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