1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/convertnode.hpp" 46 #include "opto/divnode.hpp" 47 #include "opto/escape.hpp" 48 #include "opto/idealGraphPrinter.hpp" 49 #include "opto/loopnode.hpp" 50 #include "opto/machnode.hpp" 51 #include "opto/macro.hpp" 52 #include "opto/matcher.hpp" 53 #include "opto/mathexactnode.hpp" 54 #include "opto/memnode.hpp" 55 #include "opto/mulnode.hpp" 56 #include "opto/narrowptrnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/sharedRuntime.hpp" 69 #include "runtime/signature.hpp" 70 #include "runtime/stubRoutines.hpp" 71 #include "runtime/timer.hpp" 72 #include "utilities/copy.hpp" 73 74 75 // -------------------- Compile::mach_constant_base_node ----------------------- 76 // Constant table base node singleton. 77 MachConstantBaseNode* Compile::mach_constant_base_node() { 78 if (_mach_constant_base_node == NULL) { 79 _mach_constant_base_node = new MachConstantBaseNode(); 80 _mach_constant_base_node->add_req(C->root()); 81 } 82 return _mach_constant_base_node; 83 } 84 85 86 /// Support for intrinsics. 87 88 // Return the index at which m must be inserted (or already exists). 89 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 90 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 91 #ifdef ASSERT 92 for (int i = 1; i < _intrinsics->length(); i++) { 93 CallGenerator* cg1 = _intrinsics->at(i-1); 94 CallGenerator* cg2 = _intrinsics->at(i); 95 assert(cg1->method() != cg2->method() 96 ? cg1->method() < cg2->method() 97 : cg1->is_virtual() < cg2->is_virtual(), 98 "compiler intrinsics list must stay sorted"); 99 } 100 #endif 101 // Binary search sorted list, in decreasing intervals [lo, hi]. 102 int lo = 0, hi = _intrinsics->length()-1; 103 while (lo <= hi) { 104 int mid = (uint)(hi + lo) / 2; 105 ciMethod* mid_m = _intrinsics->at(mid)->method(); 106 if (m < mid_m) { 107 hi = mid-1; 108 } else if (m > mid_m) { 109 lo = mid+1; 110 } else { 111 // look at minor sort key 112 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 113 if (is_virtual < mid_virt) { 114 hi = mid-1; 115 } else if (is_virtual > mid_virt) { 116 lo = mid+1; 117 } else { 118 return mid; // exact match 119 } 120 } 121 } 122 return lo; // inexact match 123 } 124 125 void Compile::register_intrinsic(CallGenerator* cg) { 126 if (_intrinsics == NULL) { 127 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 128 } 129 // This code is stolen from ciObjectFactory::insert. 130 // Really, GrowableArray should have methods for 131 // insert_at, remove_at, and binary_search. 132 int len = _intrinsics->length(); 133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 134 if (index == len) { 135 _intrinsics->append(cg); 136 } else { 137 #ifdef ASSERT 138 CallGenerator* oldcg = _intrinsics->at(index); 139 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 140 #endif 141 _intrinsics->append(_intrinsics->at(len-1)); 142 int pos; 143 for (pos = len-2; pos >= index; pos--) { 144 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 145 } 146 _intrinsics->at_put(index, cg); 147 } 148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 149 } 150 151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 152 assert(m->is_loaded(), "don't try this on unloaded methods"); 153 if (_intrinsics != NULL) { 154 int index = intrinsic_insertion_index(m, is_virtual); 155 if (index < _intrinsics->length() 156 && _intrinsics->at(index)->method() == m 157 && _intrinsics->at(index)->is_virtual() == is_virtual) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = unique(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless expensive node 406 for (int i = C->expensive_count()-1; i >= 0; i--) { 407 Node* n = C->expensive_node(i); 408 if (!useful.member(n)) { 409 remove_expensive_node(n); 410 } 411 } 412 // clean up the late inline lists 413 remove_useless_late_inlines(&_string_late_inlines, useful); 414 remove_useless_late_inlines(&_boxing_late_inlines, useful); 415 remove_useless_late_inlines(&_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 //------------------------------frame_size_in_words----------------------------- 420 // frame_slots in units of words 421 int Compile::frame_size_in_words() const { 422 // shift is 0 in LP32 and 1 in LP64 423 const int shift = (LogBytesPerWord - LogBytesPerInt); 424 int words = _frame_slots >> shift; 425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 426 return words; 427 } 428 429 // To bang the stack of this compiled method we use the stack size 430 // that the interpreter would need in case of a deoptimization. This 431 // removes the need to bang the stack in the deoptimization blob which 432 // in turn simplifies stack overflow handling. 433 int Compile::bang_size_in_bytes() const { 434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 457 compile->clone_map().set_clone_idx(0); 458 compile->set_type_hwm(NULL); 459 compile->set_type_last_size(0); 460 compile->set_last_tf(NULL, NULL); 461 compile->set_indexSet_arena(NULL); 462 compile->set_indexSet_free_block_list(NULL); 463 compile->init_type_arena(); 464 Type::Initialize(compile); 465 _compile->set_scratch_buffer_blob(NULL); 466 _compile->begin_method(); 467 _compile->clone_map().set_debug(_compile->has_method() && _compile->method_has_option(_compile->clone_map().debug_option_name)); 468 } 469 CompileWrapper::~CompileWrapper() { 470 _compile->end_method(); 471 if (_compile->scratch_buffer_blob() != NULL) 472 BufferBlob::free(_compile->scratch_buffer_blob()); 473 _compile->env()->set_compiler_data(NULL); 474 } 475 476 477 //----------------------------print_compile_messages--------------------------- 478 void Compile::print_compile_messages() { 479 #ifndef PRODUCT 480 // Check if recompiling 481 if (_subsume_loads == false && PrintOpto) { 482 // Recompiling without allowing machine instructions to subsume loads 483 tty->print_cr("*********************************************************"); 484 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 485 tty->print_cr("*********************************************************"); 486 } 487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 488 // Recompiling without escape analysis 489 tty->print_cr("*********************************************************"); 490 tty->print_cr("** Bailout: Recompile without escape analysis **"); 491 tty->print_cr("*********************************************************"); 492 } 493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 494 // Recompiling without boxing elimination 495 tty->print_cr("*********************************************************"); 496 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 497 tty->print_cr("*********************************************************"); 498 } 499 if (env()->break_at_compile()) { 500 // Open the debugger when compiling this method. 501 tty->print("### Breaking when compiling: "); 502 method()->print_short_name(); 503 tty->cr(); 504 BREAKPOINT; 505 } 506 507 if( PrintOpto ) { 508 if (is_osr_compilation()) { 509 tty->print("[OSR]%3d", _compile_id); 510 } else { 511 tty->print("%3d", _compile_id); 512 } 513 } 514 #endif 515 } 516 517 518 //-----------------------init_scratch_buffer_blob------------------------------ 519 // Construct a temporary BufferBlob and cache it for this compile. 520 void Compile::init_scratch_buffer_blob(int const_size) { 521 // If there is already a scratch buffer blob allocated and the 522 // constant section is big enough, use it. Otherwise free the 523 // current and allocate a new one. 524 BufferBlob* blob = scratch_buffer_blob(); 525 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 526 // Use the current blob. 527 } else { 528 if (blob != NULL) { 529 BufferBlob::free(blob); 530 } 531 532 ResourceMark rm; 533 _scratch_const_size = const_size; 534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 535 blob = BufferBlob::create("Compile::scratch_buffer", size); 536 // Record the buffer blob for next time. 537 set_scratch_buffer_blob(blob); 538 // Have we run out of code space? 539 if (scratch_buffer_blob() == NULL) { 540 // Let CompilerBroker disable further compilations. 541 record_failure("Not enough space for scratch buffer in CodeCache"); 542 return; 543 } 544 } 545 546 // Initialize the relocation buffers 547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 548 set_scratch_locs_memory(locs_buf); 549 } 550 551 552 //-----------------------scratch_emit_size------------------------------------- 553 // Helper function that computes size by emitting code 554 uint Compile::scratch_emit_size(const Node* n) { 555 // Start scratch_emit_size section. 556 set_in_scratch_emit_size(true); 557 558 // Emit into a trash buffer and count bytes emitted. 559 // This is a pretty expensive way to compute a size, 560 // but it works well enough if seldom used. 561 // All common fixed-size instructions are given a size 562 // method by the AD file. 563 // Note that the scratch buffer blob and locs memory are 564 // allocated at the beginning of the compile task, and 565 // may be shared by several calls to scratch_emit_size. 566 // The allocation of the scratch buffer blob is particularly 567 // expensive, since it has to grab the code cache lock. 568 BufferBlob* blob = this->scratch_buffer_blob(); 569 assert(blob != NULL, "Initialize BufferBlob at start"); 570 assert(blob->size() > MAX_inst_size, "sanity"); 571 relocInfo* locs_buf = scratch_locs_memory(); 572 address blob_begin = blob->content_begin(); 573 address blob_end = (address)locs_buf; 574 assert(blob->content_contains(blob_end), "sanity"); 575 CodeBuffer buf(blob_begin, blob_end - blob_begin); 576 buf.initialize_consts_size(_scratch_const_size); 577 buf.initialize_stubs_size(MAX_stubs_size); 578 assert(locs_buf != NULL, "sanity"); 579 int lsize = MAX_locs_size / 3; 580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 583 584 // Do the emission. 585 586 Label fakeL; // Fake label for branch instructions. 587 Label* saveL = NULL; 588 uint save_bnum = 0; 589 bool is_branch = n->is_MachBranch(); 590 if (is_branch) { 591 MacroAssembler masm(&buf); 592 masm.bind(fakeL); 593 n->as_MachBranch()->save_label(&saveL, &save_bnum); 594 n->as_MachBranch()->label_set(&fakeL, 0); 595 } 596 n->emit(buf, this->regalloc()); 597 if (is_branch) // Restore label. 598 n->as_MachBranch()->label_set(saveL, save_bnum); 599 600 // End scratch_emit_size section. 601 set_in_scratch_emit_size(false); 602 603 return buf.insts_size(); 604 } 605 606 607 // ============================================================================ 608 //------------------------------Compile standard------------------------------- 609 debug_only( int Compile::_debug_idx = 100000; ) 610 611 // Compile a method. entry_bci is -1 for normal compilations and indicates 612 // the continuation bci for on stack replacement. 613 614 615 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 616 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 617 : Phase(Compiler), 618 _env(ci_env), 619 _log(ci_env->log()), 620 _compile_id(ci_env->compile_id()), 621 _save_argument_registers(false), 622 _stub_name(NULL), 623 _stub_function(NULL), 624 _stub_entry_point(NULL), 625 _method(target), 626 _entry_bci(osr_bci), 627 _initial_gvn(NULL), 628 _for_igvn(NULL), 629 _warm_calls(NULL), 630 _subsume_loads(subsume_loads), 631 _do_escape_analysis(do_escape_analysis), 632 _eliminate_boxing(eliminate_boxing), 633 _failure_reason(NULL), 634 _code_buffer("Compile::Fill_buffer"), 635 _orig_pc_slot(0), 636 _orig_pc_slot_offset_in_bytes(0), 637 _has_method_handle_invokes(false), 638 _mach_constant_base_node(NULL), 639 _node_bundling_limit(0), 640 _node_bundling_base(NULL), 641 _java_calls(0), 642 _inner_loops(0), 643 _scratch_const_size(-1), 644 _in_scratch_emit_size(false), 645 _dead_node_list(comp_arena()), 646 _dead_node_count(0), 647 #ifndef PRODUCT 648 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 649 _in_dump_cnt(0), 650 _printer(IdealGraphPrinter::printer()), 651 #endif 652 _congraph(NULL), 653 _comp_arena(mtCompiler), 654 _node_arena(mtCompiler), 655 _old_arena(mtCompiler), 656 _Compile_types(mtCompiler), 657 _replay_inline_data(NULL), 658 _late_inlines(comp_arena(), 2, 0, NULL), 659 _string_late_inlines(comp_arena(), 2, 0, NULL), 660 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 661 _late_inlines_pos(0), 662 _number_of_mh_late_inlines(0), 663 _inlining_progress(false), 664 _inlining_incrementally(false), 665 _print_inlining_list(NULL), 666 _print_inlining_stream(NULL), 667 _print_inlining_idx(0), 668 _print_inlining_output(NULL), 669 _interpreter_frame_size(0), 670 _max_node_limit(MaxNodeLimit) { 671 C = this; 672 673 CompileWrapper cw(this); 674 675 if (CITimeVerbose) { 676 tty->print(" "); 677 target->holder()->name()->print(); 678 tty->print("."); 679 target->print_short_name(); 680 tty->print(" "); 681 } 682 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 683 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 684 685 #ifndef PRODUCT 686 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 687 if (!print_opto_assembly) { 688 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 689 if (print_assembly && !Disassembler::can_decode()) { 690 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 691 print_opto_assembly = true; 692 } 693 } 694 set_print_assembly(print_opto_assembly); 695 set_parsed_irreducible_loop(false); 696 697 if (method()->has_option("ReplayInline")) { 698 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 699 } 700 #endif 701 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 702 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 703 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 704 705 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 706 // Make sure the method being compiled gets its own MDO, 707 // so we can at least track the decompile_count(). 708 // Need MDO to record RTM code generation state. 709 method()->ensure_method_data(); 710 } 711 712 Init(::AliasLevel); 713 714 715 print_compile_messages(); 716 717 _ilt = InlineTree::build_inline_tree_root(); 718 719 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 720 assert(num_alias_types() >= AliasIdxRaw, ""); 721 722 #define MINIMUM_NODE_HASH 1023 723 // Node list that Iterative GVN will start with 724 Unique_Node_List for_igvn(comp_arena()); 725 set_for_igvn(&for_igvn); 726 727 // GVN that will be run immediately on new nodes 728 uint estimated_size = method()->code_size()*4+64; 729 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 730 PhaseGVN gvn(node_arena(), estimated_size); 731 set_initial_gvn(&gvn); 732 733 print_inlining_init(); 734 { // Scope for timing the parser 735 TracePhase tp("parse", &timers[_t_parser]); 736 737 // Put top into the hash table ASAP. 738 initial_gvn()->transform_no_reclaim(top()); 739 740 // Set up tf(), start(), and find a CallGenerator. 741 CallGenerator* cg = NULL; 742 if (is_osr_compilation()) { 743 const TypeTuple *domain = StartOSRNode::osr_domain(); 744 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 745 init_tf(TypeFunc::make(domain, range)); 746 StartNode* s = new StartOSRNode(root(), domain); 747 initial_gvn()->set_type_bottom(s); 748 init_start(s); 749 cg = CallGenerator::for_osr(method(), entry_bci()); 750 } else { 751 // Normal case. 752 init_tf(TypeFunc::make(method())); 753 StartNode* s = new StartNode(root(), tf()->domain()); 754 initial_gvn()->set_type_bottom(s); 755 init_start(s); 756 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 757 // With java.lang.ref.reference.get() we must go through the 758 // intrinsic when G1 is enabled - even when get() is the root 759 // method of the compile - so that, if necessary, the value in 760 // the referent field of the reference object gets recorded by 761 // the pre-barrier code. 762 // Specifically, if G1 is enabled, the value in the referent 763 // field is recorded by the G1 SATB pre barrier. This will 764 // result in the referent being marked live and the reference 765 // object removed from the list of discovered references during 766 // reference processing. 767 cg = find_intrinsic(method(), false); 768 } 769 if (cg == NULL) { 770 float past_uses = method()->interpreter_invocation_count(); 771 float expected_uses = past_uses; 772 cg = CallGenerator::for_inline(method(), expected_uses); 773 } 774 } 775 if (failing()) return; 776 if (cg == NULL) { 777 record_method_not_compilable_all_tiers("cannot parse method"); 778 return; 779 } 780 JVMState* jvms = build_start_state(start(), tf()); 781 if ((jvms = cg->generate(jvms)) == NULL) { 782 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 783 record_method_not_compilable("method parse failed"); 784 } 785 return; 786 } 787 GraphKit kit(jvms); 788 789 if (!kit.stopped()) { 790 // Accept return values, and transfer control we know not where. 791 // This is done by a special, unique ReturnNode bound to root. 792 return_values(kit.jvms()); 793 } 794 795 if (kit.has_exceptions()) { 796 // Any exceptions that escape from this call must be rethrown 797 // to whatever caller is dynamically above us on the stack. 798 // This is done by a special, unique RethrowNode bound to root. 799 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 800 } 801 802 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 803 804 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 805 inline_string_calls(true); 806 } 807 808 if (failing()) return; 809 810 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 811 812 // Remove clutter produced by parsing. 813 if (!failing()) { 814 ResourceMark rm; 815 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 816 } 817 } 818 819 // Note: Large methods are capped off in do_one_bytecode(). 820 if (failing()) return; 821 822 // After parsing, node notes are no longer automagic. 823 // They must be propagated by register_new_node_with_optimizer(), 824 // clone(), or the like. 825 set_default_node_notes(NULL); 826 827 for (;;) { 828 int successes = Inline_Warm(); 829 if (failing()) return; 830 if (successes == 0) break; 831 } 832 833 // Drain the list. 834 Finish_Warm(); 835 #ifndef PRODUCT 836 if (_printer && _printer->should_print(_method)) { 837 _printer->print_inlining(this); 838 } 839 #endif 840 841 if (failing()) return; 842 NOT_PRODUCT( verify_graph_edges(); ) 843 844 // Now optimize 845 Optimize(); 846 if (failing()) return; 847 NOT_PRODUCT( verify_graph_edges(); ) 848 849 #ifndef PRODUCT 850 if (PrintIdeal) { 851 ttyLocker ttyl; // keep the following output all in one block 852 // This output goes directly to the tty, not the compiler log. 853 // To enable tools to match it up with the compilation activity, 854 // be sure to tag this tty output with the compile ID. 855 if (xtty != NULL) { 856 xtty->head("ideal compile_id='%d'%s", compile_id(), 857 is_osr_compilation() ? " compile_kind='osr'" : 858 ""); 859 } 860 root()->dump(9999); 861 if (xtty != NULL) { 862 xtty->tail("ideal"); 863 } 864 } 865 #endif 866 867 NOT_PRODUCT( verify_barriers(); ) 868 869 // Dump compilation data to replay it. 870 if (method()->has_option("DumpReplay")) { 871 env()->dump_replay_data(_compile_id); 872 } 873 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 874 env()->dump_inline_data(_compile_id); 875 } 876 877 // Now that we know the size of all the monitors we can add a fixed slot 878 // for the original deopt pc. 879 880 _orig_pc_slot = fixed_slots(); 881 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 882 set_fixed_slots(next_slot); 883 884 // Compute when to use implicit null checks. Used by matching trap based 885 // nodes and NullCheck optimization. 886 set_allowed_deopt_reasons(); 887 888 // Now generate code 889 Code_Gen(); 890 if (failing()) return; 891 892 // Check if we want to skip execution of all compiled code. 893 { 894 #ifndef PRODUCT 895 if (OptoNoExecute) { 896 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 897 return; 898 } 899 #endif 900 TracePhase tp("install_code", &timers[_t_registerMethod]); 901 902 if (is_osr_compilation()) { 903 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 904 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 905 } else { 906 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 907 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 908 } 909 910 env()->register_method(_method, _entry_bci, 911 &_code_offsets, 912 _orig_pc_slot_offset_in_bytes, 913 code_buffer(), 914 frame_size_in_words(), _oop_map_set, 915 &_handler_table, &_inc_table, 916 compiler, 917 env()->comp_level(), 918 has_unsafe_access(), 919 SharedRuntime::is_wide_vector(max_vector_size()), 920 rtm_state() 921 ); 922 923 if (log() != NULL) // Print code cache state into compiler log 924 log()->code_cache_state(); 925 } 926 } 927 928 //------------------------------Compile---------------------------------------- 929 // Compile a runtime stub 930 Compile::Compile( ciEnv* ci_env, 931 TypeFunc_generator generator, 932 address stub_function, 933 const char *stub_name, 934 int is_fancy_jump, 935 bool pass_tls, 936 bool save_arg_registers, 937 bool return_pc ) 938 : Phase(Compiler), 939 _env(ci_env), 940 _log(ci_env->log()), 941 _compile_id(0), 942 _save_argument_registers(save_arg_registers), 943 _method(NULL), 944 _stub_name(stub_name), 945 _stub_function(stub_function), 946 _stub_entry_point(NULL), 947 _entry_bci(InvocationEntryBci), 948 _initial_gvn(NULL), 949 _for_igvn(NULL), 950 _warm_calls(NULL), 951 _orig_pc_slot(0), 952 _orig_pc_slot_offset_in_bytes(0), 953 _subsume_loads(true), 954 _do_escape_analysis(false), 955 _eliminate_boxing(false), 956 _failure_reason(NULL), 957 _code_buffer("Compile::Fill_buffer"), 958 _has_method_handle_invokes(false), 959 _mach_constant_base_node(NULL), 960 _node_bundling_limit(0), 961 _node_bundling_base(NULL), 962 _java_calls(0), 963 _inner_loops(0), 964 #ifndef PRODUCT 965 _trace_opto_output(TraceOptoOutput), 966 _in_dump_cnt(0), 967 _printer(NULL), 968 #endif 969 _comp_arena(mtCompiler), 970 _node_arena(mtCompiler), 971 _old_arena(mtCompiler), 972 _Compile_types(mtCompiler), 973 _dead_node_list(comp_arena()), 974 _dead_node_count(0), 975 _congraph(NULL), 976 _replay_inline_data(NULL), 977 _number_of_mh_late_inlines(0), 978 _inlining_progress(false), 979 _inlining_incrementally(false), 980 _print_inlining_list(NULL), 981 _print_inlining_stream(NULL), 982 _print_inlining_idx(0), 983 _print_inlining_output(NULL), 984 _allowed_reasons(0), 985 _interpreter_frame_size(0), 986 _max_node_limit(MaxNodeLimit) { 987 C = this; 988 989 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 990 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 991 992 #ifndef PRODUCT 993 set_print_assembly(PrintFrameConverterAssembly); 994 set_parsed_irreducible_loop(false); 995 #endif 996 set_has_irreducible_loop(false); // no loops 997 998 CompileWrapper cw(this); 999 Init(/*AliasLevel=*/ 0); 1000 init_tf((*generator)()); 1001 1002 { 1003 // The following is a dummy for the sake of GraphKit::gen_stub 1004 Unique_Node_List for_igvn(comp_arena()); 1005 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1006 PhaseGVN gvn(Thread::current()->resource_area(),255); 1007 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1008 gvn.transform_no_reclaim(top()); 1009 1010 GraphKit kit; 1011 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1012 } 1013 1014 NOT_PRODUCT( verify_graph_edges(); ) 1015 Code_Gen(); 1016 if (failing()) return; 1017 1018 1019 // Entry point will be accessed using compile->stub_entry_point(); 1020 if (code_buffer() == NULL) { 1021 Matcher::soft_match_failure(); 1022 } else { 1023 if (PrintAssembly && (WizardMode || Verbose)) 1024 tty->print_cr("### Stub::%s", stub_name); 1025 1026 if (!failing()) { 1027 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1028 1029 // Make the NMethod 1030 // For now we mark the frame as never safe for profile stackwalking 1031 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1032 code_buffer(), 1033 CodeOffsets::frame_never_safe, 1034 // _code_offsets.value(CodeOffsets::Frame_Complete), 1035 frame_size_in_words(), 1036 _oop_map_set, 1037 save_arg_registers); 1038 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1039 1040 _stub_entry_point = rs->entry_point(); 1041 } 1042 } 1043 } 1044 1045 //------------------------------Init------------------------------------------- 1046 // Prepare for a single compilation 1047 void Compile::Init(int aliaslevel) { 1048 _unique = 0; 1049 _regalloc = NULL; 1050 1051 _tf = NULL; // filled in later 1052 _top = NULL; // cached later 1053 _matcher = NULL; // filled in later 1054 _cfg = NULL; // filled in later 1055 1056 set_24_bit_selection_and_mode(Use24BitFP, false); 1057 1058 _node_note_array = NULL; 1059 _default_node_notes = NULL; 1060 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1061 1062 _immutable_memory = NULL; // filled in at first inquiry 1063 1064 // Globally visible Nodes 1065 // First set TOP to NULL to give safe behavior during creation of RootNode 1066 set_cached_top_node(NULL); 1067 set_root(new RootNode()); 1068 // Now that you have a Root to point to, create the real TOP 1069 set_cached_top_node( new ConNode(Type::TOP) ); 1070 set_recent_alloc(NULL, NULL); 1071 1072 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1073 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1074 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1075 env()->set_dependencies(new Dependencies(env())); 1076 1077 _fixed_slots = 0; 1078 set_has_split_ifs(false); 1079 set_has_loops(has_method() && method()->has_loops()); // first approximation 1080 set_has_stringbuilder(false); 1081 set_has_boxed_value(false); 1082 _trap_can_recompile = false; // no traps emitted yet 1083 _major_progress = true; // start out assuming good things will happen 1084 set_has_unsafe_access(false); 1085 set_max_vector_size(0); 1086 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1087 set_decompile_count(0); 1088 1089 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1090 set_num_loop_opts(LoopOptsCount); 1091 set_do_inlining(Inline); 1092 set_max_inline_size(MaxInlineSize); 1093 set_freq_inline_size(FreqInlineSize); 1094 set_do_scheduling(OptoScheduling); 1095 set_do_count_invocations(false); 1096 set_do_method_data_update(false); 1097 1098 set_do_vector_loop(false); 1099 1100 bool do_vector = false; 1101 if (AllowVectorizeOnDemand) { 1102 if (has_method() && (method()->has_option("Vectorize") || method()->has_option("VectorizeDebug"))) { 1103 set_do_vector_loop(true); 1104 } else if (has_method() && method()->name() != 0 && 1105 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1106 set_do_vector_loop(true); 1107 } 1108 #ifndef PRODUCT 1109 if (do_vector_loop() && Verbose) { 1110 tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii()); 1111 } 1112 #endif 1113 } 1114 1115 set_age_code(has_method() && method()->profile_aging()); 1116 set_rtm_state(NoRTM); // No RTM lock eliding by default 1117 method_has_option_value("MaxNodeLimit", _max_node_limit); 1118 #if INCLUDE_RTM_OPT 1119 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1120 int rtm_state = method()->method_data()->rtm_state(); 1121 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1122 // Don't generate RTM lock eliding code. 1123 set_rtm_state(NoRTM); 1124 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1125 // Generate RTM lock eliding code without abort ratio calculation code. 1126 set_rtm_state(UseRTM); 1127 } else if (UseRTMDeopt) { 1128 // Generate RTM lock eliding code and include abort ratio calculation 1129 // code if UseRTMDeopt is on. 1130 set_rtm_state(ProfileRTM); 1131 } 1132 } 1133 #endif 1134 if (debug_info()->recording_non_safepoints()) { 1135 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1136 (comp_arena(), 8, 0, NULL)); 1137 set_default_node_notes(Node_Notes::make(this)); 1138 } 1139 1140 // // -- Initialize types before each compile -- 1141 // // Update cached type information 1142 // if( _method && _method->constants() ) 1143 // Type::update_loaded_types(_method, _method->constants()); 1144 1145 // Init alias_type map. 1146 if (!_do_escape_analysis && aliaslevel == 3) 1147 aliaslevel = 2; // No unique types without escape analysis 1148 _AliasLevel = aliaslevel; 1149 const int grow_ats = 16; 1150 _max_alias_types = grow_ats; 1151 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1152 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1153 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1154 { 1155 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1156 } 1157 // Initialize the first few types. 1158 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1159 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1160 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1161 _num_alias_types = AliasIdxRaw+1; 1162 // Zero out the alias type cache. 1163 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1164 // A NULL adr_type hits in the cache right away. Preload the right answer. 1165 probe_alias_cache(NULL)->_index = AliasIdxTop; 1166 1167 _intrinsics = NULL; 1168 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1169 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1170 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1171 register_library_intrinsics(); 1172 } 1173 1174 //---------------------------init_start---------------------------------------- 1175 // Install the StartNode on this compile object. 1176 void Compile::init_start(StartNode* s) { 1177 if (failing()) 1178 return; // already failing 1179 assert(s == start(), ""); 1180 } 1181 1182 /** 1183 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1184 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1185 * the ideal graph. 1186 */ 1187 StartNode* Compile::start() const { 1188 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1189 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1190 Node* start = root()->fast_out(i); 1191 if (start->is_Start()) { 1192 return start->as_Start(); 1193 } 1194 } 1195 fatal("Did not find Start node!"); 1196 return NULL; 1197 } 1198 1199 //-------------------------------immutable_memory------------------------------------- 1200 // Access immutable memory 1201 Node* Compile::immutable_memory() { 1202 if (_immutable_memory != NULL) { 1203 return _immutable_memory; 1204 } 1205 StartNode* s = start(); 1206 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1207 Node *p = s->fast_out(i); 1208 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1209 _immutable_memory = p; 1210 return _immutable_memory; 1211 } 1212 } 1213 ShouldNotReachHere(); 1214 return NULL; 1215 } 1216 1217 //----------------------set_cached_top_node------------------------------------ 1218 // Install the cached top node, and make sure Node::is_top works correctly. 1219 void Compile::set_cached_top_node(Node* tn) { 1220 if (tn != NULL) verify_top(tn); 1221 Node* old_top = _top; 1222 _top = tn; 1223 // Calling Node::setup_is_top allows the nodes the chance to adjust 1224 // their _out arrays. 1225 if (_top != NULL) _top->setup_is_top(); 1226 if (old_top != NULL) old_top->setup_is_top(); 1227 assert(_top == NULL || top()->is_top(), ""); 1228 } 1229 1230 #ifdef ASSERT 1231 uint Compile::count_live_nodes_by_graph_walk() { 1232 Unique_Node_List useful(comp_arena()); 1233 // Get useful node list by walking the graph. 1234 identify_useful_nodes(useful); 1235 return useful.size(); 1236 } 1237 1238 void Compile::print_missing_nodes() { 1239 1240 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1241 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1242 return; 1243 } 1244 1245 // This is an expensive function. It is executed only when the user 1246 // specifies VerifyIdealNodeCount option or otherwise knows the 1247 // additional work that needs to be done to identify reachable nodes 1248 // by walking the flow graph and find the missing ones using 1249 // _dead_node_list. 1250 1251 Unique_Node_List useful(comp_arena()); 1252 // Get useful node list by walking the graph. 1253 identify_useful_nodes(useful); 1254 1255 uint l_nodes = C->live_nodes(); 1256 uint l_nodes_by_walk = useful.size(); 1257 1258 if (l_nodes != l_nodes_by_walk) { 1259 if (_log != NULL) { 1260 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1261 _log->stamp(); 1262 _log->end_head(); 1263 } 1264 VectorSet& useful_member_set = useful.member_set(); 1265 int last_idx = l_nodes_by_walk; 1266 for (int i = 0; i < last_idx; i++) { 1267 if (useful_member_set.test(i)) { 1268 if (_dead_node_list.test(i)) { 1269 if (_log != NULL) { 1270 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1271 } 1272 if (PrintIdealNodeCount) { 1273 // Print the log message to tty 1274 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1275 useful.at(i)->dump(); 1276 } 1277 } 1278 } 1279 else if (! _dead_node_list.test(i)) { 1280 if (_log != NULL) { 1281 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1282 } 1283 if (PrintIdealNodeCount) { 1284 // Print the log message to tty 1285 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1286 } 1287 } 1288 } 1289 if (_log != NULL) { 1290 _log->tail("mismatched_nodes"); 1291 } 1292 } 1293 } 1294 void Compile::record_modified_node(Node* n) { 1295 if (_modified_nodes != NULL && !_inlining_incrementally && 1296 n->outcnt() != 0 && !n->is_Con()) { 1297 _modified_nodes->push(n); 1298 } 1299 } 1300 1301 void Compile::remove_modified_node(Node* n) { 1302 if (_modified_nodes != NULL) { 1303 _modified_nodes->remove(n); 1304 } 1305 } 1306 #endif 1307 1308 #ifndef PRODUCT 1309 void Compile::verify_top(Node* tn) const { 1310 if (tn != NULL) { 1311 assert(tn->is_Con(), "top node must be a constant"); 1312 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1313 assert(tn->in(0) != NULL, "must have live top node"); 1314 } 1315 } 1316 #endif 1317 1318 1319 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1320 1321 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1322 guarantee(arr != NULL, ""); 1323 int num_blocks = arr->length(); 1324 if (grow_by < num_blocks) grow_by = num_blocks; 1325 int num_notes = grow_by * _node_notes_block_size; 1326 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1327 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1328 while (num_notes > 0) { 1329 arr->append(notes); 1330 notes += _node_notes_block_size; 1331 num_notes -= _node_notes_block_size; 1332 } 1333 assert(num_notes == 0, "exact multiple, please"); 1334 } 1335 1336 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1337 if (source == NULL || dest == NULL) return false; 1338 1339 if (dest->is_Con()) 1340 return false; // Do not push debug info onto constants. 1341 1342 #ifdef ASSERT 1343 // Leave a bread crumb trail pointing to the original node: 1344 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1345 dest->set_debug_orig(source); 1346 } 1347 #endif 1348 1349 if (node_note_array() == NULL) 1350 return false; // Not collecting any notes now. 1351 1352 // This is a copy onto a pre-existing node, which may already have notes. 1353 // If both nodes have notes, do not overwrite any pre-existing notes. 1354 Node_Notes* source_notes = node_notes_at(source->_idx); 1355 if (source_notes == NULL || source_notes->is_clear()) return false; 1356 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1357 if (dest_notes == NULL || dest_notes->is_clear()) { 1358 return set_node_notes_at(dest->_idx, source_notes); 1359 } 1360 1361 Node_Notes merged_notes = (*source_notes); 1362 // The order of operations here ensures that dest notes will win... 1363 merged_notes.update_from(dest_notes); 1364 return set_node_notes_at(dest->_idx, &merged_notes); 1365 } 1366 1367 1368 //--------------------------allow_range_check_smearing------------------------- 1369 // Gating condition for coalescing similar range checks. 1370 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1371 // single covering check that is at least as strong as any of them. 1372 // If the optimization succeeds, the simplified (strengthened) range check 1373 // will always succeed. If it fails, we will deopt, and then give up 1374 // on the optimization. 1375 bool Compile::allow_range_check_smearing() const { 1376 // If this method has already thrown a range-check, 1377 // assume it was because we already tried range smearing 1378 // and it failed. 1379 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1380 return !already_trapped; 1381 } 1382 1383 1384 //------------------------------flatten_alias_type----------------------------- 1385 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1386 int offset = tj->offset(); 1387 TypePtr::PTR ptr = tj->ptr(); 1388 1389 // Known instance (scalarizable allocation) alias only with itself. 1390 bool is_known_inst = tj->isa_oopptr() != NULL && 1391 tj->is_oopptr()->is_known_instance(); 1392 1393 // Process weird unsafe references. 1394 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1395 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1396 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1397 tj = TypeOopPtr::BOTTOM; 1398 ptr = tj->ptr(); 1399 offset = tj->offset(); 1400 } 1401 1402 // Array pointers need some flattening 1403 const TypeAryPtr *ta = tj->isa_aryptr(); 1404 if (ta && ta->is_stable()) { 1405 // Erase stability property for alias analysis. 1406 tj = ta = ta->cast_to_stable(false); 1407 } 1408 if( ta && is_known_inst ) { 1409 if ( offset != Type::OffsetBot && 1410 offset > arrayOopDesc::length_offset_in_bytes() ) { 1411 offset = Type::OffsetBot; // Flatten constant access into array body only 1412 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1413 } 1414 } else if( ta && _AliasLevel >= 2 ) { 1415 // For arrays indexed by constant indices, we flatten the alias 1416 // space to include all of the array body. Only the header, klass 1417 // and array length can be accessed un-aliased. 1418 if( offset != Type::OffsetBot ) { 1419 if( ta->const_oop() ) { // MethodData* or Method* 1420 offset = Type::OffsetBot; // Flatten constant access into array body 1421 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1422 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1423 // range is OK as-is. 1424 tj = ta = TypeAryPtr::RANGE; 1425 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1426 tj = TypeInstPtr::KLASS; // all klass loads look alike 1427 ta = TypeAryPtr::RANGE; // generic ignored junk 1428 ptr = TypePtr::BotPTR; 1429 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1430 tj = TypeInstPtr::MARK; 1431 ta = TypeAryPtr::RANGE; // generic ignored junk 1432 ptr = TypePtr::BotPTR; 1433 } else { // Random constant offset into array body 1434 offset = Type::OffsetBot; // Flatten constant access into array body 1435 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1436 } 1437 } 1438 // Arrays of fixed size alias with arrays of unknown size. 1439 if (ta->size() != TypeInt::POS) { 1440 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1441 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1442 } 1443 // Arrays of known objects become arrays of unknown objects. 1444 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1445 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1446 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1447 } 1448 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1449 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1450 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1451 } 1452 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1453 // cannot be distinguished by bytecode alone. 1454 if (ta->elem() == TypeInt::BOOL) { 1455 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1456 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1457 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1458 } 1459 // During the 2nd round of IterGVN, NotNull castings are removed. 1460 // Make sure the Bottom and NotNull variants alias the same. 1461 // Also, make sure exact and non-exact variants alias the same. 1462 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1463 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1464 } 1465 } 1466 1467 // Oop pointers need some flattening 1468 const TypeInstPtr *to = tj->isa_instptr(); 1469 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1470 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1471 if( ptr == TypePtr::Constant ) { 1472 if (to->klass() != ciEnv::current()->Class_klass() || 1473 offset < k->size_helper() * wordSize) { 1474 // No constant oop pointers (such as Strings); they alias with 1475 // unknown strings. 1476 assert(!is_known_inst, "not scalarizable allocation"); 1477 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1478 } 1479 } else if( is_known_inst ) { 1480 tj = to; // Keep NotNull and klass_is_exact for instance type 1481 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1482 // During the 2nd round of IterGVN, NotNull castings are removed. 1483 // Make sure the Bottom and NotNull variants alias the same. 1484 // Also, make sure exact and non-exact variants alias the same. 1485 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1486 } 1487 if (to->speculative() != NULL) { 1488 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1489 } 1490 // Canonicalize the holder of this field 1491 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1492 // First handle header references such as a LoadKlassNode, even if the 1493 // object's klass is unloaded at compile time (4965979). 1494 if (!is_known_inst) { // Do it only for non-instance types 1495 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1496 } 1497 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1498 // Static fields are in the space above the normal instance 1499 // fields in the java.lang.Class instance. 1500 if (to->klass() != ciEnv::current()->Class_klass()) { 1501 to = NULL; 1502 tj = TypeOopPtr::BOTTOM; 1503 offset = tj->offset(); 1504 } 1505 } else { 1506 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1507 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1508 if( is_known_inst ) { 1509 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1510 } else { 1511 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1512 } 1513 } 1514 } 1515 } 1516 1517 // Klass pointers to object array klasses need some flattening 1518 const TypeKlassPtr *tk = tj->isa_klassptr(); 1519 if( tk ) { 1520 // If we are referencing a field within a Klass, we need 1521 // to assume the worst case of an Object. Both exact and 1522 // inexact types must flatten to the same alias class so 1523 // use NotNull as the PTR. 1524 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1525 1526 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1527 TypeKlassPtr::OBJECT->klass(), 1528 offset); 1529 } 1530 1531 ciKlass* klass = tk->klass(); 1532 if( klass->is_obj_array_klass() ) { 1533 ciKlass* k = TypeAryPtr::OOPS->klass(); 1534 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1535 k = TypeInstPtr::BOTTOM->klass(); 1536 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1537 } 1538 1539 // Check for precise loads from the primary supertype array and force them 1540 // to the supertype cache alias index. Check for generic array loads from 1541 // the primary supertype array and also force them to the supertype cache 1542 // alias index. Since the same load can reach both, we need to merge 1543 // these 2 disparate memories into the same alias class. Since the 1544 // primary supertype array is read-only, there's no chance of confusion 1545 // where we bypass an array load and an array store. 1546 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1547 if (offset == Type::OffsetBot || 1548 (offset >= primary_supers_offset && 1549 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1550 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1551 offset = in_bytes(Klass::secondary_super_cache_offset()); 1552 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1553 } 1554 } 1555 1556 // Flatten all Raw pointers together. 1557 if (tj->base() == Type::RawPtr) 1558 tj = TypeRawPtr::BOTTOM; 1559 1560 if (tj->base() == Type::AnyPtr) 1561 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1562 1563 // Flatten all to bottom for now 1564 switch( _AliasLevel ) { 1565 case 0: 1566 tj = TypePtr::BOTTOM; 1567 break; 1568 case 1: // Flatten to: oop, static, field or array 1569 switch (tj->base()) { 1570 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1571 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1572 case Type::AryPtr: // do not distinguish arrays at all 1573 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1574 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1575 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1576 default: ShouldNotReachHere(); 1577 } 1578 break; 1579 case 2: // No collapsing at level 2; keep all splits 1580 case 3: // No collapsing at level 3; keep all splits 1581 break; 1582 default: 1583 Unimplemented(); 1584 } 1585 1586 offset = tj->offset(); 1587 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1588 1589 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1590 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1591 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1592 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1593 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1594 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1595 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1596 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1597 assert( tj->ptr() != TypePtr::TopPTR && 1598 tj->ptr() != TypePtr::AnyNull && 1599 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1600 // assert( tj->ptr() != TypePtr::Constant || 1601 // tj->base() == Type::RawPtr || 1602 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1603 1604 return tj; 1605 } 1606 1607 void Compile::AliasType::Init(int i, const TypePtr* at) { 1608 _index = i; 1609 _adr_type = at; 1610 _field = NULL; 1611 _element = NULL; 1612 _is_rewritable = true; // default 1613 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1614 if (atoop != NULL && atoop->is_known_instance()) { 1615 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1616 _general_index = Compile::current()->get_alias_index(gt); 1617 } else { 1618 _general_index = 0; 1619 } 1620 } 1621 1622 //---------------------------------print_on------------------------------------ 1623 #ifndef PRODUCT 1624 void Compile::AliasType::print_on(outputStream* st) { 1625 if (index() < 10) 1626 st->print("@ <%d> ", index()); 1627 else st->print("@ <%d>", index()); 1628 st->print(is_rewritable() ? " " : " RO"); 1629 int offset = adr_type()->offset(); 1630 if (offset == Type::OffsetBot) 1631 st->print(" +any"); 1632 else st->print(" +%-3d", offset); 1633 st->print(" in "); 1634 adr_type()->dump_on(st); 1635 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1636 if (field() != NULL && tjp) { 1637 if (tjp->klass() != field()->holder() || 1638 tjp->offset() != field()->offset_in_bytes()) { 1639 st->print(" != "); 1640 field()->print(); 1641 st->print(" ***"); 1642 } 1643 } 1644 } 1645 1646 void print_alias_types() { 1647 Compile* C = Compile::current(); 1648 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1649 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1650 C->alias_type(idx)->print_on(tty); 1651 tty->cr(); 1652 } 1653 } 1654 #endif 1655 1656 1657 //----------------------------probe_alias_cache-------------------------------- 1658 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1659 intptr_t key = (intptr_t) adr_type; 1660 key ^= key >> logAliasCacheSize; 1661 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1662 } 1663 1664 1665 //-----------------------------grow_alias_types-------------------------------- 1666 void Compile::grow_alias_types() { 1667 const int old_ats = _max_alias_types; // how many before? 1668 const int new_ats = old_ats; // how many more? 1669 const int grow_ats = old_ats+new_ats; // how many now? 1670 _max_alias_types = grow_ats; 1671 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1672 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1673 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1674 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1675 } 1676 1677 1678 //--------------------------------find_alias_type------------------------------ 1679 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1680 if (_AliasLevel == 0) 1681 return alias_type(AliasIdxBot); 1682 1683 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1684 if (ace->_adr_type == adr_type) { 1685 return alias_type(ace->_index); 1686 } 1687 1688 // Handle special cases. 1689 if (adr_type == NULL) return alias_type(AliasIdxTop); 1690 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1691 1692 // Do it the slow way. 1693 const TypePtr* flat = flatten_alias_type(adr_type); 1694 1695 #ifdef ASSERT 1696 assert(flat == flatten_alias_type(flat), "idempotent"); 1697 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1698 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1699 const TypeOopPtr* foop = flat->is_oopptr(); 1700 // Scalarizable allocations have exact klass always. 1701 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1702 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1703 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1704 } 1705 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1706 #endif 1707 1708 int idx = AliasIdxTop; 1709 for (int i = 0; i < num_alias_types(); i++) { 1710 if (alias_type(i)->adr_type() == flat) { 1711 idx = i; 1712 break; 1713 } 1714 } 1715 1716 if (idx == AliasIdxTop) { 1717 if (no_create) return NULL; 1718 // Grow the array if necessary. 1719 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1720 // Add a new alias type. 1721 idx = _num_alias_types++; 1722 _alias_types[idx]->Init(idx, flat); 1723 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1724 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1725 if (flat->isa_instptr()) { 1726 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1727 && flat->is_instptr()->klass() == env()->Class_klass()) 1728 alias_type(idx)->set_rewritable(false); 1729 } 1730 if (flat->isa_aryptr()) { 1731 #ifdef ASSERT 1732 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1733 // (T_BYTE has the weakest alignment and size restrictions...) 1734 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1735 #endif 1736 if (flat->offset() == TypePtr::OffsetBot) { 1737 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1738 } 1739 } 1740 if (flat->isa_klassptr()) { 1741 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1742 alias_type(idx)->set_rewritable(false); 1743 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1744 alias_type(idx)->set_rewritable(false); 1745 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1746 alias_type(idx)->set_rewritable(false); 1747 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1748 alias_type(idx)->set_rewritable(false); 1749 } 1750 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1751 // but the base pointer type is not distinctive enough to identify 1752 // references into JavaThread.) 1753 1754 // Check for final fields. 1755 const TypeInstPtr* tinst = flat->isa_instptr(); 1756 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1757 ciField* field; 1758 if (tinst->const_oop() != NULL && 1759 tinst->klass() == ciEnv::current()->Class_klass() && 1760 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1761 // static field 1762 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1763 field = k->get_field_by_offset(tinst->offset(), true); 1764 } else { 1765 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1766 field = k->get_field_by_offset(tinst->offset(), false); 1767 } 1768 assert(field == NULL || 1769 original_field == NULL || 1770 (field->holder() == original_field->holder() && 1771 field->offset() == original_field->offset() && 1772 field->is_static() == original_field->is_static()), "wrong field?"); 1773 // Set field() and is_rewritable() attributes. 1774 if (field != NULL) alias_type(idx)->set_field(field); 1775 } 1776 } 1777 1778 // Fill the cache for next time. 1779 ace->_adr_type = adr_type; 1780 ace->_index = idx; 1781 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1782 1783 // Might as well try to fill the cache for the flattened version, too. 1784 AliasCacheEntry* face = probe_alias_cache(flat); 1785 if (face->_adr_type == NULL) { 1786 face->_adr_type = flat; 1787 face->_index = idx; 1788 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1789 } 1790 1791 return alias_type(idx); 1792 } 1793 1794 1795 Compile::AliasType* Compile::alias_type(ciField* field) { 1796 const TypeOopPtr* t; 1797 if (field->is_static()) 1798 t = TypeInstPtr::make(field->holder()->java_mirror()); 1799 else 1800 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1801 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1802 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1803 return atp; 1804 } 1805 1806 1807 //------------------------------have_alias_type-------------------------------- 1808 bool Compile::have_alias_type(const TypePtr* adr_type) { 1809 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1810 if (ace->_adr_type == adr_type) { 1811 return true; 1812 } 1813 1814 // Handle special cases. 1815 if (adr_type == NULL) return true; 1816 if (adr_type == TypePtr::BOTTOM) return true; 1817 1818 return find_alias_type(adr_type, true, NULL) != NULL; 1819 } 1820 1821 //-----------------------------must_alias-------------------------------------- 1822 // True if all values of the given address type are in the given alias category. 1823 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1824 if (alias_idx == AliasIdxBot) return true; // the universal category 1825 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1826 if (alias_idx == AliasIdxTop) return false; // the empty category 1827 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1828 1829 // the only remaining possible overlap is identity 1830 int adr_idx = get_alias_index(adr_type); 1831 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1832 assert(adr_idx == alias_idx || 1833 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1834 && adr_type != TypeOopPtr::BOTTOM), 1835 "should not be testing for overlap with an unsafe pointer"); 1836 return adr_idx == alias_idx; 1837 } 1838 1839 //------------------------------can_alias-------------------------------------- 1840 // True if any values of the given address type are in the given alias category. 1841 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1842 if (alias_idx == AliasIdxTop) return false; // the empty category 1843 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1844 if (alias_idx == AliasIdxBot) return true; // the universal category 1845 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1846 1847 // the only remaining possible overlap is identity 1848 int adr_idx = get_alias_index(adr_type); 1849 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1850 return adr_idx == alias_idx; 1851 } 1852 1853 1854 1855 //---------------------------pop_warm_call------------------------------------- 1856 WarmCallInfo* Compile::pop_warm_call() { 1857 WarmCallInfo* wci = _warm_calls; 1858 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1859 return wci; 1860 } 1861 1862 //----------------------------Inline_Warm-------------------------------------- 1863 int Compile::Inline_Warm() { 1864 // If there is room, try to inline some more warm call sites. 1865 // %%% Do a graph index compaction pass when we think we're out of space? 1866 if (!InlineWarmCalls) return 0; 1867 1868 int calls_made_hot = 0; 1869 int room_to_grow = NodeCountInliningCutoff - unique(); 1870 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1871 int amount_grown = 0; 1872 WarmCallInfo* call; 1873 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1874 int est_size = (int)call->size(); 1875 if (est_size > (room_to_grow - amount_grown)) { 1876 // This one won't fit anyway. Get rid of it. 1877 call->make_cold(); 1878 continue; 1879 } 1880 call->make_hot(); 1881 calls_made_hot++; 1882 amount_grown += est_size; 1883 amount_to_grow -= est_size; 1884 } 1885 1886 if (calls_made_hot > 0) set_major_progress(); 1887 return calls_made_hot; 1888 } 1889 1890 1891 //----------------------------Finish_Warm-------------------------------------- 1892 void Compile::Finish_Warm() { 1893 if (!InlineWarmCalls) return; 1894 if (failing()) return; 1895 if (warm_calls() == NULL) return; 1896 1897 // Clean up loose ends, if we are out of space for inlining. 1898 WarmCallInfo* call; 1899 while ((call = pop_warm_call()) != NULL) { 1900 call->make_cold(); 1901 } 1902 } 1903 1904 //---------------------cleanup_loop_predicates----------------------- 1905 // Remove the opaque nodes that protect the predicates so that all unused 1906 // checks and uncommon_traps will be eliminated from the ideal graph 1907 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1908 if (predicate_count()==0) return; 1909 for (int i = predicate_count(); i > 0; i--) { 1910 Node * n = predicate_opaque1_node(i-1); 1911 assert(n->Opcode() == Op_Opaque1, "must be"); 1912 igvn.replace_node(n, n->in(1)); 1913 } 1914 assert(predicate_count()==0, "should be clean!"); 1915 } 1916 1917 // StringOpts and late inlining of string methods 1918 void Compile::inline_string_calls(bool parse_time) { 1919 { 1920 // remove useless nodes to make the usage analysis simpler 1921 ResourceMark rm; 1922 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1923 } 1924 1925 { 1926 ResourceMark rm; 1927 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1928 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1929 print_method(PHASE_AFTER_STRINGOPTS, 3); 1930 } 1931 1932 // now inline anything that we skipped the first time around 1933 if (!parse_time) { 1934 _late_inlines_pos = _late_inlines.length(); 1935 } 1936 1937 while (_string_late_inlines.length() > 0) { 1938 CallGenerator* cg = _string_late_inlines.pop(); 1939 cg->do_late_inline(); 1940 if (failing()) return; 1941 } 1942 _string_late_inlines.trunc_to(0); 1943 } 1944 1945 // Late inlining of boxing methods 1946 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1947 if (_boxing_late_inlines.length() > 0) { 1948 assert(has_boxed_value(), "inconsistent"); 1949 1950 PhaseGVN* gvn = initial_gvn(); 1951 set_inlining_incrementally(true); 1952 1953 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1954 for_igvn()->clear(); 1955 gvn->replace_with(&igvn); 1956 1957 _late_inlines_pos = _late_inlines.length(); 1958 1959 while (_boxing_late_inlines.length() > 0) { 1960 CallGenerator* cg = _boxing_late_inlines.pop(); 1961 cg->do_late_inline(); 1962 if (failing()) return; 1963 } 1964 _boxing_late_inlines.trunc_to(0); 1965 1966 { 1967 ResourceMark rm; 1968 PhaseRemoveUseless pru(gvn, for_igvn()); 1969 } 1970 1971 igvn = PhaseIterGVN(gvn); 1972 igvn.optimize(); 1973 1974 set_inlining_progress(false); 1975 set_inlining_incrementally(false); 1976 } 1977 } 1978 1979 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1980 assert(IncrementalInline, "incremental inlining should be on"); 1981 PhaseGVN* gvn = initial_gvn(); 1982 1983 set_inlining_progress(false); 1984 for_igvn()->clear(); 1985 gvn->replace_with(&igvn); 1986 1987 { 1988 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1989 int i = 0; 1990 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1991 CallGenerator* cg = _late_inlines.at(i); 1992 _late_inlines_pos = i+1; 1993 cg->do_late_inline(); 1994 if (failing()) return; 1995 } 1996 int j = 0; 1997 for (; i < _late_inlines.length(); i++, j++) { 1998 _late_inlines.at_put(j, _late_inlines.at(i)); 1999 } 2000 _late_inlines.trunc_to(j); 2001 } 2002 2003 { 2004 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2005 ResourceMark rm; 2006 PhaseRemoveUseless pru(gvn, for_igvn()); 2007 } 2008 2009 { 2010 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2011 igvn = PhaseIterGVN(gvn); 2012 } 2013 } 2014 2015 // Perform incremental inlining until bound on number of live nodes is reached 2016 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2017 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2018 2019 PhaseGVN* gvn = initial_gvn(); 2020 2021 set_inlining_incrementally(true); 2022 set_inlining_progress(true); 2023 uint low_live_nodes = 0; 2024 2025 while(inlining_progress() && _late_inlines.length() > 0) { 2026 2027 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2028 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2029 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2030 // PhaseIdealLoop is expensive so we only try it once we are 2031 // out of live nodes and we only try it again if the previous 2032 // helped got the number of nodes down significantly 2033 PhaseIdealLoop ideal_loop( igvn, false, true ); 2034 if (failing()) return; 2035 low_live_nodes = live_nodes(); 2036 _major_progress = true; 2037 } 2038 2039 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2040 break; 2041 } 2042 } 2043 2044 inline_incrementally_one(igvn); 2045 2046 if (failing()) return; 2047 2048 { 2049 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2050 igvn.optimize(); 2051 } 2052 2053 if (failing()) return; 2054 } 2055 2056 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2057 2058 if (_string_late_inlines.length() > 0) { 2059 assert(has_stringbuilder(), "inconsistent"); 2060 for_igvn()->clear(); 2061 initial_gvn()->replace_with(&igvn); 2062 2063 inline_string_calls(false); 2064 2065 if (failing()) return; 2066 2067 { 2068 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2069 ResourceMark rm; 2070 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2071 } 2072 2073 { 2074 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2075 igvn = PhaseIterGVN(gvn); 2076 igvn.optimize(); 2077 } 2078 } 2079 2080 set_inlining_incrementally(false); 2081 } 2082 2083 2084 //------------------------------Optimize--------------------------------------- 2085 // Given a graph, optimize it. 2086 void Compile::Optimize() { 2087 TracePhase tp("optimizer", &timers[_t_optimizer]); 2088 2089 #ifndef PRODUCT 2090 if (env()->break_at_compile()) { 2091 BREAKPOINT; 2092 } 2093 2094 #endif 2095 2096 ResourceMark rm; 2097 int loop_opts_cnt; 2098 2099 print_inlining_reinit(); 2100 2101 NOT_PRODUCT( verify_graph_edges(); ) 2102 2103 print_method(PHASE_AFTER_PARSING); 2104 2105 { 2106 // Iterative Global Value Numbering, including ideal transforms 2107 // Initialize IterGVN with types and values from parse-time GVN 2108 PhaseIterGVN igvn(initial_gvn()); 2109 #ifdef ASSERT 2110 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2111 #endif 2112 { 2113 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2114 igvn.optimize(); 2115 } 2116 2117 print_method(PHASE_ITER_GVN1, 2); 2118 2119 if (failing()) return; 2120 2121 inline_incrementally(igvn); 2122 2123 print_method(PHASE_INCREMENTAL_INLINE, 2); 2124 2125 if (failing()) return; 2126 2127 if (eliminate_boxing()) { 2128 // Inline valueOf() methods now. 2129 inline_boxing_calls(igvn); 2130 2131 if (AlwaysIncrementalInline) { 2132 inline_incrementally(igvn); 2133 } 2134 2135 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2136 2137 if (failing()) return; 2138 } 2139 2140 // Remove the speculative part of types and clean up the graph from 2141 // the extra CastPP nodes whose only purpose is to carry them. Do 2142 // that early so that optimizations are not disrupted by the extra 2143 // CastPP nodes. 2144 remove_speculative_types(igvn); 2145 2146 // No more new expensive nodes will be added to the list from here 2147 // so keep only the actual candidates for optimizations. 2148 cleanup_expensive_nodes(igvn); 2149 2150 // Perform escape analysis 2151 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2152 if (has_loops()) { 2153 // Cleanup graph (remove dead nodes). 2154 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2155 PhaseIdealLoop ideal_loop( igvn, false, true ); 2156 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2157 if (failing()) return; 2158 } 2159 ConnectionGraph::do_analysis(this, &igvn); 2160 2161 if (failing()) return; 2162 2163 // Optimize out fields loads from scalar replaceable allocations. 2164 igvn.optimize(); 2165 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2166 2167 if (failing()) return; 2168 2169 if (congraph() != NULL && macro_count() > 0) { 2170 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2171 PhaseMacroExpand mexp(igvn); 2172 mexp.eliminate_macro_nodes(); 2173 igvn.set_delay_transform(false); 2174 2175 igvn.optimize(); 2176 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2177 2178 if (failing()) return; 2179 } 2180 } 2181 2182 // Loop transforms on the ideal graph. Range Check Elimination, 2183 // peeling, unrolling, etc. 2184 2185 // Set loop opts counter 2186 loop_opts_cnt = num_loop_opts(); 2187 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2188 { 2189 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2190 PhaseIdealLoop ideal_loop( igvn, true ); 2191 loop_opts_cnt--; 2192 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2193 if (failing()) return; 2194 } 2195 // Loop opts pass if partial peeling occurred in previous pass 2196 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2197 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2198 PhaseIdealLoop ideal_loop( igvn, false ); 2199 loop_opts_cnt--; 2200 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2201 if (failing()) return; 2202 } 2203 // Loop opts pass for loop-unrolling before CCP 2204 if(major_progress() && (loop_opts_cnt > 0)) { 2205 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2206 PhaseIdealLoop ideal_loop( igvn, false ); 2207 loop_opts_cnt--; 2208 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2209 } 2210 if (!failing()) { 2211 // Verify that last round of loop opts produced a valid graph 2212 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2213 PhaseIdealLoop::verify(igvn); 2214 } 2215 } 2216 if (failing()) return; 2217 2218 // Conditional Constant Propagation; 2219 PhaseCCP ccp( &igvn ); 2220 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2221 { 2222 TracePhase tp("ccp", &timers[_t_ccp]); 2223 ccp.do_transform(); 2224 } 2225 print_method(PHASE_CPP1, 2); 2226 2227 assert( true, "Break here to ccp.dump_old2new_map()"); 2228 2229 // Iterative Global Value Numbering, including ideal transforms 2230 { 2231 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2232 igvn = ccp; 2233 igvn.optimize(); 2234 } 2235 2236 print_method(PHASE_ITER_GVN2, 2); 2237 2238 if (failing()) return; 2239 2240 // Loop transforms on the ideal graph. Range Check Elimination, 2241 // peeling, unrolling, etc. 2242 if(loop_opts_cnt > 0) { 2243 debug_only( int cnt = 0; ); 2244 while(major_progress() && (loop_opts_cnt > 0)) { 2245 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2246 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2247 PhaseIdealLoop ideal_loop( igvn, true); 2248 loop_opts_cnt--; 2249 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2250 if (failing()) return; 2251 } 2252 } 2253 2254 { 2255 // Verify that all previous optimizations produced a valid graph 2256 // at least to this point, even if no loop optimizations were done. 2257 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2258 PhaseIdealLoop::verify(igvn); 2259 } 2260 2261 { 2262 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2263 PhaseMacroExpand mex(igvn); 2264 if (mex.expand_macro_nodes()) { 2265 assert(failing(), "must bail out w/ explicit message"); 2266 return; 2267 } 2268 } 2269 2270 DEBUG_ONLY( _modified_nodes = NULL; ) 2271 } // (End scope of igvn; run destructor if necessary for asserts.) 2272 2273 process_print_inlining(); 2274 // A method with only infinite loops has no edges entering loops from root 2275 { 2276 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2277 if (final_graph_reshaping()) { 2278 assert(failing(), "must bail out w/ explicit message"); 2279 return; 2280 } 2281 } 2282 2283 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2284 } 2285 2286 2287 //------------------------------Code_Gen--------------------------------------- 2288 // Given a graph, generate code for it 2289 void Compile::Code_Gen() { 2290 if (failing()) { 2291 return; 2292 } 2293 2294 // Perform instruction selection. You might think we could reclaim Matcher 2295 // memory PDQ, but actually the Matcher is used in generating spill code. 2296 // Internals of the Matcher (including some VectorSets) must remain live 2297 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2298 // set a bit in reclaimed memory. 2299 2300 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2301 // nodes. Mapping is only valid at the root of each matched subtree. 2302 NOT_PRODUCT( verify_graph_edges(); ) 2303 2304 Matcher matcher; 2305 _matcher = &matcher; 2306 { 2307 TracePhase tp("matcher", &timers[_t_matcher]); 2308 matcher.match(); 2309 } 2310 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2311 // nodes. Mapping is only valid at the root of each matched subtree. 2312 NOT_PRODUCT( verify_graph_edges(); ) 2313 2314 // If you have too many nodes, or if matching has failed, bail out 2315 check_node_count(0, "out of nodes matching instructions"); 2316 if (failing()) { 2317 return; 2318 } 2319 2320 // Build a proper-looking CFG 2321 PhaseCFG cfg(node_arena(), root(), matcher); 2322 _cfg = &cfg; 2323 { 2324 TracePhase tp("scheduler", &timers[_t_scheduler]); 2325 bool success = cfg.do_global_code_motion(); 2326 if (!success) { 2327 return; 2328 } 2329 2330 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2331 NOT_PRODUCT( verify_graph_edges(); ) 2332 debug_only( cfg.verify(); ) 2333 } 2334 2335 PhaseChaitin regalloc(unique(), cfg, matcher); 2336 _regalloc = ®alloc; 2337 { 2338 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2339 // Perform register allocation. After Chaitin, use-def chains are 2340 // no longer accurate (at spill code) and so must be ignored. 2341 // Node->LRG->reg mappings are still accurate. 2342 _regalloc->Register_Allocate(); 2343 2344 // Bail out if the allocator builds too many nodes 2345 if (failing()) { 2346 return; 2347 } 2348 } 2349 2350 // Prior to register allocation we kept empty basic blocks in case the 2351 // the allocator needed a place to spill. After register allocation we 2352 // are not adding any new instructions. If any basic block is empty, we 2353 // can now safely remove it. 2354 { 2355 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2356 cfg.remove_empty_blocks(); 2357 if (do_freq_based_layout()) { 2358 PhaseBlockLayout layout(cfg); 2359 } else { 2360 cfg.set_loop_alignment(); 2361 } 2362 cfg.fixup_flow(); 2363 } 2364 2365 // Apply peephole optimizations 2366 if( OptoPeephole ) { 2367 TracePhase tp("peephole", &timers[_t_peephole]); 2368 PhasePeephole peep( _regalloc, cfg); 2369 peep.do_transform(); 2370 } 2371 2372 // Do late expand if CPU requires this. 2373 if (Matcher::require_postalloc_expand) { 2374 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2375 cfg.postalloc_expand(_regalloc); 2376 } 2377 2378 // Convert Nodes to instruction bits in a buffer 2379 { 2380 TraceTime tp("output", &timers[_t_output], CITime); 2381 Output(); 2382 } 2383 2384 print_method(PHASE_FINAL_CODE); 2385 2386 // He's dead, Jim. 2387 _cfg = (PhaseCFG*)0xdeadbeef; 2388 _regalloc = (PhaseChaitin*)0xdeadbeef; 2389 } 2390 2391 2392 //------------------------------dump_asm--------------------------------------- 2393 // Dump formatted assembly 2394 #ifndef PRODUCT 2395 void Compile::dump_asm(int *pcs, uint pc_limit) { 2396 bool cut_short = false; 2397 tty->print_cr("#"); 2398 tty->print("# "); _tf->dump(); tty->cr(); 2399 tty->print_cr("#"); 2400 2401 // For all blocks 2402 int pc = 0x0; // Program counter 2403 char starts_bundle = ' '; 2404 _regalloc->dump_frame(); 2405 2406 Node *n = NULL; 2407 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2408 if (VMThread::should_terminate()) { 2409 cut_short = true; 2410 break; 2411 } 2412 Block* block = _cfg->get_block(i); 2413 if (block->is_connector() && !Verbose) { 2414 continue; 2415 } 2416 n = block->head(); 2417 if (pcs && n->_idx < pc_limit) { 2418 tty->print("%3.3x ", pcs[n->_idx]); 2419 } else { 2420 tty->print(" "); 2421 } 2422 block->dump_head(_cfg); 2423 if (block->is_connector()) { 2424 tty->print_cr(" # Empty connector block"); 2425 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2426 tty->print_cr(" # Block is sole successor of call"); 2427 } 2428 2429 // For all instructions 2430 Node *delay = NULL; 2431 for (uint j = 0; j < block->number_of_nodes(); j++) { 2432 if (VMThread::should_terminate()) { 2433 cut_short = true; 2434 break; 2435 } 2436 n = block->get_node(j); 2437 if (valid_bundle_info(n)) { 2438 Bundle* bundle = node_bundling(n); 2439 if (bundle->used_in_unconditional_delay()) { 2440 delay = n; 2441 continue; 2442 } 2443 if (bundle->starts_bundle()) { 2444 starts_bundle = '+'; 2445 } 2446 } 2447 2448 if (WizardMode) { 2449 n->dump(); 2450 } 2451 2452 if( !n->is_Region() && // Dont print in the Assembly 2453 !n->is_Phi() && // a few noisely useless nodes 2454 !n->is_Proj() && 2455 !n->is_MachTemp() && 2456 !n->is_SafePointScalarObject() && 2457 !n->is_Catch() && // Would be nice to print exception table targets 2458 !n->is_MergeMem() && // Not very interesting 2459 !n->is_top() && // Debug info table constants 2460 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2461 ) { 2462 if (pcs && n->_idx < pc_limit) 2463 tty->print("%3.3x", pcs[n->_idx]); 2464 else 2465 tty->print(" "); 2466 tty->print(" %c ", starts_bundle); 2467 starts_bundle = ' '; 2468 tty->print("\t"); 2469 n->format(_regalloc, tty); 2470 tty->cr(); 2471 } 2472 2473 // If we have an instruction with a delay slot, and have seen a delay, 2474 // then back up and print it 2475 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2476 assert(delay != NULL, "no unconditional delay instruction"); 2477 if (WizardMode) delay->dump(); 2478 2479 if (node_bundling(delay)->starts_bundle()) 2480 starts_bundle = '+'; 2481 if (pcs && n->_idx < pc_limit) 2482 tty->print("%3.3x", pcs[n->_idx]); 2483 else 2484 tty->print(" "); 2485 tty->print(" %c ", starts_bundle); 2486 starts_bundle = ' '; 2487 tty->print("\t"); 2488 delay->format(_regalloc, tty); 2489 tty->cr(); 2490 delay = NULL; 2491 } 2492 2493 // Dump the exception table as well 2494 if( n->is_Catch() && (Verbose || WizardMode) ) { 2495 // Print the exception table for this offset 2496 _handler_table.print_subtable_for(pc); 2497 } 2498 } 2499 2500 if (pcs && n->_idx < pc_limit) 2501 tty->print_cr("%3.3x", pcs[n->_idx]); 2502 else 2503 tty->cr(); 2504 2505 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2506 2507 } // End of per-block dump 2508 tty->cr(); 2509 2510 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2511 } 2512 #endif 2513 2514 //------------------------------Final_Reshape_Counts--------------------------- 2515 // This class defines counters to help identify when a method 2516 // may/must be executed using hardware with only 24-bit precision. 2517 struct Final_Reshape_Counts : public StackObj { 2518 int _call_count; // count non-inlined 'common' calls 2519 int _float_count; // count float ops requiring 24-bit precision 2520 int _double_count; // count double ops requiring more precision 2521 int _java_call_count; // count non-inlined 'java' calls 2522 int _inner_loop_count; // count loops which need alignment 2523 VectorSet _visited; // Visitation flags 2524 Node_List _tests; // Set of IfNodes & PCTableNodes 2525 2526 Final_Reshape_Counts() : 2527 _call_count(0), _float_count(0), _double_count(0), 2528 _java_call_count(0), _inner_loop_count(0), 2529 _visited( Thread::current()->resource_area() ) { } 2530 2531 void inc_call_count () { _call_count ++; } 2532 void inc_float_count () { _float_count ++; } 2533 void inc_double_count() { _double_count++; } 2534 void inc_java_call_count() { _java_call_count++; } 2535 void inc_inner_loop_count() { _inner_loop_count++; } 2536 2537 int get_call_count () const { return _call_count ; } 2538 int get_float_count () const { return _float_count ; } 2539 int get_double_count() const { return _double_count; } 2540 int get_java_call_count() const { return _java_call_count; } 2541 int get_inner_loop_count() const { return _inner_loop_count; } 2542 }; 2543 2544 #ifdef ASSERT 2545 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2546 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2547 // Make sure the offset goes inside the instance layout. 2548 return k->contains_field_offset(tp->offset()); 2549 // Note that OffsetBot and OffsetTop are very negative. 2550 } 2551 #endif 2552 2553 // Eliminate trivially redundant StoreCMs and accumulate their 2554 // precedence edges. 2555 void Compile::eliminate_redundant_card_marks(Node* n) { 2556 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2557 if (n->in(MemNode::Address)->outcnt() > 1) { 2558 // There are multiple users of the same address so it might be 2559 // possible to eliminate some of the StoreCMs 2560 Node* mem = n->in(MemNode::Memory); 2561 Node* adr = n->in(MemNode::Address); 2562 Node* val = n->in(MemNode::ValueIn); 2563 Node* prev = n; 2564 bool done = false; 2565 // Walk the chain of StoreCMs eliminating ones that match. As 2566 // long as it's a chain of single users then the optimization is 2567 // safe. Eliminating partially redundant StoreCMs would require 2568 // cloning copies down the other paths. 2569 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2570 if (adr == mem->in(MemNode::Address) && 2571 val == mem->in(MemNode::ValueIn)) { 2572 // redundant StoreCM 2573 if (mem->req() > MemNode::OopStore) { 2574 // Hasn't been processed by this code yet. 2575 n->add_prec(mem->in(MemNode::OopStore)); 2576 } else { 2577 // Already converted to precedence edge 2578 for (uint i = mem->req(); i < mem->len(); i++) { 2579 // Accumulate any precedence edges 2580 if (mem->in(i) != NULL) { 2581 n->add_prec(mem->in(i)); 2582 } 2583 } 2584 // Everything above this point has been processed. 2585 done = true; 2586 } 2587 // Eliminate the previous StoreCM 2588 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2589 assert(mem->outcnt() == 0, "should be dead"); 2590 mem->disconnect_inputs(NULL, this); 2591 } else { 2592 prev = mem; 2593 } 2594 mem = prev->in(MemNode::Memory); 2595 } 2596 } 2597 } 2598 2599 //------------------------------final_graph_reshaping_impl---------------------- 2600 // Implement items 1-5 from final_graph_reshaping below. 2601 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2602 2603 if ( n->outcnt() == 0 ) return; // dead node 2604 uint nop = n->Opcode(); 2605 2606 // Check for 2-input instruction with "last use" on right input. 2607 // Swap to left input. Implements item (2). 2608 if( n->req() == 3 && // two-input instruction 2609 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2610 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2611 n->in(2)->outcnt() == 1 &&// right use IS a last use 2612 !n->in(2)->is_Con() ) { // right use is not a constant 2613 // Check for commutative opcode 2614 switch( nop ) { 2615 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2616 case Op_MaxI: case Op_MinI: 2617 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2618 case Op_AndL: case Op_XorL: case Op_OrL: 2619 case Op_AndI: case Op_XorI: case Op_OrI: { 2620 // Move "last use" input to left by swapping inputs 2621 n->swap_edges(1, 2); 2622 break; 2623 } 2624 default: 2625 break; 2626 } 2627 } 2628 2629 #ifdef ASSERT 2630 if( n->is_Mem() ) { 2631 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2632 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2633 // oop will be recorded in oop map if load crosses safepoint 2634 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2635 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2636 "raw memory operations should have control edge"); 2637 } 2638 #endif 2639 // Count FPU ops and common calls, implements item (3) 2640 switch( nop ) { 2641 // Count all float operations that may use FPU 2642 case Op_AddF: 2643 case Op_SubF: 2644 case Op_MulF: 2645 case Op_DivF: 2646 case Op_NegF: 2647 case Op_ModF: 2648 case Op_ConvI2F: 2649 case Op_ConF: 2650 case Op_CmpF: 2651 case Op_CmpF3: 2652 // case Op_ConvL2F: // longs are split into 32-bit halves 2653 frc.inc_float_count(); 2654 break; 2655 2656 case Op_ConvF2D: 2657 case Op_ConvD2F: 2658 frc.inc_float_count(); 2659 frc.inc_double_count(); 2660 break; 2661 2662 // Count all double operations that may use FPU 2663 case Op_AddD: 2664 case Op_SubD: 2665 case Op_MulD: 2666 case Op_DivD: 2667 case Op_NegD: 2668 case Op_ModD: 2669 case Op_ConvI2D: 2670 case Op_ConvD2I: 2671 // case Op_ConvL2D: // handled by leaf call 2672 // case Op_ConvD2L: // handled by leaf call 2673 case Op_ConD: 2674 case Op_CmpD: 2675 case Op_CmpD3: 2676 frc.inc_double_count(); 2677 break; 2678 case Op_Opaque1: // Remove Opaque Nodes before matching 2679 case Op_Opaque2: // Remove Opaque Nodes before matching 2680 case Op_Opaque3: 2681 n->subsume_by(n->in(1), this); 2682 break; 2683 case Op_CallStaticJava: 2684 case Op_CallJava: 2685 case Op_CallDynamicJava: 2686 frc.inc_java_call_count(); // Count java call site; 2687 case Op_CallRuntime: 2688 case Op_CallLeaf: 2689 case Op_CallLeafNoFP: { 2690 assert( n->is_Call(), "" ); 2691 CallNode *call = n->as_Call(); 2692 // Count call sites where the FP mode bit would have to be flipped. 2693 // Do not count uncommon runtime calls: 2694 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2695 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2696 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2697 frc.inc_call_count(); // Count the call site 2698 } else { // See if uncommon argument is shared 2699 Node *n = call->in(TypeFunc::Parms); 2700 int nop = n->Opcode(); 2701 // Clone shared simple arguments to uncommon calls, item (1). 2702 if( n->outcnt() > 1 && 2703 !n->is_Proj() && 2704 nop != Op_CreateEx && 2705 nop != Op_CheckCastPP && 2706 nop != Op_DecodeN && 2707 nop != Op_DecodeNKlass && 2708 !n->is_Mem() ) { 2709 Node *x = n->clone(); 2710 call->set_req( TypeFunc::Parms, x ); 2711 } 2712 } 2713 break; 2714 } 2715 2716 case Op_StoreD: 2717 case Op_LoadD: 2718 case Op_LoadD_unaligned: 2719 frc.inc_double_count(); 2720 goto handle_mem; 2721 case Op_StoreF: 2722 case Op_LoadF: 2723 frc.inc_float_count(); 2724 goto handle_mem; 2725 2726 case Op_StoreCM: 2727 { 2728 // Convert OopStore dependence into precedence edge 2729 Node* prec = n->in(MemNode::OopStore); 2730 n->del_req(MemNode::OopStore); 2731 n->add_prec(prec); 2732 eliminate_redundant_card_marks(n); 2733 } 2734 2735 // fall through 2736 2737 case Op_StoreB: 2738 case Op_StoreC: 2739 case Op_StorePConditional: 2740 case Op_StoreI: 2741 case Op_StoreL: 2742 case Op_StoreIConditional: 2743 case Op_StoreLConditional: 2744 case Op_CompareAndSwapI: 2745 case Op_CompareAndSwapL: 2746 case Op_CompareAndSwapP: 2747 case Op_CompareAndSwapN: 2748 case Op_GetAndAddI: 2749 case Op_GetAndAddL: 2750 case Op_GetAndSetI: 2751 case Op_GetAndSetL: 2752 case Op_GetAndSetP: 2753 case Op_GetAndSetN: 2754 case Op_StoreP: 2755 case Op_StoreN: 2756 case Op_StoreNKlass: 2757 case Op_LoadB: 2758 case Op_LoadUB: 2759 case Op_LoadUS: 2760 case Op_LoadI: 2761 case Op_LoadKlass: 2762 case Op_LoadNKlass: 2763 case Op_LoadL: 2764 case Op_LoadL_unaligned: 2765 case Op_LoadPLocked: 2766 case Op_LoadP: 2767 case Op_LoadN: 2768 case Op_LoadRange: 2769 case Op_LoadS: { 2770 handle_mem: 2771 #ifdef ASSERT 2772 if( VerifyOptoOopOffsets ) { 2773 assert( n->is_Mem(), "" ); 2774 MemNode *mem = (MemNode*)n; 2775 // Check to see if address types have grounded out somehow. 2776 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2777 assert( !tp || oop_offset_is_sane(tp), "" ); 2778 } 2779 #endif 2780 break; 2781 } 2782 2783 case Op_AddP: { // Assert sane base pointers 2784 Node *addp = n->in(AddPNode::Address); 2785 assert( !addp->is_AddP() || 2786 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2787 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2788 "Base pointers must match" ); 2789 #ifdef _LP64 2790 if ((UseCompressedOops || UseCompressedClassPointers) && 2791 addp->Opcode() == Op_ConP && 2792 addp == n->in(AddPNode::Base) && 2793 n->in(AddPNode::Offset)->is_Con()) { 2794 // Use addressing with narrow klass to load with offset on x86. 2795 // On sparc loading 32-bits constant and decoding it have less 2796 // instructions (4) then load 64-bits constant (7). 2797 // Do this transformation here since IGVN will convert ConN back to ConP. 2798 const Type* t = addp->bottom_type(); 2799 if (t->isa_oopptr() || t->isa_klassptr()) { 2800 Node* nn = NULL; 2801 2802 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2803 2804 // Look for existing ConN node of the same exact type. 2805 Node* r = root(); 2806 uint cnt = r->outcnt(); 2807 for (uint i = 0; i < cnt; i++) { 2808 Node* m = r->raw_out(i); 2809 if (m!= NULL && m->Opcode() == op && 2810 m->bottom_type()->make_ptr() == t) { 2811 nn = m; 2812 break; 2813 } 2814 } 2815 if (nn != NULL) { 2816 // Decode a narrow oop to match address 2817 // [R12 + narrow_oop_reg<<3 + offset] 2818 if (t->isa_oopptr()) { 2819 nn = new DecodeNNode(nn, t); 2820 } else { 2821 nn = new DecodeNKlassNode(nn, t); 2822 } 2823 n->set_req(AddPNode::Base, nn); 2824 n->set_req(AddPNode::Address, nn); 2825 if (addp->outcnt() == 0) { 2826 addp->disconnect_inputs(NULL, this); 2827 } 2828 } 2829 } 2830 } 2831 #endif 2832 break; 2833 } 2834 2835 case Op_CastPP: { 2836 // Remove CastPP nodes to gain more freedom during scheduling but 2837 // keep the dependency they encode as control or precedence edges 2838 // (if control is set already) on memory operations. Some CastPP 2839 // nodes don't have a control (don't carry a dependency): skip 2840 // those. 2841 if (n->in(0) != NULL) { 2842 ResourceMark rm; 2843 Unique_Node_List wq; 2844 wq.push(n); 2845 for (uint next = 0; next < wq.size(); ++next) { 2846 Node *m = wq.at(next); 2847 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2848 Node* use = m->fast_out(i); 2849 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2850 use->ensure_control_or_add_prec(n->in(0)); 2851 } else if (use->in(0) == NULL) { 2852 switch(use->Opcode()) { 2853 case Op_AddP: 2854 case Op_DecodeN: 2855 case Op_DecodeNKlass: 2856 case Op_CheckCastPP: 2857 case Op_CastPP: 2858 wq.push(use); 2859 break; 2860 } 2861 } 2862 } 2863 } 2864 } 2865 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2866 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2867 Node* in1 = n->in(1); 2868 const Type* t = n->bottom_type(); 2869 Node* new_in1 = in1->clone(); 2870 new_in1->as_DecodeN()->set_type(t); 2871 2872 if (!Matcher::narrow_oop_use_complex_address()) { 2873 // 2874 // x86, ARM and friends can handle 2 adds in addressing mode 2875 // and Matcher can fold a DecodeN node into address by using 2876 // a narrow oop directly and do implicit NULL check in address: 2877 // 2878 // [R12 + narrow_oop_reg<<3 + offset] 2879 // NullCheck narrow_oop_reg 2880 // 2881 // On other platforms (Sparc) we have to keep new DecodeN node and 2882 // use it to do implicit NULL check in address: 2883 // 2884 // decode_not_null narrow_oop_reg, base_reg 2885 // [base_reg + offset] 2886 // NullCheck base_reg 2887 // 2888 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2889 // to keep the information to which NULL check the new DecodeN node 2890 // corresponds to use it as value in implicit_null_check(). 2891 // 2892 new_in1->set_req(0, n->in(0)); 2893 } 2894 2895 n->subsume_by(new_in1, this); 2896 if (in1->outcnt() == 0) { 2897 in1->disconnect_inputs(NULL, this); 2898 } 2899 } else { 2900 n->subsume_by(n->in(1), this); 2901 if (n->outcnt() == 0) { 2902 n->disconnect_inputs(NULL, this); 2903 } 2904 } 2905 break; 2906 } 2907 #ifdef _LP64 2908 case Op_CmpP: 2909 // Do this transformation here to preserve CmpPNode::sub() and 2910 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2911 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2912 Node* in1 = n->in(1); 2913 Node* in2 = n->in(2); 2914 if (!in1->is_DecodeNarrowPtr()) { 2915 in2 = in1; 2916 in1 = n->in(2); 2917 } 2918 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2919 2920 Node* new_in2 = NULL; 2921 if (in2->is_DecodeNarrowPtr()) { 2922 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2923 new_in2 = in2->in(1); 2924 } else if (in2->Opcode() == Op_ConP) { 2925 const Type* t = in2->bottom_type(); 2926 if (t == TypePtr::NULL_PTR) { 2927 assert(in1->is_DecodeN(), "compare klass to null?"); 2928 // Don't convert CmpP null check into CmpN if compressed 2929 // oops implicit null check is not generated. 2930 // This will allow to generate normal oop implicit null check. 2931 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2932 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2933 // 2934 // This transformation together with CastPP transformation above 2935 // will generated code for implicit NULL checks for compressed oops. 2936 // 2937 // The original code after Optimize() 2938 // 2939 // LoadN memory, narrow_oop_reg 2940 // decode narrow_oop_reg, base_reg 2941 // CmpP base_reg, NULL 2942 // CastPP base_reg // NotNull 2943 // Load [base_reg + offset], val_reg 2944 // 2945 // after these transformations will be 2946 // 2947 // LoadN memory, narrow_oop_reg 2948 // CmpN narrow_oop_reg, NULL 2949 // decode_not_null narrow_oop_reg, base_reg 2950 // Load [base_reg + offset], val_reg 2951 // 2952 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2953 // since narrow oops can be used in debug info now (see the code in 2954 // final_graph_reshaping_walk()). 2955 // 2956 // At the end the code will be matched to 2957 // on x86: 2958 // 2959 // Load_narrow_oop memory, narrow_oop_reg 2960 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2961 // NullCheck narrow_oop_reg 2962 // 2963 // and on sparc: 2964 // 2965 // Load_narrow_oop memory, narrow_oop_reg 2966 // decode_not_null narrow_oop_reg, base_reg 2967 // Load [base_reg + offset], val_reg 2968 // NullCheck base_reg 2969 // 2970 } else if (t->isa_oopptr()) { 2971 new_in2 = ConNode::make(t->make_narrowoop()); 2972 } else if (t->isa_klassptr()) { 2973 new_in2 = ConNode::make(t->make_narrowklass()); 2974 } 2975 } 2976 if (new_in2 != NULL) { 2977 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2978 n->subsume_by(cmpN, this); 2979 if (in1->outcnt() == 0) { 2980 in1->disconnect_inputs(NULL, this); 2981 } 2982 if (in2->outcnt() == 0) { 2983 in2->disconnect_inputs(NULL, this); 2984 } 2985 } 2986 } 2987 break; 2988 2989 case Op_DecodeN: 2990 case Op_DecodeNKlass: 2991 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2992 // DecodeN could be pinned when it can't be fold into 2993 // an address expression, see the code for Op_CastPP above. 2994 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2995 break; 2996 2997 case Op_EncodeP: 2998 case Op_EncodePKlass: { 2999 Node* in1 = n->in(1); 3000 if (in1->is_DecodeNarrowPtr()) { 3001 n->subsume_by(in1->in(1), this); 3002 } else if (in1->Opcode() == Op_ConP) { 3003 const Type* t = in1->bottom_type(); 3004 if (t == TypePtr::NULL_PTR) { 3005 assert(t->isa_oopptr(), "null klass?"); 3006 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3007 } else if (t->isa_oopptr()) { 3008 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3009 } else if (t->isa_klassptr()) { 3010 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3011 } 3012 } 3013 if (in1->outcnt() == 0) { 3014 in1->disconnect_inputs(NULL, this); 3015 } 3016 break; 3017 } 3018 3019 case Op_Proj: { 3020 if (OptimizeStringConcat) { 3021 ProjNode* p = n->as_Proj(); 3022 if (p->_is_io_use) { 3023 // Separate projections were used for the exception path which 3024 // are normally removed by a late inline. If it wasn't inlined 3025 // then they will hang around and should just be replaced with 3026 // the original one. 3027 Node* proj = NULL; 3028 // Replace with just one 3029 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3030 Node *use = i.get(); 3031 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3032 proj = use; 3033 break; 3034 } 3035 } 3036 assert(proj != NULL, "must be found"); 3037 p->subsume_by(proj, this); 3038 } 3039 } 3040 break; 3041 } 3042 3043 case Op_Phi: 3044 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3045 // The EncodeP optimization may create Phi with the same edges 3046 // for all paths. It is not handled well by Register Allocator. 3047 Node* unique_in = n->in(1); 3048 assert(unique_in != NULL, ""); 3049 uint cnt = n->req(); 3050 for (uint i = 2; i < cnt; i++) { 3051 Node* m = n->in(i); 3052 assert(m != NULL, ""); 3053 if (unique_in != m) 3054 unique_in = NULL; 3055 } 3056 if (unique_in != NULL) { 3057 n->subsume_by(unique_in, this); 3058 } 3059 } 3060 break; 3061 3062 #endif 3063 3064 case Op_ModI: 3065 if (UseDivMod) { 3066 // Check if a%b and a/b both exist 3067 Node* d = n->find_similar(Op_DivI); 3068 if (d) { 3069 // Replace them with a fused divmod if supported 3070 if (Matcher::has_match_rule(Op_DivModI)) { 3071 DivModINode* divmod = DivModINode::make(n); 3072 d->subsume_by(divmod->div_proj(), this); 3073 n->subsume_by(divmod->mod_proj(), this); 3074 } else { 3075 // replace a%b with a-((a/b)*b) 3076 Node* mult = new MulINode(d, d->in(2)); 3077 Node* sub = new SubINode(d->in(1), mult); 3078 n->subsume_by(sub, this); 3079 } 3080 } 3081 } 3082 break; 3083 3084 case Op_ModL: 3085 if (UseDivMod) { 3086 // Check if a%b and a/b both exist 3087 Node* d = n->find_similar(Op_DivL); 3088 if (d) { 3089 // Replace them with a fused divmod if supported 3090 if (Matcher::has_match_rule(Op_DivModL)) { 3091 DivModLNode* divmod = DivModLNode::make(n); 3092 d->subsume_by(divmod->div_proj(), this); 3093 n->subsume_by(divmod->mod_proj(), this); 3094 } else { 3095 // replace a%b with a-((a/b)*b) 3096 Node* mult = new MulLNode(d, d->in(2)); 3097 Node* sub = new SubLNode(d->in(1), mult); 3098 n->subsume_by(sub, this); 3099 } 3100 } 3101 } 3102 break; 3103 3104 case Op_LoadVector: 3105 case Op_StoreVector: 3106 break; 3107 3108 case Op_AddReductionVI: 3109 case Op_AddReductionVL: 3110 case Op_AddReductionVF: 3111 case Op_AddReductionVD: 3112 case Op_MulReductionVI: 3113 case Op_MulReductionVF: 3114 case Op_MulReductionVD: 3115 break; 3116 3117 case Op_PackB: 3118 case Op_PackS: 3119 case Op_PackI: 3120 case Op_PackF: 3121 case Op_PackL: 3122 case Op_PackD: 3123 if (n->req()-1 > 2) { 3124 // Replace many operand PackNodes with a binary tree for matching 3125 PackNode* p = (PackNode*) n; 3126 Node* btp = p->binary_tree_pack(1, n->req()); 3127 n->subsume_by(btp, this); 3128 } 3129 break; 3130 case Op_Loop: 3131 case Op_CountedLoop: 3132 if (n->as_Loop()->is_inner_loop()) { 3133 frc.inc_inner_loop_count(); 3134 } 3135 break; 3136 case Op_LShiftI: 3137 case Op_RShiftI: 3138 case Op_URShiftI: 3139 case Op_LShiftL: 3140 case Op_RShiftL: 3141 case Op_URShiftL: 3142 if (Matcher::need_masked_shift_count) { 3143 // The cpu's shift instructions don't restrict the count to the 3144 // lower 5/6 bits. We need to do the masking ourselves. 3145 Node* in2 = n->in(2); 3146 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3147 const TypeInt* t = in2->find_int_type(); 3148 if (t != NULL && t->is_con()) { 3149 juint shift = t->get_con(); 3150 if (shift > mask) { // Unsigned cmp 3151 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3152 } 3153 } else { 3154 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3155 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3156 n->set_req(2, shift); 3157 } 3158 } 3159 if (in2->outcnt() == 0) { // Remove dead node 3160 in2->disconnect_inputs(NULL, this); 3161 } 3162 } 3163 break; 3164 case Op_MemBarStoreStore: 3165 case Op_MemBarRelease: 3166 // Break the link with AllocateNode: it is no longer useful and 3167 // confuses register allocation. 3168 if (n->req() > MemBarNode::Precedent) { 3169 n->set_req(MemBarNode::Precedent, top()); 3170 } 3171 break; 3172 default: 3173 assert( !n->is_Call(), "" ); 3174 assert( !n->is_Mem(), "" ); 3175 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3176 break; 3177 } 3178 3179 // Collect CFG split points 3180 if (n->is_MultiBranch()) 3181 frc._tests.push(n); 3182 } 3183 3184 //------------------------------final_graph_reshaping_walk--------------------- 3185 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3186 // requires that the walk visits a node's inputs before visiting the node. 3187 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3188 ResourceArea *area = Thread::current()->resource_area(); 3189 Unique_Node_List sfpt(area); 3190 3191 frc._visited.set(root->_idx); // first, mark node as visited 3192 uint cnt = root->req(); 3193 Node *n = root; 3194 uint i = 0; 3195 while (true) { 3196 if (i < cnt) { 3197 // Place all non-visited non-null inputs onto stack 3198 Node* m = n->in(i); 3199 ++i; 3200 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3201 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3202 // compute worst case interpreter size in case of a deoptimization 3203 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3204 3205 sfpt.push(m); 3206 } 3207 cnt = m->req(); 3208 nstack.push(n, i); // put on stack parent and next input's index 3209 n = m; 3210 i = 0; 3211 } 3212 } else { 3213 // Now do post-visit work 3214 final_graph_reshaping_impl( n, frc ); 3215 if (nstack.is_empty()) 3216 break; // finished 3217 n = nstack.node(); // Get node from stack 3218 cnt = n->req(); 3219 i = nstack.index(); 3220 nstack.pop(); // Shift to the next node on stack 3221 } 3222 } 3223 3224 // Skip next transformation if compressed oops are not used. 3225 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3226 (!UseCompressedOops && !UseCompressedClassPointers)) 3227 return; 3228 3229 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3230 // It could be done for an uncommon traps or any safepoints/calls 3231 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3232 while (sfpt.size() > 0) { 3233 n = sfpt.pop(); 3234 JVMState *jvms = n->as_SafePoint()->jvms(); 3235 assert(jvms != NULL, "sanity"); 3236 int start = jvms->debug_start(); 3237 int end = n->req(); 3238 bool is_uncommon = (n->is_CallStaticJava() && 3239 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3240 for (int j = start; j < end; j++) { 3241 Node* in = n->in(j); 3242 if (in->is_DecodeNarrowPtr()) { 3243 bool safe_to_skip = true; 3244 if (!is_uncommon ) { 3245 // Is it safe to skip? 3246 for (uint i = 0; i < in->outcnt(); i++) { 3247 Node* u = in->raw_out(i); 3248 if (!u->is_SafePoint() || 3249 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3250 safe_to_skip = false; 3251 } 3252 } 3253 } 3254 if (safe_to_skip) { 3255 n->set_req(j, in->in(1)); 3256 } 3257 if (in->outcnt() == 0) { 3258 in->disconnect_inputs(NULL, this); 3259 } 3260 } 3261 } 3262 } 3263 } 3264 3265 //------------------------------final_graph_reshaping-------------------------- 3266 // Final Graph Reshaping. 3267 // 3268 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3269 // and not commoned up and forced early. Must come after regular 3270 // optimizations to avoid GVN undoing the cloning. Clone constant 3271 // inputs to Loop Phis; these will be split by the allocator anyways. 3272 // Remove Opaque nodes. 3273 // (2) Move last-uses by commutative operations to the left input to encourage 3274 // Intel update-in-place two-address operations and better register usage 3275 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3276 // calls canonicalizing them back. 3277 // (3) Count the number of double-precision FP ops, single-precision FP ops 3278 // and call sites. On Intel, we can get correct rounding either by 3279 // forcing singles to memory (requires extra stores and loads after each 3280 // FP bytecode) or we can set a rounding mode bit (requires setting and 3281 // clearing the mode bit around call sites). The mode bit is only used 3282 // if the relative frequency of single FP ops to calls is low enough. 3283 // This is a key transform for SPEC mpeg_audio. 3284 // (4) Detect infinite loops; blobs of code reachable from above but not 3285 // below. Several of the Code_Gen algorithms fail on such code shapes, 3286 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3287 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3288 // Detection is by looking for IfNodes where only 1 projection is 3289 // reachable from below or CatchNodes missing some targets. 3290 // (5) Assert for insane oop offsets in debug mode. 3291 3292 bool Compile::final_graph_reshaping() { 3293 // an infinite loop may have been eliminated by the optimizer, 3294 // in which case the graph will be empty. 3295 if (root()->req() == 1) { 3296 record_method_not_compilable("trivial infinite loop"); 3297 return true; 3298 } 3299 3300 // Expensive nodes have their control input set to prevent the GVN 3301 // from freely commoning them. There's no GVN beyond this point so 3302 // no need to keep the control input. We want the expensive nodes to 3303 // be freely moved to the least frequent code path by gcm. 3304 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3305 for (int i = 0; i < expensive_count(); i++) { 3306 _expensive_nodes->at(i)->set_req(0, NULL); 3307 } 3308 3309 Final_Reshape_Counts frc; 3310 3311 // Visit everybody reachable! 3312 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3313 Node_Stack nstack(unique() >> 1); 3314 final_graph_reshaping_walk(nstack, root(), frc); 3315 3316 // Check for unreachable (from below) code (i.e., infinite loops). 3317 for( uint i = 0; i < frc._tests.size(); i++ ) { 3318 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3319 // Get number of CFG targets. 3320 // Note that PCTables include exception targets after calls. 3321 uint required_outcnt = n->required_outcnt(); 3322 if (n->outcnt() != required_outcnt) { 3323 // Check for a few special cases. Rethrow Nodes never take the 3324 // 'fall-thru' path, so expected kids is 1 less. 3325 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3326 if (n->in(0)->in(0)->is_Call()) { 3327 CallNode *call = n->in(0)->in(0)->as_Call(); 3328 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3329 required_outcnt--; // Rethrow always has 1 less kid 3330 } else if (call->req() > TypeFunc::Parms && 3331 call->is_CallDynamicJava()) { 3332 // Check for null receiver. In such case, the optimizer has 3333 // detected that the virtual call will always result in a null 3334 // pointer exception. The fall-through projection of this CatchNode 3335 // will not be populated. 3336 Node *arg0 = call->in(TypeFunc::Parms); 3337 if (arg0->is_Type() && 3338 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3339 required_outcnt--; 3340 } 3341 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3342 call->req() > TypeFunc::Parms+1 && 3343 call->is_CallStaticJava()) { 3344 // Check for negative array length. In such case, the optimizer has 3345 // detected that the allocation attempt will always result in an 3346 // exception. There is no fall-through projection of this CatchNode . 3347 Node *arg1 = call->in(TypeFunc::Parms+1); 3348 if (arg1->is_Type() && 3349 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3350 required_outcnt--; 3351 } 3352 } 3353 } 3354 } 3355 // Recheck with a better notion of 'required_outcnt' 3356 if (n->outcnt() != required_outcnt) { 3357 record_method_not_compilable("malformed control flow"); 3358 return true; // Not all targets reachable! 3359 } 3360 } 3361 // Check that I actually visited all kids. Unreached kids 3362 // must be infinite loops. 3363 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3364 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3365 record_method_not_compilable("infinite loop"); 3366 return true; // Found unvisited kid; must be unreach 3367 } 3368 } 3369 3370 // If original bytecodes contained a mixture of floats and doubles 3371 // check if the optimizer has made it homogenous, item (3). 3372 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3373 frc.get_float_count() > 32 && 3374 frc.get_double_count() == 0 && 3375 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3376 set_24_bit_selection_and_mode( false, true ); 3377 } 3378 3379 set_java_calls(frc.get_java_call_count()); 3380 set_inner_loops(frc.get_inner_loop_count()); 3381 3382 // No infinite loops, no reason to bail out. 3383 return false; 3384 } 3385 3386 //-----------------------------too_many_traps---------------------------------- 3387 // Report if there are too many traps at the current method and bci. 3388 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3389 bool Compile::too_many_traps(ciMethod* method, 3390 int bci, 3391 Deoptimization::DeoptReason reason) { 3392 if (method->has_injected_profile()) { 3393 return false; 3394 } 3395 ciMethodData* md = method->method_data(); 3396 if (md->is_empty()) { 3397 // Assume the trap has not occurred, or that it occurred only 3398 // because of a transient condition during start-up in the interpreter. 3399 return false; 3400 } 3401 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3402 if (md->has_trap_at(bci, m, reason) != 0) { 3403 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3404 // Also, if there are multiple reasons, or if there is no per-BCI record, 3405 // assume the worst. 3406 if (log()) 3407 log()->elem("observe trap='%s' count='%d'", 3408 Deoptimization::trap_reason_name(reason), 3409 md->trap_count(reason)); 3410 return true; 3411 } else { 3412 // Ignore method/bci and see if there have been too many globally. 3413 return too_many_traps(reason, md); 3414 } 3415 } 3416 3417 // Less-accurate variant which does not require a method and bci. 3418 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3419 ciMethodData* logmd) { 3420 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3421 // Too many traps globally. 3422 // Note that we use cumulative trap_count, not just md->trap_count. 3423 if (log()) { 3424 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3425 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3426 Deoptimization::trap_reason_name(reason), 3427 mcount, trap_count(reason)); 3428 } 3429 return true; 3430 } else { 3431 // The coast is clear. 3432 return false; 3433 } 3434 } 3435 3436 //--------------------------too_many_recompiles-------------------------------- 3437 // Report if there are too many recompiles at the current method and bci. 3438 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3439 // Is not eager to return true, since this will cause the compiler to use 3440 // Action_none for a trap point, to avoid too many recompilations. 3441 bool Compile::too_many_recompiles(ciMethod* method, 3442 int bci, 3443 Deoptimization::DeoptReason reason) { 3444 if (method->has_injected_profile()) { 3445 return false; 3446 } 3447 ciMethodData* md = method->method_data(); 3448 if (md->is_empty()) { 3449 // Assume the trap has not occurred, or that it occurred only 3450 // because of a transient condition during start-up in the interpreter. 3451 return false; 3452 } 3453 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3454 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3455 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3456 Deoptimization::DeoptReason per_bc_reason 3457 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3458 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3459 if ((per_bc_reason == Deoptimization::Reason_none 3460 || md->has_trap_at(bci, m, reason) != 0) 3461 // The trap frequency measure we care about is the recompile count: 3462 && md->trap_recompiled_at(bci, m) 3463 && md->overflow_recompile_count() >= bc_cutoff) { 3464 // Do not emit a trap here if it has already caused recompilations. 3465 // Also, if there are multiple reasons, or if there is no per-BCI record, 3466 // assume the worst. 3467 if (log()) 3468 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3469 Deoptimization::trap_reason_name(reason), 3470 md->trap_count(reason), 3471 md->overflow_recompile_count()); 3472 return true; 3473 } else if (trap_count(reason) != 0 3474 && decompile_count() >= m_cutoff) { 3475 // Too many recompiles globally, and we have seen this sort of trap. 3476 // Use cumulative decompile_count, not just md->decompile_count. 3477 if (log()) 3478 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3479 Deoptimization::trap_reason_name(reason), 3480 md->trap_count(reason), trap_count(reason), 3481 md->decompile_count(), decompile_count()); 3482 return true; 3483 } else { 3484 // The coast is clear. 3485 return false; 3486 } 3487 } 3488 3489 // Compute when not to trap. Used by matching trap based nodes and 3490 // NullCheck optimization. 3491 void Compile::set_allowed_deopt_reasons() { 3492 _allowed_reasons = 0; 3493 if (is_method_compilation()) { 3494 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3495 assert(rs < BitsPerInt, "recode bit map"); 3496 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3497 _allowed_reasons |= nth_bit(rs); 3498 } 3499 } 3500 } 3501 } 3502 3503 #ifndef PRODUCT 3504 //------------------------------verify_graph_edges--------------------------- 3505 // Walk the Graph and verify that there is a one-to-one correspondence 3506 // between Use-Def edges and Def-Use edges in the graph. 3507 void Compile::verify_graph_edges(bool no_dead_code) { 3508 if (VerifyGraphEdges) { 3509 ResourceArea *area = Thread::current()->resource_area(); 3510 Unique_Node_List visited(area); 3511 // Call recursive graph walk to check edges 3512 _root->verify_edges(visited); 3513 if (no_dead_code) { 3514 // Now make sure that no visited node is used by an unvisited node. 3515 bool dead_nodes = false; 3516 Unique_Node_List checked(area); 3517 while (visited.size() > 0) { 3518 Node* n = visited.pop(); 3519 checked.push(n); 3520 for (uint i = 0; i < n->outcnt(); i++) { 3521 Node* use = n->raw_out(i); 3522 if (checked.member(use)) continue; // already checked 3523 if (visited.member(use)) continue; // already in the graph 3524 if (use->is_Con()) continue; // a dead ConNode is OK 3525 // At this point, we have found a dead node which is DU-reachable. 3526 if (!dead_nodes) { 3527 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3528 dead_nodes = true; 3529 } 3530 use->dump(2); 3531 tty->print_cr("---"); 3532 checked.push(use); // No repeats; pretend it is now checked. 3533 } 3534 } 3535 assert(!dead_nodes, "using nodes must be reachable from root"); 3536 } 3537 } 3538 } 3539 3540 // Verify GC barriers consistency 3541 // Currently supported: 3542 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3543 void Compile::verify_barriers() { 3544 if (UseG1GC) { 3545 // Verify G1 pre-barriers 3546 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3547 3548 ResourceArea *area = Thread::current()->resource_area(); 3549 Unique_Node_List visited(area); 3550 Node_List worklist(area); 3551 // We're going to walk control flow backwards starting from the Root 3552 worklist.push(_root); 3553 while (worklist.size() > 0) { 3554 Node* x = worklist.pop(); 3555 if (x == NULL || x == top()) continue; 3556 if (visited.member(x)) { 3557 continue; 3558 } else { 3559 visited.push(x); 3560 } 3561 3562 if (x->is_Region()) { 3563 for (uint i = 1; i < x->req(); i++) { 3564 worklist.push(x->in(i)); 3565 } 3566 } else { 3567 worklist.push(x->in(0)); 3568 // We are looking for the pattern: 3569 // /->ThreadLocal 3570 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3571 // \->ConI(0) 3572 // We want to verify that the If and the LoadB have the same control 3573 // See GraphKit::g1_write_barrier_pre() 3574 if (x->is_If()) { 3575 IfNode *iff = x->as_If(); 3576 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3577 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3578 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3579 && cmp->in(1)->is_Load()) { 3580 LoadNode* load = cmp->in(1)->as_Load(); 3581 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3582 && load->in(2)->in(3)->is_Con() 3583 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3584 3585 Node* if_ctrl = iff->in(0); 3586 Node* load_ctrl = load->in(0); 3587 3588 if (if_ctrl != load_ctrl) { 3589 // Skip possible CProj->NeverBranch in infinite loops 3590 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3591 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3592 if_ctrl = if_ctrl->in(0)->in(0); 3593 } 3594 } 3595 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3596 } 3597 } 3598 } 3599 } 3600 } 3601 } 3602 } 3603 } 3604 3605 #endif 3606 3607 // The Compile object keeps track of failure reasons separately from the ciEnv. 3608 // This is required because there is not quite a 1-1 relation between the 3609 // ciEnv and its compilation task and the Compile object. Note that one 3610 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3611 // to backtrack and retry without subsuming loads. Other than this backtracking 3612 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3613 // by the logic in C2Compiler. 3614 void Compile::record_failure(const char* reason) { 3615 if (log() != NULL) { 3616 log()->elem("failure reason='%s' phase='compile'", reason); 3617 } 3618 if (_failure_reason == NULL) { 3619 // Record the first failure reason. 3620 _failure_reason = reason; 3621 } 3622 3623 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3624 C->print_method(PHASE_FAILURE); 3625 } 3626 _root = NULL; // flush the graph, too 3627 } 3628 3629 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3630 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3631 _phase_name(name), _dolog(CITimeVerbose) 3632 { 3633 if (_dolog) { 3634 C = Compile::current(); 3635 _log = C->log(); 3636 } else { 3637 C = NULL; 3638 _log = NULL; 3639 } 3640 if (_log != NULL) { 3641 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3642 _log->stamp(); 3643 _log->end_head(); 3644 } 3645 } 3646 3647 Compile::TracePhase::~TracePhase() { 3648 3649 C = Compile::current(); 3650 if (_dolog) { 3651 _log = C->log(); 3652 } else { 3653 _log = NULL; 3654 } 3655 3656 #ifdef ASSERT 3657 if (PrintIdealNodeCount) { 3658 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3659 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3660 } 3661 3662 if (VerifyIdealNodeCount) { 3663 Compile::current()->print_missing_nodes(); 3664 } 3665 #endif 3666 3667 if (_log != NULL) { 3668 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3669 } 3670 } 3671 3672 //============================================================================= 3673 // Two Constant's are equal when the type and the value are equal. 3674 bool Compile::Constant::operator==(const Constant& other) { 3675 if (type() != other.type() ) return false; 3676 if (can_be_reused() != other.can_be_reused()) return false; 3677 // For floating point values we compare the bit pattern. 3678 switch (type()) { 3679 case T_FLOAT: return (_v._value.i == other._v._value.i); 3680 case T_LONG: 3681 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3682 case T_OBJECT: 3683 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3684 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3685 case T_METADATA: return (_v._metadata == other._v._metadata); 3686 default: ShouldNotReachHere(); 3687 } 3688 return false; 3689 } 3690 3691 static int type_to_size_in_bytes(BasicType t) { 3692 switch (t) { 3693 case T_LONG: return sizeof(jlong ); 3694 case T_FLOAT: return sizeof(jfloat ); 3695 case T_DOUBLE: return sizeof(jdouble); 3696 case T_METADATA: return sizeof(Metadata*); 3697 // We use T_VOID as marker for jump-table entries (labels) which 3698 // need an internal word relocation. 3699 case T_VOID: 3700 case T_ADDRESS: 3701 case T_OBJECT: return sizeof(jobject); 3702 } 3703 3704 ShouldNotReachHere(); 3705 return -1; 3706 } 3707 3708 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3709 // sort descending 3710 if (a->freq() > b->freq()) return -1; 3711 if (a->freq() < b->freq()) return 1; 3712 return 0; 3713 } 3714 3715 void Compile::ConstantTable::calculate_offsets_and_size() { 3716 // First, sort the array by frequencies. 3717 _constants.sort(qsort_comparator); 3718 3719 #ifdef ASSERT 3720 // Make sure all jump-table entries were sorted to the end of the 3721 // array (they have a negative frequency). 3722 bool found_void = false; 3723 for (int i = 0; i < _constants.length(); i++) { 3724 Constant con = _constants.at(i); 3725 if (con.type() == T_VOID) 3726 found_void = true; // jump-tables 3727 else 3728 assert(!found_void, "wrong sorting"); 3729 } 3730 #endif 3731 3732 int offset = 0; 3733 for (int i = 0; i < _constants.length(); i++) { 3734 Constant* con = _constants.adr_at(i); 3735 3736 // Align offset for type. 3737 int typesize = type_to_size_in_bytes(con->type()); 3738 offset = align_size_up(offset, typesize); 3739 con->set_offset(offset); // set constant's offset 3740 3741 if (con->type() == T_VOID) { 3742 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3743 offset = offset + typesize * n->outcnt(); // expand jump-table 3744 } else { 3745 offset = offset + typesize; 3746 } 3747 } 3748 3749 // Align size up to the next section start (which is insts; see 3750 // CodeBuffer::align_at_start). 3751 assert(_size == -1, "already set?"); 3752 _size = align_size_up(offset, CodeEntryAlignment); 3753 } 3754 3755 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3756 MacroAssembler _masm(&cb); 3757 for (int i = 0; i < _constants.length(); i++) { 3758 Constant con = _constants.at(i); 3759 address constant_addr; 3760 switch (con.type()) { 3761 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3762 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3763 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3764 case T_OBJECT: { 3765 jobject obj = con.get_jobject(); 3766 int oop_index = _masm.oop_recorder()->find_index(obj); 3767 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3768 break; 3769 } 3770 case T_ADDRESS: { 3771 address addr = (address) con.get_jobject(); 3772 constant_addr = _masm.address_constant(addr); 3773 break; 3774 } 3775 // We use T_VOID as marker for jump-table entries (labels) which 3776 // need an internal word relocation. 3777 case T_VOID: { 3778 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3779 // Fill the jump-table with a dummy word. The real value is 3780 // filled in later in fill_jump_table. 3781 address dummy = (address) n; 3782 constant_addr = _masm.address_constant(dummy); 3783 // Expand jump-table 3784 for (uint i = 1; i < n->outcnt(); i++) { 3785 address temp_addr = _masm.address_constant(dummy + i); 3786 assert(temp_addr, "consts section too small"); 3787 } 3788 break; 3789 } 3790 case T_METADATA: { 3791 Metadata* obj = con.get_metadata(); 3792 int metadata_index = _masm.oop_recorder()->find_index(obj); 3793 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3794 break; 3795 } 3796 default: ShouldNotReachHere(); 3797 } 3798 assert(constant_addr, "consts section too small"); 3799 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3800 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3801 } 3802 } 3803 3804 int Compile::ConstantTable::find_offset(Constant& con) const { 3805 int idx = _constants.find(con); 3806 assert(idx != -1, "constant must be in constant table"); 3807 int offset = _constants.at(idx).offset(); 3808 assert(offset != -1, "constant table not emitted yet?"); 3809 return offset; 3810 } 3811 3812 void Compile::ConstantTable::add(Constant& con) { 3813 if (con.can_be_reused()) { 3814 int idx = _constants.find(con); 3815 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3816 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3817 return; 3818 } 3819 } 3820 (void) _constants.append(con); 3821 } 3822 3823 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3824 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3825 Constant con(type, value, b->_freq); 3826 add(con); 3827 return con; 3828 } 3829 3830 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3831 Constant con(metadata); 3832 add(con); 3833 return con; 3834 } 3835 3836 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3837 jvalue value; 3838 BasicType type = oper->type()->basic_type(); 3839 switch (type) { 3840 case T_LONG: value.j = oper->constantL(); break; 3841 case T_FLOAT: value.f = oper->constantF(); break; 3842 case T_DOUBLE: value.d = oper->constantD(); break; 3843 case T_OBJECT: 3844 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3845 case T_METADATA: return add((Metadata*)oper->constant()); break; 3846 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3847 } 3848 return add(n, type, value); 3849 } 3850 3851 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3852 jvalue value; 3853 // We can use the node pointer here to identify the right jump-table 3854 // as this method is called from Compile::Fill_buffer right before 3855 // the MachNodes are emitted and the jump-table is filled (means the 3856 // MachNode pointers do not change anymore). 3857 value.l = (jobject) n; 3858 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3859 add(con); 3860 return con; 3861 } 3862 3863 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3864 // If called from Compile::scratch_emit_size do nothing. 3865 if (Compile::current()->in_scratch_emit_size()) return; 3866 3867 assert(labels.is_nonempty(), "must be"); 3868 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3869 3870 // Since MachConstantNode::constant_offset() also contains 3871 // table_base_offset() we need to subtract the table_base_offset() 3872 // to get the plain offset into the constant table. 3873 int offset = n->constant_offset() - table_base_offset(); 3874 3875 MacroAssembler _masm(&cb); 3876 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3877 3878 for (uint i = 0; i < n->outcnt(); i++) { 3879 address* constant_addr = &jump_table_base[i]; 3880 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))); 3881 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3882 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3883 } 3884 } 3885 3886 //----------------------------static_subtype_check----------------------------- 3887 // Shortcut important common cases when superklass is exact: 3888 // (0) superklass is java.lang.Object (can occur in reflective code) 3889 // (1) subklass is already limited to a subtype of superklass => always ok 3890 // (2) subklass does not overlap with superklass => always fail 3891 // (3) superklass has NO subtypes and we can check with a simple compare. 3892 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3893 if (StressReflectiveCode) { 3894 return SSC_full_test; // Let caller generate the general case. 3895 } 3896 3897 if (superk == env()->Object_klass()) { 3898 return SSC_always_true; // (0) this test cannot fail 3899 } 3900 3901 ciType* superelem = superk; 3902 if (superelem->is_array_klass()) 3903 superelem = superelem->as_array_klass()->base_element_type(); 3904 3905 if (!subk->is_interface()) { // cannot trust static interface types yet 3906 if (subk->is_subtype_of(superk)) { 3907 return SSC_always_true; // (1) false path dead; no dynamic test needed 3908 } 3909 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3910 !superk->is_subtype_of(subk)) { 3911 return SSC_always_false; 3912 } 3913 } 3914 3915 // If casting to an instance klass, it must have no subtypes 3916 if (superk->is_interface()) { 3917 // Cannot trust interfaces yet. 3918 // %%% S.B. superk->nof_implementors() == 1 3919 } else if (superelem->is_instance_klass()) { 3920 ciInstanceKlass* ik = superelem->as_instance_klass(); 3921 if (!ik->has_subklass() && !ik->is_interface()) { 3922 if (!ik->is_final()) { 3923 // Add a dependency if there is a chance of a later subclass. 3924 dependencies()->assert_leaf_type(ik); 3925 } 3926 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3927 } 3928 } else { 3929 // A primitive array type has no subtypes. 3930 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3931 } 3932 3933 return SSC_full_test; 3934 } 3935 3936 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3937 #ifdef _LP64 3938 // The scaled index operand to AddP must be a clean 64-bit value. 3939 // Java allows a 32-bit int to be incremented to a negative 3940 // value, which appears in a 64-bit register as a large 3941 // positive number. Using that large positive number as an 3942 // operand in pointer arithmetic has bad consequences. 3943 // On the other hand, 32-bit overflow is rare, and the possibility 3944 // can often be excluded, if we annotate the ConvI2L node with 3945 // a type assertion that its value is known to be a small positive 3946 // number. (The prior range check has ensured this.) 3947 // This assertion is used by ConvI2LNode::Ideal. 3948 int index_max = max_jint - 1; // array size is max_jint, index is one less 3949 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3950 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3951 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3952 #endif 3953 return idx; 3954 } 3955 3956 // The message about the current inlining is accumulated in 3957 // _print_inlining_stream and transfered into the _print_inlining_list 3958 // once we know whether inlining succeeds or not. For regular 3959 // inlining, messages are appended to the buffer pointed by 3960 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3961 // a new buffer is added after _print_inlining_idx in the list. This 3962 // way we can update the inlining message for late inlining call site 3963 // when the inlining is attempted again. 3964 void Compile::print_inlining_init() { 3965 if (print_inlining() || print_intrinsics()) { 3966 _print_inlining_stream = new stringStream(); 3967 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3968 } 3969 } 3970 3971 void Compile::print_inlining_reinit() { 3972 if (print_inlining() || print_intrinsics()) { 3973 // Re allocate buffer when we change ResourceMark 3974 _print_inlining_stream = new stringStream(); 3975 } 3976 } 3977 3978 void Compile::print_inlining_reset() { 3979 _print_inlining_stream->reset(); 3980 } 3981 3982 void Compile::print_inlining_commit() { 3983 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3984 // Transfer the message from _print_inlining_stream to the current 3985 // _print_inlining_list buffer and clear _print_inlining_stream. 3986 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3987 print_inlining_reset(); 3988 } 3989 3990 void Compile::print_inlining_push() { 3991 // Add new buffer to the _print_inlining_list at current position 3992 _print_inlining_idx++; 3993 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3994 } 3995 3996 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3997 return _print_inlining_list->at(_print_inlining_idx); 3998 } 3999 4000 void Compile::print_inlining_update(CallGenerator* cg) { 4001 if (print_inlining() || print_intrinsics()) { 4002 if (!cg->is_late_inline()) { 4003 if (print_inlining_current().cg() != NULL) { 4004 print_inlining_push(); 4005 } 4006 print_inlining_commit(); 4007 } else { 4008 if (print_inlining_current().cg() != cg && 4009 (print_inlining_current().cg() != NULL || 4010 print_inlining_current().ss()->size() != 0)) { 4011 print_inlining_push(); 4012 } 4013 print_inlining_commit(); 4014 print_inlining_current().set_cg(cg); 4015 } 4016 } 4017 } 4018 4019 void Compile::print_inlining_move_to(CallGenerator* cg) { 4020 // We resume inlining at a late inlining call site. Locate the 4021 // corresponding inlining buffer so that we can update it. 4022 if (print_inlining()) { 4023 for (int i = 0; i < _print_inlining_list->length(); i++) { 4024 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4025 _print_inlining_idx = i; 4026 return; 4027 } 4028 } 4029 ShouldNotReachHere(); 4030 } 4031 } 4032 4033 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4034 if (print_inlining()) { 4035 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4036 assert(print_inlining_current().cg() == cg, "wrong entry"); 4037 // replace message with new message 4038 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4039 print_inlining_commit(); 4040 print_inlining_current().set_cg(cg); 4041 } 4042 } 4043 4044 void Compile::print_inlining_assert_ready() { 4045 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4046 } 4047 4048 void Compile::process_print_inlining() { 4049 bool do_print_inlining = print_inlining() || print_intrinsics(); 4050 if (do_print_inlining || log() != NULL) { 4051 // Print inlining message for candidates that we couldn't inline 4052 // for lack of space 4053 for (int i = 0; i < _late_inlines.length(); i++) { 4054 CallGenerator* cg = _late_inlines.at(i); 4055 if (!cg->is_mh_late_inline()) { 4056 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4057 if (do_print_inlining) { 4058 cg->print_inlining_late(msg); 4059 } 4060 log_late_inline_failure(cg, msg); 4061 } 4062 } 4063 } 4064 if (do_print_inlining) { 4065 ResourceMark rm; 4066 stringStream ss; 4067 for (int i = 0; i < _print_inlining_list->length(); i++) { 4068 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4069 } 4070 size_t end = ss.size(); 4071 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4072 strncpy(_print_inlining_output, ss.base(), end+1); 4073 _print_inlining_output[end] = 0; 4074 } 4075 } 4076 4077 void Compile::dump_print_inlining() { 4078 if (_print_inlining_output != NULL) { 4079 tty->print_raw(_print_inlining_output); 4080 } 4081 } 4082 4083 void Compile::log_late_inline(CallGenerator* cg) { 4084 if (log() != NULL) { 4085 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4086 cg->unique_id()); 4087 JVMState* p = cg->call_node()->jvms(); 4088 while (p != NULL) { 4089 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4090 p = p->caller(); 4091 } 4092 log()->tail("late_inline"); 4093 } 4094 } 4095 4096 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4097 log_late_inline(cg); 4098 if (log() != NULL) { 4099 log()->inline_fail(msg); 4100 } 4101 } 4102 4103 void Compile::log_inline_id(CallGenerator* cg) { 4104 if (log() != NULL) { 4105 // The LogCompilation tool needs a unique way to identify late 4106 // inline call sites. This id must be unique for this call site in 4107 // this compilation. Try to have it unique across compilations as 4108 // well because it can be convenient when grepping through the log 4109 // file. 4110 // Distinguish OSR compilations from others in case CICountOSR is 4111 // on. 4112 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4113 cg->set_unique_id(id); 4114 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4115 } 4116 } 4117 4118 void Compile::log_inline_failure(const char* msg) { 4119 if (C->log() != NULL) { 4120 C->log()->inline_fail(msg); 4121 } 4122 } 4123 4124 4125 // Dump inlining replay data to the stream. 4126 // Don't change thread state and acquire any locks. 4127 void Compile::dump_inline_data(outputStream* out) { 4128 InlineTree* inl_tree = ilt(); 4129 if (inl_tree != NULL) { 4130 out->print(" inline %d", inl_tree->count()); 4131 inl_tree->dump_replay_data(out); 4132 } 4133 } 4134 4135 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4136 if (n1->Opcode() < n2->Opcode()) return -1; 4137 else if (n1->Opcode() > n2->Opcode()) return 1; 4138 4139 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())); 4140 for (uint i = 1; i < n1->req(); i++) { 4141 if (n1->in(i) < n2->in(i)) return -1; 4142 else if (n1->in(i) > n2->in(i)) return 1; 4143 } 4144 4145 return 0; 4146 } 4147 4148 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4149 Node* n1 = *n1p; 4150 Node* n2 = *n2p; 4151 4152 return cmp_expensive_nodes(n1, n2); 4153 } 4154 4155 void Compile::sort_expensive_nodes() { 4156 if (!expensive_nodes_sorted()) { 4157 _expensive_nodes->sort(cmp_expensive_nodes); 4158 } 4159 } 4160 4161 bool Compile::expensive_nodes_sorted() const { 4162 for (int i = 1; i < _expensive_nodes->length(); i++) { 4163 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4164 return false; 4165 } 4166 } 4167 return true; 4168 } 4169 4170 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4171 if (_expensive_nodes->length() == 0) { 4172 return false; 4173 } 4174 4175 assert(OptimizeExpensiveOps, "optimization off?"); 4176 4177 // Take this opportunity to remove dead nodes from the list 4178 int j = 0; 4179 for (int i = 0; i < _expensive_nodes->length(); i++) { 4180 Node* n = _expensive_nodes->at(i); 4181 if (!n->is_unreachable(igvn)) { 4182 assert(n->is_expensive(), "should be expensive"); 4183 _expensive_nodes->at_put(j, n); 4184 j++; 4185 } 4186 } 4187 _expensive_nodes->trunc_to(j); 4188 4189 // Then sort the list so that similar nodes are next to each other 4190 // and check for at least two nodes of identical kind with same data 4191 // inputs. 4192 sort_expensive_nodes(); 4193 4194 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4195 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4196 return true; 4197 } 4198 } 4199 4200 return false; 4201 } 4202 4203 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4204 if (_expensive_nodes->length() == 0) { 4205 return; 4206 } 4207 4208 assert(OptimizeExpensiveOps, "optimization off?"); 4209 4210 // Sort to bring similar nodes next to each other and clear the 4211 // control input of nodes for which there's only a single copy. 4212 sort_expensive_nodes(); 4213 4214 int j = 0; 4215 int identical = 0; 4216 int i = 0; 4217 bool modified = false; 4218 for (; i < _expensive_nodes->length()-1; i++) { 4219 assert(j <= i, "can't write beyond current index"); 4220 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4221 identical++; 4222 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4223 continue; 4224 } 4225 if (identical > 0) { 4226 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4227 identical = 0; 4228 } else { 4229 Node* n = _expensive_nodes->at(i); 4230 igvn.replace_input_of(n, 0, NULL); 4231 igvn.hash_insert(n); 4232 modified = true; 4233 } 4234 } 4235 if (identical > 0) { 4236 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4237 } else if (_expensive_nodes->length() >= 1) { 4238 Node* n = _expensive_nodes->at(i); 4239 igvn.replace_input_of(n, 0, NULL); 4240 igvn.hash_insert(n); 4241 modified = true; 4242 } 4243 _expensive_nodes->trunc_to(j); 4244 if (modified) { 4245 igvn.optimize(); 4246 } 4247 } 4248 4249 void Compile::add_expensive_node(Node * n) { 4250 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4251 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4252 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4253 if (OptimizeExpensiveOps) { 4254 _expensive_nodes->append(n); 4255 } else { 4256 // Clear control input and let IGVN optimize expensive nodes if 4257 // OptimizeExpensiveOps is off. 4258 n->set_req(0, NULL); 4259 } 4260 } 4261 4262 /** 4263 * Remove the speculative part of types and clean up the graph 4264 */ 4265 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4266 if (UseTypeSpeculation) { 4267 Unique_Node_List worklist; 4268 worklist.push(root()); 4269 int modified = 0; 4270 // Go over all type nodes that carry a speculative type, drop the 4271 // speculative part of the type and enqueue the node for an igvn 4272 // which may optimize it out. 4273 for (uint next = 0; next < worklist.size(); ++next) { 4274 Node *n = worklist.at(next); 4275 if (n->is_Type()) { 4276 TypeNode* tn = n->as_Type(); 4277 const Type* t = tn->type(); 4278 const Type* t_no_spec = t->remove_speculative(); 4279 if (t_no_spec != t) { 4280 bool in_hash = igvn.hash_delete(n); 4281 assert(in_hash, "node should be in igvn hash table"); 4282 tn->set_type(t_no_spec); 4283 igvn.hash_insert(n); 4284 igvn._worklist.push(n); // give it a chance to go away 4285 modified++; 4286 } 4287 } 4288 uint max = n->len(); 4289 for( uint i = 0; i < max; ++i ) { 4290 Node *m = n->in(i); 4291 if (not_a_node(m)) continue; 4292 worklist.push(m); 4293 } 4294 } 4295 // Drop the speculative part of all types in the igvn's type table 4296 igvn.remove_speculative_types(); 4297 if (modified > 0) { 4298 igvn.optimize(); 4299 } 4300 #ifdef ASSERT 4301 // Verify that after the IGVN is over no speculative type has resurfaced 4302 worklist.clear(); 4303 worklist.push(root()); 4304 for (uint next = 0; next < worklist.size(); ++next) { 4305 Node *n = worklist.at(next); 4306 const Type* t = igvn.type_or_null(n); 4307 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4308 if (n->is_Type()) { 4309 t = n->as_Type()->type(); 4310 assert(t == t->remove_speculative(), "no more speculative types"); 4311 } 4312 uint max = n->len(); 4313 for( uint i = 0; i < max; ++i ) { 4314 Node *m = n->in(i); 4315 if (not_a_node(m)) continue; 4316 worklist.push(m); 4317 } 4318 } 4319 igvn.check_no_speculative_types(); 4320 #endif 4321 } 4322 } 4323 4324 // Auxiliary method to support randomized stressing/fuzzing. 4325 // 4326 // This method can be called the arbitrary number of times, with current count 4327 // as the argument. The logic allows selecting a single candidate from the 4328 // running list of candidates as follows: 4329 // int count = 0; 4330 // Cand* selected = null; 4331 // while(cand = cand->next()) { 4332 // if (randomized_select(++count)) { 4333 // selected = cand; 4334 // } 4335 // } 4336 // 4337 // Including count equalizes the chances any candidate is "selected". 4338 // This is useful when we don't have the complete list of candidates to choose 4339 // from uniformly. In this case, we need to adjust the randomicity of the 4340 // selection, or else we will end up biasing the selection towards the latter 4341 // candidates. 4342 // 4343 // Quick back-envelope calculation shows that for the list of n candidates 4344 // the equal probability for the candidate to persist as "best" can be 4345 // achieved by replacing it with "next" k-th candidate with the probability 4346 // of 1/k. It can be easily shown that by the end of the run, the 4347 // probability for any candidate is converged to 1/n, thus giving the 4348 // uniform distribution among all the candidates. 4349 // 4350 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4351 #define RANDOMIZED_DOMAIN_POW 29 4352 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4353 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4354 bool Compile::randomized_select(int count) { 4355 assert(count > 0, "only positive"); 4356 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4357 } 4358 4359 const char* CloneMap::debug_option_name = "CloneMapDebug"; 4360 CloneMap& Compile::clone_map() { return _clone_map; } 4361 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4362 4363 void NodeCloneInfo::dump() const { 4364 tty->print(" {%d:%d} ", idx(), gen()); 4365 } 4366 4367 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4368 uint64_t val = value(old->_idx); 4369 NodeCloneInfo cio(val); 4370 assert(val != 0, "old node should be in the map"); 4371 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4372 insert(nnn->_idx, cin.get()); 4373 #ifndef PRODUCT 4374 if (is_debug()) { 4375 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4376 } 4377 #endif 4378 } 4379 4380 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4381 NodeCloneInfo cio(value(old->_idx)); 4382 if (cio.get() == 0) { 4383 cio.set(old->_idx, 0); 4384 insert(old->_idx, cio.get()); 4385 #ifndef PRODUCT 4386 if (is_debug()) { 4387 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4388 } 4389 #endif 4390 } 4391 clone(old, nnn, gen); 4392 } 4393 4394 int CloneMap::max_gen() const { 4395 int g = 0; 4396 DictI di(_dict); 4397 for(; di.test(); ++di) { 4398 int t = gen(di._key); 4399 if (g < t) { 4400 g = t; 4401 #ifndef PRODUCT 4402 if (is_debug()) { 4403 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4404 } 4405 #endif 4406 } 4407 } 4408 return g; 4409 } 4410 4411 void CloneMap::dump(node_idx_t key) const { 4412 uint64_t val = value(key); 4413 if (val != 0) { 4414 NodeCloneInfo ni(val); 4415 ni.dump(); 4416 } 4417 }