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