1 /* 2 * Copyright (c) 1997, 2013, 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 "classfile/systemDictionary.hpp" 29 #include "code/exceptionHandlerTable.hpp" 30 #include "code/nmethod.hpp" 31 #include "compiler/compileLog.hpp" 32 #include "compiler/disassembler.hpp" 33 #include "compiler/oopMap.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/block.hpp" 36 #include "opto/c2compiler.hpp" 37 #include "opto/callGenerator.hpp" 38 #include "opto/callnode.hpp" 39 #include "opto/cfgnode.hpp" 40 #include "opto/chaitin.hpp" 41 #include "opto/compile.hpp" 42 #include "opto/connode.hpp" 43 #include "opto/divnode.hpp" 44 #include "opto/escape.hpp" 45 #include "opto/idealGraphPrinter.hpp" 46 #include "opto/loopnode.hpp" 47 #include "opto/machnode.hpp" 48 #include "opto/macro.hpp" 49 #include "opto/matcher.hpp" 50 #include "opto/mathexactnode.hpp" 51 #include "opto/memnode.hpp" 52 #include "opto/mulnode.hpp" 53 #include "opto/node.hpp" 54 #include "opto/opcodes.hpp" 55 #include "opto/output.hpp" 56 #include "opto/parse.hpp" 57 #include "opto/phaseX.hpp" 58 #include "opto/rootnode.hpp" 59 #include "opto/runtime.hpp" 60 #include "opto/stringopts.hpp" 61 #include "opto/type.hpp" 62 #include "opto/vectornode.hpp" 63 #include "runtime/arguments.hpp" 64 #include "runtime/signature.hpp" 65 #include "runtime/stubRoutines.hpp" 66 #include "runtime/timer.hpp" 67 #include "trace/tracing.hpp" 68 #include "utilities/copy.hpp" 69 #ifdef TARGET_ARCH_MODEL_x86_32 70 # include "adfiles/ad_x86_32.hpp" 71 #endif 72 #ifdef TARGET_ARCH_MODEL_x86_64 73 # include "adfiles/ad_x86_64.hpp" 74 #endif 75 #ifdef TARGET_ARCH_MODEL_sparc 76 # include "adfiles/ad_sparc.hpp" 77 #endif 78 #ifdef TARGET_ARCH_MODEL_zero 79 # include "adfiles/ad_zero.hpp" 80 #endif 81 #ifdef TARGET_ARCH_MODEL_arm 82 # include "adfiles/ad_arm.hpp" 83 #endif 84 #ifdef TARGET_ARCH_MODEL_ppc_32 85 # include "adfiles/ad_ppc_32.hpp" 86 #endif 87 #ifdef TARGET_ARCH_MODEL_ppc_64 88 # include "adfiles/ad_ppc_64.hpp" 89 #endif 90 91 92 // -------------------- Compile::mach_constant_base_node ----------------------- 93 // Constant table base node singleton. 94 MachConstantBaseNode* Compile::mach_constant_base_node() { 95 if (_mach_constant_base_node == NULL) { 96 _mach_constant_base_node = new (C) MachConstantBaseNode(); 97 _mach_constant_base_node->add_req(C->root()); 98 } 99 return _mach_constant_base_node; 100 } 101 102 103 /// Support for intrinsics. 104 105 // Return the index at which m must be inserted (or already exists). 106 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 107 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 108 #ifdef ASSERT 109 for (int i = 1; i < _intrinsics->length(); i++) { 110 CallGenerator* cg1 = _intrinsics->at(i-1); 111 CallGenerator* cg2 = _intrinsics->at(i); 112 assert(cg1->method() != cg2->method() 113 ? cg1->method() < cg2->method() 114 : cg1->is_virtual() < cg2->is_virtual(), 115 "compiler intrinsics list must stay sorted"); 116 } 117 #endif 118 // Binary search sorted list, in decreasing intervals [lo, hi]. 119 int lo = 0, hi = _intrinsics->length()-1; 120 while (lo <= hi) { 121 int mid = (uint)(hi + lo) / 2; 122 ciMethod* mid_m = _intrinsics->at(mid)->method(); 123 if (m < mid_m) { 124 hi = mid-1; 125 } else if (m > mid_m) { 126 lo = mid+1; 127 } else { 128 // look at minor sort key 129 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 130 if (is_virtual < mid_virt) { 131 hi = mid-1; 132 } else if (is_virtual > mid_virt) { 133 lo = mid+1; 134 } else { 135 return mid; // exact match 136 } 137 } 138 } 139 return lo; // inexact match 140 } 141 142 void Compile::register_intrinsic(CallGenerator* cg) { 143 if (_intrinsics == NULL) { 144 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 145 } 146 // This code is stolen from ciObjectFactory::insert. 147 // Really, GrowableArray should have methods for 148 // insert_at, remove_at, and binary_search. 149 int len = _intrinsics->length(); 150 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 151 if (index == len) { 152 _intrinsics->append(cg); 153 } else { 154 #ifdef ASSERT 155 CallGenerator* oldcg = _intrinsics->at(index); 156 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 157 #endif 158 _intrinsics->append(_intrinsics->at(len-1)); 159 int pos; 160 for (pos = len-2; pos >= index; pos--) { 161 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 162 } 163 _intrinsics->at_put(index, cg); 164 } 165 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 166 } 167 168 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 169 assert(m->is_loaded(), "don't try this on unloaded methods"); 170 if (_intrinsics != NULL) { 171 int index = intrinsic_insertion_index(m, is_virtual); 172 if (index < _intrinsics->length() 173 && _intrinsics->at(index)->method() == m 174 && _intrinsics->at(index)->is_virtual() == is_virtual) { 175 return _intrinsics->at(index); 176 } 177 } 178 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 179 if (m->intrinsic_id() != vmIntrinsics::_none && 180 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 181 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 182 if (cg != NULL) { 183 // Save it for next time: 184 register_intrinsic(cg); 185 return cg; 186 } else { 187 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 188 } 189 } 190 return NULL; 191 } 192 193 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 194 // in library_call.cpp. 195 196 197 #ifndef PRODUCT 198 // statistics gathering... 199 200 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 201 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 202 203 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 204 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 205 int oflags = _intrinsic_hist_flags[id]; 206 assert(flags != 0, "what happened?"); 207 if (is_virtual) { 208 flags |= _intrinsic_virtual; 209 } 210 bool changed = (flags != oflags); 211 if ((flags & _intrinsic_worked) != 0) { 212 juint count = (_intrinsic_hist_count[id] += 1); 213 if (count == 1) { 214 changed = true; // first time 215 } 216 // increment the overall count also: 217 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 218 } 219 if (changed) { 220 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 221 // Something changed about the intrinsic's virtuality. 222 if ((flags & _intrinsic_virtual) != 0) { 223 // This is the first use of this intrinsic as a virtual call. 224 if (oflags != 0) { 225 // We already saw it as a non-virtual, so note both cases. 226 flags |= _intrinsic_both; 227 } 228 } else if ((oflags & _intrinsic_both) == 0) { 229 // This is the first use of this intrinsic as a non-virtual 230 flags |= _intrinsic_both; 231 } 232 } 233 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 234 } 235 // update the overall flags also: 236 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 237 return changed; 238 } 239 240 static char* format_flags(int flags, char* buf) { 241 buf[0] = 0; 242 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 243 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 244 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 245 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 246 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 247 if (buf[0] == 0) strcat(buf, ","); 248 assert(buf[0] == ',', "must be"); 249 return &buf[1]; 250 } 251 252 void Compile::print_intrinsic_statistics() { 253 char flagsbuf[100]; 254 ttyLocker ttyl; 255 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 256 tty->print_cr("Compiler intrinsic usage:"); 257 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 258 if (total == 0) total = 1; // avoid div0 in case of no successes 259 #define PRINT_STAT_LINE(name, c, f) \ 260 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 261 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 262 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 263 int flags = _intrinsic_hist_flags[id]; 264 juint count = _intrinsic_hist_count[id]; 265 if ((flags | count) != 0) { 266 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 267 } 268 } 269 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 270 if (xtty != NULL) xtty->tail("statistics"); 271 } 272 273 void Compile::print_statistics() { 274 { ttyLocker ttyl; 275 if (xtty != NULL) xtty->head("statistics type='opto'"); 276 Parse::print_statistics(); 277 PhaseCCP::print_statistics(); 278 PhaseRegAlloc::print_statistics(); 279 Scheduling::print_statistics(); 280 PhasePeephole::print_statistics(); 281 PhaseIdealLoop::print_statistics(); 282 if (xtty != NULL) xtty->tail("statistics"); 283 } 284 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 285 // put this under its own <statistics> element. 286 print_intrinsic_statistics(); 287 } 288 } 289 #endif //PRODUCT 290 291 // Support for bundling info 292 Bundle* Compile::node_bundling(const Node *n) { 293 assert(valid_bundle_info(n), "oob"); 294 return &_node_bundling_base[n->_idx]; 295 } 296 297 bool Compile::valid_bundle_info(const Node *n) { 298 return (_node_bundling_limit > n->_idx); 299 } 300 301 302 void Compile::gvn_replace_by(Node* n, Node* nn) { 303 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 304 Node* use = n->last_out(i); 305 bool is_in_table = initial_gvn()->hash_delete(use); 306 uint uses_found = 0; 307 for (uint j = 0; j < use->len(); j++) { 308 if (use->in(j) == n) { 309 if (j < use->req()) 310 use->set_req(j, nn); 311 else 312 use->set_prec(j, nn); 313 uses_found++; 314 } 315 } 316 if (is_in_table) { 317 // reinsert into table 318 initial_gvn()->hash_find_insert(use); 319 } 320 record_for_igvn(use); 321 i -= uses_found; // we deleted 1 or more copies of this edge 322 } 323 } 324 325 326 static inline bool not_a_node(const Node* n) { 327 if (n == NULL) return true; 328 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 329 if (*(address*)n == badAddress) return true; // kill by Node::destruct 330 return false; 331 } 332 333 // Identify all nodes that are reachable from below, useful. 334 // Use breadth-first pass that records state in a Unique_Node_List, 335 // recursive traversal is slower. 336 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 337 int estimated_worklist_size = unique(); 338 useful.map( estimated_worklist_size, NULL ); // preallocate space 339 340 // Initialize worklist 341 if (root() != NULL) { useful.push(root()); } 342 // If 'top' is cached, declare it useful to preserve cached node 343 if( cached_top_node() ) { useful.push(cached_top_node()); } 344 345 // Push all useful nodes onto the list, breadthfirst 346 for( uint next = 0; next < useful.size(); ++next ) { 347 assert( next < unique(), "Unique useful nodes < total nodes"); 348 Node *n = useful.at(next); 349 uint max = n->len(); 350 for( uint i = 0; i < max; ++i ) { 351 Node *m = n->in(i); 352 if (not_a_node(m)) continue; 353 useful.push(m); 354 } 355 } 356 } 357 358 // Update dead_node_list with any missing dead nodes using useful 359 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 360 void Compile::update_dead_node_list(Unique_Node_List &useful) { 361 uint max_idx = unique(); 362 VectorSet& useful_node_set = useful.member_set(); 363 364 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 365 // If node with index node_idx is not in useful set, 366 // mark it as dead in dead node list. 367 if (! useful_node_set.test(node_idx) ) { 368 record_dead_node(node_idx); 369 } 370 } 371 } 372 373 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 374 int shift = 0; 375 for (int i = 0; i < inlines->length(); i++) { 376 CallGenerator* cg = inlines->at(i); 377 CallNode* call = cg->call_node(); 378 if (shift > 0) { 379 inlines->at_put(i-shift, cg); 380 } 381 if (!useful.member(call)) { 382 shift++; 383 } 384 } 385 inlines->trunc_to(inlines->length()-shift); 386 } 387 388 // Disconnect all useless nodes by disconnecting those at the boundary. 389 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 390 uint next = 0; 391 while (next < useful.size()) { 392 Node *n = useful.at(next++); 393 // Use raw traversal of out edges since this code removes out edges 394 int max = n->outcnt(); 395 for (int j = 0; j < max; ++j) { 396 Node* child = n->raw_out(j); 397 if (! useful.member(child)) { 398 assert(!child->is_top() || child != top(), 399 "If top is cached in Compile object it is in useful list"); 400 // Only need to remove this out-edge to the useless node 401 n->raw_del_out(j); 402 --j; 403 --max; 404 } 405 } 406 if (n->outcnt() == 1 && n->has_special_unique_user()) { 407 record_for_igvn(n->unique_out()); 408 } 409 } 410 // Remove useless macro and predicate opaq nodes 411 for (int i = C->macro_count()-1; i >= 0; i--) { 412 Node* n = C->macro_node(i); 413 if (!useful.member(n)) { 414 remove_macro_node(n); 415 } 416 } 417 // Remove useless expensive node 418 for (int i = C->expensive_count()-1; i >= 0; i--) { 419 Node* n = C->expensive_node(i); 420 if (!useful.member(n)) { 421 remove_expensive_node(n); 422 } 423 } 424 // clean up the late inline lists 425 remove_useless_late_inlines(&_string_late_inlines, useful); 426 remove_useless_late_inlines(&_boxing_late_inlines, useful); 427 remove_useless_late_inlines(&_late_inlines, useful); 428 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 429 } 430 431 //------------------------------frame_size_in_words----------------------------- 432 // frame_slots in units of words 433 int Compile::frame_size_in_words() const { 434 // shift is 0 in LP32 and 1 in LP64 435 const int shift = (LogBytesPerWord - LogBytesPerInt); 436 int words = _frame_slots >> shift; 437 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 438 return words; 439 } 440 441 // ============================================================================ 442 //------------------------------CompileWrapper--------------------------------- 443 class CompileWrapper : public StackObj { 444 Compile *const _compile; 445 public: 446 CompileWrapper(Compile* compile); 447 448 ~CompileWrapper(); 449 }; 450 451 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 452 // the Compile* pointer is stored in the current ciEnv: 453 ciEnv* env = compile->env(); 454 assert(env == ciEnv::current(), "must already be a ciEnv active"); 455 assert(env->compiler_data() == NULL, "compile already active?"); 456 env->set_compiler_data(compile); 457 assert(compile == Compile::current(), "sanity"); 458 459 compile->set_type_dict(NULL); 460 compile->set_type_hwm(NULL); 461 compile->set_type_last_size(0); 462 compile->set_last_tf(NULL, NULL); 463 compile->set_indexSet_arena(NULL); 464 compile->set_indexSet_free_block_list(NULL); 465 compile->init_type_arena(); 466 Type::Initialize(compile); 467 _compile->set_scratch_buffer_blob(NULL); 468 _compile->begin_method(); 469 } 470 CompileWrapper::~CompileWrapper() { 471 _compile->end_method(); 472 if (_compile->scratch_buffer_blob() != NULL) 473 BufferBlob::free(_compile->scratch_buffer_blob()); 474 _compile->env()->set_compiler_data(NULL); 475 } 476 477 478 //----------------------------print_compile_messages--------------------------- 479 void Compile::print_compile_messages() { 480 #ifndef PRODUCT 481 // Check if recompiling 482 if (_subsume_loads == false && PrintOpto) { 483 // Recompiling without allowing machine instructions to subsume loads 484 tty->print_cr("*********************************************************"); 485 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 486 tty->print_cr("*********************************************************"); 487 } 488 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 489 // Recompiling without escape analysis 490 tty->print_cr("*********************************************************"); 491 tty->print_cr("** Bailout: Recompile without escape analysis **"); 492 tty->print_cr("*********************************************************"); 493 } 494 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 495 // Recompiling without boxing elimination 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (env()->break_at_compile()) { 501 // Open the debugger when compiling this method. 502 tty->print("### Breaking when compiling: "); 503 method()->print_short_name(); 504 tty->cr(); 505 BREAKPOINT; 506 } 507 508 if( PrintOpto ) { 509 if (is_osr_compilation()) { 510 tty->print("[OSR]%3d", _compile_id); 511 } else { 512 tty->print("%3d", _compile_id); 513 } 514 } 515 #endif 516 } 517 518 519 //-----------------------init_scratch_buffer_blob------------------------------ 520 // Construct a temporary BufferBlob and cache it for this compile. 521 void Compile::init_scratch_buffer_blob(int const_size) { 522 // If there is already a scratch buffer blob allocated and the 523 // constant section is big enough, use it. Otherwise free the 524 // current and allocate a new one. 525 BufferBlob* blob = scratch_buffer_blob(); 526 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 527 // Use the current blob. 528 } else { 529 if (blob != NULL) { 530 BufferBlob::free(blob); 531 } 532 533 ResourceMark rm; 534 _scratch_const_size = const_size; 535 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 536 blob = BufferBlob::create("Compile::scratch_buffer", size); 537 // Record the buffer blob for next time. 538 set_scratch_buffer_blob(blob); 539 // Have we run out of code space? 540 if (scratch_buffer_blob() == NULL) { 541 // Let CompilerBroker disable further compilations. 542 record_failure("Not enough space for scratch buffer in CodeCache"); 543 return; 544 } 545 } 546 547 // Initialize the relocation buffers 548 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 549 set_scratch_locs_memory(locs_buf); 550 } 551 552 553 //-----------------------scratch_emit_size------------------------------------- 554 // Helper function that computes size by emitting code 555 uint Compile::scratch_emit_size(const Node* n) { 556 // Start scratch_emit_size section. 557 set_in_scratch_emit_size(true); 558 559 // Emit into a trash buffer and count bytes emitted. 560 // This is a pretty expensive way to compute a size, 561 // but it works well enough if seldom used. 562 // All common fixed-size instructions are given a size 563 // method by the AD file. 564 // Note that the scratch buffer blob and locs memory are 565 // allocated at the beginning of the compile task, and 566 // may be shared by several calls to scratch_emit_size. 567 // The allocation of the scratch buffer blob is particularly 568 // expensive, since it has to grab the code cache lock. 569 BufferBlob* blob = this->scratch_buffer_blob(); 570 assert(blob != NULL, "Initialize BufferBlob at start"); 571 assert(blob->size() > MAX_inst_size, "sanity"); 572 relocInfo* locs_buf = scratch_locs_memory(); 573 address blob_begin = blob->content_begin(); 574 address blob_end = (address)locs_buf; 575 assert(blob->content_contains(blob_end), "sanity"); 576 CodeBuffer buf(blob_begin, blob_end - blob_begin); 577 buf.initialize_consts_size(_scratch_const_size); 578 buf.initialize_stubs_size(MAX_stubs_size); 579 assert(locs_buf != NULL, "sanity"); 580 int lsize = MAX_locs_size / 3; 581 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 582 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 583 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 584 585 // Do the emission. 586 587 Label fakeL; // Fake label for branch instructions. 588 Label* saveL = NULL; 589 uint save_bnum = 0; 590 bool is_branch = n->is_MachBranch(); 591 if (is_branch) { 592 MacroAssembler masm(&buf); 593 masm.bind(fakeL); 594 n->as_MachBranch()->save_label(&saveL, &save_bnum); 595 n->as_MachBranch()->label_set(&fakeL, 0); 596 } 597 n->emit(buf, this->regalloc()); 598 if (is_branch) // Restore label. 599 n->as_MachBranch()->label_set(saveL, save_bnum); 600 601 // End scratch_emit_size section. 602 set_in_scratch_emit_size(false); 603 604 return buf.insts_size(); 605 } 606 607 608 // ============================================================================ 609 //------------------------------Compile standard------------------------------- 610 debug_only( int Compile::_debug_idx = 100000; ) 611 612 // Compile a method. entry_bci is -1 for normal compilations and indicates 613 // the continuation bci for on stack replacement. 614 615 616 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 617 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 618 : Phase(Compiler), 619 _env(ci_env), 620 _log(ci_env->log()), 621 _compile_id(ci_env->compile_id()), 622 _save_argument_registers(false), 623 _stub_name(NULL), 624 _stub_function(NULL), 625 _stub_entry_point(NULL), 626 _method(target), 627 _entry_bci(osr_bci), 628 _initial_gvn(NULL), 629 _for_igvn(NULL), 630 _warm_calls(NULL), 631 _subsume_loads(subsume_loads), 632 _do_escape_analysis(do_escape_analysis), 633 _eliminate_boxing(eliminate_boxing), 634 _failure_reason(NULL), 635 _code_buffer("Compile::Fill_buffer"), 636 _orig_pc_slot(0), 637 _orig_pc_slot_offset_in_bytes(0), 638 _has_method_handle_invokes(false), 639 _mach_constant_base_node(NULL), 640 _node_bundling_limit(0), 641 _node_bundling_base(NULL), 642 _java_calls(0), 643 _inner_loops(0), 644 _scratch_const_size(-1), 645 _in_scratch_emit_size(false), 646 _dead_node_list(comp_arena()), 647 _dead_node_count(0), 648 #ifndef PRODUCT 649 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 650 _printer(IdealGraphPrinter::printer()), 651 #endif 652 _congraph(NULL), 653 _late_inlines(comp_arena(), 2, 0, NULL), 654 _string_late_inlines(comp_arena(), 2, 0, NULL), 655 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 656 _late_inlines_pos(0), 657 _number_of_mh_late_inlines(0), 658 _inlining_progress(false), 659 _inlining_incrementally(false), 660 _print_inlining_list(NULL), 661 _print_inlining_idx(0), 662 _preserve_jvm_state(0) { 663 C = this; 664 665 CompileWrapper cw(this); 666 #ifndef PRODUCT 667 if (TimeCompiler2) { 668 tty->print(" "); 669 target->holder()->name()->print(); 670 tty->print("."); 671 target->print_short_name(); 672 tty->print(" "); 673 } 674 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 675 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 676 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 677 if (!print_opto_assembly) { 678 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 679 if (print_assembly && !Disassembler::can_decode()) { 680 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 681 print_opto_assembly = true; 682 } 683 } 684 set_print_assembly(print_opto_assembly); 685 set_parsed_irreducible_loop(false); 686 #endif 687 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 688 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 689 690 if (ProfileTraps) { 691 // Make sure the method being compiled gets its own MDO, 692 // so we can at least track the decompile_count(). 693 method()->ensure_method_data(); 694 } 695 696 Init(::AliasLevel); 697 698 699 print_compile_messages(); 700 701 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) 702 _ilt = InlineTree::build_inline_tree_root(); 703 else 704 _ilt = NULL; 705 706 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 707 assert(num_alias_types() >= AliasIdxRaw, ""); 708 709 #define MINIMUM_NODE_HASH 1023 710 // Node list that Iterative GVN will start with 711 Unique_Node_List for_igvn(comp_arena()); 712 set_for_igvn(&for_igvn); 713 714 // GVN that will be run immediately on new nodes 715 uint estimated_size = method()->code_size()*4+64; 716 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 717 PhaseGVN gvn(node_arena(), estimated_size); 718 set_initial_gvn(&gvn); 719 720 if (print_inlining() || print_intrinsics()) { 721 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 722 } 723 { // Scope for timing the parser 724 TracePhase t3("parse", &_t_parser, true); 725 726 // Put top into the hash table ASAP. 727 initial_gvn()->transform_no_reclaim(top()); 728 729 // Set up tf(), start(), and find a CallGenerator. 730 CallGenerator* cg = NULL; 731 if (is_osr_compilation()) { 732 const TypeTuple *domain = StartOSRNode::osr_domain(); 733 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 734 init_tf(TypeFunc::make(domain, range)); 735 StartNode* s = new (this) StartOSRNode(root(), domain); 736 initial_gvn()->set_type_bottom(s); 737 init_start(s); 738 cg = CallGenerator::for_osr(method(), entry_bci()); 739 } else { 740 // Normal case. 741 init_tf(TypeFunc::make(method())); 742 StartNode* s = new (this) StartNode(root(), tf()->domain()); 743 initial_gvn()->set_type_bottom(s); 744 init_start(s); 745 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 746 // With java.lang.ref.reference.get() we must go through the 747 // intrinsic when G1 is enabled - even when get() is the root 748 // method of the compile - so that, if necessary, the value in 749 // the referent field of the reference object gets recorded by 750 // the pre-barrier code. 751 // Specifically, if G1 is enabled, the value in the referent 752 // field is recorded by the G1 SATB pre barrier. This will 753 // result in the referent being marked live and the reference 754 // object removed from the list of discovered references during 755 // reference processing. 756 cg = find_intrinsic(method(), false); 757 } 758 if (cg == NULL) { 759 float past_uses = method()->interpreter_invocation_count(); 760 float expected_uses = past_uses; 761 cg = CallGenerator::for_inline(method(), expected_uses); 762 } 763 } 764 if (failing()) return; 765 if (cg == NULL) { 766 record_method_not_compilable_all_tiers("cannot parse method"); 767 return; 768 } 769 JVMState* jvms = build_start_state(start(), tf()); 770 if ((jvms = cg->generate(jvms, NULL)) == NULL) { 771 record_method_not_compilable("method parse failed"); 772 return; 773 } 774 GraphKit kit(jvms); 775 776 if (!kit.stopped()) { 777 // Accept return values, and transfer control we know not where. 778 // This is done by a special, unique ReturnNode bound to root. 779 return_values(kit.jvms()); 780 } 781 782 if (kit.has_exceptions()) { 783 // Any exceptions that escape from this call must be rethrown 784 // to whatever caller is dynamically above us on the stack. 785 // This is done by a special, unique RethrowNode bound to root. 786 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 787 } 788 789 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 790 791 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 792 inline_string_calls(true); 793 } 794 795 if (failing()) return; 796 797 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 798 799 // Remove clutter produced by parsing. 800 if (!failing()) { 801 ResourceMark rm; 802 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 803 } 804 } 805 806 // Note: Large methods are capped off in do_one_bytecode(). 807 if (failing()) return; 808 809 // After parsing, node notes are no longer automagic. 810 // They must be propagated by register_new_node_with_optimizer(), 811 // clone(), or the like. 812 set_default_node_notes(NULL); 813 814 for (;;) { 815 int successes = Inline_Warm(); 816 if (failing()) return; 817 if (successes == 0) break; 818 } 819 820 // Drain the list. 821 Finish_Warm(); 822 #ifndef PRODUCT 823 if (_printer) { 824 _printer->print_inlining(this); 825 } 826 #endif 827 828 if (failing()) return; 829 NOT_PRODUCT( verify_graph_edges(); ) 830 831 // Now optimize 832 Optimize(); 833 if (failing()) return; 834 NOT_PRODUCT( verify_graph_edges(); ) 835 836 #ifndef PRODUCT 837 if (PrintIdeal) { 838 ttyLocker ttyl; // keep the following output all in one block 839 // This output goes directly to the tty, not the compiler log. 840 // To enable tools to match it up with the compilation activity, 841 // be sure to tag this tty output with the compile ID. 842 if (xtty != NULL) { 843 xtty->head("ideal compile_id='%d'%s", compile_id(), 844 is_osr_compilation() ? " compile_kind='osr'" : 845 ""); 846 } 847 root()->dump(9999); 848 if (xtty != NULL) { 849 xtty->tail("ideal"); 850 } 851 } 852 #endif 853 854 NOT_PRODUCT( verify_barriers(); ) 855 // Now that we know the size of all the monitors we can add a fixed slot 856 // for the original deopt pc. 857 858 _orig_pc_slot = fixed_slots(); 859 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 860 set_fixed_slots(next_slot); 861 862 // Now generate code 863 Code_Gen(); 864 if (failing()) return; 865 866 // Check if we want to skip execution of all compiled code. 867 { 868 #ifndef PRODUCT 869 if (OptoNoExecute) { 870 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 871 return; 872 } 873 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 874 #endif 875 876 if (is_osr_compilation()) { 877 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 878 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 879 } else { 880 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 881 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 882 } 883 884 env()->register_method(_method, _entry_bci, 885 &_code_offsets, 886 _orig_pc_slot_offset_in_bytes, 887 code_buffer(), 888 frame_size_in_words(), _oop_map_set, 889 &_handler_table, &_inc_table, 890 compiler, 891 env()->comp_level(), 892 has_unsafe_access(), 893 SharedRuntime::is_wide_vector(max_vector_size()) 894 ); 895 896 if (log() != NULL) // Print code cache state into compiler log 897 log()->code_cache_state(); 898 } 899 } 900 901 //------------------------------Compile---------------------------------------- 902 // Compile a runtime stub 903 Compile::Compile( ciEnv* ci_env, 904 TypeFunc_generator generator, 905 address stub_function, 906 const char *stub_name, 907 int is_fancy_jump, 908 bool pass_tls, 909 bool save_arg_registers, 910 bool return_pc ) 911 : Phase(Compiler), 912 _env(ci_env), 913 _log(ci_env->log()), 914 _compile_id(0), 915 _save_argument_registers(save_arg_registers), 916 _method(NULL), 917 _stub_name(stub_name), 918 _stub_function(stub_function), 919 _stub_entry_point(NULL), 920 _entry_bci(InvocationEntryBci), 921 _initial_gvn(NULL), 922 _for_igvn(NULL), 923 _warm_calls(NULL), 924 _orig_pc_slot(0), 925 _orig_pc_slot_offset_in_bytes(0), 926 _subsume_loads(true), 927 _do_escape_analysis(false), 928 _eliminate_boxing(false), 929 _failure_reason(NULL), 930 _code_buffer("Compile::Fill_buffer"), 931 _has_method_handle_invokes(false), 932 _mach_constant_base_node(NULL), 933 _node_bundling_limit(0), 934 _node_bundling_base(NULL), 935 _java_calls(0), 936 _inner_loops(0), 937 #ifndef PRODUCT 938 _trace_opto_output(TraceOptoOutput), 939 _printer(NULL), 940 #endif 941 _dead_node_list(comp_arena()), 942 _dead_node_count(0), 943 _congraph(NULL), 944 _number_of_mh_late_inlines(0), 945 _inlining_progress(false), 946 _inlining_incrementally(false), 947 _print_inlining_list(NULL), 948 _print_inlining_idx(0), 949 _preserve_jvm_state(0) { 950 C = this; 951 952 #ifndef PRODUCT 953 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 954 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 955 set_print_assembly(PrintFrameConverterAssembly); 956 set_parsed_irreducible_loop(false); 957 #endif 958 CompileWrapper cw(this); 959 Init(/*AliasLevel=*/ 0); 960 init_tf((*generator)()); 961 962 { 963 // The following is a dummy for the sake of GraphKit::gen_stub 964 Unique_Node_List for_igvn(comp_arena()); 965 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 966 PhaseGVN gvn(Thread::current()->resource_area(),255); 967 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 968 gvn.transform_no_reclaim(top()); 969 970 GraphKit kit; 971 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 972 } 973 974 NOT_PRODUCT( verify_graph_edges(); ) 975 Code_Gen(); 976 if (failing()) return; 977 978 979 // Entry point will be accessed using compile->stub_entry_point(); 980 if (code_buffer() == NULL) { 981 Matcher::soft_match_failure(); 982 } else { 983 if (PrintAssembly && (WizardMode || Verbose)) 984 tty->print_cr("### Stub::%s", stub_name); 985 986 if (!failing()) { 987 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 988 989 // Make the NMethod 990 // For now we mark the frame as never safe for profile stackwalking 991 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 992 code_buffer(), 993 CodeOffsets::frame_never_safe, 994 // _code_offsets.value(CodeOffsets::Frame_Complete), 995 frame_size_in_words(), 996 _oop_map_set, 997 save_arg_registers); 998 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 999 1000 _stub_entry_point = rs->entry_point(); 1001 } 1002 } 1003 } 1004 1005 //------------------------------Init------------------------------------------- 1006 // Prepare for a single compilation 1007 void Compile::Init(int aliaslevel) { 1008 _unique = 0; 1009 _regalloc = NULL; 1010 1011 _tf = NULL; // filled in later 1012 _top = NULL; // cached later 1013 _matcher = NULL; // filled in later 1014 _cfg = NULL; // filled in later 1015 1016 set_24_bit_selection_and_mode(Use24BitFP, false); 1017 1018 _node_note_array = NULL; 1019 _default_node_notes = NULL; 1020 1021 _immutable_memory = NULL; // filled in at first inquiry 1022 1023 // Globally visible Nodes 1024 // First set TOP to NULL to give safe behavior during creation of RootNode 1025 set_cached_top_node(NULL); 1026 set_root(new (this) RootNode()); 1027 // Now that you have a Root to point to, create the real TOP 1028 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1029 set_recent_alloc(NULL, NULL); 1030 1031 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1032 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1033 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1034 env()->set_dependencies(new Dependencies(env())); 1035 1036 _fixed_slots = 0; 1037 set_has_split_ifs(false); 1038 set_has_loops(has_method() && method()->has_loops()); // first approximation 1039 set_has_stringbuilder(false); 1040 set_has_boxed_value(false); 1041 _trap_can_recompile = false; // no traps emitted yet 1042 _major_progress = true; // start out assuming good things will happen 1043 set_has_unsafe_access(false); 1044 set_max_vector_size(0); 1045 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1046 set_decompile_count(0); 1047 1048 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1049 set_num_loop_opts(LoopOptsCount); 1050 set_do_inlining(Inline); 1051 set_max_inline_size(MaxInlineSize); 1052 set_freq_inline_size(FreqInlineSize); 1053 set_do_scheduling(OptoScheduling); 1054 set_do_count_invocations(false); 1055 set_do_method_data_update(false); 1056 1057 if (debug_info()->recording_non_safepoints()) { 1058 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1059 (comp_arena(), 8, 0, NULL)); 1060 set_default_node_notes(Node_Notes::make(this)); 1061 } 1062 1063 // // -- Initialize types before each compile -- 1064 // // Update cached type information 1065 // if( _method && _method->constants() ) 1066 // Type::update_loaded_types(_method, _method->constants()); 1067 1068 // Init alias_type map. 1069 if (!_do_escape_analysis && aliaslevel == 3) 1070 aliaslevel = 2; // No unique types without escape analysis 1071 _AliasLevel = aliaslevel; 1072 const int grow_ats = 16; 1073 _max_alias_types = grow_ats; 1074 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1075 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1076 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1077 { 1078 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1079 } 1080 // Initialize the first few types. 1081 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1082 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1083 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1084 _num_alias_types = AliasIdxRaw+1; 1085 // Zero out the alias type cache. 1086 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1087 // A NULL adr_type hits in the cache right away. Preload the right answer. 1088 probe_alias_cache(NULL)->_index = AliasIdxTop; 1089 1090 _intrinsics = NULL; 1091 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1092 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1093 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1094 register_library_intrinsics(); 1095 } 1096 1097 //---------------------------init_start---------------------------------------- 1098 // Install the StartNode on this compile object. 1099 void Compile::init_start(StartNode* s) { 1100 if (failing()) 1101 return; // already failing 1102 assert(s == start(), ""); 1103 } 1104 1105 StartNode* Compile::start() const { 1106 assert(!failing(), ""); 1107 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1108 Node* start = root()->fast_out(i); 1109 if( start->is_Start() ) 1110 return start->as_Start(); 1111 } 1112 ShouldNotReachHere(); 1113 return NULL; 1114 } 1115 1116 //-------------------------------immutable_memory------------------------------------- 1117 // Access immutable memory 1118 Node* Compile::immutable_memory() { 1119 if (_immutable_memory != NULL) { 1120 return _immutable_memory; 1121 } 1122 StartNode* s = start(); 1123 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1124 Node *p = s->fast_out(i); 1125 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1126 _immutable_memory = p; 1127 return _immutable_memory; 1128 } 1129 } 1130 ShouldNotReachHere(); 1131 return NULL; 1132 } 1133 1134 //----------------------set_cached_top_node------------------------------------ 1135 // Install the cached top node, and make sure Node::is_top works correctly. 1136 void Compile::set_cached_top_node(Node* tn) { 1137 if (tn != NULL) verify_top(tn); 1138 Node* old_top = _top; 1139 _top = tn; 1140 // Calling Node::setup_is_top allows the nodes the chance to adjust 1141 // their _out arrays. 1142 if (_top != NULL) _top->setup_is_top(); 1143 if (old_top != NULL) old_top->setup_is_top(); 1144 assert(_top == NULL || top()->is_top(), ""); 1145 } 1146 1147 #ifdef ASSERT 1148 uint Compile::count_live_nodes_by_graph_walk() { 1149 Unique_Node_List useful(comp_arena()); 1150 // Get useful node list by walking the graph. 1151 identify_useful_nodes(useful); 1152 return useful.size(); 1153 } 1154 1155 void Compile::print_missing_nodes() { 1156 1157 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1158 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1159 return; 1160 } 1161 1162 // This is an expensive function. It is executed only when the user 1163 // specifies VerifyIdealNodeCount option or otherwise knows the 1164 // additional work that needs to be done to identify reachable nodes 1165 // by walking the flow graph and find the missing ones using 1166 // _dead_node_list. 1167 1168 Unique_Node_List useful(comp_arena()); 1169 // Get useful node list by walking the graph. 1170 identify_useful_nodes(useful); 1171 1172 uint l_nodes = C->live_nodes(); 1173 uint l_nodes_by_walk = useful.size(); 1174 1175 if (l_nodes != l_nodes_by_walk) { 1176 if (_log != NULL) { 1177 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1178 _log->stamp(); 1179 _log->end_head(); 1180 } 1181 VectorSet& useful_member_set = useful.member_set(); 1182 int last_idx = l_nodes_by_walk; 1183 for (int i = 0; i < last_idx; i++) { 1184 if (useful_member_set.test(i)) { 1185 if (_dead_node_list.test(i)) { 1186 if (_log != NULL) { 1187 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1188 } 1189 if (PrintIdealNodeCount) { 1190 // Print the log message to tty 1191 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1192 useful.at(i)->dump(); 1193 } 1194 } 1195 } 1196 else if (! _dead_node_list.test(i)) { 1197 if (_log != NULL) { 1198 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1199 } 1200 if (PrintIdealNodeCount) { 1201 // Print the log message to tty 1202 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1203 } 1204 } 1205 } 1206 if (_log != NULL) { 1207 _log->tail("mismatched_nodes"); 1208 } 1209 } 1210 } 1211 #endif 1212 1213 #ifndef PRODUCT 1214 void Compile::verify_top(Node* tn) const { 1215 if (tn != NULL) { 1216 assert(tn->is_Con(), "top node must be a constant"); 1217 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1218 assert(tn->in(0) != NULL, "must have live top node"); 1219 } 1220 } 1221 #endif 1222 1223 1224 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1225 1226 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1227 guarantee(arr != NULL, ""); 1228 int num_blocks = arr->length(); 1229 if (grow_by < num_blocks) grow_by = num_blocks; 1230 int num_notes = grow_by * _node_notes_block_size; 1231 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1232 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1233 while (num_notes > 0) { 1234 arr->append(notes); 1235 notes += _node_notes_block_size; 1236 num_notes -= _node_notes_block_size; 1237 } 1238 assert(num_notes == 0, "exact multiple, please"); 1239 } 1240 1241 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1242 if (source == NULL || dest == NULL) return false; 1243 1244 if (dest->is_Con()) 1245 return false; // Do not push debug info onto constants. 1246 1247 #ifdef ASSERT 1248 // Leave a bread crumb trail pointing to the original node: 1249 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1250 dest->set_debug_orig(source); 1251 } 1252 #endif 1253 1254 if (node_note_array() == NULL) 1255 return false; // Not collecting any notes now. 1256 1257 // This is a copy onto a pre-existing node, which may already have notes. 1258 // If both nodes have notes, do not overwrite any pre-existing notes. 1259 Node_Notes* source_notes = node_notes_at(source->_idx); 1260 if (source_notes == NULL || source_notes->is_clear()) return false; 1261 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1262 if (dest_notes == NULL || dest_notes->is_clear()) { 1263 return set_node_notes_at(dest->_idx, source_notes); 1264 } 1265 1266 Node_Notes merged_notes = (*source_notes); 1267 // The order of operations here ensures that dest notes will win... 1268 merged_notes.update_from(dest_notes); 1269 return set_node_notes_at(dest->_idx, &merged_notes); 1270 } 1271 1272 1273 //--------------------------allow_range_check_smearing------------------------- 1274 // Gating condition for coalescing similar range checks. 1275 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1276 // single covering check that is at least as strong as any of them. 1277 // If the optimization succeeds, the simplified (strengthened) range check 1278 // will always succeed. If it fails, we will deopt, and then give up 1279 // on the optimization. 1280 bool Compile::allow_range_check_smearing() const { 1281 // If this method has already thrown a range-check, 1282 // assume it was because we already tried range smearing 1283 // and it failed. 1284 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1285 return !already_trapped; 1286 } 1287 1288 1289 //------------------------------flatten_alias_type----------------------------- 1290 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1291 int offset = tj->offset(); 1292 TypePtr::PTR ptr = tj->ptr(); 1293 1294 // Known instance (scalarizable allocation) alias only with itself. 1295 bool is_known_inst = tj->isa_oopptr() != NULL && 1296 tj->is_oopptr()->is_known_instance(); 1297 1298 // Process weird unsafe references. 1299 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1300 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1301 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1302 tj = TypeOopPtr::BOTTOM; 1303 ptr = tj->ptr(); 1304 offset = tj->offset(); 1305 } 1306 1307 // Array pointers need some flattening 1308 const TypeAryPtr *ta = tj->isa_aryptr(); 1309 if (ta && ta->is_stable()) { 1310 // Erase stability property for alias analysis. 1311 tj = ta = ta->cast_to_stable(false); 1312 } 1313 if( ta && is_known_inst ) { 1314 if ( offset != Type::OffsetBot && 1315 offset > arrayOopDesc::length_offset_in_bytes() ) { 1316 offset = Type::OffsetBot; // Flatten constant access into array body only 1317 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1318 } 1319 } else if( ta && _AliasLevel >= 2 ) { 1320 // For arrays indexed by constant indices, we flatten the alias 1321 // space to include all of the array body. Only the header, klass 1322 // and array length can be accessed un-aliased. 1323 if( offset != Type::OffsetBot ) { 1324 if( ta->const_oop() ) { // MethodData* or Method* 1325 offset = Type::OffsetBot; // Flatten constant access into array body 1326 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1327 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1328 // range is OK as-is. 1329 tj = ta = TypeAryPtr::RANGE; 1330 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1331 tj = TypeInstPtr::KLASS; // all klass loads look alike 1332 ta = TypeAryPtr::RANGE; // generic ignored junk 1333 ptr = TypePtr::BotPTR; 1334 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1335 tj = TypeInstPtr::MARK; 1336 ta = TypeAryPtr::RANGE; // generic ignored junk 1337 ptr = TypePtr::BotPTR; 1338 } else { // Random constant offset into array body 1339 offset = Type::OffsetBot; // Flatten constant access into array body 1340 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1341 } 1342 } 1343 // Arrays of fixed size alias with arrays of unknown size. 1344 if (ta->size() != TypeInt::POS) { 1345 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1346 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1347 } 1348 // Arrays of known objects become arrays of unknown objects. 1349 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1350 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1351 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1352 } 1353 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1354 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1355 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1356 } 1357 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1358 // cannot be distinguished by bytecode alone. 1359 if (ta->elem() == TypeInt::BOOL) { 1360 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1361 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1362 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1363 } 1364 // During the 2nd round of IterGVN, NotNull castings are removed. 1365 // Make sure the Bottom and NotNull variants alias the same. 1366 // Also, make sure exact and non-exact variants alias the same. 1367 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1368 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1369 } 1370 } 1371 1372 // Oop pointers need some flattening 1373 const TypeInstPtr *to = tj->isa_instptr(); 1374 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1375 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1376 if( ptr == TypePtr::Constant ) { 1377 if (to->klass() != ciEnv::current()->Class_klass() || 1378 offset < k->size_helper() * wordSize) { 1379 // No constant oop pointers (such as Strings); they alias with 1380 // unknown strings. 1381 assert(!is_known_inst, "not scalarizable allocation"); 1382 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1383 } 1384 } else if( is_known_inst ) { 1385 tj = to; // Keep NotNull and klass_is_exact for instance type 1386 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1387 // During the 2nd round of IterGVN, NotNull castings are removed. 1388 // Make sure the Bottom and NotNull variants alias the same. 1389 // Also, make sure exact and non-exact variants alias the same. 1390 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1391 } 1392 if (to->speculative() != NULL) { 1393 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1394 } 1395 // Canonicalize the holder of this field 1396 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1397 // First handle header references such as a LoadKlassNode, even if the 1398 // object's klass is unloaded at compile time (4965979). 1399 if (!is_known_inst) { // Do it only for non-instance types 1400 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1401 } 1402 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1403 // Static fields are in the space above the normal instance 1404 // fields in the java.lang.Class instance. 1405 if (to->klass() != ciEnv::current()->Class_klass()) { 1406 to = NULL; 1407 tj = TypeOopPtr::BOTTOM; 1408 offset = tj->offset(); 1409 } 1410 } else { 1411 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1412 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1413 if( is_known_inst ) { 1414 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1415 } else { 1416 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1417 } 1418 } 1419 } 1420 } 1421 1422 // Klass pointers to object array klasses need some flattening 1423 const TypeKlassPtr *tk = tj->isa_klassptr(); 1424 if( tk ) { 1425 // If we are referencing a field within a Klass, we need 1426 // to assume the worst case of an Object. Both exact and 1427 // inexact types must flatten to the same alias class so 1428 // use NotNull as the PTR. 1429 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1430 1431 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1432 TypeKlassPtr::OBJECT->klass(), 1433 offset); 1434 } 1435 1436 ciKlass* klass = tk->klass(); 1437 if( klass->is_obj_array_klass() ) { 1438 ciKlass* k = TypeAryPtr::OOPS->klass(); 1439 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1440 k = TypeInstPtr::BOTTOM->klass(); 1441 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1442 } 1443 1444 // Check for precise loads from the primary supertype array and force them 1445 // to the supertype cache alias index. Check for generic array loads from 1446 // the primary supertype array and also force them to the supertype cache 1447 // alias index. Since the same load can reach both, we need to merge 1448 // these 2 disparate memories into the same alias class. Since the 1449 // primary supertype array is read-only, there's no chance of confusion 1450 // where we bypass an array load and an array store. 1451 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1452 if (offset == Type::OffsetBot || 1453 (offset >= primary_supers_offset && 1454 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1455 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1456 offset = in_bytes(Klass::secondary_super_cache_offset()); 1457 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1458 } 1459 } 1460 1461 // Flatten all Raw pointers together. 1462 if (tj->base() == Type::RawPtr) 1463 tj = TypeRawPtr::BOTTOM; 1464 1465 if (tj->base() == Type::AnyPtr) 1466 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1467 1468 // Flatten all to bottom for now 1469 switch( _AliasLevel ) { 1470 case 0: 1471 tj = TypePtr::BOTTOM; 1472 break; 1473 case 1: // Flatten to: oop, static, field or array 1474 switch (tj->base()) { 1475 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1476 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1477 case Type::AryPtr: // do not distinguish arrays at all 1478 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1479 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1480 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1481 default: ShouldNotReachHere(); 1482 } 1483 break; 1484 case 2: // No collapsing at level 2; keep all splits 1485 case 3: // No collapsing at level 3; keep all splits 1486 break; 1487 default: 1488 Unimplemented(); 1489 } 1490 1491 offset = tj->offset(); 1492 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1493 1494 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1495 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1496 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1497 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1498 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1499 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1500 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1501 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1502 assert( tj->ptr() != TypePtr::TopPTR && 1503 tj->ptr() != TypePtr::AnyNull && 1504 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1505 // assert( tj->ptr() != TypePtr::Constant || 1506 // tj->base() == Type::RawPtr || 1507 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1508 1509 return tj; 1510 } 1511 1512 void Compile::AliasType::Init(int i, const TypePtr* at) { 1513 _index = i; 1514 _adr_type = at; 1515 _field = NULL; 1516 _element = NULL; 1517 _is_rewritable = true; // default 1518 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1519 if (atoop != NULL && atoop->is_known_instance()) { 1520 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1521 _general_index = Compile::current()->get_alias_index(gt); 1522 } else { 1523 _general_index = 0; 1524 } 1525 } 1526 1527 //---------------------------------print_on------------------------------------ 1528 #ifndef PRODUCT 1529 void Compile::AliasType::print_on(outputStream* st) { 1530 if (index() < 10) 1531 st->print("@ <%d> ", index()); 1532 else st->print("@ <%d>", index()); 1533 st->print(is_rewritable() ? " " : " RO"); 1534 int offset = adr_type()->offset(); 1535 if (offset == Type::OffsetBot) 1536 st->print(" +any"); 1537 else st->print(" +%-3d", offset); 1538 st->print(" in "); 1539 adr_type()->dump_on(st); 1540 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1541 if (field() != NULL && tjp) { 1542 if (tjp->klass() != field()->holder() || 1543 tjp->offset() != field()->offset_in_bytes()) { 1544 st->print(" != "); 1545 field()->print(); 1546 st->print(" ***"); 1547 } 1548 } 1549 } 1550 1551 void print_alias_types() { 1552 Compile* C = Compile::current(); 1553 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1554 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1555 C->alias_type(idx)->print_on(tty); 1556 tty->cr(); 1557 } 1558 } 1559 #endif 1560 1561 1562 //----------------------------probe_alias_cache-------------------------------- 1563 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1564 intptr_t key = (intptr_t) adr_type; 1565 key ^= key >> logAliasCacheSize; 1566 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1567 } 1568 1569 1570 //-----------------------------grow_alias_types-------------------------------- 1571 void Compile::grow_alias_types() { 1572 const int old_ats = _max_alias_types; // how many before? 1573 const int new_ats = old_ats; // how many more? 1574 const int grow_ats = old_ats+new_ats; // how many now? 1575 _max_alias_types = grow_ats; 1576 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1577 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1578 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1579 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1580 } 1581 1582 1583 //--------------------------------find_alias_type------------------------------ 1584 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1585 if (_AliasLevel == 0) 1586 return alias_type(AliasIdxBot); 1587 1588 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1589 if (ace->_adr_type == adr_type) { 1590 return alias_type(ace->_index); 1591 } 1592 1593 // Handle special cases. 1594 if (adr_type == NULL) return alias_type(AliasIdxTop); 1595 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1596 1597 // Do it the slow way. 1598 const TypePtr* flat = flatten_alias_type(adr_type); 1599 1600 #ifdef ASSERT 1601 assert(flat == flatten_alias_type(flat), "idempotent"); 1602 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1603 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1604 const TypeOopPtr* foop = flat->is_oopptr(); 1605 // Scalarizable allocations have exact klass always. 1606 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1607 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1608 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1609 } 1610 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1611 #endif 1612 1613 int idx = AliasIdxTop; 1614 for (int i = 0; i < num_alias_types(); i++) { 1615 if (alias_type(i)->adr_type() == flat) { 1616 idx = i; 1617 break; 1618 } 1619 } 1620 1621 if (idx == AliasIdxTop) { 1622 if (no_create) return NULL; 1623 // Grow the array if necessary. 1624 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1625 // Add a new alias type. 1626 idx = _num_alias_types++; 1627 _alias_types[idx]->Init(idx, flat); 1628 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1629 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1630 if (flat->isa_instptr()) { 1631 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1632 && flat->is_instptr()->klass() == env()->Class_klass()) 1633 alias_type(idx)->set_rewritable(false); 1634 } 1635 if (flat->isa_aryptr()) { 1636 #ifdef ASSERT 1637 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1638 // (T_BYTE has the weakest alignment and size restrictions...) 1639 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1640 #endif 1641 if (flat->offset() == TypePtr::OffsetBot) { 1642 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1643 } 1644 } 1645 if (flat->isa_klassptr()) { 1646 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1647 alias_type(idx)->set_rewritable(false); 1648 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1649 alias_type(idx)->set_rewritable(false); 1650 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1651 alias_type(idx)->set_rewritable(false); 1652 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1653 alias_type(idx)->set_rewritable(false); 1654 } 1655 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1656 // but the base pointer type is not distinctive enough to identify 1657 // references into JavaThread.) 1658 1659 // Check for final fields. 1660 const TypeInstPtr* tinst = flat->isa_instptr(); 1661 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1662 ciField* field; 1663 if (tinst->const_oop() != NULL && 1664 tinst->klass() == ciEnv::current()->Class_klass() && 1665 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1666 // static field 1667 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1668 field = k->get_field_by_offset(tinst->offset(), true); 1669 } else { 1670 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1671 field = k->get_field_by_offset(tinst->offset(), false); 1672 } 1673 assert(field == NULL || 1674 original_field == NULL || 1675 (field->holder() == original_field->holder() && 1676 field->offset() == original_field->offset() && 1677 field->is_static() == original_field->is_static()), "wrong field?"); 1678 // Set field() and is_rewritable() attributes. 1679 if (field != NULL) alias_type(idx)->set_field(field); 1680 } 1681 } 1682 1683 // Fill the cache for next time. 1684 ace->_adr_type = adr_type; 1685 ace->_index = idx; 1686 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1687 1688 // Might as well try to fill the cache for the flattened version, too. 1689 AliasCacheEntry* face = probe_alias_cache(flat); 1690 if (face->_adr_type == NULL) { 1691 face->_adr_type = flat; 1692 face->_index = idx; 1693 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1694 } 1695 1696 return alias_type(idx); 1697 } 1698 1699 1700 Compile::AliasType* Compile::alias_type(ciField* field) { 1701 const TypeOopPtr* t; 1702 if (field->is_static()) 1703 t = TypeInstPtr::make(field->holder()->java_mirror()); 1704 else 1705 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1706 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1707 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1708 return atp; 1709 } 1710 1711 1712 //------------------------------have_alias_type-------------------------------- 1713 bool Compile::have_alias_type(const TypePtr* adr_type) { 1714 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1715 if (ace->_adr_type == adr_type) { 1716 return true; 1717 } 1718 1719 // Handle special cases. 1720 if (adr_type == NULL) return true; 1721 if (adr_type == TypePtr::BOTTOM) return true; 1722 1723 return find_alias_type(adr_type, true, NULL) != NULL; 1724 } 1725 1726 //-----------------------------must_alias-------------------------------------- 1727 // True if all values of the given address type are in the given alias category. 1728 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1729 if (alias_idx == AliasIdxBot) return true; // the universal category 1730 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1731 if (alias_idx == AliasIdxTop) return false; // the empty category 1732 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1733 1734 // the only remaining possible overlap is identity 1735 int adr_idx = get_alias_index(adr_type); 1736 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1737 assert(adr_idx == alias_idx || 1738 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1739 && adr_type != TypeOopPtr::BOTTOM), 1740 "should not be testing for overlap with an unsafe pointer"); 1741 return adr_idx == alias_idx; 1742 } 1743 1744 //------------------------------can_alias-------------------------------------- 1745 // True if any values of the given address type are in the given alias category. 1746 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1747 if (alias_idx == AliasIdxTop) return false; // the empty category 1748 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1749 if (alias_idx == AliasIdxBot) return true; // the universal category 1750 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1751 1752 // the only remaining possible overlap is identity 1753 int adr_idx = get_alias_index(adr_type); 1754 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1755 return adr_idx == alias_idx; 1756 } 1757 1758 1759 1760 //---------------------------pop_warm_call------------------------------------- 1761 WarmCallInfo* Compile::pop_warm_call() { 1762 WarmCallInfo* wci = _warm_calls; 1763 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1764 return wci; 1765 } 1766 1767 //----------------------------Inline_Warm-------------------------------------- 1768 int Compile::Inline_Warm() { 1769 // If there is room, try to inline some more warm call sites. 1770 // %%% Do a graph index compaction pass when we think we're out of space? 1771 if (!InlineWarmCalls) return 0; 1772 1773 int calls_made_hot = 0; 1774 int room_to_grow = NodeCountInliningCutoff - unique(); 1775 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1776 int amount_grown = 0; 1777 WarmCallInfo* call; 1778 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1779 int est_size = (int)call->size(); 1780 if (est_size > (room_to_grow - amount_grown)) { 1781 // This one won't fit anyway. Get rid of it. 1782 call->make_cold(); 1783 continue; 1784 } 1785 call->make_hot(); 1786 calls_made_hot++; 1787 amount_grown += est_size; 1788 amount_to_grow -= est_size; 1789 } 1790 1791 if (calls_made_hot > 0) set_major_progress(); 1792 return calls_made_hot; 1793 } 1794 1795 1796 //----------------------------Finish_Warm-------------------------------------- 1797 void Compile::Finish_Warm() { 1798 if (!InlineWarmCalls) return; 1799 if (failing()) return; 1800 if (warm_calls() == NULL) return; 1801 1802 // Clean up loose ends, if we are out of space for inlining. 1803 WarmCallInfo* call; 1804 while ((call = pop_warm_call()) != NULL) { 1805 call->make_cold(); 1806 } 1807 } 1808 1809 //---------------------cleanup_loop_predicates----------------------- 1810 // Remove the opaque nodes that protect the predicates so that all unused 1811 // checks and uncommon_traps will be eliminated from the ideal graph 1812 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1813 if (predicate_count()==0) return; 1814 for (int i = predicate_count(); i > 0; i--) { 1815 Node * n = predicate_opaque1_node(i-1); 1816 assert(n->Opcode() == Op_Opaque1, "must be"); 1817 igvn.replace_node(n, n->in(1)); 1818 } 1819 assert(predicate_count()==0, "should be clean!"); 1820 } 1821 1822 // StringOpts and late inlining of string methods 1823 void Compile::inline_string_calls(bool parse_time) { 1824 { 1825 // remove useless nodes to make the usage analysis simpler 1826 ResourceMark rm; 1827 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1828 } 1829 1830 { 1831 ResourceMark rm; 1832 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1833 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1834 print_method(PHASE_AFTER_STRINGOPTS, 3); 1835 } 1836 1837 // now inline anything that we skipped the first time around 1838 if (!parse_time) { 1839 _late_inlines_pos = _late_inlines.length(); 1840 } 1841 1842 while (_string_late_inlines.length() > 0) { 1843 CallGenerator* cg = _string_late_inlines.pop(); 1844 cg->do_late_inline(); 1845 if (failing()) return; 1846 } 1847 _string_late_inlines.trunc_to(0); 1848 } 1849 1850 // Late inlining of boxing methods 1851 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1852 if (_boxing_late_inlines.length() > 0) { 1853 assert(has_boxed_value(), "inconsistent"); 1854 1855 PhaseGVN* gvn = initial_gvn(); 1856 set_inlining_incrementally(true); 1857 1858 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1859 for_igvn()->clear(); 1860 gvn->replace_with(&igvn); 1861 1862 while (_boxing_late_inlines.length() > 0) { 1863 CallGenerator* cg = _boxing_late_inlines.pop(); 1864 cg->do_late_inline(); 1865 if (failing()) return; 1866 } 1867 _boxing_late_inlines.trunc_to(0); 1868 1869 { 1870 ResourceMark rm; 1871 PhaseRemoveUseless pru(gvn, for_igvn()); 1872 } 1873 1874 igvn = PhaseIterGVN(gvn); 1875 igvn.optimize(); 1876 1877 set_inlining_progress(false); 1878 set_inlining_incrementally(false); 1879 } 1880 } 1881 1882 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1883 assert(IncrementalInline, "incremental inlining should be on"); 1884 PhaseGVN* gvn = initial_gvn(); 1885 1886 set_inlining_progress(false); 1887 for_igvn()->clear(); 1888 gvn->replace_with(&igvn); 1889 1890 int i = 0; 1891 1892 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1893 CallGenerator* cg = _late_inlines.at(i); 1894 _late_inlines_pos = i+1; 1895 cg->do_late_inline(); 1896 if (failing()) return; 1897 } 1898 int j = 0; 1899 for (; i < _late_inlines.length(); i++, j++) { 1900 _late_inlines.at_put(j, _late_inlines.at(i)); 1901 } 1902 _late_inlines.trunc_to(j); 1903 1904 { 1905 ResourceMark rm; 1906 PhaseRemoveUseless pru(gvn, for_igvn()); 1907 } 1908 1909 igvn = PhaseIterGVN(gvn); 1910 } 1911 1912 // Perform incremental inlining until bound on number of live nodes is reached 1913 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1914 PhaseGVN* gvn = initial_gvn(); 1915 1916 set_inlining_incrementally(true); 1917 set_inlining_progress(true); 1918 uint low_live_nodes = 0; 1919 1920 while(inlining_progress() && _late_inlines.length() > 0) { 1921 1922 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1923 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1924 // PhaseIdealLoop is expensive so we only try it once we are 1925 // out of loop and we only try it again if the previous helped 1926 // got the number of nodes down significantly 1927 PhaseIdealLoop ideal_loop( igvn, false, true ); 1928 if (failing()) return; 1929 low_live_nodes = live_nodes(); 1930 _major_progress = true; 1931 } 1932 1933 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1934 break; 1935 } 1936 } 1937 1938 inline_incrementally_one(igvn); 1939 1940 if (failing()) return; 1941 1942 igvn.optimize(); 1943 1944 if (failing()) return; 1945 } 1946 1947 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1948 1949 if (_string_late_inlines.length() > 0) { 1950 assert(has_stringbuilder(), "inconsistent"); 1951 for_igvn()->clear(); 1952 initial_gvn()->replace_with(&igvn); 1953 1954 inline_string_calls(false); 1955 1956 if (failing()) return; 1957 1958 { 1959 ResourceMark rm; 1960 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1961 } 1962 1963 igvn = PhaseIterGVN(gvn); 1964 1965 igvn.optimize(); 1966 } 1967 1968 set_inlining_incrementally(false); 1969 } 1970 1971 1972 //------------------------------Optimize--------------------------------------- 1973 // Given a graph, optimize it. 1974 void Compile::Optimize() { 1975 TracePhase t1("optimizer", &_t_optimizer, true); 1976 1977 #ifndef PRODUCT 1978 if (env()->break_at_compile()) { 1979 BREAKPOINT; 1980 } 1981 1982 #endif 1983 1984 ResourceMark rm; 1985 int loop_opts_cnt; 1986 1987 NOT_PRODUCT( verify_graph_edges(); ) 1988 1989 print_method(PHASE_AFTER_PARSING); 1990 1991 { 1992 // Iterative Global Value Numbering, including ideal transforms 1993 // Initialize IterGVN with types and values from parse-time GVN 1994 PhaseIterGVN igvn(initial_gvn()); 1995 { 1996 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 1997 igvn.optimize(); 1998 } 1999 2000 print_method(PHASE_ITER_GVN1, 2); 2001 2002 if (failing()) return; 2003 2004 { 2005 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2006 inline_incrementally(igvn); 2007 } 2008 2009 print_method(PHASE_INCREMENTAL_INLINE, 2); 2010 2011 if (failing()) return; 2012 2013 if (eliminate_boxing()) { 2014 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2015 // Inline valueOf() methods now. 2016 inline_boxing_calls(igvn); 2017 2018 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2019 2020 if (failing()) return; 2021 } 2022 2023 // Remove the speculative part of types and clean up the graph from 2024 // the extra CastPP nodes whose only purpose is to carry them. Do 2025 // that early so that optimizations are not disrupted by the extra 2026 // CastPP nodes. 2027 remove_speculative_types(igvn); 2028 2029 // No more new expensive nodes will be added to the list from here 2030 // so keep only the actual candidates for optimizations. 2031 cleanup_expensive_nodes(igvn); 2032 2033 // Perform escape analysis 2034 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2035 if (has_loops()) { 2036 // Cleanup graph (remove dead nodes). 2037 TracePhase t2("idealLoop", &_t_idealLoop, true); 2038 PhaseIdealLoop ideal_loop( igvn, false, true ); 2039 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2040 if (failing()) return; 2041 } 2042 ConnectionGraph::do_analysis(this, &igvn); 2043 2044 if (failing()) return; 2045 2046 // Optimize out fields loads from scalar replaceable allocations. 2047 igvn.optimize(); 2048 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2049 2050 if (failing()) return; 2051 2052 if (congraph() != NULL && macro_count() > 0) { 2053 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2054 PhaseMacroExpand mexp(igvn); 2055 mexp.eliminate_macro_nodes(); 2056 igvn.set_delay_transform(false); 2057 2058 igvn.optimize(); 2059 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2060 2061 if (failing()) return; 2062 } 2063 } 2064 2065 // Loop transforms on the ideal graph. Range Check Elimination, 2066 // peeling, unrolling, etc. 2067 2068 // Set loop opts counter 2069 loop_opts_cnt = num_loop_opts(); 2070 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2071 { 2072 TracePhase t2("idealLoop", &_t_idealLoop, true); 2073 PhaseIdealLoop ideal_loop( igvn, true ); 2074 loop_opts_cnt--; 2075 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2076 if (failing()) return; 2077 } 2078 // Loop opts pass if partial peeling occurred in previous pass 2079 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2080 TracePhase t3("idealLoop", &_t_idealLoop, true); 2081 PhaseIdealLoop ideal_loop( igvn, false ); 2082 loop_opts_cnt--; 2083 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2084 if (failing()) return; 2085 } 2086 // Loop opts pass for loop-unrolling before CCP 2087 if(major_progress() && (loop_opts_cnt > 0)) { 2088 TracePhase t4("idealLoop", &_t_idealLoop, true); 2089 PhaseIdealLoop ideal_loop( igvn, false ); 2090 loop_opts_cnt--; 2091 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2092 } 2093 if (!failing()) { 2094 // Verify that last round of loop opts produced a valid graph 2095 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2096 PhaseIdealLoop::verify(igvn); 2097 } 2098 } 2099 if (failing()) return; 2100 2101 // Conditional Constant Propagation; 2102 PhaseCCP ccp( &igvn ); 2103 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2104 { 2105 TracePhase t2("ccp", &_t_ccp, true); 2106 ccp.do_transform(); 2107 } 2108 print_method(PHASE_CPP1, 2); 2109 2110 assert( true, "Break here to ccp.dump_old2new_map()"); 2111 2112 // Iterative Global Value Numbering, including ideal transforms 2113 { 2114 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2115 igvn = ccp; 2116 igvn.optimize(); 2117 } 2118 2119 print_method(PHASE_ITER_GVN2, 2); 2120 2121 if (failing()) return; 2122 2123 // Loop transforms on the ideal graph. Range Check Elimination, 2124 // peeling, unrolling, etc. 2125 if(loop_opts_cnt > 0) { 2126 debug_only( int cnt = 0; ); 2127 while(major_progress() && (loop_opts_cnt > 0)) { 2128 TracePhase t2("idealLoop", &_t_idealLoop, true); 2129 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2130 PhaseIdealLoop ideal_loop( igvn, true); 2131 loop_opts_cnt--; 2132 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2133 if (failing()) return; 2134 } 2135 } 2136 2137 { 2138 // Verify that all previous optimizations produced a valid graph 2139 // at least to this point, even if no loop optimizations were done. 2140 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2141 PhaseIdealLoop::verify(igvn); 2142 } 2143 2144 { 2145 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2146 PhaseMacroExpand mex(igvn); 2147 if (mex.expand_macro_nodes()) { 2148 assert(failing(), "must bail out w/ explicit message"); 2149 return; 2150 } 2151 } 2152 2153 } // (End scope of igvn; run destructor if necessary for asserts.) 2154 2155 dump_inlining(); 2156 // A method with only infinite loops has no edges entering loops from root 2157 { 2158 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2159 if (final_graph_reshaping()) { 2160 assert(failing(), "must bail out w/ explicit message"); 2161 return; 2162 } 2163 } 2164 2165 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2166 } 2167 2168 2169 //------------------------------Code_Gen--------------------------------------- 2170 // Given a graph, generate code for it 2171 void Compile::Code_Gen() { 2172 if (failing()) { 2173 return; 2174 } 2175 2176 // Perform instruction selection. You might think we could reclaim Matcher 2177 // memory PDQ, but actually the Matcher is used in generating spill code. 2178 // Internals of the Matcher (including some VectorSets) must remain live 2179 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2180 // set a bit in reclaimed memory. 2181 2182 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2183 // nodes. Mapping is only valid at the root of each matched subtree. 2184 NOT_PRODUCT( verify_graph_edges(); ) 2185 2186 Matcher matcher; 2187 _matcher = &matcher; 2188 { 2189 TracePhase t2("matcher", &_t_matcher, true); 2190 matcher.match(); 2191 } 2192 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2193 // nodes. Mapping is only valid at the root of each matched subtree. 2194 NOT_PRODUCT( verify_graph_edges(); ) 2195 2196 // If you have too many nodes, or if matching has failed, bail out 2197 check_node_count(0, "out of nodes matching instructions"); 2198 if (failing()) { 2199 return; 2200 } 2201 2202 // Build a proper-looking CFG 2203 PhaseCFG cfg(node_arena(), root(), matcher); 2204 _cfg = &cfg; 2205 { 2206 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2207 bool success = cfg.do_global_code_motion(); 2208 if (!success) { 2209 return; 2210 } 2211 2212 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2213 NOT_PRODUCT( verify_graph_edges(); ) 2214 debug_only( cfg.verify(); ) 2215 } 2216 2217 PhaseChaitin regalloc(unique(), cfg, matcher); 2218 _regalloc = ®alloc; 2219 { 2220 TracePhase t2("regalloc", &_t_registerAllocation, true); 2221 // Perform register allocation. After Chaitin, use-def chains are 2222 // no longer accurate (at spill code) and so must be ignored. 2223 // Node->LRG->reg mappings are still accurate. 2224 _regalloc->Register_Allocate(); 2225 2226 // Bail out if the allocator builds too many nodes 2227 if (failing()) { 2228 return; 2229 } 2230 } 2231 2232 // Prior to register allocation we kept empty basic blocks in case the 2233 // the allocator needed a place to spill. After register allocation we 2234 // are not adding any new instructions. If any basic block is empty, we 2235 // can now safely remove it. 2236 { 2237 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2238 cfg.remove_empty_blocks(); 2239 if (do_freq_based_layout()) { 2240 PhaseBlockLayout layout(cfg); 2241 } else { 2242 cfg.set_loop_alignment(); 2243 } 2244 cfg.fixup_flow(); 2245 } 2246 2247 // Apply peephole optimizations 2248 if( OptoPeephole ) { 2249 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2250 PhasePeephole peep( _regalloc, cfg); 2251 peep.do_transform(); 2252 } 2253 2254 // Do late expand if CPU requires this. 2255 if (Matcher::require_postalloc_expand) { 2256 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2257 cfg.postalloc_expand(_regalloc); 2258 } 2259 2260 // Convert Nodes to instruction bits in a buffer 2261 { 2262 // %%%% workspace merge brought two timers together for one job 2263 TracePhase t2a("output", &_t_output, true); 2264 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2265 Output(); 2266 } 2267 2268 print_method(PHASE_FINAL_CODE); 2269 2270 // He's dead, Jim. 2271 _cfg = (PhaseCFG*)0xdeadbeef; 2272 _regalloc = (PhaseChaitin*)0xdeadbeef; 2273 } 2274 2275 2276 //------------------------------dump_asm--------------------------------------- 2277 // Dump formatted assembly 2278 #ifndef PRODUCT 2279 void Compile::dump_asm(int *pcs, uint pc_limit) { 2280 bool cut_short = false; 2281 tty->print_cr("#"); 2282 tty->print("# "); _tf->dump(); tty->cr(); 2283 tty->print_cr("#"); 2284 2285 // For all blocks 2286 int pc = 0x0; // Program counter 2287 char starts_bundle = ' '; 2288 _regalloc->dump_frame(); 2289 2290 Node *n = NULL; 2291 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2292 if (VMThread::should_terminate()) { 2293 cut_short = true; 2294 break; 2295 } 2296 Block* block = _cfg->get_block(i); 2297 if (block->is_connector() && !Verbose) { 2298 continue; 2299 } 2300 n = block->head(); 2301 if (pcs && n->_idx < pc_limit) { 2302 tty->print("%3.3x ", pcs[n->_idx]); 2303 } else { 2304 tty->print(" "); 2305 } 2306 block->dump_head(_cfg); 2307 if (block->is_connector()) { 2308 tty->print_cr(" # Empty connector block"); 2309 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2310 tty->print_cr(" # Block is sole successor of call"); 2311 } 2312 2313 // For all instructions 2314 Node *delay = NULL; 2315 for (uint j = 0; j < block->number_of_nodes(); j++) { 2316 if (VMThread::should_terminate()) { 2317 cut_short = true; 2318 break; 2319 } 2320 n = block->get_node(j); 2321 if (valid_bundle_info(n)) { 2322 Bundle* bundle = node_bundling(n); 2323 if (bundle->used_in_unconditional_delay()) { 2324 delay = n; 2325 continue; 2326 } 2327 if (bundle->starts_bundle()) { 2328 starts_bundle = '+'; 2329 } 2330 } 2331 2332 if (WizardMode) { 2333 n->dump(); 2334 } 2335 2336 if( !n->is_Region() && // Dont print in the Assembly 2337 !n->is_Phi() && // a few noisely useless nodes 2338 !n->is_Proj() && 2339 !n->is_MachTemp() && 2340 !n->is_SafePointScalarObject() && 2341 !n->is_Catch() && // Would be nice to print exception table targets 2342 !n->is_MergeMem() && // Not very interesting 2343 !n->is_top() && // Debug info table constants 2344 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2345 ) { 2346 if (pcs && n->_idx < pc_limit) 2347 tty->print("%3.3x", pcs[n->_idx]); 2348 else 2349 tty->print(" "); 2350 tty->print(" %c ", starts_bundle); 2351 starts_bundle = ' '; 2352 tty->print("\t"); 2353 n->format(_regalloc, tty); 2354 tty->cr(); 2355 } 2356 2357 // If we have an instruction with a delay slot, and have seen a delay, 2358 // then back up and print it 2359 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2360 assert(delay != NULL, "no unconditional delay instruction"); 2361 if (WizardMode) delay->dump(); 2362 2363 if (node_bundling(delay)->starts_bundle()) 2364 starts_bundle = '+'; 2365 if (pcs && n->_idx < pc_limit) 2366 tty->print("%3.3x", pcs[n->_idx]); 2367 else 2368 tty->print(" "); 2369 tty->print(" %c ", starts_bundle); 2370 starts_bundle = ' '; 2371 tty->print("\t"); 2372 delay->format(_regalloc, tty); 2373 tty->print_cr(""); 2374 delay = NULL; 2375 } 2376 2377 // Dump the exception table as well 2378 if( n->is_Catch() && (Verbose || WizardMode) ) { 2379 // Print the exception table for this offset 2380 _handler_table.print_subtable_for(pc); 2381 } 2382 } 2383 2384 if (pcs && n->_idx < pc_limit) 2385 tty->print_cr("%3.3x", pcs[n->_idx]); 2386 else 2387 tty->print_cr(""); 2388 2389 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2390 2391 } // End of per-block dump 2392 tty->print_cr(""); 2393 2394 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2395 } 2396 #endif 2397 2398 //------------------------------Final_Reshape_Counts--------------------------- 2399 // This class defines counters to help identify when a method 2400 // may/must be executed using hardware with only 24-bit precision. 2401 struct Final_Reshape_Counts : public StackObj { 2402 int _call_count; // count non-inlined 'common' calls 2403 int _float_count; // count float ops requiring 24-bit precision 2404 int _double_count; // count double ops requiring more precision 2405 int _java_call_count; // count non-inlined 'java' calls 2406 int _inner_loop_count; // count loops which need alignment 2407 VectorSet _visited; // Visitation flags 2408 Node_List _tests; // Set of IfNodes & PCTableNodes 2409 2410 Final_Reshape_Counts() : 2411 _call_count(0), _float_count(0), _double_count(0), 2412 _java_call_count(0), _inner_loop_count(0), 2413 _visited( Thread::current()->resource_area() ) { } 2414 2415 void inc_call_count () { _call_count ++; } 2416 void inc_float_count () { _float_count ++; } 2417 void inc_double_count() { _double_count++; } 2418 void inc_java_call_count() { _java_call_count++; } 2419 void inc_inner_loop_count() { _inner_loop_count++; } 2420 2421 int get_call_count () const { return _call_count ; } 2422 int get_float_count () const { return _float_count ; } 2423 int get_double_count() const { return _double_count; } 2424 int get_java_call_count() const { return _java_call_count; } 2425 int get_inner_loop_count() const { return _inner_loop_count; } 2426 }; 2427 2428 #ifdef ASSERT 2429 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2430 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2431 // Make sure the offset goes inside the instance layout. 2432 return k->contains_field_offset(tp->offset()); 2433 // Note that OffsetBot and OffsetTop are very negative. 2434 } 2435 #endif 2436 2437 // Eliminate trivially redundant StoreCMs and accumulate their 2438 // precedence edges. 2439 void Compile::eliminate_redundant_card_marks(Node* n) { 2440 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2441 if (n->in(MemNode::Address)->outcnt() > 1) { 2442 // There are multiple users of the same address so it might be 2443 // possible to eliminate some of the StoreCMs 2444 Node* mem = n->in(MemNode::Memory); 2445 Node* adr = n->in(MemNode::Address); 2446 Node* val = n->in(MemNode::ValueIn); 2447 Node* prev = n; 2448 bool done = false; 2449 // Walk the chain of StoreCMs eliminating ones that match. As 2450 // long as it's a chain of single users then the optimization is 2451 // safe. Eliminating partially redundant StoreCMs would require 2452 // cloning copies down the other paths. 2453 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2454 if (adr == mem->in(MemNode::Address) && 2455 val == mem->in(MemNode::ValueIn)) { 2456 // redundant StoreCM 2457 if (mem->req() > MemNode::OopStore) { 2458 // Hasn't been processed by this code yet. 2459 n->add_prec(mem->in(MemNode::OopStore)); 2460 } else { 2461 // Already converted to precedence edge 2462 for (uint i = mem->req(); i < mem->len(); i++) { 2463 // Accumulate any precedence edges 2464 if (mem->in(i) != NULL) { 2465 n->add_prec(mem->in(i)); 2466 } 2467 } 2468 // Everything above this point has been processed. 2469 done = true; 2470 } 2471 // Eliminate the previous StoreCM 2472 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2473 assert(mem->outcnt() == 0, "should be dead"); 2474 mem->disconnect_inputs(NULL, this); 2475 } else { 2476 prev = mem; 2477 } 2478 mem = prev->in(MemNode::Memory); 2479 } 2480 } 2481 } 2482 2483 //------------------------------final_graph_reshaping_impl---------------------- 2484 // Implement items 1-5 from final_graph_reshaping below. 2485 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2486 2487 if ( n->outcnt() == 0 ) return; // dead node 2488 uint nop = n->Opcode(); 2489 2490 // Check for 2-input instruction with "last use" on right input. 2491 // Swap to left input. Implements item (2). 2492 if( n->req() == 3 && // two-input instruction 2493 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2494 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2495 n->in(2)->outcnt() == 1 &&// right use IS a last use 2496 !n->in(2)->is_Con() ) { // right use is not a constant 2497 // Check for commutative opcode 2498 switch( nop ) { 2499 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2500 case Op_MaxI: case Op_MinI: 2501 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2502 case Op_AndL: case Op_XorL: case Op_OrL: 2503 case Op_AndI: case Op_XorI: case Op_OrI: { 2504 // Move "last use" input to left by swapping inputs 2505 n->swap_edges(1, 2); 2506 break; 2507 } 2508 default: 2509 break; 2510 } 2511 } 2512 2513 #ifdef ASSERT 2514 if( n->is_Mem() ) { 2515 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2516 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2517 // oop will be recorded in oop map if load crosses safepoint 2518 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2519 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2520 "raw memory operations should have control edge"); 2521 } 2522 #endif 2523 // Count FPU ops and common calls, implements item (3) 2524 switch( nop ) { 2525 // Count all float operations that may use FPU 2526 case Op_AddF: 2527 case Op_SubF: 2528 case Op_MulF: 2529 case Op_DivF: 2530 case Op_NegF: 2531 case Op_ModF: 2532 case Op_ConvI2F: 2533 case Op_ConF: 2534 case Op_CmpF: 2535 case Op_CmpF3: 2536 // case Op_ConvL2F: // longs are split into 32-bit halves 2537 frc.inc_float_count(); 2538 break; 2539 2540 case Op_ConvF2D: 2541 case Op_ConvD2F: 2542 frc.inc_float_count(); 2543 frc.inc_double_count(); 2544 break; 2545 2546 // Count all double operations that may use FPU 2547 case Op_AddD: 2548 case Op_SubD: 2549 case Op_MulD: 2550 case Op_DivD: 2551 case Op_NegD: 2552 case Op_ModD: 2553 case Op_ConvI2D: 2554 case Op_ConvD2I: 2555 // case Op_ConvL2D: // handled by leaf call 2556 // case Op_ConvD2L: // handled by leaf call 2557 case Op_ConD: 2558 case Op_CmpD: 2559 case Op_CmpD3: 2560 frc.inc_double_count(); 2561 break; 2562 case Op_Opaque1: // Remove Opaque Nodes before matching 2563 case Op_Opaque2: // Remove Opaque Nodes before matching 2564 n->subsume_by(n->in(1), this); 2565 break; 2566 case Op_CallStaticJava: 2567 case Op_CallJava: 2568 case Op_CallDynamicJava: 2569 frc.inc_java_call_count(); // Count java call site; 2570 case Op_CallRuntime: 2571 case Op_CallLeaf: 2572 case Op_CallLeafNoFP: { 2573 assert( n->is_Call(), "" ); 2574 CallNode *call = n->as_Call(); 2575 // Count call sites where the FP mode bit would have to be flipped. 2576 // Do not count uncommon runtime calls: 2577 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2578 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2579 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2580 frc.inc_call_count(); // Count the call site 2581 } else { // See if uncommon argument is shared 2582 Node *n = call->in(TypeFunc::Parms); 2583 int nop = n->Opcode(); 2584 // Clone shared simple arguments to uncommon calls, item (1). 2585 if( n->outcnt() > 1 && 2586 !n->is_Proj() && 2587 nop != Op_CreateEx && 2588 nop != Op_CheckCastPP && 2589 nop != Op_DecodeN && 2590 nop != Op_DecodeNKlass && 2591 !n->is_Mem() ) { 2592 Node *x = n->clone(); 2593 call->set_req( TypeFunc::Parms, x ); 2594 } 2595 } 2596 break; 2597 } 2598 2599 case Op_StoreD: 2600 case Op_LoadD: 2601 case Op_LoadD_unaligned: 2602 frc.inc_double_count(); 2603 goto handle_mem; 2604 case Op_StoreF: 2605 case Op_LoadF: 2606 frc.inc_float_count(); 2607 goto handle_mem; 2608 2609 case Op_StoreCM: 2610 { 2611 // Convert OopStore dependence into precedence edge 2612 Node* prec = n->in(MemNode::OopStore); 2613 n->del_req(MemNode::OopStore); 2614 n->add_prec(prec); 2615 eliminate_redundant_card_marks(n); 2616 } 2617 2618 // fall through 2619 2620 case Op_StoreB: 2621 case Op_StoreC: 2622 case Op_StorePConditional: 2623 case Op_StoreI: 2624 case Op_StoreL: 2625 case Op_StoreIConditional: 2626 case Op_StoreLConditional: 2627 case Op_CompareAndSwapI: 2628 case Op_CompareAndSwapL: 2629 case Op_CompareAndSwapP: 2630 case Op_CompareAndSwapN: 2631 case Op_GetAndAddI: 2632 case Op_GetAndAddL: 2633 case Op_GetAndSetI: 2634 case Op_GetAndSetL: 2635 case Op_GetAndSetP: 2636 case Op_GetAndSetN: 2637 case Op_StoreP: 2638 case Op_StoreN: 2639 case Op_StoreNKlass: 2640 case Op_LoadB: 2641 case Op_LoadUB: 2642 case Op_LoadUS: 2643 case Op_LoadI: 2644 case Op_LoadKlass: 2645 case Op_LoadNKlass: 2646 case Op_LoadL: 2647 case Op_LoadL_unaligned: 2648 case Op_LoadPLocked: 2649 case Op_LoadP: 2650 case Op_LoadN: 2651 case Op_LoadRange: 2652 case Op_LoadS: { 2653 handle_mem: 2654 #ifdef ASSERT 2655 if( VerifyOptoOopOffsets ) { 2656 assert( n->is_Mem(), "" ); 2657 MemNode *mem = (MemNode*)n; 2658 // Check to see if address types have grounded out somehow. 2659 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2660 assert( !tp || oop_offset_is_sane(tp), "" ); 2661 } 2662 #endif 2663 break; 2664 } 2665 2666 case Op_AddP: { // Assert sane base pointers 2667 Node *addp = n->in(AddPNode::Address); 2668 assert( !addp->is_AddP() || 2669 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2670 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2671 "Base pointers must match" ); 2672 #ifdef _LP64 2673 if ((UseCompressedOops || UseCompressedClassPointers) && 2674 addp->Opcode() == Op_ConP && 2675 addp == n->in(AddPNode::Base) && 2676 n->in(AddPNode::Offset)->is_Con()) { 2677 // Use addressing with narrow klass to load with offset on x86. 2678 // On sparc loading 32-bits constant and decoding it have less 2679 // instructions (4) then load 64-bits constant (7). 2680 // Do this transformation here since IGVN will convert ConN back to ConP. 2681 const Type* t = addp->bottom_type(); 2682 if (t->isa_oopptr() || t->isa_klassptr()) { 2683 Node* nn = NULL; 2684 2685 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2686 2687 // Look for existing ConN node of the same exact type. 2688 Node* r = root(); 2689 uint cnt = r->outcnt(); 2690 for (uint i = 0; i < cnt; i++) { 2691 Node* m = r->raw_out(i); 2692 if (m!= NULL && m->Opcode() == op && 2693 m->bottom_type()->make_ptr() == t) { 2694 nn = m; 2695 break; 2696 } 2697 } 2698 if (nn != NULL) { 2699 // Decode a narrow oop to match address 2700 // [R12 + narrow_oop_reg<<3 + offset] 2701 if (t->isa_oopptr()) { 2702 nn = new (this) DecodeNNode(nn, t); 2703 } else { 2704 nn = new (this) DecodeNKlassNode(nn, t); 2705 } 2706 n->set_req(AddPNode::Base, nn); 2707 n->set_req(AddPNode::Address, nn); 2708 if (addp->outcnt() == 0) { 2709 addp->disconnect_inputs(NULL, this); 2710 } 2711 } 2712 } 2713 } 2714 #endif 2715 break; 2716 } 2717 2718 #ifdef _LP64 2719 case Op_CastPP: 2720 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2721 Node* in1 = n->in(1); 2722 const Type* t = n->bottom_type(); 2723 Node* new_in1 = in1->clone(); 2724 new_in1->as_DecodeN()->set_type(t); 2725 2726 if (!Matcher::narrow_oop_use_complex_address()) { 2727 // 2728 // x86, ARM and friends can handle 2 adds in addressing mode 2729 // and Matcher can fold a DecodeN node into address by using 2730 // a narrow oop directly and do implicit NULL check in address: 2731 // 2732 // [R12 + narrow_oop_reg<<3 + offset] 2733 // NullCheck narrow_oop_reg 2734 // 2735 // On other platforms (Sparc) we have to keep new DecodeN node and 2736 // use it to do implicit NULL check in address: 2737 // 2738 // decode_not_null narrow_oop_reg, base_reg 2739 // [base_reg + offset] 2740 // NullCheck base_reg 2741 // 2742 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2743 // to keep the information to which NULL check the new DecodeN node 2744 // corresponds to use it as value in implicit_null_check(). 2745 // 2746 new_in1->set_req(0, n->in(0)); 2747 } 2748 2749 n->subsume_by(new_in1, this); 2750 if (in1->outcnt() == 0) { 2751 in1->disconnect_inputs(NULL, this); 2752 } 2753 } 2754 break; 2755 2756 case Op_CmpP: 2757 // Do this transformation here to preserve CmpPNode::sub() and 2758 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2759 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2760 Node* in1 = n->in(1); 2761 Node* in2 = n->in(2); 2762 if (!in1->is_DecodeNarrowPtr()) { 2763 in2 = in1; 2764 in1 = n->in(2); 2765 } 2766 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2767 2768 Node* new_in2 = NULL; 2769 if (in2->is_DecodeNarrowPtr()) { 2770 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2771 new_in2 = in2->in(1); 2772 } else if (in2->Opcode() == Op_ConP) { 2773 const Type* t = in2->bottom_type(); 2774 if (t == TypePtr::NULL_PTR) { 2775 assert(in1->is_DecodeN(), "compare klass to null?"); 2776 // Don't convert CmpP null check into CmpN if compressed 2777 // oops implicit null check is not generated. 2778 // This will allow to generate normal oop implicit null check. 2779 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2780 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2781 // 2782 // This transformation together with CastPP transformation above 2783 // will generated code for implicit NULL checks for compressed oops. 2784 // 2785 // The original code after Optimize() 2786 // 2787 // LoadN memory, narrow_oop_reg 2788 // decode narrow_oop_reg, base_reg 2789 // CmpP base_reg, NULL 2790 // CastPP base_reg // NotNull 2791 // Load [base_reg + offset], val_reg 2792 // 2793 // after these transformations will be 2794 // 2795 // LoadN memory, narrow_oop_reg 2796 // CmpN narrow_oop_reg, NULL 2797 // decode_not_null narrow_oop_reg, base_reg 2798 // Load [base_reg + offset], val_reg 2799 // 2800 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2801 // since narrow oops can be used in debug info now (see the code in 2802 // final_graph_reshaping_walk()). 2803 // 2804 // At the end the code will be matched to 2805 // on x86: 2806 // 2807 // Load_narrow_oop memory, narrow_oop_reg 2808 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2809 // NullCheck narrow_oop_reg 2810 // 2811 // and on sparc: 2812 // 2813 // Load_narrow_oop memory, narrow_oop_reg 2814 // decode_not_null narrow_oop_reg, base_reg 2815 // Load [base_reg + offset], val_reg 2816 // NullCheck base_reg 2817 // 2818 } else if (t->isa_oopptr()) { 2819 new_in2 = ConNode::make(this, t->make_narrowoop()); 2820 } else if (t->isa_klassptr()) { 2821 new_in2 = ConNode::make(this, t->make_narrowklass()); 2822 } 2823 } 2824 if (new_in2 != NULL) { 2825 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2826 n->subsume_by(cmpN, this); 2827 if (in1->outcnt() == 0) { 2828 in1->disconnect_inputs(NULL, this); 2829 } 2830 if (in2->outcnt() == 0) { 2831 in2->disconnect_inputs(NULL, this); 2832 } 2833 } 2834 } 2835 break; 2836 2837 case Op_DecodeN: 2838 case Op_DecodeNKlass: 2839 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2840 // DecodeN could be pinned when it can't be fold into 2841 // an address expression, see the code for Op_CastPP above. 2842 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2843 break; 2844 2845 case Op_EncodeP: 2846 case Op_EncodePKlass: { 2847 Node* in1 = n->in(1); 2848 if (in1->is_DecodeNarrowPtr()) { 2849 n->subsume_by(in1->in(1), this); 2850 } else if (in1->Opcode() == Op_ConP) { 2851 const Type* t = in1->bottom_type(); 2852 if (t == TypePtr::NULL_PTR) { 2853 assert(t->isa_oopptr(), "null klass?"); 2854 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2855 } else if (t->isa_oopptr()) { 2856 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2857 } else if (t->isa_klassptr()) { 2858 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2859 } 2860 } 2861 if (in1->outcnt() == 0) { 2862 in1->disconnect_inputs(NULL, this); 2863 } 2864 break; 2865 } 2866 2867 case Op_Proj: { 2868 if (OptimizeStringConcat) { 2869 ProjNode* p = n->as_Proj(); 2870 if (p->_is_io_use) { 2871 // Separate projections were used for the exception path which 2872 // are normally removed by a late inline. If it wasn't inlined 2873 // then they will hang around and should just be replaced with 2874 // the original one. 2875 Node* proj = NULL; 2876 // Replace with just one 2877 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2878 Node *use = i.get(); 2879 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2880 proj = use; 2881 break; 2882 } 2883 } 2884 assert(proj != NULL, "must be found"); 2885 p->subsume_by(proj, this); 2886 } 2887 } 2888 break; 2889 } 2890 2891 case Op_Phi: 2892 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2893 // The EncodeP optimization may create Phi with the same edges 2894 // for all paths. It is not handled well by Register Allocator. 2895 Node* unique_in = n->in(1); 2896 assert(unique_in != NULL, ""); 2897 uint cnt = n->req(); 2898 for (uint i = 2; i < cnt; i++) { 2899 Node* m = n->in(i); 2900 assert(m != NULL, ""); 2901 if (unique_in != m) 2902 unique_in = NULL; 2903 } 2904 if (unique_in != NULL) { 2905 n->subsume_by(unique_in, this); 2906 } 2907 } 2908 break; 2909 2910 #endif 2911 2912 case Op_ModI: 2913 if (UseDivMod) { 2914 // Check if a%b and a/b both exist 2915 Node* d = n->find_similar(Op_DivI); 2916 if (d) { 2917 // Replace them with a fused divmod if supported 2918 if (Matcher::has_match_rule(Op_DivModI)) { 2919 DivModINode* divmod = DivModINode::make(this, n); 2920 d->subsume_by(divmod->div_proj(), this); 2921 n->subsume_by(divmod->mod_proj(), this); 2922 } else { 2923 // replace a%b with a-((a/b)*b) 2924 Node* mult = new (this) MulINode(d, d->in(2)); 2925 Node* sub = new (this) SubINode(d->in(1), mult); 2926 n->subsume_by(sub, this); 2927 } 2928 } 2929 } 2930 break; 2931 2932 case Op_ModL: 2933 if (UseDivMod) { 2934 // Check if a%b and a/b both exist 2935 Node* d = n->find_similar(Op_DivL); 2936 if (d) { 2937 // Replace them with a fused divmod if supported 2938 if (Matcher::has_match_rule(Op_DivModL)) { 2939 DivModLNode* divmod = DivModLNode::make(this, n); 2940 d->subsume_by(divmod->div_proj(), this); 2941 n->subsume_by(divmod->mod_proj(), this); 2942 } else { 2943 // replace a%b with a-((a/b)*b) 2944 Node* mult = new (this) MulLNode(d, d->in(2)); 2945 Node* sub = new (this) SubLNode(d->in(1), mult); 2946 n->subsume_by(sub, this); 2947 } 2948 } 2949 } 2950 break; 2951 2952 case Op_LoadVector: 2953 case Op_StoreVector: 2954 break; 2955 2956 case Op_PackB: 2957 case Op_PackS: 2958 case Op_PackI: 2959 case Op_PackF: 2960 case Op_PackL: 2961 case Op_PackD: 2962 if (n->req()-1 > 2) { 2963 // Replace many operand PackNodes with a binary tree for matching 2964 PackNode* p = (PackNode*) n; 2965 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2966 n->subsume_by(btp, this); 2967 } 2968 break; 2969 case Op_Loop: 2970 case Op_CountedLoop: 2971 if (n->as_Loop()->is_inner_loop()) { 2972 frc.inc_inner_loop_count(); 2973 } 2974 break; 2975 case Op_LShiftI: 2976 case Op_RShiftI: 2977 case Op_URShiftI: 2978 case Op_LShiftL: 2979 case Op_RShiftL: 2980 case Op_URShiftL: 2981 if (Matcher::need_masked_shift_count) { 2982 // The cpu's shift instructions don't restrict the count to the 2983 // lower 5/6 bits. We need to do the masking ourselves. 2984 Node* in2 = n->in(2); 2985 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 2986 const TypeInt* t = in2->find_int_type(); 2987 if (t != NULL && t->is_con()) { 2988 juint shift = t->get_con(); 2989 if (shift > mask) { // Unsigned cmp 2990 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 2991 } 2992 } else { 2993 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 2994 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 2995 n->set_req(2, shift); 2996 } 2997 } 2998 if (in2->outcnt() == 0) { // Remove dead node 2999 in2->disconnect_inputs(NULL, this); 3000 } 3001 } 3002 break; 3003 case Op_MemBarStoreStore: 3004 case Op_MemBarRelease: 3005 // Break the link with AllocateNode: it is no longer useful and 3006 // confuses register allocation. 3007 if (n->req() > MemBarNode::Precedent) { 3008 n->set_req(MemBarNode::Precedent, top()); 3009 } 3010 break; 3011 // Must set a control edge on all nodes that produce a FlagsProj 3012 // so they can't escape the block that consumes the flags. 3013 // Must also set the non throwing branch as the control 3014 // for all nodes that depends on the result. Unless the node 3015 // already have a control that isn't the control of the 3016 // flag producer 3017 case Op_FlagsProj: 3018 { 3019 MathExactNode* math = (MathExactNode*) n->in(0); 3020 Node* ctrl = math->control_node(); 3021 Node* non_throwing = math->non_throwing_branch(); 3022 math->set_req(0, ctrl); 3023 3024 Node* result = math->result_node(); 3025 if (result != NULL) { 3026 for (DUIterator_Fast jmax, j = result->fast_outs(jmax); j < jmax; j++) { 3027 Node* out = result->fast_out(j); 3028 // Phi nodes shouldn't be moved. They would only match below if they 3029 // had the same control as the MathExactNode. The only time that 3030 // would happen is if the Phi is also an input to the MathExact 3031 // 3032 // Cmp nodes shouldn't have control set at all. 3033 if (out->is_Phi() || 3034 out->is_Cmp()) { 3035 continue; 3036 } 3037 3038 if (out->in(0) == NULL) { 3039 out->set_req(0, non_throwing); 3040 } else if (out->in(0) == ctrl) { 3041 out->set_req(0, non_throwing); 3042 } 3043 } 3044 } 3045 } 3046 break; 3047 default: 3048 assert( !n->is_Call(), "" ); 3049 assert( !n->is_Mem(), "" ); 3050 break; 3051 } 3052 3053 // Collect CFG split points 3054 if (n->is_MultiBranch()) 3055 frc._tests.push(n); 3056 } 3057 3058 //------------------------------final_graph_reshaping_walk--------------------- 3059 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3060 // requires that the walk visits a node's inputs before visiting the node. 3061 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3062 ResourceArea *area = Thread::current()->resource_area(); 3063 Unique_Node_List sfpt(area); 3064 3065 frc._visited.set(root->_idx); // first, mark node as visited 3066 uint cnt = root->req(); 3067 Node *n = root; 3068 uint i = 0; 3069 while (true) { 3070 if (i < cnt) { 3071 // Place all non-visited non-null inputs onto stack 3072 Node* m = n->in(i); 3073 ++i; 3074 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3075 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) 3076 sfpt.push(m); 3077 cnt = m->req(); 3078 nstack.push(n, i); // put on stack parent and next input's index 3079 n = m; 3080 i = 0; 3081 } 3082 } else { 3083 // Now do post-visit work 3084 final_graph_reshaping_impl( n, frc ); 3085 if (nstack.is_empty()) 3086 break; // finished 3087 n = nstack.node(); // Get node from stack 3088 cnt = n->req(); 3089 i = nstack.index(); 3090 nstack.pop(); // Shift to the next node on stack 3091 } 3092 } 3093 3094 // Skip next transformation if compressed oops are not used. 3095 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3096 (!UseCompressedOops && !UseCompressedClassPointers)) 3097 return; 3098 3099 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3100 // It could be done for an uncommon traps or any safepoints/calls 3101 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3102 while (sfpt.size() > 0) { 3103 n = sfpt.pop(); 3104 JVMState *jvms = n->as_SafePoint()->jvms(); 3105 assert(jvms != NULL, "sanity"); 3106 int start = jvms->debug_start(); 3107 int end = n->req(); 3108 bool is_uncommon = (n->is_CallStaticJava() && 3109 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3110 for (int j = start; j < end; j++) { 3111 Node* in = n->in(j); 3112 if (in->is_DecodeNarrowPtr()) { 3113 bool safe_to_skip = true; 3114 if (!is_uncommon ) { 3115 // Is it safe to skip? 3116 for (uint i = 0; i < in->outcnt(); i++) { 3117 Node* u = in->raw_out(i); 3118 if (!u->is_SafePoint() || 3119 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3120 safe_to_skip = false; 3121 } 3122 } 3123 } 3124 if (safe_to_skip) { 3125 n->set_req(j, in->in(1)); 3126 } 3127 if (in->outcnt() == 0) { 3128 in->disconnect_inputs(NULL, this); 3129 } 3130 } 3131 } 3132 } 3133 } 3134 3135 //------------------------------final_graph_reshaping-------------------------- 3136 // Final Graph Reshaping. 3137 // 3138 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3139 // and not commoned up and forced early. Must come after regular 3140 // optimizations to avoid GVN undoing the cloning. Clone constant 3141 // inputs to Loop Phis; these will be split by the allocator anyways. 3142 // Remove Opaque nodes. 3143 // (2) Move last-uses by commutative operations to the left input to encourage 3144 // Intel update-in-place two-address operations and better register usage 3145 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3146 // calls canonicalizing them back. 3147 // (3) Count the number of double-precision FP ops, single-precision FP ops 3148 // and call sites. On Intel, we can get correct rounding either by 3149 // forcing singles to memory (requires extra stores and loads after each 3150 // FP bytecode) or we can set a rounding mode bit (requires setting and 3151 // clearing the mode bit around call sites). The mode bit is only used 3152 // if the relative frequency of single FP ops to calls is low enough. 3153 // This is a key transform for SPEC mpeg_audio. 3154 // (4) Detect infinite loops; blobs of code reachable from above but not 3155 // below. Several of the Code_Gen algorithms fail on such code shapes, 3156 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3157 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3158 // Detection is by looking for IfNodes where only 1 projection is 3159 // reachable from below or CatchNodes missing some targets. 3160 // (5) Assert for insane oop offsets in debug mode. 3161 3162 bool Compile::final_graph_reshaping() { 3163 // an infinite loop may have been eliminated by the optimizer, 3164 // in which case the graph will be empty. 3165 if (root()->req() == 1) { 3166 record_method_not_compilable("trivial infinite loop"); 3167 return true; 3168 } 3169 3170 // Expensive nodes have their control input set to prevent the GVN 3171 // from freely commoning them. There's no GVN beyond this point so 3172 // no need to keep the control input. We want the expensive nodes to 3173 // be freely moved to the least frequent code path by gcm. 3174 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3175 for (int i = 0; i < expensive_count(); i++) { 3176 _expensive_nodes->at(i)->set_req(0, NULL); 3177 } 3178 3179 Final_Reshape_Counts frc; 3180 3181 // Visit everybody reachable! 3182 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3183 Node_Stack nstack(unique() >> 1); 3184 final_graph_reshaping_walk(nstack, root(), frc); 3185 3186 // Check for unreachable (from below) code (i.e., infinite loops). 3187 for( uint i = 0; i < frc._tests.size(); i++ ) { 3188 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3189 // Get number of CFG targets. 3190 // Note that PCTables include exception targets after calls. 3191 uint required_outcnt = n->required_outcnt(); 3192 if (n->outcnt() != required_outcnt) { 3193 // Check for a few special cases. Rethrow Nodes never take the 3194 // 'fall-thru' path, so expected kids is 1 less. 3195 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3196 if (n->in(0)->in(0)->is_Call()) { 3197 CallNode *call = n->in(0)->in(0)->as_Call(); 3198 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3199 required_outcnt--; // Rethrow always has 1 less kid 3200 } else if (call->req() > TypeFunc::Parms && 3201 call->is_CallDynamicJava()) { 3202 // Check for null receiver. In such case, the optimizer has 3203 // detected that the virtual call will always result in a null 3204 // pointer exception. The fall-through projection of this CatchNode 3205 // will not be populated. 3206 Node *arg0 = call->in(TypeFunc::Parms); 3207 if (arg0->is_Type() && 3208 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3209 required_outcnt--; 3210 } 3211 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3212 call->req() > TypeFunc::Parms+1 && 3213 call->is_CallStaticJava()) { 3214 // Check for negative array length. In such case, the optimizer has 3215 // detected that the allocation attempt will always result in an 3216 // exception. There is no fall-through projection of this CatchNode . 3217 Node *arg1 = call->in(TypeFunc::Parms+1); 3218 if (arg1->is_Type() && 3219 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3220 required_outcnt--; 3221 } 3222 } 3223 } 3224 } 3225 // Recheck with a better notion of 'required_outcnt' 3226 if (n->outcnt() != required_outcnt) { 3227 record_method_not_compilable("malformed control flow"); 3228 return true; // Not all targets reachable! 3229 } 3230 } 3231 // Check that I actually visited all kids. Unreached kids 3232 // must be infinite loops. 3233 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3234 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3235 record_method_not_compilable("infinite loop"); 3236 return true; // Found unvisited kid; must be unreach 3237 } 3238 } 3239 3240 // If original bytecodes contained a mixture of floats and doubles 3241 // check if the optimizer has made it homogenous, item (3). 3242 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3243 frc.get_float_count() > 32 && 3244 frc.get_double_count() == 0 && 3245 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3246 set_24_bit_selection_and_mode( false, true ); 3247 } 3248 3249 set_java_calls(frc.get_java_call_count()); 3250 set_inner_loops(frc.get_inner_loop_count()); 3251 3252 // No infinite loops, no reason to bail out. 3253 return false; 3254 } 3255 3256 //-----------------------------too_many_traps---------------------------------- 3257 // Report if there are too many traps at the current method and bci. 3258 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3259 bool Compile::too_many_traps(ciMethod* method, 3260 int bci, 3261 Deoptimization::DeoptReason reason) { 3262 ciMethodData* md = method->method_data(); 3263 if (md->is_empty()) { 3264 // Assume the trap has not occurred, or that it occurred only 3265 // because of a transient condition during start-up in the interpreter. 3266 return false; 3267 } 3268 if (md->has_trap_at(bci, reason) != 0) { 3269 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3270 // Also, if there are multiple reasons, or if there is no per-BCI record, 3271 // assume the worst. 3272 if (log()) 3273 log()->elem("observe trap='%s' count='%d'", 3274 Deoptimization::trap_reason_name(reason), 3275 md->trap_count(reason)); 3276 return true; 3277 } else { 3278 // Ignore method/bci and see if there have been too many globally. 3279 return too_many_traps(reason, md); 3280 } 3281 } 3282 3283 // Less-accurate variant which does not require a method and bci. 3284 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3285 ciMethodData* logmd) { 3286 if (trap_count(reason) >= (uint)PerMethodTrapLimit) { 3287 // Too many traps globally. 3288 // Note that we use cumulative trap_count, not just md->trap_count. 3289 if (log()) { 3290 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3291 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3292 Deoptimization::trap_reason_name(reason), 3293 mcount, trap_count(reason)); 3294 } 3295 return true; 3296 } else { 3297 // The coast is clear. 3298 return false; 3299 } 3300 } 3301 3302 //--------------------------too_many_recompiles-------------------------------- 3303 // Report if there are too many recompiles at the current method and bci. 3304 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3305 // Is not eager to return true, since this will cause the compiler to use 3306 // Action_none for a trap point, to avoid too many recompilations. 3307 bool Compile::too_many_recompiles(ciMethod* method, 3308 int bci, 3309 Deoptimization::DeoptReason reason) { 3310 ciMethodData* md = method->method_data(); 3311 if (md->is_empty()) { 3312 // Assume the trap has not occurred, or that it occurred only 3313 // because of a transient condition during start-up in the interpreter. 3314 return false; 3315 } 3316 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3317 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3318 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3319 Deoptimization::DeoptReason per_bc_reason 3320 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3321 if ((per_bc_reason == Deoptimization::Reason_none 3322 || md->has_trap_at(bci, reason) != 0) 3323 // The trap frequency measure we care about is the recompile count: 3324 && md->trap_recompiled_at(bci) 3325 && md->overflow_recompile_count() >= bc_cutoff) { 3326 // Do not emit a trap here if it has already caused recompilations. 3327 // Also, if there are multiple reasons, or if there is no per-BCI record, 3328 // assume the worst. 3329 if (log()) 3330 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3331 Deoptimization::trap_reason_name(reason), 3332 md->trap_count(reason), 3333 md->overflow_recompile_count()); 3334 return true; 3335 } else if (trap_count(reason) != 0 3336 && decompile_count() >= m_cutoff) { 3337 // Too many recompiles globally, and we have seen this sort of trap. 3338 // Use cumulative decompile_count, not just md->decompile_count. 3339 if (log()) 3340 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3341 Deoptimization::trap_reason_name(reason), 3342 md->trap_count(reason), trap_count(reason), 3343 md->decompile_count(), decompile_count()); 3344 return true; 3345 } else { 3346 // The coast is clear. 3347 return false; 3348 } 3349 } 3350 3351 3352 #ifndef PRODUCT 3353 //------------------------------verify_graph_edges--------------------------- 3354 // Walk the Graph and verify that there is a one-to-one correspondence 3355 // between Use-Def edges and Def-Use edges in the graph. 3356 void Compile::verify_graph_edges(bool no_dead_code) { 3357 if (VerifyGraphEdges) { 3358 ResourceArea *area = Thread::current()->resource_area(); 3359 Unique_Node_List visited(area); 3360 // Call recursive graph walk to check edges 3361 _root->verify_edges(visited); 3362 if (no_dead_code) { 3363 // Now make sure that no visited node is used by an unvisited node. 3364 bool dead_nodes = 0; 3365 Unique_Node_List checked(area); 3366 while (visited.size() > 0) { 3367 Node* n = visited.pop(); 3368 checked.push(n); 3369 for (uint i = 0; i < n->outcnt(); i++) { 3370 Node* use = n->raw_out(i); 3371 if (checked.member(use)) continue; // already checked 3372 if (visited.member(use)) continue; // already in the graph 3373 if (use->is_Con()) continue; // a dead ConNode is OK 3374 // At this point, we have found a dead node which is DU-reachable. 3375 if (dead_nodes++ == 0) 3376 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3377 use->dump(2); 3378 tty->print_cr("---"); 3379 checked.push(use); // No repeats; pretend it is now checked. 3380 } 3381 } 3382 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3383 } 3384 } 3385 } 3386 3387 // Verify GC barriers consistency 3388 // Currently supported: 3389 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3390 void Compile::verify_barriers() { 3391 if (UseG1GC) { 3392 // Verify G1 pre-barriers 3393 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3394 3395 ResourceArea *area = Thread::current()->resource_area(); 3396 Unique_Node_List visited(area); 3397 Node_List worklist(area); 3398 // We're going to walk control flow backwards starting from the Root 3399 worklist.push(_root); 3400 while (worklist.size() > 0) { 3401 Node* x = worklist.pop(); 3402 if (x == NULL || x == top()) continue; 3403 if (visited.member(x)) { 3404 continue; 3405 } else { 3406 visited.push(x); 3407 } 3408 3409 if (x->is_Region()) { 3410 for (uint i = 1; i < x->req(); i++) { 3411 worklist.push(x->in(i)); 3412 } 3413 } else { 3414 worklist.push(x->in(0)); 3415 // We are looking for the pattern: 3416 // /->ThreadLocal 3417 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3418 // \->ConI(0) 3419 // We want to verify that the If and the LoadB have the same control 3420 // See GraphKit::g1_write_barrier_pre() 3421 if (x->is_If()) { 3422 IfNode *iff = x->as_If(); 3423 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3424 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3425 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3426 && cmp->in(1)->is_Load()) { 3427 LoadNode* load = cmp->in(1)->as_Load(); 3428 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3429 && load->in(2)->in(3)->is_Con() 3430 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3431 3432 Node* if_ctrl = iff->in(0); 3433 Node* load_ctrl = load->in(0); 3434 3435 if (if_ctrl != load_ctrl) { 3436 // Skip possible CProj->NeverBranch in infinite loops 3437 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3438 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3439 if_ctrl = if_ctrl->in(0)->in(0); 3440 } 3441 } 3442 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3443 } 3444 } 3445 } 3446 } 3447 } 3448 } 3449 } 3450 } 3451 3452 #endif 3453 3454 // The Compile object keeps track of failure reasons separately from the ciEnv. 3455 // This is required because there is not quite a 1-1 relation between the 3456 // ciEnv and its compilation task and the Compile object. Note that one 3457 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3458 // to backtrack and retry without subsuming loads. Other than this backtracking 3459 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3460 // by the logic in C2Compiler. 3461 void Compile::record_failure(const char* reason) { 3462 if (log() != NULL) { 3463 log()->elem("failure reason='%s' phase='compile'", reason); 3464 } 3465 if (_failure_reason == NULL) { 3466 // Record the first failure reason. 3467 _failure_reason = reason; 3468 } 3469 3470 EventCompilerFailure event; 3471 if (event.should_commit()) { 3472 event.set_compileID(Compile::compile_id()); 3473 event.set_failure(reason); 3474 event.commit(); 3475 } 3476 3477 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3478 C->print_method(PHASE_FAILURE); 3479 } 3480 _root = NULL; // flush the graph, too 3481 } 3482 3483 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3484 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3485 _phase_name(name), _dolog(dolog) 3486 { 3487 if (dolog) { 3488 C = Compile::current(); 3489 _log = C->log(); 3490 } else { 3491 C = NULL; 3492 _log = NULL; 3493 } 3494 if (_log != NULL) { 3495 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3496 _log->stamp(); 3497 _log->end_head(); 3498 } 3499 } 3500 3501 Compile::TracePhase::~TracePhase() { 3502 3503 C = Compile::current(); 3504 if (_dolog) { 3505 _log = C->log(); 3506 } else { 3507 _log = NULL; 3508 } 3509 3510 #ifdef ASSERT 3511 if (PrintIdealNodeCount) { 3512 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3513 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3514 } 3515 3516 if (VerifyIdealNodeCount) { 3517 Compile::current()->print_missing_nodes(); 3518 } 3519 #endif 3520 3521 if (_log != NULL) { 3522 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3523 } 3524 } 3525 3526 //============================================================================= 3527 // Two Constant's are equal when the type and the value are equal. 3528 bool Compile::Constant::operator==(const Constant& other) { 3529 if (type() != other.type() ) return false; 3530 if (can_be_reused() != other.can_be_reused()) return false; 3531 // For floating point values we compare the bit pattern. 3532 switch (type()) { 3533 case T_FLOAT: return (_v._value.i == other._v._value.i); 3534 case T_LONG: 3535 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3536 case T_OBJECT: 3537 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3538 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3539 case T_METADATA: return (_v._metadata == other._v._metadata); 3540 default: ShouldNotReachHere(); 3541 } 3542 return false; 3543 } 3544 3545 static int type_to_size_in_bytes(BasicType t) { 3546 switch (t) { 3547 case T_LONG: return sizeof(jlong ); 3548 case T_FLOAT: return sizeof(jfloat ); 3549 case T_DOUBLE: return sizeof(jdouble); 3550 case T_METADATA: return sizeof(Metadata*); 3551 // We use T_VOID as marker for jump-table entries (labels) which 3552 // need an internal word relocation. 3553 case T_VOID: 3554 case T_ADDRESS: 3555 case T_OBJECT: return sizeof(jobject); 3556 } 3557 3558 ShouldNotReachHere(); 3559 return -1; 3560 } 3561 3562 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3563 // sort descending 3564 if (a->freq() > b->freq()) return -1; 3565 if (a->freq() < b->freq()) return 1; 3566 return 0; 3567 } 3568 3569 void Compile::ConstantTable::calculate_offsets_and_size() { 3570 // First, sort the array by frequencies. 3571 _constants.sort(qsort_comparator); 3572 3573 #ifdef ASSERT 3574 // Make sure all jump-table entries were sorted to the end of the 3575 // array (they have a negative frequency). 3576 bool found_void = false; 3577 for (int i = 0; i < _constants.length(); i++) { 3578 Constant con = _constants.at(i); 3579 if (con.type() == T_VOID) 3580 found_void = true; // jump-tables 3581 else 3582 assert(!found_void, "wrong sorting"); 3583 } 3584 #endif 3585 3586 int offset = 0; 3587 for (int i = 0; i < _constants.length(); i++) { 3588 Constant* con = _constants.adr_at(i); 3589 3590 // Align offset for type. 3591 int typesize = type_to_size_in_bytes(con->type()); 3592 offset = align_size_up(offset, typesize); 3593 con->set_offset(offset); // set constant's offset 3594 3595 if (con->type() == T_VOID) { 3596 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3597 offset = offset + typesize * n->outcnt(); // expand jump-table 3598 } else { 3599 offset = offset + typesize; 3600 } 3601 } 3602 3603 // Align size up to the next section start (which is insts; see 3604 // CodeBuffer::align_at_start). 3605 assert(_size == -1, "already set?"); 3606 _size = align_size_up(offset, CodeEntryAlignment); 3607 } 3608 3609 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3610 MacroAssembler _masm(&cb); 3611 for (int i = 0; i < _constants.length(); i++) { 3612 Constant con = _constants.at(i); 3613 address constant_addr; 3614 switch (con.type()) { 3615 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3616 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3617 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3618 case T_OBJECT: { 3619 jobject obj = con.get_jobject(); 3620 int oop_index = _masm.oop_recorder()->find_index(obj); 3621 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3622 break; 3623 } 3624 case T_ADDRESS: { 3625 address addr = (address) con.get_jobject(); 3626 constant_addr = _masm.address_constant(addr); 3627 break; 3628 } 3629 // We use T_VOID as marker for jump-table entries (labels) which 3630 // need an internal word relocation. 3631 case T_VOID: { 3632 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3633 // Fill the jump-table with a dummy word. The real value is 3634 // filled in later in fill_jump_table. 3635 address dummy = (address) n; 3636 constant_addr = _masm.address_constant(dummy); 3637 // Expand jump-table 3638 for (uint i = 1; i < n->outcnt(); i++) { 3639 address temp_addr = _masm.address_constant(dummy + i); 3640 assert(temp_addr, "consts section too small"); 3641 } 3642 break; 3643 } 3644 case T_METADATA: { 3645 Metadata* obj = con.get_metadata(); 3646 int metadata_index = _masm.oop_recorder()->find_index(obj); 3647 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3648 break; 3649 } 3650 default: ShouldNotReachHere(); 3651 } 3652 assert(constant_addr, "consts section too small"); 3653 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3654 } 3655 } 3656 3657 int Compile::ConstantTable::find_offset(Constant& con) const { 3658 int idx = _constants.find(con); 3659 assert(idx != -1, "constant must be in constant table"); 3660 int offset = _constants.at(idx).offset(); 3661 assert(offset != -1, "constant table not emitted yet?"); 3662 return offset; 3663 } 3664 3665 void Compile::ConstantTable::add(Constant& con) { 3666 if (con.can_be_reused()) { 3667 int idx = _constants.find(con); 3668 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3669 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3670 return; 3671 } 3672 } 3673 (void) _constants.append(con); 3674 } 3675 3676 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3677 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3678 Constant con(type, value, b->_freq); 3679 add(con); 3680 return con; 3681 } 3682 3683 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3684 Constant con(metadata); 3685 add(con); 3686 return con; 3687 } 3688 3689 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3690 jvalue value; 3691 BasicType type = oper->type()->basic_type(); 3692 switch (type) { 3693 case T_LONG: value.j = oper->constantL(); break; 3694 case T_FLOAT: value.f = oper->constantF(); break; 3695 case T_DOUBLE: value.d = oper->constantD(); break; 3696 case T_OBJECT: 3697 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3698 case T_METADATA: return add((Metadata*)oper->constant()); break; 3699 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3700 } 3701 return add(n, type, value); 3702 } 3703 3704 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3705 jvalue value; 3706 // We can use the node pointer here to identify the right jump-table 3707 // as this method is called from Compile::Fill_buffer right before 3708 // the MachNodes are emitted and the jump-table is filled (means the 3709 // MachNode pointers do not change anymore). 3710 value.l = (jobject) n; 3711 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3712 add(con); 3713 return con; 3714 } 3715 3716 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3717 // If called from Compile::scratch_emit_size do nothing. 3718 if (Compile::current()->in_scratch_emit_size()) return; 3719 3720 assert(labels.is_nonempty(), "must be"); 3721 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3722 3723 // Since MachConstantNode::constant_offset() also contains 3724 // table_base_offset() we need to subtract the table_base_offset() 3725 // to get the plain offset into the constant table. 3726 int offset = n->constant_offset() - table_base_offset(); 3727 3728 MacroAssembler _masm(&cb); 3729 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3730 3731 for (uint i = 0; i < n->outcnt(); i++) { 3732 address* constant_addr = &jump_table_base[i]; 3733 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); 3734 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3735 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3736 } 3737 } 3738 3739 void Compile::dump_inlining() { 3740 if (print_inlining() || print_intrinsics()) { 3741 // Print inlining message for candidates that we couldn't inline 3742 // for lack of space or non constant receiver 3743 for (int i = 0; i < _late_inlines.length(); i++) { 3744 CallGenerator* cg = _late_inlines.at(i); 3745 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3746 } 3747 Unique_Node_List useful; 3748 useful.push(root()); 3749 for (uint next = 0; next < useful.size(); ++next) { 3750 Node* n = useful.at(next); 3751 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3752 CallNode* call = n->as_Call(); 3753 CallGenerator* cg = call->generator(); 3754 cg->print_inlining_late("receiver not constant"); 3755 } 3756 uint max = n->len(); 3757 for ( uint i = 0; i < max; ++i ) { 3758 Node *m = n->in(i); 3759 if ( m == NULL ) continue; 3760 useful.push(m); 3761 } 3762 } 3763 for (int i = 0; i < _print_inlining_list->length(); i++) { 3764 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string()); 3765 } 3766 } 3767 } 3768 3769 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3770 if (n1->Opcode() < n2->Opcode()) return -1; 3771 else if (n1->Opcode() > n2->Opcode()) return 1; 3772 3773 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())); 3774 for (uint i = 1; i < n1->req(); i++) { 3775 if (n1->in(i) < n2->in(i)) return -1; 3776 else if (n1->in(i) > n2->in(i)) return 1; 3777 } 3778 3779 return 0; 3780 } 3781 3782 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3783 Node* n1 = *n1p; 3784 Node* n2 = *n2p; 3785 3786 return cmp_expensive_nodes(n1, n2); 3787 } 3788 3789 void Compile::sort_expensive_nodes() { 3790 if (!expensive_nodes_sorted()) { 3791 _expensive_nodes->sort(cmp_expensive_nodes); 3792 } 3793 } 3794 3795 bool Compile::expensive_nodes_sorted() const { 3796 for (int i = 1; i < _expensive_nodes->length(); i++) { 3797 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3798 return false; 3799 } 3800 } 3801 return true; 3802 } 3803 3804 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3805 if (_expensive_nodes->length() == 0) { 3806 return false; 3807 } 3808 3809 assert(OptimizeExpensiveOps, "optimization off?"); 3810 3811 // Take this opportunity to remove dead nodes from the list 3812 int j = 0; 3813 for (int i = 0; i < _expensive_nodes->length(); i++) { 3814 Node* n = _expensive_nodes->at(i); 3815 if (!n->is_unreachable(igvn)) { 3816 assert(n->is_expensive(), "should be expensive"); 3817 _expensive_nodes->at_put(j, n); 3818 j++; 3819 } 3820 } 3821 _expensive_nodes->trunc_to(j); 3822 3823 // Then sort the list so that similar nodes are next to each other 3824 // and check for at least two nodes of identical kind with same data 3825 // inputs. 3826 sort_expensive_nodes(); 3827 3828 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3829 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3830 return true; 3831 } 3832 } 3833 3834 return false; 3835 } 3836 3837 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3838 if (_expensive_nodes->length() == 0) { 3839 return; 3840 } 3841 3842 assert(OptimizeExpensiveOps, "optimization off?"); 3843 3844 // Sort to bring similar nodes next to each other and clear the 3845 // control input of nodes for which there's only a single copy. 3846 sort_expensive_nodes(); 3847 3848 int j = 0; 3849 int identical = 0; 3850 int i = 0; 3851 for (; i < _expensive_nodes->length()-1; i++) { 3852 assert(j <= i, "can't write beyond current index"); 3853 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3854 identical++; 3855 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3856 continue; 3857 } 3858 if (identical > 0) { 3859 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3860 identical = 0; 3861 } else { 3862 Node* n = _expensive_nodes->at(i); 3863 igvn.hash_delete(n); 3864 n->set_req(0, NULL); 3865 igvn.hash_insert(n); 3866 } 3867 } 3868 if (identical > 0) { 3869 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3870 } else if (_expensive_nodes->length() >= 1) { 3871 Node* n = _expensive_nodes->at(i); 3872 igvn.hash_delete(n); 3873 n->set_req(0, NULL); 3874 igvn.hash_insert(n); 3875 } 3876 _expensive_nodes->trunc_to(j); 3877 } 3878 3879 void Compile::add_expensive_node(Node * n) { 3880 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3881 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3882 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3883 if (OptimizeExpensiveOps) { 3884 _expensive_nodes->append(n); 3885 } else { 3886 // Clear control input and let IGVN optimize expensive nodes if 3887 // OptimizeExpensiveOps is off. 3888 n->set_req(0, NULL); 3889 } 3890 } 3891 3892 /** 3893 * Remove the speculative part of types and clean up the graph 3894 */ 3895 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 3896 if (UseTypeSpeculation) { 3897 Unique_Node_List worklist; 3898 worklist.push(root()); 3899 int modified = 0; 3900 // Go over all type nodes that carry a speculative type, drop the 3901 // speculative part of the type and enqueue the node for an igvn 3902 // which may optimize it out. 3903 for (uint next = 0; next < worklist.size(); ++next) { 3904 Node *n = worklist.at(next); 3905 if (n->is_Type() && n->as_Type()->type()->isa_oopptr() != NULL && 3906 n->as_Type()->type()->is_oopptr()->speculative() != NULL) { 3907 TypeNode* tn = n->as_Type(); 3908 const TypeOopPtr* t = tn->type()->is_oopptr(); 3909 bool in_hash = igvn.hash_delete(n); 3910 assert(in_hash, "node should be in igvn hash table"); 3911 tn->set_type(t->remove_speculative()); 3912 igvn.hash_insert(n); 3913 igvn._worklist.push(n); // give it a chance to go away 3914 modified++; 3915 } 3916 uint max = n->len(); 3917 for( uint i = 0; i < max; ++i ) { 3918 Node *m = n->in(i); 3919 if (not_a_node(m)) continue; 3920 worklist.push(m); 3921 } 3922 } 3923 // Drop the speculative part of all types in the igvn's type table 3924 igvn.remove_speculative_types(); 3925 if (modified > 0) { 3926 igvn.optimize(); 3927 } 3928 } 3929 } 3930 3931 // Auxiliary method to support randomized stressing/fuzzing. 3932 // 3933 // This method can be called the arbitrary number of times, with current count 3934 // as the argument. The logic allows selecting a single candidate from the 3935 // running list of candidates as follows: 3936 // int count = 0; 3937 // Cand* selected = null; 3938 // while(cand = cand->next()) { 3939 // if (randomized_select(++count)) { 3940 // selected = cand; 3941 // } 3942 // } 3943 // 3944 // Including count equalizes the chances any candidate is "selected". 3945 // This is useful when we don't have the complete list of candidates to choose 3946 // from uniformly. In this case, we need to adjust the randomicity of the 3947 // selection, or else we will end up biasing the selection towards the latter 3948 // candidates. 3949 // 3950 // Quick back-envelope calculation shows that for the list of n candidates 3951 // the equal probability for the candidate to persist as "best" can be 3952 // achieved by replacing it with "next" k-th candidate with the probability 3953 // of 1/k. It can be easily shown that by the end of the run, the 3954 // probability for any candidate is converged to 1/n, thus giving the 3955 // uniform distribution among all the candidates. 3956 // 3957 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 3958 #define RANDOMIZED_DOMAIN_POW 29 3959 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 3960 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 3961 bool Compile::randomized_select(int count) { 3962 assert(count > 0, "only positive"); 3963 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 3964 }