1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileBroker.hpp" 33 #include "compiler/compileLog.hpp" 34 #include "compiler/disassembler.hpp" 35 #include "compiler/oopMap.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/convertnode.hpp" 46 #include "opto/divnode.hpp" 47 #include "opto/escape.hpp" 48 #include "opto/idealGraphPrinter.hpp" 49 #include "opto/loopnode.hpp" 50 #include "opto/machnode.hpp" 51 #include "opto/macro.hpp" 52 #include "opto/matcher.hpp" 53 #include "opto/mathexactnode.hpp" 54 #include "opto/memnode.hpp" 55 #include "opto/mulnode.hpp" 56 #include "opto/narrowptrnode.hpp" 57 #include "opto/node.hpp" 58 #include "opto/opcodes.hpp" 59 #include "opto/output.hpp" 60 #include "opto/parse.hpp" 61 #include "opto/phaseX.hpp" 62 #include "opto/rootnode.hpp" 63 #include "opto/runtime.hpp" 64 #include "opto/stringopts.hpp" 65 #include "opto/type.hpp" 66 #include "opto/vectornode.hpp" 67 #include "runtime/arguments.hpp" 68 #include "runtime/sharedRuntime.hpp" 69 #include "runtime/signature.hpp" 70 #include "runtime/stubRoutines.hpp" 71 #include "runtime/timer.hpp" 72 #include "utilities/copy.hpp" 73 74 75 // -------------------- Compile::mach_constant_base_node ----------------------- 76 // Constant table base node singleton. 77 MachConstantBaseNode* Compile::mach_constant_base_node() { 78 if (_mach_constant_base_node == NULL) { 79 _mach_constant_base_node = new MachConstantBaseNode(); 80 _mach_constant_base_node->add_req(C->root()); 81 } 82 return _mach_constant_base_node; 83 } 84 85 86 /// Support for intrinsics. 87 88 // Return the index at which m must be inserted (or already exists). 89 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 90 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 91 #ifdef ASSERT 92 for (int i = 1; i < _intrinsics->length(); i++) { 93 CallGenerator* cg1 = _intrinsics->at(i-1); 94 CallGenerator* cg2 = _intrinsics->at(i); 95 assert(cg1->method() != cg2->method() 96 ? cg1->method() < cg2->method() 97 : cg1->is_virtual() < cg2->is_virtual(), 98 "compiler intrinsics list must stay sorted"); 99 } 100 #endif 101 // Binary search sorted list, in decreasing intervals [lo, hi]. 102 int lo = 0, hi = _intrinsics->length()-1; 103 while (lo <= hi) { 104 int mid = (uint)(hi + lo) / 2; 105 ciMethod* mid_m = _intrinsics->at(mid)->method(); 106 if (m < mid_m) { 107 hi = mid-1; 108 } else if (m > mid_m) { 109 lo = mid+1; 110 } else { 111 // look at minor sort key 112 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 113 if (is_virtual < mid_virt) { 114 hi = mid-1; 115 } else if (is_virtual > mid_virt) { 116 lo = mid+1; 117 } else { 118 return mid; // exact match 119 } 120 } 121 } 122 return lo; // inexact match 123 } 124 125 void Compile::register_intrinsic(CallGenerator* cg) { 126 if (_intrinsics == NULL) { 127 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 128 } 129 // This code is stolen from ciObjectFactory::insert. 130 // Really, GrowableArray should have methods for 131 // insert_at, remove_at, and binary_search. 132 int len = _intrinsics->length(); 133 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 134 if (index == len) { 135 _intrinsics->append(cg); 136 } else { 137 #ifdef ASSERT 138 CallGenerator* oldcg = _intrinsics->at(index); 139 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 140 #endif 141 _intrinsics->append(_intrinsics->at(len-1)); 142 int pos; 143 for (pos = len-2; pos >= index; pos--) { 144 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 145 } 146 _intrinsics->at_put(index, cg); 147 } 148 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 149 } 150 151 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 152 assert(m->is_loaded(), "don't try this on unloaded methods"); 153 if (_intrinsics != NULL) { 154 int index = intrinsic_insertion_index(m, is_virtual); 155 if (index < _intrinsics->length() 156 && _intrinsics->at(index)->method() == m 157 && _intrinsics->at(index)->is_virtual() == is_virtual) { 158 return _intrinsics->at(index); 159 } 160 } 161 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 162 if (m->intrinsic_id() != vmIntrinsics::_none && 163 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 164 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 165 if (cg != NULL) { 166 // Save it for next time: 167 register_intrinsic(cg); 168 return cg; 169 } else { 170 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 171 } 172 } 173 return NULL; 174 } 175 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 177 // in library_call.cpp. 178 179 180 #ifndef PRODUCT 181 // statistics gathering... 182 183 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 185 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 187 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 188 int oflags = _intrinsic_hist_flags[id]; 189 assert(flags != 0, "what happened?"); 190 if (is_virtual) { 191 flags |= _intrinsic_virtual; 192 } 193 bool changed = (flags != oflags); 194 if ((flags & _intrinsic_worked) != 0) { 195 juint count = (_intrinsic_hist_count[id] += 1); 196 if (count == 1) { 197 changed = true; // first time 198 } 199 // increment the overall count also: 200 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 201 } 202 if (changed) { 203 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 204 // Something changed about the intrinsic's virtuality. 205 if ((flags & _intrinsic_virtual) != 0) { 206 // This is the first use of this intrinsic as a virtual call. 207 if (oflags != 0) { 208 // We already saw it as a non-virtual, so note both cases. 209 flags |= _intrinsic_both; 210 } 211 } else if ((oflags & _intrinsic_both) == 0) { 212 // This is the first use of this intrinsic as a non-virtual 213 flags |= _intrinsic_both; 214 } 215 } 216 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 217 } 218 // update the overall flags also: 219 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 220 return changed; 221 } 222 223 static char* format_flags(int flags, char* buf) { 224 buf[0] = 0; 225 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 226 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 227 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 228 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 229 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 230 if (buf[0] == 0) strcat(buf, ","); 231 assert(buf[0] == ',', "must be"); 232 return &buf[1]; 233 } 234 235 void Compile::print_intrinsic_statistics() { 236 char flagsbuf[100]; 237 ttyLocker ttyl; 238 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 239 tty->print_cr("Compiler intrinsic usage:"); 240 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 241 if (total == 0) total = 1; // avoid div0 in case of no successes 242 #define PRINT_STAT_LINE(name, c, f) \ 243 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 244 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 245 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 246 int flags = _intrinsic_hist_flags[id]; 247 juint count = _intrinsic_hist_count[id]; 248 if ((flags | count) != 0) { 249 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 250 } 251 } 252 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 253 if (xtty != NULL) xtty->tail("statistics"); 254 } 255 256 void Compile::print_statistics() { 257 { ttyLocker ttyl; 258 if (xtty != NULL) xtty->head("statistics type='opto'"); 259 Parse::print_statistics(); 260 PhaseCCP::print_statistics(); 261 PhaseRegAlloc::print_statistics(); 262 Scheduling::print_statistics(); 263 PhasePeephole::print_statistics(); 264 PhaseIdealLoop::print_statistics(); 265 if (xtty != NULL) xtty->tail("statistics"); 266 } 267 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 268 // put this under its own <statistics> element. 269 print_intrinsic_statistics(); 270 } 271 } 272 #endif //PRODUCT 273 274 // Support for bundling info 275 Bundle* Compile::node_bundling(const Node *n) { 276 assert(valid_bundle_info(n), "oob"); 277 return &_node_bundling_base[n->_idx]; 278 } 279 280 bool Compile::valid_bundle_info(const Node *n) { 281 return (_node_bundling_limit > n->_idx); 282 } 283 284 285 void Compile::gvn_replace_by(Node* n, Node* nn) { 286 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 287 Node* use = n->last_out(i); 288 bool is_in_table = initial_gvn()->hash_delete(use); 289 uint uses_found = 0; 290 for (uint j = 0; j < use->len(); j++) { 291 if (use->in(j) == n) { 292 if (j < use->req()) 293 use->set_req(j, nn); 294 else 295 use->set_prec(j, nn); 296 uses_found++; 297 } 298 } 299 if (is_in_table) { 300 // reinsert into table 301 initial_gvn()->hash_find_insert(use); 302 } 303 record_for_igvn(use); 304 i -= uses_found; // we deleted 1 or more copies of this edge 305 } 306 } 307 308 309 static inline bool not_a_node(const Node* n) { 310 if (n == NULL) return true; 311 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 312 if (*(address*)n == badAddress) return true; // kill by Node::destruct 313 return false; 314 } 315 316 // Identify all nodes that are reachable from below, useful. 317 // Use breadth-first pass that records state in a Unique_Node_List, 318 // recursive traversal is slower. 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 320 int estimated_worklist_size = unique(); 321 useful.map( estimated_worklist_size, NULL ); // preallocate space 322 323 // Initialize worklist 324 if (root() != NULL) { useful.push(root()); } 325 // If 'top' is cached, declare it useful to preserve cached node 326 if( cached_top_node() ) { useful.push(cached_top_node()); } 327 328 // Push all useful nodes onto the list, breadthfirst 329 for( uint next = 0; next < useful.size(); ++next ) { 330 assert( next < unique(), "Unique useful nodes < total nodes"); 331 Node *n = useful.at(next); 332 uint max = n->len(); 333 for( uint i = 0; i < max; ++i ) { 334 Node *m = n->in(i); 335 if (not_a_node(m)) continue; 336 useful.push(m); 337 } 338 } 339 } 340 341 // Update dead_node_list with any missing dead nodes using useful 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 343 void Compile::update_dead_node_list(Unique_Node_List &useful) { 344 uint max_idx = unique(); 345 VectorSet& useful_node_set = useful.member_set(); 346 347 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 348 // If node with index node_idx is not in useful set, 349 // mark it as dead in dead node list. 350 if (! useful_node_set.test(node_idx) ) { 351 record_dead_node(node_idx); 352 } 353 } 354 } 355 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 357 int shift = 0; 358 for (int i = 0; i < inlines->length(); i++) { 359 CallGenerator* cg = inlines->at(i); 360 CallNode* call = cg->call_node(); 361 if (shift > 0) { 362 inlines->at_put(i-shift, cg); 363 } 364 if (!useful.member(call)) { 365 shift++; 366 } 367 } 368 inlines->trunc_to(inlines->length()-shift); 369 } 370 371 // Disconnect all useless nodes by disconnecting those at the boundary. 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 373 uint next = 0; 374 while (next < useful.size()) { 375 Node *n = useful.at(next++); 376 if (n->is_SafePoint()) { 377 // We're done with a parsing phase. Replaced nodes are not valid 378 // beyond that point. 379 n->as_SafePoint()->delete_replaced_nodes(); 380 } 381 // Use raw traversal of out edges since this code removes out edges 382 int max = n->outcnt(); 383 for (int j = 0; j < max; ++j) { 384 Node* child = n->raw_out(j); 385 if (! useful.member(child)) { 386 assert(!child->is_top() || child != top(), 387 "If top is cached in Compile object it is in useful list"); 388 // Only need to remove this out-edge to the useless node 389 n->raw_del_out(j); 390 --j; 391 --max; 392 } 393 } 394 if (n->outcnt() == 1 && n->has_special_unique_user()) { 395 record_for_igvn(n->unique_out()); 396 } 397 } 398 // Remove useless macro and predicate opaq nodes 399 for (int i = C->macro_count()-1; i >= 0; i--) { 400 Node* n = C->macro_node(i); 401 if (!useful.member(n)) { 402 remove_macro_node(n); 403 } 404 } 405 // Remove useless expensive node 406 for (int i = C->expensive_count()-1; i >= 0; i--) { 407 Node* n = C->expensive_node(i); 408 if (!useful.member(n)) { 409 remove_expensive_node(n); 410 } 411 } 412 // clean up the late inline lists 413 remove_useless_late_inlines(&_string_late_inlines, useful); 414 remove_useless_late_inlines(&_boxing_late_inlines, useful); 415 remove_useless_late_inlines(&_late_inlines, useful); 416 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 417 } 418 419 //------------------------------frame_size_in_words----------------------------- 420 // frame_slots in units of words 421 int Compile::frame_size_in_words() const { 422 // shift is 0 in LP32 and 1 in LP64 423 const int shift = (LogBytesPerWord - LogBytesPerInt); 424 int words = _frame_slots >> shift; 425 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 426 return words; 427 } 428 429 // To bang the stack of this compiled method we use the stack size 430 // that the interpreter would need in case of a deoptimization. This 431 // removes the need to bang the stack in the deoptimization blob which 432 // in turn simplifies stack overflow handling. 433 int Compile::bang_size_in_bytes() const { 434 return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size); 435 } 436 437 // ============================================================================ 438 //------------------------------CompileWrapper--------------------------------- 439 class CompileWrapper : public StackObj { 440 Compile *const _compile; 441 public: 442 CompileWrapper(Compile* compile); 443 444 ~CompileWrapper(); 445 }; 446 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 448 // the Compile* pointer is stored in the current ciEnv: 449 ciEnv* env = compile->env(); 450 assert(env == ciEnv::current(), "must already be a ciEnv active"); 451 assert(env->compiler_data() == NULL, "compile already active?"); 452 env->set_compiler_data(compile); 453 assert(compile == Compile::current(), "sanity"); 454 455 compile->set_type_dict(NULL); 456 compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena())); 457 compile->clone_map().set_clone_idx(0); 458 compile->set_type_hwm(NULL); 459 compile->set_type_last_size(0); 460 compile->set_last_tf(NULL, NULL); 461 compile->set_indexSet_arena(NULL); 462 compile->set_indexSet_free_block_list(NULL); 463 compile->init_type_arena(); 464 Type::Initialize(compile); 465 _compile->set_scratch_buffer_blob(NULL); 466 _compile->begin_method(); 467 _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption); 468 } 469 CompileWrapper::~CompileWrapper() { 470 _compile->end_method(); 471 if (_compile->scratch_buffer_blob() != NULL) 472 BufferBlob::free(_compile->scratch_buffer_blob()); 473 _compile->env()->set_compiler_data(NULL); 474 } 475 476 477 //----------------------------print_compile_messages--------------------------- 478 void Compile::print_compile_messages() { 479 #ifndef PRODUCT 480 // Check if recompiling 481 if (_subsume_loads == false && PrintOpto) { 482 // Recompiling without allowing machine instructions to subsume loads 483 tty->print_cr("*********************************************************"); 484 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 485 tty->print_cr("*********************************************************"); 486 } 487 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 488 // Recompiling without escape analysis 489 tty->print_cr("*********************************************************"); 490 tty->print_cr("** Bailout: Recompile without escape analysis **"); 491 tty->print_cr("*********************************************************"); 492 } 493 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 494 // Recompiling without boxing elimination 495 tty->print_cr("*********************************************************"); 496 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 497 tty->print_cr("*********************************************************"); 498 } 499 if (C->directive()->BreakAtCompileOption) { 500 // Open the debugger when compiling this method. 501 tty->print("### Breaking when compiling: "); 502 method()->print_short_name(); 503 tty->cr(); 504 BREAKPOINT; 505 } 506 507 if( PrintOpto ) { 508 if (is_osr_compilation()) { 509 tty->print("[OSR]%3d", _compile_id); 510 } else { 511 tty->print("%3d", _compile_id); 512 } 513 } 514 #endif 515 } 516 517 518 //-----------------------init_scratch_buffer_blob------------------------------ 519 // Construct a temporary BufferBlob and cache it for this compile. 520 void Compile::init_scratch_buffer_blob(int const_size) { 521 // If there is already a scratch buffer blob allocated and the 522 // constant section is big enough, use it. Otherwise free the 523 // current and allocate a new one. 524 BufferBlob* blob = scratch_buffer_blob(); 525 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 526 // Use the current blob. 527 } else { 528 if (blob != NULL) { 529 BufferBlob::free(blob); 530 } 531 532 ResourceMark rm; 533 _scratch_const_size = const_size; 534 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 535 blob = BufferBlob::create("Compile::scratch_buffer", size); 536 // Record the buffer blob for next time. 537 set_scratch_buffer_blob(blob); 538 // Have we run out of code space? 539 if (scratch_buffer_blob() == NULL) { 540 // Let CompilerBroker disable further compilations. 541 record_failure("Not enough space for scratch buffer in CodeCache"); 542 return; 543 } 544 } 545 546 // Initialize the relocation buffers 547 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 548 set_scratch_locs_memory(locs_buf); 549 } 550 551 552 //-----------------------scratch_emit_size------------------------------------- 553 // Helper function that computes size by emitting code 554 uint Compile::scratch_emit_size(const Node* n) { 555 // Start scratch_emit_size section. 556 set_in_scratch_emit_size(true); 557 558 // Emit into a trash buffer and count bytes emitted. 559 // This is a pretty expensive way to compute a size, 560 // but it works well enough if seldom used. 561 // All common fixed-size instructions are given a size 562 // method by the AD file. 563 // Note that the scratch buffer blob and locs memory are 564 // allocated at the beginning of the compile task, and 565 // may be shared by several calls to scratch_emit_size. 566 // The allocation of the scratch buffer blob is particularly 567 // expensive, since it has to grab the code cache lock. 568 BufferBlob* blob = this->scratch_buffer_blob(); 569 assert(blob != NULL, "Initialize BufferBlob at start"); 570 assert(blob->size() > MAX_inst_size, "sanity"); 571 relocInfo* locs_buf = scratch_locs_memory(); 572 address blob_begin = blob->content_begin(); 573 address blob_end = (address)locs_buf; 574 assert(blob->content_contains(blob_end), "sanity"); 575 CodeBuffer buf(blob_begin, blob_end - blob_begin); 576 buf.initialize_consts_size(_scratch_const_size); 577 buf.initialize_stubs_size(MAX_stubs_size); 578 assert(locs_buf != NULL, "sanity"); 579 int lsize = MAX_locs_size / 3; 580 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 581 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 582 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 583 584 // Do the emission. 585 586 Label fakeL; // Fake label for branch instructions. 587 Label* saveL = NULL; 588 uint save_bnum = 0; 589 bool is_branch = n->is_MachBranch(); 590 if (is_branch) { 591 MacroAssembler masm(&buf); 592 masm.bind(fakeL); 593 n->as_MachBranch()->save_label(&saveL, &save_bnum); 594 n->as_MachBranch()->label_set(&fakeL, 0); 595 } 596 n->emit(buf, this->regalloc()); 597 598 // Emitting into the scratch buffer should not fail 599 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 600 601 if (is_branch) // Restore label. 602 n->as_MachBranch()->label_set(saveL, save_bnum); 603 604 // End scratch_emit_size section. 605 set_in_scratch_emit_size(false); 606 607 return buf.insts_size(); 608 } 609 610 611 // ============================================================================ 612 //------------------------------Compile standard------------------------------- 613 debug_only( int Compile::_debug_idx = 100000; ) 614 615 // Compile a method. entry_bci is -1 for normal compilations and indicates 616 // the continuation bci for on stack replacement. 617 618 619 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 620 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive) 621 : Phase(Compiler), 622 _env(ci_env), 623 _directive(directive), 624 _log(ci_env->log()), 625 _compile_id(ci_env->compile_id()), 626 _save_argument_registers(false), 627 _stub_name(NULL), 628 _stub_function(NULL), 629 _stub_entry_point(NULL), 630 _method(target), 631 _entry_bci(osr_bci), 632 _initial_gvn(NULL), 633 _for_igvn(NULL), 634 _warm_calls(NULL), 635 _subsume_loads(subsume_loads), 636 _do_escape_analysis(do_escape_analysis), 637 _eliminate_boxing(eliminate_boxing), 638 _failure_reason(NULL), 639 _code_buffer("Compile::Fill_buffer"), 640 _orig_pc_slot(0), 641 _orig_pc_slot_offset_in_bytes(0), 642 _has_method_handle_invokes(false), 643 _mach_constant_base_node(NULL), 644 _node_bundling_limit(0), 645 _node_bundling_base(NULL), 646 _java_calls(0), 647 _inner_loops(0), 648 _scratch_const_size(-1), 649 _in_scratch_emit_size(false), 650 _dead_node_list(comp_arena()), 651 _dead_node_count(0), 652 #ifndef PRODUCT 653 _trace_opto_output(directive->TraceOptoOutputOption), 654 _in_dump_cnt(0), 655 _printer(IdealGraphPrinter::printer()), 656 #endif 657 _congraph(NULL), 658 _comp_arena(mtCompiler), 659 _node_arena(mtCompiler), 660 _old_arena(mtCompiler), 661 _Compile_types(mtCompiler), 662 _replay_inline_data(NULL), 663 _late_inlines(comp_arena(), 2, 0, NULL), 664 _string_late_inlines(comp_arena(), 2, 0, NULL), 665 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 666 _late_inlines_pos(0), 667 _number_of_mh_late_inlines(0), 668 _inlining_progress(false), 669 _inlining_incrementally(false), 670 _print_inlining_list(NULL), 671 _print_inlining_stream(NULL), 672 _print_inlining_idx(0), 673 _print_inlining_output(NULL), 674 _interpreter_frame_size(0), 675 _max_node_limit(MaxNodeLimit) { 676 C = this; 677 #ifndef PRODUCT 678 if (_printer != NULL) { 679 _printer->set_compile(this); 680 } 681 #endif 682 CompileWrapper cw(this); 683 684 if (CITimeVerbose) { 685 tty->print(" "); 686 target->holder()->name()->print(); 687 tty->print("."); 688 target->print_short_name(); 689 tty->print(" "); 690 } 691 TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose); 692 TraceTime t2(NULL, &_t_methodCompilation, CITime, false); 693 694 #ifndef PRODUCT 695 bool print_opto_assembly = directive->PrintOptoAssemblyOption; 696 if (!print_opto_assembly) { 697 bool print_assembly = directive->PrintAssemblyOption; 698 if (print_assembly && !Disassembler::can_decode()) { 699 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 700 print_opto_assembly = true; 701 } 702 } 703 set_print_assembly(print_opto_assembly); 704 set_parsed_irreducible_loop(false); 705 706 if (directive->ReplayInlineOption) { 707 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 708 } 709 #endif 710 set_print_inlining(directive->PrintInliningOption NOT_PRODUCT( || PrintOptoInlining)); 711 set_print_intrinsics(directive->PrintIntrinsicsOption); 712 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 713 714 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 715 // Make sure the method being compiled gets its own MDO, 716 // so we can at least track the decompile_count(). 717 // Need MDO to record RTM code generation state. 718 method()->ensure_method_data(); 719 } 720 721 Init(::AliasLevel); 722 723 724 print_compile_messages(); 725 726 _ilt = InlineTree::build_inline_tree_root(); 727 728 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 729 assert(num_alias_types() >= AliasIdxRaw, ""); 730 731 #define MINIMUM_NODE_HASH 1023 732 // Node list that Iterative GVN will start with 733 Unique_Node_List for_igvn(comp_arena()); 734 set_for_igvn(&for_igvn); 735 736 // GVN that will be run immediately on new nodes 737 uint estimated_size = method()->code_size()*4+64; 738 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 739 PhaseGVN gvn(node_arena(), estimated_size); 740 set_initial_gvn(&gvn); 741 742 print_inlining_init(); 743 { // Scope for timing the parser 744 TracePhase tp("parse", &timers[_t_parser]); 745 746 // Put top into the hash table ASAP. 747 initial_gvn()->transform_no_reclaim(top()); 748 749 // Set up tf(), start(), and find a CallGenerator. 750 CallGenerator* cg = NULL; 751 if (is_osr_compilation()) { 752 const TypeTuple *domain = StartOSRNode::osr_domain(); 753 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 754 init_tf(TypeFunc::make(domain, range)); 755 StartNode* s = new StartOSRNode(root(), domain); 756 initial_gvn()->set_type_bottom(s); 757 init_start(s); 758 cg = CallGenerator::for_osr(method(), entry_bci()); 759 } else { 760 // Normal case. 761 init_tf(TypeFunc::make(method())); 762 StartNode* s = new StartNode(root(), tf()->domain()); 763 initial_gvn()->set_type_bottom(s); 764 init_start(s); 765 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 766 // With java.lang.ref.reference.get() we must go through the 767 // intrinsic when G1 is enabled - even when get() is the root 768 // method of the compile - so that, if necessary, the value in 769 // the referent field of the reference object gets recorded by 770 // the pre-barrier code. 771 // Specifically, if G1 is enabled, the value in the referent 772 // field is recorded by the G1 SATB pre barrier. This will 773 // result in the referent being marked live and the reference 774 // object removed from the list of discovered references during 775 // reference processing. 776 cg = find_intrinsic(method(), false); 777 } 778 if (cg == NULL) { 779 float past_uses = method()->interpreter_invocation_count(); 780 float expected_uses = past_uses; 781 cg = CallGenerator::for_inline(method(), expected_uses); 782 } 783 } 784 if (failing()) return; 785 if (cg == NULL) { 786 record_method_not_compilable_all_tiers("cannot parse method"); 787 return; 788 } 789 JVMState* jvms = build_start_state(start(), tf()); 790 if ((jvms = cg->generate(jvms)) == NULL) { 791 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 792 record_method_not_compilable("method parse failed"); 793 } 794 return; 795 } 796 GraphKit kit(jvms); 797 798 if (!kit.stopped()) { 799 // Accept return values, and transfer control we know not where. 800 // This is done by a special, unique ReturnNode bound to root. 801 return_values(kit.jvms()); 802 } 803 804 if (kit.has_exceptions()) { 805 // Any exceptions that escape from this call must be rethrown 806 // to whatever caller is dynamically above us on the stack. 807 // This is done by a special, unique RethrowNode bound to root. 808 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 809 } 810 811 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 812 813 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 814 inline_string_calls(true); 815 } 816 817 if (failing()) return; 818 819 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 820 821 // Remove clutter produced by parsing. 822 if (!failing()) { 823 ResourceMark rm; 824 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 825 } 826 } 827 828 // Note: Large methods are capped off in do_one_bytecode(). 829 if (failing()) return; 830 831 // After parsing, node notes are no longer automagic. 832 // They must be propagated by register_new_node_with_optimizer(), 833 // clone(), or the like. 834 set_default_node_notes(NULL); 835 836 for (;;) { 837 int successes = Inline_Warm(); 838 if (failing()) return; 839 if (successes == 0) break; 840 } 841 842 // Drain the list. 843 Finish_Warm(); 844 #ifndef PRODUCT 845 if (_printer && _printer->should_print(1)) { 846 _printer->print_inlining(); 847 } 848 #endif 849 850 if (failing()) return; 851 NOT_PRODUCT( verify_graph_edges(); ) 852 853 // Now optimize 854 Optimize(); 855 if (failing()) return; 856 NOT_PRODUCT( verify_graph_edges(); ) 857 858 #ifndef PRODUCT 859 if (PrintIdeal) { 860 ttyLocker ttyl; // keep the following output all in one block 861 // This output goes directly to the tty, not the compiler log. 862 // To enable tools to match it up with the compilation activity, 863 // be sure to tag this tty output with the compile ID. 864 if (xtty != NULL) { 865 xtty->head("ideal compile_id='%d'%s", compile_id(), 866 is_osr_compilation() ? " compile_kind='osr'" : 867 ""); 868 } 869 root()->dump(9999); 870 if (xtty != NULL) { 871 xtty->tail("ideal"); 872 } 873 } 874 #endif 875 876 NOT_PRODUCT( verify_barriers(); ) 877 878 // Dump compilation data to replay it. 879 if (directive->DumpReplayOption) { 880 env()->dump_replay_data(_compile_id); 881 } 882 if (directive->DumpInlineOption && (ilt() != NULL)) { 883 env()->dump_inline_data(_compile_id); 884 } 885 886 // Now that we know the size of all the monitors we can add a fixed slot 887 // for the original deopt pc. 888 889 _orig_pc_slot = fixed_slots(); 890 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 891 set_fixed_slots(next_slot); 892 893 // Compute when to use implicit null checks. Used by matching trap based 894 // nodes and NullCheck optimization. 895 set_allowed_deopt_reasons(); 896 897 // Now generate code 898 Code_Gen(); 899 if (failing()) return; 900 901 // Check if we want to skip execution of all compiled code. 902 { 903 #ifndef PRODUCT 904 if (OptoNoExecute) { 905 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 906 return; 907 } 908 #endif 909 TracePhase tp("install_code", &timers[_t_registerMethod]); 910 911 if (is_osr_compilation()) { 912 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 913 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 914 } else { 915 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 916 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 917 } 918 919 env()->register_method(_method, _entry_bci, 920 &_code_offsets, 921 _orig_pc_slot_offset_in_bytes, 922 code_buffer(), 923 frame_size_in_words(), _oop_map_set, 924 &_handler_table, &_inc_table, 925 compiler, 926 has_unsafe_access(), 927 SharedRuntime::is_wide_vector(max_vector_size()), 928 _directive, 929 rtm_state() 930 ); 931 932 if (log() != NULL) // Print code cache state into compiler log 933 log()->code_cache_state(); 934 } 935 } 936 937 //------------------------------Compile---------------------------------------- 938 // Compile a runtime stub 939 Compile::Compile( ciEnv* ci_env, 940 TypeFunc_generator generator, 941 address stub_function, 942 const char *stub_name, 943 int is_fancy_jump, 944 bool pass_tls, 945 bool save_arg_registers, 946 bool return_pc, 947 DirectiveSet* directive) 948 : Phase(Compiler), 949 _env(ci_env), 950 _directive(directive), 951 _log(ci_env->log()), 952 _compile_id(0), 953 _save_argument_registers(save_arg_registers), 954 _method(NULL), 955 _stub_name(stub_name), 956 _stub_function(stub_function), 957 _stub_entry_point(NULL), 958 _entry_bci(InvocationEntryBci), 959 _initial_gvn(NULL), 960 _for_igvn(NULL), 961 _warm_calls(NULL), 962 _orig_pc_slot(0), 963 _orig_pc_slot_offset_in_bytes(0), 964 _subsume_loads(true), 965 _do_escape_analysis(false), 966 _eliminate_boxing(false), 967 _failure_reason(NULL), 968 _code_buffer("Compile::Fill_buffer"), 969 _has_method_handle_invokes(false), 970 _mach_constant_base_node(NULL), 971 _node_bundling_limit(0), 972 _node_bundling_base(NULL), 973 _java_calls(0), 974 _inner_loops(0), 975 #ifndef PRODUCT 976 _trace_opto_output(TraceOptoOutput), 977 _in_dump_cnt(0), 978 _printer(NULL), 979 #endif 980 _comp_arena(mtCompiler), 981 _node_arena(mtCompiler), 982 _old_arena(mtCompiler), 983 _Compile_types(mtCompiler), 984 _dead_node_list(comp_arena()), 985 _dead_node_count(0), 986 _congraph(NULL), 987 _replay_inline_data(NULL), 988 _number_of_mh_late_inlines(0), 989 _inlining_progress(false), 990 _inlining_incrementally(false), 991 _print_inlining_list(NULL), 992 _print_inlining_stream(NULL), 993 _print_inlining_idx(0), 994 _print_inlining_output(NULL), 995 _allowed_reasons(0), 996 _interpreter_frame_size(0), 997 _max_node_limit(MaxNodeLimit) { 998 C = this; 999 1000 TraceTime t1(NULL, &_t_totalCompilation, CITime, false); 1001 TraceTime t2(NULL, &_t_stubCompilation, CITime, false); 1002 1003 #ifndef PRODUCT 1004 set_print_assembly(PrintFrameConverterAssembly); 1005 set_parsed_irreducible_loop(false); 1006 #endif 1007 set_has_irreducible_loop(false); // no loops 1008 1009 CompileWrapper cw(this); 1010 Init(/*AliasLevel=*/ 0); 1011 init_tf((*generator)()); 1012 1013 { 1014 // The following is a dummy for the sake of GraphKit::gen_stub 1015 Unique_Node_List for_igvn(comp_arena()); 1016 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1017 PhaseGVN gvn(Thread::current()->resource_area(),255); 1018 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1019 gvn.transform_no_reclaim(top()); 1020 1021 GraphKit kit; 1022 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1023 } 1024 1025 NOT_PRODUCT( verify_graph_edges(); ) 1026 Code_Gen(); 1027 if (failing()) return; 1028 1029 1030 // Entry point will be accessed using compile->stub_entry_point(); 1031 if (code_buffer() == NULL) { 1032 Matcher::soft_match_failure(); 1033 } else { 1034 if (PrintAssembly && (WizardMode || Verbose)) 1035 tty->print_cr("### Stub::%s", stub_name); 1036 1037 if (!failing()) { 1038 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1039 1040 // Make the NMethod 1041 // For now we mark the frame as never safe for profile stackwalking 1042 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1043 code_buffer(), 1044 CodeOffsets::frame_never_safe, 1045 // _code_offsets.value(CodeOffsets::Frame_Complete), 1046 frame_size_in_words(), 1047 _oop_map_set, 1048 save_arg_registers); 1049 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1050 1051 _stub_entry_point = rs->entry_point(); 1052 } 1053 } 1054 } 1055 1056 //------------------------------Init------------------------------------------- 1057 // Prepare for a single compilation 1058 void Compile::Init(int aliaslevel) { 1059 _unique = 0; 1060 _regalloc = NULL; 1061 1062 _tf = NULL; // filled in later 1063 _top = NULL; // cached later 1064 _matcher = NULL; // filled in later 1065 _cfg = NULL; // filled in later 1066 1067 set_24_bit_selection_and_mode(Use24BitFP, false); 1068 1069 _node_note_array = NULL; 1070 _default_node_notes = NULL; 1071 DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize() 1072 1073 _immutable_memory = NULL; // filled in at first inquiry 1074 1075 // Globally visible Nodes 1076 // First set TOP to NULL to give safe behavior during creation of RootNode 1077 set_cached_top_node(NULL); 1078 set_root(new RootNode()); 1079 // Now that you have a Root to point to, create the real TOP 1080 set_cached_top_node( new ConNode(Type::TOP) ); 1081 set_recent_alloc(NULL, NULL); 1082 1083 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1084 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1085 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1086 env()->set_dependencies(new Dependencies(env())); 1087 1088 _fixed_slots = 0; 1089 set_has_split_ifs(false); 1090 set_has_loops(has_method() && method()->has_loops()); // first approximation 1091 set_has_stringbuilder(false); 1092 set_has_boxed_value(false); 1093 _trap_can_recompile = false; // no traps emitted yet 1094 _major_progress = true; // start out assuming good things will happen 1095 set_has_unsafe_access(false); 1096 set_max_vector_size(0); 1097 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1098 set_decompile_count(0); 1099 1100 set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption); 1101 set_num_loop_opts(LoopOptsCount); 1102 set_do_inlining(Inline); 1103 set_max_inline_size(MaxInlineSize); 1104 set_freq_inline_size(FreqInlineSize); 1105 set_do_scheduling(OptoScheduling); 1106 set_do_count_invocations(false); 1107 set_do_method_data_update(false); 1108 1109 set_do_vector_loop(false); 1110 1111 bool do_vector = false; 1112 if (AllowVectorizeOnDemand) { 1113 if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) { 1114 set_do_vector_loop(true); 1115 } else if (has_method() && method()->name() != 0 && 1116 method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) { 1117 set_do_vector_loop(true); 1118 } 1119 #ifndef PRODUCT 1120 if (do_vector_loop() && Verbose) { 1121 tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n", method()->name()->as_quoted_ascii()); 1122 } 1123 #endif 1124 } 1125 1126 set_age_code(has_method() && method()->profile_aging()); 1127 set_rtm_state(NoRTM); // No RTM lock eliding by default 1128 _max_node_limit = _directive->MaxNodeLimitOption; 1129 1130 #if INCLUDE_RTM_OPT 1131 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1132 int rtm_state = method()->method_data()->rtm_state(); 1133 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1134 // Don't generate RTM lock eliding code. 1135 set_rtm_state(NoRTM); 1136 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1137 // Generate RTM lock eliding code without abort ratio calculation code. 1138 set_rtm_state(UseRTM); 1139 } else if (UseRTMDeopt) { 1140 // Generate RTM lock eliding code and include abort ratio calculation 1141 // code if UseRTMDeopt is on. 1142 set_rtm_state(ProfileRTM); 1143 } 1144 } 1145 #endif 1146 if (debug_info()->recording_non_safepoints()) { 1147 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1148 (comp_arena(), 8, 0, NULL)); 1149 set_default_node_notes(Node_Notes::make(this)); 1150 } 1151 1152 // // -- Initialize types before each compile -- 1153 // // Update cached type information 1154 // if( _method && _method->constants() ) 1155 // Type::update_loaded_types(_method, _method->constants()); 1156 1157 // Init alias_type map. 1158 if (!_do_escape_analysis && aliaslevel == 3) 1159 aliaslevel = 2; // No unique types without escape analysis 1160 _AliasLevel = aliaslevel; 1161 const int grow_ats = 16; 1162 _max_alias_types = grow_ats; 1163 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1164 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1165 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1166 { 1167 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1168 } 1169 // Initialize the first few types. 1170 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1171 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1172 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1173 _num_alias_types = AliasIdxRaw+1; 1174 // Zero out the alias type cache. 1175 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1176 // A NULL adr_type hits in the cache right away. Preload the right answer. 1177 probe_alias_cache(NULL)->_index = AliasIdxTop; 1178 1179 _intrinsics = NULL; 1180 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1181 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1182 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1183 register_library_intrinsics(); 1184 } 1185 1186 //---------------------------init_start---------------------------------------- 1187 // Install the StartNode on this compile object. 1188 void Compile::init_start(StartNode* s) { 1189 if (failing()) 1190 return; // already failing 1191 assert(s == start(), ""); 1192 } 1193 1194 /** 1195 * Return the 'StartNode'. We must not have a pending failure, since the ideal graph 1196 * can be in an inconsistent state, i.e., we can get segmentation faults when traversing 1197 * the ideal graph. 1198 */ 1199 StartNode* Compile::start() const { 1200 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 1201 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1202 Node* start = root()->fast_out(i); 1203 if (start->is_Start()) { 1204 return start->as_Start(); 1205 } 1206 } 1207 fatal("Did not find Start node!"); 1208 return NULL; 1209 } 1210 1211 //-------------------------------immutable_memory------------------------------------- 1212 // Access immutable memory 1213 Node* Compile::immutable_memory() { 1214 if (_immutable_memory != NULL) { 1215 return _immutable_memory; 1216 } 1217 StartNode* s = start(); 1218 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1219 Node *p = s->fast_out(i); 1220 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1221 _immutable_memory = p; 1222 return _immutable_memory; 1223 } 1224 } 1225 ShouldNotReachHere(); 1226 return NULL; 1227 } 1228 1229 //----------------------set_cached_top_node------------------------------------ 1230 // Install the cached top node, and make sure Node::is_top works correctly. 1231 void Compile::set_cached_top_node(Node* tn) { 1232 if (tn != NULL) verify_top(tn); 1233 Node* old_top = _top; 1234 _top = tn; 1235 // Calling Node::setup_is_top allows the nodes the chance to adjust 1236 // their _out arrays. 1237 if (_top != NULL) _top->setup_is_top(); 1238 if (old_top != NULL) old_top->setup_is_top(); 1239 assert(_top == NULL || top()->is_top(), ""); 1240 } 1241 1242 #ifdef ASSERT 1243 uint Compile::count_live_nodes_by_graph_walk() { 1244 Unique_Node_List useful(comp_arena()); 1245 // Get useful node list by walking the graph. 1246 identify_useful_nodes(useful); 1247 return useful.size(); 1248 } 1249 1250 void Compile::print_missing_nodes() { 1251 1252 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1253 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1254 return; 1255 } 1256 1257 // This is an expensive function. It is executed only when the user 1258 // specifies VerifyIdealNodeCount option or otherwise knows the 1259 // additional work that needs to be done to identify reachable nodes 1260 // by walking the flow graph and find the missing ones using 1261 // _dead_node_list. 1262 1263 Unique_Node_List useful(comp_arena()); 1264 // Get useful node list by walking the graph. 1265 identify_useful_nodes(useful); 1266 1267 uint l_nodes = C->live_nodes(); 1268 uint l_nodes_by_walk = useful.size(); 1269 1270 if (l_nodes != l_nodes_by_walk) { 1271 if (_log != NULL) { 1272 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1273 _log->stamp(); 1274 _log->end_head(); 1275 } 1276 VectorSet& useful_member_set = useful.member_set(); 1277 int last_idx = l_nodes_by_walk; 1278 for (int i = 0; i < last_idx; i++) { 1279 if (useful_member_set.test(i)) { 1280 if (_dead_node_list.test(i)) { 1281 if (_log != NULL) { 1282 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1283 } 1284 if (PrintIdealNodeCount) { 1285 // Print the log message to tty 1286 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1287 useful.at(i)->dump(); 1288 } 1289 } 1290 } 1291 else if (! _dead_node_list.test(i)) { 1292 if (_log != NULL) { 1293 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1294 } 1295 if (PrintIdealNodeCount) { 1296 // Print the log message to tty 1297 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1298 } 1299 } 1300 } 1301 if (_log != NULL) { 1302 _log->tail("mismatched_nodes"); 1303 } 1304 } 1305 } 1306 void Compile::record_modified_node(Node* n) { 1307 if (_modified_nodes != NULL && !_inlining_incrementally && 1308 n->outcnt() != 0 && !n->is_Con()) { 1309 _modified_nodes->push(n); 1310 } 1311 } 1312 1313 void Compile::remove_modified_node(Node* n) { 1314 if (_modified_nodes != NULL) { 1315 _modified_nodes->remove(n); 1316 } 1317 } 1318 #endif 1319 1320 #ifndef PRODUCT 1321 void Compile::verify_top(Node* tn) const { 1322 if (tn != NULL) { 1323 assert(tn->is_Con(), "top node must be a constant"); 1324 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1325 assert(tn->in(0) != NULL, "must have live top node"); 1326 } 1327 } 1328 #endif 1329 1330 1331 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1332 1333 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1334 guarantee(arr != NULL, ""); 1335 int num_blocks = arr->length(); 1336 if (grow_by < num_blocks) grow_by = num_blocks; 1337 int num_notes = grow_by * _node_notes_block_size; 1338 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1339 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1340 while (num_notes > 0) { 1341 arr->append(notes); 1342 notes += _node_notes_block_size; 1343 num_notes -= _node_notes_block_size; 1344 } 1345 assert(num_notes == 0, "exact multiple, please"); 1346 } 1347 1348 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1349 if (source == NULL || dest == NULL) return false; 1350 1351 if (dest->is_Con()) 1352 return false; // Do not push debug info onto constants. 1353 1354 #ifdef ASSERT 1355 // Leave a bread crumb trail pointing to the original node: 1356 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1357 dest->set_debug_orig(source); 1358 } 1359 #endif 1360 1361 if (node_note_array() == NULL) 1362 return false; // Not collecting any notes now. 1363 1364 // This is a copy onto a pre-existing node, which may already have notes. 1365 // If both nodes have notes, do not overwrite any pre-existing notes. 1366 Node_Notes* source_notes = node_notes_at(source->_idx); 1367 if (source_notes == NULL || source_notes->is_clear()) return false; 1368 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1369 if (dest_notes == NULL || dest_notes->is_clear()) { 1370 return set_node_notes_at(dest->_idx, source_notes); 1371 } 1372 1373 Node_Notes merged_notes = (*source_notes); 1374 // The order of operations here ensures that dest notes will win... 1375 merged_notes.update_from(dest_notes); 1376 return set_node_notes_at(dest->_idx, &merged_notes); 1377 } 1378 1379 1380 //--------------------------allow_range_check_smearing------------------------- 1381 // Gating condition for coalescing similar range checks. 1382 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1383 // single covering check that is at least as strong as any of them. 1384 // If the optimization succeeds, the simplified (strengthened) range check 1385 // will always succeed. If it fails, we will deopt, and then give up 1386 // on the optimization. 1387 bool Compile::allow_range_check_smearing() const { 1388 // If this method has already thrown a range-check, 1389 // assume it was because we already tried range smearing 1390 // and it failed. 1391 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1392 return !already_trapped; 1393 } 1394 1395 1396 //------------------------------flatten_alias_type----------------------------- 1397 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1398 int offset = tj->offset(); 1399 TypePtr::PTR ptr = tj->ptr(); 1400 1401 // Known instance (scalarizable allocation) alias only with itself. 1402 bool is_known_inst = tj->isa_oopptr() != NULL && 1403 tj->is_oopptr()->is_known_instance(); 1404 1405 // Process weird unsafe references. 1406 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1407 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1408 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1409 tj = TypeOopPtr::BOTTOM; 1410 ptr = tj->ptr(); 1411 offset = tj->offset(); 1412 } 1413 1414 // Array pointers need some flattening 1415 const TypeAryPtr *ta = tj->isa_aryptr(); 1416 if (ta && ta->is_stable()) { 1417 // Erase stability property for alias analysis. 1418 tj = ta = ta->cast_to_stable(false); 1419 } 1420 if( ta && is_known_inst ) { 1421 if ( offset != Type::OffsetBot && 1422 offset > arrayOopDesc::length_offset_in_bytes() ) { 1423 offset = Type::OffsetBot; // Flatten constant access into array body only 1424 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1425 } 1426 } else if( ta && _AliasLevel >= 2 ) { 1427 // For arrays indexed by constant indices, we flatten the alias 1428 // space to include all of the array body. Only the header, klass 1429 // and array length can be accessed un-aliased. 1430 if( offset != Type::OffsetBot ) { 1431 if( ta->const_oop() ) { // MethodData* or Method* 1432 offset = Type::OffsetBot; // Flatten constant access into array body 1433 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1434 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1435 // range is OK as-is. 1436 tj = ta = TypeAryPtr::RANGE; 1437 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1438 tj = TypeInstPtr::KLASS; // all klass loads look alike 1439 ta = TypeAryPtr::RANGE; // generic ignored junk 1440 ptr = TypePtr::BotPTR; 1441 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1442 tj = TypeInstPtr::MARK; 1443 ta = TypeAryPtr::RANGE; // generic ignored junk 1444 ptr = TypePtr::BotPTR; 1445 } else { // Random constant offset into array body 1446 offset = Type::OffsetBot; // Flatten constant access into array body 1447 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1448 } 1449 } 1450 // Arrays of fixed size alias with arrays of unknown size. 1451 if (ta->size() != TypeInt::POS) { 1452 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1453 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1454 } 1455 // Arrays of known objects become arrays of unknown objects. 1456 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1457 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1458 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1459 } 1460 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1461 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1462 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1463 } 1464 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1465 // cannot be distinguished by bytecode alone. 1466 if (ta->elem() == TypeInt::BOOL) { 1467 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1468 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1469 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1470 } 1471 // During the 2nd round of IterGVN, NotNull castings are removed. 1472 // Make sure the Bottom and NotNull variants alias the same. 1473 // Also, make sure exact and non-exact variants alias the same. 1474 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1475 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1476 } 1477 } 1478 1479 // Oop pointers need some flattening 1480 const TypeInstPtr *to = tj->isa_instptr(); 1481 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1482 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1483 if( ptr == TypePtr::Constant ) { 1484 if (to->klass() != ciEnv::current()->Class_klass() || 1485 offset < k->size_helper() * wordSize) { 1486 // No constant oop pointers (such as Strings); they alias with 1487 // unknown strings. 1488 assert(!is_known_inst, "not scalarizable allocation"); 1489 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1490 } 1491 } else if( is_known_inst ) { 1492 tj = to; // Keep NotNull and klass_is_exact for instance type 1493 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1494 // During the 2nd round of IterGVN, NotNull castings are removed. 1495 // Make sure the Bottom and NotNull variants alias the same. 1496 // Also, make sure exact and non-exact variants alias the same. 1497 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1498 } 1499 if (to->speculative() != NULL) { 1500 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1501 } 1502 // Canonicalize the holder of this field 1503 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1504 // First handle header references such as a LoadKlassNode, even if the 1505 // object's klass is unloaded at compile time (4965979). 1506 if (!is_known_inst) { // Do it only for non-instance types 1507 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1508 } 1509 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1510 // Static fields are in the space above the normal instance 1511 // fields in the java.lang.Class instance. 1512 if (to->klass() != ciEnv::current()->Class_klass()) { 1513 to = NULL; 1514 tj = TypeOopPtr::BOTTOM; 1515 offset = tj->offset(); 1516 } 1517 } else { 1518 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1519 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1520 if( is_known_inst ) { 1521 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1522 } else { 1523 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1524 } 1525 } 1526 } 1527 } 1528 1529 // Klass pointers to object array klasses need some flattening 1530 const TypeKlassPtr *tk = tj->isa_klassptr(); 1531 if( tk ) { 1532 // If we are referencing a field within a Klass, we need 1533 // to assume the worst case of an Object. Both exact and 1534 // inexact types must flatten to the same alias class so 1535 // use NotNull as the PTR. 1536 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1537 1538 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1539 TypeKlassPtr::OBJECT->klass(), 1540 offset); 1541 } 1542 1543 ciKlass* klass = tk->klass(); 1544 if( klass->is_obj_array_klass() ) { 1545 ciKlass* k = TypeAryPtr::OOPS->klass(); 1546 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1547 k = TypeInstPtr::BOTTOM->klass(); 1548 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1549 } 1550 1551 // Check for precise loads from the primary supertype array and force them 1552 // to the supertype cache alias index. Check for generic array loads from 1553 // the primary supertype array and also force them to the supertype cache 1554 // alias index. Since the same load can reach both, we need to merge 1555 // these 2 disparate memories into the same alias class. Since the 1556 // primary supertype array is read-only, there's no chance of confusion 1557 // where we bypass an array load and an array store. 1558 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1559 if (offset == Type::OffsetBot || 1560 (offset >= primary_supers_offset && 1561 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1562 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1563 offset = in_bytes(Klass::secondary_super_cache_offset()); 1564 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1565 } 1566 } 1567 1568 // Flatten all Raw pointers together. 1569 if (tj->base() == Type::RawPtr) 1570 tj = TypeRawPtr::BOTTOM; 1571 1572 if (tj->base() == Type::AnyPtr) 1573 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1574 1575 // Flatten all to bottom for now 1576 switch( _AliasLevel ) { 1577 case 0: 1578 tj = TypePtr::BOTTOM; 1579 break; 1580 case 1: // Flatten to: oop, static, field or array 1581 switch (tj->base()) { 1582 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1583 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1584 case Type::AryPtr: // do not distinguish arrays at all 1585 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1586 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1587 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1588 default: ShouldNotReachHere(); 1589 } 1590 break; 1591 case 2: // No collapsing at level 2; keep all splits 1592 case 3: // No collapsing at level 3; keep all splits 1593 break; 1594 default: 1595 Unimplemented(); 1596 } 1597 1598 offset = tj->offset(); 1599 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1600 1601 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1602 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1603 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1604 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1605 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1606 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1607 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1608 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1609 assert( tj->ptr() != TypePtr::TopPTR && 1610 tj->ptr() != TypePtr::AnyNull && 1611 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1612 // assert( tj->ptr() != TypePtr::Constant || 1613 // tj->base() == Type::RawPtr || 1614 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1615 1616 return tj; 1617 } 1618 1619 void Compile::AliasType::Init(int i, const TypePtr* at) { 1620 _index = i; 1621 _adr_type = at; 1622 _field = NULL; 1623 _element = NULL; 1624 _is_rewritable = true; // default 1625 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1626 if (atoop != NULL && atoop->is_known_instance()) { 1627 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1628 _general_index = Compile::current()->get_alias_index(gt); 1629 } else { 1630 _general_index = 0; 1631 } 1632 } 1633 1634 //---------------------------------print_on------------------------------------ 1635 #ifndef PRODUCT 1636 void Compile::AliasType::print_on(outputStream* st) { 1637 if (index() < 10) 1638 st->print("@ <%d> ", index()); 1639 else st->print("@ <%d>", index()); 1640 st->print(is_rewritable() ? " " : " RO"); 1641 int offset = adr_type()->offset(); 1642 if (offset == Type::OffsetBot) 1643 st->print(" +any"); 1644 else st->print(" +%-3d", offset); 1645 st->print(" in "); 1646 adr_type()->dump_on(st); 1647 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1648 if (field() != NULL && tjp) { 1649 if (tjp->klass() != field()->holder() || 1650 tjp->offset() != field()->offset_in_bytes()) { 1651 st->print(" != "); 1652 field()->print(); 1653 st->print(" ***"); 1654 } 1655 } 1656 } 1657 1658 void print_alias_types() { 1659 Compile* C = Compile::current(); 1660 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1661 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1662 C->alias_type(idx)->print_on(tty); 1663 tty->cr(); 1664 } 1665 } 1666 #endif 1667 1668 1669 //----------------------------probe_alias_cache-------------------------------- 1670 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1671 intptr_t key = (intptr_t) adr_type; 1672 key ^= key >> logAliasCacheSize; 1673 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1674 } 1675 1676 1677 //-----------------------------grow_alias_types-------------------------------- 1678 void Compile::grow_alias_types() { 1679 const int old_ats = _max_alias_types; // how many before? 1680 const int new_ats = old_ats; // how many more? 1681 const int grow_ats = old_ats+new_ats; // how many now? 1682 _max_alias_types = grow_ats; 1683 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1684 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1685 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1686 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1687 } 1688 1689 1690 //--------------------------------find_alias_type------------------------------ 1691 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1692 if (_AliasLevel == 0) 1693 return alias_type(AliasIdxBot); 1694 1695 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1696 if (ace->_adr_type == adr_type) { 1697 return alias_type(ace->_index); 1698 } 1699 1700 // Handle special cases. 1701 if (adr_type == NULL) return alias_type(AliasIdxTop); 1702 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1703 1704 // Do it the slow way. 1705 const TypePtr* flat = flatten_alias_type(adr_type); 1706 1707 #ifdef ASSERT 1708 assert(flat == flatten_alias_type(flat), "idempotent"); 1709 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1710 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1711 const TypeOopPtr* foop = flat->is_oopptr(); 1712 // Scalarizable allocations have exact klass always. 1713 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1714 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1715 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1716 } 1717 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1718 #endif 1719 1720 int idx = AliasIdxTop; 1721 for (int i = 0; i < num_alias_types(); i++) { 1722 if (alias_type(i)->adr_type() == flat) { 1723 idx = i; 1724 break; 1725 } 1726 } 1727 1728 if (idx == AliasIdxTop) { 1729 if (no_create) return NULL; 1730 // Grow the array if necessary. 1731 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1732 // Add a new alias type. 1733 idx = _num_alias_types++; 1734 _alias_types[idx]->Init(idx, flat); 1735 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1736 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1737 if (flat->isa_instptr()) { 1738 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1739 && flat->is_instptr()->klass() == env()->Class_klass()) 1740 alias_type(idx)->set_rewritable(false); 1741 } 1742 if (flat->isa_aryptr()) { 1743 #ifdef ASSERT 1744 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1745 // (T_BYTE has the weakest alignment and size restrictions...) 1746 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1747 #endif 1748 if (flat->offset() == TypePtr::OffsetBot) { 1749 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1750 } 1751 } 1752 if (flat->isa_klassptr()) { 1753 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1754 alias_type(idx)->set_rewritable(false); 1755 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1756 alias_type(idx)->set_rewritable(false); 1757 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1758 alias_type(idx)->set_rewritable(false); 1759 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1760 alias_type(idx)->set_rewritable(false); 1761 } 1762 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1763 // but the base pointer type is not distinctive enough to identify 1764 // references into JavaThread.) 1765 1766 // Check for final fields. 1767 const TypeInstPtr* tinst = flat->isa_instptr(); 1768 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1769 ciField* field; 1770 if (tinst->const_oop() != NULL && 1771 tinst->klass() == ciEnv::current()->Class_klass() && 1772 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1773 // static field 1774 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1775 field = k->get_field_by_offset(tinst->offset(), true); 1776 } else { 1777 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1778 field = k->get_field_by_offset(tinst->offset(), false); 1779 } 1780 assert(field == NULL || 1781 original_field == NULL || 1782 (field->holder() == original_field->holder() && 1783 field->offset() == original_field->offset() && 1784 field->is_static() == original_field->is_static()), "wrong field?"); 1785 // Set field() and is_rewritable() attributes. 1786 if (field != NULL) alias_type(idx)->set_field(field); 1787 } 1788 } 1789 1790 // Fill the cache for next time. 1791 ace->_adr_type = adr_type; 1792 ace->_index = idx; 1793 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1794 1795 // Might as well try to fill the cache for the flattened version, too. 1796 AliasCacheEntry* face = probe_alias_cache(flat); 1797 if (face->_adr_type == NULL) { 1798 face->_adr_type = flat; 1799 face->_index = idx; 1800 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1801 } 1802 1803 return alias_type(idx); 1804 } 1805 1806 1807 Compile::AliasType* Compile::alias_type(ciField* field) { 1808 const TypeOopPtr* t; 1809 if (field->is_static()) 1810 t = TypeInstPtr::make(field->holder()->java_mirror()); 1811 else 1812 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1813 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1814 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1815 return atp; 1816 } 1817 1818 1819 //------------------------------have_alias_type-------------------------------- 1820 bool Compile::have_alias_type(const TypePtr* adr_type) { 1821 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1822 if (ace->_adr_type == adr_type) { 1823 return true; 1824 } 1825 1826 // Handle special cases. 1827 if (adr_type == NULL) return true; 1828 if (adr_type == TypePtr::BOTTOM) return true; 1829 1830 return find_alias_type(adr_type, true, NULL) != NULL; 1831 } 1832 1833 //-----------------------------must_alias-------------------------------------- 1834 // True if all values of the given address type are in the given alias category. 1835 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1836 if (alias_idx == AliasIdxBot) return true; // the universal category 1837 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1838 if (alias_idx == AliasIdxTop) return false; // the empty category 1839 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1840 1841 // the only remaining possible overlap is identity 1842 int adr_idx = get_alias_index(adr_type); 1843 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1844 assert(adr_idx == alias_idx || 1845 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1846 && adr_type != TypeOopPtr::BOTTOM), 1847 "should not be testing for overlap with an unsafe pointer"); 1848 return adr_idx == alias_idx; 1849 } 1850 1851 //------------------------------can_alias-------------------------------------- 1852 // True if any values of the given address type are in the given alias category. 1853 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1854 if (alias_idx == AliasIdxTop) return false; // the empty category 1855 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1856 if (alias_idx == AliasIdxBot) return true; // the universal category 1857 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1858 1859 // the only remaining possible overlap is identity 1860 int adr_idx = get_alias_index(adr_type); 1861 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1862 return adr_idx == alias_idx; 1863 } 1864 1865 1866 1867 //---------------------------pop_warm_call------------------------------------- 1868 WarmCallInfo* Compile::pop_warm_call() { 1869 WarmCallInfo* wci = _warm_calls; 1870 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1871 return wci; 1872 } 1873 1874 //----------------------------Inline_Warm-------------------------------------- 1875 int Compile::Inline_Warm() { 1876 // If there is room, try to inline some more warm call sites. 1877 // %%% Do a graph index compaction pass when we think we're out of space? 1878 if (!InlineWarmCalls) return 0; 1879 1880 int calls_made_hot = 0; 1881 int room_to_grow = NodeCountInliningCutoff - unique(); 1882 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1883 int amount_grown = 0; 1884 WarmCallInfo* call; 1885 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1886 int est_size = (int)call->size(); 1887 if (est_size > (room_to_grow - amount_grown)) { 1888 // This one won't fit anyway. Get rid of it. 1889 call->make_cold(); 1890 continue; 1891 } 1892 call->make_hot(); 1893 calls_made_hot++; 1894 amount_grown += est_size; 1895 amount_to_grow -= est_size; 1896 } 1897 1898 if (calls_made_hot > 0) set_major_progress(); 1899 return calls_made_hot; 1900 } 1901 1902 1903 //----------------------------Finish_Warm-------------------------------------- 1904 void Compile::Finish_Warm() { 1905 if (!InlineWarmCalls) return; 1906 if (failing()) return; 1907 if (warm_calls() == NULL) return; 1908 1909 // Clean up loose ends, if we are out of space for inlining. 1910 WarmCallInfo* call; 1911 while ((call = pop_warm_call()) != NULL) { 1912 call->make_cold(); 1913 } 1914 } 1915 1916 //---------------------cleanup_loop_predicates----------------------- 1917 // Remove the opaque nodes that protect the predicates so that all unused 1918 // checks and uncommon_traps will be eliminated from the ideal graph 1919 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1920 if (predicate_count()==0) return; 1921 for (int i = predicate_count(); i > 0; i--) { 1922 Node * n = predicate_opaque1_node(i-1); 1923 assert(n->Opcode() == Op_Opaque1, "must be"); 1924 igvn.replace_node(n, n->in(1)); 1925 } 1926 assert(predicate_count()==0, "should be clean!"); 1927 } 1928 1929 // StringOpts and late inlining of string methods 1930 void Compile::inline_string_calls(bool parse_time) { 1931 { 1932 // remove useless nodes to make the usage analysis simpler 1933 ResourceMark rm; 1934 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1935 } 1936 1937 { 1938 ResourceMark rm; 1939 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1940 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1941 print_method(PHASE_AFTER_STRINGOPTS, 3); 1942 } 1943 1944 // now inline anything that we skipped the first time around 1945 if (!parse_time) { 1946 _late_inlines_pos = _late_inlines.length(); 1947 } 1948 1949 while (_string_late_inlines.length() > 0) { 1950 CallGenerator* cg = _string_late_inlines.pop(); 1951 cg->do_late_inline(); 1952 if (failing()) return; 1953 } 1954 _string_late_inlines.trunc_to(0); 1955 } 1956 1957 // Late inlining of boxing methods 1958 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1959 if (_boxing_late_inlines.length() > 0) { 1960 assert(has_boxed_value(), "inconsistent"); 1961 1962 PhaseGVN* gvn = initial_gvn(); 1963 set_inlining_incrementally(true); 1964 1965 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1966 for_igvn()->clear(); 1967 gvn->replace_with(&igvn); 1968 1969 _late_inlines_pos = _late_inlines.length(); 1970 1971 while (_boxing_late_inlines.length() > 0) { 1972 CallGenerator* cg = _boxing_late_inlines.pop(); 1973 cg->do_late_inline(); 1974 if (failing()) return; 1975 } 1976 _boxing_late_inlines.trunc_to(0); 1977 1978 { 1979 ResourceMark rm; 1980 PhaseRemoveUseless pru(gvn, for_igvn()); 1981 } 1982 1983 igvn = PhaseIterGVN(gvn); 1984 igvn.optimize(); 1985 1986 set_inlining_progress(false); 1987 set_inlining_incrementally(false); 1988 } 1989 } 1990 1991 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1992 assert(IncrementalInline, "incremental inlining should be on"); 1993 PhaseGVN* gvn = initial_gvn(); 1994 1995 set_inlining_progress(false); 1996 for_igvn()->clear(); 1997 gvn->replace_with(&igvn); 1998 1999 { 2000 TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]); 2001 int i = 0; 2002 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2003 CallGenerator* cg = _late_inlines.at(i); 2004 _late_inlines_pos = i+1; 2005 cg->do_late_inline(); 2006 if (failing()) return; 2007 } 2008 int j = 0; 2009 for (; i < _late_inlines.length(); i++, j++) { 2010 _late_inlines.at_put(j, _late_inlines.at(i)); 2011 } 2012 _late_inlines.trunc_to(j); 2013 } 2014 2015 { 2016 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2017 ResourceMark rm; 2018 PhaseRemoveUseless pru(gvn, for_igvn()); 2019 } 2020 2021 { 2022 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2023 igvn = PhaseIterGVN(gvn); 2024 } 2025 } 2026 2027 // Perform incremental inlining until bound on number of live nodes is reached 2028 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2029 TracePhase tp("incrementalInline", &timers[_t_incrInline]); 2030 2031 PhaseGVN* gvn = initial_gvn(); 2032 2033 set_inlining_incrementally(true); 2034 set_inlining_progress(true); 2035 uint low_live_nodes = 0; 2036 2037 while(inlining_progress() && _late_inlines.length() > 0) { 2038 2039 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2040 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2041 TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]); 2042 // PhaseIdealLoop is expensive so we only try it once we are 2043 // out of live nodes and we only try it again if the previous 2044 // helped got the number of nodes down significantly 2045 PhaseIdealLoop ideal_loop( igvn, false, true ); 2046 if (failing()) return; 2047 low_live_nodes = live_nodes(); 2048 _major_progress = true; 2049 } 2050 2051 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2052 break; 2053 } 2054 } 2055 2056 inline_incrementally_one(igvn); 2057 2058 if (failing()) return; 2059 2060 { 2061 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2062 igvn.optimize(); 2063 } 2064 2065 if (failing()) return; 2066 } 2067 2068 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2069 2070 if (_string_late_inlines.length() > 0) { 2071 assert(has_stringbuilder(), "inconsistent"); 2072 for_igvn()->clear(); 2073 initial_gvn()->replace_with(&igvn); 2074 2075 inline_string_calls(false); 2076 2077 if (failing()) return; 2078 2079 { 2080 TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]); 2081 ResourceMark rm; 2082 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2083 } 2084 2085 { 2086 TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]); 2087 igvn = PhaseIterGVN(gvn); 2088 igvn.optimize(); 2089 } 2090 } 2091 2092 set_inlining_incrementally(false); 2093 } 2094 2095 2096 //------------------------------Optimize--------------------------------------- 2097 // Given a graph, optimize it. 2098 void Compile::Optimize() { 2099 TracePhase tp("optimizer", &timers[_t_optimizer]); 2100 2101 #ifndef PRODUCT 2102 if (_directive->BreakAtCompileOption) { 2103 BREAKPOINT; 2104 } 2105 2106 #endif 2107 2108 ResourceMark rm; 2109 int loop_opts_cnt; 2110 2111 print_inlining_reinit(); 2112 2113 NOT_PRODUCT( verify_graph_edges(); ) 2114 2115 print_method(PHASE_AFTER_PARSING); 2116 2117 { 2118 // Iterative Global Value Numbering, including ideal transforms 2119 // Initialize IterGVN with types and values from parse-time GVN 2120 PhaseIterGVN igvn(initial_gvn()); 2121 #ifdef ASSERT 2122 _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena()); 2123 #endif 2124 { 2125 TracePhase tp("iterGVN", &timers[_t_iterGVN]); 2126 igvn.optimize(); 2127 } 2128 2129 print_method(PHASE_ITER_GVN1, 2); 2130 2131 if (failing()) return; 2132 2133 inline_incrementally(igvn); 2134 2135 print_method(PHASE_INCREMENTAL_INLINE, 2); 2136 2137 if (failing()) return; 2138 2139 if (eliminate_boxing()) { 2140 // Inline valueOf() methods now. 2141 inline_boxing_calls(igvn); 2142 2143 if (AlwaysIncrementalInline) { 2144 inline_incrementally(igvn); 2145 } 2146 2147 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2148 2149 if (failing()) return; 2150 } 2151 2152 // Remove the speculative part of types and clean up the graph from 2153 // the extra CastPP nodes whose only purpose is to carry them. Do 2154 // that early so that optimizations are not disrupted by the extra 2155 // CastPP nodes. 2156 remove_speculative_types(igvn); 2157 2158 // No more new expensive nodes will be added to the list from here 2159 // so keep only the actual candidates for optimizations. 2160 cleanup_expensive_nodes(igvn); 2161 2162 // Perform escape analysis 2163 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2164 if (has_loops()) { 2165 // Cleanup graph (remove dead nodes). 2166 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2167 PhaseIdealLoop ideal_loop( igvn, false, true ); 2168 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2169 if (failing()) return; 2170 } 2171 ConnectionGraph::do_analysis(this, &igvn); 2172 2173 if (failing()) return; 2174 2175 // Optimize out fields loads from scalar replaceable allocations. 2176 igvn.optimize(); 2177 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2178 2179 if (failing()) return; 2180 2181 if (congraph() != NULL && macro_count() > 0) { 2182 TracePhase tp("macroEliminate", &timers[_t_macroEliminate]); 2183 PhaseMacroExpand mexp(igvn); 2184 mexp.eliminate_macro_nodes(); 2185 igvn.set_delay_transform(false); 2186 2187 igvn.optimize(); 2188 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2189 2190 if (failing()) return; 2191 } 2192 } 2193 2194 // Loop transforms on the ideal graph. Range Check Elimination, 2195 // peeling, unrolling, etc. 2196 2197 // Set loop opts counter 2198 loop_opts_cnt = num_loop_opts(); 2199 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2200 { 2201 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2202 PhaseIdealLoop ideal_loop( igvn, true ); 2203 loop_opts_cnt--; 2204 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2205 if (failing()) return; 2206 } 2207 // Loop opts pass if partial peeling occurred in previous pass 2208 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2209 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2210 PhaseIdealLoop ideal_loop( igvn, false ); 2211 loop_opts_cnt--; 2212 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2213 if (failing()) return; 2214 } 2215 // Loop opts pass for loop-unrolling before CCP 2216 if(major_progress() && (loop_opts_cnt > 0)) { 2217 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2218 PhaseIdealLoop ideal_loop( igvn, false ); 2219 loop_opts_cnt--; 2220 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2221 } 2222 if (!failing()) { 2223 // Verify that last round of loop opts produced a valid graph 2224 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2225 PhaseIdealLoop::verify(igvn); 2226 } 2227 } 2228 if (failing()) return; 2229 2230 // Conditional Constant Propagation; 2231 PhaseCCP ccp( &igvn ); 2232 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2233 { 2234 TracePhase tp("ccp", &timers[_t_ccp]); 2235 ccp.do_transform(); 2236 } 2237 print_method(PHASE_CPP1, 2); 2238 2239 assert( true, "Break here to ccp.dump_old2new_map()"); 2240 2241 // Iterative Global Value Numbering, including ideal transforms 2242 { 2243 TracePhase tp("iterGVN2", &timers[_t_iterGVN2]); 2244 igvn = ccp; 2245 igvn.optimize(); 2246 } 2247 2248 print_method(PHASE_ITER_GVN2, 2); 2249 2250 if (failing()) return; 2251 2252 // Loop transforms on the ideal graph. Range Check Elimination, 2253 // peeling, unrolling, etc. 2254 if(loop_opts_cnt > 0) { 2255 debug_only( int cnt = 0; ); 2256 while(major_progress() && (loop_opts_cnt > 0)) { 2257 TracePhase tp("idealLoop", &timers[_t_idealLoop]); 2258 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2259 PhaseIdealLoop ideal_loop( igvn, true); 2260 loop_opts_cnt--; 2261 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2262 if (failing()) return; 2263 } 2264 } 2265 2266 { 2267 // Verify that all previous optimizations produced a valid graph 2268 // at least to this point, even if no loop optimizations were done. 2269 TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]); 2270 PhaseIdealLoop::verify(igvn); 2271 } 2272 2273 { 2274 TracePhase tp("macroExpand", &timers[_t_macroExpand]); 2275 PhaseMacroExpand mex(igvn); 2276 if (mex.expand_macro_nodes()) { 2277 assert(failing(), "must bail out w/ explicit message"); 2278 return; 2279 } 2280 } 2281 2282 DEBUG_ONLY( _modified_nodes = NULL; ) 2283 } // (End scope of igvn; run destructor if necessary for asserts.) 2284 2285 process_print_inlining(); 2286 // A method with only infinite loops has no edges entering loops from root 2287 { 2288 TracePhase tp("graphReshape", &timers[_t_graphReshaping]); 2289 if (final_graph_reshaping()) { 2290 assert(failing(), "must bail out w/ explicit message"); 2291 return; 2292 } 2293 } 2294 2295 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2296 } 2297 2298 2299 //------------------------------Code_Gen--------------------------------------- 2300 // Given a graph, generate code for it 2301 void Compile::Code_Gen() { 2302 if (failing()) { 2303 return; 2304 } 2305 2306 // Perform instruction selection. You might think we could reclaim Matcher 2307 // memory PDQ, but actually the Matcher is used in generating spill code. 2308 // Internals of the Matcher (including some VectorSets) must remain live 2309 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2310 // set a bit in reclaimed memory. 2311 2312 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2313 // nodes. Mapping is only valid at the root of each matched subtree. 2314 NOT_PRODUCT( verify_graph_edges(); ) 2315 2316 Matcher matcher; 2317 _matcher = &matcher; 2318 { 2319 TracePhase tp("matcher", &timers[_t_matcher]); 2320 matcher.match(); 2321 } 2322 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2323 // nodes. Mapping is only valid at the root of each matched subtree. 2324 NOT_PRODUCT( verify_graph_edges(); ) 2325 2326 // If you have too many nodes, or if matching has failed, bail out 2327 check_node_count(0, "out of nodes matching instructions"); 2328 if (failing()) { 2329 return; 2330 } 2331 2332 // Build a proper-looking CFG 2333 PhaseCFG cfg(node_arena(), root(), matcher); 2334 _cfg = &cfg; 2335 { 2336 TracePhase tp("scheduler", &timers[_t_scheduler]); 2337 bool success = cfg.do_global_code_motion(); 2338 if (!success) { 2339 return; 2340 } 2341 2342 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2343 NOT_PRODUCT( verify_graph_edges(); ) 2344 debug_only( cfg.verify(); ) 2345 } 2346 2347 PhaseChaitin regalloc(unique(), cfg, matcher, false); 2348 _regalloc = ®alloc; 2349 { 2350 TracePhase tp("regalloc", &timers[_t_registerAllocation]); 2351 // Perform register allocation. After Chaitin, use-def chains are 2352 // no longer accurate (at spill code) and so must be ignored. 2353 // Node->LRG->reg mappings are still accurate. 2354 _regalloc->Register_Allocate(); 2355 2356 // Bail out if the allocator builds too many nodes 2357 if (failing()) { 2358 return; 2359 } 2360 } 2361 2362 // Prior to register allocation we kept empty basic blocks in case the 2363 // the allocator needed a place to spill. After register allocation we 2364 // are not adding any new instructions. If any basic block is empty, we 2365 // can now safely remove it. 2366 { 2367 TracePhase tp("blockOrdering", &timers[_t_blockOrdering]); 2368 cfg.remove_empty_blocks(); 2369 if (do_freq_based_layout()) { 2370 PhaseBlockLayout layout(cfg); 2371 } else { 2372 cfg.set_loop_alignment(); 2373 } 2374 cfg.fixup_flow(); 2375 } 2376 2377 // Apply peephole optimizations 2378 if( OptoPeephole ) { 2379 TracePhase tp("peephole", &timers[_t_peephole]); 2380 PhasePeephole peep( _regalloc, cfg); 2381 peep.do_transform(); 2382 } 2383 2384 // Do late expand if CPU requires this. 2385 if (Matcher::require_postalloc_expand) { 2386 TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]); 2387 cfg.postalloc_expand(_regalloc); 2388 } 2389 2390 // Convert Nodes to instruction bits in a buffer 2391 { 2392 TraceTime tp("output", &timers[_t_output], CITime); 2393 Output(); 2394 } 2395 2396 print_method(PHASE_FINAL_CODE); 2397 2398 // He's dead, Jim. 2399 _cfg = (PhaseCFG*)0xdeadbeef; 2400 _regalloc = (PhaseChaitin*)0xdeadbeef; 2401 } 2402 2403 2404 //------------------------------dump_asm--------------------------------------- 2405 // Dump formatted assembly 2406 #ifndef PRODUCT 2407 void Compile::dump_asm(int *pcs, uint pc_limit) { 2408 bool cut_short = false; 2409 tty->print_cr("#"); 2410 tty->print("# "); _tf->dump(); tty->cr(); 2411 tty->print_cr("#"); 2412 2413 // For all blocks 2414 int pc = 0x0; // Program counter 2415 char starts_bundle = ' '; 2416 _regalloc->dump_frame(); 2417 2418 Node *n = NULL; 2419 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2420 if (VMThread::should_terminate()) { 2421 cut_short = true; 2422 break; 2423 } 2424 Block* block = _cfg->get_block(i); 2425 if (block->is_connector() && !Verbose) { 2426 continue; 2427 } 2428 n = block->head(); 2429 if (pcs && n->_idx < pc_limit) { 2430 tty->print("%3.3x ", pcs[n->_idx]); 2431 } else { 2432 tty->print(" "); 2433 } 2434 block->dump_head(_cfg); 2435 if (block->is_connector()) { 2436 tty->print_cr(" # Empty connector block"); 2437 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2438 tty->print_cr(" # Block is sole successor of call"); 2439 } 2440 2441 // For all instructions 2442 Node *delay = NULL; 2443 for (uint j = 0; j < block->number_of_nodes(); j++) { 2444 if (VMThread::should_terminate()) { 2445 cut_short = true; 2446 break; 2447 } 2448 n = block->get_node(j); 2449 if (valid_bundle_info(n)) { 2450 Bundle* bundle = node_bundling(n); 2451 if (bundle->used_in_unconditional_delay()) { 2452 delay = n; 2453 continue; 2454 } 2455 if (bundle->starts_bundle()) { 2456 starts_bundle = '+'; 2457 } 2458 } 2459 2460 if (WizardMode) { 2461 n->dump(); 2462 } 2463 2464 if( !n->is_Region() && // Dont print in the Assembly 2465 !n->is_Phi() && // a few noisely useless nodes 2466 !n->is_Proj() && 2467 !n->is_MachTemp() && 2468 !n->is_SafePointScalarObject() && 2469 !n->is_Catch() && // Would be nice to print exception table targets 2470 !n->is_MergeMem() && // Not very interesting 2471 !n->is_top() && // Debug info table constants 2472 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2473 ) { 2474 if (pcs && n->_idx < pc_limit) 2475 tty->print("%3.3x", pcs[n->_idx]); 2476 else 2477 tty->print(" "); 2478 tty->print(" %c ", starts_bundle); 2479 starts_bundle = ' '; 2480 tty->print("\t"); 2481 n->format(_regalloc, tty); 2482 tty->cr(); 2483 } 2484 2485 // If we have an instruction with a delay slot, and have seen a delay, 2486 // then back up and print it 2487 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2488 assert(delay != NULL, "no unconditional delay instruction"); 2489 if (WizardMode) delay->dump(); 2490 2491 if (node_bundling(delay)->starts_bundle()) 2492 starts_bundle = '+'; 2493 if (pcs && n->_idx < pc_limit) 2494 tty->print("%3.3x", pcs[n->_idx]); 2495 else 2496 tty->print(" "); 2497 tty->print(" %c ", starts_bundle); 2498 starts_bundle = ' '; 2499 tty->print("\t"); 2500 delay->format(_regalloc, tty); 2501 tty->cr(); 2502 delay = NULL; 2503 } 2504 2505 // Dump the exception table as well 2506 if( n->is_Catch() && (Verbose || WizardMode) ) { 2507 // Print the exception table for this offset 2508 _handler_table.print_subtable_for(pc); 2509 } 2510 } 2511 2512 if (pcs && n->_idx < pc_limit) 2513 tty->print_cr("%3.3x", pcs[n->_idx]); 2514 else 2515 tty->cr(); 2516 2517 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2518 2519 } // End of per-block dump 2520 tty->cr(); 2521 2522 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2523 } 2524 #endif 2525 2526 //------------------------------Final_Reshape_Counts--------------------------- 2527 // This class defines counters to help identify when a method 2528 // may/must be executed using hardware with only 24-bit precision. 2529 struct Final_Reshape_Counts : public StackObj { 2530 int _call_count; // count non-inlined 'common' calls 2531 int _float_count; // count float ops requiring 24-bit precision 2532 int _double_count; // count double ops requiring more precision 2533 int _java_call_count; // count non-inlined 'java' calls 2534 int _inner_loop_count; // count loops which need alignment 2535 VectorSet _visited; // Visitation flags 2536 Node_List _tests; // Set of IfNodes & PCTableNodes 2537 2538 Final_Reshape_Counts() : 2539 _call_count(0), _float_count(0), _double_count(0), 2540 _java_call_count(0), _inner_loop_count(0), 2541 _visited( Thread::current()->resource_area() ) { } 2542 2543 void inc_call_count () { _call_count ++; } 2544 void inc_float_count () { _float_count ++; } 2545 void inc_double_count() { _double_count++; } 2546 void inc_java_call_count() { _java_call_count++; } 2547 void inc_inner_loop_count() { _inner_loop_count++; } 2548 2549 int get_call_count () const { return _call_count ; } 2550 int get_float_count () const { return _float_count ; } 2551 int get_double_count() const { return _double_count; } 2552 int get_java_call_count() const { return _java_call_count; } 2553 int get_inner_loop_count() const { return _inner_loop_count; } 2554 }; 2555 2556 #ifdef ASSERT 2557 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2558 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2559 // Make sure the offset goes inside the instance layout. 2560 return k->contains_field_offset(tp->offset()); 2561 // Note that OffsetBot and OffsetTop are very negative. 2562 } 2563 #endif 2564 2565 // Eliminate trivially redundant StoreCMs and accumulate their 2566 // precedence edges. 2567 void Compile::eliminate_redundant_card_marks(Node* n) { 2568 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2569 if (n->in(MemNode::Address)->outcnt() > 1) { 2570 // There are multiple users of the same address so it might be 2571 // possible to eliminate some of the StoreCMs 2572 Node* mem = n->in(MemNode::Memory); 2573 Node* adr = n->in(MemNode::Address); 2574 Node* val = n->in(MemNode::ValueIn); 2575 Node* prev = n; 2576 bool done = false; 2577 // Walk the chain of StoreCMs eliminating ones that match. As 2578 // long as it's a chain of single users then the optimization is 2579 // safe. Eliminating partially redundant StoreCMs would require 2580 // cloning copies down the other paths. 2581 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2582 if (adr == mem->in(MemNode::Address) && 2583 val == mem->in(MemNode::ValueIn)) { 2584 // redundant StoreCM 2585 if (mem->req() > MemNode::OopStore) { 2586 // Hasn't been processed by this code yet. 2587 n->add_prec(mem->in(MemNode::OopStore)); 2588 } else { 2589 // Already converted to precedence edge 2590 for (uint i = mem->req(); i < mem->len(); i++) { 2591 // Accumulate any precedence edges 2592 if (mem->in(i) != NULL) { 2593 n->add_prec(mem->in(i)); 2594 } 2595 } 2596 // Everything above this point has been processed. 2597 done = true; 2598 } 2599 // Eliminate the previous StoreCM 2600 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2601 assert(mem->outcnt() == 0, "should be dead"); 2602 mem->disconnect_inputs(NULL, this); 2603 } else { 2604 prev = mem; 2605 } 2606 mem = prev->in(MemNode::Memory); 2607 } 2608 } 2609 } 2610 2611 //------------------------------final_graph_reshaping_impl---------------------- 2612 // Implement items 1-5 from final_graph_reshaping below. 2613 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2614 2615 if ( n->outcnt() == 0 ) return; // dead node 2616 uint nop = n->Opcode(); 2617 2618 // Check for 2-input instruction with "last use" on right input. 2619 // Swap to left input. Implements item (2). 2620 if( n->req() == 3 && // two-input instruction 2621 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2622 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2623 n->in(2)->outcnt() == 1 &&// right use IS a last use 2624 !n->in(2)->is_Con() ) { // right use is not a constant 2625 // Check for commutative opcode 2626 switch( nop ) { 2627 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2628 case Op_MaxI: case Op_MinI: 2629 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2630 case Op_AndL: case Op_XorL: case Op_OrL: 2631 case Op_AndI: case Op_XorI: case Op_OrI: { 2632 // Move "last use" input to left by swapping inputs 2633 n->swap_edges(1, 2); 2634 break; 2635 } 2636 default: 2637 break; 2638 } 2639 } 2640 2641 #ifdef ASSERT 2642 if( n->is_Mem() ) { 2643 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2644 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2645 // oop will be recorded in oop map if load crosses safepoint 2646 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2647 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2648 "raw memory operations should have control edge"); 2649 } 2650 #endif 2651 // Count FPU ops and common calls, implements item (3) 2652 switch( nop ) { 2653 // Count all float operations that may use FPU 2654 case Op_AddF: 2655 case Op_SubF: 2656 case Op_MulF: 2657 case Op_DivF: 2658 case Op_NegF: 2659 case Op_ModF: 2660 case Op_ConvI2F: 2661 case Op_ConF: 2662 case Op_CmpF: 2663 case Op_CmpF3: 2664 // case Op_ConvL2F: // longs are split into 32-bit halves 2665 frc.inc_float_count(); 2666 break; 2667 2668 case Op_ConvF2D: 2669 case Op_ConvD2F: 2670 frc.inc_float_count(); 2671 frc.inc_double_count(); 2672 break; 2673 2674 // Count all double operations that may use FPU 2675 case Op_AddD: 2676 case Op_SubD: 2677 case Op_MulD: 2678 case Op_DivD: 2679 case Op_NegD: 2680 case Op_ModD: 2681 case Op_ConvI2D: 2682 case Op_ConvD2I: 2683 // case Op_ConvL2D: // handled by leaf call 2684 // case Op_ConvD2L: // handled by leaf call 2685 case Op_ConD: 2686 case Op_CmpD: 2687 case Op_CmpD3: 2688 frc.inc_double_count(); 2689 break; 2690 case Op_Opaque1: // Remove Opaque Nodes before matching 2691 case Op_Opaque2: // Remove Opaque Nodes before matching 2692 case Op_Opaque3: 2693 n->subsume_by(n->in(1), this); 2694 break; 2695 case Op_CallStaticJava: 2696 case Op_CallJava: 2697 case Op_CallDynamicJava: 2698 frc.inc_java_call_count(); // Count java call site; 2699 case Op_CallRuntime: 2700 case Op_CallLeaf: 2701 case Op_CallLeafNoFP: { 2702 assert( n->is_Call(), "" ); 2703 CallNode *call = n->as_Call(); 2704 // Count call sites where the FP mode bit would have to be flipped. 2705 // Do not count uncommon runtime calls: 2706 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2707 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2708 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2709 frc.inc_call_count(); // Count the call site 2710 } else { // See if uncommon argument is shared 2711 Node *n = call->in(TypeFunc::Parms); 2712 int nop = n->Opcode(); 2713 // Clone shared simple arguments to uncommon calls, item (1). 2714 if( n->outcnt() > 1 && 2715 !n->is_Proj() && 2716 nop != Op_CreateEx && 2717 nop != Op_CheckCastPP && 2718 nop != Op_DecodeN && 2719 nop != Op_DecodeNKlass && 2720 !n->is_Mem() ) { 2721 Node *x = n->clone(); 2722 call->set_req( TypeFunc::Parms, x ); 2723 } 2724 } 2725 break; 2726 } 2727 2728 case Op_StoreD: 2729 case Op_LoadD: 2730 case Op_LoadD_unaligned: 2731 frc.inc_double_count(); 2732 goto handle_mem; 2733 case Op_StoreF: 2734 case Op_LoadF: 2735 frc.inc_float_count(); 2736 goto handle_mem; 2737 2738 case Op_StoreCM: 2739 { 2740 // Convert OopStore dependence into precedence edge 2741 Node* prec = n->in(MemNode::OopStore); 2742 n->del_req(MemNode::OopStore); 2743 n->add_prec(prec); 2744 eliminate_redundant_card_marks(n); 2745 } 2746 2747 // fall through 2748 2749 case Op_StoreB: 2750 case Op_StoreC: 2751 case Op_StorePConditional: 2752 case Op_StoreI: 2753 case Op_StoreL: 2754 case Op_StoreIConditional: 2755 case Op_StoreLConditional: 2756 case Op_CompareAndSwapI: 2757 case Op_CompareAndSwapL: 2758 case Op_CompareAndSwapP: 2759 case Op_CompareAndSwapN: 2760 case Op_GetAndAddI: 2761 case Op_GetAndAddL: 2762 case Op_GetAndSetI: 2763 case Op_GetAndSetL: 2764 case Op_GetAndSetP: 2765 case Op_GetAndSetN: 2766 case Op_StoreP: 2767 case Op_StoreN: 2768 case Op_StoreNKlass: 2769 case Op_LoadB: 2770 case Op_LoadUB: 2771 case Op_LoadUS: 2772 case Op_LoadI: 2773 case Op_LoadKlass: 2774 case Op_LoadNKlass: 2775 case Op_LoadL: 2776 case Op_LoadL_unaligned: 2777 case Op_LoadPLocked: 2778 case Op_LoadP: 2779 case Op_LoadN: 2780 case Op_LoadRange: 2781 case Op_LoadS: { 2782 handle_mem: 2783 #ifdef ASSERT 2784 if( VerifyOptoOopOffsets ) { 2785 assert( n->is_Mem(), "" ); 2786 MemNode *mem = (MemNode*)n; 2787 // Check to see if address types have grounded out somehow. 2788 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2789 assert( !tp || oop_offset_is_sane(tp), "" ); 2790 } 2791 #endif 2792 break; 2793 } 2794 2795 case Op_AddP: { // Assert sane base pointers 2796 Node *addp = n->in(AddPNode::Address); 2797 assert( !addp->is_AddP() || 2798 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2799 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2800 "Base pointers must match" ); 2801 #ifdef _LP64 2802 if ((UseCompressedOops || UseCompressedClassPointers) && 2803 addp->Opcode() == Op_ConP && 2804 addp == n->in(AddPNode::Base) && 2805 n->in(AddPNode::Offset)->is_Con()) { 2806 // Use addressing with narrow klass to load with offset on x86. 2807 // On sparc loading 32-bits constant and decoding it have less 2808 // instructions (4) then load 64-bits constant (7). 2809 // Do this transformation here since IGVN will convert ConN back to ConP. 2810 const Type* t = addp->bottom_type(); 2811 if (t->isa_oopptr() || t->isa_klassptr()) { 2812 Node* nn = NULL; 2813 2814 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2815 2816 // Look for existing ConN node of the same exact type. 2817 Node* r = root(); 2818 uint cnt = r->outcnt(); 2819 for (uint i = 0; i < cnt; i++) { 2820 Node* m = r->raw_out(i); 2821 if (m!= NULL && m->Opcode() == op && 2822 m->bottom_type()->make_ptr() == t) { 2823 nn = m; 2824 break; 2825 } 2826 } 2827 if (nn != NULL) { 2828 // Decode a narrow oop to match address 2829 // [R12 + narrow_oop_reg<<3 + offset] 2830 if (t->isa_oopptr()) { 2831 nn = new DecodeNNode(nn, t); 2832 } else { 2833 nn = new DecodeNKlassNode(nn, t); 2834 } 2835 n->set_req(AddPNode::Base, nn); 2836 n->set_req(AddPNode::Address, nn); 2837 if (addp->outcnt() == 0) { 2838 addp->disconnect_inputs(NULL, this); 2839 } 2840 } 2841 } 2842 } 2843 #endif 2844 break; 2845 } 2846 2847 case Op_CastPP: { 2848 // Remove CastPP nodes to gain more freedom during scheduling but 2849 // keep the dependency they encode as control or precedence edges 2850 // (if control is set already) on memory operations. Some CastPP 2851 // nodes don't have a control (don't carry a dependency): skip 2852 // those. 2853 if (n->in(0) != NULL) { 2854 ResourceMark rm; 2855 Unique_Node_List wq; 2856 wq.push(n); 2857 for (uint next = 0; next < wq.size(); ++next) { 2858 Node *m = wq.at(next); 2859 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2860 Node* use = m->fast_out(i); 2861 if (use->is_Mem() || use->is_EncodeNarrowPtr()) { 2862 use->ensure_control_or_add_prec(n->in(0)); 2863 } else if (use->in(0) == NULL) { 2864 switch(use->Opcode()) { 2865 case Op_AddP: 2866 case Op_DecodeN: 2867 case Op_DecodeNKlass: 2868 case Op_CheckCastPP: 2869 case Op_CastPP: 2870 wq.push(use); 2871 break; 2872 } 2873 } 2874 } 2875 } 2876 } 2877 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2878 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2879 Node* in1 = n->in(1); 2880 const Type* t = n->bottom_type(); 2881 Node* new_in1 = in1->clone(); 2882 new_in1->as_DecodeN()->set_type(t); 2883 2884 if (!Matcher::narrow_oop_use_complex_address()) { 2885 // 2886 // x86, ARM and friends can handle 2 adds in addressing mode 2887 // and Matcher can fold a DecodeN node into address by using 2888 // a narrow oop directly and do implicit NULL check in address: 2889 // 2890 // [R12 + narrow_oop_reg<<3 + offset] 2891 // NullCheck narrow_oop_reg 2892 // 2893 // On other platforms (Sparc) we have to keep new DecodeN node and 2894 // use it to do implicit NULL check in address: 2895 // 2896 // decode_not_null narrow_oop_reg, base_reg 2897 // [base_reg + offset] 2898 // NullCheck base_reg 2899 // 2900 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2901 // to keep the information to which NULL check the new DecodeN node 2902 // corresponds to use it as value in implicit_null_check(). 2903 // 2904 new_in1->set_req(0, n->in(0)); 2905 } 2906 2907 n->subsume_by(new_in1, this); 2908 if (in1->outcnt() == 0) { 2909 in1->disconnect_inputs(NULL, this); 2910 } 2911 } else { 2912 n->subsume_by(n->in(1), this); 2913 if (n->outcnt() == 0) { 2914 n->disconnect_inputs(NULL, this); 2915 } 2916 } 2917 break; 2918 } 2919 #ifdef _LP64 2920 case Op_CmpP: 2921 // Do this transformation here to preserve CmpPNode::sub() and 2922 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2923 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2924 Node* in1 = n->in(1); 2925 Node* in2 = n->in(2); 2926 if (!in1->is_DecodeNarrowPtr()) { 2927 in2 = in1; 2928 in1 = n->in(2); 2929 } 2930 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2931 2932 Node* new_in2 = NULL; 2933 if (in2->is_DecodeNarrowPtr()) { 2934 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2935 new_in2 = in2->in(1); 2936 } else if (in2->Opcode() == Op_ConP) { 2937 const Type* t = in2->bottom_type(); 2938 if (t == TypePtr::NULL_PTR) { 2939 assert(in1->is_DecodeN(), "compare klass to null?"); 2940 // Don't convert CmpP null check into CmpN if compressed 2941 // oops implicit null check is not generated. 2942 // This will allow to generate normal oop implicit null check. 2943 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2944 new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR); 2945 // 2946 // This transformation together with CastPP transformation above 2947 // will generated code for implicit NULL checks for compressed oops. 2948 // 2949 // The original code after Optimize() 2950 // 2951 // LoadN memory, narrow_oop_reg 2952 // decode narrow_oop_reg, base_reg 2953 // CmpP base_reg, NULL 2954 // CastPP base_reg // NotNull 2955 // Load [base_reg + offset], val_reg 2956 // 2957 // after these transformations will be 2958 // 2959 // LoadN memory, narrow_oop_reg 2960 // CmpN narrow_oop_reg, NULL 2961 // decode_not_null narrow_oop_reg, base_reg 2962 // Load [base_reg + offset], val_reg 2963 // 2964 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2965 // since narrow oops can be used in debug info now (see the code in 2966 // final_graph_reshaping_walk()). 2967 // 2968 // At the end the code will be matched to 2969 // on x86: 2970 // 2971 // Load_narrow_oop memory, narrow_oop_reg 2972 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2973 // NullCheck narrow_oop_reg 2974 // 2975 // and on sparc: 2976 // 2977 // Load_narrow_oop memory, narrow_oop_reg 2978 // decode_not_null narrow_oop_reg, base_reg 2979 // Load [base_reg + offset], val_reg 2980 // NullCheck base_reg 2981 // 2982 } else if (t->isa_oopptr()) { 2983 new_in2 = ConNode::make(t->make_narrowoop()); 2984 } else if (t->isa_klassptr()) { 2985 new_in2 = ConNode::make(t->make_narrowklass()); 2986 } 2987 } 2988 if (new_in2 != NULL) { 2989 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2990 n->subsume_by(cmpN, this); 2991 if (in1->outcnt() == 0) { 2992 in1->disconnect_inputs(NULL, this); 2993 } 2994 if (in2->outcnt() == 0) { 2995 in2->disconnect_inputs(NULL, this); 2996 } 2997 } 2998 } 2999 break; 3000 3001 case Op_DecodeN: 3002 case Op_DecodeNKlass: 3003 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3004 // DecodeN could be pinned when it can't be fold into 3005 // an address expression, see the code for Op_CastPP above. 3006 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3007 break; 3008 3009 case Op_EncodeP: 3010 case Op_EncodePKlass: { 3011 Node* in1 = n->in(1); 3012 if (in1->is_DecodeNarrowPtr()) { 3013 n->subsume_by(in1->in(1), this); 3014 } else if (in1->Opcode() == Op_ConP) { 3015 const Type* t = in1->bottom_type(); 3016 if (t == TypePtr::NULL_PTR) { 3017 assert(t->isa_oopptr(), "null klass?"); 3018 n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this); 3019 } else if (t->isa_oopptr()) { 3020 n->subsume_by(ConNode::make(t->make_narrowoop()), this); 3021 } else if (t->isa_klassptr()) { 3022 n->subsume_by(ConNode::make(t->make_narrowklass()), this); 3023 } 3024 } 3025 if (in1->outcnt() == 0) { 3026 in1->disconnect_inputs(NULL, this); 3027 } 3028 break; 3029 } 3030 3031 case Op_Proj: { 3032 if (OptimizeStringConcat) { 3033 ProjNode* p = n->as_Proj(); 3034 if (p->_is_io_use) { 3035 // Separate projections were used for the exception path which 3036 // are normally removed by a late inline. If it wasn't inlined 3037 // then they will hang around and should just be replaced with 3038 // the original one. 3039 Node* proj = NULL; 3040 // Replace with just one 3041 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3042 Node *use = i.get(); 3043 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3044 proj = use; 3045 break; 3046 } 3047 } 3048 assert(proj != NULL, "must be found"); 3049 p->subsume_by(proj, this); 3050 } 3051 } 3052 break; 3053 } 3054 3055 case Op_Phi: 3056 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3057 // The EncodeP optimization may create Phi with the same edges 3058 // for all paths. It is not handled well by Register Allocator. 3059 Node* unique_in = n->in(1); 3060 assert(unique_in != NULL, ""); 3061 uint cnt = n->req(); 3062 for (uint i = 2; i < cnt; i++) { 3063 Node* m = n->in(i); 3064 assert(m != NULL, ""); 3065 if (unique_in != m) 3066 unique_in = NULL; 3067 } 3068 if (unique_in != NULL) { 3069 n->subsume_by(unique_in, this); 3070 } 3071 } 3072 break; 3073 3074 #endif 3075 3076 case Op_ModI: 3077 if (UseDivMod) { 3078 // Check if a%b and a/b both exist 3079 Node* d = n->find_similar(Op_DivI); 3080 if (d) { 3081 // Replace them with a fused divmod if supported 3082 if (Matcher::has_match_rule(Op_DivModI)) { 3083 DivModINode* divmod = DivModINode::make(n); 3084 d->subsume_by(divmod->div_proj(), this); 3085 n->subsume_by(divmod->mod_proj(), this); 3086 } else { 3087 // replace a%b with a-((a/b)*b) 3088 Node* mult = new MulINode(d, d->in(2)); 3089 Node* sub = new SubINode(d->in(1), mult); 3090 n->subsume_by(sub, this); 3091 } 3092 } 3093 } 3094 break; 3095 3096 case Op_ModL: 3097 if (UseDivMod) { 3098 // Check if a%b and a/b both exist 3099 Node* d = n->find_similar(Op_DivL); 3100 if (d) { 3101 // Replace them with a fused divmod if supported 3102 if (Matcher::has_match_rule(Op_DivModL)) { 3103 DivModLNode* divmod = DivModLNode::make(n); 3104 d->subsume_by(divmod->div_proj(), this); 3105 n->subsume_by(divmod->mod_proj(), this); 3106 } else { 3107 // replace a%b with a-((a/b)*b) 3108 Node* mult = new MulLNode(d, d->in(2)); 3109 Node* sub = new SubLNode(d->in(1), mult); 3110 n->subsume_by(sub, this); 3111 } 3112 } 3113 } 3114 break; 3115 3116 case Op_LoadVector: 3117 case Op_StoreVector: 3118 break; 3119 3120 case Op_AddReductionVI: 3121 case Op_AddReductionVL: 3122 case Op_AddReductionVF: 3123 case Op_AddReductionVD: 3124 case Op_MulReductionVI: 3125 case Op_MulReductionVL: 3126 case Op_MulReductionVF: 3127 case Op_MulReductionVD: 3128 break; 3129 3130 case Op_PackB: 3131 case Op_PackS: 3132 case Op_PackI: 3133 case Op_PackF: 3134 case Op_PackL: 3135 case Op_PackD: 3136 if (n->req()-1 > 2) { 3137 // Replace many operand PackNodes with a binary tree for matching 3138 PackNode* p = (PackNode*) n; 3139 Node* btp = p->binary_tree_pack(1, n->req()); 3140 n->subsume_by(btp, this); 3141 } 3142 break; 3143 case Op_Loop: 3144 case Op_CountedLoop: 3145 if (n->as_Loop()->is_inner_loop()) { 3146 frc.inc_inner_loop_count(); 3147 } 3148 break; 3149 case Op_LShiftI: 3150 case Op_RShiftI: 3151 case Op_URShiftI: 3152 case Op_LShiftL: 3153 case Op_RShiftL: 3154 case Op_URShiftL: 3155 if (Matcher::need_masked_shift_count) { 3156 // The cpu's shift instructions don't restrict the count to the 3157 // lower 5/6 bits. We need to do the masking ourselves. 3158 Node* in2 = n->in(2); 3159 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3160 const TypeInt* t = in2->find_int_type(); 3161 if (t != NULL && t->is_con()) { 3162 juint shift = t->get_con(); 3163 if (shift > mask) { // Unsigned cmp 3164 n->set_req(2, ConNode::make(TypeInt::make(shift & mask))); 3165 } 3166 } else { 3167 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3168 Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask))); 3169 n->set_req(2, shift); 3170 } 3171 } 3172 if (in2->outcnt() == 0) { // Remove dead node 3173 in2->disconnect_inputs(NULL, this); 3174 } 3175 } 3176 break; 3177 case Op_MemBarStoreStore: 3178 case Op_MemBarRelease: 3179 // Break the link with AllocateNode: it is no longer useful and 3180 // confuses register allocation. 3181 if (n->req() > MemBarNode::Precedent) { 3182 n->set_req(MemBarNode::Precedent, top()); 3183 } 3184 break; 3185 default: 3186 assert( !n->is_Call(), "" ); 3187 assert( !n->is_Mem(), "" ); 3188 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3189 break; 3190 } 3191 3192 // Collect CFG split points 3193 if (n->is_MultiBranch()) 3194 frc._tests.push(n); 3195 } 3196 3197 //------------------------------final_graph_reshaping_walk--------------------- 3198 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3199 // requires that the walk visits a node's inputs before visiting the node. 3200 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3201 ResourceArea *area = Thread::current()->resource_area(); 3202 Unique_Node_List sfpt(area); 3203 3204 frc._visited.set(root->_idx); // first, mark node as visited 3205 uint cnt = root->req(); 3206 Node *n = root; 3207 uint i = 0; 3208 while (true) { 3209 if (i < cnt) { 3210 // Place all non-visited non-null inputs onto stack 3211 Node* m = n->in(i); 3212 ++i; 3213 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3214 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3215 // compute worst case interpreter size in case of a deoptimization 3216 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3217 3218 sfpt.push(m); 3219 } 3220 cnt = m->req(); 3221 nstack.push(n, i); // put on stack parent and next input's index 3222 n = m; 3223 i = 0; 3224 } 3225 } else { 3226 // Now do post-visit work 3227 final_graph_reshaping_impl( n, frc ); 3228 if (nstack.is_empty()) 3229 break; // finished 3230 n = nstack.node(); // Get node from stack 3231 cnt = n->req(); 3232 i = nstack.index(); 3233 nstack.pop(); // Shift to the next node on stack 3234 } 3235 } 3236 3237 // Skip next transformation if compressed oops are not used. 3238 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3239 (!UseCompressedOops && !UseCompressedClassPointers)) 3240 return; 3241 3242 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3243 // It could be done for an uncommon traps or any safepoints/calls 3244 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3245 while (sfpt.size() > 0) { 3246 n = sfpt.pop(); 3247 JVMState *jvms = n->as_SafePoint()->jvms(); 3248 assert(jvms != NULL, "sanity"); 3249 int start = jvms->debug_start(); 3250 int end = n->req(); 3251 bool is_uncommon = (n->is_CallStaticJava() && 3252 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3253 for (int j = start; j < end; j++) { 3254 Node* in = n->in(j); 3255 if (in->is_DecodeNarrowPtr()) { 3256 bool safe_to_skip = true; 3257 if (!is_uncommon ) { 3258 // Is it safe to skip? 3259 for (uint i = 0; i < in->outcnt(); i++) { 3260 Node* u = in->raw_out(i); 3261 if (!u->is_SafePoint() || 3262 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3263 safe_to_skip = false; 3264 } 3265 } 3266 } 3267 if (safe_to_skip) { 3268 n->set_req(j, in->in(1)); 3269 } 3270 if (in->outcnt() == 0) { 3271 in->disconnect_inputs(NULL, this); 3272 } 3273 } 3274 } 3275 } 3276 } 3277 3278 //------------------------------final_graph_reshaping-------------------------- 3279 // Final Graph Reshaping. 3280 // 3281 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3282 // and not commoned up and forced early. Must come after regular 3283 // optimizations to avoid GVN undoing the cloning. Clone constant 3284 // inputs to Loop Phis; these will be split by the allocator anyways. 3285 // Remove Opaque nodes. 3286 // (2) Move last-uses by commutative operations to the left input to encourage 3287 // Intel update-in-place two-address operations and better register usage 3288 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3289 // calls canonicalizing them back. 3290 // (3) Count the number of double-precision FP ops, single-precision FP ops 3291 // and call sites. On Intel, we can get correct rounding either by 3292 // forcing singles to memory (requires extra stores and loads after each 3293 // FP bytecode) or we can set a rounding mode bit (requires setting and 3294 // clearing the mode bit around call sites). The mode bit is only used 3295 // if the relative frequency of single FP ops to calls is low enough. 3296 // This is a key transform for SPEC mpeg_audio. 3297 // (4) Detect infinite loops; blobs of code reachable from above but not 3298 // below. Several of the Code_Gen algorithms fail on such code shapes, 3299 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3300 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3301 // Detection is by looking for IfNodes where only 1 projection is 3302 // reachable from below or CatchNodes missing some targets. 3303 // (5) Assert for insane oop offsets in debug mode. 3304 3305 bool Compile::final_graph_reshaping() { 3306 // an infinite loop may have been eliminated by the optimizer, 3307 // in which case the graph will be empty. 3308 if (root()->req() == 1) { 3309 record_method_not_compilable("trivial infinite loop"); 3310 return true; 3311 } 3312 3313 // Expensive nodes have their control input set to prevent the GVN 3314 // from freely commoning them. There's no GVN beyond this point so 3315 // no need to keep the control input. We want the expensive nodes to 3316 // be freely moved to the least frequent code path by gcm. 3317 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3318 for (int i = 0; i < expensive_count(); i++) { 3319 _expensive_nodes->at(i)->set_req(0, NULL); 3320 } 3321 3322 Final_Reshape_Counts frc; 3323 3324 // Visit everybody reachable! 3325 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3326 Node_Stack nstack(live_nodes() >> 1); 3327 final_graph_reshaping_walk(nstack, root(), frc); 3328 3329 // Check for unreachable (from below) code (i.e., infinite loops). 3330 for( uint i = 0; i < frc._tests.size(); i++ ) { 3331 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3332 // Get number of CFG targets. 3333 // Note that PCTables include exception targets after calls. 3334 uint required_outcnt = n->required_outcnt(); 3335 if (n->outcnt() != required_outcnt) { 3336 // Check for a few special cases. Rethrow Nodes never take the 3337 // 'fall-thru' path, so expected kids is 1 less. 3338 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3339 if (n->in(0)->in(0)->is_Call()) { 3340 CallNode *call = n->in(0)->in(0)->as_Call(); 3341 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3342 required_outcnt--; // Rethrow always has 1 less kid 3343 } else if (call->req() > TypeFunc::Parms && 3344 call->is_CallDynamicJava()) { 3345 // Check for null receiver. In such case, the optimizer has 3346 // detected that the virtual call will always result in a null 3347 // pointer exception. The fall-through projection of this CatchNode 3348 // will not be populated. 3349 Node *arg0 = call->in(TypeFunc::Parms); 3350 if (arg0->is_Type() && 3351 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3352 required_outcnt--; 3353 } 3354 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3355 call->req() > TypeFunc::Parms+1 && 3356 call->is_CallStaticJava()) { 3357 // Check for negative array length. In such case, the optimizer has 3358 // detected that the allocation attempt will always result in an 3359 // exception. There is no fall-through projection of this CatchNode . 3360 Node *arg1 = call->in(TypeFunc::Parms+1); 3361 if (arg1->is_Type() && 3362 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3363 required_outcnt--; 3364 } 3365 } 3366 } 3367 } 3368 // Recheck with a better notion of 'required_outcnt' 3369 if (n->outcnt() != required_outcnt) { 3370 record_method_not_compilable("malformed control flow"); 3371 return true; // Not all targets reachable! 3372 } 3373 } 3374 // Check that I actually visited all kids. Unreached kids 3375 // must be infinite loops. 3376 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3377 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3378 record_method_not_compilable("infinite loop"); 3379 return true; // Found unvisited kid; must be unreach 3380 } 3381 } 3382 3383 // If original bytecodes contained a mixture of floats and doubles 3384 // check if the optimizer has made it homogenous, item (3). 3385 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3386 frc.get_float_count() > 32 && 3387 frc.get_double_count() == 0 && 3388 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3389 set_24_bit_selection_and_mode( false, true ); 3390 } 3391 3392 set_java_calls(frc.get_java_call_count()); 3393 set_inner_loops(frc.get_inner_loop_count()); 3394 3395 // No infinite loops, no reason to bail out. 3396 return false; 3397 } 3398 3399 //-----------------------------too_many_traps---------------------------------- 3400 // Report if there are too many traps at the current method and bci. 3401 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3402 bool Compile::too_many_traps(ciMethod* method, 3403 int bci, 3404 Deoptimization::DeoptReason reason) { 3405 ciMethodData* md = method->method_data(); 3406 if (md->is_empty()) { 3407 // Assume the trap has not occurred, or that it occurred only 3408 // because of a transient condition during start-up in the interpreter. 3409 return false; 3410 } 3411 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3412 if (md->has_trap_at(bci, m, reason) != 0) { 3413 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3414 // Also, if there are multiple reasons, or if there is no per-BCI record, 3415 // assume the worst. 3416 if (log()) 3417 log()->elem("observe trap='%s' count='%d'", 3418 Deoptimization::trap_reason_name(reason), 3419 md->trap_count(reason)); 3420 return true; 3421 } else { 3422 // Ignore method/bci and see if there have been too many globally. 3423 return too_many_traps(reason, md); 3424 } 3425 } 3426 3427 // Less-accurate variant which does not require a method and bci. 3428 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3429 ciMethodData* logmd) { 3430 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3431 // Too many traps globally. 3432 // Note that we use cumulative trap_count, not just md->trap_count. 3433 if (log()) { 3434 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3435 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3436 Deoptimization::trap_reason_name(reason), 3437 mcount, trap_count(reason)); 3438 } 3439 return true; 3440 } else { 3441 // The coast is clear. 3442 return false; 3443 } 3444 } 3445 3446 //--------------------------too_many_recompiles-------------------------------- 3447 // Report if there are too many recompiles at the current method and bci. 3448 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3449 // Is not eager to return true, since this will cause the compiler to use 3450 // Action_none for a trap point, to avoid too many recompilations. 3451 bool Compile::too_many_recompiles(ciMethod* method, 3452 int bci, 3453 Deoptimization::DeoptReason reason) { 3454 ciMethodData* md = method->method_data(); 3455 if (md->is_empty()) { 3456 // Assume the trap has not occurred, or that it occurred only 3457 // because of a transient condition during start-up in the interpreter. 3458 return false; 3459 } 3460 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3461 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3462 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3463 Deoptimization::DeoptReason per_bc_reason 3464 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3465 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3466 if ((per_bc_reason == Deoptimization::Reason_none 3467 || md->has_trap_at(bci, m, reason) != 0) 3468 // The trap frequency measure we care about is the recompile count: 3469 && md->trap_recompiled_at(bci, m) 3470 && md->overflow_recompile_count() >= bc_cutoff) { 3471 // Do not emit a trap here if it has already caused recompilations. 3472 // Also, if there are multiple reasons, or if there is no per-BCI record, 3473 // assume the worst. 3474 if (log()) 3475 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3476 Deoptimization::trap_reason_name(reason), 3477 md->trap_count(reason), 3478 md->overflow_recompile_count()); 3479 return true; 3480 } else if (trap_count(reason) != 0 3481 && decompile_count() >= m_cutoff) { 3482 // Too many recompiles globally, and we have seen this sort of trap. 3483 // Use cumulative decompile_count, not just md->decompile_count. 3484 if (log()) 3485 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3486 Deoptimization::trap_reason_name(reason), 3487 md->trap_count(reason), trap_count(reason), 3488 md->decompile_count(), decompile_count()); 3489 return true; 3490 } else { 3491 // The coast is clear. 3492 return false; 3493 } 3494 } 3495 3496 // Compute when not to trap. Used by matching trap based nodes and 3497 // NullCheck optimization. 3498 void Compile::set_allowed_deopt_reasons() { 3499 _allowed_reasons = 0; 3500 if (is_method_compilation()) { 3501 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3502 assert(rs < BitsPerInt, "recode bit map"); 3503 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3504 _allowed_reasons |= nth_bit(rs); 3505 } 3506 } 3507 } 3508 } 3509 3510 #ifndef PRODUCT 3511 //------------------------------verify_graph_edges--------------------------- 3512 // Walk the Graph and verify that there is a one-to-one correspondence 3513 // between Use-Def edges and Def-Use edges in the graph. 3514 void Compile::verify_graph_edges(bool no_dead_code) { 3515 if (VerifyGraphEdges) { 3516 ResourceArea *area = Thread::current()->resource_area(); 3517 Unique_Node_List visited(area); 3518 // Call recursive graph walk to check edges 3519 _root->verify_edges(visited); 3520 if (no_dead_code) { 3521 // Now make sure that no visited node is used by an unvisited node. 3522 bool dead_nodes = false; 3523 Unique_Node_List checked(area); 3524 while (visited.size() > 0) { 3525 Node* n = visited.pop(); 3526 checked.push(n); 3527 for (uint i = 0; i < n->outcnt(); i++) { 3528 Node* use = n->raw_out(i); 3529 if (checked.member(use)) continue; // already checked 3530 if (visited.member(use)) continue; // already in the graph 3531 if (use->is_Con()) continue; // a dead ConNode is OK 3532 // At this point, we have found a dead node which is DU-reachable. 3533 if (!dead_nodes) { 3534 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3535 dead_nodes = true; 3536 } 3537 use->dump(2); 3538 tty->print_cr("---"); 3539 checked.push(use); // No repeats; pretend it is now checked. 3540 } 3541 } 3542 assert(!dead_nodes, "using nodes must be reachable from root"); 3543 } 3544 } 3545 } 3546 3547 // Verify GC barriers consistency 3548 // Currently supported: 3549 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3550 void Compile::verify_barriers() { 3551 if (UseG1GC) { 3552 // Verify G1 pre-barriers 3553 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3554 3555 ResourceArea *area = Thread::current()->resource_area(); 3556 Unique_Node_List visited(area); 3557 Node_List worklist(area); 3558 // We're going to walk control flow backwards starting from the Root 3559 worklist.push(_root); 3560 while (worklist.size() > 0) { 3561 Node* x = worklist.pop(); 3562 if (x == NULL || x == top()) continue; 3563 if (visited.member(x)) { 3564 continue; 3565 } else { 3566 visited.push(x); 3567 } 3568 3569 if (x->is_Region()) { 3570 for (uint i = 1; i < x->req(); i++) { 3571 worklist.push(x->in(i)); 3572 } 3573 } else { 3574 worklist.push(x->in(0)); 3575 // We are looking for the pattern: 3576 // /->ThreadLocal 3577 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3578 // \->ConI(0) 3579 // We want to verify that the If and the LoadB have the same control 3580 // See GraphKit::g1_write_barrier_pre() 3581 if (x->is_If()) { 3582 IfNode *iff = x->as_If(); 3583 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3584 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3585 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3586 && cmp->in(1)->is_Load()) { 3587 LoadNode* load = cmp->in(1)->as_Load(); 3588 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3589 && load->in(2)->in(3)->is_Con() 3590 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3591 3592 Node* if_ctrl = iff->in(0); 3593 Node* load_ctrl = load->in(0); 3594 3595 if (if_ctrl != load_ctrl) { 3596 // Skip possible CProj->NeverBranch in infinite loops 3597 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3598 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3599 if_ctrl = if_ctrl->in(0)->in(0); 3600 } 3601 } 3602 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3603 } 3604 } 3605 } 3606 } 3607 } 3608 } 3609 } 3610 } 3611 3612 #endif 3613 3614 // The Compile object keeps track of failure reasons separately from the ciEnv. 3615 // This is required because there is not quite a 1-1 relation between the 3616 // ciEnv and its compilation task and the Compile object. Note that one 3617 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3618 // to backtrack and retry without subsuming loads. Other than this backtracking 3619 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3620 // by the logic in C2Compiler. 3621 void Compile::record_failure(const char* reason) { 3622 if (log() != NULL) { 3623 log()->elem("failure reason='%s' phase='compile'", reason); 3624 } 3625 if (_failure_reason == NULL) { 3626 // Record the first failure reason. 3627 _failure_reason = reason; 3628 } 3629 3630 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3631 C->print_method(PHASE_FAILURE); 3632 } 3633 _root = NULL; // flush the graph, too 3634 } 3635 3636 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator) 3637 : TraceTime(name, accumulator, CITime, CITimeVerbose), 3638 _phase_name(name), _dolog(CITimeVerbose) 3639 { 3640 if (_dolog) { 3641 C = Compile::current(); 3642 _log = C->log(); 3643 } else { 3644 C = NULL; 3645 _log = NULL; 3646 } 3647 if (_log != NULL) { 3648 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3649 _log->stamp(); 3650 _log->end_head(); 3651 } 3652 } 3653 3654 Compile::TracePhase::~TracePhase() { 3655 3656 C = Compile::current(); 3657 if (_dolog) { 3658 _log = C->log(); 3659 } else { 3660 _log = NULL; 3661 } 3662 3663 #ifdef ASSERT 3664 if (PrintIdealNodeCount) { 3665 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3666 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3667 } 3668 3669 if (VerifyIdealNodeCount) { 3670 Compile::current()->print_missing_nodes(); 3671 } 3672 #endif 3673 3674 if (_log != NULL) { 3675 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3676 } 3677 } 3678 3679 //============================================================================= 3680 // Two Constant's are equal when the type and the value are equal. 3681 bool Compile::Constant::operator==(const Constant& other) { 3682 if (type() != other.type() ) return false; 3683 if (can_be_reused() != other.can_be_reused()) return false; 3684 // For floating point values we compare the bit pattern. 3685 switch (type()) { 3686 case T_FLOAT: return (_v._value.i == other._v._value.i); 3687 case T_LONG: 3688 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3689 case T_OBJECT: 3690 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3691 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3692 case T_METADATA: return (_v._metadata == other._v._metadata); 3693 default: ShouldNotReachHere(); 3694 } 3695 return false; 3696 } 3697 3698 static int type_to_size_in_bytes(BasicType t) { 3699 switch (t) { 3700 case T_LONG: return sizeof(jlong ); 3701 case T_FLOAT: return sizeof(jfloat ); 3702 case T_DOUBLE: return sizeof(jdouble); 3703 case T_METADATA: return sizeof(Metadata*); 3704 // We use T_VOID as marker for jump-table entries (labels) which 3705 // need an internal word relocation. 3706 case T_VOID: 3707 case T_ADDRESS: 3708 case T_OBJECT: return sizeof(jobject); 3709 } 3710 3711 ShouldNotReachHere(); 3712 return -1; 3713 } 3714 3715 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3716 // sort descending 3717 if (a->freq() > b->freq()) return -1; 3718 if (a->freq() < b->freq()) return 1; 3719 return 0; 3720 } 3721 3722 void Compile::ConstantTable::calculate_offsets_and_size() { 3723 // First, sort the array by frequencies. 3724 _constants.sort(qsort_comparator); 3725 3726 #ifdef ASSERT 3727 // Make sure all jump-table entries were sorted to the end of the 3728 // array (they have a negative frequency). 3729 bool found_void = false; 3730 for (int i = 0; i < _constants.length(); i++) { 3731 Constant con = _constants.at(i); 3732 if (con.type() == T_VOID) 3733 found_void = true; // jump-tables 3734 else 3735 assert(!found_void, "wrong sorting"); 3736 } 3737 #endif 3738 3739 int offset = 0; 3740 for (int i = 0; i < _constants.length(); i++) { 3741 Constant* con = _constants.adr_at(i); 3742 3743 // Align offset for type. 3744 int typesize = type_to_size_in_bytes(con->type()); 3745 offset = align_size_up(offset, typesize); 3746 con->set_offset(offset); // set constant's offset 3747 3748 if (con->type() == T_VOID) { 3749 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3750 offset = offset + typesize * n->outcnt(); // expand jump-table 3751 } else { 3752 offset = offset + typesize; 3753 } 3754 } 3755 3756 // Align size up to the next section start (which is insts; see 3757 // CodeBuffer::align_at_start). 3758 assert(_size == -1, "already set?"); 3759 _size = align_size_up(offset, CodeEntryAlignment); 3760 } 3761 3762 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3763 MacroAssembler _masm(&cb); 3764 for (int i = 0; i < _constants.length(); i++) { 3765 Constant con = _constants.at(i); 3766 address constant_addr; 3767 switch (con.type()) { 3768 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3769 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3770 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3771 case T_OBJECT: { 3772 jobject obj = con.get_jobject(); 3773 int oop_index = _masm.oop_recorder()->find_index(obj); 3774 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3775 break; 3776 } 3777 case T_ADDRESS: { 3778 address addr = (address) con.get_jobject(); 3779 constant_addr = _masm.address_constant(addr); 3780 break; 3781 } 3782 // We use T_VOID as marker for jump-table entries (labels) which 3783 // need an internal word relocation. 3784 case T_VOID: { 3785 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3786 // Fill the jump-table with a dummy word. The real value is 3787 // filled in later in fill_jump_table. 3788 address dummy = (address) n; 3789 constant_addr = _masm.address_constant(dummy); 3790 // Expand jump-table 3791 for (uint i = 1; i < n->outcnt(); i++) { 3792 address temp_addr = _masm.address_constant(dummy + i); 3793 assert(temp_addr, "consts section too small"); 3794 } 3795 break; 3796 } 3797 case T_METADATA: { 3798 Metadata* obj = con.get_metadata(); 3799 int metadata_index = _masm.oop_recorder()->find_index(obj); 3800 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3801 break; 3802 } 3803 default: ShouldNotReachHere(); 3804 } 3805 assert(constant_addr, "consts section too small"); 3806 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3807 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3808 } 3809 } 3810 3811 int Compile::ConstantTable::find_offset(Constant& con) const { 3812 int idx = _constants.find(con); 3813 assert(idx != -1, "constant must be in constant table"); 3814 int offset = _constants.at(idx).offset(); 3815 assert(offset != -1, "constant table not emitted yet?"); 3816 return offset; 3817 } 3818 3819 void Compile::ConstantTable::add(Constant& con) { 3820 if (con.can_be_reused()) { 3821 int idx = _constants.find(con); 3822 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3823 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3824 return; 3825 } 3826 } 3827 (void) _constants.append(con); 3828 } 3829 3830 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3831 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3832 Constant con(type, value, b->_freq); 3833 add(con); 3834 return con; 3835 } 3836 3837 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3838 Constant con(metadata); 3839 add(con); 3840 return con; 3841 } 3842 3843 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3844 jvalue value; 3845 BasicType type = oper->type()->basic_type(); 3846 switch (type) { 3847 case T_LONG: value.j = oper->constantL(); break; 3848 case T_FLOAT: value.f = oper->constantF(); break; 3849 case T_DOUBLE: value.d = oper->constantD(); break; 3850 case T_OBJECT: 3851 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3852 case T_METADATA: return add((Metadata*)oper->constant()); break; 3853 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3854 } 3855 return add(n, type, value); 3856 } 3857 3858 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3859 jvalue value; 3860 // We can use the node pointer here to identify the right jump-table 3861 // as this method is called from Compile::Fill_buffer right before 3862 // the MachNodes are emitted and the jump-table is filled (means the 3863 // MachNode pointers do not change anymore). 3864 value.l = (jobject) n; 3865 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3866 add(con); 3867 return con; 3868 } 3869 3870 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3871 // If called from Compile::scratch_emit_size do nothing. 3872 if (Compile::current()->in_scratch_emit_size()) return; 3873 3874 assert(labels.is_nonempty(), "must be"); 3875 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3876 3877 // Since MachConstantNode::constant_offset() also contains 3878 // table_base_offset() we need to subtract the table_base_offset() 3879 // to get the plain offset into the constant table. 3880 int offset = n->constant_offset() - table_base_offset(); 3881 3882 MacroAssembler _masm(&cb); 3883 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3884 3885 for (uint i = 0; i < n->outcnt(); i++) { 3886 address* constant_addr = &jump_table_base[i]; 3887 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))); 3888 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3889 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3890 } 3891 } 3892 3893 //----------------------------static_subtype_check----------------------------- 3894 // Shortcut important common cases when superklass is exact: 3895 // (0) superklass is java.lang.Object (can occur in reflective code) 3896 // (1) subklass is already limited to a subtype of superklass => always ok 3897 // (2) subklass does not overlap with superklass => always fail 3898 // (3) superklass has NO subtypes and we can check with a simple compare. 3899 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3900 if (StressReflectiveCode) { 3901 return SSC_full_test; // Let caller generate the general case. 3902 } 3903 3904 if (superk == env()->Object_klass()) { 3905 return SSC_always_true; // (0) this test cannot fail 3906 } 3907 3908 ciType* superelem = superk; 3909 if (superelem->is_array_klass()) 3910 superelem = superelem->as_array_klass()->base_element_type(); 3911 3912 if (!subk->is_interface()) { // cannot trust static interface types yet 3913 if (subk->is_subtype_of(superk)) { 3914 return SSC_always_true; // (1) false path dead; no dynamic test needed 3915 } 3916 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3917 !superk->is_subtype_of(subk)) { 3918 return SSC_always_false; 3919 } 3920 } 3921 3922 // If casting to an instance klass, it must have no subtypes 3923 if (superk->is_interface()) { 3924 // Cannot trust interfaces yet. 3925 // %%% S.B. superk->nof_implementors() == 1 3926 } else if (superelem->is_instance_klass()) { 3927 ciInstanceKlass* ik = superelem->as_instance_klass(); 3928 if (!ik->has_subklass() && !ik->is_interface()) { 3929 if (!ik->is_final()) { 3930 // Add a dependency if there is a chance of a later subclass. 3931 dependencies()->assert_leaf_type(ik); 3932 } 3933 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3934 } 3935 } else { 3936 // A primitive array type has no subtypes. 3937 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3938 } 3939 3940 return SSC_full_test; 3941 } 3942 3943 Node* Compile::conv_I2X_index(PhaseGVN *phase, Node* idx, const TypeInt* sizetype) { 3944 #ifdef _LP64 3945 // The scaled index operand to AddP must be a clean 64-bit value. 3946 // Java allows a 32-bit int to be incremented to a negative 3947 // value, which appears in a 64-bit register as a large 3948 // positive number. Using that large positive number as an 3949 // operand in pointer arithmetic has bad consequences. 3950 // On the other hand, 32-bit overflow is rare, and the possibility 3951 // can often be excluded, if we annotate the ConvI2L node with 3952 // a type assertion that its value is known to be a small positive 3953 // number. (The prior range check has ensured this.) 3954 // This assertion is used by ConvI2LNode::Ideal. 3955 int index_max = max_jint - 1; // array size is max_jint, index is one less 3956 if (sizetype != NULL) index_max = sizetype->_hi - 1; 3957 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax); 3958 idx = phase->transform(new ConvI2LNode(idx, lidxtype)); 3959 #endif 3960 return idx; 3961 } 3962 3963 // The message about the current inlining is accumulated in 3964 // _print_inlining_stream and transfered into the _print_inlining_list 3965 // once we know whether inlining succeeds or not. For regular 3966 // inlining, messages are appended to the buffer pointed by 3967 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3968 // a new buffer is added after _print_inlining_idx in the list. This 3969 // way we can update the inlining message for late inlining call site 3970 // when the inlining is attempted again. 3971 void Compile::print_inlining_init() { 3972 if (print_inlining() || print_intrinsics()) { 3973 _print_inlining_stream = new stringStream(); 3974 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3975 } 3976 } 3977 3978 void Compile::print_inlining_reinit() { 3979 if (print_inlining() || print_intrinsics()) { 3980 // Re allocate buffer when we change ResourceMark 3981 _print_inlining_stream = new stringStream(); 3982 } 3983 } 3984 3985 void Compile::print_inlining_reset() { 3986 _print_inlining_stream->reset(); 3987 } 3988 3989 void Compile::print_inlining_commit() { 3990 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3991 // Transfer the message from _print_inlining_stream to the current 3992 // _print_inlining_list buffer and clear _print_inlining_stream. 3993 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3994 print_inlining_reset(); 3995 } 3996 3997 void Compile::print_inlining_push() { 3998 // Add new buffer to the _print_inlining_list at current position 3999 _print_inlining_idx++; 4000 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 4001 } 4002 4003 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 4004 return _print_inlining_list->at(_print_inlining_idx); 4005 } 4006 4007 void Compile::print_inlining_update(CallGenerator* cg) { 4008 if (print_inlining() || print_intrinsics()) { 4009 if (!cg->is_late_inline()) { 4010 if (print_inlining_current().cg() != NULL) { 4011 print_inlining_push(); 4012 } 4013 print_inlining_commit(); 4014 } else { 4015 if (print_inlining_current().cg() != cg && 4016 (print_inlining_current().cg() != NULL || 4017 print_inlining_current().ss()->size() != 0)) { 4018 print_inlining_push(); 4019 } 4020 print_inlining_commit(); 4021 print_inlining_current().set_cg(cg); 4022 } 4023 } 4024 } 4025 4026 void Compile::print_inlining_move_to(CallGenerator* cg) { 4027 // We resume inlining at a late inlining call site. Locate the 4028 // corresponding inlining buffer so that we can update it. 4029 if (print_inlining()) { 4030 for (int i = 0; i < _print_inlining_list->length(); i++) { 4031 if (_print_inlining_list->adr_at(i)->cg() == cg) { 4032 _print_inlining_idx = i; 4033 return; 4034 } 4035 } 4036 ShouldNotReachHere(); 4037 } 4038 } 4039 4040 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 4041 if (print_inlining()) { 4042 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 4043 assert(print_inlining_current().cg() == cg, "wrong entry"); 4044 // replace message with new message 4045 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 4046 print_inlining_commit(); 4047 print_inlining_current().set_cg(cg); 4048 } 4049 } 4050 4051 void Compile::print_inlining_assert_ready() { 4052 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 4053 } 4054 4055 void Compile::process_print_inlining() { 4056 bool do_print_inlining = print_inlining() || print_intrinsics(); 4057 if (do_print_inlining || log() != NULL) { 4058 // Print inlining message for candidates that we couldn't inline 4059 // for lack of space 4060 for (int i = 0; i < _late_inlines.length(); i++) { 4061 CallGenerator* cg = _late_inlines.at(i); 4062 if (!cg->is_mh_late_inline()) { 4063 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 4064 if (do_print_inlining) { 4065 cg->print_inlining_late(msg); 4066 } 4067 log_late_inline_failure(cg, msg); 4068 } 4069 } 4070 } 4071 if (do_print_inlining) { 4072 ResourceMark rm; 4073 stringStream ss; 4074 for (int i = 0; i < _print_inlining_list->length(); i++) { 4075 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4076 } 4077 size_t end = ss.size(); 4078 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 4079 strncpy(_print_inlining_output, ss.base(), end+1); 4080 _print_inlining_output[end] = 0; 4081 } 4082 } 4083 4084 void Compile::dump_print_inlining() { 4085 if (_print_inlining_output != NULL) { 4086 tty->print_raw(_print_inlining_output); 4087 } 4088 } 4089 4090 void Compile::log_late_inline(CallGenerator* cg) { 4091 if (log() != NULL) { 4092 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 4093 cg->unique_id()); 4094 JVMState* p = cg->call_node()->jvms(); 4095 while (p != NULL) { 4096 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 4097 p = p->caller(); 4098 } 4099 log()->tail("late_inline"); 4100 } 4101 } 4102 4103 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 4104 log_late_inline(cg); 4105 if (log() != NULL) { 4106 log()->inline_fail(msg); 4107 } 4108 } 4109 4110 void Compile::log_inline_id(CallGenerator* cg) { 4111 if (log() != NULL) { 4112 // The LogCompilation tool needs a unique way to identify late 4113 // inline call sites. This id must be unique for this call site in 4114 // this compilation. Try to have it unique across compilations as 4115 // well because it can be convenient when grepping through the log 4116 // file. 4117 // Distinguish OSR compilations from others in case CICountOSR is 4118 // on. 4119 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 4120 cg->set_unique_id(id); 4121 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 4122 } 4123 } 4124 4125 void Compile::log_inline_failure(const char* msg) { 4126 if (C->log() != NULL) { 4127 C->log()->inline_fail(msg); 4128 } 4129 } 4130 4131 4132 // Dump inlining replay data to the stream. 4133 // Don't change thread state and acquire any locks. 4134 void Compile::dump_inline_data(outputStream* out) { 4135 InlineTree* inl_tree = ilt(); 4136 if (inl_tree != NULL) { 4137 out->print(" inline %d", inl_tree->count()); 4138 inl_tree->dump_replay_data(out); 4139 } 4140 } 4141 4142 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4143 if (n1->Opcode() < n2->Opcode()) return -1; 4144 else if (n1->Opcode() > n2->Opcode()) return 1; 4145 4146 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())); 4147 for (uint i = 1; i < n1->req(); i++) { 4148 if (n1->in(i) < n2->in(i)) return -1; 4149 else if (n1->in(i) > n2->in(i)) return 1; 4150 } 4151 4152 return 0; 4153 } 4154 4155 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4156 Node* n1 = *n1p; 4157 Node* n2 = *n2p; 4158 4159 return cmp_expensive_nodes(n1, n2); 4160 } 4161 4162 void Compile::sort_expensive_nodes() { 4163 if (!expensive_nodes_sorted()) { 4164 _expensive_nodes->sort(cmp_expensive_nodes); 4165 } 4166 } 4167 4168 bool Compile::expensive_nodes_sorted() const { 4169 for (int i = 1; i < _expensive_nodes->length(); i++) { 4170 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4171 return false; 4172 } 4173 } 4174 return true; 4175 } 4176 4177 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4178 if (_expensive_nodes->length() == 0) { 4179 return false; 4180 } 4181 4182 assert(OptimizeExpensiveOps, "optimization off?"); 4183 4184 // Take this opportunity to remove dead nodes from the list 4185 int j = 0; 4186 for (int i = 0; i < _expensive_nodes->length(); i++) { 4187 Node* n = _expensive_nodes->at(i); 4188 if (!n->is_unreachable(igvn)) { 4189 assert(n->is_expensive(), "should be expensive"); 4190 _expensive_nodes->at_put(j, n); 4191 j++; 4192 } 4193 } 4194 _expensive_nodes->trunc_to(j); 4195 4196 // Then sort the list so that similar nodes are next to each other 4197 // and check for at least two nodes of identical kind with same data 4198 // inputs. 4199 sort_expensive_nodes(); 4200 4201 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4202 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4203 return true; 4204 } 4205 } 4206 4207 return false; 4208 } 4209 4210 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4211 if (_expensive_nodes->length() == 0) { 4212 return; 4213 } 4214 4215 assert(OptimizeExpensiveOps, "optimization off?"); 4216 4217 // Sort to bring similar nodes next to each other and clear the 4218 // control input of nodes for which there's only a single copy. 4219 sort_expensive_nodes(); 4220 4221 int j = 0; 4222 int identical = 0; 4223 int i = 0; 4224 bool modified = false; 4225 for (; i < _expensive_nodes->length()-1; i++) { 4226 assert(j <= i, "can't write beyond current index"); 4227 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4228 identical++; 4229 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4230 continue; 4231 } 4232 if (identical > 0) { 4233 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4234 identical = 0; 4235 } else { 4236 Node* n = _expensive_nodes->at(i); 4237 igvn.replace_input_of(n, 0, NULL); 4238 igvn.hash_insert(n); 4239 modified = true; 4240 } 4241 } 4242 if (identical > 0) { 4243 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4244 } else if (_expensive_nodes->length() >= 1) { 4245 Node* n = _expensive_nodes->at(i); 4246 igvn.replace_input_of(n, 0, NULL); 4247 igvn.hash_insert(n); 4248 modified = true; 4249 } 4250 _expensive_nodes->trunc_to(j); 4251 if (modified) { 4252 igvn.optimize(); 4253 } 4254 } 4255 4256 void Compile::add_expensive_node(Node * n) { 4257 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4258 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4259 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4260 if (OptimizeExpensiveOps) { 4261 _expensive_nodes->append(n); 4262 } else { 4263 // Clear control input and let IGVN optimize expensive nodes if 4264 // OptimizeExpensiveOps is off. 4265 n->set_req(0, NULL); 4266 } 4267 } 4268 4269 /** 4270 * Remove the speculative part of types and clean up the graph 4271 */ 4272 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4273 if (UseTypeSpeculation) { 4274 Unique_Node_List worklist; 4275 worklist.push(root()); 4276 int modified = 0; 4277 // Go over all type nodes that carry a speculative type, drop the 4278 // speculative part of the type and enqueue the node for an igvn 4279 // which may optimize it out. 4280 for (uint next = 0; next < worklist.size(); ++next) { 4281 Node *n = worklist.at(next); 4282 if (n->is_Type()) { 4283 TypeNode* tn = n->as_Type(); 4284 const Type* t = tn->type(); 4285 const Type* t_no_spec = t->remove_speculative(); 4286 if (t_no_spec != t) { 4287 bool in_hash = igvn.hash_delete(n); 4288 assert(in_hash, "node should be in igvn hash table"); 4289 tn->set_type(t_no_spec); 4290 igvn.hash_insert(n); 4291 igvn._worklist.push(n); // give it a chance to go away 4292 modified++; 4293 } 4294 } 4295 uint max = n->len(); 4296 for( uint i = 0; i < max; ++i ) { 4297 Node *m = n->in(i); 4298 if (not_a_node(m)) continue; 4299 worklist.push(m); 4300 } 4301 } 4302 // Drop the speculative part of all types in the igvn's type table 4303 igvn.remove_speculative_types(); 4304 if (modified > 0) { 4305 igvn.optimize(); 4306 } 4307 #ifdef ASSERT 4308 // Verify that after the IGVN is over no speculative type has resurfaced 4309 worklist.clear(); 4310 worklist.push(root()); 4311 for (uint next = 0; next < worklist.size(); ++next) { 4312 Node *n = worklist.at(next); 4313 const Type* t = igvn.type_or_null(n); 4314 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4315 if (n->is_Type()) { 4316 t = n->as_Type()->type(); 4317 assert(t == t->remove_speculative(), "no more speculative types"); 4318 } 4319 uint max = n->len(); 4320 for( uint i = 0; i < max; ++i ) { 4321 Node *m = n->in(i); 4322 if (not_a_node(m)) continue; 4323 worklist.push(m); 4324 } 4325 } 4326 igvn.check_no_speculative_types(); 4327 #endif 4328 } 4329 } 4330 4331 // Auxiliary method to support randomized stressing/fuzzing. 4332 // 4333 // This method can be called the arbitrary number of times, with current count 4334 // as the argument. The logic allows selecting a single candidate from the 4335 // running list of candidates as follows: 4336 // int count = 0; 4337 // Cand* selected = null; 4338 // while(cand = cand->next()) { 4339 // if (randomized_select(++count)) { 4340 // selected = cand; 4341 // } 4342 // } 4343 // 4344 // Including count equalizes the chances any candidate is "selected". 4345 // This is useful when we don't have the complete list of candidates to choose 4346 // from uniformly. In this case, we need to adjust the randomicity of the 4347 // selection, or else we will end up biasing the selection towards the latter 4348 // candidates. 4349 // 4350 // Quick back-envelope calculation shows that for the list of n candidates 4351 // the equal probability for the candidate to persist as "best" can be 4352 // achieved by replacing it with "next" k-th candidate with the probability 4353 // of 1/k. It can be easily shown that by the end of the run, the 4354 // probability for any candidate is converged to 1/n, thus giving the 4355 // uniform distribution among all the candidates. 4356 // 4357 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4358 #define RANDOMIZED_DOMAIN_POW 29 4359 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4360 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4361 bool Compile::randomized_select(int count) { 4362 assert(count > 0, "only positive"); 4363 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4364 } 4365 4366 CloneMap& Compile::clone_map() { return _clone_map; } 4367 void Compile::set_clone_map(Dict* d) { _clone_map._dict = d; } 4368 4369 void NodeCloneInfo::dump() const { 4370 tty->print(" {%d:%d} ", idx(), gen()); 4371 } 4372 4373 void CloneMap::clone(Node* old, Node* nnn, int gen) { 4374 uint64_t val = value(old->_idx); 4375 NodeCloneInfo cio(val); 4376 assert(val != 0, "old node should be in the map"); 4377 NodeCloneInfo cin(cio.idx(), gen + cio.gen()); 4378 insert(nnn->_idx, cin.get()); 4379 #ifndef PRODUCT 4380 if (is_debug()) { 4381 tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen()); 4382 } 4383 #endif 4384 } 4385 4386 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) { 4387 NodeCloneInfo cio(value(old->_idx)); 4388 if (cio.get() == 0) { 4389 cio.set(old->_idx, 0); 4390 insert(old->_idx, cio.get()); 4391 #ifndef PRODUCT 4392 if (is_debug()) { 4393 tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen()); 4394 } 4395 #endif 4396 } 4397 clone(old, nnn, gen); 4398 } 4399 4400 int CloneMap::max_gen() const { 4401 int g = 0; 4402 DictI di(_dict); 4403 for(; di.test(); ++di) { 4404 int t = gen(di._key); 4405 if (g < t) { 4406 g = t; 4407 #ifndef PRODUCT 4408 if (is_debug()) { 4409 tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key)); 4410 } 4411 #endif 4412 } 4413 } 4414 return g; 4415 } 4416 4417 void CloneMap::dump(node_idx_t key) const { 4418 uint64_t val = value(key); 4419 if (val != 0) { 4420 NodeCloneInfo ni(val); 4421 ni.dump(); 4422 } 4423 }