1 /* 2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "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 #ifdef TARGET_ARCH_MODEL_x86_32 71 # include "adfiles/ad_x86_32.hpp" 72 #endif 73 #ifdef TARGET_ARCH_MODEL_x86_64 74 # include "adfiles/ad_x86_64.hpp" 75 #endif 76 #ifdef TARGET_ARCH_MODEL_sparc 77 # include "adfiles/ad_sparc.hpp" 78 #endif 79 #ifdef TARGET_ARCH_MODEL_zero 80 # include "adfiles/ad_zero.hpp" 81 #endif 82 #ifdef TARGET_ARCH_MODEL_arm 83 # include "adfiles/ad_arm.hpp" 84 #endif 85 #ifdef TARGET_ARCH_MODEL_ppc_32 86 # include "adfiles/ad_ppc_32.hpp" 87 #endif 88 #ifdef TARGET_ARCH_MODEL_ppc_64 89 # include "adfiles/ad_ppc_64.hpp" 90 #endif 91 92 93 // -------------------- Compile::mach_constant_base_node ----------------------- 94 // Constant table base node singleton. 95 MachConstantBaseNode* Compile::mach_constant_base_node() { 96 if (_mach_constant_base_node == NULL) { 97 _mach_constant_base_node = new (C) MachConstantBaseNode(); 98 _mach_constant_base_node->add_req(C->root()); 99 } 100 return _mach_constant_base_node; 101 } 102 103 104 /// Support for intrinsics. 105 106 // Return the index at which m must be inserted (or already exists). 107 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 108 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 109 #ifdef ASSERT 110 for (int i = 1; i < _intrinsics->length(); i++) { 111 CallGenerator* cg1 = _intrinsics->at(i-1); 112 CallGenerator* cg2 = _intrinsics->at(i); 113 assert(cg1->method() != cg2->method() 114 ? cg1->method() < cg2->method() 115 : cg1->is_virtual() < cg2->is_virtual(), 116 "compiler intrinsics list must stay sorted"); 117 } 118 #endif 119 // Binary search sorted list, in decreasing intervals [lo, hi]. 120 int lo = 0, hi = _intrinsics->length()-1; 121 while (lo <= hi) { 122 int mid = (uint)(hi + lo) / 2; 123 ciMethod* mid_m = _intrinsics->at(mid)->method(); 124 if (m < mid_m) { 125 hi = mid-1; 126 } else if (m > mid_m) { 127 lo = mid+1; 128 } else { 129 // look at minor sort key 130 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 131 if (is_virtual < mid_virt) { 132 hi = mid-1; 133 } else if (is_virtual > mid_virt) { 134 lo = mid+1; 135 } else { 136 return mid; // exact match 137 } 138 } 139 } 140 return lo; // inexact match 141 } 142 143 void Compile::register_intrinsic(CallGenerator* cg) { 144 if (_intrinsics == NULL) { 145 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 146 } 147 // This code is stolen from ciObjectFactory::insert. 148 // Really, GrowableArray should have methods for 149 // insert_at, remove_at, and binary_search. 150 int len = _intrinsics->length(); 151 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 152 if (index == len) { 153 _intrinsics->append(cg); 154 } else { 155 #ifdef ASSERT 156 CallGenerator* oldcg = _intrinsics->at(index); 157 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 158 #endif 159 _intrinsics->append(_intrinsics->at(len-1)); 160 int pos; 161 for (pos = len-2; pos >= index; pos--) { 162 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 163 } 164 _intrinsics->at_put(index, cg); 165 } 166 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 167 } 168 169 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 170 assert(m->is_loaded(), "don't try this on unloaded methods"); 171 if (_intrinsics != NULL) { 172 int index = intrinsic_insertion_index(m, is_virtual); 173 if (index < _intrinsics->length() 174 && _intrinsics->at(index)->method() == m 175 && _intrinsics->at(index)->is_virtual() == is_virtual) { 176 return _intrinsics->at(index); 177 } 178 } 179 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 180 if (m->intrinsic_id() != vmIntrinsics::_none && 181 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 182 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 183 if (cg != NULL) { 184 // Save it for next time: 185 register_intrinsic(cg); 186 return cg; 187 } else { 188 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 189 } 190 } 191 return NULL; 192 } 193 194 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 195 // in library_call.cpp. 196 197 198 #ifndef PRODUCT 199 // statistics gathering... 200 201 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 202 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 203 204 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 205 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 206 int oflags = _intrinsic_hist_flags[id]; 207 assert(flags != 0, "what happened?"); 208 if (is_virtual) { 209 flags |= _intrinsic_virtual; 210 } 211 bool changed = (flags != oflags); 212 if ((flags & _intrinsic_worked) != 0) { 213 juint count = (_intrinsic_hist_count[id] += 1); 214 if (count == 1) { 215 changed = true; // first time 216 } 217 // increment the overall count also: 218 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 219 } 220 if (changed) { 221 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 222 // Something changed about the intrinsic's virtuality. 223 if ((flags & _intrinsic_virtual) != 0) { 224 // This is the first use of this intrinsic as a virtual call. 225 if (oflags != 0) { 226 // We already saw it as a non-virtual, so note both cases. 227 flags |= _intrinsic_both; 228 } 229 } else if ((oflags & _intrinsic_both) == 0) { 230 // This is the first use of this intrinsic as a non-virtual 231 flags |= _intrinsic_both; 232 } 233 } 234 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 235 } 236 // update the overall flags also: 237 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 238 return changed; 239 } 240 241 static char* format_flags(int flags, char* buf) { 242 buf[0] = 0; 243 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 244 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 245 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 246 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 247 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 248 if (buf[0] == 0) strcat(buf, ","); 249 assert(buf[0] == ',', "must be"); 250 return &buf[1]; 251 } 252 253 void Compile::print_intrinsic_statistics() { 254 char flagsbuf[100]; 255 ttyLocker ttyl; 256 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 257 tty->print_cr("Compiler intrinsic usage:"); 258 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 259 if (total == 0) total = 1; // avoid div0 in case of no successes 260 #define PRINT_STAT_LINE(name, c, f) \ 261 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 262 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 263 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 264 int flags = _intrinsic_hist_flags[id]; 265 juint count = _intrinsic_hist_count[id]; 266 if ((flags | count) != 0) { 267 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 268 } 269 } 270 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 271 if (xtty != NULL) xtty->tail("statistics"); 272 } 273 274 void Compile::print_statistics() { 275 { ttyLocker ttyl; 276 if (xtty != NULL) xtty->head("statistics type='opto'"); 277 Parse::print_statistics(); 278 PhaseCCP::print_statistics(); 279 PhaseRegAlloc::print_statistics(); 280 Scheduling::print_statistics(); 281 PhasePeephole::print_statistics(); 282 PhaseIdealLoop::print_statistics(); 283 if (xtty != NULL) xtty->tail("statistics"); 284 } 285 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 286 // put this under its own <statistics> element. 287 print_intrinsic_statistics(); 288 } 289 } 290 #endif //PRODUCT 291 292 // Support for bundling info 293 Bundle* Compile::node_bundling(const Node *n) { 294 assert(valid_bundle_info(n), "oob"); 295 return &_node_bundling_base[n->_idx]; 296 } 297 298 bool Compile::valid_bundle_info(const Node *n) { 299 return (_node_bundling_limit > n->_idx); 300 } 301 302 303 void Compile::gvn_replace_by(Node* n, Node* nn) { 304 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 305 Node* use = n->last_out(i); 306 bool is_in_table = initial_gvn()->hash_delete(use); 307 uint uses_found = 0; 308 for (uint j = 0; j < use->len(); j++) { 309 if (use->in(j) == n) { 310 if (j < use->req()) 311 use->set_req(j, nn); 312 else 313 use->set_prec(j, nn); 314 uses_found++; 315 } 316 } 317 if (is_in_table) { 318 // reinsert into table 319 initial_gvn()->hash_find_insert(use); 320 } 321 record_for_igvn(use); 322 i -= uses_found; // we deleted 1 or more copies of this edge 323 } 324 } 325 326 327 static inline bool not_a_node(const Node* n) { 328 if (n == NULL) return true; 329 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 330 if (*(address*)n == badAddress) return true; // kill by Node::destruct 331 return false; 332 } 333 334 // Identify all nodes that are reachable from below, useful. 335 // Use breadth-first pass that records state in a Unique_Node_List, 336 // recursive traversal is slower. 337 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 338 int estimated_worklist_size = unique(); 339 useful.map( estimated_worklist_size, NULL ); // preallocate space 340 341 // Initialize worklist 342 if (root() != NULL) { useful.push(root()); } 343 // If 'top' is cached, declare it useful to preserve cached node 344 if( cached_top_node() ) { useful.push(cached_top_node()); } 345 346 // Push all useful nodes onto the list, breadthfirst 347 for( uint next = 0; next < useful.size(); ++next ) { 348 assert( next < unique(), "Unique useful nodes < total nodes"); 349 Node *n = useful.at(next); 350 uint max = n->len(); 351 for( uint i = 0; i < max; ++i ) { 352 Node *m = n->in(i); 353 if (not_a_node(m)) continue; 354 useful.push(m); 355 } 356 } 357 } 358 359 // Update dead_node_list with any missing dead nodes using useful 360 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 361 void Compile::update_dead_node_list(Unique_Node_List &useful) { 362 uint max_idx = unique(); 363 VectorSet& useful_node_set = useful.member_set(); 364 365 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 366 // If node with index node_idx is not in useful set, 367 // mark it as dead in dead node list. 368 if (! useful_node_set.test(node_idx) ) { 369 record_dead_node(node_idx); 370 } 371 } 372 } 373 374 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 375 int shift = 0; 376 for (int i = 0; i < inlines->length(); i++) { 377 CallGenerator* cg = inlines->at(i); 378 CallNode* call = cg->call_node(); 379 if (shift > 0) { 380 inlines->at_put(i-shift, cg); 381 } 382 if (!useful.member(call)) { 383 shift++; 384 } 385 } 386 inlines->trunc_to(inlines->length()-shift); 387 } 388 389 // Disconnect all useless nodes by disconnecting those at the boundary. 390 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 391 uint next = 0; 392 while (next < useful.size()) { 393 Node *n = useful.at(next++); 394 // Use raw traversal of out edges since this code removes out edges 395 int max = n->outcnt(); 396 for (int j = 0; j < max; ++j) { 397 Node* child = n->raw_out(j); 398 if (! useful.member(child)) { 399 assert(!child->is_top() || child != top(), 400 "If top is cached in Compile object it is in useful list"); 401 // Only need to remove this out-edge to the useless node 402 n->raw_del_out(j); 403 --j; 404 --max; 405 } 406 } 407 if (n->outcnt() == 1 && n->has_special_unique_user()) { 408 record_for_igvn(n->unique_out()); 409 } 410 } 411 // Remove useless macro and predicate opaq nodes 412 for (int i = C->macro_count()-1; i >= 0; i--) { 413 Node* n = C->macro_node(i); 414 if (!useful.member(n)) { 415 remove_macro_node(n); 416 } 417 } 418 // Remove useless expensive node 419 for (int i = C->expensive_count()-1; i >= 0; i--) { 420 Node* n = C->expensive_node(i); 421 if (!useful.member(n)) { 422 remove_expensive_node(n); 423 } 424 } 425 // clean up the late inline lists 426 remove_useless_late_inlines(&_string_late_inlines, useful); 427 remove_useless_late_inlines(&_boxing_late_inlines, useful); 428 remove_useless_late_inlines(&_late_inlines, useful); 429 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 430 } 431 432 //------------------------------frame_size_in_words----------------------------- 433 // frame_slots in units of words 434 int Compile::frame_size_in_words() const { 435 // shift is 0 in LP32 and 1 in LP64 436 const int shift = (LogBytesPerWord - LogBytesPerInt); 437 int words = _frame_slots >> shift; 438 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 439 return words; 440 } 441 442 // To bang the stack of this compiled method we use the stack size 443 // that the interpreter would need in case of a deoptimization. This 444 // removes the need to bang the stack in the deoptimization blob which 445 // in turn simplifies stack overflow handling. 446 int Compile::bang_size_in_bytes() const { 447 int callee_locals = method() != NULL ? method()->max_locals() : 0; 448 int interpreter_frame_size = _interpreter_frame_size; 449 return MAX2(interpreter_frame_size, frame_size_in_bytes()); 450 } 451 452 // ============================================================================ 453 //------------------------------CompileWrapper--------------------------------- 454 class CompileWrapper : public StackObj { 455 Compile *const _compile; 456 public: 457 CompileWrapper(Compile* compile); 458 459 ~CompileWrapper(); 460 }; 461 462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 463 // the Compile* pointer is stored in the current ciEnv: 464 ciEnv* env = compile->env(); 465 assert(env == ciEnv::current(), "must already be a ciEnv active"); 466 assert(env->compiler_data() == NULL, "compile already active?"); 467 env->set_compiler_data(compile); 468 assert(compile == Compile::current(), "sanity"); 469 470 compile->set_type_dict(NULL); 471 compile->set_type_hwm(NULL); 472 compile->set_type_last_size(0); 473 compile->set_last_tf(NULL, NULL); 474 compile->set_indexSet_arena(NULL); 475 compile->set_indexSet_free_block_list(NULL); 476 compile->init_type_arena(); 477 Type::Initialize(compile); 478 _compile->set_scratch_buffer_blob(NULL); 479 _compile->begin_method(); 480 } 481 CompileWrapper::~CompileWrapper() { 482 _compile->end_method(); 483 if (_compile->scratch_buffer_blob() != NULL) 484 BufferBlob::free(_compile->scratch_buffer_blob()); 485 _compile->env()->set_compiler_data(NULL); 486 } 487 488 489 //----------------------------print_compile_messages--------------------------- 490 void Compile::print_compile_messages() { 491 #ifndef PRODUCT 492 // Check if recompiling 493 if (_subsume_loads == false && PrintOpto) { 494 // Recompiling without allowing machine instructions to subsume loads 495 tty->print_cr("*********************************************************"); 496 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 497 tty->print_cr("*********************************************************"); 498 } 499 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 500 // Recompiling without escape analysis 501 tty->print_cr("*********************************************************"); 502 tty->print_cr("** Bailout: Recompile without escape analysis **"); 503 tty->print_cr("*********************************************************"); 504 } 505 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 506 // Recompiling without boxing elimination 507 tty->print_cr("*********************************************************"); 508 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 509 tty->print_cr("*********************************************************"); 510 } 511 if (env()->break_at_compile()) { 512 // Open the debugger when compiling this method. 513 tty->print("### Breaking when compiling: "); 514 method()->print_short_name(); 515 tty->cr(); 516 BREAKPOINT; 517 } 518 519 if( PrintOpto ) { 520 if (is_osr_compilation()) { 521 tty->print("[OSR]%3d", _compile_id); 522 } else { 523 tty->print("%3d", _compile_id); 524 } 525 } 526 #endif 527 } 528 529 530 //-----------------------init_scratch_buffer_blob------------------------------ 531 // Construct a temporary BufferBlob and cache it for this compile. 532 void Compile::init_scratch_buffer_blob(int const_size) { 533 // If there is already a scratch buffer blob allocated and the 534 // constant section is big enough, use it. Otherwise free the 535 // current and allocate a new one. 536 BufferBlob* blob = scratch_buffer_blob(); 537 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 538 // Use the current blob. 539 } else { 540 if (blob != NULL) { 541 BufferBlob::free(blob); 542 } 543 544 ResourceMark rm; 545 _scratch_const_size = const_size; 546 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 547 blob = BufferBlob::create("Compile::scratch_buffer", size); 548 // Record the buffer blob for next time. 549 set_scratch_buffer_blob(blob); 550 // Have we run out of code space? 551 if (scratch_buffer_blob() == NULL) { 552 // Let CompilerBroker disable further compilations. 553 record_failure("Not enough space for scratch buffer in CodeCache"); 554 return; 555 } 556 } 557 558 // Initialize the relocation buffers 559 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 560 set_scratch_locs_memory(locs_buf); 561 } 562 563 564 //-----------------------scratch_emit_size------------------------------------- 565 // Helper function that computes size by emitting code 566 uint Compile::scratch_emit_size(const Node* n) { 567 // Start scratch_emit_size section. 568 set_in_scratch_emit_size(true); 569 570 // Emit into a trash buffer and count bytes emitted. 571 // This is a pretty expensive way to compute a size, 572 // but it works well enough if seldom used. 573 // All common fixed-size instructions are given a size 574 // method by the AD file. 575 // Note that the scratch buffer blob and locs memory are 576 // allocated at the beginning of the compile task, and 577 // may be shared by several calls to scratch_emit_size. 578 // The allocation of the scratch buffer blob is particularly 579 // expensive, since it has to grab the code cache lock. 580 BufferBlob* blob = this->scratch_buffer_blob(); 581 assert(blob != NULL, "Initialize BufferBlob at start"); 582 assert(blob->size() > MAX_inst_size, "sanity"); 583 relocInfo* locs_buf = scratch_locs_memory(); 584 address blob_begin = blob->content_begin(); 585 address blob_end = (address)locs_buf; 586 assert(blob->content_contains(blob_end), "sanity"); 587 CodeBuffer buf(blob_begin, blob_end - blob_begin); 588 buf.initialize_consts_size(_scratch_const_size); 589 buf.initialize_stubs_size(MAX_stubs_size); 590 assert(locs_buf != NULL, "sanity"); 591 int lsize = MAX_locs_size / 3; 592 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 593 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 594 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 595 596 // Do the emission. 597 598 Label fakeL; // Fake label for branch instructions. 599 Label* saveL = NULL; 600 uint save_bnum = 0; 601 bool is_branch = n->is_MachBranch(); 602 if (is_branch) { 603 MacroAssembler masm(&buf); 604 masm.bind(fakeL); 605 n->as_MachBranch()->save_label(&saveL, &save_bnum); 606 n->as_MachBranch()->label_set(&fakeL, 0); 607 } 608 n->emit(buf, this->regalloc()); 609 if (is_branch) // Restore label. 610 n->as_MachBranch()->label_set(saveL, save_bnum); 611 612 // End scratch_emit_size section. 613 set_in_scratch_emit_size(false); 614 615 return buf.insts_size(); 616 } 617 618 619 // ============================================================================ 620 //------------------------------Compile standard------------------------------- 621 debug_only( int Compile::_debug_idx = 100000; ) 622 623 // Compile a method. entry_bci is -1 for normal compilations and indicates 624 // the continuation bci for on stack replacement. 625 626 627 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 628 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 629 : Phase(Compiler), 630 _env(ci_env), 631 _log(ci_env->log()), 632 _compile_id(ci_env->compile_id()), 633 _save_argument_registers(false), 634 _stub_name(NULL), 635 _stub_function(NULL), 636 _stub_entry_point(NULL), 637 _method(target), 638 _entry_bci(osr_bci), 639 _initial_gvn(NULL), 640 _for_igvn(NULL), 641 _warm_calls(NULL), 642 _subsume_loads(subsume_loads), 643 _do_escape_analysis(do_escape_analysis), 644 _eliminate_boxing(eliminate_boxing), 645 _failure_reason(NULL), 646 _code_buffer("Compile::Fill_buffer"), 647 _orig_pc_slot(0), 648 _orig_pc_slot_offset_in_bytes(0), 649 _has_method_handle_invokes(false), 650 _mach_constant_base_node(NULL), 651 _node_bundling_limit(0), 652 _node_bundling_base(NULL), 653 _java_calls(0), 654 _inner_loops(0), 655 _scratch_const_size(-1), 656 _in_scratch_emit_size(false), 657 _dead_node_list(comp_arena()), 658 _dead_node_count(0), 659 #ifndef PRODUCT 660 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 661 _in_dump_cnt(0), 662 _printer(IdealGraphPrinter::printer()), 663 #endif 664 _congraph(NULL), 665 _replay_inline_data(NULL), 666 _late_inlines(comp_arena(), 2, 0, NULL), 667 _string_late_inlines(comp_arena(), 2, 0, NULL), 668 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 669 _late_inlines_pos(0), 670 _number_of_mh_late_inlines(0), 671 _inlining_progress(false), 672 _inlining_incrementally(false), 673 _print_inlining_list(NULL), 674 _print_inlining_idx(0), 675 _preserve_jvm_state(0), 676 _interpreter_frame_size(0) { 677 C = this; 678 679 CompileWrapper cw(this); 680 #ifndef PRODUCT 681 if (TimeCompiler2) { 682 tty->print(" "); 683 target->holder()->name()->print(); 684 tty->print("."); 685 target->print_short_name(); 686 tty->print(" "); 687 } 688 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 689 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 690 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 691 if (!print_opto_assembly) { 692 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 693 if (print_assembly && !Disassembler::can_decode()) { 694 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 695 print_opto_assembly = true; 696 } 697 } 698 set_print_assembly(print_opto_assembly); 699 set_parsed_irreducible_loop(false); 700 701 if (method()->has_option("ReplayInline")) { 702 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 703 } 704 #endif 705 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 706 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 707 708 if (ProfileTraps) { 709 // Make sure the method being compiled gets its own MDO, 710 // so we can at least track the decompile_count(). 711 method()->ensure_method_data(); 712 } 713 714 Init(::AliasLevel); 715 716 717 print_compile_messages(); 718 719 _ilt = InlineTree::build_inline_tree_root(); 720 721 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 722 assert(num_alias_types() >= AliasIdxRaw, ""); 723 724 #define MINIMUM_NODE_HASH 1023 725 // Node list that Iterative GVN will start with 726 Unique_Node_List for_igvn(comp_arena()); 727 set_for_igvn(&for_igvn); 728 729 // GVN that will be run immediately on new nodes 730 uint estimated_size = method()->code_size()*4+64; 731 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 732 PhaseGVN gvn(node_arena(), estimated_size); 733 set_initial_gvn(&gvn); 734 735 if (print_inlining() || print_intrinsics()) { 736 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 737 } 738 { // Scope for timing the parser 739 TracePhase t3("parse", &_t_parser, true); 740 741 // Put top into the hash table ASAP. 742 initial_gvn()->transform_no_reclaim(top()); 743 744 // Set up tf(), start(), and find a CallGenerator. 745 CallGenerator* cg = NULL; 746 if (is_osr_compilation()) { 747 const TypeTuple *domain = StartOSRNode::osr_domain(); 748 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 749 init_tf(TypeFunc::make(domain, range)); 750 StartNode* s = new (this) StartOSRNode(root(), domain); 751 initial_gvn()->set_type_bottom(s); 752 init_start(s); 753 cg = CallGenerator::for_osr(method(), entry_bci()); 754 } else { 755 // Normal case. 756 init_tf(TypeFunc::make(method())); 757 StartNode* s = new (this) StartNode(root(), tf()->domain()); 758 initial_gvn()->set_type_bottom(s); 759 init_start(s); 760 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 761 // With java.lang.ref.reference.get() we must go through the 762 // intrinsic when G1 is enabled - even when get() is the root 763 // method of the compile - so that, if necessary, the value in 764 // the referent field of the reference object gets recorded by 765 // the pre-barrier code. 766 // Specifically, if G1 is enabled, the value in the referent 767 // field is recorded by the G1 SATB pre barrier. This will 768 // result in the referent being marked live and the reference 769 // object removed from the list of discovered references during 770 // reference processing. 771 cg = find_intrinsic(method(), false); 772 } 773 if (cg == NULL) { 774 float past_uses = method()->interpreter_invocation_count(); 775 float expected_uses = past_uses; 776 cg = CallGenerator::for_inline(method(), expected_uses); 777 } 778 } 779 if (failing()) return; 780 if (cg == NULL) { 781 record_method_not_compilable_all_tiers("cannot parse method"); 782 return; 783 } 784 JVMState* jvms = build_start_state(start(), tf()); 785 if ((jvms = cg->generate(jvms, NULL)) == NULL) { 786 record_method_not_compilable("method parse failed"); 787 return; 788 } 789 GraphKit kit(jvms); 790 791 if (!kit.stopped()) { 792 // Accept return values, and transfer control we know not where. 793 // This is done by a special, unique ReturnNode bound to root. 794 return_values(kit.jvms()); 795 } 796 797 if (kit.has_exceptions()) { 798 // Any exceptions that escape from this call must be rethrown 799 // to whatever caller is dynamically above us on the stack. 800 // This is done by a special, unique RethrowNode bound to root. 801 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 802 } 803 804 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 805 806 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 807 inline_string_calls(true); 808 } 809 810 if (failing()) return; 811 812 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 813 814 // Remove clutter produced by parsing. 815 if (!failing()) { 816 ResourceMark rm; 817 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 818 } 819 } 820 821 // Note: Large methods are capped off in do_one_bytecode(). 822 if (failing()) return; 823 824 // After parsing, node notes are no longer automagic. 825 // They must be propagated by register_new_node_with_optimizer(), 826 // clone(), or the like. 827 set_default_node_notes(NULL); 828 829 for (;;) { 830 int successes = Inline_Warm(); 831 if (failing()) return; 832 if (successes == 0) break; 833 } 834 835 // Drain the list. 836 Finish_Warm(); 837 #ifndef PRODUCT 838 if (_printer) { 839 _printer->print_inlining(this); 840 } 841 #endif 842 843 if (failing()) return; 844 NOT_PRODUCT( verify_graph_edges(); ) 845 846 // Now optimize 847 Optimize(); 848 if (failing()) return; 849 NOT_PRODUCT( verify_graph_edges(); ) 850 851 #ifndef PRODUCT 852 if (PrintIdeal) { 853 ttyLocker ttyl; // keep the following output all in one block 854 // This output goes directly to the tty, not the compiler log. 855 // To enable tools to match it up with the compilation activity, 856 // be sure to tag this tty output with the compile ID. 857 if (xtty != NULL) { 858 xtty->head("ideal compile_id='%d'%s", compile_id(), 859 is_osr_compilation() ? " compile_kind='osr'" : 860 ""); 861 } 862 root()->dump(9999); 863 if (xtty != NULL) { 864 xtty->tail("ideal"); 865 } 866 } 867 #endif 868 869 NOT_PRODUCT( verify_barriers(); ) 870 871 // Dump compilation data to replay it. 872 if (method()->has_option("DumpReplay")) { 873 env()->dump_replay_data(_compile_id); 874 } 875 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 876 env()->dump_inline_data(_compile_id); 877 } 878 879 // Now that we know the size of all the monitors we can add a fixed slot 880 // for the original deopt pc. 881 882 _orig_pc_slot = fixed_slots(); 883 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 884 set_fixed_slots(next_slot); 885 886 // Compute when to use implicit null checks. Used by matching trap based 887 // nodes and NullCheck optimization. 888 set_allowed_deopt_reasons(); 889 890 // Now generate code 891 Code_Gen(); 892 if (failing()) return; 893 894 // Check if we want to skip execution of all compiled code. 895 { 896 #ifndef PRODUCT 897 if (OptoNoExecute) { 898 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 899 return; 900 } 901 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 902 #endif 903 904 if (is_osr_compilation()) { 905 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 906 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 907 } else { 908 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 909 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 910 } 911 912 env()->register_method(_method, _entry_bci, 913 &_code_offsets, 914 _orig_pc_slot_offset_in_bytes, 915 code_buffer(), 916 frame_size_in_words(), _oop_map_set, 917 &_handler_table, &_inc_table, 918 compiler, 919 env()->comp_level(), 920 has_unsafe_access(), 921 SharedRuntime::is_wide_vector(max_vector_size()) 922 ); 923 924 if (log() != NULL) // Print code cache state into compiler log 925 log()->code_cache_state(); 926 } 927 } 928 929 //------------------------------Compile---------------------------------------- 930 // Compile a runtime stub 931 Compile::Compile( ciEnv* ci_env, 932 TypeFunc_generator generator, 933 address stub_function, 934 const char *stub_name, 935 int is_fancy_jump, 936 bool pass_tls, 937 bool save_arg_registers, 938 bool return_pc ) 939 : Phase(Compiler), 940 _env(ci_env), 941 _log(ci_env->log()), 942 _compile_id(0), 943 _save_argument_registers(save_arg_registers), 944 _method(NULL), 945 _stub_name(stub_name), 946 _stub_function(stub_function), 947 _stub_entry_point(NULL), 948 _entry_bci(InvocationEntryBci), 949 _initial_gvn(NULL), 950 _for_igvn(NULL), 951 _warm_calls(NULL), 952 _orig_pc_slot(0), 953 _orig_pc_slot_offset_in_bytes(0), 954 _subsume_loads(true), 955 _do_escape_analysis(false), 956 _eliminate_boxing(false), 957 _failure_reason(NULL), 958 _code_buffer("Compile::Fill_buffer"), 959 _has_method_handle_invokes(false), 960 _mach_constant_base_node(NULL), 961 _node_bundling_limit(0), 962 _node_bundling_base(NULL), 963 _java_calls(0), 964 _inner_loops(0), 965 #ifndef PRODUCT 966 _trace_opto_output(TraceOptoOutput), 967 _in_dump_cnt(0), 968 _printer(NULL), 969 #endif 970 _dead_node_list(comp_arena()), 971 _dead_node_count(0), 972 _congraph(NULL), 973 _replay_inline_data(NULL), 974 _number_of_mh_late_inlines(0), 975 _inlining_progress(false), 976 _inlining_incrementally(false), 977 _print_inlining_list(NULL), 978 _print_inlining_idx(0), 979 _preserve_jvm_state(0), 980 _allowed_reasons(0), 981 _interpreter_frame_size(0) { 982 C = this; 983 984 #ifndef PRODUCT 985 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 986 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 987 set_print_assembly(PrintFrameConverterAssembly); 988 set_parsed_irreducible_loop(false); 989 #endif 990 CompileWrapper cw(this); 991 Init(/*AliasLevel=*/ 0); 992 init_tf((*generator)()); 993 994 { 995 // The following is a dummy for the sake of GraphKit::gen_stub 996 Unique_Node_List for_igvn(comp_arena()); 997 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 998 PhaseGVN gvn(Thread::current()->resource_area(),255); 999 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1000 gvn.transform_no_reclaim(top()); 1001 1002 GraphKit kit; 1003 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1004 } 1005 1006 NOT_PRODUCT( verify_graph_edges(); ) 1007 Code_Gen(); 1008 if (failing()) return; 1009 1010 1011 // Entry point will be accessed using compile->stub_entry_point(); 1012 if (code_buffer() == NULL) { 1013 Matcher::soft_match_failure(); 1014 } else { 1015 if (PrintAssembly && (WizardMode || Verbose)) 1016 tty->print_cr("### Stub::%s", stub_name); 1017 1018 if (!failing()) { 1019 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1020 1021 // Make the NMethod 1022 // For now we mark the frame as never safe for profile stackwalking 1023 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1024 code_buffer(), 1025 CodeOffsets::frame_never_safe, 1026 // _code_offsets.value(CodeOffsets::Frame_Complete), 1027 frame_size_in_words(), 1028 _oop_map_set, 1029 save_arg_registers); 1030 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1031 1032 _stub_entry_point = rs->entry_point(); 1033 } 1034 } 1035 } 1036 1037 //------------------------------Init------------------------------------------- 1038 // Prepare for a single compilation 1039 void Compile::Init(int aliaslevel) { 1040 _unique = 0; 1041 _regalloc = NULL; 1042 1043 _tf = NULL; // filled in later 1044 _top = NULL; // cached later 1045 _matcher = NULL; // filled in later 1046 _cfg = NULL; // filled in later 1047 1048 set_24_bit_selection_and_mode(Use24BitFP, false); 1049 1050 _node_note_array = NULL; 1051 _default_node_notes = NULL; 1052 1053 _immutable_memory = NULL; // filled in at first inquiry 1054 1055 // Globally visible Nodes 1056 // First set TOP to NULL to give safe behavior during creation of RootNode 1057 set_cached_top_node(NULL); 1058 set_root(new (this) RootNode()); 1059 // Now that you have a Root to point to, create the real TOP 1060 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1061 set_recent_alloc(NULL, NULL); 1062 1063 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1064 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1065 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1066 env()->set_dependencies(new Dependencies(env())); 1067 1068 _fixed_slots = 0; 1069 set_has_split_ifs(false); 1070 set_has_loops(has_method() && method()->has_loops()); // first approximation 1071 set_has_stringbuilder(false); 1072 set_has_boxed_value(false); 1073 _trap_can_recompile = false; // no traps emitted yet 1074 _major_progress = true; // start out assuming good things will happen 1075 set_has_unsafe_access(false); 1076 set_max_vector_size(0); 1077 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1078 set_decompile_count(0); 1079 1080 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1081 set_num_loop_opts(LoopOptsCount); 1082 set_do_inlining(Inline); 1083 set_max_inline_size(MaxInlineSize); 1084 set_freq_inline_size(FreqInlineSize); 1085 set_do_scheduling(OptoScheduling); 1086 set_do_count_invocations(false); 1087 set_do_method_data_update(false); 1088 1089 if (debug_info()->recording_non_safepoints()) { 1090 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1091 (comp_arena(), 8, 0, NULL)); 1092 set_default_node_notes(Node_Notes::make(this)); 1093 } 1094 1095 // // -- Initialize types before each compile -- 1096 // // Update cached type information 1097 // if( _method && _method->constants() ) 1098 // Type::update_loaded_types(_method, _method->constants()); 1099 1100 // Init alias_type map. 1101 if (!_do_escape_analysis && aliaslevel == 3) 1102 aliaslevel = 2; // No unique types without escape analysis 1103 _AliasLevel = aliaslevel; 1104 const int grow_ats = 16; 1105 _max_alias_types = grow_ats; 1106 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1107 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1108 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1109 { 1110 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1111 } 1112 // Initialize the first few types. 1113 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1114 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1115 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1116 _num_alias_types = AliasIdxRaw+1; 1117 // Zero out the alias type cache. 1118 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1119 // A NULL adr_type hits in the cache right away. Preload the right answer. 1120 probe_alias_cache(NULL)->_index = AliasIdxTop; 1121 1122 _intrinsics = NULL; 1123 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1124 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1125 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1126 register_library_intrinsics(); 1127 } 1128 1129 //---------------------------init_start---------------------------------------- 1130 // Install the StartNode on this compile object. 1131 void Compile::init_start(StartNode* s) { 1132 if (failing()) 1133 return; // already failing 1134 assert(s == start(), ""); 1135 } 1136 1137 StartNode* Compile::start() const { 1138 assert(!failing(), ""); 1139 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1140 Node* start = root()->fast_out(i); 1141 if( start->is_Start() ) 1142 return start->as_Start(); 1143 } 1144 ShouldNotReachHere(); 1145 return NULL; 1146 } 1147 1148 //-------------------------------immutable_memory------------------------------------- 1149 // Access immutable memory 1150 Node* Compile::immutable_memory() { 1151 if (_immutable_memory != NULL) { 1152 return _immutable_memory; 1153 } 1154 StartNode* s = start(); 1155 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1156 Node *p = s->fast_out(i); 1157 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1158 _immutable_memory = p; 1159 return _immutable_memory; 1160 } 1161 } 1162 ShouldNotReachHere(); 1163 return NULL; 1164 } 1165 1166 //----------------------set_cached_top_node------------------------------------ 1167 // Install the cached top node, and make sure Node::is_top works correctly. 1168 void Compile::set_cached_top_node(Node* tn) { 1169 if (tn != NULL) verify_top(tn); 1170 Node* old_top = _top; 1171 _top = tn; 1172 // Calling Node::setup_is_top allows the nodes the chance to adjust 1173 // their _out arrays. 1174 if (_top != NULL) _top->setup_is_top(); 1175 if (old_top != NULL) old_top->setup_is_top(); 1176 assert(_top == NULL || top()->is_top(), ""); 1177 } 1178 1179 #ifdef ASSERT 1180 uint Compile::count_live_nodes_by_graph_walk() { 1181 Unique_Node_List useful(comp_arena()); 1182 // Get useful node list by walking the graph. 1183 identify_useful_nodes(useful); 1184 return useful.size(); 1185 } 1186 1187 void Compile::print_missing_nodes() { 1188 1189 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1190 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1191 return; 1192 } 1193 1194 // This is an expensive function. It is executed only when the user 1195 // specifies VerifyIdealNodeCount option or otherwise knows the 1196 // additional work that needs to be done to identify reachable nodes 1197 // by walking the flow graph and find the missing ones using 1198 // _dead_node_list. 1199 1200 Unique_Node_List useful(comp_arena()); 1201 // Get useful node list by walking the graph. 1202 identify_useful_nodes(useful); 1203 1204 uint l_nodes = C->live_nodes(); 1205 uint l_nodes_by_walk = useful.size(); 1206 1207 if (l_nodes != l_nodes_by_walk) { 1208 if (_log != NULL) { 1209 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1210 _log->stamp(); 1211 _log->end_head(); 1212 } 1213 VectorSet& useful_member_set = useful.member_set(); 1214 int last_idx = l_nodes_by_walk; 1215 for (int i = 0; i < last_idx; i++) { 1216 if (useful_member_set.test(i)) { 1217 if (_dead_node_list.test(i)) { 1218 if (_log != NULL) { 1219 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1220 } 1221 if (PrintIdealNodeCount) { 1222 // Print the log message to tty 1223 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1224 useful.at(i)->dump(); 1225 } 1226 } 1227 } 1228 else if (! _dead_node_list.test(i)) { 1229 if (_log != NULL) { 1230 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1231 } 1232 if (PrintIdealNodeCount) { 1233 // Print the log message to tty 1234 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1235 } 1236 } 1237 } 1238 if (_log != NULL) { 1239 _log->tail("mismatched_nodes"); 1240 } 1241 } 1242 } 1243 #endif 1244 1245 #ifndef PRODUCT 1246 void Compile::verify_top(Node* tn) const { 1247 if (tn != NULL) { 1248 assert(tn->is_Con(), "top node must be a constant"); 1249 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1250 assert(tn->in(0) != NULL, "must have live top node"); 1251 } 1252 } 1253 #endif 1254 1255 1256 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1257 1258 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1259 guarantee(arr != NULL, ""); 1260 int num_blocks = arr->length(); 1261 if (grow_by < num_blocks) grow_by = num_blocks; 1262 int num_notes = grow_by * _node_notes_block_size; 1263 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1264 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1265 while (num_notes > 0) { 1266 arr->append(notes); 1267 notes += _node_notes_block_size; 1268 num_notes -= _node_notes_block_size; 1269 } 1270 assert(num_notes == 0, "exact multiple, please"); 1271 } 1272 1273 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1274 if (source == NULL || dest == NULL) return false; 1275 1276 if (dest->is_Con()) 1277 return false; // Do not push debug info onto constants. 1278 1279 #ifdef ASSERT 1280 // Leave a bread crumb trail pointing to the original node: 1281 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1282 dest->set_debug_orig(source); 1283 } 1284 #endif 1285 1286 if (node_note_array() == NULL) 1287 return false; // Not collecting any notes now. 1288 1289 // This is a copy onto a pre-existing node, which may already have notes. 1290 // If both nodes have notes, do not overwrite any pre-existing notes. 1291 Node_Notes* source_notes = node_notes_at(source->_idx); 1292 if (source_notes == NULL || source_notes->is_clear()) return false; 1293 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1294 if (dest_notes == NULL || dest_notes->is_clear()) { 1295 return set_node_notes_at(dest->_idx, source_notes); 1296 } 1297 1298 Node_Notes merged_notes = (*source_notes); 1299 // The order of operations here ensures that dest notes will win... 1300 merged_notes.update_from(dest_notes); 1301 return set_node_notes_at(dest->_idx, &merged_notes); 1302 } 1303 1304 1305 //--------------------------allow_range_check_smearing------------------------- 1306 // Gating condition for coalescing similar range checks. 1307 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1308 // single covering check that is at least as strong as any of them. 1309 // If the optimization succeeds, the simplified (strengthened) range check 1310 // will always succeed. If it fails, we will deopt, and then give up 1311 // on the optimization. 1312 bool Compile::allow_range_check_smearing() const { 1313 // If this method has already thrown a range-check, 1314 // assume it was because we already tried range smearing 1315 // and it failed. 1316 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1317 return !already_trapped; 1318 } 1319 1320 1321 //------------------------------flatten_alias_type----------------------------- 1322 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1323 int offset = tj->offset(); 1324 TypePtr::PTR ptr = tj->ptr(); 1325 1326 // Known instance (scalarizable allocation) alias only with itself. 1327 bool is_known_inst = tj->isa_oopptr() != NULL && 1328 tj->is_oopptr()->is_known_instance(); 1329 1330 // Process weird unsafe references. 1331 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1332 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1333 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1334 tj = TypeOopPtr::BOTTOM; 1335 ptr = tj->ptr(); 1336 offset = tj->offset(); 1337 } 1338 1339 // Array pointers need some flattening 1340 const TypeAryPtr *ta = tj->isa_aryptr(); 1341 if (ta && ta->is_stable()) { 1342 // Erase stability property for alias analysis. 1343 tj = ta = ta->cast_to_stable(false); 1344 } 1345 if( ta && is_known_inst ) { 1346 if ( offset != Type::OffsetBot && 1347 offset > arrayOopDesc::length_offset_in_bytes() ) { 1348 offset = Type::OffsetBot; // Flatten constant access into array body only 1349 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1350 } 1351 } else if( ta && _AliasLevel >= 2 ) { 1352 // For arrays indexed by constant indices, we flatten the alias 1353 // space to include all of the array body. Only the header, klass 1354 // and array length can be accessed un-aliased. 1355 if( offset != Type::OffsetBot ) { 1356 if( ta->const_oop() ) { // MethodData* or Method* 1357 offset = Type::OffsetBot; // Flatten constant access into array body 1358 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1359 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1360 // range is OK as-is. 1361 tj = ta = TypeAryPtr::RANGE; 1362 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1363 tj = TypeInstPtr::KLASS; // all klass loads look alike 1364 ta = TypeAryPtr::RANGE; // generic ignored junk 1365 ptr = TypePtr::BotPTR; 1366 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1367 tj = TypeInstPtr::MARK; 1368 ta = TypeAryPtr::RANGE; // generic ignored junk 1369 ptr = TypePtr::BotPTR; 1370 } else { // Random constant offset into array body 1371 offset = Type::OffsetBot; // Flatten constant access into array body 1372 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1373 } 1374 } 1375 // Arrays of fixed size alias with arrays of unknown size. 1376 if (ta->size() != TypeInt::POS) { 1377 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1378 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1379 } 1380 // Arrays of known objects become arrays of unknown objects. 1381 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1382 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1383 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1384 } 1385 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1386 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1387 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1388 } 1389 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1390 // cannot be distinguished by bytecode alone. 1391 if (ta->elem() == TypeInt::BOOL) { 1392 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1393 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1394 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1395 } 1396 // During the 2nd round of IterGVN, NotNull castings are removed. 1397 // Make sure the Bottom and NotNull variants alias the same. 1398 // Also, make sure exact and non-exact variants alias the same. 1399 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1400 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1401 } 1402 } 1403 1404 // Oop pointers need some flattening 1405 const TypeInstPtr *to = tj->isa_instptr(); 1406 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1407 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1408 if( ptr == TypePtr::Constant ) { 1409 if (to->klass() != ciEnv::current()->Class_klass() || 1410 offset < k->size_helper() * wordSize) { 1411 // No constant oop pointers (such as Strings); they alias with 1412 // unknown strings. 1413 assert(!is_known_inst, "not scalarizable allocation"); 1414 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1415 } 1416 } else if( is_known_inst ) { 1417 tj = to; // Keep NotNull and klass_is_exact for instance type 1418 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1419 // During the 2nd round of IterGVN, NotNull castings are removed. 1420 // Make sure the Bottom and NotNull variants alias the same. 1421 // Also, make sure exact and non-exact variants alias the same. 1422 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1423 } 1424 if (to->speculative() != NULL) { 1425 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1426 } 1427 // Canonicalize the holder of this field 1428 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1429 // First handle header references such as a LoadKlassNode, even if the 1430 // object's klass is unloaded at compile time (4965979). 1431 if (!is_known_inst) { // Do it only for non-instance types 1432 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1433 } 1434 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1435 // Static fields are in the space above the normal instance 1436 // fields in the java.lang.Class instance. 1437 if (to->klass() != ciEnv::current()->Class_klass()) { 1438 to = NULL; 1439 tj = TypeOopPtr::BOTTOM; 1440 offset = tj->offset(); 1441 } 1442 } else { 1443 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1444 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1445 if( is_known_inst ) { 1446 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1447 } else { 1448 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1449 } 1450 } 1451 } 1452 } 1453 1454 // Klass pointers to object array klasses need some flattening 1455 const TypeKlassPtr *tk = tj->isa_klassptr(); 1456 if( tk ) { 1457 // If we are referencing a field within a Klass, we need 1458 // to assume the worst case of an Object. Both exact and 1459 // inexact types must flatten to the same alias class so 1460 // use NotNull as the PTR. 1461 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1462 1463 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1464 TypeKlassPtr::OBJECT->klass(), 1465 offset); 1466 } 1467 1468 ciKlass* klass = tk->klass(); 1469 if( klass->is_obj_array_klass() ) { 1470 ciKlass* k = TypeAryPtr::OOPS->klass(); 1471 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1472 k = TypeInstPtr::BOTTOM->klass(); 1473 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1474 } 1475 1476 // Check for precise loads from the primary supertype array and force them 1477 // to the supertype cache alias index. Check for generic array loads from 1478 // the primary supertype array and also force them to the supertype cache 1479 // alias index. Since the same load can reach both, we need to merge 1480 // these 2 disparate memories into the same alias class. Since the 1481 // primary supertype array is read-only, there's no chance of confusion 1482 // where we bypass an array load and an array store. 1483 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1484 if (offset == Type::OffsetBot || 1485 (offset >= primary_supers_offset && 1486 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1487 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1488 offset = in_bytes(Klass::secondary_super_cache_offset()); 1489 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1490 } 1491 } 1492 1493 // Flatten all Raw pointers together. 1494 if (tj->base() == Type::RawPtr) 1495 tj = TypeRawPtr::BOTTOM; 1496 1497 if (tj->base() == Type::AnyPtr) 1498 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1499 1500 // Flatten all to bottom for now 1501 switch( _AliasLevel ) { 1502 case 0: 1503 tj = TypePtr::BOTTOM; 1504 break; 1505 case 1: // Flatten to: oop, static, field or array 1506 switch (tj->base()) { 1507 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1508 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1509 case Type::AryPtr: // do not distinguish arrays at all 1510 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1511 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1512 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1513 default: ShouldNotReachHere(); 1514 } 1515 break; 1516 case 2: // No collapsing at level 2; keep all splits 1517 case 3: // No collapsing at level 3; keep all splits 1518 break; 1519 default: 1520 Unimplemented(); 1521 } 1522 1523 offset = tj->offset(); 1524 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1525 1526 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1527 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1528 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1529 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1530 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1531 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1532 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1533 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1534 assert( tj->ptr() != TypePtr::TopPTR && 1535 tj->ptr() != TypePtr::AnyNull && 1536 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1537 // assert( tj->ptr() != TypePtr::Constant || 1538 // tj->base() == Type::RawPtr || 1539 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1540 1541 return tj; 1542 } 1543 1544 void Compile::AliasType::Init(int i, const TypePtr* at) { 1545 _index = i; 1546 _adr_type = at; 1547 _field = NULL; 1548 _element = NULL; 1549 _is_rewritable = true; // default 1550 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1551 if (atoop != NULL && atoop->is_known_instance()) { 1552 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1553 _general_index = Compile::current()->get_alias_index(gt); 1554 } else { 1555 _general_index = 0; 1556 } 1557 } 1558 1559 //---------------------------------print_on------------------------------------ 1560 #ifndef PRODUCT 1561 void Compile::AliasType::print_on(outputStream* st) { 1562 if (index() < 10) 1563 st->print("@ <%d> ", index()); 1564 else st->print("@ <%d>", index()); 1565 st->print(is_rewritable() ? " " : " RO"); 1566 int offset = adr_type()->offset(); 1567 if (offset == Type::OffsetBot) 1568 st->print(" +any"); 1569 else st->print(" +%-3d", offset); 1570 st->print(" in "); 1571 adr_type()->dump_on(st); 1572 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1573 if (field() != NULL && tjp) { 1574 if (tjp->klass() != field()->holder() || 1575 tjp->offset() != field()->offset_in_bytes()) { 1576 st->print(" != "); 1577 field()->print(); 1578 st->print(" ***"); 1579 } 1580 } 1581 } 1582 1583 void print_alias_types() { 1584 Compile* C = Compile::current(); 1585 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1586 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1587 C->alias_type(idx)->print_on(tty); 1588 tty->cr(); 1589 } 1590 } 1591 #endif 1592 1593 1594 //----------------------------probe_alias_cache-------------------------------- 1595 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1596 intptr_t key = (intptr_t) adr_type; 1597 key ^= key >> logAliasCacheSize; 1598 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1599 } 1600 1601 1602 //-----------------------------grow_alias_types-------------------------------- 1603 void Compile::grow_alias_types() { 1604 const int old_ats = _max_alias_types; // how many before? 1605 const int new_ats = old_ats; // how many more? 1606 const int grow_ats = old_ats+new_ats; // how many now? 1607 _max_alias_types = grow_ats; 1608 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1609 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1610 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1611 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1612 } 1613 1614 1615 //--------------------------------find_alias_type------------------------------ 1616 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1617 if (_AliasLevel == 0) 1618 return alias_type(AliasIdxBot); 1619 1620 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1621 if (ace->_adr_type == adr_type) { 1622 return alias_type(ace->_index); 1623 } 1624 1625 // Handle special cases. 1626 if (adr_type == NULL) return alias_type(AliasIdxTop); 1627 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1628 1629 // Do it the slow way. 1630 const TypePtr* flat = flatten_alias_type(adr_type); 1631 1632 #ifdef ASSERT 1633 assert(flat == flatten_alias_type(flat), "idempotent"); 1634 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1635 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1636 const TypeOopPtr* foop = flat->is_oopptr(); 1637 // Scalarizable allocations have exact klass always. 1638 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1639 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1640 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1641 } 1642 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1643 #endif 1644 1645 int idx = AliasIdxTop; 1646 for (int i = 0; i < num_alias_types(); i++) { 1647 if (alias_type(i)->adr_type() == flat) { 1648 idx = i; 1649 break; 1650 } 1651 } 1652 1653 if (idx == AliasIdxTop) { 1654 if (no_create) return NULL; 1655 // Grow the array if necessary. 1656 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1657 // Add a new alias type. 1658 idx = _num_alias_types++; 1659 _alias_types[idx]->Init(idx, flat); 1660 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1661 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1662 if (flat->isa_instptr()) { 1663 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1664 && flat->is_instptr()->klass() == env()->Class_klass()) 1665 alias_type(idx)->set_rewritable(false); 1666 } 1667 if (flat->isa_aryptr()) { 1668 #ifdef ASSERT 1669 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1670 // (T_BYTE has the weakest alignment and size restrictions...) 1671 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1672 #endif 1673 if (flat->offset() == TypePtr::OffsetBot) { 1674 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1675 } 1676 } 1677 if (flat->isa_klassptr()) { 1678 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1679 alias_type(idx)->set_rewritable(false); 1680 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1681 alias_type(idx)->set_rewritable(false); 1682 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1683 alias_type(idx)->set_rewritable(false); 1684 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1685 alias_type(idx)->set_rewritable(false); 1686 } 1687 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1688 // but the base pointer type is not distinctive enough to identify 1689 // references into JavaThread.) 1690 1691 // Check for final fields. 1692 const TypeInstPtr* tinst = flat->isa_instptr(); 1693 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1694 ciField* field; 1695 if (tinst->const_oop() != NULL && 1696 tinst->klass() == ciEnv::current()->Class_klass() && 1697 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1698 // static field 1699 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1700 field = k->get_field_by_offset(tinst->offset(), true); 1701 } else { 1702 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1703 field = k->get_field_by_offset(tinst->offset(), false); 1704 } 1705 assert(field == NULL || 1706 original_field == NULL || 1707 (field->holder() == original_field->holder() && 1708 field->offset() == original_field->offset() && 1709 field->is_static() == original_field->is_static()), "wrong field?"); 1710 // Set field() and is_rewritable() attributes. 1711 if (field != NULL) alias_type(idx)->set_field(field); 1712 } 1713 } 1714 1715 // Fill the cache for next time. 1716 ace->_adr_type = adr_type; 1717 ace->_index = idx; 1718 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1719 1720 // Might as well try to fill the cache for the flattened version, too. 1721 AliasCacheEntry* face = probe_alias_cache(flat); 1722 if (face->_adr_type == NULL) { 1723 face->_adr_type = flat; 1724 face->_index = idx; 1725 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1726 } 1727 1728 return alias_type(idx); 1729 } 1730 1731 1732 Compile::AliasType* Compile::alias_type(ciField* field) { 1733 const TypeOopPtr* t; 1734 if (field->is_static()) 1735 t = TypeInstPtr::make(field->holder()->java_mirror()); 1736 else 1737 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1738 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1739 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1740 return atp; 1741 } 1742 1743 1744 //------------------------------have_alias_type-------------------------------- 1745 bool Compile::have_alias_type(const TypePtr* adr_type) { 1746 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1747 if (ace->_adr_type == adr_type) { 1748 return true; 1749 } 1750 1751 // Handle special cases. 1752 if (adr_type == NULL) return true; 1753 if (adr_type == TypePtr::BOTTOM) return true; 1754 1755 return find_alias_type(adr_type, true, NULL) != NULL; 1756 } 1757 1758 //-----------------------------must_alias-------------------------------------- 1759 // True if all values of the given address type are in the given alias category. 1760 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1761 if (alias_idx == AliasIdxBot) return true; // the universal category 1762 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1763 if (alias_idx == AliasIdxTop) return false; // the empty category 1764 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1765 1766 // the only remaining possible overlap is identity 1767 int adr_idx = get_alias_index(adr_type); 1768 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1769 assert(adr_idx == alias_idx || 1770 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1771 && adr_type != TypeOopPtr::BOTTOM), 1772 "should not be testing for overlap with an unsafe pointer"); 1773 return adr_idx == alias_idx; 1774 } 1775 1776 //------------------------------can_alias-------------------------------------- 1777 // True if any values of the given address type are in the given alias category. 1778 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1779 if (alias_idx == AliasIdxTop) return false; // the empty category 1780 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1781 if (alias_idx == AliasIdxBot) return true; // the universal category 1782 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1783 1784 // the only remaining possible overlap is identity 1785 int adr_idx = get_alias_index(adr_type); 1786 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1787 return adr_idx == alias_idx; 1788 } 1789 1790 1791 1792 //---------------------------pop_warm_call------------------------------------- 1793 WarmCallInfo* Compile::pop_warm_call() { 1794 WarmCallInfo* wci = _warm_calls; 1795 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1796 return wci; 1797 } 1798 1799 //----------------------------Inline_Warm-------------------------------------- 1800 int Compile::Inline_Warm() { 1801 // If there is room, try to inline some more warm call sites. 1802 // %%% Do a graph index compaction pass when we think we're out of space? 1803 if (!InlineWarmCalls) return 0; 1804 1805 int calls_made_hot = 0; 1806 int room_to_grow = NodeCountInliningCutoff - unique(); 1807 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1808 int amount_grown = 0; 1809 WarmCallInfo* call; 1810 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1811 int est_size = (int)call->size(); 1812 if (est_size > (room_to_grow - amount_grown)) { 1813 // This one won't fit anyway. Get rid of it. 1814 call->make_cold(); 1815 continue; 1816 } 1817 call->make_hot(); 1818 calls_made_hot++; 1819 amount_grown += est_size; 1820 amount_to_grow -= est_size; 1821 } 1822 1823 if (calls_made_hot > 0) set_major_progress(); 1824 return calls_made_hot; 1825 } 1826 1827 1828 //----------------------------Finish_Warm-------------------------------------- 1829 void Compile::Finish_Warm() { 1830 if (!InlineWarmCalls) return; 1831 if (failing()) return; 1832 if (warm_calls() == NULL) return; 1833 1834 // Clean up loose ends, if we are out of space for inlining. 1835 WarmCallInfo* call; 1836 while ((call = pop_warm_call()) != NULL) { 1837 call->make_cold(); 1838 } 1839 } 1840 1841 //---------------------cleanup_loop_predicates----------------------- 1842 // Remove the opaque nodes that protect the predicates so that all unused 1843 // checks and uncommon_traps will be eliminated from the ideal graph 1844 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1845 if (predicate_count()==0) return; 1846 for (int i = predicate_count(); i > 0; i--) { 1847 Node * n = predicate_opaque1_node(i-1); 1848 assert(n->Opcode() == Op_Opaque1, "must be"); 1849 igvn.replace_node(n, n->in(1)); 1850 } 1851 assert(predicate_count()==0, "should be clean!"); 1852 } 1853 1854 // StringOpts and late inlining of string methods 1855 void Compile::inline_string_calls(bool parse_time) { 1856 { 1857 // remove useless nodes to make the usage analysis simpler 1858 ResourceMark rm; 1859 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1860 } 1861 1862 { 1863 ResourceMark rm; 1864 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1865 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1866 print_method(PHASE_AFTER_STRINGOPTS, 3); 1867 } 1868 1869 // now inline anything that we skipped the first time around 1870 if (!parse_time) { 1871 _late_inlines_pos = _late_inlines.length(); 1872 } 1873 1874 while (_string_late_inlines.length() > 0) { 1875 CallGenerator* cg = _string_late_inlines.pop(); 1876 cg->do_late_inline(); 1877 if (failing()) return; 1878 } 1879 _string_late_inlines.trunc_to(0); 1880 } 1881 1882 // Late inlining of boxing methods 1883 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1884 if (_boxing_late_inlines.length() > 0) { 1885 assert(has_boxed_value(), "inconsistent"); 1886 1887 PhaseGVN* gvn = initial_gvn(); 1888 set_inlining_incrementally(true); 1889 1890 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1891 for_igvn()->clear(); 1892 gvn->replace_with(&igvn); 1893 1894 while (_boxing_late_inlines.length() > 0) { 1895 CallGenerator* cg = _boxing_late_inlines.pop(); 1896 cg->do_late_inline(); 1897 if (failing()) return; 1898 } 1899 _boxing_late_inlines.trunc_to(0); 1900 1901 { 1902 ResourceMark rm; 1903 PhaseRemoveUseless pru(gvn, for_igvn()); 1904 } 1905 1906 igvn = PhaseIterGVN(gvn); 1907 igvn.optimize(); 1908 1909 set_inlining_progress(false); 1910 set_inlining_incrementally(false); 1911 } 1912 } 1913 1914 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1915 assert(IncrementalInline, "incremental inlining should be on"); 1916 PhaseGVN* gvn = initial_gvn(); 1917 1918 set_inlining_progress(false); 1919 for_igvn()->clear(); 1920 gvn->replace_with(&igvn); 1921 1922 int i = 0; 1923 1924 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1925 CallGenerator* cg = _late_inlines.at(i); 1926 _late_inlines_pos = i+1; 1927 cg->do_late_inline(); 1928 if (failing()) return; 1929 } 1930 int j = 0; 1931 for (; i < _late_inlines.length(); i++, j++) { 1932 _late_inlines.at_put(j, _late_inlines.at(i)); 1933 } 1934 _late_inlines.trunc_to(j); 1935 1936 { 1937 ResourceMark rm; 1938 PhaseRemoveUseless pru(gvn, for_igvn()); 1939 } 1940 1941 igvn = PhaseIterGVN(gvn); 1942 } 1943 1944 // Perform incremental inlining until bound on number of live nodes is reached 1945 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1946 PhaseGVN* gvn = initial_gvn(); 1947 1948 set_inlining_incrementally(true); 1949 set_inlining_progress(true); 1950 uint low_live_nodes = 0; 1951 1952 while(inlining_progress() && _late_inlines.length() > 0) { 1953 1954 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1955 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1956 // PhaseIdealLoop is expensive so we only try it once we are 1957 // out of loop and we only try it again if the previous helped 1958 // got the number of nodes down significantly 1959 PhaseIdealLoop ideal_loop( igvn, false, true ); 1960 if (failing()) return; 1961 low_live_nodes = live_nodes(); 1962 _major_progress = true; 1963 } 1964 1965 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1966 break; 1967 } 1968 } 1969 1970 inline_incrementally_one(igvn); 1971 1972 if (failing()) return; 1973 1974 igvn.optimize(); 1975 1976 if (failing()) return; 1977 } 1978 1979 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1980 1981 if (_string_late_inlines.length() > 0) { 1982 assert(has_stringbuilder(), "inconsistent"); 1983 for_igvn()->clear(); 1984 initial_gvn()->replace_with(&igvn); 1985 1986 inline_string_calls(false); 1987 1988 if (failing()) return; 1989 1990 { 1991 ResourceMark rm; 1992 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1993 } 1994 1995 igvn = PhaseIterGVN(gvn); 1996 1997 igvn.optimize(); 1998 } 1999 2000 set_inlining_incrementally(false); 2001 } 2002 2003 2004 //------------------------------Optimize--------------------------------------- 2005 // Given a graph, optimize it. 2006 void Compile::Optimize() { 2007 TracePhase t1("optimizer", &_t_optimizer, true); 2008 2009 #ifndef PRODUCT 2010 if (env()->break_at_compile()) { 2011 BREAKPOINT; 2012 } 2013 2014 #endif 2015 2016 ResourceMark rm; 2017 int loop_opts_cnt; 2018 2019 NOT_PRODUCT( verify_graph_edges(); ) 2020 2021 print_method(PHASE_AFTER_PARSING); 2022 2023 { 2024 // Iterative Global Value Numbering, including ideal transforms 2025 // Initialize IterGVN with types and values from parse-time GVN 2026 PhaseIterGVN igvn(initial_gvn()); 2027 { 2028 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2029 igvn.optimize(); 2030 } 2031 2032 print_method(PHASE_ITER_GVN1, 2); 2033 2034 if (failing()) return; 2035 2036 { 2037 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2038 inline_incrementally(igvn); 2039 } 2040 2041 print_method(PHASE_INCREMENTAL_INLINE, 2); 2042 2043 if (failing()) return; 2044 2045 if (eliminate_boxing()) { 2046 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2047 // Inline valueOf() methods now. 2048 inline_boxing_calls(igvn); 2049 2050 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2051 2052 if (failing()) return; 2053 } 2054 2055 // Remove the speculative part of types and clean up the graph from 2056 // the extra CastPP nodes whose only purpose is to carry them. Do 2057 // that early so that optimizations are not disrupted by the extra 2058 // CastPP nodes. 2059 remove_speculative_types(igvn); 2060 2061 // No more new expensive nodes will be added to the list from here 2062 // so keep only the actual candidates for optimizations. 2063 cleanup_expensive_nodes(igvn); 2064 2065 // Perform escape analysis 2066 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2067 if (has_loops()) { 2068 // Cleanup graph (remove dead nodes). 2069 TracePhase t2("idealLoop", &_t_idealLoop, true); 2070 PhaseIdealLoop ideal_loop( igvn, false, true ); 2071 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2072 if (failing()) return; 2073 } 2074 ConnectionGraph::do_analysis(this, &igvn); 2075 2076 if (failing()) return; 2077 2078 // Optimize out fields loads from scalar replaceable allocations. 2079 igvn.optimize(); 2080 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2081 2082 if (failing()) return; 2083 2084 if (congraph() != NULL && macro_count() > 0) { 2085 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2086 PhaseMacroExpand mexp(igvn); 2087 mexp.eliminate_macro_nodes(); 2088 igvn.set_delay_transform(false); 2089 2090 igvn.optimize(); 2091 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2092 2093 if (failing()) return; 2094 } 2095 } 2096 2097 // Loop transforms on the ideal graph. Range Check Elimination, 2098 // peeling, unrolling, etc. 2099 2100 // Set loop opts counter 2101 loop_opts_cnt = num_loop_opts(); 2102 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2103 { 2104 TracePhase t2("idealLoop", &_t_idealLoop, true); 2105 PhaseIdealLoop ideal_loop( igvn, true ); 2106 loop_opts_cnt--; 2107 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2108 if (failing()) return; 2109 } 2110 // Loop opts pass if partial peeling occurred in previous pass 2111 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2112 TracePhase t3("idealLoop", &_t_idealLoop, true); 2113 PhaseIdealLoop ideal_loop( igvn, false ); 2114 loop_opts_cnt--; 2115 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2116 if (failing()) return; 2117 } 2118 // Loop opts pass for loop-unrolling before CCP 2119 if(major_progress() && (loop_opts_cnt > 0)) { 2120 TracePhase t4("idealLoop", &_t_idealLoop, true); 2121 PhaseIdealLoop ideal_loop( igvn, false ); 2122 loop_opts_cnt--; 2123 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2124 } 2125 if (!failing()) { 2126 // Verify that last round of loop opts produced a valid graph 2127 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2128 PhaseIdealLoop::verify(igvn); 2129 } 2130 } 2131 if (failing()) return; 2132 2133 // Conditional Constant Propagation; 2134 PhaseCCP ccp( &igvn ); 2135 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2136 { 2137 TracePhase t2("ccp", &_t_ccp, true); 2138 ccp.do_transform(); 2139 } 2140 print_method(PHASE_CPP1, 2); 2141 2142 assert( true, "Break here to ccp.dump_old2new_map()"); 2143 2144 // Iterative Global Value Numbering, including ideal transforms 2145 { 2146 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2147 igvn = ccp; 2148 igvn.optimize(); 2149 } 2150 2151 print_method(PHASE_ITER_GVN2, 2); 2152 2153 if (failing()) return; 2154 2155 // Loop transforms on the ideal graph. Range Check Elimination, 2156 // peeling, unrolling, etc. 2157 if(loop_opts_cnt > 0) { 2158 debug_only( int cnt = 0; ); 2159 while(major_progress() && (loop_opts_cnt > 0)) { 2160 TracePhase t2("idealLoop", &_t_idealLoop, true); 2161 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2162 PhaseIdealLoop ideal_loop( igvn, true); 2163 loop_opts_cnt--; 2164 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2165 if (failing()) return; 2166 } 2167 } 2168 2169 { 2170 // Verify that all previous optimizations produced a valid graph 2171 // at least to this point, even if no loop optimizations were done. 2172 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2173 PhaseIdealLoop::verify(igvn); 2174 } 2175 2176 { 2177 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2178 PhaseMacroExpand mex(igvn); 2179 if (mex.expand_macro_nodes()) { 2180 assert(failing(), "must bail out w/ explicit message"); 2181 return; 2182 } 2183 } 2184 2185 } // (End scope of igvn; run destructor if necessary for asserts.) 2186 2187 dump_inlining(); 2188 // A method with only infinite loops has no edges entering loops from root 2189 { 2190 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2191 if (final_graph_reshaping()) { 2192 assert(failing(), "must bail out w/ explicit message"); 2193 return; 2194 } 2195 } 2196 2197 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2198 } 2199 2200 2201 //------------------------------Code_Gen--------------------------------------- 2202 // Given a graph, generate code for it 2203 void Compile::Code_Gen() { 2204 if (failing()) { 2205 return; 2206 } 2207 2208 // Perform instruction selection. You might think we could reclaim Matcher 2209 // memory PDQ, but actually the Matcher is used in generating spill code. 2210 // Internals of the Matcher (including some VectorSets) must remain live 2211 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2212 // set a bit in reclaimed memory. 2213 2214 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2215 // nodes. Mapping is only valid at the root of each matched subtree. 2216 NOT_PRODUCT( verify_graph_edges(); ) 2217 2218 Matcher matcher; 2219 _matcher = &matcher; 2220 { 2221 TracePhase t2("matcher", &_t_matcher, true); 2222 matcher.match(); 2223 } 2224 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2225 // nodes. Mapping is only valid at the root of each matched subtree. 2226 NOT_PRODUCT( verify_graph_edges(); ) 2227 2228 // If you have too many nodes, or if matching has failed, bail out 2229 check_node_count(0, "out of nodes matching instructions"); 2230 if (failing()) { 2231 return; 2232 } 2233 2234 // Build a proper-looking CFG 2235 PhaseCFG cfg(node_arena(), root(), matcher); 2236 _cfg = &cfg; 2237 { 2238 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2239 bool success = cfg.do_global_code_motion(); 2240 if (!success) { 2241 return; 2242 } 2243 2244 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2245 NOT_PRODUCT( verify_graph_edges(); ) 2246 debug_only( cfg.verify(); ) 2247 } 2248 2249 PhaseChaitin regalloc(unique(), cfg, matcher); 2250 _regalloc = ®alloc; 2251 { 2252 TracePhase t2("regalloc", &_t_registerAllocation, true); 2253 // Perform register allocation. After Chaitin, use-def chains are 2254 // no longer accurate (at spill code) and so must be ignored. 2255 // Node->LRG->reg mappings are still accurate. 2256 _regalloc->Register_Allocate(); 2257 2258 // Bail out if the allocator builds too many nodes 2259 if (failing()) { 2260 return; 2261 } 2262 } 2263 2264 // Prior to register allocation we kept empty basic blocks in case the 2265 // the allocator needed a place to spill. After register allocation we 2266 // are not adding any new instructions. If any basic block is empty, we 2267 // can now safely remove it. 2268 { 2269 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2270 cfg.remove_empty_blocks(); 2271 if (do_freq_based_layout()) { 2272 PhaseBlockLayout layout(cfg); 2273 } else { 2274 cfg.set_loop_alignment(); 2275 } 2276 cfg.fixup_flow(); 2277 } 2278 2279 // Apply peephole optimizations 2280 if( OptoPeephole ) { 2281 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2282 PhasePeephole peep( _regalloc, cfg); 2283 peep.do_transform(); 2284 } 2285 2286 // Do late expand if CPU requires this. 2287 if (Matcher::require_postalloc_expand) { 2288 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2289 cfg.postalloc_expand(_regalloc); 2290 } 2291 2292 // Convert Nodes to instruction bits in a buffer 2293 { 2294 // %%%% workspace merge brought two timers together for one job 2295 TracePhase t2a("output", &_t_output, true); 2296 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2297 Output(); 2298 } 2299 2300 print_method(PHASE_FINAL_CODE); 2301 2302 // He's dead, Jim. 2303 _cfg = (PhaseCFG*)0xdeadbeef; 2304 _regalloc = (PhaseChaitin*)0xdeadbeef; 2305 } 2306 2307 2308 //------------------------------dump_asm--------------------------------------- 2309 // Dump formatted assembly 2310 #ifndef PRODUCT 2311 void Compile::dump_asm(int *pcs, uint pc_limit) { 2312 bool cut_short = false; 2313 tty->print_cr("#"); 2314 tty->print("# "); _tf->dump(); tty->cr(); 2315 tty->print_cr("#"); 2316 2317 // For all blocks 2318 int pc = 0x0; // Program counter 2319 char starts_bundle = ' '; 2320 _regalloc->dump_frame(); 2321 2322 Node *n = NULL; 2323 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2324 if (VMThread::should_terminate()) { 2325 cut_short = true; 2326 break; 2327 } 2328 Block* block = _cfg->get_block(i); 2329 if (block->is_connector() && !Verbose) { 2330 continue; 2331 } 2332 n = block->head(); 2333 if (pcs && n->_idx < pc_limit) { 2334 tty->print("%3.3x ", pcs[n->_idx]); 2335 } else { 2336 tty->print(" "); 2337 } 2338 block->dump_head(_cfg); 2339 if (block->is_connector()) { 2340 tty->print_cr(" # Empty connector block"); 2341 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2342 tty->print_cr(" # Block is sole successor of call"); 2343 } 2344 2345 // For all instructions 2346 Node *delay = NULL; 2347 for (uint j = 0; j < block->number_of_nodes(); j++) { 2348 if (VMThread::should_terminate()) { 2349 cut_short = true; 2350 break; 2351 } 2352 n = block->get_node(j); 2353 if (valid_bundle_info(n)) { 2354 Bundle* bundle = node_bundling(n); 2355 if (bundle->used_in_unconditional_delay()) { 2356 delay = n; 2357 continue; 2358 } 2359 if (bundle->starts_bundle()) { 2360 starts_bundle = '+'; 2361 } 2362 } 2363 2364 if (WizardMode) { 2365 n->dump(); 2366 } 2367 2368 if( !n->is_Region() && // Dont print in the Assembly 2369 !n->is_Phi() && // a few noisely useless nodes 2370 !n->is_Proj() && 2371 !n->is_MachTemp() && 2372 !n->is_SafePointScalarObject() && 2373 !n->is_Catch() && // Would be nice to print exception table targets 2374 !n->is_MergeMem() && // Not very interesting 2375 !n->is_top() && // Debug info table constants 2376 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2377 ) { 2378 if (pcs && n->_idx < pc_limit) 2379 tty->print("%3.3x", pcs[n->_idx]); 2380 else 2381 tty->print(" "); 2382 tty->print(" %c ", starts_bundle); 2383 starts_bundle = ' '; 2384 tty->print("\t"); 2385 n->format(_regalloc, tty); 2386 tty->cr(); 2387 } 2388 2389 // If we have an instruction with a delay slot, and have seen a delay, 2390 // then back up and print it 2391 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2392 assert(delay != NULL, "no unconditional delay instruction"); 2393 if (WizardMode) delay->dump(); 2394 2395 if (node_bundling(delay)->starts_bundle()) 2396 starts_bundle = '+'; 2397 if (pcs && n->_idx < pc_limit) 2398 tty->print("%3.3x", pcs[n->_idx]); 2399 else 2400 tty->print(" "); 2401 tty->print(" %c ", starts_bundle); 2402 starts_bundle = ' '; 2403 tty->print("\t"); 2404 delay->format(_regalloc, tty); 2405 tty->print_cr(""); 2406 delay = NULL; 2407 } 2408 2409 // Dump the exception table as well 2410 if( n->is_Catch() && (Verbose || WizardMode) ) { 2411 // Print the exception table for this offset 2412 _handler_table.print_subtable_for(pc); 2413 } 2414 } 2415 2416 if (pcs && n->_idx < pc_limit) 2417 tty->print_cr("%3.3x", pcs[n->_idx]); 2418 else 2419 tty->print_cr(""); 2420 2421 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2422 2423 } // End of per-block dump 2424 tty->print_cr(""); 2425 2426 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2427 } 2428 #endif 2429 2430 //------------------------------Final_Reshape_Counts--------------------------- 2431 // This class defines counters to help identify when a method 2432 // may/must be executed using hardware with only 24-bit precision. 2433 struct Final_Reshape_Counts : public StackObj { 2434 int _call_count; // count non-inlined 'common' calls 2435 int _float_count; // count float ops requiring 24-bit precision 2436 int _double_count; // count double ops requiring more precision 2437 int _java_call_count; // count non-inlined 'java' calls 2438 int _inner_loop_count; // count loops which need alignment 2439 VectorSet _visited; // Visitation flags 2440 Node_List _tests; // Set of IfNodes & PCTableNodes 2441 2442 Final_Reshape_Counts() : 2443 _call_count(0), _float_count(0), _double_count(0), 2444 _java_call_count(0), _inner_loop_count(0), 2445 _visited( Thread::current()->resource_area() ) { } 2446 2447 void inc_call_count () { _call_count ++; } 2448 void inc_float_count () { _float_count ++; } 2449 void inc_double_count() { _double_count++; } 2450 void inc_java_call_count() { _java_call_count++; } 2451 void inc_inner_loop_count() { _inner_loop_count++; } 2452 2453 int get_call_count () const { return _call_count ; } 2454 int get_float_count () const { return _float_count ; } 2455 int get_double_count() const { return _double_count; } 2456 int get_java_call_count() const { return _java_call_count; } 2457 int get_inner_loop_count() const { return _inner_loop_count; } 2458 }; 2459 2460 #ifdef ASSERT 2461 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2462 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2463 // Make sure the offset goes inside the instance layout. 2464 return k->contains_field_offset(tp->offset()); 2465 // Note that OffsetBot and OffsetTop are very negative. 2466 } 2467 #endif 2468 2469 // Eliminate trivially redundant StoreCMs and accumulate their 2470 // precedence edges. 2471 void Compile::eliminate_redundant_card_marks(Node* n) { 2472 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2473 if (n->in(MemNode::Address)->outcnt() > 1) { 2474 // There are multiple users of the same address so it might be 2475 // possible to eliminate some of the StoreCMs 2476 Node* mem = n->in(MemNode::Memory); 2477 Node* adr = n->in(MemNode::Address); 2478 Node* val = n->in(MemNode::ValueIn); 2479 Node* prev = n; 2480 bool done = false; 2481 // Walk the chain of StoreCMs eliminating ones that match. As 2482 // long as it's a chain of single users then the optimization is 2483 // safe. Eliminating partially redundant StoreCMs would require 2484 // cloning copies down the other paths. 2485 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2486 if (adr == mem->in(MemNode::Address) && 2487 val == mem->in(MemNode::ValueIn)) { 2488 // redundant StoreCM 2489 if (mem->req() > MemNode::OopStore) { 2490 // Hasn't been processed by this code yet. 2491 n->add_prec(mem->in(MemNode::OopStore)); 2492 } else { 2493 // Already converted to precedence edge 2494 for (uint i = mem->req(); i < mem->len(); i++) { 2495 // Accumulate any precedence edges 2496 if (mem->in(i) != NULL) { 2497 n->add_prec(mem->in(i)); 2498 } 2499 } 2500 // Everything above this point has been processed. 2501 done = true; 2502 } 2503 // Eliminate the previous StoreCM 2504 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2505 assert(mem->outcnt() == 0, "should be dead"); 2506 mem->disconnect_inputs(NULL, this); 2507 } else { 2508 prev = mem; 2509 } 2510 mem = prev->in(MemNode::Memory); 2511 } 2512 } 2513 } 2514 2515 //------------------------------final_graph_reshaping_impl---------------------- 2516 // Implement items 1-5 from final_graph_reshaping below. 2517 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2518 2519 if ( n->outcnt() == 0 ) return; // dead node 2520 uint nop = n->Opcode(); 2521 2522 // Check for 2-input instruction with "last use" on right input. 2523 // Swap to left input. Implements item (2). 2524 if( n->req() == 3 && // two-input instruction 2525 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2526 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2527 n->in(2)->outcnt() == 1 &&// right use IS a last use 2528 !n->in(2)->is_Con() ) { // right use is not a constant 2529 // Check for commutative opcode 2530 switch( nop ) { 2531 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2532 case Op_MaxI: case Op_MinI: 2533 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2534 case Op_AndL: case Op_XorL: case Op_OrL: 2535 case Op_AndI: case Op_XorI: case Op_OrI: { 2536 // Move "last use" input to left by swapping inputs 2537 n->swap_edges(1, 2); 2538 break; 2539 } 2540 default: 2541 break; 2542 } 2543 } 2544 2545 #ifdef ASSERT 2546 if( n->is_Mem() ) { 2547 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2548 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2549 // oop will be recorded in oop map if load crosses safepoint 2550 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2551 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2552 "raw memory operations should have control edge"); 2553 } 2554 #endif 2555 // Count FPU ops and common calls, implements item (3) 2556 switch( nop ) { 2557 // Count all float operations that may use FPU 2558 case Op_AddF: 2559 case Op_SubF: 2560 case Op_MulF: 2561 case Op_DivF: 2562 case Op_NegF: 2563 case Op_ModF: 2564 case Op_ConvI2F: 2565 case Op_ConF: 2566 case Op_CmpF: 2567 case Op_CmpF3: 2568 // case Op_ConvL2F: // longs are split into 32-bit halves 2569 frc.inc_float_count(); 2570 break; 2571 2572 case Op_ConvF2D: 2573 case Op_ConvD2F: 2574 frc.inc_float_count(); 2575 frc.inc_double_count(); 2576 break; 2577 2578 // Count all double operations that may use FPU 2579 case Op_AddD: 2580 case Op_SubD: 2581 case Op_MulD: 2582 case Op_DivD: 2583 case Op_NegD: 2584 case Op_ModD: 2585 case Op_ConvI2D: 2586 case Op_ConvD2I: 2587 // case Op_ConvL2D: // handled by leaf call 2588 // case Op_ConvD2L: // handled by leaf call 2589 case Op_ConD: 2590 case Op_CmpD: 2591 case Op_CmpD3: 2592 frc.inc_double_count(); 2593 break; 2594 case Op_Opaque1: // Remove Opaque Nodes before matching 2595 case Op_Opaque2: // Remove Opaque Nodes before matching 2596 n->subsume_by(n->in(1), this); 2597 break; 2598 case Op_CallStaticJava: 2599 case Op_CallJava: 2600 case Op_CallDynamicJava: 2601 frc.inc_java_call_count(); // Count java call site; 2602 case Op_CallRuntime: 2603 case Op_CallLeaf: 2604 case Op_CallLeafNoFP: { 2605 assert( n->is_Call(), "" ); 2606 CallNode *call = n->as_Call(); 2607 // Count call sites where the FP mode bit would have to be flipped. 2608 // Do not count uncommon runtime calls: 2609 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2610 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2611 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2612 frc.inc_call_count(); // Count the call site 2613 } else { // See if uncommon argument is shared 2614 Node *n = call->in(TypeFunc::Parms); 2615 int nop = n->Opcode(); 2616 // Clone shared simple arguments to uncommon calls, item (1). 2617 if( n->outcnt() > 1 && 2618 !n->is_Proj() && 2619 nop != Op_CreateEx && 2620 nop != Op_CheckCastPP && 2621 nop != Op_DecodeN && 2622 nop != Op_DecodeNKlass && 2623 !n->is_Mem() ) { 2624 Node *x = n->clone(); 2625 call->set_req( TypeFunc::Parms, x ); 2626 } 2627 } 2628 break; 2629 } 2630 2631 case Op_StoreD: 2632 case Op_LoadD: 2633 case Op_LoadD_unaligned: 2634 frc.inc_double_count(); 2635 goto handle_mem; 2636 case Op_StoreF: 2637 case Op_LoadF: 2638 frc.inc_float_count(); 2639 goto handle_mem; 2640 2641 case Op_StoreCM: 2642 { 2643 // Convert OopStore dependence into precedence edge 2644 Node* prec = n->in(MemNode::OopStore); 2645 n->del_req(MemNode::OopStore); 2646 n->add_prec(prec); 2647 eliminate_redundant_card_marks(n); 2648 } 2649 2650 // fall through 2651 2652 case Op_StoreB: 2653 case Op_StoreC: 2654 case Op_StorePConditional: 2655 case Op_StoreI: 2656 case Op_StoreL: 2657 case Op_StoreIConditional: 2658 case Op_StoreLConditional: 2659 case Op_CompareAndSwapI: 2660 case Op_CompareAndSwapL: 2661 case Op_CompareAndSwapP: 2662 case Op_CompareAndSwapN: 2663 case Op_GetAndAddI: 2664 case Op_GetAndAddL: 2665 case Op_GetAndSetI: 2666 case Op_GetAndSetL: 2667 case Op_GetAndSetP: 2668 case Op_GetAndSetN: 2669 case Op_StoreP: 2670 case Op_StoreN: 2671 case Op_StoreNKlass: 2672 case Op_LoadB: 2673 case Op_LoadUB: 2674 case Op_LoadUS: 2675 case Op_LoadI: 2676 case Op_LoadKlass: 2677 case Op_LoadNKlass: 2678 case Op_LoadL: 2679 case Op_LoadL_unaligned: 2680 case Op_LoadPLocked: 2681 case Op_LoadP: 2682 case Op_LoadN: 2683 case Op_LoadRange: 2684 case Op_LoadS: { 2685 handle_mem: 2686 #ifdef ASSERT 2687 if( VerifyOptoOopOffsets ) { 2688 assert( n->is_Mem(), "" ); 2689 MemNode *mem = (MemNode*)n; 2690 // Check to see if address types have grounded out somehow. 2691 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2692 assert( !tp || oop_offset_is_sane(tp), "" ); 2693 } 2694 #endif 2695 break; 2696 } 2697 2698 case Op_AddP: { // Assert sane base pointers 2699 Node *addp = n->in(AddPNode::Address); 2700 assert( !addp->is_AddP() || 2701 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2702 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2703 "Base pointers must match" ); 2704 #ifdef _LP64 2705 if ((UseCompressedOops || UseCompressedClassPointers) && 2706 addp->Opcode() == Op_ConP && 2707 addp == n->in(AddPNode::Base) && 2708 n->in(AddPNode::Offset)->is_Con()) { 2709 // Use addressing with narrow klass to load with offset on x86. 2710 // On sparc loading 32-bits constant and decoding it have less 2711 // instructions (4) then load 64-bits constant (7). 2712 // Do this transformation here since IGVN will convert ConN back to ConP. 2713 const Type* t = addp->bottom_type(); 2714 if (t->isa_oopptr() || t->isa_klassptr()) { 2715 Node* nn = NULL; 2716 2717 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2718 2719 // Look for existing ConN node of the same exact type. 2720 Node* r = root(); 2721 uint cnt = r->outcnt(); 2722 for (uint i = 0; i < cnt; i++) { 2723 Node* m = r->raw_out(i); 2724 if (m!= NULL && m->Opcode() == op && 2725 m->bottom_type()->make_ptr() == t) { 2726 nn = m; 2727 break; 2728 } 2729 } 2730 if (nn != NULL) { 2731 // Decode a narrow oop to match address 2732 // [R12 + narrow_oop_reg<<3 + offset] 2733 if (t->isa_oopptr()) { 2734 nn = new (this) DecodeNNode(nn, t); 2735 } else { 2736 nn = new (this) DecodeNKlassNode(nn, t); 2737 } 2738 n->set_req(AddPNode::Base, nn); 2739 n->set_req(AddPNode::Address, nn); 2740 if (addp->outcnt() == 0) { 2741 addp->disconnect_inputs(NULL, this); 2742 } 2743 } 2744 } 2745 } 2746 #endif 2747 break; 2748 } 2749 2750 #ifdef _LP64 2751 case Op_CastPP: 2752 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2753 Node* in1 = n->in(1); 2754 const Type* t = n->bottom_type(); 2755 Node* new_in1 = in1->clone(); 2756 new_in1->as_DecodeN()->set_type(t); 2757 2758 if (!Matcher::narrow_oop_use_complex_address()) { 2759 // 2760 // x86, ARM and friends can handle 2 adds in addressing mode 2761 // and Matcher can fold a DecodeN node into address by using 2762 // a narrow oop directly and do implicit NULL check in address: 2763 // 2764 // [R12 + narrow_oop_reg<<3 + offset] 2765 // NullCheck narrow_oop_reg 2766 // 2767 // On other platforms (Sparc) we have to keep new DecodeN node and 2768 // use it to do implicit NULL check in address: 2769 // 2770 // decode_not_null narrow_oop_reg, base_reg 2771 // [base_reg + offset] 2772 // NullCheck base_reg 2773 // 2774 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2775 // to keep the information to which NULL check the new DecodeN node 2776 // corresponds to use it as value in implicit_null_check(). 2777 // 2778 new_in1->set_req(0, n->in(0)); 2779 } 2780 2781 n->subsume_by(new_in1, this); 2782 if (in1->outcnt() == 0) { 2783 in1->disconnect_inputs(NULL, this); 2784 } 2785 } 2786 break; 2787 2788 case Op_CmpP: 2789 // Do this transformation here to preserve CmpPNode::sub() and 2790 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2791 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2792 Node* in1 = n->in(1); 2793 Node* in2 = n->in(2); 2794 if (!in1->is_DecodeNarrowPtr()) { 2795 in2 = in1; 2796 in1 = n->in(2); 2797 } 2798 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2799 2800 Node* new_in2 = NULL; 2801 if (in2->is_DecodeNarrowPtr()) { 2802 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2803 new_in2 = in2->in(1); 2804 } else if (in2->Opcode() == Op_ConP) { 2805 const Type* t = in2->bottom_type(); 2806 if (t == TypePtr::NULL_PTR) { 2807 assert(in1->is_DecodeN(), "compare klass to null?"); 2808 // Don't convert CmpP null check into CmpN if compressed 2809 // oops implicit null check is not generated. 2810 // This will allow to generate normal oop implicit null check. 2811 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2812 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2813 // 2814 // This transformation together with CastPP transformation above 2815 // will generated code for implicit NULL checks for compressed oops. 2816 // 2817 // The original code after Optimize() 2818 // 2819 // LoadN memory, narrow_oop_reg 2820 // decode narrow_oop_reg, base_reg 2821 // CmpP base_reg, NULL 2822 // CastPP base_reg // NotNull 2823 // Load [base_reg + offset], val_reg 2824 // 2825 // after these transformations will be 2826 // 2827 // LoadN memory, narrow_oop_reg 2828 // CmpN narrow_oop_reg, NULL 2829 // decode_not_null narrow_oop_reg, base_reg 2830 // Load [base_reg + offset], val_reg 2831 // 2832 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2833 // since narrow oops can be used in debug info now (see the code in 2834 // final_graph_reshaping_walk()). 2835 // 2836 // At the end the code will be matched to 2837 // on x86: 2838 // 2839 // Load_narrow_oop memory, narrow_oop_reg 2840 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2841 // NullCheck narrow_oop_reg 2842 // 2843 // and on sparc: 2844 // 2845 // Load_narrow_oop memory, narrow_oop_reg 2846 // decode_not_null narrow_oop_reg, base_reg 2847 // Load [base_reg + offset], val_reg 2848 // NullCheck base_reg 2849 // 2850 } else if (t->isa_oopptr()) { 2851 new_in2 = ConNode::make(this, t->make_narrowoop()); 2852 } else if (t->isa_klassptr()) { 2853 new_in2 = ConNode::make(this, t->make_narrowklass()); 2854 } 2855 } 2856 if (new_in2 != NULL) { 2857 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 2858 n->subsume_by(cmpN, this); 2859 if (in1->outcnt() == 0) { 2860 in1->disconnect_inputs(NULL, this); 2861 } 2862 if (in2->outcnt() == 0) { 2863 in2->disconnect_inputs(NULL, this); 2864 } 2865 } 2866 } 2867 break; 2868 2869 case Op_DecodeN: 2870 case Op_DecodeNKlass: 2871 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2872 // DecodeN could be pinned when it can't be fold into 2873 // an address expression, see the code for Op_CastPP above. 2874 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2875 break; 2876 2877 case Op_EncodeP: 2878 case Op_EncodePKlass: { 2879 Node* in1 = n->in(1); 2880 if (in1->is_DecodeNarrowPtr()) { 2881 n->subsume_by(in1->in(1), this); 2882 } else if (in1->Opcode() == Op_ConP) { 2883 const Type* t = in1->bottom_type(); 2884 if (t == TypePtr::NULL_PTR) { 2885 assert(t->isa_oopptr(), "null klass?"); 2886 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2887 } else if (t->isa_oopptr()) { 2888 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2889 } else if (t->isa_klassptr()) { 2890 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2891 } 2892 } 2893 if (in1->outcnt() == 0) { 2894 in1->disconnect_inputs(NULL, this); 2895 } 2896 break; 2897 } 2898 2899 case Op_Proj: { 2900 if (OptimizeStringConcat) { 2901 ProjNode* p = n->as_Proj(); 2902 if (p->_is_io_use) { 2903 // Separate projections were used for the exception path which 2904 // are normally removed by a late inline. If it wasn't inlined 2905 // then they will hang around and should just be replaced with 2906 // the original one. 2907 Node* proj = NULL; 2908 // Replace with just one 2909 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2910 Node *use = i.get(); 2911 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2912 proj = use; 2913 break; 2914 } 2915 } 2916 assert(proj != NULL, "must be found"); 2917 p->subsume_by(proj, this); 2918 } 2919 } 2920 break; 2921 } 2922 2923 case Op_Phi: 2924 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2925 // The EncodeP optimization may create Phi with the same edges 2926 // for all paths. It is not handled well by Register Allocator. 2927 Node* unique_in = n->in(1); 2928 assert(unique_in != NULL, ""); 2929 uint cnt = n->req(); 2930 for (uint i = 2; i < cnt; i++) { 2931 Node* m = n->in(i); 2932 assert(m != NULL, ""); 2933 if (unique_in != m) 2934 unique_in = NULL; 2935 } 2936 if (unique_in != NULL) { 2937 n->subsume_by(unique_in, this); 2938 } 2939 } 2940 break; 2941 2942 #endif 2943 2944 case Op_ModI: 2945 if (UseDivMod) { 2946 // Check if a%b and a/b both exist 2947 Node* d = n->find_similar(Op_DivI); 2948 if (d) { 2949 // Replace them with a fused divmod if supported 2950 if (Matcher::has_match_rule(Op_DivModI)) { 2951 DivModINode* divmod = DivModINode::make(this, n); 2952 d->subsume_by(divmod->div_proj(), this); 2953 n->subsume_by(divmod->mod_proj(), this); 2954 } else { 2955 // replace a%b with a-((a/b)*b) 2956 Node* mult = new (this) MulINode(d, d->in(2)); 2957 Node* sub = new (this) SubINode(d->in(1), mult); 2958 n->subsume_by(sub, this); 2959 } 2960 } 2961 } 2962 break; 2963 2964 case Op_ModL: 2965 if (UseDivMod) { 2966 // Check if a%b and a/b both exist 2967 Node* d = n->find_similar(Op_DivL); 2968 if (d) { 2969 // Replace them with a fused divmod if supported 2970 if (Matcher::has_match_rule(Op_DivModL)) { 2971 DivModLNode* divmod = DivModLNode::make(this, n); 2972 d->subsume_by(divmod->div_proj(), this); 2973 n->subsume_by(divmod->mod_proj(), this); 2974 } else { 2975 // replace a%b with a-((a/b)*b) 2976 Node* mult = new (this) MulLNode(d, d->in(2)); 2977 Node* sub = new (this) SubLNode(d->in(1), mult); 2978 n->subsume_by(sub, this); 2979 } 2980 } 2981 } 2982 break; 2983 2984 case Op_LoadVector: 2985 case Op_StoreVector: 2986 break; 2987 2988 case Op_PackB: 2989 case Op_PackS: 2990 case Op_PackI: 2991 case Op_PackF: 2992 case Op_PackL: 2993 case Op_PackD: 2994 if (n->req()-1 > 2) { 2995 // Replace many operand PackNodes with a binary tree for matching 2996 PackNode* p = (PackNode*) n; 2997 Node* btp = p->binary_tree_pack(this, 1, n->req()); 2998 n->subsume_by(btp, this); 2999 } 3000 break; 3001 case Op_Loop: 3002 case Op_CountedLoop: 3003 if (n->as_Loop()->is_inner_loop()) { 3004 frc.inc_inner_loop_count(); 3005 } 3006 break; 3007 case Op_LShiftI: 3008 case Op_RShiftI: 3009 case Op_URShiftI: 3010 case Op_LShiftL: 3011 case Op_RShiftL: 3012 case Op_URShiftL: 3013 if (Matcher::need_masked_shift_count) { 3014 // The cpu's shift instructions don't restrict the count to the 3015 // lower 5/6 bits. We need to do the masking ourselves. 3016 Node* in2 = n->in(2); 3017 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3018 const TypeInt* t = in2->find_int_type(); 3019 if (t != NULL && t->is_con()) { 3020 juint shift = t->get_con(); 3021 if (shift > mask) { // Unsigned cmp 3022 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3023 } 3024 } else { 3025 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3026 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3027 n->set_req(2, shift); 3028 } 3029 } 3030 if (in2->outcnt() == 0) { // Remove dead node 3031 in2->disconnect_inputs(NULL, this); 3032 } 3033 } 3034 break; 3035 case Op_MemBarStoreStore: 3036 case Op_MemBarRelease: 3037 // Break the link with AllocateNode: it is no longer useful and 3038 // confuses register allocation. 3039 if (n->req() > MemBarNode::Precedent) { 3040 n->set_req(MemBarNode::Precedent, top()); 3041 } 3042 break; 3043 default: 3044 assert( !n->is_Call(), "" ); 3045 assert( !n->is_Mem(), "" ); 3046 break; 3047 } 3048 3049 // Collect CFG split points 3050 if (n->is_MultiBranch()) 3051 frc._tests.push(n); 3052 } 3053 3054 //------------------------------final_graph_reshaping_walk--------------------- 3055 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3056 // requires that the walk visits a node's inputs before visiting the node. 3057 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3058 ResourceArea *area = Thread::current()->resource_area(); 3059 Unique_Node_List sfpt(area); 3060 3061 frc._visited.set(root->_idx); // first, mark node as visited 3062 uint cnt = root->req(); 3063 Node *n = root; 3064 uint i = 0; 3065 while (true) { 3066 if (i < cnt) { 3067 // Place all non-visited non-null inputs onto stack 3068 Node* m = n->in(i); 3069 ++i; 3070 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3071 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3072 // compute worst case interpreter size in case of a deoptimization 3073 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3074 3075 sfpt.push(m); 3076 } 3077 cnt = m->req(); 3078 nstack.push(n, i); // put on stack parent and next input's index 3079 n = m; 3080 i = 0; 3081 } 3082 } else { 3083 // Now do post-visit work 3084 final_graph_reshaping_impl( n, frc ); 3085 if (nstack.is_empty()) 3086 break; // finished 3087 n = nstack.node(); // Get node from stack 3088 cnt = n->req(); 3089 i = nstack.index(); 3090 nstack.pop(); // Shift to the next node on stack 3091 } 3092 } 3093 3094 // Skip next transformation if compressed oops are not used. 3095 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3096 (!UseCompressedOops && !UseCompressedClassPointers)) 3097 return; 3098 3099 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3100 // It could be done for an uncommon traps or any safepoints/calls 3101 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3102 while (sfpt.size() > 0) { 3103 n = sfpt.pop(); 3104 JVMState *jvms = n->as_SafePoint()->jvms(); 3105 assert(jvms != NULL, "sanity"); 3106 int start = jvms->debug_start(); 3107 int end = n->req(); 3108 bool is_uncommon = (n->is_CallStaticJava() && 3109 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3110 for (int j = start; j < end; j++) { 3111 Node* in = n->in(j); 3112 if (in->is_DecodeNarrowPtr()) { 3113 bool safe_to_skip = true; 3114 if (!is_uncommon ) { 3115 // Is it safe to skip? 3116 for (uint i = 0; i < in->outcnt(); i++) { 3117 Node* u = in->raw_out(i); 3118 if (!u->is_SafePoint() || 3119 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3120 safe_to_skip = false; 3121 } 3122 } 3123 } 3124 if (safe_to_skip) { 3125 n->set_req(j, in->in(1)); 3126 } 3127 if (in->outcnt() == 0) { 3128 in->disconnect_inputs(NULL, this); 3129 } 3130 } 3131 } 3132 } 3133 } 3134 3135 //------------------------------final_graph_reshaping-------------------------- 3136 // Final Graph Reshaping. 3137 // 3138 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3139 // and not commoned up and forced early. Must come after regular 3140 // optimizations to avoid GVN undoing the cloning. Clone constant 3141 // inputs to Loop Phis; these will be split by the allocator anyways. 3142 // Remove Opaque nodes. 3143 // (2) Move last-uses by commutative operations to the left input to encourage 3144 // Intel update-in-place two-address operations and better register usage 3145 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3146 // calls canonicalizing them back. 3147 // (3) Count the number of double-precision FP ops, single-precision FP ops 3148 // and call sites. On Intel, we can get correct rounding either by 3149 // forcing singles to memory (requires extra stores and loads after each 3150 // FP bytecode) or we can set a rounding mode bit (requires setting and 3151 // clearing the mode bit around call sites). The mode bit is only used 3152 // if the relative frequency of single FP ops to calls is low enough. 3153 // This is a key transform for SPEC mpeg_audio. 3154 // (4) Detect infinite loops; blobs of code reachable from above but not 3155 // below. Several of the Code_Gen algorithms fail on such code shapes, 3156 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3157 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3158 // Detection is by looking for IfNodes where only 1 projection is 3159 // reachable from below or CatchNodes missing some targets. 3160 // (5) Assert for insane oop offsets in debug mode. 3161 3162 bool Compile::final_graph_reshaping() { 3163 // an infinite loop may have been eliminated by the optimizer, 3164 // in which case the graph will be empty. 3165 if (root()->req() == 1) { 3166 record_method_not_compilable("trivial infinite loop"); 3167 return true; 3168 } 3169 3170 // Expensive nodes have their control input set to prevent the GVN 3171 // from freely commoning them. There's no GVN beyond this point so 3172 // no need to keep the control input. We want the expensive nodes to 3173 // be freely moved to the least frequent code path by gcm. 3174 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3175 for (int i = 0; i < expensive_count(); i++) { 3176 _expensive_nodes->at(i)->set_req(0, NULL); 3177 } 3178 3179 Final_Reshape_Counts frc; 3180 3181 // Visit everybody reachable! 3182 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3183 Node_Stack nstack(unique() >> 1); 3184 final_graph_reshaping_walk(nstack, root(), frc); 3185 3186 // Check for unreachable (from below) code (i.e., infinite loops). 3187 for( uint i = 0; i < frc._tests.size(); i++ ) { 3188 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3189 // Get number of CFG targets. 3190 // Note that PCTables include exception targets after calls. 3191 uint required_outcnt = n->required_outcnt(); 3192 if (n->outcnt() != required_outcnt) { 3193 // Check for a few special cases. Rethrow Nodes never take the 3194 // 'fall-thru' path, so expected kids is 1 less. 3195 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3196 if (n->in(0)->in(0)->is_Call()) { 3197 CallNode *call = n->in(0)->in(0)->as_Call(); 3198 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3199 required_outcnt--; // Rethrow always has 1 less kid 3200 } else if (call->req() > TypeFunc::Parms && 3201 call->is_CallDynamicJava()) { 3202 // Check for null receiver. In such case, the optimizer has 3203 // detected that the virtual call will always result in a null 3204 // pointer exception. The fall-through projection of this CatchNode 3205 // will not be populated. 3206 Node *arg0 = call->in(TypeFunc::Parms); 3207 if (arg0->is_Type() && 3208 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3209 required_outcnt--; 3210 } 3211 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3212 call->req() > TypeFunc::Parms+1 && 3213 call->is_CallStaticJava()) { 3214 // Check for negative array length. In such case, the optimizer has 3215 // detected that the allocation attempt will always result in an 3216 // exception. There is no fall-through projection of this CatchNode . 3217 Node *arg1 = call->in(TypeFunc::Parms+1); 3218 if (arg1->is_Type() && 3219 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3220 required_outcnt--; 3221 } 3222 } 3223 } 3224 } 3225 // Recheck with a better notion of 'required_outcnt' 3226 if (n->outcnt() != required_outcnt) { 3227 record_method_not_compilable("malformed control flow"); 3228 return true; // Not all targets reachable! 3229 } 3230 } 3231 // Check that I actually visited all kids. Unreached kids 3232 // must be infinite loops. 3233 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3234 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3235 record_method_not_compilable("infinite loop"); 3236 return true; // Found unvisited kid; must be unreach 3237 } 3238 } 3239 3240 // If original bytecodes contained a mixture of floats and doubles 3241 // check if the optimizer has made it homogenous, item (3). 3242 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3243 frc.get_float_count() > 32 && 3244 frc.get_double_count() == 0 && 3245 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3246 set_24_bit_selection_and_mode( false, true ); 3247 } 3248 3249 set_java_calls(frc.get_java_call_count()); 3250 set_inner_loops(frc.get_inner_loop_count()); 3251 3252 // No infinite loops, no reason to bail out. 3253 return false; 3254 } 3255 3256 //-----------------------------too_many_traps---------------------------------- 3257 // Report if there are too many traps at the current method and bci. 3258 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3259 bool Compile::too_many_traps(ciMethod* method, 3260 int bci, 3261 Deoptimization::DeoptReason reason) { 3262 ciMethodData* md = method->method_data(); 3263 if (md->is_empty()) { 3264 // Assume the trap has not occurred, or that it occurred only 3265 // because of a transient condition during start-up in the interpreter. 3266 return false; 3267 } 3268 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3269 if (md->has_trap_at(bci, m, reason) != 0) { 3270 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3271 // Also, if there are multiple reasons, or if there is no per-BCI record, 3272 // assume the worst. 3273 if (log()) 3274 log()->elem("observe trap='%s' count='%d'", 3275 Deoptimization::trap_reason_name(reason), 3276 md->trap_count(reason)); 3277 return true; 3278 } else { 3279 // Ignore method/bci and see if there have been too many globally. 3280 return too_many_traps(reason, md); 3281 } 3282 } 3283 3284 // Less-accurate variant which does not require a method and bci. 3285 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3286 ciMethodData* logmd) { 3287 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3288 // Too many traps globally. 3289 // Note that we use cumulative trap_count, not just md->trap_count. 3290 if (log()) { 3291 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3292 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3293 Deoptimization::trap_reason_name(reason), 3294 mcount, trap_count(reason)); 3295 } 3296 return true; 3297 } else { 3298 // The coast is clear. 3299 return false; 3300 } 3301 } 3302 3303 //--------------------------too_many_recompiles-------------------------------- 3304 // Report if there are too many recompiles at the current method and bci. 3305 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3306 // Is not eager to return true, since this will cause the compiler to use 3307 // Action_none for a trap point, to avoid too many recompilations. 3308 bool Compile::too_many_recompiles(ciMethod* method, 3309 int bci, 3310 Deoptimization::DeoptReason reason) { 3311 ciMethodData* md = method->method_data(); 3312 if (md->is_empty()) { 3313 // Assume the trap has not occurred, or that it occurred only 3314 // because of a transient condition during start-up in the interpreter. 3315 return false; 3316 } 3317 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3318 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3319 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3320 Deoptimization::DeoptReason per_bc_reason 3321 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3322 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3323 if ((per_bc_reason == Deoptimization::Reason_none 3324 || md->has_trap_at(bci, m, reason) != 0) 3325 // The trap frequency measure we care about is the recompile count: 3326 && md->trap_recompiled_at(bci, m) 3327 && md->overflow_recompile_count() >= bc_cutoff) { 3328 // Do not emit a trap here if it has already caused recompilations. 3329 // Also, if there are multiple reasons, or if there is no per-BCI record, 3330 // assume the worst. 3331 if (log()) 3332 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3333 Deoptimization::trap_reason_name(reason), 3334 md->trap_count(reason), 3335 md->overflow_recompile_count()); 3336 return true; 3337 } else if (trap_count(reason) != 0 3338 && decompile_count() >= m_cutoff) { 3339 // Too many recompiles globally, and we have seen this sort of trap. 3340 // Use cumulative decompile_count, not just md->decompile_count. 3341 if (log()) 3342 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3343 Deoptimization::trap_reason_name(reason), 3344 md->trap_count(reason), trap_count(reason), 3345 md->decompile_count(), decompile_count()); 3346 return true; 3347 } else { 3348 // The coast is clear. 3349 return false; 3350 } 3351 } 3352 3353 // Compute when not to trap. Used by matching trap based nodes and 3354 // NullCheck optimization. 3355 void Compile::set_allowed_deopt_reasons() { 3356 _allowed_reasons = 0; 3357 if (is_method_compilation()) { 3358 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3359 assert(rs < BitsPerInt, "recode bit map"); 3360 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3361 _allowed_reasons |= nth_bit(rs); 3362 } 3363 } 3364 } 3365 } 3366 3367 #ifndef PRODUCT 3368 //------------------------------verify_graph_edges--------------------------- 3369 // Walk the Graph and verify that there is a one-to-one correspondence 3370 // between Use-Def edges and Def-Use edges in the graph. 3371 void Compile::verify_graph_edges(bool no_dead_code) { 3372 if (VerifyGraphEdges) { 3373 ResourceArea *area = Thread::current()->resource_area(); 3374 Unique_Node_List visited(area); 3375 // Call recursive graph walk to check edges 3376 _root->verify_edges(visited); 3377 if (no_dead_code) { 3378 // Now make sure that no visited node is used by an unvisited node. 3379 bool dead_nodes = 0; 3380 Unique_Node_List checked(area); 3381 while (visited.size() > 0) { 3382 Node* n = visited.pop(); 3383 checked.push(n); 3384 for (uint i = 0; i < n->outcnt(); i++) { 3385 Node* use = n->raw_out(i); 3386 if (checked.member(use)) continue; // already checked 3387 if (visited.member(use)) continue; // already in the graph 3388 if (use->is_Con()) continue; // a dead ConNode is OK 3389 // At this point, we have found a dead node which is DU-reachable. 3390 if (dead_nodes++ == 0) 3391 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3392 use->dump(2); 3393 tty->print_cr("---"); 3394 checked.push(use); // No repeats; pretend it is now checked. 3395 } 3396 } 3397 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3398 } 3399 } 3400 } 3401 3402 // Verify GC barriers consistency 3403 // Currently supported: 3404 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3405 void Compile::verify_barriers() { 3406 if (UseG1GC) { 3407 // Verify G1 pre-barriers 3408 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3409 3410 ResourceArea *area = Thread::current()->resource_area(); 3411 Unique_Node_List visited(area); 3412 Node_List worklist(area); 3413 // We're going to walk control flow backwards starting from the Root 3414 worklist.push(_root); 3415 while (worklist.size() > 0) { 3416 Node* x = worklist.pop(); 3417 if (x == NULL || x == top()) continue; 3418 if (visited.member(x)) { 3419 continue; 3420 } else { 3421 visited.push(x); 3422 } 3423 3424 if (x->is_Region()) { 3425 for (uint i = 1; i < x->req(); i++) { 3426 worklist.push(x->in(i)); 3427 } 3428 } else { 3429 worklist.push(x->in(0)); 3430 // We are looking for the pattern: 3431 // /->ThreadLocal 3432 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3433 // \->ConI(0) 3434 // We want to verify that the If and the LoadB have the same control 3435 // See GraphKit::g1_write_barrier_pre() 3436 if (x->is_If()) { 3437 IfNode *iff = x->as_If(); 3438 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3439 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3440 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3441 && cmp->in(1)->is_Load()) { 3442 LoadNode* load = cmp->in(1)->as_Load(); 3443 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3444 && load->in(2)->in(3)->is_Con() 3445 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3446 3447 Node* if_ctrl = iff->in(0); 3448 Node* load_ctrl = load->in(0); 3449 3450 if (if_ctrl != load_ctrl) { 3451 // Skip possible CProj->NeverBranch in infinite loops 3452 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3453 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3454 if_ctrl = if_ctrl->in(0)->in(0); 3455 } 3456 } 3457 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3458 } 3459 } 3460 } 3461 } 3462 } 3463 } 3464 } 3465 } 3466 3467 #endif 3468 3469 // The Compile object keeps track of failure reasons separately from the ciEnv. 3470 // This is required because there is not quite a 1-1 relation between the 3471 // ciEnv and its compilation task and the Compile object. Note that one 3472 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3473 // to backtrack and retry without subsuming loads. Other than this backtracking 3474 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3475 // by the logic in C2Compiler. 3476 void Compile::record_failure(const char* reason) { 3477 if (log() != NULL) { 3478 log()->elem("failure reason='%s' phase='compile'", reason); 3479 } 3480 if (_failure_reason == NULL) { 3481 // Record the first failure reason. 3482 _failure_reason = reason; 3483 } 3484 3485 EventCompilerFailure event; 3486 if (event.should_commit()) { 3487 event.set_compileID(Compile::compile_id()); 3488 event.set_failure(reason); 3489 event.commit(); 3490 } 3491 3492 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3493 C->print_method(PHASE_FAILURE); 3494 } 3495 _root = NULL; // flush the graph, too 3496 } 3497 3498 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3499 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3500 _phase_name(name), _dolog(dolog) 3501 { 3502 if (dolog) { 3503 C = Compile::current(); 3504 _log = C->log(); 3505 } else { 3506 C = NULL; 3507 _log = NULL; 3508 } 3509 if (_log != NULL) { 3510 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3511 _log->stamp(); 3512 _log->end_head(); 3513 } 3514 } 3515 3516 Compile::TracePhase::~TracePhase() { 3517 3518 C = Compile::current(); 3519 if (_dolog) { 3520 _log = C->log(); 3521 } else { 3522 _log = NULL; 3523 } 3524 3525 #ifdef ASSERT 3526 if (PrintIdealNodeCount) { 3527 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3528 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3529 } 3530 3531 if (VerifyIdealNodeCount) { 3532 Compile::current()->print_missing_nodes(); 3533 } 3534 #endif 3535 3536 if (_log != NULL) { 3537 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3538 } 3539 } 3540 3541 //============================================================================= 3542 // Two Constant's are equal when the type and the value are equal. 3543 bool Compile::Constant::operator==(const Constant& other) { 3544 if (type() != other.type() ) return false; 3545 if (can_be_reused() != other.can_be_reused()) return false; 3546 // For floating point values we compare the bit pattern. 3547 switch (type()) { 3548 case T_FLOAT: return (_v._value.i == other._v._value.i); 3549 case T_LONG: 3550 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3551 case T_OBJECT: 3552 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3553 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3554 case T_METADATA: return (_v._metadata == other._v._metadata); 3555 default: ShouldNotReachHere(); 3556 } 3557 return false; 3558 } 3559 3560 static int type_to_size_in_bytes(BasicType t) { 3561 switch (t) { 3562 case T_LONG: return sizeof(jlong ); 3563 case T_FLOAT: return sizeof(jfloat ); 3564 case T_DOUBLE: return sizeof(jdouble); 3565 case T_METADATA: return sizeof(Metadata*); 3566 // We use T_VOID as marker for jump-table entries (labels) which 3567 // need an internal word relocation. 3568 case T_VOID: 3569 case T_ADDRESS: 3570 case T_OBJECT: return sizeof(jobject); 3571 } 3572 3573 ShouldNotReachHere(); 3574 return -1; 3575 } 3576 3577 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3578 // sort descending 3579 if (a->freq() > b->freq()) return -1; 3580 if (a->freq() < b->freq()) return 1; 3581 return 0; 3582 } 3583 3584 void Compile::ConstantTable::calculate_offsets_and_size() { 3585 // First, sort the array by frequencies. 3586 _constants.sort(qsort_comparator); 3587 3588 #ifdef ASSERT 3589 // Make sure all jump-table entries were sorted to the end of the 3590 // array (they have a negative frequency). 3591 bool found_void = false; 3592 for (int i = 0; i < _constants.length(); i++) { 3593 Constant con = _constants.at(i); 3594 if (con.type() == T_VOID) 3595 found_void = true; // jump-tables 3596 else 3597 assert(!found_void, "wrong sorting"); 3598 } 3599 #endif 3600 3601 int offset = 0; 3602 for (int i = 0; i < _constants.length(); i++) { 3603 Constant* con = _constants.adr_at(i); 3604 3605 // Align offset for type. 3606 int typesize = type_to_size_in_bytes(con->type()); 3607 offset = align_size_up(offset, typesize); 3608 con->set_offset(offset); // set constant's offset 3609 3610 if (con->type() == T_VOID) { 3611 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3612 offset = offset + typesize * n->outcnt(); // expand jump-table 3613 } else { 3614 offset = offset + typesize; 3615 } 3616 } 3617 3618 // Align size up to the next section start (which is insts; see 3619 // CodeBuffer::align_at_start). 3620 assert(_size == -1, "already set?"); 3621 _size = align_size_up(offset, CodeEntryAlignment); 3622 } 3623 3624 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3625 MacroAssembler _masm(&cb); 3626 for (int i = 0; i < _constants.length(); i++) { 3627 Constant con = _constants.at(i); 3628 address constant_addr; 3629 switch (con.type()) { 3630 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3631 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3632 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3633 case T_OBJECT: { 3634 jobject obj = con.get_jobject(); 3635 int oop_index = _masm.oop_recorder()->find_index(obj); 3636 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3637 break; 3638 } 3639 case T_ADDRESS: { 3640 address addr = (address) con.get_jobject(); 3641 constant_addr = _masm.address_constant(addr); 3642 break; 3643 } 3644 // We use T_VOID as marker for jump-table entries (labels) which 3645 // need an internal word relocation. 3646 case T_VOID: { 3647 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3648 // Fill the jump-table with a dummy word. The real value is 3649 // filled in later in fill_jump_table. 3650 address dummy = (address) n; 3651 constant_addr = _masm.address_constant(dummy); 3652 // Expand jump-table 3653 for (uint i = 1; i < n->outcnt(); i++) { 3654 address temp_addr = _masm.address_constant(dummy + i); 3655 assert(temp_addr, "consts section too small"); 3656 } 3657 break; 3658 } 3659 case T_METADATA: { 3660 Metadata* obj = con.get_metadata(); 3661 int metadata_index = _masm.oop_recorder()->find_index(obj); 3662 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3663 break; 3664 } 3665 default: ShouldNotReachHere(); 3666 } 3667 assert(constant_addr, "consts section too small"); 3668 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); 3669 } 3670 } 3671 3672 int Compile::ConstantTable::find_offset(Constant& con) const { 3673 int idx = _constants.find(con); 3674 assert(idx != -1, "constant must be in constant table"); 3675 int offset = _constants.at(idx).offset(); 3676 assert(offset != -1, "constant table not emitted yet?"); 3677 return offset; 3678 } 3679 3680 void Compile::ConstantTable::add(Constant& con) { 3681 if (con.can_be_reused()) { 3682 int idx = _constants.find(con); 3683 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3684 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3685 return; 3686 } 3687 } 3688 (void) _constants.append(con); 3689 } 3690 3691 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3692 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3693 Constant con(type, value, b->_freq); 3694 add(con); 3695 return con; 3696 } 3697 3698 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3699 Constant con(metadata); 3700 add(con); 3701 return con; 3702 } 3703 3704 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3705 jvalue value; 3706 BasicType type = oper->type()->basic_type(); 3707 switch (type) { 3708 case T_LONG: value.j = oper->constantL(); break; 3709 case T_FLOAT: value.f = oper->constantF(); break; 3710 case T_DOUBLE: value.d = oper->constantD(); break; 3711 case T_OBJECT: 3712 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3713 case T_METADATA: return add((Metadata*)oper->constant()); break; 3714 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3715 } 3716 return add(n, type, value); 3717 } 3718 3719 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3720 jvalue value; 3721 // We can use the node pointer here to identify the right jump-table 3722 // as this method is called from Compile::Fill_buffer right before 3723 // the MachNodes are emitted and the jump-table is filled (means the 3724 // MachNode pointers do not change anymore). 3725 value.l = (jobject) n; 3726 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3727 add(con); 3728 return con; 3729 } 3730 3731 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3732 // If called from Compile::scratch_emit_size do nothing. 3733 if (Compile::current()->in_scratch_emit_size()) return; 3734 3735 assert(labels.is_nonempty(), "must be"); 3736 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3737 3738 // Since MachConstantNode::constant_offset() also contains 3739 // table_base_offset() we need to subtract the table_base_offset() 3740 // to get the plain offset into the constant table. 3741 int offset = n->constant_offset() - table_base_offset(); 3742 3743 MacroAssembler _masm(&cb); 3744 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3745 3746 for (uint i = 0; i < n->outcnt(); i++) { 3747 address* constant_addr = &jump_table_base[i]; 3748 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); 3749 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3750 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3751 } 3752 } 3753 3754 void Compile::dump_inlining() { 3755 if (print_inlining() || print_intrinsics()) { 3756 // Print inlining message for candidates that we couldn't inline 3757 // for lack of space or non constant receiver 3758 for (int i = 0; i < _late_inlines.length(); i++) { 3759 CallGenerator* cg = _late_inlines.at(i); 3760 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3761 } 3762 Unique_Node_List useful; 3763 useful.push(root()); 3764 for (uint next = 0; next < useful.size(); ++next) { 3765 Node* n = useful.at(next); 3766 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3767 CallNode* call = n->as_Call(); 3768 CallGenerator* cg = call->generator(); 3769 cg->print_inlining_late("receiver not constant"); 3770 } 3771 uint max = n->len(); 3772 for ( uint i = 0; i < max; ++i ) { 3773 Node *m = n->in(i); 3774 if ( m == NULL ) continue; 3775 useful.push(m); 3776 } 3777 } 3778 for (int i = 0; i < _print_inlining_list->length(); i++) { 3779 tty->print(_print_inlining_list->adr_at(i)->ss()->as_string()); 3780 } 3781 } 3782 } 3783 3784 // Dump inlining replay data to the stream. 3785 // Don't change thread state and acquire any locks. 3786 void Compile::dump_inline_data(outputStream* out) { 3787 InlineTree* inl_tree = ilt(); 3788 if (inl_tree != NULL) { 3789 out->print(" inline %d", inl_tree->count()); 3790 inl_tree->dump_replay_data(out); 3791 } 3792 } 3793 3794 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 3795 if (n1->Opcode() < n2->Opcode()) return -1; 3796 else if (n1->Opcode() > n2->Opcode()) return 1; 3797 3798 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())); 3799 for (uint i = 1; i < n1->req(); i++) { 3800 if (n1->in(i) < n2->in(i)) return -1; 3801 else if (n1->in(i) > n2->in(i)) return 1; 3802 } 3803 3804 return 0; 3805 } 3806 3807 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 3808 Node* n1 = *n1p; 3809 Node* n2 = *n2p; 3810 3811 return cmp_expensive_nodes(n1, n2); 3812 } 3813 3814 void Compile::sort_expensive_nodes() { 3815 if (!expensive_nodes_sorted()) { 3816 _expensive_nodes->sort(cmp_expensive_nodes); 3817 } 3818 } 3819 3820 bool Compile::expensive_nodes_sorted() const { 3821 for (int i = 1; i < _expensive_nodes->length(); i++) { 3822 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 3823 return false; 3824 } 3825 } 3826 return true; 3827 } 3828 3829 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 3830 if (_expensive_nodes->length() == 0) { 3831 return false; 3832 } 3833 3834 assert(OptimizeExpensiveOps, "optimization off?"); 3835 3836 // Take this opportunity to remove dead nodes from the list 3837 int j = 0; 3838 for (int i = 0; i < _expensive_nodes->length(); i++) { 3839 Node* n = _expensive_nodes->at(i); 3840 if (!n->is_unreachable(igvn)) { 3841 assert(n->is_expensive(), "should be expensive"); 3842 _expensive_nodes->at_put(j, n); 3843 j++; 3844 } 3845 } 3846 _expensive_nodes->trunc_to(j); 3847 3848 // Then sort the list so that similar nodes are next to each other 3849 // and check for at least two nodes of identical kind with same data 3850 // inputs. 3851 sort_expensive_nodes(); 3852 3853 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 3854 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 3855 return true; 3856 } 3857 } 3858 3859 return false; 3860 } 3861 3862 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 3863 if (_expensive_nodes->length() == 0) { 3864 return; 3865 } 3866 3867 assert(OptimizeExpensiveOps, "optimization off?"); 3868 3869 // Sort to bring similar nodes next to each other and clear the 3870 // control input of nodes for which there's only a single copy. 3871 sort_expensive_nodes(); 3872 3873 int j = 0; 3874 int identical = 0; 3875 int i = 0; 3876 for (; i < _expensive_nodes->length()-1; i++) { 3877 assert(j <= i, "can't write beyond current index"); 3878 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 3879 identical++; 3880 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3881 continue; 3882 } 3883 if (identical > 0) { 3884 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3885 identical = 0; 3886 } else { 3887 Node* n = _expensive_nodes->at(i); 3888 igvn.hash_delete(n); 3889 n->set_req(0, NULL); 3890 igvn.hash_insert(n); 3891 } 3892 } 3893 if (identical > 0) { 3894 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 3895 } else if (_expensive_nodes->length() >= 1) { 3896 Node* n = _expensive_nodes->at(i); 3897 igvn.hash_delete(n); 3898 n->set_req(0, NULL); 3899 igvn.hash_insert(n); 3900 } 3901 _expensive_nodes->trunc_to(j); 3902 } 3903 3904 void Compile::add_expensive_node(Node * n) { 3905 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 3906 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 3907 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 3908 if (OptimizeExpensiveOps) { 3909 _expensive_nodes->append(n); 3910 } else { 3911 // Clear control input and let IGVN optimize expensive nodes if 3912 // OptimizeExpensiveOps is off. 3913 n->set_req(0, NULL); 3914 } 3915 } 3916 3917 /** 3918 * Remove the speculative part of types and clean up the graph 3919 */ 3920 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 3921 if (UseTypeSpeculation) { 3922 Unique_Node_List worklist; 3923 worklist.push(root()); 3924 int modified = 0; 3925 // Go over all type nodes that carry a speculative type, drop the 3926 // speculative part of the type and enqueue the node for an igvn 3927 // which may optimize it out. 3928 for (uint next = 0; next < worklist.size(); ++next) { 3929 Node *n = worklist.at(next); 3930 if (n->is_Type()) { 3931 TypeNode* tn = n->as_Type(); 3932 const Type* t = tn->type(); 3933 const Type* t_no_spec = t->remove_speculative(); 3934 if (t_no_spec != t) { 3935 bool in_hash = igvn.hash_delete(n); 3936 assert(in_hash, "node should be in igvn hash table"); 3937 tn->set_type(t_no_spec); 3938 igvn.hash_insert(n); 3939 igvn._worklist.push(n); // give it a chance to go away 3940 modified++; 3941 } 3942 } 3943 uint max = n->len(); 3944 for( uint i = 0; i < max; ++i ) { 3945 Node *m = n->in(i); 3946 if (not_a_node(m)) continue; 3947 worklist.push(m); 3948 } 3949 } 3950 // Drop the speculative part of all types in the igvn's type table 3951 igvn.remove_speculative_types(); 3952 if (modified > 0) { 3953 igvn.optimize(); 3954 } 3955 #ifdef ASSERT 3956 // Verify that after the IGVN is over no speculative type has resurfaced 3957 worklist.clear(); 3958 worklist.push(root()); 3959 for (uint next = 0; next < worklist.size(); ++next) { 3960 Node *n = worklist.at(next); 3961 const Type* t = igvn.type(n); 3962 assert(t == t->remove_speculative(), "no more speculative types"); 3963 if (n->is_Type()) { 3964 t = n->as_Type()->type(); 3965 assert(t == t->remove_speculative(), "no more speculative types"); 3966 } 3967 uint max = n->len(); 3968 for( uint i = 0; i < max; ++i ) { 3969 Node *m = n->in(i); 3970 if (not_a_node(m)) continue; 3971 worklist.push(m); 3972 } 3973 } 3974 igvn.check_no_speculative_types(); 3975 #endif 3976 } 3977 } 3978 3979 // Auxiliary method to support randomized stressing/fuzzing. 3980 // 3981 // This method can be called the arbitrary number of times, with current count 3982 // as the argument. The logic allows selecting a single candidate from the 3983 // running list of candidates as follows: 3984 // int count = 0; 3985 // Cand* selected = null; 3986 // while(cand = cand->next()) { 3987 // if (randomized_select(++count)) { 3988 // selected = cand; 3989 // } 3990 // } 3991 // 3992 // Including count equalizes the chances any candidate is "selected". 3993 // This is useful when we don't have the complete list of candidates to choose 3994 // from uniformly. In this case, we need to adjust the randomicity of the 3995 // selection, or else we will end up biasing the selection towards the latter 3996 // candidates. 3997 // 3998 // Quick back-envelope calculation shows that for the list of n candidates 3999 // the equal probability for the candidate to persist as "best" can be 4000 // achieved by replacing it with "next" k-th candidate with the probability 4001 // of 1/k. It can be easily shown that by the end of the run, the 4002 // probability for any candidate is converged to 1/n, thus giving the 4003 // uniform distribution among all the candidates. 4004 // 4005 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4006 #define RANDOMIZED_DOMAIN_POW 29 4007 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4008 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4009 bool Compile::randomized_select(int count) { 4010 assert(count > 0, "only positive"); 4011 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4012 }