1 /* 2 * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "ci/ciReplay.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/exceptionHandlerTable.hpp" 31 #include "code/nmethod.hpp" 32 #include "compiler/compileLog.hpp" 33 #include "compiler/disassembler.hpp" 34 #include "compiler/oopMap.hpp" 35 #include "opto/addnode.hpp" 36 #include "opto/block.hpp" 37 #include "opto/c2compiler.hpp" 38 #include "opto/callGenerator.hpp" 39 #include "opto/callnode.hpp" 40 #include "opto/cfgnode.hpp" 41 #include "opto/chaitin.hpp" 42 #include "opto/compile.hpp" 43 #include "opto/connode.hpp" 44 #include "opto/divnode.hpp" 45 #include "opto/escape.hpp" 46 #include "opto/idealGraphPrinter.hpp" 47 #include "opto/loopnode.hpp" 48 #include "opto/machnode.hpp" 49 #include "opto/macro.hpp" 50 #include "opto/matcher.hpp" 51 #include "opto/mathexactnode.hpp" 52 #include "opto/memnode.hpp" 53 #include "opto/mulnode.hpp" 54 #include "opto/narrowptrnode.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/stringopts.hpp" 63 #include "opto/type.hpp" 64 #include "opto/vectornode.hpp" 65 #include "runtime/arguments.hpp" 66 #include "runtime/signature.hpp" 67 #include "runtime/stubRoutines.hpp" 68 #include "runtime/timer.hpp" 69 #include "trace/tracing.hpp" 70 #include "utilities/copy.hpp" 71 #ifdef TARGET_ARCH_MODEL_x86_32 72 # include "adfiles/ad_x86_32.hpp" 73 #endif 74 #ifdef TARGET_ARCH_MODEL_x86_64 75 # include "adfiles/ad_x86_64.hpp" 76 #endif 77 #ifdef TARGET_ARCH_MODEL_sparc 78 # include "adfiles/ad_sparc.hpp" 79 #endif 80 #ifdef TARGET_ARCH_MODEL_zero 81 # include "adfiles/ad_zero.hpp" 82 #endif 83 #ifdef TARGET_ARCH_MODEL_arm 84 # include "adfiles/ad_arm.hpp" 85 #endif 86 #ifdef TARGET_ARCH_MODEL_ppc_32 87 # include "adfiles/ad_ppc_32.hpp" 88 #endif 89 #ifdef TARGET_ARCH_MODEL_ppc_64 90 # include "adfiles/ad_ppc_64.hpp" 91 #endif 92 93 94 // -------------------- Compile::mach_constant_base_node ----------------------- 95 // Constant table base node singleton. 96 MachConstantBaseNode* Compile::mach_constant_base_node() { 97 if (_mach_constant_base_node == NULL) { 98 _mach_constant_base_node = new MachConstantBaseNode(); 99 _mach_constant_base_node->add_req(C->root()); 100 } 101 return _mach_constant_base_node; 102 } 103 104 105 /// Support for intrinsics. 106 107 // Return the index at which m must be inserted (or already exists). 108 // The sort order is by the address of the ciMethod, with is_virtual as minor key. 109 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { 110 #ifdef ASSERT 111 for (int i = 1; i < _intrinsics->length(); i++) { 112 CallGenerator* cg1 = _intrinsics->at(i-1); 113 CallGenerator* cg2 = _intrinsics->at(i); 114 assert(cg1->method() != cg2->method() 115 ? cg1->method() < cg2->method() 116 : cg1->is_virtual() < cg2->is_virtual(), 117 "compiler intrinsics list must stay sorted"); 118 } 119 #endif 120 // Binary search sorted list, in decreasing intervals [lo, hi]. 121 int lo = 0, hi = _intrinsics->length()-1; 122 while (lo <= hi) { 123 int mid = (uint)(hi + lo) / 2; 124 ciMethod* mid_m = _intrinsics->at(mid)->method(); 125 if (m < mid_m) { 126 hi = mid-1; 127 } else if (m > mid_m) { 128 lo = mid+1; 129 } else { 130 // look at minor sort key 131 bool mid_virt = _intrinsics->at(mid)->is_virtual(); 132 if (is_virtual < mid_virt) { 133 hi = mid-1; 134 } else if (is_virtual > mid_virt) { 135 lo = mid+1; 136 } else { 137 return mid; // exact match 138 } 139 } 140 } 141 return lo; // inexact match 142 } 143 144 void Compile::register_intrinsic(CallGenerator* cg) { 145 if (_intrinsics == NULL) { 146 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL); 147 } 148 // This code is stolen from ciObjectFactory::insert. 149 // Really, GrowableArray should have methods for 150 // insert_at, remove_at, and binary_search. 151 int len = _intrinsics->length(); 152 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); 153 if (index == len) { 154 _intrinsics->append(cg); 155 } else { 156 #ifdef ASSERT 157 CallGenerator* oldcg = _intrinsics->at(index); 158 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); 159 #endif 160 _intrinsics->append(_intrinsics->at(len-1)); 161 int pos; 162 for (pos = len-2; pos >= index; pos--) { 163 _intrinsics->at_put(pos+1,_intrinsics->at(pos)); 164 } 165 _intrinsics->at_put(index, cg); 166 } 167 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); 168 } 169 170 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { 171 assert(m->is_loaded(), "don't try this on unloaded methods"); 172 if (_intrinsics != NULL) { 173 int index = intrinsic_insertion_index(m, is_virtual); 174 if (index < _intrinsics->length() 175 && _intrinsics->at(index)->method() == m 176 && _intrinsics->at(index)->is_virtual() == is_virtual) { 177 return _intrinsics->at(index); 178 } 179 } 180 // Lazily create intrinsics for intrinsic IDs well-known in the runtime. 181 if (m->intrinsic_id() != vmIntrinsics::_none && 182 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { 183 CallGenerator* cg = make_vm_intrinsic(m, is_virtual); 184 if (cg != NULL) { 185 // Save it for next time: 186 register_intrinsic(cg); 187 return cg; 188 } else { 189 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); 190 } 191 } 192 return NULL; 193 } 194 195 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined 196 // in library_call.cpp. 197 198 199 #ifndef PRODUCT 200 // statistics gathering... 201 202 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; 203 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; 204 205 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { 206 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); 207 int oflags = _intrinsic_hist_flags[id]; 208 assert(flags != 0, "what happened?"); 209 if (is_virtual) { 210 flags |= _intrinsic_virtual; 211 } 212 bool changed = (flags != oflags); 213 if ((flags & _intrinsic_worked) != 0) { 214 juint count = (_intrinsic_hist_count[id] += 1); 215 if (count == 1) { 216 changed = true; // first time 217 } 218 // increment the overall count also: 219 _intrinsic_hist_count[vmIntrinsics::_none] += 1; 220 } 221 if (changed) { 222 if (((oflags ^ flags) & _intrinsic_virtual) != 0) { 223 // Something changed about the intrinsic's virtuality. 224 if ((flags & _intrinsic_virtual) != 0) { 225 // This is the first use of this intrinsic as a virtual call. 226 if (oflags != 0) { 227 // We already saw it as a non-virtual, so note both cases. 228 flags |= _intrinsic_both; 229 } 230 } else if ((oflags & _intrinsic_both) == 0) { 231 // This is the first use of this intrinsic as a non-virtual 232 flags |= _intrinsic_both; 233 } 234 } 235 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); 236 } 237 // update the overall flags also: 238 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; 239 return changed; 240 } 241 242 static char* format_flags(int flags, char* buf) { 243 buf[0] = 0; 244 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); 245 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); 246 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); 247 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); 248 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); 249 if (buf[0] == 0) strcat(buf, ","); 250 assert(buf[0] == ',', "must be"); 251 return &buf[1]; 252 } 253 254 void Compile::print_intrinsic_statistics() { 255 char flagsbuf[100]; 256 ttyLocker ttyl; 257 if (xtty != NULL) xtty->head("statistics type='intrinsic'"); 258 tty->print_cr("Compiler intrinsic usage:"); 259 juint total = _intrinsic_hist_count[vmIntrinsics::_none]; 260 if (total == 0) total = 1; // avoid div0 in case of no successes 261 #define PRINT_STAT_LINE(name, c, f) \ 262 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); 263 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { 264 vmIntrinsics::ID id = (vmIntrinsics::ID) index; 265 int flags = _intrinsic_hist_flags[id]; 266 juint count = _intrinsic_hist_count[id]; 267 if ((flags | count) != 0) { 268 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); 269 } 270 } 271 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); 272 if (xtty != NULL) xtty->tail("statistics"); 273 } 274 275 void Compile::print_statistics() { 276 { ttyLocker ttyl; 277 if (xtty != NULL) xtty->head("statistics type='opto'"); 278 Parse::print_statistics(); 279 PhaseCCP::print_statistics(); 280 PhaseRegAlloc::print_statistics(); 281 Scheduling::print_statistics(); 282 PhasePeephole::print_statistics(); 283 PhaseIdealLoop::print_statistics(); 284 if (xtty != NULL) xtty->tail("statistics"); 285 } 286 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { 287 // put this under its own <statistics> element. 288 print_intrinsic_statistics(); 289 } 290 } 291 #endif //PRODUCT 292 293 // Support for bundling info 294 Bundle* Compile::node_bundling(const Node *n) { 295 assert(valid_bundle_info(n), "oob"); 296 return &_node_bundling_base[n->_idx]; 297 } 298 299 bool Compile::valid_bundle_info(const Node *n) { 300 return (_node_bundling_limit > n->_idx); 301 } 302 303 304 void Compile::gvn_replace_by(Node* n, Node* nn) { 305 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { 306 Node* use = n->last_out(i); 307 bool is_in_table = initial_gvn()->hash_delete(use); 308 uint uses_found = 0; 309 for (uint j = 0; j < use->len(); j++) { 310 if (use->in(j) == n) { 311 if (j < use->req()) 312 use->set_req(j, nn); 313 else 314 use->set_prec(j, nn); 315 uses_found++; 316 } 317 } 318 if (is_in_table) { 319 // reinsert into table 320 initial_gvn()->hash_find_insert(use); 321 } 322 record_for_igvn(use); 323 i -= uses_found; // we deleted 1 or more copies of this edge 324 } 325 } 326 327 328 static inline bool not_a_node(const Node* n) { 329 if (n == NULL) return true; 330 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. 331 if (*(address*)n == badAddress) return true; // kill by Node::destruct 332 return false; 333 } 334 335 // Identify all nodes that are reachable from below, useful. 336 // Use breadth-first pass that records state in a Unique_Node_List, 337 // recursive traversal is slower. 338 void Compile::identify_useful_nodes(Unique_Node_List &useful) { 339 int estimated_worklist_size = unique(); 340 useful.map( estimated_worklist_size, NULL ); // preallocate space 341 342 // Initialize worklist 343 if (root() != NULL) { useful.push(root()); } 344 // If 'top' is cached, declare it useful to preserve cached node 345 if( cached_top_node() ) { useful.push(cached_top_node()); } 346 347 // Push all useful nodes onto the list, breadthfirst 348 for( uint next = 0; next < useful.size(); ++next ) { 349 assert( next < unique(), "Unique useful nodes < total nodes"); 350 Node *n = useful.at(next); 351 uint max = n->len(); 352 for( uint i = 0; i < max; ++i ) { 353 Node *m = n->in(i); 354 if (not_a_node(m)) continue; 355 useful.push(m); 356 } 357 } 358 } 359 360 // Update dead_node_list with any missing dead nodes using useful 361 // list. Consider all non-useful nodes to be useless i.e., dead nodes. 362 void Compile::update_dead_node_list(Unique_Node_List &useful) { 363 uint max_idx = unique(); 364 VectorSet& useful_node_set = useful.member_set(); 365 366 for (uint node_idx = 0; node_idx < max_idx; node_idx++) { 367 // If node with index node_idx is not in useful set, 368 // mark it as dead in dead node list. 369 if (! useful_node_set.test(node_idx) ) { 370 record_dead_node(node_idx); 371 } 372 } 373 } 374 375 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) { 376 int shift = 0; 377 for (int i = 0; i < inlines->length(); i++) { 378 CallGenerator* cg = inlines->at(i); 379 CallNode* call = cg->call_node(); 380 if (shift > 0) { 381 inlines->at_put(i-shift, cg); 382 } 383 if (!useful.member(call)) { 384 shift++; 385 } 386 } 387 inlines->trunc_to(inlines->length()-shift); 388 } 389 390 // Disconnect all useless nodes by disconnecting those at the boundary. 391 void Compile::remove_useless_nodes(Unique_Node_List &useful) { 392 uint next = 0; 393 while (next < useful.size()) { 394 Node *n = useful.at(next++); 395 if (n->is_SafePoint()) { 396 // We're done with a parsing phase. Replaced nodes are not valid 397 // beyond that point. 398 n->as_SafePoint()->delete_replaced_nodes(); 399 } 400 // Use raw traversal of out edges since this code removes out edges 401 int max = n->outcnt(); 402 for (int j = 0; j < max; ++j) { 403 Node* child = n->raw_out(j); 404 if (! useful.member(child)) { 405 assert(!child->is_top() || child != top(), 406 "If top is cached in Compile object it is in useful list"); 407 // Only need to remove this out-edge to the useless node 408 n->raw_del_out(j); 409 --j; 410 --max; 411 } 412 } 413 if (n->outcnt() == 1 && n->has_special_unique_user()) { 414 record_for_igvn(n->unique_out()); 415 } 416 } 417 // Remove useless macro and predicate opaq nodes 418 for (int i = C->macro_count()-1; i >= 0; i--) { 419 Node* n = C->macro_node(i); 420 if (!useful.member(n)) { 421 remove_macro_node(n); 422 } 423 } 424 // Remove useless expensive node 425 for (int i = C->expensive_count()-1; i >= 0; i--) { 426 Node* n = C->expensive_node(i); 427 if (!useful.member(n)) { 428 remove_expensive_node(n); 429 } 430 } 431 // clean up the late inline lists 432 remove_useless_late_inlines(&_string_late_inlines, useful); 433 remove_useless_late_inlines(&_boxing_late_inlines, useful); 434 remove_useless_late_inlines(&_late_inlines, useful); 435 debug_only(verify_graph_edges(true/*check for no_dead_code*/);) 436 } 437 438 //------------------------------frame_size_in_words----------------------------- 439 // frame_slots in units of words 440 int Compile::frame_size_in_words() const { 441 // shift is 0 in LP32 and 1 in LP64 442 const int shift = (LogBytesPerWord - LogBytesPerInt); 443 int words = _frame_slots >> shift; 444 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); 445 return words; 446 } 447 448 // To bang the stack of this compiled method we use the stack size 449 // that the interpreter would need in case of a deoptimization. This 450 // removes the need to bang the stack in the deoptimization blob which 451 // in turn simplifies stack overflow handling. 452 int Compile::bang_size_in_bytes() const { 453 return MAX2(_interpreter_frame_size, frame_size_in_bytes()); 454 } 455 456 // ============================================================================ 457 //------------------------------CompileWrapper--------------------------------- 458 class CompileWrapper : public StackObj { 459 Compile *const _compile; 460 public: 461 CompileWrapper(Compile* compile); 462 463 ~CompileWrapper(); 464 }; 465 466 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { 467 // the Compile* pointer is stored in the current ciEnv: 468 ciEnv* env = compile->env(); 469 assert(env == ciEnv::current(), "must already be a ciEnv active"); 470 assert(env->compiler_data() == NULL, "compile already active?"); 471 env->set_compiler_data(compile); 472 assert(compile == Compile::current(), "sanity"); 473 474 compile->set_type_dict(NULL); 475 compile->set_type_hwm(NULL); 476 compile->set_type_last_size(0); 477 compile->set_last_tf(NULL, NULL); 478 compile->set_indexSet_arena(NULL); 479 compile->set_indexSet_free_block_list(NULL); 480 compile->init_type_arena(); 481 Type::Initialize(compile); 482 _compile->set_scratch_buffer_blob(NULL); 483 _compile->begin_method(); 484 } 485 CompileWrapper::~CompileWrapper() { 486 _compile->end_method(); 487 if (_compile->scratch_buffer_blob() != NULL) 488 BufferBlob::free(_compile->scratch_buffer_blob()); 489 _compile->env()->set_compiler_data(NULL); 490 } 491 492 493 //----------------------------print_compile_messages--------------------------- 494 void Compile::print_compile_messages() { 495 #ifndef PRODUCT 496 // Check if recompiling 497 if (_subsume_loads == false && PrintOpto) { 498 // Recompiling without allowing machine instructions to subsume loads 499 tty->print_cr("*********************************************************"); 500 tty->print_cr("** Bailout: Recompile without subsuming loads **"); 501 tty->print_cr("*********************************************************"); 502 } 503 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { 504 // Recompiling without escape analysis 505 tty->print_cr("*********************************************************"); 506 tty->print_cr("** Bailout: Recompile without escape analysis **"); 507 tty->print_cr("*********************************************************"); 508 } 509 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { 510 // Recompiling without boxing elimination 511 tty->print_cr("*********************************************************"); 512 tty->print_cr("** Bailout: Recompile without boxing elimination **"); 513 tty->print_cr("*********************************************************"); 514 } 515 if (env()->break_at_compile()) { 516 // Open the debugger when compiling this method. 517 tty->print("### Breaking when compiling: "); 518 method()->print_short_name(); 519 tty->cr(); 520 BREAKPOINT; 521 } 522 523 if( PrintOpto ) { 524 if (is_osr_compilation()) { 525 tty->print("[OSR]%3d", _compile_id); 526 } else { 527 tty->print("%3d", _compile_id); 528 } 529 } 530 #endif 531 } 532 533 534 //-----------------------init_scratch_buffer_blob------------------------------ 535 // Construct a temporary BufferBlob and cache it for this compile. 536 void Compile::init_scratch_buffer_blob(int const_size) { 537 // If there is already a scratch buffer blob allocated and the 538 // constant section is big enough, use it. Otherwise free the 539 // current and allocate a new one. 540 BufferBlob* blob = scratch_buffer_blob(); 541 if ((blob != NULL) && (const_size <= _scratch_const_size)) { 542 // Use the current blob. 543 } else { 544 if (blob != NULL) { 545 BufferBlob::free(blob); 546 } 547 548 ResourceMark rm; 549 _scratch_const_size = const_size; 550 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); 551 blob = BufferBlob::create("Compile::scratch_buffer", size); 552 // Record the buffer blob for next time. 553 set_scratch_buffer_blob(blob); 554 // Have we run out of code space? 555 if (scratch_buffer_blob() == NULL) { 556 // Let CompilerBroker disable further compilations. 557 record_failure("Not enough space for scratch buffer in CodeCache"); 558 return; 559 } 560 } 561 562 // Initialize the relocation buffers 563 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; 564 set_scratch_locs_memory(locs_buf); 565 } 566 567 568 //-----------------------scratch_emit_size------------------------------------- 569 // Helper function that computes size by emitting code 570 uint Compile::scratch_emit_size(const Node* n) { 571 // Start scratch_emit_size section. 572 set_in_scratch_emit_size(true); 573 574 // Emit into a trash buffer and count bytes emitted. 575 // This is a pretty expensive way to compute a size, 576 // but it works well enough if seldom used. 577 // All common fixed-size instructions are given a size 578 // method by the AD file. 579 // Note that the scratch buffer blob and locs memory are 580 // allocated at the beginning of the compile task, and 581 // may be shared by several calls to scratch_emit_size. 582 // The allocation of the scratch buffer blob is particularly 583 // expensive, since it has to grab the code cache lock. 584 BufferBlob* blob = this->scratch_buffer_blob(); 585 assert(blob != NULL, "Initialize BufferBlob at start"); 586 assert(blob->size() > MAX_inst_size, "sanity"); 587 relocInfo* locs_buf = scratch_locs_memory(); 588 address blob_begin = blob->content_begin(); 589 address blob_end = (address)locs_buf; 590 assert(blob->content_contains(blob_end), "sanity"); 591 CodeBuffer buf(blob_begin, blob_end - blob_begin); 592 buf.initialize_consts_size(_scratch_const_size); 593 buf.initialize_stubs_size(MAX_stubs_size); 594 assert(locs_buf != NULL, "sanity"); 595 int lsize = MAX_locs_size / 3; 596 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); 597 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); 598 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); 599 600 // Do the emission. 601 602 Label fakeL; // Fake label for branch instructions. 603 Label* saveL = NULL; 604 uint save_bnum = 0; 605 bool is_branch = n->is_MachBranch(); 606 if (is_branch) { 607 MacroAssembler masm(&buf); 608 masm.bind(fakeL); 609 n->as_MachBranch()->save_label(&saveL, &save_bnum); 610 n->as_MachBranch()->label_set(&fakeL, 0); 611 } 612 n->emit(buf, this->regalloc()); 613 if (is_branch) // Restore label. 614 n->as_MachBranch()->label_set(saveL, save_bnum); 615 616 // End scratch_emit_size section. 617 set_in_scratch_emit_size(false); 618 619 return buf.insts_size(); 620 } 621 622 623 // ============================================================================ 624 //------------------------------Compile standard------------------------------- 625 debug_only( int Compile::_debug_idx = 100000; ) 626 627 // Compile a method. entry_bci is -1 for normal compilations and indicates 628 // the continuation bci for on stack replacement. 629 630 631 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, 632 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) 633 : Phase(Compiler), 634 _env(ci_env), 635 _log(ci_env->log()), 636 _compile_id(ci_env->compile_id()), 637 _save_argument_registers(false), 638 _stub_name(NULL), 639 _stub_function(NULL), 640 _stub_entry_point(NULL), 641 _method(target), 642 _entry_bci(osr_bci), 643 _initial_gvn(NULL), 644 _for_igvn(NULL), 645 _warm_calls(NULL), 646 _subsume_loads(subsume_loads), 647 _do_escape_analysis(do_escape_analysis), 648 _eliminate_boxing(eliminate_boxing), 649 _failure_reason(NULL), 650 _code_buffer("Compile::Fill_buffer"), 651 _orig_pc_slot(0), 652 _orig_pc_slot_offset_in_bytes(0), 653 _has_method_handle_invokes(false), 654 _mach_constant_base_node(NULL), 655 _node_bundling_limit(0), 656 _node_bundling_base(NULL), 657 _java_calls(0), 658 _inner_loops(0), 659 _scratch_const_size(-1), 660 _in_scratch_emit_size(false), 661 _dead_node_list(comp_arena()), 662 _dead_node_count(0), 663 #ifndef PRODUCT 664 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), 665 _in_dump_cnt(0), 666 _printer(IdealGraphPrinter::printer()), 667 #endif 668 _congraph(NULL), 669 _replay_inline_data(NULL), 670 _late_inlines(comp_arena(), 2, 0, NULL), 671 _string_late_inlines(comp_arena(), 2, 0, NULL), 672 _boxing_late_inlines(comp_arena(), 2, 0, NULL), 673 _late_inlines_pos(0), 674 _number_of_mh_late_inlines(0), 675 _inlining_progress(false), 676 _inlining_incrementally(false), 677 _print_inlining_list(NULL), 678 _print_inlining_stream(NULL), 679 _print_inlining_idx(0), 680 _print_inlining_output(NULL), 681 _interpreter_frame_size(0) { 682 C = this; 683 684 CompileWrapper cw(this); 685 #ifndef PRODUCT 686 if (TimeCompiler2) { 687 tty->print(" "); 688 target->holder()->name()->print(); 689 tty->print("."); 690 target->print_short_name(); 691 tty->print(" "); 692 } 693 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); 694 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); 695 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); 696 if (!print_opto_assembly) { 697 bool print_assembly = (PrintAssembly || _method->should_print_assembly()); 698 if (print_assembly && !Disassembler::can_decode()) { 699 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); 700 print_opto_assembly = true; 701 } 702 } 703 set_print_assembly(print_opto_assembly); 704 set_parsed_irreducible_loop(false); 705 706 if (method()->has_option("ReplayInline")) { 707 _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level()); 708 } 709 #endif 710 set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining)); 711 set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics")); 712 set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it 713 714 if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) { 715 // Make sure the method being compiled gets its own MDO, 716 // so we can at least track the decompile_count(). 717 // Need MDO to record RTM code generation state. 718 method()->ensure_method_data(); 719 } 720 721 Init(::AliasLevel); 722 723 724 print_compile_messages(); 725 726 _ilt = InlineTree::build_inline_tree_root(); 727 728 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice 729 assert(num_alias_types() >= AliasIdxRaw, ""); 730 731 #define MINIMUM_NODE_HASH 1023 732 // Node list that Iterative GVN will start with 733 Unique_Node_List for_igvn(comp_arena()); 734 set_for_igvn(&for_igvn); 735 736 // GVN that will be run immediately on new nodes 737 uint estimated_size = method()->code_size()*4+64; 738 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); 739 PhaseGVN gvn(node_arena(), estimated_size); 740 set_initial_gvn(&gvn); 741 742 print_inlining_init(); 743 { // Scope for timing the parser 744 TracePhase t3("parse", &_t_parser, true); 745 746 // Put top into the hash table ASAP. 747 initial_gvn()->transform_no_reclaim(top()); 748 749 // Set up tf(), start(), and find a CallGenerator. 750 CallGenerator* cg = NULL; 751 if (is_osr_compilation()) { 752 const TypeTuple *domain = StartOSRNode::osr_domain(); 753 const TypeTuple *range = TypeTuple::make_range(method()->signature()); 754 init_tf(TypeFunc::make(domain, range)); 755 StartNode* s = new StartOSRNode(root(), domain); 756 initial_gvn()->set_type_bottom(s); 757 init_start(s); 758 cg = CallGenerator::for_osr(method(), entry_bci()); 759 } else { 760 // Normal case. 761 init_tf(TypeFunc::make(method())); 762 StartNode* s = new StartNode(root(), tf()->domain()); 763 initial_gvn()->set_type_bottom(s); 764 init_start(s); 765 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { 766 // With java.lang.ref.reference.get() we must go through the 767 // intrinsic when G1 is enabled - even when get() is the root 768 // method of the compile - so that, if necessary, the value in 769 // the referent field of the reference object gets recorded by 770 // the pre-barrier code. 771 // Specifically, if G1 is enabled, the value in the referent 772 // field is recorded by the G1 SATB pre barrier. This will 773 // result in the referent being marked live and the reference 774 // object removed from the list of discovered references during 775 // reference processing. 776 cg = find_intrinsic(method(), false); 777 } 778 if (cg == NULL) { 779 float past_uses = method()->interpreter_invocation_count(); 780 float expected_uses = past_uses; 781 cg = CallGenerator::for_inline(method(), expected_uses); 782 } 783 } 784 if (failing()) return; 785 if (cg == NULL) { 786 record_method_not_compilable_all_tiers("cannot parse method"); 787 return; 788 } 789 JVMState* jvms = build_start_state(start(), tf()); 790 if ((jvms = cg->generate(jvms)) == NULL) { 791 record_method_not_compilable("method parse failed"); 792 return; 793 } 794 GraphKit kit(jvms); 795 796 if (!kit.stopped()) { 797 // Accept return values, and transfer control we know not where. 798 // This is done by a special, unique ReturnNode bound to root. 799 return_values(kit.jvms()); 800 } 801 802 if (kit.has_exceptions()) { 803 // Any exceptions that escape from this call must be rethrown 804 // to whatever caller is dynamically above us on the stack. 805 // This is done by a special, unique RethrowNode bound to root. 806 rethrow_exceptions(kit.transfer_exceptions_into_jvms()); 807 } 808 809 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); 810 811 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { 812 inline_string_calls(true); 813 } 814 815 if (failing()) return; 816 817 print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 818 819 // Remove clutter produced by parsing. 820 if (!failing()) { 821 ResourceMark rm; 822 PhaseRemoveUseless pru(initial_gvn(), &for_igvn); 823 } 824 } 825 826 // Note: Large methods are capped off in do_one_bytecode(). 827 if (failing()) return; 828 829 // After parsing, node notes are no longer automagic. 830 // They must be propagated by register_new_node_with_optimizer(), 831 // clone(), or the like. 832 set_default_node_notes(NULL); 833 834 for (;;) { 835 int successes = Inline_Warm(); 836 if (failing()) return; 837 if (successes == 0) break; 838 } 839 840 // Drain the list. 841 Finish_Warm(); 842 #ifndef PRODUCT 843 if (_printer) { 844 _printer->print_inlining(this); 845 } 846 #endif 847 848 if (failing()) return; 849 NOT_PRODUCT( verify_graph_edges(); ) 850 851 // Now optimize 852 Optimize(); 853 if (failing()) return; 854 NOT_PRODUCT( verify_graph_edges(); ) 855 856 #ifndef PRODUCT 857 if (PrintIdeal) { 858 ttyLocker ttyl; // keep the following output all in one block 859 // This output goes directly to the tty, not the compiler log. 860 // To enable tools to match it up with the compilation activity, 861 // be sure to tag this tty output with the compile ID. 862 if (xtty != NULL) { 863 xtty->head("ideal compile_id='%d'%s", compile_id(), 864 is_osr_compilation() ? " compile_kind='osr'" : 865 ""); 866 } 867 root()->dump(9999); 868 if (xtty != NULL) { 869 xtty->tail("ideal"); 870 } 871 } 872 #endif 873 874 NOT_PRODUCT( verify_barriers(); ) 875 876 // Dump compilation data to replay it. 877 if (method()->has_option("DumpReplay")) { 878 env()->dump_replay_data(_compile_id); 879 } 880 if (method()->has_option("DumpInline") && (ilt() != NULL)) { 881 env()->dump_inline_data(_compile_id); 882 } 883 884 // Now that we know the size of all the monitors we can add a fixed slot 885 // for the original deopt pc. 886 887 _orig_pc_slot = fixed_slots(); 888 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); 889 set_fixed_slots(next_slot); 890 891 // Compute when to use implicit null checks. Used by matching trap based 892 // nodes and NullCheck optimization. 893 set_allowed_deopt_reasons(); 894 895 // Now generate code 896 Code_Gen(); 897 if (failing()) return; 898 899 // Check if we want to skip execution of all compiled code. 900 { 901 #ifndef PRODUCT 902 if (OptoNoExecute) { 903 record_method_not_compilable("+OptoNoExecute"); // Flag as failed 904 return; 905 } 906 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); 907 #endif 908 909 if (is_osr_compilation()) { 910 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); 911 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); 912 } else { 913 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); 914 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); 915 } 916 917 env()->register_method(_method, _entry_bci, 918 &_code_offsets, 919 _orig_pc_slot_offset_in_bytes, 920 code_buffer(), 921 frame_size_in_words(), _oop_map_set, 922 &_handler_table, &_inc_table, 923 compiler, 924 env()->comp_level(), 925 has_unsafe_access(), 926 SharedRuntime::is_wide_vector(max_vector_size()), 927 rtm_state() 928 ); 929 930 if (log() != NULL) // Print code cache state into compiler log 931 log()->code_cache_state(); 932 } 933 } 934 935 //------------------------------Compile---------------------------------------- 936 // Compile a runtime stub 937 Compile::Compile( ciEnv* ci_env, 938 TypeFunc_generator generator, 939 address stub_function, 940 const char *stub_name, 941 int is_fancy_jump, 942 bool pass_tls, 943 bool save_arg_registers, 944 bool return_pc ) 945 : Phase(Compiler), 946 _env(ci_env), 947 _log(ci_env->log()), 948 _compile_id(0), 949 _save_argument_registers(save_arg_registers), 950 _method(NULL), 951 _stub_name(stub_name), 952 _stub_function(stub_function), 953 _stub_entry_point(NULL), 954 _entry_bci(InvocationEntryBci), 955 _initial_gvn(NULL), 956 _for_igvn(NULL), 957 _warm_calls(NULL), 958 _orig_pc_slot(0), 959 _orig_pc_slot_offset_in_bytes(0), 960 _subsume_loads(true), 961 _do_escape_analysis(false), 962 _eliminate_boxing(false), 963 _failure_reason(NULL), 964 _code_buffer("Compile::Fill_buffer"), 965 _has_method_handle_invokes(false), 966 _mach_constant_base_node(NULL), 967 _node_bundling_limit(0), 968 _node_bundling_base(NULL), 969 _java_calls(0), 970 _inner_loops(0), 971 #ifndef PRODUCT 972 _trace_opto_output(TraceOptoOutput), 973 _in_dump_cnt(0), 974 _printer(NULL), 975 #endif 976 _dead_node_list(comp_arena()), 977 _dead_node_count(0), 978 _congraph(NULL), 979 _replay_inline_data(NULL), 980 _number_of_mh_late_inlines(0), 981 _inlining_progress(false), 982 _inlining_incrementally(false), 983 _print_inlining_list(NULL), 984 _print_inlining_stream(NULL), 985 _print_inlining_idx(0), 986 _print_inlining_output(NULL), 987 _allowed_reasons(0), 988 _interpreter_frame_size(0) { 989 C = this; 990 991 #ifndef PRODUCT 992 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); 993 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); 994 set_print_assembly(PrintFrameConverterAssembly); 995 set_parsed_irreducible_loop(false); 996 #endif 997 set_has_irreducible_loop(false); // no loops 998 999 CompileWrapper cw(this); 1000 Init(/*AliasLevel=*/ 0); 1001 init_tf((*generator)()); 1002 1003 { 1004 // The following is a dummy for the sake of GraphKit::gen_stub 1005 Unique_Node_List for_igvn(comp_arena()); 1006 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this 1007 PhaseGVN gvn(Thread::current()->resource_area(),255); 1008 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively 1009 gvn.transform_no_reclaim(top()); 1010 1011 GraphKit kit; 1012 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); 1013 } 1014 1015 NOT_PRODUCT( verify_graph_edges(); ) 1016 Code_Gen(); 1017 if (failing()) return; 1018 1019 1020 // Entry point will be accessed using compile->stub_entry_point(); 1021 if (code_buffer() == NULL) { 1022 Matcher::soft_match_failure(); 1023 } else { 1024 if (PrintAssembly && (WizardMode || Verbose)) 1025 tty->print_cr("### Stub::%s", stub_name); 1026 1027 if (!failing()) { 1028 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); 1029 1030 // Make the NMethod 1031 // For now we mark the frame as never safe for profile stackwalking 1032 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, 1033 code_buffer(), 1034 CodeOffsets::frame_never_safe, 1035 // _code_offsets.value(CodeOffsets::Frame_Complete), 1036 frame_size_in_words(), 1037 _oop_map_set, 1038 save_arg_registers); 1039 assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); 1040 1041 _stub_entry_point = rs->entry_point(); 1042 } 1043 } 1044 } 1045 1046 //------------------------------Init------------------------------------------- 1047 // Prepare for a single compilation 1048 void Compile::Init(int aliaslevel) { 1049 _unique = 0; 1050 _regalloc = NULL; 1051 1052 _tf = NULL; // filled in later 1053 _top = NULL; // cached later 1054 _matcher = NULL; // filled in later 1055 _cfg = NULL; // filled in later 1056 1057 set_24_bit_selection_and_mode(Use24BitFP, false); 1058 1059 _node_note_array = NULL; 1060 _default_node_notes = NULL; 1061 1062 _immutable_memory = NULL; // filled in at first inquiry 1063 1064 // Globally visible Nodes 1065 // First set TOP to NULL to give safe behavior during creation of RootNode 1066 set_cached_top_node(NULL); 1067 set_root(new RootNode()); 1068 // Now that you have a Root to point to, create the real TOP 1069 set_cached_top_node( new ConNode(Type::TOP) ); 1070 set_recent_alloc(NULL, NULL); 1071 1072 // Create Debug Information Recorder to record scopes, oopmaps, etc. 1073 env()->set_oop_recorder(new OopRecorder(env()->arena())); 1074 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); 1075 env()->set_dependencies(new Dependencies(env())); 1076 1077 _fixed_slots = 0; 1078 set_has_split_ifs(false); 1079 set_has_loops(has_method() && method()->has_loops()); // first approximation 1080 set_has_stringbuilder(false); 1081 set_has_boxed_value(false); 1082 _trap_can_recompile = false; // no traps emitted yet 1083 _major_progress = true; // start out assuming good things will happen 1084 set_has_unsafe_access(false); 1085 set_max_vector_size(0); 1086 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); 1087 set_decompile_count(0); 1088 1089 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); 1090 set_num_loop_opts(LoopOptsCount); 1091 set_do_inlining(Inline); 1092 set_max_inline_size(MaxInlineSize); 1093 set_freq_inline_size(FreqInlineSize); 1094 set_do_scheduling(OptoScheduling); 1095 set_do_count_invocations(false); 1096 set_do_method_data_update(false); 1097 set_age_code(has_method() && method()->profile_aging()); 1098 set_rtm_state(NoRTM); // No RTM lock eliding by default 1099 #if INCLUDE_RTM_OPT 1100 if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) { 1101 int rtm_state = method()->method_data()->rtm_state(); 1102 if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) { 1103 // Don't generate RTM lock eliding code. 1104 set_rtm_state(NoRTM); 1105 } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) { 1106 // Generate RTM lock eliding code without abort ratio calculation code. 1107 set_rtm_state(UseRTM); 1108 } else if (UseRTMDeopt) { 1109 // Generate RTM lock eliding code and include abort ratio calculation 1110 // code if UseRTMDeopt is on. 1111 set_rtm_state(ProfileRTM); 1112 } 1113 } 1114 #endif 1115 if (debug_info()->recording_non_safepoints()) { 1116 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*> 1117 (comp_arena(), 8, 0, NULL)); 1118 set_default_node_notes(Node_Notes::make(this)); 1119 } 1120 1121 // // -- Initialize types before each compile -- 1122 // // Update cached type information 1123 // if( _method && _method->constants() ) 1124 // Type::update_loaded_types(_method, _method->constants()); 1125 1126 // Init alias_type map. 1127 if (!_do_escape_analysis && aliaslevel == 3) 1128 aliaslevel = 2; // No unique types without escape analysis 1129 _AliasLevel = aliaslevel; 1130 const int grow_ats = 16; 1131 _max_alias_types = grow_ats; 1132 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); 1133 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); 1134 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); 1135 { 1136 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; 1137 } 1138 // Initialize the first few types. 1139 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); 1140 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); 1141 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); 1142 _num_alias_types = AliasIdxRaw+1; 1143 // Zero out the alias type cache. 1144 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); 1145 // A NULL adr_type hits in the cache right away. Preload the right answer. 1146 probe_alias_cache(NULL)->_index = AliasIdxTop; 1147 1148 _intrinsics = NULL; 1149 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1150 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1151 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL); 1152 register_library_intrinsics(); 1153 } 1154 1155 //---------------------------init_start---------------------------------------- 1156 // Install the StartNode on this compile object. 1157 void Compile::init_start(StartNode* s) { 1158 if (failing()) 1159 return; // already failing 1160 assert(s == start(), ""); 1161 } 1162 1163 StartNode* Compile::start() const { 1164 assert(!failing(), ""); 1165 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { 1166 Node* start = root()->fast_out(i); 1167 if( start->is_Start() ) 1168 return start->as_Start(); 1169 } 1170 fatal("Did not find Start node!"); 1171 return NULL; 1172 } 1173 1174 //-------------------------------immutable_memory------------------------------------- 1175 // Access immutable memory 1176 Node* Compile::immutable_memory() { 1177 if (_immutable_memory != NULL) { 1178 return _immutable_memory; 1179 } 1180 StartNode* s = start(); 1181 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { 1182 Node *p = s->fast_out(i); 1183 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { 1184 _immutable_memory = p; 1185 return _immutable_memory; 1186 } 1187 } 1188 ShouldNotReachHere(); 1189 return NULL; 1190 } 1191 1192 //----------------------set_cached_top_node------------------------------------ 1193 // Install the cached top node, and make sure Node::is_top works correctly. 1194 void Compile::set_cached_top_node(Node* tn) { 1195 if (tn != NULL) verify_top(tn); 1196 Node* old_top = _top; 1197 _top = tn; 1198 // Calling Node::setup_is_top allows the nodes the chance to adjust 1199 // their _out arrays. 1200 if (_top != NULL) _top->setup_is_top(); 1201 if (old_top != NULL) old_top->setup_is_top(); 1202 assert(_top == NULL || top()->is_top(), ""); 1203 } 1204 1205 #ifdef ASSERT 1206 uint Compile::count_live_nodes_by_graph_walk() { 1207 Unique_Node_List useful(comp_arena()); 1208 // Get useful node list by walking the graph. 1209 identify_useful_nodes(useful); 1210 return useful.size(); 1211 } 1212 1213 void Compile::print_missing_nodes() { 1214 1215 // Return if CompileLog is NULL and PrintIdealNodeCount is false. 1216 if ((_log == NULL) && (! PrintIdealNodeCount)) { 1217 return; 1218 } 1219 1220 // This is an expensive function. It is executed only when the user 1221 // specifies VerifyIdealNodeCount option or otherwise knows the 1222 // additional work that needs to be done to identify reachable nodes 1223 // by walking the flow graph and find the missing ones using 1224 // _dead_node_list. 1225 1226 Unique_Node_List useful(comp_arena()); 1227 // Get useful node list by walking the graph. 1228 identify_useful_nodes(useful); 1229 1230 uint l_nodes = C->live_nodes(); 1231 uint l_nodes_by_walk = useful.size(); 1232 1233 if (l_nodes != l_nodes_by_walk) { 1234 if (_log != NULL) { 1235 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); 1236 _log->stamp(); 1237 _log->end_head(); 1238 } 1239 VectorSet& useful_member_set = useful.member_set(); 1240 int last_idx = l_nodes_by_walk; 1241 for (int i = 0; i < last_idx; i++) { 1242 if (useful_member_set.test(i)) { 1243 if (_dead_node_list.test(i)) { 1244 if (_log != NULL) { 1245 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); 1246 } 1247 if (PrintIdealNodeCount) { 1248 // Print the log message to tty 1249 tty->print_cr("mismatched_node idx='%d' both live and dead'", i); 1250 useful.at(i)->dump(); 1251 } 1252 } 1253 } 1254 else if (! _dead_node_list.test(i)) { 1255 if (_log != NULL) { 1256 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); 1257 } 1258 if (PrintIdealNodeCount) { 1259 // Print the log message to tty 1260 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); 1261 } 1262 } 1263 } 1264 if (_log != NULL) { 1265 _log->tail("mismatched_nodes"); 1266 } 1267 } 1268 } 1269 #endif 1270 1271 #ifndef PRODUCT 1272 void Compile::verify_top(Node* tn) const { 1273 if (tn != NULL) { 1274 assert(tn->is_Con(), "top node must be a constant"); 1275 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); 1276 assert(tn->in(0) != NULL, "must have live top node"); 1277 } 1278 } 1279 #endif 1280 1281 1282 ///-------------------Managing Per-Node Debug & Profile Info------------------- 1283 1284 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) { 1285 guarantee(arr != NULL, ""); 1286 int num_blocks = arr->length(); 1287 if (grow_by < num_blocks) grow_by = num_blocks; 1288 int num_notes = grow_by * _node_notes_block_size; 1289 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); 1290 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); 1291 while (num_notes > 0) { 1292 arr->append(notes); 1293 notes += _node_notes_block_size; 1294 num_notes -= _node_notes_block_size; 1295 } 1296 assert(num_notes == 0, "exact multiple, please"); 1297 } 1298 1299 bool Compile::copy_node_notes_to(Node* dest, Node* source) { 1300 if (source == NULL || dest == NULL) return false; 1301 1302 if (dest->is_Con()) 1303 return false; // Do not push debug info onto constants. 1304 1305 #ifdef ASSERT 1306 // Leave a bread crumb trail pointing to the original node: 1307 if (dest != NULL && dest != source && dest->debug_orig() == NULL) { 1308 dest->set_debug_orig(source); 1309 } 1310 #endif 1311 1312 if (node_note_array() == NULL) 1313 return false; // Not collecting any notes now. 1314 1315 // This is a copy onto a pre-existing node, which may already have notes. 1316 // If both nodes have notes, do not overwrite any pre-existing notes. 1317 Node_Notes* source_notes = node_notes_at(source->_idx); 1318 if (source_notes == NULL || source_notes->is_clear()) return false; 1319 Node_Notes* dest_notes = node_notes_at(dest->_idx); 1320 if (dest_notes == NULL || dest_notes->is_clear()) { 1321 return set_node_notes_at(dest->_idx, source_notes); 1322 } 1323 1324 Node_Notes merged_notes = (*source_notes); 1325 // The order of operations here ensures that dest notes will win... 1326 merged_notes.update_from(dest_notes); 1327 return set_node_notes_at(dest->_idx, &merged_notes); 1328 } 1329 1330 1331 //--------------------------allow_range_check_smearing------------------------- 1332 // Gating condition for coalescing similar range checks. 1333 // Sometimes we try 'speculatively' replacing a series of a range checks by a 1334 // single covering check that is at least as strong as any of them. 1335 // If the optimization succeeds, the simplified (strengthened) range check 1336 // will always succeed. If it fails, we will deopt, and then give up 1337 // on the optimization. 1338 bool Compile::allow_range_check_smearing() const { 1339 // If this method has already thrown a range-check, 1340 // assume it was because we already tried range smearing 1341 // and it failed. 1342 uint already_trapped = trap_count(Deoptimization::Reason_range_check); 1343 return !already_trapped; 1344 } 1345 1346 1347 //------------------------------flatten_alias_type----------------------------- 1348 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { 1349 int offset = tj->offset(); 1350 TypePtr::PTR ptr = tj->ptr(); 1351 1352 // Known instance (scalarizable allocation) alias only with itself. 1353 bool is_known_inst = tj->isa_oopptr() != NULL && 1354 tj->is_oopptr()->is_known_instance(); 1355 1356 // Process weird unsafe references. 1357 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { 1358 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); 1359 assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); 1360 tj = TypeOopPtr::BOTTOM; 1361 ptr = tj->ptr(); 1362 offset = tj->offset(); 1363 } 1364 1365 // Array pointers need some flattening 1366 const TypeAryPtr *ta = tj->isa_aryptr(); 1367 if (ta && ta->is_stable()) { 1368 // Erase stability property for alias analysis. 1369 tj = ta = ta->cast_to_stable(false); 1370 } 1371 if( ta && is_known_inst ) { 1372 if ( offset != Type::OffsetBot && 1373 offset > arrayOopDesc::length_offset_in_bytes() ) { 1374 offset = Type::OffsetBot; // Flatten constant access into array body only 1375 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); 1376 } 1377 } else if( ta && _AliasLevel >= 2 ) { 1378 // For arrays indexed by constant indices, we flatten the alias 1379 // space to include all of the array body. Only the header, klass 1380 // and array length can be accessed un-aliased. 1381 if( offset != Type::OffsetBot ) { 1382 if( ta->const_oop() ) { // MethodData* or Method* 1383 offset = Type::OffsetBot; // Flatten constant access into array body 1384 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); 1385 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { 1386 // range is OK as-is. 1387 tj = ta = TypeAryPtr::RANGE; 1388 } else if( offset == oopDesc::klass_offset_in_bytes() ) { 1389 tj = TypeInstPtr::KLASS; // all klass loads look alike 1390 ta = TypeAryPtr::RANGE; // generic ignored junk 1391 ptr = TypePtr::BotPTR; 1392 } else if( offset == oopDesc::mark_offset_in_bytes() ) { 1393 tj = TypeInstPtr::MARK; 1394 ta = TypeAryPtr::RANGE; // generic ignored junk 1395 ptr = TypePtr::BotPTR; 1396 } else { // Random constant offset into array body 1397 offset = Type::OffsetBot; // Flatten constant access into array body 1398 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); 1399 } 1400 } 1401 // Arrays of fixed size alias with arrays of unknown size. 1402 if (ta->size() != TypeInt::POS) { 1403 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); 1404 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); 1405 } 1406 // Arrays of known objects become arrays of unknown objects. 1407 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { 1408 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); 1409 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1410 } 1411 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { 1412 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); 1413 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); 1414 } 1415 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so 1416 // cannot be distinguished by bytecode alone. 1417 if (ta->elem() == TypeInt::BOOL) { 1418 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); 1419 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); 1420 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); 1421 } 1422 // During the 2nd round of IterGVN, NotNull castings are removed. 1423 // Make sure the Bottom and NotNull variants alias the same. 1424 // Also, make sure exact and non-exact variants alias the same. 1425 if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) { 1426 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); 1427 } 1428 } 1429 1430 // Oop pointers need some flattening 1431 const TypeInstPtr *to = tj->isa_instptr(); 1432 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { 1433 ciInstanceKlass *k = to->klass()->as_instance_klass(); 1434 if( ptr == TypePtr::Constant ) { 1435 if (to->klass() != ciEnv::current()->Class_klass() || 1436 offset < k->size_helper() * wordSize) { 1437 // No constant oop pointers (such as Strings); they alias with 1438 // unknown strings. 1439 assert(!is_known_inst, "not scalarizable allocation"); 1440 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1441 } 1442 } else if( is_known_inst ) { 1443 tj = to; // Keep NotNull and klass_is_exact for instance type 1444 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { 1445 // During the 2nd round of IterGVN, NotNull castings are removed. 1446 // Make sure the Bottom and NotNull variants alias the same. 1447 // Also, make sure exact and non-exact variants alias the same. 1448 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); 1449 } 1450 if (to->speculative() != NULL) { 1451 tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id()); 1452 } 1453 // Canonicalize the holder of this field 1454 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { 1455 // First handle header references such as a LoadKlassNode, even if the 1456 // object's klass is unloaded at compile time (4965979). 1457 if (!is_known_inst) { // Do it only for non-instance types 1458 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); 1459 } 1460 } else if (offset < 0 || offset >= k->size_helper() * wordSize) { 1461 // Static fields are in the space above the normal instance 1462 // fields in the java.lang.Class instance. 1463 if (to->klass() != ciEnv::current()->Class_klass()) { 1464 to = NULL; 1465 tj = TypeOopPtr::BOTTOM; 1466 offset = tj->offset(); 1467 } 1468 } else { 1469 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); 1470 if (!k->equals(canonical_holder) || tj->offset() != offset) { 1471 if( is_known_inst ) { 1472 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); 1473 } else { 1474 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); 1475 } 1476 } 1477 } 1478 } 1479 1480 // Klass pointers to object array klasses need some flattening 1481 const TypeKlassPtr *tk = tj->isa_klassptr(); 1482 if( tk ) { 1483 // If we are referencing a field within a Klass, we need 1484 // to assume the worst case of an Object. Both exact and 1485 // inexact types must flatten to the same alias class so 1486 // use NotNull as the PTR. 1487 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { 1488 1489 tj = tk = TypeKlassPtr::make(TypePtr::NotNull, 1490 TypeKlassPtr::OBJECT->klass(), 1491 offset); 1492 } 1493 1494 ciKlass* klass = tk->klass(); 1495 if( klass->is_obj_array_klass() ) { 1496 ciKlass* k = TypeAryPtr::OOPS->klass(); 1497 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs 1498 k = TypeInstPtr::BOTTOM->klass(); 1499 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); 1500 } 1501 1502 // Check for precise loads from the primary supertype array and force them 1503 // to the supertype cache alias index. Check for generic array loads from 1504 // the primary supertype array and also force them to the supertype cache 1505 // alias index. Since the same load can reach both, we need to merge 1506 // these 2 disparate memories into the same alias class. Since the 1507 // primary supertype array is read-only, there's no chance of confusion 1508 // where we bypass an array load and an array store. 1509 int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); 1510 if (offset == Type::OffsetBot || 1511 (offset >= primary_supers_offset && 1512 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || 1513 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { 1514 offset = in_bytes(Klass::secondary_super_cache_offset()); 1515 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); 1516 } 1517 } 1518 1519 // Flatten all Raw pointers together. 1520 if (tj->base() == Type::RawPtr) 1521 tj = TypeRawPtr::BOTTOM; 1522 1523 if (tj->base() == Type::AnyPtr) 1524 tj = TypePtr::BOTTOM; // An error, which the caller must check for. 1525 1526 // Flatten all to bottom for now 1527 switch( _AliasLevel ) { 1528 case 0: 1529 tj = TypePtr::BOTTOM; 1530 break; 1531 case 1: // Flatten to: oop, static, field or array 1532 switch (tj->base()) { 1533 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; 1534 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; 1535 case Type::AryPtr: // do not distinguish arrays at all 1536 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; 1537 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; 1538 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it 1539 default: ShouldNotReachHere(); 1540 } 1541 break; 1542 case 2: // No collapsing at level 2; keep all splits 1543 case 3: // No collapsing at level 3; keep all splits 1544 break; 1545 default: 1546 Unimplemented(); 1547 } 1548 1549 offset = tj->offset(); 1550 assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); 1551 1552 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || 1553 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || 1554 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || 1555 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || 1556 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || 1557 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || 1558 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , 1559 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); 1560 assert( tj->ptr() != TypePtr::TopPTR && 1561 tj->ptr() != TypePtr::AnyNull && 1562 tj->ptr() != TypePtr::Null, "No imprecise addresses" ); 1563 // assert( tj->ptr() != TypePtr::Constant || 1564 // tj->base() == Type::RawPtr || 1565 // tj->base() == Type::KlassPtr, "No constant oop addresses" ); 1566 1567 return tj; 1568 } 1569 1570 void Compile::AliasType::Init(int i, const TypePtr* at) { 1571 _index = i; 1572 _adr_type = at; 1573 _field = NULL; 1574 _element = NULL; 1575 _is_rewritable = true; // default 1576 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; 1577 if (atoop != NULL && atoop->is_known_instance()) { 1578 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); 1579 _general_index = Compile::current()->get_alias_index(gt); 1580 } else { 1581 _general_index = 0; 1582 } 1583 } 1584 1585 //---------------------------------print_on------------------------------------ 1586 #ifndef PRODUCT 1587 void Compile::AliasType::print_on(outputStream* st) { 1588 if (index() < 10) 1589 st->print("@ <%d> ", index()); 1590 else st->print("@ <%d>", index()); 1591 st->print(is_rewritable() ? " " : " RO"); 1592 int offset = adr_type()->offset(); 1593 if (offset == Type::OffsetBot) 1594 st->print(" +any"); 1595 else st->print(" +%-3d", offset); 1596 st->print(" in "); 1597 adr_type()->dump_on(st); 1598 const TypeOopPtr* tjp = adr_type()->isa_oopptr(); 1599 if (field() != NULL && tjp) { 1600 if (tjp->klass() != field()->holder() || 1601 tjp->offset() != field()->offset_in_bytes()) { 1602 st->print(" != "); 1603 field()->print(); 1604 st->print(" ***"); 1605 } 1606 } 1607 } 1608 1609 void print_alias_types() { 1610 Compile* C = Compile::current(); 1611 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); 1612 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { 1613 C->alias_type(idx)->print_on(tty); 1614 tty->cr(); 1615 } 1616 } 1617 #endif 1618 1619 1620 //----------------------------probe_alias_cache-------------------------------- 1621 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { 1622 intptr_t key = (intptr_t) adr_type; 1623 key ^= key >> logAliasCacheSize; 1624 return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; 1625 } 1626 1627 1628 //-----------------------------grow_alias_types-------------------------------- 1629 void Compile::grow_alias_types() { 1630 const int old_ats = _max_alias_types; // how many before? 1631 const int new_ats = old_ats; // how many more? 1632 const int grow_ats = old_ats+new_ats; // how many now? 1633 _max_alias_types = grow_ats; 1634 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); 1635 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); 1636 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); 1637 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; 1638 } 1639 1640 1641 //--------------------------------find_alias_type------------------------------ 1642 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { 1643 if (_AliasLevel == 0) 1644 return alias_type(AliasIdxBot); 1645 1646 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1647 if (ace->_adr_type == adr_type) { 1648 return alias_type(ace->_index); 1649 } 1650 1651 // Handle special cases. 1652 if (adr_type == NULL) return alias_type(AliasIdxTop); 1653 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); 1654 1655 // Do it the slow way. 1656 const TypePtr* flat = flatten_alias_type(adr_type); 1657 1658 #ifdef ASSERT 1659 assert(flat == flatten_alias_type(flat), "idempotent"); 1660 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); 1661 if (flat->isa_oopptr() && !flat->isa_klassptr()) { 1662 const TypeOopPtr* foop = flat->is_oopptr(); 1663 // Scalarizable allocations have exact klass always. 1664 bool exact = !foop->klass_is_exact() || foop->is_known_instance(); 1665 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); 1666 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); 1667 } 1668 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); 1669 #endif 1670 1671 int idx = AliasIdxTop; 1672 for (int i = 0; i < num_alias_types(); i++) { 1673 if (alias_type(i)->adr_type() == flat) { 1674 idx = i; 1675 break; 1676 } 1677 } 1678 1679 if (idx == AliasIdxTop) { 1680 if (no_create) return NULL; 1681 // Grow the array if necessary. 1682 if (_num_alias_types == _max_alias_types) grow_alias_types(); 1683 // Add a new alias type. 1684 idx = _num_alias_types++; 1685 _alias_types[idx]->Init(idx, flat); 1686 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); 1687 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); 1688 if (flat->isa_instptr()) { 1689 if (flat->offset() == java_lang_Class::klass_offset_in_bytes() 1690 && flat->is_instptr()->klass() == env()->Class_klass()) 1691 alias_type(idx)->set_rewritable(false); 1692 } 1693 if (flat->isa_aryptr()) { 1694 #ifdef ASSERT 1695 const int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE); 1696 // (T_BYTE has the weakest alignment and size restrictions...) 1697 assert(flat->offset() < header_size_min, "array body reference must be OffsetBot"); 1698 #endif 1699 if (flat->offset() == TypePtr::OffsetBot) { 1700 alias_type(idx)->set_element(flat->is_aryptr()->elem()); 1701 } 1702 } 1703 if (flat->isa_klassptr()) { 1704 if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) 1705 alias_type(idx)->set_rewritable(false); 1706 if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) 1707 alias_type(idx)->set_rewritable(false); 1708 if (flat->offset() == in_bytes(Klass::access_flags_offset())) 1709 alias_type(idx)->set_rewritable(false); 1710 if (flat->offset() == in_bytes(Klass::java_mirror_offset())) 1711 alias_type(idx)->set_rewritable(false); 1712 } 1713 // %%% (We would like to finalize JavaThread::threadObj_offset(), 1714 // but the base pointer type is not distinctive enough to identify 1715 // references into JavaThread.) 1716 1717 // Check for final fields. 1718 const TypeInstPtr* tinst = flat->isa_instptr(); 1719 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { 1720 ciField* field; 1721 if (tinst->const_oop() != NULL && 1722 tinst->klass() == ciEnv::current()->Class_klass() && 1723 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { 1724 // static field 1725 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); 1726 field = k->get_field_by_offset(tinst->offset(), true); 1727 } else { 1728 ciInstanceKlass *k = tinst->klass()->as_instance_klass(); 1729 field = k->get_field_by_offset(tinst->offset(), false); 1730 } 1731 assert(field == NULL || 1732 original_field == NULL || 1733 (field->holder() == original_field->holder() && 1734 field->offset() == original_field->offset() && 1735 field->is_static() == original_field->is_static()), "wrong field?"); 1736 // Set field() and is_rewritable() attributes. 1737 if (field != NULL) alias_type(idx)->set_field(field); 1738 } 1739 } 1740 1741 // Fill the cache for next time. 1742 ace->_adr_type = adr_type; 1743 ace->_index = idx; 1744 assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); 1745 1746 // Might as well try to fill the cache for the flattened version, too. 1747 AliasCacheEntry* face = probe_alias_cache(flat); 1748 if (face->_adr_type == NULL) { 1749 face->_adr_type = flat; 1750 face->_index = idx; 1751 assert(alias_type(flat) == alias_type(idx), "flat type must work too"); 1752 } 1753 1754 return alias_type(idx); 1755 } 1756 1757 1758 Compile::AliasType* Compile::alias_type(ciField* field) { 1759 const TypeOopPtr* t; 1760 if (field->is_static()) 1761 t = TypeInstPtr::make(field->holder()->java_mirror()); 1762 else 1763 t = TypeOopPtr::make_from_klass_raw(field->holder()); 1764 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); 1765 assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct"); 1766 return atp; 1767 } 1768 1769 1770 //------------------------------have_alias_type-------------------------------- 1771 bool Compile::have_alias_type(const TypePtr* adr_type) { 1772 AliasCacheEntry* ace = probe_alias_cache(adr_type); 1773 if (ace->_adr_type == adr_type) { 1774 return true; 1775 } 1776 1777 // Handle special cases. 1778 if (adr_type == NULL) return true; 1779 if (adr_type == TypePtr::BOTTOM) return true; 1780 1781 return find_alias_type(adr_type, true, NULL) != NULL; 1782 } 1783 1784 //-----------------------------must_alias-------------------------------------- 1785 // True if all values of the given address type are in the given alias category. 1786 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { 1787 if (alias_idx == AliasIdxBot) return true; // the universal category 1788 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP 1789 if (alias_idx == AliasIdxTop) return false; // the empty category 1790 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins 1791 1792 // the only remaining possible overlap is identity 1793 int adr_idx = get_alias_index(adr_type); 1794 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1795 assert(adr_idx == alias_idx || 1796 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM 1797 && adr_type != TypeOopPtr::BOTTOM), 1798 "should not be testing for overlap with an unsafe pointer"); 1799 return adr_idx == alias_idx; 1800 } 1801 1802 //------------------------------can_alias-------------------------------------- 1803 // True if any values of the given address type are in the given alias category. 1804 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { 1805 if (alias_idx == AliasIdxTop) return false; // the empty category 1806 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP 1807 if (alias_idx == AliasIdxBot) return true; // the universal category 1808 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins 1809 1810 // the only remaining possible overlap is identity 1811 int adr_idx = get_alias_index(adr_type); 1812 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); 1813 return adr_idx == alias_idx; 1814 } 1815 1816 1817 1818 //---------------------------pop_warm_call------------------------------------- 1819 WarmCallInfo* Compile::pop_warm_call() { 1820 WarmCallInfo* wci = _warm_calls; 1821 if (wci != NULL) _warm_calls = wci->remove_from(wci); 1822 return wci; 1823 } 1824 1825 //----------------------------Inline_Warm-------------------------------------- 1826 int Compile::Inline_Warm() { 1827 // If there is room, try to inline some more warm call sites. 1828 // %%% Do a graph index compaction pass when we think we're out of space? 1829 if (!InlineWarmCalls) return 0; 1830 1831 int calls_made_hot = 0; 1832 int room_to_grow = NodeCountInliningCutoff - unique(); 1833 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); 1834 int amount_grown = 0; 1835 WarmCallInfo* call; 1836 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { 1837 int est_size = (int)call->size(); 1838 if (est_size > (room_to_grow - amount_grown)) { 1839 // This one won't fit anyway. Get rid of it. 1840 call->make_cold(); 1841 continue; 1842 } 1843 call->make_hot(); 1844 calls_made_hot++; 1845 amount_grown += est_size; 1846 amount_to_grow -= est_size; 1847 } 1848 1849 if (calls_made_hot > 0) set_major_progress(); 1850 return calls_made_hot; 1851 } 1852 1853 1854 //----------------------------Finish_Warm-------------------------------------- 1855 void Compile::Finish_Warm() { 1856 if (!InlineWarmCalls) return; 1857 if (failing()) return; 1858 if (warm_calls() == NULL) return; 1859 1860 // Clean up loose ends, if we are out of space for inlining. 1861 WarmCallInfo* call; 1862 while ((call = pop_warm_call()) != NULL) { 1863 call->make_cold(); 1864 } 1865 } 1866 1867 //---------------------cleanup_loop_predicates----------------------- 1868 // Remove the opaque nodes that protect the predicates so that all unused 1869 // checks and uncommon_traps will be eliminated from the ideal graph 1870 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { 1871 if (predicate_count()==0) return; 1872 for (int i = predicate_count(); i > 0; i--) { 1873 Node * n = predicate_opaque1_node(i-1); 1874 assert(n->Opcode() == Op_Opaque1, "must be"); 1875 igvn.replace_node(n, n->in(1)); 1876 } 1877 assert(predicate_count()==0, "should be clean!"); 1878 } 1879 1880 // StringOpts and late inlining of string methods 1881 void Compile::inline_string_calls(bool parse_time) { 1882 { 1883 // remove useless nodes to make the usage analysis simpler 1884 ResourceMark rm; 1885 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 1886 } 1887 1888 { 1889 ResourceMark rm; 1890 print_method(PHASE_BEFORE_STRINGOPTS, 3); 1891 PhaseStringOpts pso(initial_gvn(), for_igvn()); 1892 print_method(PHASE_AFTER_STRINGOPTS, 3); 1893 } 1894 1895 // now inline anything that we skipped the first time around 1896 if (!parse_time) { 1897 _late_inlines_pos = _late_inlines.length(); 1898 } 1899 1900 while (_string_late_inlines.length() > 0) { 1901 CallGenerator* cg = _string_late_inlines.pop(); 1902 cg->do_late_inline(); 1903 if (failing()) return; 1904 } 1905 _string_late_inlines.trunc_to(0); 1906 } 1907 1908 // Late inlining of boxing methods 1909 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { 1910 if (_boxing_late_inlines.length() > 0) { 1911 assert(has_boxed_value(), "inconsistent"); 1912 1913 PhaseGVN* gvn = initial_gvn(); 1914 set_inlining_incrementally(true); 1915 1916 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 1917 for_igvn()->clear(); 1918 gvn->replace_with(&igvn); 1919 1920 _late_inlines_pos = _late_inlines.length(); 1921 1922 while (_boxing_late_inlines.length() > 0) { 1923 CallGenerator* cg = _boxing_late_inlines.pop(); 1924 cg->do_late_inline(); 1925 if (failing()) return; 1926 } 1927 _boxing_late_inlines.trunc_to(0); 1928 1929 { 1930 ResourceMark rm; 1931 PhaseRemoveUseless pru(gvn, for_igvn()); 1932 } 1933 1934 igvn = PhaseIterGVN(gvn); 1935 igvn.optimize(); 1936 1937 set_inlining_progress(false); 1938 set_inlining_incrementally(false); 1939 } 1940 } 1941 1942 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { 1943 assert(IncrementalInline, "incremental inlining should be on"); 1944 PhaseGVN* gvn = initial_gvn(); 1945 1946 set_inlining_progress(false); 1947 for_igvn()->clear(); 1948 gvn->replace_with(&igvn); 1949 1950 int i = 0; 1951 1952 for (; i <_late_inlines.length() && !inlining_progress(); i++) { 1953 CallGenerator* cg = _late_inlines.at(i); 1954 _late_inlines_pos = i+1; 1955 cg->do_late_inline(); 1956 if (failing()) return; 1957 } 1958 int j = 0; 1959 for (; i < _late_inlines.length(); i++, j++) { 1960 _late_inlines.at_put(j, _late_inlines.at(i)); 1961 } 1962 _late_inlines.trunc_to(j); 1963 1964 { 1965 ResourceMark rm; 1966 PhaseRemoveUseless pru(gvn, for_igvn()); 1967 } 1968 1969 igvn = PhaseIterGVN(gvn); 1970 } 1971 1972 // Perform incremental inlining until bound on number of live nodes is reached 1973 void Compile::inline_incrementally(PhaseIterGVN& igvn) { 1974 PhaseGVN* gvn = initial_gvn(); 1975 1976 set_inlining_incrementally(true); 1977 set_inlining_progress(true); 1978 uint low_live_nodes = 0; 1979 1980 while(inlining_progress() && _late_inlines.length() > 0) { 1981 1982 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1983 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { 1984 // PhaseIdealLoop is expensive so we only try it once we are 1985 // out of live nodes and we only try it again if the previous 1986 // helped got the number of nodes down significantly 1987 PhaseIdealLoop ideal_loop( igvn, false, true ); 1988 if (failing()) return; 1989 low_live_nodes = live_nodes(); 1990 _major_progress = true; 1991 } 1992 1993 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { 1994 break; 1995 } 1996 } 1997 1998 inline_incrementally_one(igvn); 1999 2000 if (failing()) return; 2001 2002 igvn.optimize(); 2003 2004 if (failing()) return; 2005 } 2006 2007 assert( igvn._worklist.size() == 0, "should be done with igvn" ); 2008 2009 if (_string_late_inlines.length() > 0) { 2010 assert(has_stringbuilder(), "inconsistent"); 2011 for_igvn()->clear(); 2012 initial_gvn()->replace_with(&igvn); 2013 2014 inline_string_calls(false); 2015 2016 if (failing()) return; 2017 2018 { 2019 ResourceMark rm; 2020 PhaseRemoveUseless pru(initial_gvn(), for_igvn()); 2021 } 2022 2023 igvn = PhaseIterGVN(gvn); 2024 2025 igvn.optimize(); 2026 } 2027 2028 set_inlining_incrementally(false); 2029 } 2030 2031 2032 //------------------------------Optimize--------------------------------------- 2033 // Given a graph, optimize it. 2034 void Compile::Optimize() { 2035 TracePhase t1("optimizer", &_t_optimizer, true); 2036 2037 #ifndef PRODUCT 2038 if (env()->break_at_compile()) { 2039 BREAKPOINT; 2040 } 2041 2042 #endif 2043 2044 ResourceMark rm; 2045 int loop_opts_cnt; 2046 2047 print_inlining_reinit(); 2048 2049 NOT_PRODUCT( verify_graph_edges(); ) 2050 2051 print_method(PHASE_AFTER_PARSING); 2052 2053 { 2054 // Iterative Global Value Numbering, including ideal transforms 2055 // Initialize IterGVN with types and values from parse-time GVN 2056 PhaseIterGVN igvn(initial_gvn()); 2057 { 2058 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) 2059 igvn.optimize(); 2060 } 2061 2062 print_method(PHASE_ITER_GVN1, 2); 2063 2064 if (failing()) return; 2065 2066 { 2067 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2068 inline_incrementally(igvn); 2069 } 2070 2071 print_method(PHASE_INCREMENTAL_INLINE, 2); 2072 2073 if (failing()) return; 2074 2075 if (eliminate_boxing()) { 2076 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) 2077 // Inline valueOf() methods now. 2078 inline_boxing_calls(igvn); 2079 2080 if (AlwaysIncrementalInline) { 2081 inline_incrementally(igvn); 2082 } 2083 2084 print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); 2085 2086 if (failing()) return; 2087 } 2088 2089 // Remove the speculative part of types and clean up the graph from 2090 // the extra CastPP nodes whose only purpose is to carry them. Do 2091 // that early so that optimizations are not disrupted by the extra 2092 // CastPP nodes. 2093 remove_speculative_types(igvn); 2094 2095 // No more new expensive nodes will be added to the list from here 2096 // so keep only the actual candidates for optimizations. 2097 cleanup_expensive_nodes(igvn); 2098 2099 // Perform escape analysis 2100 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { 2101 if (has_loops()) { 2102 // Cleanup graph (remove dead nodes). 2103 TracePhase t2("idealLoop", &_t_idealLoop, true); 2104 PhaseIdealLoop ideal_loop( igvn, false, true ); 2105 if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); 2106 if (failing()) return; 2107 } 2108 ConnectionGraph::do_analysis(this, &igvn); 2109 2110 if (failing()) return; 2111 2112 // Optimize out fields loads from scalar replaceable allocations. 2113 igvn.optimize(); 2114 print_method(PHASE_ITER_GVN_AFTER_EA, 2); 2115 2116 if (failing()) return; 2117 2118 if (congraph() != NULL && macro_count() > 0) { 2119 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) 2120 PhaseMacroExpand mexp(igvn); 2121 mexp.eliminate_macro_nodes(); 2122 igvn.set_delay_transform(false); 2123 2124 igvn.optimize(); 2125 print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); 2126 2127 if (failing()) return; 2128 } 2129 } 2130 2131 // Loop transforms on the ideal graph. Range Check Elimination, 2132 // peeling, unrolling, etc. 2133 2134 // Set loop opts counter 2135 loop_opts_cnt = num_loop_opts(); 2136 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { 2137 { 2138 TracePhase t2("idealLoop", &_t_idealLoop, true); 2139 PhaseIdealLoop ideal_loop( igvn, true ); 2140 loop_opts_cnt--; 2141 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); 2142 if (failing()) return; 2143 } 2144 // Loop opts pass if partial peeling occurred in previous pass 2145 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { 2146 TracePhase t3("idealLoop", &_t_idealLoop, true); 2147 PhaseIdealLoop ideal_loop( igvn, false ); 2148 loop_opts_cnt--; 2149 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); 2150 if (failing()) return; 2151 } 2152 // Loop opts pass for loop-unrolling before CCP 2153 if(major_progress() && (loop_opts_cnt > 0)) { 2154 TracePhase t4("idealLoop", &_t_idealLoop, true); 2155 PhaseIdealLoop ideal_loop( igvn, false ); 2156 loop_opts_cnt--; 2157 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); 2158 } 2159 if (!failing()) { 2160 // Verify that last round of loop opts produced a valid graph 2161 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2162 PhaseIdealLoop::verify(igvn); 2163 } 2164 } 2165 if (failing()) return; 2166 2167 // Conditional Constant Propagation; 2168 PhaseCCP ccp( &igvn ); 2169 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); 2170 { 2171 TracePhase t2("ccp", &_t_ccp, true); 2172 ccp.do_transform(); 2173 } 2174 print_method(PHASE_CPP1, 2); 2175 2176 assert( true, "Break here to ccp.dump_old2new_map()"); 2177 2178 // Iterative Global Value Numbering, including ideal transforms 2179 { 2180 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) 2181 igvn = ccp; 2182 igvn.optimize(); 2183 } 2184 2185 print_method(PHASE_ITER_GVN2, 2); 2186 2187 if (failing()) return; 2188 2189 // Loop transforms on the ideal graph. Range Check Elimination, 2190 // peeling, unrolling, etc. 2191 if(loop_opts_cnt > 0) { 2192 debug_only( int cnt = 0; ); 2193 while(major_progress() && (loop_opts_cnt > 0)) { 2194 TracePhase t2("idealLoop", &_t_idealLoop, true); 2195 assert( cnt++ < 40, "infinite cycle in loop optimization" ); 2196 PhaseIdealLoop ideal_loop( igvn, true); 2197 loop_opts_cnt--; 2198 if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); 2199 if (failing()) return; 2200 } 2201 } 2202 2203 { 2204 // Verify that all previous optimizations produced a valid graph 2205 // at least to this point, even if no loop optimizations were done. 2206 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) 2207 PhaseIdealLoop::verify(igvn); 2208 } 2209 2210 { 2211 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) 2212 PhaseMacroExpand mex(igvn); 2213 if (mex.expand_macro_nodes()) { 2214 assert(failing(), "must bail out w/ explicit message"); 2215 return; 2216 } 2217 } 2218 2219 } // (End scope of igvn; run destructor if necessary for asserts.) 2220 2221 process_print_inlining(); 2222 // A method with only infinite loops has no edges entering loops from root 2223 { 2224 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) 2225 if (final_graph_reshaping()) { 2226 assert(failing(), "must bail out w/ explicit message"); 2227 return; 2228 } 2229 } 2230 2231 print_method(PHASE_OPTIMIZE_FINISHED, 2); 2232 } 2233 2234 2235 //------------------------------Code_Gen--------------------------------------- 2236 // Given a graph, generate code for it 2237 void Compile::Code_Gen() { 2238 if (failing()) { 2239 return; 2240 } 2241 2242 // Perform instruction selection. You might think we could reclaim Matcher 2243 // memory PDQ, but actually the Matcher is used in generating spill code. 2244 // Internals of the Matcher (including some VectorSets) must remain live 2245 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage 2246 // set a bit in reclaimed memory. 2247 2248 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2249 // nodes. Mapping is only valid at the root of each matched subtree. 2250 NOT_PRODUCT( verify_graph_edges(); ) 2251 2252 Matcher matcher; 2253 _matcher = &matcher; 2254 { 2255 TracePhase t2("matcher", &_t_matcher, true); 2256 matcher.match(); 2257 } 2258 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine 2259 // nodes. Mapping is only valid at the root of each matched subtree. 2260 NOT_PRODUCT( verify_graph_edges(); ) 2261 2262 // If you have too many nodes, or if matching has failed, bail out 2263 check_node_count(0, "out of nodes matching instructions"); 2264 if (failing()) { 2265 return; 2266 } 2267 2268 // Build a proper-looking CFG 2269 PhaseCFG cfg(node_arena(), root(), matcher); 2270 _cfg = &cfg; 2271 { 2272 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) 2273 bool success = cfg.do_global_code_motion(); 2274 if (!success) { 2275 return; 2276 } 2277 2278 print_method(PHASE_GLOBAL_CODE_MOTION, 2); 2279 NOT_PRODUCT( verify_graph_edges(); ) 2280 debug_only( cfg.verify(); ) 2281 } 2282 2283 PhaseChaitin regalloc(unique(), cfg, matcher); 2284 _regalloc = ®alloc; 2285 { 2286 TracePhase t2("regalloc", &_t_registerAllocation, true); 2287 // Perform register allocation. After Chaitin, use-def chains are 2288 // no longer accurate (at spill code) and so must be ignored. 2289 // Node->LRG->reg mappings are still accurate. 2290 _regalloc->Register_Allocate(); 2291 2292 // Bail out if the allocator builds too many nodes 2293 if (failing()) { 2294 return; 2295 } 2296 } 2297 2298 // Prior to register allocation we kept empty basic blocks in case the 2299 // the allocator needed a place to spill. After register allocation we 2300 // are not adding any new instructions. If any basic block is empty, we 2301 // can now safely remove it. 2302 { 2303 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) 2304 cfg.remove_empty_blocks(); 2305 if (do_freq_based_layout()) { 2306 PhaseBlockLayout layout(cfg); 2307 } else { 2308 cfg.set_loop_alignment(); 2309 } 2310 cfg.fixup_flow(); 2311 } 2312 2313 // Apply peephole optimizations 2314 if( OptoPeephole ) { 2315 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) 2316 PhasePeephole peep( _regalloc, cfg); 2317 peep.do_transform(); 2318 } 2319 2320 // Do late expand if CPU requires this. 2321 if (Matcher::require_postalloc_expand) { 2322 NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true)); 2323 cfg.postalloc_expand(_regalloc); 2324 } 2325 2326 // Convert Nodes to instruction bits in a buffer 2327 { 2328 // %%%% workspace merge brought two timers together for one job 2329 TracePhase t2a("output", &_t_output, true); 2330 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) 2331 Output(); 2332 } 2333 2334 print_method(PHASE_FINAL_CODE); 2335 2336 // He's dead, Jim. 2337 _cfg = (PhaseCFG*)0xdeadbeef; 2338 _regalloc = (PhaseChaitin*)0xdeadbeef; 2339 } 2340 2341 2342 //------------------------------dump_asm--------------------------------------- 2343 // Dump formatted assembly 2344 #ifndef PRODUCT 2345 void Compile::dump_asm(int *pcs, uint pc_limit) { 2346 bool cut_short = false; 2347 tty->print_cr("#"); 2348 tty->print("# "); _tf->dump(); tty->cr(); 2349 tty->print_cr("#"); 2350 2351 // For all blocks 2352 int pc = 0x0; // Program counter 2353 char starts_bundle = ' '; 2354 _regalloc->dump_frame(); 2355 2356 Node *n = NULL; 2357 for (uint i = 0; i < _cfg->number_of_blocks(); i++) { 2358 if (VMThread::should_terminate()) { 2359 cut_short = true; 2360 break; 2361 } 2362 Block* block = _cfg->get_block(i); 2363 if (block->is_connector() && !Verbose) { 2364 continue; 2365 } 2366 n = block->head(); 2367 if (pcs && n->_idx < pc_limit) { 2368 tty->print("%3.3x ", pcs[n->_idx]); 2369 } else { 2370 tty->print(" "); 2371 } 2372 block->dump_head(_cfg); 2373 if (block->is_connector()) { 2374 tty->print_cr(" # Empty connector block"); 2375 } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { 2376 tty->print_cr(" # Block is sole successor of call"); 2377 } 2378 2379 // For all instructions 2380 Node *delay = NULL; 2381 for (uint j = 0; j < block->number_of_nodes(); j++) { 2382 if (VMThread::should_terminate()) { 2383 cut_short = true; 2384 break; 2385 } 2386 n = block->get_node(j); 2387 if (valid_bundle_info(n)) { 2388 Bundle* bundle = node_bundling(n); 2389 if (bundle->used_in_unconditional_delay()) { 2390 delay = n; 2391 continue; 2392 } 2393 if (bundle->starts_bundle()) { 2394 starts_bundle = '+'; 2395 } 2396 } 2397 2398 if (WizardMode) { 2399 n->dump(); 2400 } 2401 2402 if( !n->is_Region() && // Dont print in the Assembly 2403 !n->is_Phi() && // a few noisely useless nodes 2404 !n->is_Proj() && 2405 !n->is_MachTemp() && 2406 !n->is_SafePointScalarObject() && 2407 !n->is_Catch() && // Would be nice to print exception table targets 2408 !n->is_MergeMem() && // Not very interesting 2409 !n->is_top() && // Debug info table constants 2410 !(n->is_Con() && !n->is_Mach())// Debug info table constants 2411 ) { 2412 if (pcs && n->_idx < pc_limit) 2413 tty->print("%3.3x", pcs[n->_idx]); 2414 else 2415 tty->print(" "); 2416 tty->print(" %c ", starts_bundle); 2417 starts_bundle = ' '; 2418 tty->print("\t"); 2419 n->format(_regalloc, tty); 2420 tty->cr(); 2421 } 2422 2423 // If we have an instruction with a delay slot, and have seen a delay, 2424 // then back up and print it 2425 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { 2426 assert(delay != NULL, "no unconditional delay instruction"); 2427 if (WizardMode) delay->dump(); 2428 2429 if (node_bundling(delay)->starts_bundle()) 2430 starts_bundle = '+'; 2431 if (pcs && n->_idx < pc_limit) 2432 tty->print("%3.3x", pcs[n->_idx]); 2433 else 2434 tty->print(" "); 2435 tty->print(" %c ", starts_bundle); 2436 starts_bundle = ' '; 2437 tty->print("\t"); 2438 delay->format(_regalloc, tty); 2439 tty->cr(); 2440 delay = NULL; 2441 } 2442 2443 // Dump the exception table as well 2444 if( n->is_Catch() && (Verbose || WizardMode) ) { 2445 // Print the exception table for this offset 2446 _handler_table.print_subtable_for(pc); 2447 } 2448 } 2449 2450 if (pcs && n->_idx < pc_limit) 2451 tty->print_cr("%3.3x", pcs[n->_idx]); 2452 else 2453 tty->cr(); 2454 2455 assert(cut_short || delay == NULL, "no unconditional delay branch"); 2456 2457 } // End of per-block dump 2458 tty->cr(); 2459 2460 if (cut_short) tty->print_cr("*** disassembly is cut short ***"); 2461 } 2462 #endif 2463 2464 //------------------------------Final_Reshape_Counts--------------------------- 2465 // This class defines counters to help identify when a method 2466 // may/must be executed using hardware with only 24-bit precision. 2467 struct Final_Reshape_Counts : public StackObj { 2468 int _call_count; // count non-inlined 'common' calls 2469 int _float_count; // count float ops requiring 24-bit precision 2470 int _double_count; // count double ops requiring more precision 2471 int _java_call_count; // count non-inlined 'java' calls 2472 int _inner_loop_count; // count loops which need alignment 2473 VectorSet _visited; // Visitation flags 2474 Node_List _tests; // Set of IfNodes & PCTableNodes 2475 2476 Final_Reshape_Counts() : 2477 _call_count(0), _float_count(0), _double_count(0), 2478 _java_call_count(0), _inner_loop_count(0), 2479 _visited( Thread::current()->resource_area() ) { } 2480 2481 void inc_call_count () { _call_count ++; } 2482 void inc_float_count () { _float_count ++; } 2483 void inc_double_count() { _double_count++; } 2484 void inc_java_call_count() { _java_call_count++; } 2485 void inc_inner_loop_count() { _inner_loop_count++; } 2486 2487 int get_call_count () const { return _call_count ; } 2488 int get_float_count () const { return _float_count ; } 2489 int get_double_count() const { return _double_count; } 2490 int get_java_call_count() const { return _java_call_count; } 2491 int get_inner_loop_count() const { return _inner_loop_count; } 2492 }; 2493 2494 #ifdef ASSERT 2495 static bool oop_offset_is_sane(const TypeInstPtr* tp) { 2496 ciInstanceKlass *k = tp->klass()->as_instance_klass(); 2497 // Make sure the offset goes inside the instance layout. 2498 return k->contains_field_offset(tp->offset()); 2499 // Note that OffsetBot and OffsetTop are very negative. 2500 } 2501 #endif 2502 2503 // Eliminate trivially redundant StoreCMs and accumulate their 2504 // precedence edges. 2505 void Compile::eliminate_redundant_card_marks(Node* n) { 2506 assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); 2507 if (n->in(MemNode::Address)->outcnt() > 1) { 2508 // There are multiple users of the same address so it might be 2509 // possible to eliminate some of the StoreCMs 2510 Node* mem = n->in(MemNode::Memory); 2511 Node* adr = n->in(MemNode::Address); 2512 Node* val = n->in(MemNode::ValueIn); 2513 Node* prev = n; 2514 bool done = false; 2515 // Walk the chain of StoreCMs eliminating ones that match. As 2516 // long as it's a chain of single users then the optimization is 2517 // safe. Eliminating partially redundant StoreCMs would require 2518 // cloning copies down the other paths. 2519 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { 2520 if (adr == mem->in(MemNode::Address) && 2521 val == mem->in(MemNode::ValueIn)) { 2522 // redundant StoreCM 2523 if (mem->req() > MemNode::OopStore) { 2524 // Hasn't been processed by this code yet. 2525 n->add_prec(mem->in(MemNode::OopStore)); 2526 } else { 2527 // Already converted to precedence edge 2528 for (uint i = mem->req(); i < mem->len(); i++) { 2529 // Accumulate any precedence edges 2530 if (mem->in(i) != NULL) { 2531 n->add_prec(mem->in(i)); 2532 } 2533 } 2534 // Everything above this point has been processed. 2535 done = true; 2536 } 2537 // Eliminate the previous StoreCM 2538 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); 2539 assert(mem->outcnt() == 0, "should be dead"); 2540 mem->disconnect_inputs(NULL, this); 2541 } else { 2542 prev = mem; 2543 } 2544 mem = prev->in(MemNode::Memory); 2545 } 2546 } 2547 } 2548 2549 //------------------------------final_graph_reshaping_impl---------------------- 2550 // Implement items 1-5 from final_graph_reshaping below. 2551 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { 2552 2553 if ( n->outcnt() == 0 ) return; // dead node 2554 uint nop = n->Opcode(); 2555 2556 // Check for 2-input instruction with "last use" on right input. 2557 // Swap to left input. Implements item (2). 2558 if( n->req() == 3 && // two-input instruction 2559 n->in(1)->outcnt() > 1 && // left use is NOT a last use 2560 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop 2561 n->in(2)->outcnt() == 1 &&// right use IS a last use 2562 !n->in(2)->is_Con() ) { // right use is not a constant 2563 // Check for commutative opcode 2564 switch( nop ) { 2565 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: 2566 case Op_MaxI: case Op_MinI: 2567 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: 2568 case Op_AndL: case Op_XorL: case Op_OrL: 2569 case Op_AndI: case Op_XorI: case Op_OrI: { 2570 // Move "last use" input to left by swapping inputs 2571 n->swap_edges(1, 2); 2572 break; 2573 } 2574 default: 2575 break; 2576 } 2577 } 2578 2579 #ifdef ASSERT 2580 if( n->is_Mem() ) { 2581 int alias_idx = get_alias_index(n->as_Mem()->adr_type()); 2582 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || 2583 // oop will be recorded in oop map if load crosses safepoint 2584 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || 2585 LoadNode::is_immutable_value(n->in(MemNode::Address))), 2586 "raw memory operations should have control edge"); 2587 } 2588 #endif 2589 // Count FPU ops and common calls, implements item (3) 2590 switch( nop ) { 2591 // Count all float operations that may use FPU 2592 case Op_AddF: 2593 case Op_SubF: 2594 case Op_MulF: 2595 case Op_DivF: 2596 case Op_NegF: 2597 case Op_ModF: 2598 case Op_ConvI2F: 2599 case Op_ConF: 2600 case Op_CmpF: 2601 case Op_CmpF3: 2602 // case Op_ConvL2F: // longs are split into 32-bit halves 2603 frc.inc_float_count(); 2604 break; 2605 2606 case Op_ConvF2D: 2607 case Op_ConvD2F: 2608 frc.inc_float_count(); 2609 frc.inc_double_count(); 2610 break; 2611 2612 // Count all double operations that may use FPU 2613 case Op_AddD: 2614 case Op_SubD: 2615 case Op_MulD: 2616 case Op_DivD: 2617 case Op_NegD: 2618 case Op_ModD: 2619 case Op_ConvI2D: 2620 case Op_ConvD2I: 2621 // case Op_ConvL2D: // handled by leaf call 2622 // case Op_ConvD2L: // handled by leaf call 2623 case Op_ConD: 2624 case Op_CmpD: 2625 case Op_CmpD3: 2626 frc.inc_double_count(); 2627 break; 2628 case Op_Opaque1: // Remove Opaque Nodes before matching 2629 case Op_Opaque2: // Remove Opaque Nodes before matching 2630 case Op_Opaque3: 2631 n->subsume_by(n->in(1), this); 2632 break; 2633 case Op_CallStaticJava: 2634 case Op_CallJava: 2635 case Op_CallDynamicJava: 2636 frc.inc_java_call_count(); // Count java call site; 2637 case Op_CallRuntime: 2638 case Op_CallLeaf: 2639 case Op_CallLeafNoFP: { 2640 assert( n->is_Call(), "" ); 2641 CallNode *call = n->as_Call(); 2642 // Count call sites where the FP mode bit would have to be flipped. 2643 // Do not count uncommon runtime calls: 2644 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, 2645 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... 2646 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { 2647 frc.inc_call_count(); // Count the call site 2648 } else { // See if uncommon argument is shared 2649 Node *n = call->in(TypeFunc::Parms); 2650 int nop = n->Opcode(); 2651 // Clone shared simple arguments to uncommon calls, item (1). 2652 if( n->outcnt() > 1 && 2653 !n->is_Proj() && 2654 nop != Op_CreateEx && 2655 nop != Op_CheckCastPP && 2656 nop != Op_DecodeN && 2657 nop != Op_DecodeNKlass && 2658 !n->is_Mem() ) { 2659 Node *x = n->clone(); 2660 call->set_req( TypeFunc::Parms, x ); 2661 } 2662 } 2663 break; 2664 } 2665 2666 case Op_StoreD: 2667 case Op_LoadD: 2668 case Op_LoadD_unaligned: 2669 frc.inc_double_count(); 2670 goto handle_mem; 2671 case Op_StoreF: 2672 case Op_LoadF: 2673 frc.inc_float_count(); 2674 goto handle_mem; 2675 2676 case Op_StoreCM: 2677 { 2678 // Convert OopStore dependence into precedence edge 2679 Node* prec = n->in(MemNode::OopStore); 2680 n->del_req(MemNode::OopStore); 2681 n->add_prec(prec); 2682 eliminate_redundant_card_marks(n); 2683 } 2684 2685 // fall through 2686 2687 case Op_StoreB: 2688 case Op_StoreC: 2689 case Op_StorePConditional: 2690 case Op_StoreI: 2691 case Op_StoreL: 2692 case Op_StoreIConditional: 2693 case Op_StoreLConditional: 2694 case Op_CompareAndSwapI: 2695 case Op_CompareAndSwapL: 2696 case Op_CompareAndSwapP: 2697 case Op_CompareAndSwapN: 2698 case Op_GetAndAddI: 2699 case Op_GetAndAddL: 2700 case Op_GetAndSetI: 2701 case Op_GetAndSetL: 2702 case Op_GetAndSetP: 2703 case Op_GetAndSetN: 2704 case Op_StoreP: 2705 case Op_StoreN: 2706 case Op_StoreNKlass: 2707 case Op_LoadB: 2708 case Op_LoadUB: 2709 case Op_LoadUS: 2710 case Op_LoadI: 2711 case Op_LoadKlass: 2712 case Op_LoadNKlass: 2713 case Op_LoadL: 2714 case Op_LoadL_unaligned: 2715 case Op_LoadPLocked: 2716 case Op_LoadP: 2717 case Op_LoadN: 2718 case Op_LoadRange: 2719 case Op_LoadS: { 2720 handle_mem: 2721 #ifdef ASSERT 2722 if( VerifyOptoOopOffsets ) { 2723 assert( n->is_Mem(), "" ); 2724 MemNode *mem = (MemNode*)n; 2725 // Check to see if address types have grounded out somehow. 2726 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); 2727 assert( !tp || oop_offset_is_sane(tp), "" ); 2728 } 2729 #endif 2730 break; 2731 } 2732 2733 case Op_AddP: { // Assert sane base pointers 2734 Node *addp = n->in(AddPNode::Address); 2735 assert( !addp->is_AddP() || 2736 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation 2737 addp->in(AddPNode::Base) == n->in(AddPNode::Base), 2738 "Base pointers must match" ); 2739 #ifdef _LP64 2740 if ((UseCompressedOops || UseCompressedClassPointers) && 2741 addp->Opcode() == Op_ConP && 2742 addp == n->in(AddPNode::Base) && 2743 n->in(AddPNode::Offset)->is_Con()) { 2744 // Use addressing with narrow klass to load with offset on x86. 2745 // On sparc loading 32-bits constant and decoding it have less 2746 // instructions (4) then load 64-bits constant (7). 2747 // Do this transformation here since IGVN will convert ConN back to ConP. 2748 const Type* t = addp->bottom_type(); 2749 if (t->isa_oopptr() || t->isa_klassptr()) { 2750 Node* nn = NULL; 2751 2752 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; 2753 2754 // Look for existing ConN node of the same exact type. 2755 Node* r = root(); 2756 uint cnt = r->outcnt(); 2757 for (uint i = 0; i < cnt; i++) { 2758 Node* m = r->raw_out(i); 2759 if (m!= NULL && m->Opcode() == op && 2760 m->bottom_type()->make_ptr() == t) { 2761 nn = m; 2762 break; 2763 } 2764 } 2765 if (nn != NULL) { 2766 // Decode a narrow oop to match address 2767 // [R12 + narrow_oop_reg<<3 + offset] 2768 if (t->isa_oopptr()) { 2769 nn = new DecodeNNode(nn, t); 2770 } else { 2771 nn = new DecodeNKlassNode(nn, t); 2772 } 2773 n->set_req(AddPNode::Base, nn); 2774 n->set_req(AddPNode::Address, nn); 2775 if (addp->outcnt() == 0) { 2776 addp->disconnect_inputs(NULL, this); 2777 } 2778 } 2779 } 2780 } 2781 #endif 2782 break; 2783 } 2784 2785 #ifdef _LP64 2786 case Op_CastPP: 2787 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { 2788 Node* in1 = n->in(1); 2789 const Type* t = n->bottom_type(); 2790 Node* new_in1 = in1->clone(); 2791 new_in1->as_DecodeN()->set_type(t); 2792 2793 if (!Matcher::narrow_oop_use_complex_address()) { 2794 // 2795 // x86, ARM and friends can handle 2 adds in addressing mode 2796 // and Matcher can fold a DecodeN node into address by using 2797 // a narrow oop directly and do implicit NULL check in address: 2798 // 2799 // [R12 + narrow_oop_reg<<3 + offset] 2800 // NullCheck narrow_oop_reg 2801 // 2802 // On other platforms (Sparc) we have to keep new DecodeN node and 2803 // use it to do implicit NULL check in address: 2804 // 2805 // decode_not_null narrow_oop_reg, base_reg 2806 // [base_reg + offset] 2807 // NullCheck base_reg 2808 // 2809 // Pin the new DecodeN node to non-null path on these platform (Sparc) 2810 // to keep the information to which NULL check the new DecodeN node 2811 // corresponds to use it as value in implicit_null_check(). 2812 // 2813 new_in1->set_req(0, n->in(0)); 2814 } 2815 2816 n->subsume_by(new_in1, this); 2817 if (in1->outcnt() == 0) { 2818 in1->disconnect_inputs(NULL, this); 2819 } 2820 } 2821 break; 2822 2823 case Op_CmpP: 2824 // Do this transformation here to preserve CmpPNode::sub() and 2825 // other TypePtr related Ideal optimizations (for example, ptr nullness). 2826 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { 2827 Node* in1 = n->in(1); 2828 Node* in2 = n->in(2); 2829 if (!in1->is_DecodeNarrowPtr()) { 2830 in2 = in1; 2831 in1 = n->in(2); 2832 } 2833 assert(in1->is_DecodeNarrowPtr(), "sanity"); 2834 2835 Node* new_in2 = NULL; 2836 if (in2->is_DecodeNarrowPtr()) { 2837 assert(in2->Opcode() == in1->Opcode(), "must be same node type"); 2838 new_in2 = in2->in(1); 2839 } else if (in2->Opcode() == Op_ConP) { 2840 const Type* t = in2->bottom_type(); 2841 if (t == TypePtr::NULL_PTR) { 2842 assert(in1->is_DecodeN(), "compare klass to null?"); 2843 // Don't convert CmpP null check into CmpN if compressed 2844 // oops implicit null check is not generated. 2845 // This will allow to generate normal oop implicit null check. 2846 if (Matcher::gen_narrow_oop_implicit_null_checks()) 2847 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); 2848 // 2849 // This transformation together with CastPP transformation above 2850 // will generated code for implicit NULL checks for compressed oops. 2851 // 2852 // The original code after Optimize() 2853 // 2854 // LoadN memory, narrow_oop_reg 2855 // decode narrow_oop_reg, base_reg 2856 // CmpP base_reg, NULL 2857 // CastPP base_reg // NotNull 2858 // Load [base_reg + offset], val_reg 2859 // 2860 // after these transformations will be 2861 // 2862 // LoadN memory, narrow_oop_reg 2863 // CmpN narrow_oop_reg, NULL 2864 // decode_not_null narrow_oop_reg, base_reg 2865 // Load [base_reg + offset], val_reg 2866 // 2867 // and the uncommon path (== NULL) will use narrow_oop_reg directly 2868 // since narrow oops can be used in debug info now (see the code in 2869 // final_graph_reshaping_walk()). 2870 // 2871 // At the end the code will be matched to 2872 // on x86: 2873 // 2874 // Load_narrow_oop memory, narrow_oop_reg 2875 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg 2876 // NullCheck narrow_oop_reg 2877 // 2878 // and on sparc: 2879 // 2880 // Load_narrow_oop memory, narrow_oop_reg 2881 // decode_not_null narrow_oop_reg, base_reg 2882 // Load [base_reg + offset], val_reg 2883 // NullCheck base_reg 2884 // 2885 } else if (t->isa_oopptr()) { 2886 new_in2 = ConNode::make(this, t->make_narrowoop()); 2887 } else if (t->isa_klassptr()) { 2888 new_in2 = ConNode::make(this, t->make_narrowklass()); 2889 } 2890 } 2891 if (new_in2 != NULL) { 2892 Node* cmpN = new CmpNNode(in1->in(1), new_in2); 2893 n->subsume_by(cmpN, this); 2894 if (in1->outcnt() == 0) { 2895 in1->disconnect_inputs(NULL, this); 2896 } 2897 if (in2->outcnt() == 0) { 2898 in2->disconnect_inputs(NULL, this); 2899 } 2900 } 2901 } 2902 break; 2903 2904 case Op_DecodeN: 2905 case Op_DecodeNKlass: 2906 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); 2907 // DecodeN could be pinned when it can't be fold into 2908 // an address expression, see the code for Op_CastPP above. 2909 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); 2910 break; 2911 2912 case Op_EncodeP: 2913 case Op_EncodePKlass: { 2914 Node* in1 = n->in(1); 2915 if (in1->is_DecodeNarrowPtr()) { 2916 n->subsume_by(in1->in(1), this); 2917 } else if (in1->Opcode() == Op_ConP) { 2918 const Type* t = in1->bottom_type(); 2919 if (t == TypePtr::NULL_PTR) { 2920 assert(t->isa_oopptr(), "null klass?"); 2921 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); 2922 } else if (t->isa_oopptr()) { 2923 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); 2924 } else if (t->isa_klassptr()) { 2925 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); 2926 } 2927 } 2928 if (in1->outcnt() == 0) { 2929 in1->disconnect_inputs(NULL, this); 2930 } 2931 break; 2932 } 2933 2934 case Op_Proj: { 2935 if (OptimizeStringConcat) { 2936 ProjNode* p = n->as_Proj(); 2937 if (p->_is_io_use) { 2938 // Separate projections were used for the exception path which 2939 // are normally removed by a late inline. If it wasn't inlined 2940 // then they will hang around and should just be replaced with 2941 // the original one. 2942 Node* proj = NULL; 2943 // Replace with just one 2944 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { 2945 Node *use = i.get(); 2946 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { 2947 proj = use; 2948 break; 2949 } 2950 } 2951 assert(proj != NULL, "must be found"); 2952 p->subsume_by(proj, this); 2953 } 2954 } 2955 break; 2956 } 2957 2958 case Op_Phi: 2959 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { 2960 // The EncodeP optimization may create Phi with the same edges 2961 // for all paths. It is not handled well by Register Allocator. 2962 Node* unique_in = n->in(1); 2963 assert(unique_in != NULL, ""); 2964 uint cnt = n->req(); 2965 for (uint i = 2; i < cnt; i++) { 2966 Node* m = n->in(i); 2967 assert(m != NULL, ""); 2968 if (unique_in != m) 2969 unique_in = NULL; 2970 } 2971 if (unique_in != NULL) { 2972 n->subsume_by(unique_in, this); 2973 } 2974 } 2975 break; 2976 2977 #endif 2978 2979 case Op_ModI: 2980 if (UseDivMod) { 2981 // Check if a%b and a/b both exist 2982 Node* d = n->find_similar(Op_DivI); 2983 if (d) { 2984 // Replace them with a fused divmod if supported 2985 if (Matcher::has_match_rule(Op_DivModI)) { 2986 DivModINode* divmod = DivModINode::make(this, n); 2987 d->subsume_by(divmod->div_proj(), this); 2988 n->subsume_by(divmod->mod_proj(), this); 2989 } else { 2990 // replace a%b with a-((a/b)*b) 2991 Node* mult = new MulINode(d, d->in(2)); 2992 Node* sub = new SubINode(d->in(1), mult); 2993 n->subsume_by(sub, this); 2994 } 2995 } 2996 } 2997 break; 2998 2999 case Op_ModL: 3000 if (UseDivMod) { 3001 // Check if a%b and a/b both exist 3002 Node* d = n->find_similar(Op_DivL); 3003 if (d) { 3004 // Replace them with a fused divmod if supported 3005 if (Matcher::has_match_rule(Op_DivModL)) { 3006 DivModLNode* divmod = DivModLNode::make(this, n); 3007 d->subsume_by(divmod->div_proj(), this); 3008 n->subsume_by(divmod->mod_proj(), this); 3009 } else { 3010 // replace a%b with a-((a/b)*b) 3011 Node* mult = new MulLNode(d, d->in(2)); 3012 Node* sub = new SubLNode(d->in(1), mult); 3013 n->subsume_by(sub, this); 3014 } 3015 } 3016 } 3017 break; 3018 3019 case Op_LoadVector: 3020 case Op_StoreVector: 3021 break; 3022 3023 case Op_PackB: 3024 case Op_PackS: 3025 case Op_PackI: 3026 case Op_PackF: 3027 case Op_PackL: 3028 case Op_PackD: 3029 if (n->req()-1 > 2) { 3030 // Replace many operand PackNodes with a binary tree for matching 3031 PackNode* p = (PackNode*) n; 3032 Node* btp = p->binary_tree_pack(this, 1, n->req()); 3033 n->subsume_by(btp, this); 3034 } 3035 break; 3036 case Op_Loop: 3037 case Op_CountedLoop: 3038 if (n->as_Loop()->is_inner_loop()) { 3039 frc.inc_inner_loop_count(); 3040 } 3041 break; 3042 case Op_LShiftI: 3043 case Op_RShiftI: 3044 case Op_URShiftI: 3045 case Op_LShiftL: 3046 case Op_RShiftL: 3047 case Op_URShiftL: 3048 if (Matcher::need_masked_shift_count) { 3049 // The cpu's shift instructions don't restrict the count to the 3050 // lower 5/6 bits. We need to do the masking ourselves. 3051 Node* in2 = n->in(2); 3052 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); 3053 const TypeInt* t = in2->find_int_type(); 3054 if (t != NULL && t->is_con()) { 3055 juint shift = t->get_con(); 3056 if (shift > mask) { // Unsigned cmp 3057 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); 3058 } 3059 } else { 3060 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { 3061 Node* shift = new AndINode(in2, ConNode::make(this, TypeInt::make(mask))); 3062 n->set_req(2, shift); 3063 } 3064 } 3065 if (in2->outcnt() == 0) { // Remove dead node 3066 in2->disconnect_inputs(NULL, this); 3067 } 3068 } 3069 break; 3070 case Op_MemBarStoreStore: 3071 case Op_MemBarRelease: 3072 // Break the link with AllocateNode: it is no longer useful and 3073 // confuses register allocation. 3074 if (n->req() > MemBarNode::Precedent) { 3075 n->set_req(MemBarNode::Precedent, top()); 3076 } 3077 break; 3078 default: 3079 assert( !n->is_Call(), "" ); 3080 assert( !n->is_Mem(), "" ); 3081 break; 3082 } 3083 3084 // Collect CFG split points 3085 if (n->is_MultiBranch()) 3086 frc._tests.push(n); 3087 } 3088 3089 //------------------------------final_graph_reshaping_walk--------------------- 3090 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), 3091 // requires that the walk visits a node's inputs before visiting the node. 3092 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { 3093 ResourceArea *area = Thread::current()->resource_area(); 3094 Unique_Node_List sfpt(area); 3095 3096 frc._visited.set(root->_idx); // first, mark node as visited 3097 uint cnt = root->req(); 3098 Node *n = root; 3099 uint i = 0; 3100 while (true) { 3101 if (i < cnt) { 3102 // Place all non-visited non-null inputs onto stack 3103 Node* m = n->in(i); 3104 ++i; 3105 if (m != NULL && !frc._visited.test_set(m->_idx)) { 3106 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) { 3107 // compute worst case interpreter size in case of a deoptimization 3108 update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size()); 3109 3110 sfpt.push(m); 3111 } 3112 cnt = m->req(); 3113 nstack.push(n, i); // put on stack parent and next input's index 3114 n = m; 3115 i = 0; 3116 } 3117 } else { 3118 // Now do post-visit work 3119 final_graph_reshaping_impl( n, frc ); 3120 if (nstack.is_empty()) 3121 break; // finished 3122 n = nstack.node(); // Get node from stack 3123 cnt = n->req(); 3124 i = nstack.index(); 3125 nstack.pop(); // Shift to the next node on stack 3126 } 3127 } 3128 3129 // Skip next transformation if compressed oops are not used. 3130 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || 3131 (!UseCompressedOops && !UseCompressedClassPointers)) 3132 return; 3133 3134 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. 3135 // It could be done for an uncommon traps or any safepoints/calls 3136 // if the DecodeN/DecodeNKlass node is referenced only in a debug info. 3137 while (sfpt.size() > 0) { 3138 n = sfpt.pop(); 3139 JVMState *jvms = n->as_SafePoint()->jvms(); 3140 assert(jvms != NULL, "sanity"); 3141 int start = jvms->debug_start(); 3142 int end = n->req(); 3143 bool is_uncommon = (n->is_CallStaticJava() && 3144 n->as_CallStaticJava()->uncommon_trap_request() != 0); 3145 for (int j = start; j < end; j++) { 3146 Node* in = n->in(j); 3147 if (in->is_DecodeNarrowPtr()) { 3148 bool safe_to_skip = true; 3149 if (!is_uncommon ) { 3150 // Is it safe to skip? 3151 for (uint i = 0; i < in->outcnt(); i++) { 3152 Node* u = in->raw_out(i); 3153 if (!u->is_SafePoint() || 3154 u->is_Call() && u->as_Call()->has_non_debug_use(n)) { 3155 safe_to_skip = false; 3156 } 3157 } 3158 } 3159 if (safe_to_skip) { 3160 n->set_req(j, in->in(1)); 3161 } 3162 if (in->outcnt() == 0) { 3163 in->disconnect_inputs(NULL, this); 3164 } 3165 } 3166 } 3167 } 3168 } 3169 3170 //------------------------------final_graph_reshaping-------------------------- 3171 // Final Graph Reshaping. 3172 // 3173 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late 3174 // and not commoned up and forced early. Must come after regular 3175 // optimizations to avoid GVN undoing the cloning. Clone constant 3176 // inputs to Loop Phis; these will be split by the allocator anyways. 3177 // Remove Opaque nodes. 3178 // (2) Move last-uses by commutative operations to the left input to encourage 3179 // Intel update-in-place two-address operations and better register usage 3180 // on RISCs. Must come after regular optimizations to avoid GVN Ideal 3181 // calls canonicalizing them back. 3182 // (3) Count the number of double-precision FP ops, single-precision FP ops 3183 // and call sites. On Intel, we can get correct rounding either by 3184 // forcing singles to memory (requires extra stores and loads after each 3185 // FP bytecode) or we can set a rounding mode bit (requires setting and 3186 // clearing the mode bit around call sites). The mode bit is only used 3187 // if the relative frequency of single FP ops to calls is low enough. 3188 // This is a key transform for SPEC mpeg_audio. 3189 // (4) Detect infinite loops; blobs of code reachable from above but not 3190 // below. Several of the Code_Gen algorithms fail on such code shapes, 3191 // so we simply bail out. Happens a lot in ZKM.jar, but also happens 3192 // from time to time in other codes (such as -Xcomp finalizer loops, etc). 3193 // Detection is by looking for IfNodes where only 1 projection is 3194 // reachable from below or CatchNodes missing some targets. 3195 // (5) Assert for insane oop offsets in debug mode. 3196 3197 bool Compile::final_graph_reshaping() { 3198 // an infinite loop may have been eliminated by the optimizer, 3199 // in which case the graph will be empty. 3200 if (root()->req() == 1) { 3201 record_method_not_compilable("trivial infinite loop"); 3202 return true; 3203 } 3204 3205 // Expensive nodes have their control input set to prevent the GVN 3206 // from freely commoning them. There's no GVN beyond this point so 3207 // no need to keep the control input. We want the expensive nodes to 3208 // be freely moved to the least frequent code path by gcm. 3209 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); 3210 for (int i = 0; i < expensive_count(); i++) { 3211 _expensive_nodes->at(i)->set_req(0, NULL); 3212 } 3213 3214 Final_Reshape_Counts frc; 3215 3216 // Visit everybody reachable! 3217 // Allocate stack of size C->unique()/2 to avoid frequent realloc 3218 Node_Stack nstack(unique() >> 1); 3219 final_graph_reshaping_walk(nstack, root(), frc); 3220 3221 // Check for unreachable (from below) code (i.e., infinite loops). 3222 for( uint i = 0; i < frc._tests.size(); i++ ) { 3223 MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); 3224 // Get number of CFG targets. 3225 // Note that PCTables include exception targets after calls. 3226 uint required_outcnt = n->required_outcnt(); 3227 if (n->outcnt() != required_outcnt) { 3228 // Check for a few special cases. Rethrow Nodes never take the 3229 // 'fall-thru' path, so expected kids is 1 less. 3230 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { 3231 if (n->in(0)->in(0)->is_Call()) { 3232 CallNode *call = n->in(0)->in(0)->as_Call(); 3233 if (call->entry_point() == OptoRuntime::rethrow_stub()) { 3234 required_outcnt--; // Rethrow always has 1 less kid 3235 } else if (call->req() > TypeFunc::Parms && 3236 call->is_CallDynamicJava()) { 3237 // Check for null receiver. In such case, the optimizer has 3238 // detected that the virtual call will always result in a null 3239 // pointer exception. The fall-through projection of this CatchNode 3240 // will not be populated. 3241 Node *arg0 = call->in(TypeFunc::Parms); 3242 if (arg0->is_Type() && 3243 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { 3244 required_outcnt--; 3245 } 3246 } else if (call->entry_point() == OptoRuntime::new_array_Java() && 3247 call->req() > TypeFunc::Parms+1 && 3248 call->is_CallStaticJava()) { 3249 // Check for negative array length. In such case, the optimizer has 3250 // detected that the allocation attempt will always result in an 3251 // exception. There is no fall-through projection of this CatchNode . 3252 Node *arg1 = call->in(TypeFunc::Parms+1); 3253 if (arg1->is_Type() && 3254 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { 3255 required_outcnt--; 3256 } 3257 } 3258 } 3259 } 3260 // Recheck with a better notion of 'required_outcnt' 3261 if (n->outcnt() != required_outcnt) { 3262 record_method_not_compilable("malformed control flow"); 3263 return true; // Not all targets reachable! 3264 } 3265 } 3266 // Check that I actually visited all kids. Unreached kids 3267 // must be infinite loops. 3268 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) 3269 if (!frc._visited.test(n->fast_out(j)->_idx)) { 3270 record_method_not_compilable("infinite loop"); 3271 return true; // Found unvisited kid; must be unreach 3272 } 3273 } 3274 3275 // If original bytecodes contained a mixture of floats and doubles 3276 // check if the optimizer has made it homogenous, item (3). 3277 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && 3278 frc.get_float_count() > 32 && 3279 frc.get_double_count() == 0 && 3280 (10 * frc.get_call_count() < frc.get_float_count()) ) { 3281 set_24_bit_selection_and_mode( false, true ); 3282 } 3283 3284 set_java_calls(frc.get_java_call_count()); 3285 set_inner_loops(frc.get_inner_loop_count()); 3286 3287 // No infinite loops, no reason to bail out. 3288 return false; 3289 } 3290 3291 //-----------------------------too_many_traps---------------------------------- 3292 // Report if there are too many traps at the current method and bci. 3293 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. 3294 bool Compile::too_many_traps(ciMethod* method, 3295 int bci, 3296 Deoptimization::DeoptReason reason) { 3297 ciMethodData* md = method->method_data(); 3298 if (md->is_empty()) { 3299 // Assume the trap has not occurred, or that it occurred only 3300 // because of a transient condition during start-up in the interpreter. 3301 return false; 3302 } 3303 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3304 if (md->has_trap_at(bci, m, reason) != 0) { 3305 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. 3306 // Also, if there are multiple reasons, or if there is no per-BCI record, 3307 // assume the worst. 3308 if (log()) 3309 log()->elem("observe trap='%s' count='%d'", 3310 Deoptimization::trap_reason_name(reason), 3311 md->trap_count(reason)); 3312 return true; 3313 } else { 3314 // Ignore method/bci and see if there have been too many globally. 3315 return too_many_traps(reason, md); 3316 } 3317 } 3318 3319 // Less-accurate variant which does not require a method and bci. 3320 bool Compile::too_many_traps(Deoptimization::DeoptReason reason, 3321 ciMethodData* logmd) { 3322 if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) { 3323 // Too many traps globally. 3324 // Note that we use cumulative trap_count, not just md->trap_count. 3325 if (log()) { 3326 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); 3327 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", 3328 Deoptimization::trap_reason_name(reason), 3329 mcount, trap_count(reason)); 3330 } 3331 return true; 3332 } else { 3333 // The coast is clear. 3334 return false; 3335 } 3336 } 3337 3338 //--------------------------too_many_recompiles-------------------------------- 3339 // Report if there are too many recompiles at the current method and bci. 3340 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. 3341 // Is not eager to return true, since this will cause the compiler to use 3342 // Action_none for a trap point, to avoid too many recompilations. 3343 bool Compile::too_many_recompiles(ciMethod* method, 3344 int bci, 3345 Deoptimization::DeoptReason reason) { 3346 ciMethodData* md = method->method_data(); 3347 if (md->is_empty()) { 3348 // Assume the trap has not occurred, or that it occurred only 3349 // because of a transient condition during start-up in the interpreter. 3350 return false; 3351 } 3352 // Pick a cutoff point well within PerBytecodeRecompilationCutoff. 3353 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; 3354 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero 3355 Deoptimization::DeoptReason per_bc_reason 3356 = Deoptimization::reason_recorded_per_bytecode_if_any(reason); 3357 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL; 3358 if ((per_bc_reason == Deoptimization::Reason_none 3359 || md->has_trap_at(bci, m, reason) != 0) 3360 // The trap frequency measure we care about is the recompile count: 3361 && md->trap_recompiled_at(bci, m) 3362 && md->overflow_recompile_count() >= bc_cutoff) { 3363 // Do not emit a trap here if it has already caused recompilations. 3364 // Also, if there are multiple reasons, or if there is no per-BCI record, 3365 // assume the worst. 3366 if (log()) 3367 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", 3368 Deoptimization::trap_reason_name(reason), 3369 md->trap_count(reason), 3370 md->overflow_recompile_count()); 3371 return true; 3372 } else if (trap_count(reason) != 0 3373 && decompile_count() >= m_cutoff) { 3374 // Too many recompiles globally, and we have seen this sort of trap. 3375 // Use cumulative decompile_count, not just md->decompile_count. 3376 if (log()) 3377 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", 3378 Deoptimization::trap_reason_name(reason), 3379 md->trap_count(reason), trap_count(reason), 3380 md->decompile_count(), decompile_count()); 3381 return true; 3382 } else { 3383 // The coast is clear. 3384 return false; 3385 } 3386 } 3387 3388 // Compute when not to trap. Used by matching trap based nodes and 3389 // NullCheck optimization. 3390 void Compile::set_allowed_deopt_reasons() { 3391 _allowed_reasons = 0; 3392 if (is_method_compilation()) { 3393 for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) { 3394 assert(rs < BitsPerInt, "recode bit map"); 3395 if (!too_many_traps((Deoptimization::DeoptReason) rs)) { 3396 _allowed_reasons |= nth_bit(rs); 3397 } 3398 } 3399 } 3400 } 3401 3402 #ifndef PRODUCT 3403 //------------------------------verify_graph_edges--------------------------- 3404 // Walk the Graph and verify that there is a one-to-one correspondence 3405 // between Use-Def edges and Def-Use edges in the graph. 3406 void Compile::verify_graph_edges(bool no_dead_code) { 3407 if (VerifyGraphEdges) { 3408 ResourceArea *area = Thread::current()->resource_area(); 3409 Unique_Node_List visited(area); 3410 // Call recursive graph walk to check edges 3411 _root->verify_edges(visited); 3412 if (no_dead_code) { 3413 // Now make sure that no visited node is used by an unvisited node. 3414 bool dead_nodes = 0; 3415 Unique_Node_List checked(area); 3416 while (visited.size() > 0) { 3417 Node* n = visited.pop(); 3418 checked.push(n); 3419 for (uint i = 0; i < n->outcnt(); i++) { 3420 Node* use = n->raw_out(i); 3421 if (checked.member(use)) continue; // already checked 3422 if (visited.member(use)) continue; // already in the graph 3423 if (use->is_Con()) continue; // a dead ConNode is OK 3424 // At this point, we have found a dead node which is DU-reachable. 3425 if (dead_nodes++ == 0) 3426 tty->print_cr("*** Dead nodes reachable via DU edges:"); 3427 use->dump(2); 3428 tty->print_cr("---"); 3429 checked.push(use); // No repeats; pretend it is now checked. 3430 } 3431 } 3432 assert(dead_nodes == 0, "using nodes must be reachable from root"); 3433 } 3434 } 3435 } 3436 3437 // Verify GC barriers consistency 3438 // Currently supported: 3439 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre()) 3440 void Compile::verify_barriers() { 3441 if (UseG1GC) { 3442 // Verify G1 pre-barriers 3443 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active()); 3444 3445 ResourceArea *area = Thread::current()->resource_area(); 3446 Unique_Node_List visited(area); 3447 Node_List worklist(area); 3448 // We're going to walk control flow backwards starting from the Root 3449 worklist.push(_root); 3450 while (worklist.size() > 0) { 3451 Node* x = worklist.pop(); 3452 if (x == NULL || x == top()) continue; 3453 if (visited.member(x)) { 3454 continue; 3455 } else { 3456 visited.push(x); 3457 } 3458 3459 if (x->is_Region()) { 3460 for (uint i = 1; i < x->req(); i++) { 3461 worklist.push(x->in(i)); 3462 } 3463 } else { 3464 worklist.push(x->in(0)); 3465 // We are looking for the pattern: 3466 // /->ThreadLocal 3467 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset) 3468 // \->ConI(0) 3469 // We want to verify that the If and the LoadB have the same control 3470 // See GraphKit::g1_write_barrier_pre() 3471 if (x->is_If()) { 3472 IfNode *iff = x->as_If(); 3473 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) { 3474 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp(); 3475 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0 3476 && cmp->in(1)->is_Load()) { 3477 LoadNode* load = cmp->in(1)->as_Load(); 3478 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal 3479 && load->in(2)->in(3)->is_Con() 3480 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) { 3481 3482 Node* if_ctrl = iff->in(0); 3483 Node* load_ctrl = load->in(0); 3484 3485 if (if_ctrl != load_ctrl) { 3486 // Skip possible CProj->NeverBranch in infinite loops 3487 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj) 3488 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) { 3489 if_ctrl = if_ctrl->in(0)->in(0); 3490 } 3491 } 3492 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match"); 3493 } 3494 } 3495 } 3496 } 3497 } 3498 } 3499 } 3500 } 3501 3502 #endif 3503 3504 // The Compile object keeps track of failure reasons separately from the ciEnv. 3505 // This is required because there is not quite a 1-1 relation between the 3506 // ciEnv and its compilation task and the Compile object. Note that one 3507 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides 3508 // to backtrack and retry without subsuming loads. Other than this backtracking 3509 // behavior, the Compile's failure reason is quietly copied up to the ciEnv 3510 // by the logic in C2Compiler. 3511 void Compile::record_failure(const char* reason) { 3512 if (log() != NULL) { 3513 log()->elem("failure reason='%s' phase='compile'", reason); 3514 } 3515 if (_failure_reason == NULL) { 3516 // Record the first failure reason. 3517 _failure_reason = reason; 3518 } 3519 3520 EventCompilerFailure event; 3521 if (event.should_commit()) { 3522 event.set_compileID(Compile::compile_id()); 3523 event.set_failure(reason); 3524 event.commit(); 3525 } 3526 3527 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { 3528 C->print_method(PHASE_FAILURE); 3529 } 3530 _root = NULL; // flush the graph, too 3531 } 3532 3533 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) 3534 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), 3535 _phase_name(name), _dolog(dolog) 3536 { 3537 if (dolog) { 3538 C = Compile::current(); 3539 _log = C->log(); 3540 } else { 3541 C = NULL; 3542 _log = NULL; 3543 } 3544 if (_log != NULL) { 3545 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3546 _log->stamp(); 3547 _log->end_head(); 3548 } 3549 } 3550 3551 Compile::TracePhase::~TracePhase() { 3552 3553 C = Compile::current(); 3554 if (_dolog) { 3555 _log = C->log(); 3556 } else { 3557 _log = NULL; 3558 } 3559 3560 #ifdef ASSERT 3561 if (PrintIdealNodeCount) { 3562 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", 3563 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); 3564 } 3565 3566 if (VerifyIdealNodeCount) { 3567 Compile::current()->print_missing_nodes(); 3568 } 3569 #endif 3570 3571 if (_log != NULL) { 3572 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); 3573 } 3574 } 3575 3576 //============================================================================= 3577 // Two Constant's are equal when the type and the value are equal. 3578 bool Compile::Constant::operator==(const Constant& other) { 3579 if (type() != other.type() ) return false; 3580 if (can_be_reused() != other.can_be_reused()) return false; 3581 // For floating point values we compare the bit pattern. 3582 switch (type()) { 3583 case T_FLOAT: return (_v._value.i == other._v._value.i); 3584 case T_LONG: 3585 case T_DOUBLE: return (_v._value.j == other._v._value.j); 3586 case T_OBJECT: 3587 case T_ADDRESS: return (_v._value.l == other._v._value.l); 3588 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries 3589 case T_METADATA: return (_v._metadata == other._v._metadata); 3590 default: ShouldNotReachHere(); 3591 } 3592 return false; 3593 } 3594 3595 static int type_to_size_in_bytes(BasicType t) { 3596 switch (t) { 3597 case T_LONG: return sizeof(jlong ); 3598 case T_FLOAT: return sizeof(jfloat ); 3599 case T_DOUBLE: return sizeof(jdouble); 3600 case T_METADATA: return sizeof(Metadata*); 3601 // We use T_VOID as marker for jump-table entries (labels) which 3602 // need an internal word relocation. 3603 case T_VOID: 3604 case T_ADDRESS: 3605 case T_OBJECT: return sizeof(jobject); 3606 } 3607 3608 ShouldNotReachHere(); 3609 return -1; 3610 } 3611 3612 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { 3613 // sort descending 3614 if (a->freq() > b->freq()) return -1; 3615 if (a->freq() < b->freq()) return 1; 3616 return 0; 3617 } 3618 3619 void Compile::ConstantTable::calculate_offsets_and_size() { 3620 // First, sort the array by frequencies. 3621 _constants.sort(qsort_comparator); 3622 3623 #ifdef ASSERT 3624 // Make sure all jump-table entries were sorted to the end of the 3625 // array (they have a negative frequency). 3626 bool found_void = false; 3627 for (int i = 0; i < _constants.length(); i++) { 3628 Constant con = _constants.at(i); 3629 if (con.type() == T_VOID) 3630 found_void = true; // jump-tables 3631 else 3632 assert(!found_void, "wrong sorting"); 3633 } 3634 #endif 3635 3636 int offset = 0; 3637 for (int i = 0; i < _constants.length(); i++) { 3638 Constant* con = _constants.adr_at(i); 3639 3640 // Align offset for type. 3641 int typesize = type_to_size_in_bytes(con->type()); 3642 offset = align_size_up(offset, typesize); 3643 con->set_offset(offset); // set constant's offset 3644 3645 if (con->type() == T_VOID) { 3646 MachConstantNode* n = (MachConstantNode*) con->get_jobject(); 3647 offset = offset + typesize * n->outcnt(); // expand jump-table 3648 } else { 3649 offset = offset + typesize; 3650 } 3651 } 3652 3653 // Align size up to the next section start (which is insts; see 3654 // CodeBuffer::align_at_start). 3655 assert(_size == -1, "already set?"); 3656 _size = align_size_up(offset, CodeEntryAlignment); 3657 } 3658 3659 void Compile::ConstantTable::emit(CodeBuffer& cb) { 3660 MacroAssembler _masm(&cb); 3661 for (int i = 0; i < _constants.length(); i++) { 3662 Constant con = _constants.at(i); 3663 address constant_addr; 3664 switch (con.type()) { 3665 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; 3666 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; 3667 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; 3668 case T_OBJECT: { 3669 jobject obj = con.get_jobject(); 3670 int oop_index = _masm.oop_recorder()->find_index(obj); 3671 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); 3672 break; 3673 } 3674 case T_ADDRESS: { 3675 address addr = (address) con.get_jobject(); 3676 constant_addr = _masm.address_constant(addr); 3677 break; 3678 } 3679 // We use T_VOID as marker for jump-table entries (labels) which 3680 // need an internal word relocation. 3681 case T_VOID: { 3682 MachConstantNode* n = (MachConstantNode*) con.get_jobject(); 3683 // Fill the jump-table with a dummy word. The real value is 3684 // filled in later in fill_jump_table. 3685 address dummy = (address) n; 3686 constant_addr = _masm.address_constant(dummy); 3687 // Expand jump-table 3688 for (uint i = 1; i < n->outcnt(); i++) { 3689 address temp_addr = _masm.address_constant(dummy + i); 3690 assert(temp_addr, "consts section too small"); 3691 } 3692 break; 3693 } 3694 case T_METADATA: { 3695 Metadata* obj = con.get_metadata(); 3696 int metadata_index = _masm.oop_recorder()->find_index(obj); 3697 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); 3698 break; 3699 } 3700 default: ShouldNotReachHere(); 3701 } 3702 assert(constant_addr, "consts section too small"); 3703 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), 3704 err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()))); 3705 } 3706 } 3707 3708 int Compile::ConstantTable::find_offset(Constant& con) const { 3709 int idx = _constants.find(con); 3710 assert(idx != -1, "constant must be in constant table"); 3711 int offset = _constants.at(idx).offset(); 3712 assert(offset != -1, "constant table not emitted yet?"); 3713 return offset; 3714 } 3715 3716 void Compile::ConstantTable::add(Constant& con) { 3717 if (con.can_be_reused()) { 3718 int idx = _constants.find(con); 3719 if (idx != -1 && _constants.at(idx).can_be_reused()) { 3720 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value 3721 return; 3722 } 3723 } 3724 (void) _constants.append(con); 3725 } 3726 3727 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { 3728 Block* b = Compile::current()->cfg()->get_block_for_node(n); 3729 Constant con(type, value, b->_freq); 3730 add(con); 3731 return con; 3732 } 3733 3734 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { 3735 Constant con(metadata); 3736 add(con); 3737 return con; 3738 } 3739 3740 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { 3741 jvalue value; 3742 BasicType type = oper->type()->basic_type(); 3743 switch (type) { 3744 case T_LONG: value.j = oper->constantL(); break; 3745 case T_FLOAT: value.f = oper->constantF(); break; 3746 case T_DOUBLE: value.d = oper->constantD(); break; 3747 case T_OBJECT: 3748 case T_ADDRESS: value.l = (jobject) oper->constant(); break; 3749 case T_METADATA: return add((Metadata*)oper->constant()); break; 3750 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); 3751 } 3752 return add(n, type, value); 3753 } 3754 3755 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { 3756 jvalue value; 3757 // We can use the node pointer here to identify the right jump-table 3758 // as this method is called from Compile::Fill_buffer right before 3759 // the MachNodes are emitted and the jump-table is filled (means the 3760 // MachNode pointers do not change anymore). 3761 value.l = (jobject) n; 3762 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. 3763 add(con); 3764 return con; 3765 } 3766 3767 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const { 3768 // If called from Compile::scratch_emit_size do nothing. 3769 if (Compile::current()->in_scratch_emit_size()) return; 3770 3771 assert(labels.is_nonempty(), "must be"); 3772 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); 3773 3774 // Since MachConstantNode::constant_offset() also contains 3775 // table_base_offset() we need to subtract the table_base_offset() 3776 // to get the plain offset into the constant table. 3777 int offset = n->constant_offset() - table_base_offset(); 3778 3779 MacroAssembler _masm(&cb); 3780 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); 3781 3782 for (uint i = 0; i < n->outcnt(); i++) { 3783 address* constant_addr = &jump_table_base[i]; 3784 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))); 3785 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); 3786 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); 3787 } 3788 } 3789 3790 //----------------------------static_subtype_check----------------------------- 3791 // Shortcut important common cases when superklass is exact: 3792 // (0) superklass is java.lang.Object (can occur in reflective code) 3793 // (1) subklass is already limited to a subtype of superklass => always ok 3794 // (2) subklass does not overlap with superklass => always fail 3795 // (3) superklass has NO subtypes and we can check with a simple compare. 3796 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) { 3797 if (StressReflectiveCode) { 3798 return SSC_full_test; // Let caller generate the general case. 3799 } 3800 3801 if (superk == env()->Object_klass()) { 3802 return SSC_always_true; // (0) this test cannot fail 3803 } 3804 3805 ciType* superelem = superk; 3806 if (superelem->is_array_klass()) 3807 superelem = superelem->as_array_klass()->base_element_type(); 3808 3809 if (!subk->is_interface()) { // cannot trust static interface types yet 3810 if (subk->is_subtype_of(superk)) { 3811 return SSC_always_true; // (1) false path dead; no dynamic test needed 3812 } 3813 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) && 3814 !superk->is_subtype_of(subk)) { 3815 return SSC_always_false; 3816 } 3817 } 3818 3819 // If casting to an instance klass, it must have no subtypes 3820 if (superk->is_interface()) { 3821 // Cannot trust interfaces yet. 3822 // %%% S.B. superk->nof_implementors() == 1 3823 } else if (superelem->is_instance_klass()) { 3824 ciInstanceKlass* ik = superelem->as_instance_klass(); 3825 if (!ik->has_subklass() && !ik->is_interface()) { 3826 if (!ik->is_final()) { 3827 // Add a dependency if there is a chance of a later subclass. 3828 dependencies()->assert_leaf_type(ik); 3829 } 3830 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3831 } 3832 } else { 3833 // A primitive array type has no subtypes. 3834 return SSC_easy_test; // (3) caller can do a simple ptr comparison 3835 } 3836 3837 return SSC_full_test; 3838 } 3839 3840 // The message about the current inlining is accumulated in 3841 // _print_inlining_stream and transfered into the _print_inlining_list 3842 // once we know whether inlining succeeds or not. For regular 3843 // inlining, messages are appended to the buffer pointed by 3844 // _print_inlining_idx in the _print_inlining_list. For late inlining, 3845 // a new buffer is added after _print_inlining_idx in the list. This 3846 // way we can update the inlining message for late inlining call site 3847 // when the inlining is attempted again. 3848 void Compile::print_inlining_init() { 3849 if (print_inlining() || print_intrinsics()) { 3850 _print_inlining_stream = new stringStream(); 3851 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer()); 3852 } 3853 } 3854 3855 void Compile::print_inlining_reinit() { 3856 if (print_inlining() || print_intrinsics()) { 3857 // Re allocate buffer when we change ResourceMark 3858 _print_inlining_stream = new stringStream(); 3859 } 3860 } 3861 3862 void Compile::print_inlining_reset() { 3863 _print_inlining_stream->reset(); 3864 } 3865 3866 void Compile::print_inlining_commit() { 3867 assert(print_inlining() || print_intrinsics(), "PrintInlining off?"); 3868 // Transfer the message from _print_inlining_stream to the current 3869 // _print_inlining_list buffer and clear _print_inlining_stream. 3870 _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size()); 3871 print_inlining_reset(); 3872 } 3873 3874 void Compile::print_inlining_push() { 3875 // Add new buffer to the _print_inlining_list at current position 3876 _print_inlining_idx++; 3877 _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer()); 3878 } 3879 3880 Compile::PrintInliningBuffer& Compile::print_inlining_current() { 3881 return _print_inlining_list->at(_print_inlining_idx); 3882 } 3883 3884 void Compile::print_inlining_update(CallGenerator* cg) { 3885 if (print_inlining() || print_intrinsics()) { 3886 if (!cg->is_late_inline()) { 3887 if (print_inlining_current().cg() != NULL) { 3888 print_inlining_push(); 3889 } 3890 print_inlining_commit(); 3891 } else { 3892 if (print_inlining_current().cg() != cg && 3893 (print_inlining_current().cg() != NULL || 3894 print_inlining_current().ss()->size() != 0)) { 3895 print_inlining_push(); 3896 } 3897 print_inlining_commit(); 3898 print_inlining_current().set_cg(cg); 3899 } 3900 } 3901 } 3902 3903 void Compile::print_inlining_move_to(CallGenerator* cg) { 3904 // We resume inlining at a late inlining call site. Locate the 3905 // corresponding inlining buffer so that we can update it. 3906 if (print_inlining()) { 3907 for (int i = 0; i < _print_inlining_list->length(); i++) { 3908 if (_print_inlining_list->adr_at(i)->cg() == cg) { 3909 _print_inlining_idx = i; 3910 return; 3911 } 3912 } 3913 ShouldNotReachHere(); 3914 } 3915 } 3916 3917 void Compile::print_inlining_update_delayed(CallGenerator* cg) { 3918 if (print_inlining()) { 3919 assert(_print_inlining_stream->size() > 0, "missing inlining msg"); 3920 assert(print_inlining_current().cg() == cg, "wrong entry"); 3921 // replace message with new message 3922 _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer()); 3923 print_inlining_commit(); 3924 print_inlining_current().set_cg(cg); 3925 } 3926 } 3927 3928 void Compile::print_inlining_assert_ready() { 3929 assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data"); 3930 } 3931 3932 void Compile::process_print_inlining() { 3933 bool do_print_inlining = print_inlining() || print_intrinsics(); 3934 if (do_print_inlining || log() != NULL) { 3935 // Print inlining message for candidates that we couldn't inline 3936 // for lack of space 3937 for (int i = 0; i < _late_inlines.length(); i++) { 3938 CallGenerator* cg = _late_inlines.at(i); 3939 if (!cg->is_mh_late_inline()) { 3940 const char* msg = "live nodes > LiveNodeCountInliningCutoff"; 3941 if (do_print_inlining) { 3942 cg->print_inlining_late(msg); 3943 } 3944 log_late_inline_failure(cg, msg); 3945 } 3946 } 3947 } 3948 if (do_print_inlining) { 3949 ResourceMark rm; 3950 stringStream ss; 3951 for (int i = 0; i < _print_inlining_list->length(); i++) { 3952 ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string()); 3953 } 3954 size_t end = ss.size(); 3955 _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1); 3956 strncpy(_print_inlining_output, ss.base(), end+1); 3957 _print_inlining_output[end] = 0; 3958 } 3959 } 3960 3961 void Compile::dump_print_inlining() { 3962 if (_print_inlining_output != NULL) { 3963 tty->print_raw(_print_inlining_output); 3964 } 3965 } 3966 3967 void Compile::log_late_inline(CallGenerator* cg) { 3968 if (log() != NULL) { 3969 log()->head("late_inline method='%d' inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()), 3970 cg->unique_id()); 3971 JVMState* p = cg->call_node()->jvms(); 3972 while (p != NULL) { 3973 log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method())); 3974 p = p->caller(); 3975 } 3976 log()->tail("late_inline"); 3977 } 3978 } 3979 3980 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) { 3981 log_late_inline(cg); 3982 if (log() != NULL) { 3983 log()->inline_fail(msg); 3984 } 3985 } 3986 3987 void Compile::log_inline_id(CallGenerator* cg) { 3988 if (log() != NULL) { 3989 // The LogCompilation tool needs a unique way to identify late 3990 // inline call sites. This id must be unique for this call site in 3991 // this compilation. Try to have it unique across compilations as 3992 // well because it can be convenient when grepping through the log 3993 // file. 3994 // Distinguish OSR compilations from others in case CICountOSR is 3995 // on. 3996 jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0); 3997 cg->set_unique_id(id); 3998 log()->elem("inline_id id='" JLONG_FORMAT "'", id); 3999 } 4000 } 4001 4002 void Compile::log_inline_failure(const char* msg) { 4003 if (C->log() != NULL) { 4004 C->log()->inline_fail(msg); 4005 } 4006 } 4007 4008 4009 // Dump inlining replay data to the stream. 4010 // Don't change thread state and acquire any locks. 4011 void Compile::dump_inline_data(outputStream* out) { 4012 InlineTree* inl_tree = ilt(); 4013 if (inl_tree != NULL) { 4014 out->print(" inline %d", inl_tree->count()); 4015 inl_tree->dump_replay_data(out); 4016 } 4017 } 4018 4019 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { 4020 if (n1->Opcode() < n2->Opcode()) return -1; 4021 else if (n1->Opcode() > n2->Opcode()) return 1; 4022 4023 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())); 4024 for (uint i = 1; i < n1->req(); i++) { 4025 if (n1->in(i) < n2->in(i)) return -1; 4026 else if (n1->in(i) > n2->in(i)) return 1; 4027 } 4028 4029 return 0; 4030 } 4031 4032 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { 4033 Node* n1 = *n1p; 4034 Node* n2 = *n2p; 4035 4036 return cmp_expensive_nodes(n1, n2); 4037 } 4038 4039 void Compile::sort_expensive_nodes() { 4040 if (!expensive_nodes_sorted()) { 4041 _expensive_nodes->sort(cmp_expensive_nodes); 4042 } 4043 } 4044 4045 bool Compile::expensive_nodes_sorted() const { 4046 for (int i = 1; i < _expensive_nodes->length(); i++) { 4047 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { 4048 return false; 4049 } 4050 } 4051 return true; 4052 } 4053 4054 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { 4055 if (_expensive_nodes->length() == 0) { 4056 return false; 4057 } 4058 4059 assert(OptimizeExpensiveOps, "optimization off?"); 4060 4061 // Take this opportunity to remove dead nodes from the list 4062 int j = 0; 4063 for (int i = 0; i < _expensive_nodes->length(); i++) { 4064 Node* n = _expensive_nodes->at(i); 4065 if (!n->is_unreachable(igvn)) { 4066 assert(n->is_expensive(), "should be expensive"); 4067 _expensive_nodes->at_put(j, n); 4068 j++; 4069 } 4070 } 4071 _expensive_nodes->trunc_to(j); 4072 4073 // Then sort the list so that similar nodes are next to each other 4074 // and check for at least two nodes of identical kind with same data 4075 // inputs. 4076 sort_expensive_nodes(); 4077 4078 for (int i = 0; i < _expensive_nodes->length()-1; i++) { 4079 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { 4080 return true; 4081 } 4082 } 4083 4084 return false; 4085 } 4086 4087 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { 4088 if (_expensive_nodes->length() == 0) { 4089 return; 4090 } 4091 4092 assert(OptimizeExpensiveOps, "optimization off?"); 4093 4094 // Sort to bring similar nodes next to each other and clear the 4095 // control input of nodes for which there's only a single copy. 4096 sort_expensive_nodes(); 4097 4098 int j = 0; 4099 int identical = 0; 4100 int i = 0; 4101 for (; i < _expensive_nodes->length()-1; i++) { 4102 assert(j <= i, "can't write beyond current index"); 4103 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { 4104 identical++; 4105 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4106 continue; 4107 } 4108 if (identical > 0) { 4109 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4110 identical = 0; 4111 } else { 4112 Node* n = _expensive_nodes->at(i); 4113 igvn.hash_delete(n); 4114 n->set_req(0, NULL); 4115 igvn.hash_insert(n); 4116 } 4117 } 4118 if (identical > 0) { 4119 _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); 4120 } else if (_expensive_nodes->length() >= 1) { 4121 Node* n = _expensive_nodes->at(i); 4122 igvn.hash_delete(n); 4123 n->set_req(0, NULL); 4124 igvn.hash_insert(n); 4125 } 4126 _expensive_nodes->trunc_to(j); 4127 } 4128 4129 void Compile::add_expensive_node(Node * n) { 4130 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); 4131 assert(n->is_expensive(), "expensive nodes with non-null control here only"); 4132 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); 4133 if (OptimizeExpensiveOps) { 4134 _expensive_nodes->append(n); 4135 } else { 4136 // Clear control input and let IGVN optimize expensive nodes if 4137 // OptimizeExpensiveOps is off. 4138 n->set_req(0, NULL); 4139 } 4140 } 4141 4142 /** 4143 * Remove the speculative part of types and clean up the graph 4144 */ 4145 void Compile::remove_speculative_types(PhaseIterGVN &igvn) { 4146 if (UseTypeSpeculation) { 4147 Unique_Node_List worklist; 4148 worklist.push(root()); 4149 int modified = 0; 4150 // Go over all type nodes that carry a speculative type, drop the 4151 // speculative part of the type and enqueue the node for an igvn 4152 // which may optimize it out. 4153 for (uint next = 0; next < worklist.size(); ++next) { 4154 Node *n = worklist.at(next); 4155 if (n->is_Type()) { 4156 TypeNode* tn = n->as_Type(); 4157 const Type* t = tn->type(); 4158 const Type* t_no_spec = t->remove_speculative(); 4159 if (t_no_spec != t) { 4160 bool in_hash = igvn.hash_delete(n); 4161 assert(in_hash, "node should be in igvn hash table"); 4162 tn->set_type(t_no_spec); 4163 igvn.hash_insert(n); 4164 igvn._worklist.push(n); // give it a chance to go away 4165 modified++; 4166 } 4167 } 4168 uint max = n->len(); 4169 for( uint i = 0; i < max; ++i ) { 4170 Node *m = n->in(i); 4171 if (not_a_node(m)) continue; 4172 worklist.push(m); 4173 } 4174 } 4175 // Drop the speculative part of all types in the igvn's type table 4176 igvn.remove_speculative_types(); 4177 if (modified > 0) { 4178 igvn.optimize(); 4179 } 4180 #ifdef ASSERT 4181 // Verify that after the IGVN is over no speculative type has resurfaced 4182 worklist.clear(); 4183 worklist.push(root()); 4184 for (uint next = 0; next < worklist.size(); ++next) { 4185 Node *n = worklist.at(next); 4186 const Type* t = igvn.type_or_null(n); 4187 assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types"); 4188 if (n->is_Type()) { 4189 t = n->as_Type()->type(); 4190 assert(t == t->remove_speculative(), "no more speculative types"); 4191 } 4192 uint max = n->len(); 4193 for( uint i = 0; i < max; ++i ) { 4194 Node *m = n->in(i); 4195 if (not_a_node(m)) continue; 4196 worklist.push(m); 4197 } 4198 } 4199 igvn.check_no_speculative_types(); 4200 #endif 4201 } 4202 } 4203 4204 // Auxiliary method to support randomized stressing/fuzzing. 4205 // 4206 // This method can be called the arbitrary number of times, with current count 4207 // as the argument. The logic allows selecting a single candidate from the 4208 // running list of candidates as follows: 4209 // int count = 0; 4210 // Cand* selected = null; 4211 // while(cand = cand->next()) { 4212 // if (randomized_select(++count)) { 4213 // selected = cand; 4214 // } 4215 // } 4216 // 4217 // Including count equalizes the chances any candidate is "selected". 4218 // This is useful when we don't have the complete list of candidates to choose 4219 // from uniformly. In this case, we need to adjust the randomicity of the 4220 // selection, or else we will end up biasing the selection towards the latter 4221 // candidates. 4222 // 4223 // Quick back-envelope calculation shows that for the list of n candidates 4224 // the equal probability for the candidate to persist as "best" can be 4225 // achieved by replacing it with "next" k-th candidate with the probability 4226 // of 1/k. It can be easily shown that by the end of the run, the 4227 // probability for any candidate is converged to 1/n, thus giving the 4228 // uniform distribution among all the candidates. 4229 // 4230 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. 4231 #define RANDOMIZED_DOMAIN_POW 29 4232 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) 4233 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) 4234 bool Compile::randomized_select(int count) { 4235 assert(count > 0, "only positive"); 4236 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); 4237 }