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