1 /* 2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "gc_implementation/shenandoah/brooksPointer.hpp" 36 #include "opto/addnode.hpp" 37 #include "opto/block.hpp" 38 #include "opto/c2compiler.hpp" 39 #include "opto/callGenerator.hpp" 40 #include "opto/callnode.hpp" 41 #include "opto/cfgnode.hpp" 42 #include "opto/chaitin.hpp" 43 #include "opto/compile.hpp" 44 #include "opto/connode.hpp" 45 #include "opto/divnode.hpp" 46 #include "opto/escape.hpp" 47 #include "opto/idealGraphPrinter.hpp" 48 #include "opto/loopnode.hpp" 49 #include "opto/machnode.hpp" 50 #include "opto/macro.hpp" 51 #include "opto/matcher.hpp" 52 #include "opto/mathexactnode.hpp" 53 #include "opto/memnode.hpp" 54 #include "opto/mulnode.hpp" 55 #include "opto/node.hpp" 56 #include "opto/opcodes.hpp" 57 #include "opto/output.hpp" 58 #include "opto/parse.hpp" 59 #include "opto/phaseX.hpp" 60 #include "opto/rootnode.hpp" 61 #include "opto/runtime.hpp" 62 #include "opto/shenandoahSupport.hpp" 63 #include "opto/stringopts.hpp" 64 #include "opto/type.hpp" 65 #include "opto/vectornode.hpp" 66 #include "runtime/arguments.hpp" 67 #include "runtime/signature.hpp" 68 #include "runtime/stubRoutines.hpp" 69 #include "runtime/timer.hpp" 70 #include "trace/tracing.hpp" 71 #include "utilities/copy.hpp" 72 #if defined AD_MD_HPP 73 # include AD_MD_HPP 74 #elif defined TARGET_ARCH_MODEL_x86_32 75 # include "adfiles/ad_x86_32.hpp" 76 #elif defined TARGET_ARCH_MODEL_x86_64 77 # include "adfiles/ad_x86_64.hpp" 78 #elif defined TARGET_ARCH_MODEL_aarch64 79 # include "adfiles/ad_aarch64.hpp" 80 #elif defined TARGET_ARCH_MODEL_sparc 81 # include "adfiles/ad_sparc.hpp" 82 #elif defined TARGET_ARCH_MODEL_zero 83 # include "adfiles/ad_zero.hpp" 84 #elif defined TARGET_ARCH_MODEL_ppc_64 85 # include "adfiles/ad_ppc_64.hpp" 86 #endif 87 88 #ifdef BUILTIN_SIM 89 #include "../../../../../../simulator/simulator.hpp" 90 #endif 91 92 // -------------------- Compile::mach_constant_base_node ----------------------- 93 // Constant table base node singleton. 94 MachConstantBaseNode* Compile::mach_constant_base_node() { 95 if (_mach_constant_base_node == NULL) { 96 _mach_constant_base_node = new (C) MachConstantBaseNode(); 97 _mach_constant_base_node->add_req(C->root()); 98 } 99 return _mach_constant_base_node; 100 } 101 102 103 /// Support for intrinsics. 104 105 // Return the index at which m must be inserted (or already exists). 106 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 107 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 108 #ifdef ASSERT 109 for (int i = 1; i < _intrinsics->length(); i++) { 110 CallGenerator* cg1 = _intrinsics->at(i-1); 111 CallGenerator* cg2 = _intrinsics->at(i); 112 assert(cg1->method() != cg2->method() 113 ? cg1->method() < cg2->method() 114 : cg1->is_virtual() < cg2->is_virtual(), 115 "compiler intrinsics list must stay sorted"); 116 } 117 #endif 118 // Binary search sorted list, in decreasing intervals [lo, hi]. 119 int lo = 0, hi = _intrinsics->length()-1; 120 while (lo <= hi) { 121 int mid = (uint)(hi + lo) / 2; 122 ciMethod* mid_m = _intrinsics->at(mid)->method(); 123 if (m < mid_m) { 124 hi = mid-1; 125 } else if (m > mid_m) { 126 lo = mid+1; 127 } else { 128 // look at minor sort key 129 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 130 if (is_virtual < mid_virt) { 131 hi = mid-1; 132 } else if (is_virtual > mid_virt) { 133 lo = mid+1; 134 } else { 135 return mid; // exact match 136 } 137 } 138 } 139 return lo; // inexact match 140 } 141 142 void Compile::register_intrinsic(CallGenerator* cg) { 143 if (_intrinsics == NULL) { 144 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 145 } 146 // This code is stolen from ciObjectFactory::insert. 147 // Really, GrowableArray should have methods for 148 // insert_at, remove_at, and binary_search. 149 int len = _intrinsics->length(); 150 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 151 if (index == len) { 152 _intrinsics->append(cg); 153 } else { 154 #ifdef ASSERT 155 CallGenerator* oldcg = _intrinsics->at(index); 156 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 157 #endif 158 _intrinsics->append(_intrinsics->at(len-1)); 159 int pos; 160 for (pos = len-2; pos >= index; pos--) { 161 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 162 } 163 _intrinsics->at_put(index, cg); 164 } 165 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 166 } 167 168 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 169 assert(m->is_loaded(), "don't try this on unloaded methods"); 170 if (_intrinsics != NULL) { 171 int index = intrinsic_insertion_index(m, is_virtual); 172 if (index < _intrinsics->length() 173 && _intrinsics->at(index)->method() == m 174 && _intrinsics->at(index)->is_virtual() == is_virtual) { 175 return _intrinsics->at(index); 176 } 177 } 178 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 179 if (m->intrinsic_id() != vmIntrinsics::_none && 180 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 181 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 182 if (cg != NULL) { 183 // Save it for next time: 184 register_intrinsic(cg); 185 return cg; 186 } else { 187 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 188 } 189 } 190 return NULL; 191 } 192 193 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 194 // in library_call.cpp. 195 196 197 #ifndef PRODUCT 198 // statistics gathering... 199 200 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 201 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 202 203 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 204 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 205 int oflags = _intrinsic_hist_flags[id]; 206 assert(flags != 0, "what happened?"); 207 if (is_virtual) { 208 flags |= _intrinsic_virtual; 209 } 210 bool changed = (flags != oflags); 211 if ((flags & _intrinsic_worked) != 0) { 212 juint count = (_intrinsic_hist_count[id] += 1); 213 if (count == 1) { 214 changed = true; // first time 215 } 216 // increment the overall count also: 217 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 218 } 219 if (changed) { 220 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 221 // Something changed about the intrinsic's virtuality. 222 if ((flags & _intrinsic_virtual) != 0) { 223 // This is the first use of this intrinsic as a virtual call. 224 if (oflags != 0) { 225 // We already saw it as a non-virtual, so note both cases. 226 flags |= _intrinsic_both; 227 } 228 } else if ((oflags & _intrinsic_both) == 0) { 229 // This is the first use of this intrinsic as a non-virtual 230 flags |= _intrinsic_both; 231 } 232 } 233 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 234 } 235 // update the overall flags also: 236 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 237 return changed; 238 } 239 240 static char* format_flags(int flags, char* buf) { 241 buf[0] = 0; 242 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 243 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 244 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 245 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 246 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 247 if (buf[0] == 0) strcat(buf, ","); 248 assert(buf[0] == ',', "must be"); 249 return &buf[1]; 250 } 251 252 void Compile::print_intrinsic_statistics() { 253 char flagsbuf[100]; 254 ttyLocker ttyl; 255 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 256 tty->print_cr("Compiler intrinsic usage:"); 257 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 258 if (total == 0) total = 1; // avoid div0 in case of no successes 259 #define PRINT_STAT_LINE(name, c, f) \ 260 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 261 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 262 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 263 int flags = _intrinsic_hist_flags[id]; 264 juint count = _intrinsic_hist_count[id]; 265 if ((flags | count) != 0) { 266 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 267 } 268 } 269 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 270 if (xtty != NULL) xtty->tail("statistics"); 271 } 272 273 void Compile::print_statistics() { 274 { ttyLocker ttyl; 275 if (xtty != NULL) xtty->head("statistics type='opto'"); 276 Parse::print_statistics(); 277 PhaseCCP::print_statistics(); 278 PhaseRegAlloc::print_statistics(); 279 Scheduling::print_statistics(); 280 PhasePeephole::print_statistics(); 281 PhaseIdealLoop::print_statistics(); 282 if (xtty != NULL) xtty->tail("statistics"); 283 } 284 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 285 // put this under its own <statistics> element. 286 print_intrinsic_statistics(); 287 } 288 } 289 #endif //PRODUCT 290 291 // Support for bundling info 292 Bundle* Compile::node_bundling(const Node *n) { 293 assert(valid_bundle_info(n), "oob"); 294 return &_node_bundling_base[n->_idx]; 295 } 296 297 bool Compile::valid_bundle_info(const Node *n) { 298 return (_node_bundling_limit > n->_idx); 299 } 300 301 302 void Compile::gvn_replace_by(Node* n, Node* nn) { 303 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 304 Node* use = n->last_out(i); 305 bool is_in_table = initial_gvn()->hash_delete(use); 306 uint uses_found = 0; 307 for (uint j = 0; j < use->len(); j++) { 308 if (use->in(j) == n) { 309 if (j < use->req()) 310 use->set_req(j, nn); 311 else 312 use->set_prec(j, nn); 313 uses_found++; 314 } 315 } 316 if (is_in_table) { 317 // reinsert into table 318 initial_gvn()->hash_find_insert(use); 319 } 320 record_for_igvn(use); 321 i -= uses_found; // we deleted 1 or more copies of this edge 322 } 323 } 324 325 326 static inline bool not_a_node(const Node* n) { 327 if (n == NULL) return true; 328 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 329 if (*(address*)n == badAddress) return true; // kill by Node::destruct 330 return false; 331 } 332 333 // Identify all nodes that are reachable from below, useful. 334 // Use breadth-first pass that records state in a Unique_Node_List, 335 // recursive traversal is slower. 336 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 337 int estimated_worklist_size = live_nodes(); 338 useful.map( estimated_worklist_size, NULL ); // preallocate space 339 340 // Initialize worklist 341 if (root() != NULL) { useful.push(root()); } 342 // If 'top' is cached, declare it useful to preserve cached node 343 if( cached_top_node() ) { useful.push(cached_top_node()); } 344 345 // Push all useful nodes onto the list, breadthfirst 346 for( uint next = 0; next < useful.size(); ++next ) { 347 assert( next < unique(), "Unique useful nodes < total nodes"); 348 Node *n = useful.at(next); 349 uint max = n->len(); 350 for( uint i = 0; i < max; ++i ) { 351 Node *m = n->in(i); 352 if (not_a_node(m)) continue; 353 useful.push(m); 354 } 355 } 356 } 357 358 // Update dead_node_list with any missing dead nodes using useful 359 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 360 void Compile::update_dead_node_list(Unique_Node_List &useful) { 361 uint max_idx = unique(); 362 VectorSet& useful_node_set = useful.member_set(); 363 364 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 365 // If node with index node_idx is not in useful set, 366 // mark it as dead in dead node list. 367 if (! useful_node_set.test(node_idx) ) { 368 record_dead_node(node_idx); 369 } 370 } 371 } 372 373 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 374 int shift = 0; 375 for (int i = 0; i < inlines->length(); i++) { 376 CallGenerator* cg = inlines->at(i); 377 CallNode* call = cg->call_node(); 378 if (shift > 0) { 379 inlines->at_put(i-shift, cg); 380 } 381 if (!useful.member(call)) { 382 shift++; 383 } 384 } 385 inlines->trunc_to(inlines->length()-shift); 386 } 387 388 // Disconnect all useless nodes by disconnecting those at the boundary. 389 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 390 uint next = 0; 391 while (next < useful.size()) { 392 Node *n = useful.at(next++); 393 if (n->is_SafePoint()) { 394 // We're done with a parsing phase. Replaced nodes are not valid 395 // beyond that point. 396 n->as_SafePoint()->delete_replaced_nodes(); 397 } 398 // Use raw traversal of out edges since this code removes out edges 399 int max = n->outcnt(); 400 for (int j = 0; j < max; ++j) { 401 Node* child = n->raw_out(j); 402 if (! useful.member(child)) { 403 assert(!child->is_top() || child != top(), 404 "If top is cached in Compile object it is in useful list"); 405 // Only need to remove this out-edge to the useless node 406 n->raw_del_out(j); 407 --j; 408 --max; 409 } 410 } 411 if (n->outcnt() == 1 && n->has_special_unique_user()) { 412 record_for_igvn(n->unique_out()); 413 } 414 if (n->Opcode() == Op_AddP && CallLeafNode::has_only_g1_wb_pre_uses(n)) { 415 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 416 record_for_igvn(n->fast_out(i)); 417 } 418 } 419 } 420 // Remove useless macro and predicate opaq nodes 421 for (int i = C->macro_count()-1; i >= 0; i--) { 422 Node* n = C->macro_node(i); 423 if (!useful.member(n)) { 424 remove_macro_node(n); 425 } 426 } 427 // Remove useless CastII nodes with range check dependency 428 for (int i = range_check_cast_count() - 1; i >= 0; i--) { 429 Node* cast = range_check_cast_node(i); 430 if (!useful.member(cast)) { 431 remove_range_check_cast(cast); 432 } 433 } 434 // Remove useless expensive node 435 for (int i = C->expensive_count()-1; i >= 0; i--) { 436 Node* n = C->expensive_node(i); 437 if (!useful.member(n)) { 438 remove_expensive_node(n); 439 } 440 } 441 for (int i = C->shenandoah_barriers_count()-1; i >= 0; i--) { 442 Node* n = C->shenandoah_barrier(i); 443 if (!useful.member(n)) { 444 remove_shenandoah_barrier(n->as_ShenandoahBarrier()); 445 } 446 } 447 // clean up the late inline lists 448 remove_useless_late_inlines(&_string_late_inlines, useful); 449 remove_useless_late_inlines(&_boxing_late_inlines, useful); 450 remove_useless_late_inlines(&_late_inlines, useful); 451 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 452 } 453 454 //------------------------------frame_size_in_words----------------------------- 455 // frame_slots in units of words 456 int Compile::frame_size_in_words() const { 457 // shift is 0 in LP32 and 1 in LP64 458 const int shift = (LogBytesPerWord - LogBytesPerInt); 459 int words = _frame_slots >> shift; 460 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 461 return words; 462 } 463 464 // To bang the stack of this compiled method we use the stack size 465 // that the interpreter would need in case of a deoptimization. This 466 // removes the need to bang the stack in the deoptimization blob which 467 // in turn simplifies stack overflow handling. 468 int Compile::bang_size_in_bytes() const { 469 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 470 } 471 472 // ============================================================================ 473 //------------------------------CompileWrapper--------------------------------- 474 class CompileWrapper : public StackObj { 475 Compile *const _compile; 476 public: 477 CompileWrapper(Compile* compile); 478 479 ~CompileWrapper(); 480 }; 481 482 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 483 // the Compile* pointer is stored in the current ciEnv: 484 ciEnv* env = compile->env(); 485 assert(env == ciEnv::current(), "must already be a ciEnv active"); 486 assert(env->compiler_data() == NULL, "compile already active?"); 487 env->set_compiler_data(compile); 488 assert(compile == Compile::current(), "sanity"); 489 490 compile->set_type_dict(NULL); 491 compile->set_type_hwm(NULL); 492 compile->set_type_last_size(0); 493 compile->set_last_tf(NULL, NULL); 494 compile->set_indexSet_arena(NULL); 495 compile->set_indexSet_free_block_list(NULL); 496 compile->init_type_arena(); 497 Type::Initialize(compile); 498 _compile->set_scratch_buffer_blob(NULL); 499 _compile->begin_method(); 500 } 501 CompileWrapper::~CompileWrapper() { 502 _compile->end_method(); 503 if (_compile->scratch_buffer_blob() != NULL) 504 BufferBlob::free(_compile->scratch_buffer_blob()); 505 _compile->env()->set_compiler_data(NULL); 506 } 507 508 509 //----------------------------print_compile_messages--------------------------- 510 void Compile::print_compile_messages() { 511 #ifndef PRODUCT 512 // Check if recompiling 513 if (_subsume_loads == false && PrintOpto) { 514 // Recompiling without allowing machine instructions to subsume loads 515 tty->print_cr("*********************************************************"); 516 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 517 tty->print_cr("*********************************************************"); 518 } 519 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 520 // Recompiling without escape analysis 521 tty->print_cr("*********************************************************"); 522 tty->print_cr("** Bailout: Recompile without escape analysis **"); 523 tty->print_cr("*********************************************************"); 524 } 525 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 526 // Recompiling without boxing elimination 527 tty->print_cr("*********************************************************"); 528 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 529 tty->print_cr("*********************************************************"); 530 } 531 if (env()->break_at_compile()) { 532 // Open the debugger when compiling this method. 533 tty->print("### Breaking when compiling: "); 534 method()->print_short_name(); 535 tty->cr(); 536 BREAKPOINT; 537 } 538 539 if( PrintOpto ) { 540 if (is_osr_compilation()) { 541 tty->print("[OSR]%3d", _compile_id); 542 } else { 543 tty->print("%3d", _compile_id); 544 } 545 } 546 #endif 547 } 548 549 550 //-----------------------init_scratch_buffer_blob------------------------------ 551 // Construct a temporary BufferBlob and cache it for this compile. 552 void Compile::init_scratch_buffer_blob(int const_size) { 553 // If there is already a scratch buffer blob allocated and the 554 // constant section is big enough, use it. Otherwise free the 555 // current and allocate a new one. 556 BufferBlob* blob = scratch_buffer_blob(); 557 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 558 // Use the current blob. 559 } else { 560 if (blob != NULL) { 561 BufferBlob::free(blob); 562 } 563 564 ResourceMark rm; 565 _scratch_const_size = const_size; 566 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 567 blob = BufferBlob::create("Compile::scratch_buffer", size); 568 // Record the buffer blob for next time. 569 set_scratch_buffer_blob(blob); 570 // Have we run out of code space? 571 if (scratch_buffer_blob() == NULL) { 572 // Let CompilerBroker disable further compilations. 573 record_failure("Not enough space for scratch buffer in CodeCache"); 574 return; 575 } 576 } 577 578 // Initialize the relocation buffers 579 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 580 set_scratch_locs_memory(locs_buf); 581 } 582 583 584 //-----------------------scratch_emit_size------------------------------------- 585 // Helper function that computes size by emitting code 586 uint Compile::scratch_emit_size(const Node* n) { 587 // Start scratch_emit_size section. 588 set_in_scratch_emit_size(true); 589 590 // Emit into a trash buffer and count bytes emitted. 591 // This is a pretty expensive way to compute a size, 592 // but it works well enough if seldom used. 593 // All common fixed-size instructions are given a size 594 // method by the AD file. 595 // Note that the scratch buffer blob and locs memory are 596 // allocated at the beginning of the compile task, and 597 // may be shared by several calls to scratch_emit_size. 598 // The allocation of the scratch buffer blob is particularly 599 // expensive, since it has to grab the code cache lock. 600 BufferBlob* blob = this->scratch_buffer_blob(); 601 assert(blob != NULL, "Initialize BufferBlob at start"); 602 assert(blob->size() > MAX_inst_size, "sanity"); 603 relocInfo* locs_buf = scratch_locs_memory(); 604 address blob_begin = blob->content_begin(); 605 address blob_end = (address)locs_buf; 606 assert(blob->content_contains(blob_end), "sanity"); 607 CodeBuffer buf(blob_begin, blob_end - blob_begin); 608 buf.initialize_consts_size(_scratch_const_size); 609 buf.initialize_stubs_size(MAX_stubs_size); 610 assert(locs_buf != NULL, "sanity"); 611 int lsize = MAX_locs_size / 3; 612 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 613 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 614 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 615 616 // Do the emission. 617 618 Label fakeL; // Fake label for branch instructions. 619 Label* saveL = NULL; 620 uint save_bnum = 0; 621 bool is_branch = n->is_MachBranch(); 622 if (is_branch) { 623 MacroAssembler masm(&buf); 624 masm.bind(fakeL); 625 n->as_MachBranch()->save_label(&saveL, &save_bnum); 626 n->as_MachBranch()->label_set(&fakeL, 0); 627 } 628 n->emit(buf, this->regalloc()); 629 630 // Emitting into the scratch buffer should not fail 631 assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason())); 632 633 if (is_branch) // Restore label. 634 n->as_MachBranch()->label_set(saveL, save_bnum); 635 636 // End scratch_emit_size section. 637 set_in_scratch_emit_size(false); 638 639 return buf.insts_size(); 640 } 641 642 643 // ============================================================================ 644 //------------------------------Compile standard------------------------------- 645 debug_only( int Compile::_debug_idx = 100000; ) 646 647 // Compile a method. entry_bci is -1 for normal compilations and indicates 648 // the continuation bci for on stack replacement. 649 650 651 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 652 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 653 : Phase(Compiler), 654 _env(ci_env), 655 _log(ci_env->log()), 656 _compile_id(ci_env->compile_id()), 657 _save_argument_registers(false), 658 _stub_name(NULL), 659 _stub_function(NULL), 660 _stub_entry_point(NULL), 661 _method(target), 662 _entry_bci(osr_bci), 663 _initial_gvn(NULL), 664 _for_igvn(NULL), 665 _warm_calls(NULL), 666 _subsume_loads(subsume_loads), 667 _do_escape_analysis(do_escape_analysis), 668 _eliminate_boxing(eliminate_boxing), 669 _failure_reason(NULL), 670 _code_buffer("Compile::Fill_buffer"), 671 _orig_pc_slot(0), 672 _orig_pc_slot_offset_in_bytes(0), 673 _has_method_handle_invokes(false), 674 _mach_constant_base_node(NULL), 675 _node_bundling_limit(0), 676 _node_bundling_base(NULL), 677 _java_calls(0), 678 _inner_loops(0), 679 _scratch_const_size(-1), 680 _in_scratch_emit_size(false), 681 _dead_node_list(comp_arena()), 682 _dead_node_count(0), 683 #ifndef PRODUCT 684 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 685 _in_dump_cnt(0), 686 _printer(IdealGraphPrinter::printer()), 687 #endif 688 _congraph(NULL), 689 _comp_arena(mtCompiler), 690 _node_arena(mtCompiler), 691 _old_arena(mtCompiler), 692 _Compile_types(mtCompiler), 693 _replay_inline_data(NULL), 694 _late_inlines(comp_arena(), 2, 0, NULL), 695 _string_late_inlines(comp_arena(), 2, 0, NULL), 696 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 697 _late_inlines_pos(0), 698 _number_of_mh_late_inlines(0), 699 _inlining_progress(false), 700 _inlining_incrementally(false), 701 _print_inlining_list(NULL), 702 _print_inlining_idx(0), 703 _interpreter_frame_size(0), 704 _max_node_limit(MaxNodeLimit) { 705 C = this; 706 707 CompileWrapper cw(this); 708 #ifndef PRODUCT 709 if (TimeCompiler2) { 710 tty->print(" "); 711 target->holder()->name()->print(); 712 tty->print("."); 713 target->print_short_name(); 714 tty->print(" "); 715 } 716 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 717 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 718 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 719 if (!print_opto_assembly) { 720 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 721 if (print_assembly && !Disassembler::can_decode()) { 722 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 723 print_opto_assembly = true; 724 } 725 } 726 set_print_assembly(print_opto_assembly); 727 set_parsed_irreducible_loop(false); 728 729 if (method()->has_option("ReplayInline")) { 730 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 731 } 732 #endif 733 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 734 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 735 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 736 737 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 738 // Make sure the method being compiled gets its own MDO, 739 // so we can at least track the decompile_count(). 740 // Need MDO to record RTM code generation state. 741 method()->ensure_method_data(); 742 } 743 744 Init(::AliasLevel); 745 746 747 print_compile_messages(); 748 749 _ilt = InlineTree::build_inline_tree_root(); 750 751 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 752 assert(num_alias_types() >= AliasIdxRaw, ""); 753 754 #define MINIMUM_NODE_HASH 1023 755 // Node list that Iterative GVN will start with 756 Unique_Node_List for_igvn(comp_arena()); 757 set_for_igvn(&for_igvn); 758 759 // GVN that will be run immediately on new nodes 760 uint estimated_size = method()->code_size()*4+64; 761 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 762 PhaseGVN gvn(node_arena(), estimated_size); 763 set_initial_gvn(&gvn); 764 765 if (print_inlining() || print_intrinsics()) { 766 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 767 } 768 { // Scope for timing the parser 769 TracePhase t3("parse", &_t_parser, true); 770 771 // Put top into the hash table ASAP. 772 initial_gvn()->transform_no_reclaim(top()); 773 774 // Set up tf(), start(), and find a CallGenerator. 775 CallGenerator* cg = NULL; 776 if (is_osr_compilation()) { 777 const TypeTuple *domain = StartOSRNode::osr_domain(); 778 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 779 init_tf(TypeFunc::make(domain, range)); 780 StartNode* s = new (this) StartOSRNode(root(), domain); 781 initial_gvn()->set_type_bottom(s); 782 init_start(s); 783 cg = CallGenerator::for_osr(method(), entry_bci()); 784 } else { 785 // Normal case. 786 init_tf(TypeFunc::make(method())); 787 StartNode* s = new (this) StartNode(root(), tf()->domain()); 788 initial_gvn()->set_type_bottom(s); 789 init_start(s); 790 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && (UseG1GC || UseShenandoahGC)) { 791 // With java.lang.ref.reference.get() we must go through the 792 // intrinsic when G1 is enabled - even when get() is the root 793 // method of the compile - so that, if necessary, the value in 794 // the referent field of the reference object gets recorded by 795 // the pre-barrier code. 796 // Specifically, if G1 is enabled, the value in the referent 797 // field is recorded by the G1 SATB pre barrier. This will 798 // result in the referent being marked live and the reference 799 // object removed from the list of discovered references during 800 // reference processing. 801 cg = find_intrinsic(method(), false); 802 } 803 if (cg == NULL) { 804 float past_uses = method()->interpreter_invocation_count(); 805 float expected_uses = past_uses; 806 cg = CallGenerator::for_inline(method(), expected_uses); 807 } 808 } 809 if (failing()) return; 810 if (cg == NULL) { 811 record_method_not_compilable_all_tiers("cannot parse method"); 812 return; 813 } 814 JVMState* jvms = build_start_state(start(), tf()); 815 if ((jvms = cg->generate(jvms)) == NULL) { 816 if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) { 817 record_method_not_compilable("method parse failed"); 818 } 819 return; 820 } 821 GraphKit kit(jvms); 822 823 if (!kit.stopped()) { 824 // Accept return values, and transfer control we know not where. 825 // This is done by a special, unique ReturnNode bound to root. 826 return_values(kit.jvms()); 827 } 828 829 if (kit.has_exceptions()) { 830 // Any exceptions that escape from this call must be rethrown 831 // to whatever caller is dynamically above us on the stack. 832 // This is done by a special, unique RethrowNode bound to root. 833 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 834 } 835 836 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 837 838 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 839 inline_string_calls(true); 840 } 841 842 if (failing()) return; 843 844 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 845 846 // Remove clutter produced by parsing. 847 if (!failing()) { 848 ResourceMark rm; 849 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 850 } 851 } 852 853 // Note: Large methods are capped off in do_one_bytecode(). 854 if (failing()) return; 855 856 // After parsing, node notes are no longer automagic. 857 // They must be propagated by register_new_node_with_optimizer(), 858 // clone(), or the like. 859 set_default_node_notes(NULL); 860 861 for (;;) { 862 int successes = Inline_Warm(); 863 if (failing()) return; 864 if (successes == 0) break; 865 } 866 867 // Drain the list. 868 Finish_Warm(); 869 #ifndef PRODUCT 870 if (_printer) { 871 _printer->print_inlining(this); 872 } 873 #endif 874 875 if (failing()) return; 876 NOT_PRODUCT( verify_graph_edges(); ) 877 878 // Now optimize 879 Optimize(); 880 if (failing()) return; 881 NOT_PRODUCT( verify_graph_edges(); ) 882 883 #ifndef PRODUCT 884 if (PrintIdeal) { 885 ttyLocker ttyl; // keep the following output all in one block 886 // This output goes directly to the tty, not the compiler log. 887 // To enable tools to match it up with the compilation activity, 888 // be sure to tag this tty output with the compile ID. 889 if (xtty != NULL) { 890 xtty->head("ideal compile_id='%d'%s", compile_id(), 891 is_osr_compilation() ? " compile_kind='osr'" : 892 ""); 893 } 894 root()->dump(9999); 895 if (xtty != NULL) { 896 xtty->tail("ideal"); 897 } 898 } 899 #endif 900 901 NOT_PRODUCT( verify_barriers(); ) 902 903 // Dump compilation data to replay it. 904 if (method()->has_option("DumpReplay")) { 905 env()->dump_replay_data(_compile_id); 906 } 907 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 908 env()->dump_inline_data(_compile_id); 909 } 910 911 // Now that we know the size of all the monitors we can add a fixed slot 912 // for the original deopt pc. 913 914 _orig_pc_slot = fixed_slots(); 915 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 916 set_fixed_slots(next_slot); 917 918 // Compute when to use implicit null checks. Used by matching trap based 919 // nodes and NullCheck optimization. 920 set_allowed_deopt_reasons(); 921 922 // Now generate code 923 Code_Gen(); 924 if (failing()) return; 925 926 // Check if we want to skip execution of all compiled code. 927 { 928 #ifndef PRODUCT 929 if (OptoNoExecute) { 930 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 931 return; 932 } 933 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 934 #endif 935 936 if (is_osr_compilation()) { 937 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 938 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 939 } else { 940 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 941 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 942 } 943 944 #ifdef BUILTIN_SIM 945 char *method_name = NULL; 946 AArch64Simulator *sim = NULL; 947 size_t len = 65536; 948 if (NotifySimulator) { 949 method_name = new char[len]; 950 } 951 if (method_name) { 952 unsigned char *entry = code_buffer()->insts_begin(); 953 stringStream st(method_name, 400); 954 if (_entry_bci != InvocationEntryBci) { 955 st.print("osr:"); 956 } 957 _method->holder()->name()->print_symbol_on(&st); 958 // convert '/' separators in class name into '.' separator 959 for (unsigned i = 0; i < len; i++) { 960 if (method_name[i] == '/') { 961 method_name[i] = '.'; 962 } else if (method_name[i] == '\0') { 963 break; 964 } 965 } 966 st.print("."); 967 _method->name()->print_symbol_on(&st); 968 _method->signature()->as_symbol()->print_symbol_on(&st); 969 sim = AArch64Simulator::get_current(UseSimulatorCache, DisableBCCheck); 970 sim->notifyCompile(method_name, entry); 971 sim->notifyRelocate(entry, 0); 972 } 973 #endif 974 975 env()->register_method(_method, _entry_bci, 976 &_code_offsets, 977 _orig_pc_slot_offset_in_bytes, 978 code_buffer(), 979 frame_size_in_words(), _oop_map_set, 980 &_handler_table, &_inc_table, 981 compiler, 982 env()->comp_level(), 983 has_unsafe_access(), 984 SharedRuntime::is_wide_vector(max_vector_size()), 985 rtm_state() 986 ); 987 988 if (log() != NULL) // Print code cache state into compiler log 989 log()->code_cache_state(); 990 } 991 } 992 993 //------------------------------Compile---------------------------------------- 994 // Compile a runtime stub 995 Compile::Compile( ciEnv* ci_env, 996 TypeFunc_generator generator, 997 address stub_function, 998 const char *stub_name, 999 int is_fancy_jump, 1000 bool pass_tls, 1001 bool save_arg_registers, 1002 bool return_pc ) 1003 : Phase(Compiler), 1004 _env(ci_env), 1005 _log(ci_env->log()), 1006 _compile_id(0), 1007 _save_argument_registers(save_arg_registers), 1008 _method(NULL), 1009 _stub_name(stub_name), 1010 _stub_function(stub_function), 1011 _stub_entry_point(NULL), 1012 _entry_bci(InvocationEntryBci), 1013 _initial_gvn(NULL), 1014 _for_igvn(NULL), 1015 _warm_calls(NULL), 1016 _orig_pc_slot(0), 1017 _orig_pc_slot_offset_in_bytes(0), 1018 _subsume_loads(true), 1019 _do_escape_analysis(false), 1020 _eliminate_boxing(false), 1021 _failure_reason(NULL), 1022 _code_buffer("Compile::Fill_buffer"), 1023 _has_method_handle_invokes(false), 1024 _mach_constant_base_node(NULL), 1025 _node_bundling_limit(0), 1026 _node_bundling_base(NULL), 1027 _java_calls(0), 1028 _inner_loops(0), 1029 #ifndef PRODUCT 1030 _trace_opto_output(TraceOptoOutput), 1031 _in_dump_cnt(0), 1032 _printer(NULL), 1033 #endif 1034 _comp_arena(mtCompiler), 1035 _node_arena(mtCompiler), 1036 _old_arena(mtCompiler), 1037 _Compile_types(mtCompiler), 1038 _dead_node_list(comp_arena()), 1039 _dead_node_count(0), 1040 _congraph(NULL), 1041 _replay_inline_data(NULL), 1042 _number_of_mh_late_inlines(0), 1043 _inlining_progress(false), 1044 _inlining_incrementally(false), 1045 _print_inlining_list(NULL), 1046 _print_inlining_idx(0), 1047 _allowed_reasons(0), 1048 _interpreter_frame_size(0), 1049 _max_node_limit(MaxNodeLimit) { 1050 C = this; 1051 1052 #ifndef PRODUCT 1053 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 1054 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 1055 set_print_assembly(PrintFrameConverterAssembly); 1056 set_parsed_irreducible_loop(false); 1057 #endif 1058 set_has_irreducible_loop(false); // no loops 1059 1060 CompileWrapper cw(this); 1061 Init(/*AliasLevel=*/ 0); 1062 init_tf((*generator)()); 1063 1064 { 1065 // The following is a dummy for the sake of GraphKit::gen_stub 1066 Unique_Node_List for_igvn(comp_arena()); 1067 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1068 PhaseGVN gvn(Thread::current()->resource_area(),255); 1069 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1070 gvn.transform_no_reclaim(top()); 1071 1072 GraphKit kit; 1073 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1074 } 1075 1076 NOT_PRODUCT( verify_graph_edges(); ) 1077 Code_Gen(); 1078 if (failing()) return; 1079 1080 1081 // Entry point will be accessed using compile->stub_entry_point(); 1082 if (code_buffer() == NULL) { 1083 Matcher::soft_match_failure(); 1084 } else { 1085 if (PrintAssembly && (WizardMode || Verbose)) 1086 tty->print_cr("### Stub::%s", stub_name); 1087 1088 if (!failing()) { 1089 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1090 1091 // Make the NMethod 1092 // For now we mark the frame as never safe for profile stackwalking 1093 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1094 code_buffer(), 1095 CodeOffsets::frame_never_safe, 1096 // _code_offsets.value(CodeOffsets::Frame_Complete), 1097 frame_size_in_words(), 1098 _oop_map_set, 1099 save_arg_registers); 1100 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1101 1102 _stub_entry_point = rs->entry_point(); 1103 } 1104 } 1105 } 1106 1107 //------------------------------Init------------------------------------------- 1108 // Prepare for a single compilation 1109 void Compile::Init(int aliaslevel) { 1110 _unique = 0; 1111 _regalloc = NULL; 1112 1113 _tf = NULL; // filled in later 1114 _top = NULL; // cached later 1115 _matcher = NULL; // filled in later 1116 _cfg = NULL; // filled in later 1117 1118 set_24_bit_selection_and_mode(Use24BitFP, false); 1119 1120 _node_note_array = NULL; 1121 _default_node_notes = NULL; 1122 1123 _immutable_memory = NULL; // filled in at first inquiry 1124 1125 // Globally visible Nodes 1126 // First set TOP to NULL to give safe behavior during creation of RootNode 1127 set_cached_top_node(NULL); 1128 set_root(new (this) RootNode()); 1129 // Now that you have a Root to point to, create the real TOP 1130 set_cached_top_node( new (this) ConNode(Type::TOP) ); 1131 set_recent_alloc(NULL, NULL); 1132 1133 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1134 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1135 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1136 env()->set_dependencies(new Dependencies(env())); 1137 1138 _fixed_slots = 0; 1139 set_has_split_ifs(false); 1140 set_has_loops(has_method() && method()->has_loops()); // first approximation 1141 set_has_stringbuilder(false); 1142 set_has_boxed_value(false); 1143 _trap_can_recompile = false; // no traps emitted yet 1144 _major_progress = true; // start out assuming good things will happen 1145 set_has_unsafe_access(false); 1146 set_max_vector_size(0); 1147 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1148 set_decompile_count(0); 1149 1150 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1151 set_num_loop_opts(LoopOptsCount); 1152 set_do_inlining(Inline); 1153 set_max_inline_size(MaxInlineSize); 1154 set_freq_inline_size(FreqInlineSize); 1155 set_do_scheduling(OptoScheduling); 1156 set_do_count_invocations(false); 1157 set_do_method_data_update(false); 1158 set_rtm_state(NoRTM); // No RTM lock eliding by default 1159 method_has_option_value("MaxNodeLimit", _max_node_limit); 1160 #if INCLUDE_RTM_OPT 1161 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1162 int rtm_state = method()->method_data()->rtm_state(); 1163 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1164 // Don't generate RTM lock eliding code. 1165 set_rtm_state(NoRTM); 1166 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1167 // Generate RTM lock eliding code without abort ratio calculation code. 1168 set_rtm_state(UseRTM); 1169 } else if (UseRTMDeopt) { 1170 // Generate RTM lock eliding code and include abort ratio calculation 1171 // code if UseRTMDeopt is on. 1172 set_rtm_state(ProfileRTM); 1173 } 1174 } 1175 #endif 1176 if (debug_info()->recording_non_safepoints()) { 1177 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1178 (comp_arena(), 8, 0, NULL)); 1179 set_default_node_notes(Node_Notes::make(this)); 1180 } 1181 1182 // // -- Initialize types before each compile -- 1183 // // Update cached type information 1184 // if( _method && _method->constants() ) 1185 // Type::update_loaded_types(_method, _method->constants()); 1186 1187 // Init alias_type map. 1188 if (!_do_escape_analysis && aliaslevel == 3) 1189 aliaslevel = 2; // No unique types without escape analysis 1190 _AliasLevel = aliaslevel; 1191 const int grow_ats = 16; 1192 _max_alias_types = grow_ats; 1193 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1194 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1195 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1196 { 1197 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1198 } 1199 // Initialize the first few types. 1200 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1201 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1202 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1203 _num_alias_types = AliasIdxRaw+1; 1204 // Zero out the alias type cache. 1205 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1206 // A NULL adr_type hits in the cache right away. Preload the right answer. 1207 probe_alias_cache(NULL)->_index = AliasIdxTop; 1208 1209 _intrinsics = NULL; 1210 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1211 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1212 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1213 _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1214 _shenandoah_barriers = new(comp_arena()) GrowableArray<ShenandoahBarrierNode*>(comp_arena(), 8, 0, NULL); 1215 register_library_intrinsics(); 1216 } 1217 1218 //---------------------------init_start---------------------------------------- 1219 // Install the StartNode on this compile object. 1220 void Compile::init_start(StartNode* s) { 1221 if (failing()) 1222 return; // already failing 1223 assert(s == start(), ""); 1224 } 1225 1226 StartNode* Compile::start() const { 1227 assert(!failing(), ""); 1228 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1229 Node* start = root()->fast_out(i); 1230 if( start->is_Start() ) 1231 return start->as_Start(); 1232 } 1233 fatal("Did not find Start node!"); 1234 return NULL; 1235 } 1236 1237 //-------------------------------immutable_memory------------------------------------- 1238 // Access immutable memory 1239 Node* Compile::immutable_memory() { 1240 if (_immutable_memory != NULL) { 1241 return _immutable_memory; 1242 } 1243 StartNode* s = start(); 1244 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1245 Node *p = s->fast_out(i); 1246 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1247 _immutable_memory = p; 1248 return _immutable_memory; 1249 } 1250 } 1251 ShouldNotReachHere(); 1252 return NULL; 1253 } 1254 1255 //----------------------set_cached_top_node------------------------------------ 1256 // Install the cached top node, and make sure Node::is_top works correctly. 1257 void Compile::set_cached_top_node(Node* tn) { 1258 if (tn != NULL) verify_top(tn); 1259 Node* old_top = _top; 1260 _top = tn; 1261 // Calling Node::setup_is_top allows the nodes the chance to adjust 1262 // their _out arrays. 1263 if (_top != NULL) _top->setup_is_top(); 1264 if (old_top != NULL) old_top->setup_is_top(); 1265 assert(_top == NULL || top()->is_top(), ""); 1266 } 1267 1268 #ifdef ASSERT 1269 uint Compile::count_live_nodes_by_graph_walk() { 1270 Unique_Node_List useful(comp_arena()); 1271 // Get useful node list by walking the graph. 1272 identify_useful_nodes(useful); 1273 return useful.size(); 1274 } 1275 1276 void Compile::print_missing_nodes() { 1277 1278 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1279 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1280 return; 1281 } 1282 1283 // This is an expensive function. It is executed only when the user 1284 // specifies VerifyIdealNodeCount option or otherwise knows the 1285 // additional work that needs to be done to identify reachable nodes 1286 // by walking the flow graph and find the missing ones using 1287 // _dead_node_list. 1288 1289 Unique_Node_List useful(comp_arena()); 1290 // Get useful node list by walking the graph. 1291 identify_useful_nodes(useful); 1292 1293 uint l_nodes = C->live_nodes(); 1294 uint l_nodes_by_walk = useful.size(); 1295 1296 if (l_nodes != l_nodes_by_walk) { 1297 if (_log != NULL) { 1298 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1299 _log->stamp(); 1300 _log->end_head(); 1301 } 1302 VectorSet& useful_member_set = useful.member_set(); 1303 int last_idx = l_nodes_by_walk; 1304 for (int i = 0; i < last_idx; i++) { 1305 if (useful_member_set.test(i)) { 1306 if (_dead_node_list.test(i)) { 1307 if (_log != NULL) { 1308 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1309 } 1310 if (PrintIdealNodeCount) { 1311 // Print the log message to tty 1312 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1313 useful.at(i)->dump(); 1314 } 1315 } 1316 } 1317 else if (! _dead_node_list.test(i)) { 1318 if (_log != NULL) { 1319 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1320 } 1321 if (PrintIdealNodeCount) { 1322 // Print the log message to tty 1323 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1324 } 1325 } 1326 } 1327 if (_log != NULL) { 1328 _log->tail("mismatched_nodes"); 1329 } 1330 } 1331 } 1332 #endif 1333 1334 #ifndef PRODUCT 1335 void Compile::verify_top(Node* tn) const { 1336 if (tn != NULL) { 1337 assert(tn->is_Con(), "top node must be a constant"); 1338 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1339 assert(tn->in(0) != NULL, "must have live top node"); 1340 } 1341 } 1342 #endif 1343 1344 1345 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1346 1347 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1348 guarantee(arr != NULL, ""); 1349 int num_blocks = arr->length(); 1350 if (grow_by < num_blocks) grow_by = num_blocks; 1351 int num_notes = grow_by * _node_notes_block_size; 1352 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1353 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1354 while (num_notes > 0) { 1355 arr->append(notes); 1356 notes += _node_notes_block_size; 1357 num_notes -= _node_notes_block_size; 1358 } 1359 assert(num_notes == 0, "exact multiple, please"); 1360 } 1361 1362 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1363 if (source == NULL || dest == NULL) return false; 1364 1365 if (dest->is_Con()) 1366 return false; // Do not push debug info onto constants. 1367 1368 #ifdef ASSERT 1369 // Leave a bread crumb trail pointing to the original node: 1370 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1371 dest->set_debug_orig(source); 1372 } 1373 #endif 1374 1375 if (node_note_array() == NULL) 1376 return false; // Not collecting any notes now. 1377 1378 // This is a copy onto a pre-existing node, which may already have notes. 1379 // If both nodes have notes, do not overwrite any pre-existing notes. 1380 Node_Notes* source_notes = node_notes_at(source->_idx); 1381 if (source_notes == NULL || source_notes->is_clear()) return false; 1382 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1383 if (dest_notes == NULL || dest_notes->is_clear()) { 1384 return set_node_notes_at(dest->_idx, source_notes); 1385 } 1386 1387 Node_Notes merged_notes = (*source_notes); 1388 // The order of operations here ensures that dest notes will win... 1389 merged_notes.update_from(dest_notes); 1390 return set_node_notes_at(dest->_idx, &merged_notes); 1391 } 1392 1393 1394 //--------------------------allow_range_check_smearing------------------------- 1395 // Gating condition for coalescing similar range checks. 1396 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1397 // single covering check that is at least as strong as any of them. 1398 // If the optimization succeeds, the simplified (strengthened) range check 1399 // will always succeed. If it fails, we will deopt, and then give up 1400 // on the optimization. 1401 bool Compile::allow_range_check_smearing() const { 1402 // If this method has already thrown a range-check, 1403 // assume it was because we already tried range smearing 1404 // and it failed. 1405 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1406 return !already_trapped; 1407 } 1408 1409 1410 //------------------------------flatten_alias_type----------------------------- 1411 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1412 int offset = tj->offset(); 1413 TypePtr::PTR ptr = tj->ptr(); 1414 1415 // Known instance (scalarizable allocation) alias only with itself. 1416 bool is_known_inst = tj->isa_oopptr() != NULL && 1417 tj->is_oopptr()->is_known_instance(); 1418 1419 // Process weird unsafe references. 1420 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1421 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1422 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1423 tj = TypeOopPtr::BOTTOM; 1424 ptr = tj->ptr(); 1425 offset = tj->offset(); 1426 } 1427 1428 // Array pointers need some flattening 1429 const TypeAryPtr *ta = tj->isa_aryptr(); 1430 if (ta && ta->is_stable()) { 1431 // Erase stability property for alias analysis. 1432 tj = ta = ta->cast_to_stable(false); 1433 } 1434 if( ta && is_known_inst ) { 1435 if ( offset != Type::OffsetBot && 1436 offset > arrayOopDesc::length_offset_in_bytes() ) { 1437 offset = Type::OffsetBot; // Flatten constant access into array body only 1438 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1439 } 1440 } else if( ta && _AliasLevel >= 2 ) { 1441 // For arrays indexed by constant indices, we flatten the alias 1442 // space to include all of the array body. Only the header, klass 1443 // and array length can be accessed un-aliased. 1444 if( offset != Type::OffsetBot ) { 1445 if( ta->const_oop() ) { // MethodData* or Method* 1446 offset = Type::OffsetBot; // Flatten constant access into array body 1447 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1448 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1449 // range is OK as-is. 1450 tj = ta = TypeAryPtr::RANGE; 1451 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1452 tj = TypeInstPtr::KLASS; // all klass loads look alike 1453 ta = TypeAryPtr::RANGE; // generic ignored junk 1454 ptr = TypePtr::BotPTR; 1455 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1456 tj = TypeInstPtr::MARK; 1457 ta = TypeAryPtr::RANGE; // generic ignored junk 1458 ptr = TypePtr::BotPTR; 1459 } else if (offset == BrooksPointer::byte_offset() && UseShenandoahGC) { 1460 // Need to distinguish brooks ptr as is. 1461 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1462 } else { // Random constant offset into array body 1463 offset = Type::OffsetBot; // Flatten constant access into array body 1464 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1465 } 1466 } 1467 // Arrays of fixed size alias with arrays of unknown size. 1468 if (ta->size() != TypeInt::POS) { 1469 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1470 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1471 } 1472 // Arrays of known objects become arrays of unknown objects. 1473 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1474 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1475 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1476 } 1477 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1478 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1479 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1480 } 1481 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1482 // cannot be distinguished by bytecode alone. 1483 if (ta->elem() == TypeInt::BOOL) { 1484 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1485 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1486 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1487 } 1488 // During the 2nd round of IterGVN, NotNull castings are removed. 1489 // Make sure the Bottom and NotNull variants alias the same. 1490 // Also, make sure exact and non-exact variants alias the same. 1491 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1492 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1493 } 1494 } 1495 1496 // Oop pointers need some flattening 1497 const TypeInstPtr *to = tj->isa_instptr(); 1498 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1499 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1500 if( ptr == TypePtr::Constant ) { 1501 if (to->klass() != ciEnv::current()->Class_klass() || 1502 offset < k->size_helper() * wordSize) { 1503 // No constant oop pointers (such as Strings); they alias with 1504 // unknown strings. 1505 assert(!is_known_inst, "not scalarizable allocation"); 1506 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1507 } 1508 } else if( is_known_inst ) { 1509 tj = to; // Keep NotNull and klass_is_exact for instance type 1510 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1511 // During the 2nd round of IterGVN, NotNull castings are removed. 1512 // Make sure the Bottom and NotNull variants alias the same. 1513 // Also, make sure exact and non-exact variants alias the same. 1514 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1515 } 1516 if (to->speculative() != NULL) { 1517 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1518 } 1519 // Canonicalize the holder of this field 1520 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1521 // First handle header references such as a LoadKlassNode, even if the 1522 // object's klass is unloaded at compile time (4965979). 1523 if (!is_known_inst) { // Do it only for non-instance types 1524 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1525 } 1526 } else if ((offset != BrooksPointer::byte_offset() || !UseShenandoahGC) && (offset < 0 || offset >= k->size_helper() * wordSize)) { 1527 // Static fields are in the space above the normal instance 1528 // fields in the java.lang.Class instance. 1529 if (to->klass() != ciEnv::current()->Class_klass()) { 1530 to = NULL; 1531 tj = TypeOopPtr::BOTTOM; 1532 offset = tj->offset(); 1533 } 1534 } else { 1535 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1536 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1537 if( is_known_inst ) { 1538 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1539 } else { 1540 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1541 } 1542 } 1543 } 1544 } 1545 1546 // Klass pointers to object array klasses need some flattening 1547 const TypeKlassPtr *tk = tj->isa_klassptr(); 1548 if( tk ) { 1549 // If we are referencing a field within a Klass, we need 1550 // to assume the worst case of an Object. Both exact and 1551 // inexact types must flatten to the same alias class so 1552 // use NotNull as the PTR. 1553 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1554 1555 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1556 TypeKlassPtr::OBJECT->klass(), 1557 offset); 1558 } 1559 1560 ciKlass* klass = tk->klass(); 1561 if( klass->is_obj_array_klass() ) { 1562 ciKlass* k = TypeAryPtr::OOPS->klass(); 1563 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1564 k = TypeInstPtr::BOTTOM->klass(); 1565 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1566 } 1567 1568 // Check for precise loads from the primary supertype array and force them 1569 // to the supertype cache alias index. Check for generic array loads from 1570 // the primary supertype array and also force them to the supertype cache 1571 // alias index. Since the same load can reach both, we need to merge 1572 // these 2 disparate memories into the same alias class. Since the 1573 // primary supertype array is read-only, there's no chance of confusion 1574 // where we bypass an array load and an array store. 1575 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1576 if (offset == Type::OffsetBot || 1577 (offset >= primary_supers_offset && 1578 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1579 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1580 offset = in_bytes(Klass::secondary_super_cache_offset()); 1581 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1582 } 1583 } 1584 1585 // Flatten all Raw pointers together. 1586 if (tj->base() == Type::RawPtr) 1587 tj = TypeRawPtr::BOTTOM; 1588 1589 if (tj->base() == Type::AnyPtr) 1590 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1591 1592 // Flatten all to bottom for now 1593 switch( _AliasLevel ) { 1594 case 0: 1595 tj = TypePtr::BOTTOM; 1596 break; 1597 case 1: // Flatten to: oop, static, field or array 1598 switch (tj->base()) { 1599 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1600 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1601 case Type::AryPtr: // do not distinguish arrays at all 1602 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1603 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1604 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1605 default: ShouldNotReachHere(); 1606 } 1607 break; 1608 case 2: // No collapsing at level 2; keep all splits 1609 case 3: // No collapsing at level 3; keep all splits 1610 break; 1611 default: 1612 Unimplemented(); 1613 } 1614 1615 offset = tj->offset(); 1616 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1617 1618 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1619 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1620 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1621 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1622 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1623 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1624 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) || 1625 (offset == BrooksPointer::byte_offset() && tj->base() == Type::AryPtr && UseShenandoahGC), 1626 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1627 assert( tj->ptr() != TypePtr::TopPTR && 1628 tj->ptr() != TypePtr::AnyNull && 1629 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1630 // assert( tj->ptr() != TypePtr::Constant || 1631 // tj->base() == Type::RawPtr || 1632 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1633 1634 return tj; 1635 } 1636 1637 void Compile::AliasType::Init(int i, const TypePtr* at) { 1638 _index = i; 1639 _adr_type = at; 1640 _field = NULL; 1641 _element = NULL; 1642 _is_rewritable = true; // default 1643 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1644 if (atoop != NULL && atoop->is_known_instance()) { 1645 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1646 _general_index = Compile::current()->get_alias_index(gt); 1647 } else { 1648 _general_index = 0; 1649 } 1650 } 1651 1652 BasicType Compile::AliasType::basic_type() const { 1653 if (element() != NULL) { 1654 const Type* element = adr_type()->is_aryptr()->elem(); 1655 return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type(); 1656 } if (field() != NULL) { 1657 return field()->layout_type(); 1658 } else { 1659 return T_ILLEGAL; // unknown 1660 } 1661 } 1662 1663 //---------------------------------print_on------------------------------------ 1664 #ifndef PRODUCT 1665 void Compile::AliasType::print_on(outputStream* st) { 1666 if (index() < 10) 1667 st->print("@ <%d> ", index()); 1668 else st->print("@ <%d>", index()); 1669 st->print(is_rewritable() ? " " : " RO"); 1670 int offset = adr_type()->offset(); 1671 if (offset == Type::OffsetBot) 1672 st->print(" +any"); 1673 else st->print(" +%-3d", offset); 1674 st->print(" in "); 1675 adr_type()->dump_on(st); 1676 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1677 if (field() != NULL && tjp) { 1678 if (tjp->klass() != field()->holder() || 1679 tjp->offset() != field()->offset_in_bytes()) { 1680 st->print(" != "); 1681 field()->print(); 1682 st->print(" ***"); 1683 } 1684 } 1685 } 1686 1687 void print_alias_types() { 1688 Compile* C = Compile::current(); 1689 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1690 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1691 C->alias_type(idx)->print_on(tty); 1692 tty->cr(); 1693 } 1694 } 1695 #endif 1696 1697 1698 //----------------------------probe_alias_cache-------------------------------- 1699 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1700 intptr_t key = (intptr_t) adr_type; 1701 key ^= key >> logAliasCacheSize; 1702 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1703 } 1704 1705 1706 //-----------------------------grow_alias_types-------------------------------- 1707 void Compile::grow_alias_types() { 1708 const int old_ats = _max_alias_types; // how many before? 1709 const int new_ats = old_ats; // how many more? 1710 const int grow_ats = old_ats+new_ats; // how many now? 1711 _max_alias_types = grow_ats; 1712 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1713 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1714 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1715 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1716 } 1717 1718 1719 //--------------------------------find_alias_type------------------------------ 1720 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1721 if (_AliasLevel == 0) 1722 return alias_type(AliasIdxBot); 1723 1724 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1725 if (ace->_adr_type == adr_type) { 1726 return alias_type(ace->_index); 1727 } 1728 1729 // Handle special cases. 1730 if (adr_type == NULL) return alias_type(AliasIdxTop); 1731 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1732 1733 // Do it the slow way. 1734 const TypePtr* flat = flatten_alias_type(adr_type); 1735 1736 #ifdef ASSERT 1737 assert(flat == flatten_alias_type(flat), "idempotent"); 1738 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1739 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1740 const TypeOopPtr* foop = flat->is_oopptr(); 1741 // Scalarizable allocations have exact klass always. 1742 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1743 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1744 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1745 } 1746 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1747 #endif 1748 1749 int idx = AliasIdxTop; 1750 for (int i = 0; i < num_alias_types(); i++) { 1751 if (alias_type(i)->adr_type() == flat) { 1752 idx = i; 1753 break; 1754 } 1755 } 1756 1757 if (idx == AliasIdxTop) { 1758 if (no_create) return NULL; 1759 // Grow the array if necessary. 1760 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1761 // Add a new alias type. 1762 idx = _num_alias_types++; 1763 _alias_types[idx]->Init(idx, flat); 1764 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1765 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1766 if (flat->isa_instptr()) { 1767 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1768 && flat->is_instptr()->klass() == env()->Class_klass()) 1769 alias_type(idx)->set_rewritable(false); 1770 } 1771 if (flat->isa_aryptr()) { 1772 #ifdef ASSERT 1773 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1774 // (T_BYTE has the weakest alignment and size restrictions...) 1775 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1776 #endif 1777 if (flat->offset() == TypePtr::OffsetBot) { 1778 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1779 } 1780 } 1781 if (flat->isa_klassptr()) { 1782 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1783 alias_type(idx)->set_rewritable(false); 1784 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1785 alias_type(idx)->set_rewritable(false); 1786 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1787 alias_type(idx)->set_rewritable(false); 1788 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1789 alias_type(idx)->set_rewritable(false); 1790 } 1791 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1792 // but the base pointer type is not distinctive enough to identify 1793 // references into JavaThread.) 1794 1795 // Check for final fields. 1796 const TypeInstPtr* tinst = flat->isa_instptr(); 1797 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1798 ciField* field; 1799 if (tinst->const_oop() != NULL && 1800 tinst->klass() == ciEnv::current()->Class_klass() && 1801 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1802 // static field 1803 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1804 field = k->get_field_by_offset(tinst->offset(), true); 1805 } else { 1806 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1807 field = k->get_field_by_offset(tinst->offset(), false); 1808 } 1809 assert(field == NULL || 1810 original_field == NULL || 1811 (field->holder() == original_field->holder() && 1812 field->offset() == original_field->offset() && 1813 field->is_static() == original_field->is_static()), "wrong field?"); 1814 // Set field() and is_rewritable() attributes. 1815 if (field != NULL) alias_type(idx)->set_field(field); 1816 } 1817 } 1818 1819 // Fill the cache for next time. 1820 ace->_adr_type = adr_type; 1821 ace->_index = idx; 1822 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1823 1824 // Might as well try to fill the cache for the flattened version, too. 1825 AliasCacheEntry* face = probe_alias_cache(flat); 1826 if (face->_adr_type == NULL) { 1827 face->_adr_type = flat; 1828 face->_index = idx; 1829 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1830 } 1831 1832 return alias_type(idx); 1833 } 1834 1835 1836 Compile::AliasType* Compile::alias_type(ciField* field) { 1837 const TypeOopPtr* t; 1838 if (field->is_static()) 1839 t = TypeInstPtr::make(field->holder()->java_mirror()); 1840 else 1841 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1842 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1843 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1844 return atp; 1845 } 1846 1847 1848 //------------------------------have_alias_type-------------------------------- 1849 bool Compile::have_alias_type(const TypePtr* adr_type) { 1850 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1851 if (ace->_adr_type == adr_type) { 1852 return true; 1853 } 1854 1855 // Handle special cases. 1856 if (adr_type == NULL) return true; 1857 if (adr_type == TypePtr::BOTTOM) return true; 1858 1859 return find_alias_type(adr_type, true, NULL) != NULL; 1860 } 1861 1862 //-----------------------------must_alias-------------------------------------- 1863 // True if all values of the given address type are in the given alias category. 1864 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1865 if (alias_idx == AliasIdxBot) return true; // the universal category 1866 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1867 if (alias_idx == AliasIdxTop) return false; // the empty category 1868 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1869 1870 // the only remaining possible overlap is identity 1871 int adr_idx = get_alias_index(adr_type); 1872 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1873 assert(adr_idx == alias_idx || 1874 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1875 && adr_type != TypeOopPtr::BOTTOM), 1876 "should not be testing for overlap with an unsafe pointer"); 1877 return adr_idx == alias_idx; 1878 } 1879 1880 //------------------------------can_alias-------------------------------------- 1881 // True if any values of the given address type are in the given alias category. 1882 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1883 if (alias_idx == AliasIdxTop) return false; // the empty category 1884 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1885 if (alias_idx == AliasIdxBot) return true; // the universal category 1886 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1887 1888 // the only remaining possible overlap is identity 1889 int adr_idx = get_alias_index(adr_type); 1890 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1891 return adr_idx == alias_idx; 1892 } 1893 1894 1895 1896 //---------------------------pop_warm_call------------------------------------- 1897 WarmCallInfo* Compile::pop_warm_call() { 1898 WarmCallInfo* wci = _warm_calls; 1899 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1900 return wci; 1901 } 1902 1903 //----------------------------Inline_Warm-------------------------------------- 1904 int Compile::Inline_Warm() { 1905 // If there is room, try to inline some more warm call sites. 1906 // %%% Do a graph index compaction pass when we think we're out of space? 1907 if (!InlineWarmCalls) return 0; 1908 1909 int calls_made_hot = 0; 1910 int room_to_grow = NodeCountInliningCutoff - unique(); 1911 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1912 int amount_grown = 0; 1913 WarmCallInfo* call; 1914 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1915 int est_size = (int)call->size(); 1916 if (est_size > (room_to_grow - amount_grown)) { 1917 // This one won't fit anyway. Get rid of it. 1918 call->make_cold(); 1919 continue; 1920 } 1921 call->make_hot(); 1922 calls_made_hot++; 1923 amount_grown += est_size; 1924 amount_to_grow -= est_size; 1925 } 1926 1927 if (calls_made_hot > 0) set_major_progress(); 1928 return calls_made_hot; 1929 } 1930 1931 1932 //----------------------------Finish_Warm-------------------------------------- 1933 void Compile::Finish_Warm() { 1934 if (!InlineWarmCalls) return; 1935 if (failing()) return; 1936 if (warm_calls() == NULL) return; 1937 1938 // Clean up loose ends, if we are out of space for inlining. 1939 WarmCallInfo* call; 1940 while ((call = pop_warm_call()) != NULL) { 1941 call->make_cold(); 1942 } 1943 } 1944 1945 //---------------------cleanup_loop_predicates----------------------- 1946 // Remove the opaque nodes that protect the predicates so that all unused 1947 // checks and uncommon_traps will be eliminated from the ideal graph 1948 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1949 if (predicate_count()==0) return; 1950 for (int i = predicate_count(); i > 0; i--) { 1951 Node * n = predicate_opaque1_node(i-1); 1952 assert(n->Opcode() == Op_Opaque1, "must be"); 1953 igvn.replace_node(n, n->in(1)); 1954 } 1955 assert(predicate_count()==0, "should be clean!"); 1956 } 1957 1958 void Compile::add_range_check_cast(Node* n) { 1959 assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1960 assert(!_range_check_casts->contains(n), "duplicate entry in range check casts"); 1961 _range_check_casts->append(n); 1962 } 1963 1964 // Remove all range check dependent CastIINodes. 1965 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) { 1966 for (int i = range_check_cast_count(); i > 0; i--) { 1967 Node* cast = range_check_cast_node(i-1); 1968 assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency"); 1969 igvn.replace_node(cast, cast->in(1)); 1970 } 1971 assert(range_check_cast_count() == 0, "should be empty"); 1972 } 1973 1974 // StringOpts and late inlining of string methods 1975 void Compile::inline_string_calls(bool parse_time) { 1976 { 1977 // remove useless nodes to make the usage analysis simpler 1978 ResourceMark rm; 1979 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1980 } 1981 1982 { 1983 ResourceMark rm; 1984 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1985 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1986 print_method(PHASE_AFTER_STRINGOPTS, 3); 1987 } 1988 1989 // now inline anything that we skipped the first time around 1990 if (!parse_time) { 1991 _late_inlines_pos = _late_inlines.length(); 1992 } 1993 1994 while (_string_late_inlines.length() > 0) { 1995 CallGenerator* cg = _string_late_inlines.pop(); 1996 cg->do_late_inline(); 1997 if (failing()) return; 1998 } 1999 _string_late_inlines.trunc_to(0); 2000 } 2001 2002 // Late inlining of boxing methods 2003 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 2004 if (_boxing_late_inlines.length() > 0) { 2005 assert(has_boxed_value(), "inconsistent"); 2006 2007 PhaseGVN* gvn = initial_gvn(); 2008 set_inlining_incrementally(true); 2009 2010 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2011 for_igvn()->clear(); 2012 gvn->replace_with(&igvn); 2013 2014 _late_inlines_pos = _late_inlines.length(); 2015 2016 while (_boxing_late_inlines.length() > 0) { 2017 CallGenerator* cg = _boxing_late_inlines.pop(); 2018 cg->do_late_inline(); 2019 if (failing()) return; 2020 } 2021 _boxing_late_inlines.trunc_to(0); 2022 2023 { 2024 ResourceMark rm; 2025 PhaseRemoveUseless pru(gvn, for_igvn()); 2026 } 2027 2028 igvn = PhaseIterGVN(gvn); 2029 igvn.optimize(); 2030 2031 set_inlining_progress(false); 2032 set_inlining_incrementally(false); 2033 } 2034 } 2035 2036 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 2037 assert(IncrementalInline, "incremental inlining should be on"); 2038 PhaseGVN* gvn = initial_gvn(); 2039 2040 set_inlining_progress(false); 2041 for_igvn()->clear(); 2042 gvn->replace_with(&igvn); 2043 2044 int i = 0; 2045 2046 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 2047 CallGenerator* cg = _late_inlines.at(i); 2048 _late_inlines_pos = i+1; 2049 cg->do_late_inline(); 2050 if (failing()) return; 2051 } 2052 int j = 0; 2053 for (; i < _late_inlines.length(); i++, j++) { 2054 _late_inlines.at_put(j, _late_inlines.at(i)); 2055 } 2056 _late_inlines.trunc_to(j); 2057 2058 { 2059 ResourceMark rm; 2060 PhaseRemoveUseless pru(gvn, for_igvn()); 2061 } 2062 2063 igvn = PhaseIterGVN(gvn); 2064 } 2065 2066 // Perform incremental inlining until bound on number of live nodes is reached 2067 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 2068 PhaseGVN* gvn = initial_gvn(); 2069 2070 set_inlining_incrementally(true); 2071 set_inlining_progress(true); 2072 uint low_live_nodes = 0; 2073 2074 while(inlining_progress() && _late_inlines.length() > 0) { 2075 2076 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2077 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 2078 // PhaseIdealLoop is expensive so we only try it once we are 2079 // out of live nodes and we only try it again if the previous 2080 // helped got the number of nodes down significantly 2081 PhaseIdealLoop ideal_loop( igvn, false, true ); 2082 if (failing()) return; 2083 low_live_nodes = live_nodes(); 2084 _major_progress = true; 2085 } 2086 2087 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 2088 break; 2089 } 2090 } 2091 2092 inline_incrementally_one(igvn); 2093 2094 if (failing()) return; 2095 2096 igvn.optimize(); 2097 2098 if (failing()) return; 2099 } 2100 2101 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2102 2103 if (_string_late_inlines.length() > 0) { 2104 assert(has_stringbuilder(), "inconsistent"); 2105 for_igvn()->clear(); 2106 initial_gvn()->replace_with(&igvn); 2107 2108 inline_string_calls(false); 2109 2110 if (failing()) return; 2111 2112 { 2113 ResourceMark rm; 2114 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2115 } 2116 2117 igvn = PhaseIterGVN(gvn); 2118 2119 igvn.optimize(); 2120 } 2121 2122 set_inlining_incrementally(false); 2123 } 2124 2125 2126 //------------------------------Optimize--------------------------------------- 2127 // Given a graph, optimize it. 2128 void Compile::Optimize() { 2129 TracePhase t1("optimizer", &_t_optimizer, true); 2130 2131 #ifndef PRODUCT 2132 if (env()->break_at_compile()) { 2133 BREAKPOINT; 2134 } 2135 2136 #endif 2137 2138 ResourceMark rm; 2139 int loop_opts_cnt; 2140 2141 NOT_PRODUCT( verify_graph_edges(); ) 2142 2143 print_method(PHASE_AFTER_PARSING); 2144 2145 { 2146 // Iterative Global Value Numbering, including ideal transforms 2147 // Initialize IterGVN with types and values from parse-time GVN 2148 PhaseIterGVN igvn(initial_gvn()); 2149 { 2150 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2151 igvn.optimize(); 2152 } 2153 2154 print_method(PHASE_ITER_GVN1, 2); 2155 2156 if (failing()) return; 2157 2158 { 2159 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2160 inline_incrementally(igvn); 2161 } 2162 2163 print_method(PHASE_INCREMENTAL_INLINE, 2); 2164 2165 if (failing()) return; 2166 2167 if (eliminate_boxing()) { 2168 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2169 // Inline valueOf() methods now. 2170 inline_boxing_calls(igvn); 2171 2172 if (AlwaysIncrementalInline) { 2173 inline_incrementally(igvn); 2174 } 2175 2176 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2177 2178 if (failing()) return; 2179 } 2180 2181 // Remove the speculative part of types and clean up the graph from 2182 // the extra CastPP nodes whose only purpose is to carry them. Do 2183 // that early so that optimizations are not disrupted by the extra 2184 // CastPP nodes. 2185 remove_speculative_types(igvn); 2186 2187 // No more new expensive nodes will be added to the list from here 2188 // so keep only the actual candidates for optimizations. 2189 cleanup_expensive_nodes(igvn); 2190 2191 if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) { 2192 NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);) 2193 initial_gvn()->replace_with(&igvn); 2194 for_igvn()->clear(); 2195 Unique_Node_List new_worklist(C->comp_arena()); 2196 { 2197 ResourceMark rm; 2198 PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist); 2199 } 2200 set_for_igvn(&new_worklist); 2201 igvn = PhaseIterGVN(initial_gvn()); 2202 igvn.optimize(); 2203 } 2204 2205 // Perform escape analysis 2206 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2207 if (has_loops()) { 2208 // Cleanup graph (remove dead nodes). 2209 TracePhase t2("idealLoop", &_t_idealLoop, true); 2210 PhaseIdealLoop ideal_loop( igvn, false, true ); 2211 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2212 if (failing()) return; 2213 } 2214 ConnectionGraph::do_analysis(this, &igvn); 2215 2216 if (failing()) return; 2217 2218 // Optimize out fields loads from scalar replaceable allocations. 2219 igvn.optimize(); 2220 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2221 2222 if (failing()) return; 2223 2224 if (congraph() != NULL && macro_count() > 0) { 2225 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2226 PhaseMacroExpand mexp(igvn); 2227 mexp.eliminate_macro_nodes(); 2228 igvn.set_delay_transform(false); 2229 2230 igvn.optimize(); 2231 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2232 2233 if (failing()) return; 2234 } 2235 } 2236 2237 // Loop transforms on the ideal graph. Range Check Elimination, 2238 // peeling, unrolling, etc. 2239 2240 // Set loop opts counter 2241 loop_opts_cnt = num_loop_opts(); 2242 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2243 { 2244 TracePhase t2("idealLoop", &_t_idealLoop, true); 2245 PhaseIdealLoop ideal_loop( igvn, true ); 2246 loop_opts_cnt--; 2247 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2248 if (failing()) return; 2249 } 2250 // Loop opts pass if partial peeling occurred in previous pass 2251 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2252 TracePhase t3("idealLoop", &_t_idealLoop, true); 2253 PhaseIdealLoop ideal_loop( igvn, false ); 2254 loop_opts_cnt--; 2255 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2256 if (failing()) return; 2257 } 2258 // Loop opts pass for loop-unrolling before CCP 2259 if(major_progress() && (loop_opts_cnt > 0)) { 2260 TracePhase t4("idealLoop", &_t_idealLoop, true); 2261 PhaseIdealLoop ideal_loop( igvn, false ); 2262 loop_opts_cnt--; 2263 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2264 } 2265 if (!failing()) { 2266 // Verify that last round of loop opts produced a valid graph 2267 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2268 PhaseIdealLoop::verify(igvn); 2269 } 2270 } 2271 if (failing()) return; 2272 2273 // Conditional Constant Propagation; 2274 PhaseCCP ccp( &igvn ); 2275 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2276 { 2277 TracePhase t2("ccp", &_t_ccp, true); 2278 ccp.do_transform(); 2279 } 2280 print_method(PHASE_CPP1, 2); 2281 2282 assert( true, "Break here to ccp.dump_old2new_map()"); 2283 2284 // Iterative Global Value Numbering, including ideal transforms 2285 { 2286 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2287 igvn = ccp; 2288 igvn.optimize(); 2289 } 2290 2291 print_method(PHASE_ITER_GVN2, 2); 2292 2293 if (failing()) return; 2294 2295 // Loop transforms on the ideal graph. Range Check Elimination, 2296 // peeling, unrolling, etc. 2297 if(loop_opts_cnt > 0) { 2298 debug_only( int cnt = 0; ); 2299 while(major_progress() && (loop_opts_cnt > 0)) { 2300 TracePhase t2("idealLoop", &_t_idealLoop, true); 2301 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2302 PhaseIdealLoop ideal_loop( igvn, true); 2303 loop_opts_cnt--; 2304 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2305 if (failing()) return; 2306 } 2307 } 2308 2309 { 2310 // Verify that all previous optimizations produced a valid graph 2311 // at least to this point, even if no loop optimizations were done. 2312 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2313 PhaseIdealLoop::verify(igvn); 2314 } 2315 2316 if (range_check_cast_count() > 0) { 2317 // No more loop optimizations. Remove all range check dependent CastIINodes. 2318 C->remove_range_check_casts(igvn); 2319 igvn.optimize(); 2320 } 2321 2322 #ifdef ASSERT 2323 if (UseShenandoahGC && ShenandoahVerifyOptoBarriers) { 2324 ShenandoahBarrierNode::verify(C->root()); 2325 } 2326 #endif 2327 2328 { 2329 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2330 PhaseMacroExpand mex(igvn); 2331 if (mex.expand_macro_nodes()) { 2332 assert(failing(), "must bail out w/ explicit message"); 2333 return; 2334 } 2335 } 2336 2337 if (UseShenandoahGC && ShenandoahWriteBarrierToIR) { 2338 if (shenandoah_barriers_count() > 0) { 2339 C->clear_major_progress(); 2340 PhaseIdealLoop ideal_loop(igvn, false, true); 2341 if (failing()) return; 2342 PhaseIdealLoop::verify(igvn); 2343 DEBUG_ONLY(ShenandoahBarrierNode::verify_raw_mem(C->root());) 2344 } 2345 } 2346 2347 } // (End scope of igvn; run destructor if necessary for asserts.) 2348 2349 dump_inlining(); 2350 // A method with only infinite loops has no edges entering loops from root 2351 { 2352 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2353 if (final_graph_reshaping()) { 2354 assert(failing(), "must bail out w/ explicit message"); 2355 return; 2356 } 2357 } 2358 2359 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2360 } 2361 2362 2363 //------------------------------Code_Gen--------------------------------------- 2364 // Given a graph, generate code for it 2365 void Compile::Code_Gen() { 2366 if (failing()) { 2367 return; 2368 } 2369 2370 // Perform instruction selection. You might think we could reclaim Matcher 2371 // memory PDQ, but actually the Matcher is used in generating spill code. 2372 // Internals of the Matcher (including some VectorSets) must remain live 2373 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2374 // set a bit in reclaimed memory. 2375 2376 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2377 // nodes. Mapping is only valid at the root of each matched subtree. 2378 NOT_PRODUCT( verify_graph_edges(); ) 2379 2380 Matcher matcher; 2381 _matcher = &matcher; 2382 { 2383 TracePhase t2("matcher", &_t_matcher, true); 2384 matcher.match(); 2385 } 2386 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2387 // nodes. Mapping is only valid at the root of each matched subtree. 2388 NOT_PRODUCT( verify_graph_edges(); ) 2389 2390 // If you have too many nodes, or if matching has failed, bail out 2391 check_node_count(0, "out of nodes matching instructions"); 2392 if (failing()) { 2393 return; 2394 } 2395 2396 // Build a proper-looking CFG 2397 PhaseCFG cfg(node_arena(), root(), matcher); 2398 _cfg = &cfg; 2399 { 2400 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2401 bool success = cfg.do_global_code_motion(); 2402 if (!success) { 2403 return; 2404 } 2405 2406 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2407 NOT_PRODUCT( verify_graph_edges(); ) 2408 debug_only( cfg.verify(); ) 2409 } 2410 2411 PhaseChaitin regalloc(unique(), cfg, matcher); 2412 _regalloc = ®alloc; 2413 { 2414 TracePhase t2("regalloc", &_t_registerAllocation, true); 2415 // Perform register allocation. After Chaitin, use-def chains are 2416 // no longer accurate (at spill code) and so must be ignored. 2417 // Node->LRG->reg mappings are still accurate. 2418 _regalloc->Register_Allocate(); 2419 2420 // Bail out if the allocator builds too many nodes 2421 if (failing()) { 2422 return; 2423 } 2424 } 2425 2426 // Prior to register allocation we kept empty basic blocks in case the 2427 // the allocator needed a place to spill. After register allocation we 2428 // are not adding any new instructions. If any basic block is empty, we 2429 // can now safely remove it. 2430 { 2431 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2432 cfg.remove_empty_blocks(); 2433 if (do_freq_based_layout()) { 2434 PhaseBlockLayout layout(cfg); 2435 } else { 2436 cfg.set_loop_alignment(); 2437 } 2438 cfg.fixup_flow(); 2439 } 2440 2441 // Apply peephole optimizations 2442 if( OptoPeephole ) { 2443 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2444 PhasePeephole peep( _regalloc, cfg); 2445 peep.do_transform(); 2446 } 2447 2448 // Do late expand if CPU requires this. 2449 if (Matcher::require_postalloc_expand) { 2450 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2451 cfg.postalloc_expand(_regalloc); 2452 } 2453 2454 // Convert Nodes to instruction bits in a buffer 2455 { 2456 // %%%% workspace merge brought two timers together for one job 2457 TracePhase t2a("output", &_t_output, true); 2458 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2459 Output(); 2460 } 2461 2462 print_method(PHASE_FINAL_CODE); 2463 2464 // He's dead, Jim. 2465 _cfg = (PhaseCFG*)0xdeadbeef; 2466 _regalloc = (PhaseChaitin*)0xdeadbeef; 2467 } 2468 2469 2470 //------------------------------dump_asm--------------------------------------- 2471 // Dump formatted assembly 2472 #ifndef PRODUCT 2473 void Compile::dump_asm(int *pcs, uint pc_limit) { 2474 bool cut_short = false; 2475 tty->print_cr("#"); 2476 tty->print("# "); _tf->dump(); tty->cr(); 2477 tty->print_cr("#"); 2478 2479 // For all blocks 2480 int pc = 0x0; // Program counter 2481 char starts_bundle = ' '; 2482 _regalloc->dump_frame(); 2483 2484 Node *n = NULL; 2485 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2486 if (VMThread::should_terminate()) { 2487 cut_short = true; 2488 break; 2489 } 2490 Block* block = _cfg->get_block(i); 2491 if (block->is_connector() && !Verbose) { 2492 continue; 2493 } 2494 n = block->head(); 2495 if (pcs && n->_idx < pc_limit) { 2496 tty->print("%3.3x ", pcs[n->_idx]); 2497 } else { 2498 tty->print(" "); 2499 } 2500 block->dump_head(_cfg); 2501 if (block->is_connector()) { 2502 tty->print_cr(" # Empty connector block"); 2503 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2504 tty->print_cr(" # Block is sole successor of call"); 2505 } 2506 2507 // For all instructions 2508 Node *delay = NULL; 2509 for (uint j = 0; j < block->number_of_nodes(); j++) { 2510 if (VMThread::should_terminate()) { 2511 cut_short = true; 2512 break; 2513 } 2514 n = block->get_node(j); 2515 if (valid_bundle_info(n)) { 2516 Bundle* bundle = node_bundling(n); 2517 if (bundle->used_in_unconditional_delay()) { 2518 delay = n; 2519 continue; 2520 } 2521 if (bundle->starts_bundle()) { 2522 starts_bundle = '+'; 2523 } 2524 } 2525 2526 if (WizardMode) { 2527 n->dump(); 2528 } 2529 2530 if( !n->is_Region() && // Dont print in the Assembly 2531 !n->is_Phi() && // a few noisely useless nodes 2532 !n->is_Proj() && 2533 !n->is_MachTemp() && 2534 !n->is_SafePointScalarObject() && 2535 !n->is_Catch() && // Would be nice to print exception table targets 2536 !n->is_MergeMem() && // Not very interesting 2537 !n->is_top() && // Debug info table constants 2538 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2539 ) { 2540 if (pcs && n->_idx < pc_limit) 2541 tty->print("%3.3x", pcs[n->_idx]); 2542 else 2543 tty->print(" "); 2544 tty->print(" %c ", starts_bundle); 2545 starts_bundle = ' '; 2546 tty->print("\t"); 2547 n->format(_regalloc, tty); 2548 tty->cr(); 2549 } 2550 2551 // If we have an instruction with a delay slot, and have seen a delay, 2552 // then back up and print it 2553 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2554 assert(delay != NULL, "no unconditional delay instruction"); 2555 if (WizardMode) delay->dump(); 2556 2557 if (node_bundling(delay)->starts_bundle()) 2558 starts_bundle = '+'; 2559 if (pcs && n->_idx < pc_limit) 2560 tty->print("%3.3x", pcs[n->_idx]); 2561 else 2562 tty->print(" "); 2563 tty->print(" %c ", starts_bundle); 2564 starts_bundle = ' '; 2565 tty->print("\t"); 2566 delay->format(_regalloc, tty); 2567 tty->cr(); 2568 delay = NULL; 2569 } 2570 2571 // Dump the exception table as well 2572 if( n->is_Catch() && (Verbose || WizardMode) ) { 2573 // Print the exception table for this offset 2574 _handler_table.print_subtable_for(pc); 2575 } 2576 } 2577 2578 if (pcs && n->_idx < pc_limit) 2579 tty->print_cr("%3.3x", pcs[n->_idx]); 2580 else 2581 tty->cr(); 2582 2583 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2584 2585 } // End of per-block dump 2586 tty->cr(); 2587 2588 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2589 } 2590 #endif 2591 2592 //------------------------------Final_Reshape_Counts--------------------------- 2593 // This class defines counters to help identify when a method 2594 // may/must be executed using hardware with only 24-bit precision. 2595 struct Final_Reshape_Counts : public StackObj { 2596 int _call_count; // count non-inlined 'common' calls 2597 int _float_count; // count float ops requiring 24-bit precision 2598 int _double_count; // count double ops requiring more precision 2599 int _java_call_count; // count non-inlined 'java' calls 2600 int _inner_loop_count; // count loops which need alignment 2601 VectorSet _visited; // Visitation flags 2602 Node_List _tests; // Set of IfNodes & PCTableNodes 2603 2604 Final_Reshape_Counts() : 2605 _call_count(0), _float_count(0), _double_count(0), 2606 _java_call_count(0), _inner_loop_count(0), 2607 _visited( Thread::current()->resource_area() ) { } 2608 2609 void inc_call_count () { _call_count ++; } 2610 void inc_float_count () { _float_count ++; } 2611 void inc_double_count() { _double_count++; } 2612 void inc_java_call_count() { _java_call_count++; } 2613 void inc_inner_loop_count() { _inner_loop_count++; } 2614 2615 int get_call_count () const { return _call_count ; } 2616 int get_float_count () const { return _float_count ; } 2617 int get_double_count() const { return _double_count; } 2618 int get_java_call_count() const { return _java_call_count; } 2619 int get_inner_loop_count() const { return _inner_loop_count; } 2620 }; 2621 2622 #ifdef ASSERT 2623 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2624 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2625 // Make sure the offset goes inside the instance layout. 2626 return k->contains_field_offset(tp->offset()); 2627 // Note that OffsetBot and OffsetTop are very negative. 2628 } 2629 #endif 2630 2631 // Eliminate trivially redundant StoreCMs and accumulate their 2632 // precedence edges. 2633 void Compile::eliminate_redundant_card_marks(Node* n) { 2634 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2635 if (n->in(MemNode::Address)->outcnt() > 1) { 2636 // There are multiple users of the same address so it might be 2637 // possible to eliminate some of the StoreCMs 2638 Node* mem = n->in(MemNode::Memory); 2639 Node* adr = n->in(MemNode::Address); 2640 Node* val = n->in(MemNode::ValueIn); 2641 Node* prev = n; 2642 bool done = false; 2643 // Walk the chain of StoreCMs eliminating ones that match. As 2644 // long as it's a chain of single users then the optimization is 2645 // safe. Eliminating partially redundant StoreCMs would require 2646 // cloning copies down the other paths. 2647 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2648 if (adr == mem->in(MemNode::Address) && 2649 val == mem->in(MemNode::ValueIn)) { 2650 // redundant StoreCM 2651 if (mem->req() > MemNode::OopStore) { 2652 // Hasn't been processed by this code yet. 2653 n->add_prec(mem->in(MemNode::OopStore)); 2654 } else { 2655 // Already converted to precedence edge 2656 for (uint i = mem->req(); i < mem->len(); i++) { 2657 // Accumulate any precedence edges 2658 if (mem->in(i) != NULL) { 2659 n->add_prec(mem->in(i)); 2660 } 2661 } 2662 // Everything above this point has been processed. 2663 done = true; 2664 } 2665 // Eliminate the previous StoreCM 2666 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2667 assert(mem->outcnt() == 0, "should be dead"); 2668 mem->disconnect_inputs(NULL, this); 2669 } else { 2670 prev = mem; 2671 } 2672 mem = prev->in(MemNode::Memory); 2673 } 2674 } 2675 } 2676 2677 //------------------------------final_graph_reshaping_impl---------------------- 2678 // Implement items 1-5 from final_graph_reshaping below. 2679 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2680 2681 if ( n->outcnt() == 0 ) return; // dead node 2682 uint nop = n->Opcode(); 2683 2684 // Check for 2-input instruction with "last use" on right input. 2685 // Swap to left input. Implements item (2). 2686 if( n->req() == 3 && // two-input instruction 2687 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2688 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2689 n->in(2)->outcnt() == 1 &&// right use IS a last use 2690 !n->in(2)->is_Con() ) { // right use is not a constant 2691 // Check for commutative opcode 2692 switch( nop ) { 2693 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2694 case Op_MaxI: case Op_MinI: 2695 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2696 case Op_AndL: case Op_XorL: case Op_OrL: 2697 case Op_AndI: case Op_XorI: case Op_OrI: { 2698 // Move "last use" input to left by swapping inputs 2699 n->swap_edges(1, 2); 2700 break; 2701 } 2702 default: 2703 break; 2704 } 2705 } 2706 2707 #ifdef ASSERT 2708 if( n->is_Mem() ) { 2709 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2710 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2711 // oop will be recorded in oop map if load crosses safepoint 2712 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2713 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2714 "raw memory operations should have control edge"); 2715 } 2716 #endif 2717 // Count FPU ops and common calls, implements item (3) 2718 switch( nop ) { 2719 // Count all float operations that may use FPU 2720 case Op_AddF: 2721 case Op_SubF: 2722 case Op_MulF: 2723 case Op_DivF: 2724 case Op_NegF: 2725 case Op_ModF: 2726 case Op_ConvI2F: 2727 case Op_ConF: 2728 case Op_CmpF: 2729 case Op_CmpF3: 2730 // case Op_ConvL2F: // longs are split into 32-bit halves 2731 frc.inc_float_count(); 2732 break; 2733 2734 case Op_ConvF2D: 2735 case Op_ConvD2F: 2736 frc.inc_float_count(); 2737 frc.inc_double_count(); 2738 break; 2739 2740 // Count all double operations that may use FPU 2741 case Op_AddD: 2742 case Op_SubD: 2743 case Op_MulD: 2744 case Op_DivD: 2745 case Op_NegD: 2746 case Op_ModD: 2747 case Op_ConvI2D: 2748 case Op_ConvD2I: 2749 // case Op_ConvL2D: // handled by leaf call 2750 // case Op_ConvD2L: // handled by leaf call 2751 case Op_ConD: 2752 case Op_CmpD: 2753 case Op_CmpD3: 2754 frc.inc_double_count(); 2755 break; 2756 case Op_Opaque1: // Remove Opaque Nodes before matching 2757 case Op_Opaque2: // Remove Opaque Nodes before matching 2758 case Op_Opaque3: 2759 n->subsume_by(n->in(1), this); 2760 break; 2761 case Op_CallStaticJava: 2762 case Op_CallJava: 2763 case Op_CallDynamicJava: 2764 frc.inc_java_call_count(); // Count java call site; 2765 case Op_CallRuntime: 2766 case Op_CallLeaf: 2767 case Op_CallLeafNoFP: { 2768 assert( n->is_Call(), "" ); 2769 CallNode *call = n->as_Call(); 2770 if (UseShenandoahGC && call->is_g1_wb_pre_call()) { 2771 uint cnt = OptoRuntime::g1_wb_pre_Type()->domain()->cnt(); 2772 if (call->req() > cnt) { 2773 assert(call->req() == cnt+1, "only one extra input"); 2774 Node* addp = call->in(cnt); 2775 assert(!CallLeafNode::has_only_g1_wb_pre_uses(addp), "useless address computation?"); 2776 call->del_req(cnt); 2777 } 2778 } 2779 // Count call sites where the FP mode bit would have to be flipped. 2780 // Do not count uncommon runtime calls: 2781 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2782 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2783 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2784 frc.inc_call_count(); // Count the call site 2785 } else { // See if uncommon argument is shared 2786 Node *n = call->in(TypeFunc::Parms); 2787 int nop = n->Opcode(); 2788 // Clone shared simple arguments to uncommon calls, item (1). 2789 if( n->outcnt() > 1 && 2790 !n->is_Proj() && 2791 nop != Op_CreateEx && 2792 nop != Op_CheckCastPP && 2793 nop != Op_DecodeN && 2794 nop != Op_DecodeNKlass && 2795 !n->is_Mem() ) { 2796 Node *x = n->clone(); 2797 call->set_req( TypeFunc::Parms, x ); 2798 } 2799 } 2800 break; 2801 } 2802 2803 case Op_StoreD: 2804 case Op_LoadD: 2805 case Op_LoadD_unaligned: 2806 frc.inc_double_count(); 2807 goto handle_mem; 2808 case Op_StoreF: 2809 case Op_LoadF: 2810 frc.inc_float_count(); 2811 goto handle_mem; 2812 2813 case Op_StoreCM: 2814 { 2815 // Convert OopStore dependence into precedence edge 2816 Node* prec = n->in(MemNode::OopStore); 2817 n->del_req(MemNode::OopStore); 2818 n->add_prec(prec); 2819 eliminate_redundant_card_marks(n); 2820 } 2821 2822 // fall through 2823 2824 case Op_StoreB: 2825 case Op_StoreC: 2826 case Op_StorePConditional: 2827 case Op_StoreI: 2828 case Op_StoreL: 2829 case Op_StoreIConditional: 2830 case Op_StoreLConditional: 2831 case Op_CompareAndSwapI: 2832 case Op_CompareAndSwapL: 2833 case Op_CompareAndSwapP: 2834 case Op_CompareAndSwapN: 2835 case Op_GetAndAddI: 2836 case Op_GetAndAddL: 2837 case Op_GetAndSetI: 2838 case Op_GetAndSetL: 2839 case Op_GetAndSetP: 2840 case Op_GetAndSetN: 2841 case Op_StoreP: 2842 case Op_StoreN: 2843 case Op_StoreNKlass: 2844 case Op_LoadB: 2845 case Op_LoadUB: 2846 case Op_LoadUS: 2847 case Op_LoadI: 2848 case Op_LoadKlass: 2849 case Op_LoadNKlass: 2850 case Op_LoadL: 2851 case Op_LoadL_unaligned: 2852 case Op_LoadPLocked: 2853 case Op_LoadP: 2854 case Op_LoadN: 2855 case Op_LoadRange: 2856 case Op_LoadS: { 2857 handle_mem: 2858 #ifdef ASSERT 2859 if( VerifyOptoOopOffsets ) { 2860 assert( n->is_Mem(), "" ); 2861 MemNode *mem = (MemNode*)n; 2862 // Check to see if address types have grounded out somehow. 2863 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2864 assert( !tp || oop_offset_is_sane(tp), "" ); 2865 } 2866 #endif 2867 break; 2868 } 2869 2870 case Op_AddP: { // Assert sane base pointers 2871 Node *addp = n->in(AddPNode::Address); 2872 assert( !addp->is_AddP() || 2873 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2874 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2875 "Base pointers must match" ); 2876 #ifdef _LP64 2877 if ((UseCompressedOops || UseCompressedClassPointers) && 2878 addp->Opcode() == Op_ConP && 2879 addp == n->in(AddPNode::Base) && 2880 n->in(AddPNode::Offset)->is_Con()) { 2881 // Use addressing with narrow klass to load with offset on x86. 2882 // On sparc loading 32-bits constant and decoding it have less 2883 // instructions (4) then load 64-bits constant (7). 2884 // Do this transformation here since IGVN will convert ConN back to ConP. 2885 const Type* t = addp->bottom_type(); 2886 if (t->isa_oopptr() || t->isa_klassptr()) { 2887 Node* nn = NULL; 2888 2889 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2890 2891 // Look for existing ConN node of the same exact type. 2892 Node* r = root(); 2893 uint cnt = r->outcnt(); 2894 for (uint i = 0; i < cnt; i++) { 2895 Node* m = r->raw_out(i); 2896 if (m!= NULL && m->Opcode() == op && 2897 m->bottom_type()->make_ptr() == t) { 2898 nn = m; 2899 break; 2900 } 2901 } 2902 if (nn != NULL) { 2903 // Decode a narrow oop to match address 2904 // [R12 + narrow_oop_reg<<3 + offset] 2905 if (t->isa_oopptr()) { 2906 nn = new (this) DecodeNNode(nn, t); 2907 } else { 2908 nn = new (this) DecodeNKlassNode(nn, t); 2909 } 2910 n->set_req(AddPNode::Base, nn); 2911 n->set_req(AddPNode::Address, nn); 2912 if (addp->outcnt() == 0) { 2913 addp->disconnect_inputs(NULL, this); 2914 } 2915 } 2916 } 2917 } 2918 #endif 2919 break; 2920 } 2921 2922 case Op_CastPP: { 2923 // Remove CastPP nodes to gain more freedom during scheduling but 2924 // keep the dependency they encode as control or precedence edges 2925 // (if control is set already) on memory operations. Some CastPP 2926 // nodes don't have a control (don't carry a dependency): skip 2927 // those. 2928 if (n->in(0) != NULL) { 2929 ResourceMark rm; 2930 Unique_Node_List wq; 2931 wq.push(n); 2932 for (uint next = 0; next < wq.size(); ++next) { 2933 Node *m = wq.at(next); 2934 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) { 2935 Node* use = m->fast_out(i); 2936 if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) { 2937 use->ensure_control_or_add_prec(n->in(0)); 2938 } else if (use->in(0) == NULL) { 2939 switch(use->Opcode()) { 2940 case Op_AddP: 2941 case Op_DecodeN: 2942 case Op_DecodeNKlass: 2943 case Op_CheckCastPP: 2944 case Op_CastPP: 2945 wq.push(use); 2946 break; 2947 } 2948 } 2949 } 2950 } 2951 } 2952 const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false); 2953 if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2954 Node* in1 = n->in(1); 2955 const Type* t = n->bottom_type(); 2956 Node* new_in1 = in1->clone(); 2957 new_in1->as_DecodeN()->set_type(t); 2958 2959 if (!Matcher::narrow_oop_use_complex_address()) { 2960 // 2961 // x86, ARM and friends can handle 2 adds in addressing mode 2962 // and Matcher can fold a DecodeN node into address by using 2963 // a narrow oop directly and do implicit NULL check in address: 2964 // 2965 // [R12 + narrow_oop_reg<<3 + offset] 2966 // NullCheck narrow_oop_reg 2967 // 2968 // On other platforms (Sparc) we have to keep new DecodeN node and 2969 // use it to do implicit NULL check in address: 2970 // 2971 // decode_not_null narrow_oop_reg, base_reg 2972 // [base_reg + offset] 2973 // NullCheck base_reg 2974 // 2975 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2976 // to keep the information to which NULL check the new DecodeN node 2977 // corresponds to use it as value in implicit_null_check(). 2978 // 2979 new_in1->set_req(0, n->in(0)); 2980 } 2981 2982 n->subsume_by(new_in1, this); 2983 if (in1->outcnt() == 0) { 2984 in1->disconnect_inputs(NULL, this); 2985 } 2986 } else { 2987 n->subsume_by(n->in(1), this); 2988 if (n->outcnt() == 0) { 2989 n->disconnect_inputs(NULL, this); 2990 } 2991 } 2992 break; 2993 } 2994 #ifdef _LP64 2995 case Op_CmpP: 2996 // Do this transformation here to preserve CmpPNode::sub() and 2997 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2998 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2999 Node* in1 = n->in(1); 3000 Node* in2 = n->in(2); 3001 if (!in1->is_DecodeNarrowPtr()) { 3002 in2 = in1; 3003 in1 = n->in(2); 3004 } 3005 assert(in1->is_DecodeNarrowPtr(), "sanity"); 3006 3007 Node* new_in2 = NULL; 3008 if (in2->is_DecodeNarrowPtr()) { 3009 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 3010 new_in2 = in2->in(1); 3011 } else if (in2->Opcode() == Op_ConP) { 3012 const Type* t = in2->bottom_type(); 3013 if (t == TypePtr::NULL_PTR) { 3014 assert(in1->is_DecodeN(), "compare klass to null?"); 3015 // Don't convert CmpP null check into CmpN if compressed 3016 // oops implicit null check is not generated. 3017 // This will allow to generate normal oop implicit null check. 3018 if (Matcher::gen_narrow_oop_implicit_null_checks()) 3019 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 3020 // 3021 // This transformation together with CastPP transformation above 3022 // will generated code for implicit NULL checks for compressed oops. 3023 // 3024 // The original code after Optimize() 3025 // 3026 // LoadN memory, narrow_oop_reg 3027 // decode narrow_oop_reg, base_reg 3028 // CmpP base_reg, NULL 3029 // CastPP base_reg // NotNull 3030 // Load [base_reg + offset], val_reg 3031 // 3032 // after these transformations will be 3033 // 3034 // LoadN memory, narrow_oop_reg 3035 // CmpN narrow_oop_reg, NULL 3036 // decode_not_null narrow_oop_reg, base_reg 3037 // Load [base_reg + offset], val_reg 3038 // 3039 // and the uncommon path (== NULL) will use narrow_oop_reg directly 3040 // since narrow oops can be used in debug info now (see the code in 3041 // final_graph_reshaping_walk()). 3042 // 3043 // At the end the code will be matched to 3044 // on x86: 3045 // 3046 // Load_narrow_oop memory, narrow_oop_reg 3047 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 3048 // NullCheck narrow_oop_reg 3049 // 3050 // and on sparc: 3051 // 3052 // Load_narrow_oop memory, narrow_oop_reg 3053 // decode_not_null narrow_oop_reg, base_reg 3054 // Load [base_reg + offset], val_reg 3055 // NullCheck base_reg 3056 // 3057 } else if (t->isa_oopptr()) { 3058 new_in2 = ConNode::make(this, t->make_narrowoop()); 3059 } else if (t->isa_klassptr()) { 3060 new_in2 = ConNode::make(this, t->make_narrowklass()); 3061 } 3062 } 3063 if (new_in2 != NULL) { 3064 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); 3065 n->subsume_by(cmpN, this); 3066 if (in1->outcnt() == 0) { 3067 in1->disconnect_inputs(NULL, this); 3068 } 3069 if (in2->outcnt() == 0) { 3070 in2->disconnect_inputs(NULL, this); 3071 } 3072 } 3073 } 3074 break; 3075 3076 case Op_DecodeN: 3077 case Op_DecodeNKlass: 3078 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 3079 // DecodeN could be pinned when it can't be fold into 3080 // an address expression, see the code for Op_CastPP above. 3081 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 3082 break; 3083 3084 case Op_EncodeP: 3085 case Op_EncodePKlass: { 3086 Node* in1 = n->in(1); 3087 if (in1->is_DecodeNarrowPtr()) { 3088 n->subsume_by(in1->in(1), this); 3089 } else if (in1->Opcode() == Op_ConP) { 3090 const Type* t = in1->bottom_type(); 3091 if (t == TypePtr::NULL_PTR) { 3092 assert(t->isa_oopptr(), "null klass?"); 3093 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 3094 } else if (t->isa_oopptr()) { 3095 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 3096 } else if (t->isa_klassptr()) { 3097 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 3098 } 3099 } 3100 if (in1->outcnt() == 0) { 3101 in1->disconnect_inputs(NULL, this); 3102 } 3103 break; 3104 } 3105 3106 case Op_Proj: { 3107 if (OptimizeStringConcat) { 3108 ProjNode* p = n->as_Proj(); 3109 if (p->_is_io_use) { 3110 // Separate projections were used for the exception path which 3111 // are normally removed by a late inline. If it wasn't inlined 3112 // then they will hang around and should just be replaced with 3113 // the original one. 3114 Node* proj = NULL; 3115 // Replace with just one 3116 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 3117 Node *use = i.get(); 3118 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 3119 proj = use; 3120 break; 3121 } 3122 } 3123 assert(proj != NULL, "must be found"); 3124 p->subsume_by(proj, this); 3125 } 3126 } 3127 break; 3128 } 3129 3130 case Op_Phi: 3131 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 3132 // The EncodeP optimization may create Phi with the same edges 3133 // for all paths. It is not handled well by Register Allocator. 3134 Node* unique_in = n->in(1); 3135 assert(unique_in != NULL, ""); 3136 uint cnt = n->req(); 3137 for (uint i = 2; i < cnt; i++) { 3138 Node* m = n->in(i); 3139 assert(m != NULL, ""); 3140 if (unique_in != m) 3141 unique_in = NULL; 3142 } 3143 if (unique_in != NULL) { 3144 n->subsume_by(unique_in, this); 3145 } 3146 } 3147 break; 3148 3149 #endif 3150 3151 #ifdef ASSERT 3152 case Op_CastII: 3153 // Verify that all range check dependent CastII nodes were removed. 3154 if (n->isa_CastII()->has_range_check()) { 3155 n->dump(3); 3156 assert(false, "Range check dependent CastII node was not removed"); 3157 } 3158 break; 3159 #endif 3160 3161 case Op_ModI: 3162 if (UseDivMod) { 3163 // Check if a%b and a/b both exist 3164 Node* d = n->find_similar(Op_DivI); 3165 if (d) { 3166 // Replace them with a fused divmod if supported 3167 if (Matcher::has_match_rule(Op_DivModI)) { 3168 DivModINode* divmod = DivModINode::make(this, n); 3169 d->subsume_by(divmod->div_proj(), this); 3170 n->subsume_by(divmod->mod_proj(), this); 3171 } else { 3172 // replace a%b with a-((a/b)*b) 3173 Node* mult = new (this) MulINode(d, d->in(2)); 3174 Node* sub = new (this) SubINode(d->in(1), mult); 3175 n->subsume_by(sub, this); 3176 } 3177 } 3178 } 3179 break; 3180 3181 case Op_ModL: 3182 if (UseDivMod) { 3183 // Check if a%b and a/b both exist 3184 Node* d = n->find_similar(Op_DivL); 3185 if (d) { 3186 // Replace them with a fused divmod if supported 3187 if (Matcher::has_match_rule(Op_DivModL)) { 3188 DivModLNode* divmod = DivModLNode::make(this, n); 3189 d->subsume_by(divmod->div_proj(), this); 3190 n->subsume_by(divmod->mod_proj(), this); 3191 } else { 3192 // replace a%b with a-((a/b)*b) 3193 Node* mult = new (this) MulLNode(d, d->in(2)); 3194 Node* sub = new (this) SubLNode(d->in(1), mult); 3195 n->subsume_by(sub, this); 3196 } 3197 } 3198 } 3199 break; 3200 3201 case Op_LoadVector: 3202 case Op_StoreVector: 3203 break; 3204 3205 case Op_PackB: 3206 case Op_PackS: 3207 case Op_PackI: 3208 case Op_PackF: 3209 case Op_PackL: 3210 case Op_PackD: 3211 if (n->req()-1 > 2) { 3212 // Replace many operand PackNodes with a binary tree for matching 3213 PackNode* p = (PackNode*) n; 3214 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3215 n->subsume_by(btp, this); 3216 } 3217 break; 3218 case Op_Loop: 3219 case Op_CountedLoop: 3220 if (n->as_Loop()->is_inner_loop()) { 3221 frc.inc_inner_loop_count(); 3222 } 3223 break; 3224 case Op_LShiftI: 3225 case Op_RShiftI: 3226 case Op_URShiftI: 3227 case Op_LShiftL: 3228 case Op_RShiftL: 3229 case Op_URShiftL: 3230 if (Matcher::need_masked_shift_count) { 3231 // The cpu's shift instructions don't restrict the count to the 3232 // lower 5/6 bits. We need to do the masking ourselves. 3233 Node* in2 = n->in(2); 3234 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3235 const TypeInt* t = in2->find_int_type(); 3236 if (t != NULL && t->is_con()) { 3237 juint shift = t->get_con(); 3238 if (shift > mask) { // Unsigned cmp 3239 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3240 } 3241 } else { 3242 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3243 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3244 n->set_req(2, shift); 3245 } 3246 } 3247 if (in2->outcnt() == 0) { // Remove dead node 3248 in2->disconnect_inputs(NULL, this); 3249 } 3250 } 3251 break; 3252 case Op_MemBarStoreStore: 3253 case Op_MemBarRelease: 3254 // Break the link with AllocateNode: it is no longer useful and 3255 // confuses register allocation. 3256 if (n->req() > MemBarNode::Precedent) { 3257 n->set_req(MemBarNode::Precedent, top()); 3258 } 3259 break; 3260 case Op_ShenandoahReadBarrier: 3261 break; 3262 case Op_ShenandoahWriteBarrier: 3263 assert(!ShenandoahWriteBarrierToIR, "should have been expanded already"); 3264 break; 3265 default: 3266 assert( !n->is_Call(), "" ); 3267 assert( !n->is_Mem(), "" ); 3268 assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN"); 3269 break; 3270 } 3271 3272 // Collect CFG split points 3273 if (n->is_MultiBranch()) 3274 frc._tests.push(n); 3275 } 3276 3277 //------------------------------final_graph_reshaping_walk--------------------- 3278 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3279 // requires that the walk visits a node's inputs before visiting the node. 3280 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3281 ResourceArea *area = Thread::current()->resource_area(); 3282 Unique_Node_List sfpt(area); 3283 3284 frc._visited.set(root->_idx); // first, mark node as visited 3285 uint cnt = root->req(); 3286 Node *n = root; 3287 uint i = 0; 3288 while (true) { 3289 if (i < cnt) { 3290 // Place all non-visited non-null inputs onto stack 3291 Node* m = n->in(i); 3292 ++i; 3293 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3294 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3295 // compute worst case interpreter size in case of a deoptimization 3296 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3297 3298 sfpt.push(m); 3299 } 3300 cnt = m->req(); 3301 nstack.push(n, i); // put on stack parent and next input's index 3302 n = m; 3303 i = 0; 3304 } 3305 } else { 3306 // Now do post-visit work 3307 final_graph_reshaping_impl( n, frc ); 3308 if (nstack.is_empty()) 3309 break; // finished 3310 n = nstack.node(); // Get node from stack 3311 cnt = n->req(); 3312 i = nstack.index(); 3313 nstack.pop(); // Shift to the next node on stack 3314 } 3315 } 3316 3317 // Skip next transformation if compressed oops are not used. 3318 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3319 (!UseCompressedOops && !UseCompressedClassPointers)) 3320 return; 3321 3322 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3323 // It could be done for an uncommon traps or any safepoints/calls 3324 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3325 while (sfpt.size() > 0) { 3326 n = sfpt.pop(); 3327 JVMState *jvms = n->as_SafePoint()->jvms(); 3328 assert(jvms != NULL, "sanity"); 3329 int start = jvms->debug_start(); 3330 int end = n->req(); 3331 bool is_uncommon = (n->is_CallStaticJava() && 3332 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3333 for (int j = start; j < end; j++) { 3334 Node* in = n->in(j); 3335 if (in->is_DecodeNarrowPtr()) { 3336 bool safe_to_skip = true; 3337 if (!is_uncommon ) { 3338 // Is it safe to skip? 3339 for (uint i = 0; i < in->outcnt(); i++) { 3340 Node* u = in->raw_out(i); 3341 if (!u->is_SafePoint() || 3342 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3343 safe_to_skip = false; 3344 } 3345 } 3346 } 3347 if (safe_to_skip) { 3348 n->set_req(j, in->in(1)); 3349 } 3350 if (in->outcnt() == 0) { 3351 in->disconnect_inputs(NULL, this); 3352 } 3353 } 3354 } 3355 } 3356 } 3357 3358 //------------------------------final_graph_reshaping-------------------------- 3359 // Final Graph Reshaping. 3360 // 3361 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3362 // and not commoned up and forced early. Must come after regular 3363 // optimizations to avoid GVN undoing the cloning. Clone constant 3364 // inputs to Loop Phis; these will be split by the allocator anyways. 3365 // Remove Opaque nodes. 3366 // (2) Move last-uses by commutative operations to the left input to encourage 3367 // Intel update-in-place two-address operations and better register usage 3368 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3369 // calls canonicalizing them back. 3370 // (3) Count the number of double-precision FP ops, single-precision FP ops 3371 // and call sites. On Intel, we can get correct rounding either by 3372 // forcing singles to memory (requires extra stores and loads after each 3373 // FP bytecode) or we can set a rounding mode bit (requires setting and 3374 // clearing the mode bit around call sites). The mode bit is only used 3375 // if the relative frequency of single FP ops to calls is low enough. 3376 // This is a key transform for SPEC mpeg_audio. 3377 // (4) Detect infinite loops; blobs of code reachable from above but not 3378 // below. Several of the Code_Gen algorithms fail on such code shapes, 3379 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3380 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3381 // Detection is by looking for IfNodes where only 1 projection is 3382 // reachable from below or CatchNodes missing some targets. 3383 // (5) Assert for insane oop offsets in debug mode. 3384 3385 bool Compile::final_graph_reshaping() { 3386 // an infinite loop may have been eliminated by the optimizer, 3387 // in which case the graph will be empty. 3388 if (root()->req() == 1) { 3389 record_method_not_compilable("trivial infinite loop"); 3390 return true; 3391 } 3392 3393 // Expensive nodes have their control input set to prevent the GVN 3394 // from freely commoning them. There's no GVN beyond this point so 3395 // no need to keep the control input. We want the expensive nodes to 3396 // be freely moved to the least frequent code path by gcm. 3397 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3398 for (int i = 0; i < expensive_count(); i++) { 3399 _expensive_nodes->at(i)->set_req(0, NULL); 3400 } 3401 3402 Final_Reshape_Counts frc; 3403 3404 // Visit everybody reachable! 3405 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3406 Node_Stack nstack(live_nodes() >> 1); 3407 final_graph_reshaping_walk(nstack, root(), frc); 3408 3409 // Check for unreachable (from below) code (i.e., infinite loops). 3410 for( uint i = 0; i < frc._tests.size(); i++ ) { 3411 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3412 // Get number of CFG targets. 3413 // Note that PCTables include exception targets after calls. 3414 uint required_outcnt = n->required_outcnt(); 3415 if (n->outcnt() != required_outcnt) { 3416 // Check for a few special cases. Rethrow Nodes never take the 3417 // 'fall-thru' path, so expected kids is 1 less. 3418 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3419 if (n->in(0)->in(0)->is_Call()) { 3420 CallNode *call = n->in(0)->in(0)->as_Call(); 3421 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3422 required_outcnt--; // Rethrow always has 1 less kid 3423 } else if (call->req() > TypeFunc::Parms && 3424 call->is_CallDynamicJava()) { 3425 // Check for null receiver. In such case, the optimizer has 3426 // detected that the virtual call will always result in a null 3427 // pointer exception. The fall-through projection of this CatchNode 3428 // will not be populated. 3429 Node *arg0 = call->in(TypeFunc::Parms); 3430 if (arg0->is_Type() && 3431 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3432 required_outcnt--; 3433 } 3434 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3435 call->req() > TypeFunc::Parms+1 && 3436 call->is_CallStaticJava()) { 3437 // Check for negative array length. In such case, the optimizer has 3438 // detected that the allocation attempt will always result in an 3439 // exception. There is no fall-through projection of this CatchNode . 3440 Node *arg1 = call->in(TypeFunc::Parms+1); 3441 if (arg1->is_Type() && 3442 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3443 required_outcnt--; 3444 } 3445 } 3446 } 3447 } 3448 // Recheck with a better notion of 'required_outcnt' 3449 if (n->outcnt() != required_outcnt) { 3450 record_method_not_compilable("malformed control flow"); 3451 return true; // Not all targets reachable! 3452 } 3453 } 3454 // Check that I actually visited all kids. Unreached kids 3455 // must be infinite loops. 3456 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3457 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3458 record_method_not_compilable("infinite loop"); 3459 return true; // Found unvisited kid; must be unreach 3460 } 3461 } 3462 3463 // If original bytecodes contained a mixture of floats and doubles 3464 // check if the optimizer has made it homogenous, item (3). 3465 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3466 frc.get_float_count() > 32 && 3467 frc.get_double_count() == 0 && 3468 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3469 set_24_bit_selection_and_mode( false, true ); 3470 } 3471 3472 set_java_calls(frc.get_java_call_count()); 3473 set_inner_loops(frc.get_inner_loop_count()); 3474 3475 // No infinite loops, no reason to bail out. 3476 return false; 3477 } 3478 3479 //-----------------------------too_many_traps---------------------------------- 3480 // Report if there are too many traps at the current method and bci. 3481 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3482 bool Compile::too_many_traps(ciMethod* method, 3483 int bci, 3484 Deoptimization::DeoptReason reason) { 3485 ciMethodData* md = method->method_data(); 3486 if (md->is_empty()) { 3487 // Assume the trap has not occurred, or that it occurred only 3488 // because of a transient condition during start-up in the interpreter. 3489 return false; 3490 } 3491 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3492 if (md->has_trap_at(bci, m, reason) != 0) { 3493 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3494 // Also, if there are multiple reasons, or if there is no per-BCI record, 3495 // assume the worst. 3496 if (log()) 3497 log()->elem("observe trap='%s' count='%d'", 3498 Deoptimization::trap_reason_name(reason), 3499 md->trap_count(reason)); 3500 return true; 3501 } else { 3502 // Ignore method/bci and see if there have been too many globally. 3503 return too_many_traps(reason, md); 3504 } 3505 } 3506 3507 // Less-accurate variant which does not require a method and bci. 3508 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3509 ciMethodData* logmd) { 3510 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3511 // Too many traps globally. 3512 // Note that we use cumulative trap_count, not just md->trap_count. 3513 if (log()) { 3514 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3515 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3516 Deoptimization::trap_reason_name(reason), 3517 mcount, trap_count(reason)); 3518 } 3519 return true; 3520 } else { 3521 // The coast is clear. 3522 return false; 3523 } 3524 } 3525 3526 //--------------------------too_many_recompiles-------------------------------- 3527 // Report if there are too many recompiles at the current method and bci. 3528 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3529 // Is not eager to return true, since this will cause the compiler to use 3530 // Action_none for a trap point, to avoid too many recompilations. 3531 bool Compile::too_many_recompiles(ciMethod* method, 3532 int bci, 3533 Deoptimization::DeoptReason reason) { 3534 ciMethodData* md = method->method_data(); 3535 if (md->is_empty()) { 3536 // Assume the trap has not occurred, or that it occurred only 3537 // because of a transient condition during start-up in the interpreter. 3538 return false; 3539 } 3540 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3541 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3542 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3543 Deoptimization::DeoptReason per_bc_reason 3544 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3545 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3546 if ((per_bc_reason == Deoptimization::Reason_none 3547 || md->has_trap_at(bci, m, reason) != 0) 3548 // The trap frequency measure we care about is the recompile count: 3549 && md->trap_recompiled_at(bci, m) 3550 && md->overflow_recompile_count() >= bc_cutoff) { 3551 // Do not emit a trap here if it has already caused recompilations. 3552 // Also, if there are multiple reasons, or if there is no per-BCI record, 3553 // assume the worst. 3554 if (log()) 3555 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3556 Deoptimization::trap_reason_name(reason), 3557 md->trap_count(reason), 3558 md->overflow_recompile_count()); 3559 return true; 3560 } else if (trap_count(reason) != 0 3561 && decompile_count() >= m_cutoff) { 3562 // Too many recompiles globally, and we have seen this sort of trap. 3563 // Use cumulative decompile_count, not just md->decompile_count. 3564 if (log()) 3565 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3566 Deoptimization::trap_reason_name(reason), 3567 md->trap_count(reason), trap_count(reason), 3568 md->decompile_count(), decompile_count()); 3569 return true; 3570 } else { 3571 // The coast is clear. 3572 return false; 3573 } 3574 } 3575 3576 // Compute when not to trap. Used by matching trap based nodes and 3577 // NullCheck optimization. 3578 void Compile::set_allowed_deopt_reasons() { 3579 _allowed_reasons = 0; 3580 if (is_method_compilation()) { 3581 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3582 assert(rs < BitsPerInt, "recode bit map"); 3583 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3584 _allowed_reasons |= nth_bit(rs); 3585 } 3586 } 3587 } 3588 } 3589 3590 #ifndef PRODUCT 3591 //------------------------------verify_graph_edges--------------------------- 3592 // Walk the Graph and verify that there is a one-to-one correspondence 3593 // between Use-Def edges and Def-Use edges in the graph. 3594 void Compile::verify_graph_edges(bool no_dead_code) { 3595 if (VerifyGraphEdges) { 3596 ResourceArea *area = Thread::current()->resource_area(); 3597 Unique_Node_List visited(area); 3598 // Call recursive graph walk to check edges 3599 _root->verify_edges(visited); 3600 if (no_dead_code) { 3601 // Now make sure that no visited node is used by an unvisited node. 3602 bool dead_nodes = 0; 3603 Unique_Node_List checked(area); 3604 while (visited.size() > 0) { 3605 Node* n = visited.pop(); 3606 checked.push(n); 3607 for (uint i = 0; i < n->outcnt(); i++) { 3608 Node* use = n->raw_out(i); 3609 if (checked.member(use)) continue; // already checked 3610 if (visited.member(use)) continue; // already in the graph 3611 if (use->is_Con()) continue; // a dead ConNode is OK 3612 // At this point, we have found a dead node which is DU-reachable. 3613 if (dead_nodes++ == 0) 3614 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3615 use->dump(2); 3616 tty->print_cr("---"); 3617 checked.push(use); // No repeats; pretend it is now checked. 3618 } 3619 } 3620 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3621 } 3622 } 3623 } 3624 3625 // Verify GC barriers consistency 3626 // Currently supported: 3627 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3628 void Compile::verify_barriers() { 3629 if (UseG1GC || UseShenandoahGC) { 3630 // Verify G1 pre-barriers 3631 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3632 3633 ResourceArea *area = Thread::current()->resource_area(); 3634 Unique_Node_List visited(area); 3635 Node_List worklist(area); 3636 // We're going to walk control flow backwards starting from the Root 3637 worklist.push(_root); 3638 while (worklist.size() > 0) { 3639 Node* x = worklist.pop(); 3640 if (x == NULL || x == top()) continue; 3641 if (visited.member(x)) { 3642 continue; 3643 } else { 3644 visited.push(x); 3645 } 3646 3647 if (x->is_Region()) { 3648 for (uint i = 1; i < x->req(); i++) { 3649 worklist.push(x->in(i)); 3650 } 3651 } else { 3652 worklist.push(x->in(0)); 3653 // We are looking for the pattern: 3654 // /->ThreadLocal 3655 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3656 // \->ConI(0) 3657 // We want to verify that the If and the LoadB have the same control 3658 // See GraphKit::g1_write_barrier_pre() 3659 if (x->is_If()) { 3660 IfNode *iff = x->as_If(); 3661 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3662 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3663 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3664 && cmp->in(1)->is_Load()) { 3665 LoadNode* load = cmp->in(1)->as_Load(); 3666 if (load->is_g1_marking_load()) { 3667 3668 Node* if_ctrl = iff->in(0); 3669 Node* load_ctrl = load->in(0); 3670 3671 if (if_ctrl != load_ctrl) { 3672 // Skip possible CProj->NeverBranch in infinite loops 3673 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3674 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3675 if_ctrl = if_ctrl->in(0)->in(0); 3676 } 3677 } 3678 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3679 } 3680 } 3681 } 3682 } 3683 } 3684 } 3685 } 3686 } 3687 3688 #endif 3689 3690 // The Compile object keeps track of failure reasons separately from the ciEnv. 3691 // This is required because there is not quite a 1-1 relation between the 3692 // ciEnv and its compilation task and the Compile object. Note that one 3693 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3694 // to backtrack and retry without subsuming loads. Other than this backtracking 3695 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3696 // by the logic in C2Compiler. 3697 void Compile::record_failure(const char* reason) { 3698 if (log() != NULL) { 3699 log()->elem("failure reason='%s' phase='compile'", reason); 3700 } 3701 if (_failure_reason == NULL) { 3702 // Record the first failure reason. 3703 _failure_reason = reason; 3704 } 3705 3706 EventCompilerFailure event; 3707 if (event.should_commit()) { 3708 event.set_compileID(Compile::compile_id()); 3709 event.set_failure(reason); 3710 event.commit(); 3711 } 3712 3713 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3714 C->print_method(PHASE_FAILURE); 3715 } 3716 _root = NULL; // flush the graph, too 3717 } 3718 3719 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3720 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3721 _phase_name(name), _dolog(dolog) 3722 { 3723 if (dolog) { 3724 C = Compile::current(); 3725 _log = C->log(); 3726 } else { 3727 C = NULL; 3728 _log = NULL; 3729 } 3730 if (_log != NULL) { 3731 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3732 _log->stamp(); 3733 _log->end_head(); 3734 } 3735 } 3736 3737 Compile::TracePhase::~TracePhase() { 3738 3739 C = Compile::current(); 3740 if (_dolog) { 3741 _log = C->log(); 3742 } else { 3743 _log = NULL; 3744 } 3745 3746 #ifdef ASSERT 3747 if (PrintIdealNodeCount) { 3748 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3749 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3750 } 3751 3752 if (VerifyIdealNodeCount) { 3753 Compile::current()->print_missing_nodes(); 3754 } 3755 #endif 3756 3757 if (_log != NULL) { 3758 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3759 } 3760 } 3761 3762 //============================================================================= 3763 // Two Constant's are equal when the type and the value are equal. 3764 bool Compile::Constant::operator==(const Constant& other) { 3765 if (type() != other.type() ) return false; 3766 if (can_be_reused() != other.can_be_reused()) return false; 3767 // For floating point values we compare the bit pattern. 3768 switch (type()) { 3769 case T_FLOAT: return (_v._value.i == other._v._value.i); 3770 case T_LONG: 3771 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3772 case T_OBJECT: 3773 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3774 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3775 case T_METADATA: return (_v._metadata == other._v._metadata); 3776 default: ShouldNotReachHere(); 3777 } 3778 return false; 3779 } 3780 3781 static int type_to_size_in_bytes(BasicType t) { 3782 switch (t) { 3783 case T_LONG: return sizeof(jlong ); 3784 case T_FLOAT: return sizeof(jfloat ); 3785 case T_DOUBLE: return sizeof(jdouble); 3786 case T_METADATA: return sizeof(Metadata*); 3787 // We use T_VOID as marker for jump-table entries (labels) which 3788 // need an internal word relocation. 3789 case T_VOID: 3790 case T_ADDRESS: 3791 case T_OBJECT: return sizeof(jobject); 3792 } 3793 3794 ShouldNotReachHere(); 3795 return -1; 3796 } 3797 3798 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3799 // sort descending 3800 if (a->freq() > b->freq()) return -1; 3801 if (a->freq() < b->freq()) return 1; 3802 return 0; 3803 } 3804 3805 void Compile::ConstantTable::calculate_offsets_and_size() { 3806 // First, sort the array by frequencies. 3807 _constants.sort(qsort_comparator); 3808 3809 #ifdef ASSERT 3810 // Make sure all jump-table entries were sorted to the end of the 3811 // array (they have a negative frequency). 3812 bool found_void = false; 3813 for (int i = 0; i < _constants.length(); i++) { 3814 Constant con = _constants.at(i); 3815 if (con.type() == T_VOID) 3816 found_void = true; // jump-tables 3817 else 3818 assert(!found_void, "wrong sorting"); 3819 } 3820 #endif 3821 3822 int offset = 0; 3823 for (int i = 0; i < _constants.length(); i++) { 3824 Constant* con = _constants.adr_at(i); 3825 3826 // Align offset for type. 3827 int typesize = type_to_size_in_bytes(con->type()); 3828 offset = align_size_up(offset, typesize); 3829 con->set_offset(offset); // set constant's offset 3830 3831 if (con->type() == T_VOID) { 3832 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3833 offset = offset + typesize * n->outcnt(); // expand jump-table 3834 } else { 3835 offset = offset + typesize; 3836 } 3837 } 3838 3839 // Align size up to the next section start (which is insts; see 3840 // CodeBuffer::align_at_start). 3841 assert(_size == -1, "already set?"); 3842 _size = align_size_up(offset, CodeEntryAlignment); 3843 } 3844 3845 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3846 MacroAssembler _masm(&cb); 3847 for (int i = 0; i < _constants.length(); i++) { 3848 Constant con = _constants.at(i); 3849 address constant_addr = NULL; 3850 switch (con.type()) { 3851 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3852 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3853 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3854 case T_OBJECT: { 3855 jobject obj = con.get_jobject(); 3856 int oop_index = _masm.oop_recorder()->find_index(obj); 3857 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3858 break; 3859 } 3860 case T_ADDRESS: { 3861 address addr = (address) con.get_jobject(); 3862 constant_addr = _masm.address_constant(addr); 3863 break; 3864 } 3865 // We use T_VOID as marker for jump-table entries (labels) which 3866 // need an internal word relocation. 3867 case T_VOID: { 3868 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3869 // Fill the jump-table with a dummy word. The real value is 3870 // filled in later in fill_jump_table. 3871 address dummy = (address) n; 3872 constant_addr = _masm.address_constant(dummy); 3873 // Expand jump-table 3874 for (uint i = 1; i < n->outcnt(); i++) { 3875 address temp_addr = _masm.address_constant(dummy + i); 3876 assert(temp_addr, "consts section too small"); 3877 } 3878 break; 3879 } 3880 case T_METADATA: { 3881 Metadata* obj = con.get_metadata(); 3882 int metadata_index = _masm.oop_recorder()->find_index(obj); 3883 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3884 break; 3885 } 3886 default: ShouldNotReachHere(); 3887 } 3888 assert(constant_addr, "consts section too small"); 3889 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3890 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3891 } 3892 } 3893 3894 int Compile::ConstantTable::find_offset(Constant& con) const { 3895 int idx = _constants.find(con); 3896 assert(idx != -1, "constant must be in constant table"); 3897 int offset = _constants.at(idx).offset(); 3898 assert(offset != -1, "constant table not emitted yet?"); 3899 return offset; 3900 } 3901 3902 void Compile::ConstantTable::add(Constant& con) { 3903 if (con.can_be_reused()) { 3904 int idx = _constants.find(con); 3905 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3906 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3907 return; 3908 } 3909 } 3910 (void) _constants.append(con); 3911 } 3912 3913 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3914 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3915 Constant con(type, value, b->_freq); 3916 add(con); 3917 return con; 3918 } 3919 3920 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3921 Constant con(metadata); 3922 add(con); 3923 return con; 3924 } 3925 3926 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3927 jvalue value; 3928 BasicType type = oper->type()->basic_type(); 3929 switch (type) { 3930 case T_LONG: value.j = oper->constantL(); break; 3931 case T_FLOAT: value.f = oper->constantF(); break; 3932 case T_DOUBLE: value.d = oper->constantD(); break; 3933 case T_OBJECT: 3934 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3935 case T_METADATA: return add((Metadata*)oper->constant()); break; 3936 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3937 } 3938 return add(n, type, value); 3939 } 3940 3941 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3942 jvalue value; 3943 // We can use the node pointer here to identify the right jump-table 3944 // as this method is called from Compile::Fill_buffer right before 3945 // the MachNodes are emitted and the jump-table is filled (means the 3946 // MachNode pointers do not change anymore). 3947 value.l = (jobject) n; 3948 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3949 add(con); 3950 return con; 3951 } 3952 3953 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3954 // If called from Compile::scratch_emit_size do nothing. 3955 if (Compile::current()->in_scratch_emit_size()) return; 3956 3957 assert(labels.is_nonempty(), "must be"); 3958 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3959 3960 // Since MachConstantNode::constant_offset() also contains 3961 // table_base_offset() we need to subtract the table_base_offset() 3962 // to get the plain offset into the constant table. 3963 int offset = n->constant_offset() - table_base_offset(); 3964 3965 MacroAssembler _masm(&cb); 3966 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3967 3968 for (uint i = 0; i < n->outcnt(); i++) { 3969 address* constant_addr = &jump_table_base[i]; 3970 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))); 3971 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3972 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3973 } 3974 } 3975 3976 void Compile::dump_inlining() { 3977 if (print_inlining() || print_intrinsics()) { 3978 // Print inlining message for candidates that we couldn't inline 3979 // for lack of space or non constant receiver 3980 for (int i = 0; i < _late_inlines.length(); i++) { 3981 CallGenerator* cg = _late_inlines.at(i); 3982 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); 3983 } 3984 Unique_Node_List useful; 3985 useful.push(root()); 3986 for (uint next = 0; next < useful.size(); ++next) { 3987 Node* n = useful.at(next); 3988 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { 3989 CallNode* call = n->as_Call(); 3990 CallGenerator* cg = call->generator(); 3991 cg->print_inlining_late("receiver not constant"); 3992 } 3993 uint max = n->len(); 3994 for ( uint i = 0; i < max; ++i ) { 3995 Node *m = n->in(i); 3996 if ( m == NULL ) continue; 3997 useful.push(m); 3998 } 3999 } 4000 for (int i = 0; i < _print_inlining_list->length(); i++) { 4001 tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 4002 } 4003 } 4004 } 4005 4006 // Dump inlining replay data to the stream. 4007 // Don't change thread state and acquire any locks. 4008 void Compile::dump_inline_data(outputStream* out) { 4009 InlineTree* inl_tree = ilt(); 4010 if (inl_tree != NULL) { 4011 out->print(" inline %d", inl_tree->count()); 4012 inl_tree->dump_replay_data(out); 4013 } 4014 } 4015 4016 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4017 if (n1->Opcode() < n2->Opcode()) return -1; 4018 else if (n1->Opcode() > n2->Opcode()) return 1; 4019 4020 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())); 4021 for (uint i = 1; i < n1->req(); i++) { 4022 if (n1->in(i) < n2->in(i)) return -1; 4023 else if (n1->in(i) > n2->in(i)) return 1; 4024 } 4025 4026 return 0; 4027 } 4028 4029 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4030 Node* n1 = *n1p; 4031 Node* n2 = *n2p; 4032 4033 return cmp_expensive_nodes(n1, n2); 4034 } 4035 4036 void Compile::sort_expensive_nodes() { 4037 if (!expensive_nodes_sorted()) { 4038 _expensive_nodes->sort(cmp_expensive_nodes); 4039 } 4040 } 4041 4042 bool Compile::expensive_nodes_sorted() const { 4043 for (int i = 1; i < _expensive_nodes->length(); i++) { 4044 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4045 return false; 4046 } 4047 } 4048 return true; 4049 } 4050 4051 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4052 if (_expensive_nodes->length() == 0) { 4053 return false; 4054 } 4055 4056 assert(OptimizeExpensiveOps, "optimization off?"); 4057 4058 // Take this opportunity to remove dead nodes from the list 4059 int j = 0; 4060 for (int i = 0; i < _expensive_nodes->length(); i++) { 4061 Node* n = _expensive_nodes->at(i); 4062 if (!n->is_unreachable(igvn)) { 4063 assert(n->is_expensive(), "should be expensive"); 4064 _expensive_nodes->at_put(j, n); 4065 j++; 4066 } 4067 } 4068 _expensive_nodes->trunc_to(j); 4069 4070 // Then sort the list so that similar nodes are next to each other 4071 // and check for at least two nodes of identical kind with same data 4072 // inputs. 4073 sort_expensive_nodes(); 4074 4075 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4076 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4077 return true; 4078 } 4079 } 4080 4081 return false; 4082 } 4083 4084 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4085 if (_expensive_nodes->length() == 0) { 4086 return; 4087 } 4088 4089 assert(OptimizeExpensiveOps, "optimization off?"); 4090 4091 // Sort to bring similar nodes next to each other and clear the 4092 // control input of nodes for which there's only a single copy. 4093 sort_expensive_nodes(); 4094 4095 int j = 0; 4096 int identical = 0; 4097 int i = 0; 4098 for (; i < _expensive_nodes->length()-1; i++) { 4099 assert(j <= i, "can't write beyond current index"); 4100 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4101 identical++; 4102 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4103 continue; 4104 } 4105 if (identical > 0) { 4106 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4107 identical = 0; 4108 } else { 4109 Node* n = _expensive_nodes->at(i); 4110 igvn.hash_delete(n); 4111 n->set_req(0, NULL); 4112 igvn.hash_insert(n); 4113 } 4114 } 4115 if (identical > 0) { 4116 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4117 } else if (_expensive_nodes->length() >= 1) { 4118 Node* n = _expensive_nodes->at(i); 4119 igvn.hash_delete(n); 4120 n->set_req(0, NULL); 4121 igvn.hash_insert(n); 4122 } 4123 _expensive_nodes->trunc_to(j); 4124 } 4125 4126 void Compile::add_expensive_node(Node * n) { 4127 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4128 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4129 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4130 if (OptimizeExpensiveOps) { 4131 _expensive_nodes->append(n); 4132 } else { 4133 // Clear control input and let IGVN optimize expensive nodes if 4134 // OptimizeExpensiveOps is off. 4135 n->set_req(0, NULL); 4136 } 4137 } 4138 4139 /** 4140 * Remove the speculative part of types and clean up the graph 4141 */ 4142 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4143 if (UseTypeSpeculation) { 4144 Unique_Node_List worklist; 4145 worklist.push(root()); 4146 int modified = 0; 4147 // Go over all type nodes that carry a speculative type, drop the 4148 // speculative part of the type and enqueue the node for an igvn 4149 // which may optimize it out. 4150 for (uint next = 0; next < worklist.size(); ++next) { 4151 Node *n = worklist.at(next); 4152 if (n->is_Type()) { 4153 TypeNode* tn = n->as_Type(); 4154 const Type* t = tn->type(); 4155 const Type* t_no_spec = t->remove_speculative(); 4156 if (t_no_spec != t) { 4157 bool in_hash = igvn.hash_delete(n); 4158 assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table"); 4159 tn->set_type(t_no_spec); 4160 igvn.hash_insert(n); 4161 igvn._worklist.push(n); // give it a chance to go away 4162 modified++; 4163 } 4164 } 4165 uint max = n->len(); 4166 for( uint i = 0; i < max; ++i ) { 4167 Node *m = n->in(i); 4168 if (not_a_node(m)) continue; 4169 worklist.push(m); 4170 } 4171 } 4172 // Drop the speculative part of all types in the igvn's type table 4173 igvn.remove_speculative_types(); 4174 if (modified > 0) { 4175 igvn.optimize(); 4176 } 4177 #ifdef ASSERT 4178 // Verify that after the IGVN is over no speculative type has resurfaced 4179 worklist.clear(); 4180 worklist.push(root()); 4181 for (uint next = 0; next < worklist.size(); ++next) { 4182 Node *n = worklist.at(next); 4183 const Type* t = igvn.type_or_null(n); 4184 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4185 if (n->is_Type()) { 4186 t = n->as_Type()->type(); 4187 assert(t == t->remove_speculative(), "no more speculative types"); 4188 } 4189 uint max = n->len(); 4190 for( uint i = 0; i < max; ++i ) { 4191 Node *m = n->in(i); 4192 if (not_a_node(m)) continue; 4193 worklist.push(m); 4194 } 4195 } 4196 igvn.check_no_speculative_types(); 4197 #endif 4198 } 4199 } 4200 4201 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check) 4202 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) { 4203 if (ctrl != NULL) { 4204 // Express control dependency by a CastII node with a narrow type. 4205 value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */); 4206 // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L 4207 // node from floating above the range check during loop optimizations. Otherwise, the 4208 // ConvI2L node may be eliminated independently of the range check, causing the data path 4209 // to become TOP while the control path is still there (although it's unreachable). 4210 value->set_req(0, ctrl); 4211 // Save CastII node to remove it after loop optimizations. 4212 phase->C->add_range_check_cast(value); 4213 value = phase->transform(value); 4214 } 4215 const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen); 4216 return phase->transform(new (phase->C) ConvI2LNode(value, ltype)); 4217 } 4218 4219 // Auxiliary method to support randomized stressing/fuzzing. 4220 // 4221 // This method can be called the arbitrary number of times, with current count 4222 // as the argument. The logic allows selecting a single candidate from the 4223 // running list of candidates as follows: 4224 // int count = 0; 4225 // Cand* selected = null; 4226 // while(cand = cand->next()) { 4227 // if (randomized_select(++count)) { 4228 // selected = cand; 4229 // } 4230 // } 4231 // 4232 // Including count equalizes the chances any candidate is "selected". 4233 // This is useful when we don't have the complete list of candidates to choose 4234 // from uniformly. In this case, we need to adjust the randomicity of the 4235 // selection, or else we will end up biasing the selection towards the latter 4236 // candidates. 4237 // 4238 // Quick back-envelope calculation shows that for the list of n candidates 4239 // the equal probability for the candidate to persist as "best" can be 4240 // achieved by replacing it with "next" k-th candidate with the probability 4241 // of 1/k. It can be easily shown that by the end of the run, the 4242 // probability for any candidate is converged to 1/n, thus giving the 4243 // uniform distribution among all the candidates. 4244 // 4245 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4246 #define RANDOMIZED_DOMAIN_POW 29 4247 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4248 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4249 bool Compile::randomized_select(int count) { 4250 assert(count > 0, "only positive"); 4251 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4252 } 4253 4254 void Compile::shenandoah_eliminate_g1_wb_pre(Node* call, PhaseIterGVN* igvn) { 4255 assert(UseShenandoahGC && call->is_g1_wb_pre_call(), ""); 4256 Node* c = call->as_Call()->proj_out(TypeFunc::Control); 4257 c = c->unique_ctrl_out(); 4258 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 4259 c = c->unique_ctrl_out(); 4260 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 4261 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 4262 assert(iff->is_If(), "expect test"); 4263 if (!iff->is_shenandoah_marking_if(igvn)) { 4264 c = c->unique_ctrl_out(); 4265 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?"); 4266 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0); 4267 assert(iff->is_shenandoah_marking_if(igvn), "expect marking test"); 4268 } 4269 Node* cmpx = iff->in(1)->in(1); 4270 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ)); 4271 igvn->rehash_node_delayed(call); 4272 call->del_req(call->req()-1); 4273 }