1 /* 2 * Copyright (c) 1997, 2015, 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/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/convertnode.hpp" 46 #include "opto/divnode.hpp" 47 #include "opto/escape.hpp" 48 #include "opto/idealGraphPrinter.hpp" 49 #include "opto/loopnode.hpp" 50 #include "opto/machnode.hpp" 51 #include "opto/macro.hpp" 52 #include "opto/matcher.hpp" 53 #include "opto/mathexactnode.hpp" 54 #include "opto/memnode.hpp" 55 #include "opto/mulnode.hpp" 56 #include "opto/narrowptrnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/sharedRuntime.hpp" 69 #include "runtime/signature.hpp" 70 #include "runtime/stubRoutines.hpp" 71 #include "runtime/timer.hpp" 72 #include "utilities/copy.hpp" 73 74 75 // -------------------- Compile::mach_constant_base_node ----------------------- 76 // Constant table base node singleton. 77 MachConstantBaseNode* Compile::mach_constant_base_node() { 78 if (_mach_constant_base_node == NULL) { 79 _mach_constant_base_node = new MachConstantBaseNode(); 80 _mach_constant_base_node->add_req(C->root()); 81 } 82 return _mach_constant_base_node; 83 } 84 85 86 /// Support for intrinsics. 87 88 // Return the index at which m must be inserted (or already exists). 89 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 90 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 91 #ifdef ASSERT 92 for (int i = 1; i < _intrinsics->length(); i++) { 93 CallGenerator* cg1 = _intrinsics->at(i-1); 94 CallGenerator* cg2 = _intrinsics->at(i); 95 assert(cg1->method() != cg2->method() 96 ? cg1->method() < cg2->method() 97 : cg1->is_virtual() < cg2->is_virtual(), 98 "compiler intrinsics list must stay sorted"); 99 } 100 #endif 101 // Binary search sorted list, in decreasing intervals [lo, hi]. 102 int lo = 0, hi = _intrinsics->length()-1; 103 while (lo <= hi) { 104 int mid = (uint)(hi + lo) / 2; 105 ciMethod* mid_m = _intrinsics->at(mid)->method(); 106 if (m < mid_m) { 107 hi = mid-1; 108 } else if (m > mid_m) { 109 lo = mid+1; 110 } else { 111 // look at minor sort key 112 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 113 if (is_virtual < mid_virt) { 114 hi = mid-1; 115 } else if (is_virtual > mid_virt) { 116 lo = mid+1; 117 } else { 118 return mid; // exact match 119 } 120 } 121 } 122 return lo; // inexact match 123 } 124 125 void Compile::register_intrinsic(CallGenerator* cg) { 126 if (_intrinsics == NULL) { 127 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 128 } 129 // This code is stolen from ciObjectFactory::insert. 130 // Really, GrowableArray should have methods for 131 // insert_at, remove_at, and binary_search. 132 int len = _intrinsics->length(); 133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 134 if (index == len) { 135 _intrinsics->append(cg); 136 } else { 137 #ifdef ASSERT 138 CallGenerator* oldcg = _intrinsics->at(index); 139 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 140 #endif 141 _intrinsics->append(_intrinsics->at(len-1)); 142 int pos; 143 for (pos = len-2; pos >= index; pos--) { 144 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 145 } 146 _intrinsics->at_put(index, cg); 147 } 148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 149 } 150 151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 152 assert(m->is_loaded(), "don't try this on unloaded methods"); 153 if (_intrinsics != NULL) { 154 int index = intrinsic_insertion_index(m, is_virtual); 155 if (index < _intrinsics->length() 156 && _intrinsics->at(index)->method() == m 157 && _intrinsics->at(index)->is_virtual() == is_virtual) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = unique(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless expensive node 406 for (int i = C->expensive_count()-1; i >= 0; i--) { 407 Node* n = C->expensive_node(i); 408 if (!useful.member(n)) { 409 remove_expensive_node(n); 410 } 411 } 412 // clean up the late inline lists 413 remove_useless_late_inlines(&_string_late_inlines, useful); 414 remove_useless_late_inlines(&_boxing_late_inlines, useful); 415 remove_useless_late_inlines(&_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 //------------------------------frame_size_in_words----------------------------- 420 // frame_slots in units of words 421 int Compile::frame_size_in_words() const { 422 // shift is 0 in LP32 and 1 in LP64 423 const int shift = (LogBytesPerWord - LogBytesPerInt); 424 int words = _frame_slots >> shift; 425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 426 return words; 427 } 428 429 // To bang the stack of this compiled method we use the stack size 430 // that the interpreter would need in case of a deoptimization. This 431 // removes the need to bang the stack in the deoptimization blob which 432 // in turn simplifies stack overflow handling. 433 int Compile::bang_size_in_bytes() const { 434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_type_hwm(NULL); 457 compile->set_type_last_size(0); 458 compile->set_last_tf(NULL, NULL); 459 compile->set_indexSet_arena(NULL); 460 compile->set_indexSet_free_block_list(NULL); 461 compile->init_type_arena(); 462 Type::Initialize(compile); 463 _compile->set_scratch_buffer_blob(NULL); 464 _compile->begin_method(); 465 } 466 CompileWrapper::~CompileWrapper() { 467 _compile->end_method(); 468 if (_compile->scratch_buffer_blob() != NULL) 469 BufferBlob::free(_compile->scratch_buffer_blob()); 470 _compile->env()->set_compiler_data(NULL); 471 } 472 473 474 //----------------------------print_compile_messages--------------------------- 475 void Compile::print_compile_messages() { 476 #ifndef PRODUCT 477 // Check if recompiling 478 if (_subsume_loads == false && PrintOpto) { 479 // Recompiling without allowing machine instructions to subsume loads 480 tty->print_cr("*********************************************************"); 481 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 482 tty->print_cr("*********************************************************"); 483 } 484 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 485 // Recompiling without escape analysis 486 tty->print_cr("*********************************************************"); 487 tty->print_cr("** Bailout: Recompile without escape analysis **"); 488 tty->print_cr("*********************************************************"); 489 } 490 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 491 // Recompiling without boxing elimination 492 tty->print_cr("*********************************************************"); 493 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 494 tty->print_cr("*********************************************************"); 495 } 496 if (env()->break_at_compile()) { 497 // Open the debugger when compiling this method. 498 tty->print("### Breaking when compiling: "); 499 method()->print_short_name(); 500 tty->cr(); 501 BREAKPOINT; 502 } 503 504 if( PrintOpto ) { 505 if (is_osr_compilation()) { 506 tty->print("[OSR]%3d", _compile_id); 507 } else { 508 tty->print("%3d", _compile_id); 509 } 510 } 511 #endif 512 } 513 514 515 //-----------------------init_scratch_buffer_blob------------------------------ 516 // Construct a temporary BufferBlob and cache it for this compile. 517 void Compile::init_scratch_buffer_blob(int const_size) { 518 // If there is already a scratch buffer blob allocated and the 519 // constant section is big enough, use it. Otherwise free the 520 // current and allocate a new one. 521 BufferBlob* blob = scratch_buffer_blob(); 522 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 523 // Use the current blob. 524 } else { 525 if (blob != NULL) { 526 BufferBlob::free(blob); 527 } 528 529 ResourceMark rm; 530 _scratch_const_size = const_size; 531 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 532 blob = BufferBlob::create("Compile::scratch_buffer", size); 533 // Record the buffer blob for next time. 534 set_scratch_buffer_blob(blob); 535 // Have we run out of code space? 536 if (scratch_buffer_blob() == NULL) { 537 // Let CompilerBroker disable further compilations. 538 record_failure("Not enough space for scratch buffer in CodeCache"); 539 return; 540 } 541 } 542 543 // Initialize the relocation buffers 544 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 545 set_scratch_locs_memory(locs_buf); 546 } 547 548 549 //-----------------------scratch_emit_size------------------------------------- 550 // Helper function that computes size by emitting code 551 uint Compile::scratch_emit_size(const Node* n) { 552 // Start scratch_emit_size section. 553 set_in_scratch_emit_size(true); 554 555 // Emit into a trash buffer and count bytes emitted. 556 // This is a pretty expensive way to compute a size, 557 // but it works well enough if seldom used. 558 // All common fixed-size instructions are given a size 559 // method by the AD file. 560 // Note that the scratch buffer blob and locs memory are 561 // allocated at the beginning of the compile task, and 562 // may be shared by several calls to scratch_emit_size. 563 // The allocation of the scratch buffer blob is particularly 564 // expensive, since it has to grab the code cache lock. 565 BufferBlob* blob = this->scratch_buffer_blob(); 566 assert(blob != NULL, "Initialize BufferBlob at start"); 567 assert(blob->size() > MAX_inst_size, "sanity"); 568 relocInfo* locs_buf = scratch_locs_memory(); 569 address blob_begin = blob->content_begin(); 570 address blob_end = (address)locs_buf; 571 assert(blob->content_contains(blob_end), "sanity"); 572 CodeBuffer buf(blob_begin, blob_end - blob_begin); 573 buf.initialize_consts_size(_scratch_const_size); 574 buf.initialize_stubs_size(MAX_stubs_size); 575 assert(locs_buf != NULL, "sanity"); 576 int lsize = MAX_locs_size / 3; 577 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 578 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 579 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 580 581 // Do the emission. 582 583 Label fakeL; // Fake label for branch instructions. 584 Label* saveL = NULL; 585 uint save_bnum = 0; 586 bool is_branch = n->is_MachBranch(); 587 if (is_branch) { 588 MacroAssembler masm(&buf); 589 masm.bind(fakeL); 590 n->as_MachBranch()->save_label(&saveL, &save_bnum); 591 n->as_MachBranch()->label_set(&fakeL, 0); 592 } 593 n->emit(buf, this->regalloc()); 594 if (is_branch) // Restore label. 595 n->as_MachBranch()->label_set(saveL, save_bnum); 596 597 // End scratch_emit_size section. 598 set_in_scratch_emit_size(false); 599 600 return buf.insts_size(); 601 } 602 603 604 // ============================================================================ 605 //------------------------------Compile standard------------------------------- 606 debug_only( int Compile::_debug_idx = 100000; ) 607 608 // Compile a method. entry_bci is -1 for normal compilations and indicates 609 // the continuation bci for on stack replacement. 610 611 612 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 613 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 614 : Phase(Compiler), 615 _env(ci_env), 616 _log(ci_env->log()), 617 _compile_id(ci_env->compile_id()), 618 _save_argument_registers(false), 619 _stub_name(NULL), 620 _stub_function(NULL), 621 _stub_entry_point(NULL), 622 _method(target), 623 _entry_bci(osr_bci), 624 _initial_gvn(NULL), 625 _for_igvn(NULL), 626 _warm_calls(NULL), 627 _subsume_loads(subsume_loads), 628 _do_escape_analysis(do_escape_analysis), 629 _eliminate_boxing(eliminate_boxing), 630 _failure_reason(NULL), 631 _code_buffer("Compile::Fill_buffer"), 632 _orig_pc_slot(0), 633 _orig_pc_slot_offset_in_bytes(0), 634 _has_method_handle_invokes(false), 635 _mach_constant_base_node(NULL), 636 _node_bundling_limit(0), 637 _node_bundling_base(NULL), 638 _java_calls(0), 639 _inner_loops(0), 640 _scratch_const_size(-1), 641 _in_scratch_emit_size(false), 642 _dead_node_list(comp_arena()), 643 _dead_node_count(0), 644 #ifndef PRODUCT 645 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 646 _in_dump_cnt(0), 647 _printer(IdealGraphPrinter::printer()), 648 #endif 649 _congraph(NULL), 650 _comp_arena(mtCompiler), 651 _node_arena(mtCompiler), 652 _old_arena(mtCompiler), 653 _Compile_types(mtCompiler), 654 _replay_inline_data(NULL), 655 _late_inlines(comp_arena(), 2, 0, NULL), 656 _string_late_inlines(comp_arena(), 2, 0, NULL), 657 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 658 _late_inlines_pos(0), 659 _number_of_mh_late_inlines(0), 660 _inlining_progress(false), 661 _inlining_incrementally(false), 662 _print_inlining_list(NULL), 663 _print_inlining_stream(NULL), 664 _print_inlining_idx(0), 665 _print_inlining_output(NULL), 666 _interpreter_frame_size(0), 667 _max_node_limit(MaxNodeLimit) { 668 C = this; 669 670 CompileWrapper cw(this); 671 672 if (CITimeVerbose) { 673 tty->print(" "); 674 target->holder()->name()->print(); 675 tty->print("."); 676 target->print_short_name(); 677 tty->print(" "); 678 } 679 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 680 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 681 682 #ifndef PRODUCT 683 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 684 if (!print_opto_assembly) { 685 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 686 if (print_assembly && !Disassembler::can_decode()) { 687 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 688 print_opto_assembly = true; 689 } 690 } 691 set_print_assembly(print_opto_assembly); 692 set_parsed_irreducible_loop(false); 693 694 if (method()->has_option("ReplayInline")) { 695 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 696 } 697 #endif 698 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 699 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 700 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 701 702 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 703 // Make sure the method being compiled gets its own MDO, 704 // so we can at least track the decompile_count(). 705 // Need MDO to record RTM code generation state. 706 method()->ensure_method_data(); 707 } 708 709 Init(::AliasLevel); 710 711 712 print_compile_messages(); 713 714 _ilt = InlineTree::build_inline_tree_root(); 715 716 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 717 assert(num_alias_types() >= AliasIdxRaw, ""); 718 719 #define MINIMUM_NODE_HASH 1023 720 // Node list that Iterative GVN will start with 721 Unique_Node_List for_igvn(comp_arena()); 722 set_for_igvn(&for_igvn); 723 724 // GVN that will be run immediately on new nodes 725 uint estimated_size = method()->code_size()*4+64; 726 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 727 PhaseGVN gvn(node_arena(), estimated_size); 728 set_initial_gvn(&gvn); 729 730 print_inlining_init(); 731 { // Scope for timing the parser 732 TracePhase tp("parse", &timers[_t_parser]); 733 734 // Put top into the hash table ASAP. 735 initial_gvn()->transform_no_reclaim(top()); 736 737 // Set up tf(), start(), and find a CallGenerator. 738 CallGenerator* cg = NULL; 739 if (is_osr_compilation()) { 740 const TypeTuple *domain = StartOSRNode::osr_domain(); 741 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 742 init_tf(TypeFunc::make(domain, range)); 743 StartNode* s = new StartOSRNode(root(), domain); 744 initial_gvn()->set_type_bottom(s); 745 init_start(s); 746 cg = CallGenerator::for_osr(method(), entry_bci()); 747 } else { 748 // Normal case. 749 init_tf(TypeFunc::make(method())); 750 StartNode* s = new StartNode(root(), tf()->domain()); 751 initial_gvn()->set_type_bottom(s); 752 init_start(s); 753 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 754 // With java.lang.ref.reference.get() we must go through the 755 // intrinsic when G1 is enabled - even when get() is the root 756 // method of the compile - so that, if necessary, the value in 757 // the referent field of the reference object gets recorded by 758 // the pre-barrier code. 759 // Specifically, if G1 is enabled, the value in the referent 760 // field is recorded by the G1 SATB pre barrier. This will 761 // result in the referent being marked live and the reference 762 // object removed from the list of discovered references during 763 // reference processing. 764 cg = find_intrinsic(method(), false); 765 } 766 if (cg == NULL) { 767 float past_uses = method()->interpreter_invocation_count(); 768 float expected_uses = past_uses; 769 cg = CallGenerator::for_inline(method(), expected_uses); 770 } 771 } 772 if (failing()) return; 773 if (cg == NULL) { 774 record_method_not_compilable_all_tiers("cannot parse method"); 775 return; 776 } 777 JVMState* jvms = build_start_state(start(), tf()); 778 if ((jvms = cg->generate(jvms)) == NULL) { 779 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 780 record_method_not_compilable("method parse failed"); 781 } 782 return; 783 } 784 GraphKit kit(jvms); 785 786 if (!kit.stopped()) { 787 // Accept return values, and transfer control we know not where. 788 // This is done by a special, unique ReturnNode bound to root. 789 return_values(kit.jvms()); 790 } 791 792 if (kit.has_exceptions()) { 793 // Any exceptions that escape from this call must be rethrown 794 // to whatever caller is dynamically above us on the stack. 795 // This is done by a special, unique RethrowNode bound to root. 796 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 797 } 798 799 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 800 801 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 802 inline_string_calls(true); 803 } 804 805 if (failing()) return; 806 807 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 808 809 // Remove clutter produced by parsing. 810 if (!failing()) { 811 ResourceMark rm; 812 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 813 } 814 } 815 816 // Note: Large methods are capped off in do_one_bytecode(). 817 if (failing()) return; 818 819 // After parsing, node notes are no longer automagic. 820 // They must be propagated by register_new_node_with_optimizer(), 821 // clone(), or the like. 822 set_default_node_notes(NULL); 823 824 for (;;) { 825 int successes = Inline_Warm(); 826 if (failing()) return; 827 if (successes == 0) break; 828 } 829 830 // Drain the list. 831 Finish_Warm(); 832 #ifndef PRODUCT 833 if (_printer && _printer->should_print(_method)) { 834 _printer->print_inlining(this); 835 } 836 #endif 837 838 if (failing()) return; 839 NOT_PRODUCT( verify_graph_edges(); ) 840 841 // Now optimize 842 Optimize(); 843 if (failing()) return; 844 NOT_PRODUCT( verify_graph_edges(); ) 845 846 #ifndef PRODUCT 847 if (PrintIdeal) { 848 ttyLocker ttyl; // keep the following output all in one block 849 // This output goes directly to the tty, not the compiler log. 850 // To enable tools to match it up with the compilation activity, 851 // be sure to tag this tty output with the compile ID. 852 if (xtty != NULL) { 853 xtty->head("ideal compile_id='%d'%s", compile_id(), 854 is_osr_compilation() ? " compile_kind='osr'" : 855 ""); 856 } 857 root()->dump(9999); 858 if (xtty != NULL) { 859 xtty->tail("ideal"); 860 } 861 } 862 #endif 863 864 NOT_PRODUCT( verify_barriers(); ) 865 866 // Dump compilation data to replay it. 867 if (method()->has_option("DumpReplay")) { 868 env()->dump_replay_data(_compile_id); 869 } 870 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 871 env()->dump_inline_data(_compile_id); 872 } 873 874 // Now that we know the size of all the monitors we can add a fixed slot 875 // for the original deopt pc. 876 877 _orig_pc_slot = fixed_slots(); 878 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 879 set_fixed_slots(next_slot); 880 881 // Compute when to use implicit null checks. Used by matching trap based 882 // nodes and NullCheck optimization. 883 set_allowed_deopt_reasons(); 884 885 // Now generate code 886 Code_Gen(); 887 if (failing()) return; 888 889 // Check if we want to skip execution of all compiled code. 890 { 891 #ifndef PRODUCT 892 if (OptoNoExecute) { 893 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 894 return; 895 } 896 #endif 897 TracePhase tp("install_code", &timers[_t_registerMethod]); 898 899 if (is_osr_compilation()) { 900 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 901 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 902 } else { 903 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 904 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 905 } 906 907 env()->register_method(_method, _entry_bci, 908 &_code_offsets, 909 _orig_pc_slot_offset_in_bytes, 910 code_buffer(), 911 frame_size_in_words(), _oop_map_set, 912 &_handler_table, &_inc_table, 913 compiler, 914 env()->comp_level(), 915 has_unsafe_access(), 916 SharedRuntime::is_wide_vector(max_vector_size()), 917 rtm_state() 918 ); 919 920 if (log() != NULL) // Print code cache state into compiler log 921 log()->code_cache_state(); 922 } 923 } 924 925 //------------------------------Compile---------------------------------------- 926 // Compile a runtime stub 927 Compile::Compile( ciEnv* ci_env, 928 TypeFunc_generator generator, 929 address stub_function, 930 const char *stub_name, 931 int is_fancy_jump, 932 bool pass_tls, 933 bool save_arg_registers, 934 bool return_pc ) 935 : Phase(Compiler), 936 _env(ci_env), 937 _log(ci_env->log()), 938 _compile_id(0), 939 _save_argument_registers(save_arg_registers), 940 _method(NULL), 941 _stub_name(stub_name), 942 _stub_function(stub_function), 943 _stub_entry_point(NULL), 944 _entry_bci(InvocationEntryBci), 945 _initial_gvn(NULL), 946 _for_igvn(NULL), 947 _warm_calls(NULL), 948 _orig_pc_slot(0), 949 _orig_pc_slot_offset_in_bytes(0), 950 _subsume_loads(true), 951 _do_escape_analysis(false), 952 _eliminate_boxing(false), 953 _failure_reason(NULL), 954 _code_buffer("Compile::Fill_buffer"), 955 _has_method_handle_invokes(false), 956 _mach_constant_base_node(NULL), 957 _node_bundling_limit(0), 958 _node_bundling_base(NULL), 959 _java_calls(0), 960 _inner_loops(0), 961 #ifndef PRODUCT 962 _trace_opto_output(TraceOptoOutput), 963 _in_dump_cnt(0), 964 _printer(NULL), 965 #endif 966 _comp_arena(mtCompiler), 967 _node_arena(mtCompiler), 968 _old_arena(mtCompiler), 969 _Compile_types(mtCompiler), 970 _dead_node_list(comp_arena()), 971 _dead_node_count(0), 972 _congraph(NULL), 973 _replay_inline_data(NULL), 974 _number_of_mh_late_inlines(0), 975 _inlining_progress(false), 976 _inlining_incrementally(false), 977 _print_inlining_list(NULL), 978 _print_inlining_stream(NULL), 979 _print_inlining_idx(0), 980 _print_inlining_output(NULL), 981 _allowed_reasons(0), 982 _interpreter_frame_size(0), 983 _max_node_limit(MaxNodeLimit) { 984 C = this; 985 986 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 987 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 988 989 #ifndef PRODUCT 990 set_print_assembly(PrintFrameConverterAssembly); 991 set_parsed_irreducible_loop(false); 992 #endif 993 set_has_irreducible_loop(false); // no loops 994 995 CompileWrapper cw(this); 996 Init(/*AliasLevel=*/ 0); 997 init_tf((*generator)()); 998 999 { 1000 // The following is a dummy for the sake of GraphKit::gen_stub 1001 Unique_Node_List for_igvn(comp_arena()); 1002 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1003 PhaseGVN gvn(Thread::current()->resource_area(),255); 1004 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1005 gvn.transform_no_reclaim(top()); 1006 1007 GraphKit kit; 1008 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1009 } 1010 1011 NOT_PRODUCT( verify_graph_edges(); ) 1012 Code_Gen(); 1013 if (failing()) return; 1014 1015 1016 // Entry point will be accessed using compile->stub_entry_point(); 1017 if (code_buffer() == NULL) { 1018 Matcher::soft_match_failure(); 1019 } else { 1020 if (PrintAssembly && (WizardMode || Verbose)) 1021 tty->print_cr("### Stub::%s", stub_name); 1022 1023 if (!failing()) { 1024 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1025 1026 // Make the NMethod 1027 // For now we mark the frame as never safe for profile stackwalking 1028 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1029 code_buffer(), 1030 CodeOffsets::frame_never_safe, 1031 // _code_offsets.value(CodeOffsets::Frame_Complete), 1032 frame_size_in_words(), 1033 _oop_map_set, 1034 save_arg_registers); 1035 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1036 1037 _stub_entry_point = rs->entry_point(); 1038 } 1039 } 1040 } 1041 1042 //------------------------------Init------------------------------------------- 1043 // Prepare for a single compilation 1044 void Compile::Init(int aliaslevel) { 1045 _unique = 0; 1046 _regalloc = NULL; 1047 1048 _tf = NULL; // filled in later 1049 _top = NULL; // cached later 1050 _matcher = NULL; // filled in later 1051 _cfg = NULL; // filled in later 1052 1053 set_24_bit_selection_and_mode(Use24BitFP, false); 1054 1055 _node_note_array = NULL; 1056 _default_node_notes = NULL; 1057 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1058 1059 _immutable_memory = NULL; // filled in at first inquiry 1060 1061 // Globally visible Nodes 1062 // First set TOP to NULL to give safe behavior during creation of RootNode 1063 set_cached_top_node(NULL); 1064 set_root(new RootNode()); 1065 // Now that you have a Root to point to, create the real TOP 1066 set_cached_top_node( new ConNode(Type::TOP) ); 1067 set_recent_alloc(NULL, NULL); 1068 1069 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1070 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1071 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1072 env()->set_dependencies(new Dependencies(env())); 1073 1074 _fixed_slots = 0; 1075 set_has_split_ifs(false); 1076 set_has_loops(has_method() && method()->has_loops()); // first approximation 1077 set_has_stringbuilder(false); 1078 set_has_boxed_value(false); 1079 _trap_can_recompile = false; // no traps emitted yet 1080 _major_progress = true; // start out assuming good things will happen 1081 set_has_unsafe_access(false); 1082 set_max_vector_size(0); 1083 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1084 set_decompile_count(0); 1085 1086 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1087 set_num_loop_opts(LoopOptsCount); 1088 set_do_inlining(Inline); 1089 set_max_inline_size(MaxInlineSize); 1090 set_freq_inline_size(FreqInlineSize); 1091 set_do_scheduling(OptoScheduling); 1092 set_do_count_invocations(false); 1093 set_do_method_data_update(false); 1094 set_age_code(has_method() && method()->profile_aging()); 1095 set_rtm_state(NoRTM); // No RTM lock eliding by default 1096 method_has_option_value("MaxNodeLimit", _max_node_limit); 1097 #if INCLUDE_RTM_OPT 1098 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1099 int rtm_state = method()->method_data()->rtm_state(); 1100 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1101 // Don't generate RTM lock eliding code. 1102 set_rtm_state(NoRTM); 1103 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1104 // Generate RTM lock eliding code without abort ratio calculation code. 1105 set_rtm_state(UseRTM); 1106 } else if (UseRTMDeopt) { 1107 // Generate RTM lock eliding code and include abort ratio calculation 1108 // code if UseRTMDeopt is on. 1109 set_rtm_state(ProfileRTM); 1110 } 1111 } 1112 #endif 1113 if (debug_info()->recording_non_safepoints()) { 1114 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1115 (comp_arena(), 8, 0, NULL)); 1116 set_default_node_notes(Node_Notes::make(this)); 1117 } 1118 1119 // // -- Initialize types before each compile -- 1120 // // Update cached type information 1121 // if( _method && _method->constants() ) 1122 // Type::update_loaded_types(_method, _method->constants()); 1123 1124 // Init alias_type map. 1125 if (!_do_escape_analysis && aliaslevel == 3) 1126 aliaslevel = 2; // No unique types without escape analysis 1127 _AliasLevel = aliaslevel; 1128 const int grow_ats = 16; 1129 _max_alias_types = grow_ats; 1130 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1131 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1132 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1133 { 1134 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1135 } 1136 // Initialize the first few types. 1137 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1138 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1139 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1140 _num_alias_types = AliasIdxRaw+1; 1141 // Zero out the alias type cache. 1142 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1143 // A NULL adr_type hits in the cache right away. Preload the right answer. 1144 probe_alias_cache(NULL)->_index = AliasIdxTop; 1145 1146 _intrinsics = NULL; 1147 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1148 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1149 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1150 register_library_intrinsics(); 1151 } 1152 1153 //---------------------------init_start---------------------------------------- 1154 // Install the StartNode on this compile object. 1155 void Compile::init_start(StartNode* s) { 1156 if (failing()) 1157 return; // already failing 1158 assert(s == start(), ""); 1159 } 1160 1161 /** 1162 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1163 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1164 * the ideal graph. 1165 */ 1166 StartNode* Compile::start() const { 1167 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1168 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1169 Node* start = root()->fast_out(i); 1170 if (start->is_Start()) { 1171 return start->as_Start(); 1172 } 1173 } 1174 fatal("Did not find Start node!"); 1175 return NULL; 1176 } 1177 1178 //-------------------------------immutable_memory------------------------------------- 1179 // Access immutable memory 1180 Node* Compile::immutable_memory() { 1181 if (_immutable_memory != NULL) { 1182 return _immutable_memory; 1183 } 1184 StartNode* s = start(); 1185 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1186 Node *p = s->fast_out(i); 1187 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1188 _immutable_memory = p; 1189 return _immutable_memory; 1190 } 1191 } 1192 ShouldNotReachHere(); 1193 return NULL; 1194 } 1195 1196 //----------------------set_cached_top_node------------------------------------ 1197 // Install the cached top node, and make sure Node::is_top works correctly. 1198 void Compile::set_cached_top_node(Node* tn) { 1199 if (tn != NULL) verify_top(tn); 1200 Node* old_top = _top; 1201 _top = tn; 1202 // Calling Node::setup_is_top allows the nodes the chance to adjust 1203 // their _out arrays. 1204 if (_top != NULL) _top->setup_is_top(); 1205 if (old_top != NULL) old_top->setup_is_top(); 1206 assert(_top == NULL || top()->is_top(), ""); 1207 } 1208 1209 #ifdef ASSERT 1210 uint Compile::count_live_nodes_by_graph_walk() { 1211 Unique_Node_List useful(comp_arena()); 1212 // Get useful node list by walking the graph. 1213 identify_useful_nodes(useful); 1214 return useful.size(); 1215 } 1216 1217 void Compile::print_missing_nodes() { 1218 1219 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1220 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1221 return; 1222 } 1223 1224 // This is an expensive function. It is executed only when the user 1225 // specifies VerifyIdealNodeCount option or otherwise knows the 1226 // additional work that needs to be done to identify reachable nodes 1227 // by walking the flow graph and find the missing ones using 1228 // _dead_node_list. 1229 1230 Unique_Node_List useful(comp_arena()); 1231 // Get useful node list by walking the graph. 1232 identify_useful_nodes(useful); 1233 1234 uint l_nodes = C->live_nodes(); 1235 uint l_nodes_by_walk = useful.size(); 1236 1237 if (l_nodes != l_nodes_by_walk) { 1238 if (_log != NULL) { 1239 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1240 _log->stamp(); 1241 _log->end_head(); 1242 } 1243 VectorSet& useful_member_set = useful.member_set(); 1244 int last_idx = l_nodes_by_walk; 1245 for (int i = 0; i < last_idx; i++) { 1246 if (useful_member_set.test(i)) { 1247 if (_dead_node_list.test(i)) { 1248 if (_log != NULL) { 1249 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1250 } 1251 if (PrintIdealNodeCount) { 1252 // Print the log message to tty 1253 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1254 useful.at(i)->dump(); 1255 } 1256 } 1257 } 1258 else if (! _dead_node_list.test(i)) { 1259 if (_log != NULL) { 1260 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1261 } 1262 if (PrintIdealNodeCount) { 1263 // Print the log message to tty 1264 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1265 } 1266 } 1267 } 1268 if (_log != NULL) { 1269 _log->tail("mismatched_nodes"); 1270 } 1271 } 1272 } 1273 void Compile::record_modified_node(Node* n) { 1274 if (_modified_nodes != NULL && !_inlining_incrementally && 1275 n->outcnt() != 0 && !n->is_Con()) { 1276 _modified_nodes->push(n); 1277 } 1278 } 1279 1280 void Compile::remove_modified_node(Node* n) { 1281 if (_modified_nodes != NULL) { 1282 _modified_nodes->remove(n); 1283 } 1284 } 1285 #endif 1286 1287 #ifndef PRODUCT 1288 void Compile::verify_top(Node* tn) const { 1289 if (tn != NULL) { 1290 assert(tn->is_Con(), "top node must be a constant"); 1291 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1292 assert(tn->in(0) != NULL, "must have live top node"); 1293 } 1294 } 1295 #endif 1296 1297 1298 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1299 1300 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1301 guarantee(arr != NULL, ""); 1302 int num_blocks = arr->length(); 1303 if (grow_by < num_blocks) grow_by = num_blocks; 1304 int num_notes = grow_by * _node_notes_block_size; 1305 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1306 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1307 while (num_notes > 0) { 1308 arr->append(notes); 1309 notes += _node_notes_block_size; 1310 num_notes -= _node_notes_block_size; 1311 } 1312 assert(num_notes == 0, "exact multiple, please"); 1313 } 1314 1315 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1316 if (source == NULL || dest == NULL) return false; 1317 1318 if (dest->is_Con()) 1319 return false; // Do not push debug info onto constants. 1320 1321 #ifdef ASSERT 1322 // Leave a bread crumb trail pointing to the original node: 1323 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1324 dest->set_debug_orig(source); 1325 } 1326 #endif 1327 1328 if (node_note_array() == NULL) 1329 return false; // Not collecting any notes now. 1330 1331 // This is a copy onto a pre-existing node, which may already have notes. 1332 // If both nodes have notes, do not overwrite any pre-existing notes. 1333 Node_Notes* source_notes = node_notes_at(source->_idx); 1334 if (source_notes == NULL || source_notes->is_clear()) return false; 1335 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1336 if (dest_notes == NULL || dest_notes->is_clear()) { 1337 return set_node_notes_at(dest->_idx, source_notes); 1338 } 1339 1340 Node_Notes merged_notes = (*source_notes); 1341 // The order of operations here ensures that dest notes will win... 1342 merged_notes.update_from(dest_notes); 1343 return set_node_notes_at(dest->_idx, &merged_notes); 1344 } 1345 1346 1347 //--------------------------allow_range_check_smearing------------------------- 1348 // Gating condition for coalescing similar range checks. 1349 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1350 // single covering check that is at least as strong as any of them. 1351 // If the optimization succeeds, the simplified (strengthened) range check 1352 // will always succeed. If it fails, we will deopt, and then give up 1353 // on the optimization. 1354 bool Compile::allow_range_check_smearing() const { 1355 // If this method has already thrown a range-check, 1356 // assume it was because we already tried range smearing 1357 // and it failed. 1358 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1359 return !already_trapped; 1360 } 1361 1362 1363 //------------------------------flatten_alias_type----------------------------- 1364 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1365 int offset = tj->offset(); 1366 TypePtr::PTR ptr = tj->ptr(); 1367 1368 // Known instance (scalarizable allocation) alias only with itself. 1369 bool is_known_inst = tj->isa_oopptr() != NULL && 1370 tj->is_oopptr()->is_known_instance(); 1371 1372 // Process weird unsafe references. 1373 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1374 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1375 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1376 tj = TypeOopPtr::BOTTOM; 1377 ptr = tj->ptr(); 1378 offset = tj->offset(); 1379 } 1380 1381 // Array pointers need some flattening 1382 const TypeAryPtr *ta = tj->isa_aryptr(); 1383 if (ta && ta->is_stable()) { 1384 // Erase stability property for alias analysis. 1385 tj = ta = ta->cast_to_stable(false); 1386 } 1387 if( ta && is_known_inst ) { 1388 if ( offset != Type::OffsetBot && 1389 offset > arrayOopDesc::length_offset_in_bytes() ) { 1390 offset = Type::OffsetBot; // Flatten constant access into array body only 1391 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1392 } 1393 } else if( ta && _AliasLevel >= 2 ) { 1394 // For arrays indexed by constant indices, we flatten the alias 1395 // space to include all of the array body. Only the header, klass 1396 // and array length can be accessed un-aliased. 1397 if( offset != Type::OffsetBot ) { 1398 if( ta->const_oop() ) { // MethodData* or Method* 1399 offset = Type::OffsetBot; // Flatten constant access into array body 1400 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1401 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1402 // range is OK as-is. 1403 tj = ta = TypeAryPtr::RANGE; 1404 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1405 tj = TypeInstPtr::KLASS; // all klass loads look alike 1406 ta = TypeAryPtr::RANGE; // generic ignored junk 1407 ptr = TypePtr::BotPTR; 1408 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1409 tj = TypeInstPtr::MARK; 1410 ta = TypeAryPtr::RANGE; // generic ignored junk 1411 ptr = TypePtr::BotPTR; 1412 } else { // Random constant offset into array body 1413 offset = Type::OffsetBot; // Flatten constant access into array body 1414 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1415 } 1416 } 1417 // Arrays of fixed size alias with arrays of unknown size. 1418 if (ta->size() != TypeInt::POS) { 1419 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1420 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1421 } 1422 // Arrays of known objects become arrays of unknown objects. 1423 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1424 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1425 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1426 } 1427 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1428 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1429 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1430 } 1431 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1432 // cannot be distinguished by bytecode alone. 1433 if (ta->elem() == TypeInt::BOOL) { 1434 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1435 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1436 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1437 } 1438 // During the 2nd round of IterGVN, NotNull castings are removed. 1439 // Make sure the Bottom and NotNull variants alias the same. 1440 // Also, make sure exact and non-exact variants alias the same. 1441 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1442 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1443 } 1444 } 1445 1446 // Oop pointers need some flattening 1447 const TypeInstPtr *to = tj->isa_instptr(); 1448 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1449 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1450 if( ptr == TypePtr::Constant ) { 1451 if (to->klass() != ciEnv::current()->Class_klass() || 1452 offset < k->size_helper() * wordSize) { 1453 // No constant oop pointers (such as Strings); they alias with 1454 // unknown strings. 1455 assert(!is_known_inst, "not scalarizable allocation"); 1456 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1457 } 1458 } else if( is_known_inst ) { 1459 tj = to; // Keep NotNull and klass_is_exact for instance type 1460 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1461 // During the 2nd round of IterGVN, NotNull castings are removed. 1462 // Make sure the Bottom and NotNull variants alias the same. 1463 // Also, make sure exact and non-exact variants alias the same. 1464 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1465 } 1466 if (to->speculative() != NULL) { 1467 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1468 } 1469 // Canonicalize the holder of this field 1470 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1471 // First handle header references such as a LoadKlassNode, even if the 1472 // object's klass is unloaded at compile time (4965979). 1473 if (!is_known_inst) { // Do it only for non-instance types 1474 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1475 } 1476 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1477 // Static fields are in the space above the normal instance 1478 // fields in the java.lang.Class instance. 1479 if (to->klass() != ciEnv::current()->Class_klass()) { 1480 to = NULL; 1481 tj = TypeOopPtr::BOTTOM; 1482 offset = tj->offset(); 1483 } 1484 } else { 1485 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1486 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1487 if( is_known_inst ) { 1488 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1489 } else { 1490 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1491 } 1492 } 1493 } 1494 } 1495 1496 // Klass pointers to object array klasses need some flattening 1497 const TypeKlassPtr *tk = tj->isa_klassptr(); 1498 if( tk ) { 1499 // If we are referencing a field within a Klass, we need 1500 // to assume the worst case of an Object. Both exact and 1501 // inexact types must flatten to the same alias class so 1502 // use NotNull as the PTR. 1503 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1504 1505 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1506 TypeKlassPtr::OBJECT->klass(), 1507 offset); 1508 } 1509 1510 ciKlass* klass = tk->klass(); 1511 if( klass->is_obj_array_klass() ) { 1512 ciKlass* k = TypeAryPtr::OOPS->klass(); 1513 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1514 k = TypeInstPtr::BOTTOM->klass(); 1515 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1516 } 1517 1518 // Check for precise loads from the primary supertype array and force them 1519 // to the supertype cache alias index. Check for generic array loads from 1520 // the primary supertype array and also force them to the supertype cache 1521 // alias index. Since the same load can reach both, we need to merge 1522 // these 2 disparate memories into the same alias class. Since the 1523 // primary supertype array is read-only, there's no chance of confusion 1524 // where we bypass an array load and an array store. 1525 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1526 if (offset == Type::OffsetBot || 1527 (offset >= primary_supers_offset && 1528 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1529 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1530 offset = in_bytes(Klass::secondary_super_cache_offset()); 1531 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1532 } 1533 } 1534 1535 // Flatten all Raw pointers together. 1536 if (tj->base() == Type::RawPtr) 1537 tj = TypeRawPtr::BOTTOM; 1538 1539 if (tj->base() == Type::AnyPtr) 1540 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1541 1542 // Flatten all to bottom for now 1543 switch( _AliasLevel ) { 1544 case 0: 1545 tj = TypePtr::BOTTOM; 1546 break; 1547 case 1: // Flatten to: oop, static, field or array 1548 switch (tj->base()) { 1549 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1550 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1551 case Type::AryPtr: // do not distinguish arrays at all 1552 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1553 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1554 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1555 default: ShouldNotReachHere(); 1556 } 1557 break; 1558 case 2: // No collapsing at level 2; keep all splits 1559 case 3: // No collapsing at level 3; keep all splits 1560 break; 1561 default: 1562 Unimplemented(); 1563 } 1564 1565 offset = tj->offset(); 1566 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1567 1568 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1569 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1570 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1571 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1572 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1573 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1574 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1575 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1576 assert( tj->ptr() != TypePtr::TopPTR && 1577 tj->ptr() != TypePtr::AnyNull && 1578 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1579 // assert( tj->ptr() != TypePtr::Constant || 1580 // tj->base() == Type::RawPtr || 1581 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1582 1583 return tj; 1584 } 1585 1586 void Compile::AliasType::Init(int i, const TypePtr* at) { 1587 _index = i; 1588 _adr_type = at; 1589 _field = NULL; 1590 _element = NULL; 1591 _is_rewritable = true; // default 1592 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1593 if (atoop != NULL && atoop->is_known_instance()) { 1594 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1595 _general_index = Compile::current()->get_alias_index(gt); 1596 } else { 1597 _general_index = 0; 1598 } 1599 } 1600 1601 //---------------------------------print_on------------------------------------ 1602 #ifndef PRODUCT 1603 void Compile::AliasType::print_on(outputStream* st) { 1604 if (index() < 10) 1605 st->print("@ <%d> ", index()); 1606 else st->print("@ <%d>", index()); 1607 st->print(is_rewritable() ? " " : " RO"); 1608 int offset = adr_type()->offset(); 1609 if (offset == Type::OffsetBot) 1610 st->print(" +any"); 1611 else st->print(" +%-3d", offset); 1612 st->print(" in "); 1613 adr_type()->dump_on(st); 1614 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1615 if (field() != NULL && tjp) { 1616 if (tjp->klass() != field()->holder() || 1617 tjp->offset() != field()->offset_in_bytes()) { 1618 st->print(" != "); 1619 field()->print(); 1620 st->print(" ***"); 1621 } 1622 } 1623 } 1624 1625 void print_alias_types() { 1626 Compile* C = Compile::current(); 1627 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1628 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1629 C->alias_type(idx)->print_on(tty); 1630 tty->cr(); 1631 } 1632 } 1633 #endif 1634 1635 1636 //----------------------------probe_alias_cache-------------------------------- 1637 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1638 intptr_t key = (intptr_t) adr_type; 1639 key ^= key >> logAliasCacheSize; 1640 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1641 } 1642 1643 1644 //-----------------------------grow_alias_types-------------------------------- 1645 void Compile::grow_alias_types() { 1646 const int old_ats = _max_alias_types; // how many before? 1647 const int new_ats = old_ats; // how many more? 1648 const int grow_ats = old_ats+new_ats; // how many now? 1649 _max_alias_types = grow_ats; 1650 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1651 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1652 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1653 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1654 } 1655 1656 1657 //--------------------------------find_alias_type------------------------------ 1658 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1659 if (_AliasLevel == 0) 1660 return alias_type(AliasIdxBot); 1661 1662 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1663 if (ace->_adr_type == adr_type) { 1664 return alias_type(ace->_index); 1665 } 1666 1667 // Handle special cases. 1668 if (adr_type == NULL) return alias_type(AliasIdxTop); 1669 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1670 1671 // Do it the slow way. 1672 const TypePtr* flat = flatten_alias_type(adr_type); 1673 1674 #ifdef ASSERT 1675 assert(flat == flatten_alias_type(flat), "idempotent"); 1676 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1677 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1678 const TypeOopPtr* foop = flat->is_oopptr(); 1679 // Scalarizable allocations have exact klass always. 1680 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1681 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1682 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1683 } 1684 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1685 #endif 1686 1687 int idx = AliasIdxTop; 1688 for (int i = 0; i < num_alias_types(); i++) { 1689 if (alias_type(i)->adr_type() == flat) { 1690 idx = i; 1691 break; 1692 } 1693 } 1694 1695 if (idx == AliasIdxTop) { 1696 if (no_create) return NULL; 1697 // Grow the array if necessary. 1698 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1699 // Add a new alias type. 1700 idx = _num_alias_types++; 1701 _alias_types[idx]->Init(idx, flat); 1702 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1703 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1704 if (flat->isa_instptr()) { 1705 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1706 && flat->is_instptr()->klass() == env()->Class_klass()) 1707 alias_type(idx)->set_rewritable(false); 1708 } 1709 if (flat->isa_aryptr()) { 1710 #ifdef ASSERT 1711 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1712 // (T_BYTE has the weakest alignment and size restrictions...) 1713 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1714 #endif 1715 if (flat->offset() == TypePtr::OffsetBot) { 1716 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1717 } 1718 } 1719 if (flat->isa_klassptr()) { 1720 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1721 alias_type(idx)->set_rewritable(false); 1722 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1723 alias_type(idx)->set_rewritable(false); 1724 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1725 alias_type(idx)->set_rewritable(false); 1726 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1727 alias_type(idx)->set_rewritable(false); 1728 } 1729 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1730 // but the base pointer type is not distinctive enough to identify 1731 // references into JavaThread.) 1732 1733 // Check for final fields. 1734 const TypeInstPtr* tinst = flat->isa_instptr(); 1735 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1736 ciField* field; 1737 if (tinst->const_oop() != NULL && 1738 tinst->klass() == ciEnv::current()->Class_klass() && 1739 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1740 // static field 1741 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1742 field = k->get_field_by_offset(tinst->offset(), true); 1743 } else { 1744 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1745 field = k->get_field_by_offset(tinst->offset(), false); 1746 } 1747 assert(field == NULL || 1748 original_field == NULL || 1749 (field->holder() == original_field->holder() && 1750 field->offset() == original_field->offset() && 1751 field->is_static() == original_field->is_static()), "wrong field?"); 1752 // Set field() and is_rewritable() attributes. 1753 if (field != NULL) alias_type(idx)->set_field(field); 1754 } 1755 } 1756 1757 // Fill the cache for next time. 1758 ace->_adr_type = adr_type; 1759 ace->_index = idx; 1760 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1761 1762 // Might as well try to fill the cache for the flattened version, too. 1763 AliasCacheEntry* face = probe_alias_cache(flat); 1764 if (face->_adr_type == NULL) { 1765 face->_adr_type = flat; 1766 face->_index = idx; 1767 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1768 } 1769 1770 return alias_type(idx); 1771 } 1772 1773 1774 Compile::AliasType* Compile::alias_type(ciField* field) { 1775 const TypeOopPtr* t; 1776 if (field->is_static()) 1777 t = TypeInstPtr::make(field->holder()->java_mirror()); 1778 else 1779 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1780 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1781 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1782 return atp; 1783 } 1784 1785 1786 //------------------------------have_alias_type-------------------------------- 1787 bool Compile::have_alias_type(const TypePtr* adr_type) { 1788 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1789 if (ace->_adr_type == adr_type) { 1790 return true; 1791 } 1792 1793 // Handle special cases. 1794 if (adr_type == NULL) return true; 1795 if (adr_type == TypePtr::BOTTOM) return true; 1796 1797 return find_alias_type(adr_type, true, NULL) != NULL; 1798 } 1799 1800 //-----------------------------must_alias-------------------------------------- 1801 // True if all values of the given address type are in the given alias category. 1802 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1803 if (alias_idx == AliasIdxBot) return true; // the universal category 1804 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1805 if (alias_idx == AliasIdxTop) return false; // the empty category 1806 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1807 1808 // the only remaining possible overlap is identity 1809 int adr_idx = get_alias_index(adr_type); 1810 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1811 assert(adr_idx == alias_idx || 1812 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1813 && adr_type != TypeOopPtr::BOTTOM), 1814 "should not be testing for overlap with an unsafe pointer"); 1815 return adr_idx == alias_idx; 1816 } 1817 1818 //------------------------------can_alias-------------------------------------- 1819 // True if any values of the given address type are in the given alias category. 1820 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1821 if (alias_idx == AliasIdxTop) return false; // the empty category 1822 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1823 if (alias_idx == AliasIdxBot) return true; // the universal category 1824 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1825 1826 // the only remaining possible overlap is identity 1827 int adr_idx = get_alias_index(adr_type); 1828 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1829 return adr_idx == alias_idx; 1830 } 1831 1832 1833 1834 //---------------------------pop_warm_call------------------------------------- 1835 WarmCallInfo* Compile::pop_warm_call() { 1836 WarmCallInfo* wci = _warm_calls; 1837 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1838 return wci; 1839 } 1840 1841 //----------------------------Inline_Warm-------------------------------------- 1842 int Compile::Inline_Warm() { 1843 // If there is room, try to inline some more warm call sites. 1844 // %%% Do a graph index compaction pass when we think we're out of space? 1845 if (!InlineWarmCalls) return 0; 1846 1847 int calls_made_hot = 0; 1848 int room_to_grow = NodeCountInliningCutoff - unique(); 1849 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1850 int amount_grown = 0; 1851 WarmCallInfo* call; 1852 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1853 int est_size = (int)call->size(); 1854 if (est_size > (room_to_grow - amount_grown)) { 1855 // This one won't fit anyway. Get rid of it. 1856 call->make_cold(); 1857 continue; 1858 } 1859 call->make_hot(); 1860 calls_made_hot++; 1861 amount_grown += est_size; 1862 amount_to_grow -= est_size; 1863 } 1864 1865 if (calls_made_hot > 0) set_major_progress(); 1866 return calls_made_hot; 1867 } 1868 1869 1870 //----------------------------Finish_Warm-------------------------------------- 1871 void Compile::Finish_Warm() { 1872 if (!InlineWarmCalls) return; 1873 if (failing()) return; 1874 if (warm_calls() == NULL) return; 1875 1876 // Clean up loose ends, if we are out of space for inlining. 1877 WarmCallInfo* call; 1878 while ((call = pop_warm_call()) != NULL) { 1879 call->make_cold(); 1880 } 1881 } 1882 1883 //---------------------cleanup_loop_predicates----------------------- 1884 // Remove the opaque nodes that protect the predicates so that all unused 1885 // checks and uncommon_traps will be eliminated from the ideal graph 1886 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1887 if (predicate_count()==0) return; 1888 for (int i = predicate_count(); i > 0; i--) { 1889 Node * n = predicate_opaque1_node(i-1); 1890 assert(n->Opcode() == Op_Opaque1, "must be"); 1891 igvn.replace_node(n, n->in(1)); 1892 } 1893 assert(predicate_count()==0, "should be clean!"); 1894 } 1895 1896 // StringOpts and late inlining of string methods 1897 void Compile::inline_string_calls(bool parse_time) { 1898 { 1899 // remove useless nodes to make the usage analysis simpler 1900 ResourceMark rm; 1901 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1902 } 1903 1904 { 1905 ResourceMark rm; 1906 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1907 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1908 print_method(PHASE_AFTER_STRINGOPTS, 3); 1909 } 1910 1911 // now inline anything that we skipped the first time around 1912 if (!parse_time) { 1913 _late_inlines_pos = _late_inlines.length(); 1914 } 1915 1916 while (_string_late_inlines.length() > 0) { 1917 CallGenerator* cg = _string_late_inlines.pop(); 1918 cg->do_late_inline(); 1919 if (failing()) return; 1920 } 1921 _string_late_inlines.trunc_to(0); 1922 } 1923 1924 // Late inlining of boxing methods 1925 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1926 if (_boxing_late_inlines.length() > 0) { 1927 assert(has_boxed_value(), "inconsistent"); 1928 1929 PhaseGVN* gvn = initial_gvn(); 1930 set_inlining_incrementally(true); 1931 1932 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1933 for_igvn()->clear(); 1934 gvn->replace_with(&igvn); 1935 1936 _late_inlines_pos = _late_inlines.length(); 1937 1938 while (_boxing_late_inlines.length() > 0) { 1939 CallGenerator* cg = _boxing_late_inlines.pop(); 1940 cg->do_late_inline(); 1941 if (failing()) return; 1942 } 1943 _boxing_late_inlines.trunc_to(0); 1944 1945 { 1946 ResourceMark rm; 1947 PhaseRemoveUseless pru(gvn, for_igvn()); 1948 } 1949 1950 igvn = PhaseIterGVN(gvn); 1951 igvn.optimize(); 1952 1953 set_inlining_progress(false); 1954 set_inlining_incrementally(false); 1955 } 1956 } 1957 1958 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1959 assert(IncrementalInline, "incremental inlining should be on"); 1960 PhaseGVN* gvn = initial_gvn(); 1961 1962 set_inlining_progress(false); 1963 for_igvn()->clear(); 1964 gvn->replace_with(&igvn); 1965 1966 { 1967 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1968 int i = 0; 1969 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1970 CallGenerator* cg = _late_inlines.at(i); 1971 _late_inlines_pos = i+1; 1972 cg->do_late_inline(); 1973 if (failing()) return; 1974 } 1975 int j = 0; 1976 for (; i < _late_inlines.length(); i++, j++) { 1977 _late_inlines.at_put(j, _late_inlines.at(i)); 1978 } 1979 _late_inlines.trunc_to(j); 1980 } 1981 1982 { 1983 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1984 ResourceMark rm; 1985 PhaseRemoveUseless pru(gvn, for_igvn()); 1986 } 1987 1988 { 1989 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1990 igvn = PhaseIterGVN(gvn); 1991 } 1992 } 1993 1994 // Perform incremental inlining until bound on number of live nodes is reached 1995 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1996 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1997 1998 PhaseGVN* gvn = initial_gvn(); 1999 2000 set_inlining_incrementally(true); 2001 set_inlining_progress(true); 2002 uint low_live_nodes = 0; 2003 2004 while(inlining_progress() && _late_inlines.length() > 0) { 2005 2006 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2007 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2008 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2009 // PhaseIdealLoop is expensive so we only try it once we are 2010 // out of live nodes and we only try it again if the previous 2011 // helped got the number of nodes down significantly 2012 PhaseIdealLoop ideal_loop( igvn, false, true ); 2013 if (failing()) return; 2014 low_live_nodes = live_nodes(); 2015 _major_progress = true; 2016 } 2017 2018 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2019 break; 2020 } 2021 } 2022 2023 inline_incrementally_one(igvn); 2024 2025 if (failing()) return; 2026 2027 { 2028 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2029 igvn.optimize(); 2030 } 2031 2032 if (failing()) return; 2033 } 2034 2035 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2036 2037 if (_string_late_inlines.length() > 0) { 2038 assert(has_stringbuilder(), "inconsistent"); 2039 for_igvn()->clear(); 2040 initial_gvn()->replace_with(&igvn); 2041 2042 inline_string_calls(false); 2043 2044 if (failing()) return; 2045 2046 { 2047 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2048 ResourceMark rm; 2049 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2050 } 2051 2052 { 2053 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2054 igvn = PhaseIterGVN(gvn); 2055 igvn.optimize(); 2056 } 2057 } 2058 2059 set_inlining_incrementally(false); 2060 } 2061 2062 2063 //------------------------------Optimize--------------------------------------- 2064 // Given a graph, optimize it. 2065 void Compile::Optimize() { 2066 TracePhase tp("optimizer", &timers[_t_optimizer]); 2067 2068 #ifndef PRODUCT 2069 if (env()->break_at_compile()) { 2070 BREAKPOINT; 2071 } 2072 2073 #endif 2074 2075 ResourceMark rm; 2076 int loop_opts_cnt; 2077 2078 print_inlining_reinit(); 2079 2080 NOT_PRODUCT( verify_graph_edges(); ) 2081 2082 print_method(PHASE_AFTER_PARSING); 2083 2084 { 2085 // Iterative Global Value Numbering, including ideal transforms 2086 // Initialize IterGVN with types and values from parse-time GVN 2087 PhaseIterGVN igvn(initial_gvn()); 2088 #ifdef ASSERT 2089 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2090 #endif 2091 { 2092 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2093 igvn.optimize(); 2094 } 2095 2096 print_method(PHASE_ITER_GVN1, 2); 2097 2098 if (failing()) return; 2099 2100 inline_incrementally(igvn); 2101 2102 print_method(PHASE_INCREMENTAL_INLINE, 2); 2103 2104 if (failing()) return; 2105 2106 if (eliminate_boxing()) { 2107 // Inline valueOf() methods now. 2108 inline_boxing_calls(igvn); 2109 2110 if (AlwaysIncrementalInline) { 2111 inline_incrementally(igvn); 2112 } 2113 2114 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2115 2116 if (failing()) return; 2117 } 2118 2119 // Remove the speculative part of types and clean up the graph from 2120 // the extra CastPP nodes whose only purpose is to carry them. Do 2121 // that early so that optimizations are not disrupted by the extra 2122 // CastPP nodes. 2123 remove_speculative_types(igvn); 2124 2125 // No more new expensive nodes will be added to the list from here 2126 // so keep only the actual candidates for optimizations. 2127 cleanup_expensive_nodes(igvn); 2128 2129 // Perform escape analysis 2130 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2131 if (has_loops()) { 2132 // Cleanup graph (remove dead nodes). 2133 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2134 PhaseIdealLoop ideal_loop( igvn, false, true ); 2135 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2136 if (failing()) return; 2137 } 2138 ConnectionGraph::do_analysis(this, &igvn); 2139 2140 if (failing()) return; 2141 2142 // Optimize out fields loads from scalar replaceable allocations. 2143 igvn.optimize(); 2144 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2145 2146 if (failing()) return; 2147 2148 if (congraph() != NULL && macro_count() > 0) { 2149 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2150 PhaseMacroExpand mexp(igvn); 2151 mexp.eliminate_macro_nodes(); 2152 igvn.set_delay_transform(false); 2153 2154 igvn.optimize(); 2155 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2156 2157 if (failing()) return; 2158 } 2159 } 2160 2161 // Loop transforms on the ideal graph. Range Check Elimination, 2162 // peeling, unrolling, etc. 2163 2164 // Set loop opts counter 2165 loop_opts_cnt = num_loop_opts(); 2166 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2167 { 2168 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2169 PhaseIdealLoop ideal_loop( igvn, true ); 2170 loop_opts_cnt--; 2171 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2172 if (failing()) return; 2173 } 2174 // Loop opts pass if partial peeling occurred in previous pass 2175 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2176 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2177 PhaseIdealLoop ideal_loop( igvn, false ); 2178 loop_opts_cnt--; 2179 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2180 if (failing()) return; 2181 } 2182 // Loop opts pass for loop-unrolling before CCP 2183 if(major_progress() && (loop_opts_cnt > 0)) { 2184 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2185 PhaseIdealLoop ideal_loop( igvn, false ); 2186 loop_opts_cnt--; 2187 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2188 } 2189 if (!failing()) { 2190 // Verify that last round of loop opts produced a valid graph 2191 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2192 PhaseIdealLoop::verify(igvn); 2193 } 2194 } 2195 if (failing()) return; 2196 2197 // Conditional Constant Propagation; 2198 PhaseCCP ccp( &igvn ); 2199 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2200 { 2201 TracePhase tp("ccp", &timers[_t_ccp]); 2202 ccp.do_transform(); 2203 } 2204 print_method(PHASE_CPP1, 2); 2205 2206 assert( true, "Break here to ccp.dump_old2new_map()"); 2207 2208 // Iterative Global Value Numbering, including ideal transforms 2209 { 2210 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2211 igvn = ccp; 2212 igvn.optimize(); 2213 } 2214 2215 print_method(PHASE_ITER_GVN2, 2); 2216 2217 if (failing()) return; 2218 2219 // Loop transforms on the ideal graph. Range Check Elimination, 2220 // peeling, unrolling, etc. 2221 if(loop_opts_cnt > 0) { 2222 debug_only( int cnt = 0; ); 2223 while(major_progress() && (loop_opts_cnt > 0)) { 2224 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2225 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2226 PhaseIdealLoop ideal_loop( igvn, true); 2227 loop_opts_cnt--; 2228 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2229 if (failing()) return; 2230 } 2231 } 2232 2233 { 2234 // Verify that all previous optimizations produced a valid graph 2235 // at least to this point, even if no loop optimizations were done. 2236 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2237 PhaseIdealLoop::verify(igvn); 2238 } 2239 2240 { 2241 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2242 PhaseMacroExpand mex(igvn); 2243 if (mex.expand_macro_nodes()) { 2244 assert(failing(), "must bail out w/ explicit message"); 2245 return; 2246 } 2247 } 2248 2249 DEBUG_ONLY( _modified_nodes = NULL; ) 2250 } // (End scope of igvn; run destructor if necessary for asserts.) 2251 2252 process_print_inlining(); 2253 // A method with only infinite loops has no edges entering loops from root 2254 { 2255 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2256 if (final_graph_reshaping()) { 2257 assert(failing(), "must bail out w/ explicit message"); 2258 return; 2259 } 2260 } 2261 2262 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2263 } 2264 2265 2266 //------------------------------Code_Gen--------------------------------------- 2267 // Given a graph, generate code for it 2268 void Compile::Code_Gen() { 2269 if (failing()) { 2270 return; 2271 } 2272 2273 // Perform instruction selection. You might think we could reclaim Matcher 2274 // memory PDQ, but actually the Matcher is used in generating spill code. 2275 // Internals of the Matcher (including some VectorSets) must remain live 2276 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2277 // set a bit in reclaimed memory. 2278 2279 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2280 // nodes. Mapping is only valid at the root of each matched subtree. 2281 NOT_PRODUCT( verify_graph_edges(); ) 2282 2283 Matcher matcher; 2284 _matcher = &matcher; 2285 { 2286 TracePhase tp("matcher", &timers[_t_matcher]); 2287 matcher.match(); 2288 } 2289 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2290 // nodes. Mapping is only valid at the root of each matched subtree. 2291 NOT_PRODUCT( verify_graph_edges(); ) 2292 2293 // If you have too many nodes, or if matching has failed, bail out 2294 check_node_count(0, "out of nodes matching instructions"); 2295 if (failing()) { 2296 return; 2297 } 2298 2299 // Build a proper-looking CFG 2300 PhaseCFG cfg(node_arena(), root(), matcher); 2301 _cfg = &cfg; 2302 { 2303 TracePhase tp("scheduler", &timers[_t_scheduler]); 2304 bool success = cfg.do_global_code_motion(); 2305 if (!success) { 2306 return; 2307 } 2308 2309 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2310 NOT_PRODUCT( verify_graph_edges(); ) 2311 debug_only( cfg.verify(); ) 2312 } 2313 2314 PhaseChaitin regalloc(unique(), cfg, matcher); 2315 _regalloc = ®alloc; 2316 { 2317 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2318 // Perform register allocation. After Chaitin, use-def chains are 2319 // no longer accurate (at spill code) and so must be ignored. 2320 // Node->LRG->reg mappings are still accurate. 2321 _regalloc->Register_Allocate(); 2322 2323 // Bail out if the allocator builds too many nodes 2324 if (failing()) { 2325 return; 2326 } 2327 } 2328 2329 // Prior to register allocation we kept empty basic blocks in case the 2330 // the allocator needed a place to spill. After register allocation we 2331 // are not adding any new instructions. If any basic block is empty, we 2332 // can now safely remove it. 2333 { 2334 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2335 cfg.remove_empty_blocks(); 2336 if (do_freq_based_layout()) { 2337 PhaseBlockLayout layout(cfg); 2338 } else { 2339 cfg.set_loop_alignment(); 2340 } 2341 cfg.fixup_flow(); 2342 } 2343 2344 // Apply peephole optimizations 2345 if( OptoPeephole ) { 2346 TracePhase tp("peephole", &timers[_t_peephole]); 2347 PhasePeephole peep( _regalloc, cfg); 2348 peep.do_transform(); 2349 } 2350 2351 // Do late expand if CPU requires this. 2352 if (Matcher::require_postalloc_expand) { 2353 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2354 cfg.postalloc_expand(_regalloc); 2355 } 2356 2357 // Convert Nodes to instruction bits in a buffer 2358 { 2359 TraceTime tp("output", &timers[_t_output], CITime); 2360 Output(); 2361 } 2362 2363 print_method(PHASE_FINAL_CODE); 2364 2365 // He's dead, Jim. 2366 _cfg = (PhaseCFG*)0xdeadbeef; 2367 _regalloc = (PhaseChaitin*)0xdeadbeef; 2368 } 2369 2370 2371 //------------------------------dump_asm--------------------------------------- 2372 // Dump formatted assembly 2373 #ifndef PRODUCT 2374 void Compile::dump_asm(int *pcs, uint pc_limit) { 2375 bool cut_short = false; 2376 tty->print_cr("#"); 2377 tty->print("# "); _tf->dump(); tty->cr(); 2378 tty->print_cr("#"); 2379 2380 // For all blocks 2381 int pc = 0x0; // Program counter 2382 char starts_bundle = ' '; 2383 _regalloc->dump_frame(); 2384 2385 Node *n = NULL; 2386 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2387 if (VMThread::should_terminate()) { 2388 cut_short = true; 2389 break; 2390 } 2391 Block* block = _cfg->get_block(i); 2392 if (block->is_connector() && !Verbose) { 2393 continue; 2394 } 2395 n = block->head(); 2396 if (pcs && n->_idx < pc_limit) { 2397 tty->print("%3.3x ", pcs[n->_idx]); 2398 } else { 2399 tty->print(" "); 2400 } 2401 block->dump_head(_cfg); 2402 if (block->is_connector()) { 2403 tty->print_cr(" # Empty connector block"); 2404 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2405 tty->print_cr(" # Block is sole successor of call"); 2406 } 2407 2408 // For all instructions 2409 Node *delay = NULL; 2410 for (uint j = 0; j < block->number_of_nodes(); j++) { 2411 if (VMThread::should_terminate()) { 2412 cut_short = true; 2413 break; 2414 } 2415 n = block->get_node(j); 2416 if (valid_bundle_info(n)) { 2417 Bundle* bundle = node_bundling(n); 2418 if (bundle->used_in_unconditional_delay()) { 2419 delay = n; 2420 continue; 2421 } 2422 if (bundle->starts_bundle()) { 2423 starts_bundle = '+'; 2424 } 2425 } 2426 2427 if (WizardMode) { 2428 n->dump(); 2429 } 2430 2431 if( !n->is_Region() && // Dont print in the Assembly 2432 !n->is_Phi() && // a few noisely useless nodes 2433 !n->is_Proj() && 2434 !n->is_MachTemp() && 2435 !n->is_SafePointScalarObject() && 2436 !n->is_Catch() && // Would be nice to print exception table targets 2437 !n->is_MergeMem() && // Not very interesting 2438 !n->is_top() && // Debug info table constants 2439 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2440 ) { 2441 if (pcs && n->_idx < pc_limit) 2442 tty->print("%3.3x", pcs[n->_idx]); 2443 else 2444 tty->print(" "); 2445 tty->print(" %c ", starts_bundle); 2446 starts_bundle = ' '; 2447 tty->print("\t"); 2448 n->format(_regalloc, tty); 2449 tty->cr(); 2450 } 2451 2452 // If we have an instruction with a delay slot, and have seen a delay, 2453 // then back up and print it 2454 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2455 assert(delay != NULL, "no unconditional delay instruction"); 2456 if (WizardMode) delay->dump(); 2457 2458 if (node_bundling(delay)->starts_bundle()) 2459 starts_bundle = '+'; 2460 if (pcs && n->_idx < pc_limit) 2461 tty->print("%3.3x", pcs[n->_idx]); 2462 else 2463 tty->print(" "); 2464 tty->print(" %c ", starts_bundle); 2465 starts_bundle = ' '; 2466 tty->print("\t"); 2467 delay->format(_regalloc, tty); 2468 tty->cr(); 2469 delay = NULL; 2470 } 2471 2472 // Dump the exception table as well 2473 if( n->is_Catch() && (Verbose || WizardMode) ) { 2474 // Print the exception table for this offset 2475 _handler_table.print_subtable_for(pc); 2476 } 2477 } 2478 2479 if (pcs && n->_idx < pc_limit) 2480 tty->print_cr("%3.3x", pcs[n->_idx]); 2481 else 2482 tty->cr(); 2483 2484 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2485 2486 } // End of per-block dump 2487 tty->cr(); 2488 2489 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2490 } 2491 #endif 2492 2493 //------------------------------Final_Reshape_Counts--------------------------- 2494 // This class defines counters to help identify when a method 2495 // may/must be executed using hardware with only 24-bit precision. 2496 struct Final_Reshape_Counts : public StackObj { 2497 int _call_count; // count non-inlined 'common' calls 2498 int _float_count; // count float ops requiring 24-bit precision 2499 int _double_count; // count double ops requiring more precision 2500 int _java_call_count; // count non-inlined 'java' calls 2501 int _inner_loop_count; // count loops which need alignment 2502 VectorSet _visited; // Visitation flags 2503 Node_List _tests; // Set of IfNodes & PCTableNodes 2504 2505 Final_Reshape_Counts() : 2506 _call_count(0), _float_count(0), _double_count(0), 2507 _java_call_count(0), _inner_loop_count(0), 2508 _visited( Thread::current()->resource_area() ) { } 2509 2510 void inc_call_count () { _call_count ++; } 2511 void inc_float_count () { _float_count ++; } 2512 void inc_double_count() { _double_count++; } 2513 void inc_java_call_count() { _java_call_count++; } 2514 void inc_inner_loop_count() { _inner_loop_count++; } 2515 2516 int get_call_count () const { return _call_count ; } 2517 int get_float_count () const { return _float_count ; } 2518 int get_double_count() const { return _double_count; } 2519 int get_java_call_count() const { return _java_call_count; } 2520 int get_inner_loop_count() const { return _inner_loop_count; } 2521 }; 2522 2523 #ifdef ASSERT 2524 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2525 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2526 // Make sure the offset goes inside the instance layout. 2527 return k->contains_field_offset(tp->offset()); 2528 // Note that OffsetBot and OffsetTop are very negative. 2529 } 2530 #endif 2531 2532 // Eliminate trivially redundant StoreCMs and accumulate their 2533 // precedence edges. 2534 void Compile::eliminate_redundant_card_marks(Node* n) { 2535 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2536 if (n->in(MemNode::Address)->outcnt() > 1) { 2537 // There are multiple users of the same address so it might be 2538 // possible to eliminate some of the StoreCMs 2539 Node* mem = n->in(MemNode::Memory); 2540 Node* adr = n->in(MemNode::Address); 2541 Node* val = n->in(MemNode::ValueIn); 2542 Node* prev = n; 2543 bool done = false; 2544 // Walk the chain of StoreCMs eliminating ones that match. As 2545 // long as it's a chain of single users then the optimization is 2546 // safe. Eliminating partially redundant StoreCMs would require 2547 // cloning copies down the other paths. 2548 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2549 if (adr == mem->in(MemNode::Address) && 2550 val == mem->in(MemNode::ValueIn)) { 2551 // redundant StoreCM 2552 if (mem->req() > MemNode::OopStore) { 2553 // Hasn't been processed by this code yet. 2554 n->add_prec(mem->in(MemNode::OopStore)); 2555 } else { 2556 // Already converted to precedence edge 2557 for (uint i = mem->req(); i < mem->len(); i++) { 2558 // Accumulate any precedence edges 2559 if (mem->in(i) != NULL) { 2560 n->add_prec(mem->in(i)); 2561 } 2562 } 2563 // Everything above this point has been processed. 2564 done = true; 2565 } 2566 // Eliminate the previous StoreCM 2567 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2568 assert(mem->outcnt() == 0, "should be dead"); 2569 mem->disconnect_inputs(NULL, this); 2570 } else { 2571 prev = mem; 2572 } 2573 mem = prev->in(MemNode::Memory); 2574 } 2575 } 2576 } 2577 2578 //------------------------------final_graph_reshaping_impl---------------------- 2579 // Implement items 1-5 from final_graph_reshaping below. 2580 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2581 2582 if ( n->outcnt() == 0 ) return; // dead node 2583 uint nop = n->Opcode(); 2584 2585 // Check for 2-input instruction with "last use" on right input. 2586 // Swap to left input. Implements item (2). 2587 if( n->req() == 3 && // two-input instruction 2588 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2589 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2590 n->in(2)->outcnt() == 1 &&// right use IS a last use 2591 !n->in(2)->is_Con() ) { // right use is not a constant 2592 // Check for commutative opcode 2593 switch( nop ) { 2594 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2595 case Op_MaxI: case Op_MinI: 2596 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2597 case Op_AndL: case Op_XorL: case Op_OrL: 2598 case Op_AndI: case Op_XorI: case Op_OrI: { 2599 // Move "last use" input to left by swapping inputs 2600 n->swap_edges(1, 2); 2601 break; 2602 } 2603 default: 2604 break; 2605 } 2606 } 2607 2608 #ifdef ASSERT 2609 if( n->is_Mem() ) { 2610 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2611 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2612 // oop will be recorded in oop map if load crosses safepoint 2613 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2614 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2615 "raw memory operations should have control edge"); 2616 } 2617 #endif 2618 // Count FPU ops and common calls, implements item (3) 2619 switch( nop ) { 2620 // Count all float operations that may use FPU 2621 case Op_AddF: 2622 case Op_SubF: 2623 case Op_MulF: 2624 case Op_DivF: 2625 case Op_NegF: 2626 case Op_ModF: 2627 case Op_ConvI2F: 2628 case Op_ConF: 2629 case Op_CmpF: 2630 case Op_CmpF3: 2631 // case Op_ConvL2F: // longs are split into 32-bit halves 2632 frc.inc_float_count(); 2633 break; 2634 2635 case Op_ConvF2D: 2636 case Op_ConvD2F: 2637 frc.inc_float_count(); 2638 frc.inc_double_count(); 2639 break; 2640 2641 // Count all double operations that may use FPU 2642 case Op_AddD: 2643 case Op_SubD: 2644 case Op_MulD: 2645 case Op_DivD: 2646 case Op_NegD: 2647 case Op_ModD: 2648 case Op_ConvI2D: 2649 case Op_ConvD2I: 2650 // case Op_ConvL2D: // handled by leaf call 2651 // case Op_ConvD2L: // handled by leaf call 2652 case Op_ConD: 2653 case Op_CmpD: 2654 case Op_CmpD3: 2655 frc.inc_double_count(); 2656 break; 2657 case Op_Opaque1: // Remove Opaque Nodes before matching 2658 case Op_Opaque2: // Remove Opaque Nodes before matching 2659 case Op_Opaque3: 2660 n->subsume_by(n->in(1), this); 2661 break; 2662 case Op_CallStaticJava: 2663 case Op_CallJava: 2664 case Op_CallDynamicJava: 2665 frc.inc_java_call_count(); // Count java call site; 2666 case Op_CallRuntime: 2667 case Op_CallLeaf: 2668 case Op_CallLeafNoFP: { 2669 assert( n->is_Call(), "" ); 2670 CallNode *call = n->as_Call(); 2671 // Count call sites where the FP mode bit would have to be flipped. 2672 // Do not count uncommon runtime calls: 2673 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2674 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2675 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2676 frc.inc_call_count(); // Count the call site 2677 } else { // See if uncommon argument is shared 2678 Node *n = call->in(TypeFunc::Parms); 2679 int nop = n->Opcode(); 2680 // Clone shared simple arguments to uncommon calls, item (1). 2681 if( n->outcnt() > 1 && 2682 !n->is_Proj() && 2683 nop != Op_CreateEx && 2684 nop != Op_CheckCastPP && 2685 nop != Op_DecodeN && 2686 nop != Op_DecodeNKlass && 2687 !n->is_Mem() ) { 2688 Node *x = n->clone(); 2689 call->set_req( TypeFunc::Parms, x ); 2690 } 2691 } 2692 break; 2693 } 2694 2695 case Op_StoreD: 2696 case Op_LoadD: 2697 case Op_LoadD_unaligned: 2698 frc.inc_double_count(); 2699 goto handle_mem; 2700 case Op_StoreF: 2701 case Op_LoadF: 2702 frc.inc_float_count(); 2703 goto handle_mem; 2704 2705 case Op_StoreCM: 2706 { 2707 // Convert OopStore dependence into precedence edge 2708 Node* prec = n->in(MemNode::OopStore); 2709 n->del_req(MemNode::OopStore); 2710 n->add_prec(prec); 2711 eliminate_redundant_card_marks(n); 2712 } 2713 2714 // fall through 2715 2716 case Op_StoreB: 2717 case Op_StoreC: 2718 case Op_StorePConditional: 2719 case Op_StoreI: 2720 case Op_StoreL: 2721 case Op_StoreIConditional: 2722 case Op_StoreLConditional: 2723 case Op_CompareAndSwapI: 2724 case Op_CompareAndSwapL: 2725 case Op_CompareAndSwapP: 2726 case Op_CompareAndSwapN: 2727 case Op_GetAndAddI: 2728 case Op_GetAndAddL: 2729 case Op_GetAndSetI: 2730 case Op_GetAndSetL: 2731 case Op_GetAndSetP: 2732 case Op_GetAndSetN: 2733 case Op_StoreP: 2734 case Op_StoreN: 2735 case Op_StoreNKlass: 2736 case Op_LoadB: 2737 case Op_LoadUB: 2738 case Op_LoadUS: 2739 case Op_LoadI: 2740 case Op_LoadKlass: 2741 case Op_LoadNKlass: 2742 case Op_LoadL: 2743 case Op_LoadL_unaligned: 2744 case Op_LoadPLocked: 2745 case Op_LoadP: 2746 case Op_LoadN: 2747 case Op_LoadRange: 2748 case Op_LoadS: { 2749 handle_mem: 2750 #ifdef ASSERT 2751 if( VerifyOptoOopOffsets ) { 2752 assert( n->is_Mem(), "" ); 2753 MemNode *mem = (MemNode*)n; 2754 // Check to see if address types have grounded out somehow. 2755 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2756 assert( !tp || oop_offset_is_sane(tp), "" ); 2757 } 2758 #endif 2759 break; 2760 } 2761 2762 case Op_AddP: { // Assert sane base pointers 2763 Node *addp = n->in(AddPNode::Address); 2764 assert( !addp->is_AddP() || 2765 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2766 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2767 "Base pointers must match" ); 2768 #ifdef _LP64 2769 if ((UseCompressedOops || UseCompressedClassPointers) && 2770 addp->Opcode() == Op_ConP && 2771 addp == n->in(AddPNode::Base) && 2772 n->in(AddPNode::Offset)->is_Con()) { 2773 // Use addressing with narrow klass to load with offset on x86. 2774 // On sparc loading 32-bits constant and decoding it have less 2775 // instructions (4) then load 64-bits constant (7). 2776 // Do this transformation here since IGVN will convert ConN back to ConP. 2777 const Type* t = addp->bottom_type(); 2778 if (t->isa_oopptr() || t->isa_klassptr()) { 2779 Node* nn = NULL; 2780 2781 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2782 2783 // Look for existing ConN node of the same exact type. 2784 Node* r = root(); 2785 uint cnt = r->outcnt(); 2786 for (uint i = 0; i < cnt; i++) { 2787 Node* m = r->raw_out(i); 2788 if (m!= NULL && m->Opcode() == op && 2789 m->bottom_type()->make_ptr() == t) { 2790 nn = m; 2791 break; 2792 } 2793 } 2794 if (nn != NULL) { 2795 // Decode a narrow oop to match address 2796 // [R12 + narrow_oop_reg<<3 + offset] 2797 if (t->isa_oopptr()) { 2798 nn = new DecodeNNode(nn, t); 2799 } else { 2800 nn = new DecodeNKlassNode(nn, t); 2801 } 2802 n->set_req(AddPNode::Base, nn); 2803 n->set_req(AddPNode::Address, nn); 2804 if (addp->outcnt() == 0) { 2805 addp->disconnect_inputs(NULL, this); 2806 } 2807 } 2808 } 2809 } 2810 #endif 2811 break; 2812 } 2813 2814 #ifdef _LP64 2815 case Op_CastPP: 2816 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2817 Node* in1 = n->in(1); 2818 const Type* t = n->bottom_type(); 2819 Node* new_in1 = in1->clone(); 2820 new_in1->as_DecodeN()->set_type(t); 2821 2822 if (!Matcher::narrow_oop_use_complex_address()) { 2823 // 2824 // x86, ARM and friends can handle 2 adds in addressing mode 2825 // and Matcher can fold a DecodeN node into address by using 2826 // a narrow oop directly and do implicit NULL check in address: 2827 // 2828 // [R12 + narrow_oop_reg<<3 + offset] 2829 // NullCheck narrow_oop_reg 2830 // 2831 // On other platforms (Sparc) we have to keep new DecodeN node and 2832 // use it to do implicit NULL check in address: 2833 // 2834 // decode_not_null narrow_oop_reg, base_reg 2835 // [base_reg + offset] 2836 // NullCheck base_reg 2837 // 2838 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2839 // to keep the information to which NULL check the new DecodeN node 2840 // corresponds to use it as value in implicit_null_check(). 2841 // 2842 new_in1->set_req(0, n->in(0)); 2843 } 2844 2845 n->subsume_by(new_in1, this); 2846 if (in1->outcnt() == 0) { 2847 in1->disconnect_inputs(NULL, this); 2848 } 2849 } 2850 break; 2851 2852 case Op_CmpP: 2853 // Do this transformation here to preserve CmpPNode::sub() and 2854 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2855 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2856 Node* in1 = n->in(1); 2857 Node* in2 = n->in(2); 2858 if (!in1->is_DecodeNarrowPtr()) { 2859 in2 = in1; 2860 in1 = n->in(2); 2861 } 2862 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2863 2864 Node* new_in2 = NULL; 2865 if (in2->is_DecodeNarrowPtr()) { 2866 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2867 new_in2 = in2->in(1); 2868 } else if (in2->Opcode() == Op_ConP) { 2869 const Type* t = in2->bottom_type(); 2870 if (t == TypePtr::NULL_PTR) { 2871 assert(in1->is_DecodeN(), "compare klass to null?"); 2872 // Don't convert CmpP null check into CmpN if compressed 2873 // oops implicit null check is not generated. 2874 // This will allow to generate normal oop implicit null check. 2875 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2876 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2877 // 2878 // This transformation together with CastPP transformation above 2879 // will generated code for implicit NULL checks for compressed oops. 2880 // 2881 // The original code after Optimize() 2882 // 2883 // LoadN memory, narrow_oop_reg 2884 // decode narrow_oop_reg, base_reg 2885 // CmpP base_reg, NULL 2886 // CastPP base_reg // NotNull 2887 // Load [base_reg + offset], val_reg 2888 // 2889 // after these transformations will be 2890 // 2891 // LoadN memory, narrow_oop_reg 2892 // CmpN narrow_oop_reg, NULL 2893 // decode_not_null narrow_oop_reg, base_reg 2894 // Load [base_reg + offset], val_reg 2895 // 2896 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2897 // since narrow oops can be used in debug info now (see the code in 2898 // final_graph_reshaping_walk()). 2899 // 2900 // At the end the code will be matched to 2901 // on x86: 2902 // 2903 // Load_narrow_oop memory, narrow_oop_reg 2904 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2905 // NullCheck narrow_oop_reg 2906 // 2907 // and on sparc: 2908 // 2909 // Load_narrow_oop memory, narrow_oop_reg 2910 // decode_not_null narrow_oop_reg, base_reg 2911 // Load [base_reg + offset], val_reg 2912 // NullCheck base_reg 2913 // 2914 } else if (t->isa_oopptr()) { 2915 new_in2 = ConNode::make(t->make_narrowoop()); 2916 } else if (t->isa_klassptr()) { 2917 new_in2 = ConNode::make(t->make_narrowklass()); 2918 } 2919 } 2920 if (new_in2 != NULL) { 2921 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2922 n->subsume_by(cmpN, this); 2923 if (in1->outcnt() == 0) { 2924 in1->disconnect_inputs(NULL, this); 2925 } 2926 if (in2->outcnt() == 0) { 2927 in2->disconnect_inputs(NULL, this); 2928 } 2929 } 2930 } 2931 break; 2932 2933 case Op_DecodeN: 2934 case Op_DecodeNKlass: 2935 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2936 // DecodeN could be pinned when it can't be fold into 2937 // an address expression, see the code for Op_CastPP above. 2938 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2939 break; 2940 2941 case Op_EncodeP: 2942 case Op_EncodePKlass: { 2943 Node* in1 = n->in(1); 2944 if (in1->is_DecodeNarrowPtr()) { 2945 n->subsume_by(in1->in(1), this); 2946 } else if (in1->Opcode() == Op_ConP) { 2947 const Type* t = in1->bottom_type(); 2948 if (t == TypePtr::NULL_PTR) { 2949 assert(t->isa_oopptr(), "null klass?"); 2950 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 2951 } else if (t->isa_oopptr()) { 2952 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 2953 } else if (t->isa_klassptr()) { 2954 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 2955 } 2956 } 2957 if (in1->outcnt() == 0) { 2958 in1->disconnect_inputs(NULL, this); 2959 } 2960 break; 2961 } 2962 2963 case Op_Proj: { 2964 if (OptimizeStringConcat) { 2965 ProjNode* p = n->as_Proj(); 2966 if (p->_is_io_use) { 2967 // Separate projections were used for the exception path which 2968 // are normally removed by a late inline. If it wasn't inlined 2969 // then they will hang around and should just be replaced with 2970 // the original one. 2971 Node* proj = NULL; 2972 // Replace with just one 2973 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2974 Node *use = i.get(); 2975 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2976 proj = use; 2977 break; 2978 } 2979 } 2980 assert(proj != NULL, "must be found"); 2981 p->subsume_by(proj, this); 2982 } 2983 } 2984 break; 2985 } 2986 2987 case Op_Phi: 2988 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2989 // The EncodeP optimization may create Phi with the same edges 2990 // for all paths. It is not handled well by Register Allocator. 2991 Node* unique_in = n->in(1); 2992 assert(unique_in != NULL, ""); 2993 uint cnt = n->req(); 2994 for (uint i = 2; i < cnt; i++) { 2995 Node* m = n->in(i); 2996 assert(m != NULL, ""); 2997 if (unique_in != m) 2998 unique_in = NULL; 2999 } 3000 if (unique_in != NULL) { 3001 n->subsume_by(unique_in, this); 3002 } 3003 } 3004 break; 3005 3006 #endif 3007 3008 case Op_ModI: 3009 if (UseDivMod) { 3010 // Check if a%b and a/b both exist 3011 Node* d = n->find_similar(Op_DivI); 3012 if (d) { 3013 // Replace them with a fused divmod if supported 3014 if (Matcher::has_match_rule(Op_DivModI)) { 3015 DivModINode* divmod = DivModINode::make(n); 3016 d->subsume_by(divmod->div_proj(), this); 3017 n->subsume_by(divmod->mod_proj(), this); 3018 } else { 3019 // replace a%b with a-((a/b)*b) 3020 Node* mult = new MulINode(d, d->in(2)); 3021 Node* sub = new SubINode(d->in(1), mult); 3022 n->subsume_by(sub, this); 3023 } 3024 } 3025 } 3026 break; 3027 3028 case Op_ModL: 3029 if (UseDivMod) { 3030 // Check if a%b and a/b both exist 3031 Node* d = n->find_similar(Op_DivL); 3032 if (d) { 3033 // Replace them with a fused divmod if supported 3034 if (Matcher::has_match_rule(Op_DivModL)) { 3035 DivModLNode* divmod = DivModLNode::make(n); 3036 d->subsume_by(divmod->div_proj(), this); 3037 n->subsume_by(divmod->mod_proj(), this); 3038 } else { 3039 // replace a%b with a-((a/b)*b) 3040 Node* mult = new MulLNode(d, d->in(2)); 3041 Node* sub = new SubLNode(d->in(1), mult); 3042 n->subsume_by(sub, this); 3043 } 3044 } 3045 } 3046 break; 3047 3048 case Op_LoadVector: 3049 case Op_StoreVector: 3050 break; 3051 3052 case Op_PackB: 3053 case Op_PackS: 3054 case Op_PackI: 3055 case Op_PackF: 3056 case Op_PackL: 3057 case Op_PackD: 3058 if (n->req()-1 > 2) { 3059 // Replace many operand PackNodes with a binary tree for matching 3060 PackNode* p = (PackNode*) n; 3061 Node* btp = p->binary_tree_pack(1, n->req()); 3062 n->subsume_by(btp, this); 3063 } 3064 break; 3065 case Op_Loop: 3066 case Op_CountedLoop: 3067 if (n->as_Loop()->is_inner_loop()) { 3068 frc.inc_inner_loop_count(); 3069 } 3070 break; 3071 case Op_LShiftI: 3072 case Op_RShiftI: 3073 case Op_URShiftI: 3074 case Op_LShiftL: 3075 case Op_RShiftL: 3076 case Op_URShiftL: 3077 if (Matcher::need_masked_shift_count) { 3078 // The cpu's shift instructions don't restrict the count to the 3079 // lower 5/6 bits. We need to do the masking ourselves. 3080 Node* in2 = n->in(2); 3081 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3082 const TypeInt* t = in2->find_int_type(); 3083 if (t != NULL && t->is_con()) { 3084 juint shift = t->get_con(); 3085 if (shift > mask) { // Unsigned cmp 3086 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3087 } 3088 } else { 3089 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3090 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3091 n->set_req(2, shift); 3092 } 3093 } 3094 if (in2->outcnt() == 0) { // Remove dead node 3095 in2->disconnect_inputs(NULL, this); 3096 } 3097 } 3098 break; 3099 case Op_MemBarStoreStore: 3100 case Op_MemBarRelease: 3101 // Break the link with AllocateNode: it is no longer useful and 3102 // confuses register allocation. 3103 if (n->req() > MemBarNode::Precedent) { 3104 n->set_req(MemBarNode::Precedent, top()); 3105 } 3106 break; 3107 default: 3108 assert( !n->is_Call(), "" ); 3109 assert( !n->is_Mem(), "" ); 3110 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3111 break; 3112 } 3113 3114 // Collect CFG split points 3115 if (n->is_MultiBranch()) 3116 frc._tests.push(n); 3117 } 3118 3119 //------------------------------final_graph_reshaping_walk--------------------- 3120 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3121 // requires that the walk visits a node's inputs before visiting the node. 3122 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3123 ResourceArea *area = Thread::current()->resource_area(); 3124 Unique_Node_List sfpt(area); 3125 3126 frc._visited.set(root->_idx); // first, mark node as visited 3127 uint cnt = root->req(); 3128 Node *n = root; 3129 uint i = 0; 3130 while (true) { 3131 if (i < cnt) { 3132 // Place all non-visited non-null inputs onto stack 3133 Node* m = n->in(i); 3134 ++i; 3135 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3136 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3137 // compute worst case interpreter size in case of a deoptimization 3138 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3139 3140 sfpt.push(m); 3141 } 3142 cnt = m->req(); 3143 nstack.push(n, i); // put on stack parent and next input's index 3144 n = m; 3145 i = 0; 3146 } 3147 } else { 3148 // Now do post-visit work 3149 final_graph_reshaping_impl( n, frc ); 3150 if (nstack.is_empty()) 3151 break; // finished 3152 n = nstack.node(); // Get node from stack 3153 cnt = n->req(); 3154 i = nstack.index(); 3155 nstack.pop(); // Shift to the next node on stack 3156 } 3157 } 3158 3159 // Skip next transformation if compressed oops are not used. 3160 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3161 (!UseCompressedOops && !UseCompressedClassPointers)) 3162 return; 3163 3164 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3165 // It could be done for an uncommon traps or any safepoints/calls 3166 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3167 while (sfpt.size() > 0) { 3168 n = sfpt.pop(); 3169 JVMState *jvms = n->as_SafePoint()->jvms(); 3170 assert(jvms != NULL, "sanity"); 3171 int start = jvms->debug_start(); 3172 int end = n->req(); 3173 bool is_uncommon = (n->is_CallStaticJava() && 3174 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3175 for (int j = start; j < end; j++) { 3176 Node* in = n->in(j); 3177 if (in->is_DecodeNarrowPtr()) { 3178 bool safe_to_skip = true; 3179 if (!is_uncommon ) { 3180 // Is it safe to skip? 3181 for (uint i = 0; i < in->outcnt(); i++) { 3182 Node* u = in->raw_out(i); 3183 if (!u->is_SafePoint() || 3184 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3185 safe_to_skip = false; 3186 } 3187 } 3188 } 3189 if (safe_to_skip) { 3190 n->set_req(j, in->in(1)); 3191 } 3192 if (in->outcnt() == 0) { 3193 in->disconnect_inputs(NULL, this); 3194 } 3195 } 3196 } 3197 } 3198 } 3199 3200 //------------------------------final_graph_reshaping-------------------------- 3201 // Final Graph Reshaping. 3202 // 3203 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3204 // and not commoned up and forced early. Must come after regular 3205 // optimizations to avoid GVN undoing the cloning. Clone constant 3206 // inputs to Loop Phis; these will be split by the allocator anyways. 3207 // Remove Opaque nodes. 3208 // (2) Move last-uses by commutative operations to the left input to encourage 3209 // Intel update-in-place two-address operations and better register usage 3210 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3211 // calls canonicalizing them back. 3212 // (3) Count the number of double-precision FP ops, single-precision FP ops 3213 // and call sites. On Intel, we can get correct rounding either by 3214 // forcing singles to memory (requires extra stores and loads after each 3215 // FP bytecode) or we can set a rounding mode bit (requires setting and 3216 // clearing the mode bit around call sites). The mode bit is only used 3217 // if the relative frequency of single FP ops to calls is low enough. 3218 // This is a key transform for SPEC mpeg_audio. 3219 // (4) Detect infinite loops; blobs of code reachable from above but not 3220 // below. Several of the Code_Gen algorithms fail on such code shapes, 3221 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3222 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3223 // Detection is by looking for IfNodes where only 1 projection is 3224 // reachable from below or CatchNodes missing some targets. 3225 // (5) Assert for insane oop offsets in debug mode. 3226 3227 bool Compile::final_graph_reshaping() { 3228 // an infinite loop may have been eliminated by the optimizer, 3229 // in which case the graph will be empty. 3230 if (root()->req() == 1) { 3231 record_method_not_compilable("trivial infinite loop"); 3232 return true; 3233 } 3234 3235 // Expensive nodes have their control input set to prevent the GVN 3236 // from freely commoning them. There's no GVN beyond this point so 3237 // no need to keep the control input. We want the expensive nodes to 3238 // be freely moved to the least frequent code path by gcm. 3239 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3240 for (int i = 0; i < expensive_count(); i++) { 3241 _expensive_nodes->at(i)->set_req(0, NULL); 3242 } 3243 3244 Final_Reshape_Counts frc; 3245 3246 // Visit everybody reachable! 3247 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3248 Node_Stack nstack(unique() >> 1); 3249 final_graph_reshaping_walk(nstack, root(), frc); 3250 3251 // Check for unreachable (from below) code (i.e., infinite loops). 3252 for( uint i = 0; i < frc._tests.size(); i++ ) { 3253 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3254 // Get number of CFG targets. 3255 // Note that PCTables include exception targets after calls. 3256 uint required_outcnt = n->required_outcnt(); 3257 if (n->outcnt() != required_outcnt) { 3258 // Check for a few special cases. Rethrow Nodes never take the 3259 // 'fall-thru' path, so expected kids is 1 less. 3260 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3261 if (n->in(0)->in(0)->is_Call()) { 3262 CallNode *call = n->in(0)->in(0)->as_Call(); 3263 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3264 required_outcnt--; // Rethrow always has 1 less kid 3265 } else if (call->req() > TypeFunc::Parms && 3266 call->is_CallDynamicJava()) { 3267 // Check for null receiver. In such case, the optimizer has 3268 // detected that the virtual call will always result in a null 3269 // pointer exception. The fall-through projection of this CatchNode 3270 // will not be populated. 3271 Node *arg0 = call->in(TypeFunc::Parms); 3272 if (arg0->is_Type() && 3273 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3274 required_outcnt--; 3275 } 3276 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3277 call->req() > TypeFunc::Parms+1 && 3278 call->is_CallStaticJava()) { 3279 // Check for negative array length. In such case, the optimizer has 3280 // detected that the allocation attempt will always result in an 3281 // exception. There is no fall-through projection of this CatchNode . 3282 Node *arg1 = call->in(TypeFunc::Parms+1); 3283 if (arg1->is_Type() && 3284 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3285 required_outcnt--; 3286 } 3287 } 3288 } 3289 } 3290 // Recheck with a better notion of 'required_outcnt' 3291 if (n->outcnt() != required_outcnt) { 3292 record_method_not_compilable("malformed control flow"); 3293 return true; // Not all targets reachable! 3294 } 3295 } 3296 // Check that I actually visited all kids. Unreached kids 3297 // must be infinite loops. 3298 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3299 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3300 record_method_not_compilable("infinite loop"); 3301 return true; // Found unvisited kid; must be unreach 3302 } 3303 } 3304 3305 // If original bytecodes contained a mixture of floats and doubles 3306 // check if the optimizer has made it homogenous, item (3). 3307 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3308 frc.get_float_count() > 32 && 3309 frc.get_double_count() == 0 && 3310 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3311 set_24_bit_selection_and_mode( false, true ); 3312 } 3313 3314 set_java_calls(frc.get_java_call_count()); 3315 set_inner_loops(frc.get_inner_loop_count()); 3316 3317 // No infinite loops, no reason to bail out. 3318 return false; 3319 } 3320 3321 //-----------------------------too_many_traps---------------------------------- 3322 // Report if there are too many traps at the current method and bci. 3323 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3324 bool Compile::too_many_traps(ciMethod* method, 3325 int bci, 3326 Deoptimization::DeoptReason reason) { 3327 if (method->has_injected_profile()) { 3328 return false; 3329 } 3330 ciMethodData* md = method->method_data(); 3331 if (md->is_empty()) { 3332 // Assume the trap has not occurred, or that it occurred only 3333 // because of a transient condition during start-up in the interpreter. 3334 return false; 3335 } 3336 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3337 if (md->has_trap_at(bci, m, reason) != 0) { 3338 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3339 // Also, if there are multiple reasons, or if there is no per-BCI record, 3340 // assume the worst. 3341 if (log()) 3342 log()->elem("observe trap='%s' count='%d'", 3343 Deoptimization::trap_reason_name(reason), 3344 md->trap_count(reason)); 3345 return true; 3346 } else { 3347 // Ignore method/bci and see if there have been too many globally. 3348 return too_many_traps(reason, md); 3349 } 3350 } 3351 3352 // Less-accurate variant which does not require a method and bci. 3353 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3354 ciMethodData* logmd) { 3355 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3356 // Too many traps globally. 3357 // Note that we use cumulative trap_count, not just md->trap_count. 3358 if (log()) { 3359 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3360 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3361 Deoptimization::trap_reason_name(reason), 3362 mcount, trap_count(reason)); 3363 } 3364 return true; 3365 } else { 3366 // The coast is clear. 3367 return false; 3368 } 3369 } 3370 3371 //--------------------------too_many_recompiles-------------------------------- 3372 // Report if there are too many recompiles at the current method and bci. 3373 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3374 // Is not eager to return true, since this will cause the compiler to use 3375 // Action_none for a trap point, to avoid too many recompilations. 3376 bool Compile::too_many_recompiles(ciMethod* method, 3377 int bci, 3378 Deoptimization::DeoptReason reason) { 3379 if (method->has_injected_profile()) { 3380 return false; 3381 } 3382 ciMethodData* md = method->method_data(); 3383 if (md->is_empty()) { 3384 // Assume the trap has not occurred, or that it occurred only 3385 // because of a transient condition during start-up in the interpreter. 3386 return false; 3387 } 3388 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3389 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3390 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3391 Deoptimization::DeoptReason per_bc_reason 3392 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3393 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3394 if ((per_bc_reason == Deoptimization::Reason_none 3395 || md->has_trap_at(bci, m, reason) != 0) 3396 // The trap frequency measure we care about is the recompile count: 3397 && md->trap_recompiled_at(bci, m) 3398 && md->overflow_recompile_count() >= bc_cutoff) { 3399 // Do not emit a trap here if it has already caused recompilations. 3400 // Also, if there are multiple reasons, or if there is no per-BCI record, 3401 // assume the worst. 3402 if (log()) 3403 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3404 Deoptimization::trap_reason_name(reason), 3405 md->trap_count(reason), 3406 md->overflow_recompile_count()); 3407 return true; 3408 } else if (trap_count(reason) != 0 3409 && decompile_count() >= m_cutoff) { 3410 // Too many recompiles globally, and we have seen this sort of trap. 3411 // Use cumulative decompile_count, not just md->decompile_count. 3412 if (log()) 3413 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3414 Deoptimization::trap_reason_name(reason), 3415 md->trap_count(reason), trap_count(reason), 3416 md->decompile_count(), decompile_count()); 3417 return true; 3418 } else { 3419 // The coast is clear. 3420 return false; 3421 } 3422 } 3423 3424 // Compute when not to trap. Used by matching trap based nodes and 3425 // NullCheck optimization. 3426 void Compile::set_allowed_deopt_reasons() { 3427 _allowed_reasons = 0; 3428 if (is_method_compilation()) { 3429 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3430 assert(rs < BitsPerInt, "recode bit map"); 3431 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3432 _allowed_reasons |= nth_bit(rs); 3433 } 3434 } 3435 } 3436 } 3437 3438 #ifndef PRODUCT 3439 //------------------------------verify_graph_edges--------------------------- 3440 // Walk the Graph and verify that there is a one-to-one correspondence 3441 // between Use-Def edges and Def-Use edges in the graph. 3442 void Compile::verify_graph_edges(bool no_dead_code) { 3443 if (VerifyGraphEdges) { 3444 ResourceArea *area = Thread::current()->resource_area(); 3445 Unique_Node_List visited(area); 3446 // Call recursive graph walk to check edges 3447 _root->verify_edges(visited); 3448 if (no_dead_code) { 3449 // Now make sure that no visited node is used by an unvisited node. 3450 bool dead_nodes = false; 3451 Unique_Node_List checked(area); 3452 while (visited.size() > 0) { 3453 Node* n = visited.pop(); 3454 checked.push(n); 3455 for (uint i = 0; i < n->outcnt(); i++) { 3456 Node* use = n->raw_out(i); 3457 if (checked.member(use)) continue; // already checked 3458 if (visited.member(use)) continue; // already in the graph 3459 if (use->is_Con()) continue; // a dead ConNode is OK 3460 // At this point, we have found a dead node which is DU-reachable. 3461 if (!dead_nodes) { 3462 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3463 dead_nodes = true; 3464 } 3465 use->dump(2); 3466 tty->print_cr("---"); 3467 checked.push(use); // No repeats; pretend it is now checked. 3468 } 3469 } 3470 assert(!dead_nodes, "using nodes must be reachable from root"); 3471 } 3472 } 3473 } 3474 3475 // Verify GC barriers consistency 3476 // Currently supported: 3477 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3478 void Compile::verify_barriers() { 3479 if (UseG1GC) { 3480 // Verify G1 pre-barriers 3481 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3482 3483 ResourceArea *area = Thread::current()->resource_area(); 3484 Unique_Node_List visited(area); 3485 Node_List worklist(area); 3486 // We're going to walk control flow backwards starting from the Root 3487 worklist.push(_root); 3488 while (worklist.size() > 0) { 3489 Node* x = worklist.pop(); 3490 if (x == NULL || x == top()) continue; 3491 if (visited.member(x)) { 3492 continue; 3493 } else { 3494 visited.push(x); 3495 } 3496 3497 if (x->is_Region()) { 3498 for (uint i = 1; i < x->req(); i++) { 3499 worklist.push(x->in(i)); 3500 } 3501 } else { 3502 worklist.push(x->in(0)); 3503 // We are looking for the pattern: 3504 // /->ThreadLocal 3505 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3506 // \->ConI(0) 3507 // We want to verify that the If and the LoadB have the same control 3508 // See GraphKit::g1_write_barrier_pre() 3509 if (x->is_If()) { 3510 IfNode *iff = x->as_If(); 3511 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3512 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3513 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3514 && cmp->in(1)->is_Load()) { 3515 LoadNode* load = cmp->in(1)->as_Load(); 3516 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3517 && load->in(2)->in(3)->is_Con() 3518 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3519 3520 Node* if_ctrl = iff->in(0); 3521 Node* load_ctrl = load->in(0); 3522 3523 if (if_ctrl != load_ctrl) { 3524 // Skip possible CProj->NeverBranch in infinite loops 3525 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3526 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3527 if_ctrl = if_ctrl->in(0)->in(0); 3528 } 3529 } 3530 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3531 } 3532 } 3533 } 3534 } 3535 } 3536 } 3537 } 3538 } 3539 3540 #endif 3541 3542 // The Compile object keeps track of failure reasons separately from the ciEnv. 3543 // This is required because there is not quite a 1-1 relation between the 3544 // ciEnv and its compilation task and the Compile object. Note that one 3545 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3546 // to backtrack and retry without subsuming loads. Other than this backtracking 3547 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3548 // by the logic in C2Compiler. 3549 void Compile::record_failure(const char* reason) { 3550 if (log() != NULL) { 3551 log()->elem("failure reason='%s' phase='compile'", reason); 3552 } 3553 if (_failure_reason == NULL) { 3554 // Record the first failure reason. 3555 _failure_reason = reason; 3556 } 3557 3558 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3559 C->print_method(PHASE_FAILURE); 3560 } 3561 _root = NULL; // flush the graph, too 3562 } 3563 3564 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3565 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3566 _phase_name(name), _dolog(CITimeVerbose) 3567 { 3568 if (_dolog) { 3569 C = Compile::current(); 3570 _log = C->log(); 3571 } else { 3572 C = NULL; 3573 _log = NULL; 3574 } 3575 if (_log != NULL) { 3576 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3577 _log->stamp(); 3578 _log->end_head(); 3579 } 3580 } 3581 3582 Compile::TracePhase::~TracePhase() { 3583 3584 C = Compile::current(); 3585 if (_dolog) { 3586 _log = C->log(); 3587 } else { 3588 _log = NULL; 3589 } 3590 3591 #ifdef ASSERT 3592 if (PrintIdealNodeCount) { 3593 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3594 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3595 } 3596 3597 if (VerifyIdealNodeCount) { 3598 Compile::current()->print_missing_nodes(); 3599 } 3600 #endif 3601 3602 if (_log != NULL) { 3603 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3604 } 3605 } 3606 3607 //============================================================================= 3608 // Two Constant's are equal when the type and the value are equal. 3609 bool Compile::Constant::operator==(const Constant& other) { 3610 if (type() != other.type() ) return false; 3611 if (can_be_reused() != other.can_be_reused()) return false; 3612 // For floating point values we compare the bit pattern. 3613 switch (type()) { 3614 case T_FLOAT: return (_v._value.i == other._v._value.i); 3615 case T_LONG: 3616 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3617 case T_OBJECT: 3618 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3619 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3620 case T_METADATA: return (_v._metadata == other._v._metadata); 3621 default: ShouldNotReachHere(); 3622 } 3623 return false; 3624 } 3625 3626 static int type_to_size_in_bytes(BasicType t) { 3627 switch (t) { 3628 case T_LONG: return sizeof(jlong ); 3629 case T_FLOAT: return sizeof(jfloat ); 3630 case T_DOUBLE: return sizeof(jdouble); 3631 case T_METADATA: return sizeof(Metadata*); 3632 // We use T_VOID as marker for jump-table entries (labels) which 3633 // need an internal word relocation. 3634 case T_VOID: 3635 case T_ADDRESS: 3636 case T_OBJECT: return sizeof(jobject); 3637 } 3638 3639 ShouldNotReachHere(); 3640 return -1; 3641 } 3642 3643 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3644 // sort descending 3645 if (a->freq() > b->freq()) return -1; 3646 if (a->freq() < b->freq()) return 1; 3647 return 0; 3648 } 3649 3650 void Compile::ConstantTable::calculate_offsets_and_size() { 3651 // First, sort the array by frequencies. 3652 _constants.sort(qsort_comparator); 3653 3654 #ifdef ASSERT 3655 // Make sure all jump-table entries were sorted to the end of the 3656 // array (they have a negative frequency). 3657 bool found_void = false; 3658 for (int i = 0; i < _constants.length(); i++) { 3659 Constant con = _constants.at(i); 3660 if (con.type() == T_VOID) 3661 found_void = true; // jump-tables 3662 else 3663 assert(!found_void, "wrong sorting"); 3664 } 3665 #endif 3666 3667 int offset = 0; 3668 for (int i = 0; i < _constants.length(); i++) { 3669 Constant* con = _constants.adr_at(i); 3670 3671 // Align offset for type. 3672 int typesize = type_to_size_in_bytes(con->type()); 3673 offset = align_size_up(offset, typesize); 3674 con->set_offset(offset); // set constant's offset 3675 3676 if (con->type() == T_VOID) { 3677 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3678 offset = offset + typesize * n->outcnt(); // expand jump-table 3679 } else { 3680 offset = offset + typesize; 3681 } 3682 } 3683 3684 // Align size up to the next section start (which is insts; see 3685 // CodeBuffer::align_at_start). 3686 assert(_size == -1, "already set?"); 3687 _size = align_size_up(offset, CodeEntryAlignment); 3688 } 3689 3690 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3691 MacroAssembler _masm(&cb); 3692 for (int i = 0; i < _constants.length(); i++) { 3693 Constant con = _constants.at(i); 3694 address constant_addr; 3695 switch (con.type()) { 3696 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3697 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3698 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3699 case T_OBJECT: { 3700 jobject obj = con.get_jobject(); 3701 int oop_index = _masm.oop_recorder()->find_index(obj); 3702 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3703 break; 3704 } 3705 case T_ADDRESS: { 3706 address addr = (address) con.get_jobject(); 3707 constant_addr = _masm.address_constant(addr); 3708 break; 3709 } 3710 // We use T_VOID as marker for jump-table entries (labels) which 3711 // need an internal word relocation. 3712 case T_VOID: { 3713 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3714 // Fill the jump-table with a dummy word. The real value is 3715 // filled in later in fill_jump_table. 3716 address dummy = (address) n; 3717 constant_addr = _masm.address_constant(dummy); 3718 // Expand jump-table 3719 for (uint i = 1; i < n->outcnt(); i++) { 3720 address temp_addr = _masm.address_constant(dummy + i); 3721 assert(temp_addr, "consts section too small"); 3722 } 3723 break; 3724 } 3725 case T_METADATA: { 3726 Metadata* obj = con.get_metadata(); 3727 int metadata_index = _masm.oop_recorder()->find_index(obj); 3728 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3729 break; 3730 } 3731 default: ShouldNotReachHere(); 3732 } 3733 assert(constant_addr, "consts section too small"); 3734 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3735 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3736 } 3737 } 3738 3739 int Compile::ConstantTable::find_offset(Constant& con) const { 3740 int idx = _constants.find(con); 3741 assert(idx != -1, "constant must be in constant table"); 3742 int offset = _constants.at(idx).offset(); 3743 assert(offset != -1, "constant table not emitted yet?"); 3744 return offset; 3745 } 3746 3747 void Compile::ConstantTable::add(Constant& con) { 3748 if (con.can_be_reused()) { 3749 int idx = _constants.find(con); 3750 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3751 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3752 return; 3753 } 3754 } 3755 (void) _constants.append(con); 3756 } 3757 3758 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3759 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3760 Constant con(type, value, b->_freq); 3761 add(con); 3762 return con; 3763 } 3764 3765 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3766 Constant con(metadata); 3767 add(con); 3768 return con; 3769 } 3770 3771 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3772 jvalue value; 3773 BasicType type = oper->type()->basic_type(); 3774 switch (type) { 3775 case T_LONG: value.j = oper->constantL(); break; 3776 case T_FLOAT: value.f = oper->constantF(); break; 3777 case T_DOUBLE: value.d = oper->constantD(); break; 3778 case T_OBJECT: 3779 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3780 case T_METADATA: return add((Metadata*)oper->constant()); break; 3781 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3782 } 3783 return add(n, type, value); 3784 } 3785 3786 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3787 jvalue value; 3788 // We can use the node pointer here to identify the right jump-table 3789 // as this method is called from Compile::Fill_buffer right before 3790 // the MachNodes are emitted and the jump-table is filled (means the 3791 // MachNode pointers do not change anymore). 3792 value.l = (jobject) n; 3793 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3794 add(con); 3795 return con; 3796 } 3797 3798 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3799 // If called from Compile::scratch_emit_size do nothing. 3800 if (Compile::current()->in_scratch_emit_size()) return; 3801 3802 assert(labels.is_nonempty(), "must be"); 3803 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3804 3805 // Since MachConstantNode::constant_offset() also contains 3806 // table_base_offset() we need to subtract the table_base_offset() 3807 // to get the plain offset into the constant table. 3808 int offset = n->constant_offset() - table_base_offset(); 3809 3810 MacroAssembler _masm(&cb); 3811 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3812 3813 for (uint i = 0; i < n->outcnt(); i++) { 3814 address* constant_addr = &jump_table_base[i]; 3815 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i))); 3816 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3817 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3818 } 3819 } 3820 3821 //----------------------------static_subtype_check----------------------------- 3822 // Shortcut important common cases when superklass is exact: 3823 // (0) superklass is java.lang.Object (can occur in reflective code) 3824 // (1) subklass is already limited to a subtype of superklass => always ok 3825 // (2) subklass does not overlap with superklass => always fail 3826 // (3) superklass has NO subtypes and we can check with a simple compare. 3827 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3828 if (StressReflectiveCode) { 3829 return SSC_full_test; // Let caller generate the general case. 3830 } 3831 3832 if (superk == env()->Object_klass()) { 3833 return SSC_always_true; // (0) this test cannot fail 3834 } 3835 3836 ciType* superelem = superk; 3837 if (superelem->is_array_klass()) 3838 superelem = superelem->as_array_klass()->base_element_type(); 3839 3840 if (!subk->is_interface()) { // cannot trust static interface types yet 3841 if (subk->is_subtype_of(superk)) { 3842 return SSC_always_true; // (1) false path dead; no dynamic test needed 3843 } 3844 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3845 !superk->is_subtype_of(subk)) { 3846 return SSC_always_false; 3847 } 3848 } 3849 3850 // If casting to an instance klass, it must have no subtypes 3851 if (superk->is_interface()) { 3852 // Cannot trust interfaces yet. 3853 // %%% S.B. superk->nof_implementors() == 1 3854 } else if (superelem->is_instance_klass()) { 3855 ciInstanceKlass* ik = superelem->as_instance_klass(); 3856 if (!ik->has_subklass() && !ik->is_interface()) { 3857 if (!ik->is_final()) { 3858 // Add a dependency if there is a chance of a later subclass. 3859 dependencies()->assert_leaf_type(ik); 3860 } 3861 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3862 } 3863 } else { 3864 // A primitive array type has no subtypes. 3865 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3866 } 3867 3868 return SSC_full_test; 3869 } 3870 3871 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3872 #ifdef _LP64 3873 // The scaled index operand to AddP must be a clean 64-bit value. 3874 // Java allows a 32-bit int to be incremented to a negative 3875 // value, which appears in a 64-bit register as a large 3876 // positive number. Using that large positive number as an 3877 // operand in pointer arithmetic has bad consequences. 3878 // On the other hand, 32-bit overflow is rare, and the possibility 3879 // can often be excluded, if we annotate the ConvI2L node with 3880 // a type assertion that its value is known to be a small positive 3881 // number. (The prior range check has ensured this.) 3882 // This assertion is used by ConvI2LNode::Ideal. 3883 int index_max = max_jint - 1; // array size is max_jint, index is one less 3884 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3885 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3886 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3887 #endif 3888 return idx; 3889 } 3890 3891 // The message about the current inlining is accumulated in 3892 // _print_inlining_stream and transfered into the _print_inlining_list 3893 // once we know whether inlining succeeds or not. For regular 3894 // inlining, messages are appended to the buffer pointed by 3895 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3896 // a new buffer is added after _print_inlining_idx in the list. This 3897 // way we can update the inlining message for late inlining call site 3898 // when the inlining is attempted again. 3899 void Compile::print_inlining_init() { 3900 if (print_inlining() || print_intrinsics()) { 3901 _print_inlining_stream = new stringStream(); 3902 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3903 } 3904 } 3905 3906 void Compile::print_inlining_reinit() { 3907 if (print_inlining() || print_intrinsics()) { 3908 // Re allocate buffer when we change ResourceMark 3909 _print_inlining_stream = new stringStream(); 3910 } 3911 } 3912 3913 void Compile::print_inlining_reset() { 3914 _print_inlining_stream->reset(); 3915 } 3916 3917 void Compile::print_inlining_commit() { 3918 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3919 // Transfer the message from _print_inlining_stream to the current 3920 // _print_inlining_list buffer and clear _print_inlining_stream. 3921 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3922 print_inlining_reset(); 3923 } 3924 3925 void Compile::print_inlining_push() { 3926 // Add new buffer to the _print_inlining_list at current position 3927 _print_inlining_idx++; 3928 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3929 } 3930 3931 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3932 return _print_inlining_list->at(_print_inlining_idx); 3933 } 3934 3935 void Compile::print_inlining_update(CallGenerator* cg) { 3936 if (print_inlining() || print_intrinsics()) { 3937 if (!cg->is_late_inline()) { 3938 if (print_inlining_current().cg() != NULL) { 3939 print_inlining_push(); 3940 } 3941 print_inlining_commit(); 3942 } else { 3943 if (print_inlining_current().cg() != cg && 3944 (print_inlining_current().cg() != NULL || 3945 print_inlining_current().ss()->size() != 0)) { 3946 print_inlining_push(); 3947 } 3948 print_inlining_commit(); 3949 print_inlining_current().set_cg(cg); 3950 } 3951 } 3952 } 3953 3954 void Compile::print_inlining_move_to(CallGenerator* cg) { 3955 // We resume inlining at a late inlining call site. Locate the 3956 // corresponding inlining buffer so that we can update it. 3957 if (print_inlining()) { 3958 for (int i = 0; i < _print_inlining_list->length(); i++) { 3959 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3960 _print_inlining_idx = i; 3961 return; 3962 } 3963 } 3964 ShouldNotReachHere(); 3965 } 3966 } 3967 3968 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3969 if (print_inlining()) { 3970 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3971 assert(print_inlining_current().cg() == cg, "wrong entry"); 3972 // replace message with new message 3973 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3974 print_inlining_commit(); 3975 print_inlining_current().set_cg(cg); 3976 } 3977 } 3978 3979 void Compile::print_inlining_assert_ready() { 3980 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3981 } 3982 3983 void Compile::process_print_inlining() { 3984 bool do_print_inlining = print_inlining() || print_intrinsics(); 3985 if (do_print_inlining || log() != NULL) { 3986 // Print inlining message for candidates that we couldn't inline 3987 // for lack of space 3988 for (int i = 0; i < _late_inlines.length(); i++) { 3989 CallGenerator* cg = _late_inlines.at(i); 3990 if (!cg->is_mh_late_inline()) { 3991 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3992 if (do_print_inlining) { 3993 cg->print_inlining_late(msg); 3994 } 3995 log_late_inline_failure(cg, msg); 3996 } 3997 } 3998 } 3999 if (do_print_inlining) { 4000 ResourceMark rm; 4001 stringStream ss; 4002 for (int i = 0; i < _print_inlining_list->length(); i++) { 4003 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4004 } 4005 size_t end = ss.size(); 4006 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4007 strncpy(_print_inlining_output, ss.base(), end+1); 4008 _print_inlining_output[end] = 0; 4009 } 4010 } 4011 4012 void Compile::dump_print_inlining() { 4013 if (_print_inlining_output != NULL) { 4014 tty->print_raw(_print_inlining_output); 4015 } 4016 } 4017 4018 void Compile::log_late_inline(CallGenerator* cg) { 4019 if (log() != NULL) { 4020 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4021 cg->unique_id()); 4022 JVMState* p = cg->call_node()->jvms(); 4023 while (p != NULL) { 4024 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4025 p = p->caller(); 4026 } 4027 log()->tail("late_inline"); 4028 } 4029 } 4030 4031 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4032 log_late_inline(cg); 4033 if (log() != NULL) { 4034 log()->inline_fail(msg); 4035 } 4036 } 4037 4038 void Compile::log_inline_id(CallGenerator* cg) { 4039 if (log() != NULL) { 4040 // The LogCompilation tool needs a unique way to identify late 4041 // inline call sites. This id must be unique for this call site in 4042 // this compilation. Try to have it unique across compilations as 4043 // well because it can be convenient when grepping through the log 4044 // file. 4045 // Distinguish OSR compilations from others in case CICountOSR is 4046 // on. 4047 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4048 cg->set_unique_id(id); 4049 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4050 } 4051 } 4052 4053 void Compile::log_inline_failure(const char* msg) { 4054 if (C->log() != NULL) { 4055 C->log()->inline_fail(msg); 4056 } 4057 } 4058 4059 4060 // Dump inlining replay data to the stream. 4061 // Don't change thread state and acquire any locks. 4062 void Compile::dump_inline_data(outputStream* out) { 4063 InlineTree* inl_tree = ilt(); 4064 if (inl_tree != NULL) { 4065 out->print(" inline %d", inl_tree->count()); 4066 inl_tree->dump_replay_data(out); 4067 } 4068 } 4069 4070 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4071 if (n1->Opcode() < n2->Opcode()) return -1; 4072 else if (n1->Opcode() > n2->Opcode()) return 1; 4073 4074 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())); 4075 for (uint i = 1; i < n1->req(); i++) { 4076 if (n1->in(i) < n2->in(i)) return -1; 4077 else if (n1->in(i) > n2->in(i)) return 1; 4078 } 4079 4080 return 0; 4081 } 4082 4083 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4084 Node* n1 = *n1p; 4085 Node* n2 = *n2p; 4086 4087 return cmp_expensive_nodes(n1, n2); 4088 } 4089 4090 void Compile::sort_expensive_nodes() { 4091 if (!expensive_nodes_sorted()) { 4092 _expensive_nodes->sort(cmp_expensive_nodes); 4093 } 4094 } 4095 4096 bool Compile::expensive_nodes_sorted() const { 4097 for (int i = 1; i < _expensive_nodes->length(); i++) { 4098 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4099 return false; 4100 } 4101 } 4102 return true; 4103 } 4104 4105 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4106 if (_expensive_nodes->length() == 0) { 4107 return false; 4108 } 4109 4110 assert(OptimizeExpensiveOps, "optimization off?"); 4111 4112 // Take this opportunity to remove dead nodes from the list 4113 int j = 0; 4114 for (int i = 0; i < _expensive_nodes->length(); i++) { 4115 Node* n = _expensive_nodes->at(i); 4116 if (!n->is_unreachable(igvn)) { 4117 assert(n->is_expensive(), "should be expensive"); 4118 _expensive_nodes->at_put(j, n); 4119 j++; 4120 } 4121 } 4122 _expensive_nodes->trunc_to(j); 4123 4124 // Then sort the list so that similar nodes are next to each other 4125 // and check for at least two nodes of identical kind with same data 4126 // inputs. 4127 sort_expensive_nodes(); 4128 4129 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4130 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4131 return true; 4132 } 4133 } 4134 4135 return false; 4136 } 4137 4138 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4139 if (_expensive_nodes->length() == 0) { 4140 return; 4141 } 4142 4143 assert(OptimizeExpensiveOps, "optimization off?"); 4144 4145 // Sort to bring similar nodes next to each other and clear the 4146 // control input of nodes for which there's only a single copy. 4147 sort_expensive_nodes(); 4148 4149 int j = 0; 4150 int identical = 0; 4151 int i = 0; 4152 bool modified = false; 4153 for (; i < _expensive_nodes->length()-1; i++) { 4154 assert(j <= i, "can't write beyond current index"); 4155 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4156 identical++; 4157 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4158 continue; 4159 } 4160 if (identical > 0) { 4161 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4162 identical = 0; 4163 } else { 4164 Node* n = _expensive_nodes->at(i); 4165 igvn.replace_input_of(n, 0, NULL); 4166 igvn.hash_insert(n); 4167 modified = true; 4168 } 4169 } 4170 if (identical > 0) { 4171 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4172 } else if (_expensive_nodes->length() >= 1) { 4173 Node* n = _expensive_nodes->at(i); 4174 igvn.replace_input_of(n, 0, NULL); 4175 igvn.hash_insert(n); 4176 modified = true; 4177 } 4178 _expensive_nodes->trunc_to(j); 4179 if (modified) { 4180 igvn.optimize(); 4181 } 4182 } 4183 4184 void Compile::add_expensive_node(Node * n) { 4185 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4186 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4187 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4188 if (OptimizeExpensiveOps) { 4189 _expensive_nodes->append(n); 4190 } else { 4191 // Clear control input and let IGVN optimize expensive nodes if 4192 // OptimizeExpensiveOps is off. 4193 n->set_req(0, NULL); 4194 } 4195 } 4196 4197 /** 4198 * Remove the speculative part of types and clean up the graph 4199 */ 4200 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4201 if (UseTypeSpeculation) { 4202 Unique_Node_List worklist; 4203 worklist.push(root()); 4204 int modified = 0; 4205 // Go over all type nodes that carry a speculative type, drop the 4206 // speculative part of the type and enqueue the node for an igvn 4207 // which may optimize it out. 4208 for (uint next = 0; next < worklist.size(); ++next) { 4209 Node *n = worklist.at(next); 4210 if (n->is_Type()) { 4211 TypeNode* tn = n->as_Type(); 4212 const Type* t = tn->type(); 4213 const Type* t_no_spec = t->remove_speculative(); 4214 if (t_no_spec != t) { 4215 bool in_hash = igvn.hash_delete(n); 4216 assert(in_hash, "node should be in igvn hash table"); 4217 tn->set_type(t_no_spec); 4218 igvn.hash_insert(n); 4219 igvn._worklist.push(n); // give it a chance to go away 4220 modified++; 4221 } 4222 } 4223 uint max = n->len(); 4224 for( uint i = 0; i < max; ++i ) { 4225 Node *m = n->in(i); 4226 if (not_a_node(m)) continue; 4227 worklist.push(m); 4228 } 4229 } 4230 // Drop the speculative part of all types in the igvn's type table 4231 igvn.remove_speculative_types(); 4232 if (modified > 0) { 4233 igvn.optimize(); 4234 } 4235 #ifdef ASSERT 4236 // Verify that after the IGVN is over no speculative type has resurfaced 4237 worklist.clear(); 4238 worklist.push(root()); 4239 for (uint next = 0; next < worklist.size(); ++next) { 4240 Node *n = worklist.at(next); 4241 const Type* t = igvn.type_or_null(n); 4242 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4243 if (n->is_Type()) { 4244 t = n->as_Type()->type(); 4245 assert(t == t->remove_speculative(), "no more speculative types"); 4246 } 4247 uint max = n->len(); 4248 for( uint i = 0; i < max; ++i ) { 4249 Node *m = n->in(i); 4250 if (not_a_node(m)) continue; 4251 worklist.push(m); 4252 } 4253 } 4254 igvn.check_no_speculative_types(); 4255 #endif 4256 } 4257 } 4258 4259 // Auxiliary method to support randomized stressing/fuzzing. 4260 // 4261 // This method can be called the arbitrary number of times, with current count 4262 // as the argument. The logic allows selecting a single candidate from the 4263 // running list of candidates as follows: 4264 // int count = 0; 4265 // Cand* selected = null; 4266 // while(cand = cand->next()) { 4267 // if (randomized_select(++count)) { 4268 // selected = cand; 4269 // } 4270 // } 4271 // 4272 // Including count equalizes the chances any candidate is "selected". 4273 // This is useful when we don't have the complete list of candidates to choose 4274 // from uniformly. In this case, we need to adjust the randomicity of the 4275 // selection, or else we will end up biasing the selection towards the latter 4276 // candidates. 4277 // 4278 // Quick back-envelope calculation shows that for the list of n candidates 4279 // the equal probability for the candidate to persist as "best" can be 4280 // achieved by replacing it with "next" k-th candidate with the probability 4281 // of 1/k. It can be easily shown that by the end of the run, the 4282 // probability for any candidate is converged to 1/n, thus giving the 4283 // uniform distribution among all the candidates. 4284 // 4285 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4286 #define RANDOMIZED_DOMAIN_POW 29 4287 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4288 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4289 bool Compile::randomized_select(int count) { 4290 assert(count > 0, "only positive"); 4291 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4292 }