1 /* 2 * Copyright (c) 1998, 2019, 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 "ci/ciMethodData.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "interpreter/linkResolver.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "memory/universe.hpp" 33 #include "oops/oop.inline.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/divnode.hpp" 38 #include "opto/idealGraphPrinter.hpp" 39 #include "opto/idealKit.hpp" 40 #include "opto/matcher.hpp" 41 #include "opto/memnode.hpp" 42 #include "opto/mulnode.hpp" 43 #include "opto/opaquenode.hpp" 44 #include "opto/parse.hpp" 45 #include "opto/runtime.hpp" 46 #include "opto/valuetypenode.hpp" 47 #include "runtime/deoptimization.hpp" 48 #include "runtime/sharedRuntime.hpp" 49 50 #ifndef PRODUCT 51 extern int explicit_null_checks_inserted, 52 explicit_null_checks_elided; 53 #endif 54 55 //---------------------------------array_load---------------------------------- 56 void Parse::array_load(BasicType bt) { 57 const Type* elemtype = Type::TOP; 58 Node* adr = array_addressing(bt, 0, &elemtype); 59 if (stopped()) return; // guaranteed null or range check 60 61 Node* idx = pop(); 62 Node* ary = pop(); 63 64 // Handle value type arrays 65 const TypeOopPtr* elemptr = elemtype->make_oopptr(); 66 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr(); 67 if (elemtype->isa_valuetype() != NULL) { 68 // Load from flattened value type array 69 Node* vt = ValueTypeNode::make_from_flattened(this, elemtype->value_klass(), ary, adr); 70 push(vt); 71 return; 72 } else if (elemptr != NULL && elemptr->is_valuetypeptr() && !elemptr->maybe_null()) { 73 // Load from non-flattened but flattenable value type array (elements can never be null) 74 bt = T_VALUETYPE; 75 } else if (ValueArrayFlatten && elemptr != NULL && elemptr->can_be_value_type() && 76 ((!elemptr->is_valuetypeptr() && !ary_t->klass_is_exact()) || (elemptr->is_valuetypeptr() && elemptr->value_klass()->flatten_array()))) { 77 // Cannot statically determine if array is flattened, emit runtime check 78 IdealKit ideal(this); 79 IdealVariable res(ideal); 80 ideal.declarations_done(); 81 Node* kls = load_object_klass(ary); 82 Node* tag = load_lh_array_tag(kls); 83 ideal.if_then(tag, BoolTest::ne, intcon(Klass::_lh_array_tag_vt_value)); { 84 // non-flattened 85 sync_kit(ideal); 86 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 87 Node* ld = access_load_at(ary, adr, adr_type, elemptr, bt, 88 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD); 89 ideal.sync_kit(this); 90 ideal.set(res, ld); 91 } ideal.else_(); { 92 // flattened 93 sync_kit(ideal); 94 if (elemptr->is_valuetypeptr()) { 95 // Element type is known, cast and load from flattened representation 96 assert(elemptr->maybe_null(), "must be nullable"); 97 ciValueKlass* vk = elemptr->value_klass(); 98 assert(vk->flatten_array(), "must be flattenable"); 99 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* never_null */ true); 100 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr(); 101 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, arytype)); 102 adr = array_element_address(cast, idx, T_VALUETYPE, ary_t->size(), control()); 103 Node* vt = ValueTypeNode::make_from_flattened(this, vk, cast, adr)->allocate(this, false, false)->get_oop(); 104 ideal.set(res, vt); 105 } else { 106 // Element type is unknown, emit runtime call 107 assert(!ary_t->klass_is_exact(), "should not have exact type here"); 108 Node* k_adr = basic_plus_adr(kls, in_bytes(ArrayKlass::element_klass_offset())); 109 Node* elem_klass = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS)); 110 Node* obj_size = NULL; 111 kill_dead_locals(); 112 inc_sp(2); 113 Node* alloc_obj = new_instance(elem_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 114 dec_sp(2); 115 116 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 117 assert(alloc->maybe_set_complete(&_gvn), ""); 118 alloc->initialization()->set_complete_with_arraycopy(); 119 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 120 // Unknown value type might contain reference fields 121 if (!bs->array_copy_requires_gc_barriers(false, T_OBJECT, false, BarrierSetC2::Parsing)) { 122 int base_off = sizeof(instanceOopDesc); 123 Node* dst_base = basic_plus_adr(alloc_obj, base_off); 124 Node* countx = obj_size; 125 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 126 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong))); 127 128 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place"); 129 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset())); 130 Node* elem_shift = make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered); 131 uint header = arrayOopDesc::base_offset_in_bytes(T_VALUETYPE); 132 Node* base = basic_plus_adr(ary, header); 133 idx = Compile::conv_I2X_index(&_gvn, idx, TypeInt::POS, control()); 134 Node* scale = _gvn.transform(new LShiftXNode(idx, elem_shift)); 135 Node* adr = basic_plus_adr(ary, base, scale); 136 137 access_clone(adr, dst_base, countx, false); 138 } else { 139 ideal.sync_kit(this); 140 ideal.make_leaf_call(OptoRuntime::load_unknown_value_Type(), 141 CAST_FROM_FN_PTR(address, OptoRuntime::load_unknown_value), 142 "load_unknown_value", 143 ary, idx, alloc_obj); 144 sync_kit(ideal); 145 } 146 147 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress)); 148 149 ideal.set(res, alloc_obj); 150 } 151 ideal.sync_kit(this); 152 } ideal.end_if(); 153 sync_kit(ideal); 154 push_node(bt, _gvn.transform(ideal.value(res))); 155 return; 156 } 157 158 if (elemtype == TypeInt::BOOL) { 159 bt = T_BOOLEAN; 160 } else if (bt == T_OBJECT) { 161 elemtype = ary_t->elem()->make_oopptr(); 162 } 163 164 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 165 Node* ld = access_load_at(ary, adr, adr_type, elemtype, bt, 166 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD); 167 if (bt == T_VALUETYPE) { 168 // Loading a non-flattened (but flattenable) value type from an array 169 assert(!gvn().type(ld)->maybe_null(), "value type array elements should never be null"); 170 if (elemptr->value_klass()->is_scalarizable()) { 171 ld = ValueTypeNode::make_from_oop(this, ld, elemptr->value_klass()); 172 } 173 } 174 175 push_node(bt, ld); 176 } 177 178 179 //--------------------------------array_store---------------------------------- 180 void Parse::array_store(BasicType bt) { 181 const Type* elemtype = Type::TOP; 182 Node* adr = array_addressing(bt, type2size[bt], &elemtype); 183 if (stopped()) return; // guaranteed null or range check 184 Node* cast_val = NULL; 185 if (bt == T_OBJECT) { 186 cast_val = array_store_check(); 187 if (stopped()) return; 188 } 189 Node* val = pop_node(bt); // Value to store 190 Node* idx = pop(); // Index in the array 191 Node* ary = pop(); // The array itself 192 193 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr(); 194 if (bt == T_OBJECT) { 195 const TypeOopPtr* elemptr = elemtype->make_oopptr(); 196 const Type* val_t = _gvn.type(val); 197 if (elemtype->isa_valuetype() != NULL) { 198 // Store to flattened value type array 199 if (!cast_val->is_ValueType()) { 200 inc_sp(3); 201 cast_val = null_check(cast_val); 202 if (stopped()) return; 203 dec_sp(3); 204 cast_val = ValueTypeNode::make_from_oop(this, cast_val, elemtype->value_klass()); 205 } 206 cast_val->as_ValueType()->store_flattened(this, ary, adr); 207 return; 208 } else if (elemptr->is_valuetypeptr() && !elemptr->maybe_null()) { 209 // Store to non-flattened but flattenable value type array (elements can never be null) 210 if (!cast_val->is_ValueType()) { 211 inc_sp(3); 212 cast_val = null_check(cast_val); 213 if (stopped()) return; 214 dec_sp(3); 215 } 216 } else if (elemptr->can_be_value_type() && (!ary_t->klass_is_exact() || elemptr->is_valuetypeptr()) && 217 (val->is_ValueType() || val_t == TypePtr::NULL_PTR || val_t->is_oopptr()->can_be_value_type())) { 218 // Cannot statically determine if array is flattened, emit runtime check 219 ciValueKlass* vk = NULL; 220 // Try to determine the value klass 221 if (val->is_ValueType()) { 222 vk = val_t->value_klass(); 223 } else if (elemptr->is_valuetypeptr()) { 224 vk = elemptr->value_klass(); 225 } 226 if (ValueArrayFlatten && (vk == NULL || vk->flatten_array())) { 227 IdealKit ideal(this); 228 Node* kls = load_object_klass(ary); 229 Node* layout_val = load_lh_array_tag(kls); 230 ideal.if_then(layout_val, BoolTest::ne, intcon(Klass::_lh_array_tag_vt_value)); { 231 // non-flattened 232 sync_kit(ideal); 233 gen_value_array_null_guard(ary, val, 3); 234 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 235 elemtype = ary_t->elem()->make_oopptr(); 236 access_store_at(ary, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false, false); 237 ideal.sync_kit(this); 238 } ideal.else_(); { 239 // flattened 240 if (!val->is_ValueType() && TypePtr::NULL_PTR->higher_equal(val_t)) { 241 // Add null check 242 sync_kit(ideal); 243 Node* null_ctl = top(); 244 val = null_check_oop(val, &null_ctl); 245 if (null_ctl != top()) { 246 PreserveJVMState pjvms(this); 247 inc_sp(3); 248 set_control(null_ctl); 249 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none); 250 dec_sp(3); 251 } 252 ideal.sync_kit(this); 253 } 254 if (vk != NULL && !stopped()) { 255 // Element type is known, cast and store to flattened representation 256 sync_kit(ideal); 257 assert(vk->flatten_array(), "must be flattenable"); 258 assert(elemptr->maybe_null(), "must be nullable"); 259 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* never_null */ true); 260 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr(); 261 ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype)); 262 adr = array_element_address(ary, idx, T_OBJECT, arytype->size(), control()); 263 if (!val->is_ValueType()) { 264 assert(!gvn().type(val)->maybe_null(), "value type array elements should never be null"); 265 val = ValueTypeNode::make_from_oop(this, val, vk); 266 } 267 val->as_ValueType()->store_flattened(this, ary, adr); 268 ideal.sync_kit(this); 269 } else if (!ideal.ctrl()->is_top()) { 270 // Element type is unknown, emit runtime call 271 assert(!ary_t->klass_is_exact(), "should not have exact type here"); 272 ideal.make_leaf_call(OptoRuntime::store_unknown_value_Type(), 273 CAST_FROM_FN_PTR(address, OptoRuntime::store_unknown_value), 274 "store_unknown_value", 275 val, ary, idx); 276 } 277 } ideal.end_if(); 278 sync_kit(ideal); 279 return; 280 } else { 281 gen_value_array_null_guard(ary, val, 3); 282 } 283 } 284 } 285 286 if (elemtype == TypeInt::BOOL) { 287 bt = T_BOOLEAN; 288 } else if (bt == T_OBJECT) { 289 elemtype = ary_t->elem()->make_oopptr(); 290 } 291 292 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 293 294 access_store_at(ary, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY); 295 } 296 297 298 //------------------------------array_addressing------------------------------- 299 // Pull array and index from the stack. Compute pointer-to-element. 300 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 301 Node *idx = peek(0+vals); // Get from stack without popping 302 Node *ary = peek(1+vals); // in case of exception 303 304 // Null check the array base, with correct stack contents 305 ary = null_check(ary, T_ARRAY); 306 // Compile-time detect of null-exception? 307 if (stopped()) return top(); 308 309 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 310 const TypeInt* sizetype = arytype->size(); 311 const Type* elemtype = arytype->elem(); 312 313 if (UseUniqueSubclasses && result2 != NULL) { 314 const Type* el = elemtype->make_ptr(); 315 if (el && el->isa_instptr()) { 316 const TypeInstPtr* toop = el->is_instptr(); 317 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 318 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 319 const Type* subklass = Type::get_const_type(toop->klass()); 320 elemtype = subklass->join_speculative(el); 321 } 322 } 323 } 324 325 // Check for big class initializers with all constant offsets 326 // feeding into a known-size array. 327 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 328 // See if the highest idx value is less than the lowest array bound, 329 // and if the idx value cannot be negative: 330 bool need_range_check = true; 331 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 332 need_range_check = false; 333 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 334 } 335 336 ciKlass * arytype_klass = arytype->klass(); 337 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 338 // Only fails for some -Xcomp runs 339 // The class is unloaded. We have to run this bytecode in the interpreter. 340 uncommon_trap(Deoptimization::Reason_unloaded, 341 Deoptimization::Action_reinterpret, 342 arytype->klass(), "!loaded array"); 343 return top(); 344 } 345 346 // Do the range check 347 if (GenerateRangeChecks && need_range_check) { 348 Node* tst; 349 if (sizetype->_hi <= 0) { 350 // The greatest array bound is negative, so we can conclude that we're 351 // compiling unreachable code, but the unsigned compare trick used below 352 // only works with non-negative lengths. Instead, hack "tst" to be zero so 353 // the uncommon_trap path will always be taken. 354 tst = _gvn.intcon(0); 355 } else { 356 // Range is constant in array-oop, so we can use the original state of mem 357 Node* len = load_array_length(ary); 358 359 // Test length vs index (standard trick using unsigned compare) 360 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 361 BoolTest::mask btest = BoolTest::lt; 362 tst = _gvn.transform( new BoolNode(chk, btest) ); 363 } 364 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 365 _gvn.set_type(rc, rc->Value(&_gvn)); 366 if (!tst->is_Con()) { 367 record_for_igvn(rc); 368 } 369 set_control(_gvn.transform(new IfTrueNode(rc))); 370 // Branch to failure if out of bounds 371 { 372 PreserveJVMState pjvms(this); 373 set_control(_gvn.transform(new IfFalseNode(rc))); 374 if (C->allow_range_check_smearing()) { 375 // Do not use builtin_throw, since range checks are sometimes 376 // made more stringent by an optimistic transformation. 377 // This creates "tentative" range checks at this point, 378 // which are not guaranteed to throw exceptions. 379 // See IfNode::Ideal, is_range_check, adjust_check. 380 uncommon_trap(Deoptimization::Reason_range_check, 381 Deoptimization::Action_make_not_entrant, 382 NULL, "range_check"); 383 } else { 384 // If we have already recompiled with the range-check-widening 385 // heroic optimization turned off, then we must really be throwing 386 // range check exceptions. 387 builtin_throw(Deoptimization::Reason_range_check, idx); 388 } 389 } 390 } 391 // Check for always knowing you are throwing a range-check exception 392 if (stopped()) return top(); 393 394 // Make array address computation control dependent to prevent it 395 // from floating above the range check during loop optimizations. 396 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 397 398 if (result2 != NULL) *result2 = elemtype; 399 400 assert(ptr != top(), "top should go hand-in-hand with stopped"); 401 402 return ptr; 403 } 404 405 406 // returns IfNode 407 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 408 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 409 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 410 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 411 return iff; 412 } 413 414 // return Region node 415 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 416 Node *region = new RegionNode(3); // 2 results 417 record_for_igvn(region); 418 region->init_req(1, iffalse); 419 region->init_req(2, iftrue ); 420 _gvn.set_type(region, Type::CONTROL); 421 region = _gvn.transform(region); 422 set_control (region); 423 return region; 424 } 425 426 // sentinel value for the target bci to mark never taken branches 427 // (according to profiling) 428 static const int never_reached = INT_MAX; 429 430 //------------------------------helper for tableswitch------------------------- 431 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 432 // True branch, use existing map info 433 { PreserveJVMState pjvms(this); 434 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 435 set_control( iftrue ); 436 if (unc) { 437 repush_if_args(); 438 uncommon_trap(Deoptimization::Reason_unstable_if, 439 Deoptimization::Action_reinterpret, 440 NULL, 441 "taken always"); 442 } else { 443 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 444 profile_switch_case(prof_table_index); 445 merge_new_path(dest_bci_if_true); 446 } 447 } 448 449 // False branch 450 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 451 set_control( iffalse ); 452 } 453 454 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 455 // True branch, use existing map info 456 { PreserveJVMState pjvms(this); 457 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 458 set_control( iffalse ); 459 if (unc) { 460 repush_if_args(); 461 uncommon_trap(Deoptimization::Reason_unstable_if, 462 Deoptimization::Action_reinterpret, 463 NULL, 464 "taken never"); 465 } else { 466 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 467 profile_switch_case(prof_table_index); 468 merge_new_path(dest_bci_if_true); 469 } 470 } 471 472 // False branch 473 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 474 set_control( iftrue ); 475 } 476 477 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index, bool unc) { 478 // False branch, use existing map and control() 479 if (unc) { 480 repush_if_args(); 481 uncommon_trap(Deoptimization::Reason_unstable_if, 482 Deoptimization::Action_reinterpret, 483 NULL, 484 "taken never"); 485 } else { 486 assert(dest_bci != never_reached, "inconsistent dest"); 487 profile_switch_case(prof_table_index); 488 merge_new_path(dest_bci); 489 } 490 } 491 492 493 extern "C" { 494 static int jint_cmp(const void *i, const void *j) { 495 int a = *(jint *)i; 496 int b = *(jint *)j; 497 return a > b ? 1 : a < b ? -1 : 0; 498 } 499 } 500 501 502 // Default value for methodData switch indexing. Must be a negative value to avoid 503 // conflict with any legal switch index. 504 #define NullTableIndex -1 505 506 class SwitchRange : public StackObj { 507 // a range of integers coupled with a bci destination 508 jint _lo; // inclusive lower limit 509 jint _hi; // inclusive upper limit 510 int _dest; 511 int _table_index; // index into method data table 512 float _cnt; // how many times this range was hit according to profiling 513 514 public: 515 jint lo() const { return _lo; } 516 jint hi() const { return _hi; } 517 int dest() const { return _dest; } 518 int table_index() const { return _table_index; } 519 bool is_singleton() const { return _lo == _hi; } 520 float cnt() const { return _cnt; } 521 522 void setRange(jint lo, jint hi, int dest, int table_index, float cnt) { 523 assert(lo <= hi, "must be a non-empty range"); 524 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; _cnt = cnt; 525 assert(_cnt >= 0, ""); 526 } 527 bool adjoinRange(jint lo, jint hi, int dest, int table_index, float cnt, bool trim_ranges) { 528 assert(lo <= hi, "must be a non-empty range"); 529 if (lo == _hi+1 && table_index == _table_index) { 530 // see merge_ranges() comment below 531 if (trim_ranges) { 532 if (cnt == 0) { 533 if (_cnt != 0) { 534 return false; 535 } 536 if (dest != _dest) { 537 _dest = never_reached; 538 } 539 } else { 540 if (_cnt == 0) { 541 return false; 542 } 543 if (dest != _dest) { 544 return false; 545 } 546 } 547 } else { 548 if (dest != _dest) { 549 return false; 550 } 551 } 552 _hi = hi; 553 _cnt += cnt; 554 return true; 555 } 556 return false; 557 } 558 559 void set (jint value, int dest, int table_index, float cnt) { 560 setRange(value, value, dest, table_index, cnt); 561 } 562 bool adjoin(jint value, int dest, int table_index, float cnt, bool trim_ranges) { 563 return adjoinRange(value, value, dest, table_index, cnt, trim_ranges); 564 } 565 bool adjoin(SwitchRange& other) { 566 return adjoinRange(other._lo, other._hi, other._dest, other._table_index, other._cnt, false); 567 } 568 569 void print() { 570 if (is_singleton()) 571 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 572 else if (lo() == min_jint) 573 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 574 else if (hi() == max_jint) 575 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 576 else 577 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 578 } 579 }; 580 581 // We try to minimize the number of ranges and the size of the taken 582 // ones using profiling data. When ranges are created, 583 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 584 // if both were never hit or both were hit to build longer unreached 585 // ranges. Here, we now merge adjoining ranges with the same 586 // destination and finally set destination of unreached ranges to the 587 // special value never_reached because it can help minimize the number 588 // of tests that are necessary. 589 // 590 // For instance: 591 // [0, 1] to target1 sometimes taken 592 // [1, 2] to target1 never taken 593 // [2, 3] to target2 never taken 594 // would lead to: 595 // [0, 1] to target1 sometimes taken 596 // [1, 3] never taken 597 // 598 // (first 2 ranges to target1 are not merged) 599 static void merge_ranges(SwitchRange* ranges, int& rp) { 600 if (rp == 0) { 601 return; 602 } 603 int shift = 0; 604 for (int j = 0; j < rp; j++) { 605 SwitchRange& r1 = ranges[j-shift]; 606 SwitchRange& r2 = ranges[j+1]; 607 if (r1.adjoin(r2)) { 608 shift++; 609 } else if (shift > 0) { 610 ranges[j+1-shift] = r2; 611 } 612 } 613 rp -= shift; 614 for (int j = 0; j <= rp; j++) { 615 SwitchRange& r = ranges[j]; 616 if (r.cnt() == 0 && r.dest() != never_reached) { 617 r.setRange(r.lo(), r.hi(), never_reached, r.table_index(), r.cnt()); 618 } 619 } 620 } 621 622 //-------------------------------do_tableswitch-------------------------------- 623 void Parse::do_tableswitch() { 624 Node* lookup = pop(); 625 // Get information about tableswitch 626 int default_dest = iter().get_dest_table(0); 627 int lo_index = iter().get_int_table(1); 628 int hi_index = iter().get_int_table(2); 629 int len = hi_index - lo_index + 1; 630 631 if (len < 1) { 632 // If this is a backward branch, add safepoint 633 maybe_add_safepoint(default_dest); 634 merge(default_dest); 635 return; 636 } 637 638 ciMethodData* methodData = method()->method_data(); 639 ciMultiBranchData* profile = NULL; 640 if (methodData->is_mature() && UseSwitchProfiling) { 641 ciProfileData* data = methodData->bci_to_data(bci()); 642 if (data != NULL && data->is_MultiBranchData()) { 643 profile = (ciMultiBranchData*)data; 644 } 645 } 646 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 647 648 // generate decision tree, using trichotomy when possible 649 int rnum = len+2; 650 bool makes_backward_branch = false; 651 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 652 int rp = -1; 653 if (lo_index != min_jint) { 654 uint cnt = 1; 655 if (profile != NULL) { 656 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1); 657 } 658 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex, cnt); 659 } 660 for (int j = 0; j < len; j++) { 661 jint match_int = lo_index+j; 662 int dest = iter().get_dest_table(j+3); 663 makes_backward_branch |= (dest <= bci()); 664 int table_index = method_data_update() ? j : NullTableIndex; 665 uint cnt = 1; 666 if (profile != NULL) { 667 cnt = profile->count_at(j); 668 } 669 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 670 ranges[++rp].set(match_int, dest, table_index, cnt); 671 } 672 } 673 jint highest = lo_index+(len-1); 674 assert(ranges[rp].hi() == highest, ""); 675 if (highest != max_jint) { 676 uint cnt = 1; 677 if (profile != NULL) { 678 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1); 679 } 680 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, cnt, trim_ranges)) { 681 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, cnt); 682 } 683 } 684 assert(rp < len+2, "not too many ranges"); 685 686 if (trim_ranges) { 687 merge_ranges(ranges, rp); 688 } 689 690 // Safepoint in case if backward branch observed 691 if( makes_backward_branch && UseLoopSafepoints ) 692 add_safepoint(); 693 694 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 695 } 696 697 698 //------------------------------do_lookupswitch-------------------------------- 699 void Parse::do_lookupswitch() { 700 Node *lookup = pop(); // lookup value 701 // Get information about lookupswitch 702 int default_dest = iter().get_dest_table(0); 703 int len = iter().get_int_table(1); 704 705 if (len < 1) { // If this is a backward branch, add safepoint 706 maybe_add_safepoint(default_dest); 707 merge(default_dest); 708 return; 709 } 710 711 ciMethodData* methodData = method()->method_data(); 712 ciMultiBranchData* profile = NULL; 713 if (methodData->is_mature() && UseSwitchProfiling) { 714 ciProfileData* data = methodData->bci_to_data(bci()); 715 if (data != NULL && data->is_MultiBranchData()) { 716 profile = (ciMultiBranchData*)data; 717 } 718 } 719 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 720 721 // generate decision tree, using trichotomy when possible 722 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 723 { 724 for (int j = 0; j < len; j++) { 725 table[3*j+0] = iter().get_int_table(2+2*j); 726 table[3*j+1] = iter().get_dest_table(2+2*j+1); 727 table[3*j+2] = profile == NULL ? 1 : profile->count_at(j); 728 } 729 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 730 } 731 732 float defaults = 0; 733 jint prev = min_jint; 734 for (int j = 0; j < len; j++) { 735 jint match_int = table[3*j+0]; 736 if (match_int != prev) { 737 defaults += (float)match_int - prev; 738 } 739 prev = match_int+1; 740 } 741 if (prev-1 != max_jint) { 742 defaults += (float)max_jint - prev + 1; 743 } 744 float default_cnt = 1; 745 if (profile != NULL) { 746 default_cnt = profile->default_count()/defaults; 747 } 748 749 int rnum = len*2+1; 750 bool makes_backward_branch = false; 751 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 752 int rp = -1; 753 for (int j = 0; j < len; j++) { 754 jint match_int = table[3*j+0]; 755 int dest = table[3*j+1]; 756 int cnt = table[3*j+2]; 757 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 758 int table_index = method_data_update() ? j : NullTableIndex; 759 makes_backward_branch |= (dest <= bci()); 760 float c = default_cnt * ((float)match_int - next_lo); 761 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, NullTableIndex, c, trim_ranges))) { 762 assert(default_dest != never_reached, "sentinel value for dead destinations"); 763 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex, c); 764 } 765 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 766 assert(dest != never_reached, "sentinel value for dead destinations"); 767 ranges[++rp].set(match_int, dest, table_index, cnt); 768 } 769 } 770 jint highest = table[3*(len-1)]; 771 assert(ranges[rp].hi() == highest, ""); 772 if (highest != max_jint && 773 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest), trim_ranges)) { 774 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest)); 775 } 776 assert(rp < rnum, "not too many ranges"); 777 778 if (trim_ranges) { 779 merge_ranges(ranges, rp); 780 } 781 782 // Safepoint in case backward branch observed 783 if (makes_backward_branch && UseLoopSafepoints) 784 add_safepoint(); 785 786 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 787 } 788 789 static float if_prob(float taken_cnt, float total_cnt) { 790 assert(taken_cnt <= total_cnt, ""); 791 if (total_cnt == 0) { 792 return PROB_FAIR; 793 } 794 float p = taken_cnt / total_cnt; 795 return MIN2(MAX2(p, PROB_MIN), PROB_MAX); 796 } 797 798 static float if_cnt(float cnt) { 799 if (cnt == 0) { 800 return COUNT_UNKNOWN; 801 } 802 return cnt; 803 } 804 805 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 806 float total_cnt = 0; 807 for (SwitchRange* sr = lo; sr <= hi; sr++) { 808 total_cnt += sr->cnt(); 809 } 810 return total_cnt; 811 } 812 813 class SwitchRanges : public ResourceObj { 814 public: 815 SwitchRange* _lo; 816 SwitchRange* _hi; 817 SwitchRange* _mid; 818 float _cost; 819 820 enum { 821 Start, 822 LeftDone, 823 RightDone, 824 Done 825 } _state; 826 827 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 828 : _lo(lo), _hi(hi), _mid(NULL), 829 _cost(0), _state(Start) { 830 } 831 832 SwitchRanges() 833 : _lo(NULL), _hi(NULL), _mid(NULL), 834 _cost(0), _state(Start) {} 835 }; 836 837 // Estimate cost of performing a binary search on lo..hi 838 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 839 GrowableArray<SwitchRanges> tree; 840 SwitchRanges root(lo, hi); 841 tree.push(root); 842 843 float cost = 0; 844 do { 845 SwitchRanges& r = *tree.adr_at(tree.length()-1); 846 if (r._hi != r._lo) { 847 if (r._mid == NULL) { 848 float r_cnt = sum_of_cnts(r._lo, r._hi); 849 850 if (r_cnt == 0) { 851 tree.pop(); 852 cost = 0; 853 continue; 854 } 855 856 SwitchRange* mid = NULL; 857 mid = r._lo; 858 for (float cnt = 0; ; ) { 859 assert(mid <= r._hi, "out of bounds"); 860 cnt += mid->cnt(); 861 if (cnt > r_cnt / 2) { 862 break; 863 } 864 mid++; 865 } 866 assert(mid <= r._hi, "out of bounds"); 867 r._mid = mid; 868 r._cost = r_cnt / total_cnt; 869 } 870 r._cost += cost; 871 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 872 cost = 0; 873 r._state = SwitchRanges::LeftDone; 874 tree.push(SwitchRanges(r._lo, r._mid-1)); 875 } else if (r._state < SwitchRanges::RightDone) { 876 cost = 0; 877 r._state = SwitchRanges::RightDone; 878 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 879 } else { 880 tree.pop(); 881 cost = r._cost; 882 } 883 } else { 884 tree.pop(); 885 cost = r._cost; 886 } 887 } while (tree.length() > 0); 888 889 890 return cost; 891 } 892 893 // It sometimes pays off to test most common ranges before the binary search 894 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 895 uint nr = hi - lo + 1; 896 float total_cnt = sum_of_cnts(lo, hi); 897 898 float min = compute_tree_cost(lo, hi, total_cnt); 899 float extra = 1; 900 float sub = 0; 901 902 SwitchRange* array1 = lo; 903 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 904 905 SwitchRange* ranges = NULL; 906 907 while (nr >= 2) { 908 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 909 ranges = (lo == array1) ? array2 : array1; 910 911 // Find highest frequency range 912 SwitchRange* candidate = lo; 913 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 914 if (sr->cnt() > candidate->cnt()) { 915 candidate = sr; 916 } 917 } 918 SwitchRange most_freq = *candidate; 919 if (most_freq.cnt() == 0) { 920 break; 921 } 922 923 // Copy remaining ranges into another array 924 int shift = 0; 925 for (uint i = 0; i < nr; i++) { 926 SwitchRange* sr = &lo[i]; 927 if (sr != candidate) { 928 ranges[i-shift] = *sr; 929 } else { 930 shift++; 931 if (i > 0 && i < nr-1) { 932 SwitchRange prev = lo[i-1]; 933 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.table_index(), prev.cnt()); 934 if (prev.adjoin(lo[i+1])) { 935 shift++; 936 i++; 937 } 938 ranges[i-shift] = prev; 939 } 940 } 941 } 942 nr -= shift; 943 944 // Evaluate cost of testing the most common range and performing a 945 // binary search on the other ranges 946 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 947 if (cost >= min) { 948 break; 949 } 950 // swap arrays 951 lo = &ranges[0]; 952 hi = &ranges[nr-1]; 953 954 // It pays off: emit the test for the most common range 955 assert(most_freq.cnt() > 0, "must be taken"); 956 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 957 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo()))); 958 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 959 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 960 jump_if_true_fork(iff, most_freq.dest(), most_freq.table_index(), false); 961 962 sub += most_freq.cnt() / total_cnt; 963 extra += 1 - sub; 964 min = cost; 965 } 966 } 967 968 //----------------------------create_jump_tables------------------------------- 969 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 970 // Are jumptables enabled 971 if (!UseJumpTables) return false; 972 973 // Are jumptables supported 974 if (!Matcher::has_match_rule(Op_Jump)) return false; 975 976 // Don't make jump table if profiling 977 if (method_data_update()) return false; 978 979 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 980 981 // Decide if a guard is needed to lop off big ranges at either (or 982 // both) end(s) of the input set. We'll call this the default target 983 // even though we can't be sure that it is the true "default". 984 985 bool needs_guard = false; 986 int default_dest; 987 int64_t total_outlier_size = 0; 988 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 989 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 990 991 if (lo->dest() == hi->dest()) { 992 total_outlier_size = hi_size + lo_size; 993 default_dest = lo->dest(); 994 } else if (lo_size > hi_size) { 995 total_outlier_size = lo_size; 996 default_dest = lo->dest(); 997 } else { 998 total_outlier_size = hi_size; 999 default_dest = hi->dest(); 1000 } 1001 1002 float total = sum_of_cnts(lo, hi); 1003 float cost = compute_tree_cost(lo, hi, total); 1004 1005 // If a guard test will eliminate very sparse end ranges, then 1006 // it is worth the cost of an extra jump. 1007 float trimmed_cnt = 0; 1008 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 1009 needs_guard = true; 1010 if (default_dest == lo->dest()) { 1011 trimmed_cnt += lo->cnt(); 1012 lo++; 1013 } 1014 if (default_dest == hi->dest()) { 1015 trimmed_cnt += hi->cnt(); 1016 hi--; 1017 } 1018 } 1019 1020 // Find the total number of cases and ranges 1021 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 1022 int num_range = hi - lo + 1; 1023 1024 // Don't create table if: too large, too small, or too sparse. 1025 if (num_cases > MaxJumpTableSize) 1026 return false; 1027 if (UseSwitchProfiling) { 1028 // MinJumpTableSize is set so with a well balanced binary tree, 1029 // when the number of ranges is MinJumpTableSize, it's cheaper to 1030 // go through a JumpNode that a tree of IfNodes. Average cost of a 1031 // tree of IfNodes with MinJumpTableSize is 1032 // log2f(MinJumpTableSize) comparisons. So if the cost computed 1033 // from profile data is less than log2f(MinJumpTableSize) then 1034 // going with the binary search is cheaper. 1035 if (cost < log2f(MinJumpTableSize)) { 1036 return false; 1037 } 1038 } else { 1039 if (num_cases < MinJumpTableSize) 1040 return false; 1041 } 1042 if (num_cases > (MaxJumpTableSparseness * num_range)) 1043 return false; 1044 1045 // Normalize table lookups to zero 1046 int lowval = lo->lo(); 1047 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 1048 1049 // Generate a guard to protect against input keyvals that aren't 1050 // in the switch domain. 1051 if (needs_guard) { 1052 Node* size = _gvn.intcon(num_cases); 1053 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 1054 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 1055 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 1056 jump_if_true_fork(iff, default_dest, NullTableIndex, trim_ranges && trimmed_cnt == 0); 1057 1058 total -= trimmed_cnt; 1059 } 1060 1061 // Create an ideal node JumpTable that has projections 1062 // of all possible ranges for a switch statement 1063 // The key_val input must be converted to a pointer offset and scaled. 1064 // Compare Parse::array_addressing above. 1065 1066 // Clean the 32-bit int into a real 64-bit offset. 1067 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 1068 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 1069 // Make I2L conversion control dependent to prevent it from 1070 // floating above the range check during loop optimizations. 1071 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 1072 1073 // Shift the value by wordsize so we have an index into the table, rather 1074 // than a switch value 1075 Node *shiftWord = _gvn.MakeConX(wordSize); 1076 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 1077 1078 // Create the JumpNode 1079 Arena* arena = C->comp_arena(); 1080 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 1081 int i = 0; 1082 if (total == 0) { 1083 for (SwitchRange* r = lo; r <= hi; r++) { 1084 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 1085 probs[i] = 1.0F / num_cases; 1086 } 1087 } 1088 } else { 1089 for (SwitchRange* r = lo; r <= hi; r++) { 1090 float prob = r->cnt()/total; 1091 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 1092 probs[i] = prob / (r->hi() - r->lo() + 1); 1093 } 1094 } 1095 } 1096 1097 ciMethodData* methodData = method()->method_data(); 1098 ciMultiBranchData* profile = NULL; 1099 if (methodData->is_mature()) { 1100 ciProfileData* data = methodData->bci_to_data(bci()); 1101 if (data != NULL && data->is_MultiBranchData()) { 1102 profile = (ciMultiBranchData*)data; 1103 } 1104 } 1105 1106 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total)); 1107 1108 // These are the switch destinations hanging off the jumpnode 1109 i = 0; 1110 for (SwitchRange* r = lo; r <= hi; r++) { 1111 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 1112 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 1113 { 1114 PreserveJVMState pjvms(this); 1115 set_control(input); 1116 jump_if_always_fork(r->dest(), r->table_index(), trim_ranges && r->cnt() == 0); 1117 } 1118 } 1119 } 1120 assert(i == num_cases, "miscount of cases"); 1121 stop_and_kill_map(); // no more uses for this JVMS 1122 return true; 1123 } 1124 1125 //----------------------------jump_switch_ranges------------------------------- 1126 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 1127 Block* switch_block = block(); 1128 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 1129 1130 if (switch_depth == 0) { 1131 // Do special processing for the top-level call. 1132 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 1133 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 1134 1135 // Decrement pred-numbers for the unique set of nodes. 1136 #ifdef ASSERT 1137 if (!trim_ranges) { 1138 // Ensure that the block's successors are a (duplicate-free) set. 1139 int successors_counted = 0; // block occurrences in [hi..lo] 1140 int unique_successors = switch_block->num_successors(); 1141 for (int i = 0; i < unique_successors; i++) { 1142 Block* target = switch_block->successor_at(i); 1143 1144 // Check that the set of successors is the same in both places. 1145 int successors_found = 0; 1146 for (SwitchRange* p = lo; p <= hi; p++) { 1147 if (p->dest() == target->start()) successors_found++; 1148 } 1149 assert(successors_found > 0, "successor must be known"); 1150 successors_counted += successors_found; 1151 } 1152 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 1153 } 1154 #endif 1155 1156 // Maybe prune the inputs, based on the type of key_val. 1157 jint min_val = min_jint; 1158 jint max_val = max_jint; 1159 const TypeInt* ti = key_val->bottom_type()->isa_int(); 1160 if (ti != NULL) { 1161 min_val = ti->_lo; 1162 max_val = ti->_hi; 1163 assert(min_val <= max_val, "invalid int type"); 1164 } 1165 while (lo->hi() < min_val) { 1166 lo++; 1167 } 1168 if (lo->lo() < min_val) { 1169 lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index(), lo->cnt()); 1170 } 1171 while (hi->lo() > max_val) { 1172 hi--; 1173 } 1174 if (hi->hi() > max_val) { 1175 hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index(), hi->cnt()); 1176 } 1177 1178 linear_search_switch_ranges(key_val, lo, hi); 1179 } 1180 1181 #ifndef PRODUCT 1182 if (switch_depth == 0) { 1183 _max_switch_depth = 0; 1184 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 1185 } 1186 #endif 1187 1188 assert(lo <= hi, "must be a non-empty set of ranges"); 1189 if (lo == hi) { 1190 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1191 } else { 1192 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 1193 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 1194 1195 if (create_jump_tables(key_val, lo, hi)) return; 1196 1197 SwitchRange* mid = NULL; 1198 float total_cnt = sum_of_cnts(lo, hi); 1199 1200 int nr = hi - lo + 1; 1201 if (UseSwitchProfiling) { 1202 // Don't keep the binary search tree balanced: pick up mid point 1203 // that split frequencies in half. 1204 float cnt = 0; 1205 for (SwitchRange* sr = lo; sr <= hi; sr++) { 1206 cnt += sr->cnt(); 1207 if (cnt >= total_cnt / 2) { 1208 mid = sr; 1209 break; 1210 } 1211 } 1212 } else { 1213 mid = lo + nr/2; 1214 1215 // if there is an easy choice, pivot at a singleton: 1216 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1217 1218 assert(lo < mid && mid <= hi, "good pivot choice"); 1219 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1220 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1221 } 1222 1223 1224 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1225 1226 if (mid->is_singleton()) { 1227 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1228 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index(), trim_ranges && mid->cnt() == 0); 1229 1230 // Special Case: If there are exactly three ranges, and the high 1231 // and low range each go to the same place, omit the "gt" test, 1232 // since it will not discriminate anything. 1233 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1234 1235 // if there is a higher range, test for it and process it: 1236 if (mid < hi && !eq_test_only) { 1237 // two comparisons of same values--should enable 1 test for 2 branches 1238 // Use BoolTest::le instead of BoolTest::gt 1239 float cnt = sum_of_cnts(lo, mid-1); 1240 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le, if_prob(cnt, total_cnt), if_cnt(cnt)); 1241 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 1242 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 1243 { PreserveJVMState pjvms(this); 1244 set_control(iffalse); 1245 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1246 } 1247 set_control(iftrue); 1248 } 1249 1250 } else { 1251 // mid is a range, not a singleton, so treat mid..hi as a unit 1252 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1253 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, mid == lo ? BoolTest::gt : BoolTest::ge, if_prob(cnt, total_cnt), if_cnt(cnt)); 1254 1255 // if there is a higher range, test for it and process it: 1256 if (mid == hi) { 1257 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index(), trim_ranges && cnt == 0); 1258 } else { 1259 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1260 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1261 { PreserveJVMState pjvms(this); 1262 set_control(iftrue); 1263 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1264 } 1265 set_control(iffalse); 1266 } 1267 } 1268 1269 // in any case, process the lower range 1270 if (mid == lo) { 1271 if (mid->is_singleton()) { 1272 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1273 } else { 1274 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1275 } 1276 } else { 1277 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1278 } 1279 } 1280 1281 // Decrease pred_count for each successor after all is done. 1282 if (switch_depth == 0) { 1283 int unique_successors = switch_block->num_successors(); 1284 for (int i = 0; i < unique_successors; i++) { 1285 Block* target = switch_block->successor_at(i); 1286 // Throw away the pre-allocated path for each unique successor. 1287 target->next_path_num(); 1288 } 1289 } 1290 1291 #ifndef PRODUCT 1292 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1293 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1294 SwitchRange* r; 1295 int nsing = 0; 1296 for( r = lo; r <= hi; r++ ) { 1297 if( r->is_singleton() ) nsing++; 1298 } 1299 tty->print(">>> "); 1300 _method->print_short_name(); 1301 tty->print_cr(" switch decision tree"); 1302 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1303 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1304 if (_max_switch_depth > _est_switch_depth) { 1305 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1306 } 1307 tty->print(" "); 1308 for( r = lo; r <= hi; r++ ) { 1309 r->print(); 1310 } 1311 tty->cr(); 1312 } 1313 #endif 1314 } 1315 1316 void Parse::modf() { 1317 Node *f2 = pop(); 1318 Node *f1 = pop(); 1319 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1320 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1321 "frem", NULL, //no memory effects 1322 f1, f2); 1323 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1324 1325 push(res); 1326 } 1327 1328 void Parse::modd() { 1329 Node *d2 = pop_pair(); 1330 Node *d1 = pop_pair(); 1331 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1332 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1333 "drem", NULL, //no memory effects 1334 d1, top(), d2, top()); 1335 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1336 1337 #ifdef ASSERT 1338 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 1339 assert(res_top == top(), "second value must be top"); 1340 #endif 1341 1342 push_pair(res_d); 1343 } 1344 1345 void Parse::l2f() { 1346 Node* f2 = pop(); 1347 Node* f1 = pop(); 1348 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1349 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1350 "l2f", NULL, //no memory effects 1351 f1, f2); 1352 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1353 1354 push(res); 1355 } 1356 1357 void Parse::do_irem() { 1358 // Must keep both values on the expression-stack during null-check 1359 zero_check_int(peek()); 1360 // Compile-time detect of null-exception? 1361 if (stopped()) return; 1362 1363 Node* b = pop(); 1364 Node* a = pop(); 1365 1366 const Type *t = _gvn.type(b); 1367 if (t != Type::TOP) { 1368 const TypeInt *ti = t->is_int(); 1369 if (ti->is_con()) { 1370 int divisor = ti->get_con(); 1371 // check for positive power of 2 1372 if (divisor > 0 && 1373 (divisor & ~(divisor-1)) == divisor) { 1374 // yes ! 1375 Node *mask = _gvn.intcon((divisor - 1)); 1376 // Sigh, must handle negative dividends 1377 Node *zero = _gvn.intcon(0); 1378 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN); 1379 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 1380 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 1381 Node *reg = jump_if_join(ift, iff); 1382 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 1383 // Negative path; negate/and/negate 1384 Node *neg = _gvn.transform( new SubINode(zero, a) ); 1385 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 1386 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 1387 phi->init_req(1, negn); 1388 // Fast positive case 1389 Node *andx = _gvn.transform( new AndINode(a, mask) ); 1390 phi->init_req(2, andx); 1391 // Push the merge 1392 push( _gvn.transform(phi) ); 1393 return; 1394 } 1395 } 1396 } 1397 // Default case 1398 push( _gvn.transform( new ModINode(control(),a,b) ) ); 1399 } 1400 1401 // Handle jsr and jsr_w bytecode 1402 void Parse::do_jsr() { 1403 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1404 1405 // Store information about current state, tagged with new _jsr_bci 1406 int return_bci = iter().next_bci(); 1407 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1408 1409 // Update method data 1410 profile_taken_branch(jsr_bci); 1411 1412 // The way we do things now, there is only one successor block 1413 // for the jsr, because the target code is cloned by ciTypeFlow. 1414 Block* target = successor_for_bci(jsr_bci); 1415 1416 // What got pushed? 1417 const Type* ret_addr = target->peek(); 1418 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1419 1420 // Effect on jsr on stack 1421 push(_gvn.makecon(ret_addr)); 1422 1423 // Flow to the jsr. 1424 merge(jsr_bci); 1425 } 1426 1427 // Handle ret bytecode 1428 void Parse::do_ret() { 1429 // Find to whom we return. 1430 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1431 Block* target = block()->successor_at(0); 1432 assert(!target->is_ready(), "our arrival must be expected"); 1433 profile_ret(target->flow()->start()); 1434 int pnum = target->next_path_num(); 1435 merge_common(target, pnum); 1436 } 1437 1438 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1439 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1440 // Only ::eq and ::ne are supported for profile injection. 1441 return false; 1442 } 1443 if (test->is_Cmp() && 1444 test->in(1)->Opcode() == Op_ProfileBoolean) { 1445 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1446 int false_cnt = profile->false_count(); 1447 int true_cnt = profile->true_count(); 1448 1449 // Counts matching depends on the actual test operation (::eq or ::ne). 1450 // No need to scale the counts because profile injection was designed 1451 // to feed exact counts into VM. 1452 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1453 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1454 1455 profile->consume(); 1456 return true; 1457 } 1458 return false; 1459 } 1460 //--------------------------dynamic_branch_prediction-------------------------- 1461 // Try to gather dynamic branch prediction behavior. Return a probability 1462 // of the branch being taken and set the "cnt" field. Returns a -1.0 1463 // if we need to use static prediction for some reason. 1464 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1465 ResourceMark rm; 1466 1467 cnt = COUNT_UNKNOWN; 1468 1469 int taken = 0; 1470 int not_taken = 0; 1471 1472 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1473 1474 if (use_mdo) { 1475 // Use MethodData information if it is available 1476 // FIXME: free the ProfileData structure 1477 ciMethodData* methodData = method()->method_data(); 1478 if (!methodData->is_mature()) return PROB_UNKNOWN; 1479 ciProfileData* data = methodData->bci_to_data(bci()); 1480 if (data == NULL) { 1481 return PROB_UNKNOWN; 1482 } 1483 if (!data->is_JumpData()) return PROB_UNKNOWN; 1484 1485 // get taken and not taken values 1486 taken = data->as_JumpData()->taken(); 1487 not_taken = 0; 1488 if (data->is_BranchData()) { 1489 not_taken = data->as_BranchData()->not_taken(); 1490 } 1491 1492 // scale the counts to be commensurate with invocation counts: 1493 taken = method()->scale_count(taken); 1494 not_taken = method()->scale_count(not_taken); 1495 } 1496 1497 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1498 // We also check that individual counters are positive first, otherwise the sum can become positive. 1499 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 1500 if (C->log() != NULL) { 1501 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1502 } 1503 return PROB_UNKNOWN; 1504 } 1505 1506 // Compute frequency that we arrive here 1507 float sum = taken + not_taken; 1508 // Adjust, if this block is a cloned private block but the 1509 // Jump counts are shared. Taken the private counts for 1510 // just this path instead of the shared counts. 1511 if( block()->count() > 0 ) 1512 sum = block()->count(); 1513 cnt = sum / FreqCountInvocations; 1514 1515 // Pin probability to sane limits 1516 float prob; 1517 if( !taken ) 1518 prob = (0+PROB_MIN) / 2; 1519 else if( !not_taken ) 1520 prob = (1+PROB_MAX) / 2; 1521 else { // Compute probability of true path 1522 prob = (float)taken / (float)(taken + not_taken); 1523 if (prob > PROB_MAX) prob = PROB_MAX; 1524 if (prob < PROB_MIN) prob = PROB_MIN; 1525 } 1526 1527 assert((cnt > 0.0f) && (prob > 0.0f), 1528 "Bad frequency assignment in if"); 1529 1530 if (C->log() != NULL) { 1531 const char* prob_str = NULL; 1532 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1533 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1534 char prob_str_buf[30]; 1535 if (prob_str == NULL) { 1536 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob); 1537 prob_str = prob_str_buf; 1538 } 1539 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1540 iter().get_dest(), taken, not_taken, cnt, prob_str); 1541 } 1542 return prob; 1543 } 1544 1545 //-----------------------------branch_prediction------------------------------- 1546 float Parse::branch_prediction(float& cnt, 1547 BoolTest::mask btest, 1548 int target_bci, 1549 Node* test) { 1550 float prob = dynamic_branch_prediction(cnt, btest, test); 1551 // If prob is unknown, switch to static prediction 1552 if (prob != PROB_UNKNOWN) return prob; 1553 1554 prob = PROB_FAIR; // Set default value 1555 if (btest == BoolTest::eq) // Exactly equal test? 1556 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1557 else if (btest == BoolTest::ne) 1558 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1559 1560 // If this is a conditional test guarding a backwards branch, 1561 // assume its a loop-back edge. Make it a likely taken branch. 1562 if (target_bci < bci()) { 1563 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1564 // Since it's an OSR, we probably have profile data, but since 1565 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1566 // Let's make a special check here for completely zero counts. 1567 ciMethodData* methodData = method()->method_data(); 1568 if (!methodData->is_empty()) { 1569 ciProfileData* data = methodData->bci_to_data(bci()); 1570 // Only stop for truly zero counts, which mean an unknown part 1571 // of the OSR-ed method, and we want to deopt to gather more stats. 1572 // If you have ANY counts, then this loop is simply 'cold' relative 1573 // to the OSR loop. 1574 if (data == NULL || 1575 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1576 // This is the only way to return PROB_UNKNOWN: 1577 return PROB_UNKNOWN; 1578 } 1579 } 1580 } 1581 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1582 } 1583 1584 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1585 return prob; 1586 } 1587 1588 // The magic constants are chosen so as to match the output of 1589 // branch_prediction() when the profile reports a zero taken count. 1590 // It is important to distinguish zero counts unambiguously, because 1591 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1592 // very small but nonzero probabilities, which if confused with zero 1593 // counts would keep the program recompiling indefinitely. 1594 bool Parse::seems_never_taken(float prob) const { 1595 return prob < PROB_MIN; 1596 } 1597 1598 // True if the comparison seems to be the kind that will not change its 1599 // statistics from true to false. See comments in adjust_map_after_if. 1600 // This question is only asked along paths which are already 1601 // classifed as untaken (by seems_never_taken), so really, 1602 // if a path is never taken, its controlling comparison is 1603 // already acting in a stable fashion. If the comparison 1604 // seems stable, we will put an expensive uncommon trap 1605 // on the untaken path. 1606 bool Parse::seems_stable_comparison() const { 1607 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 1608 return false; 1609 } 1610 return true; 1611 } 1612 1613 //-------------------------------repush_if_args-------------------------------- 1614 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1615 inline int Parse::repush_if_args() { 1616 if (PrintOpto && WizardMode) { 1617 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1618 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1619 method()->print_name(); tty->cr(); 1620 } 1621 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1622 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1623 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1624 assert(argument(0) != NULL, "must exist"); 1625 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 1626 inc_sp(bc_depth); 1627 return bc_depth; 1628 } 1629 1630 //----------------------------------do_ifnull---------------------------------- 1631 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1632 int target_bci = iter().get_dest(); 1633 1634 Block* branch_block = successor_for_bci(target_bci); 1635 Block* next_block = successor_for_bci(iter().next_bci()); 1636 1637 float cnt; 1638 float prob = branch_prediction(cnt, btest, target_bci, c); 1639 if (prob == PROB_UNKNOWN) { 1640 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1641 if (PrintOpto && Verbose) { 1642 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1643 } 1644 repush_if_args(); // to gather stats on loop 1645 // We need to mark this branch as taken so that if we recompile we will 1646 // see that it is possible. In the tiered system the interpreter doesn't 1647 // do profiling and by the time we get to the lower tier from the interpreter 1648 // the path may be cold again. Make sure it doesn't look untaken 1649 profile_taken_branch(target_bci, !ProfileInterpreter); 1650 uncommon_trap(Deoptimization::Reason_unreached, 1651 Deoptimization::Action_reinterpret, 1652 NULL, "cold"); 1653 if (C->eliminate_boxing()) { 1654 // Mark the successor blocks as parsed 1655 branch_block->next_path_num(); 1656 next_block->next_path_num(); 1657 } 1658 return; 1659 } 1660 1661 NOT_PRODUCT(explicit_null_checks_inserted++); 1662 1663 // Generate real control flow 1664 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1665 1666 // Sanity check the probability value 1667 assert(prob > 0.0f,"Bad probability in Parser"); 1668 // Need xform to put node in hash table 1669 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1670 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1671 // True branch 1672 { PreserveJVMState pjvms(this); 1673 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1674 set_control(iftrue); 1675 1676 if (stopped()) { // Path is dead? 1677 NOT_PRODUCT(explicit_null_checks_elided++); 1678 if (C->eliminate_boxing()) { 1679 // Mark the successor block as parsed 1680 branch_block->next_path_num(); 1681 } 1682 } else { // Path is live. 1683 // Update method data 1684 profile_taken_branch(target_bci); 1685 adjust_map_after_if(btest, c, prob, branch_block); 1686 if (!stopped()) { 1687 merge(target_bci); 1688 } 1689 } 1690 } 1691 1692 // False branch 1693 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1694 set_control(iffalse); 1695 1696 if (stopped()) { // Path is dead? 1697 NOT_PRODUCT(explicit_null_checks_elided++); 1698 if (C->eliminate_boxing()) { 1699 // Mark the successor block as parsed 1700 next_block->next_path_num(); 1701 } 1702 } else { // Path is live. 1703 // Update method data 1704 profile_not_taken_branch(); 1705 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block); 1706 } 1707 } 1708 1709 //------------------------------------do_if------------------------------------ 1710 void Parse::do_if(BoolTest::mask btest, Node* c, bool new_path, Node** ctrl_taken) { 1711 int target_bci = iter().get_dest(); 1712 1713 Block* branch_block = successor_for_bci(target_bci); 1714 Block* next_block = successor_for_bci(iter().next_bci()); 1715 1716 float cnt; 1717 float prob = branch_prediction(cnt, btest, target_bci, c); 1718 float untaken_prob = 1.0 - prob; 1719 1720 if (prob == PROB_UNKNOWN) { 1721 if (PrintOpto && Verbose) { 1722 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1723 } 1724 repush_if_args(); // to gather stats on loop 1725 // We need to mark this branch as taken so that if we recompile we will 1726 // see that it is possible. In the tiered system the interpreter doesn't 1727 // do profiling and by the time we get to the lower tier from the interpreter 1728 // the path may be cold again. Make sure it doesn't look untaken 1729 profile_taken_branch(target_bci, !ProfileInterpreter); 1730 uncommon_trap(Deoptimization::Reason_unreached, 1731 Deoptimization::Action_reinterpret, 1732 NULL, "cold"); 1733 if (C->eliminate_boxing()) { 1734 // Mark the successor blocks as parsed 1735 branch_block->next_path_num(); 1736 next_block->next_path_num(); 1737 } 1738 return; 1739 } 1740 1741 // Sanity check the probability value 1742 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1743 1744 bool taken_if_true = true; 1745 // Convert BoolTest to canonical form: 1746 if (!BoolTest(btest).is_canonical()) { 1747 btest = BoolTest(btest).negate(); 1748 taken_if_true = false; 1749 // prob is NOT updated here; it remains the probability of the taken 1750 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1751 } 1752 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1753 1754 Node* tst0 = new BoolNode(c, btest); 1755 Node* tst = _gvn.transform(tst0); 1756 BoolTest::mask taken_btest = BoolTest::illegal; 1757 BoolTest::mask untaken_btest = BoolTest::illegal; 1758 1759 if (tst->is_Bool()) { 1760 // Refresh c from the transformed bool node, since it may be 1761 // simpler than the original c. Also re-canonicalize btest. 1762 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1763 // That can arise from statements like: if (x instanceof C) ... 1764 if (tst != tst0) { 1765 // Canonicalize one more time since transform can change it. 1766 btest = tst->as_Bool()->_test._test; 1767 if (!BoolTest(btest).is_canonical()) { 1768 // Reverse edges one more time... 1769 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1770 btest = tst->as_Bool()->_test._test; 1771 assert(BoolTest(btest).is_canonical(), "sanity"); 1772 taken_if_true = !taken_if_true; 1773 } 1774 c = tst->in(1); 1775 } 1776 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1777 taken_btest = taken_if_true ? btest : neg_btest; 1778 untaken_btest = taken_if_true ? neg_btest : btest; 1779 } 1780 1781 // Generate real control flow 1782 float true_prob = (taken_if_true ? prob : untaken_prob); 1783 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1784 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1785 Node* taken_branch = new IfTrueNode(iff); 1786 Node* untaken_branch = new IfFalseNode(iff); 1787 if (!taken_if_true) { // Finish conversion to canonical form 1788 Node* tmp = taken_branch; 1789 taken_branch = untaken_branch; 1790 untaken_branch = tmp; 1791 } 1792 1793 // Branch is taken: 1794 { PreserveJVMState pjvms(this); 1795 taken_branch = _gvn.transform(taken_branch); 1796 set_control(taken_branch); 1797 1798 if (stopped()) { 1799 if (C->eliminate_boxing() && !new_path) { 1800 // Mark the successor block as parsed (if we haven't created a new path) 1801 branch_block->next_path_num(); 1802 } 1803 } else { 1804 // Update method data 1805 profile_taken_branch(target_bci); 1806 adjust_map_after_if(taken_btest, c, prob, branch_block); 1807 if (!stopped()) { 1808 if (new_path) { 1809 // Merge by using a new path 1810 merge_new_path(target_bci); 1811 } else if (ctrl_taken != NULL) { 1812 // Don't merge but save taken branch to be wired by caller 1813 *ctrl_taken = control(); 1814 } else { 1815 merge(target_bci); 1816 } 1817 } 1818 } 1819 } 1820 1821 untaken_branch = _gvn.transform(untaken_branch); 1822 set_control(untaken_branch); 1823 1824 // Branch not taken. 1825 if (stopped() && ctrl_taken == NULL) { 1826 if (C->eliminate_boxing()) { 1827 // Mark the successor block as parsed (if caller does not re-wire control flow) 1828 next_block->next_path_num(); 1829 } 1830 } else { 1831 // Update method data 1832 profile_not_taken_branch(); 1833 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block); 1834 } 1835 } 1836 1837 void Parse::do_acmp(BoolTest::mask btest, Node* a, Node* b) { 1838 ciMethod* subst_method = ciEnv::current()->ValueBootstrapMethods_klass()->find_method(ciSymbol::isSubstitutable_name(), ciSymbol::object_object_boolean_signature()); 1839 // If current method is ValueBootstrapMethods::isSubstitutable(), 1840 // compile the acmp as a regular pointer comparison otherwise we 1841 // could call ValueBootstrapMethods::isSubstitutable() back 1842 if (ACmpOnValues == 0 || method() == subst_method) { 1843 Node* cmp = CmpP(a, b); 1844 cmp = optimize_cmp_with_klass(cmp); 1845 do_if(btest, cmp); 1846 return; 1847 } 1848 1849 if (ACmpOnValues == 3) { 1850 // Substitutability test 1851 if (a->is_ValueType()) { 1852 inc_sp(2); 1853 a = a->as_ValueType()->allocate(this, true)->get_oop(); 1854 dec_sp(2); 1855 } 1856 if (b->is_ValueType()) { 1857 inc_sp(2); 1858 b = b->as_ValueType()->allocate(this, true)->get_oop(); 1859 dec_sp(2); 1860 } 1861 1862 const TypeOopPtr* ta = _gvn.type(a)->isa_oopptr(); 1863 const TypeOopPtr* tb = _gvn.type(b)->isa_oopptr(); 1864 1865 if (ta == NULL || !ta->can_be_value_type_raw() || 1866 tb == NULL || !tb->can_be_value_type_raw()) { 1867 Node* cmp = CmpP(a, b); 1868 cmp = optimize_cmp_with_klass(cmp); 1869 do_if(btest, cmp); 1870 return; 1871 } 1872 1873 Node* cmp = CmpP(a, b); 1874 cmp = optimize_cmp_with_klass(cmp); 1875 Node* eq_region = NULL; 1876 if (btest == BoolTest::eq) { 1877 do_if(btest, cmp, true); 1878 if (stopped()) { 1879 return; 1880 } 1881 } else { 1882 assert(btest == BoolTest::ne, "only eq or ne"); 1883 Node* is_not_equal = NULL; 1884 eq_region = new RegionNode(3); 1885 { 1886 PreserveJVMState pjvms(this); 1887 do_if(btest, cmp, false, &is_not_equal); 1888 if (!stopped()) { 1889 eq_region->init_req(1, control()); 1890 } 1891 } 1892 if (is_not_equal == NULL || is_not_equal->is_top()) { 1893 record_for_igvn(eq_region); 1894 set_control(_gvn.transform(eq_region)); 1895 return; 1896 } 1897 set_control(is_not_equal); 1898 } 1899 // Pointers not equal, check for values 1900 Node* ne_region = new RegionNode(6); 1901 inc_sp(2); 1902 Node* null_ctl = top(); 1903 Node* not_null_a = null_check_oop(a, &null_ctl, !too_many_traps(Deoptimization::Reason_null_check), false, false); 1904 dec_sp(2); 1905 ne_region->init_req(1, null_ctl); 1906 if (stopped()) { 1907 record_for_igvn(ne_region); 1908 set_control(_gvn.transform(ne_region)); 1909 if (btest == BoolTest::ne) { 1910 { 1911 PreserveJVMState pjvms(this); 1912 int target_bci = iter().get_dest(); 1913 merge(target_bci); 1914 } 1915 record_for_igvn(eq_region); 1916 set_control(_gvn.transform(eq_region)); 1917 } 1918 return; 1919 } 1920 1921 Node* is_value = is_always_locked(not_null_a); 1922 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern); 1923 Node* is_value_cmp = _gvn.transform(new CmpXNode(is_value, value_mask)); 1924 Node* is_value_bol = _gvn.transform(new BoolNode(is_value_cmp, BoolTest::ne)); 1925 IfNode* is_value_iff = create_and_map_if(control(), is_value_bol, PROB_FAIR, COUNT_UNKNOWN); 1926 Node* not_value = _gvn.transform(new IfTrueNode(is_value_iff)); 1927 set_control(_gvn.transform(new IfFalseNode(is_value_iff))); 1928 ne_region->init_req(2, not_value); 1929 1930 // One of the 2 pointers refers to a value, check if both are of 1931 // the same class 1932 inc_sp(2); 1933 null_ctl = top(); 1934 Node* not_null_b = null_check_oop(b, &null_ctl, !too_many_traps(Deoptimization::Reason_null_check), false, false); 1935 dec_sp(2); 1936 ne_region->init_req(3, null_ctl); 1937 if (stopped()) { 1938 record_for_igvn(ne_region); 1939 set_control(_gvn.transform(ne_region)); 1940 if (btest == BoolTest::ne) { 1941 { 1942 PreserveJVMState pjvms(this); 1943 int target_bci = iter().get_dest(); 1944 merge(target_bci); 1945 } 1946 record_for_igvn(eq_region); 1947 set_control(_gvn.transform(eq_region)); 1948 } 1949 return; 1950 } 1951 Node* kls_a = load_object_klass(not_null_a); 1952 Node* kls_b = load_object_klass(not_null_b); 1953 Node* kls_cmp = CmpP(kls_a, kls_b); 1954 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne)); 1955 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN); 1956 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff)); 1957 set_control(_gvn.transform(new IfFalseNode(kls_iff))); 1958 ne_region->init_req(4, kls_ne); 1959 1960 if (stopped()) { 1961 record_for_igvn(ne_region); 1962 set_control(_gvn.transform(ne_region)); 1963 if (btest == BoolTest::ne) { 1964 { 1965 PreserveJVMState pjvms(this); 1966 int target_bci = iter().get_dest(); 1967 merge(target_bci); 1968 } 1969 record_for_igvn(eq_region); 1970 set_control(_gvn.transform(eq_region)); 1971 } 1972 return; 1973 } 1974 // Both are values of the same class, we need to perform a 1975 // substitutability test. Delegate to 1976 // ValueBootstrapMethods::isSubstitutable(). 1977 1978 Node* ne_io_phi = PhiNode::make(ne_region, i_o()); 1979 Node* mem = reset_memory(); 1980 Node* ne_mem_phi = PhiNode::make(ne_region, mem); 1981 1982 Node* eq_io_phi = NULL; 1983 Node* eq_mem_phi = NULL; 1984 if (eq_region != NULL) { 1985 eq_io_phi = PhiNode::make(eq_region, i_o()); 1986 eq_mem_phi = PhiNode::make(eq_region, mem); 1987 } 1988 1989 set_all_memory(mem); 1990 1991 kill_dead_locals(); 1992 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method, bci()); 1993 call->set_override_symbolic_info(true); 1994 call->init_req(TypeFunc::Parms, not_null_a); 1995 call->init_req(TypeFunc::Parms+1, not_null_b); 1996 inc_sp(2); 1997 set_edges_for_java_call(call, false, false); 1998 Node* ret = set_results_for_java_call(call, false, true); 1999 dec_sp(2); 2000 2001 // Test the return value of ValueBootstrapMethods::isSubstitutable() 2002 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1))); 2003 if (btest == BoolTest::eq) { 2004 do_if(btest, subst_cmp); 2005 } else { 2006 assert(btest == BoolTest::ne, "only eq or ne"); 2007 Node* is_not_equal = NULL; 2008 { 2009 PreserveJVMState pjvms(this); 2010 do_if(btest, subst_cmp, false, &is_not_equal); 2011 if (!stopped()) { 2012 eq_region->init_req(2, control()); 2013 eq_io_phi->init_req(2, i_o()); 2014 eq_mem_phi->init_req(2, reset_memory()); 2015 } 2016 } 2017 set_control(is_not_equal); 2018 } 2019 ne_region->init_req(5, control()); 2020 ne_io_phi->init_req(5, i_o()); 2021 ne_mem_phi->init_req(5, reset_memory()); 2022 2023 record_for_igvn(ne_region); 2024 set_control(_gvn.transform(ne_region)); 2025 set_i_o(_gvn.transform(ne_io_phi)); 2026 set_all_memory(_gvn.transform(ne_mem_phi)); 2027 2028 if (btest == BoolTest::ne) { 2029 { 2030 PreserveJVMState pjvms(this); 2031 int target_bci = iter().get_dest(); 2032 merge(target_bci); 2033 } 2034 2035 record_for_igvn(eq_region); 2036 set_control(_gvn.transform(eq_region)); 2037 set_i_o(_gvn.transform(eq_io_phi)); 2038 set_all_memory(_gvn.transform(eq_mem_phi)); 2039 } 2040 2041 return; 2042 } 2043 // In the case were both operands might be value types, we need to 2044 // use the new acmp implementation. Otherwise, i.e. if one operand 2045 // is not a value type, we can use the old acmp implementation. 2046 Node* cmp = C->optimize_acmp(&_gvn, a, b); 2047 if (cmp != NULL) { 2048 // Use optimized/old acmp 2049 cmp = optimize_cmp_with_klass(_gvn.transform(cmp)); 2050 do_if(btest, cmp); 2051 return; 2052 } 2053 2054 Node* ctrl = NULL; 2055 bool safe_for_replace = true; 2056 if (ACmpOnValues != 1) { 2057 // Emit old acmp before new acmp for quick a != b check 2058 cmp = CmpP(a, b); 2059 cmp = optimize_cmp_with_klass(_gvn.transform(cmp)); 2060 if (btest == BoolTest::ne) { 2061 do_if(btest, cmp, true); 2062 if (stopped()) { 2063 return; // Never equal 2064 } 2065 } else if (btest == BoolTest::eq) { 2066 Node* is_equal = NULL; 2067 { 2068 PreserveJVMState pjvms(this); 2069 do_if(btest, cmp, false, &is_equal); 2070 if (!stopped()) { 2071 // Not equal, skip valuetype check 2072 ctrl = new RegionNode(3); 2073 ctrl->init_req(1, control()); 2074 _gvn.set_type(ctrl, Type::CONTROL); 2075 record_for_igvn(ctrl); 2076 safe_for_replace = false; 2077 } 2078 } 2079 if (is_equal == NULL) { 2080 assert(ctrl != NULL, "no control left"); 2081 set_control(_gvn.transform(ctrl)); 2082 return; // Never equal 2083 } 2084 set_control(is_equal); 2085 } 2086 } 2087 2088 // Null check operand before loading the is_value bit 2089 bool speculate = false; 2090 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(b))) { 2091 // Operand 'b' is never null, swap operands to avoid null check 2092 swap(a, b); 2093 } else if (!too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2094 // Speculate on non-nullness of one operand 2095 if (!_gvn.type(a)->speculative_maybe_null()) { 2096 speculate = true; 2097 } else if (!_gvn.type(b)->speculative_maybe_null()) { 2098 speculate = true; 2099 swap(a, b); 2100 } 2101 } 2102 inc_sp(2); 2103 Node* null_ctl = top(); 2104 Node* not_null_a = null_check_oop(a, &null_ctl, speculate, safe_for_replace, speculate); 2105 assert(!stopped(), "operand is always null"); 2106 dec_sp(2); 2107 Node* region = new RegionNode(2); 2108 Node* is_value = new PhiNode(region, TypeX_X); 2109 if (null_ctl != top()) { 2110 assert(!speculate, "should never be null"); 2111 region->add_req(null_ctl); 2112 is_value->add_req(_gvn.MakeConX(0)); 2113 } 2114 2115 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern); 2116 if (ACmpOnValues == 1) { 2117 Node* mark_addr = basic_plus_adr(not_null_a, oopDesc::mark_offset_in_bytes()); 2118 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 2119 Node* not_mark = _gvn.transform(new XorXNode(mark, _gvn.MakeConX(-1))); 2120 Node* andn = _gvn.transform(new AndXNode(not_mark, value_mask)); 2121 Node* neg_if_value = _gvn.transform(new SubXNode(andn, _gvn.MakeConX(1))); 2122 is_value->init_req(1, _gvn.transform(new RShiftXNode(neg_if_value, _gvn.intcon(63)))); 2123 } else { 2124 is_value->init_req(1, is_always_locked(not_null_a)); 2125 } 2126 region->init_req(1, control()); 2127 2128 set_control(_gvn.transform(region)); 2129 is_value = _gvn.transform(is_value); 2130 2131 if (ACmpOnValues == 1) { 2132 // Perturbe oop if operand is a value type to make comparison fail 2133 Node* pert = _gvn.transform(new AddPNode(a, a, is_value)); 2134 cmp = _gvn.transform(new CmpPNode(pert, b)); 2135 } else { 2136 // Check for a value type because we already know that operands are equal 2137 cmp = _gvn.transform(new CmpXNode(is_value, value_mask)); 2138 btest = (btest == BoolTest::eq) ? BoolTest::ne : BoolTest::eq; 2139 } 2140 cmp = optimize_cmp_with_klass(cmp); 2141 do_if(btest, cmp); 2142 2143 if (ctrl != NULL) { 2144 ctrl->init_req(2, control()); 2145 set_control(_gvn.transform(ctrl)); 2146 } 2147 } 2148 2149 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 2150 // Don't want to speculate on uncommon traps when running with -Xcomp 2151 if (!UseInterpreter) { 2152 return false; 2153 } 2154 return (seems_never_taken(prob) && seems_stable_comparison()); 2155 } 2156 2157 void Parse::maybe_add_predicate_after_if(Block* path) { 2158 if (path->is_SEL_head() && path->preds_parsed() == 0) { 2159 // Add predicates at bci of if dominating the loop so traps can be 2160 // recorded on the if's profile data 2161 int bc_depth = repush_if_args(); 2162 add_predicate(); 2163 dec_sp(bc_depth); 2164 path->set_has_predicates(); 2165 } 2166 } 2167 2168 2169 //----------------------------adjust_map_after_if------------------------------ 2170 // Adjust the JVM state to reflect the result of taking this path. 2171 // Basically, it means inspecting the CmpNode controlling this 2172 // branch, seeing how it constrains a tested value, and then 2173 // deciding if it's worth our while to encode this constraint 2174 // as graph nodes in the current abstract interpretation map. 2175 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) { 2176 if (!c->is_Cmp()) { 2177 maybe_add_predicate_after_if(path); 2178 return; 2179 } 2180 2181 if (stopped() || btest == BoolTest::illegal) { 2182 return; // nothing to do 2183 } 2184 2185 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 2186 2187 if (path_is_suitable_for_uncommon_trap(prob)) { 2188 repush_if_args(); 2189 uncommon_trap(Deoptimization::Reason_unstable_if, 2190 Deoptimization::Action_reinterpret, 2191 NULL, 2192 (is_fallthrough ? "taken always" : "taken never")); 2193 return; 2194 } 2195 2196 Node* val = c->in(1); 2197 Node* con = c->in(2); 2198 const Type* tcon = _gvn.type(con); 2199 const Type* tval = _gvn.type(val); 2200 bool have_con = tcon->singleton(); 2201 if (tval->singleton()) { 2202 if (!have_con) { 2203 // Swap, so constant is in con. 2204 con = val; 2205 tcon = tval; 2206 val = c->in(2); 2207 tval = _gvn.type(val); 2208 btest = BoolTest(btest).commute(); 2209 have_con = true; 2210 } else { 2211 // Do we have two constants? Then leave well enough alone. 2212 have_con = false; 2213 } 2214 } 2215 if (!have_con) { // remaining adjustments need a con 2216 maybe_add_predicate_after_if(path); 2217 return; 2218 } 2219 2220 sharpen_type_after_if(btest, con, tcon, val, tval); 2221 maybe_add_predicate_after_if(path); 2222 } 2223 2224 2225 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 2226 Node* ldk; 2227 if (n->is_DecodeNKlass()) { 2228 if (n->in(1)->Opcode() != Op_LoadNKlass) { 2229 return NULL; 2230 } else { 2231 ldk = n->in(1); 2232 } 2233 } else if (n->Opcode() != Op_LoadKlass) { 2234 return NULL; 2235 } else { 2236 ldk = n; 2237 } 2238 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 2239 2240 Node* adr = ldk->in(MemNode::Address); 2241 intptr_t off = 0; 2242 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 2243 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 2244 return NULL; 2245 const TypePtr* tp = gvn->type(obj)->is_ptr(); 2246 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 2247 return NULL; 2248 2249 return obj; 2250 } 2251 2252 void Parse::sharpen_type_after_if(BoolTest::mask btest, 2253 Node* con, const Type* tcon, 2254 Node* val, const Type* tval) { 2255 // Look for opportunities to sharpen the type of a node 2256 // whose klass is compared with a constant klass. 2257 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 2258 Node* obj = extract_obj_from_klass_load(&_gvn, val); 2259 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 2260 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 2261 // Found: 2262 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 2263 // or the narrowOop equivalent. 2264 const Type* obj_type = _gvn.type(obj); 2265 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 2266 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 2267 tboth->higher_equal(obj_type)) { 2268 // obj has to be of the exact type Foo if the CmpP succeeds. 2269 int obj_in_map = map()->find_edge(obj); 2270 JVMState* jvms = this->jvms(); 2271 if (obj_in_map >= 0 && 2272 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 2273 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 2274 const Type* tcc = ccast->as_Type()->type(); 2275 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 2276 // Delay transform() call to allow recovery of pre-cast value 2277 // at the control merge. 2278 _gvn.set_type_bottom(ccast); 2279 record_for_igvn(ccast); 2280 // Here's the payoff. 2281 replace_in_map(obj, ccast); 2282 } 2283 } 2284 } 2285 } 2286 2287 int val_in_map = map()->find_edge(val); 2288 if (val_in_map < 0) return; // replace_in_map would be useless 2289 { 2290 JVMState* jvms = this->jvms(); 2291 if (!(jvms->is_loc(val_in_map) || 2292 jvms->is_stk(val_in_map))) 2293 return; // again, it would be useless 2294 } 2295 2296 // Check for a comparison to a constant, and "know" that the compared 2297 // value is constrained on this path. 2298 assert(tcon->singleton(), ""); 2299 ConstraintCastNode* ccast = NULL; 2300 Node* cast = NULL; 2301 2302 switch (btest) { 2303 case BoolTest::eq: // Constant test? 2304 { 2305 const Type* tboth = tcon->join_speculative(tval); 2306 if (tboth == tval) break; // Nothing to gain. 2307 if (tcon->isa_int()) { 2308 ccast = new CastIINode(val, tboth); 2309 } else if (tcon == TypePtr::NULL_PTR) { 2310 // Cast to null, but keep the pointer identity temporarily live. 2311 ccast = new CastPPNode(val, tboth); 2312 } else { 2313 const TypeF* tf = tcon->isa_float_constant(); 2314 const TypeD* td = tcon->isa_double_constant(); 2315 // Exclude tests vs float/double 0 as these could be 2316 // either +0 or -0. Just because you are equal to +0 2317 // doesn't mean you ARE +0! 2318 // Note, following code also replaces Long and Oop values. 2319 if ((!tf || tf->_f != 0.0) && 2320 (!td || td->_d != 0.0)) 2321 cast = con; // Replace non-constant val by con. 2322 } 2323 } 2324 break; 2325 2326 case BoolTest::ne: 2327 if (tcon == TypePtr::NULL_PTR) { 2328 cast = cast_not_null(val, false); 2329 } 2330 break; 2331 2332 default: 2333 // (At this point we could record int range types with CastII.) 2334 break; 2335 } 2336 2337 if (ccast != NULL) { 2338 const Type* tcc = ccast->as_Type()->type(); 2339 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 2340 // Delay transform() call to allow recovery of pre-cast value 2341 // at the control merge. 2342 ccast->set_req(0, control()); 2343 _gvn.set_type_bottom(ccast); 2344 record_for_igvn(ccast); 2345 cast = ccast; 2346 } 2347 2348 if (cast != NULL) { // Here's the payoff. 2349 replace_in_map(val, cast); 2350 } 2351 } 2352 2353 /** 2354 * Use speculative type to optimize CmpP node: if comparison is 2355 * against the low level class, cast the object to the speculative 2356 * type if any. CmpP should then go away. 2357 * 2358 * @param c expected CmpP node 2359 * @return result of CmpP on object casted to speculative type 2360 * 2361 */ 2362 Node* Parse::optimize_cmp_with_klass(Node* c) { 2363 // If this is transformed by the _gvn to a comparison with the low 2364 // level klass then we may be able to use speculation 2365 if (c->Opcode() == Op_CmpP && 2366 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 2367 c->in(2)->is_Con()) { 2368 Node* load_klass = NULL; 2369 Node* decode = NULL; 2370 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 2371 decode = c->in(1); 2372 load_klass = c->in(1)->in(1); 2373 } else { 2374 load_klass = c->in(1); 2375 } 2376 if (load_klass->in(2)->is_AddP()) { 2377 Node* addp = load_klass->in(2); 2378 Node* obj = addp->in(AddPNode::Address); 2379 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2380 if (obj_type->speculative_type_not_null() != NULL) { 2381 ciKlass* k = obj_type->speculative_type(); 2382 inc_sp(2); 2383 obj = maybe_cast_profiled_obj(obj, k); 2384 dec_sp(2); 2385 if (obj->is_ValueType()) { 2386 assert(obj->as_ValueType()->is_allocated(&_gvn), "must be allocated"); 2387 obj = obj->as_ValueType()->get_oop(); 2388 } 2389 // Make the CmpP use the casted obj 2390 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 2391 load_klass = load_klass->clone(); 2392 load_klass->set_req(2, addp); 2393 load_klass = _gvn.transform(load_klass); 2394 if (decode != NULL) { 2395 decode = decode->clone(); 2396 decode->set_req(1, load_klass); 2397 load_klass = _gvn.transform(decode); 2398 } 2399 c = c->clone(); 2400 c->set_req(1, load_klass); 2401 c = _gvn.transform(c); 2402 } 2403 } 2404 } 2405 return c; 2406 } 2407 2408 //------------------------------do_one_bytecode-------------------------------- 2409 // Parse this bytecode, and alter the Parsers JVM->Node mapping 2410 void Parse::do_one_bytecode() { 2411 Node *a, *b, *c, *d; // Handy temps 2412 BoolTest::mask btest; 2413 int i; 2414 2415 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 2416 2417 if (C->check_node_count(NodeLimitFudgeFactor * 5, 2418 "out of nodes parsing method")) { 2419 return; 2420 } 2421 2422 #ifdef ASSERT 2423 // for setting breakpoints 2424 if (TraceOptoParse) { 2425 tty->print(" @"); 2426 dump_bci(bci()); 2427 tty->cr(); 2428 } 2429 #endif 2430 2431 switch (bc()) { 2432 case Bytecodes::_nop: 2433 // do nothing 2434 break; 2435 case Bytecodes::_lconst_0: 2436 push_pair(longcon(0)); 2437 break; 2438 2439 case Bytecodes::_lconst_1: 2440 push_pair(longcon(1)); 2441 break; 2442 2443 case Bytecodes::_fconst_0: 2444 push(zerocon(T_FLOAT)); 2445 break; 2446 2447 case Bytecodes::_fconst_1: 2448 push(makecon(TypeF::ONE)); 2449 break; 2450 2451 case Bytecodes::_fconst_2: 2452 push(makecon(TypeF::make(2.0f))); 2453 break; 2454 2455 case Bytecodes::_dconst_0: 2456 push_pair(zerocon(T_DOUBLE)); 2457 break; 2458 2459 case Bytecodes::_dconst_1: 2460 push_pair(makecon(TypeD::ONE)); 2461 break; 2462 2463 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 2464 case Bytecodes::_iconst_0: push(intcon( 0)); break; 2465 case Bytecodes::_iconst_1: push(intcon( 1)); break; 2466 case Bytecodes::_iconst_2: push(intcon( 2)); break; 2467 case Bytecodes::_iconst_3: push(intcon( 3)); break; 2468 case Bytecodes::_iconst_4: push(intcon( 4)); break; 2469 case Bytecodes::_iconst_5: push(intcon( 5)); break; 2470 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 2471 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 2472 case Bytecodes::_aconst_null: push(null()); break; 2473 case Bytecodes::_ldc: 2474 case Bytecodes::_ldc_w: 2475 case Bytecodes::_ldc2_w: 2476 // If the constant is unresolved, run this BC once in the interpreter. 2477 { 2478 ciConstant constant = iter().get_constant(); 2479 if (!constant.is_valid() || 2480 (constant.basic_type() == T_OBJECT && 2481 !constant.as_object()->is_loaded())) { 2482 int index = iter().get_constant_pool_index(); 2483 constantTag tag = iter().get_constant_pool_tag(index); 2484 uncommon_trap(Deoptimization::make_trap_request 2485 (Deoptimization::Reason_unloaded, 2486 Deoptimization::Action_reinterpret, 2487 index), 2488 NULL, tag.internal_name()); 2489 break; 2490 } 2491 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 2492 "must be java_mirror of klass"); 2493 const Type* con_type = Type::make_from_constant(constant); 2494 if (con_type != NULL) { 2495 push_node(con_type->basic_type(), makecon(con_type)); 2496 } 2497 } 2498 2499 break; 2500 2501 case Bytecodes::_aload_0: 2502 push( local(0) ); 2503 break; 2504 case Bytecodes::_aload_1: 2505 push( local(1) ); 2506 break; 2507 case Bytecodes::_aload_2: 2508 push( local(2) ); 2509 break; 2510 case Bytecodes::_aload_3: 2511 push( local(3) ); 2512 break; 2513 case Bytecodes::_aload: 2514 push( local(iter().get_index()) ); 2515 break; 2516 2517 case Bytecodes::_fload_0: 2518 case Bytecodes::_iload_0: 2519 push( local(0) ); 2520 break; 2521 case Bytecodes::_fload_1: 2522 case Bytecodes::_iload_1: 2523 push( local(1) ); 2524 break; 2525 case Bytecodes::_fload_2: 2526 case Bytecodes::_iload_2: 2527 push( local(2) ); 2528 break; 2529 case Bytecodes::_fload_3: 2530 case Bytecodes::_iload_3: 2531 push( local(3) ); 2532 break; 2533 case Bytecodes::_fload: 2534 case Bytecodes::_iload: 2535 push( local(iter().get_index()) ); 2536 break; 2537 case Bytecodes::_lload_0: 2538 push_pair_local( 0 ); 2539 break; 2540 case Bytecodes::_lload_1: 2541 push_pair_local( 1 ); 2542 break; 2543 case Bytecodes::_lload_2: 2544 push_pair_local( 2 ); 2545 break; 2546 case Bytecodes::_lload_3: 2547 push_pair_local( 3 ); 2548 break; 2549 case Bytecodes::_lload: 2550 push_pair_local( iter().get_index() ); 2551 break; 2552 2553 case Bytecodes::_dload_0: 2554 push_pair_local(0); 2555 break; 2556 case Bytecodes::_dload_1: 2557 push_pair_local(1); 2558 break; 2559 case Bytecodes::_dload_2: 2560 push_pair_local(2); 2561 break; 2562 case Bytecodes::_dload_3: 2563 push_pair_local(3); 2564 break; 2565 case Bytecodes::_dload: 2566 push_pair_local(iter().get_index()); 2567 break; 2568 case Bytecodes::_fstore_0: 2569 case Bytecodes::_istore_0: 2570 case Bytecodes::_astore_0: 2571 set_local( 0, pop() ); 2572 break; 2573 case Bytecodes::_fstore_1: 2574 case Bytecodes::_istore_1: 2575 case Bytecodes::_astore_1: 2576 set_local( 1, pop() ); 2577 break; 2578 case Bytecodes::_fstore_2: 2579 case Bytecodes::_istore_2: 2580 case Bytecodes::_astore_2: 2581 set_local( 2, pop() ); 2582 break; 2583 case Bytecodes::_fstore_3: 2584 case Bytecodes::_istore_3: 2585 case Bytecodes::_astore_3: 2586 set_local( 3, pop() ); 2587 break; 2588 case Bytecodes::_fstore: 2589 case Bytecodes::_istore: 2590 case Bytecodes::_astore: 2591 set_local( iter().get_index(), pop() ); 2592 break; 2593 // long stores 2594 case Bytecodes::_lstore_0: 2595 set_pair_local( 0, pop_pair() ); 2596 break; 2597 case Bytecodes::_lstore_1: 2598 set_pair_local( 1, pop_pair() ); 2599 break; 2600 case Bytecodes::_lstore_2: 2601 set_pair_local( 2, pop_pair() ); 2602 break; 2603 case Bytecodes::_lstore_3: 2604 set_pair_local( 3, pop_pair() ); 2605 break; 2606 case Bytecodes::_lstore: 2607 set_pair_local( iter().get_index(), pop_pair() ); 2608 break; 2609 2610 // double stores 2611 case Bytecodes::_dstore_0: 2612 set_pair_local( 0, dstore_rounding(pop_pair()) ); 2613 break; 2614 case Bytecodes::_dstore_1: 2615 set_pair_local( 1, dstore_rounding(pop_pair()) ); 2616 break; 2617 case Bytecodes::_dstore_2: 2618 set_pair_local( 2, dstore_rounding(pop_pair()) ); 2619 break; 2620 case Bytecodes::_dstore_3: 2621 set_pair_local( 3, dstore_rounding(pop_pair()) ); 2622 break; 2623 case Bytecodes::_dstore: 2624 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 2625 break; 2626 2627 case Bytecodes::_pop: dec_sp(1); break; 2628 case Bytecodes::_pop2: dec_sp(2); break; 2629 case Bytecodes::_swap: 2630 a = pop(); 2631 b = pop(); 2632 push(a); 2633 push(b); 2634 break; 2635 case Bytecodes::_dup: 2636 a = pop(); 2637 push(a); 2638 push(a); 2639 break; 2640 case Bytecodes::_dup_x1: 2641 a = pop(); 2642 b = pop(); 2643 push( a ); 2644 push( b ); 2645 push( a ); 2646 break; 2647 case Bytecodes::_dup_x2: 2648 a = pop(); 2649 b = pop(); 2650 c = pop(); 2651 push( a ); 2652 push( c ); 2653 push( b ); 2654 push( a ); 2655 break; 2656 case Bytecodes::_dup2: 2657 a = pop(); 2658 b = pop(); 2659 push( b ); 2660 push( a ); 2661 push( b ); 2662 push( a ); 2663 break; 2664 2665 case Bytecodes::_dup2_x1: 2666 // before: .. c, b, a 2667 // after: .. b, a, c, b, a 2668 // not tested 2669 a = pop(); 2670 b = pop(); 2671 c = pop(); 2672 push( b ); 2673 push( a ); 2674 push( c ); 2675 push( b ); 2676 push( a ); 2677 break; 2678 case Bytecodes::_dup2_x2: 2679 // before: .. d, c, b, a 2680 // after: .. b, a, d, c, b, a 2681 // not tested 2682 a = pop(); 2683 b = pop(); 2684 c = pop(); 2685 d = pop(); 2686 push( b ); 2687 push( a ); 2688 push( d ); 2689 push( c ); 2690 push( b ); 2691 push( a ); 2692 break; 2693 2694 case Bytecodes::_arraylength: { 2695 // Must do null-check with value on expression stack 2696 Node *ary = null_check(peek(), T_ARRAY); 2697 // Compile-time detect of null-exception? 2698 if (stopped()) return; 2699 a = pop(); 2700 push(load_array_length(a)); 2701 break; 2702 } 2703 2704 case Bytecodes::_baload: array_load(T_BYTE); break; 2705 case Bytecodes::_caload: array_load(T_CHAR); break; 2706 case Bytecodes::_iaload: array_load(T_INT); break; 2707 case Bytecodes::_saload: array_load(T_SHORT); break; 2708 case Bytecodes::_faload: array_load(T_FLOAT); break; 2709 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2710 case Bytecodes::_laload: array_load(T_LONG); break; 2711 case Bytecodes::_daload: array_load(T_DOUBLE); break; 2712 case Bytecodes::_bastore: array_store(T_BYTE); break; 2713 case Bytecodes::_castore: array_store(T_CHAR); break; 2714 case Bytecodes::_iastore: array_store(T_INT); break; 2715 case Bytecodes::_sastore: array_store(T_SHORT); break; 2716 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2717 case Bytecodes::_aastore: array_store(T_OBJECT); break; 2718 case Bytecodes::_lastore: array_store(T_LONG); break; 2719 case Bytecodes::_dastore: array_store(T_DOUBLE); break; 2720 2721 case Bytecodes::_getfield: 2722 do_getfield(); 2723 break; 2724 2725 case Bytecodes::_getstatic: 2726 do_getstatic(); 2727 break; 2728 2729 case Bytecodes::_putfield: 2730 do_putfield(); 2731 break; 2732 2733 case Bytecodes::_putstatic: 2734 do_putstatic(); 2735 break; 2736 2737 case Bytecodes::_irem: 2738 do_irem(); 2739 break; 2740 case Bytecodes::_idiv: 2741 // Must keep both values on the expression-stack during null-check 2742 zero_check_int(peek()); 2743 // Compile-time detect of null-exception? 2744 if (stopped()) return; 2745 b = pop(); 2746 a = pop(); 2747 push( _gvn.transform( new DivINode(control(),a,b) ) ); 2748 break; 2749 case Bytecodes::_imul: 2750 b = pop(); a = pop(); 2751 push( _gvn.transform( new MulINode(a,b) ) ); 2752 break; 2753 case Bytecodes::_iadd: 2754 b = pop(); a = pop(); 2755 push( _gvn.transform( new AddINode(a,b) ) ); 2756 break; 2757 case Bytecodes::_ineg: 2758 a = pop(); 2759 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 2760 break; 2761 case Bytecodes::_isub: 2762 b = pop(); a = pop(); 2763 push( _gvn.transform( new SubINode(a,b) ) ); 2764 break; 2765 case Bytecodes::_iand: 2766 b = pop(); a = pop(); 2767 push( _gvn.transform( new AndINode(a,b) ) ); 2768 break; 2769 case Bytecodes::_ior: 2770 b = pop(); a = pop(); 2771 push( _gvn.transform( new OrINode(a,b) ) ); 2772 break; 2773 case Bytecodes::_ixor: 2774 b = pop(); a = pop(); 2775 push( _gvn.transform( new XorINode(a,b) ) ); 2776 break; 2777 case Bytecodes::_ishl: 2778 b = pop(); a = pop(); 2779 push( _gvn.transform( new LShiftINode(a,b) ) ); 2780 break; 2781 case Bytecodes::_ishr: 2782 b = pop(); a = pop(); 2783 push( _gvn.transform( new RShiftINode(a,b) ) ); 2784 break; 2785 case Bytecodes::_iushr: 2786 b = pop(); a = pop(); 2787 push( _gvn.transform( new URShiftINode(a,b) ) ); 2788 break; 2789 2790 case Bytecodes::_fneg: 2791 a = pop(); 2792 b = _gvn.transform(new NegFNode (a)); 2793 push(b); 2794 break; 2795 2796 case Bytecodes::_fsub: 2797 b = pop(); 2798 a = pop(); 2799 c = _gvn.transform( new SubFNode(a,b) ); 2800 d = precision_rounding(c); 2801 push( d ); 2802 break; 2803 2804 case Bytecodes::_fadd: 2805 b = pop(); 2806 a = pop(); 2807 c = _gvn.transform( new AddFNode(a,b) ); 2808 d = precision_rounding(c); 2809 push( d ); 2810 break; 2811 2812 case Bytecodes::_fmul: 2813 b = pop(); 2814 a = pop(); 2815 c = _gvn.transform( new MulFNode(a,b) ); 2816 d = precision_rounding(c); 2817 push( d ); 2818 break; 2819 2820 case Bytecodes::_fdiv: 2821 b = pop(); 2822 a = pop(); 2823 c = _gvn.transform( new DivFNode(0,a,b) ); 2824 d = precision_rounding(c); 2825 push( d ); 2826 break; 2827 2828 case Bytecodes::_frem: 2829 if (Matcher::has_match_rule(Op_ModF)) { 2830 // Generate a ModF node. 2831 b = pop(); 2832 a = pop(); 2833 c = _gvn.transform( new ModFNode(0,a,b) ); 2834 d = precision_rounding(c); 2835 push( d ); 2836 } 2837 else { 2838 // Generate a call. 2839 modf(); 2840 } 2841 break; 2842 2843 case Bytecodes::_fcmpl: 2844 b = pop(); 2845 a = pop(); 2846 c = _gvn.transform( new CmpF3Node( a, b)); 2847 push(c); 2848 break; 2849 case Bytecodes::_fcmpg: 2850 b = pop(); 2851 a = pop(); 2852 2853 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 2854 // which negates the result sign except for unordered. Flip the unordered 2855 // as well by using CmpF3 which implements unordered-lesser instead of 2856 // unordered-greater semantics. Finally, commute the result bits. Result 2857 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 2858 c = _gvn.transform( new CmpF3Node( b, a)); 2859 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2860 push(c); 2861 break; 2862 2863 case Bytecodes::_f2i: 2864 a = pop(); 2865 push(_gvn.transform(new ConvF2INode(a))); 2866 break; 2867 2868 case Bytecodes::_d2i: 2869 a = pop_pair(); 2870 b = _gvn.transform(new ConvD2INode(a)); 2871 push( b ); 2872 break; 2873 2874 case Bytecodes::_f2d: 2875 a = pop(); 2876 b = _gvn.transform( new ConvF2DNode(a)); 2877 push_pair( b ); 2878 break; 2879 2880 case Bytecodes::_d2f: 2881 a = pop_pair(); 2882 b = _gvn.transform( new ConvD2FNode(a)); 2883 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 2884 //b = _gvn.transform(new RoundFloatNode(0, b) ); 2885 push( b ); 2886 break; 2887 2888 case Bytecodes::_l2f: 2889 if (Matcher::convL2FSupported()) { 2890 a = pop_pair(); 2891 b = _gvn.transform( new ConvL2FNode(a)); 2892 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 2893 // Rather than storing the result into an FP register then pushing 2894 // out to memory to round, the machine instruction that implements 2895 // ConvL2D is responsible for rounding. 2896 // c = precision_rounding(b); 2897 c = _gvn.transform(b); 2898 push(c); 2899 } else { 2900 l2f(); 2901 } 2902 break; 2903 2904 case Bytecodes::_l2d: 2905 a = pop_pair(); 2906 b = _gvn.transform( new ConvL2DNode(a)); 2907 // For i486.ad, rounding is always necessary (see _l2f above). 2908 // c = dprecision_rounding(b); 2909 c = _gvn.transform(b); 2910 push_pair(c); 2911 break; 2912 2913 case Bytecodes::_f2l: 2914 a = pop(); 2915 b = _gvn.transform( new ConvF2LNode(a)); 2916 push_pair(b); 2917 break; 2918 2919 case Bytecodes::_d2l: 2920 a = pop_pair(); 2921 b = _gvn.transform( new ConvD2LNode(a)); 2922 push_pair(b); 2923 break; 2924 2925 case Bytecodes::_dsub: 2926 b = pop_pair(); 2927 a = pop_pair(); 2928 c = _gvn.transform( new SubDNode(a,b) ); 2929 d = dprecision_rounding(c); 2930 push_pair( d ); 2931 break; 2932 2933 case Bytecodes::_dadd: 2934 b = pop_pair(); 2935 a = pop_pair(); 2936 c = _gvn.transform( new AddDNode(a,b) ); 2937 d = dprecision_rounding(c); 2938 push_pair( d ); 2939 break; 2940 2941 case Bytecodes::_dmul: 2942 b = pop_pair(); 2943 a = pop_pair(); 2944 c = _gvn.transform( new MulDNode(a,b) ); 2945 d = dprecision_rounding(c); 2946 push_pair( d ); 2947 break; 2948 2949 case Bytecodes::_ddiv: 2950 b = pop_pair(); 2951 a = pop_pair(); 2952 c = _gvn.transform( new DivDNode(0,a,b) ); 2953 d = dprecision_rounding(c); 2954 push_pair( d ); 2955 break; 2956 2957 case Bytecodes::_dneg: 2958 a = pop_pair(); 2959 b = _gvn.transform(new NegDNode (a)); 2960 push_pair(b); 2961 break; 2962 2963 case Bytecodes::_drem: 2964 if (Matcher::has_match_rule(Op_ModD)) { 2965 // Generate a ModD node. 2966 b = pop_pair(); 2967 a = pop_pair(); 2968 // a % b 2969 2970 c = _gvn.transform( new ModDNode(0,a,b) ); 2971 d = dprecision_rounding(c); 2972 push_pair( d ); 2973 } 2974 else { 2975 // Generate a call. 2976 modd(); 2977 } 2978 break; 2979 2980 case Bytecodes::_dcmpl: 2981 b = pop_pair(); 2982 a = pop_pair(); 2983 c = _gvn.transform( new CmpD3Node( a, b)); 2984 push(c); 2985 break; 2986 2987 case Bytecodes::_dcmpg: 2988 b = pop_pair(); 2989 a = pop_pair(); 2990 // Same as dcmpl but need to flip the unordered case. 2991 // Commute the inputs, which negates the result sign except for unordered. 2992 // Flip the unordered as well by using CmpD3 which implements 2993 // unordered-lesser instead of unordered-greater semantics. 2994 // Finally, negate the result bits. Result is same as using a 2995 // CmpD3Greater except we did it with CmpD3 alone. 2996 c = _gvn.transform( new CmpD3Node( b, a)); 2997 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2998 push(c); 2999 break; 3000 3001 3002 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 3003 case Bytecodes::_land: 3004 b = pop_pair(); 3005 a = pop_pair(); 3006 c = _gvn.transform( new AndLNode(a,b) ); 3007 push_pair(c); 3008 break; 3009 case Bytecodes::_lor: 3010 b = pop_pair(); 3011 a = pop_pair(); 3012 c = _gvn.transform( new OrLNode(a,b) ); 3013 push_pair(c); 3014 break; 3015 case Bytecodes::_lxor: 3016 b = pop_pair(); 3017 a = pop_pair(); 3018 c = _gvn.transform( new XorLNode(a,b) ); 3019 push_pair(c); 3020 break; 3021 3022 case Bytecodes::_lshl: 3023 b = pop(); // the shift count 3024 a = pop_pair(); // value to be shifted 3025 c = _gvn.transform( new LShiftLNode(a,b) ); 3026 push_pair(c); 3027 break; 3028 case Bytecodes::_lshr: 3029 b = pop(); // the shift count 3030 a = pop_pair(); // value to be shifted 3031 c = _gvn.transform( new RShiftLNode(a,b) ); 3032 push_pair(c); 3033 break; 3034 case Bytecodes::_lushr: 3035 b = pop(); // the shift count 3036 a = pop_pair(); // value to be shifted 3037 c = _gvn.transform( new URShiftLNode(a,b) ); 3038 push_pair(c); 3039 break; 3040 case Bytecodes::_lmul: 3041 b = pop_pair(); 3042 a = pop_pair(); 3043 c = _gvn.transform( new MulLNode(a,b) ); 3044 push_pair(c); 3045 break; 3046 3047 case Bytecodes::_lrem: 3048 // Must keep both values on the expression-stack during null-check 3049 assert(peek(0) == top(), "long word order"); 3050 zero_check_long(peek(1)); 3051 // Compile-time detect of null-exception? 3052 if (stopped()) return; 3053 b = pop_pair(); 3054 a = pop_pair(); 3055 c = _gvn.transform( new ModLNode(control(),a,b) ); 3056 push_pair(c); 3057 break; 3058 3059 case Bytecodes::_ldiv: 3060 // Must keep both values on the expression-stack during null-check 3061 assert(peek(0) == top(), "long word order"); 3062 zero_check_long(peek(1)); 3063 // Compile-time detect of null-exception? 3064 if (stopped()) return; 3065 b = pop_pair(); 3066 a = pop_pair(); 3067 c = _gvn.transform( new DivLNode(control(),a,b) ); 3068 push_pair(c); 3069 break; 3070 3071 case Bytecodes::_ladd: 3072 b = pop_pair(); 3073 a = pop_pair(); 3074 c = _gvn.transform( new AddLNode(a,b) ); 3075 push_pair(c); 3076 break; 3077 case Bytecodes::_lsub: 3078 b = pop_pair(); 3079 a = pop_pair(); 3080 c = _gvn.transform( new SubLNode(a,b) ); 3081 push_pair(c); 3082 break; 3083 case Bytecodes::_lcmp: 3084 // Safepoints are now inserted _before_ branches. The long-compare 3085 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 3086 // slew of control flow. These are usually followed by a CmpI vs zero and 3087 // a branch; this pattern then optimizes to the obvious long-compare and 3088 // branch. However, if the branch is backwards there's a Safepoint 3089 // inserted. The inserted Safepoint captures the JVM state at the 3090 // pre-branch point, i.e. it captures the 3-way value. Thus if a 3091 // long-compare is used to control a loop the debug info will force 3092 // computation of the 3-way value, even though the generated code uses a 3093 // long-compare and branch. We try to rectify the situation by inserting 3094 // a SafePoint here and have it dominate and kill the safepoint added at a 3095 // following backwards branch. At this point the JVM state merely holds 2 3096 // longs but not the 3-way value. 3097 if( UseLoopSafepoints ) { 3098 switch( iter().next_bc() ) { 3099 case Bytecodes::_ifgt: 3100 case Bytecodes::_iflt: 3101 case Bytecodes::_ifge: 3102 case Bytecodes::_ifle: 3103 case Bytecodes::_ifne: 3104 case Bytecodes::_ifeq: 3105 // If this is a backwards branch in the bytecodes, add Safepoint 3106 maybe_add_safepoint(iter().next_get_dest()); 3107 default: 3108 break; 3109 } 3110 } 3111 b = pop_pair(); 3112 a = pop_pair(); 3113 c = _gvn.transform( new CmpL3Node( a, b )); 3114 push(c); 3115 break; 3116 3117 case Bytecodes::_lneg: 3118 a = pop_pair(); 3119 b = _gvn.transform( new SubLNode(longcon(0),a)); 3120 push_pair(b); 3121 break; 3122 case Bytecodes::_l2i: 3123 a = pop_pair(); 3124 push( _gvn.transform( new ConvL2INode(a))); 3125 break; 3126 case Bytecodes::_i2l: 3127 a = pop(); 3128 b = _gvn.transform( new ConvI2LNode(a)); 3129 push_pair(b); 3130 break; 3131 case Bytecodes::_i2b: 3132 // Sign extend 3133 a = pop(); 3134 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 3135 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 3136 push( a ); 3137 break; 3138 case Bytecodes::_i2s: 3139 a = pop(); 3140 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 3141 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 3142 push( a ); 3143 break; 3144 case Bytecodes::_i2c: 3145 a = pop(); 3146 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 3147 break; 3148 3149 case Bytecodes::_i2f: 3150 a = pop(); 3151 b = _gvn.transform( new ConvI2FNode(a) ) ; 3152 c = precision_rounding(b); 3153 push (b); 3154 break; 3155 3156 case Bytecodes::_i2d: 3157 a = pop(); 3158 b = _gvn.transform( new ConvI2DNode(a)); 3159 push_pair(b); 3160 break; 3161 3162 case Bytecodes::_iinc: // Increment local 3163 i = iter().get_index(); // Get local index 3164 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 3165 break; 3166 3167 // Exit points of synchronized methods must have an unlock node 3168 case Bytecodes::_return: 3169 return_current(NULL); 3170 break; 3171 3172 case Bytecodes::_ireturn: 3173 case Bytecodes::_areturn: 3174 case Bytecodes::_freturn: 3175 return_current(pop()); 3176 break; 3177 case Bytecodes::_lreturn: 3178 return_current(pop_pair()); 3179 break; 3180 case Bytecodes::_dreturn: 3181 return_current(pop_pair()); 3182 break; 3183 3184 case Bytecodes::_athrow: 3185 // null exception oop throws NULL pointer exception 3186 null_check(peek()); 3187 if (stopped()) return; 3188 // Hook the thrown exception directly to subsequent handlers. 3189 if (BailoutToInterpreterForThrows) { 3190 // Keep method interpreted from now on. 3191 uncommon_trap(Deoptimization::Reason_unhandled, 3192 Deoptimization::Action_make_not_compilable); 3193 return; 3194 } 3195 if (env()->jvmti_can_post_on_exceptions()) { 3196 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 3197 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 3198 } 3199 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 3200 add_exception_state(make_exception_state(peek())); 3201 break; 3202 3203 case Bytecodes::_goto: // fall through 3204 case Bytecodes::_goto_w: { 3205 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 3206 3207 // If this is a backwards branch in the bytecodes, add Safepoint 3208 maybe_add_safepoint(target_bci); 3209 3210 // Update method data 3211 profile_taken_branch(target_bci); 3212 3213 // Merge the current control into the target basic block 3214 merge(target_bci); 3215 3216 // See if we can get some profile data and hand it off to the next block 3217 Block *target_block = block()->successor_for_bci(target_bci); 3218 if (target_block->pred_count() != 1) break; 3219 ciMethodData* methodData = method()->method_data(); 3220 if (!methodData->is_mature()) break; 3221 ciProfileData* data = methodData->bci_to_data(bci()); 3222 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch"); 3223 int taken = ((ciJumpData*)data)->taken(); 3224 taken = method()->scale_count(taken); 3225 target_block->set_count(taken); 3226 break; 3227 } 3228 3229 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 3230 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 3231 handle_if_null: 3232 // If this is a backwards branch in the bytecodes, add Safepoint 3233 maybe_add_safepoint(iter().get_dest()); 3234 a = null(); 3235 b = pop(); 3236 if (b->is_ValueType()) { 3237 // Return constant false because 'b' is always non-null 3238 c = _gvn.makecon(TypeInt::CC_GT); 3239 } else { 3240 if (!_gvn.type(b)->speculative_maybe_null() && 3241 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 3242 inc_sp(1); 3243 Node* null_ctl = top(); 3244 b = null_check_oop(b, &null_ctl, true, true, true); 3245 assert(null_ctl->is_top(), "no null control here"); 3246 dec_sp(1); 3247 } else if (_gvn.type(b)->speculative_always_null() && 3248 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 3249 inc_sp(1); 3250 b = null_assert(b); 3251 dec_sp(1); 3252 } 3253 c = _gvn.transform( new CmpPNode(b, a) ); 3254 } 3255 do_ifnull(btest, c); 3256 break; 3257 3258 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 3259 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 3260 handle_if_acmp: 3261 // If this is a backwards branch in the bytecodes, add Safepoint 3262 maybe_add_safepoint(iter().get_dest()); 3263 a = access_resolve(pop(), 0); 3264 b = access_resolve(pop(), 0); 3265 do_acmp(btest, a, b); 3266 break; 3267 3268 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 3269 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 3270 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 3271 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 3272 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 3273 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 3274 handle_ifxx: 3275 // If this is a backwards branch in the bytecodes, add Safepoint 3276 maybe_add_safepoint(iter().get_dest()); 3277 a = _gvn.intcon(0); 3278 b = pop(); 3279 c = _gvn.transform( new CmpINode(b, a) ); 3280 do_if(btest, c); 3281 break; 3282 3283 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 3284 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 3285 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 3286 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 3287 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 3288 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 3289 handle_if_icmp: 3290 // If this is a backwards branch in the bytecodes, add Safepoint 3291 maybe_add_safepoint(iter().get_dest()); 3292 a = pop(); 3293 b = pop(); 3294 c = _gvn.transform( new CmpINode( b, a ) ); 3295 do_if(btest, c); 3296 break; 3297 3298 case Bytecodes::_tableswitch: 3299 do_tableswitch(); 3300 break; 3301 3302 case Bytecodes::_lookupswitch: 3303 do_lookupswitch(); 3304 break; 3305 3306 case Bytecodes::_invokestatic: 3307 case Bytecodes::_invokedynamic: 3308 case Bytecodes::_invokespecial: 3309 case Bytecodes::_invokevirtual: 3310 case Bytecodes::_invokeinterface: 3311 do_call(); 3312 break; 3313 case Bytecodes::_checkcast: 3314 do_checkcast(); 3315 break; 3316 case Bytecodes::_instanceof: 3317 do_instanceof(); 3318 break; 3319 case Bytecodes::_anewarray: 3320 do_newarray(); 3321 break; 3322 case Bytecodes::_newarray: 3323 do_newarray((BasicType)iter().get_index()); 3324 break; 3325 case Bytecodes::_multianewarray: 3326 do_multianewarray(); 3327 break; 3328 case Bytecodes::_new: 3329 do_new(); 3330 break; 3331 case Bytecodes::_defaultvalue: 3332 do_defaultvalue(); 3333 break; 3334 case Bytecodes::_withfield: 3335 do_withfield(); 3336 break; 3337 3338 case Bytecodes::_jsr: 3339 case Bytecodes::_jsr_w: 3340 do_jsr(); 3341 break; 3342 3343 case Bytecodes::_ret: 3344 do_ret(); 3345 break; 3346 3347 3348 case Bytecodes::_monitorenter: 3349 do_monitor_enter(); 3350 break; 3351 3352 case Bytecodes::_monitorexit: 3353 do_monitor_exit(); 3354 break; 3355 3356 case Bytecodes::_breakpoint: 3357 // Breakpoint set concurrently to compile 3358 // %%% use an uncommon trap? 3359 C->record_failure("breakpoint in method"); 3360 return; 3361 3362 default: 3363 #ifndef PRODUCT 3364 map()->dump(99); 3365 #endif 3366 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 3367 ShouldNotReachHere(); 3368 } 3369 3370 #ifndef PRODUCT 3371 IdealGraphPrinter *printer = C->printer(); 3372 if (printer && printer->should_print(1)) { 3373 char buffer[256]; 3374 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 3375 bool old = printer->traverse_outs(); 3376 printer->set_traverse_outs(true); 3377 printer->print_method(buffer, 4); 3378 printer->set_traverse_outs(old); 3379 } 3380 #endif 3381 }