1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "opto/addnode.hpp" 36 #include "opto/block.hpp" 37 #include "opto/c2compiler.hpp" 38 #include "opto/callGenerator.hpp" 39 #include "opto/callnode.hpp" 40 #include "opto/cfgnode.hpp" 41 #include "opto/chaitin.hpp" 42 #include "opto/compile.hpp" 43 #include "opto/connode.hpp" 44 #include "opto/divnode.hpp" 45 #include "opto/escape.hpp" 46 #include "opto/idealGraphPrinter.hpp" 47 #include "opto/loopnode.hpp" 48 #include "opto/machnode.hpp" 49 #include "opto/macro.hpp" 50 #include "opto/matcher.hpp" 51 #include "opto/mathexactnode.hpp" 52 #include "opto/memnode.hpp" 53 #include "opto/mulnode.hpp" 54 #include "opto/node.hpp" 55 #include "opto/opcodes.hpp" 56 #include "opto/output.hpp" 57 #include "opto/parse.hpp" 58 #include "opto/phaseX.hpp" 59 #include "opto/rootnode.hpp" 60 #include "opto/runtime.hpp" 61 #include "opto/stringopts.hpp" 62 #include "opto/type.hpp" 63 #include "opto/vectornode.hpp" 64 #include "runtime/arguments.hpp" 65 #include "runtime/signature.hpp" 66 #include "runtime/stubRoutines.hpp" 67 #include "runtime/timer.hpp" 68 #include "trace/tracing.hpp" 69 #include "utilities/copy.hpp" 70 #if defined AD_MD_HPP 71 # include AD_MD_HPP 72 #elif defined TARGET_ARCH_MODEL_x86_32 73 # include "adfiles/ad_x86_32.hpp" 74 #elif defined TARGET_ARCH_MODEL_x86_64 75 # include "adfiles/ad_x86_64.hpp" 76 #elif defined TARGET_ARCH_MODEL_sparc 77 # include "adfiles/ad_sparc.hpp" 78 #elif defined TARGET_ARCH_MODEL_zero 79 # include "adfiles/ad_zero.hpp" 80 #elif defined TARGET_ARCH_MODEL_ppc_64 81 # include "adfiles/ad_ppc_64.hpp" 82 #endif 83 84 85 // -------------------- Compile::mach_constant_base_node ----------------------- 86 // Constant table base node singleton. 87 MachConstantBaseNode* Compile::mach_constant_base_node() { 88 if (_mach_constant_base_node == NULL) { 89 _mach_constant_base_node = new (C) MachConstantBaseNode(); 90 _mach_constant_base_node->add_req(C->root()); 91 } 92 return _mach_constant_base_node; 93 } 94 95 96 /// Support for intrinsics. 97 98 // Return the index at which m must be inserted (or already exists). 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 101 #ifdef ASSERT 102 for (int i = 1; i < _intrinsics->length(); i++) { 103 CallGenerator* cg1 = _intrinsics->at(i-1); 104 CallGenerator* cg2 = _intrinsics->at(i); 105 assert(cg1->method() != cg2->method() 106 ? cg1->method() < cg2->method() 107 : cg1->is_virtual() < cg2->is_virtual(), 108 "compiler intrinsics list must stay sorted"); 109 } 110 #endif 111 // Binary search sorted list, in decreasing intervals [lo, hi]. 112 int lo = 0, hi = _intrinsics->length()-1; 113 while (lo <= hi) { 114 int mid = (uint)(hi + lo) / 2; 115 ciMethod* mid_m = _intrinsics->at(mid)->method(); 116 if (m < mid_m) { 117 hi = mid-1; 118 } else if (m > mid_m) { 119 lo = mid+1; 120 } else { 121 // look at minor sort key 122 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 123 if (is_virtual < mid_virt) { 124 hi = mid-1; 125 } else if (is_virtual > mid_virt) { 126 lo = mid+1; 127 } else { 128 return mid; // exact match 129 } 130 } 131 } 132 return lo; // inexact match 133 } 134 135 void Compile::register_intrinsic(CallGenerator* cg) { 136 if (_intrinsics == NULL) { 137 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 138 } 139 // This code is stolen from ciObjectFactory::insert. 140 // Really, GrowableArray should have methods for 141 // insert_at, remove_at, and binary_search. 142 int len = _intrinsics->length(); 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 144 if (index == len) { 145 _intrinsics->append(cg); 146 } else { 147 #ifdef ASSERT 148 CallGenerator* oldcg = _intrinsics->at(index); 149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 150 #endif 151 _intrinsics->append(_intrinsics->at(len-1)); 152 int pos; 153 for (pos = len-2; pos >= index; pos--) { 154 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 155 } 156 _intrinsics->at_put(index, cg); 157 } 158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 159 } 160 161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 162 assert(m->is_loaded(), "don't try this on unloaded methods"); 163 if (_intrinsics != NULL) { 164 int index = intrinsic_insertion_index(m, is_virtual); 165 if (index < _intrinsics->length() 166 && _intrinsics->at(index)->method() == m 167 && _intrinsics->at(index)->is_virtual() == is_virtual) { 168 return _intrinsics->at(index); 169 } 170 } 171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 172 if (m->intrinsic_id() != vmIntrinsics::_none && 173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 175 if (cg != NULL) { 176 // Save it for next time: 177 register_intrinsic(cg); 178 return cg; 179 } else { 180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 181 } 182 } 183 return NULL; 184 } 185 186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 187 // in library_call.cpp. 188 189 190 #ifndef PRODUCT 191 // statistics gathering... 192 193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 195 196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 198 int oflags = _intrinsic_hist_flags[id]; 199 assert(flags != 0, "what happened?"); 200 if (is_virtual) { 201 flags |= _intrinsic_virtual; 202 } 203 bool changed = (flags != oflags); 204 if ((flags & _intrinsic_worked) != 0) { 205 juint count = (_intrinsic_hist_count[id] += 1); 206 if (count == 1) { 207 changed = true; // first time 208 } 209 // increment the overall count also: 210 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 211 } 212 if (changed) { 213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 214 // Something changed about the intrinsic's virtuality. 215 if ((flags & _intrinsic_virtual) != 0) { 216 // This is the first use of this intrinsic as a virtual call. 217 if (oflags != 0) { 218 // We already saw it as a non-virtual, so note both cases. 219 flags |= _intrinsic_both; 220 } 221 } else if ((oflags & _intrinsic_both) == 0) { 222 // This is the first use of this intrinsic as a non-virtual 223 flags |= _intrinsic_both; 224 } 225 } 226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 227 } 228 // update the overall flags also: 229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 230 return changed; 231 } 232 233 static char* format_flags(int flags, char* buf) { 234 buf[0] = 0; 235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 240 if (buf[0] == 0) strcat(buf, ","); 241 assert(buf[0] == ',', "must be"); 242 return &buf[1]; 243 } 244 245 void Compile::print_intrinsic_statistics() { 246 char flagsbuf[100]; 247 ttyLocker ttyl; 248 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 249 tty->print_cr("Compiler intrinsic usage:"); 250 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 251 if (total == 0) total = 1; // avoid div0 in case of no successes 252 #define PRINT_STAT_LINE(name, c, f) \ 253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 255 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 256 int flags = _intrinsic_hist_flags[id]; 257 juint count = _intrinsic_hist_count[id]; 258 if ((flags | count) != 0) { 259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 260 } 261 } 262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 263 if (xtty != NULL) xtty->tail("statistics"); 264 } 265 266 void Compile::print_statistics() { 267 { ttyLocker ttyl; 268 if (xtty != NULL) xtty->head("statistics type='opto'"); 269 Parse::print_statistics(); 270 PhaseCCP::print_statistics(); 271 PhaseRegAlloc::print_statistics(); 272 Scheduling::print_statistics(); 273 PhasePeephole::print_statistics(); 274 PhaseIdealLoop::print_statistics(); 275 if (xtty != NULL) xtty->tail("statistics"); 276 } 277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 278 // put this under its own <statistics> element. 279 print_intrinsic_statistics(); 280 } 281 } 282 #endif //PRODUCT 283 284 // Support for bundling info 285 Bundle* Compile::node_bundling(const Node *n) { 286 assert(valid_bundle_info(n), "oob"); 287 return &_node_bundling_base[n->_idx]; 288 } 289 290 bool Compile::valid_bundle_info(const Node *n) { 291 return (_node_bundling_limit > n->_idx); 292 } 293 294 295 void Compile::gvn_replace_by(Node* n, Node* nn) { 296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 297 Node* use = n->last_out(i); 298 bool is_in_table = initial_gvn()->hash_delete(use); 299 uint uses_found = 0; 300 for (uint j = 0; j < use->len(); j++) { 301 if (use->in(j) == n) { 302 if (j < use->req()) 303 use->set_req(j, nn); 304 else 305 use->set_prec(j, nn); 306 uses_found++; 307 } 308 } 309 if (is_in_table) { 310 // reinsert into table 311 initial_gvn()->hash_find_insert(use); 312 } 313 record_for_igvn(use); 314 i -= uses_found; // we deleted 1 or more copies of this edge 315 } 316 } 317 318 319 static inline bool not_a_node(const Node* n) { 320 if (n == NULL) return true; 321 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 322 if (*(address*)n == badAddress) return true; // kill by Node::destruct 323 return false; 324 } 325 326 // Identify all nodes that are reachable from below, useful. 327 // Use breadth-first pass that records state in a Unique_Node_List, 328 // recursive traversal is slower. 329 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 330 int estimated_worklist_size = live_nodes(); 331 useful.map( estimated_worklist_size, NULL ); // preallocate space 332 333 // Initialize worklist 334 if (root() != NULL) { useful.push(root()); } 335 // If 'top' is cached, declare it useful to preserve cached node 336 if( cached_top_node() ) { useful.push(cached_top_node()); } 337 338 // Push all useful nodes onto the list, breadthfirst 339 for( uint next = 0; next < useful.size(); ++next ) { 340 assert( next < unique(), "Unique useful nodes < total nodes"); 341 Node *n = useful.at(next); 342 uint max = n->len(); 343 for( uint i = 0; i < max; ++i ) { 344 Node *m = n->in(i); 345 if (not_a_node(m)) continue; 346 useful.push(m); 347 } 348 } 349 } 350 351 // Update dead_node_list with any missing dead nodes using useful 352 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 353 void Compile::update_dead_node_list(Unique_Node_List &useful) { 354 uint max_idx = unique(); 355 VectorSet& useful_node_set = useful.member_set(); 356 357 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 358 // If node with index node_idx is not in useful set, 359 // mark it as dead in dead node list. 360 if (! useful_node_set.test(node_idx) ) { 361 record_dead_node(node_idx); 362 } 363 } 364 } 365 366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 367 int shift = 0; 368 for (int i = 0; i < inlines->length(); i++) { 369 CallGenerator* cg = inlines->at(i); 370 CallNode* call = cg->call_node(); 371 if (shift > 0) { 372 inlines->at_put(i-shift, cg); 373 } 374 if (!useful.member(call)) { 375 shift++; 376 } 377 } 378 inlines->trunc_to(inlines->length()-shift); 379 } 380 381 // Disconnect all useless nodes by disconnecting those at the boundary. 382 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 383 uint next = 0; 384 while (next < useful.size()) { 385 Node *n = useful.at(next++); 386 if (n->is_SafePoint()) { 387 // We're done with a parsing phase. Replaced nodes are not valid 388 // beyond that point. 389 n->as_SafePoint()->delete_replaced_nodes(); 390 } 391 // Use raw traversal of out edges since this code removes out edges 392 int max = n->outcnt(); 393 for (int j = 0; j < max; ++j) { 394 Node* child = n->raw_out(j); 395 if (! useful.member(child)) { 396 assert(!child->is_top() || child != top(), 397 "If top is cached in Compile object it is in useful list"); 398 // Only need to remove this out-edge to the useless node 399 n->raw_del_out(j); 400 --j; 401 --max; 402 } 403 } 404 if (n->outcnt() == 1 && n->has_special_unique_user()) { 405 record_for_igvn(n->unique_out()); 406 } 407 } 408 // Remove useless macro and predicate opaq nodes 409 for (int i = C->macro_count()-1; i >= 0; i--) { 410 Node* n = C->macro_node(i); 411 if (!useful.member(n)) { 412 remove_macro_node(n); 413 } 414 } 415 // Remove useless expensive node 416 for (int i = C->expensive_count()-1; i >= 0; i--) { 417 Node* n = C->expensive_node(i); 418 if (!useful.member(n)) { 419 remove_expensive_node(n); 420 } 421 } 422 // clean up the late inline lists 423 remove_useless_late_inlines(&_string_late_inlines, useful); 424 remove_useless_late_inlines(&_boxing_late_inlines, useful); 425 remove_useless_late_inlines(&_late_inlines, useful); 426 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 427 } 428 429 //------------------------------frame_size_in_words----------------------------- 430 // frame_slots in units of words 431 int Compile::frame_size_in_words() const { 432 // shift is 0 in LP32 and 1 in LP64 433 const int shift = (LogBytesPerWord - LogBytesPerInt); 434 int words = _frame_slots >> shift; 435 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 436 return words; 437 } 438 439 // To bang the stack of this compiled method we use the stack size 440 // that the interpreter would need in case of a deoptimization. This 441 // removes the need to bang the stack in the deoptimization blob which 442 // in turn simplifies stack overflow handling. 443 int Compile::bang_size_in_bytes() const { 444 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 445 } 446 447 // ============================================================================ 448 //------------------------------CompileWrapper--------------------------------- 449 class CompileWrapper : public StackObj { 450 Compile *const _compile; 451 public: 452 CompileWrapper(Compile* compile); 453 454 ~CompileWrapper(); 455 }; 456 457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 458 // the Compile* pointer is stored in the current ciEnv: 459 ciEnv* env = compile->env(); 460 assert(env == ciEnv::current(), "must already be a ciEnv active"); 461 assert(env->compiler_data() == NULL, "compile already active?"); 462 env->set_compiler_data(compile); 463 assert(compile == Compile::current(), "sanity"); 464 465 compile->set_type_dict(NULL); 466 compile->set_type_hwm(NULL); 467 compile->set_type_last_size(0); 468 compile->set_last_tf(NULL, NULL); 469 compile->set_indexSet_arena(NULL); 470 compile->set_indexSet_free_block_list(NULL); 471 compile->init_type_arena(); 472 Type::Initialize(compile); 473 _compile->set_scratch_buffer_blob(NULL); 474 _compile->begin_method(); 475 } 476 CompileWrapper::~CompileWrapper() { 477 _compile->end_method(); 478 if (_compile->scratch_buffer_blob() != NULL) 479 BufferBlob::free(_compile->scratch_buffer_blob()); 480 _compile->env()->set_compiler_data(NULL); 481 } 482 483 484 //----------------------------print_compile_messages--------------------------- 485 void Compile::print_compile_messages() { 486 #ifndef PRODUCT 487 // Check if recompiling 488 if (_subsume_loads == false && PrintOpto) { 489 // Recompiling without allowing machine instructions to subsume loads 490 tty->print_cr("*********************************************************"); 491 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 492 tty->print_cr("*********************************************************"); 493 } 494 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 495 // Recompiling without escape analysis 496 tty->print_cr("*********************************************************"); 497 tty->print_cr("** Bailout: Recompile without escape analysis **"); 498 tty->print_cr("*********************************************************"); 499 } 500 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 501 // Recompiling without boxing elimination 502 tty->print_cr("*********************************************************"); 503 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 504 tty->print_cr("*********************************************************"); 505 } 506 if (env()->break_at_compile()) { 507 // Open the debugger when compiling this method. 508 tty->print("### Breaking when compiling: "); 509 method()->print_short_name(); 510 tty->cr(); 511 BREAKPOINT; 512 } 513 514 if( PrintOpto ) { 515 if (is_osr_compilation()) { 516 tty->print("[OSR]%3d", _compile_id); 517 } else { 518 tty->print("%3d", _compile_id); 519 } 520 } 521 #endif 522 } 523 524 525 //-----------------------init_scratch_buffer_blob------------------------------ 526 // Construct a temporary BufferBlob and cache it for this compile. 527 void Compile::init_scratch_buffer_blob(int const_size) { 528 // If there is already a scratch buffer blob allocated and the 529 // constant section is big enough, use it. Otherwise free the 530 // current and allocate a new one. 531 BufferBlob* blob = scratch_buffer_blob(); 532 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 533 // Use the current blob. 534 } else { 535 if (blob != NULL) { 536 BufferBlob::free(blob); 537 } 538 539 ResourceMark rm; 540 _scratch_const_size = const_size; 541 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 542 blob = BufferBlob::create("Compile::scratch_buffer", size); 543 // Record the buffer blob for next time. 544 set_scratch_buffer_blob(blob); 545 // Have we run out of code space? 546 if (scratch_buffer_blob() == NULL) { 547 // Let CompilerBroker disable further compilations. 548 record_failure("Not enough space for scratch buffer in CodeCache"); 549 return; 550 } 551 } 552 553 // Initialize the relocation buffers 554 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 555 set_scratch_locs_memory(locs_buf); 556 } 557 558 559 //-----------------------scratch_emit_size------------------------------------- 560 // Helper function that computes size by emitting code 561 uint Compile::scratch_emit_size(const Node* n) { 562 // Start scratch_emit_size section. 563 set_in_scratch_emit_size(true); 564 565 // Emit into a trash buffer and count bytes emitted. 566 // This is a pretty expensive way to compute a size, 567 // but it works well enough if seldom used. 568 // All common fixed-size instructions are given a size 569 // method by the AD file. 570 // Note that the scratch buffer blob and locs memory are 571 // allocated at the beginning of the compile task, and 572 // may be shared by several calls to scratch_emit_size. 573 // The allocation of the scratch buffer blob is particularly 574 // expensive, since it has to grab the code cache lock. 575 BufferBlob* blob = this->scratch_buffer_blob(); 576 assert(blob != NULL, "Initialize BufferBlob at start"); 577 assert(blob->size() > MAX_inst_size, "sanity"); 578 relocInfo* locs_buf = scratch_locs_memory(); 579 address blob_begin = blob->content_begin(); 580 address blob_end = (address)locs_buf; 581 assert(blob->content_contains(blob_end), "sanity"); 582 CodeBuffer buf(blob_begin, blob_end - blob_begin); 583 buf.initialize_consts_size(_scratch_const_size); 584 buf.initialize_stubs_size(MAX_stubs_size); 585 assert(locs_buf != NULL, "sanity"); 586 int lsize = MAX_locs_size / 3; 587 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 588 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 589 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 590 591 // Do the emission. 592 593 Label fakeL; // Fake label for branch instructions. 594 Label* saveL = NULL; 595 uint save_bnum = 0; 596 bool is_branch = n->is_MachBranch(); 597 if (is_branch) { 598 MacroAssembler masm(&buf); 599 masm.bind(fakeL); 600 n->as_MachBranch()->save_label(&saveL, &save_bnum); 601 n->as_MachBranch()->label_set(&fakeL, 0); 602 } 603 n->emit(buf, this->regalloc()); 604 if (is_branch) // Restore label. 605 n->as_MachBranch()->label_set(saveL, save_bnum); 606 607 // End scratch_emit_size section. 608 set_in_scratch_emit_size(false); 609 610 return buf.insts_size(); 611 } 612 613 614 // ============================================================================ 615 //------------------------------Compile standard------------------------------- 616 debug_only( int Compile::_debug_idx = 100000; ) 617 618 // Compile a method. entry_bci is -1 for normal compilations and indicates 619 // the continuation bci for on stack replacement. 620 621 622 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 623 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 624 : Phase(Compiler), 625 _env(ci_env), 626 _log(ci_env->log()), 627 _compile_id(ci_env->compile_id()), 628 _save_argument_registers(false), 629 _stub_name(NULL), 630 _stub_function(NULL), 631 _stub_entry_point(NULL), 632 _method(target), 633 _entry_bci(osr_bci), 634 _initial_gvn(NULL), 635 _for_igvn(NULL), 636 _warm_calls(NULL), 637 _subsume_loads(subsume_loads), 638 _do_escape_analysis(do_escape_analysis), 639 _eliminate_boxing(eliminate_boxing), 640 _failure_reason(NULL), 641 _code_buffer("Compile::Fill_buffer"), 642 _orig_pc_slot(0), 643 _orig_pc_slot_offset_in_bytes(0), 644 _has_method_handle_invokes(false), 645 _mach_constant_base_node(NULL), 646 _node_bundling_limit(0), 647 _node_bundling_base(NULL), 648 _java_calls(0), 649 _inner_loops(0), 650 _scratch_const_size(-1), 651 _in_scratch_emit_size(false), 652 _dead_node_list(comp_arena()), 653 _dead_node_count(0), 654 #ifndef PRODUCT 655 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 656 _in_dump_cnt(0), 657 _printer(IdealGraphPrinter::printer()), 658 #endif 659 _congraph(NULL), 660 _comp_arena(mtCompiler), 661 _node_arena(mtCompiler), 662 _old_arena(mtCompiler), 663 _Compile_types(mtCompiler), 664 _replay_inline_data(NULL), 665 _late_inlines(comp_arena(), 2, 0, NULL), 666 _string_late_inlines(comp_arena(), 2, 0, NULL), 667 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 668 _late_inlines_pos(0), 669 _number_of_mh_late_inlines(0), 670 _inlining_progress(false), 671 _inlining_incrementally(false), 672 _print_inlining_list(NULL), 673 _print_inlining_idx(0), 674 _interpreter_frame_size(0), 675 _max_node_limit(MaxNodeLimit) { 676 C = this; 677 678 CompileWrapper cw(this); 679 #ifndef PRODUCT 680 if (TimeCompiler2) { 681 tty->print(" "); 682 target->holder()->name()->print(); 683 tty->print("."); 684 target->print_short_name(); 685 tty->print(" "); 686 } 687 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 688 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 689 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 690 if (!print_opto_assembly) { 691 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 692 if (print_assembly && !Disassembler::can_decode()) { 693 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 694 print_opto_assembly = true; 695 } 696 } 697 set_print_assembly(print_opto_assembly); 698 set_parsed_irreducible_loop(false); 699 700 if (method()->has_option("ReplayInline")) { 701 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 702 } 703 #endif 704 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 705 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 706 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 707 708 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 709 // Make sure the method being compiled gets its own MDO, 710 // so we can at least track the decompile_count(). 711 // Need MDO to record RTM code generation state. 712 method()->ensure_method_data(); 713 } 714 715 Init(::AliasLevel); 716 717 718 print_compile_messages(); 719 720 _ilt = InlineTree::build_inline_tree_root(); 721 722 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 723 assert(num_alias_types() >= AliasIdxRaw, ""); 724 725 #define MINIMUM_NODE_HASH 1023 726 // Node list that Iterative GVN will start with 727 Unique_Node_List for_igvn(comp_arena()); 728 set_for_igvn(&for_igvn); 729 730 // GVN that will be run immediately on new nodes 731 uint estimated_size = method()->code_size()*4+64; 732 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 733 PhaseGVN gvn(node_arena(), estimated_size); 734 set_initial_gvn(&gvn); 735 736 if (print_inlining() || print_intrinsics()) { 737 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 738 } 739 { // Scope for timing the parser 740 TracePhase t3("parse", &_t_parser, true); 741 742 // Put top into the hash table ASAP. 743 initial_gvn()->transform_no_reclaim(top()); 744 745 // Set up tf(), start(), and find a CallGenerator. 746 CallGenerator* cg = NULL; 747 if (is_osr_compilation()) { 748 const TypeTuple *domain = StartOSRNode::osr_domain(); 749 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 750 init_tf(TypeFunc::make(domain, range)); 751 StartNode* s = new (this) StartOSRNode(root(), domain); 752 initial_gvn()->set_type_bottom(s); 753 init_start(s); 754 cg = CallGenerator::for_osr(method(), entry_bci()); 755 } else { 756 // Normal case. 757 init_tf(TypeFunc::make(method())); 758 StartNode* s = new (this) StartNode(root(), tf()->domain()); 759 initial_gvn()->set_type_bottom(s); 760 init_start(s); 761 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 762 // With java.lang.ref.reference.get() we must go through the 763 // intrinsic when G1 is enabled - even when get() is the root 764 // method of the compile - so that, if necessary, the value in 765 // the referent field of the reference object gets recorded by 766 // the pre-barrier code. 767 // Specifically, if G1 is enabled, the value in the referent 768 // field is recorded by the G1 SATB pre barrier. This will 769 // result in the referent being marked live and the reference 770 // object removed from the list of discovered references during 771 // reference processing. 772 cg = find_intrinsic(method(), false); 773 } 774 if (cg == NULL) { 775 float past_uses = method()->interpreter_invocation_count(); 776 float expected_uses = past_uses; 777 cg = CallGenerator::for_inline(method(), expected_uses); 778 } 779 } 780 if (failing()) return; 781 if (cg == NULL) { 782 record_method_not_compilable_all_tiers("cannot parse method"); 783 return; 784 } 785 JVMState* jvms = build_start_state(start(), tf()); 786 if ((jvms = cg->generate(jvms)) == NULL) { 787 record_method_not_compilable("method parse failed"); 788 return; 789 } 790 GraphKit kit(jvms); 791 792 if (!kit.stopped()) { 793 // Accept return values, and transfer control we know not where. 794 // This is done by a special, unique ReturnNode bound to root. 795 return_values(kit.jvms()); 796 } 797 798 if (kit.has_exceptions()) { 799 // Any exceptions that escape from this call must be rethrown 800 // to whatever caller is dynamically above us on the stack. 801 // This is done by a special, unique RethrowNode bound to root. 802 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 803 } 804 805 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 806 807 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 808 inline_string_calls(true); 809 } 810 811 if (failing()) return; 812 813 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 814 815 // Remove clutter produced by parsing. 816 if (!failing()) { 817 ResourceMark rm; 818 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 819 } 820 } 821 822 // Note: Large methods are capped off in do_one_bytecode(). 823 if (failing()) return; 824 825 // After parsing, node notes are no longer automagic. 826 // They must be propagated by register_new_node_with_optimizer(), 827 // clone(), or the like. 828 set_default_node_notes(NULL); 829 830 for (;;) { 831 int successes = Inline_Warm(); 832 if (failing()) return; 833 if (successes == 0) break; 834 } 835 836 // Drain the list. 837 Finish_Warm(); 838 #ifndef PRODUCT 839 if (_printer) { 840 _printer->print_inlining(this); 841 } 842 #endif 843 844 if (failing()) return; 845 NOT_PRODUCT( verify_graph_edges(); ) 846 847 // Now optimize 848 Optimize(); 849 if (failing()) return; 850 NOT_PRODUCT( verify_graph_edges(); ) 851 852 #ifndef PRODUCT 853 if (PrintIdeal) { 854 ttyLocker ttyl; // keep the following output all in one block 855 // This output goes directly to the tty, not the compiler log. 856 // To enable tools to match it up with the compilation activity, 857 // be sure to tag this tty output with the compile ID. 858 if (xtty != NULL) { 859 xtty->head("ideal compile_id='%d'%s", compile_id(), 860 is_osr_compilation() ? " compile_kind='osr'" : 861 ""); 862 } 863 root()->dump(9999); 864 if (xtty != NULL) { 865 xtty->tail("ideal"); 866 } 867 } 868 #endif 869 870 NOT_PRODUCT( verify_barriers(); ) 871 872 // Dump compilation data to replay it. 873 if (method()->has_option("DumpReplay")) { 874 env()->dump_replay_data(_compile_id); 875 } 876 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 877 env()->dump_inline_data(_compile_id); 878 } 879 880 // Now that we know the size of all the monitors we can add a fixed slot 881 // for the original deopt pc. 882 883 _orig_pc_slot = fixed_slots(); 884 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 885 set_fixed_slots(next_slot); 886 887 // Compute when to use implicit null checks. Used by matching trap based 888 // nodes and NullCheck optimization. 889 set_allowed_deopt_reasons(); 890 891 // Now generate code 892 Code_Gen(); 893 if (failing()) return; 894 895 // Check if we want to skip execution of all compiled code. 896 { 897 #ifndef PRODUCT 898 if (OptoNoExecute) { 899 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 900 return; 901 } 902 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 903 #endif 904 905 if (is_osr_compilation()) { 906 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 907 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 908 } else { 909 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 910 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 911 } 912 913 env()->register_method(_method, _entry_bci, 914 &_code_offsets, 915 _orig_pc_slot_offset_in_bytes, 916 code_buffer(), 917 frame_size_in_words(), _oop_map_set, 918 &_handler_table, &_inc_table, 919 compiler, 920 env()->comp_level(), 921 has_unsafe_access(), 922 SharedRuntime::is_wide_vector(max_vector_size()), 923 rtm_state() 924 ); 925 926 if (log() != NULL) // Print code cache state into compiler log 927 log()->code_cache_state(); 928 } 929 } 930 931 //------------------------------Compile---------------------------------------- 932 // Compile a runtime stub 933 Compile::Compile( ciEnv* ci_env, 934 TypeFunc_generator generator, 935 address stub_function, 936 const char *stub_name, 937 int is_fancy_jump, 938 bool pass_tls, 939 bool save_arg_registers, 940 bool return_pc ) 941 : Phase(Compiler), 942 _env(ci_env), 943 _log(ci_env->log()), 944 _compile_id(0), 945 _save_argument_registers(save_arg_registers), 946 _method(NULL), 947 _stub_name(stub_name), 948 _stub_function(stub_function), 949 _stub_entry_point(NULL), 950 _entry_bci(InvocationEntryBci), 951 _initial_gvn(NULL), 952 _for_igvn(NULL), 953 _warm_calls(NULL), 954 _orig_pc_slot(0), 955 _orig_pc_slot_offset_in_bytes(0), 956 _subsume_loads(true), 957 _do_escape_analysis(false), 958 _eliminate_boxing(false), 959 _failure_reason(NULL), 960 _code_buffer("Compile::Fill_buffer"), 961 _has_method_handle_invokes(false), 962 _mach_constant_base_node(NULL), 963 _node_bundling_limit(0), 964 _node_bundling_base(NULL), 965 _java_calls(0), 966 _inner_loops(0), 967 #ifndef PRODUCT 968 _trace_opto_output(TraceOptoOutput), 969 _in_dump_cnt(0), 970 _printer(NULL), 971 #endif 972 _comp_arena(mtCompiler), 973 _node_arena(mtCompiler), 974 _old_arena(mtCompiler), 975 _Compile_types(mtCompiler), 976 _dead_node_list(comp_arena()), 977 _dead_node_count(0), 978 _congraph(NULL), 979 _replay_inline_data(NULL), 980 _number_of_mh_late_inlines(0), 981 _inlining_progress(false), 982 _inlining_incrementally(false), 983 _print_inlining_list(NULL), 984 _print_inlining_idx(0), 985 _allowed_reasons(0), 986 _interpreter_frame_size(0), 987 _max_node_limit(MaxNodeLimit) { 988 C = this; 989 990 #ifndef PRODUCT 991 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 992 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 993 set_print_assembly(PrintFrameConverterAssembly); 994 set_parsed_irreducible_loop(false); 995 #endif 996 set_has_irreducible_loop(false); // no loops 997 998 CompileWrapper cw(this); 999 Init(/*AliasLevel=*/ 0); 1000 init_tf((*generator)()); 1001 1002 { 1003 // The following is a dummy for the sake of GraphKit::gen_stub 1004 Unique_Node_List for_igvn(comp_arena()); 1005 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1006 PhaseGVN gvn(Thread::current()->resource_area(),255); 1007 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1008 gvn.transform_no_reclaim(top()); 1009 1010 GraphKit kit; 1011 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1012 } 1013 1014 NOT_PRODUCT( verify_graph_edges(); ) 1015 Code_Gen(); 1016 if (failing()) return; 1017 1018 1019 // Entry point will be accessed using compile->stub_entry_point(); 1020 if (code_buffer() == NULL) { 1021 Matcher::soft_match_failure(); 1022 } else { 1023 if (PrintAssembly && (WizardMode || Verbose)) 1024 tty->print_cr("### Stub::%s", stub_name); 1025 1026 if (!failing()) { 1027 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1028 1029 // Make the NMethod 1030 // For now we mark the frame as never safe for profile stackwalking 1031 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1032 code_buffer(), 1033 CodeOffsets::frame_never_safe, 1034 // _code_offsets.value(CodeOffsets::Frame_Complete), 1035 frame_size_in_words(), 1036 _oop_map_set, 1037 save_arg_registers); 1038 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1039 1040 _stub_entry_point = rs->entry_point(); 1041 } 1042 } 1043 } 1044 1045 //------------------------------Init------------------------------------------- 1046 // Prepare for a single compilation 1047 void Compile::Init(int aliaslevel) { 1048 _unique = 0; 1049 _regalloc = NULL; 1050 1051 _tf = NULL; // filled in later 1052 _top = NULL; // cached later 1053 _matcher = NULL; // filled in later 1054 _cfg = NULL; // filled in later 1055 1056 set_24_bit_selection_and_mode(Use24BitFP, false); 1057 1058 _node_note_array = NULL; 1059 _default_node_notes = NULL; 1060 1061 _immutable_memory = NULL; // filled in at first inquiry 1062 1063 // Globally visible Nodes 1064 // First set TOP to NULL to give safe behavior during creation of RootNode 1065 set_cached_top_node(NULL); 1066 set_root(new (this) RootNode()); 1067 // Now that you have a Root to point to, create the real TOP 1068 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1069 set_recent_alloc(NULL, NULL); 1070 1071 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1072 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1073 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1074 env()->set_dependencies(new Dependencies(env())); 1075 1076 _fixed_slots = 0; 1077 set_has_split_ifs(false); 1078 set_has_loops(has_method() && method()->has_loops()); // first approximation 1079 set_has_stringbuilder(false); 1080 set_has_boxed_value(false); 1081 _trap_can_recompile = false; // no traps emitted yet 1082 _major_progress = true; // start out assuming good things will happen 1083 set_has_unsafe_access(false); 1084 set_max_vector_size(0); 1085 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1086 set_decompile_count(0); 1087 1088 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1089 set_num_loop_opts(LoopOptsCount); 1090 set_do_inlining(Inline); 1091 set_max_inline_size(MaxInlineSize); 1092 set_freq_inline_size(FreqInlineSize); 1093 set_do_scheduling(OptoScheduling); 1094 set_do_count_invocations(false); 1095 set_do_method_data_update(false); 1096 set_rtm_state(NoRTM); // No RTM lock eliding by default 1097 method_has_option_value("MaxNodeLimit", _max_node_limit); 1098 #if INCLUDE_RTM_OPT 1099 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1100 int rtm_state = method()->method_data()->rtm_state(); 1101 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1102 // Don't generate RTM lock eliding code. 1103 set_rtm_state(NoRTM); 1104 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1105 // Generate RTM lock eliding code without abort ratio calculation code. 1106 set_rtm_state(UseRTM); 1107 } else if (UseRTMDeopt) { 1108 // Generate RTM lock eliding code and include abort ratio calculation 1109 // code if UseRTMDeopt is on. 1110 set_rtm_state(ProfileRTM); 1111 } 1112 } 1113 #endif 1114 if (debug_info()->recording_non_safepoints()) { 1115 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1116 (comp_arena(), 8, 0, NULL)); 1117 set_default_node_notes(Node_Notes::make(this)); 1118 } 1119 1120 // // -- Initialize types before each compile -- 1121 // // Update cached type information 1122 // if( _method && _method->constants() ) 1123 // Type::update_loaded_types(_method, _method->constants()); 1124 1125 // Init alias_type map. 1126 if (!_do_escape_analysis && aliaslevel == 3) 1127 aliaslevel = 2; // No unique types without escape analysis 1128 _AliasLevel = aliaslevel; 1129 const int grow_ats = 16; 1130 _max_alias_types = grow_ats; 1131 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1132 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1133 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1134 { 1135 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1136 } 1137 // Initialize the first few types. 1138 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1139 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1140 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1141 _num_alias_types = AliasIdxRaw+1; 1142 // Zero out the alias type cache. 1143 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1144 // A NULL adr_type hits in the cache right away. Preload the right answer. 1145 probe_alias_cache(NULL)->_index = AliasIdxTop; 1146 1147 _intrinsics = NULL; 1148 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1149 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1150 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1151 register_library_intrinsics(); 1152 } 1153 1154 //---------------------------init_start---------------------------------------- 1155 // Install the StartNode on this compile object. 1156 void Compile::init_start(StartNode* s) { 1157 if (failing()) 1158 return; // already failing 1159 assert(s == start(), ""); 1160 } 1161 1162 StartNode* Compile::start() const { 1163 assert(!failing(), ""); 1164 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1165 Node* start = root()->fast_out(i); 1166 if( start->is_Start() ) 1167 return start->as_Start(); 1168 } 1169 fatal("Did not find Start node!"); 1170 return NULL; 1171 } 1172 1173 //-------------------------------immutable_memory------------------------------------- 1174 // Access immutable memory 1175 Node* Compile::immutable_memory() { 1176 if (_immutable_memory != NULL) { 1177 return _immutable_memory; 1178 } 1179 StartNode* s = start(); 1180 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1181 Node *p = s->fast_out(i); 1182 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1183 _immutable_memory = p; 1184 return _immutable_memory; 1185 } 1186 } 1187 ShouldNotReachHere(); 1188 return NULL; 1189 } 1190 1191 //----------------------set_cached_top_node------------------------------------ 1192 // Install the cached top node, and make sure Node::is_top works correctly. 1193 void Compile::set_cached_top_node(Node* tn) { 1194 if (tn != NULL) verify_top(tn); 1195 Node* old_top = _top; 1196 _top = tn; 1197 // Calling Node::setup_is_top allows the nodes the chance to adjust 1198 // their _out arrays. 1199 if (_top != NULL) _top->setup_is_top(); 1200 if (old_top != NULL) old_top->setup_is_top(); 1201 assert(_top == NULL || top()->is_top(), ""); 1202 } 1203 1204 #ifdef ASSERT 1205 uint Compile::count_live_nodes_by_graph_walk() { 1206 Unique_Node_List useful(comp_arena()); 1207 // Get useful node list by walking the graph. 1208 identify_useful_nodes(useful); 1209 return useful.size(); 1210 } 1211 1212 void Compile::print_missing_nodes() { 1213 1214 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1215 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1216 return; 1217 } 1218 1219 // This is an expensive function. It is executed only when the user 1220 // specifies VerifyIdealNodeCount option or otherwise knows the 1221 // additional work that needs to be done to identify reachable nodes 1222 // by walking the flow graph and find the missing ones using 1223 // _dead_node_list. 1224 1225 Unique_Node_List useful(comp_arena()); 1226 // Get useful node list by walking the graph. 1227 identify_useful_nodes(useful); 1228 1229 uint l_nodes = C->live_nodes(); 1230 uint l_nodes_by_walk = useful.size(); 1231 1232 if (l_nodes != l_nodes_by_walk) { 1233 if (_log != NULL) { 1234 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1235 _log->stamp(); 1236 _log->end_head(); 1237 } 1238 VectorSet& useful_member_set = useful.member_set(); 1239 int last_idx = l_nodes_by_walk; 1240 for (int i = 0; i < last_idx; i++) { 1241 if (useful_member_set.test(i)) { 1242 if (_dead_node_list.test(i)) { 1243 if (_log != NULL) { 1244 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1245 } 1246 if (PrintIdealNodeCount) { 1247 // Print the log message to tty 1248 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1249 useful.at(i)->dump(); 1250 } 1251 } 1252 } 1253 else if (! _dead_node_list.test(i)) { 1254 if (_log != NULL) { 1255 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1256 } 1257 if (PrintIdealNodeCount) { 1258 // Print the log message to tty 1259 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1260 } 1261 } 1262 } 1263 if (_log != NULL) { 1264 _log->tail("mismatched_nodes"); 1265 } 1266 } 1267 } 1268 #endif 1269 1270 #ifndef PRODUCT 1271 void Compile::verify_top(Node* tn) const { 1272 if (tn != NULL) { 1273 assert(tn->is_Con(), "top node must be a constant"); 1274 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1275 assert(tn->in(0) != NULL, "must have live top node"); 1276 } 1277 } 1278 #endif 1279 1280 1281 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1282 1283 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1284 guarantee(arr != NULL, ""); 1285 int num_blocks = arr->length(); 1286 if (grow_by < num_blocks) grow_by = num_blocks; 1287 int num_notes = grow_by * _node_notes_block_size; 1288 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1289 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1290 while (num_notes > 0) { 1291 arr->append(notes); 1292 notes += _node_notes_block_size; 1293 num_notes -= _node_notes_block_size; 1294 } 1295 assert(num_notes == 0, "exact multiple, please"); 1296 } 1297 1298 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1299 if (source == NULL || dest == NULL) return false; 1300 1301 if (dest->is_Con()) 1302 return false; // Do not push debug info onto constants. 1303 1304 #ifdef ASSERT 1305 // Leave a bread crumb trail pointing to the original node: 1306 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1307 dest->set_debug_orig(source); 1308 } 1309 #endif 1310 1311 if (node_note_array() == NULL) 1312 return false; // Not collecting any notes now. 1313 1314 // This is a copy onto a pre-existing node, which may already have notes. 1315 // If both nodes have notes, do not overwrite any pre-existing notes. 1316 Node_Notes* source_notes = node_notes_at(source->_idx); 1317 if (source_notes == NULL || source_notes->is_clear()) return false; 1318 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1319 if (dest_notes == NULL || dest_notes->is_clear()) { 1320 return set_node_notes_at(dest->_idx, source_notes); 1321 } 1322 1323 Node_Notes merged_notes = (*source_notes); 1324 // The order of operations here ensures that dest notes will win... 1325 merged_notes.update_from(dest_notes); 1326 return set_node_notes_at(dest->_idx, &merged_notes); 1327 } 1328 1329 1330 //--------------------------allow_range_check_smearing------------------------- 1331 // Gating condition for coalescing similar range checks. 1332 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1333 // single covering check that is at least as strong as any of them. 1334 // If the optimization succeeds, the simplified (strengthened) range check 1335 // will always succeed. If it fails, we will deopt, and then give up 1336 // on the optimization. 1337 bool Compile::allow_range_check_smearing() const { 1338 // If this method has already thrown a range-check, 1339 // assume it was because we already tried range smearing 1340 // and it failed. 1341 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1342 return !already_trapped; 1343 } 1344 1345 1346 //------------------------------flatten_alias_type----------------------------- 1347 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1348 int offset = tj->offset(); 1349 TypePtr::PTR ptr = tj->ptr(); 1350 1351 // Known instance (scalarizable allocation) alias only with itself. 1352 bool is_known_inst = tj->isa_oopptr() != NULL && 1353 tj->is_oopptr()->is_known_instance(); 1354 1355 // Process weird unsafe references. 1356 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1357 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1358 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1359 tj = TypeOopPtr::BOTTOM; 1360 ptr = tj->ptr(); 1361 offset = tj->offset(); 1362 } 1363 1364 // Array pointers need some flattening 1365 const TypeAryPtr *ta = tj->isa_aryptr(); 1366 if (ta && ta->is_stable()) { 1367 // Erase stability property for alias analysis. 1368 tj = ta = ta->cast_to_stable(false); 1369 } 1370 if( ta && is_known_inst ) { 1371 if ( offset != Type::OffsetBot && 1372 offset > arrayOopDesc::length_offset_in_bytes() ) { 1373 offset = Type::OffsetBot; // Flatten constant access into array body only 1374 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1375 } 1376 } else if( ta && _AliasLevel >= 2 ) { 1377 // For arrays indexed by constant indices, we flatten the alias 1378 // space to include all of the array body. Only the header, klass 1379 // and array length can be accessed un-aliased. 1380 if( offset != Type::OffsetBot ) { 1381 if( ta->const_oop() ) { // MethodData* or Method* 1382 offset = Type::OffsetBot; // Flatten constant access into array body 1383 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1384 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1385 // range is OK as-is. 1386 tj = ta = TypeAryPtr::RANGE; 1387 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1388 tj = TypeInstPtr::KLASS; // all klass loads look alike 1389 ta = TypeAryPtr::RANGE; // generic ignored junk 1390 ptr = TypePtr::BotPTR; 1391 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1392 tj = TypeInstPtr::MARK; 1393 ta = TypeAryPtr::RANGE; // generic ignored junk 1394 ptr = TypePtr::BotPTR; 1395 } else { // Random constant offset into array body 1396 offset = Type::OffsetBot; // Flatten constant access into array body 1397 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1398 } 1399 } 1400 // Arrays of fixed size alias with arrays of unknown size. 1401 if (ta->size() != TypeInt::POS) { 1402 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1403 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1404 } 1405 // Arrays of known objects become arrays of unknown objects. 1406 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1407 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1408 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1409 } 1410 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1411 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1412 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1413 } 1414 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1415 // cannot be distinguished by bytecode alone. 1416 if (ta->elem() == TypeInt::BOOL) { 1417 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1418 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1419 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1420 } 1421 // During the 2nd round of IterGVN, NotNull castings are removed. 1422 // Make sure the Bottom and NotNull variants alias the same. 1423 // Also, make sure exact and non-exact variants alias the same. 1424 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1425 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1426 } 1427 } 1428 1429 // Oop pointers need some flattening 1430 const TypeInstPtr *to = tj->isa_instptr(); 1431 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1432 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1433 if( ptr == TypePtr::Constant ) { 1434 if (to->klass() != ciEnv::current()->Class_klass() || 1435 offset < k->size_helper() * wordSize) { 1436 // No constant oop pointers (such as Strings); they alias with 1437 // unknown strings. 1438 assert(!is_known_inst, "not scalarizable allocation"); 1439 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1440 } 1441 } else if( is_known_inst ) { 1442 tj = to; // Keep NotNull and klass_is_exact for instance type 1443 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1444 // During the 2nd round of IterGVN, NotNull castings are removed. 1445 // Make sure the Bottom and NotNull variants alias the same. 1446 // Also, make sure exact and non-exact variants alias the same. 1447 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1448 } 1449 if (to->speculative() != NULL) { 1450 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1451 } 1452 // Canonicalize the holder of this field 1453 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1454 // First handle header references such as a LoadKlassNode, even if the 1455 // object's klass is unloaded at compile time (4965979). 1456 if (!is_known_inst) { // Do it only for non-instance types 1457 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1458 } 1459 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1460 // Static fields are in the space above the normal instance 1461 // fields in the java.lang.Class instance. 1462 if (to->klass() != ciEnv::current()->Class_klass()) { 1463 to = NULL; 1464 tj = TypeOopPtr::BOTTOM; 1465 offset = tj->offset(); 1466 } 1467 } else { 1468 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1469 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1470 if( is_known_inst ) { 1471 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1472 } else { 1473 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1474 } 1475 } 1476 } 1477 } 1478 1479 // Klass pointers to object array klasses need some flattening 1480 const TypeKlassPtr *tk = tj->isa_klassptr(); 1481 if( tk ) { 1482 // If we are referencing a field within a Klass, we need 1483 // to assume the worst case of an Object. Both exact and 1484 // inexact types must flatten to the same alias class so 1485 // use NotNull as the PTR. 1486 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1487 1488 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1489 TypeKlassPtr::OBJECT->klass(), 1490 offset); 1491 } 1492 1493 ciKlass* klass = tk->klass(); 1494 if( klass->is_obj_array_klass() ) { 1495 ciKlass* k = TypeAryPtr::OOPS->klass(); 1496 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1497 k = TypeInstPtr::BOTTOM->klass(); 1498 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1499 } 1500 1501 // Check for precise loads from the primary supertype array and force them 1502 // to the supertype cache alias index. Check for generic array loads from 1503 // the primary supertype array and also force them to the supertype cache 1504 // alias index. Since the same load can reach both, we need to merge 1505 // these 2 disparate memories into the same alias class. Since the 1506 // primary supertype array is read-only, there's no chance of confusion 1507 // where we bypass an array load and an array store. 1508 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1509 if (offset == Type::OffsetBot || 1510 (offset >= primary_supers_offset && 1511 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1512 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1513 offset = in_bytes(Klass::secondary_super_cache_offset()); 1514 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1515 } 1516 } 1517 1518 // Flatten all Raw pointers together. 1519 if (tj->base() == Type::RawPtr) 1520 tj = TypeRawPtr::BOTTOM; 1521 1522 if (tj->base() == Type::AnyPtr) 1523 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1524 1525 // Flatten all to bottom for now 1526 switch( _AliasLevel ) { 1527 case 0: 1528 tj = TypePtr::BOTTOM; 1529 break; 1530 case 1: // Flatten to: oop, static, field or array 1531 switch (tj->base()) { 1532 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1533 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1534 case Type::AryPtr: // do not distinguish arrays at all 1535 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1536 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1537 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1538 default: ShouldNotReachHere(); 1539 } 1540 break; 1541 case 2: // No collapsing at level 2; keep all splits 1542 case 3: // No collapsing at level 3; keep all splits 1543 break; 1544 default: 1545 Unimplemented(); 1546 } 1547 1548 offset = tj->offset(); 1549 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1550 1551 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1552 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1553 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1554 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1555 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1556 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1557 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1558 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1559 assert( tj->ptr() != TypePtr::TopPTR && 1560 tj->ptr() != TypePtr::AnyNull && 1561 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1562 // assert( tj->ptr() != TypePtr::Constant || 1563 // tj->base() == Type::RawPtr || 1564 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1565 1566 return tj; 1567 } 1568 1569 void Compile::AliasType::Init(int i, const TypePtr* at) { 1570 _index = i; 1571 _adr_type = at; 1572 _field = NULL; 1573 _element = NULL; 1574 _is_rewritable = true; // default 1575 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1576 if (atoop != NULL && atoop->is_known_instance()) { 1577 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1578 _general_index = Compile::current()->get_alias_index(gt); 1579 } else { 1580 _general_index = 0; 1581 } 1582 } 1583 1584 //---------------------------------print_on------------------------------------ 1585 #ifndef PRODUCT 1586 void Compile::AliasType::print_on(outputStream* st) { 1587 if (index() < 10) 1588 st->print("@ <%d> ", index()); 1589 else st->print("@ <%d>", index()); 1590 st->print(is_rewritable() ? " " : " RO"); 1591 int offset = adr_type()->offset(); 1592 if (offset == Type::OffsetBot) 1593 st->print(" +any"); 1594 else st->print(" +%-3d", offset); 1595 st->print(" in "); 1596 adr_type()->dump_on(st); 1597 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1598 if (field() != NULL && tjp) { 1599 if (tjp->klass() != field()->holder() || 1600 tjp->offset() != field()->offset_in_bytes()) { 1601 st->print(" != "); 1602 field()->print(); 1603 st->print(" ***"); 1604 } 1605 } 1606 } 1607 1608 void print_alias_types() { 1609 Compile* C = Compile::current(); 1610 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1611 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1612 C->alias_type(idx)->print_on(tty); 1613 tty->cr(); 1614 } 1615 } 1616 #endif 1617 1618 1619 //----------------------------probe_alias_cache-------------------------------- 1620 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1621 intptr_t key = (intptr_t) adr_type; 1622 key ^= key >> logAliasCacheSize; 1623 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1624 } 1625 1626 1627 //-----------------------------grow_alias_types-------------------------------- 1628 void Compile::grow_alias_types() { 1629 const int old_ats = _max_alias_types; // how many before? 1630 const int new_ats = old_ats; // how many more? 1631 const int grow_ats = old_ats+new_ats; // how many now? 1632 _max_alias_types = grow_ats; 1633 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1634 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1635 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1636 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1637 } 1638 1639 1640 //--------------------------------find_alias_type------------------------------ 1641 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1642 if (_AliasLevel == 0) 1643 return alias_type(AliasIdxBot); 1644 1645 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1646 if (ace->_adr_type == adr_type) { 1647 return alias_type(ace->_index); 1648 } 1649 1650 // Handle special cases. 1651 if (adr_type == NULL) return alias_type(AliasIdxTop); 1652 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1653 1654 // Do it the slow way. 1655 const TypePtr* flat = flatten_alias_type(adr_type); 1656 1657 #ifdef ASSERT 1658 assert(flat == flatten_alias_type(flat), "idempotent"); 1659 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1660 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1661 const TypeOopPtr* foop = flat->is_oopptr(); 1662 // Scalarizable allocations have exact klass always. 1663 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1664 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1665 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1666 } 1667 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1668 #endif 1669 1670 int idx = AliasIdxTop; 1671 for (int i = 0; i < num_alias_types(); i++) { 1672 if (alias_type(i)->adr_type() == flat) { 1673 idx = i; 1674 break; 1675 } 1676 } 1677 1678 if (idx == AliasIdxTop) { 1679 if (no_create) return NULL; 1680 // Grow the array if necessary. 1681 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1682 // Add a new alias type. 1683 idx = _num_alias_types++; 1684 _alias_types[idx]->Init(idx, flat); 1685 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1686 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1687 if (flat->isa_instptr()) { 1688 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1689 && flat->is_instptr()->klass() == env()->Class_klass()) 1690 alias_type(idx)->set_rewritable(false); 1691 } 1692 if (flat->isa_aryptr()) { 1693 #ifdef ASSERT 1694 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1695 // (T_BYTE has the weakest alignment and size restrictions...) 1696 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1697 #endif 1698 if (flat->offset() == TypePtr::OffsetBot) { 1699 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1700 } 1701 } 1702 if (flat->isa_klassptr()) { 1703 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1704 alias_type(idx)->set_rewritable(false); 1705 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1706 alias_type(idx)->set_rewritable(false); 1707 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1708 alias_type(idx)->set_rewritable(false); 1709 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1710 alias_type(idx)->set_rewritable(false); 1711 } 1712 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1713 // but the base pointer type is not distinctive enough to identify 1714 // references into JavaThread.) 1715 1716 // Check for final fields. 1717 const TypeInstPtr* tinst = flat->isa_instptr(); 1718 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1719 ciField* field; 1720 if (tinst->const_oop() != NULL && 1721 tinst->klass() == ciEnv::current()->Class_klass() && 1722 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1723 // static field 1724 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1725 field = k->get_field_by_offset(tinst->offset(), true); 1726 } else { 1727 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1728 field = k->get_field_by_offset(tinst->offset(), false); 1729 } 1730 assert(field == NULL || 1731 original_field == NULL || 1732 (field->holder() == original_field->holder() && 1733 field->offset() == original_field->offset() && 1734 field->is_static() == original_field->is_static()), "wrong field?"); 1735 // Set field() and is_rewritable() attributes. 1736 if (field != NULL) alias_type(idx)->set_field(field); 1737 } 1738 } 1739 1740 // Fill the cache for next time. 1741 ace->_adr_type = adr_type; 1742 ace->_index = idx; 1743 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1744 1745 // Might as well try to fill the cache for the flattened version, too. 1746 AliasCacheEntry* face = probe_alias_cache(flat); 1747 if (face->_adr_type == NULL) { 1748 face->_adr_type = flat; 1749 face->_index = idx; 1750 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1751 } 1752 1753 return alias_type(idx); 1754 } 1755 1756 1757 Compile::AliasType* Compile::alias_type(ciField* field) { 1758 const TypeOopPtr* t; 1759 if (field->is_static()) 1760 t = TypeInstPtr::make(field->holder()->java_mirror()); 1761 else 1762 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1763 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1764 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1765 return atp; 1766 } 1767 1768 1769 //------------------------------have_alias_type-------------------------------- 1770 bool Compile::have_alias_type(const TypePtr* adr_type) { 1771 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1772 if (ace->_adr_type == adr_type) { 1773 return true; 1774 } 1775 1776 // Handle special cases. 1777 if (adr_type == NULL) return true; 1778 if (adr_type == TypePtr::BOTTOM) return true; 1779 1780 return find_alias_type(adr_type, true, NULL) != NULL; 1781 } 1782 1783 //-----------------------------must_alias-------------------------------------- 1784 // True if all values of the given address type are in the given alias category. 1785 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1786 if (alias_idx == AliasIdxBot) return true; // the universal category 1787 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1788 if (alias_idx == AliasIdxTop) return false; // the empty category 1789 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1790 1791 // the only remaining possible overlap is identity 1792 int adr_idx = get_alias_index(adr_type); 1793 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1794 assert(adr_idx == alias_idx || 1795 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1796 && adr_type != TypeOopPtr::BOTTOM), 1797 "should not be testing for overlap with an unsafe pointer"); 1798 return adr_idx == alias_idx; 1799 } 1800 1801 //------------------------------can_alias-------------------------------------- 1802 // True if any values of the given address type are in the given alias category. 1803 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1804 if (alias_idx == AliasIdxTop) return false; // the empty category 1805 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1806 if (alias_idx == AliasIdxBot) return true; // the universal category 1807 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1808 1809 // the only remaining possible overlap is identity 1810 int adr_idx = get_alias_index(adr_type); 1811 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1812 return adr_idx == alias_idx; 1813 } 1814 1815 1816 1817 //---------------------------pop_warm_call------------------------------------- 1818 WarmCallInfo* Compile::pop_warm_call() { 1819 WarmCallInfo* wci = _warm_calls; 1820 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1821 return wci; 1822 } 1823 1824 //----------------------------Inline_Warm-------------------------------------- 1825 int Compile::Inline_Warm() { 1826 // If there is room, try to inline some more warm call sites. 1827 // %%% Do a graph index compaction pass when we think we're out of space? 1828 if (!InlineWarmCalls) return 0; 1829 1830 int calls_made_hot = 0; 1831 int room_to_grow = NodeCountInliningCutoff - unique(); 1832 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1833 int amount_grown = 0; 1834 WarmCallInfo* call; 1835 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1836 int est_size = (int)call->size(); 1837 if (est_size > (room_to_grow - amount_grown)) { 1838 // This one won't fit anyway. Get rid of it. 1839 call->make_cold(); 1840 continue; 1841 } 1842 call->make_hot(); 1843 calls_made_hot++; 1844 amount_grown += est_size; 1845 amount_to_grow -= est_size; 1846 } 1847 1848 if (calls_made_hot > 0) set_major_progress(); 1849 return calls_made_hot; 1850 } 1851 1852 1853 //----------------------------Finish_Warm-------------------------------------- 1854 void Compile::Finish_Warm() { 1855 if (!InlineWarmCalls) return; 1856 if (failing()) return; 1857 if (warm_calls() == NULL) return; 1858 1859 // Clean up loose ends, if we are out of space for inlining. 1860 WarmCallInfo* call; 1861 while ((call = pop_warm_call()) != NULL) { 1862 call->make_cold(); 1863 } 1864 } 1865 1866 //---------------------cleanup_loop_predicates----------------------- 1867 // Remove the opaque nodes that protect the predicates so that all unused 1868 // checks and uncommon_traps will be eliminated from the ideal graph 1869 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1870 if (predicate_count()==0) return; 1871 for (int i = predicate_count(); i > 0; i--) { 1872 Node * n = predicate_opaque1_node(i-1); 1873 assert(n->Opcode() == Op_Opaque1, "must be"); 1874 igvn.replace_node(n, n->in(1)); 1875 } 1876 assert(predicate_count()==0, "should be clean!"); 1877 } 1878 1879 // StringOpts and late inlining of string methods 1880 void Compile::inline_string_calls(bool parse_time) { 1881 { 1882 // remove useless nodes to make the usage analysis simpler 1883 ResourceMark rm; 1884 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1885 } 1886 1887 { 1888 ResourceMark rm; 1889 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1890 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1891 print_method(PHASE_AFTER_STRINGOPTS, 3); 1892 } 1893 1894 // now inline anything that we skipped the first time around 1895 if (!parse_time) { 1896 _late_inlines_pos = _late_inlines.length(); 1897 } 1898 1899 while (_string_late_inlines.length() > 0) { 1900 CallGenerator* cg = _string_late_inlines.pop(); 1901 cg->do_late_inline(); 1902 if (failing()) return; 1903 } 1904 _string_late_inlines.trunc_to(0); 1905 } 1906 1907 // Late inlining of boxing methods 1908 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1909 if (_boxing_late_inlines.length() > 0) { 1910 assert(has_boxed_value(), "inconsistent"); 1911 1912 PhaseGVN* gvn = initial_gvn(); 1913 set_inlining_incrementally(true); 1914 1915 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1916 for_igvn()->clear(); 1917 gvn->replace_with(&igvn); 1918 1919 _late_inlines_pos = _late_inlines.length(); 1920 1921 while (_boxing_late_inlines.length() > 0) { 1922 CallGenerator* cg = _boxing_late_inlines.pop(); 1923 cg->do_late_inline(); 1924 if (failing()) return; 1925 } 1926 _boxing_late_inlines.trunc_to(0); 1927 1928 { 1929 ResourceMark rm; 1930 PhaseRemoveUseless pru(gvn, for_igvn()); 1931 } 1932 1933 igvn = PhaseIterGVN(gvn); 1934 igvn.optimize(); 1935 1936 set_inlining_progress(false); 1937 set_inlining_incrementally(false); 1938 } 1939 } 1940 1941 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1942 assert(IncrementalInline, "incremental inlining should be on"); 1943 PhaseGVN* gvn = initial_gvn(); 1944 1945 set_inlining_progress(false); 1946 for_igvn()->clear(); 1947 gvn->replace_with(&igvn); 1948 1949 int i = 0; 1950 1951 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1952 CallGenerator* cg = _late_inlines.at(i); 1953 _late_inlines_pos = i+1; 1954 cg->do_late_inline(); 1955 if (failing()) return; 1956 } 1957 int j = 0; 1958 for (; i < _late_inlines.length(); i++, j++) { 1959 _late_inlines.at_put(j, _late_inlines.at(i)); 1960 } 1961 _late_inlines.trunc_to(j); 1962 1963 { 1964 ResourceMark rm; 1965 PhaseRemoveUseless pru(gvn, for_igvn()); 1966 } 1967 1968 igvn = PhaseIterGVN(gvn); 1969 } 1970 1971 // Perform incremental inlining until bound on number of live nodes is reached 1972 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1973 PhaseGVN* gvn = initial_gvn(); 1974 1975 set_inlining_incrementally(true); 1976 set_inlining_progress(true); 1977 uint low_live_nodes = 0; 1978 1979 while(inlining_progress() && _late_inlines.length() > 0) { 1980 1981 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1982 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1983 // PhaseIdealLoop is expensive so we only try it once we are 1984 // out of live nodes and we only try it again if the previous 1985 // helped got the number of nodes down significantly 1986 PhaseIdealLoop ideal_loop( igvn, false, true ); 1987 if (failing()) return; 1988 low_live_nodes = live_nodes(); 1989 _major_progress = true; 1990 } 1991 1992 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1993 break; 1994 } 1995 } 1996 1997 inline_incrementally_one(igvn); 1998 1999 if (failing()) return; 2000 2001 igvn.optimize(); 2002 2003 if (failing()) return; 2004 } 2005 2006 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2007 2008 if (_string_late_inlines.length() > 0) { 2009 assert(has_stringbuilder(), "inconsistent"); 2010 for_igvn()->clear(); 2011 initial_gvn()->replace_with(&igvn); 2012 2013 inline_string_calls(false); 2014 2015 if (failing()) return; 2016 2017 { 2018 ResourceMark rm; 2019 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2020 } 2021 2022 igvn = PhaseIterGVN(gvn); 2023 2024 igvn.optimize(); 2025 } 2026 2027 set_inlining_incrementally(false); 2028 } 2029 2030 2031 //------------------------------Optimize--------------------------------------- 2032 // Given a graph, optimize it. 2033 void Compile::Optimize() { 2034 TracePhase t1("optimizer", &_t_optimizer, true); 2035 2036 #ifndef PRODUCT 2037 if (env()->break_at_compile()) { 2038 BREAKPOINT; 2039 } 2040 2041 #endif 2042 2043 ResourceMark rm; 2044 int loop_opts_cnt; 2045 2046 NOT_PRODUCT( verify_graph_edges(); ) 2047 2048 print_method(PHASE_AFTER_PARSING); 2049 2050 { 2051 // Iterative Global Value Numbering, including ideal transforms 2052 // Initialize IterGVN with types and values from parse-time GVN 2053 PhaseIterGVN igvn(initial_gvn()); 2054 { 2055 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2056 igvn.optimize(); 2057 } 2058 2059 print_method(PHASE_ITER_GVN1, 2); 2060 2061 if (failing()) return; 2062 2063 { 2064 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2065 inline_incrementally(igvn); 2066 } 2067 2068 print_method(PHASE_INCREMENTAL_INLINE, 2); 2069 2070 if (failing()) return; 2071 2072 if (eliminate_boxing()) { 2073 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2074 // Inline valueOf() methods now. 2075 inline_boxing_calls(igvn); 2076 2077 if (AlwaysIncrementalInline) { 2078 inline_incrementally(igvn); 2079 } 2080 2081 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2082 2083 if (failing()) return; 2084 } 2085 2086 // Remove the speculative part of types and clean up the graph from 2087 // the extra CastPP nodes whose only purpose is to carry them. Do 2088 // that early so that optimizations are not disrupted by the extra 2089 // CastPP nodes. 2090 remove_speculative_types(igvn); 2091 2092 // No more new expensive nodes will be added to the list from here 2093 // so keep only the actual candidates for optimizations. 2094 cleanup_expensive_nodes(igvn); 2095 2096 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2097 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);) 2098 initial_gvn()->replace_with(&igvn); 2099 for_igvn()->clear(); 2100 Unique_Node_List new_worklist(C->comp_arena()); 2101 { 2102 ResourceMark rm; 2103 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2104 } 2105 set_for_igvn(&new_worklist); 2106 igvn = PhaseIterGVN(initial_gvn()); 2107 igvn.optimize(); 2108 } 2109 2110 // Perform escape analysis 2111 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2112 if (has_loops()) { 2113 // Cleanup graph (remove dead nodes). 2114 TracePhase t2("idealLoop", &_t_idealLoop, true); 2115 PhaseIdealLoop ideal_loop( igvn, false, true ); 2116 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2117 if (failing()) return; 2118 } 2119 ConnectionGraph::do_analysis(this, &igvn); 2120 2121 if (failing()) return; 2122 2123 // Optimize out fields loads from scalar replaceable allocations. 2124 igvn.optimize(); 2125 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2126 2127 if (failing()) return; 2128 2129 if (congraph() != NULL && macro_count() > 0) { 2130 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2131 PhaseMacroExpand mexp(igvn); 2132 mexp.eliminate_macro_nodes(); 2133 igvn.set_delay_transform(false); 2134 2135 igvn.optimize(); 2136 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2137 2138 if (failing()) return; 2139 } 2140 } 2141 2142 // Loop transforms on the ideal graph. Range Check Elimination, 2143 // peeling, unrolling, etc. 2144 2145 // Set loop opts counter 2146 loop_opts_cnt = num_loop_opts(); 2147 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2148 { 2149 TracePhase t2("idealLoop", &_t_idealLoop, true); 2150 PhaseIdealLoop ideal_loop( igvn, true ); 2151 loop_opts_cnt--; 2152 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2153 if (failing()) return; 2154 } 2155 // Loop opts pass if partial peeling occurred in previous pass 2156 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2157 TracePhase t3("idealLoop", &_t_idealLoop, true); 2158 PhaseIdealLoop ideal_loop( igvn, false ); 2159 loop_opts_cnt--; 2160 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2161 if (failing()) return; 2162 } 2163 // Loop opts pass for loop-unrolling before CCP 2164 if(major_progress() && (loop_opts_cnt > 0)) { 2165 TracePhase t4("idealLoop", &_t_idealLoop, true); 2166 PhaseIdealLoop ideal_loop( igvn, false ); 2167 loop_opts_cnt--; 2168 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2169 } 2170 if (!failing()) { 2171 // Verify that last round of loop opts produced a valid graph 2172 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2173 PhaseIdealLoop::verify(igvn); 2174 } 2175 } 2176 if (failing()) return; 2177 2178 // Conditional Constant Propagation; 2179 PhaseCCP ccp( &igvn ); 2180 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2181 { 2182 TracePhase t2("ccp", &_t_ccp, true); 2183 ccp.do_transform(); 2184 } 2185 print_method(PHASE_CPP1, 2); 2186 2187 assert( true, "Break here to ccp.dump_old2new_map()"); 2188 2189 // Iterative Global Value Numbering, including ideal transforms 2190 { 2191 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2192 igvn = ccp; 2193 igvn.optimize(); 2194 } 2195 2196 print_method(PHASE_ITER_GVN2, 2); 2197 2198 if (failing()) return; 2199 2200 // Loop transforms on the ideal graph. Range Check Elimination, 2201 // peeling, unrolling, etc. 2202 if(loop_opts_cnt > 0) { 2203 debug_only( int cnt = 0; ); 2204 while(major_progress() && (loop_opts_cnt > 0)) { 2205 TracePhase t2("idealLoop", &_t_idealLoop, true); 2206 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2207 PhaseIdealLoop ideal_loop( igvn, true); 2208 loop_opts_cnt--; 2209 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2210 if (failing()) return; 2211 } 2212 } 2213 2214 { 2215 // Verify that all previous optimizations produced a valid graph 2216 // at least to this point, even if no loop optimizations were done. 2217 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2218 PhaseIdealLoop::verify(igvn); 2219 } 2220 2221 { 2222 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2223 PhaseMacroExpand mex(igvn); 2224 if (mex.expand_macro_nodes()) { 2225 assert(failing(), "must bail out w/ explicit message"); 2226 return; 2227 } 2228 } 2229 2230 } // (End scope of igvn; run destructor if necessary for asserts.) 2231 2232 dump_inlining(); 2233 // A method with only infinite loops has no edges entering loops from root 2234 { 2235 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2236 if (final_graph_reshaping()) { 2237 assert(failing(), "must bail out w/ explicit message"); 2238 return; 2239 } 2240 } 2241 2242 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2243 } 2244 2245 2246 //------------------------------Code_Gen--------------------------------------- 2247 // Given a graph, generate code for it 2248 void Compile::Code_Gen() { 2249 if (failing()) { 2250 return; 2251 } 2252 2253 // Perform instruction selection. You might think we could reclaim Matcher 2254 // memory PDQ, but actually the Matcher is used in generating spill code. 2255 // Internals of the Matcher (including some VectorSets) must remain live 2256 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2257 // set a bit in reclaimed memory. 2258 2259 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2260 // nodes. Mapping is only valid at the root of each matched subtree. 2261 NOT_PRODUCT( verify_graph_edges(); ) 2262 2263 Matcher matcher; 2264 _matcher = &matcher; 2265 { 2266 TracePhase t2("matcher", &_t_matcher, true); 2267 matcher.match(); 2268 } 2269 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2270 // nodes. Mapping is only valid at the root of each matched subtree. 2271 NOT_PRODUCT( verify_graph_edges(); ) 2272 2273 // If you have too many nodes, or if matching has failed, bail out 2274 check_node_count(0, "out of nodes matching instructions"); 2275 if (failing()) { 2276 return; 2277 } 2278 2279 // Build a proper-looking CFG 2280 PhaseCFG cfg(node_arena(), root(), matcher); 2281 _cfg = &cfg; 2282 { 2283 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2284 bool success = cfg.do_global_code_motion(); 2285 if (!success) { 2286 return; 2287 } 2288 2289 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2290 NOT_PRODUCT( verify_graph_edges(); ) 2291 debug_only( cfg.verify(); ) 2292 } 2293 2294 PhaseChaitin regalloc(unique(), cfg, matcher); 2295 _regalloc = ®alloc; 2296 { 2297 TracePhase t2("regalloc", &_t_registerAllocation, true); 2298 // Perform register allocation. After Chaitin, use-def chains are 2299 // no longer accurate (at spill code) and so must be ignored. 2300 // Node->LRG->reg mappings are still accurate. 2301 _regalloc->Register_Allocate(); 2302 2303 // Bail out if the allocator builds too many nodes 2304 if (failing()) { 2305 return; 2306 } 2307 } 2308 2309 // Prior to register allocation we kept empty basic blocks in case the 2310 // the allocator needed a place to spill. After register allocation we 2311 // are not adding any new instructions. If any basic block is empty, we 2312 // can now safely remove it. 2313 { 2314 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2315 cfg.remove_empty_blocks(); 2316 if (do_freq_based_layout()) { 2317 PhaseBlockLayout layout(cfg); 2318 } else { 2319 cfg.set_loop_alignment(); 2320 } 2321 cfg.fixup_flow(); 2322 } 2323 2324 // Apply peephole optimizations 2325 if( OptoPeephole ) { 2326 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2327 PhasePeephole peep( _regalloc, cfg); 2328 peep.do_transform(); 2329 } 2330 2331 // Do late expand if CPU requires this. 2332 if (Matcher::require_postalloc_expand) { 2333 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2334 cfg.postalloc_expand(_regalloc); 2335 } 2336 2337 // Convert Nodes to instruction bits in a buffer 2338 { 2339 // %%%% workspace merge brought two timers together for one job 2340 TracePhase t2a("output", &_t_output, true); 2341 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2342 Output(); 2343 } 2344 2345 print_method(PHASE_FINAL_CODE); 2346 2347 // He's dead, Jim. 2348 _cfg = (PhaseCFG*)0xdeadbeef; 2349 _regalloc = (PhaseChaitin*)0xdeadbeef; 2350 } 2351 2352 2353 //------------------------------dump_asm--------------------------------------- 2354 // Dump formatted assembly 2355 #ifndef PRODUCT 2356 void Compile::dump_asm(int *pcs, uint pc_limit) { 2357 bool cut_short = false; 2358 tty->print_cr("#"); 2359 tty->print("# "); _tf->dump(); tty->cr(); 2360 tty->print_cr("#"); 2361 2362 // For all blocks 2363 int pc = 0x0; // Program counter 2364 char starts_bundle = ' '; 2365 _regalloc->dump_frame(); 2366 2367 Node *n = NULL; 2368 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2369 if (VMThread::should_terminate()) { 2370 cut_short = true; 2371 break; 2372 } 2373 Block* block = _cfg->get_block(i); 2374 if (block->is_connector() && !Verbose) { 2375 continue; 2376 } 2377 n = block->head(); 2378 if (pcs && n->_idx < pc_limit) { 2379 tty->print("%3.3x ", pcs[n->_idx]); 2380 } else { 2381 tty->print(" "); 2382 } 2383 block->dump_head(_cfg); 2384 if (block->is_connector()) { 2385 tty->print_cr(" # Empty connector block"); 2386 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2387 tty->print_cr(" # Block is sole successor of call"); 2388 } 2389 2390 // For all instructions 2391 Node *delay = NULL; 2392 for (uint j = 0; j < block->number_of_nodes(); j++) { 2393 if (VMThread::should_terminate()) { 2394 cut_short = true; 2395 break; 2396 } 2397 n = block->get_node(j); 2398 if (valid_bundle_info(n)) { 2399 Bundle* bundle = node_bundling(n); 2400 if (bundle->used_in_unconditional_delay()) { 2401 delay = n; 2402 continue; 2403 } 2404 if (bundle->starts_bundle()) { 2405 starts_bundle = '+'; 2406 } 2407 } 2408 2409 if (WizardMode) { 2410 n->dump(); 2411 } 2412 2413 if( !n->is_Region() && // Dont print in the Assembly 2414 !n->is_Phi() && // a few noisely useless nodes 2415 !n->is_Proj() && 2416 !n->is_MachTemp() && 2417 !n->is_SafePointScalarObject() && 2418 !n->is_Catch() && // Would be nice to print exception table targets 2419 !n->is_MergeMem() && // Not very interesting 2420 !n->is_top() && // Debug info table constants 2421 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2422 ) { 2423 if (pcs && n->_idx < pc_limit) 2424 tty->print("%3.3x", pcs[n->_idx]); 2425 else 2426 tty->print(" "); 2427 tty->print(" %c ", starts_bundle); 2428 starts_bundle = ' '; 2429 tty->print("\t"); 2430 n->format(_regalloc, tty); 2431 tty->cr(); 2432 } 2433 2434 // If we have an instruction with a delay slot, and have seen a delay, 2435 // then back up and print it 2436 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2437 assert(delay != NULL, "no unconditional delay instruction"); 2438 if (WizardMode) delay->dump(); 2439 2440 if (node_bundling(delay)->starts_bundle()) 2441 starts_bundle = '+'; 2442 if (pcs && n->_idx < pc_limit) 2443 tty->print("%3.3x", pcs[n->_idx]); 2444 else 2445 tty->print(" "); 2446 tty->print(" %c ", starts_bundle); 2447 starts_bundle = ' '; 2448 tty->print("\t"); 2449 delay->format(_regalloc, tty); 2450 tty->cr(); 2451 delay = NULL; 2452 } 2453 2454 // Dump the exception table as well 2455 if( n->is_Catch() && (Verbose || WizardMode) ) { 2456 // Print the exception table for this offset 2457 _handler_table.print_subtable_for(pc); 2458 } 2459 } 2460 2461 if (pcs && n->_idx < pc_limit) 2462 tty->print_cr("%3.3x", pcs[n->_idx]); 2463 else 2464 tty->cr(); 2465 2466 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2467 2468 } // End of per-block dump 2469 tty->cr(); 2470 2471 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2472 } 2473 #endif 2474 2475 //------------------------------Final_Reshape_Counts--------------------------- 2476 // This class defines counters to help identify when a method 2477 // may/must be executed using hardware with only 24-bit precision. 2478 struct Final_Reshape_Counts : public StackObj { 2479 int _call_count; // count non-inlined 'common' calls 2480 int _float_count; // count float ops requiring 24-bit precision 2481 int _double_count; // count double ops requiring more precision 2482 int _java_call_count; // count non-inlined 'java' calls 2483 int _inner_loop_count; // count loops which need alignment 2484 VectorSet _visited; // Visitation flags 2485 Node_List _tests; // Set of IfNodes & PCTableNodes 2486 2487 Final_Reshape_Counts() : 2488 _call_count(0), _float_count(0), _double_count(0), 2489 _java_call_count(0), _inner_loop_count(0), 2490 _visited( Thread::current()->resource_area() ) { } 2491 2492 void inc_call_count () { _call_count ++; } 2493 void inc_float_count () { _float_count ++; } 2494 void inc_double_count() { _double_count++; } 2495 void inc_java_call_count() { _java_call_count++; } 2496 void inc_inner_loop_count() { _inner_loop_count++; } 2497 2498 int get_call_count () const { return _call_count ; } 2499 int get_float_count () const { return _float_count ; } 2500 int get_double_count() const { return _double_count; } 2501 int get_java_call_count() const { return _java_call_count; } 2502 int get_inner_loop_count() const { return _inner_loop_count; } 2503 }; 2504 2505 #ifdef ASSERT 2506 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2507 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2508 // Make sure the offset goes inside the instance layout. 2509 return k->contains_field_offset(tp->offset()); 2510 // Note that OffsetBot and OffsetTop are very negative. 2511 } 2512 #endif 2513 2514 // Eliminate trivially redundant StoreCMs and accumulate their 2515 // precedence edges. 2516 void Compile::eliminate_redundant_card_marks(Node* n) { 2517 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2518 if (n->in(MemNode::Address)->outcnt() > 1) { 2519 // There are multiple users of the same address so it might be 2520 // possible to eliminate some of the StoreCMs 2521 Node* mem = n->in(MemNode::Memory); 2522 Node* adr = n->in(MemNode::Address); 2523 Node* val = n->in(MemNode::ValueIn); 2524 Node* prev = n; 2525 bool done = false; 2526 // Walk the chain of StoreCMs eliminating ones that match. As 2527 // long as it's a chain of single users then the optimization is 2528 // safe. Eliminating partially redundant StoreCMs would require 2529 // cloning copies down the other paths. 2530 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2531 if (adr == mem->in(MemNode::Address) && 2532 val == mem->in(MemNode::ValueIn)) { 2533 // redundant StoreCM 2534 if (mem->req() > MemNode::OopStore) { 2535 // Hasn't been processed by this code yet. 2536 n->add_prec(mem->in(MemNode::OopStore)); 2537 } else { 2538 // Already converted to precedence edge 2539 for (uint i = mem->req(); i < mem->len(); i++) { 2540 // Accumulate any precedence edges 2541 if (mem->in(i) != NULL) { 2542 n->add_prec(mem->in(i)); 2543 } 2544 } 2545 // Everything above this point has been processed. 2546 done = true; 2547 } 2548 // Eliminate the previous StoreCM 2549 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2550 assert(mem->outcnt() == 0, "should be dead"); 2551 mem->disconnect_inputs(NULL, this); 2552 } else { 2553 prev = mem; 2554 } 2555 mem = prev->in(MemNode::Memory); 2556 } 2557 } 2558 } 2559 2560 //------------------------------final_graph_reshaping_impl---------------------- 2561 // Implement items 1-5 from final_graph_reshaping below. 2562 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2563 2564 if ( n->outcnt() == 0 ) return; // dead node 2565 uint nop = n->Opcode(); 2566 2567 // Check for 2-input instruction with "last use" on right input. 2568 // Swap to left input. Implements item (2). 2569 if( n->req() == 3 && // two-input instruction 2570 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2571 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2572 n->in(2)->outcnt() == 1 &&// right use IS a last use 2573 !n->in(2)->is_Con() ) { // right use is not a constant 2574 // Check for commutative opcode 2575 switch( nop ) { 2576 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2577 case Op_MaxI: case Op_MinI: 2578 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2579 case Op_AndL: case Op_XorL: case Op_OrL: 2580 case Op_AndI: case Op_XorI: case Op_OrI: { 2581 // Move "last use" input to left by swapping inputs 2582 n->swap_edges(1, 2); 2583 break; 2584 } 2585 default: 2586 break; 2587 } 2588 } 2589 2590 #ifdef ASSERT 2591 if( n->is_Mem() ) { 2592 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2593 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2594 // oop will be recorded in oop map if load crosses safepoint 2595 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2596 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2597 "raw memory operations should have control edge"); 2598 } 2599 #endif 2600 // Count FPU ops and common calls, implements item (3) 2601 switch( nop ) { 2602 // Count all float operations that may use FPU 2603 case Op_AddF: 2604 case Op_SubF: 2605 case Op_MulF: 2606 case Op_DivF: 2607 case Op_NegF: 2608 case Op_ModF: 2609 case Op_ConvI2F: 2610 case Op_ConF: 2611 case Op_CmpF: 2612 case Op_CmpF3: 2613 // case Op_ConvL2F: // longs are split into 32-bit halves 2614 frc.inc_float_count(); 2615 break; 2616 2617 case Op_ConvF2D: 2618 case Op_ConvD2F: 2619 frc.inc_float_count(); 2620 frc.inc_double_count(); 2621 break; 2622 2623 // Count all double operations that may use FPU 2624 case Op_AddD: 2625 case Op_SubD: 2626 case Op_MulD: 2627 case Op_DivD: 2628 case Op_NegD: 2629 case Op_ModD: 2630 case Op_ConvI2D: 2631 case Op_ConvD2I: 2632 // case Op_ConvL2D: // handled by leaf call 2633 // case Op_ConvD2L: // handled by leaf call 2634 case Op_ConD: 2635 case Op_CmpD: 2636 case Op_CmpD3: 2637 frc.inc_double_count(); 2638 break; 2639 case Op_Opaque1: // Remove Opaque Nodes before matching 2640 case Op_Opaque2: // Remove Opaque Nodes before matching 2641 case Op_Opaque3: 2642 n->subsume_by(n->in(1), this); 2643 break; 2644 case Op_CallStaticJava: 2645 case Op_CallJava: 2646 case Op_CallDynamicJava: 2647 frc.inc_java_call_count(); // Count java call site; 2648 case Op_CallRuntime: 2649 case Op_CallLeaf: 2650 case Op_CallLeafNoFP: { 2651 assert( n->is_Call(), "" ); 2652 CallNode *call = n->as_Call(); 2653 // Count call sites where the FP mode bit would have to be flipped. 2654 // Do not count uncommon runtime calls: 2655 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2656 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2657 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2658 frc.inc_call_count(); // Count the call site 2659 } else { // See if uncommon argument is shared 2660 Node *n = call->in(TypeFunc::Parms); 2661 int nop = n->Opcode(); 2662 // Clone shared simple arguments to uncommon calls, item (1). 2663 if( n->outcnt() > 1 && 2664 !n->is_Proj() && 2665 nop != Op_CreateEx && 2666 nop != Op_CheckCastPP && 2667 nop != Op_DecodeN && 2668 nop != Op_DecodeNKlass && 2669 !n->is_Mem() ) { 2670 Node *x = n->clone(); 2671 call->set_req( TypeFunc::Parms, x ); 2672 } 2673 } 2674 break; 2675 } 2676 2677 case Op_StoreD: 2678 case Op_LoadD: 2679 case Op_LoadD_unaligned: 2680 frc.inc_double_count(); 2681 goto handle_mem; 2682 case Op_StoreF: 2683 case Op_LoadF: 2684 frc.inc_float_count(); 2685 goto handle_mem; 2686 2687 case Op_StoreCM: 2688 { 2689 // Convert OopStore dependence into precedence edge 2690 Node* prec = n->in(MemNode::OopStore); 2691 n->del_req(MemNode::OopStore); 2692 n->add_prec(prec); 2693 eliminate_redundant_card_marks(n); 2694 } 2695 2696 // fall through 2697 2698 case Op_StoreB: 2699 case Op_StoreC: 2700 case Op_StorePConditional: 2701 case Op_StoreI: 2702 case Op_StoreL: 2703 case Op_StoreIConditional: 2704 case Op_StoreLConditional: 2705 case Op_CompareAndSwapI: 2706 case Op_CompareAndSwapL: 2707 case Op_CompareAndSwapP: 2708 case Op_CompareAndSwapN: 2709 case Op_GetAndAddI: 2710 case Op_GetAndAddL: 2711 case Op_GetAndSetI: 2712 case Op_GetAndSetL: 2713 case Op_GetAndSetP: 2714 case Op_GetAndSetN: 2715 case Op_StoreP: 2716 case Op_StoreN: 2717 case Op_StoreNKlass: 2718 case Op_LoadB: 2719 case Op_LoadUB: 2720 case Op_LoadUS: 2721 case Op_LoadI: 2722 case Op_LoadKlass: 2723 case Op_LoadNKlass: 2724 case Op_LoadL: 2725 case Op_LoadL_unaligned: 2726 case Op_LoadPLocked: 2727 case Op_LoadP: 2728 case Op_LoadN: 2729 case Op_LoadRange: 2730 case Op_LoadS: { 2731 handle_mem: 2732 #ifdef ASSERT 2733 if( VerifyOptoOopOffsets ) { 2734 assert( n->is_Mem(), "" ); 2735 MemNode *mem = (MemNode*)n; 2736 // Check to see if address types have grounded out somehow. 2737 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2738 assert( !tp || oop_offset_is_sane(tp), "" ); 2739 } 2740 #endif 2741 break; 2742 } 2743 2744 case Op_AddP: { // Assert sane base pointers 2745 Node *addp = n->in(AddPNode::Address); 2746 assert( !addp->is_AddP() || 2747 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2748 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2749 "Base pointers must match" ); 2750 #ifdef _LP64 2751 if ((UseCompressedOops || UseCompressedClassPointers) && 2752 addp->Opcode() == Op_ConP && 2753 addp == n->in(AddPNode::Base) && 2754 n->in(AddPNode::Offset)->is_Con()) { 2755 // Use addressing with narrow klass to load with offset on x86. 2756 // On sparc loading 32-bits constant and decoding it have less 2757 // instructions (4) then load 64-bits constant (7). 2758 // Do this transformation here since IGVN will convert ConN back to ConP. 2759 const Type* t = addp->bottom_type(); 2760 if (t->isa_oopptr() || t->isa_klassptr()) { 2761 Node* nn = NULL; 2762 2763 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2764 2765 // Look for existing ConN node of the same exact type. 2766 Node* r = root(); 2767 uint cnt = r->outcnt(); 2768 for (uint i = 0; i < cnt; i++) { 2769 Node* m = r->raw_out(i); 2770 if (m!= NULL && m->Opcode() == op && 2771 m->bottom_type()->make_ptr() == t) { 2772 nn = m; 2773 break; 2774 } 2775 } 2776 if (nn != NULL) { 2777 // Decode a narrow oop to match address 2778 // [R12 + narrow_oop_reg<<3 + offset] 2779 if (t->isa_oopptr()) { 2780 nn = new (this) DecodeNNode(nn, t); 2781 } else { 2782 nn = new (this) DecodeNKlassNode(nn, t); 2783 } 2784 n->set_req(AddPNode::Base, nn); 2785 n->set_req(AddPNode::Address, nn); 2786 if (addp->outcnt() == 0) { 2787 addp->disconnect_inputs(NULL, this); 2788 } 2789 } 2790 } 2791 } 2792 #endif 2793 break; 2794 } 2795 2796 #ifdef _LP64 2797 case Op_CastPP: 2798 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2799 Node* in1 = n->in(1); 2800 const Type* t = n->bottom_type(); 2801 Node* new_in1 = in1->clone(); 2802 new_in1->as_DecodeN()->set_type(t); 2803 2804 if (!Matcher::narrow_oop_use_complex_address()) { 2805 // 2806 // x86, ARM and friends can handle 2 adds in addressing mode 2807 // and Matcher can fold a DecodeN node into address by using 2808 // a narrow oop directly and do implicit NULL check in address: 2809 // 2810 // [R12 + narrow_oop_reg<<3 + offset] 2811 // NullCheck narrow_oop_reg 2812 // 2813 // On other platforms (Sparc) we have to keep new DecodeN node and 2814 // use it to do implicit NULL check in address: 2815 // 2816 // decode_not_null narrow_oop_reg, base_reg 2817 // [base_reg + offset] 2818 // NullCheck base_reg 2819 // 2820 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2821 // to keep the information to which NULL check the new DecodeN node 2822 // corresponds to use it as value in implicit_null_check(). 2823 // 2824 new_in1->set_req(0, n->in(0)); 2825 } 2826 2827 n->subsume_by(new_in1, this); 2828 if (in1->outcnt() == 0) { 2829 in1->disconnect_inputs(NULL, this); 2830 } 2831 } 2832 break; 2833 2834 case Op_CmpP: 2835 // Do this transformation here to preserve CmpPNode::sub() and 2836 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2837 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2838 Node* in1 = n->in(1); 2839 Node* in2 = n->in(2); 2840 if (!in1->is_DecodeNarrowPtr()) { 2841 in2 = in1; 2842 in1 = n->in(2); 2843 } 2844 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2845 2846 Node* new_in2 = NULL; 2847 if (in2->is_DecodeNarrowPtr()) { 2848 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2849 new_in2 = in2->in(1); 2850 } else if (in2->Opcode() == Op_ConP) { 2851 const Type* t = in2->bottom_type(); 2852 if (t == TypePtr::NULL_PTR) { 2853 assert(in1->is_DecodeN(), "compare klass to null?"); 2854 // Don't convert CmpP null check into CmpN if compressed 2855 // oops implicit null check is not generated. 2856 // This will allow to generate normal oop implicit null check. 2857 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2858 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2859 // 2860 // This transformation together with CastPP transformation above 2861 // will generated code for implicit NULL checks for compressed oops. 2862 // 2863 // The original code after Optimize() 2864 // 2865 // LoadN memory, narrow_oop_reg 2866 // decode narrow_oop_reg, base_reg 2867 // CmpP base_reg, NULL 2868 // CastPP base_reg // NotNull 2869 // Load [base_reg + offset], val_reg 2870 // 2871 // after these transformations will be 2872 // 2873 // LoadN memory, narrow_oop_reg 2874 // CmpN narrow_oop_reg, NULL 2875 // decode_not_null narrow_oop_reg, base_reg 2876 // Load [base_reg + offset], val_reg 2877 // 2878 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2879 // since narrow oops can be used in debug info now (see the code in 2880 // final_graph_reshaping_walk()). 2881 // 2882 // At the end the code will be matched to 2883 // on x86: 2884 // 2885 // Load_narrow_oop memory, narrow_oop_reg 2886 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2887 // NullCheck narrow_oop_reg 2888 // 2889 // and on sparc: 2890 // 2891 // Load_narrow_oop memory, narrow_oop_reg 2892 // decode_not_null narrow_oop_reg, base_reg 2893 // Load [base_reg + offset], val_reg 2894 // NullCheck base_reg 2895 // 2896 } else if (t->isa_oopptr()) { 2897 new_in2 = ConNode::make(this, t->make_narrowoop()); 2898 } else if (t->isa_klassptr()) { 2899 new_in2 = ConNode::make(this, t->make_narrowklass()); 2900 } 2901 } 2902 if (new_in2 != NULL) { 2903 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2904 n->subsume_by(cmpN, this); 2905 if (in1->outcnt() == 0) { 2906 in1->disconnect_inputs(NULL, this); 2907 } 2908 if (in2->outcnt() == 0) { 2909 in2->disconnect_inputs(NULL, this); 2910 } 2911 } 2912 } 2913 break; 2914 2915 case Op_DecodeN: 2916 case Op_DecodeNKlass: 2917 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2918 // DecodeN could be pinned when it can't be fold into 2919 // an address expression, see the code for Op_CastPP above. 2920 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2921 break; 2922 2923 case Op_EncodeP: 2924 case Op_EncodePKlass: { 2925 Node* in1 = n->in(1); 2926 if (in1->is_DecodeNarrowPtr()) { 2927 n->subsume_by(in1->in(1), this); 2928 } else if (in1->Opcode() == Op_ConP) { 2929 const Type* t = in1->bottom_type(); 2930 if (t == TypePtr::NULL_PTR) { 2931 assert(t->isa_oopptr(), "null klass?"); 2932 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2933 } else if (t->isa_oopptr()) { 2934 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2935 } else if (t->isa_klassptr()) { 2936 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2937 } 2938 } 2939 if (in1->outcnt() == 0) { 2940 in1->disconnect_inputs(NULL, this); 2941 } 2942 break; 2943 } 2944 2945 case Op_Proj: { 2946 if (OptimizeStringConcat) { 2947 ProjNode* p = n->as_Proj(); 2948 if (p->_is_io_use) { 2949 // Separate projections were used for the exception path which 2950 // are normally removed by a late inline. If it wasn't inlined 2951 // then they will hang around and should just be replaced with 2952 // the original one. 2953 Node* proj = NULL; 2954 // Replace with just one 2955 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2956 Node *use = i.get(); 2957 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2958 proj = use; 2959 break; 2960 } 2961 } 2962 assert(proj != NULL, "must be found"); 2963 p->subsume_by(proj, this); 2964 } 2965 } 2966 break; 2967 } 2968 2969 case Op_Phi: 2970 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2971 // The EncodeP optimization may create Phi with the same edges 2972 // for all paths. It is not handled well by Register Allocator. 2973 Node* unique_in = n->in(1); 2974 assert(unique_in != NULL, ""); 2975 uint cnt = n->req(); 2976 for (uint i = 2; i < cnt; i++) { 2977 Node* m = n->in(i); 2978 assert(m != NULL, ""); 2979 if (unique_in != m) 2980 unique_in = NULL; 2981 } 2982 if (unique_in != NULL) { 2983 n->subsume_by(unique_in, this); 2984 } 2985 } 2986 break; 2987 2988 #endif 2989 2990 case Op_ModI: 2991 if (UseDivMod) { 2992 // Check if a%b and a/b both exist 2993 Node* d = n->find_similar(Op_DivI); 2994 if (d) { 2995 // Replace them with a fused divmod if supported 2996 if (Matcher::has_match_rule(Op_DivModI)) { 2997 DivModINode* divmod = DivModINode::make(this, n); 2998 d->subsume_by(divmod->div_proj(), this); 2999 n->subsume_by(divmod->mod_proj(), this); 3000 } else { 3001 // replace a%b with a-((a/b)*b) 3002 Node* mult = new (this) MulINode(d, d->in(2)); 3003 Node* sub = new (this) SubINode(d->in(1), mult); 3004 n->subsume_by(sub, this); 3005 } 3006 } 3007 } 3008 break; 3009 3010 case Op_ModL: 3011 if (UseDivMod) { 3012 // Check if a%b and a/b both exist 3013 Node* d = n->find_similar(Op_DivL); 3014 if (d) { 3015 // Replace them with a fused divmod if supported 3016 if (Matcher::has_match_rule(Op_DivModL)) { 3017 DivModLNode* divmod = DivModLNode::make(this, n); 3018 d->subsume_by(divmod->div_proj(), this); 3019 n->subsume_by(divmod->mod_proj(), this); 3020 } else { 3021 // replace a%b with a-((a/b)*b) 3022 Node* mult = new (this) MulLNode(d, d->in(2)); 3023 Node* sub = new (this) SubLNode(d->in(1), mult); 3024 n->subsume_by(sub, this); 3025 } 3026 } 3027 } 3028 break; 3029 3030 case Op_LoadVector: 3031 case Op_StoreVector: 3032 break; 3033 3034 case Op_PackB: 3035 case Op_PackS: 3036 case Op_PackI: 3037 case Op_PackF: 3038 case Op_PackL: 3039 case Op_PackD: 3040 if (n->req()-1 > 2) { 3041 // Replace many operand PackNodes with a binary tree for matching 3042 PackNode* p = (PackNode*) n; 3043 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3044 n->subsume_by(btp, this); 3045 } 3046 break; 3047 case Op_Loop: 3048 case Op_CountedLoop: 3049 if (n->as_Loop()->is_inner_loop()) { 3050 frc.inc_inner_loop_count(); 3051 } 3052 break; 3053 case Op_LShiftI: 3054 case Op_RShiftI: 3055 case Op_URShiftI: 3056 case Op_LShiftL: 3057 case Op_RShiftL: 3058 case Op_URShiftL: 3059 if (Matcher::need_masked_shift_count) { 3060 // The cpu's shift instructions don't restrict the count to the 3061 // lower 5/6 bits. We need to do the masking ourselves. 3062 Node* in2 = n->in(2); 3063 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3064 const TypeInt* t = in2->find_int_type(); 3065 if (t != NULL && t->is_con()) { 3066 juint shift = t->get_con(); 3067 if (shift > mask) { // Unsigned cmp 3068 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3069 } 3070 } else { 3071 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3072 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3073 n->set_req(2, shift); 3074 } 3075 } 3076 if (in2->outcnt() == 0) { // Remove dead node 3077 in2->disconnect_inputs(NULL, this); 3078 } 3079 } 3080 break; 3081 case Op_MemBarStoreStore: 3082 case Op_MemBarRelease: 3083 // Break the link with AllocateNode: it is no longer useful and 3084 // confuses register allocation. 3085 if (n->req() > MemBarNode::Precedent) { 3086 n->set_req(MemBarNode::Precedent, top()); 3087 } 3088 break; 3089 default: 3090 assert( !n->is_Call(), "" ); 3091 assert( !n->is_Mem(), "" ); 3092 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3093 break; 3094 } 3095 3096 // Collect CFG split points 3097 if (n->is_MultiBranch()) 3098 frc._tests.push(n); 3099 } 3100 3101 //------------------------------final_graph_reshaping_walk--------------------- 3102 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3103 // requires that the walk visits a node's inputs before visiting the node. 3104 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3105 ResourceArea *area = Thread::current()->resource_area(); 3106 Unique_Node_List sfpt(area); 3107 3108 frc._visited.set(root->_idx); // first, mark node as visited 3109 uint cnt = root->req(); 3110 Node *n = root; 3111 uint i = 0; 3112 while (true) { 3113 if (i < cnt) { 3114 // Place all non-visited non-null inputs onto stack 3115 Node* m = n->in(i); 3116 ++i; 3117 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3118 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3119 // compute worst case interpreter size in case of a deoptimization 3120 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3121 3122 sfpt.push(m); 3123 } 3124 cnt = m->req(); 3125 nstack.push(n, i); // put on stack parent and next input's index 3126 n = m; 3127 i = 0; 3128 } 3129 } else { 3130 // Now do post-visit work 3131 final_graph_reshaping_impl( n, frc ); 3132 if (nstack.is_empty()) 3133 break; // finished 3134 n = nstack.node(); // Get node from stack 3135 cnt = n->req(); 3136 i = nstack.index(); 3137 nstack.pop(); // Shift to the next node on stack 3138 } 3139 } 3140 3141 // Skip next transformation if compressed oops are not used. 3142 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3143 (!UseCompressedOops && !UseCompressedClassPointers)) 3144 return; 3145 3146 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3147 // It could be done for an uncommon traps or any safepoints/calls 3148 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3149 while (sfpt.size() > 0) { 3150 n = sfpt.pop(); 3151 JVMState *jvms = n->as_SafePoint()->jvms(); 3152 assert(jvms != NULL, "sanity"); 3153 int start = jvms->debug_start(); 3154 int end = n->req(); 3155 bool is_uncommon = (n->is_CallStaticJava() && 3156 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3157 for (int j = start; j < end; j++) { 3158 Node* in = n->in(j); 3159 if (in->is_DecodeNarrowPtr()) { 3160 bool safe_to_skip = true; 3161 if (!is_uncommon ) { 3162 // Is it safe to skip? 3163 for (uint i = 0; i < in->outcnt(); i++) { 3164 Node* u = in->raw_out(i); 3165 if (!u->is_SafePoint() || 3166 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3167 safe_to_skip = false; 3168 } 3169 } 3170 } 3171 if (safe_to_skip) { 3172 n->set_req(j, in->in(1)); 3173 } 3174 if (in->outcnt() == 0) { 3175 in->disconnect_inputs(NULL, this); 3176 } 3177 } 3178 } 3179 } 3180 } 3181 3182 //------------------------------final_graph_reshaping-------------------------- 3183 // Final Graph Reshaping. 3184 // 3185 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3186 // and not commoned up and forced early. Must come after regular 3187 // optimizations to avoid GVN undoing the cloning. Clone constant 3188 // inputs to Loop Phis; these will be split by the allocator anyways. 3189 // Remove Opaque nodes. 3190 // (2) Move last-uses by commutative operations to the left input to encourage 3191 // Intel update-in-place two-address operations and better register usage 3192 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3193 // calls canonicalizing them back. 3194 // (3) Count the number of double-precision FP ops, single-precision FP ops 3195 // and call sites. On Intel, we can get correct rounding either by 3196 // forcing singles to memory (requires extra stores and loads after each 3197 // FP bytecode) or we can set a rounding mode bit (requires setting and 3198 // clearing the mode bit around call sites). The mode bit is only used 3199 // if the relative frequency of single FP ops to calls is low enough. 3200 // This is a key transform for SPEC mpeg_audio. 3201 // (4) Detect infinite loops; blobs of code reachable from above but not 3202 // below. Several of the Code_Gen algorithms fail on such code shapes, 3203 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3204 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3205 // Detection is by looking for IfNodes where only 1 projection is 3206 // reachable from below or CatchNodes missing some targets. 3207 // (5) Assert for insane oop offsets in debug mode. 3208 3209 bool Compile::final_graph_reshaping() { 3210 // an infinite loop may have been eliminated by the optimizer, 3211 // in which case the graph will be empty. 3212 if (root()->req() == 1) { 3213 record_method_not_compilable("trivial infinite loop"); 3214 return true; 3215 } 3216 3217 // Expensive nodes have their control input set to prevent the GVN 3218 // from freely commoning them. There's no GVN beyond this point so 3219 // no need to keep the control input. We want the expensive nodes to 3220 // be freely moved to the least frequent code path by gcm. 3221 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3222 for (int i = 0; i < expensive_count(); i++) { 3223 _expensive_nodes->at(i)->set_req(0, NULL); 3224 } 3225 3226 Final_Reshape_Counts frc; 3227 3228 // Visit everybody reachable! 3229 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3230 Node_Stack nstack(live_nodes() >> 1); 3231 final_graph_reshaping_walk(nstack, root(), frc); 3232 3233 // Check for unreachable (from below) code (i.e., infinite loops). 3234 for( uint i = 0; i < frc._tests.size(); i++ ) { 3235 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3236 // Get number of CFG targets. 3237 // Note that PCTables include exception targets after calls. 3238 uint required_outcnt = n->required_outcnt(); 3239 if (n->outcnt() != required_outcnt) { 3240 // Check for a few special cases. Rethrow Nodes never take the 3241 // 'fall-thru' path, so expected kids is 1 less. 3242 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3243 if (n->in(0)->in(0)->is_Call()) { 3244 CallNode *call = n->in(0)->in(0)->as_Call(); 3245 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3246 required_outcnt--; // Rethrow always has 1 less kid 3247 } else if (call->req() > TypeFunc::Parms && 3248 call->is_CallDynamicJava()) { 3249 // Check for null receiver. In such case, the optimizer has 3250 // detected that the virtual call will always result in a null 3251 // pointer exception. The fall-through projection of this CatchNode 3252 // will not be populated. 3253 Node *arg0 = call->in(TypeFunc::Parms); 3254 if (arg0->is_Type() && 3255 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3256 required_outcnt--; 3257 } 3258 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3259 call->req() > TypeFunc::Parms+1 && 3260 call->is_CallStaticJava()) { 3261 // Check for negative array length. In such case, the optimizer has 3262 // detected that the allocation attempt will always result in an 3263 // exception. There is no fall-through projection of this CatchNode . 3264 Node *arg1 = call->in(TypeFunc::Parms+1); 3265 if (arg1->is_Type() && 3266 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3267 required_outcnt--; 3268 } 3269 } 3270 } 3271 } 3272 // Recheck with a better notion of 'required_outcnt' 3273 if (n->outcnt() != required_outcnt) { 3274 record_method_not_compilable("malformed control flow"); 3275 return true; // Not all targets reachable! 3276 } 3277 } 3278 // Check that I actually visited all kids. Unreached kids 3279 // must be infinite loops. 3280 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3281 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3282 record_method_not_compilable("infinite loop"); 3283 return true; // Found unvisited kid; must be unreach 3284 } 3285 } 3286 3287 // If original bytecodes contained a mixture of floats and doubles 3288 // check if the optimizer has made it homogenous, item (3). 3289 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3290 frc.get_float_count() > 32 && 3291 frc.get_double_count() == 0 && 3292 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3293 set_24_bit_selection_and_mode( false, true ); 3294 } 3295 3296 set_java_calls(frc.get_java_call_count()); 3297 set_inner_loops(frc.get_inner_loop_count()); 3298 3299 // No infinite loops, no reason to bail out. 3300 return false; 3301 } 3302 3303 //-----------------------------too_many_traps---------------------------------- 3304 // Report if there are too many traps at the current method and bci. 3305 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3306 bool Compile::too_many_traps(ciMethod* method, 3307 int bci, 3308 Deoptimization::DeoptReason reason) { 3309 ciMethodData* md = method->method_data(); 3310 if (md->is_empty()) { 3311 // Assume the trap has not occurred, or that it occurred only 3312 // because of a transient condition during start-up in the interpreter. 3313 return false; 3314 } 3315 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3316 if (md->has_trap_at(bci, m, reason) != 0) { 3317 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3318 // Also, if there are multiple reasons, or if there is no per-BCI record, 3319 // assume the worst. 3320 if (log()) 3321 log()->elem("observe trap='%s' count='%d'", 3322 Deoptimization::trap_reason_name(reason), 3323 md->trap_count(reason)); 3324 return true; 3325 } else { 3326 // Ignore method/bci and see if there have been too many globally. 3327 return too_many_traps(reason, md); 3328 } 3329 } 3330 3331 // Less-accurate variant which does not require a method and bci. 3332 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3333 ciMethodData* logmd) { 3334 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3335 // Too many traps globally. 3336 // Note that we use cumulative trap_count, not just md->trap_count. 3337 if (log()) { 3338 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3339 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3340 Deoptimization::trap_reason_name(reason), 3341 mcount, trap_count(reason)); 3342 } 3343 return true; 3344 } else { 3345 // The coast is clear. 3346 return false; 3347 } 3348 } 3349 3350 //--------------------------too_many_recompiles-------------------------------- 3351 // Report if there are too many recompiles at the current method and bci. 3352 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3353 // Is not eager to return true, since this will cause the compiler to use 3354 // Action_none for a trap point, to avoid too many recompilations. 3355 bool Compile::too_many_recompiles(ciMethod* method, 3356 int bci, 3357 Deoptimization::DeoptReason reason) { 3358 ciMethodData* md = method->method_data(); 3359 if (md->is_empty()) { 3360 // Assume the trap has not occurred, or that it occurred only 3361 // because of a transient condition during start-up in the interpreter. 3362 return false; 3363 } 3364 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3365 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3366 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3367 Deoptimization::DeoptReason per_bc_reason 3368 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3369 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3370 if ((per_bc_reason == Deoptimization::Reason_none 3371 || md->has_trap_at(bci, m, reason) != 0) 3372 // The trap frequency measure we care about is the recompile count: 3373 && md->trap_recompiled_at(bci, m) 3374 && md->overflow_recompile_count() >= bc_cutoff) { 3375 // Do not emit a trap here if it has already caused recompilations. 3376 // Also, if there are multiple reasons, or if there is no per-BCI record, 3377 // assume the worst. 3378 if (log()) 3379 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3380 Deoptimization::trap_reason_name(reason), 3381 md->trap_count(reason), 3382 md->overflow_recompile_count()); 3383 return true; 3384 } else if (trap_count(reason) != 0 3385 && decompile_count() >= m_cutoff) { 3386 // Too many recompiles globally, and we have seen this sort of trap. 3387 // Use cumulative decompile_count, not just md->decompile_count. 3388 if (log()) 3389 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3390 Deoptimization::trap_reason_name(reason), 3391 md->trap_count(reason), trap_count(reason), 3392 md->decompile_count(), decompile_count()); 3393 return true; 3394 } else { 3395 // The coast is clear. 3396 return false; 3397 } 3398 } 3399 3400 // Compute when not to trap. Used by matching trap based nodes and 3401 // NullCheck optimization. 3402 void Compile::set_allowed_deopt_reasons() { 3403 _allowed_reasons = 0; 3404 if (is_method_compilation()) { 3405 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3406 assert(rs < BitsPerInt, "recode bit map"); 3407 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3408 _allowed_reasons |= nth_bit(rs); 3409 } 3410 } 3411 } 3412 } 3413 3414 #ifndef PRODUCT 3415 //------------------------------verify_graph_edges--------------------------- 3416 // Walk the Graph and verify that there is a one-to-one correspondence 3417 // between Use-Def edges and Def-Use edges in the graph. 3418 void Compile::verify_graph_edges(bool no_dead_code) { 3419 if (VerifyGraphEdges) { 3420 ResourceArea *area = Thread::current()->resource_area(); 3421 Unique_Node_List visited(area); 3422 // Call recursive graph walk to check edges 3423 _root->verify_edges(visited); 3424 if (no_dead_code) { 3425 // Now make sure that no visited node is used by an unvisited node. 3426 bool dead_nodes = 0; 3427 Unique_Node_List checked(area); 3428 while (visited.size() > 0) { 3429 Node* n = visited.pop(); 3430 checked.push(n); 3431 for (uint i = 0; i < n->outcnt(); i++) { 3432 Node* use = n->raw_out(i); 3433 if (checked.member(use)) continue; // already checked 3434 if (visited.member(use)) continue; // already in the graph 3435 if (use->is_Con()) continue; // a dead ConNode is OK 3436 // At this point, we have found a dead node which is DU-reachable. 3437 if (dead_nodes++ == 0) 3438 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3439 use->dump(2); 3440 tty->print_cr("---"); 3441 checked.push(use); // No repeats; pretend it is now checked. 3442 } 3443 } 3444 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3445 } 3446 } 3447 } 3448 3449 // Verify GC barriers consistency 3450 // Currently supported: 3451 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3452 void Compile::verify_barriers() { 3453 if (UseG1GC) { 3454 // Verify G1 pre-barriers 3455 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3456 3457 ResourceArea *area = Thread::current()->resource_area(); 3458 Unique_Node_List visited(area); 3459 Node_List worklist(area); 3460 // We're going to walk control flow backwards starting from the Root 3461 worklist.push(_root); 3462 while (worklist.size() > 0) { 3463 Node* x = worklist.pop(); 3464 if (x == NULL || x == top()) continue; 3465 if (visited.member(x)) { 3466 continue; 3467 } else { 3468 visited.push(x); 3469 } 3470 3471 if (x->is_Region()) { 3472 for (uint i = 1; i < x->req(); i++) { 3473 worklist.push(x->in(i)); 3474 } 3475 } else { 3476 worklist.push(x->in(0)); 3477 // We are looking for the pattern: 3478 // /->ThreadLocal 3479 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3480 // \->ConI(0) 3481 // We want to verify that the If and the LoadB have the same control 3482 // See GraphKit::g1_write_barrier_pre() 3483 if (x->is_If()) { 3484 IfNode *iff = x->as_If(); 3485 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3486 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3487 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3488 && cmp->in(1)->is_Load()) { 3489 LoadNode* load = cmp->in(1)->as_Load(); 3490 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3491 && load->in(2)->in(3)->is_Con() 3492 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3493 3494 Node* if_ctrl = iff->in(0); 3495 Node* load_ctrl = load->in(0); 3496 3497 if (if_ctrl != load_ctrl) { 3498 // Skip possible CProj->NeverBranch in infinite loops 3499 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3500 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3501 if_ctrl = if_ctrl->in(0)->in(0); 3502 } 3503 } 3504 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3505 } 3506 } 3507 } 3508 } 3509 } 3510 } 3511 } 3512 } 3513 3514 #endif 3515 3516 // The Compile object keeps track of failure reasons separately from the ciEnv. 3517 // This is required because there is not quite a 1-1 relation between the 3518 // ciEnv and its compilation task and the Compile object. Note that one 3519 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3520 // to backtrack and retry without subsuming loads. Other than this backtracking 3521 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3522 // by the logic in C2Compiler. 3523 void Compile::record_failure(const char* reason) { 3524 if (log() != NULL) { 3525 log()->elem("failure reason='%s' phase='compile'", reason); 3526 } 3527 if (_failure_reason == NULL) { 3528 // Record the first failure reason. 3529 _failure_reason = reason; 3530 } 3531 3532 EventCompilerFailure event; 3533 if (event.should_commit()) { 3534 event.set_compileID(Compile::compile_id()); 3535 event.set_failure(reason); 3536 event.commit(); 3537 } 3538 3539 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3540 C->print_method(PHASE_FAILURE); 3541 } 3542 _root = NULL; // flush the graph, too 3543 } 3544 3545 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3546 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3547 _phase_name(name), _dolog(dolog) 3548 { 3549 if (dolog) { 3550 C = Compile::current(); 3551 _log = C->log(); 3552 } else { 3553 C = NULL; 3554 _log = NULL; 3555 } 3556 if (_log != NULL) { 3557 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3558 _log->stamp(); 3559 _log->end_head(); 3560 } 3561 } 3562 3563 Compile::TracePhase::~TracePhase() { 3564 3565 C = Compile::current(); 3566 if (_dolog) { 3567 _log = C->log(); 3568 } else { 3569 _log = NULL; 3570 } 3571 3572 #ifdef ASSERT 3573 if (PrintIdealNodeCount) { 3574 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3575 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3576 } 3577 3578 if (VerifyIdealNodeCount) { 3579 Compile::current()->print_missing_nodes(); 3580 } 3581 #endif 3582 3583 if (_log != NULL) { 3584 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3585 } 3586 } 3587 3588 //============================================================================= 3589 // Two Constant's are equal when the type and the value are equal. 3590 bool Compile::Constant::operator==(const Constant& other) { 3591 if (type() != other.type() ) return false; 3592 if (can_be_reused() != other.can_be_reused()) return false; 3593 // For floating point values we compare the bit pattern. 3594 switch (type()) { 3595 case T_FLOAT: return (_v._value.i == other._v._value.i); 3596 case T_LONG: 3597 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3598 case T_OBJECT: 3599 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3600 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3601 case T_METADATA: return (_v._metadata == other._v._metadata); 3602 default: ShouldNotReachHere(); 3603 } 3604 return false; 3605 } 3606 3607 static int type_to_size_in_bytes(BasicType t) { 3608 switch (t) { 3609 case T_LONG: return sizeof(jlong ); 3610 case T_FLOAT: return sizeof(jfloat ); 3611 case T_DOUBLE: return sizeof(jdouble); 3612 case T_METADATA: return sizeof(Metadata*); 3613 // We use T_VOID as marker for jump-table entries (labels) which 3614 // need an internal word relocation. 3615 case T_VOID: 3616 case T_ADDRESS: 3617 case T_OBJECT: return sizeof(jobject); 3618 } 3619 3620 ShouldNotReachHere(); 3621 return -1; 3622 } 3623 3624 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3625 // sort descending 3626 if (a->freq() > b->freq()) return -1; 3627 if (a->freq() < b->freq()) return 1; 3628 return 0; 3629 } 3630 3631 void Compile::ConstantTable::calculate_offsets_and_size() { 3632 // First, sort the array by frequencies. 3633 _constants.sort(qsort_comparator); 3634 3635 #ifdef ASSERT 3636 // Make sure all jump-table entries were sorted to the end of the 3637 // array (they have a negative frequency). 3638 bool found_void = false; 3639 for (int i = 0; i < _constants.length(); i++) { 3640 Constant con = _constants.at(i); 3641 if (con.type() == T_VOID) 3642 found_void = true; // jump-tables 3643 else 3644 assert(!found_void, "wrong sorting"); 3645 } 3646 #endif 3647 3648 int offset = 0; 3649 for (int i = 0; i < _constants.length(); i++) { 3650 Constant* con = _constants.adr_at(i); 3651 3652 // Align offset for type. 3653 int typesize = type_to_size_in_bytes(con->type()); 3654 offset = align_size_up(offset, typesize); 3655 con->set_offset(offset); // set constant's offset 3656 3657 if (con->type() == T_VOID) { 3658 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3659 offset = offset + typesize * n->outcnt(); // expand jump-table 3660 } else { 3661 offset = offset + typesize; 3662 } 3663 } 3664 3665 // Align size up to the next section start (which is insts; see 3666 // CodeBuffer::align_at_start). 3667 assert(_size == -1, "already set?"); 3668 _size = align_size_up(offset, CodeEntryAlignment); 3669 } 3670 3671 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3672 MacroAssembler _masm(&cb); 3673 for (int i = 0; i < _constants.length(); i++) { 3674 Constant con = _constants.at(i); 3675 address constant_addr; 3676 switch (con.type()) { 3677 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3678 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3679 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3680 case T_OBJECT: { 3681 jobject obj = con.get_jobject(); 3682 int oop_index = _masm.oop_recorder()->find_index(obj); 3683 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3684 break; 3685 } 3686 case T_ADDRESS: { 3687 address addr = (address) con.get_jobject(); 3688 constant_addr = _masm.address_constant(addr); 3689 break; 3690 } 3691 // We use T_VOID as marker for jump-table entries (labels) which 3692 // need an internal word relocation. 3693 case T_VOID: { 3694 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3695 // Fill the jump-table with a dummy word. The real value is 3696 // filled in later in fill_jump_table. 3697 address dummy = (address) n; 3698 constant_addr = _masm.address_constant(dummy); 3699 // Expand jump-table 3700 for (uint i = 1; i < n->outcnt(); i++) { 3701 address temp_addr = _masm.address_constant(dummy + i); 3702 assert(temp_addr, "consts section too small"); 3703 } 3704 break; 3705 } 3706 case T_METADATA: { 3707 Metadata* obj = con.get_metadata(); 3708 int metadata_index = _masm.oop_recorder()->find_index(obj); 3709 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3710 break; 3711 } 3712 default: ShouldNotReachHere(); 3713 } 3714 assert(constant_addr, "consts section too small"); 3715 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3716 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3717 } 3718 } 3719 3720 int Compile::ConstantTable::find_offset(Constant& con) const { 3721 int idx = _constants.find(con); 3722 assert(idx != -1, "constant must be in constant table"); 3723 int offset = _constants.at(idx).offset(); 3724 assert(offset != -1, "constant table not emitted yet?"); 3725 return offset; 3726 } 3727 3728 void Compile::ConstantTable::add(Constant& con) { 3729 if (con.can_be_reused()) { 3730 int idx = _constants.find(con); 3731 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3732 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3733 return; 3734 } 3735 } 3736 (void) _constants.append(con); 3737 } 3738 3739 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3740 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3741 Constant con(type, value, b->_freq); 3742 add(con); 3743 return con; 3744 } 3745 3746 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3747 Constant con(metadata); 3748 add(con); 3749 return con; 3750 } 3751 3752 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3753 jvalue value; 3754 BasicType type = oper->type()->basic_type(); 3755 switch (type) { 3756 case T_LONG: value.j = oper->constantL(); break; 3757 case T_FLOAT: value.f = oper->constantF(); break; 3758 case T_DOUBLE: value.d = oper->constantD(); break; 3759 case T_OBJECT: 3760 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3761 case T_METADATA: return add((Metadata*)oper->constant()); break; 3762 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3763 } 3764 return add(n, type, value); 3765 } 3766 3767 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3768 jvalue value; 3769 // We can use the node pointer here to identify the right jump-table 3770 // as this method is called from Compile::Fill_buffer right before 3771 // the MachNodes are emitted and the jump-table is filled (means the 3772 // MachNode pointers do not change anymore). 3773 value.l = (jobject) n; 3774 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3775 add(con); 3776 return con; 3777 } 3778 3779 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3780 // If called from Compile::scratch_emit_size do nothing. 3781 if (Compile::current()->in_scratch_emit_size()) return; 3782 3783 assert(labels.is_nonempty(), "must be"); 3784 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3785 3786 // Since MachConstantNode::constant_offset() also contains 3787 // table_base_offset() we need to subtract the table_base_offset() 3788 // to get the plain offset into the constant table. 3789 int offset = n->constant_offset() - table_base_offset(); 3790 3791 MacroAssembler _masm(&cb); 3792 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3793 3794 for (uint i = 0; i < n->outcnt(); i++) { 3795 address* constant_addr = &jump_table_base[i]; 3796 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i))); 3797 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3798 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3799 } 3800 } 3801 3802 void Compile::dump_inlining() { 3803 if (print_inlining() || print_intrinsics()) { 3804 // Print inlining message for candidates that we couldn't inline 3805 // for lack of space or non constant receiver 3806 for (int i = 0; i < _late_inlines.length(); i++) { 3807 CallGenerator* cg = _late_inlines.at(i); 3808 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3809 } 3810 Unique_Node_List useful; 3811 useful.push(root()); 3812 for (uint next = 0; next < useful.size(); ++next) { 3813 Node* n = useful.at(next); 3814 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3815 CallNode* call = n->as_Call(); 3816 CallGenerator* cg = call->generator(); 3817 cg->print_inlining_late("receiver not constant"); 3818 } 3819 uint max = n->len(); 3820 for ( uint i = 0; i < max; ++i ) { 3821 Node *m = n->in(i); 3822 if ( m == NULL ) continue; 3823 useful.push(m); 3824 } 3825 } 3826 for (int i = 0; i < _print_inlining_list->length(); i++) { 3827 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3828 } 3829 } 3830 } 3831 3832 // Dump inlining replay data to the stream. 3833 // Don't change thread state and acquire any locks. 3834 void Compile::dump_inline_data(outputStream* out) { 3835 InlineTree* inl_tree = ilt(); 3836 if (inl_tree != NULL) { 3837 out->print(" inline %d", inl_tree->count()); 3838 inl_tree->dump_replay_data(out); 3839 } 3840 } 3841 3842 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3843 if (n1->Opcode() < n2->Opcode()) return -1; 3844 else if (n1->Opcode() > n2->Opcode()) return 1; 3845 3846 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())); 3847 for (uint i = 1; i < n1->req(); i++) { 3848 if (n1->in(i) < n2->in(i)) return -1; 3849 else if (n1->in(i) > n2->in(i)) return 1; 3850 } 3851 3852 return 0; 3853 } 3854 3855 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3856 Node* n1 = *n1p; 3857 Node* n2 = *n2p; 3858 3859 return cmp_expensive_nodes(n1, n2); 3860 } 3861 3862 void Compile::sort_expensive_nodes() { 3863 if (!expensive_nodes_sorted()) { 3864 _expensive_nodes->sort(cmp_expensive_nodes); 3865 } 3866 } 3867 3868 bool Compile::expensive_nodes_sorted() const { 3869 for (int i = 1; i < _expensive_nodes->length(); i++) { 3870 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3871 return false; 3872 } 3873 } 3874 return true; 3875 } 3876 3877 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3878 if (_expensive_nodes->length() == 0) { 3879 return false; 3880 } 3881 3882 assert(OptimizeExpensiveOps, "optimization off?"); 3883 3884 // Take this opportunity to remove dead nodes from the list 3885 int j = 0; 3886 for (int i = 0; i < _expensive_nodes->length(); i++) { 3887 Node* n = _expensive_nodes->at(i); 3888 if (!n->is_unreachable(igvn)) { 3889 assert(n->is_expensive(), "should be expensive"); 3890 _expensive_nodes->at_put(j, n); 3891 j++; 3892 } 3893 } 3894 _expensive_nodes->trunc_to(j); 3895 3896 // Then sort the list so that similar nodes are next to each other 3897 // and check for at least two nodes of identical kind with same data 3898 // inputs. 3899 sort_expensive_nodes(); 3900 3901 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3902 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3903 return true; 3904 } 3905 } 3906 3907 return false; 3908 } 3909 3910 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3911 if (_expensive_nodes->length() == 0) { 3912 return; 3913 } 3914 3915 assert(OptimizeExpensiveOps, "optimization off?"); 3916 3917 // Sort to bring similar nodes next to each other and clear the 3918 // control input of nodes for which there's only a single copy. 3919 sort_expensive_nodes(); 3920 3921 int j = 0; 3922 int identical = 0; 3923 int i = 0; 3924 for (; i < _expensive_nodes->length()-1; i++) { 3925 assert(j <= i, "can't write beyond current index"); 3926 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3927 identical++; 3928 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3929 continue; 3930 } 3931 if (identical > 0) { 3932 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3933 identical = 0; 3934 } else { 3935 Node* n = _expensive_nodes->at(i); 3936 igvn.hash_delete(n); 3937 n->set_req(0, NULL); 3938 igvn.hash_insert(n); 3939 } 3940 } 3941 if (identical > 0) { 3942 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3943 } else if (_expensive_nodes->length() >= 1) { 3944 Node* n = _expensive_nodes->at(i); 3945 igvn.hash_delete(n); 3946 n->set_req(0, NULL); 3947 igvn.hash_insert(n); 3948 } 3949 _expensive_nodes->trunc_to(j); 3950 } 3951 3952 void Compile::add_expensive_node(Node * n) { 3953 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3954 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3955 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3956 if (OptimizeExpensiveOps) { 3957 _expensive_nodes->append(n); 3958 } else { 3959 // Clear control input and let IGVN optimize expensive nodes if 3960 // OptimizeExpensiveOps is off. 3961 n->set_req(0, NULL); 3962 } 3963 } 3964 3965 /** 3966 * Remove the speculative part of types and clean up the graph 3967 */ 3968 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 3969 if (UseTypeSpeculation) { 3970 Unique_Node_List worklist; 3971 worklist.push(root()); 3972 int modified = 0; 3973 // Go over all type nodes that carry a speculative type, drop the 3974 // speculative part of the type and enqueue the node for an igvn 3975 // which may optimize it out. 3976 for (uint next = 0; next < worklist.size(); ++next) { 3977 Node *n = worklist.at(next); 3978 if (n->is_Type()) { 3979 TypeNode* tn = n->as_Type(); 3980 const Type* t = tn->type(); 3981 const Type* t_no_spec = t->remove_speculative(); 3982 if (t_no_spec != t) { 3983 bool in_hash = igvn.hash_delete(n); 3984 assert(in_hash, "node should be in igvn hash table"); 3985 tn->set_type(t_no_spec); 3986 igvn.hash_insert(n); 3987 igvn._worklist.push(n); // give it a chance to go away 3988 modified++; 3989 } 3990 } 3991 uint max = n->len(); 3992 for( uint i = 0; i < max; ++i ) { 3993 Node *m = n->in(i); 3994 if (not_a_node(m)) continue; 3995 worklist.push(m); 3996 } 3997 } 3998 // Drop the speculative part of all types in the igvn's type table 3999 igvn.remove_speculative_types(); 4000 if (modified > 0) { 4001 igvn.optimize(); 4002 } 4003 #ifdef ASSERT 4004 // Verify that after the IGVN is over no speculative type has resurfaced 4005 worklist.clear(); 4006 worklist.push(root()); 4007 for (uint next = 0; next < worklist.size(); ++next) { 4008 Node *n = worklist.at(next); 4009 const Type* t = igvn.type_or_null(n); 4010 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4011 if (n->is_Type()) { 4012 t = n->as_Type()->type(); 4013 assert(t == t->remove_speculative(), "no more speculative types"); 4014 } 4015 uint max = n->len(); 4016 for( uint i = 0; i < max; ++i ) { 4017 Node *m = n->in(i); 4018 if (not_a_node(m)) continue; 4019 worklist.push(m); 4020 } 4021 } 4022 igvn.check_no_speculative_types(); 4023 #endif 4024 } 4025 } 4026 4027 // Auxiliary method to support randomized stressing/fuzzing. 4028 // 4029 // This method can be called the arbitrary number of times, with current count 4030 // as the argument. The logic allows selecting a single candidate from the 4031 // running list of candidates as follows: 4032 // int count = 0; 4033 // Cand* selected = null; 4034 // while(cand = cand->next()) { 4035 // if (randomized_select(++count)) { 4036 // selected = cand; 4037 // } 4038 // } 4039 // 4040 // Including count equalizes the chances any candidate is "selected". 4041 // This is useful when we don't have the complete list of candidates to choose 4042 // from uniformly. In this case, we need to adjust the randomicity of the 4043 // selection, or else we will end up biasing the selection towards the latter 4044 // candidates. 4045 // 4046 // Quick back-envelope calculation shows that for the list of n candidates 4047 // the equal probability for the candidate to persist as "best" can be 4048 // achieved by replacing it with "next" k-th candidate with the probability 4049 // of 1/k. It can be easily shown that by the end of the run, the 4050 // probability for any candidate is converged to 1/n, thus giving the 4051 // uniform distribution among all the candidates. 4052 // 4053 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4054 #define RANDOMIZED_DOMAIN_POW 29 4055 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4056 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4057 bool Compile::randomized_select(int count) { 4058 assert(count > 0, "only positive"); 4059 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4060 }