2592 nop != Op_DecodeNKlass && 2593 !n->is_Mem() ) { 2594 Node *x = n->clone(); 2595 call->set_req( TypeFunc::Parms, x ); 2596 } 2597 } 2598 break; 2599 } 2600 2601 case Op_StoreD: 2602 case Op_LoadD: 2603 case Op_LoadD_unaligned: 2604 frc.inc_double_count(); 2605 goto handle_mem; 2606 case Op_StoreF: 2607 case Op_LoadF: 2608 frc.inc_float_count(); 2609 goto handle_mem; 2610 2611 case Op_StoreCM: 2612 { 2613 // Convert OopStore dependence into precedence edge 2614 Node* prec = n->in(MemNode::OopStore); 2615 n->del_req(MemNode::OopStore); 2616 n->add_prec(prec); 2617 eliminate_redundant_card_marks(n); 2618 } 2619 2620 // fall through 2621 2622 case Op_StoreB: 2623 case Op_StoreC: 2624 case Op_StorePConditional: 2625 case Op_StoreI: 2626 case Op_StoreL: 2627 case Op_StoreIConditional: 2628 case Op_StoreLConditional: 2629 case Op_CompareAndSwapI: 2630 case Op_CompareAndSwapL: 2631 case Op_CompareAndSwapP: 2632 case Op_CompareAndSwapN: 2633 case Op_GetAndAddI: 2634 case Op_GetAndAddL: 2635 case Op_GetAndSetI: 2636 case Op_GetAndSetL: 2637 case Op_GetAndSetP: 2638 case Op_GetAndSetN: 2639 case Op_StoreP: 2640 case Op_StoreN: 2641 case Op_StoreNKlass: 2642 case Op_LoadB: 2643 case Op_LoadUB: 2644 case Op_LoadUS: 2645 case Op_LoadI: 2646 case Op_LoadKlass: 2647 case Op_LoadNKlass: 2648 case Op_LoadL: 2649 case Op_LoadL_unaligned: 2650 case Op_LoadPLocked: 2651 case Op_LoadP: 2652 case Op_LoadN: 2653 case Op_LoadRange: 2654 case Op_LoadS: { 2655 handle_mem: 2656 #ifdef ASSERT 2657 if( VerifyOptoOopOffsets ) { 2658 assert( n->is_Mem(), "" ); 2659 MemNode *mem = (MemNode*)n; 2660 // Check to see if address types have grounded out somehow. 2661 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2662 assert( !tp || oop_offset_is_sane(tp), "" ); 2663 } 2664 #endif 2665 break; 2666 } 2667 2668 case Op_AddP: { // Assert sane base pointers 2669 Node *addp = n->in(AddPNode::Address); 2670 assert( !addp->is_AddP() || 2671 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2672 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2673 "Base pointers must match" ); 2674 #ifdef _LP64 2675 if ((UseCompressedOops || UseCompressedClassPointers) && 2676 addp->Opcode() == Op_ConP && 2677 addp == n->in(AddPNode::Base) && 2678 n->in(AddPNode::Offset)->is_Con()) { 2679 // Use addressing with narrow klass to load with offset on x86. 2680 // On sparc loading 32-bits constant and decoding it have less 2681 // instructions (4) then load 64-bits constant (7). 2682 // Do this transformation here since IGVN will convert ConN back to ConP. 2683 const Type* t = addp->bottom_type(); 2684 if (t->isa_oopptr() || t->isa_klassptr()) { 2685 Node* nn = NULL; 2686 2687 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2688 2689 // Look for existing ConN node of the same exact type. 2690 Node* r = root(); 2691 uint cnt = r->outcnt(); 2692 for (uint i = 0; i < cnt; i++) { 2693 Node* m = r->raw_out(i); 2694 if (m!= NULL && m->Opcode() == op && 2695 m->bottom_type()->make_ptr() == t) { 2696 nn = m; 2697 break; 2698 } 2699 } 2700 if (nn != NULL) { 2701 // Decode a narrow oop to match address 2702 // [R12 + narrow_oop_reg<<3 + offset] 2703 if (t->isa_oopptr()) { 2704 nn = new (this) DecodeNNode(nn, t); 2705 } else { 2706 nn = new (this) DecodeNKlassNode(nn, t); 2707 } 2708 n->set_req(AddPNode::Base, nn); 2709 n->set_req(AddPNode::Address, nn); 2710 if (addp->outcnt() == 0) { 2711 addp->disconnect_inputs(NULL, this); 2712 } 2713 } 2714 } 2715 } 2716 #endif 2717 break; 2718 } 2719 2720 #ifdef _LP64 2721 case Op_CastPP: 2722 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2723 Node* in1 = n->in(1); 2724 const Type* t = n->bottom_type(); 2725 Node* new_in1 = in1->clone(); 2726 new_in1->as_DecodeN()->set_type(t); 2727 2728 if (!Matcher::narrow_oop_use_complex_address()) { 2729 // 2730 // x86, ARM and friends can handle 2 adds in addressing mode 2731 // and Matcher can fold a DecodeN node into address by using 2732 // a narrow oop directly and do implicit NULL check in address: 2733 // 2734 // [R12 + narrow_oop_reg<<3 + offset] 2735 // NullCheck narrow_oop_reg 2736 // 2737 // On other platforms (Sparc) we have to keep new DecodeN node and 2738 // use it to do implicit NULL check in address: 2739 // 2740 // decode_not_null narrow_oop_reg, base_reg 2741 // [base_reg + offset] 2742 // NullCheck base_reg 2743 // 2744 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2745 // to keep the information to which NULL check the new DecodeN node 2746 // corresponds to use it as value in implicit_null_check(). 2747 // 2748 new_in1->set_req(0, n->in(0)); 2749 } 2750 2751 n->subsume_by(new_in1, this); 2752 if (in1->outcnt() == 0) { 2753 in1->disconnect_inputs(NULL, this); 2754 } 2755 } 2756 break; 2757 2758 case Op_CmpP: 2759 // Do this transformation here to preserve CmpPNode::sub() and 2760 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2761 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2762 Node* in1 = n->in(1); 2763 Node* in2 = n->in(2); 2764 if (!in1->is_DecodeNarrowPtr()) { 2765 in2 = in1; 2766 in1 = n->in(2); 2767 } 2768 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2769 2770 Node* new_in2 = NULL; 2771 if (in2->is_DecodeNarrowPtr()) { 2772 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2773 new_in2 = in2->in(1); 2774 } else if (in2->Opcode() == Op_ConP) { 2775 const Type* t = in2->bottom_type(); 2776 if (t == TypePtr::NULL_PTR) { 2777 assert(in1->is_DecodeN(), "compare klass to null?"); 2778 // Don't convert CmpP null check into CmpN if compressed 2779 // oops implicit null check is not generated. 2780 // This will allow to generate normal oop implicit null check. 2781 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2782 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2783 // 2784 // This transformation together with CastPP transformation above 2785 // will generated code for implicit NULL checks for compressed oops. 2786 // 2787 // The original code after Optimize() 2788 // 2789 // LoadN memory, narrow_oop_reg 2790 // decode narrow_oop_reg, base_reg 2791 // CmpP base_reg, NULL 2792 // CastPP base_reg // NotNull 2793 // Load [base_reg + offset], val_reg 2794 // 2795 // after these transformations will be 2796 // 2797 // LoadN memory, narrow_oop_reg 2798 // CmpN narrow_oop_reg, NULL 2799 // decode_not_null narrow_oop_reg, base_reg 2800 // Load [base_reg + offset], val_reg 2801 // 2802 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2803 // since narrow oops can be used in debug info now (see the code in 2804 // final_graph_reshaping_walk()). 2805 // 2806 // At the end the code will be matched to 2807 // on x86: 2808 // 2809 // Load_narrow_oop memory, narrow_oop_reg 2810 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2811 // NullCheck narrow_oop_reg 2812 // 2813 // and on sparc: 2814 // 2815 // Load_narrow_oop memory, narrow_oop_reg 2816 // decode_not_null narrow_oop_reg, base_reg 2817 // Load [base_reg + offset], val_reg 2818 // NullCheck base_reg 2819 // 2820 } else if (t->isa_oopptr()) { 2821 new_in2 = ConNode::make(this, t->make_narrowoop()); 2822 } else if (t->isa_klassptr()) { 2823 new_in2 = ConNode::make(this, t->make_narrowklass()); 2824 } 2825 } 2826 if (new_in2 != NULL) { 2827 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2828 n->subsume_by(cmpN, this); 2829 if (in1->outcnt() == 0) { 2830 in1->disconnect_inputs(NULL, this); 2831 } 2832 if (in2->outcnt() == 0) { 2833 in2->disconnect_inputs(NULL, this); 2834 } 2835 } 2836 } 2837 break; 2838 2839 case Op_DecodeN: 2840 case Op_DecodeNKlass: 2841 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2842 // DecodeN could be pinned when it can't be fold into 2843 // an address expression, see the code for Op_CastPP above. 2844 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2845 break; 2846 2847 case Op_EncodeP: 2848 case Op_EncodePKlass: { 2849 Node* in1 = n->in(1); 2850 if (in1->is_DecodeNarrowPtr()) { 2851 n->subsume_by(in1->in(1), this); 2852 } else if (in1->Opcode() == Op_ConP) { 2853 const Type* t = in1->bottom_type(); 2854 if (t == TypePtr::NULL_PTR) { 2855 assert(t->isa_oopptr(), "null klass?"); 2856 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2857 } else if (t->isa_oopptr()) { 2858 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2859 } else if (t->isa_klassptr()) { 2860 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2861 } 2862 } 2863 if (in1->outcnt() == 0) { 2864 in1->disconnect_inputs(NULL, this); 2865 } 2866 break; 2867 } 2868 2869 case Op_Proj: { 2870 if (OptimizeStringConcat) { 2871 ProjNode* p = n->as_Proj(); 2872 if (p->_is_io_use) { 2873 // Separate projections were used for the exception path which 2874 // are normally removed by a late inline. If it wasn't inlined 2875 // then they will hang around and should just be replaced with 2876 // the original one. 2877 Node* proj = NULL; 2878 // Replace with just one 2879 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2880 Node *use = i.get(); 2881 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2882 proj = use; 2883 break; 2884 } 2885 } 2886 assert(proj != NULL, "must be found"); 2887 p->subsume_by(proj, this); 2888 } 2889 } 2890 break; 2891 } 2892 2893 case Op_Phi: 2894 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2895 // The EncodeP optimization may create Phi with the same edges 2896 // for all paths. It is not handled well by Register Allocator. 2897 Node* unique_in = n->in(1); 2898 assert(unique_in != NULL, ""); 2899 uint cnt = n->req(); 2900 for (uint i = 2; i < cnt; i++) { 2901 Node* m = n->in(i); 2902 assert(m != NULL, ""); 2903 if (unique_in != m) 2904 unique_in = NULL; 2905 } 2906 if (unique_in != NULL) { 2907 n->subsume_by(unique_in, this); 2908 } 2909 } 2910 break; 2911 2912 #endif 2913 2914 case Op_ModI: 2915 if (UseDivMod) { 2916 // Check if a%b and a/b both exist 2917 Node* d = n->find_similar(Op_DivI); 2918 if (d) { 2919 // Replace them with a fused divmod if supported 2920 if (Matcher::has_match_rule(Op_DivModI)) { 2921 DivModINode* divmod = DivModINode::make(this, n); 2922 d->subsume_by(divmod->div_proj(), this); 2923 n->subsume_by(divmod->mod_proj(), this); 2924 } else { 2925 // replace a%b with a-((a/b)*b) 2926 Node* mult = new (this) MulINode(d, d->in(2)); 2927 Node* sub = new (this) SubINode(d->in(1), mult); 2928 n->subsume_by(sub, this); 2929 } 2930 } 2931 } 2932 break; 2933 2934 case Op_ModL: 2935 if (UseDivMod) { 2936 // Check if a%b and a/b both exist 2937 Node* d = n->find_similar(Op_DivL); 2938 if (d) { 2939 // Replace them with a fused divmod if supported 2940 if (Matcher::has_match_rule(Op_DivModL)) { 2941 DivModLNode* divmod = DivModLNode::make(this, n); 2942 d->subsume_by(divmod->div_proj(), this); 2943 n->subsume_by(divmod->mod_proj(), this); 2944 } else { 2945 // replace a%b with a-((a/b)*b) 2946 Node* mult = new (this) MulLNode(d, d->in(2)); 2947 Node* sub = new (this) SubLNode(d->in(1), mult); 2948 n->subsume_by(sub, this); 2949 } 2950 } 2951 } 2952 break; 2953 2954 case Op_LoadVector: 2955 case Op_StoreVector: 2956 break; 2957 2958 case Op_PackB: 2959 case Op_PackS: 2960 case Op_PackI: 2961 case Op_PackF: 2962 case Op_PackL: 2963 case Op_PackD: 2964 if (n->req()-1 > 2) { 2965 // Replace many operand PackNodes with a binary tree for matching 2966 PackNode* p = (PackNode*) n; 2967 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2968 n->subsume_by(btp, this); 2969 } 2970 break; 2971 case Op_Loop: 2972 case Op_CountedLoop: 2973 if (n->as_Loop()->is_inner_loop()) { 2974 frc.inc_inner_loop_count(); 2975 } 2976 break; 2977 case Op_LShiftI: 2978 case Op_RShiftI: 2979 case Op_URShiftI: 2980 case Op_LShiftL: 2981 case Op_RShiftL: 2982 case Op_URShiftL: 2983 if (Matcher::need_masked_shift_count) { 2984 // The cpu's shift instructions don't restrict the count to the 2985 // lower 5/6 bits. We need to do the masking ourselves. 2986 Node* in2 = n->in(2); 2987 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2988 const TypeInt* t = in2->find_int_type(); 2989 if (t != NULL && t->is_con()) { 2990 juint shift = t->get_con(); 2991 if (shift > mask) { // Unsigned cmp 2992 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2993 } 2994 } else { 2995 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2996 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2997 n->set_req(2, shift); 2998 } 2999 } 3000 if (in2->outcnt() == 0) { // Remove dead node 3001 in2->disconnect_inputs(NULL, this); 3002 } 3003 } 3004 break; 3005 case Op_MemBarStoreStore: 3006 case Op_MemBarRelease: 3007 // Break the link with AllocateNode: it is no longer useful and 3008 // confuses register allocation. 3009 if (n->req() > MemBarNode::Precedent) { 3010 n->set_req(MemBarNode::Precedent, top()); 3011 } 3012 break; 3013 // Must set a control edge on all nodes that produce a FlagsProj 3014 // so they can't escape the block that consumes the flags. 3015 // Must also set the non throwing branch as the control 3016 // for all nodes that depends on the result. Unless the node 3017 // already have a control that isn't the control of the 3018 // flag producer 3019 case Op_FlagsProj: 3020 { 3021 MathExactNode* math = (MathExactNode*) n->in(0); 3022 Node* ctrl = math->control_node(); 3023 Node* non_throwing = math->non_throwing_branch(); 3024 math->set_req(0, ctrl); 3025 3026 Node* result = math->result_node(); 3027 if (result != NULL) { 3028 for (DUIterator_Fast jmax, j = result->fast_outs(jmax); j < jmax; j++) { 3029 Node* out = result->fast_out(j); 3030 // Phi nodes shouldn't be moved. They would only match below if they 3031 // had the same control as the MathExactNode. The only time that 3032 // would happen is if the Phi is also an input to the MathExact 3033 // 3034 // Cmp nodes shouldn't have control set at all. 3035 if (out->is_Phi() || 3036 out->is_Cmp()) { 3037 continue; 3038 } 3039 3040 if (out->in(0) == NULL) { 3041 out->set_req(0, non_throwing); 3042 } else if (out->in(0) == ctrl) { 3043 out->set_req(0, non_throwing); 3044 } 3045 } 3046 } 3047 } 3048 break; 3049 default: 3050 assert( !n->is_Call(), "" ); 3051 assert( !n->is_Mem(), "" ); 3052 break; 3053 } 3054 3055 // Collect CFG split points 3056 if (n->is_MultiBranch()) 3057 frc._tests.push(n); 3058 } 3059 3060 //------------------------------final_graph_reshaping_walk--------------------- 3061 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3062 // requires that the walk visits a node's inputs before visiting the node. 3063 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3064 ResourceArea *area = Thread::current()->resource_area(); 3065 Unique_Node_List sfpt(area); 3066 3067 frc._visited.set(root->_idx); // first, mark node as visited 3068 uint cnt = root->req(); 3069 Node *n = root; 3070 uint i = 0; 3071 while (true) { 3072 if (i < cnt) { 3073 // Place all non-visited non-null inputs onto stack 3074 Node* m = n->in(i); 3075 ++i; 3076 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3077 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 3078 sfpt.push(m); 3079 cnt = m->req(); 3080 nstack.push(n, i); // put on stack parent and next input's index 3081 n = m; 3082 i = 0; 3083 } 3084 } else { 3085 // Now do post-visit work 3086 final_graph_reshaping_impl( n, frc ); 3087 if (nstack.is_empty()) 3088 break; // finished 3089 n = nstack.node(); // Get node from stack 3090 cnt = n->req(); 3091 i = nstack.index(); 3092 nstack.pop(); // Shift to the next node on stack 3093 } 3094 } 3095 3096 // Skip next transformation if compressed oops are not used. 3097 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3098 (!UseCompressedOops && !UseCompressedClassPointers)) 3099 return; 3100 3101 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3102 // It could be done for an uncommon traps or any safepoints/calls 3103 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3104 while (sfpt.size() > 0) { 3105 n = sfpt.pop(); 3106 JVMState *jvms = n->as_SafePoint()->jvms(); 3107 assert(jvms != NULL, "sanity"); 3108 int start = jvms->debug_start(); 3109 int end = n->req(); 3110 bool is_uncommon = (n->is_CallStaticJava() && 3111 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3112 for (int j = start; j < end; j++) { 3113 Node* in = n->in(j); 3114 if (in->is_DecodeNarrowPtr()) { 3115 bool safe_to_skip = true; 3116 if (!is_uncommon ) { 3117 // Is it safe to skip? 3118 for (uint i = 0; i < in->outcnt(); i++) { 3119 Node* u = in->raw_out(i); 3120 if (!u->is_SafePoint() || 3121 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3122 safe_to_skip = false; 3123 } 3124 } 3125 } 3126 if (safe_to_skip) { 3127 n->set_req(j, in->in(1)); 3128 } 3129 if (in->outcnt() == 0) { 3130 in->disconnect_inputs(NULL, this); 3131 } 3132 } 3133 } 3134 } 3135 } 3136 3137 //------------------------------final_graph_reshaping-------------------------- 3138 // Final Graph Reshaping. 3139 // 3140 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3141 // and not commoned up and forced early. Must come after regular 3142 // optimizations to avoid GVN undoing the cloning. Clone constant 3143 // inputs to Loop Phis; these will be split by the allocator anyways. 3144 // Remove Opaque nodes. 3145 // (2) Move last-uses by commutative operations to the left input to encourage 3146 // Intel update-in-place two-address operations and better register usage 3147 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3148 // calls canonicalizing them back. 3149 // (3) Count the number of double-precision FP ops, single-precision FP ops 3150 // and call sites. On Intel, we can get correct rounding either by 3151 // forcing singles to memory (requires extra stores and loads after each 3152 // FP bytecode) or we can set a rounding mode bit (requires setting and 3153 // clearing the mode bit around call sites). The mode bit is only used 3154 // if the relative frequency of single FP ops to calls is low enough. 3155 // This is a key transform for SPEC mpeg_audio. 3156 // (4) Detect infinite loops; blobs of code reachable from above but not 3157 // below. Several of the Code_Gen algorithms fail on such code shapes, 3158 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3159 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3160 // Detection is by looking for IfNodes where only 1 projection is 3161 // reachable from below or CatchNodes missing some targets. 3162 // (5) Assert for insane oop offsets in debug mode. 3163 3164 bool Compile::final_graph_reshaping() { 3165 // an infinite loop may have been eliminated by the optimizer, 3166 // in which case the graph will be empty. 3167 if (root()->req() == 1) { 3168 record_method_not_compilable("trivial infinite loop"); 3169 return true; 3170 } 3171 3172 // Expensive nodes have their control input set to prevent the GVN 3173 // from freely commoning them. There's no GVN beyond this point so 3174 // no need to keep the control input. We want the expensive nodes to 3175 // be freely moved to the least frequent code path by gcm. 3176 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3177 for (int i = 0; i < expensive_count(); i++) { 3178 _expensive_nodes->at(i)->set_req(0, NULL); 3179 } 3180 3181 Final_Reshape_Counts frc; 3182 3183 // Visit everybody reachable! 3184 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3185 Node_Stack nstack(unique() >> 1); 3186 final_graph_reshaping_walk(nstack, root(), frc); 3187 3188 // Check for unreachable (from below) code (i.e., infinite loops). 3189 for( uint i = 0; i < frc._tests.size(); i++ ) { 3190 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3191 // Get number of CFG targets. 3192 // Note that PCTables include exception targets after calls. 3193 uint required_outcnt = n->required_outcnt(); 3194 if (n->outcnt() != required_outcnt) { 3195 // Check for a few special cases. Rethrow Nodes never take the 3196 // 'fall-thru' path, so expected kids is 1 less. 3197 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3198 if (n->in(0)->in(0)->is_Call()) { 3199 CallNode *call = n->in(0)->in(0)->as_Call(); 3200 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3201 required_outcnt--; // Rethrow always has 1 less kid 3202 } else if (call->req() > TypeFunc::Parms && 3203 call->is_CallDynamicJava()) { 3204 // Check for null receiver. In such case, the optimizer has 3205 // detected that the virtual call will always result in a null 3206 // pointer exception. The fall-through projection of this CatchNode 3207 // will not be populated. 3208 Node *arg0 = call->in(TypeFunc::Parms); 3209 if (arg0->is_Type() && 3210 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3211 required_outcnt--; 3212 } 3213 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3214 call->req() > TypeFunc::Parms+1 && 3215 call->is_CallStaticJava()) { 3216 // Check for negative array length. In such case, the optimizer has 3217 // detected that the allocation attempt will always result in an 3218 // exception. There is no fall-through projection of this CatchNode . 3219 Node *arg1 = call->in(TypeFunc::Parms+1); 3220 if (arg1->is_Type() && 3221 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3222 required_outcnt--; 3223 } 3224 } 3225 } 3226 } 3227 // Recheck with a better notion of 'required_outcnt' 3228 if (n->outcnt() != required_outcnt) { 3229 record_method_not_compilable("malformed control flow"); 3230 return true; // Not all targets reachable! 3231 } 3232 } 3233 // Check that I actually visited all kids. Unreached kids 3234 // must be infinite loops. 3235 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3236 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3237 record_method_not_compilable("infinite loop"); 3238 return true; // Found unvisited kid; must be unreach 3239 } 3240 } 3241 3242 // If original bytecodes contained a mixture of floats and doubles 3243 // check if the optimizer has made it homogenous, item (3). 3244 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3245 frc.get_float_count() > 32 && 3246 frc.get_double_count() == 0 && 3247 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3248 set_24_bit_selection_and_mode( false, true ); 3249 } 3250 3251 set_java_calls(frc.get_java_call_count()); 3252 set_inner_loops(frc.get_inner_loop_count()); 3253 3254 // No infinite loops, no reason to bail out. 3255 return false; 3256 } 3257 3258 //-----------------------------too_many_traps---------------------------------- 3259 // Report if there are too many traps at the current method and bci. 3260 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3261 bool Compile::too_many_traps(ciMethod* method, 3262 int bci, 3263 Deoptimization::DeoptReason reason) { 3264 ciMethodData* md = method->method_data(); 3265 if (md->is_empty()) { 3266 // Assume the trap has not occurred, or that it occurred only 3267 // because of a transient condition during start-up in the interpreter. 3268 return false; 3269 } 3270 if (md->has_trap_at(bci, reason) != 0) { 3271 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3272 // Also, if there are multiple reasons, or if there is no per-BCI record, 3273 // assume the worst. 3274 if (log()) 3275 log()->elem("observe trap='%s' count='%d'", 3276 Deoptimization::trap_reason_name(reason), 3277 md->trap_count(reason)); 3278 return true; 3279 } else { 3280 // Ignore method/bci and see if there have been too many globally. 3281 return too_many_traps(reason, md); 3282 } 3283 } 3284 3285 // Less-accurate variant which does not require a method and bci. 3286 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3287 ciMethodData* logmd) { 3288 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 3289 // Too many traps globally. 3290 // Note that we use cumulative trap_count, not just md->trap_count. 3291 if (log()) { 3292 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3293 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3294 Deoptimization::trap_reason_name(reason), 3295 mcount, trap_count(reason)); 3296 } 3297 return true; 3298 } else { 3299 // The coast is clear. 3300 return false; 3301 } 3302 } 3303 3304 //--------------------------too_many_recompiles-------------------------------- 3305 // Report if there are too many recompiles at the current method and bci. 3306 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3307 // Is not eager to return true, since this will cause the compiler to use 3308 // Action_none for a trap point, to avoid too many recompilations. 3309 bool Compile::too_many_recompiles(ciMethod* method, 3310 int bci, 3311 Deoptimization::DeoptReason reason) { 3312 ciMethodData* md = method->method_data(); 3313 if (md->is_empty()) { 3314 // Assume the trap has not occurred, or that it occurred only 3315 // because of a transient condition during start-up in the interpreter. 3316 return false; 3317 } 3318 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3319 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3320 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3321 Deoptimization::DeoptReason per_bc_reason 3322 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3323 if ((per_bc_reason == Deoptimization::Reason_none 3324 || md->has_trap_at(bci, reason) != 0) 3325 // The trap frequency measure we care about is the recompile count: 3326 && md->trap_recompiled_at(bci) 3327 && md->overflow_recompile_count() >= bc_cutoff) { 3328 // Do not emit a trap here if it has already caused recompilations. 3329 // Also, if there are multiple reasons, or if there is no per-BCI record, 3330 // assume the worst. 3331 if (log()) 3332 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3333 Deoptimization::trap_reason_name(reason), 3334 md->trap_count(reason), 3335 md->overflow_recompile_count()); 3336 return true; 3337 } else if (trap_count(reason) != 0 3338 && decompile_count() >= m_cutoff) { 3339 // Too many recompiles globally, and we have seen this sort of trap. 3340 // Use cumulative decompile_count, not just md->decompile_count. 3341 if (log()) 3342 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3343 Deoptimization::trap_reason_name(reason), 3344 md->trap_count(reason), trap_count(reason), 3345 md->decompile_count(), decompile_count()); 3346 return true; 3347 } else { 3348 // The coast is clear. 3349 return false; 3350 } 3351 } 3352 3353 3354 #ifndef PRODUCT 3355 //------------------------------verify_graph_edges--------------------------- 3356 // Walk the Graph and verify that there is a one-to-one correspondence 3357 // between Use-Def edges and Def-Use edges in the graph. 3358 void Compile::verify_graph_edges(bool no_dead_code) { 3359 if (VerifyGraphEdges) { 3360 ResourceArea *area = Thread::current()->resource_area(); 3361 Unique_Node_List visited(area); 3362 // Call recursive graph walk to check edges 3363 _root->verify_edges(visited); 3364 if (no_dead_code) { 3365 // Now make sure that no visited node is used by an unvisited node. 3366 bool dead_nodes = 0; 3367 Unique_Node_List checked(area); 3368 while (visited.size() > 0) { 3369 Node* n = visited.pop(); 3370 checked.push(n); 3371 for (uint i = 0; i < n->outcnt(); i++) { 3372 Node* use = n->raw_out(i); 3373 if (checked.member(use)) continue; // already checked 3374 if (visited.member(use)) continue; // already in the graph 3375 if (use->is_Con()) continue; // a dead ConNode is OK 3376 // At this point, we have found a dead node which is DU-reachable. 3377 if (dead_nodes++ == 0) 3378 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3379 use->dump(2); 3380 tty->print_cr("---"); 3381 checked.push(use); // No repeats; pretend it is now checked. 3382 } 3383 } 3384 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3385 } 3386 } 3387 } 3388 3389 // Verify GC barriers consistency 3390 // Currently supported: 3391 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3392 void Compile::verify_barriers() { 3393 if (UseG1GC) { 3394 // Verify G1 pre-barriers 3395 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3396 3397 ResourceArea *area = Thread::current()->resource_area(); 3398 Unique_Node_List visited(area); 3399 Node_List worklist(area); 3400 // We're going to walk control flow backwards starting from the Root 3401 worklist.push(_root); 3402 while (worklist.size() > 0) { 3403 Node* x = worklist.pop(); 3404 if (x == NULL || x == top()) continue; 3405 if (visited.member(x)) { 3406 continue; 3407 } else { 3408 visited.push(x); 3409 } 3410 3411 if (x->is_Region()) { 3412 for (uint i = 1; i < x->req(); i++) { 3413 worklist.push(x->in(i)); 3414 } 3415 } else { 3416 worklist.push(x->in(0)); 3417 // We are looking for the pattern: 3418 // /->ThreadLocal 3419 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3420 // \->ConI(0) 3421 // We want to verify that the If and the LoadB have the same control 3422 // See GraphKit::g1_write_barrier_pre() 3423 if (x->is_If()) { 3424 IfNode *iff = x->as_If(); 3425 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3426 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3427 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3428 && cmp->in(1)->is_Load()) { 3429 LoadNode* load = cmp->in(1)->as_Load(); 3430 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3431 && load->in(2)->in(3)->is_Con() 3432 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3433 3434 Node* if_ctrl = iff->in(0); 3435 Node* load_ctrl = load->in(0); 3436 3437 if (if_ctrl != load_ctrl) { 3438 // Skip possible CProj->NeverBranch in infinite loops 3439 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3440 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3441 if_ctrl = if_ctrl->in(0)->in(0); 3442 } 3443 } 3444 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3445 } 3446 } 3447 } 3448 } 3449 } 3450 } 3451 } 3452 } 3453 3454 #endif 3455 3456 // The Compile object keeps track of failure reasons separately from the ciEnv. 3457 // This is required because there is not quite a 1-1 relation between the 3458 // ciEnv and its compilation task and the Compile object. Note that one 3459 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3460 // to backtrack and retry without subsuming loads. Other than this backtracking 3461 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3462 // by the logic in C2Compiler. 3463 void Compile::record_failure(const char* reason) { 3464 if (log() != NULL) { 3465 log()->elem("failure reason='%s' phase='compile'", reason); 3466 } 3467 if (_failure_reason == NULL) { 3468 // Record the first failure reason. 3469 _failure_reason = reason; 3470 } 3471 3472 EventCompilerFailure event; 3473 if (event.should_commit()) { 3474 event.set_compileID(Compile::compile_id()); 3475 event.set_failure(reason); 3476 event.commit(); 3477 } 3478 3479 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3480 C->print_method(PHASE_FAILURE); 3481 } 3482 _root = NULL; // flush the graph, too 3483 } 3484 3485 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3486 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3487 _phase_name(name), _dolog(dolog) 3488 { 3489 if (dolog) { 3490 C = Compile::current(); 3491 _log = C->log(); 3492 } else { 3493 C = NULL; 3494 _log = NULL; 3495 } 3496 if (_log != NULL) { 3497 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3498 _log->stamp(); 3499 _log->end_head(); 3500 } 3501 } 3502 3503 Compile::TracePhase::~TracePhase() { 3504 3505 C = Compile::current(); 3506 if (_dolog) { 3507 _log = C->log(); 3508 } else { 3509 _log = NULL; 3510 } 3511 3512 #ifdef ASSERT 3513 if (PrintIdealNodeCount) { 3514 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3515 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3516 } 3517 3518 if (VerifyIdealNodeCount) { 3519 Compile::current()->print_missing_nodes(); 3520 } 3521 #endif 3522 3523 if (_log != NULL) { 3524 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3525 } 3526 } 3527 3528 //============================================================================= 3529 // Two Constant's are equal when the type and the value are equal. 3530 bool Compile::Constant::operator==(const Constant& other) { 3531 if (type() != other.type() ) return false; 3532 if (can_be_reused() != other.can_be_reused()) return false; 3533 // For floating point values we compare the bit pattern. 3534 switch (type()) { 3535 case T_FLOAT: return (_v._value.i == other._v._value.i); 3536 case T_LONG: 3537 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3538 case T_OBJECT: 3539 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3540 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3541 case T_METADATA: return (_v._metadata == other._v._metadata); 3542 default: ShouldNotReachHere(); 3543 } 3544 return false; 3545 } 3546 3547 static int type_to_size_in_bytes(BasicType t) { 3548 switch (t) { 3549 case T_LONG: return sizeof(jlong ); 3550 case T_FLOAT: return sizeof(jfloat ); 3551 case T_DOUBLE: return sizeof(jdouble); 3552 case T_METADATA: return sizeof(Metadata*); 3553 // We use T_VOID as marker for jump-table entries (labels) which 3554 // need an internal word relocation. 3555 case T_VOID: 3556 case T_ADDRESS: 3557 case T_OBJECT: return sizeof(jobject); 3558 } 3559 3560 ShouldNotReachHere(); 3561 return -1; 3562 } 3563 3564 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3565 // sort descending 3566 if (a->freq() > b->freq()) return -1; 3567 if (a->freq() < b->freq()) return 1; 3568 return 0; 3569 } 3570 3571 void Compile::ConstantTable::calculate_offsets_and_size() { 3572 // First, sort the array by frequencies. 3573 _constants.sort(qsort_comparator); 3574 3575 #ifdef ASSERT 3576 // Make sure all jump-table entries were sorted to the end of the 3577 // array (they have a negative frequency). 3578 bool found_void = false; 3579 for (int i = 0; i < _constants.length(); i++) { 3580 Constant con = _constants.at(i); 3581 if (con.type() == T_VOID) 3582 found_void = true; // jump-tables 3583 else 3584 assert(!found_void, "wrong sorting"); 3585 } 3586 #endif 3587 3588 int offset = 0; 3589 for (int i = 0; i < _constants.length(); i++) { 3590 Constant* con = _constants.adr_at(i); 3591 3592 // Align offset for type. 3593 int typesize = type_to_size_in_bytes(con->type()); 3594 offset = align_size_up(offset, typesize); 3595 con->set_offset(offset); // set constant's offset 3596 3597 if (con->type() == T_VOID) { 3598 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3599 offset = offset + typesize * n->outcnt(); // expand jump-table 3600 } else { 3601 offset = offset + typesize; 3602 } 3603 } 3604 3605 // Align size up to the next section start (which is insts; see 3606 // CodeBuffer::align_at_start). 3607 assert(_size == -1, "already set?"); 3608 _size = align_size_up(offset, CodeEntryAlignment); 3609 } 3610 3611 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3612 MacroAssembler _masm(&cb); 3613 for (int i = 0; i < _constants.length(); i++) { 3614 Constant con = _constants.at(i); 3615 address constant_addr; 3616 switch (con.type()) { 3617 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3618 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3619 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3620 case T_OBJECT: { 3621 jobject obj = con.get_jobject(); 3622 int oop_index = _masm.oop_recorder()->find_index(obj); 3623 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3624 break; 3625 } 3626 case T_ADDRESS: { 3627 address addr = (address) con.get_jobject(); 3628 constant_addr = _masm.address_constant(addr); 3629 break; 3630 } 3631 // We use T_VOID as marker for jump-table entries (labels) which 3632 // need an internal word relocation. 3633 case T_VOID: { 3634 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3635 // Fill the jump-table with a dummy word. The real value is 3636 // filled in later in fill_jump_table. 3637 address dummy = (address) n; 3638 constant_addr = _masm.address_constant(dummy); 3639 // Expand jump-table 3640 for (uint i = 1; i < n->outcnt(); i++) { 3641 address temp_addr = _masm.address_constant(dummy + i); 3642 assert(temp_addr, "consts section too small"); 3643 } 3644 break; 3645 } 3646 case T_METADATA: { 3647 Metadata* obj = con.get_metadata(); 3648 int metadata_index = _masm.oop_recorder()->find_index(obj); 3649 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3650 break; 3651 } 3652 default: ShouldNotReachHere(); 3653 } 3654 assert(constant_addr, "consts section too small"); 3655 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3656 } 3657 } 3658 3659 int Compile::ConstantTable::find_offset(Constant& con) const { 3660 int idx = _constants.find(con); 3661 assert(idx != -1, "constant must be in constant table"); 3662 int offset = _constants.at(idx).offset(); 3663 assert(offset != -1, "constant table not emitted yet?"); 3664 return offset; 3665 } 3666 3667 void Compile::ConstantTable::add(Constant& con) { 3668 if (con.can_be_reused()) { 3669 int idx = _constants.find(con); 3670 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3671 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3672 return; 3673 } 3674 } 3675 (void) _constants.append(con); 3676 } 3677 3678 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3679 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3680 Constant con(type, value, b->_freq); 3681 add(con); 3682 return con; 3683 } 3684 3685 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3686 Constant con(metadata); 3687 add(con); 3688 return con; 3689 } 3690 3691 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3692 jvalue value; 3693 BasicType type = oper->type()->basic_type(); 3694 switch (type) { 3695 case T_LONG: value.j = oper->constantL(); break; 3696 case T_FLOAT: value.f = oper->constantF(); break; 3697 case T_DOUBLE: value.d = oper->constantD(); break; 3698 case T_OBJECT: 3699 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3700 case T_METADATA: return add((Metadata*)oper->constant()); break; 3701 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3702 } 3703 return add(n, type, value); 3704 } 3705 3706 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3707 jvalue value; 3708 // We can use the node pointer here to identify the right jump-table 3709 // as this method is called from Compile::Fill_buffer right before 3710 // the MachNodes are emitted and the jump-table is filled (means the 3711 // MachNode pointers do not change anymore). 3712 value.l = (jobject) n; 3713 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3714 add(con); 3715 return con; 3716 } 3717 3718 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3719 // If called from Compile::scratch_emit_size do nothing. 3720 if (Compile::current()->in_scratch_emit_size()) return; 3721 3722 assert(labels.is_nonempty(), "must be"); 3723 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3724 3725 // Since MachConstantNode::constant_offset() also contains 3726 // table_base_offset() we need to subtract the table_base_offset() 3727 // to get the plain offset into the constant table. 3728 int offset = n->constant_offset() - table_base_offset(); 3729 3730 MacroAssembler _masm(&cb); 3731 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3732 3733 for (uint i = 0; i < n->outcnt(); i++) { 3734 address* constant_addr = &jump_table_base[i]; 3735 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))); 3736 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3737 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3738 } 3739 } 3740 3741 void Compile::dump_inlining() { 3742 if (print_inlining() || print_intrinsics()) { 3743 // Print inlining message for candidates that we couldn't inline 3744 // for lack of space or non constant receiver 3745 for (int i = 0; i < _late_inlines.length(); i++) { 3746 CallGenerator* cg = _late_inlines.at(i); 3747 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3748 } 3749 Unique_Node_List useful; 3750 useful.push(root()); 3751 for (uint next = 0; next < useful.size(); ++next) { 3752 Node* n = useful.at(next); 3753 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3754 CallNode* call = n->as_Call(); 3755 CallGenerator* cg = call->generator(); 3756 cg->print_inlining_late("receiver not constant"); 3757 } 3758 uint max = n->len(); 3759 for ( uint i = 0; i < max; ++i ) { 3760 Node *m = n->in(i); 3761 if ( m == NULL ) continue; 3762 useful.push(m); 3763 } 3764 } 3765 for (int i = 0; i < _print_inlining_list->length(); i++) { 3766 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string()); 3767 } 3768 } 3769 } 3770 3771 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3772 if (n1->Opcode() < n2->Opcode()) return -1; 3773 else if (n1->Opcode() > n2->Opcode()) return 1; 3774 3775 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())); 3776 for (uint i = 1; i < n1->req(); i++) { 3777 if (n1->in(i) < n2->in(i)) return -1; 3778 else if (n1->in(i) > n2->in(i)) return 1; 3779 } 3780 3781 return 0; 3782 } 3783 3784 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3785 Node* n1 = *n1p; 3786 Node* n2 = *n2p; 3787 3788 return cmp_expensive_nodes(n1, n2); 3789 } 3790 3791 void Compile::sort_expensive_nodes() { 3792 if (!expensive_nodes_sorted()) { 3793 _expensive_nodes->sort(cmp_expensive_nodes); 3794 } 3795 } 3796 3797 bool Compile::expensive_nodes_sorted() const { 3798 for (int i = 1; i < _expensive_nodes->length(); i++) { 3799 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3800 return false; 3801 } 3802 } 3803 return true; 3804 } 3805 3806 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3807 if (_expensive_nodes->length() == 0) { 3808 return false; 3809 } 3810 3811 assert(OptimizeExpensiveOps, "optimization off?"); 3812 3813 // Take this opportunity to remove dead nodes from the list 3814 int j = 0; 3815 for (int i = 0; i < _expensive_nodes->length(); i++) { 3816 Node* n = _expensive_nodes->at(i); 3817 if (!n->is_unreachable(igvn)) { 3818 assert(n->is_expensive(), "should be expensive"); 3819 _expensive_nodes->at_put(j, n); 3820 j++; 3821 } 3822 } 3823 _expensive_nodes->trunc_to(j); 3824 3825 // Then sort the list so that similar nodes are next to each other 3826 // and check for at least two nodes of identical kind with same data 3827 // inputs. 3828 sort_expensive_nodes(); 3829 3830 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3831 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3832 return true; 3833 } 3834 } 3835 3836 return false; 3837 } 3838 3839 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3840 if (_expensive_nodes->length() == 0) { 3841 return; 3842 } 3843 3844 assert(OptimizeExpensiveOps, "optimization off?"); 3845 3846 // Sort to bring similar nodes next to each other and clear the 3847 // control input of nodes for which there's only a single copy. 3848 sort_expensive_nodes(); 3849 3850 int j = 0; 3851 int identical = 0; 3852 int i = 0; 3853 for (; i < _expensive_nodes->length()-1; i++) { 3854 assert(j <= i, "can't write beyond current index"); 3855 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3856 identical++; 3857 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3858 continue; 3859 } 3860 if (identical > 0) { 3861 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3862 identical = 0; 3863 } else { 3864 Node* n = _expensive_nodes->at(i); 3865 igvn.hash_delete(n); 3866 n->set_req(0, NULL); 3867 igvn.hash_insert(n); 3868 } 3869 } 3870 if (identical > 0) { 3871 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3872 } else if (_expensive_nodes->length() >= 1) { 3873 Node* n = _expensive_nodes->at(i); 3874 igvn.hash_delete(n); 3875 n->set_req(0, NULL); 3876 igvn.hash_insert(n); 3877 } 3878 _expensive_nodes->trunc_to(j); 3879 } 3880 3881 void Compile::add_expensive_node(Node * n) { 3882 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3883 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3884 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3885 if (OptimizeExpensiveOps) { 3886 _expensive_nodes->append(n); 3887 } else { 3888 // Clear control input and let IGVN optimize expensive nodes if 3889 // OptimizeExpensiveOps is off. 3890 n->set_req(0, NULL); 3891 } 3892 } 3893 3894 /** 3895 * Remove the speculative part of types and clean up the graph 3896 */ 3897 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 3898 if (UseTypeSpeculation) { 3899 Unique_Node_List worklist; 3900 worklist.push(root()); 3901 int modified = 0; 3902 // Go over all type nodes that carry a speculative type, drop the 3903 // speculative part of the type and enqueue the node for an igvn 3904 // which may optimize it out. 3905 for (uint next = 0; next < worklist.size(); ++next) { 3906 Node *n = worklist.at(next); 3907 if (n->is_Type() && n->as_Type()->type()->isa_oopptr() != NULL && 3908 n->as_Type()->type()->is_oopptr()->speculative() != NULL) { 3909 TypeNode* tn = n->as_Type(); 3910 const TypeOopPtr* t = tn->type()->is_oopptr(); 3911 bool in_hash = igvn.hash_delete(n); 3912 assert(in_hash, "node should be in igvn hash table"); 3913 tn->set_type(t->remove_speculative()); 3914 igvn.hash_insert(n); 3915 igvn._worklist.push(n); // give it a chance to go away 3916 modified++; 3917 } 3918 uint max = n->len(); 3919 for( uint i = 0; i < max; ++i ) { 3920 Node *m = n->in(i); 3921 if (not_a_node(m)) continue; 3922 worklist.push(m); 3923 } 3924 } 3925 // Drop the speculative part of all types in the igvn's type table 3926 igvn.remove_speculative_types(); 3927 if (modified > 0) { 3928 igvn.optimize(); 3929 } 3930 } 3931 } 3932 3933 // Auxiliary method to support randomized stressing/fuzzing. 3934 // 3935 // This method can be called the arbitrary number of times, with current count 3936 // as the argument. The logic allows selecting a single candidate from the 3937 // running list of candidates as follows: 3938 // int count = 0; 3939 // Cand* selected = null; 3940 // while(cand = cand->next()) { 3941 // if (randomized_select(++count)) { 3942 // selected = cand; 3943 // } 3944 // } 3945 // 3946 // Including count equalizes the chances any candidate is "selected". 3947 // This is useful when we don't have the complete list of candidates to choose 3948 // from uniformly. In this case, we need to adjust the randomicity of the 3949 // selection, or else we will end up biasing the selection towards the latter 3950 // candidates. 3951 // 3952 // Quick back-envelope calculation shows that for the list of n candidates 3953 // the equal probability for the candidate to persist as "best" can be 3954 // achieved by replacing it with "next" k-th candidate with the probability 3955 // of 1/k. It can be easily shown that by the end of the run, the 3956 // probability for any candidate is converged to 1/n, thus giving the 3957 // uniform distribution among all the candidates. 3958 // 3959 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3960 #define RANDOMIZED_DOMAIN_POW 29 3961 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3962 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3963 bool Compile::randomized_select(int count) { 3964 assert(count > 0, "only positive"); 3965 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3966 }