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/matcher.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/mulnode.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/parse.hpp" 44 #include "opto/runtime.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 #ifndef PRODUCT 49 extern int explicit_null_checks_inserted, 50 explicit_null_checks_elided; 51 #endif 52 53 //---------------------------------array_load---------------------------------- 54 void Parse::array_load(BasicType elem_type) { 55 const Type* elem = Type::TOP; 56 Node* adr = array_addressing(elem_type, 0, &elem); 57 if (stopped()) return; // guaranteed null or range check 58 dec_sp(2); // Pop array and index 59 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 60 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 61 if (UseZGC && elem_type == T_OBJECT) { 62 ld = load_barrier(ld, adr); 63 } 64 push(ld); 65 } 66 67 68 //--------------------------------array_store---------------------------------- 69 void Parse::array_store(BasicType elem_type) { 70 const Type* elem = Type::TOP; 71 Node* adr = array_addressing(elem_type, 1, &elem); 72 if (stopped()) return; // guaranteed null or range check 73 Node* val = pop(); 74 dec_sp(2); // Pop array and index 75 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 76 if (elem == TypeInt::BOOL) { 77 elem_type = T_BOOLEAN; 78 } 79 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 80 } 81 82 83 //------------------------------array_addressing------------------------------- 84 // Pull array and index from the stack. Compute pointer-to-element. 85 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 86 Node *idx = peek(0+vals); // Get from stack without popping 87 Node *ary = peek(1+vals); // in case of exception 88 89 // Null check the array base, with correct stack contents 90 ary = null_check(ary, T_ARRAY); 91 // Compile-time detect of null-exception? 92 if (stopped()) return top(); 93 94 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 95 const TypeInt* sizetype = arytype->size(); 96 const Type* elemtype = arytype->elem(); 97 98 if (UseUniqueSubclasses && result2 != NULL) { 99 const Type* el = elemtype->make_ptr(); 100 if (el && el->isa_instptr()) { 101 const TypeInstPtr* toop = el->is_instptr(); 102 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 103 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 104 const Type* subklass = Type::get_const_type(toop->klass()); 105 elemtype = subklass->join_speculative(el); 106 } 107 } 108 } 109 110 // Check for big class initializers with all constant offsets 111 // feeding into a known-size array. 112 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 113 // See if the highest idx value is less than the lowest array bound, 114 // and if the idx value cannot be negative: 115 bool need_range_check = true; 116 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 117 need_range_check = false; 118 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 119 } 120 121 ciKlass * arytype_klass = arytype->klass(); 122 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 123 // Only fails for some -Xcomp runs 124 // The class is unloaded. We have to run this bytecode in the interpreter. 125 uncommon_trap(Deoptimization::Reason_unloaded, 126 Deoptimization::Action_reinterpret, 127 arytype->klass(), "!loaded array"); 128 return top(); 129 } 130 131 // Do the range check 132 if (GenerateRangeChecks && need_range_check) { 133 Node* tst; 134 if (sizetype->_hi <= 0) { 135 // The greatest array bound is negative, so we can conclude that we're 136 // compiling unreachable code, but the unsigned compare trick used below 137 // only works with non-negative lengths. Instead, hack "tst" to be zero so 138 // the uncommon_trap path will always be taken. 139 tst = _gvn.intcon(0); 140 } else { 141 // Range is constant in array-oop, so we can use the original state of mem 142 Node* len = load_array_length(ary); 143 144 // Test length vs index (standard trick using unsigned compare) 145 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 146 BoolTest::mask btest = BoolTest::lt; 147 tst = _gvn.transform( new BoolNode(chk, btest) ); 148 } 149 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 150 _gvn.set_type(rc, rc->Value(&_gvn)); 151 if (!tst->is_Con()) { 152 record_for_igvn(rc); 153 } 154 set_control(_gvn.transform(new IfTrueNode(rc))); 155 // Branch to failure if out of bounds 156 { 157 PreserveJVMState pjvms(this); 158 set_control(_gvn.transform(new IfFalseNode(rc))); 159 if (C->allow_range_check_smearing()) { 160 // Do not use builtin_throw, since range checks are sometimes 161 // made more stringent by an optimistic transformation. 162 // This creates "tentative" range checks at this point, 163 // which are not guaranteed to throw exceptions. 164 // See IfNode::Ideal, is_range_check, adjust_check. 165 uncommon_trap(Deoptimization::Reason_range_check, 166 Deoptimization::Action_make_not_entrant, 167 NULL, "range_check"); 168 } else { 169 // If we have already recompiled with the range-check-widening 170 // heroic optimization turned off, then we must really be throwing 171 // range check exceptions. 172 builtin_throw(Deoptimization::Reason_range_check, idx); 173 } 174 } 175 } 176 // Check for always knowing you are throwing a range-check exception 177 if (stopped()) return top(); 178 179 // Make array address computation control dependent to prevent it 180 // from floating above the range check during loop optimizations. 181 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 182 183 if (result2 != NULL) *result2 = elemtype; 184 185 assert(ptr != top(), "top should go hand-in-hand with stopped"); 186 187 return ptr; 188 } 189 190 191 // returns IfNode 192 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 193 Node *cmp = _gvn.transform( new CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 194 Node *tst = _gvn.transform( new BoolNode( cmp, mask)); 195 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 196 return iff; 197 } 198 199 // return Region node 200 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 201 Node *region = new RegionNode(3); // 2 results 202 record_for_igvn(region); 203 region->init_req(1, iffalse); 204 region->init_req(2, iftrue ); 205 _gvn.set_type(region, Type::CONTROL); 206 region = _gvn.transform(region); 207 set_control (region); 208 return region; 209 } 210 211 212 //------------------------------helper for tableswitch------------------------- 213 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 214 // True branch, use existing map info 215 { PreserveJVMState pjvms(this); 216 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 217 set_control( iftrue ); 218 profile_switch_case(prof_table_index); 219 merge_new_path(dest_bci_if_true); 220 } 221 222 // False branch 223 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 224 set_control( iffalse ); 225 } 226 227 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 228 // True branch, use existing map info 229 { PreserveJVMState pjvms(this); 230 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 231 set_control( iffalse ); 232 profile_switch_case(prof_table_index); 233 merge_new_path(dest_bci_if_true); 234 } 235 236 // False branch 237 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 238 set_control( iftrue ); 239 } 240 241 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 242 // False branch, use existing map and control() 243 profile_switch_case(prof_table_index); 244 merge_new_path(dest_bci); 245 } 246 247 248 extern "C" { 249 static int jint_cmp(const void *i, const void *j) { 250 int a = *(jint *)i; 251 int b = *(jint *)j; 252 return a > b ? 1 : a < b ? -1 : 0; 253 } 254 } 255 256 257 // Default value for methodData switch indexing. Must be a negative value to avoid 258 // conflict with any legal switch index. 259 #define NullTableIndex -1 260 261 class SwitchRange : public StackObj { 262 // a range of integers coupled with a bci destination 263 jint _lo; // inclusive lower limit 264 jint _hi; // inclusive upper limit 265 int _dest; 266 int _table_index; // index into method data table 267 268 public: 269 jint lo() const { return _lo; } 270 jint hi() const { return _hi; } 271 int dest() const { return _dest; } 272 int table_index() const { return _table_index; } 273 bool is_singleton() const { return _lo == _hi; } 274 275 void setRange(jint lo, jint hi, int dest, int table_index) { 276 assert(lo <= hi, "must be a non-empty range"); 277 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 278 } 279 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 280 assert(lo <= hi, "must be a non-empty range"); 281 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 282 _hi = hi; 283 return true; 284 } 285 return false; 286 } 287 288 void set (jint value, int dest, int table_index) { 289 setRange(value, value, dest, table_index); 290 } 291 bool adjoin(jint value, int dest, int table_index) { 292 return adjoinRange(value, value, dest, table_index); 293 } 294 295 void print() { 296 if (is_singleton()) 297 tty->print(" {%d}=>%d", lo(), dest()); 298 else if (lo() == min_jint) 299 tty->print(" {..%d}=>%d", hi(), dest()); 300 else if (hi() == max_jint) 301 tty->print(" {%d..}=>%d", lo(), dest()); 302 else 303 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 304 } 305 }; 306 307 308 //-------------------------------do_tableswitch-------------------------------- 309 void Parse::do_tableswitch() { 310 Node* lookup = pop(); 311 312 // Get information about tableswitch 313 int default_dest = iter().get_dest_table(0); 314 int lo_index = iter().get_int_table(1); 315 int hi_index = iter().get_int_table(2); 316 int len = hi_index - lo_index + 1; 317 318 if (len < 1) { 319 // If this is a backward branch, add safepoint 320 maybe_add_safepoint(default_dest); 321 merge(default_dest); 322 return; 323 } 324 325 // generate decision tree, using trichotomy when possible 326 int rnum = len+2; 327 bool makes_backward_branch = false; 328 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 329 int rp = -1; 330 if (lo_index != min_jint) { 331 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 332 } 333 for (int j = 0; j < len; j++) { 334 jint match_int = lo_index+j; 335 int dest = iter().get_dest_table(j+3); 336 makes_backward_branch |= (dest <= bci()); 337 int table_index = method_data_update() ? j : NullTableIndex; 338 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 339 ranges[++rp].set(match_int, dest, table_index); 340 } 341 } 342 jint highest = lo_index+(len-1); 343 assert(ranges[rp].hi() == highest, ""); 344 if (highest != max_jint 345 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 346 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 347 } 348 assert(rp < len+2, "not too many ranges"); 349 350 // Safepoint in case if backward branch observed 351 if( makes_backward_branch && UseLoopSafepoints ) 352 add_safepoint(); 353 354 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 355 } 356 357 358 //------------------------------do_lookupswitch-------------------------------- 359 void Parse::do_lookupswitch() { 360 Node *lookup = pop(); // lookup value 361 // Get information about lookupswitch 362 int default_dest = iter().get_dest_table(0); 363 int len = iter().get_int_table(1); 364 365 if (len < 1) { // If this is a backward branch, add safepoint 366 maybe_add_safepoint(default_dest); 367 merge(default_dest); 368 return; 369 } 370 371 // generate decision tree, using trichotomy when possible 372 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 373 { 374 for( int j = 0; j < len; j++ ) { 375 table[j+j+0] = iter().get_int_table(2+j+j); 376 table[j+j+1] = iter().get_dest_table(2+j+j+1); 377 } 378 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 379 } 380 381 int rnum = len*2+1; 382 bool makes_backward_branch = false; 383 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 384 int rp = -1; 385 for( int j = 0; j < len; j++ ) { 386 jint match_int = table[j+j+0]; 387 int dest = table[j+j+1]; 388 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 389 int table_index = method_data_update() ? j : NullTableIndex; 390 makes_backward_branch |= (dest <= bci()); 391 if( match_int != next_lo ) { 392 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 393 } 394 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 395 ranges[++rp].set(match_int, dest, table_index); 396 } 397 } 398 jint highest = table[2*(len-1)]; 399 assert(ranges[rp].hi() == highest, ""); 400 if( highest != max_jint 401 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 402 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 403 } 404 assert(rp < rnum, "not too many ranges"); 405 406 // Safepoint in case backward branch observed 407 if( makes_backward_branch && UseLoopSafepoints ) 408 add_safepoint(); 409 410 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 411 } 412 413 //----------------------------create_jump_tables------------------------------- 414 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 415 // Are jumptables enabled 416 if (!UseJumpTables) return false; 417 418 // Are jumptables supported 419 if (!Matcher::has_match_rule(Op_Jump)) return false; 420 421 // Don't make jump table if profiling 422 if (method_data_update()) return false; 423 424 // Decide if a guard is needed to lop off big ranges at either (or 425 // both) end(s) of the input set. We'll call this the default target 426 // even though we can't be sure that it is the true "default". 427 428 bool needs_guard = false; 429 int default_dest; 430 int64_t total_outlier_size = 0; 431 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 432 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 433 434 if (lo->dest() == hi->dest()) { 435 total_outlier_size = hi_size + lo_size; 436 default_dest = lo->dest(); 437 } else if (lo_size > hi_size) { 438 total_outlier_size = lo_size; 439 default_dest = lo->dest(); 440 } else { 441 total_outlier_size = hi_size; 442 default_dest = hi->dest(); 443 } 444 445 // If a guard test will eliminate very sparse end ranges, then 446 // it is worth the cost of an extra jump. 447 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 448 needs_guard = true; 449 if (default_dest == lo->dest()) lo++; 450 if (default_dest == hi->dest()) hi--; 451 } 452 453 // Find the total number of cases and ranges 454 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 455 int num_range = hi - lo + 1; 456 457 // Don't create table if: too large, too small, or too sparse. 458 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 459 return false; 460 if (num_cases > (MaxJumpTableSparseness * num_range)) 461 return false; 462 463 // Normalize table lookups to zero 464 int lowval = lo->lo(); 465 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 466 467 // Generate a guard to protect against input keyvals that aren't 468 // in the switch domain. 469 if (needs_guard) { 470 Node* size = _gvn.intcon(num_cases); 471 Node* cmp = _gvn.transform( new CmpUNode(key_val, size) ); 472 Node* tst = _gvn.transform( new BoolNode(cmp, BoolTest::ge) ); 473 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 474 jump_if_true_fork(iff, default_dest, NullTableIndex); 475 } 476 477 // Create an ideal node JumpTable that has projections 478 // of all possible ranges for a switch statement 479 // The key_val input must be converted to a pointer offset and scaled. 480 // Compare Parse::array_addressing above. 481 482 // Clean the 32-bit int into a real 64-bit offset. 483 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 484 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 485 // Make I2L conversion control dependent to prevent it from 486 // floating above the range check during loop optimizations. 487 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 488 489 // Shift the value by wordsize so we have an index into the table, rather 490 // than a switch value 491 Node *shiftWord = _gvn.MakeConX(wordSize); 492 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 493 494 // Create the JumpNode 495 Node* jtn = _gvn.transform( new JumpNode(control(), key_val, num_cases) ); 496 497 // These are the switch destinations hanging off the jumpnode 498 int i = 0; 499 for (SwitchRange* r = lo; r <= hi; r++) { 500 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 501 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 502 { 503 PreserveJVMState pjvms(this); 504 set_control(input); 505 jump_if_always_fork(r->dest(), r->table_index()); 506 } 507 } 508 } 509 assert(i == num_cases, "miscount of cases"); 510 stop_and_kill_map(); // no more uses for this JVMS 511 return true; 512 } 513 514 //----------------------------jump_switch_ranges------------------------------- 515 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 516 Block* switch_block = block(); 517 518 if (switch_depth == 0) { 519 // Do special processing for the top-level call. 520 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 521 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 522 523 // Decrement pred-numbers for the unique set of nodes. 524 #ifdef ASSERT 525 // Ensure that the block's successors are a (duplicate-free) set. 526 int successors_counted = 0; // block occurrences in [hi..lo] 527 int unique_successors = switch_block->num_successors(); 528 for (int i = 0; i < unique_successors; i++) { 529 Block* target = switch_block->successor_at(i); 530 531 // Check that the set of successors is the same in both places. 532 int successors_found = 0; 533 for (SwitchRange* p = lo; p <= hi; p++) { 534 if (p->dest() == target->start()) successors_found++; 535 } 536 assert(successors_found > 0, "successor must be known"); 537 successors_counted += successors_found; 538 } 539 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 540 #endif 541 542 // Maybe prune the inputs, based on the type of key_val. 543 jint min_val = min_jint; 544 jint max_val = max_jint; 545 const TypeInt* ti = key_val->bottom_type()->isa_int(); 546 if (ti != NULL) { 547 min_val = ti->_lo; 548 max_val = ti->_hi; 549 assert(min_val <= max_val, "invalid int type"); 550 } 551 while (lo->hi() < min_val) lo++; 552 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 553 while (hi->lo() > max_val) hi--; 554 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 555 } 556 557 #ifndef PRODUCT 558 if (switch_depth == 0) { 559 _max_switch_depth = 0; 560 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 561 } 562 #endif 563 564 assert(lo <= hi, "must be a non-empty set of ranges"); 565 if (lo == hi) { 566 jump_if_always_fork(lo->dest(), lo->table_index()); 567 } else { 568 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 569 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 570 571 if (create_jump_tables(key_val, lo, hi)) return; 572 573 int nr = hi - lo + 1; 574 575 SwitchRange* mid = lo + nr/2; 576 // if there is an easy choice, pivot at a singleton: 577 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 578 579 assert(lo < mid && mid <= hi, "good pivot choice"); 580 assert(nr != 2 || mid == hi, "should pick higher of 2"); 581 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 582 583 Node *test_val = _gvn.intcon(mid->lo()); 584 585 if (mid->is_singleton()) { 586 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 587 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 588 589 // Special Case: If there are exactly three ranges, and the high 590 // and low range each go to the same place, omit the "gt" test, 591 // since it will not discriminate anything. 592 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 593 if (eq_test_only) { 594 assert(mid == hi-1, ""); 595 } 596 597 // if there is a higher range, test for it and process it: 598 if (mid < hi && !eq_test_only) { 599 // two comparisons of same values--should enable 1 test for 2 branches 600 // Use BoolTest::le instead of BoolTest::gt 601 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 602 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 603 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 604 { PreserveJVMState pjvms(this); 605 set_control(iffalse); 606 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 607 } 608 set_control(iftrue); 609 } 610 611 } else { 612 // mid is a range, not a singleton, so treat mid..hi as a unit 613 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 614 615 // if there is a higher range, test for it and process it: 616 if (mid == hi) { 617 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 618 } else { 619 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 620 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 621 { PreserveJVMState pjvms(this); 622 set_control(iftrue); 623 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 624 } 625 set_control(iffalse); 626 } 627 } 628 629 // in any case, process the lower range 630 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 631 } 632 633 // Decrease pred_count for each successor after all is done. 634 if (switch_depth == 0) { 635 int unique_successors = switch_block->num_successors(); 636 for (int i = 0; i < unique_successors; i++) { 637 Block* target = switch_block->successor_at(i); 638 // Throw away the pre-allocated path for each unique successor. 639 target->next_path_num(); 640 } 641 } 642 643 #ifndef PRODUCT 644 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 645 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 646 SwitchRange* r; 647 int nsing = 0; 648 for( r = lo; r <= hi; r++ ) { 649 if( r->is_singleton() ) nsing++; 650 } 651 tty->print(">>> "); 652 _method->print_short_name(); 653 tty->print_cr(" switch decision tree"); 654 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 655 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 656 if (_max_switch_depth > _est_switch_depth) { 657 tty->print_cr("******** BAD SWITCH DEPTH ********"); 658 } 659 tty->print(" "); 660 for( r = lo; r <= hi; r++ ) { 661 r->print(); 662 } 663 tty->cr(); 664 } 665 #endif 666 } 667 668 void Parse::modf() { 669 Node *f2 = pop(); 670 Node *f1 = pop(); 671 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 672 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 673 "frem", NULL, //no memory effects 674 f1, f2); 675 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 676 677 push(res); 678 } 679 680 void Parse::modd() { 681 Node *d2 = pop_pair(); 682 Node *d1 = pop_pair(); 683 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 684 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 685 "drem", NULL, //no memory effects 686 d1, top(), d2, top()); 687 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 688 689 #ifdef ASSERT 690 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 691 assert(res_top == top(), "second value must be top"); 692 #endif 693 694 push_pair(res_d); 695 } 696 697 void Parse::l2f() { 698 Node* f2 = pop(); 699 Node* f1 = pop(); 700 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 701 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 702 "l2f", NULL, //no memory effects 703 f1, f2); 704 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 705 706 push(res); 707 } 708 709 void Parse::do_irem() { 710 // Must keep both values on the expression-stack during null-check 711 zero_check_int(peek()); 712 // Compile-time detect of null-exception? 713 if (stopped()) return; 714 715 Node* b = pop(); 716 Node* a = pop(); 717 718 const Type *t = _gvn.type(b); 719 if (t != Type::TOP) { 720 const TypeInt *ti = t->is_int(); 721 if (ti->is_con()) { 722 int divisor = ti->get_con(); 723 // check for positive power of 2 724 if (divisor > 0 && 725 (divisor & ~(divisor-1)) == divisor) { 726 // yes ! 727 Node *mask = _gvn.intcon((divisor - 1)); 728 // Sigh, must handle negative dividends 729 Node *zero = _gvn.intcon(0); 730 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 731 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 732 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 733 Node *reg = jump_if_join(ift, iff); 734 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 735 // Negative path; negate/and/negate 736 Node *neg = _gvn.transform( new SubINode(zero, a) ); 737 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 738 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 739 phi->init_req(1, negn); 740 // Fast positive case 741 Node *andx = _gvn.transform( new AndINode(a, mask) ); 742 phi->init_req(2, andx); 743 // Push the merge 744 push( _gvn.transform(phi) ); 745 return; 746 } 747 } 748 } 749 // Default case 750 push( _gvn.transform( new ModINode(control(),a,b) ) ); 751 } 752 753 // Handle jsr and jsr_w bytecode 754 void Parse::do_jsr() { 755 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 756 757 // Store information about current state, tagged with new _jsr_bci 758 int return_bci = iter().next_bci(); 759 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 760 761 // Update method data 762 profile_taken_branch(jsr_bci); 763 764 // The way we do things now, there is only one successor block 765 // for the jsr, because the target code is cloned by ciTypeFlow. 766 Block* target = successor_for_bci(jsr_bci); 767 768 // What got pushed? 769 const Type* ret_addr = target->peek(); 770 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 771 772 // Effect on jsr on stack 773 push(_gvn.makecon(ret_addr)); 774 775 // Flow to the jsr. 776 merge(jsr_bci); 777 } 778 779 // Handle ret bytecode 780 void Parse::do_ret() { 781 // Find to whom we return. 782 assert(block()->num_successors() == 1, "a ret can only go one place now"); 783 Block* target = block()->successor_at(0); 784 assert(!target->is_ready(), "our arrival must be expected"); 785 profile_ret(target->flow()->start()); 786 int pnum = target->next_path_num(); 787 merge_common(target, pnum); 788 } 789 790 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 791 if (btest != BoolTest::eq && btest != BoolTest::ne) { 792 // Only ::eq and ::ne are supported for profile injection. 793 return false; 794 } 795 if (test->is_Cmp() && 796 test->in(1)->Opcode() == Op_ProfileBoolean) { 797 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 798 int false_cnt = profile->false_count(); 799 int true_cnt = profile->true_count(); 800 801 // Counts matching depends on the actual test operation (::eq or ::ne). 802 // No need to scale the counts because profile injection was designed 803 // to feed exact counts into VM. 804 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 805 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 806 807 profile->consume(); 808 return true; 809 } 810 return false; 811 } 812 //--------------------------dynamic_branch_prediction-------------------------- 813 // Try to gather dynamic branch prediction behavior. Return a probability 814 // of the branch being taken and set the "cnt" field. Returns a -1.0 815 // if we need to use static prediction for some reason. 816 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 817 ResourceMark rm; 818 819 cnt = COUNT_UNKNOWN; 820 821 int taken = 0; 822 int not_taken = 0; 823 824 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 825 826 if (use_mdo) { 827 // Use MethodData information if it is available 828 // FIXME: free the ProfileData structure 829 ciMethodData* methodData = method()->method_data(); 830 if (!methodData->is_mature()) return PROB_UNKNOWN; 831 ciProfileData* data = methodData->bci_to_data(bci()); 832 if (data == NULL) { 833 return PROB_UNKNOWN; 834 } 835 if (!data->is_JumpData()) return PROB_UNKNOWN; 836 837 // get taken and not taken values 838 taken = data->as_JumpData()->taken(); 839 not_taken = 0; 840 if (data->is_BranchData()) { 841 not_taken = data->as_BranchData()->not_taken(); 842 } 843 844 // scale the counts to be commensurate with invocation counts: 845 taken = method()->scale_count(taken); 846 not_taken = method()->scale_count(not_taken); 847 } 848 849 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 850 // We also check that individual counters are positive first, otherwise the sum can become positive. 851 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 852 if (C->log() != NULL) { 853 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 854 } 855 return PROB_UNKNOWN; 856 } 857 858 // Compute frequency that we arrive here 859 float sum = taken + not_taken; 860 // Adjust, if this block is a cloned private block but the 861 // Jump counts are shared. Taken the private counts for 862 // just this path instead of the shared counts. 863 if( block()->count() > 0 ) 864 sum = block()->count(); 865 cnt = sum / FreqCountInvocations; 866 867 // Pin probability to sane limits 868 float prob; 869 if( !taken ) 870 prob = (0+PROB_MIN) / 2; 871 else if( !not_taken ) 872 prob = (1+PROB_MAX) / 2; 873 else { // Compute probability of true path 874 prob = (float)taken / (float)(taken + not_taken); 875 if (prob > PROB_MAX) prob = PROB_MAX; 876 if (prob < PROB_MIN) prob = PROB_MIN; 877 } 878 879 assert((cnt > 0.0f) && (prob > 0.0f), 880 "Bad frequency assignment in if"); 881 882 if (C->log() != NULL) { 883 const char* prob_str = NULL; 884 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 885 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 886 char prob_str_buf[30]; 887 if (prob_str == NULL) { 888 sprintf(prob_str_buf, "%g", prob); 889 prob_str = prob_str_buf; 890 } 891 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 892 iter().get_dest(), taken, not_taken, cnt, prob_str); 893 } 894 return prob; 895 } 896 897 //-----------------------------branch_prediction------------------------------- 898 float Parse::branch_prediction(float& cnt, 899 BoolTest::mask btest, 900 int target_bci, 901 Node* test) { 902 float prob = dynamic_branch_prediction(cnt, btest, test); 903 // If prob is unknown, switch to static prediction 904 if (prob != PROB_UNKNOWN) return prob; 905 906 prob = PROB_FAIR; // Set default value 907 if (btest == BoolTest::eq) // Exactly equal test? 908 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 909 else if (btest == BoolTest::ne) 910 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 911 912 // If this is a conditional test guarding a backwards branch, 913 // assume its a loop-back edge. Make it a likely taken branch. 914 if (target_bci < bci()) { 915 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 916 // Since it's an OSR, we probably have profile data, but since 917 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 918 // Let's make a special check here for completely zero counts. 919 ciMethodData* methodData = method()->method_data(); 920 if (!methodData->is_empty()) { 921 ciProfileData* data = methodData->bci_to_data(bci()); 922 // Only stop for truly zero counts, which mean an unknown part 923 // of the OSR-ed method, and we want to deopt to gather more stats. 924 // If you have ANY counts, then this loop is simply 'cold' relative 925 // to the OSR loop. 926 if (data == NULL || 927 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 928 // This is the only way to return PROB_UNKNOWN: 929 return PROB_UNKNOWN; 930 } 931 } 932 } 933 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 934 } 935 936 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 937 return prob; 938 } 939 940 // The magic constants are chosen so as to match the output of 941 // branch_prediction() when the profile reports a zero taken count. 942 // It is important to distinguish zero counts unambiguously, because 943 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 944 // very small but nonzero probabilities, which if confused with zero 945 // counts would keep the program recompiling indefinitely. 946 bool Parse::seems_never_taken(float prob) const { 947 return prob < PROB_MIN; 948 } 949 950 // True if the comparison seems to be the kind that will not change its 951 // statistics from true to false. See comments in adjust_map_after_if. 952 // This question is only asked along paths which are already 953 // classifed as untaken (by seems_never_taken), so really, 954 // if a path is never taken, its controlling comparison is 955 // already acting in a stable fashion. If the comparison 956 // seems stable, we will put an expensive uncommon trap 957 // on the untaken path. 958 bool Parse::seems_stable_comparison() const { 959 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 960 return false; 961 } 962 return true; 963 } 964 965 //-------------------------------repush_if_args-------------------------------- 966 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 967 inline int Parse::repush_if_args() { 968 if (PrintOpto && WizardMode) { 969 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 970 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 971 method()->print_name(); tty->cr(); 972 } 973 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 974 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 975 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 976 assert(argument(0) != NULL, "must exist"); 977 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 978 inc_sp(bc_depth); 979 return bc_depth; 980 } 981 982 //----------------------------------do_ifnull---------------------------------- 983 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 984 int target_bci = iter().get_dest(); 985 986 Block* branch_block = successor_for_bci(target_bci); 987 Block* next_block = successor_for_bci(iter().next_bci()); 988 989 float cnt; 990 float prob = branch_prediction(cnt, btest, target_bci, c); 991 if (prob == PROB_UNKNOWN) { 992 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 993 if (PrintOpto && Verbose) { 994 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 995 } 996 repush_if_args(); // to gather stats on loop 997 // We need to mark this branch as taken so that if we recompile we will 998 // see that it is possible. In the tiered system the interpreter doesn't 999 // do profiling and by the time we get to the lower tier from the interpreter 1000 // the path may be cold again. Make sure it doesn't look untaken 1001 profile_taken_branch(target_bci, !ProfileInterpreter); 1002 uncommon_trap(Deoptimization::Reason_unreached, 1003 Deoptimization::Action_reinterpret, 1004 NULL, "cold"); 1005 if (C->eliminate_boxing()) { 1006 // Mark the successor blocks as parsed 1007 branch_block->next_path_num(); 1008 next_block->next_path_num(); 1009 } 1010 return; 1011 } 1012 1013 NOT_PRODUCT(explicit_null_checks_inserted++); 1014 1015 // Generate real control flow 1016 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1017 1018 // Sanity check the probability value 1019 assert(prob > 0.0f,"Bad probability in Parser"); 1020 // Need xform to put node in hash table 1021 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1022 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1023 // True branch 1024 { PreserveJVMState pjvms(this); 1025 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1026 set_control(iftrue); 1027 1028 if (stopped()) { // Path is dead? 1029 NOT_PRODUCT(explicit_null_checks_elided++); 1030 if (C->eliminate_boxing()) { 1031 // Mark the successor block as parsed 1032 branch_block->next_path_num(); 1033 } 1034 } else { // Path is live. 1035 // Update method data 1036 profile_taken_branch(target_bci); 1037 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1038 if (!stopped()) { 1039 merge(target_bci); 1040 } 1041 } 1042 } 1043 1044 // False branch 1045 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1046 set_control(iffalse); 1047 1048 if (stopped()) { // Path is dead? 1049 NOT_PRODUCT(explicit_null_checks_elided++); 1050 if (C->eliminate_boxing()) { 1051 // Mark the successor block as parsed 1052 next_block->next_path_num(); 1053 } 1054 } else { // Path is live. 1055 // Update method data 1056 profile_not_taken_branch(); 1057 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1058 next_block, branch_block); 1059 } 1060 } 1061 1062 //------------------------------------do_if------------------------------------ 1063 void Parse::do_if(BoolTest::mask btest, Node* c) { 1064 int target_bci = iter().get_dest(); 1065 1066 Block* branch_block = successor_for_bci(target_bci); 1067 Block* next_block = successor_for_bci(iter().next_bci()); 1068 1069 float cnt; 1070 float prob = branch_prediction(cnt, btest, target_bci, c); 1071 float untaken_prob = 1.0 - prob; 1072 1073 if (prob == PROB_UNKNOWN) { 1074 if (PrintOpto && Verbose) { 1075 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1076 } 1077 repush_if_args(); // to gather stats on loop 1078 // We need to mark this branch as taken so that if we recompile we will 1079 // see that it is possible. In the tiered system the interpreter doesn't 1080 // do profiling and by the time we get to the lower tier from the interpreter 1081 // the path may be cold again. Make sure it doesn't look untaken 1082 profile_taken_branch(target_bci, !ProfileInterpreter); 1083 uncommon_trap(Deoptimization::Reason_unreached, 1084 Deoptimization::Action_reinterpret, 1085 NULL, "cold"); 1086 if (C->eliminate_boxing()) { 1087 // Mark the successor blocks as parsed 1088 branch_block->next_path_num(); 1089 next_block->next_path_num(); 1090 } 1091 return; 1092 } 1093 1094 // Sanity check the probability value 1095 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1096 1097 bool taken_if_true = true; 1098 // Convert BoolTest to canonical form: 1099 if (!BoolTest(btest).is_canonical()) { 1100 btest = BoolTest(btest).negate(); 1101 taken_if_true = false; 1102 // prob is NOT updated here; it remains the probability of the taken 1103 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1104 } 1105 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1106 1107 Node* tst0 = new BoolNode(c, btest); 1108 Node* tst = _gvn.transform(tst0); 1109 BoolTest::mask taken_btest = BoolTest::illegal; 1110 BoolTest::mask untaken_btest = BoolTest::illegal; 1111 1112 if (tst->is_Bool()) { 1113 // Refresh c from the transformed bool node, since it may be 1114 // simpler than the original c. Also re-canonicalize btest. 1115 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1116 // That can arise from statements like: if (x instanceof C) ... 1117 if (tst != tst0) { 1118 // Canonicalize one more time since transform can change it. 1119 btest = tst->as_Bool()->_test._test; 1120 if (!BoolTest(btest).is_canonical()) { 1121 // Reverse edges one more time... 1122 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1123 btest = tst->as_Bool()->_test._test; 1124 assert(BoolTest(btest).is_canonical(), "sanity"); 1125 taken_if_true = !taken_if_true; 1126 } 1127 c = tst->in(1); 1128 } 1129 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1130 taken_btest = taken_if_true ? btest : neg_btest; 1131 untaken_btest = taken_if_true ? neg_btest : btest; 1132 } 1133 1134 // Generate real control flow 1135 float true_prob = (taken_if_true ? prob : untaken_prob); 1136 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1137 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1138 Node* taken_branch = new IfTrueNode(iff); 1139 Node* untaken_branch = new IfFalseNode(iff); 1140 if (!taken_if_true) { // Finish conversion to canonical form 1141 Node* tmp = taken_branch; 1142 taken_branch = untaken_branch; 1143 untaken_branch = tmp; 1144 } 1145 1146 // Branch is taken: 1147 { PreserveJVMState pjvms(this); 1148 taken_branch = _gvn.transform(taken_branch); 1149 set_control(taken_branch); 1150 1151 if (stopped()) { 1152 if (C->eliminate_boxing()) { 1153 // Mark the successor block as parsed 1154 branch_block->next_path_num(); 1155 } 1156 } else { 1157 // Update method data 1158 profile_taken_branch(target_bci); 1159 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1160 if (!stopped()) { 1161 merge(target_bci); 1162 } 1163 } 1164 } 1165 1166 untaken_branch = _gvn.transform(untaken_branch); 1167 set_control(untaken_branch); 1168 1169 // Branch not taken. 1170 if (stopped()) { 1171 if (C->eliminate_boxing()) { 1172 // Mark the successor block as parsed 1173 next_block->next_path_num(); 1174 } 1175 } else { 1176 // Update method data 1177 profile_not_taken_branch(); 1178 adjust_map_after_if(untaken_btest, c, untaken_prob, 1179 next_block, branch_block); 1180 } 1181 } 1182 1183 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1184 // Don't want to speculate on uncommon traps when running with -Xcomp 1185 if (!UseInterpreter) { 1186 return false; 1187 } 1188 return (seems_never_taken(prob) && seems_stable_comparison()); 1189 } 1190 1191 //----------------------------adjust_map_after_if------------------------------ 1192 // Adjust the JVM state to reflect the result of taking this path. 1193 // Basically, it means inspecting the CmpNode controlling this 1194 // branch, seeing how it constrains a tested value, and then 1195 // deciding if it's worth our while to encode this constraint 1196 // as graph nodes in the current abstract interpretation map. 1197 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1198 Block* path, Block* other_path) { 1199 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1200 return; // nothing to do 1201 1202 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1203 1204 if (path_is_suitable_for_uncommon_trap(prob)) { 1205 repush_if_args(); 1206 uncommon_trap(Deoptimization::Reason_unstable_if, 1207 Deoptimization::Action_reinterpret, 1208 NULL, 1209 (is_fallthrough ? "taken always" : "taken never")); 1210 return; 1211 } 1212 1213 Node* val = c->in(1); 1214 Node* con = c->in(2); 1215 const Type* tcon = _gvn.type(con); 1216 const Type* tval = _gvn.type(val); 1217 bool have_con = tcon->singleton(); 1218 if (tval->singleton()) { 1219 if (!have_con) { 1220 // Swap, so constant is in con. 1221 con = val; 1222 tcon = tval; 1223 val = c->in(2); 1224 tval = _gvn.type(val); 1225 btest = BoolTest(btest).commute(); 1226 have_con = true; 1227 } else { 1228 // Do we have two constants? Then leave well enough alone. 1229 have_con = false; 1230 } 1231 } 1232 if (!have_con) // remaining adjustments need a con 1233 return; 1234 1235 sharpen_type_after_if(btest, con, tcon, val, tval); 1236 } 1237 1238 1239 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1240 Node* ldk; 1241 if (n->is_DecodeNKlass()) { 1242 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1243 return NULL; 1244 } else { 1245 ldk = n->in(1); 1246 } 1247 } else if (n->Opcode() != Op_LoadKlass) { 1248 return NULL; 1249 } else { 1250 ldk = n; 1251 } 1252 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1253 1254 Node* adr = ldk->in(MemNode::Address); 1255 intptr_t off = 0; 1256 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1257 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1258 return NULL; 1259 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1260 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1261 return NULL; 1262 1263 return obj; 1264 } 1265 1266 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1267 Node* con, const Type* tcon, 1268 Node* val, const Type* tval) { 1269 // Look for opportunities to sharpen the type of a node 1270 // whose klass is compared with a constant klass. 1271 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1272 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1273 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1274 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1275 // Found: 1276 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1277 // or the narrowOop equivalent. 1278 const Type* obj_type = _gvn.type(obj); 1279 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1280 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1281 tboth->higher_equal(obj_type)) { 1282 // obj has to be of the exact type Foo if the CmpP succeeds. 1283 int obj_in_map = map()->find_edge(obj); 1284 JVMState* jvms = this->jvms(); 1285 if (obj_in_map >= 0 && 1286 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1287 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1288 const Type* tcc = ccast->as_Type()->type(); 1289 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1290 // Delay transform() call to allow recovery of pre-cast value 1291 // at the control merge. 1292 _gvn.set_type_bottom(ccast); 1293 record_for_igvn(ccast); 1294 // Here's the payoff. 1295 replace_in_map(obj, ccast); 1296 } 1297 } 1298 } 1299 } 1300 1301 int val_in_map = map()->find_edge(val); 1302 if (val_in_map < 0) return; // replace_in_map would be useless 1303 { 1304 JVMState* jvms = this->jvms(); 1305 if (!(jvms->is_loc(val_in_map) || 1306 jvms->is_stk(val_in_map))) 1307 return; // again, it would be useless 1308 } 1309 1310 // Check for a comparison to a constant, and "know" that the compared 1311 // value is constrained on this path. 1312 assert(tcon->singleton(), ""); 1313 ConstraintCastNode* ccast = NULL; 1314 Node* cast = NULL; 1315 1316 switch (btest) { 1317 case BoolTest::eq: // Constant test? 1318 { 1319 const Type* tboth = tcon->join_speculative(tval); 1320 if (tboth == tval) break; // Nothing to gain. 1321 if (tcon->isa_int()) { 1322 ccast = new CastIINode(val, tboth); 1323 } else if (tcon == TypePtr::NULL_PTR) { 1324 // Cast to null, but keep the pointer identity temporarily live. 1325 ccast = new CastPPNode(val, tboth); 1326 } else { 1327 const TypeF* tf = tcon->isa_float_constant(); 1328 const TypeD* td = tcon->isa_double_constant(); 1329 // Exclude tests vs float/double 0 as these could be 1330 // either +0 or -0. Just because you are equal to +0 1331 // doesn't mean you ARE +0! 1332 // Note, following code also replaces Long and Oop values. 1333 if ((!tf || tf->_f != 0.0) && 1334 (!td || td->_d != 0.0)) 1335 cast = con; // Replace non-constant val by con. 1336 } 1337 } 1338 break; 1339 1340 case BoolTest::ne: 1341 if (tcon == TypePtr::NULL_PTR) { 1342 cast = cast_not_null(val, false); 1343 } 1344 break; 1345 1346 default: 1347 // (At this point we could record int range types with CastII.) 1348 break; 1349 } 1350 1351 if (ccast != NULL) { 1352 const Type* tcc = ccast->as_Type()->type(); 1353 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1354 // Delay transform() call to allow recovery of pre-cast value 1355 // at the control merge. 1356 ccast->set_req(0, control()); 1357 _gvn.set_type_bottom(ccast); 1358 record_for_igvn(ccast); 1359 cast = ccast; 1360 } 1361 1362 if (cast != NULL) { // Here's the payoff. 1363 replace_in_map(val, cast); 1364 } 1365 } 1366 1367 /** 1368 * Use speculative type to optimize CmpP node: if comparison is 1369 * against the low level class, cast the object to the speculative 1370 * type if any. CmpP should then go away. 1371 * 1372 * @param c expected CmpP node 1373 * @return result of CmpP on object casted to speculative type 1374 * 1375 */ 1376 Node* Parse::optimize_cmp_with_klass(Node* c) { 1377 // If this is transformed by the _gvn to a comparison with the low 1378 // level klass then we may be able to use speculation 1379 if (c->Opcode() == Op_CmpP && 1380 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1381 c->in(2)->is_Con()) { 1382 Node* load_klass = NULL; 1383 Node* decode = NULL; 1384 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1385 decode = c->in(1); 1386 load_klass = c->in(1)->in(1); 1387 } else { 1388 load_klass = c->in(1); 1389 } 1390 if (load_klass->in(2)->is_AddP()) { 1391 Node* addp = load_klass->in(2); 1392 Node* obj = addp->in(AddPNode::Address); 1393 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1394 if (obj_type->speculative_type_not_null() != NULL) { 1395 ciKlass* k = obj_type->speculative_type(); 1396 inc_sp(2); 1397 obj = maybe_cast_profiled_obj(obj, k); 1398 dec_sp(2); 1399 // Make the CmpP use the casted obj 1400 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1401 load_klass = load_klass->clone(); 1402 load_klass->set_req(2, addp); 1403 load_klass = _gvn.transform(load_klass); 1404 if (decode != NULL) { 1405 decode = decode->clone(); 1406 decode->set_req(1, load_klass); 1407 load_klass = _gvn.transform(decode); 1408 } 1409 c = c->clone(); 1410 c->set_req(1, load_klass); 1411 c = _gvn.transform(c); 1412 } 1413 } 1414 } 1415 return c; 1416 } 1417 1418 //------------------------------do_one_bytecode-------------------------------- 1419 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1420 void Parse::do_one_bytecode() { 1421 Node *a, *b, *c, *d; // Handy temps 1422 BoolTest::mask btest; 1423 int i; 1424 1425 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1426 1427 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1428 "out of nodes parsing method")) { 1429 return; 1430 } 1431 1432 #ifdef ASSERT 1433 // for setting breakpoints 1434 if (TraceOptoParse) { 1435 tty->print(" @"); 1436 dump_bci(bci()); 1437 tty->cr(); 1438 } 1439 #endif 1440 1441 switch (bc()) { 1442 case Bytecodes::_nop: 1443 // do nothing 1444 break; 1445 case Bytecodes::_lconst_0: 1446 push_pair(longcon(0)); 1447 break; 1448 1449 case Bytecodes::_lconst_1: 1450 push_pair(longcon(1)); 1451 break; 1452 1453 case Bytecodes::_fconst_0: 1454 push(zerocon(T_FLOAT)); 1455 break; 1456 1457 case Bytecodes::_fconst_1: 1458 push(makecon(TypeF::ONE)); 1459 break; 1460 1461 case Bytecodes::_fconst_2: 1462 push(makecon(TypeF::make(2.0f))); 1463 break; 1464 1465 case Bytecodes::_dconst_0: 1466 push_pair(zerocon(T_DOUBLE)); 1467 break; 1468 1469 case Bytecodes::_dconst_1: 1470 push_pair(makecon(TypeD::ONE)); 1471 break; 1472 1473 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1474 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1475 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1476 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1477 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1478 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1479 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1480 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1481 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1482 case Bytecodes::_aconst_null: push(null()); break; 1483 case Bytecodes::_ldc: 1484 case Bytecodes::_ldc_w: 1485 case Bytecodes::_ldc2_w: 1486 // If the constant is unresolved, run this BC once in the interpreter. 1487 { 1488 ciConstant constant = iter().get_constant(); 1489 if (!constant.is_valid() || 1490 (constant.basic_type() == T_OBJECT && 1491 !constant.as_object()->is_loaded())) { 1492 int index = iter().get_constant_pool_index(); 1493 constantTag tag = iter().get_constant_pool_tag(index); 1494 uncommon_trap(Deoptimization::make_trap_request 1495 (Deoptimization::Reason_unloaded, 1496 Deoptimization::Action_reinterpret, 1497 index), 1498 NULL, tag.internal_name()); 1499 break; 1500 } 1501 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1502 "must be java_mirror of klass"); 1503 const Type* con_type = Type::make_from_constant(constant); 1504 if (con_type != NULL) { 1505 push_node(con_type->basic_type(), makecon(con_type)); 1506 } 1507 } 1508 1509 break; 1510 1511 case Bytecodes::_aload_0: 1512 push( local(0) ); 1513 break; 1514 case Bytecodes::_aload_1: 1515 push( local(1) ); 1516 break; 1517 case Bytecodes::_aload_2: 1518 push( local(2) ); 1519 break; 1520 case Bytecodes::_aload_3: 1521 push( local(3) ); 1522 break; 1523 case Bytecodes::_aload: 1524 push( local(iter().get_index()) ); 1525 break; 1526 1527 case Bytecodes::_fload_0: 1528 case Bytecodes::_iload_0: 1529 push( local(0) ); 1530 break; 1531 case Bytecodes::_fload_1: 1532 case Bytecodes::_iload_1: 1533 push( local(1) ); 1534 break; 1535 case Bytecodes::_fload_2: 1536 case Bytecodes::_iload_2: 1537 push( local(2) ); 1538 break; 1539 case Bytecodes::_fload_3: 1540 case Bytecodes::_iload_3: 1541 push( local(3) ); 1542 break; 1543 case Bytecodes::_fload: 1544 case Bytecodes::_iload: 1545 push( local(iter().get_index()) ); 1546 break; 1547 case Bytecodes::_lload_0: 1548 push_pair_local( 0 ); 1549 break; 1550 case Bytecodes::_lload_1: 1551 push_pair_local( 1 ); 1552 break; 1553 case Bytecodes::_lload_2: 1554 push_pair_local( 2 ); 1555 break; 1556 case Bytecodes::_lload_3: 1557 push_pair_local( 3 ); 1558 break; 1559 case Bytecodes::_lload: 1560 push_pair_local( iter().get_index() ); 1561 break; 1562 1563 case Bytecodes::_dload_0: 1564 push_pair_local(0); 1565 break; 1566 case Bytecodes::_dload_1: 1567 push_pair_local(1); 1568 break; 1569 case Bytecodes::_dload_2: 1570 push_pair_local(2); 1571 break; 1572 case Bytecodes::_dload_3: 1573 push_pair_local(3); 1574 break; 1575 case Bytecodes::_dload: 1576 push_pair_local(iter().get_index()); 1577 break; 1578 case Bytecodes::_fstore_0: 1579 case Bytecodes::_istore_0: 1580 case Bytecodes::_astore_0: 1581 set_local( 0, pop() ); 1582 break; 1583 case Bytecodes::_fstore_1: 1584 case Bytecodes::_istore_1: 1585 case Bytecodes::_astore_1: 1586 set_local( 1, pop() ); 1587 break; 1588 case Bytecodes::_fstore_2: 1589 case Bytecodes::_istore_2: 1590 case Bytecodes::_astore_2: 1591 set_local( 2, pop() ); 1592 break; 1593 case Bytecodes::_fstore_3: 1594 case Bytecodes::_istore_3: 1595 case Bytecodes::_astore_3: 1596 set_local( 3, pop() ); 1597 break; 1598 case Bytecodes::_fstore: 1599 case Bytecodes::_istore: 1600 case Bytecodes::_astore: 1601 set_local( iter().get_index(), pop() ); 1602 break; 1603 // long stores 1604 case Bytecodes::_lstore_0: 1605 set_pair_local( 0, pop_pair() ); 1606 break; 1607 case Bytecodes::_lstore_1: 1608 set_pair_local( 1, pop_pair() ); 1609 break; 1610 case Bytecodes::_lstore_2: 1611 set_pair_local( 2, pop_pair() ); 1612 break; 1613 case Bytecodes::_lstore_3: 1614 set_pair_local( 3, pop_pair() ); 1615 break; 1616 case Bytecodes::_lstore: 1617 set_pair_local( iter().get_index(), pop_pair() ); 1618 break; 1619 1620 // double stores 1621 case Bytecodes::_dstore_0: 1622 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1623 break; 1624 case Bytecodes::_dstore_1: 1625 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1626 break; 1627 case Bytecodes::_dstore_2: 1628 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1629 break; 1630 case Bytecodes::_dstore_3: 1631 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1632 break; 1633 case Bytecodes::_dstore: 1634 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1635 break; 1636 1637 case Bytecodes::_pop: dec_sp(1); break; 1638 case Bytecodes::_pop2: dec_sp(2); break; 1639 case Bytecodes::_swap: 1640 a = pop(); 1641 b = pop(); 1642 push(a); 1643 push(b); 1644 break; 1645 case Bytecodes::_dup: 1646 a = pop(); 1647 push(a); 1648 push(a); 1649 break; 1650 case Bytecodes::_dup_x1: 1651 a = pop(); 1652 b = pop(); 1653 push( a ); 1654 push( b ); 1655 push( a ); 1656 break; 1657 case Bytecodes::_dup_x2: 1658 a = pop(); 1659 b = pop(); 1660 c = pop(); 1661 push( a ); 1662 push( c ); 1663 push( b ); 1664 push( a ); 1665 break; 1666 case Bytecodes::_dup2: 1667 a = pop(); 1668 b = pop(); 1669 push( b ); 1670 push( a ); 1671 push( b ); 1672 push( a ); 1673 break; 1674 1675 case Bytecodes::_dup2_x1: 1676 // before: .. c, b, a 1677 // after: .. b, a, c, b, a 1678 // not tested 1679 a = pop(); 1680 b = pop(); 1681 c = pop(); 1682 push( b ); 1683 push( a ); 1684 push( c ); 1685 push( b ); 1686 push( a ); 1687 break; 1688 case Bytecodes::_dup2_x2: 1689 // before: .. d, c, b, a 1690 // after: .. b, a, d, c, b, a 1691 // not tested 1692 a = pop(); 1693 b = pop(); 1694 c = pop(); 1695 d = pop(); 1696 push( b ); 1697 push( a ); 1698 push( d ); 1699 push( c ); 1700 push( b ); 1701 push( a ); 1702 break; 1703 1704 case Bytecodes::_arraylength: { 1705 // Must do null-check with value on expression stack 1706 Node *ary = null_check(peek(), T_ARRAY); 1707 // Compile-time detect of null-exception? 1708 if (stopped()) return; 1709 a = pop(); 1710 push(load_array_length(a)); 1711 break; 1712 } 1713 1714 case Bytecodes::_baload: array_load(T_BYTE); break; 1715 case Bytecodes::_caload: array_load(T_CHAR); break; 1716 case Bytecodes::_iaload: array_load(T_INT); break; 1717 case Bytecodes::_saload: array_load(T_SHORT); break; 1718 case Bytecodes::_faload: array_load(T_FLOAT); break; 1719 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1720 case Bytecodes::_laload: { 1721 a = array_addressing(T_LONG, 0); 1722 if (stopped()) return; // guaranteed null or range check 1723 dec_sp(2); // Pop array and index 1724 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1725 break; 1726 } 1727 case Bytecodes::_daload: { 1728 a = array_addressing(T_DOUBLE, 0); 1729 if (stopped()) return; // guaranteed null or range check 1730 dec_sp(2); // Pop array and index 1731 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1732 break; 1733 } 1734 case Bytecodes::_bastore: array_store(T_BYTE); break; 1735 case Bytecodes::_castore: array_store(T_CHAR); break; 1736 case Bytecodes::_iastore: array_store(T_INT); break; 1737 case Bytecodes::_sastore: array_store(T_SHORT); break; 1738 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1739 case Bytecodes::_aastore: { 1740 d = array_addressing(T_OBJECT, 1); 1741 if (stopped()) return; // guaranteed null or range check 1742 array_store_check(); 1743 c = pop(); // Oop to store 1744 b = pop(); // index (already used) 1745 a = pop(); // the array itself 1746 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1747 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1748 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 1749 StoreNode::release_if_reference(T_OBJECT)); 1750 break; 1751 } 1752 case Bytecodes::_lastore: { 1753 a = array_addressing(T_LONG, 2); 1754 if (stopped()) return; // guaranteed null or range check 1755 c = pop_pair(); 1756 dec_sp(2); // Pop array and index 1757 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1758 break; 1759 } 1760 case Bytecodes::_dastore: { 1761 a = array_addressing(T_DOUBLE, 2); 1762 if (stopped()) return; // guaranteed null or range check 1763 c = pop_pair(); 1764 dec_sp(2); // Pop array and index 1765 c = dstore_rounding(c); 1766 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1767 break; 1768 } 1769 case Bytecodes::_getfield: 1770 do_getfield(); 1771 break; 1772 1773 case Bytecodes::_getstatic: 1774 do_getstatic(); 1775 break; 1776 1777 case Bytecodes::_putfield: 1778 do_putfield(); 1779 break; 1780 1781 case Bytecodes::_putstatic: 1782 do_putstatic(); 1783 break; 1784 1785 case Bytecodes::_irem: 1786 do_irem(); 1787 break; 1788 case Bytecodes::_idiv: 1789 // Must keep both values on the expression-stack during null-check 1790 zero_check_int(peek()); 1791 // Compile-time detect of null-exception? 1792 if (stopped()) return; 1793 b = pop(); 1794 a = pop(); 1795 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1796 break; 1797 case Bytecodes::_imul: 1798 b = pop(); a = pop(); 1799 push( _gvn.transform( new MulINode(a,b) ) ); 1800 break; 1801 case Bytecodes::_iadd: 1802 b = pop(); a = pop(); 1803 push( _gvn.transform( new AddINode(a,b) ) ); 1804 break; 1805 case Bytecodes::_ineg: 1806 a = pop(); 1807 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1808 break; 1809 case Bytecodes::_isub: 1810 b = pop(); a = pop(); 1811 push( _gvn.transform( new SubINode(a,b) ) ); 1812 break; 1813 case Bytecodes::_iand: 1814 b = pop(); a = pop(); 1815 push( _gvn.transform( new AndINode(a,b) ) ); 1816 break; 1817 case Bytecodes::_ior: 1818 b = pop(); a = pop(); 1819 push( _gvn.transform( new OrINode(a,b) ) ); 1820 break; 1821 case Bytecodes::_ixor: 1822 b = pop(); a = pop(); 1823 push( _gvn.transform( new XorINode(a,b) ) ); 1824 break; 1825 case Bytecodes::_ishl: 1826 b = pop(); a = pop(); 1827 push( _gvn.transform( new LShiftINode(a,b) ) ); 1828 break; 1829 case Bytecodes::_ishr: 1830 b = pop(); a = pop(); 1831 push( _gvn.transform( new RShiftINode(a,b) ) ); 1832 break; 1833 case Bytecodes::_iushr: 1834 b = pop(); a = pop(); 1835 push( _gvn.transform( new URShiftINode(a,b) ) ); 1836 break; 1837 1838 case Bytecodes::_fneg: 1839 a = pop(); 1840 b = _gvn.transform(new NegFNode (a)); 1841 push(b); 1842 break; 1843 1844 case Bytecodes::_fsub: 1845 b = pop(); 1846 a = pop(); 1847 c = _gvn.transform( new SubFNode(a,b) ); 1848 d = precision_rounding(c); 1849 push( d ); 1850 break; 1851 1852 case Bytecodes::_fadd: 1853 b = pop(); 1854 a = pop(); 1855 c = _gvn.transform( new AddFNode(a,b) ); 1856 d = precision_rounding(c); 1857 push( d ); 1858 break; 1859 1860 case Bytecodes::_fmul: 1861 b = pop(); 1862 a = pop(); 1863 c = _gvn.transform( new MulFNode(a,b) ); 1864 d = precision_rounding(c); 1865 push( d ); 1866 break; 1867 1868 case Bytecodes::_fdiv: 1869 b = pop(); 1870 a = pop(); 1871 c = _gvn.transform( new DivFNode(0,a,b) ); 1872 d = precision_rounding(c); 1873 push( d ); 1874 break; 1875 1876 case Bytecodes::_frem: 1877 if (Matcher::has_match_rule(Op_ModF)) { 1878 // Generate a ModF node. 1879 b = pop(); 1880 a = pop(); 1881 c = _gvn.transform( new ModFNode(0,a,b) ); 1882 d = precision_rounding(c); 1883 push( d ); 1884 } 1885 else { 1886 // Generate a call. 1887 modf(); 1888 } 1889 break; 1890 1891 case Bytecodes::_fcmpl: 1892 b = pop(); 1893 a = pop(); 1894 c = _gvn.transform( new CmpF3Node( a, b)); 1895 push(c); 1896 break; 1897 case Bytecodes::_fcmpg: 1898 b = pop(); 1899 a = pop(); 1900 1901 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1902 // which negates the result sign except for unordered. Flip the unordered 1903 // as well by using CmpF3 which implements unordered-lesser instead of 1904 // unordered-greater semantics. Finally, commute the result bits. Result 1905 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1906 c = _gvn.transform( new CmpF3Node( b, a)); 1907 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1908 push(c); 1909 break; 1910 1911 case Bytecodes::_f2i: 1912 a = pop(); 1913 push(_gvn.transform(new ConvF2INode(a))); 1914 break; 1915 1916 case Bytecodes::_d2i: 1917 a = pop_pair(); 1918 b = _gvn.transform(new ConvD2INode(a)); 1919 push( b ); 1920 break; 1921 1922 case Bytecodes::_f2d: 1923 a = pop(); 1924 b = _gvn.transform( new ConvF2DNode(a)); 1925 push_pair( b ); 1926 break; 1927 1928 case Bytecodes::_d2f: 1929 a = pop_pair(); 1930 b = _gvn.transform( new ConvD2FNode(a)); 1931 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1932 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1933 push( b ); 1934 break; 1935 1936 case Bytecodes::_l2f: 1937 if (Matcher::convL2FSupported()) { 1938 a = pop_pair(); 1939 b = _gvn.transform( new ConvL2FNode(a)); 1940 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1941 // Rather than storing the result into an FP register then pushing 1942 // out to memory to round, the machine instruction that implements 1943 // ConvL2D is responsible for rounding. 1944 // c = precision_rounding(b); 1945 c = _gvn.transform(b); 1946 push(c); 1947 } else { 1948 l2f(); 1949 } 1950 break; 1951 1952 case Bytecodes::_l2d: 1953 a = pop_pair(); 1954 b = _gvn.transform( new ConvL2DNode(a)); 1955 // For i486.ad, rounding is always necessary (see _l2f above). 1956 // c = dprecision_rounding(b); 1957 c = _gvn.transform(b); 1958 push_pair(c); 1959 break; 1960 1961 case Bytecodes::_f2l: 1962 a = pop(); 1963 b = _gvn.transform( new ConvF2LNode(a)); 1964 push_pair(b); 1965 break; 1966 1967 case Bytecodes::_d2l: 1968 a = pop_pair(); 1969 b = _gvn.transform( new ConvD2LNode(a)); 1970 push_pair(b); 1971 break; 1972 1973 case Bytecodes::_dsub: 1974 b = pop_pair(); 1975 a = pop_pair(); 1976 c = _gvn.transform( new SubDNode(a,b) ); 1977 d = dprecision_rounding(c); 1978 push_pair( d ); 1979 break; 1980 1981 case Bytecodes::_dadd: 1982 b = pop_pair(); 1983 a = pop_pair(); 1984 c = _gvn.transform( new AddDNode(a,b) ); 1985 d = dprecision_rounding(c); 1986 push_pair( d ); 1987 break; 1988 1989 case Bytecodes::_dmul: 1990 b = pop_pair(); 1991 a = pop_pair(); 1992 c = _gvn.transform( new MulDNode(a,b) ); 1993 d = dprecision_rounding(c); 1994 push_pair( d ); 1995 break; 1996 1997 case Bytecodes::_ddiv: 1998 b = pop_pair(); 1999 a = pop_pair(); 2000 c = _gvn.transform( new DivDNode(0,a,b) ); 2001 d = dprecision_rounding(c); 2002 push_pair( d ); 2003 break; 2004 2005 case Bytecodes::_dneg: 2006 a = pop_pair(); 2007 b = _gvn.transform(new NegDNode (a)); 2008 push_pair(b); 2009 break; 2010 2011 case Bytecodes::_drem: 2012 if (Matcher::has_match_rule(Op_ModD)) { 2013 // Generate a ModD node. 2014 b = pop_pair(); 2015 a = pop_pair(); 2016 // a % b 2017 2018 c = _gvn.transform( new ModDNode(0,a,b) ); 2019 d = dprecision_rounding(c); 2020 push_pair( d ); 2021 } 2022 else { 2023 // Generate a call. 2024 modd(); 2025 } 2026 break; 2027 2028 case Bytecodes::_dcmpl: 2029 b = pop_pair(); 2030 a = pop_pair(); 2031 c = _gvn.transform( new CmpD3Node( a, b)); 2032 push(c); 2033 break; 2034 2035 case Bytecodes::_dcmpg: 2036 b = pop_pair(); 2037 a = pop_pair(); 2038 // Same as dcmpl but need to flip the unordered case. 2039 // Commute the inputs, which negates the result sign except for unordered. 2040 // Flip the unordered as well by using CmpD3 which implements 2041 // unordered-lesser instead of unordered-greater semantics. 2042 // Finally, negate the result bits. Result is same as using a 2043 // CmpD3Greater except we did it with CmpD3 alone. 2044 c = _gvn.transform( new CmpD3Node( b, a)); 2045 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2046 push(c); 2047 break; 2048 2049 2050 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2051 case Bytecodes::_land: 2052 b = pop_pair(); 2053 a = pop_pair(); 2054 c = _gvn.transform( new AndLNode(a,b) ); 2055 push_pair(c); 2056 break; 2057 case Bytecodes::_lor: 2058 b = pop_pair(); 2059 a = pop_pair(); 2060 c = _gvn.transform( new OrLNode(a,b) ); 2061 push_pair(c); 2062 break; 2063 case Bytecodes::_lxor: 2064 b = pop_pair(); 2065 a = pop_pair(); 2066 c = _gvn.transform( new XorLNode(a,b) ); 2067 push_pair(c); 2068 break; 2069 2070 case Bytecodes::_lshl: 2071 b = pop(); // the shift count 2072 a = pop_pair(); // value to be shifted 2073 c = _gvn.transform( new LShiftLNode(a,b) ); 2074 push_pair(c); 2075 break; 2076 case Bytecodes::_lshr: 2077 b = pop(); // the shift count 2078 a = pop_pair(); // value to be shifted 2079 c = _gvn.transform( new RShiftLNode(a,b) ); 2080 push_pair(c); 2081 break; 2082 case Bytecodes::_lushr: 2083 b = pop(); // the shift count 2084 a = pop_pair(); // value to be shifted 2085 c = _gvn.transform( new URShiftLNode(a,b) ); 2086 push_pair(c); 2087 break; 2088 case Bytecodes::_lmul: 2089 b = pop_pair(); 2090 a = pop_pair(); 2091 c = _gvn.transform( new MulLNode(a,b) ); 2092 push_pair(c); 2093 break; 2094 2095 case Bytecodes::_lrem: 2096 // Must keep both values on the expression-stack during null-check 2097 assert(peek(0) == top(), "long word order"); 2098 zero_check_long(peek(1)); 2099 // Compile-time detect of null-exception? 2100 if (stopped()) return; 2101 b = pop_pair(); 2102 a = pop_pair(); 2103 c = _gvn.transform( new ModLNode(control(),a,b) ); 2104 push_pair(c); 2105 break; 2106 2107 case Bytecodes::_ldiv: 2108 // Must keep both values on the expression-stack during null-check 2109 assert(peek(0) == top(), "long word order"); 2110 zero_check_long(peek(1)); 2111 // Compile-time detect of null-exception? 2112 if (stopped()) return; 2113 b = pop_pair(); 2114 a = pop_pair(); 2115 c = _gvn.transform( new DivLNode(control(),a,b) ); 2116 push_pair(c); 2117 break; 2118 2119 case Bytecodes::_ladd: 2120 b = pop_pair(); 2121 a = pop_pair(); 2122 c = _gvn.transform( new AddLNode(a,b) ); 2123 push_pair(c); 2124 break; 2125 case Bytecodes::_lsub: 2126 b = pop_pair(); 2127 a = pop_pair(); 2128 c = _gvn.transform( new SubLNode(a,b) ); 2129 push_pair(c); 2130 break; 2131 case Bytecodes::_lcmp: 2132 // Safepoints are now inserted _before_ branches. The long-compare 2133 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2134 // slew of control flow. These are usually followed by a CmpI vs zero and 2135 // a branch; this pattern then optimizes to the obvious long-compare and 2136 // branch. However, if the branch is backwards there's a Safepoint 2137 // inserted. The inserted Safepoint captures the JVM state at the 2138 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2139 // long-compare is used to control a loop the debug info will force 2140 // computation of the 3-way value, even though the generated code uses a 2141 // long-compare and branch. We try to rectify the situation by inserting 2142 // a SafePoint here and have it dominate and kill the safepoint added at a 2143 // following backwards branch. At this point the JVM state merely holds 2 2144 // longs but not the 3-way value. 2145 if( UseLoopSafepoints ) { 2146 switch( iter().next_bc() ) { 2147 case Bytecodes::_ifgt: 2148 case Bytecodes::_iflt: 2149 case Bytecodes::_ifge: 2150 case Bytecodes::_ifle: 2151 case Bytecodes::_ifne: 2152 case Bytecodes::_ifeq: 2153 // If this is a backwards branch in the bytecodes, add Safepoint 2154 maybe_add_safepoint(iter().next_get_dest()); 2155 default: 2156 break; 2157 } 2158 } 2159 b = pop_pair(); 2160 a = pop_pair(); 2161 c = _gvn.transform( new CmpL3Node( a, b )); 2162 push(c); 2163 break; 2164 2165 case Bytecodes::_lneg: 2166 a = pop_pair(); 2167 b = _gvn.transform( new SubLNode(longcon(0),a)); 2168 push_pair(b); 2169 break; 2170 case Bytecodes::_l2i: 2171 a = pop_pair(); 2172 push( _gvn.transform( new ConvL2INode(a))); 2173 break; 2174 case Bytecodes::_i2l: 2175 a = pop(); 2176 b = _gvn.transform( new ConvI2LNode(a)); 2177 push_pair(b); 2178 break; 2179 case Bytecodes::_i2b: 2180 // Sign extend 2181 a = pop(); 2182 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2183 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2184 push( a ); 2185 break; 2186 case Bytecodes::_i2s: 2187 a = pop(); 2188 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2189 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2190 push( a ); 2191 break; 2192 case Bytecodes::_i2c: 2193 a = pop(); 2194 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2195 break; 2196 2197 case Bytecodes::_i2f: 2198 a = pop(); 2199 b = _gvn.transform( new ConvI2FNode(a) ) ; 2200 c = precision_rounding(b); 2201 push (b); 2202 break; 2203 2204 case Bytecodes::_i2d: 2205 a = pop(); 2206 b = _gvn.transform( new ConvI2DNode(a)); 2207 push_pair(b); 2208 break; 2209 2210 case Bytecodes::_iinc: // Increment local 2211 i = iter().get_index(); // Get local index 2212 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2213 break; 2214 2215 // Exit points of synchronized methods must have an unlock node 2216 case Bytecodes::_return: 2217 return_current(NULL); 2218 break; 2219 2220 case Bytecodes::_ireturn: 2221 case Bytecodes::_areturn: 2222 case Bytecodes::_freturn: 2223 return_current(pop()); 2224 break; 2225 case Bytecodes::_lreturn: 2226 return_current(pop_pair()); 2227 break; 2228 case Bytecodes::_dreturn: 2229 return_current(pop_pair()); 2230 break; 2231 2232 case Bytecodes::_athrow: 2233 // null exception oop throws NULL pointer exception 2234 null_check(peek()); 2235 if (stopped()) return; 2236 // Hook the thrown exception directly to subsequent handlers. 2237 if (BailoutToInterpreterForThrows) { 2238 // Keep method interpreted from now on. 2239 uncommon_trap(Deoptimization::Reason_unhandled, 2240 Deoptimization::Action_make_not_compilable); 2241 return; 2242 } 2243 if (env()->jvmti_can_post_on_exceptions()) { 2244 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2245 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2246 } 2247 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2248 add_exception_state(make_exception_state(peek())); 2249 break; 2250 2251 case Bytecodes::_goto: // fall through 2252 case Bytecodes::_goto_w: { 2253 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2254 2255 // If this is a backwards branch in the bytecodes, add Safepoint 2256 maybe_add_safepoint(target_bci); 2257 2258 // Update method data 2259 profile_taken_branch(target_bci); 2260 2261 // Merge the current control into the target basic block 2262 merge(target_bci); 2263 2264 // See if we can get some profile data and hand it off to the next block 2265 Block *target_block = block()->successor_for_bci(target_bci); 2266 if (target_block->pred_count() != 1) break; 2267 ciMethodData* methodData = method()->method_data(); 2268 if (!methodData->is_mature()) break; 2269 ciProfileData* data = methodData->bci_to_data(bci()); 2270 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch"); 2271 int taken = ((ciJumpData*)data)->taken(); 2272 taken = method()->scale_count(taken); 2273 target_block->set_count(taken); 2274 break; 2275 } 2276 2277 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2278 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2279 handle_if_null: 2280 // If this is a backwards branch in the bytecodes, add Safepoint 2281 maybe_add_safepoint(iter().get_dest()); 2282 a = null(); 2283 b = pop(); 2284 if (!_gvn.type(b)->speculative_maybe_null() && 2285 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2286 inc_sp(1); 2287 Node* null_ctl = top(); 2288 b = null_check_oop(b, &null_ctl, true, true, true); 2289 assert(null_ctl->is_top(), "no null control here"); 2290 dec_sp(1); 2291 } else if (_gvn.type(b)->speculative_always_null() && 2292 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2293 inc_sp(1); 2294 b = null_assert(b); 2295 dec_sp(1); 2296 } 2297 c = _gvn.transform( new CmpPNode(b, a) ); 2298 do_ifnull(btest, c); 2299 break; 2300 2301 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2302 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2303 handle_if_acmp: 2304 // If this is a backwards branch in the bytecodes, add Safepoint 2305 maybe_add_safepoint(iter().get_dest()); 2306 a = pop(); 2307 b = pop(); 2308 c = _gvn.transform( new CmpPNode(b, a) ); 2309 c = optimize_cmp_with_klass(c); 2310 do_if(btest, c); 2311 break; 2312 2313 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2314 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2315 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2316 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2317 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2318 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2319 handle_ifxx: 2320 // If this is a backwards branch in the bytecodes, add Safepoint 2321 maybe_add_safepoint(iter().get_dest()); 2322 a = _gvn.intcon(0); 2323 b = pop(); 2324 c = _gvn.transform( new CmpINode(b, a) ); 2325 do_if(btest, c); 2326 break; 2327 2328 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2329 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2330 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2331 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2332 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2333 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2334 handle_if_icmp: 2335 // If this is a backwards branch in the bytecodes, add Safepoint 2336 maybe_add_safepoint(iter().get_dest()); 2337 a = pop(); 2338 b = pop(); 2339 c = _gvn.transform( new CmpINode( b, a ) ); 2340 do_if(btest, c); 2341 break; 2342 2343 case Bytecodes::_tableswitch: 2344 do_tableswitch(); 2345 break; 2346 2347 case Bytecodes::_lookupswitch: 2348 do_lookupswitch(); 2349 break; 2350 2351 case Bytecodes::_invokestatic: 2352 case Bytecodes::_invokedynamic: 2353 case Bytecodes::_invokespecial: 2354 case Bytecodes::_invokevirtual: 2355 case Bytecodes::_invokeinterface: 2356 do_call(); 2357 break; 2358 case Bytecodes::_checkcast: 2359 do_checkcast(); 2360 break; 2361 case Bytecodes::_instanceof: 2362 do_instanceof(); 2363 break; 2364 case Bytecodes::_anewarray: 2365 do_anewarray(); 2366 break; 2367 case Bytecodes::_newarray: 2368 do_newarray((BasicType)iter().get_index()); 2369 break; 2370 case Bytecodes::_multianewarray: 2371 do_multianewarray(); 2372 break; 2373 case Bytecodes::_new: 2374 do_new(); 2375 break; 2376 2377 case Bytecodes::_jsr: 2378 case Bytecodes::_jsr_w: 2379 do_jsr(); 2380 break; 2381 2382 case Bytecodes::_ret: 2383 do_ret(); 2384 break; 2385 2386 2387 case Bytecodes::_monitorenter: 2388 do_monitor_enter(); 2389 break; 2390 2391 case Bytecodes::_monitorexit: 2392 do_monitor_exit(); 2393 break; 2394 2395 case Bytecodes::_breakpoint: 2396 // Breakpoint set concurrently to compile 2397 // %%% use an uncommon trap? 2398 C->record_failure("breakpoint in method"); 2399 return; 2400 2401 default: 2402 #ifndef PRODUCT 2403 map()->dump(99); 2404 #endif 2405 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2406 ShouldNotReachHere(); 2407 } 2408 2409 #ifndef PRODUCT 2410 IdealGraphPrinter *printer = C->printer(); 2411 if (printer && printer->should_print(1)) { 2412 char buffer[256]; 2413 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2414 bool old = printer->traverse_outs(); 2415 printer->set_traverse_outs(true); 2416 printer->print_method(buffer, 4); 2417 printer->set_traverse_outs(old); 2418 } 2419 #endif 2420 }