1 /* 2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "opto/addnode.hpp" 36 #include "opto/block.hpp" 37 #include "opto/c2compiler.hpp" 38 #include "opto/callGenerator.hpp" 39 #include "opto/callnode.hpp" 40 #include "opto/cfgnode.hpp" 41 #include "opto/chaitin.hpp" 42 #include "opto/compile.hpp" 43 #include "opto/connode.hpp" 44 #include "opto/divnode.hpp" 45 #include "opto/escape.hpp" 46 #include "opto/idealGraphPrinter.hpp" 47 #include "opto/loopnode.hpp" 48 #include "opto/machnode.hpp" 49 #include "opto/macro.hpp" 50 #include "opto/matcher.hpp" 51 #include "opto/mathexactnode.hpp" 52 #include "opto/memnode.hpp" 53 #include "opto/mulnode.hpp" 54 #include "opto/narrowptrnode.hpp" 55 #include "opto/node.hpp" 56 #include "opto/opcodes.hpp" 57 #include "opto/output.hpp" 58 #include "opto/parse.hpp" 59 #include "opto/phaseX.hpp" 60 #include "opto/rootnode.hpp" 61 #include "opto/runtime.hpp" 62 #include "opto/stringopts.hpp" 63 #include "opto/type.hpp" 64 #include "opto/vectornode.hpp" 65 #include "runtime/arguments.hpp" 66 #include "runtime/signature.hpp" 67 #include "runtime/stubRoutines.hpp" 68 #include "runtime/timer.hpp" 69 #include "trace/tracing.hpp" 70 #include "utilities/copy.hpp" 71 #ifdef TARGET_ARCH_MODEL_x86_32 72 # include "adfiles/ad_x86_32.hpp" 73 #endif 74 #ifdef TARGET_ARCH_MODEL_x86_64 75 # include "adfiles/ad_x86_64.hpp" 76 #endif 77 #ifdef TARGET_ARCH_MODEL_sparc 78 # include "adfiles/ad_sparc.hpp" 79 #endif 80 #ifdef TARGET_ARCH_MODEL_zero 81 # include "adfiles/ad_zero.hpp" 82 #endif 83 #ifdef TARGET_ARCH_MODEL_arm 84 # include "adfiles/ad_arm.hpp" 85 #endif 86 #ifdef TARGET_ARCH_MODEL_ppc_32 87 # include "adfiles/ad_ppc_32.hpp" 88 #endif 89 #ifdef TARGET_ARCH_MODEL_ppc_64 90 # include "adfiles/ad_ppc_64.hpp" 91 #endif 92 93 94 // -------------------- Compile::mach_constant_base_node ----------------------- 95 // Constant table base node singleton. 96 MachConstantBaseNode* Compile::mach_constant_base_node() { 97 if (_mach_constant_base_node == NULL) { 98 _mach_constant_base_node = new (C) MachConstantBaseNode(); 99 _mach_constant_base_node->add_req(C->root()); 100 } 101 return _mach_constant_base_node; 102 } 103 104 105 /// Support for intrinsics. 106 107 // Return the index at which m must be inserted (or already exists). 108 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 109 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 110 #ifdef ASSERT 111 for (int i = 1; i < _intrinsics->length(); i++) { 112 CallGenerator* cg1 = _intrinsics->at(i-1); 113 CallGenerator* cg2 = _intrinsics->at(i); 114 assert(cg1->method() != cg2->method() 115 ? cg1->method() < cg2->method() 116 : cg1->is_virtual() < cg2->is_virtual(), 117 "compiler intrinsics list must stay sorted"); 118 } 119 #endif 120 // Binary search sorted list, in decreasing intervals [lo, hi]. 121 int lo = 0, hi = _intrinsics->length()-1; 122 while (lo <= hi) { 123 int mid = (uint)(hi + lo) / 2; 124 ciMethod* mid_m = _intrinsics->at(mid)->method(); 125 if (m < mid_m) { 126 hi = mid-1; 127 } else if (m > mid_m) { 128 lo = mid+1; 129 } else { 130 // look at minor sort key 131 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 132 if (is_virtual < mid_virt) { 133 hi = mid-1; 134 } else if (is_virtual > mid_virt) { 135 lo = mid+1; 136 } else { 137 return mid; // exact match 138 } 139 } 140 } 141 return lo; // inexact match 142 } 143 144 void Compile::register_intrinsic(CallGenerator* cg) { 145 if (_intrinsics == NULL) { 146 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 147 } 148 // This code is stolen from ciObjectFactory::insert. 149 // Really, GrowableArray should have methods for 150 // insert_at, remove_at, and binary_search. 151 int len = _intrinsics->length(); 152 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 153 if (index == len) { 154 _intrinsics->append(cg); 155 } else { 156 #ifdef ASSERT 157 CallGenerator* oldcg = _intrinsics->at(index); 158 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 159 #endif 160 _intrinsics->append(_intrinsics->at(len-1)); 161 int pos; 162 for (pos = len-2; pos >= index; pos--) { 163 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 164 } 165 _intrinsics->at_put(index, cg); 166 } 167 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 168 } 169 170 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 171 assert(m->is_loaded(), "don't try this on unloaded methods"); 172 if (_intrinsics != NULL) { 173 int index = intrinsic_insertion_index(m, is_virtual); 174 if (index < _intrinsics->length() 175 && _intrinsics->at(index)->method() == m 176 && _intrinsics->at(index)->is_virtual() == is_virtual) { 177 return _intrinsics->at(index); 178 } 179 } 180 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 181 if (m->intrinsic_id() != vmIntrinsics::_none && 182 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 183 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 184 if (cg != NULL) { 185 // Save it for next time: 186 register_intrinsic(cg); 187 return cg; 188 } else { 189 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 190 } 191 } 192 return NULL; 193 } 194 195 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 196 // in library_call.cpp. 197 198 199 #ifndef PRODUCT 200 // statistics gathering... 201 202 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 203 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 204 205 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 206 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 207 int oflags = _intrinsic_hist_flags[id]; 208 assert(flags != 0, "what happened?"); 209 if (is_virtual) { 210 flags |= _intrinsic_virtual; 211 } 212 bool changed = (flags != oflags); 213 if ((flags & _intrinsic_worked) != 0) { 214 juint count = (_intrinsic_hist_count[id] += 1); 215 if (count == 1) { 216 changed = true; // first time 217 } 218 // increment the overall count also: 219 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 220 } 221 if (changed) { 222 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 223 // Something changed about the intrinsic's virtuality. 224 if ((flags & _intrinsic_virtual) != 0) { 225 // This is the first use of this intrinsic as a virtual call. 226 if (oflags != 0) { 227 // We already saw it as a non-virtual, so note both cases. 228 flags |= _intrinsic_both; 229 } 230 } else if ((oflags & _intrinsic_both) == 0) { 231 // This is the first use of this intrinsic as a non-virtual 232 flags |= _intrinsic_both; 233 } 234 } 235 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 236 } 237 // update the overall flags also: 238 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 239 return changed; 240 } 241 242 static char* format_flags(int flags, char* buf) { 243 buf[0] = 0; 244 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 245 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 246 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 247 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 248 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 249 if (buf[0] == 0) strcat(buf, ","); 250 assert(buf[0] == ',', "must be"); 251 return &buf[1]; 252 } 253 254 void Compile::print_intrinsic_statistics() { 255 char flagsbuf[100]; 256 ttyLocker ttyl; 257 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 258 tty->print_cr("Compiler intrinsic usage:"); 259 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 260 if (total == 0) total = 1; // avoid div0 in case of no successes 261 #define PRINT_STAT_LINE(name, c, f) \ 262 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 263 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 264 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 265 int flags = _intrinsic_hist_flags[id]; 266 juint count = _intrinsic_hist_count[id]; 267 if ((flags | count) != 0) { 268 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 269 } 270 } 271 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 272 if (xtty != NULL) xtty->tail("statistics"); 273 } 274 275 void Compile::print_statistics() { 276 { ttyLocker ttyl; 277 if (xtty != NULL) xtty->head("statistics type='opto'"); 278 Parse::print_statistics(); 279 PhaseCCP::print_statistics(); 280 PhaseRegAlloc::print_statistics(); 281 Scheduling::print_statistics(); 282 PhasePeephole::print_statistics(); 283 PhaseIdealLoop::print_statistics(); 284 if (xtty != NULL) xtty->tail("statistics"); 285 } 286 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 287 // put this under its own <statistics> element. 288 print_intrinsic_statistics(); 289 } 290 } 291 #endif //PRODUCT 292 293 // Support for bundling info 294 Bundle* Compile::node_bundling(const Node *n) { 295 assert(valid_bundle_info(n), "oob"); 296 return &_node_bundling_base[n->_idx]; 297 } 298 299 bool Compile::valid_bundle_info(const Node *n) { 300 return (_node_bundling_limit > n->_idx); 301 } 302 303 304 void Compile::gvn_replace_by(Node* n, Node* nn) { 305 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 306 Node* use = n->last_out(i); 307 bool is_in_table = initial_gvn()->hash_delete(use); 308 uint uses_found = 0; 309 for (uint j = 0; j < use->len(); j++) { 310 if (use->in(j) == n) { 311 if (j < use->req()) 312 use->set_req(j, nn); 313 else 314 use->set_prec(j, nn); 315 uses_found++; 316 } 317 } 318 if (is_in_table) { 319 // reinsert into table 320 initial_gvn()->hash_find_insert(use); 321 } 322 record_for_igvn(use); 323 i -= uses_found; // we deleted 1 or more copies of this edge 324 } 325 } 326 327 328 static inline bool not_a_node(const Node* n) { 329 if (n == NULL) return true; 330 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 331 if (*(address*)n == badAddress) return true; // kill by Node::destruct 332 return false; 333 } 334 335 // Identify all nodes that are reachable from below, useful. 336 // Use breadth-first pass that records state in a Unique_Node_List, 337 // recursive traversal is slower. 338 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 339 int estimated_worklist_size = unique(); 340 useful.map( estimated_worklist_size, NULL ); // preallocate space 341 342 // Initialize worklist 343 if (root() != NULL) { useful.push(root()); } 344 // If 'top' is cached, declare it useful to preserve cached node 345 if( cached_top_node() ) { useful.push(cached_top_node()); } 346 347 // Push all useful nodes onto the list, breadthfirst 348 for( uint next = 0; next < useful.size(); ++next ) { 349 assert( next < unique(), "Unique useful nodes < total nodes"); 350 Node *n = useful.at(next); 351 uint max = n->len(); 352 for( uint i = 0; i < max; ++i ) { 353 Node *m = n->in(i); 354 if (not_a_node(m)) continue; 355 useful.push(m); 356 } 357 } 358 } 359 360 // Update dead_node_list with any missing dead nodes using useful 361 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 362 void Compile::update_dead_node_list(Unique_Node_List &useful) { 363 uint max_idx = unique(); 364 VectorSet& useful_node_set = useful.member_set(); 365 366 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 367 // If node with index node_idx is not in useful set, 368 // mark it as dead in dead node list. 369 if (! useful_node_set.test(node_idx) ) { 370 record_dead_node(node_idx); 371 } 372 } 373 } 374 375 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 376 int shift = 0; 377 for (int i = 0; i < inlines->length(); i++) { 378 CallGenerator* cg = inlines->at(i); 379 CallNode* call = cg->call_node(); 380 if (shift > 0) { 381 inlines->at_put(i-shift, cg); 382 } 383 if (!useful.member(call)) { 384 shift++; 385 } 386 } 387 inlines->trunc_to(inlines->length()-shift); 388 } 389 390 // Disconnect all useless nodes by disconnecting those at the boundary. 391 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 392 uint next = 0; 393 while (next < useful.size()) { 394 Node *n = useful.at(next++); 395 // Use raw traversal of out edges since this code removes out edges 396 int max = n->outcnt(); 397 for (int j = 0; j < max; ++j) { 398 Node* child = n->raw_out(j); 399 if (! useful.member(child)) { 400 assert(!child->is_top() || child != top(), 401 "If top is cached in Compile object it is in useful list"); 402 // Only need to remove this out-edge to the useless node 403 n->raw_del_out(j); 404 --j; 405 --max; 406 } 407 } 408 if (n->outcnt() == 1 && n->has_special_unique_user()) { 409 record_for_igvn(n->unique_out()); 410 } 411 } 412 // Remove useless macro and predicate opaq nodes 413 for (int i = C->macro_count()-1; i >= 0; i--) { 414 Node* n = C->macro_node(i); 415 if (!useful.member(n)) { 416 remove_macro_node(n); 417 } 418 } 419 // Remove useless expensive node 420 for (int i = C->expensive_count()-1; i >= 0; i--) { 421 Node* n = C->expensive_node(i); 422 if (!useful.member(n)) { 423 remove_expensive_node(n); 424 } 425 } 426 // clean up the late inline lists 427 remove_useless_late_inlines(&_string_late_inlines, useful); 428 remove_useless_late_inlines(&_boxing_late_inlines, useful); 429 remove_useless_late_inlines(&_late_inlines, useful); 430 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 431 } 432 433 //------------------------------frame_size_in_words----------------------------- 434 // frame_slots in units of words 435 int Compile::frame_size_in_words() const { 436 // shift is 0 in LP32 and 1 in LP64 437 const int shift = (LogBytesPerWord - LogBytesPerInt); 438 int words = _frame_slots >> shift; 439 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 440 return words; 441 } 442 443 // ============================================================================ 444 //------------------------------CompileWrapper--------------------------------- 445 class CompileWrapper : public StackObj { 446 Compile *const _compile; 447 public: 448 CompileWrapper(Compile* compile); 449 450 ~CompileWrapper(); 451 }; 452 453 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 454 // the Compile* pointer is stored in the current ciEnv: 455 ciEnv* env = compile->env(); 456 assert(env == ciEnv::current(), "must already be a ciEnv active"); 457 assert(env->compiler_data() == NULL, "compile already active?"); 458 env->set_compiler_data(compile); 459 assert(compile == Compile::current(), "sanity"); 460 461 compile->set_type_dict(NULL); 462 compile->set_type_hwm(NULL); 463 compile->set_type_last_size(0); 464 compile->set_last_tf(NULL, NULL); 465 compile->set_indexSet_arena(NULL); 466 compile->set_indexSet_free_block_list(NULL); 467 compile->init_type_arena(); 468 Type::Initialize(compile); 469 _compile->set_scratch_buffer_blob(NULL); 470 _compile->begin_method(); 471 } 472 CompileWrapper::~CompileWrapper() { 473 _compile->end_method(); 474 if (_compile->scratch_buffer_blob() != NULL) 475 BufferBlob::free(_compile->scratch_buffer_blob()); 476 _compile->env()->set_compiler_data(NULL); 477 } 478 479 480 //----------------------------print_compile_messages--------------------------- 481 void Compile::print_compile_messages() { 482 #ifndef PRODUCT 483 // Check if recompiling 484 if (_subsume_loads == false && PrintOpto) { 485 // Recompiling without allowing machine instructions to subsume loads 486 tty->print_cr("*********************************************************"); 487 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 488 tty->print_cr("*********************************************************"); 489 } 490 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 491 // Recompiling without escape analysis 492 tty->print_cr("*********************************************************"); 493 tty->print_cr("** Bailout: Recompile without escape analysis **"); 494 tty->print_cr("*********************************************************"); 495 } 496 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 497 // Recompiling without boxing elimination 498 tty->print_cr("*********************************************************"); 499 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 500 tty->print_cr("*********************************************************"); 501 } 502 if (env()->break_at_compile()) { 503 // Open the debugger when compiling this method. 504 tty->print("### Breaking when compiling: "); 505 method()->print_short_name(); 506 tty->cr(); 507 BREAKPOINT; 508 } 509 510 if( PrintOpto ) { 511 if (is_osr_compilation()) { 512 tty->print("[OSR]%3d", _compile_id); 513 } else { 514 tty->print("%3d", _compile_id); 515 } 516 } 517 #endif 518 } 519 520 521 //-----------------------init_scratch_buffer_blob------------------------------ 522 // Construct a temporary BufferBlob and cache it for this compile. 523 void Compile::init_scratch_buffer_blob(int const_size) { 524 // If there is already a scratch buffer blob allocated and the 525 // constant section is big enough, use it. Otherwise free the 526 // current and allocate a new one. 527 BufferBlob* blob = scratch_buffer_blob(); 528 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 529 // Use the current blob. 530 } else { 531 if (blob != NULL) { 532 BufferBlob::free(blob); 533 } 534 535 ResourceMark rm; 536 _scratch_const_size = const_size; 537 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 538 blob = BufferBlob::create("Compile::scratch_buffer", size); 539 // Record the buffer blob for next time. 540 set_scratch_buffer_blob(blob); 541 // Have we run out of code space? 542 if (scratch_buffer_blob() == NULL) { 543 // Let CompilerBroker disable further compilations. 544 record_failure("Not enough space for scratch buffer in CodeCache"); 545 return; 546 } 547 } 548 549 // Initialize the relocation buffers 550 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 551 set_scratch_locs_memory(locs_buf); 552 } 553 554 555 //-----------------------scratch_emit_size------------------------------------- 556 // Helper function that computes size by emitting code 557 uint Compile::scratch_emit_size(const Node* n) { 558 // Start scratch_emit_size section. 559 set_in_scratch_emit_size(true); 560 561 // Emit into a trash buffer and count bytes emitted. 562 // This is a pretty expensive way to compute a size, 563 // but it works well enough if seldom used. 564 // All common fixed-size instructions are given a size 565 // method by the AD file. 566 // Note that the scratch buffer blob and locs memory are 567 // allocated at the beginning of the compile task, and 568 // may be shared by several calls to scratch_emit_size. 569 // The allocation of the scratch buffer blob is particularly 570 // expensive, since it has to grab the code cache lock. 571 BufferBlob* blob = this->scratch_buffer_blob(); 572 assert(blob != NULL, "Initialize BufferBlob at start"); 573 assert(blob->size() > MAX_inst_size, "sanity"); 574 relocInfo* locs_buf = scratch_locs_memory(); 575 address blob_begin = blob->content_begin(); 576 address blob_end = (address)locs_buf; 577 assert(blob->content_contains(blob_end), "sanity"); 578 CodeBuffer buf(blob_begin, blob_end - blob_begin); 579 buf.initialize_consts_size(_scratch_const_size); 580 buf.initialize_stubs_size(MAX_stubs_size); 581 assert(locs_buf != NULL, "sanity"); 582 int lsize = MAX_locs_size / 3; 583 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 584 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 585 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 586 587 // Do the emission. 588 589 Label fakeL; // Fake label for branch instructions. 590 Label* saveL = NULL; 591 uint save_bnum = 0; 592 bool is_branch = n->is_MachBranch(); 593 if (is_branch) { 594 MacroAssembler masm(&buf); 595 masm.bind(fakeL); 596 n->as_MachBranch()->save_label(&saveL, &save_bnum); 597 n->as_MachBranch()->label_set(&fakeL, 0); 598 } 599 n->emit(buf, this->regalloc()); 600 if (is_branch) // Restore label. 601 n->as_MachBranch()->label_set(saveL, save_bnum); 602 603 // End scratch_emit_size section. 604 set_in_scratch_emit_size(false); 605 606 return buf.insts_size(); 607 } 608 609 610 // ============================================================================ 611 //------------------------------Compile standard------------------------------- 612 debug_only( int Compile::_debug_idx = 100000; ) 613 614 // Compile a method. entry_bci is -1 for normal compilations and indicates 615 // the continuation bci for on stack replacement. 616 617 618 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 619 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 620 : Phase(Compiler), 621 _env(ci_env), 622 _log(ci_env->log()), 623 _compile_id(ci_env->compile_id()), 624 _save_argument_registers(false), 625 _stub_name(NULL), 626 _stub_function(NULL), 627 _stub_entry_point(NULL), 628 _method(target), 629 _entry_bci(osr_bci), 630 _initial_gvn(NULL), 631 _for_igvn(NULL), 632 _warm_calls(NULL), 633 _subsume_loads(subsume_loads), 634 _do_escape_analysis(do_escape_analysis), 635 _eliminate_boxing(eliminate_boxing), 636 _failure_reason(NULL), 637 _code_buffer("Compile::Fill_buffer"), 638 _orig_pc_slot(0), 639 _orig_pc_slot_offset_in_bytes(0), 640 _has_method_handle_invokes(false), 641 _mach_constant_base_node(NULL), 642 _node_bundling_limit(0), 643 _node_bundling_base(NULL), 644 _java_calls(0), 645 _inner_loops(0), 646 _scratch_const_size(-1), 647 _in_scratch_emit_size(false), 648 _dead_node_list(comp_arena()), 649 _dead_node_count(0), 650 #ifndef PRODUCT 651 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 652 _in_dump_cnt(0), 653 _printer(IdealGraphPrinter::printer()), 654 #endif 655 _congraph(NULL), 656 _replay_inline_data(NULL), 657 _late_inlines(comp_arena(), 2, 0, NULL), 658 _string_late_inlines(comp_arena(), 2, 0, NULL), 659 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 660 _late_inlines_pos(0), 661 _number_of_mh_late_inlines(0), 662 _inlining_progress(false), 663 _inlining_incrementally(false), 664 _print_inlining_list(NULL), 665 _print_inlining_idx(0), 666 _preserve_jvm_state(0) { 667 C = this; 668 669 CompileWrapper cw(this); 670 #ifndef PRODUCT 671 if (TimeCompiler2) { 672 tty->print(" "); 673 target->holder()->name()->print(); 674 tty->print("."); 675 target->print_short_name(); 676 tty->print(" "); 677 } 678 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 679 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 680 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 681 if (!print_opto_assembly) { 682 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 683 if (print_assembly && !Disassembler::can_decode()) { 684 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 685 print_opto_assembly = true; 686 } 687 } 688 set_print_assembly(print_opto_assembly); 689 set_parsed_irreducible_loop(false); 690 691 if (method()->has_option("ReplayInline")) { 692 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 693 } 694 #endif 695 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 696 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 697 698 if (ProfileTraps) { 699 // Make sure the method being compiled gets its own MDO, 700 // so we can at least track the decompile_count(). 701 method()->ensure_method_data(); 702 } 703 704 Init(::AliasLevel); 705 706 707 print_compile_messages(); 708 709 _ilt = InlineTree::build_inline_tree_root(); 710 711 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 712 assert(num_alias_types() >= AliasIdxRaw, ""); 713 714 #define MINIMUM_NODE_HASH 1023 715 // Node list that Iterative GVN will start with 716 Unique_Node_List for_igvn(comp_arena()); 717 set_for_igvn(&for_igvn); 718 719 // GVN that will be run immediately on new nodes 720 uint estimated_size = method()->code_size()*4+64; 721 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 722 PhaseGVN gvn(node_arena(), estimated_size); 723 set_initial_gvn(&gvn); 724 725 if (print_inlining() || print_intrinsics()) { 726 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 727 } 728 { // Scope for timing the parser 729 TracePhase t3("parse", &_t_parser, true); 730 731 // Put top into the hash table ASAP. 732 initial_gvn()->transform_no_reclaim(top()); 733 734 // Set up tf(), start(), and find a CallGenerator. 735 CallGenerator* cg = NULL; 736 if (is_osr_compilation()) { 737 const TypeTuple *domain = StartOSRNode::osr_domain(); 738 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 739 init_tf(TypeFunc::make(domain, range)); 740 StartNode* s = new (this) StartOSRNode(root(), domain); 741 initial_gvn()->set_type_bottom(s); 742 init_start(s); 743 cg = CallGenerator::for_osr(method(), entry_bci()); 744 } else { 745 // Normal case. 746 init_tf(TypeFunc::make(method())); 747 StartNode* s = new (this) StartNode(root(), tf()->domain()); 748 initial_gvn()->set_type_bottom(s); 749 init_start(s); 750 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 751 // With java.lang.ref.reference.get() we must go through the 752 // intrinsic when G1 is enabled - even when get() is the root 753 // method of the compile - so that, if necessary, the value in 754 // the referent field of the reference object gets recorded by 755 // the pre-barrier code. 756 // Specifically, if G1 is enabled, the value in the referent 757 // field is recorded by the G1 SATB pre barrier. This will 758 // result in the referent being marked live and the reference 759 // object removed from the list of discovered references during 760 // reference processing. 761 cg = find_intrinsic(method(), false); 762 } 763 if (cg == NULL) { 764 float past_uses = method()->interpreter_invocation_count(); 765 float expected_uses = past_uses; 766 cg = CallGenerator::for_inline(method(), expected_uses); 767 } 768 } 769 if (failing()) return; 770 if (cg == NULL) { 771 record_method_not_compilable_all_tiers("cannot parse method"); 772 return; 773 } 774 JVMState* jvms = build_start_state(start(), tf()); 775 if ((jvms = cg->generate(jvms, NULL)) == NULL) { 776 record_method_not_compilable("method parse failed"); 777 return; 778 } 779 GraphKit kit(jvms); 780 781 if (!kit.stopped()) { 782 // Accept return values, and transfer control we know not where. 783 // This is done by a special, unique ReturnNode bound to root. 784 return_values(kit.jvms()); 785 } 786 787 if (kit.has_exceptions()) { 788 // Any exceptions that escape from this call must be rethrown 789 // to whatever caller is dynamically above us on the stack. 790 // This is done by a special, unique RethrowNode bound to root. 791 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 792 } 793 794 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 795 796 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 797 inline_string_calls(true); 798 } 799 800 if (failing()) return; 801 802 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 803 804 // Remove clutter produced by parsing. 805 if (!failing()) { 806 ResourceMark rm; 807 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 808 } 809 } 810 811 // Note: Large methods are capped off in do_one_bytecode(). 812 if (failing()) return; 813 814 // After parsing, node notes are no longer automagic. 815 // They must be propagated by register_new_node_with_optimizer(), 816 // clone(), or the like. 817 set_default_node_notes(NULL); 818 819 for (;;) { 820 int successes = Inline_Warm(); 821 if (failing()) return; 822 if (successes == 0) break; 823 } 824 825 // Drain the list. 826 Finish_Warm(); 827 #ifndef PRODUCT 828 if (_printer) { 829 _printer->print_inlining(this); 830 } 831 #endif 832 833 if (failing()) return; 834 NOT_PRODUCT( verify_graph_edges(); ) 835 836 // Now optimize 837 Optimize(); 838 if (failing()) return; 839 NOT_PRODUCT( verify_graph_edges(); ) 840 841 #ifndef PRODUCT 842 if (PrintIdeal) { 843 ttyLocker ttyl; // keep the following output all in one block 844 // This output goes directly to the tty, not the compiler log. 845 // To enable tools to match it up with the compilation activity, 846 // be sure to tag this tty output with the compile ID. 847 if (xtty != NULL) { 848 xtty->head("ideal compile_id='%d'%s", compile_id(), 849 is_osr_compilation() ? " compile_kind='osr'" : 850 ""); 851 } 852 root()->dump(9999); 853 if (xtty != NULL) { 854 xtty->tail("ideal"); 855 } 856 } 857 #endif 858 859 NOT_PRODUCT( verify_barriers(); ) 860 861 // Dump compilation data to replay it. 862 if (method()->has_option("DumpReplay")) { 863 env()->dump_replay_data(_compile_id); 864 } 865 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 866 env()->dump_inline_data(_compile_id); 867 } 868 869 // Now that we know the size of all the monitors we can add a fixed slot 870 // for the original deopt pc. 871 872 _orig_pc_slot = fixed_slots(); 873 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 874 set_fixed_slots(next_slot); 875 876 // Compute when to use implicit null checks. Used by matching trap based 877 // nodes and NullCheck optimization. 878 set_allowed_deopt_reasons(); 879 880 // Now generate code 881 Code_Gen(); 882 if (failing()) return; 883 884 // Check if we want to skip execution of all compiled code. 885 { 886 #ifndef PRODUCT 887 if (OptoNoExecute) { 888 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 889 return; 890 } 891 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 892 #endif 893 894 if (is_osr_compilation()) { 895 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 896 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 897 } else { 898 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 899 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 900 } 901 902 env()->register_method(_method, _entry_bci, 903 &_code_offsets, 904 _orig_pc_slot_offset_in_bytes, 905 code_buffer(), 906 frame_size_in_words(), _oop_map_set, 907 &_handler_table, &_inc_table, 908 compiler, 909 env()->comp_level(), 910 has_unsafe_access(), 911 SharedRuntime::is_wide_vector(max_vector_size()) 912 ); 913 914 if (log() != NULL) // Print code cache state into compiler log 915 log()->code_cache_state(); 916 } 917 } 918 919 //------------------------------Compile---------------------------------------- 920 // Compile a runtime stub 921 Compile::Compile( ciEnv* ci_env, 922 TypeFunc_generator generator, 923 address stub_function, 924 const char *stub_name, 925 int is_fancy_jump, 926 bool pass_tls, 927 bool save_arg_registers, 928 bool return_pc ) 929 : Phase(Compiler), 930 _env(ci_env), 931 _log(ci_env->log()), 932 _compile_id(0), 933 _save_argument_registers(save_arg_registers), 934 _method(NULL), 935 _stub_name(stub_name), 936 _stub_function(stub_function), 937 _stub_entry_point(NULL), 938 _entry_bci(InvocationEntryBci), 939 _initial_gvn(NULL), 940 _for_igvn(NULL), 941 _warm_calls(NULL), 942 _orig_pc_slot(0), 943 _orig_pc_slot_offset_in_bytes(0), 944 _subsume_loads(true), 945 _do_escape_analysis(false), 946 _eliminate_boxing(false), 947 _failure_reason(NULL), 948 _code_buffer("Compile::Fill_buffer"), 949 _has_method_handle_invokes(false), 950 _mach_constant_base_node(NULL), 951 _node_bundling_limit(0), 952 _node_bundling_base(NULL), 953 _java_calls(0), 954 _inner_loops(0), 955 #ifndef PRODUCT 956 _trace_opto_output(TraceOptoOutput), 957 _in_dump_cnt(0), 958 _printer(NULL), 959 #endif 960 _dead_node_list(comp_arena()), 961 _dead_node_count(0), 962 _congraph(NULL), 963 _replay_inline_data(NULL), 964 _number_of_mh_late_inlines(0), 965 _inlining_progress(false), 966 _inlining_incrementally(false), 967 _print_inlining_list(NULL), 968 _print_inlining_idx(0), 969 _preserve_jvm_state(0), 970 _allowed_reasons(0) { 971 C = this; 972 973 #ifndef PRODUCT 974 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 975 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 976 set_print_assembly(PrintFrameConverterAssembly); 977 set_parsed_irreducible_loop(false); 978 #endif 979 CompileWrapper cw(this); 980 Init(/*AliasLevel=*/ 0); 981 init_tf((*generator)()); 982 983 { 984 // The following is a dummy for the sake of GraphKit::gen_stub 985 Unique_Node_List for_igvn(comp_arena()); 986 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 987 PhaseGVN gvn(Thread::current()->resource_area(),255); 988 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 989 gvn.transform_no_reclaim(top()); 990 991 GraphKit kit; 992 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 993 } 994 995 NOT_PRODUCT( verify_graph_edges(); ) 996 Code_Gen(); 997 if (failing()) return; 998 999 1000 // Entry point will be accessed using compile->stub_entry_point(); 1001 if (code_buffer() == NULL) { 1002 Matcher::soft_match_failure(); 1003 } else { 1004 if (PrintAssembly && (WizardMode || Verbose)) 1005 tty->print_cr("### Stub::%s", stub_name); 1006 1007 if (!failing()) { 1008 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1009 1010 // Make the NMethod 1011 // For now we mark the frame as never safe for profile stackwalking 1012 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1013 code_buffer(), 1014 CodeOffsets::frame_never_safe, 1015 // _code_offsets.value(CodeOffsets::Frame_Complete), 1016 frame_size_in_words(), 1017 _oop_map_set, 1018 save_arg_registers); 1019 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1020 1021 _stub_entry_point = rs->entry_point(); 1022 } 1023 } 1024 } 1025 1026 //------------------------------Init------------------------------------------- 1027 // Prepare for a single compilation 1028 void Compile::Init(int aliaslevel) { 1029 _unique = 0; 1030 _regalloc = NULL; 1031 1032 _tf = NULL; // filled in later 1033 _top = NULL; // cached later 1034 _matcher = NULL; // filled in later 1035 _cfg = NULL; // filled in later 1036 1037 set_24_bit_selection_and_mode(Use24BitFP, false); 1038 1039 _node_note_array = NULL; 1040 _default_node_notes = NULL; 1041 1042 _immutable_memory = NULL; // filled in at first inquiry 1043 1044 // Globally visible Nodes 1045 // First set TOP to NULL to give safe behavior during creation of RootNode 1046 set_cached_top_node(NULL); 1047 set_root(new (this) RootNode()); 1048 // Now that you have a Root to point to, create the real TOP 1049 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1050 set_recent_alloc(NULL, NULL); 1051 1052 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1053 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1054 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1055 env()->set_dependencies(new Dependencies(env())); 1056 1057 _fixed_slots = 0; 1058 set_has_split_ifs(false); 1059 set_has_loops(has_method() && method()->has_loops()); // first approximation 1060 set_has_stringbuilder(false); 1061 set_has_boxed_value(false); 1062 _trap_can_recompile = false; // no traps emitted yet 1063 _major_progress = true; // start out assuming good things will happen 1064 set_has_unsafe_access(false); 1065 set_max_vector_size(0); 1066 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1067 set_decompile_count(0); 1068 1069 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1070 set_num_loop_opts(LoopOptsCount); 1071 set_do_inlining(Inline); 1072 set_max_inline_size(MaxInlineSize); 1073 set_freq_inline_size(FreqInlineSize); 1074 set_do_scheduling(OptoScheduling); 1075 set_do_count_invocations(false); 1076 set_do_method_data_update(false); 1077 1078 if (debug_info()->recording_non_safepoints()) { 1079 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1080 (comp_arena(), 8, 0, NULL)); 1081 set_default_node_notes(Node_Notes::make(this)); 1082 } 1083 1084 // // -- Initialize types before each compile -- 1085 // // Update cached type information 1086 // if( _method && _method->constants() ) 1087 // Type::update_loaded_types(_method, _method->constants()); 1088 1089 // Init alias_type map. 1090 if (!_do_escape_analysis && aliaslevel == 3) 1091 aliaslevel = 2; // No unique types without escape analysis 1092 _AliasLevel = aliaslevel; 1093 const int grow_ats = 16; 1094 _max_alias_types = grow_ats; 1095 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1096 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1097 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1098 { 1099 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1100 } 1101 // Initialize the first few types. 1102 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1103 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1104 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1105 _num_alias_types = AliasIdxRaw+1; 1106 // Zero out the alias type cache. 1107 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1108 // A NULL adr_type hits in the cache right away. Preload the right answer. 1109 probe_alias_cache(NULL)->_index = AliasIdxTop; 1110 1111 _intrinsics = NULL; 1112 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1113 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1114 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1115 register_library_intrinsics(); 1116 } 1117 1118 //---------------------------init_start---------------------------------------- 1119 // Install the StartNode on this compile object. 1120 void Compile::init_start(StartNode* s) { 1121 if (failing()) 1122 return; // already failing 1123 assert(s == start(), ""); 1124 } 1125 1126 StartNode* Compile::start() const { 1127 assert(!failing(), ""); 1128 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1129 Node* start = root()->fast_out(i); 1130 if( start->is_Start() ) 1131 return start->as_Start(); 1132 } 1133 ShouldNotReachHere(); 1134 return NULL; 1135 } 1136 1137 //-------------------------------immutable_memory------------------------------------- 1138 // Access immutable memory 1139 Node* Compile::immutable_memory() { 1140 if (_immutable_memory != NULL) { 1141 return _immutable_memory; 1142 } 1143 StartNode* s = start(); 1144 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1145 Node *p = s->fast_out(i); 1146 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1147 _immutable_memory = p; 1148 return _immutable_memory; 1149 } 1150 } 1151 ShouldNotReachHere(); 1152 return NULL; 1153 } 1154 1155 //----------------------set_cached_top_node------------------------------------ 1156 // Install the cached top node, and make sure Node::is_top works correctly. 1157 void Compile::set_cached_top_node(Node* tn) { 1158 if (tn != NULL) verify_top(tn); 1159 Node* old_top = _top; 1160 _top = tn; 1161 // Calling Node::setup_is_top allows the nodes the chance to adjust 1162 // their _out arrays. 1163 if (_top != NULL) _top->setup_is_top(); 1164 if (old_top != NULL) old_top->setup_is_top(); 1165 assert(_top == NULL || top()->is_top(), ""); 1166 } 1167 1168 #ifdef ASSERT 1169 uint Compile::count_live_nodes_by_graph_walk() { 1170 Unique_Node_List useful(comp_arena()); 1171 // Get useful node list by walking the graph. 1172 identify_useful_nodes(useful); 1173 return useful.size(); 1174 } 1175 1176 void Compile::print_missing_nodes() { 1177 1178 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1179 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1180 return; 1181 } 1182 1183 // This is an expensive function. It is executed only when the user 1184 // specifies VerifyIdealNodeCount option or otherwise knows the 1185 // additional work that needs to be done to identify reachable nodes 1186 // by walking the flow graph and find the missing ones using 1187 // _dead_node_list. 1188 1189 Unique_Node_List useful(comp_arena()); 1190 // Get useful node list by walking the graph. 1191 identify_useful_nodes(useful); 1192 1193 uint l_nodes = C->live_nodes(); 1194 uint l_nodes_by_walk = useful.size(); 1195 1196 if (l_nodes != l_nodes_by_walk) { 1197 if (_log != NULL) { 1198 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1199 _log->stamp(); 1200 _log->end_head(); 1201 } 1202 VectorSet& useful_member_set = useful.member_set(); 1203 int last_idx = l_nodes_by_walk; 1204 for (int i = 0; i < last_idx; i++) { 1205 if (useful_member_set.test(i)) { 1206 if (_dead_node_list.test(i)) { 1207 if (_log != NULL) { 1208 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1209 } 1210 if (PrintIdealNodeCount) { 1211 // Print the log message to tty 1212 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1213 useful.at(i)->dump(); 1214 } 1215 } 1216 } 1217 else if (! _dead_node_list.test(i)) { 1218 if (_log != NULL) { 1219 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1220 } 1221 if (PrintIdealNodeCount) { 1222 // Print the log message to tty 1223 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1224 } 1225 } 1226 } 1227 if (_log != NULL) { 1228 _log->tail("mismatched_nodes"); 1229 } 1230 } 1231 } 1232 #endif 1233 1234 #ifndef PRODUCT 1235 void Compile::verify_top(Node* tn) const { 1236 if (tn != NULL) { 1237 assert(tn->is_Con(), "top node must be a constant"); 1238 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1239 assert(tn->in(0) != NULL, "must have live top node"); 1240 } 1241 } 1242 #endif 1243 1244 1245 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1246 1247 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1248 guarantee(arr != NULL, ""); 1249 int num_blocks = arr->length(); 1250 if (grow_by < num_blocks) grow_by = num_blocks; 1251 int num_notes = grow_by * _node_notes_block_size; 1252 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1253 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1254 while (num_notes > 0) { 1255 arr->append(notes); 1256 notes += _node_notes_block_size; 1257 num_notes -= _node_notes_block_size; 1258 } 1259 assert(num_notes == 0, "exact multiple, please"); 1260 } 1261 1262 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1263 if (source == NULL || dest == NULL) return false; 1264 1265 if (dest->is_Con()) 1266 return false; // Do not push debug info onto constants. 1267 1268 #ifdef ASSERT 1269 // Leave a bread crumb trail pointing to the original node: 1270 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1271 dest->set_debug_orig(source); 1272 } 1273 #endif 1274 1275 if (node_note_array() == NULL) 1276 return false; // Not collecting any notes now. 1277 1278 // This is a copy onto a pre-existing node, which may already have notes. 1279 // If both nodes have notes, do not overwrite any pre-existing notes. 1280 Node_Notes* source_notes = node_notes_at(source->_idx); 1281 if (source_notes == NULL || source_notes->is_clear()) return false; 1282 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1283 if (dest_notes == NULL || dest_notes->is_clear()) { 1284 return set_node_notes_at(dest->_idx, source_notes); 1285 } 1286 1287 Node_Notes merged_notes = (*source_notes); 1288 // The order of operations here ensures that dest notes will win... 1289 merged_notes.update_from(dest_notes); 1290 return set_node_notes_at(dest->_idx, &merged_notes); 1291 } 1292 1293 1294 //--------------------------allow_range_check_smearing------------------------- 1295 // Gating condition for coalescing similar range checks. 1296 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1297 // single covering check that is at least as strong as any of them. 1298 // If the optimization succeeds, the simplified (strengthened) range check 1299 // will always succeed. If it fails, we will deopt, and then give up 1300 // on the optimization. 1301 bool Compile::allow_range_check_smearing() const { 1302 // If this method has already thrown a range-check, 1303 // assume it was because we already tried range smearing 1304 // and it failed. 1305 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1306 return !already_trapped; 1307 } 1308 1309 1310 //------------------------------flatten_alias_type----------------------------- 1311 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1312 int offset = tj->offset(); 1313 TypePtr::PTR ptr = tj->ptr(); 1314 1315 // Known instance (scalarizable allocation) alias only with itself. 1316 bool is_known_inst = tj->isa_oopptr() != NULL && 1317 tj->is_oopptr()->is_known_instance(); 1318 1319 // Process weird unsafe references. 1320 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1321 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1322 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1323 tj = TypeOopPtr::BOTTOM; 1324 ptr = tj->ptr(); 1325 offset = tj->offset(); 1326 } 1327 1328 // Array pointers need some flattening 1329 const TypeAryPtr *ta = tj->isa_aryptr(); 1330 if (ta && ta->is_stable()) { 1331 // Erase stability property for alias analysis. 1332 tj = ta = ta->cast_to_stable(false); 1333 } 1334 if( ta && is_known_inst ) { 1335 if ( offset != Type::OffsetBot && 1336 offset > arrayOopDesc::length_offset_in_bytes() ) { 1337 offset = Type::OffsetBot; // Flatten constant access into array body only 1338 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1339 } 1340 } else if( ta && _AliasLevel >= 2 ) { 1341 // For arrays indexed by constant indices, we flatten the alias 1342 // space to include all of the array body. Only the header, klass 1343 // and array length can be accessed un-aliased. 1344 if( offset != Type::OffsetBot ) { 1345 if( ta->const_oop() ) { // MethodData* or Method* 1346 offset = Type::OffsetBot; // Flatten constant access into array body 1347 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1348 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1349 // range is OK as-is. 1350 tj = ta = TypeAryPtr::RANGE; 1351 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1352 tj = TypeInstPtr::KLASS; // all klass loads look alike 1353 ta = TypeAryPtr::RANGE; // generic ignored junk 1354 ptr = TypePtr::BotPTR; 1355 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1356 tj = TypeInstPtr::MARK; 1357 ta = TypeAryPtr::RANGE; // generic ignored junk 1358 ptr = TypePtr::BotPTR; 1359 } else { // Random constant offset into array body 1360 offset = Type::OffsetBot; // Flatten constant access into array body 1361 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1362 } 1363 } 1364 // Arrays of fixed size alias with arrays of unknown size. 1365 if (ta->size() != TypeInt::POS) { 1366 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1367 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1368 } 1369 // Arrays of known objects become arrays of unknown objects. 1370 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1371 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1372 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1373 } 1374 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1375 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1376 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1377 } 1378 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1379 // cannot be distinguished by bytecode alone. 1380 if (ta->elem() == TypeInt::BOOL) { 1381 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1382 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1383 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1384 } 1385 // During the 2nd round of IterGVN, NotNull castings are removed. 1386 // Make sure the Bottom and NotNull variants alias the same. 1387 // Also, make sure exact and non-exact variants alias the same. 1388 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1389 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1390 } 1391 } 1392 1393 // Oop pointers need some flattening 1394 const TypeInstPtr *to = tj->isa_instptr(); 1395 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1396 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1397 if( ptr == TypePtr::Constant ) { 1398 if (to->klass() != ciEnv::current()->Class_klass() || 1399 offset < k->size_helper() * wordSize) { 1400 // No constant oop pointers (such as Strings); they alias with 1401 // unknown strings. 1402 assert(!is_known_inst, "not scalarizable allocation"); 1403 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1404 } 1405 } else if( is_known_inst ) { 1406 tj = to; // Keep NotNull and klass_is_exact for instance type 1407 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1408 // During the 2nd round of IterGVN, NotNull castings are removed. 1409 // Make sure the Bottom and NotNull variants alias the same. 1410 // Also, make sure exact and non-exact variants alias the same. 1411 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1412 } 1413 if (to->speculative() != NULL) { 1414 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1415 } 1416 // Canonicalize the holder of this field 1417 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1418 // First handle header references such as a LoadKlassNode, even if the 1419 // object's klass is unloaded at compile time (4965979). 1420 if (!is_known_inst) { // Do it only for non-instance types 1421 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1422 } 1423 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1424 // Static fields are in the space above the normal instance 1425 // fields in the java.lang.Class instance. 1426 if (to->klass() != ciEnv::current()->Class_klass()) { 1427 to = NULL; 1428 tj = TypeOopPtr::BOTTOM; 1429 offset = tj->offset(); 1430 } 1431 } else { 1432 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1433 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1434 if( is_known_inst ) { 1435 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1436 } else { 1437 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1438 } 1439 } 1440 } 1441 } 1442 1443 // Klass pointers to object array klasses need some flattening 1444 const TypeKlassPtr *tk = tj->isa_klassptr(); 1445 if( tk ) { 1446 // If we are referencing a field within a Klass, we need 1447 // to assume the worst case of an Object. Both exact and 1448 // inexact types must flatten to the same alias class so 1449 // use NotNull as the PTR. 1450 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1451 1452 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1453 TypeKlassPtr::OBJECT->klass(), 1454 offset); 1455 } 1456 1457 ciKlass* klass = tk->klass(); 1458 if( klass->is_obj_array_klass() ) { 1459 ciKlass* k = TypeAryPtr::OOPS->klass(); 1460 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1461 k = TypeInstPtr::BOTTOM->klass(); 1462 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1463 } 1464 1465 // Check for precise loads from the primary supertype array and force them 1466 // to the supertype cache alias index. Check for generic array loads from 1467 // the primary supertype array and also force them to the supertype cache 1468 // alias index. Since the same load can reach both, we need to merge 1469 // these 2 disparate memories into the same alias class. Since the 1470 // primary supertype array is read-only, there's no chance of confusion 1471 // where we bypass an array load and an array store. 1472 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1473 if (offset == Type::OffsetBot || 1474 (offset >= primary_supers_offset && 1475 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1476 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1477 offset = in_bytes(Klass::secondary_super_cache_offset()); 1478 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1479 } 1480 } 1481 1482 // Flatten all Raw pointers together. 1483 if (tj->base() == Type::RawPtr) 1484 tj = TypeRawPtr::BOTTOM; 1485 1486 if (tj->base() == Type::AnyPtr) 1487 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1488 1489 // Flatten all to bottom for now 1490 switch( _AliasLevel ) { 1491 case 0: 1492 tj = TypePtr::BOTTOM; 1493 break; 1494 case 1: // Flatten to: oop, static, field or array 1495 switch (tj->base()) { 1496 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1497 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1498 case Type::AryPtr: // do not distinguish arrays at all 1499 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1500 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1501 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1502 default: ShouldNotReachHere(); 1503 } 1504 break; 1505 case 2: // No collapsing at level 2; keep all splits 1506 case 3: // No collapsing at level 3; keep all splits 1507 break; 1508 default: 1509 Unimplemented(); 1510 } 1511 1512 offset = tj->offset(); 1513 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1514 1515 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1516 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1517 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1518 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1519 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1520 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1521 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1522 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1523 assert( tj->ptr() != TypePtr::TopPTR && 1524 tj->ptr() != TypePtr::AnyNull && 1525 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1526 // assert( tj->ptr() != TypePtr::Constant || 1527 // tj->base() == Type::RawPtr || 1528 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1529 1530 return tj; 1531 } 1532 1533 void Compile::AliasType::Init(int i, const TypePtr* at) { 1534 _index = i; 1535 _adr_type = at; 1536 _field = NULL; 1537 _element = NULL; 1538 _is_rewritable = true; // default 1539 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1540 if (atoop != NULL && atoop->is_known_instance()) { 1541 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1542 _general_index = Compile::current()->get_alias_index(gt); 1543 } else { 1544 _general_index = 0; 1545 } 1546 } 1547 1548 //---------------------------------print_on------------------------------------ 1549 #ifndef PRODUCT 1550 void Compile::AliasType::print_on(outputStream* st) { 1551 if (index() < 10) 1552 st->print("@ <%d> ", index()); 1553 else st->print("@ <%d>", index()); 1554 st->print(is_rewritable() ? " " : " RO"); 1555 int offset = adr_type()->offset(); 1556 if (offset == Type::OffsetBot) 1557 st->print(" +any"); 1558 else st->print(" +%-3d", offset); 1559 st->print(" in "); 1560 adr_type()->dump_on(st); 1561 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1562 if (field() != NULL && tjp) { 1563 if (tjp->klass() != field()->holder() || 1564 tjp->offset() != field()->offset_in_bytes()) { 1565 st->print(" != "); 1566 field()->print(); 1567 st->print(" ***"); 1568 } 1569 } 1570 } 1571 1572 void print_alias_types() { 1573 Compile* C = Compile::current(); 1574 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1575 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1576 C->alias_type(idx)->print_on(tty); 1577 tty->cr(); 1578 } 1579 } 1580 #endif 1581 1582 1583 //----------------------------probe_alias_cache-------------------------------- 1584 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1585 intptr_t key = (intptr_t) adr_type; 1586 key ^= key >> logAliasCacheSize; 1587 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1588 } 1589 1590 1591 //-----------------------------grow_alias_types-------------------------------- 1592 void Compile::grow_alias_types() { 1593 const int old_ats = _max_alias_types; // how many before? 1594 const int new_ats = old_ats; // how many more? 1595 const int grow_ats = old_ats+new_ats; // how many now? 1596 _max_alias_types = grow_ats; 1597 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1598 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1599 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1600 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1601 } 1602 1603 1604 //--------------------------------find_alias_type------------------------------ 1605 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1606 if (_AliasLevel == 0) 1607 return alias_type(AliasIdxBot); 1608 1609 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1610 if (ace->_adr_type == adr_type) { 1611 return alias_type(ace->_index); 1612 } 1613 1614 // Handle special cases. 1615 if (adr_type == NULL) return alias_type(AliasIdxTop); 1616 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1617 1618 // Do it the slow way. 1619 const TypePtr* flat = flatten_alias_type(adr_type); 1620 1621 #ifdef ASSERT 1622 assert(flat == flatten_alias_type(flat), "idempotent"); 1623 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1624 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1625 const TypeOopPtr* foop = flat->is_oopptr(); 1626 // Scalarizable allocations have exact klass always. 1627 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1628 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1629 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1630 } 1631 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1632 #endif 1633 1634 int idx = AliasIdxTop; 1635 for (int i = 0; i < num_alias_types(); i++) { 1636 if (alias_type(i)->adr_type() == flat) { 1637 idx = i; 1638 break; 1639 } 1640 } 1641 1642 if (idx == AliasIdxTop) { 1643 if (no_create) return NULL; 1644 // Grow the array if necessary. 1645 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1646 // Add a new alias type. 1647 idx = _num_alias_types++; 1648 _alias_types[idx]->Init(idx, flat); 1649 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1650 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1651 if (flat->isa_instptr()) { 1652 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1653 && flat->is_instptr()->klass() == env()->Class_klass()) 1654 alias_type(idx)->set_rewritable(false); 1655 } 1656 if (flat->isa_aryptr()) { 1657 #ifdef ASSERT 1658 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1659 // (T_BYTE has the weakest alignment and size restrictions...) 1660 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1661 #endif 1662 if (flat->offset() == TypePtr::OffsetBot) { 1663 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1664 } 1665 } 1666 if (flat->isa_klassptr()) { 1667 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1668 alias_type(idx)->set_rewritable(false); 1669 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1670 alias_type(idx)->set_rewritable(false); 1671 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1672 alias_type(idx)->set_rewritable(false); 1673 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1674 alias_type(idx)->set_rewritable(false); 1675 } 1676 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1677 // but the base pointer type is not distinctive enough to identify 1678 // references into JavaThread.) 1679 1680 // Check for final fields. 1681 const TypeInstPtr* tinst = flat->isa_instptr(); 1682 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1683 ciField* field; 1684 if (tinst->const_oop() != NULL && 1685 tinst->klass() == ciEnv::current()->Class_klass() && 1686 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1687 // static field 1688 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1689 field = k->get_field_by_offset(tinst->offset(), true); 1690 } else { 1691 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1692 field = k->get_field_by_offset(tinst->offset(), false); 1693 } 1694 assert(field == NULL || 1695 original_field == NULL || 1696 (field->holder() == original_field->holder() && 1697 field->offset() == original_field->offset() && 1698 field->is_static() == original_field->is_static()), "wrong field?"); 1699 // Set field() and is_rewritable() attributes. 1700 if (field != NULL) alias_type(idx)->set_field(field); 1701 } 1702 } 1703 1704 // Fill the cache for next time. 1705 ace->_adr_type = adr_type; 1706 ace->_index = idx; 1707 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1708 1709 // Might as well try to fill the cache for the flattened version, too. 1710 AliasCacheEntry* face = probe_alias_cache(flat); 1711 if (face->_adr_type == NULL) { 1712 face->_adr_type = flat; 1713 face->_index = idx; 1714 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1715 } 1716 1717 return alias_type(idx); 1718 } 1719 1720 1721 Compile::AliasType* Compile::alias_type(ciField* field) { 1722 const TypeOopPtr* t; 1723 if (field->is_static()) 1724 t = TypeInstPtr::make(field->holder()->java_mirror()); 1725 else 1726 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1727 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1728 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1729 return atp; 1730 } 1731 1732 1733 //------------------------------have_alias_type-------------------------------- 1734 bool Compile::have_alias_type(const TypePtr* adr_type) { 1735 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1736 if (ace->_adr_type == adr_type) { 1737 return true; 1738 } 1739 1740 // Handle special cases. 1741 if (adr_type == NULL) return true; 1742 if (adr_type == TypePtr::BOTTOM) return true; 1743 1744 return find_alias_type(adr_type, true, NULL) != NULL; 1745 } 1746 1747 //-----------------------------must_alias-------------------------------------- 1748 // True if all values of the given address type are in the given alias category. 1749 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1750 if (alias_idx == AliasIdxBot) return true; // the universal category 1751 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1752 if (alias_idx == AliasIdxTop) return false; // the empty category 1753 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1754 1755 // the only remaining possible overlap is identity 1756 int adr_idx = get_alias_index(adr_type); 1757 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1758 assert(adr_idx == alias_idx || 1759 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1760 && adr_type != TypeOopPtr::BOTTOM), 1761 "should not be testing for overlap with an unsafe pointer"); 1762 return adr_idx == alias_idx; 1763 } 1764 1765 //------------------------------can_alias-------------------------------------- 1766 // True if any values of the given address type are in the given alias category. 1767 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1768 if (alias_idx == AliasIdxTop) return false; // the empty category 1769 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1770 if (alias_idx == AliasIdxBot) return true; // the universal category 1771 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1772 1773 // the only remaining possible overlap is identity 1774 int adr_idx = get_alias_index(adr_type); 1775 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1776 return adr_idx == alias_idx; 1777 } 1778 1779 1780 1781 //---------------------------pop_warm_call------------------------------------- 1782 WarmCallInfo* Compile::pop_warm_call() { 1783 WarmCallInfo* wci = _warm_calls; 1784 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1785 return wci; 1786 } 1787 1788 //----------------------------Inline_Warm-------------------------------------- 1789 int Compile::Inline_Warm() { 1790 // If there is room, try to inline some more warm call sites. 1791 // %%% Do a graph index compaction pass when we think we're out of space? 1792 if (!InlineWarmCalls) return 0; 1793 1794 int calls_made_hot = 0; 1795 int room_to_grow = NodeCountInliningCutoff - unique(); 1796 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1797 int amount_grown = 0; 1798 WarmCallInfo* call; 1799 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1800 int est_size = (int)call->size(); 1801 if (est_size > (room_to_grow - amount_grown)) { 1802 // This one won't fit anyway. Get rid of it. 1803 call->make_cold(); 1804 continue; 1805 } 1806 call->make_hot(); 1807 calls_made_hot++; 1808 amount_grown += est_size; 1809 amount_to_grow -= est_size; 1810 } 1811 1812 if (calls_made_hot > 0) set_major_progress(); 1813 return calls_made_hot; 1814 } 1815 1816 1817 //----------------------------Finish_Warm-------------------------------------- 1818 void Compile::Finish_Warm() { 1819 if (!InlineWarmCalls) return; 1820 if (failing()) return; 1821 if (warm_calls() == NULL) return; 1822 1823 // Clean up loose ends, if we are out of space for inlining. 1824 WarmCallInfo* call; 1825 while ((call = pop_warm_call()) != NULL) { 1826 call->make_cold(); 1827 } 1828 } 1829 1830 //---------------------cleanup_loop_predicates----------------------- 1831 // Remove the opaque nodes that protect the predicates so that all unused 1832 // checks and uncommon_traps will be eliminated from the ideal graph 1833 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1834 if (predicate_count()==0) return; 1835 for (int i = predicate_count(); i > 0; i--) { 1836 Node * n = predicate_opaque1_node(i-1); 1837 assert(n->Opcode() == Op_Opaque1, "must be"); 1838 igvn.replace_node(n, n->in(1)); 1839 } 1840 assert(predicate_count()==0, "should be clean!"); 1841 } 1842 1843 // StringOpts and late inlining of string methods 1844 void Compile::inline_string_calls(bool parse_time) { 1845 { 1846 // remove useless nodes to make the usage analysis simpler 1847 ResourceMark rm; 1848 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1849 } 1850 1851 { 1852 ResourceMark rm; 1853 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1854 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1855 print_method(PHASE_AFTER_STRINGOPTS, 3); 1856 } 1857 1858 // now inline anything that we skipped the first time around 1859 if (!parse_time) { 1860 _late_inlines_pos = _late_inlines.length(); 1861 } 1862 1863 while (_string_late_inlines.length() > 0) { 1864 CallGenerator* cg = _string_late_inlines.pop(); 1865 cg->do_late_inline(); 1866 if (failing()) return; 1867 } 1868 _string_late_inlines.trunc_to(0); 1869 } 1870 1871 // Late inlining of boxing methods 1872 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1873 if (_boxing_late_inlines.length() > 0) { 1874 assert(has_boxed_value(), "inconsistent"); 1875 1876 PhaseGVN* gvn = initial_gvn(); 1877 set_inlining_incrementally(true); 1878 1879 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1880 for_igvn()->clear(); 1881 gvn->replace_with(&igvn); 1882 1883 while (_boxing_late_inlines.length() > 0) { 1884 CallGenerator* cg = _boxing_late_inlines.pop(); 1885 cg->do_late_inline(); 1886 if (failing()) return; 1887 } 1888 _boxing_late_inlines.trunc_to(0); 1889 1890 { 1891 ResourceMark rm; 1892 PhaseRemoveUseless pru(gvn, for_igvn()); 1893 } 1894 1895 igvn = PhaseIterGVN(gvn); 1896 igvn.optimize(); 1897 1898 set_inlining_progress(false); 1899 set_inlining_incrementally(false); 1900 } 1901 } 1902 1903 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1904 assert(IncrementalInline, "incremental inlining should be on"); 1905 PhaseGVN* gvn = initial_gvn(); 1906 1907 set_inlining_progress(false); 1908 for_igvn()->clear(); 1909 gvn->replace_with(&igvn); 1910 1911 int i = 0; 1912 1913 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1914 CallGenerator* cg = _late_inlines.at(i); 1915 _late_inlines_pos = i+1; 1916 cg->do_late_inline(); 1917 if (failing()) return; 1918 } 1919 int j = 0; 1920 for (; i < _late_inlines.length(); i++, j++) { 1921 _late_inlines.at_put(j, _late_inlines.at(i)); 1922 } 1923 _late_inlines.trunc_to(j); 1924 1925 { 1926 ResourceMark rm; 1927 PhaseRemoveUseless pru(gvn, for_igvn()); 1928 } 1929 1930 igvn = PhaseIterGVN(gvn); 1931 } 1932 1933 // Perform incremental inlining until bound on number of live nodes is reached 1934 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1935 PhaseGVN* gvn = initial_gvn(); 1936 1937 set_inlining_incrementally(true); 1938 set_inlining_progress(true); 1939 uint low_live_nodes = 0; 1940 1941 while(inlining_progress() && _late_inlines.length() > 0) { 1942 1943 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1944 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1945 // PhaseIdealLoop is expensive so we only try it once we are 1946 // out of loop and we only try it again if the previous helped 1947 // got the number of nodes down significantly 1948 PhaseIdealLoop ideal_loop( igvn, false, true ); 1949 if (failing()) return; 1950 low_live_nodes = live_nodes(); 1951 _major_progress = true; 1952 } 1953 1954 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1955 break; 1956 } 1957 } 1958 1959 inline_incrementally_one(igvn); 1960 1961 if (failing()) return; 1962 1963 igvn.optimize(); 1964 1965 if (failing()) return; 1966 } 1967 1968 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1969 1970 if (_string_late_inlines.length() > 0) { 1971 assert(has_stringbuilder(), "inconsistent"); 1972 for_igvn()->clear(); 1973 initial_gvn()->replace_with(&igvn); 1974 1975 inline_string_calls(false); 1976 1977 if (failing()) return; 1978 1979 { 1980 ResourceMark rm; 1981 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1982 } 1983 1984 igvn = PhaseIterGVN(gvn); 1985 1986 igvn.optimize(); 1987 } 1988 1989 set_inlining_incrementally(false); 1990 } 1991 1992 1993 //------------------------------Optimize--------------------------------------- 1994 // Given a graph, optimize it. 1995 void Compile::Optimize() { 1996 TracePhase t1("optimizer", &_t_optimizer, true); 1997 1998 #ifndef PRODUCT 1999 if (env()->break_at_compile()) { 2000 BREAKPOINT; 2001 } 2002 2003 #endif 2004 2005 ResourceMark rm; 2006 int loop_opts_cnt; 2007 2008 NOT_PRODUCT( verify_graph_edges(); ) 2009 2010 print_method(PHASE_AFTER_PARSING); 2011 2012 { 2013 // Iterative Global Value Numbering, including ideal transforms 2014 // Initialize IterGVN with types and values from parse-time GVN 2015 PhaseIterGVN igvn(initial_gvn()); 2016 { 2017 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2018 igvn.optimize(); 2019 } 2020 2021 print_method(PHASE_ITER_GVN1, 2); 2022 2023 if (failing()) return; 2024 2025 { 2026 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2027 inline_incrementally(igvn); 2028 } 2029 2030 print_method(PHASE_INCREMENTAL_INLINE, 2); 2031 2032 if (failing()) return; 2033 2034 if (eliminate_boxing()) { 2035 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2036 // Inline valueOf() methods now. 2037 inline_boxing_calls(igvn); 2038 2039 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2040 2041 if (failing()) return; 2042 } 2043 2044 // Remove the speculative part of types and clean up the graph from 2045 // the extra CastPP nodes whose only purpose is to carry them. Do 2046 // that early so that optimizations are not disrupted by the extra 2047 // CastPP nodes. 2048 remove_speculative_types(igvn); 2049 2050 // No more new expensive nodes will be added to the list from here 2051 // so keep only the actual candidates for optimizations. 2052 cleanup_expensive_nodes(igvn); 2053 2054 // Perform escape analysis 2055 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2056 if (has_loops()) { 2057 // Cleanup graph (remove dead nodes). 2058 TracePhase t2("idealLoop", &_t_idealLoop, true); 2059 PhaseIdealLoop ideal_loop( igvn, false, true ); 2060 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2061 if (failing()) return; 2062 } 2063 ConnectionGraph::do_analysis(this, &igvn); 2064 2065 if (failing()) return; 2066 2067 // Optimize out fields loads from scalar replaceable allocations. 2068 igvn.optimize(); 2069 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2070 2071 if (failing()) return; 2072 2073 if (congraph() != NULL && macro_count() > 0) { 2074 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2075 PhaseMacroExpand mexp(igvn); 2076 mexp.eliminate_macro_nodes(); 2077 igvn.set_delay_transform(false); 2078 2079 igvn.optimize(); 2080 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2081 2082 if (failing()) return; 2083 } 2084 } 2085 2086 // Loop transforms on the ideal graph. Range Check Elimination, 2087 // peeling, unrolling, etc. 2088 2089 // Set loop opts counter 2090 loop_opts_cnt = num_loop_opts(); 2091 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2092 { 2093 TracePhase t2("idealLoop", &_t_idealLoop, true); 2094 PhaseIdealLoop ideal_loop( igvn, true ); 2095 loop_opts_cnt--; 2096 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2097 if (failing()) return; 2098 } 2099 // Loop opts pass if partial peeling occurred in previous pass 2100 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2101 TracePhase t3("idealLoop", &_t_idealLoop, true); 2102 PhaseIdealLoop ideal_loop( igvn, false ); 2103 loop_opts_cnt--; 2104 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2105 if (failing()) return; 2106 } 2107 // Loop opts pass for loop-unrolling before CCP 2108 if(major_progress() && (loop_opts_cnt > 0)) { 2109 TracePhase t4("idealLoop", &_t_idealLoop, true); 2110 PhaseIdealLoop ideal_loop( igvn, false ); 2111 loop_opts_cnt--; 2112 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2113 } 2114 if (!failing()) { 2115 // Verify that last round of loop opts produced a valid graph 2116 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2117 PhaseIdealLoop::verify(igvn); 2118 } 2119 } 2120 if (failing()) return; 2121 2122 // Conditional Constant Propagation; 2123 PhaseCCP ccp( &igvn ); 2124 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2125 { 2126 TracePhase t2("ccp", &_t_ccp, true); 2127 ccp.do_transform(); 2128 } 2129 print_method(PHASE_CPP1, 2); 2130 2131 assert( true, "Break here to ccp.dump_old2new_map()"); 2132 2133 // Iterative Global Value Numbering, including ideal transforms 2134 { 2135 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2136 igvn = ccp; 2137 igvn.optimize(); 2138 } 2139 2140 print_method(PHASE_ITER_GVN2, 2); 2141 2142 if (failing()) return; 2143 2144 // Loop transforms on the ideal graph. Range Check Elimination, 2145 // peeling, unrolling, etc. 2146 if(loop_opts_cnt > 0) { 2147 debug_only( int cnt = 0; ); 2148 while(major_progress() && (loop_opts_cnt > 0)) { 2149 TracePhase t2("idealLoop", &_t_idealLoop, true); 2150 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2151 PhaseIdealLoop ideal_loop( igvn, true); 2152 loop_opts_cnt--; 2153 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2154 if (failing()) return; 2155 } 2156 } 2157 2158 { 2159 // Verify that all previous optimizations produced a valid graph 2160 // at least to this point, even if no loop optimizations were done. 2161 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2162 PhaseIdealLoop::verify(igvn); 2163 } 2164 2165 { 2166 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2167 PhaseMacroExpand mex(igvn); 2168 if (mex.expand_macro_nodes()) { 2169 assert(failing(), "must bail out w/ explicit message"); 2170 return; 2171 } 2172 } 2173 2174 } // (End scope of igvn; run destructor if necessary for asserts.) 2175 2176 dump_inlining(); 2177 // A method with only infinite loops has no edges entering loops from root 2178 { 2179 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2180 if (final_graph_reshaping()) { 2181 assert(failing(), "must bail out w/ explicit message"); 2182 return; 2183 } 2184 } 2185 2186 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2187 } 2188 2189 2190 //------------------------------Code_Gen--------------------------------------- 2191 // Given a graph, generate code for it 2192 void Compile::Code_Gen() { 2193 if (failing()) { 2194 return; 2195 } 2196 2197 // Perform instruction selection. You might think we could reclaim Matcher 2198 // memory PDQ, but actually the Matcher is used in generating spill code. 2199 // Internals of the Matcher (including some VectorSets) must remain live 2200 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2201 // set a bit in reclaimed memory. 2202 2203 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2204 // nodes. Mapping is only valid at the root of each matched subtree. 2205 NOT_PRODUCT( verify_graph_edges(); ) 2206 2207 Matcher matcher; 2208 _matcher = &matcher; 2209 { 2210 TracePhase t2("matcher", &_t_matcher, true); 2211 matcher.match(); 2212 } 2213 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2214 // nodes. Mapping is only valid at the root of each matched subtree. 2215 NOT_PRODUCT( verify_graph_edges(); ) 2216 2217 // If you have too many nodes, or if matching has failed, bail out 2218 check_node_count(0, "out of nodes matching instructions"); 2219 if (failing()) { 2220 return; 2221 } 2222 2223 // Build a proper-looking CFG 2224 PhaseCFG cfg(node_arena(), root(), matcher); 2225 _cfg = &cfg; 2226 { 2227 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2228 bool success = cfg.do_global_code_motion(); 2229 if (!success) { 2230 return; 2231 } 2232 2233 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2234 NOT_PRODUCT( verify_graph_edges(); ) 2235 debug_only( cfg.verify(); ) 2236 } 2237 2238 PhaseChaitin regalloc(unique(), cfg, matcher); 2239 _regalloc = ®alloc; 2240 { 2241 TracePhase t2("regalloc", &_t_registerAllocation, true); 2242 // Perform register allocation. After Chaitin, use-def chains are 2243 // no longer accurate (at spill code) and so must be ignored. 2244 // Node->LRG->reg mappings are still accurate. 2245 _regalloc->Register_Allocate(); 2246 2247 // Bail out if the allocator builds too many nodes 2248 if (failing()) { 2249 return; 2250 } 2251 } 2252 2253 // Prior to register allocation we kept empty basic blocks in case the 2254 // the allocator needed a place to spill. After register allocation we 2255 // are not adding any new instructions. If any basic block is empty, we 2256 // can now safely remove it. 2257 { 2258 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2259 cfg.remove_empty_blocks(); 2260 if (do_freq_based_layout()) { 2261 PhaseBlockLayout layout(cfg); 2262 } else { 2263 cfg.set_loop_alignment(); 2264 } 2265 cfg.fixup_flow(); 2266 } 2267 2268 // Apply peephole optimizations 2269 if( OptoPeephole ) { 2270 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2271 PhasePeephole peep( _regalloc, cfg); 2272 peep.do_transform(); 2273 } 2274 2275 // Do late expand if CPU requires this. 2276 if (Matcher::require_postalloc_expand) { 2277 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2278 cfg.postalloc_expand(_regalloc); 2279 } 2280 2281 // Convert Nodes to instruction bits in a buffer 2282 { 2283 // %%%% workspace merge brought two timers together for one job 2284 TracePhase t2a("output", &_t_output, true); 2285 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2286 Output(); 2287 } 2288 2289 print_method(PHASE_FINAL_CODE); 2290 2291 // He's dead, Jim. 2292 _cfg = (PhaseCFG*)0xdeadbeef; 2293 _regalloc = (PhaseChaitin*)0xdeadbeef; 2294 } 2295 2296 2297 //------------------------------dump_asm--------------------------------------- 2298 // Dump formatted assembly 2299 #ifndef PRODUCT 2300 void Compile::dump_asm(int *pcs, uint pc_limit) { 2301 bool cut_short = false; 2302 tty->print_cr("#"); 2303 tty->print("# "); _tf->dump(); tty->cr(); 2304 tty->print_cr("#"); 2305 2306 // For all blocks 2307 int pc = 0x0; // Program counter 2308 char starts_bundle = ' '; 2309 _regalloc->dump_frame(); 2310 2311 Node *n = NULL; 2312 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2313 if (VMThread::should_terminate()) { 2314 cut_short = true; 2315 break; 2316 } 2317 Block* block = _cfg->get_block(i); 2318 if (block->is_connector() && !Verbose) { 2319 continue; 2320 } 2321 n = block->head(); 2322 if (pcs && n->_idx < pc_limit) { 2323 tty->print("%3.3x ", pcs[n->_idx]); 2324 } else { 2325 tty->print(" "); 2326 } 2327 block->dump_head(_cfg); 2328 if (block->is_connector()) { 2329 tty->print_cr(" # Empty connector block"); 2330 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2331 tty->print_cr(" # Block is sole successor of call"); 2332 } 2333 2334 // For all instructions 2335 Node *delay = NULL; 2336 for (uint j = 0; j < block->number_of_nodes(); j++) { 2337 if (VMThread::should_terminate()) { 2338 cut_short = true; 2339 break; 2340 } 2341 n = block->get_node(j); 2342 if (valid_bundle_info(n)) { 2343 Bundle* bundle = node_bundling(n); 2344 if (bundle->used_in_unconditional_delay()) { 2345 delay = n; 2346 continue; 2347 } 2348 if (bundle->starts_bundle()) { 2349 starts_bundle = '+'; 2350 } 2351 } 2352 2353 if (WizardMode) { 2354 n->dump(); 2355 } 2356 2357 if( !n->is_Region() && // Dont print in the Assembly 2358 !n->is_Phi() && // a few noisely useless nodes 2359 !n->is_Proj() && 2360 !n->is_MachTemp() && 2361 !n->is_SafePointScalarObject() && 2362 !n->is_Catch() && // Would be nice to print exception table targets 2363 !n->is_MergeMem() && // Not very interesting 2364 !n->is_top() && // Debug info table constants 2365 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2366 ) { 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 n->format(_regalloc, tty); 2375 tty->cr(); 2376 } 2377 2378 // If we have an instruction with a delay slot, and have seen a delay, 2379 // then back up and print it 2380 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2381 assert(delay != NULL, "no unconditional delay instruction"); 2382 if (WizardMode) delay->dump(); 2383 2384 if (node_bundling(delay)->starts_bundle()) 2385 starts_bundle = '+'; 2386 if (pcs && n->_idx < pc_limit) 2387 tty->print("%3.3x", pcs[n->_idx]); 2388 else 2389 tty->print(" "); 2390 tty->print(" %c ", starts_bundle); 2391 starts_bundle = ' '; 2392 tty->print("\t"); 2393 delay->format(_regalloc, tty); 2394 tty->print_cr(""); 2395 delay = NULL; 2396 } 2397 2398 // Dump the exception table as well 2399 if( n->is_Catch() && (Verbose || WizardMode) ) { 2400 // Print the exception table for this offset 2401 _handler_table.print_subtable_for(pc); 2402 } 2403 } 2404 2405 if (pcs && n->_idx < pc_limit) 2406 tty->print_cr("%3.3x", pcs[n->_idx]); 2407 else 2408 tty->print_cr(""); 2409 2410 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2411 2412 } // End of per-block dump 2413 tty->print_cr(""); 2414 2415 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2416 } 2417 #endif 2418 2419 //------------------------------Final_Reshape_Counts--------------------------- 2420 // This class defines counters to help identify when a method 2421 // may/must be executed using hardware with only 24-bit precision. 2422 struct Final_Reshape_Counts : public StackObj { 2423 int _call_count; // count non-inlined 'common' calls 2424 int _float_count; // count float ops requiring 24-bit precision 2425 int _double_count; // count double ops requiring more precision 2426 int _java_call_count; // count non-inlined 'java' calls 2427 int _inner_loop_count; // count loops which need alignment 2428 VectorSet _visited; // Visitation flags 2429 Node_List _tests; // Set of IfNodes & PCTableNodes 2430 2431 Final_Reshape_Counts() : 2432 _call_count(0), _float_count(0), _double_count(0), 2433 _java_call_count(0), _inner_loop_count(0), 2434 _visited( Thread::current()->resource_area() ) { } 2435 2436 void inc_call_count () { _call_count ++; } 2437 void inc_float_count () { _float_count ++; } 2438 void inc_double_count() { _double_count++; } 2439 void inc_java_call_count() { _java_call_count++; } 2440 void inc_inner_loop_count() { _inner_loop_count++; } 2441 2442 int get_call_count () const { return _call_count ; } 2443 int get_float_count () const { return _float_count ; } 2444 int get_double_count() const { return _double_count; } 2445 int get_java_call_count() const { return _java_call_count; } 2446 int get_inner_loop_count() const { return _inner_loop_count; } 2447 }; 2448 2449 #ifdef ASSERT 2450 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2451 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2452 // Make sure the offset goes inside the instance layout. 2453 return k->contains_field_offset(tp->offset()); 2454 // Note that OffsetBot and OffsetTop are very negative. 2455 } 2456 #endif 2457 2458 // Eliminate trivially redundant StoreCMs and accumulate their 2459 // precedence edges. 2460 void Compile::eliminate_redundant_card_marks(Node* n) { 2461 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2462 if (n->in(MemNode::Address)->outcnt() > 1) { 2463 // There are multiple users of the same address so it might be 2464 // possible to eliminate some of the StoreCMs 2465 Node* mem = n->in(MemNode::Memory); 2466 Node* adr = n->in(MemNode::Address); 2467 Node* val = n->in(MemNode::ValueIn); 2468 Node* prev = n; 2469 bool done = false; 2470 // Walk the chain of StoreCMs eliminating ones that match. As 2471 // long as it's a chain of single users then the optimization is 2472 // safe. Eliminating partially redundant StoreCMs would require 2473 // cloning copies down the other paths. 2474 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2475 if (adr == mem->in(MemNode::Address) && 2476 val == mem->in(MemNode::ValueIn)) { 2477 // redundant StoreCM 2478 if (mem->req() > MemNode::OopStore) { 2479 // Hasn't been processed by this code yet. 2480 n->add_prec(mem->in(MemNode::OopStore)); 2481 } else { 2482 // Already converted to precedence edge 2483 for (uint i = mem->req(); i < mem->len(); i++) { 2484 // Accumulate any precedence edges 2485 if (mem->in(i) != NULL) { 2486 n->add_prec(mem->in(i)); 2487 } 2488 } 2489 // Everything above this point has been processed. 2490 done = true; 2491 } 2492 // Eliminate the previous StoreCM 2493 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2494 assert(mem->outcnt() == 0, "should be dead"); 2495 mem->disconnect_inputs(NULL, this); 2496 } else { 2497 prev = mem; 2498 } 2499 mem = prev->in(MemNode::Memory); 2500 } 2501 } 2502 } 2503 2504 //------------------------------final_graph_reshaping_impl---------------------- 2505 // Implement items 1-5 from final_graph_reshaping below. 2506 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2507 2508 if ( n->outcnt() == 0 ) return; // dead node 2509 uint nop = n->Opcode(); 2510 2511 // Check for 2-input instruction with "last use" on right input. 2512 // Swap to left input. Implements item (2). 2513 if( n->req() == 3 && // two-input instruction 2514 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2515 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2516 n->in(2)->outcnt() == 1 &&// right use IS a last use 2517 !n->in(2)->is_Con() ) { // right use is not a constant 2518 // Check for commutative opcode 2519 switch( nop ) { 2520 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2521 case Op_MaxI: case Op_MinI: 2522 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2523 case Op_AndL: case Op_XorL: case Op_OrL: 2524 case Op_AndI: case Op_XorI: case Op_OrI: { 2525 // Move "last use" input to left by swapping inputs 2526 n->swap_edges(1, 2); 2527 break; 2528 } 2529 default: 2530 break; 2531 } 2532 } 2533 2534 #ifdef ASSERT 2535 if( n->is_Mem() ) { 2536 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2537 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2538 // oop will be recorded in oop map if load crosses safepoint 2539 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2540 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2541 "raw memory operations should have control edge"); 2542 } 2543 #endif 2544 // Count FPU ops and common calls, implements item (3) 2545 switch( nop ) { 2546 // Count all float operations that may use FPU 2547 case Op_AddF: 2548 case Op_SubF: 2549 case Op_MulF: 2550 case Op_DivF: 2551 case Op_NegF: 2552 case Op_ModF: 2553 case Op_ConvI2F: 2554 case Op_ConF: 2555 case Op_CmpF: 2556 case Op_CmpF3: 2557 // case Op_ConvL2F: // longs are split into 32-bit halves 2558 frc.inc_float_count(); 2559 break; 2560 2561 case Op_ConvF2D: 2562 case Op_ConvD2F: 2563 frc.inc_float_count(); 2564 frc.inc_double_count(); 2565 break; 2566 2567 // Count all double operations that may use FPU 2568 case Op_AddD: 2569 case Op_SubD: 2570 case Op_MulD: 2571 case Op_DivD: 2572 case Op_NegD: 2573 case Op_ModD: 2574 case Op_ConvI2D: 2575 case Op_ConvD2I: 2576 // case Op_ConvL2D: // handled by leaf call 2577 // case Op_ConvD2L: // handled by leaf call 2578 case Op_ConD: 2579 case Op_CmpD: 2580 case Op_CmpD3: 2581 frc.inc_double_count(); 2582 break; 2583 case Op_Opaque1: // Remove Opaque Nodes before matching 2584 case Op_Opaque2: // Remove Opaque Nodes before matching 2585 n->subsume_by(n->in(1), this); 2586 break; 2587 case Op_CallStaticJava: 2588 case Op_CallJava: 2589 case Op_CallDynamicJava: 2590 frc.inc_java_call_count(); // Count java call site; 2591 case Op_CallRuntime: 2592 case Op_CallLeaf: 2593 case Op_CallLeafNoFP: { 2594 assert( n->is_Call(), "" ); 2595 CallNode *call = n->as_Call(); 2596 // Count call sites where the FP mode bit would have to be flipped. 2597 // Do not count uncommon runtime calls: 2598 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2599 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2600 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2601 frc.inc_call_count(); // Count the call site 2602 } else { // See if uncommon argument is shared 2603 Node *n = call->in(TypeFunc::Parms); 2604 int nop = n->Opcode(); 2605 // Clone shared simple arguments to uncommon calls, item (1). 2606 if( n->outcnt() > 1 && 2607 !n->is_Proj() && 2608 nop != Op_CreateEx && 2609 nop != Op_CheckCastPP && 2610 nop != Op_DecodeN && 2611 nop != Op_DecodeNKlass && 2612 !n->is_Mem() ) { 2613 Node *x = n->clone(); 2614 call->set_req( TypeFunc::Parms, x ); 2615 } 2616 } 2617 break; 2618 } 2619 2620 case Op_StoreD: 2621 case Op_LoadD: 2622 case Op_LoadD_unaligned: 2623 frc.inc_double_count(); 2624 goto handle_mem; 2625 case Op_StoreF: 2626 case Op_LoadF: 2627 frc.inc_float_count(); 2628 goto handle_mem; 2629 2630 case Op_StoreCM: 2631 { 2632 // Convert OopStore dependence into precedence edge 2633 Node* prec = n->in(MemNode::OopStore); 2634 n->del_req(MemNode::OopStore); 2635 n->add_prec(prec); 2636 eliminate_redundant_card_marks(n); 2637 } 2638 2639 // fall through 2640 2641 case Op_StoreB: 2642 case Op_StoreC: 2643 case Op_StorePConditional: 2644 case Op_StoreI: 2645 case Op_StoreL: 2646 case Op_StoreIConditional: 2647 case Op_StoreLConditional: 2648 case Op_CompareAndSwapI: 2649 case Op_CompareAndSwapL: 2650 case Op_CompareAndSwapP: 2651 case Op_CompareAndSwapN: 2652 case Op_GetAndAddI: 2653 case Op_GetAndAddL: 2654 case Op_GetAndSetI: 2655 case Op_GetAndSetL: 2656 case Op_GetAndSetP: 2657 case Op_GetAndSetN: 2658 case Op_StoreP: 2659 case Op_StoreN: 2660 case Op_StoreNKlass: 2661 case Op_LoadB: 2662 case Op_LoadUB: 2663 case Op_LoadUS: 2664 case Op_LoadI: 2665 case Op_LoadKlass: 2666 case Op_LoadNKlass: 2667 case Op_LoadL: 2668 case Op_LoadL_unaligned: 2669 case Op_LoadPLocked: 2670 case Op_LoadP: 2671 case Op_LoadN: 2672 case Op_LoadRange: 2673 case Op_LoadS: { 2674 handle_mem: 2675 #ifdef ASSERT 2676 if( VerifyOptoOopOffsets ) { 2677 assert( n->is_Mem(), "" ); 2678 MemNode *mem = (MemNode*)n; 2679 // Check to see if address types have grounded out somehow. 2680 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2681 assert( !tp || oop_offset_is_sane(tp), "" ); 2682 } 2683 #endif 2684 break; 2685 } 2686 2687 case Op_AddP: { // Assert sane base pointers 2688 Node *addp = n->in(AddPNode::Address); 2689 assert( !addp->is_AddP() || 2690 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2691 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2692 "Base pointers must match" ); 2693 #ifdef _LP64 2694 if ((UseCompressedOops || UseCompressedClassPointers) && 2695 addp->Opcode() == Op_ConP && 2696 addp == n->in(AddPNode::Base) && 2697 n->in(AddPNode::Offset)->is_Con()) { 2698 // Use addressing with narrow klass to load with offset on x86. 2699 // On sparc loading 32-bits constant and decoding it have less 2700 // instructions (4) then load 64-bits constant (7). 2701 // Do this transformation here since IGVN will convert ConN back to ConP. 2702 const Type* t = addp->bottom_type(); 2703 if (t->isa_oopptr() || t->isa_klassptr()) { 2704 Node* nn = NULL; 2705 2706 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2707 2708 // Look for existing ConN node of the same exact type. 2709 Node* r = root(); 2710 uint cnt = r->outcnt(); 2711 for (uint i = 0; i < cnt; i++) { 2712 Node* m = r->raw_out(i); 2713 if (m!= NULL && m->Opcode() == op && 2714 m->bottom_type()->make_ptr() == t) { 2715 nn = m; 2716 break; 2717 } 2718 } 2719 if (nn != NULL) { 2720 // Decode a narrow oop to match address 2721 // [R12 + narrow_oop_reg<<3 + offset] 2722 if (t->isa_oopptr()) { 2723 nn = new (this) DecodeNNode(nn, t); 2724 } else { 2725 nn = new (this) DecodeNKlassNode(nn, t); 2726 } 2727 n->set_req(AddPNode::Base, nn); 2728 n->set_req(AddPNode::Address, nn); 2729 if (addp->outcnt() == 0) { 2730 addp->disconnect_inputs(NULL, this); 2731 } 2732 } 2733 } 2734 } 2735 #endif 2736 break; 2737 } 2738 2739 #ifdef _LP64 2740 case Op_CastPP: 2741 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2742 Node* in1 = n->in(1); 2743 const Type* t = n->bottom_type(); 2744 Node* new_in1 = in1->clone(); 2745 new_in1->as_DecodeN()->set_type(t); 2746 2747 if (!Matcher::narrow_oop_use_complex_address()) { 2748 // 2749 // x86, ARM and friends can handle 2 adds in addressing mode 2750 // and Matcher can fold a DecodeN node into address by using 2751 // a narrow oop directly and do implicit NULL check in address: 2752 // 2753 // [R12 + narrow_oop_reg<<3 + offset] 2754 // NullCheck narrow_oop_reg 2755 // 2756 // On other platforms (Sparc) we have to keep new DecodeN node and 2757 // use it to do implicit NULL check in address: 2758 // 2759 // decode_not_null narrow_oop_reg, base_reg 2760 // [base_reg + offset] 2761 // NullCheck base_reg 2762 // 2763 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2764 // to keep the information to which NULL check the new DecodeN node 2765 // corresponds to use it as value in implicit_null_check(). 2766 // 2767 new_in1->set_req(0, n->in(0)); 2768 } 2769 2770 n->subsume_by(new_in1, this); 2771 if (in1->outcnt() == 0) { 2772 in1->disconnect_inputs(NULL, this); 2773 } 2774 } 2775 break; 2776 2777 case Op_CmpP: 2778 // Do this transformation here to preserve CmpPNode::sub() and 2779 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2780 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2781 Node* in1 = n->in(1); 2782 Node* in2 = n->in(2); 2783 if (!in1->is_DecodeNarrowPtr()) { 2784 in2 = in1; 2785 in1 = n->in(2); 2786 } 2787 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2788 2789 Node* new_in2 = NULL; 2790 if (in2->is_DecodeNarrowPtr()) { 2791 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2792 new_in2 = in2->in(1); 2793 } else if (in2->Opcode() == Op_ConP) { 2794 const Type* t = in2->bottom_type(); 2795 if (t == TypePtr::NULL_PTR) { 2796 assert(in1->is_DecodeN(), "compare klass to null?"); 2797 // Don't convert CmpP null check into CmpN if compressed 2798 // oops implicit null check is not generated. 2799 // This will allow to generate normal oop implicit null check. 2800 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2801 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2802 // 2803 // This transformation together with CastPP transformation above 2804 // will generated code for implicit NULL checks for compressed oops. 2805 // 2806 // The original code after Optimize() 2807 // 2808 // LoadN memory, narrow_oop_reg 2809 // decode narrow_oop_reg, base_reg 2810 // CmpP base_reg, NULL 2811 // CastPP base_reg // NotNull 2812 // Load [base_reg + offset], val_reg 2813 // 2814 // after these transformations will be 2815 // 2816 // LoadN memory, narrow_oop_reg 2817 // CmpN narrow_oop_reg, NULL 2818 // decode_not_null narrow_oop_reg, base_reg 2819 // Load [base_reg + offset], val_reg 2820 // 2821 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2822 // since narrow oops can be used in debug info now (see the code in 2823 // final_graph_reshaping_walk()). 2824 // 2825 // At the end the code will be matched to 2826 // on x86: 2827 // 2828 // Load_narrow_oop memory, narrow_oop_reg 2829 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2830 // NullCheck narrow_oop_reg 2831 // 2832 // and on sparc: 2833 // 2834 // Load_narrow_oop memory, narrow_oop_reg 2835 // decode_not_null narrow_oop_reg, base_reg 2836 // Load [base_reg + offset], val_reg 2837 // NullCheck base_reg 2838 // 2839 } else if (t->isa_oopptr()) { 2840 new_in2 = ConNode::make(this, t->make_narrowoop()); 2841 } else if (t->isa_klassptr()) { 2842 new_in2 = ConNode::make(this, t->make_narrowklass()); 2843 } 2844 } 2845 if (new_in2 != NULL) { 2846 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2847 n->subsume_by(cmpN, this); 2848 if (in1->outcnt() == 0) { 2849 in1->disconnect_inputs(NULL, this); 2850 } 2851 if (in2->outcnt() == 0) { 2852 in2->disconnect_inputs(NULL, this); 2853 } 2854 } 2855 } 2856 break; 2857 2858 case Op_DecodeN: 2859 case Op_DecodeNKlass: 2860 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2861 // DecodeN could be pinned when it can't be fold into 2862 // an address expression, see the code for Op_CastPP above. 2863 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2864 break; 2865 2866 case Op_EncodeP: 2867 case Op_EncodePKlass: { 2868 Node* in1 = n->in(1); 2869 if (in1->is_DecodeNarrowPtr()) { 2870 n->subsume_by(in1->in(1), this); 2871 } else if (in1->Opcode() == Op_ConP) { 2872 const Type* t = in1->bottom_type(); 2873 if (t == TypePtr::NULL_PTR) { 2874 assert(t->isa_oopptr(), "null klass?"); 2875 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2876 } else if (t->isa_oopptr()) { 2877 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2878 } else if (t->isa_klassptr()) { 2879 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2880 } 2881 } 2882 if (in1->outcnt() == 0) { 2883 in1->disconnect_inputs(NULL, this); 2884 } 2885 break; 2886 } 2887 2888 case Op_Proj: { 2889 if (OptimizeStringConcat) { 2890 ProjNode* p = n->as_Proj(); 2891 if (p->_is_io_use) { 2892 // Separate projections were used for the exception path which 2893 // are normally removed by a late inline. If it wasn't inlined 2894 // then they will hang around and should just be replaced with 2895 // the original one. 2896 Node* proj = NULL; 2897 // Replace with just one 2898 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2899 Node *use = i.get(); 2900 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2901 proj = use; 2902 break; 2903 } 2904 } 2905 assert(proj != NULL, "must be found"); 2906 p->subsume_by(proj, this); 2907 } 2908 } 2909 break; 2910 } 2911 2912 case Op_Phi: 2913 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2914 // The EncodeP optimization may create Phi with the same edges 2915 // for all paths. It is not handled well by Register Allocator. 2916 Node* unique_in = n->in(1); 2917 assert(unique_in != NULL, ""); 2918 uint cnt = n->req(); 2919 for (uint i = 2; i < cnt; i++) { 2920 Node* m = n->in(i); 2921 assert(m != NULL, ""); 2922 if (unique_in != m) 2923 unique_in = NULL; 2924 } 2925 if (unique_in != NULL) { 2926 n->subsume_by(unique_in, this); 2927 } 2928 } 2929 break; 2930 2931 #endif 2932 2933 case Op_ModI: 2934 if (UseDivMod) { 2935 // Check if a%b and a/b both exist 2936 Node* d = n->find_similar(Op_DivI); 2937 if (d) { 2938 // Replace them with a fused divmod if supported 2939 if (Matcher::has_match_rule(Op_DivModI)) { 2940 DivModINode* divmod = DivModINode::make(this, n); 2941 d->subsume_by(divmod->div_proj(), this); 2942 n->subsume_by(divmod->mod_proj(), this); 2943 } else { 2944 // replace a%b with a-((a/b)*b) 2945 Node* mult = new (this) MulINode(d, d->in(2)); 2946 Node* sub = new (this) SubINode(d->in(1), mult); 2947 n->subsume_by(sub, this); 2948 } 2949 } 2950 } 2951 break; 2952 2953 case Op_ModL: 2954 if (UseDivMod) { 2955 // Check if a%b and a/b both exist 2956 Node* d = n->find_similar(Op_DivL); 2957 if (d) { 2958 // Replace them with a fused divmod if supported 2959 if (Matcher::has_match_rule(Op_DivModL)) { 2960 DivModLNode* divmod = DivModLNode::make(this, n); 2961 d->subsume_by(divmod->div_proj(), this); 2962 n->subsume_by(divmod->mod_proj(), this); 2963 } else { 2964 // replace a%b with a-((a/b)*b) 2965 Node* mult = new (this) MulLNode(d, d->in(2)); 2966 Node* sub = new (this) SubLNode(d->in(1), mult); 2967 n->subsume_by(sub, this); 2968 } 2969 } 2970 } 2971 break; 2972 2973 case Op_LoadVector: 2974 case Op_StoreVector: 2975 break; 2976 2977 case Op_PackB: 2978 case Op_PackS: 2979 case Op_PackI: 2980 case Op_PackF: 2981 case Op_PackL: 2982 case Op_PackD: 2983 if (n->req()-1 > 2) { 2984 // Replace many operand PackNodes with a binary tree for matching 2985 PackNode* p = (PackNode*) n; 2986 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2987 n->subsume_by(btp, this); 2988 } 2989 break; 2990 case Op_Loop: 2991 case Op_CountedLoop: 2992 if (n->as_Loop()->is_inner_loop()) { 2993 frc.inc_inner_loop_count(); 2994 } 2995 break; 2996 case Op_LShiftI: 2997 case Op_RShiftI: 2998 case Op_URShiftI: 2999 case Op_LShiftL: 3000 case Op_RShiftL: 3001 case Op_URShiftL: 3002 if (Matcher::need_masked_shift_count) { 3003 // The cpu's shift instructions don't restrict the count to the 3004 // lower 5/6 bits. We need to do the masking ourselves. 3005 Node* in2 = n->in(2); 3006 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3007 const TypeInt* t = in2->find_int_type(); 3008 if (t != NULL && t->is_con()) { 3009 juint shift = t->get_con(); 3010 if (shift > mask) { // Unsigned cmp 3011 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3012 } 3013 } else { 3014 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3015 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3016 n->set_req(2, shift); 3017 } 3018 } 3019 if (in2->outcnt() == 0) { // Remove dead node 3020 in2->disconnect_inputs(NULL, this); 3021 } 3022 } 3023 break; 3024 case Op_MemBarStoreStore: 3025 case Op_MemBarRelease: 3026 // Break the link with AllocateNode: it is no longer useful and 3027 // confuses register allocation. 3028 if (n->req() > MemBarNode::Precedent) { 3029 n->set_req(MemBarNode::Precedent, top()); 3030 } 3031 break; 3032 default: 3033 assert( !n->is_Call(), "" ); 3034 assert( !n->is_Mem(), "" ); 3035 break; 3036 } 3037 3038 // Collect CFG split points 3039 if (n->is_MultiBranch()) 3040 frc._tests.push(n); 3041 } 3042 3043 //------------------------------final_graph_reshaping_walk--------------------- 3044 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3045 // requires that the walk visits a node's inputs before visiting the node. 3046 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3047 ResourceArea *area = Thread::current()->resource_area(); 3048 Unique_Node_List sfpt(area); 3049 3050 frc._visited.set(root->_idx); // first, mark node as visited 3051 uint cnt = root->req(); 3052 Node *n = root; 3053 uint i = 0; 3054 while (true) { 3055 if (i < cnt) { 3056 // Place all non-visited non-null inputs onto stack 3057 Node* m = n->in(i); 3058 ++i; 3059 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3060 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 3061 sfpt.push(m); 3062 cnt = m->req(); 3063 nstack.push(n, i); // put on stack parent and next input's index 3064 n = m; 3065 i = 0; 3066 } 3067 } else { 3068 // Now do post-visit work 3069 final_graph_reshaping_impl( n, frc ); 3070 if (nstack.is_empty()) 3071 break; // finished 3072 n = nstack.node(); // Get node from stack 3073 cnt = n->req(); 3074 i = nstack.index(); 3075 nstack.pop(); // Shift to the next node on stack 3076 } 3077 } 3078 3079 // Skip next transformation if compressed oops are not used. 3080 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3081 (!UseCompressedOops && !UseCompressedClassPointers)) 3082 return; 3083 3084 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3085 // It could be done for an uncommon traps or any safepoints/calls 3086 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3087 while (sfpt.size() > 0) { 3088 n = sfpt.pop(); 3089 JVMState *jvms = n->as_SafePoint()->jvms(); 3090 assert(jvms != NULL, "sanity"); 3091 int start = jvms->debug_start(); 3092 int end = n->req(); 3093 bool is_uncommon = (n->is_CallStaticJava() && 3094 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3095 for (int j = start; j < end; j++) { 3096 Node* in = n->in(j); 3097 if (in->is_DecodeNarrowPtr()) { 3098 bool safe_to_skip = true; 3099 if (!is_uncommon ) { 3100 // Is it safe to skip? 3101 for (uint i = 0; i < in->outcnt(); i++) { 3102 Node* u = in->raw_out(i); 3103 if (!u->is_SafePoint() || 3104 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3105 safe_to_skip = false; 3106 } 3107 } 3108 } 3109 if (safe_to_skip) { 3110 n->set_req(j, in->in(1)); 3111 } 3112 if (in->outcnt() == 0) { 3113 in->disconnect_inputs(NULL, this); 3114 } 3115 } 3116 } 3117 } 3118 } 3119 3120 //------------------------------final_graph_reshaping-------------------------- 3121 // Final Graph Reshaping. 3122 // 3123 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3124 // and not commoned up and forced early. Must come after regular 3125 // optimizations to avoid GVN undoing the cloning. Clone constant 3126 // inputs to Loop Phis; these will be split by the allocator anyways. 3127 // Remove Opaque nodes. 3128 // (2) Move last-uses by commutative operations to the left input to encourage 3129 // Intel update-in-place two-address operations and better register usage 3130 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3131 // calls canonicalizing them back. 3132 // (3) Count the number of double-precision FP ops, single-precision FP ops 3133 // and call sites. On Intel, we can get correct rounding either by 3134 // forcing singles to memory (requires extra stores and loads after each 3135 // FP bytecode) or we can set a rounding mode bit (requires setting and 3136 // clearing the mode bit around call sites). The mode bit is only used 3137 // if the relative frequency of single FP ops to calls is low enough. 3138 // This is a key transform for SPEC mpeg_audio. 3139 // (4) Detect infinite loops; blobs of code reachable from above but not 3140 // below. Several of the Code_Gen algorithms fail on such code shapes, 3141 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3142 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3143 // Detection is by looking for IfNodes where only 1 projection is 3144 // reachable from below or CatchNodes missing some targets. 3145 // (5) Assert for insane oop offsets in debug mode. 3146 3147 bool Compile::final_graph_reshaping() { 3148 // an infinite loop may have been eliminated by the optimizer, 3149 // in which case the graph will be empty. 3150 if (root()->req() == 1) { 3151 record_method_not_compilable("trivial infinite loop"); 3152 return true; 3153 } 3154 3155 // Expensive nodes have their control input set to prevent the GVN 3156 // from freely commoning them. There's no GVN beyond this point so 3157 // no need to keep the control input. We want the expensive nodes to 3158 // be freely moved to the least frequent code path by gcm. 3159 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3160 for (int i = 0; i < expensive_count(); i++) { 3161 _expensive_nodes->at(i)->set_req(0, NULL); 3162 } 3163 3164 Final_Reshape_Counts frc; 3165 3166 // Visit everybody reachable! 3167 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3168 Node_Stack nstack(unique() >> 1); 3169 final_graph_reshaping_walk(nstack, root(), frc); 3170 3171 // Check for unreachable (from below) code (i.e., infinite loops). 3172 for( uint i = 0; i < frc._tests.size(); i++ ) { 3173 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3174 // Get number of CFG targets. 3175 // Note that PCTables include exception targets after calls. 3176 uint required_outcnt = n->required_outcnt(); 3177 if (n->outcnt() != required_outcnt) { 3178 // Check for a few special cases. Rethrow Nodes never take the 3179 // 'fall-thru' path, so expected kids is 1 less. 3180 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3181 if (n->in(0)->in(0)->is_Call()) { 3182 CallNode *call = n->in(0)->in(0)->as_Call(); 3183 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3184 required_outcnt--; // Rethrow always has 1 less kid 3185 } else if (call->req() > TypeFunc::Parms && 3186 call->is_CallDynamicJava()) { 3187 // Check for null receiver. In such case, the optimizer has 3188 // detected that the virtual call will always result in a null 3189 // pointer exception. The fall-through projection of this CatchNode 3190 // will not be populated. 3191 Node *arg0 = call->in(TypeFunc::Parms); 3192 if (arg0->is_Type() && 3193 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3194 required_outcnt--; 3195 } 3196 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3197 call->req() > TypeFunc::Parms+1 && 3198 call->is_CallStaticJava()) { 3199 // Check for negative array length. In such case, the optimizer has 3200 // detected that the allocation attempt will always result in an 3201 // exception. There is no fall-through projection of this CatchNode . 3202 Node *arg1 = call->in(TypeFunc::Parms+1); 3203 if (arg1->is_Type() && 3204 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3205 required_outcnt--; 3206 } 3207 } 3208 } 3209 } 3210 // Recheck with a better notion of 'required_outcnt' 3211 if (n->outcnt() != required_outcnt) { 3212 record_method_not_compilable("malformed control flow"); 3213 return true; // Not all targets reachable! 3214 } 3215 } 3216 // Check that I actually visited all kids. Unreached kids 3217 // must be infinite loops. 3218 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3219 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3220 record_method_not_compilable("infinite loop"); 3221 return true; // Found unvisited kid; must be unreach 3222 } 3223 } 3224 3225 // If original bytecodes contained a mixture of floats and doubles 3226 // check if the optimizer has made it homogenous, item (3). 3227 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3228 frc.get_float_count() > 32 && 3229 frc.get_double_count() == 0 && 3230 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3231 set_24_bit_selection_and_mode( false, true ); 3232 } 3233 3234 set_java_calls(frc.get_java_call_count()); 3235 set_inner_loops(frc.get_inner_loop_count()); 3236 3237 // No infinite loops, no reason to bail out. 3238 return false; 3239 } 3240 3241 //-----------------------------too_many_traps---------------------------------- 3242 // Report if there are too many traps at the current method and bci. 3243 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3244 bool Compile::too_many_traps(ciMethod* method, 3245 int bci, 3246 Deoptimization::DeoptReason reason) { 3247 ciMethodData* md = method->method_data(); 3248 if (md->is_empty()) { 3249 // Assume the trap has not occurred, or that it occurred only 3250 // because of a transient condition during start-up in the interpreter. 3251 return false; 3252 } 3253 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3254 if (md->has_trap_at(bci, m, reason) != 0) { 3255 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3256 // Also, if there are multiple reasons, or if there is no per-BCI record, 3257 // assume the worst. 3258 if (log()) 3259 log()->elem("observe trap='%s' count='%d'", 3260 Deoptimization::trap_reason_name(reason), 3261 md->trap_count(reason)); 3262 return true; 3263 } else { 3264 // Ignore method/bci and see if there have been too many globally. 3265 return too_many_traps(reason, md); 3266 } 3267 } 3268 3269 // Less-accurate variant which does not require a method and bci. 3270 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3271 ciMethodData* logmd) { 3272 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3273 // Too many traps globally. 3274 // Note that we use cumulative trap_count, not just md->trap_count. 3275 if (log()) { 3276 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3277 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3278 Deoptimization::trap_reason_name(reason), 3279 mcount, trap_count(reason)); 3280 } 3281 return true; 3282 } else { 3283 // The coast is clear. 3284 return false; 3285 } 3286 } 3287 3288 //--------------------------too_many_recompiles-------------------------------- 3289 // Report if there are too many recompiles at the current method and bci. 3290 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3291 // Is not eager to return true, since this will cause the compiler to use 3292 // Action_none for a trap point, to avoid too many recompilations. 3293 bool Compile::too_many_recompiles(ciMethod* method, 3294 int bci, 3295 Deoptimization::DeoptReason reason) { 3296 ciMethodData* md = method->method_data(); 3297 if (md->is_empty()) { 3298 // Assume the trap has not occurred, or that it occurred only 3299 // because of a transient condition during start-up in the interpreter. 3300 return false; 3301 } 3302 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3303 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3304 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3305 Deoptimization::DeoptReason per_bc_reason 3306 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3307 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3308 if ((per_bc_reason == Deoptimization::Reason_none 3309 || md->has_trap_at(bci, m, reason) != 0) 3310 // The trap frequency measure we care about is the recompile count: 3311 && md->trap_recompiled_at(bci, m) 3312 && md->overflow_recompile_count() >= bc_cutoff) { 3313 // Do not emit a trap here if it has already caused recompilations. 3314 // Also, if there are multiple reasons, or if there is no per-BCI record, 3315 // assume the worst. 3316 if (log()) 3317 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3318 Deoptimization::trap_reason_name(reason), 3319 md->trap_count(reason), 3320 md->overflow_recompile_count()); 3321 return true; 3322 } else if (trap_count(reason) != 0 3323 && decompile_count() >= m_cutoff) { 3324 // Too many recompiles globally, and we have seen this sort of trap. 3325 // Use cumulative decompile_count, not just md->decompile_count. 3326 if (log()) 3327 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3328 Deoptimization::trap_reason_name(reason), 3329 md->trap_count(reason), trap_count(reason), 3330 md->decompile_count(), decompile_count()); 3331 return true; 3332 } else { 3333 // The coast is clear. 3334 return false; 3335 } 3336 } 3337 3338 // Compute when not to trap. Used by matching trap based nodes and 3339 // NullCheck optimization. 3340 void Compile::set_allowed_deopt_reasons() { 3341 _allowed_reasons = 0; 3342 if (is_method_compilation()) { 3343 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3344 assert(rs < BitsPerInt, "recode bit map"); 3345 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3346 _allowed_reasons |= nth_bit(rs); 3347 } 3348 } 3349 } 3350 } 3351 3352 #ifndef PRODUCT 3353 //------------------------------verify_graph_edges--------------------------- 3354 // Walk the Graph and verify that there is a one-to-one correspondence 3355 // between Use-Def edges and Def-Use edges in the graph. 3356 void Compile::verify_graph_edges(bool no_dead_code) { 3357 if (VerifyGraphEdges) { 3358 ResourceArea *area = Thread::current()->resource_area(); 3359 Unique_Node_List visited(area); 3360 // Call recursive graph walk to check edges 3361 _root->verify_edges(visited); 3362 if (no_dead_code) { 3363 // Now make sure that no visited node is used by an unvisited node. 3364 bool dead_nodes = 0; 3365 Unique_Node_List checked(area); 3366 while (visited.size() > 0) { 3367 Node* n = visited.pop(); 3368 checked.push(n); 3369 for (uint i = 0; i < n->outcnt(); i++) { 3370 Node* use = n->raw_out(i); 3371 if (checked.member(use)) continue; // already checked 3372 if (visited.member(use)) continue; // already in the graph 3373 if (use->is_Con()) continue; // a dead ConNode is OK 3374 // At this point, we have found a dead node which is DU-reachable. 3375 if (dead_nodes++ == 0) 3376 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3377 use->dump(2); 3378 tty->print_cr("---"); 3379 checked.push(use); // No repeats; pretend it is now checked. 3380 } 3381 } 3382 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3383 } 3384 } 3385 } 3386 3387 // Verify GC barriers consistency 3388 // Currently supported: 3389 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3390 void Compile::verify_barriers() { 3391 if (UseG1GC) { 3392 // Verify G1 pre-barriers 3393 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3394 3395 ResourceArea *area = Thread::current()->resource_area(); 3396 Unique_Node_List visited(area); 3397 Node_List worklist(area); 3398 // We're going to walk control flow backwards starting from the Root 3399 worklist.push(_root); 3400 while (worklist.size() > 0) { 3401 Node* x = worklist.pop(); 3402 if (x == NULL || x == top()) continue; 3403 if (visited.member(x)) { 3404 continue; 3405 } else { 3406 visited.push(x); 3407 } 3408 3409 if (x->is_Region()) { 3410 for (uint i = 1; i < x->req(); i++) { 3411 worklist.push(x->in(i)); 3412 } 3413 } else { 3414 worklist.push(x->in(0)); 3415 // We are looking for the pattern: 3416 // /->ThreadLocal 3417 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3418 // \->ConI(0) 3419 // We want to verify that the If and the LoadB have the same control 3420 // See GraphKit::g1_write_barrier_pre() 3421 if (x->is_If()) { 3422 IfNode *iff = x->as_If(); 3423 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3424 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3425 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3426 && cmp->in(1)->is_Load()) { 3427 LoadNode* load = cmp->in(1)->as_Load(); 3428 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3429 && load->in(2)->in(3)->is_Con() 3430 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3431 3432 Node* if_ctrl = iff->in(0); 3433 Node* load_ctrl = load->in(0); 3434 3435 if (if_ctrl != load_ctrl) { 3436 // Skip possible CProj->NeverBranch in infinite loops 3437 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3438 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3439 if_ctrl = if_ctrl->in(0)->in(0); 3440 } 3441 } 3442 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3443 } 3444 } 3445 } 3446 } 3447 } 3448 } 3449 } 3450 } 3451 3452 #endif 3453 3454 // The Compile object keeps track of failure reasons separately from the ciEnv. 3455 // This is required because there is not quite a 1-1 relation between the 3456 // ciEnv and its compilation task and the Compile object. Note that one 3457 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3458 // to backtrack and retry without subsuming loads. Other than this backtracking 3459 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3460 // by the logic in C2Compiler. 3461 void Compile::record_failure(const char* reason) { 3462 if (log() != NULL) { 3463 log()->elem("failure reason='%s' phase='compile'", reason); 3464 } 3465 if (_failure_reason == NULL) { 3466 // Record the first failure reason. 3467 _failure_reason = reason; 3468 } 3469 3470 EventCompilerFailure event; 3471 if (event.should_commit()) { 3472 event.set_compileID(Compile::compile_id()); 3473 event.set_failure(reason); 3474 event.commit(); 3475 } 3476 3477 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3478 C->print_method(PHASE_FAILURE); 3479 } 3480 _root = NULL; // flush the graph, too 3481 } 3482 3483 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3484 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3485 _phase_name(name), _dolog(dolog) 3486 { 3487 if (dolog) { 3488 C = Compile::current(); 3489 _log = C->log(); 3490 } else { 3491 C = NULL; 3492 _log = NULL; 3493 } 3494 if (_log != NULL) { 3495 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3496 _log->stamp(); 3497 _log->end_head(); 3498 } 3499 } 3500 3501 Compile::TracePhase::~TracePhase() { 3502 3503 C = Compile::current(); 3504 if (_dolog) { 3505 _log = C->log(); 3506 } else { 3507 _log = NULL; 3508 } 3509 3510 #ifdef ASSERT 3511 if (PrintIdealNodeCount) { 3512 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3513 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3514 } 3515 3516 if (VerifyIdealNodeCount) { 3517 Compile::current()->print_missing_nodes(); 3518 } 3519 #endif 3520 3521 if (_log != NULL) { 3522 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3523 } 3524 } 3525 3526 //============================================================================= 3527 // Two Constant's are equal when the type and the value are equal. 3528 bool Compile::Constant::operator==(const Constant& other) { 3529 if (type() != other.type() ) return false; 3530 if (can_be_reused() != other.can_be_reused()) return false; 3531 // For floating point values we compare the bit pattern. 3532 switch (type()) { 3533 case T_FLOAT: return (_v._value.i == other._v._value.i); 3534 case T_LONG: 3535 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3536 case T_OBJECT: 3537 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3538 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3539 case T_METADATA: return (_v._metadata == other._v._metadata); 3540 default: ShouldNotReachHere(); 3541 } 3542 return false; 3543 } 3544 3545 static int type_to_size_in_bytes(BasicType t) { 3546 switch (t) { 3547 case T_LONG: return sizeof(jlong ); 3548 case T_FLOAT: return sizeof(jfloat ); 3549 case T_DOUBLE: return sizeof(jdouble); 3550 case T_METADATA: return sizeof(Metadata*); 3551 // We use T_VOID as marker for jump-table entries (labels) which 3552 // need an internal word relocation. 3553 case T_VOID: 3554 case T_ADDRESS: 3555 case T_OBJECT: return sizeof(jobject); 3556 } 3557 3558 ShouldNotReachHere(); 3559 return -1; 3560 } 3561 3562 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3563 // sort descending 3564 if (a->freq() > b->freq()) return -1; 3565 if (a->freq() < b->freq()) return 1; 3566 return 0; 3567 } 3568 3569 void Compile::ConstantTable::calculate_offsets_and_size() { 3570 // First, sort the array by frequencies. 3571 _constants.sort(qsort_comparator); 3572 3573 #ifdef ASSERT 3574 // Make sure all jump-table entries were sorted to the end of the 3575 // array (they have a negative frequency). 3576 bool found_void = false; 3577 for (int i = 0; i < _constants.length(); i++) { 3578 Constant con = _constants.at(i); 3579 if (con.type() == T_VOID) 3580 found_void = true; // jump-tables 3581 else 3582 assert(!found_void, "wrong sorting"); 3583 } 3584 #endif 3585 3586 int offset = 0; 3587 for (int i = 0; i < _constants.length(); i++) { 3588 Constant* con = _constants.adr_at(i); 3589 3590 // Align offset for type. 3591 int typesize = type_to_size_in_bytes(con->type()); 3592 offset = align_size_up(offset, typesize); 3593 con->set_offset(offset); // set constant's offset 3594 3595 if (con->type() == T_VOID) { 3596 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3597 offset = offset + typesize * n->outcnt(); // expand jump-table 3598 } else { 3599 offset = offset + typesize; 3600 } 3601 } 3602 3603 // Align size up to the next section start (which is insts; see 3604 // CodeBuffer::align_at_start). 3605 assert(_size == -1, "already set?"); 3606 _size = align_size_up(offset, CodeEntryAlignment); 3607 } 3608 3609 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3610 MacroAssembler _masm(&cb); 3611 for (int i = 0; i < _constants.length(); i++) { 3612 Constant con = _constants.at(i); 3613 address constant_addr; 3614 switch (con.type()) { 3615 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3616 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3617 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3618 case T_OBJECT: { 3619 jobject obj = con.get_jobject(); 3620 int oop_index = _masm.oop_recorder()->find_index(obj); 3621 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3622 break; 3623 } 3624 case T_ADDRESS: { 3625 address addr = (address) con.get_jobject(); 3626 constant_addr = _masm.address_constant(addr); 3627 break; 3628 } 3629 // We use T_VOID as marker for jump-table entries (labels) which 3630 // need an internal word relocation. 3631 case T_VOID: { 3632 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3633 // Fill the jump-table with a dummy word. The real value is 3634 // filled in later in fill_jump_table. 3635 address dummy = (address) n; 3636 constant_addr = _masm.address_constant(dummy); 3637 // Expand jump-table 3638 for (uint i = 1; i < n->outcnt(); i++) { 3639 address temp_addr = _masm.address_constant(dummy + i); 3640 assert(temp_addr, "consts section too small"); 3641 } 3642 break; 3643 } 3644 case T_METADATA: { 3645 Metadata* obj = con.get_metadata(); 3646 int metadata_index = _masm.oop_recorder()->find_index(obj); 3647 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3648 break; 3649 } 3650 default: ShouldNotReachHere(); 3651 } 3652 assert(constant_addr, "consts section too small"); 3653 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3654 } 3655 } 3656 3657 int Compile::ConstantTable::find_offset(Constant& con) const { 3658 int idx = _constants.find(con); 3659 assert(idx != -1, "constant must be in constant table"); 3660 int offset = _constants.at(idx).offset(); 3661 assert(offset != -1, "constant table not emitted yet?"); 3662 return offset; 3663 } 3664 3665 void Compile::ConstantTable::add(Constant& con) { 3666 if (con.can_be_reused()) { 3667 int idx = _constants.find(con); 3668 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3669 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3670 return; 3671 } 3672 } 3673 (void) _constants.append(con); 3674 } 3675 3676 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3677 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3678 Constant con(type, value, b->_freq); 3679 add(con); 3680 return con; 3681 } 3682 3683 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3684 Constant con(metadata); 3685 add(con); 3686 return con; 3687 } 3688 3689 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3690 jvalue value; 3691 BasicType type = oper->type()->basic_type(); 3692 switch (type) { 3693 case T_LONG: value.j = oper->constantL(); break; 3694 case T_FLOAT: value.f = oper->constantF(); break; 3695 case T_DOUBLE: value.d = oper->constantD(); break; 3696 case T_OBJECT: 3697 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3698 case T_METADATA: return add((Metadata*)oper->constant()); break; 3699 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3700 } 3701 return add(n, type, value); 3702 } 3703 3704 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3705 jvalue value; 3706 // We can use the node pointer here to identify the right jump-table 3707 // as this method is called from Compile::Fill_buffer right before 3708 // the MachNodes are emitted and the jump-table is filled (means the 3709 // MachNode pointers do not change anymore). 3710 value.l = (jobject) n; 3711 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3712 add(con); 3713 return con; 3714 } 3715 3716 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3717 // If called from Compile::scratch_emit_size do nothing. 3718 if (Compile::current()->in_scratch_emit_size()) return; 3719 3720 assert(labels.is_nonempty(), "must be"); 3721 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3722 3723 // Since MachConstantNode::constant_offset() also contains 3724 // table_base_offset() we need to subtract the table_base_offset() 3725 // to get the plain offset into the constant table. 3726 int offset = n->constant_offset() - table_base_offset(); 3727 3728 MacroAssembler _masm(&cb); 3729 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3730 3731 for (uint i = 0; i < n->outcnt(); i++) { 3732 address* constant_addr = &jump_table_base[i]; 3733 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))); 3734 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3735 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3736 } 3737 } 3738 3739 void Compile::dump_inlining() { 3740 if (print_inlining() || print_intrinsics()) { 3741 // Print inlining message for candidates that we couldn't inline 3742 // for lack of space or non constant receiver 3743 for (int i = 0; i < _late_inlines.length(); i++) { 3744 CallGenerator* cg = _late_inlines.at(i); 3745 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3746 } 3747 Unique_Node_List useful; 3748 useful.push(root()); 3749 for (uint next = 0; next < useful.size(); ++next) { 3750 Node* n = useful.at(next); 3751 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3752 CallNode* call = n->as_Call(); 3753 CallGenerator* cg = call->generator(); 3754 cg->print_inlining_late("receiver not constant"); 3755 } 3756 uint max = n->len(); 3757 for ( uint i = 0; i < max; ++i ) { 3758 Node *m = n->in(i); 3759 if ( m == NULL ) continue; 3760 useful.push(m); 3761 } 3762 } 3763 for (int i = 0; i < _print_inlining_list->length(); i++) { 3764 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string()); 3765 } 3766 } 3767 } 3768 3769 // Dump inlining replay data to the stream. 3770 // Don't change thread state and acquire any locks. 3771 void Compile::dump_inline_data(outputStream* out) { 3772 InlineTree* inl_tree = ilt(); 3773 if (inl_tree != NULL) { 3774 out->print(" inline %d", inl_tree->count()); 3775 inl_tree->dump_replay_data(out); 3776 } 3777 } 3778 3779 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3780 if (n1->Opcode() < n2->Opcode()) return -1; 3781 else if (n1->Opcode() > n2->Opcode()) return 1; 3782 3783 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())); 3784 for (uint i = 1; i < n1->req(); i++) { 3785 if (n1->in(i) < n2->in(i)) return -1; 3786 else if (n1->in(i) > n2->in(i)) return 1; 3787 } 3788 3789 return 0; 3790 } 3791 3792 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3793 Node* n1 = *n1p; 3794 Node* n2 = *n2p; 3795 3796 return cmp_expensive_nodes(n1, n2); 3797 } 3798 3799 void Compile::sort_expensive_nodes() { 3800 if (!expensive_nodes_sorted()) { 3801 _expensive_nodes->sort(cmp_expensive_nodes); 3802 } 3803 } 3804 3805 bool Compile::expensive_nodes_sorted() const { 3806 for (int i = 1; i < _expensive_nodes->length(); i++) { 3807 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3808 return false; 3809 } 3810 } 3811 return true; 3812 } 3813 3814 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3815 if (_expensive_nodes->length() == 0) { 3816 return false; 3817 } 3818 3819 assert(OptimizeExpensiveOps, "optimization off?"); 3820 3821 // Take this opportunity to remove dead nodes from the list 3822 int j = 0; 3823 for (int i = 0; i < _expensive_nodes->length(); i++) { 3824 Node* n = _expensive_nodes->at(i); 3825 if (!n->is_unreachable(igvn)) { 3826 assert(n->is_expensive(), "should be expensive"); 3827 _expensive_nodes->at_put(j, n); 3828 j++; 3829 } 3830 } 3831 _expensive_nodes->trunc_to(j); 3832 3833 // Then sort the list so that similar nodes are next to each other 3834 // and check for at least two nodes of identical kind with same data 3835 // inputs. 3836 sort_expensive_nodes(); 3837 3838 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3839 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3840 return true; 3841 } 3842 } 3843 3844 return false; 3845 } 3846 3847 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3848 if (_expensive_nodes->length() == 0) { 3849 return; 3850 } 3851 3852 assert(OptimizeExpensiveOps, "optimization off?"); 3853 3854 // Sort to bring similar nodes next to each other and clear the 3855 // control input of nodes for which there's only a single copy. 3856 sort_expensive_nodes(); 3857 3858 int j = 0; 3859 int identical = 0; 3860 int i = 0; 3861 for (; i < _expensive_nodes->length()-1; i++) { 3862 assert(j <= i, "can't write beyond current index"); 3863 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3864 identical++; 3865 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3866 continue; 3867 } 3868 if (identical > 0) { 3869 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3870 identical = 0; 3871 } else { 3872 Node* n = _expensive_nodes->at(i); 3873 igvn.hash_delete(n); 3874 n->set_req(0, NULL); 3875 igvn.hash_insert(n); 3876 } 3877 } 3878 if (identical > 0) { 3879 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3880 } else if (_expensive_nodes->length() >= 1) { 3881 Node* n = _expensive_nodes->at(i); 3882 igvn.hash_delete(n); 3883 n->set_req(0, NULL); 3884 igvn.hash_insert(n); 3885 } 3886 _expensive_nodes->trunc_to(j); 3887 } 3888 3889 void Compile::add_expensive_node(Node * n) { 3890 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3891 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3892 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3893 if (OptimizeExpensiveOps) { 3894 _expensive_nodes->append(n); 3895 } else { 3896 // Clear control input and let IGVN optimize expensive nodes if 3897 // OptimizeExpensiveOps is off. 3898 n->set_req(0, NULL); 3899 } 3900 } 3901 3902 /** 3903 * Remove the speculative part of types and clean up the graph 3904 */ 3905 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 3906 if (UseTypeSpeculation) { 3907 Unique_Node_List worklist; 3908 worklist.push(root()); 3909 int modified = 0; 3910 // Go over all type nodes that carry a speculative type, drop the 3911 // speculative part of the type and enqueue the node for an igvn 3912 // which may optimize it out. 3913 for (uint next = 0; next < worklist.size(); ++next) { 3914 Node *n = worklist.at(next); 3915 if (n->is_Type()) { 3916 TypeNode* tn = n->as_Type(); 3917 const Type* t = tn->type(); 3918 const Type* t_no_spec = t->remove_speculative(); 3919 if (t_no_spec != t) { 3920 bool in_hash = igvn.hash_delete(n); 3921 assert(in_hash, "node should be in igvn hash table"); 3922 tn->set_type(t_no_spec); 3923 igvn.hash_insert(n); 3924 igvn._worklist.push(n); // give it a chance to go away 3925 modified++; 3926 } 3927 } 3928 uint max = n->len(); 3929 for( uint i = 0; i < max; ++i ) { 3930 Node *m = n->in(i); 3931 if (not_a_node(m)) continue; 3932 worklist.push(m); 3933 } 3934 } 3935 // Drop the speculative part of all types in the igvn's type table 3936 igvn.remove_speculative_types(); 3937 if (modified > 0) { 3938 igvn.optimize(); 3939 } 3940 #ifdef ASSERT 3941 // Verify that after the IGVN is over no speculative type has resurfaced 3942 worklist.clear(); 3943 worklist.push(root()); 3944 for (uint next = 0; next < worklist.size(); ++next) { 3945 Node *n = worklist.at(next); 3946 const Type* t = igvn.type(n); 3947 assert(t == t->remove_speculative(), "no more speculative types"); 3948 if (n->is_Type()) { 3949 t = n->as_Type()->type(); 3950 assert(t == t->remove_speculative(), "no more speculative types"); 3951 } 3952 uint max = n->len(); 3953 for( uint i = 0; i < max; ++i ) { 3954 Node *m = n->in(i); 3955 if (not_a_node(m)) continue; 3956 worklist.push(m); 3957 } 3958 } 3959 igvn.check_no_speculative_types(); 3960 #endif 3961 } 3962 } 3963 3964 // Auxiliary method to support randomized stressing/fuzzing. 3965 // 3966 // This method can be called the arbitrary number of times, with current count 3967 // as the argument. The logic allows selecting a single candidate from the 3968 // running list of candidates as follows: 3969 // int count = 0; 3970 // Cand* selected = null; 3971 // while(cand = cand->next()) { 3972 // if (randomized_select(++count)) { 3973 // selected = cand; 3974 // } 3975 // } 3976 // 3977 // Including count equalizes the chances any candidate is "selected". 3978 // This is useful when we don't have the complete list of candidates to choose 3979 // from uniformly. In this case, we need to adjust the randomicity of the 3980 // selection, or else we will end up biasing the selection towards the latter 3981 // candidates. 3982 // 3983 // Quick back-envelope calculation shows that for the list of n candidates 3984 // the equal probability for the candidate to persist as "best" can be 3985 // achieved by replacing it with "next" k-th candidate with the probability 3986 // of 1/k. It can be easily shown that by the end of the run, the 3987 // probability for any candidate is converged to 1/n, thus giving the 3988 // uniform distribution among all the candidates. 3989 // 3990 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3991 #define RANDOMIZED_DOMAIN_POW 29 3992 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3993 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3994 bool Compile::randomized_select(int count) { 3995 assert(count > 0, "only positive"); 3996 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3997 }