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