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) { 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 StartNode* Compile::start() const { 1156 assert(!failing(), ""); 1157 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1158 Node* start = root()->fast_out(i); 1159 if( start->is_Start() ) 1160 return start->as_Start(); 1161 } 1162 fatal("Did not find Start node!"); 1163 return NULL; 1164 } 1165 1166 //-------------------------------immutable_memory------------------------------------- 1167 // Access immutable memory 1168 Node* Compile::immutable_memory() { 1169 if (_immutable_memory != NULL) { 1170 return _immutable_memory; 1171 } 1172 StartNode* s = start(); 1173 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1174 Node *p = s->fast_out(i); 1175 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1176 _immutable_memory = p; 1177 return _immutable_memory; 1178 } 1179 } 1180 ShouldNotReachHere(); 1181 return NULL; 1182 } 1183 1184 //----------------------set_cached_top_node------------------------------------ 1185 // Install the cached top node, and make sure Node::is_top works correctly. 1186 void Compile::set_cached_top_node(Node* tn) { 1187 if (tn != NULL) verify_top(tn); 1188 Node* old_top = _top; 1189 _top = tn; 1190 // Calling Node::setup_is_top allows the nodes the chance to adjust 1191 // their _out arrays. 1192 if (_top != NULL) _top->setup_is_top(); 1193 if (old_top != NULL) old_top->setup_is_top(); 1194 assert(_top == NULL || top()->is_top(), ""); 1195 } 1196 1197 #ifdef ASSERT 1198 uint Compile::count_live_nodes_by_graph_walk() { 1199 Unique_Node_List useful(comp_arena()); 1200 // Get useful node list by walking the graph. 1201 identify_useful_nodes(useful); 1202 return useful.size(); 1203 } 1204 1205 void Compile::print_missing_nodes() { 1206 1207 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1208 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1209 return; 1210 } 1211 1212 // This is an expensive function. It is executed only when the user 1213 // specifies VerifyIdealNodeCount option or otherwise knows the 1214 // additional work that needs to be done to identify reachable nodes 1215 // by walking the flow graph and find the missing ones using 1216 // _dead_node_list. 1217 1218 Unique_Node_List useful(comp_arena()); 1219 // Get useful node list by walking the graph. 1220 identify_useful_nodes(useful); 1221 1222 uint l_nodes = C->live_nodes(); 1223 uint l_nodes_by_walk = useful.size(); 1224 1225 if (l_nodes != l_nodes_by_walk) { 1226 if (_log != NULL) { 1227 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1228 _log->stamp(); 1229 _log->end_head(); 1230 } 1231 VectorSet& useful_member_set = useful.member_set(); 1232 int last_idx = l_nodes_by_walk; 1233 for (int i = 0; i < last_idx; i++) { 1234 if (useful_member_set.test(i)) { 1235 if (_dead_node_list.test(i)) { 1236 if (_log != NULL) { 1237 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1238 } 1239 if (PrintIdealNodeCount) { 1240 // Print the log message to tty 1241 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1242 useful.at(i)->dump(); 1243 } 1244 } 1245 } 1246 else if (! _dead_node_list.test(i)) { 1247 if (_log != NULL) { 1248 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1249 } 1250 if (PrintIdealNodeCount) { 1251 // Print the log message to tty 1252 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1253 } 1254 } 1255 } 1256 if (_log != NULL) { 1257 _log->tail("mismatched_nodes"); 1258 } 1259 } 1260 } 1261 void Compile::record_modified_node(Node* n) { 1262 if (_modified_nodes != NULL && !_inlining_incrementally && 1263 n->outcnt() != 0 && !n->is_Con()) { 1264 _modified_nodes->push(n); 1265 } 1266 } 1267 1268 void Compile::remove_modified_node(Node* n) { 1269 if (_modified_nodes != NULL) { 1270 _modified_nodes->remove(n); 1271 } 1272 } 1273 #endif 1274 1275 #ifndef PRODUCT 1276 void Compile::verify_top(Node* tn) const { 1277 if (tn != NULL) { 1278 assert(tn->is_Con(), "top node must be a constant"); 1279 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1280 assert(tn->in(0) != NULL, "must have live top node"); 1281 } 1282 } 1283 #endif 1284 1285 1286 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1287 1288 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1289 guarantee(arr != NULL, ""); 1290 int num_blocks = arr->length(); 1291 if (grow_by < num_blocks) grow_by = num_blocks; 1292 int num_notes = grow_by * _node_notes_block_size; 1293 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1294 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1295 while (num_notes > 0) { 1296 arr->append(notes); 1297 notes += _node_notes_block_size; 1298 num_notes -= _node_notes_block_size; 1299 } 1300 assert(num_notes == 0, "exact multiple, please"); 1301 } 1302 1303 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1304 if (source == NULL || dest == NULL) return false; 1305 1306 if (dest->is_Con()) 1307 return false; // Do not push debug info onto constants. 1308 1309 #ifdef ASSERT 1310 // Leave a bread crumb trail pointing to the original node: 1311 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1312 dest->set_debug_orig(source); 1313 } 1314 #endif 1315 1316 if (node_note_array() == NULL) 1317 return false; // Not collecting any notes now. 1318 1319 // This is a copy onto a pre-existing node, which may already have notes. 1320 // If both nodes have notes, do not overwrite any pre-existing notes. 1321 Node_Notes* source_notes = node_notes_at(source->_idx); 1322 if (source_notes == NULL || source_notes->is_clear()) return false; 1323 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1324 if (dest_notes == NULL || dest_notes->is_clear()) { 1325 return set_node_notes_at(dest->_idx, source_notes); 1326 } 1327 1328 Node_Notes merged_notes = (*source_notes); 1329 // The order of operations here ensures that dest notes will win... 1330 merged_notes.update_from(dest_notes); 1331 return set_node_notes_at(dest->_idx, &merged_notes); 1332 } 1333 1334 1335 //--------------------------allow_range_check_smearing------------------------- 1336 // Gating condition for coalescing similar range checks. 1337 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1338 // single covering check that is at least as strong as any of them. 1339 // If the optimization succeeds, the simplified (strengthened) range check 1340 // will always succeed. If it fails, we will deopt, and then give up 1341 // on the optimization. 1342 bool Compile::allow_range_check_smearing() const { 1343 // If this method has already thrown a range-check, 1344 // assume it was because we already tried range smearing 1345 // and it failed. 1346 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1347 return !already_trapped; 1348 } 1349 1350 1351 //------------------------------flatten_alias_type----------------------------- 1352 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1353 int offset = tj->offset(); 1354 TypePtr::PTR ptr = tj->ptr(); 1355 1356 // Known instance (scalarizable allocation) alias only with itself. 1357 bool is_known_inst = tj->isa_oopptr() != NULL && 1358 tj->is_oopptr()->is_known_instance(); 1359 1360 // Process weird unsafe references. 1361 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1362 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1363 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1364 tj = TypeOopPtr::BOTTOM; 1365 ptr = tj->ptr(); 1366 offset = tj->offset(); 1367 } 1368 1369 // Array pointers need some flattening 1370 const TypeAryPtr *ta = tj->isa_aryptr(); 1371 if (ta && ta->is_stable()) { 1372 // Erase stability property for alias analysis. 1373 tj = ta = ta->cast_to_stable(false); 1374 } 1375 if( ta && is_known_inst ) { 1376 if ( offset != Type::OffsetBot && 1377 offset > arrayOopDesc::length_offset_in_bytes() ) { 1378 offset = Type::OffsetBot; // Flatten constant access into array body only 1379 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1380 } 1381 } else if( ta && _AliasLevel >= 2 ) { 1382 // For arrays indexed by constant indices, we flatten the alias 1383 // space to include all of the array body. Only the header, klass 1384 // and array length can be accessed un-aliased. 1385 if( offset != Type::OffsetBot ) { 1386 if( ta->const_oop() ) { // MethodData* or Method* 1387 offset = Type::OffsetBot; // Flatten constant access into array body 1388 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1389 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1390 // range is OK as-is. 1391 tj = ta = TypeAryPtr::RANGE; 1392 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1393 tj = TypeInstPtr::KLASS; // all klass loads look alike 1394 ta = TypeAryPtr::RANGE; // generic ignored junk 1395 ptr = TypePtr::BotPTR; 1396 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1397 tj = TypeInstPtr::MARK; 1398 ta = TypeAryPtr::RANGE; // generic ignored junk 1399 ptr = TypePtr::BotPTR; 1400 } else { // Random constant offset into array body 1401 offset = Type::OffsetBot; // Flatten constant access into array body 1402 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1403 } 1404 } 1405 // Arrays of fixed size alias with arrays of unknown size. 1406 if (ta->size() != TypeInt::POS) { 1407 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1408 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1409 } 1410 // Arrays of known objects become arrays of unknown objects. 1411 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1412 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1413 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1414 } 1415 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1416 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1417 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1418 } 1419 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1420 // cannot be distinguished by bytecode alone. 1421 if (ta->elem() == TypeInt::BOOL) { 1422 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1423 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1424 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1425 } 1426 // During the 2nd round of IterGVN, NotNull castings are removed. 1427 // Make sure the Bottom and NotNull variants alias the same. 1428 // Also, make sure exact and non-exact variants alias the same. 1429 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1430 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1431 } 1432 } 1433 1434 // Oop pointers need some flattening 1435 const TypeInstPtr *to = tj->isa_instptr(); 1436 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1437 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1438 if( ptr == TypePtr::Constant ) { 1439 if (to->klass() != ciEnv::current()->Class_klass() || 1440 offset < k->size_helper() * wordSize) { 1441 // No constant oop pointers (such as Strings); they alias with 1442 // unknown strings. 1443 assert(!is_known_inst, "not scalarizable allocation"); 1444 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1445 } 1446 } else if( is_known_inst ) { 1447 tj = to; // Keep NotNull and klass_is_exact for instance type 1448 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1449 // During the 2nd round of IterGVN, NotNull castings are removed. 1450 // Make sure the Bottom and NotNull variants alias the same. 1451 // Also, make sure exact and non-exact variants alias the same. 1452 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1453 } 1454 if (to->speculative() != NULL) { 1455 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1456 } 1457 // Canonicalize the holder of this field 1458 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1459 // First handle header references such as a LoadKlassNode, even if the 1460 // object's klass is unloaded at compile time (4965979). 1461 if (!is_known_inst) { // Do it only for non-instance types 1462 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1463 } 1464 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1465 // Static fields are in the space above the normal instance 1466 // fields in the java.lang.Class instance. 1467 if (to->klass() != ciEnv::current()->Class_klass()) { 1468 to = NULL; 1469 tj = TypeOopPtr::BOTTOM; 1470 offset = tj->offset(); 1471 } 1472 } else { 1473 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1474 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1475 if( is_known_inst ) { 1476 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1477 } else { 1478 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1479 } 1480 } 1481 } 1482 } 1483 1484 // Klass pointers to object array klasses need some flattening 1485 const TypeKlassPtr *tk = tj->isa_klassptr(); 1486 if( tk ) { 1487 // If we are referencing a field within a Klass, we need 1488 // to assume the worst case of an Object. Both exact and 1489 // inexact types must flatten to the same alias class so 1490 // use NotNull as the PTR. 1491 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1492 1493 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1494 TypeKlassPtr::OBJECT->klass(), 1495 offset); 1496 } 1497 1498 ciKlass* klass = tk->klass(); 1499 if( klass->is_obj_array_klass() ) { 1500 ciKlass* k = TypeAryPtr::OOPS->klass(); 1501 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1502 k = TypeInstPtr::BOTTOM->klass(); 1503 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1504 } 1505 1506 // Check for precise loads from the primary supertype array and force them 1507 // to the supertype cache alias index. Check for generic array loads from 1508 // the primary supertype array and also force them to the supertype cache 1509 // alias index. Since the same load can reach both, we need to merge 1510 // these 2 disparate memories into the same alias class. Since the 1511 // primary supertype array is read-only, there's no chance of confusion 1512 // where we bypass an array load and an array store. 1513 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1514 if (offset == Type::OffsetBot || 1515 (offset >= primary_supers_offset && 1516 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1517 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1518 offset = in_bytes(Klass::secondary_super_cache_offset()); 1519 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1520 } 1521 } 1522 1523 // Flatten all Raw pointers together. 1524 if (tj->base() == Type::RawPtr) 1525 tj = TypeRawPtr::BOTTOM; 1526 1527 if (tj->base() == Type::AnyPtr) 1528 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1529 1530 // Flatten all to bottom for now 1531 switch( _AliasLevel ) { 1532 case 0: 1533 tj = TypePtr::BOTTOM; 1534 break; 1535 case 1: // Flatten to: oop, static, field or array 1536 switch (tj->base()) { 1537 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1538 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1539 case Type::AryPtr: // do not distinguish arrays at all 1540 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1541 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1542 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1543 default: ShouldNotReachHere(); 1544 } 1545 break; 1546 case 2: // No collapsing at level 2; keep all splits 1547 case 3: // No collapsing at level 3; keep all splits 1548 break; 1549 default: 1550 Unimplemented(); 1551 } 1552 1553 offset = tj->offset(); 1554 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1555 1556 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1557 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1558 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1559 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1560 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1561 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1562 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1563 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1564 assert( tj->ptr() != TypePtr::TopPTR && 1565 tj->ptr() != TypePtr::AnyNull && 1566 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1567 // assert( tj->ptr() != TypePtr::Constant || 1568 // tj->base() == Type::RawPtr || 1569 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1570 1571 return tj; 1572 } 1573 1574 void Compile::AliasType::Init(int i, const TypePtr* at) { 1575 _index = i; 1576 _adr_type = at; 1577 _field = NULL; 1578 _element = NULL; 1579 _is_rewritable = true; // default 1580 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1581 if (atoop != NULL && atoop->is_known_instance()) { 1582 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1583 _general_index = Compile::current()->get_alias_index(gt); 1584 } else { 1585 _general_index = 0; 1586 } 1587 } 1588 1589 //---------------------------------print_on------------------------------------ 1590 #ifndef PRODUCT 1591 void Compile::AliasType::print_on(outputStream* st) { 1592 if (index() < 10) 1593 st->print("@ <%d> ", index()); 1594 else st->print("@ <%d>", index()); 1595 st->print(is_rewritable() ? " " : " RO"); 1596 int offset = adr_type()->offset(); 1597 if (offset == Type::OffsetBot) 1598 st->print(" +any"); 1599 else st->print(" +%-3d", offset); 1600 st->print(" in "); 1601 adr_type()->dump_on(st); 1602 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1603 if (field() != NULL && tjp) { 1604 if (tjp->klass() != field()->holder() || 1605 tjp->offset() != field()->offset_in_bytes()) { 1606 st->print(" != "); 1607 field()->print(); 1608 st->print(" ***"); 1609 } 1610 } 1611 } 1612 1613 void print_alias_types() { 1614 Compile* C = Compile::current(); 1615 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1616 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1617 C->alias_type(idx)->print_on(tty); 1618 tty->cr(); 1619 } 1620 } 1621 #endif 1622 1623 1624 //----------------------------probe_alias_cache-------------------------------- 1625 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1626 intptr_t key = (intptr_t) adr_type; 1627 key ^= key >> logAliasCacheSize; 1628 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1629 } 1630 1631 1632 //-----------------------------grow_alias_types-------------------------------- 1633 void Compile::grow_alias_types() { 1634 const int old_ats = _max_alias_types; // how many before? 1635 const int new_ats = old_ats; // how many more? 1636 const int grow_ats = old_ats+new_ats; // how many now? 1637 _max_alias_types = grow_ats; 1638 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1639 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1640 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1641 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1642 } 1643 1644 1645 //--------------------------------find_alias_type------------------------------ 1646 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1647 if (_AliasLevel == 0) 1648 return alias_type(AliasIdxBot); 1649 1650 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1651 if (ace->_adr_type == adr_type) { 1652 return alias_type(ace->_index); 1653 } 1654 1655 // Handle special cases. 1656 if (adr_type == NULL) return alias_type(AliasIdxTop); 1657 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1658 1659 // Do it the slow way. 1660 const TypePtr* flat = flatten_alias_type(adr_type); 1661 1662 #ifdef ASSERT 1663 assert(flat == flatten_alias_type(flat), "idempotent"); 1664 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1665 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1666 const TypeOopPtr* foop = flat->is_oopptr(); 1667 // Scalarizable allocations have exact klass always. 1668 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1669 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1670 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1671 } 1672 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1673 #endif 1674 1675 int idx = AliasIdxTop; 1676 for (int i = 0; i < num_alias_types(); i++) { 1677 if (alias_type(i)->adr_type() == flat) { 1678 idx = i; 1679 break; 1680 } 1681 } 1682 1683 if (idx == AliasIdxTop) { 1684 if (no_create) return NULL; 1685 // Grow the array if necessary. 1686 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1687 // Add a new alias type. 1688 idx = _num_alias_types++; 1689 _alias_types[idx]->Init(idx, flat); 1690 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1691 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1692 if (flat->isa_instptr()) { 1693 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1694 && flat->is_instptr()->klass() == env()->Class_klass()) 1695 alias_type(idx)->set_rewritable(false); 1696 } 1697 if (flat->isa_aryptr()) { 1698 #ifdef ASSERT 1699 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1700 // (T_BYTE has the weakest alignment and size restrictions...) 1701 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1702 #endif 1703 if (flat->offset() == TypePtr::OffsetBot) { 1704 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1705 } 1706 } 1707 if (flat->isa_klassptr()) { 1708 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1709 alias_type(idx)->set_rewritable(false); 1710 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1711 alias_type(idx)->set_rewritable(false); 1712 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1713 alias_type(idx)->set_rewritable(false); 1714 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1715 alias_type(idx)->set_rewritable(false); 1716 } 1717 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1718 // but the base pointer type is not distinctive enough to identify 1719 // references into JavaThread.) 1720 1721 // Check for final fields. 1722 const TypeInstPtr* tinst = flat->isa_instptr(); 1723 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1724 ciField* field; 1725 if (tinst->const_oop() != NULL && 1726 tinst->klass() == ciEnv::current()->Class_klass() && 1727 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1728 // static field 1729 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1730 field = k->get_field_by_offset(tinst->offset(), true); 1731 } else { 1732 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1733 field = k->get_field_by_offset(tinst->offset(), false); 1734 } 1735 assert(field == NULL || 1736 original_field == NULL || 1737 (field->holder() == original_field->holder() && 1738 field->offset() == original_field->offset() && 1739 field->is_static() == original_field->is_static()), "wrong field?"); 1740 // Set field() and is_rewritable() attributes. 1741 if (field != NULL) alias_type(idx)->set_field(field); 1742 } 1743 } 1744 1745 // Fill the cache for next time. 1746 ace->_adr_type = adr_type; 1747 ace->_index = idx; 1748 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1749 1750 // Might as well try to fill the cache for the flattened version, too. 1751 AliasCacheEntry* face = probe_alias_cache(flat); 1752 if (face->_adr_type == NULL) { 1753 face->_adr_type = flat; 1754 face->_index = idx; 1755 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1756 } 1757 1758 return alias_type(idx); 1759 } 1760 1761 1762 Compile::AliasType* Compile::alias_type(ciField* field) { 1763 const TypeOopPtr* t; 1764 if (field->is_static()) 1765 t = TypeInstPtr::make(field->holder()->java_mirror()); 1766 else 1767 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1768 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1769 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1770 return atp; 1771 } 1772 1773 1774 //------------------------------have_alias_type-------------------------------- 1775 bool Compile::have_alias_type(const TypePtr* adr_type) { 1776 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1777 if (ace->_adr_type == adr_type) { 1778 return true; 1779 } 1780 1781 // Handle special cases. 1782 if (adr_type == NULL) return true; 1783 if (adr_type == TypePtr::BOTTOM) return true; 1784 1785 return find_alias_type(adr_type, true, NULL) != NULL; 1786 } 1787 1788 //-----------------------------must_alias-------------------------------------- 1789 // True if all values of the given address type are in the given alias category. 1790 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1791 if (alias_idx == AliasIdxBot) return true; // the universal category 1792 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1793 if (alias_idx == AliasIdxTop) return false; // the empty category 1794 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1795 1796 // the only remaining possible overlap is identity 1797 int adr_idx = get_alias_index(adr_type); 1798 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1799 assert(adr_idx == alias_idx || 1800 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1801 && adr_type != TypeOopPtr::BOTTOM), 1802 "should not be testing for overlap with an unsafe pointer"); 1803 return adr_idx == alias_idx; 1804 } 1805 1806 //------------------------------can_alias-------------------------------------- 1807 // True if any values of the given address type are in the given alias category. 1808 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1809 if (alias_idx == AliasIdxTop) return false; // the empty category 1810 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1811 if (alias_idx == AliasIdxBot) return true; // the universal category 1812 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1813 1814 // the only remaining possible overlap is identity 1815 int adr_idx = get_alias_index(adr_type); 1816 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1817 return adr_idx == alias_idx; 1818 } 1819 1820 1821 1822 //---------------------------pop_warm_call------------------------------------- 1823 WarmCallInfo* Compile::pop_warm_call() { 1824 WarmCallInfo* wci = _warm_calls; 1825 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1826 return wci; 1827 } 1828 1829 //----------------------------Inline_Warm-------------------------------------- 1830 int Compile::Inline_Warm() { 1831 // If there is room, try to inline some more warm call sites. 1832 // %%% Do a graph index compaction pass when we think we're out of space? 1833 if (!InlineWarmCalls) return 0; 1834 1835 int calls_made_hot = 0; 1836 int room_to_grow = NodeCountInliningCutoff - unique(); 1837 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1838 int amount_grown = 0; 1839 WarmCallInfo* call; 1840 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1841 int est_size = (int)call->size(); 1842 if (est_size > (room_to_grow - amount_grown)) { 1843 // This one won't fit anyway. Get rid of it. 1844 call->make_cold(); 1845 continue; 1846 } 1847 call->make_hot(); 1848 calls_made_hot++; 1849 amount_grown += est_size; 1850 amount_to_grow -= est_size; 1851 } 1852 1853 if (calls_made_hot > 0) set_major_progress(); 1854 return calls_made_hot; 1855 } 1856 1857 1858 //----------------------------Finish_Warm-------------------------------------- 1859 void Compile::Finish_Warm() { 1860 if (!InlineWarmCalls) return; 1861 if (failing()) return; 1862 if (warm_calls() == NULL) return; 1863 1864 // Clean up loose ends, if we are out of space for inlining. 1865 WarmCallInfo* call; 1866 while ((call = pop_warm_call()) != NULL) { 1867 call->make_cold(); 1868 } 1869 } 1870 1871 //---------------------cleanup_loop_predicates----------------------- 1872 // Remove the opaque nodes that protect the predicates so that all unused 1873 // checks and uncommon_traps will be eliminated from the ideal graph 1874 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1875 if (predicate_count()==0) return; 1876 for (int i = predicate_count(); i > 0; i--) { 1877 Node * n = predicate_opaque1_node(i-1); 1878 assert(n->Opcode() == Op_Opaque1, "must be"); 1879 igvn.replace_node(n, n->in(1)); 1880 } 1881 assert(predicate_count()==0, "should be clean!"); 1882 } 1883 1884 // StringOpts and late inlining of string methods 1885 void Compile::inline_string_calls(bool parse_time) { 1886 { 1887 // remove useless nodes to make the usage analysis simpler 1888 ResourceMark rm; 1889 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1890 } 1891 1892 { 1893 ResourceMark rm; 1894 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1895 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1896 print_method(PHASE_AFTER_STRINGOPTS, 3); 1897 } 1898 1899 // now inline anything that we skipped the first time around 1900 if (!parse_time) { 1901 _late_inlines_pos = _late_inlines.length(); 1902 } 1903 1904 while (_string_late_inlines.length() > 0) { 1905 CallGenerator* cg = _string_late_inlines.pop(); 1906 cg->do_late_inline(); 1907 if (failing()) return; 1908 } 1909 _string_late_inlines.trunc_to(0); 1910 } 1911 1912 // Late inlining of boxing methods 1913 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1914 if (_boxing_late_inlines.length() > 0) { 1915 assert(has_boxed_value(), "inconsistent"); 1916 1917 PhaseGVN* gvn = initial_gvn(); 1918 set_inlining_incrementally(true); 1919 1920 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1921 for_igvn()->clear(); 1922 gvn->replace_with(&igvn); 1923 1924 _late_inlines_pos = _late_inlines.length(); 1925 1926 while (_boxing_late_inlines.length() > 0) { 1927 CallGenerator* cg = _boxing_late_inlines.pop(); 1928 cg->do_late_inline(); 1929 if (failing()) return; 1930 } 1931 _boxing_late_inlines.trunc_to(0); 1932 1933 { 1934 ResourceMark rm; 1935 PhaseRemoveUseless pru(gvn, for_igvn()); 1936 } 1937 1938 igvn = PhaseIterGVN(gvn); 1939 igvn.optimize(); 1940 1941 set_inlining_progress(false); 1942 set_inlining_incrementally(false); 1943 } 1944 } 1945 1946 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1947 assert(IncrementalInline, "incremental inlining should be on"); 1948 PhaseGVN* gvn = initial_gvn(); 1949 1950 set_inlining_progress(false); 1951 for_igvn()->clear(); 1952 gvn->replace_with(&igvn); 1953 1954 { 1955 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 1956 int i = 0; 1957 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1958 CallGenerator* cg = _late_inlines.at(i); 1959 _late_inlines_pos = i+1; 1960 cg->do_late_inline(); 1961 if (failing()) return; 1962 } 1963 int j = 0; 1964 for (; i < _late_inlines.length(); i++, j++) { 1965 _late_inlines.at_put(j, _late_inlines.at(i)); 1966 } 1967 _late_inlines.trunc_to(j); 1968 } 1969 1970 { 1971 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 1972 ResourceMark rm; 1973 PhaseRemoveUseless pru(gvn, for_igvn()); 1974 } 1975 1976 { 1977 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 1978 igvn = PhaseIterGVN(gvn); 1979 } 1980 } 1981 1982 // Perform incremental inlining until bound on number of live nodes is reached 1983 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1984 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 1985 1986 PhaseGVN* gvn = initial_gvn(); 1987 1988 set_inlining_incrementally(true); 1989 set_inlining_progress(true); 1990 uint low_live_nodes = 0; 1991 1992 while(inlining_progress() && _late_inlines.length() > 0) { 1993 1994 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1995 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1996 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 1997 // PhaseIdealLoop is expensive so we only try it once we are 1998 // out of live nodes and we only try it again if the previous 1999 // helped got the number of nodes down significantly 2000 PhaseIdealLoop ideal_loop( igvn, false, true ); 2001 if (failing()) return; 2002 low_live_nodes = live_nodes(); 2003 _major_progress = true; 2004 } 2005 2006 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2007 break; 2008 } 2009 } 2010 2011 inline_incrementally_one(igvn); 2012 2013 if (failing()) return; 2014 2015 { 2016 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2017 igvn.optimize(); 2018 } 2019 2020 if (failing()) return; 2021 } 2022 2023 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2024 2025 if (_string_late_inlines.length() > 0) { 2026 assert(has_stringbuilder(), "inconsistent"); 2027 for_igvn()->clear(); 2028 initial_gvn()->replace_with(&igvn); 2029 2030 inline_string_calls(false); 2031 2032 if (failing()) return; 2033 2034 { 2035 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2036 ResourceMark rm; 2037 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2038 } 2039 2040 { 2041 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2042 igvn = PhaseIterGVN(gvn); 2043 igvn.optimize(); 2044 } 2045 } 2046 2047 set_inlining_incrementally(false); 2048 } 2049 2050 2051 //------------------------------Optimize--------------------------------------- 2052 // Given a graph, optimize it. 2053 void Compile::Optimize() { 2054 TracePhase tp("optimizer", &timers[_t_optimizer]); 2055 2056 #ifndef PRODUCT 2057 if (env()->break_at_compile()) { 2058 BREAKPOINT; 2059 } 2060 2061 #endif 2062 2063 ResourceMark rm; 2064 int loop_opts_cnt; 2065 2066 print_inlining_reinit(); 2067 2068 NOT_PRODUCT( verify_graph_edges(); ) 2069 2070 print_method(PHASE_AFTER_PARSING); 2071 2072 { 2073 // Iterative Global Value Numbering, including ideal transforms 2074 // Initialize IterGVN with types and values from parse-time GVN 2075 PhaseIterGVN igvn(initial_gvn()); 2076 #ifdef ASSERT 2077 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2078 #endif 2079 { 2080 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2081 igvn.optimize(); 2082 } 2083 2084 print_method(PHASE_ITER_GVN1, 2); 2085 2086 if (failing()) return; 2087 2088 inline_incrementally(igvn); 2089 2090 print_method(PHASE_INCREMENTAL_INLINE, 2); 2091 2092 if (failing()) return; 2093 2094 if (eliminate_boxing()) { 2095 // Inline valueOf() methods now. 2096 inline_boxing_calls(igvn); 2097 2098 if (AlwaysIncrementalInline) { 2099 inline_incrementally(igvn); 2100 } 2101 2102 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2103 2104 if (failing()) return; 2105 } 2106 2107 // Remove the speculative part of types and clean up the graph from 2108 // the extra CastPP nodes whose only purpose is to carry them. Do 2109 // that early so that optimizations are not disrupted by the extra 2110 // CastPP nodes. 2111 remove_speculative_types(igvn); 2112 2113 // No more new expensive nodes will be added to the list from here 2114 // so keep only the actual candidates for optimizations. 2115 cleanup_expensive_nodes(igvn); 2116 2117 // Perform escape analysis 2118 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2119 if (has_loops()) { 2120 // Cleanup graph (remove dead nodes). 2121 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2122 PhaseIdealLoop ideal_loop( igvn, false, true ); 2123 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2124 if (failing()) return; 2125 } 2126 ConnectionGraph::do_analysis(this, &igvn); 2127 2128 if (failing()) return; 2129 2130 // Optimize out fields loads from scalar replaceable allocations. 2131 igvn.optimize(); 2132 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2133 2134 if (failing()) return; 2135 2136 if (congraph() != NULL && macro_count() > 0) { 2137 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2138 PhaseMacroExpand mexp(igvn); 2139 mexp.eliminate_macro_nodes(); 2140 igvn.set_delay_transform(false); 2141 2142 igvn.optimize(); 2143 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2144 2145 if (failing()) return; 2146 } 2147 } 2148 2149 // Loop transforms on the ideal graph. Range Check Elimination, 2150 // peeling, unrolling, etc. 2151 2152 // Set loop opts counter 2153 loop_opts_cnt = num_loop_opts(); 2154 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2155 { 2156 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2157 PhaseIdealLoop ideal_loop( igvn, true ); 2158 loop_opts_cnt--; 2159 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2160 if (failing()) return; 2161 } 2162 // Loop opts pass if partial peeling occurred in previous pass 2163 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2164 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2165 PhaseIdealLoop ideal_loop( igvn, false ); 2166 loop_opts_cnt--; 2167 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2168 if (failing()) return; 2169 } 2170 // Loop opts pass for loop-unrolling before CCP 2171 if(major_progress() && (loop_opts_cnt > 0)) { 2172 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2173 PhaseIdealLoop ideal_loop( igvn, false ); 2174 loop_opts_cnt--; 2175 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2176 } 2177 if (!failing()) { 2178 // Verify that last round of loop opts produced a valid graph 2179 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2180 PhaseIdealLoop::verify(igvn); 2181 } 2182 } 2183 if (failing()) return; 2184 2185 // Conditional Constant Propagation; 2186 PhaseCCP ccp( &igvn ); 2187 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2188 { 2189 TracePhase tp("ccp", &timers[_t_ccp]); 2190 ccp.do_transform(); 2191 } 2192 print_method(PHASE_CPP1, 2); 2193 2194 assert( true, "Break here to ccp.dump_old2new_map()"); 2195 2196 // Iterative Global Value Numbering, including ideal transforms 2197 { 2198 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2199 igvn = ccp; 2200 igvn.optimize(); 2201 } 2202 2203 print_method(PHASE_ITER_GVN2, 2); 2204 2205 if (failing()) return; 2206 2207 // Loop transforms on the ideal graph. Range Check Elimination, 2208 // peeling, unrolling, etc. 2209 if(loop_opts_cnt > 0) { 2210 debug_only( int cnt = 0; ); 2211 while(major_progress() && (loop_opts_cnt > 0)) { 2212 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2213 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2214 PhaseIdealLoop ideal_loop( igvn, true); 2215 loop_opts_cnt--; 2216 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2217 if (failing()) return; 2218 } 2219 } 2220 2221 { 2222 // Verify that all previous optimizations produced a valid graph 2223 // at least to this point, even if no loop optimizations were done. 2224 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2225 PhaseIdealLoop::verify(igvn); 2226 } 2227 2228 { 2229 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2230 PhaseMacroExpand mex(igvn); 2231 if (mex.expand_macro_nodes()) { 2232 assert(failing(), "must bail out w/ explicit message"); 2233 return; 2234 } 2235 } 2236 2237 DEBUG_ONLY( _modified_nodes = NULL; ) 2238 } // (End scope of igvn; run destructor if necessary for asserts.) 2239 2240 process_print_inlining(); 2241 // A method with only infinite loops has no edges entering loops from root 2242 { 2243 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2244 if (final_graph_reshaping()) { 2245 assert(failing(), "must bail out w/ explicit message"); 2246 return; 2247 } 2248 } 2249 2250 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2251 } 2252 2253 2254 //------------------------------Code_Gen--------------------------------------- 2255 // Given a graph, generate code for it 2256 void Compile::Code_Gen() { 2257 if (failing()) { 2258 return; 2259 } 2260 2261 // Perform instruction selection. You might think we could reclaim Matcher 2262 // memory PDQ, but actually the Matcher is used in generating spill code. 2263 // Internals of the Matcher (including some VectorSets) must remain live 2264 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2265 // set a bit in reclaimed memory. 2266 2267 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2268 // nodes. Mapping is only valid at the root of each matched subtree. 2269 NOT_PRODUCT( verify_graph_edges(); ) 2270 2271 Matcher matcher; 2272 _matcher = &matcher; 2273 { 2274 TracePhase tp("matcher", &timers[_t_matcher]); 2275 matcher.match(); 2276 } 2277 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2278 // nodes. Mapping is only valid at the root of each matched subtree. 2279 NOT_PRODUCT( verify_graph_edges(); ) 2280 2281 // If you have too many nodes, or if matching has failed, bail out 2282 check_node_count(0, "out of nodes matching instructions"); 2283 if (failing()) { 2284 return; 2285 } 2286 2287 // Build a proper-looking CFG 2288 PhaseCFG cfg(node_arena(), root(), matcher); 2289 _cfg = &cfg; 2290 { 2291 TracePhase tp("scheduler", &timers[_t_scheduler]); 2292 bool success = cfg.do_global_code_motion(); 2293 if (!success) { 2294 return; 2295 } 2296 2297 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2298 NOT_PRODUCT( verify_graph_edges(); ) 2299 debug_only( cfg.verify(); ) 2300 } 2301 2302 PhaseChaitin regalloc(unique(), cfg, matcher); 2303 _regalloc = ®alloc; 2304 { 2305 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2306 // Perform register allocation. After Chaitin, use-def chains are 2307 // no longer accurate (at spill code) and so must be ignored. 2308 // Node->LRG->reg mappings are still accurate. 2309 _regalloc->Register_Allocate(); 2310 2311 // Bail out if the allocator builds too many nodes 2312 if (failing()) { 2313 return; 2314 } 2315 } 2316 2317 // Prior to register allocation we kept empty basic blocks in case the 2318 // the allocator needed a place to spill. After register allocation we 2319 // are not adding any new instructions. If any basic block is empty, we 2320 // can now safely remove it. 2321 { 2322 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2323 cfg.remove_empty_blocks(); 2324 if (do_freq_based_layout()) { 2325 PhaseBlockLayout layout(cfg); 2326 } else { 2327 cfg.set_loop_alignment(); 2328 } 2329 cfg.fixup_flow(); 2330 } 2331 2332 // Apply peephole optimizations 2333 if( OptoPeephole ) { 2334 TracePhase tp("peephole", &timers[_t_peephole]); 2335 PhasePeephole peep( _regalloc, cfg); 2336 peep.do_transform(); 2337 } 2338 2339 // Do late expand if CPU requires this. 2340 if (Matcher::require_postalloc_expand) { 2341 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2342 cfg.postalloc_expand(_regalloc); 2343 } 2344 2345 // Convert Nodes to instruction bits in a buffer 2346 { 2347 TraceTime tp("output", &timers[_t_output], CITime); 2348 Output(); 2349 } 2350 2351 print_method(PHASE_FINAL_CODE); 2352 2353 // He's dead, Jim. 2354 _cfg = (PhaseCFG*)0xdeadbeef; 2355 _regalloc = (PhaseChaitin*)0xdeadbeef; 2356 } 2357 2358 2359 //------------------------------dump_asm--------------------------------------- 2360 // Dump formatted assembly 2361 #ifndef PRODUCT 2362 void Compile::dump_asm(int *pcs, uint pc_limit) { 2363 bool cut_short = false; 2364 tty->print_cr("#"); 2365 tty->print("# "); _tf->dump(); tty->cr(); 2366 tty->print_cr("#"); 2367 2368 // For all blocks 2369 int pc = 0x0; // Program counter 2370 char starts_bundle = ' '; 2371 _regalloc->dump_frame(); 2372 2373 Node *n = NULL; 2374 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2375 if (VMThread::should_terminate()) { 2376 cut_short = true; 2377 break; 2378 } 2379 Block* block = _cfg->get_block(i); 2380 if (block->is_connector() && !Verbose) { 2381 continue; 2382 } 2383 n = block->head(); 2384 if (pcs && n->_idx < pc_limit) { 2385 tty->print("%3.3x ", pcs[n->_idx]); 2386 } else { 2387 tty->print(" "); 2388 } 2389 block->dump_head(_cfg); 2390 if (block->is_connector()) { 2391 tty->print_cr(" # Empty connector block"); 2392 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2393 tty->print_cr(" # Block is sole successor of call"); 2394 } 2395 2396 // For all instructions 2397 Node *delay = NULL; 2398 for (uint j = 0; j < block->number_of_nodes(); j++) { 2399 if (VMThread::should_terminate()) { 2400 cut_short = true; 2401 break; 2402 } 2403 n = block->get_node(j); 2404 if (valid_bundle_info(n)) { 2405 Bundle* bundle = node_bundling(n); 2406 if (bundle->used_in_unconditional_delay()) { 2407 delay = n; 2408 continue; 2409 } 2410 if (bundle->starts_bundle()) { 2411 starts_bundle = '+'; 2412 } 2413 } 2414 2415 if (WizardMode) { 2416 n->dump(); 2417 } 2418 2419 if( !n->is_Region() && // Dont print in the Assembly 2420 !n->is_Phi() && // a few noisely useless nodes 2421 !n->is_Proj() && 2422 !n->is_MachTemp() && 2423 !n->is_SafePointScalarObject() && 2424 !n->is_Catch() && // Would be nice to print exception table targets 2425 !n->is_MergeMem() && // Not very interesting 2426 !n->is_top() && // Debug info table constants 2427 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2428 ) { 2429 if (pcs && n->_idx < pc_limit) 2430 tty->print("%3.3x", pcs[n->_idx]); 2431 else 2432 tty->print(" "); 2433 tty->print(" %c ", starts_bundle); 2434 starts_bundle = ' '; 2435 tty->print("\t"); 2436 n->format(_regalloc, tty); 2437 tty->cr(); 2438 } 2439 2440 // If we have an instruction with a delay slot, and have seen a delay, 2441 // then back up and print it 2442 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2443 assert(delay != NULL, "no unconditional delay instruction"); 2444 if (WizardMode) delay->dump(); 2445 2446 if (node_bundling(delay)->starts_bundle()) 2447 starts_bundle = '+'; 2448 if (pcs && n->_idx < pc_limit) 2449 tty->print("%3.3x", pcs[n->_idx]); 2450 else 2451 tty->print(" "); 2452 tty->print(" %c ", starts_bundle); 2453 starts_bundle = ' '; 2454 tty->print("\t"); 2455 delay->format(_regalloc, tty); 2456 tty->cr(); 2457 delay = NULL; 2458 } 2459 2460 // Dump the exception table as well 2461 if( n->is_Catch() && (Verbose || WizardMode) ) { 2462 // Print the exception table for this offset 2463 _handler_table.print_subtable_for(pc); 2464 } 2465 } 2466 2467 if (pcs && n->_idx < pc_limit) 2468 tty->print_cr("%3.3x", pcs[n->_idx]); 2469 else 2470 tty->cr(); 2471 2472 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2473 2474 } // End of per-block dump 2475 tty->cr(); 2476 2477 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2478 } 2479 #endif 2480 2481 //------------------------------Final_Reshape_Counts--------------------------- 2482 // This class defines counters to help identify when a method 2483 // may/must be executed using hardware with only 24-bit precision. 2484 struct Final_Reshape_Counts : public StackObj { 2485 int _call_count; // count non-inlined 'common' calls 2486 int _float_count; // count float ops requiring 24-bit precision 2487 int _double_count; // count double ops requiring more precision 2488 int _java_call_count; // count non-inlined 'java' calls 2489 int _inner_loop_count; // count loops which need alignment 2490 VectorSet _visited; // Visitation flags 2491 Node_List _tests; // Set of IfNodes & PCTableNodes 2492 2493 Final_Reshape_Counts() : 2494 _call_count(0), _float_count(0), _double_count(0), 2495 _java_call_count(0), _inner_loop_count(0), 2496 _visited( Thread::current()->resource_area() ) { } 2497 2498 void inc_call_count () { _call_count ++; } 2499 void inc_float_count () { _float_count ++; } 2500 void inc_double_count() { _double_count++; } 2501 void inc_java_call_count() { _java_call_count++; } 2502 void inc_inner_loop_count() { _inner_loop_count++; } 2503 2504 int get_call_count () const { return _call_count ; } 2505 int get_float_count () const { return _float_count ; } 2506 int get_double_count() const { return _double_count; } 2507 int get_java_call_count() const { return _java_call_count; } 2508 int get_inner_loop_count() const { return _inner_loop_count; } 2509 }; 2510 2511 #ifdef ASSERT 2512 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2513 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2514 // Make sure the offset goes inside the instance layout. 2515 return k->contains_field_offset(tp->offset()); 2516 // Note that OffsetBot and OffsetTop are very negative. 2517 } 2518 #endif 2519 2520 // Eliminate trivially redundant StoreCMs and accumulate their 2521 // precedence edges. 2522 void Compile::eliminate_redundant_card_marks(Node* n) { 2523 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2524 if (n->in(MemNode::Address)->outcnt() > 1) { 2525 // There are multiple users of the same address so it might be 2526 // possible to eliminate some of the StoreCMs 2527 Node* mem = n->in(MemNode::Memory); 2528 Node* adr = n->in(MemNode::Address); 2529 Node* val = n->in(MemNode::ValueIn); 2530 Node* prev = n; 2531 bool done = false; 2532 // Walk the chain of StoreCMs eliminating ones that match. As 2533 // long as it's a chain of single users then the optimization is 2534 // safe. Eliminating partially redundant StoreCMs would require 2535 // cloning copies down the other paths. 2536 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2537 if (adr == mem->in(MemNode::Address) && 2538 val == mem->in(MemNode::ValueIn)) { 2539 // redundant StoreCM 2540 if (mem->req() > MemNode::OopStore) { 2541 // Hasn't been processed by this code yet. 2542 n->add_prec(mem->in(MemNode::OopStore)); 2543 } else { 2544 // Already converted to precedence edge 2545 for (uint i = mem->req(); i < mem->len(); i++) { 2546 // Accumulate any precedence edges 2547 if (mem->in(i) != NULL) { 2548 n->add_prec(mem->in(i)); 2549 } 2550 } 2551 // Everything above this point has been processed. 2552 done = true; 2553 } 2554 // Eliminate the previous StoreCM 2555 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2556 assert(mem->outcnt() == 0, "should be dead"); 2557 mem->disconnect_inputs(NULL, this); 2558 } else { 2559 prev = mem; 2560 } 2561 mem = prev->in(MemNode::Memory); 2562 } 2563 } 2564 } 2565 2566 //------------------------------final_graph_reshaping_impl---------------------- 2567 // Implement items 1-5 from final_graph_reshaping below. 2568 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2569 2570 if ( n->outcnt() == 0 ) return; // dead node 2571 uint nop = n->Opcode(); 2572 2573 // Check for 2-input instruction with "last use" on right input. 2574 // Swap to left input. Implements item (2). 2575 if( n->req() == 3 && // two-input instruction 2576 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2577 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2578 n->in(2)->outcnt() == 1 &&// right use IS a last use 2579 !n->in(2)->is_Con() ) { // right use is not a constant 2580 // Check for commutative opcode 2581 switch( nop ) { 2582 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2583 case Op_MaxI: case Op_MinI: 2584 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2585 case Op_AndL: case Op_XorL: case Op_OrL: 2586 case Op_AndI: case Op_XorI: case Op_OrI: { 2587 // Move "last use" input to left by swapping inputs 2588 n->swap_edges(1, 2); 2589 break; 2590 } 2591 default: 2592 break; 2593 } 2594 } 2595 2596 #ifdef ASSERT 2597 if( n->is_Mem() ) { 2598 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2599 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2600 // oop will be recorded in oop map if load crosses safepoint 2601 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2602 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2603 "raw memory operations should have control edge"); 2604 } 2605 #endif 2606 // Count FPU ops and common calls, implements item (3) 2607 switch( nop ) { 2608 // Count all float operations that may use FPU 2609 case Op_AddF: 2610 case Op_SubF: 2611 case Op_MulF: 2612 case Op_DivF: 2613 case Op_NegF: 2614 case Op_ModF: 2615 case Op_ConvI2F: 2616 case Op_ConF: 2617 case Op_CmpF: 2618 case Op_CmpF3: 2619 // case Op_ConvL2F: // longs are split into 32-bit halves 2620 frc.inc_float_count(); 2621 break; 2622 2623 case Op_ConvF2D: 2624 case Op_ConvD2F: 2625 frc.inc_float_count(); 2626 frc.inc_double_count(); 2627 break; 2628 2629 // Count all double operations that may use FPU 2630 case Op_AddD: 2631 case Op_SubD: 2632 case Op_MulD: 2633 case Op_DivD: 2634 case Op_NegD: 2635 case Op_ModD: 2636 case Op_ConvI2D: 2637 case Op_ConvD2I: 2638 // case Op_ConvL2D: // handled by leaf call 2639 // case Op_ConvD2L: // handled by leaf call 2640 case Op_ConD: 2641 case Op_CmpD: 2642 case Op_CmpD3: 2643 frc.inc_double_count(); 2644 break; 2645 case Op_Opaque1: // Remove Opaque Nodes before matching 2646 case Op_Opaque2: // Remove Opaque Nodes before matching 2647 case Op_Opaque3: 2648 n->subsume_by(n->in(1), this); 2649 break; 2650 case Op_CallStaticJava: 2651 case Op_CallJava: 2652 case Op_CallDynamicJava: 2653 frc.inc_java_call_count(); // Count java call site; 2654 case Op_CallRuntime: 2655 case Op_CallLeaf: 2656 case Op_CallLeafNoFP: { 2657 assert( n->is_Call(), "" ); 2658 CallNode *call = n->as_Call(); 2659 // Count call sites where the FP mode bit would have to be flipped. 2660 // Do not count uncommon runtime calls: 2661 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2662 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2663 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2664 frc.inc_call_count(); // Count the call site 2665 } else { // See if uncommon argument is shared 2666 Node *n = call->in(TypeFunc::Parms); 2667 int nop = n->Opcode(); 2668 // Clone shared simple arguments to uncommon calls, item (1). 2669 if( n->outcnt() > 1 && 2670 !n->is_Proj() && 2671 nop != Op_CreateEx && 2672 nop != Op_CheckCastPP && 2673 nop != Op_DecodeN && 2674 nop != Op_DecodeNKlass && 2675 !n->is_Mem() ) { 2676 Node *x = n->clone(); 2677 call->set_req( TypeFunc::Parms, x ); 2678 } 2679 } 2680 break; 2681 } 2682 2683 case Op_StoreD: 2684 case Op_LoadD: 2685 case Op_LoadD_unaligned: 2686 frc.inc_double_count(); 2687 goto handle_mem; 2688 case Op_StoreF: 2689 case Op_LoadF: 2690 frc.inc_float_count(); 2691 goto handle_mem; 2692 2693 case Op_StoreCM: 2694 { 2695 // Convert OopStore dependence into precedence edge 2696 Node* prec = n->in(MemNode::OopStore); 2697 n->del_req(MemNode::OopStore); 2698 n->add_prec(prec); 2699 eliminate_redundant_card_marks(n); 2700 } 2701 2702 // fall through 2703 2704 case Op_StoreB: 2705 case Op_StoreC: 2706 case Op_StorePConditional: 2707 case Op_StoreI: 2708 case Op_StoreL: 2709 case Op_StoreIConditional: 2710 case Op_StoreLConditional: 2711 case Op_CompareAndSwapI: 2712 case Op_CompareAndSwapL: 2713 case Op_CompareAndSwapP: 2714 case Op_CompareAndSwapN: 2715 case Op_GetAndAddI: 2716 case Op_GetAndAddL: 2717 case Op_GetAndSetI: 2718 case Op_GetAndSetL: 2719 case Op_GetAndSetP: 2720 case Op_GetAndSetN: 2721 case Op_StoreP: 2722 case Op_StoreN: 2723 case Op_StoreNKlass: 2724 case Op_LoadB: 2725 case Op_LoadUB: 2726 case Op_LoadUS: 2727 case Op_LoadI: 2728 case Op_LoadKlass: 2729 case Op_LoadNKlass: 2730 case Op_LoadL: 2731 case Op_LoadL_unaligned: 2732 case Op_LoadPLocked: 2733 case Op_LoadP: 2734 case Op_LoadN: 2735 case Op_LoadRange: 2736 case Op_LoadS: { 2737 handle_mem: 2738 #ifdef ASSERT 2739 if( VerifyOptoOopOffsets ) { 2740 assert( n->is_Mem(), "" ); 2741 MemNode *mem = (MemNode*)n; 2742 // Check to see if address types have grounded out somehow. 2743 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2744 assert( !tp || oop_offset_is_sane(tp), "" ); 2745 } 2746 #endif 2747 break; 2748 } 2749 2750 case Op_AddP: { // Assert sane base pointers 2751 Node *addp = n->in(AddPNode::Address); 2752 assert( !addp->is_AddP() || 2753 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2754 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2755 "Base pointers must match" ); 2756 #ifdef _LP64 2757 if ((UseCompressedOops || UseCompressedClassPointers) && 2758 addp->Opcode() == Op_ConP && 2759 addp == n->in(AddPNode::Base) && 2760 n->in(AddPNode::Offset)->is_Con()) { 2761 // Use addressing with narrow klass to load with offset on x86. 2762 // On sparc loading 32-bits constant and decoding it have less 2763 // instructions (4) then load 64-bits constant (7). 2764 // Do this transformation here since IGVN will convert ConN back to ConP. 2765 const Type* t = addp->bottom_type(); 2766 if (t->isa_oopptr() || t->isa_klassptr()) { 2767 Node* nn = NULL; 2768 2769 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2770 2771 // Look for existing ConN node of the same exact type. 2772 Node* r = root(); 2773 uint cnt = r->outcnt(); 2774 for (uint i = 0; i < cnt; i++) { 2775 Node* m = r->raw_out(i); 2776 if (m!= NULL && m->Opcode() == op && 2777 m->bottom_type()->make_ptr() == t) { 2778 nn = m; 2779 break; 2780 } 2781 } 2782 if (nn != NULL) { 2783 // Decode a narrow oop to match address 2784 // [R12 + narrow_oop_reg<<3 + offset] 2785 if (t->isa_oopptr()) { 2786 nn = new DecodeNNode(nn, t); 2787 } else { 2788 nn = new DecodeNKlassNode(nn, t); 2789 } 2790 n->set_req(AddPNode::Base, nn); 2791 n->set_req(AddPNode::Address, nn); 2792 if (addp->outcnt() == 0) { 2793 addp->disconnect_inputs(NULL, this); 2794 } 2795 } 2796 } 2797 } 2798 #endif 2799 break; 2800 } 2801 2802 #ifdef _LP64 2803 case Op_CastPP: 2804 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2805 Node* in1 = n->in(1); 2806 const Type* t = n->bottom_type(); 2807 Node* new_in1 = in1->clone(); 2808 new_in1->as_DecodeN()->set_type(t); 2809 2810 if (!Matcher::narrow_oop_use_complex_address()) { 2811 // 2812 // x86, ARM and friends can handle 2 adds in addressing mode 2813 // and Matcher can fold a DecodeN node into address by using 2814 // a narrow oop directly and do implicit NULL check in address: 2815 // 2816 // [R12 + narrow_oop_reg<<3 + offset] 2817 // NullCheck narrow_oop_reg 2818 // 2819 // On other platforms (Sparc) we have to keep new DecodeN node and 2820 // use it to do implicit NULL check in address: 2821 // 2822 // decode_not_null narrow_oop_reg, base_reg 2823 // [base_reg + offset] 2824 // NullCheck base_reg 2825 // 2826 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2827 // to keep the information to which NULL check the new DecodeN node 2828 // corresponds to use it as value in implicit_null_check(). 2829 // 2830 new_in1->set_req(0, n->in(0)); 2831 } 2832 2833 n->subsume_by(new_in1, this); 2834 if (in1->outcnt() == 0) { 2835 in1->disconnect_inputs(NULL, this); 2836 } 2837 } 2838 break; 2839 2840 case Op_CmpP: 2841 // Do this transformation here to preserve CmpPNode::sub() and 2842 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2843 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2844 Node* in1 = n->in(1); 2845 Node* in2 = n->in(2); 2846 if (!in1->is_DecodeNarrowPtr()) { 2847 in2 = in1; 2848 in1 = n->in(2); 2849 } 2850 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2851 2852 Node* new_in2 = NULL; 2853 if (in2->is_DecodeNarrowPtr()) { 2854 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2855 new_in2 = in2->in(1); 2856 } else if (in2->Opcode() == Op_ConP) { 2857 const Type* t = in2->bottom_type(); 2858 if (t == TypePtr::NULL_PTR) { 2859 assert(in1->is_DecodeN(), "compare klass to null?"); 2860 // Don't convert CmpP null check into CmpN if compressed 2861 // oops implicit null check is not generated. 2862 // This will allow to generate normal oop implicit null check. 2863 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2864 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2865 // 2866 // This transformation together with CastPP transformation above 2867 // will generated code for implicit NULL checks for compressed oops. 2868 // 2869 // The original code after Optimize() 2870 // 2871 // LoadN memory, narrow_oop_reg 2872 // decode narrow_oop_reg, base_reg 2873 // CmpP base_reg, NULL 2874 // CastPP base_reg // NotNull 2875 // Load [base_reg + offset], val_reg 2876 // 2877 // after these transformations will be 2878 // 2879 // LoadN memory, narrow_oop_reg 2880 // CmpN narrow_oop_reg, NULL 2881 // decode_not_null narrow_oop_reg, base_reg 2882 // Load [base_reg + offset], val_reg 2883 // 2884 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2885 // since narrow oops can be used in debug info now (see the code in 2886 // final_graph_reshaping_walk()). 2887 // 2888 // At the end the code will be matched to 2889 // on x86: 2890 // 2891 // Load_narrow_oop memory, narrow_oop_reg 2892 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2893 // NullCheck narrow_oop_reg 2894 // 2895 // and on sparc: 2896 // 2897 // Load_narrow_oop memory, narrow_oop_reg 2898 // decode_not_null narrow_oop_reg, base_reg 2899 // Load [base_reg + offset], val_reg 2900 // NullCheck base_reg 2901 // 2902 } else if (t->isa_oopptr()) { 2903 new_in2 = ConNode::make(t->make_narrowoop()); 2904 } else if (t->isa_klassptr()) { 2905 new_in2 = ConNode::make(t->make_narrowklass()); 2906 } 2907 } 2908 if (new_in2 != NULL) { 2909 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2910 n->subsume_by(cmpN, this); 2911 if (in1->outcnt() == 0) { 2912 in1->disconnect_inputs(NULL, this); 2913 } 2914 if (in2->outcnt() == 0) { 2915 in2->disconnect_inputs(NULL, this); 2916 } 2917 } 2918 } 2919 break; 2920 2921 case Op_DecodeN: 2922 case Op_DecodeNKlass: 2923 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2924 // DecodeN could be pinned when it can't be fold into 2925 // an address expression, see the code for Op_CastPP above. 2926 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2927 break; 2928 2929 case Op_EncodeP: 2930 case Op_EncodePKlass: { 2931 Node* in1 = n->in(1); 2932 if (in1->is_DecodeNarrowPtr()) { 2933 n->subsume_by(in1->in(1), this); 2934 } else if (in1->Opcode() == Op_ConP) { 2935 const Type* t = in1->bottom_type(); 2936 if (t == TypePtr::NULL_PTR) { 2937 assert(t->isa_oopptr(), "null klass?"); 2938 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 2939 } else if (t->isa_oopptr()) { 2940 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 2941 } else if (t->isa_klassptr()) { 2942 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 2943 } 2944 } 2945 if (in1->outcnt() == 0) { 2946 in1->disconnect_inputs(NULL, this); 2947 } 2948 break; 2949 } 2950 2951 case Op_Proj: { 2952 if (OptimizeStringConcat) { 2953 ProjNode* p = n->as_Proj(); 2954 if (p->_is_io_use) { 2955 // Separate projections were used for the exception path which 2956 // are normally removed by a late inline. If it wasn't inlined 2957 // then they will hang around and should just be replaced with 2958 // the original one. 2959 Node* proj = NULL; 2960 // Replace with just one 2961 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2962 Node *use = i.get(); 2963 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2964 proj = use; 2965 break; 2966 } 2967 } 2968 assert(proj != NULL, "must be found"); 2969 p->subsume_by(proj, this); 2970 } 2971 } 2972 break; 2973 } 2974 2975 case Op_Phi: 2976 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2977 // The EncodeP optimization may create Phi with the same edges 2978 // for all paths. It is not handled well by Register Allocator. 2979 Node* unique_in = n->in(1); 2980 assert(unique_in != NULL, ""); 2981 uint cnt = n->req(); 2982 for (uint i = 2; i < cnt; i++) { 2983 Node* m = n->in(i); 2984 assert(m != NULL, ""); 2985 if (unique_in != m) 2986 unique_in = NULL; 2987 } 2988 if (unique_in != NULL) { 2989 n->subsume_by(unique_in, this); 2990 } 2991 } 2992 break; 2993 2994 #endif 2995 2996 case Op_ModI: 2997 if (UseDivMod) { 2998 // Check if a%b and a/b both exist 2999 Node* d = n->find_similar(Op_DivI); 3000 if (d) { 3001 // Replace them with a fused divmod if supported 3002 if (Matcher::has_match_rule(Op_DivModI)) { 3003 DivModINode* divmod = DivModINode::make(n); 3004 d->subsume_by(divmod->div_proj(), this); 3005 n->subsume_by(divmod->mod_proj(), this); 3006 } else { 3007 // replace a%b with a-((a/b)*b) 3008 Node* mult = new MulINode(d, d->in(2)); 3009 Node* sub = new SubINode(d->in(1), mult); 3010 n->subsume_by(sub, this); 3011 } 3012 } 3013 } 3014 break; 3015 3016 case Op_ModL: 3017 if (UseDivMod) { 3018 // Check if a%b and a/b both exist 3019 Node* d = n->find_similar(Op_DivL); 3020 if (d) { 3021 // Replace them with a fused divmod if supported 3022 if (Matcher::has_match_rule(Op_DivModL)) { 3023 DivModLNode* divmod = DivModLNode::make(n); 3024 d->subsume_by(divmod->div_proj(), this); 3025 n->subsume_by(divmod->mod_proj(), this); 3026 } else { 3027 // replace a%b with a-((a/b)*b) 3028 Node* mult = new MulLNode(d, d->in(2)); 3029 Node* sub = new SubLNode(d->in(1), mult); 3030 n->subsume_by(sub, this); 3031 } 3032 } 3033 } 3034 break; 3035 3036 case Op_LoadVector: 3037 case Op_StoreVector: 3038 break; 3039 3040 case Op_PackB: 3041 case Op_PackS: 3042 case Op_PackI: 3043 case Op_PackF: 3044 case Op_PackL: 3045 case Op_PackD: 3046 if (n->req()-1 > 2) { 3047 // Replace many operand PackNodes with a binary tree for matching 3048 PackNode* p = (PackNode*) n; 3049 Node* btp = p->binary_tree_pack(1, n->req()); 3050 n->subsume_by(btp, this); 3051 } 3052 break; 3053 case Op_Loop: 3054 case Op_CountedLoop: 3055 if (n->as_Loop()->is_inner_loop()) { 3056 frc.inc_inner_loop_count(); 3057 } 3058 break; 3059 case Op_LShiftI: 3060 case Op_RShiftI: 3061 case Op_URShiftI: 3062 case Op_LShiftL: 3063 case Op_RShiftL: 3064 case Op_URShiftL: 3065 if (Matcher::need_masked_shift_count) { 3066 // The cpu's shift instructions don't restrict the count to the 3067 // lower 5/6 bits. We need to do the masking ourselves. 3068 Node* in2 = n->in(2); 3069 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3070 const TypeInt* t = in2->find_int_type(); 3071 if (t != NULL && t->is_con()) { 3072 juint shift = t->get_con(); 3073 if (shift > mask) { // Unsigned cmp 3074 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3075 } 3076 } else { 3077 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3078 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3079 n->set_req(2, shift); 3080 } 3081 } 3082 if (in2->outcnt() == 0) { // Remove dead node 3083 in2->disconnect_inputs(NULL, this); 3084 } 3085 } 3086 break; 3087 case Op_MemBarStoreStore: 3088 case Op_MemBarRelease: 3089 // Break the link with AllocateNode: it is no longer useful and 3090 // confuses register allocation. 3091 if (n->req() > MemBarNode::Precedent) { 3092 n->set_req(MemBarNode::Precedent, top()); 3093 } 3094 break; 3095 default: 3096 assert( !n->is_Call(), "" ); 3097 assert( !n->is_Mem(), "" ); 3098 break; 3099 } 3100 3101 // Collect CFG split points 3102 if (n->is_MultiBranch()) 3103 frc._tests.push(n); 3104 } 3105 3106 //------------------------------final_graph_reshaping_walk--------------------- 3107 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3108 // requires that the walk visits a node's inputs before visiting the node. 3109 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3110 ResourceArea *area = Thread::current()->resource_area(); 3111 Unique_Node_List sfpt(area); 3112 3113 frc._visited.set(root->_idx); // first, mark node as visited 3114 uint cnt = root->req(); 3115 Node *n = root; 3116 uint i = 0; 3117 while (true) { 3118 if (i < cnt) { 3119 // Place all non-visited non-null inputs onto stack 3120 Node* m = n->in(i); 3121 ++i; 3122 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3123 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3124 // compute worst case interpreter size in case of a deoptimization 3125 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3126 3127 sfpt.push(m); 3128 } 3129 cnt = m->req(); 3130 nstack.push(n, i); // put on stack parent and next input's index 3131 n = m; 3132 i = 0; 3133 } 3134 } else { 3135 // Now do post-visit work 3136 final_graph_reshaping_impl( n, frc ); 3137 if (nstack.is_empty()) 3138 break; // finished 3139 n = nstack.node(); // Get node from stack 3140 cnt = n->req(); 3141 i = nstack.index(); 3142 nstack.pop(); // Shift to the next node on stack 3143 } 3144 } 3145 3146 // Skip next transformation if compressed oops are not used. 3147 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3148 (!UseCompressedOops && !UseCompressedClassPointers)) 3149 return; 3150 3151 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3152 // It could be done for an uncommon traps or any safepoints/calls 3153 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3154 while (sfpt.size() > 0) { 3155 n = sfpt.pop(); 3156 JVMState *jvms = n->as_SafePoint()->jvms(); 3157 assert(jvms != NULL, "sanity"); 3158 int start = jvms->debug_start(); 3159 int end = n->req(); 3160 bool is_uncommon = (n->is_CallStaticJava() && 3161 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3162 for (int j = start; j < end; j++) { 3163 Node* in = n->in(j); 3164 if (in->is_DecodeNarrowPtr()) { 3165 bool safe_to_skip = true; 3166 if (!is_uncommon ) { 3167 // Is it safe to skip? 3168 for (uint i = 0; i < in->outcnt(); i++) { 3169 Node* u = in->raw_out(i); 3170 if (!u->is_SafePoint() || 3171 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3172 safe_to_skip = false; 3173 } 3174 } 3175 } 3176 if (safe_to_skip) { 3177 n->set_req(j, in->in(1)); 3178 } 3179 if (in->outcnt() == 0) { 3180 in->disconnect_inputs(NULL, this); 3181 } 3182 } 3183 } 3184 } 3185 } 3186 3187 //------------------------------final_graph_reshaping-------------------------- 3188 // Final Graph Reshaping. 3189 // 3190 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3191 // and not commoned up and forced early. Must come after regular 3192 // optimizations to avoid GVN undoing the cloning. Clone constant 3193 // inputs to Loop Phis; these will be split by the allocator anyways. 3194 // Remove Opaque nodes. 3195 // (2) Move last-uses by commutative operations to the left input to encourage 3196 // Intel update-in-place two-address operations and better register usage 3197 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3198 // calls canonicalizing them back. 3199 // (3) Count the number of double-precision FP ops, single-precision FP ops 3200 // and call sites. On Intel, we can get correct rounding either by 3201 // forcing singles to memory (requires extra stores and loads after each 3202 // FP bytecode) or we can set a rounding mode bit (requires setting and 3203 // clearing the mode bit around call sites). The mode bit is only used 3204 // if the relative frequency of single FP ops to calls is low enough. 3205 // This is a key transform for SPEC mpeg_audio. 3206 // (4) Detect infinite loops; blobs of code reachable from above but not 3207 // below. Several of the Code_Gen algorithms fail on such code shapes, 3208 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3209 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3210 // Detection is by looking for IfNodes where only 1 projection is 3211 // reachable from below or CatchNodes missing some targets. 3212 // (5) Assert for insane oop offsets in debug mode. 3213 3214 bool Compile::final_graph_reshaping() { 3215 // an infinite loop may have been eliminated by the optimizer, 3216 // in which case the graph will be empty. 3217 if (root()->req() == 1) { 3218 record_method_not_compilable("trivial infinite loop"); 3219 return true; 3220 } 3221 3222 // Expensive nodes have their control input set to prevent the GVN 3223 // from freely commoning them. There's no GVN beyond this point so 3224 // no need to keep the control input. We want the expensive nodes to 3225 // be freely moved to the least frequent code path by gcm. 3226 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3227 for (int i = 0; i < expensive_count(); i++) { 3228 _expensive_nodes->at(i)->set_req(0, NULL); 3229 } 3230 3231 Final_Reshape_Counts frc; 3232 3233 // Visit everybody reachable! 3234 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3235 Node_Stack nstack(unique() >> 1); 3236 final_graph_reshaping_walk(nstack, root(), frc); 3237 3238 // Check for unreachable (from below) code (i.e., infinite loops). 3239 for( uint i = 0; i < frc._tests.size(); i++ ) { 3240 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3241 // Get number of CFG targets. 3242 // Note that PCTables include exception targets after calls. 3243 uint required_outcnt = n->required_outcnt(); 3244 if (n->outcnt() != required_outcnt) { 3245 // Check for a few special cases. Rethrow Nodes never take the 3246 // 'fall-thru' path, so expected kids is 1 less. 3247 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3248 if (n->in(0)->in(0)->is_Call()) { 3249 CallNode *call = n->in(0)->in(0)->as_Call(); 3250 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3251 required_outcnt--; // Rethrow always has 1 less kid 3252 } else if (call->req() > TypeFunc::Parms && 3253 call->is_CallDynamicJava()) { 3254 // Check for null receiver. In such case, the optimizer has 3255 // detected that the virtual call will always result in a null 3256 // pointer exception. The fall-through projection of this CatchNode 3257 // will not be populated. 3258 Node *arg0 = call->in(TypeFunc::Parms); 3259 if (arg0->is_Type() && 3260 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3261 required_outcnt--; 3262 } 3263 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3264 call->req() > TypeFunc::Parms+1 && 3265 call->is_CallStaticJava()) { 3266 // Check for negative array length. In such case, the optimizer has 3267 // detected that the allocation attempt will always result in an 3268 // exception. There is no fall-through projection of this CatchNode . 3269 Node *arg1 = call->in(TypeFunc::Parms+1); 3270 if (arg1->is_Type() && 3271 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3272 required_outcnt--; 3273 } 3274 } 3275 } 3276 } 3277 // Recheck with a better notion of 'required_outcnt' 3278 if (n->outcnt() != required_outcnt) { 3279 record_method_not_compilable("malformed control flow"); 3280 return true; // Not all targets reachable! 3281 } 3282 } 3283 // Check that I actually visited all kids. Unreached kids 3284 // must be infinite loops. 3285 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3286 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3287 record_method_not_compilable("infinite loop"); 3288 return true; // Found unvisited kid; must be unreach 3289 } 3290 } 3291 3292 // If original bytecodes contained a mixture of floats and doubles 3293 // check if the optimizer has made it homogenous, item (3). 3294 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3295 frc.get_float_count() > 32 && 3296 frc.get_double_count() == 0 && 3297 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3298 set_24_bit_selection_and_mode( false, true ); 3299 } 3300 3301 set_java_calls(frc.get_java_call_count()); 3302 set_inner_loops(frc.get_inner_loop_count()); 3303 3304 // No infinite loops, no reason to bail out. 3305 return false; 3306 } 3307 3308 //-----------------------------too_many_traps---------------------------------- 3309 // Report if there are too many traps at the current method and bci. 3310 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3311 bool Compile::too_many_traps(ciMethod* method, 3312 int bci, 3313 Deoptimization::DeoptReason reason) { 3314 ciMethodData* md = method->method_data(); 3315 if (md->is_empty()) { 3316 // Assume the trap has not occurred, or that it occurred only 3317 // because of a transient condition during start-up in the interpreter. 3318 return false; 3319 } 3320 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3321 if (md->has_trap_at(bci, m, reason) != 0) { 3322 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3323 // Also, if there are multiple reasons, or if there is no per-BCI record, 3324 // assume the worst. 3325 if (log()) 3326 log()->elem("observe trap='%s' count='%d'", 3327 Deoptimization::trap_reason_name(reason), 3328 md->trap_count(reason)); 3329 return true; 3330 } else { 3331 // Ignore method/bci and see if there have been too many globally. 3332 return too_many_traps(reason, md); 3333 } 3334 } 3335 3336 // Less-accurate variant which does not require a method and bci. 3337 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3338 ciMethodData* logmd) { 3339 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3340 // Too many traps globally. 3341 // Note that we use cumulative trap_count, not just md->trap_count. 3342 if (log()) { 3343 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3344 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3345 Deoptimization::trap_reason_name(reason), 3346 mcount, trap_count(reason)); 3347 } 3348 return true; 3349 } else { 3350 // The coast is clear. 3351 return false; 3352 } 3353 } 3354 3355 //--------------------------too_many_recompiles-------------------------------- 3356 // Report if there are too many recompiles at the current method and bci. 3357 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3358 // Is not eager to return true, since this will cause the compiler to use 3359 // Action_none for a trap point, to avoid too many recompilations. 3360 bool Compile::too_many_recompiles(ciMethod* method, 3361 int bci, 3362 Deoptimization::DeoptReason reason) { 3363 ciMethodData* md = method->method_data(); 3364 if (md->is_empty()) { 3365 // Assume the trap has not occurred, or that it occurred only 3366 // because of a transient condition during start-up in the interpreter. 3367 return false; 3368 } 3369 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3370 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3371 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3372 Deoptimization::DeoptReason per_bc_reason 3373 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3374 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3375 if ((per_bc_reason == Deoptimization::Reason_none 3376 || md->has_trap_at(bci, m, reason) != 0) 3377 // The trap frequency measure we care about is the recompile count: 3378 && md->trap_recompiled_at(bci, m) 3379 && md->overflow_recompile_count() >= bc_cutoff) { 3380 // Do not emit a trap here if it has already caused recompilations. 3381 // Also, if there are multiple reasons, or if there is no per-BCI record, 3382 // assume the worst. 3383 if (log()) 3384 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3385 Deoptimization::trap_reason_name(reason), 3386 md->trap_count(reason), 3387 md->overflow_recompile_count()); 3388 return true; 3389 } else if (trap_count(reason) != 0 3390 && decompile_count() >= m_cutoff) { 3391 // Too many recompiles globally, and we have seen this sort of trap. 3392 // Use cumulative decompile_count, not just md->decompile_count. 3393 if (log()) 3394 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3395 Deoptimization::trap_reason_name(reason), 3396 md->trap_count(reason), trap_count(reason), 3397 md->decompile_count(), decompile_count()); 3398 return true; 3399 } else { 3400 // The coast is clear. 3401 return false; 3402 } 3403 } 3404 3405 // Compute when not to trap. Used by matching trap based nodes and 3406 // NullCheck optimization. 3407 void Compile::set_allowed_deopt_reasons() { 3408 _allowed_reasons = 0; 3409 if (is_method_compilation()) { 3410 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3411 assert(rs < BitsPerInt, "recode bit map"); 3412 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3413 _allowed_reasons |= nth_bit(rs); 3414 } 3415 } 3416 } 3417 } 3418 3419 #ifndef PRODUCT 3420 //------------------------------verify_graph_edges--------------------------- 3421 // Walk the Graph and verify that there is a one-to-one correspondence 3422 // between Use-Def edges and Def-Use edges in the graph. 3423 void Compile::verify_graph_edges(bool no_dead_code) { 3424 if (VerifyGraphEdges) { 3425 ResourceArea *area = Thread::current()->resource_area(); 3426 Unique_Node_List visited(area); 3427 // Call recursive graph walk to check edges 3428 _root->verify_edges(visited); 3429 if (no_dead_code) { 3430 // Now make sure that no visited node is used by an unvisited node. 3431 bool dead_nodes = false; 3432 Unique_Node_List checked(area); 3433 while (visited.size() > 0) { 3434 Node* n = visited.pop(); 3435 checked.push(n); 3436 for (uint i = 0; i < n->outcnt(); i++) { 3437 Node* use = n->raw_out(i); 3438 if (checked.member(use)) continue; // already checked 3439 if (visited.member(use)) continue; // already in the graph 3440 if (use->is_Con()) continue; // a dead ConNode is OK 3441 // At this point, we have found a dead node which is DU-reachable. 3442 if (!dead_nodes) { 3443 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3444 dead_nodes = true; 3445 } 3446 use->dump(2); 3447 tty->print_cr("---"); 3448 checked.push(use); // No repeats; pretend it is now checked. 3449 } 3450 } 3451 assert(!dead_nodes, "using nodes must be reachable from root"); 3452 } 3453 } 3454 } 3455 3456 // Verify GC barriers consistency 3457 // Currently supported: 3458 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3459 void Compile::verify_barriers() { 3460 if (UseG1GC) { 3461 // Verify G1 pre-barriers 3462 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3463 3464 ResourceArea *area = Thread::current()->resource_area(); 3465 Unique_Node_List visited(area); 3466 Node_List worklist(area); 3467 // We're going to walk control flow backwards starting from the Root 3468 worklist.push(_root); 3469 while (worklist.size() > 0) { 3470 Node* x = worklist.pop(); 3471 if (x == NULL || x == top()) continue; 3472 if (visited.member(x)) { 3473 continue; 3474 } else { 3475 visited.push(x); 3476 } 3477 3478 if (x->is_Region()) { 3479 for (uint i = 1; i < x->req(); i++) { 3480 worklist.push(x->in(i)); 3481 } 3482 } else { 3483 worklist.push(x->in(0)); 3484 // We are looking for the pattern: 3485 // /->ThreadLocal 3486 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3487 // \->ConI(0) 3488 // We want to verify that the If and the LoadB have the same control 3489 // See GraphKit::g1_write_barrier_pre() 3490 if (x->is_If()) { 3491 IfNode *iff = x->as_If(); 3492 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3493 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3494 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3495 && cmp->in(1)->is_Load()) { 3496 LoadNode* load = cmp->in(1)->as_Load(); 3497 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3498 && load->in(2)->in(3)->is_Con() 3499 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3500 3501 Node* if_ctrl = iff->in(0); 3502 Node* load_ctrl = load->in(0); 3503 3504 if (if_ctrl != load_ctrl) { 3505 // Skip possible CProj->NeverBranch in infinite loops 3506 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3507 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3508 if_ctrl = if_ctrl->in(0)->in(0); 3509 } 3510 } 3511 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3512 } 3513 } 3514 } 3515 } 3516 } 3517 } 3518 } 3519 } 3520 3521 #endif 3522 3523 // The Compile object keeps track of failure reasons separately from the ciEnv. 3524 // This is required because there is not quite a 1-1 relation between the 3525 // ciEnv and its compilation task and the Compile object. Note that one 3526 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3527 // to backtrack and retry without subsuming loads. Other than this backtracking 3528 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3529 // by the logic in C2Compiler. 3530 void Compile::record_failure(const char* reason) { 3531 if (log() != NULL) { 3532 log()->elem("failure reason='%s' phase='compile'", reason); 3533 } 3534 if (_failure_reason == NULL) { 3535 // Record the first failure reason. 3536 _failure_reason = reason; 3537 } 3538 3539 EventCompilerFailure event; 3540 if (event.should_commit()) { 3541 event.set_compileID(Compile::compile_id()); 3542 event.set_failure(reason); 3543 event.commit(); 3544 } 3545 3546 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3547 C->print_method(PHASE_FAILURE); 3548 } 3549 _root = NULL; // flush the graph, too 3550 } 3551 3552 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3553 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3554 _phase_name(name), _dolog(CITimeVerbose) 3555 { 3556 if (_dolog) { 3557 C = Compile::current(); 3558 _log = C->log(); 3559 } else { 3560 C = NULL; 3561 _log = NULL; 3562 } 3563 if (_log != NULL) { 3564 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3565 _log->stamp(); 3566 _log->end_head(); 3567 } 3568 } 3569 3570 Compile::TracePhase::~TracePhase() { 3571 3572 C = Compile::current(); 3573 if (_dolog) { 3574 _log = C->log(); 3575 } else { 3576 _log = NULL; 3577 } 3578 3579 #ifdef ASSERT 3580 if (PrintIdealNodeCount) { 3581 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3582 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3583 } 3584 3585 if (VerifyIdealNodeCount) { 3586 Compile::current()->print_missing_nodes(); 3587 } 3588 #endif 3589 3590 if (_log != NULL) { 3591 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3592 } 3593 } 3594 3595 //============================================================================= 3596 // Two Constant's are equal when the type and the value are equal. 3597 bool Compile::Constant::operator==(const Constant& other) { 3598 if (type() != other.type() ) return false; 3599 if (can_be_reused() != other.can_be_reused()) return false; 3600 // For floating point values we compare the bit pattern. 3601 switch (type()) { 3602 case T_FLOAT: return (_v._value.i == other._v._value.i); 3603 case T_LONG: 3604 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3605 case T_OBJECT: 3606 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3607 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3608 case T_METADATA: return (_v._metadata == other._v._metadata); 3609 default: ShouldNotReachHere(); 3610 } 3611 return false; 3612 } 3613 3614 static int type_to_size_in_bytes(BasicType t) { 3615 switch (t) { 3616 case T_LONG: return sizeof(jlong ); 3617 case T_FLOAT: return sizeof(jfloat ); 3618 case T_DOUBLE: return sizeof(jdouble); 3619 case T_METADATA: return sizeof(Metadata*); 3620 // We use T_VOID as marker for jump-table entries (labels) which 3621 // need an internal word relocation. 3622 case T_VOID: 3623 case T_ADDRESS: 3624 case T_OBJECT: return sizeof(jobject); 3625 } 3626 3627 ShouldNotReachHere(); 3628 return -1; 3629 } 3630 3631 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3632 // sort descending 3633 if (a->freq() > b->freq()) return -1; 3634 if (a->freq() < b->freq()) return 1; 3635 return 0; 3636 } 3637 3638 void Compile::ConstantTable::calculate_offsets_and_size() { 3639 // First, sort the array by frequencies. 3640 _constants.sort(qsort_comparator); 3641 3642 #ifdef ASSERT 3643 // Make sure all jump-table entries were sorted to the end of the 3644 // array (they have a negative frequency). 3645 bool found_void = false; 3646 for (int i = 0; i < _constants.length(); i++) { 3647 Constant con = _constants.at(i); 3648 if (con.type() == T_VOID) 3649 found_void = true; // jump-tables 3650 else 3651 assert(!found_void, "wrong sorting"); 3652 } 3653 #endif 3654 3655 int offset = 0; 3656 for (int i = 0; i < _constants.length(); i++) { 3657 Constant* con = _constants.adr_at(i); 3658 3659 // Align offset for type. 3660 int typesize = type_to_size_in_bytes(con->type()); 3661 offset = align_size_up(offset, typesize); 3662 con->set_offset(offset); // set constant's offset 3663 3664 if (con->type() == T_VOID) { 3665 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3666 offset = offset + typesize * n->outcnt(); // expand jump-table 3667 } else { 3668 offset = offset + typesize; 3669 } 3670 } 3671 3672 // Align size up to the next section start (which is insts; see 3673 // CodeBuffer::align_at_start). 3674 assert(_size == -1, "already set?"); 3675 _size = align_size_up(offset, CodeEntryAlignment); 3676 } 3677 3678 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3679 MacroAssembler _masm(&cb); 3680 for (int i = 0; i < _constants.length(); i++) { 3681 Constant con = _constants.at(i); 3682 address constant_addr; 3683 switch (con.type()) { 3684 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3685 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3686 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3687 case T_OBJECT: { 3688 jobject obj = con.get_jobject(); 3689 int oop_index = _masm.oop_recorder()->find_index(obj); 3690 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3691 break; 3692 } 3693 case T_ADDRESS: { 3694 address addr = (address) con.get_jobject(); 3695 constant_addr = _masm.address_constant(addr); 3696 break; 3697 } 3698 // We use T_VOID as marker for jump-table entries (labels) which 3699 // need an internal word relocation. 3700 case T_VOID: { 3701 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3702 // Fill the jump-table with a dummy word. The real value is 3703 // filled in later in fill_jump_table. 3704 address dummy = (address) n; 3705 constant_addr = _masm.address_constant(dummy); 3706 // Expand jump-table 3707 for (uint i = 1; i < n->outcnt(); i++) { 3708 address temp_addr = _masm.address_constant(dummy + i); 3709 assert(temp_addr, "consts section too small"); 3710 } 3711 break; 3712 } 3713 case T_METADATA: { 3714 Metadata* obj = con.get_metadata(); 3715 int metadata_index = _masm.oop_recorder()->find_index(obj); 3716 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3717 break; 3718 } 3719 default: ShouldNotReachHere(); 3720 } 3721 assert(constant_addr, "consts section too small"); 3722 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3723 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3724 } 3725 } 3726 3727 int Compile::ConstantTable::find_offset(Constant& con) const { 3728 int idx = _constants.find(con); 3729 assert(idx != -1, "constant must be in constant table"); 3730 int offset = _constants.at(idx).offset(); 3731 assert(offset != -1, "constant table not emitted yet?"); 3732 return offset; 3733 } 3734 3735 void Compile::ConstantTable::add(Constant& con) { 3736 if (con.can_be_reused()) { 3737 int idx = _constants.find(con); 3738 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3739 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3740 return; 3741 } 3742 } 3743 (void) _constants.append(con); 3744 } 3745 3746 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3747 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3748 Constant con(type, value, b->_freq); 3749 add(con); 3750 return con; 3751 } 3752 3753 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3754 Constant con(metadata); 3755 add(con); 3756 return con; 3757 } 3758 3759 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3760 jvalue value; 3761 BasicType type = oper->type()->basic_type(); 3762 switch (type) { 3763 case T_LONG: value.j = oper->constantL(); break; 3764 case T_FLOAT: value.f = oper->constantF(); break; 3765 case T_DOUBLE: value.d = oper->constantD(); break; 3766 case T_OBJECT: 3767 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3768 case T_METADATA: return add((Metadata*)oper->constant()); break; 3769 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3770 } 3771 return add(n, type, value); 3772 } 3773 3774 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3775 jvalue value; 3776 // We can use the node pointer here to identify the right jump-table 3777 // as this method is called from Compile::Fill_buffer right before 3778 // the MachNodes are emitted and the jump-table is filled (means the 3779 // MachNode pointers do not change anymore). 3780 value.l = (jobject) n; 3781 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3782 add(con); 3783 return con; 3784 } 3785 3786 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3787 // If called from Compile::scratch_emit_size do nothing. 3788 if (Compile::current()->in_scratch_emit_size()) return; 3789 3790 assert(labels.is_nonempty(), "must be"); 3791 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3792 3793 // Since MachConstantNode::constant_offset() also contains 3794 // table_base_offset() we need to subtract the table_base_offset() 3795 // to get the plain offset into the constant table. 3796 int offset = n->constant_offset() - table_base_offset(); 3797 3798 MacroAssembler _masm(&cb); 3799 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3800 3801 for (uint i = 0; i < n->outcnt(); i++) { 3802 address* constant_addr = &jump_table_base[i]; 3803 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))); 3804 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3805 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3806 } 3807 } 3808 3809 //----------------------------static_subtype_check----------------------------- 3810 // Shortcut important common cases when superklass is exact: 3811 // (0) superklass is java.lang.Object (can occur in reflective code) 3812 // (1) subklass is already limited to a subtype of superklass => always ok 3813 // (2) subklass does not overlap with superklass => always fail 3814 // (3) superklass has NO subtypes and we can check with a simple compare. 3815 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3816 if (StressReflectiveCode) { 3817 return SSC_full_test; // Let caller generate the general case. 3818 } 3819 3820 if (superk == env()->Object_klass()) { 3821 return SSC_always_true; // (0) this test cannot fail 3822 } 3823 3824 ciType* superelem = superk; 3825 if (superelem->is_array_klass()) 3826 superelem = superelem->as_array_klass()->base_element_type(); 3827 3828 if (!subk->is_interface()) { // cannot trust static interface types yet 3829 if (subk->is_subtype_of(superk)) { 3830 return SSC_always_true; // (1) false path dead; no dynamic test needed 3831 } 3832 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3833 !superk->is_subtype_of(subk)) { 3834 return SSC_always_false; 3835 } 3836 } 3837 3838 // If casting to an instance klass, it must have no subtypes 3839 if (superk->is_interface()) { 3840 // Cannot trust interfaces yet. 3841 // %%% S.B. superk->nof_implementors() == 1 3842 } else if (superelem->is_instance_klass()) { 3843 ciInstanceKlass* ik = superelem->as_instance_klass(); 3844 if (!ik->has_subklass() && !ik->is_interface()) { 3845 if (!ik->is_final()) { 3846 // Add a dependency if there is a chance of a later subclass. 3847 dependencies()->assert_leaf_type(ik); 3848 } 3849 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3850 } 3851 } else { 3852 // A primitive array type has no subtypes. 3853 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3854 } 3855 3856 return SSC_full_test; 3857 } 3858 3859 // The message about the current inlining is accumulated in 3860 // _print_inlining_stream and transfered into the _print_inlining_list 3861 // once we know whether inlining succeeds or not. For regular 3862 // inlining, messages are appended to the buffer pointed by 3863 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3864 // a new buffer is added after _print_inlining_idx in the list. This 3865 // way we can update the inlining message for late inlining call site 3866 // when the inlining is attempted again. 3867 void Compile::print_inlining_init() { 3868 if (print_inlining() || print_intrinsics()) { 3869 _print_inlining_stream = new stringStream(); 3870 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3871 } 3872 } 3873 3874 void Compile::print_inlining_reinit() { 3875 if (print_inlining() || print_intrinsics()) { 3876 // Re allocate buffer when we change ResourceMark 3877 _print_inlining_stream = new stringStream(); 3878 } 3879 } 3880 3881 void Compile::print_inlining_reset() { 3882 _print_inlining_stream->reset(); 3883 } 3884 3885 void Compile::print_inlining_commit() { 3886 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3887 // Transfer the message from _print_inlining_stream to the current 3888 // _print_inlining_list buffer and clear _print_inlining_stream. 3889 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3890 print_inlining_reset(); 3891 } 3892 3893 void Compile::print_inlining_push() { 3894 // Add new buffer to the _print_inlining_list at current position 3895 _print_inlining_idx++; 3896 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3897 } 3898 3899 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3900 return _print_inlining_list->at(_print_inlining_idx); 3901 } 3902 3903 void Compile::print_inlining_update(CallGenerator* cg) { 3904 if (print_inlining() || print_intrinsics()) { 3905 if (!cg->is_late_inline()) { 3906 if (print_inlining_current().cg() != NULL) { 3907 print_inlining_push(); 3908 } 3909 print_inlining_commit(); 3910 } else { 3911 if (print_inlining_current().cg() != cg && 3912 (print_inlining_current().cg() != NULL || 3913 print_inlining_current().ss()->size() != 0)) { 3914 print_inlining_push(); 3915 } 3916 print_inlining_commit(); 3917 print_inlining_current().set_cg(cg); 3918 } 3919 } 3920 } 3921 3922 void Compile::print_inlining_move_to(CallGenerator* cg) { 3923 // We resume inlining at a late inlining call site. Locate the 3924 // corresponding inlining buffer so that we can update it. 3925 if (print_inlining()) { 3926 for (int i = 0; i < _print_inlining_list->length(); i++) { 3927 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3928 _print_inlining_idx = i; 3929 return; 3930 } 3931 } 3932 ShouldNotReachHere(); 3933 } 3934 } 3935 3936 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3937 if (print_inlining()) { 3938 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3939 assert(print_inlining_current().cg() == cg, "wrong entry"); 3940 // replace message with new message 3941 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3942 print_inlining_commit(); 3943 print_inlining_current().set_cg(cg); 3944 } 3945 } 3946 3947 void Compile::print_inlining_assert_ready() { 3948 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3949 } 3950 3951 void Compile::process_print_inlining() { 3952 bool do_print_inlining = print_inlining() || print_intrinsics(); 3953 if (do_print_inlining || log() != NULL) { 3954 // Print inlining message for candidates that we couldn't inline 3955 // for lack of space 3956 for (int i = 0; i < _late_inlines.length(); i++) { 3957 CallGenerator* cg = _late_inlines.at(i); 3958 if (!cg->is_mh_late_inline()) { 3959 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3960 if (do_print_inlining) { 3961 cg->print_inlining_late(msg); 3962 } 3963 log_late_inline_failure(cg, msg); 3964 } 3965 } 3966 } 3967 if (do_print_inlining) { 3968 ResourceMark rm; 3969 stringStream ss; 3970 for (int i = 0; i < _print_inlining_list->length(); i++) { 3971 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3972 } 3973 size_t end = ss.size(); 3974 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 3975 strncpy(_print_inlining_output, ss.base(), end+1); 3976 _print_inlining_output[end] = 0; 3977 } 3978 } 3979 3980 void Compile::dump_print_inlining() { 3981 if (_print_inlining_output != NULL) { 3982 tty->print_raw(_print_inlining_output); 3983 } 3984 } 3985 3986 void Compile::log_late_inline(CallGenerator* cg) { 3987 if (log() != NULL) { 3988 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 3989 cg->unique_id()); 3990 JVMState* p = cg->call_node()->jvms(); 3991 while (p != NULL) { 3992 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 3993 p = p->caller(); 3994 } 3995 log()->tail("late_inline"); 3996 } 3997 } 3998 3999 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4000 log_late_inline(cg); 4001 if (log() != NULL) { 4002 log()->inline_fail(msg); 4003 } 4004 } 4005 4006 void Compile::log_inline_id(CallGenerator* cg) { 4007 if (log() != NULL) { 4008 // The LogCompilation tool needs a unique way to identify late 4009 // inline call sites. This id must be unique for this call site in 4010 // this compilation. Try to have it unique across compilations as 4011 // well because it can be convenient when grepping through the log 4012 // file. 4013 // Distinguish OSR compilations from others in case CICountOSR is 4014 // on. 4015 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4016 cg->set_unique_id(id); 4017 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4018 } 4019 } 4020 4021 void Compile::log_inline_failure(const char* msg) { 4022 if (C->log() != NULL) { 4023 C->log()->inline_fail(msg); 4024 } 4025 } 4026 4027 4028 // Dump inlining replay data to the stream. 4029 // Don't change thread state and acquire any locks. 4030 void Compile::dump_inline_data(outputStream* out) { 4031 InlineTree* inl_tree = ilt(); 4032 if (inl_tree != NULL) { 4033 out->print(" inline %d", inl_tree->count()); 4034 inl_tree->dump_replay_data(out); 4035 } 4036 } 4037 4038 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4039 if (n1->Opcode() < n2->Opcode()) return -1; 4040 else if (n1->Opcode() > n2->Opcode()) return 1; 4041 4042 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())); 4043 for (uint i = 1; i < n1->req(); i++) { 4044 if (n1->in(i) < n2->in(i)) return -1; 4045 else if (n1->in(i) > n2->in(i)) return 1; 4046 } 4047 4048 return 0; 4049 } 4050 4051 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4052 Node* n1 = *n1p; 4053 Node* n2 = *n2p; 4054 4055 return cmp_expensive_nodes(n1, n2); 4056 } 4057 4058 void Compile::sort_expensive_nodes() { 4059 if (!expensive_nodes_sorted()) { 4060 _expensive_nodes->sort(cmp_expensive_nodes); 4061 } 4062 } 4063 4064 bool Compile::expensive_nodes_sorted() const { 4065 for (int i = 1; i < _expensive_nodes->length(); i++) { 4066 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4067 return false; 4068 } 4069 } 4070 return true; 4071 } 4072 4073 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4074 if (_expensive_nodes->length() == 0) { 4075 return false; 4076 } 4077 4078 assert(OptimizeExpensiveOps, "optimization off?"); 4079 4080 // Take this opportunity to remove dead nodes from the list 4081 int j = 0; 4082 for (int i = 0; i < _expensive_nodes->length(); i++) { 4083 Node* n = _expensive_nodes->at(i); 4084 if (!n->is_unreachable(igvn)) { 4085 assert(n->is_expensive(), "should be expensive"); 4086 _expensive_nodes->at_put(j, n); 4087 j++; 4088 } 4089 } 4090 _expensive_nodes->trunc_to(j); 4091 4092 // Then sort the list so that similar nodes are next to each other 4093 // and check for at least two nodes of identical kind with same data 4094 // inputs. 4095 sort_expensive_nodes(); 4096 4097 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4098 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4099 return true; 4100 } 4101 } 4102 4103 return false; 4104 } 4105 4106 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4107 if (_expensive_nodes->length() == 0) { 4108 return; 4109 } 4110 4111 assert(OptimizeExpensiveOps, "optimization off?"); 4112 4113 // Sort to bring similar nodes next to each other and clear the 4114 // control input of nodes for which there's only a single copy. 4115 sort_expensive_nodes(); 4116 4117 int j = 0; 4118 int identical = 0; 4119 int i = 0; 4120 bool modified = false; 4121 for (; i < _expensive_nodes->length()-1; i++) { 4122 assert(j <= i, "can't write beyond current index"); 4123 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4124 identical++; 4125 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4126 continue; 4127 } 4128 if (identical > 0) { 4129 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4130 identical = 0; 4131 } else { 4132 Node* n = _expensive_nodes->at(i); 4133 igvn.replace_input_of(n, 0, NULL); 4134 igvn.hash_insert(n); 4135 modified = true; 4136 } 4137 } 4138 if (identical > 0) { 4139 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4140 } else if (_expensive_nodes->length() >= 1) { 4141 Node* n = _expensive_nodes->at(i); 4142 igvn.replace_input_of(n, 0, NULL); 4143 igvn.hash_insert(n); 4144 modified = true; 4145 } 4146 _expensive_nodes->trunc_to(j); 4147 if (modified) { 4148 igvn.optimize(); 4149 } 4150 } 4151 4152 void Compile::add_expensive_node(Node * n) { 4153 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4154 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4155 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4156 if (OptimizeExpensiveOps) { 4157 _expensive_nodes->append(n); 4158 } else { 4159 // Clear control input and let IGVN optimize expensive nodes if 4160 // OptimizeExpensiveOps is off. 4161 n->set_req(0, NULL); 4162 } 4163 } 4164 4165 /** 4166 * Remove the speculative part of types and clean up the graph 4167 */ 4168 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4169 if (UseTypeSpeculation) { 4170 Unique_Node_List worklist; 4171 worklist.push(root()); 4172 int modified = 0; 4173 // Go over all type nodes that carry a speculative type, drop the 4174 // speculative part of the type and enqueue the node for an igvn 4175 // which may optimize it out. 4176 for (uint next = 0; next < worklist.size(); ++next) { 4177 Node *n = worklist.at(next); 4178 if (n->is_Type()) { 4179 TypeNode* tn = n->as_Type(); 4180 const Type* t = tn->type(); 4181 const Type* t_no_spec = t->remove_speculative(); 4182 if (t_no_spec != t) { 4183 bool in_hash = igvn.hash_delete(n); 4184 assert(in_hash, "node should be in igvn hash table"); 4185 tn->set_type(t_no_spec); 4186 igvn.hash_insert(n); 4187 igvn._worklist.push(n); // give it a chance to go away 4188 modified++; 4189 } 4190 } 4191 uint max = n->len(); 4192 for( uint i = 0; i < max; ++i ) { 4193 Node *m = n->in(i); 4194 if (not_a_node(m)) continue; 4195 worklist.push(m); 4196 } 4197 } 4198 // Drop the speculative part of all types in the igvn's type table 4199 igvn.remove_speculative_types(); 4200 if (modified > 0) { 4201 igvn.optimize(); 4202 } 4203 #ifdef ASSERT 4204 // Verify that after the IGVN is over no speculative type has resurfaced 4205 worklist.clear(); 4206 worklist.push(root()); 4207 for (uint next = 0; next < worklist.size(); ++next) { 4208 Node *n = worklist.at(next); 4209 const Type* t = igvn.type_or_null(n); 4210 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4211 if (n->is_Type()) { 4212 t = n->as_Type()->type(); 4213 assert(t == t->remove_speculative(), "no more speculative types"); 4214 } 4215 uint max = n->len(); 4216 for( uint i = 0; i < max; ++i ) { 4217 Node *m = n->in(i); 4218 if (not_a_node(m)) continue; 4219 worklist.push(m); 4220 } 4221 } 4222 igvn.check_no_speculative_types(); 4223 #endif 4224 } 4225 } 4226 4227 // Auxiliary method to support randomized stressing/fuzzing. 4228 // 4229 // This method can be called the arbitrary number of times, with current count 4230 // as the argument. The logic allows selecting a single candidate from the 4231 // running list of candidates as follows: 4232 // int count = 0; 4233 // Cand* selected = null; 4234 // while(cand = cand->next()) { 4235 // if (randomized_select(++count)) { 4236 // selected = cand; 4237 // } 4238 // } 4239 // 4240 // Including count equalizes the chances any candidate is "selected". 4241 // This is useful when we don't have the complete list of candidates to choose 4242 // from uniformly. In this case, we need to adjust the randomicity of the 4243 // selection, or else we will end up biasing the selection towards the latter 4244 // candidates. 4245 // 4246 // Quick back-envelope calculation shows that for the list of n candidates 4247 // the equal probability for the candidate to persist as "best" can be 4248 // achieved by replacing it with "next" k-th candidate with the probability 4249 // of 1/k. It can be easily shown that by the end of the run, the 4250 // probability for any candidate is converged to 1/n, thus giving the 4251 // uniform distribution among all the candidates. 4252 // 4253 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4254 #define RANDOMIZED_DOMAIN_POW 29 4255 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4256 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4257 bool Compile::randomized_select(int count) { 4258 assert(count > 0, "only positive"); 4259 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4260 }