1 /* 2 * Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "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/universe.inline.hpp" 32 #include "oops/oop.inline.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/divnode.hpp" 37 #include "opto/idealGraphPrinter.hpp" 38 #include "opto/matcher.hpp" 39 #include "opto/memnode.hpp" 40 #include "opto/mulnode.hpp" 41 #include "opto/opaquenode.hpp" 42 #include "opto/parse.hpp" 43 #include "opto/runtime.hpp" 44 #include "runtime/deoptimization.hpp" 45 #include "runtime/sharedRuntime.hpp" 46 47 extern int explicit_null_checks_inserted, 48 explicit_null_checks_elided; 49 50 //---------------------------------array_load---------------------------------- 51 void Parse::array_load(BasicType elem_type) { 52 const Type* elem = Type::TOP; 53 Node* adr = array_addressing(elem_type, 0, &elem); 54 if (stopped()) return; // guaranteed null or range check 55 dec_sp(2); // Pop array and index 56 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 57 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 58 push(ld); 59 } 60 61 62 //--------------------------------array_store---------------------------------- 63 void Parse::array_store(BasicType elem_type) { 64 Node* adr = array_addressing(elem_type, 1); 65 if (stopped()) return; // guaranteed null or range check 66 Node* val = pop(); 67 dec_sp(2); // Pop array and index 68 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 69 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 70 } 71 72 73 //------------------------------array_addressing------------------------------- 74 // Pull array and index from the stack. Compute pointer-to-element. 75 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 76 Node *idx = peek(0+vals); // Get from stack without popping 77 Node *ary = peek(1+vals); // in case of exception 78 79 // Null check the array base, with correct stack contents 80 ary = null_check(ary, T_ARRAY); 81 // Compile-time detect of null-exception? 82 if (stopped()) return top(); 83 84 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 85 const TypeInt* sizetype = arytype->size(); 86 const Type* elemtype = arytype->elem(); 87 88 if (UseUniqueSubclasses && result2 != NULL) { 89 const Type* el = elemtype->make_ptr(); 90 if (el && el->isa_instptr()) { 91 const TypeInstPtr* toop = el->is_instptr(); 92 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 93 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 94 const Type* subklass = Type::get_const_type(toop->klass()); 95 elemtype = subklass->join_speculative(el); 96 } 97 } 98 } 99 100 // Check for big class initializers with all constant offsets 101 // feeding into a known-size array. 102 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 103 // See if the highest idx value is less than the lowest array bound, 104 // and if the idx value cannot be negative: 105 bool need_range_check = true; 106 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 107 need_range_check = false; 108 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 109 } 110 111 ciKlass * arytype_klass = arytype->klass(); 112 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 113 // Only fails for some -Xcomp runs 114 // The class is unloaded. We have to run this bytecode in the interpreter. 115 uncommon_trap(Deoptimization::Reason_unloaded, 116 Deoptimization::Action_reinterpret, 117 arytype->klass(), "!loaded array"); 118 return top(); 119 } 120 121 // Do the range check 122 if (GenerateRangeChecks && need_range_check) { 123 Node* tst; 124 if (sizetype->_hi <= 0) { 125 // The greatest array bound is negative, so we can conclude that we're 126 // compiling unreachable code, but the unsigned compare trick used below 127 // only works with non-negative lengths. Instead, hack "tst" to be zero so 128 // the uncommon_trap path will always be taken. 129 tst = _gvn.intcon(0); 130 } else { 131 // Range is constant in array-oop, so we can use the original state of mem 132 Node* len = load_array_length(ary); 133 134 // Test length vs index (standard trick using unsigned compare) 135 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 136 BoolTest::mask btest = BoolTest::lt; 137 tst = _gvn.transform( new BoolNode(chk, btest) ); 138 } 139 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 140 _gvn.set_type(rc, rc->Value(&_gvn)); 141 if (!tst->is_Con()) { 142 record_for_igvn(rc); 143 } 144 set_control(_gvn.transform(new IfTrueNode(rc))); 145 // Branch to failure if out of bounds 146 { 147 PreserveJVMState pjvms(this); 148 set_control(_gvn.transform(new IfFalseNode(rc))); 149 if (C->allow_range_check_smearing()) { 150 // Do not use builtin_throw, since range checks are sometimes 151 // made more stringent by an optimistic transformation. 152 // This creates "tentative" range checks at this point, 153 // which are not guaranteed to throw exceptions. 154 // See IfNode::Ideal, is_range_check, adjust_check. 155 uncommon_trap(Deoptimization::Reason_range_check, 156 Deoptimization::Action_make_not_entrant, 157 NULL, "range_check"); 158 } else { 159 // If we have already recompiled with the range-check-widening 160 // heroic optimization turned off, then we must really be throwing 161 // range check exceptions. 162 builtin_throw(Deoptimization::Reason_range_check, idx); 163 } 164 } 165 } 166 // Check for always knowing you are throwing a range-check exception 167 if (stopped()) return top(); 168 169 // Make array address computation control dependent to prevent it 170 // from floating above the range check during loop optimizations. 171 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 172 173 if (result2 != NULL) *result2 = elemtype; 174 175 assert(ptr != top(), "top should go hand-in-hand with stopped"); 176 177 return ptr; 178 } 179 180 181 // returns IfNode 182 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 183 Node *cmp = _gvn.transform( new CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 184 Node *tst = _gvn.transform( new BoolNode( cmp, mask)); 185 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 186 return iff; 187 } 188 189 // return Region node 190 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 191 Node *region = new RegionNode(3); // 2 results 192 record_for_igvn(region); 193 region->init_req(1, iffalse); 194 region->init_req(2, iftrue ); 195 _gvn.set_type(region, Type::CONTROL); 196 region = _gvn.transform(region); 197 set_control (region); 198 return region; 199 } 200 201 202 //------------------------------helper for tableswitch------------------------- 203 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 204 // True branch, use existing map info 205 { PreserveJVMState pjvms(this); 206 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 207 set_control( iftrue ); 208 profile_switch_case(prof_table_index); 209 merge_new_path(dest_bci_if_true); 210 } 211 212 // False branch 213 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 214 set_control( iffalse ); 215 } 216 217 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 218 // True branch, use existing map info 219 { PreserveJVMState pjvms(this); 220 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 221 set_control( iffalse ); 222 profile_switch_case(prof_table_index); 223 merge_new_path(dest_bci_if_true); 224 } 225 226 // False branch 227 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 228 set_control( iftrue ); 229 } 230 231 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 232 // False branch, use existing map and control() 233 profile_switch_case(prof_table_index); 234 merge_new_path(dest_bci); 235 } 236 237 238 extern "C" { 239 static int jint_cmp(const void *i, const void *j) { 240 int a = *(jint *)i; 241 int b = *(jint *)j; 242 return a > b ? 1 : a < b ? -1 : 0; 243 } 244 } 245 246 247 // Default value for methodData switch indexing. Must be a negative value to avoid 248 // conflict with any legal switch index. 249 #define NullTableIndex -1 250 251 class SwitchRange : public StackObj { 252 // a range of integers coupled with a bci destination 253 jint _lo; // inclusive lower limit 254 jint _hi; // inclusive upper limit 255 int _dest; 256 int _table_index; // index into method data table 257 258 public: 259 jint lo() const { return _lo; } 260 jint hi() const { return _hi; } 261 int dest() const { return _dest; } 262 int table_index() const { return _table_index; } 263 bool is_singleton() const { return _lo == _hi; } 264 265 void setRange(jint lo, jint hi, int dest, int table_index) { 266 assert(lo <= hi, "must be a non-empty range"); 267 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 268 } 269 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 270 assert(lo <= hi, "must be a non-empty range"); 271 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 272 _hi = hi; 273 return true; 274 } 275 return false; 276 } 277 278 void set (jint value, int dest, int table_index) { 279 setRange(value, value, dest, table_index); 280 } 281 bool adjoin(jint value, int dest, int table_index) { 282 return adjoinRange(value, value, dest, table_index); 283 } 284 285 void print() { 286 if (is_singleton()) 287 tty->print(" {%d}=>%d", lo(), dest()); 288 else if (lo() == min_jint) 289 tty->print(" {..%d}=>%d", hi(), dest()); 290 else if (hi() == max_jint) 291 tty->print(" {%d..}=>%d", lo(), dest()); 292 else 293 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 294 } 295 }; 296 297 298 //-------------------------------do_tableswitch-------------------------------- 299 void Parse::do_tableswitch() { 300 Node* lookup = pop(); 301 302 // Get information about tableswitch 303 int default_dest = iter().get_dest_table(0); 304 int lo_index = iter().get_int_table(1); 305 int hi_index = iter().get_int_table(2); 306 int len = hi_index - lo_index + 1; 307 308 if (len < 1) { 309 // If this is a backward branch, add safepoint 310 maybe_add_safepoint(default_dest); 311 merge(default_dest); 312 return; 313 } 314 315 // generate decision tree, using trichotomy when possible 316 int rnum = len+2; 317 bool makes_backward_branch = false; 318 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 319 int rp = -1; 320 if (lo_index != min_jint) { 321 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 322 } 323 for (int j = 0; j < len; j++) { 324 jint match_int = lo_index+j; 325 int dest = iter().get_dest_table(j+3); 326 makes_backward_branch |= (dest <= bci()); 327 int table_index = method_data_update() ? j : NullTableIndex; 328 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 329 ranges[++rp].set(match_int, dest, table_index); 330 } 331 } 332 jint highest = lo_index+(len-1); 333 assert(ranges[rp].hi() == highest, ""); 334 if (highest != max_jint 335 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 336 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 337 } 338 assert(rp < len+2, "not too many ranges"); 339 340 // Safepoint in case if backward branch observed 341 if( makes_backward_branch && UseLoopSafepoints ) 342 add_safepoint(); 343 344 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 345 } 346 347 348 //------------------------------do_lookupswitch-------------------------------- 349 void Parse::do_lookupswitch() { 350 Node *lookup = pop(); // lookup value 351 // Get information about lookupswitch 352 int default_dest = iter().get_dest_table(0); 353 int len = iter().get_int_table(1); 354 355 if (len < 1) { // If this is a backward branch, add safepoint 356 maybe_add_safepoint(default_dest); 357 merge(default_dest); 358 return; 359 } 360 361 // generate decision tree, using trichotomy when possible 362 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 363 { 364 for( int j = 0; j < len; j++ ) { 365 table[j+j+0] = iter().get_int_table(2+j+j); 366 table[j+j+1] = iter().get_dest_table(2+j+j+1); 367 } 368 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 369 } 370 371 int rnum = len*2+1; 372 bool makes_backward_branch = false; 373 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 374 int rp = -1; 375 for( int j = 0; j < len; j++ ) { 376 jint match_int = table[j+j+0]; 377 int dest = table[j+j+1]; 378 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 379 int table_index = method_data_update() ? j : NullTableIndex; 380 makes_backward_branch |= (dest <= bci()); 381 if( match_int != next_lo ) { 382 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 383 } 384 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 385 ranges[++rp].set(match_int, dest, table_index); 386 } 387 } 388 jint highest = table[2*(len-1)]; 389 assert(ranges[rp].hi() == highest, ""); 390 if( highest != max_jint 391 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 392 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 393 } 394 assert(rp < rnum, "not too many ranges"); 395 396 // Safepoint in case backward branch observed 397 if( makes_backward_branch && UseLoopSafepoints ) 398 add_safepoint(); 399 400 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 401 } 402 403 //----------------------------create_jump_tables------------------------------- 404 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 405 // Are jumptables enabled 406 if (!UseJumpTables) return false; 407 408 // Are jumptables supported 409 if (!Matcher::has_match_rule(Op_Jump)) return false; 410 411 // Don't make jump table if profiling 412 if (method_data_update()) return false; 413 414 // Decide if a guard is needed to lop off big ranges at either (or 415 // both) end(s) of the input set. We'll call this the default target 416 // even though we can't be sure that it is the true "default". 417 418 bool needs_guard = false; 419 int default_dest; 420 int64_t total_outlier_size = 0; 421 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 422 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 423 424 if (lo->dest() == hi->dest()) { 425 total_outlier_size = hi_size + lo_size; 426 default_dest = lo->dest(); 427 } else if (lo_size > hi_size) { 428 total_outlier_size = lo_size; 429 default_dest = lo->dest(); 430 } else { 431 total_outlier_size = hi_size; 432 default_dest = hi->dest(); 433 } 434 435 // If a guard test will eliminate very sparse end ranges, then 436 // it is worth the cost of an extra jump. 437 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 438 needs_guard = true; 439 if (default_dest == lo->dest()) lo++; 440 if (default_dest == hi->dest()) hi--; 441 } 442 443 // Find the total number of cases and ranges 444 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 445 int num_range = hi - lo + 1; 446 447 // Don't create table if: too large, too small, or too sparse. 448 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 449 return false; 450 if (num_cases > (MaxJumpTableSparseness * num_range)) 451 return false; 452 453 // Normalize table lookups to zero 454 int lowval = lo->lo(); 455 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 456 457 // Generate a guard to protect against input keyvals that aren't 458 // in the switch domain. 459 if (needs_guard) { 460 Node* size = _gvn.intcon(num_cases); 461 Node* cmp = _gvn.transform( new CmpUNode(key_val, size) ); 462 Node* tst = _gvn.transform( new BoolNode(cmp, BoolTest::ge) ); 463 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 464 jump_if_true_fork(iff, default_dest, NullTableIndex); 465 } 466 467 // Create an ideal node JumpTable that has projections 468 // of all possible ranges for a switch statement 469 // The key_val input must be converted to a pointer offset and scaled. 470 // Compare Parse::array_addressing above. 471 472 // Clean the 32-bit int into a real 64-bit offset. 473 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 474 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 475 // Make I2L conversion control dependent to prevent it from 476 // floating above the range check during loop optimizations. 477 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 478 479 // Shift the value by wordsize so we have an index into the table, rather 480 // than a switch value 481 Node *shiftWord = _gvn.MakeConX(wordSize); 482 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 483 484 // Create the JumpNode 485 Node* jtn = _gvn.transform( new JumpNode(control(), key_val, num_cases) ); 486 487 // These are the switch destinations hanging off the jumpnode 488 int i = 0; 489 for (SwitchRange* r = lo; r <= hi; r++) { 490 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 491 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 492 { 493 PreserveJVMState pjvms(this); 494 set_control(input); 495 jump_if_always_fork(r->dest(), r->table_index()); 496 } 497 } 498 } 499 assert(i == num_cases, "miscount of cases"); 500 stop_and_kill_map(); // no more uses for this JVMS 501 return true; 502 } 503 504 //----------------------------jump_switch_ranges------------------------------- 505 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 506 Block* switch_block = block(); 507 508 if (switch_depth == 0) { 509 // Do special processing for the top-level call. 510 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 511 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 512 513 // Decrement pred-numbers for the unique set of nodes. 514 #ifdef ASSERT 515 // Ensure that the block's successors are a (duplicate-free) set. 516 int successors_counted = 0; // block occurrences in [hi..lo] 517 int unique_successors = switch_block->num_successors(); 518 for (int i = 0; i < unique_successors; i++) { 519 Block* target = switch_block->successor_at(i); 520 521 // Check that the set of successors is the same in both places. 522 int successors_found = 0; 523 for (SwitchRange* p = lo; p <= hi; p++) { 524 if (p->dest() == target->start()) successors_found++; 525 } 526 assert(successors_found > 0, "successor must be known"); 527 successors_counted += successors_found; 528 } 529 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 530 #endif 531 532 // Maybe prune the inputs, based on the type of key_val. 533 jint min_val = min_jint; 534 jint max_val = max_jint; 535 const TypeInt* ti = key_val->bottom_type()->isa_int(); 536 if (ti != NULL) { 537 min_val = ti->_lo; 538 max_val = ti->_hi; 539 assert(min_val <= max_val, "invalid int type"); 540 } 541 while (lo->hi() < min_val) lo++; 542 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 543 while (hi->lo() > max_val) hi--; 544 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 545 } 546 547 #ifndef PRODUCT 548 if (switch_depth == 0) { 549 _max_switch_depth = 0; 550 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 551 } 552 #endif 553 554 assert(lo <= hi, "must be a non-empty set of ranges"); 555 if (lo == hi) { 556 jump_if_always_fork(lo->dest(), lo->table_index()); 557 } else { 558 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 559 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 560 561 if (create_jump_tables(key_val, lo, hi)) return; 562 563 int nr = hi - lo + 1; 564 565 SwitchRange* mid = lo + nr/2; 566 // if there is an easy choice, pivot at a singleton: 567 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 568 569 assert(lo < mid && mid <= hi, "good pivot choice"); 570 assert(nr != 2 || mid == hi, "should pick higher of 2"); 571 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 572 573 Node *test_val = _gvn.intcon(mid->lo()); 574 575 if (mid->is_singleton()) { 576 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 577 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 578 579 // Special Case: If there are exactly three ranges, and the high 580 // and low range each go to the same place, omit the "gt" test, 581 // since it will not discriminate anything. 582 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 583 if (eq_test_only) { 584 assert(mid == hi-1, ""); 585 } 586 587 // if there is a higher range, test for it and process it: 588 if (mid < hi && !eq_test_only) { 589 // two comparisons of same values--should enable 1 test for 2 branches 590 // Use BoolTest::le instead of BoolTest::gt 591 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 592 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 593 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 594 { PreserveJVMState pjvms(this); 595 set_control(iffalse); 596 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 597 } 598 set_control(iftrue); 599 } 600 601 } else { 602 // mid is a range, not a singleton, so treat mid..hi as a unit 603 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 604 605 // if there is a higher range, test for it and process it: 606 if (mid == hi) { 607 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 608 } else { 609 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 610 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 611 { PreserveJVMState pjvms(this); 612 set_control(iftrue); 613 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 614 } 615 set_control(iffalse); 616 } 617 } 618 619 // in any case, process the lower range 620 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 621 } 622 623 // Decrease pred_count for each successor after all is done. 624 if (switch_depth == 0) { 625 int unique_successors = switch_block->num_successors(); 626 for (int i = 0; i < unique_successors; i++) { 627 Block* target = switch_block->successor_at(i); 628 // Throw away the pre-allocated path for each unique successor. 629 target->next_path_num(); 630 } 631 } 632 633 #ifndef PRODUCT 634 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 635 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 636 SwitchRange* r; 637 int nsing = 0; 638 for( r = lo; r <= hi; r++ ) { 639 if( r->is_singleton() ) nsing++; 640 } 641 tty->print(">>> "); 642 _method->print_short_name(); 643 tty->print_cr(" switch decision tree"); 644 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 645 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 646 if (_max_switch_depth > _est_switch_depth) { 647 tty->print_cr("******** BAD SWITCH DEPTH ********"); 648 } 649 tty->print(" "); 650 for( r = lo; r <= hi; r++ ) { 651 r->print(); 652 } 653 tty->cr(); 654 } 655 #endif 656 } 657 658 void Parse::modf() { 659 Node *f2 = pop(); 660 Node *f1 = pop(); 661 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 662 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 663 "frem", NULL, //no memory effects 664 f1, f2); 665 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 666 667 push(res); 668 } 669 670 void Parse::modd() { 671 Node *d2 = pop_pair(); 672 Node *d1 = pop_pair(); 673 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 674 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 675 "drem", NULL, //no memory effects 676 d1, top(), d2, top()); 677 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 678 679 #ifdef ASSERT 680 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 681 assert(res_top == top(), "second value must be top"); 682 #endif 683 684 push_pair(res_d); 685 } 686 687 void Parse::l2f() { 688 Node* f2 = pop(); 689 Node* f1 = pop(); 690 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 691 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 692 "l2f", NULL, //no memory effects 693 f1, f2); 694 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 695 696 push(res); 697 } 698 699 void Parse::do_irem() { 700 // Must keep both values on the expression-stack during null-check 701 zero_check_int(peek()); 702 // Compile-time detect of null-exception? 703 if (stopped()) return; 704 705 Node* b = pop(); 706 Node* a = pop(); 707 708 const Type *t = _gvn.type(b); 709 if (t != Type::TOP) { 710 const TypeInt *ti = t->is_int(); 711 if (ti->is_con()) { 712 int divisor = ti->get_con(); 713 // check for positive power of 2 714 if (divisor > 0 && 715 (divisor & ~(divisor-1)) == divisor) { 716 // yes ! 717 Node *mask = _gvn.intcon((divisor - 1)); 718 // Sigh, must handle negative dividends 719 Node *zero = _gvn.intcon(0); 720 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 721 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 722 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 723 Node *reg = jump_if_join(ift, iff); 724 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 725 // Negative path; negate/and/negate 726 Node *neg = _gvn.transform( new SubINode(zero, a) ); 727 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 728 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 729 phi->init_req(1, negn); 730 // Fast positive case 731 Node *andx = _gvn.transform( new AndINode(a, mask) ); 732 phi->init_req(2, andx); 733 // Push the merge 734 push( _gvn.transform(phi) ); 735 return; 736 } 737 } 738 } 739 // Default case 740 push( _gvn.transform( new ModINode(control(),a,b) ) ); 741 } 742 743 // Handle jsr and jsr_w bytecode 744 void Parse::do_jsr() { 745 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 746 747 // Store information about current state, tagged with new _jsr_bci 748 int return_bci = iter().next_bci(); 749 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 750 751 // Update method data 752 profile_taken_branch(jsr_bci); 753 754 // The way we do things now, there is only one successor block 755 // for the jsr, because the target code is cloned by ciTypeFlow. 756 Block* target = successor_for_bci(jsr_bci); 757 758 // What got pushed? 759 const Type* ret_addr = target->peek(); 760 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 761 762 // Effect on jsr on stack 763 push(_gvn.makecon(ret_addr)); 764 765 // Flow to the jsr. 766 merge(jsr_bci); 767 } 768 769 // Handle ret bytecode 770 void Parse::do_ret() { 771 // Find to whom we return. 772 assert(block()->num_successors() == 1, "a ret can only go one place now"); 773 Block* target = block()->successor_at(0); 774 assert(!target->is_ready(), "our arrival must be expected"); 775 profile_ret(target->flow()->start()); 776 int pnum = target->next_path_num(); 777 merge_common(target, pnum); 778 } 779 780 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 781 if (btest != BoolTest::eq && btest != BoolTest::ne) { 782 // Only ::eq and ::ne are supported for profile injection. 783 return false; 784 } 785 if (test->is_Cmp() && 786 test->in(1)->Opcode() == Op_ProfileBoolean) { 787 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 788 int false_cnt = profile->false_count(); 789 int true_cnt = profile->true_count(); 790 791 // Counts matching depends on the actual test operation (::eq or ::ne). 792 // No need to scale the counts because profile injection was designed 793 // to feed exact counts into VM. 794 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 795 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 796 797 profile->consume(); 798 return true; 799 } 800 return false; 801 } 802 //--------------------------dynamic_branch_prediction-------------------------- 803 // Try to gather dynamic branch prediction behavior. Return a probability 804 // of the branch being taken and set the "cnt" field. Returns a -1.0 805 // if we need to use static prediction for some reason. 806 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 807 ResourceMark rm; 808 809 cnt = COUNT_UNKNOWN; 810 811 int taken = 0; 812 int not_taken = 0; 813 814 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 815 816 if (use_mdo) { 817 // Use MethodData information if it is available 818 // FIXME: free the ProfileData structure 819 ciMethodData* methodData = method()->method_data(); 820 if (!methodData->is_mature()) return PROB_UNKNOWN; 821 ciProfileData* data = methodData->bci_to_data(bci()); 822 if (!data->is_JumpData()) return PROB_UNKNOWN; 823 824 // get taken and not taken values 825 taken = data->as_JumpData()->taken(); 826 not_taken = 0; 827 if (data->is_BranchData()) { 828 not_taken = data->as_BranchData()->not_taken(); 829 } 830 831 // scale the counts to be commensurate with invocation counts: 832 taken = method()->scale_count(taken); 833 not_taken = method()->scale_count(not_taken); 834 } 835 836 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 837 // We also check that individual counters are positive first, otherwise the sum can become positive. 838 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 839 if (C->log() != NULL) { 840 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 841 } 842 return PROB_UNKNOWN; 843 } 844 845 // Compute frequency that we arrive here 846 float sum = taken + not_taken; 847 // Adjust, if this block is a cloned private block but the 848 // Jump counts are shared. Taken the private counts for 849 // just this path instead of the shared counts. 850 if( block()->count() > 0 ) 851 sum = block()->count(); 852 cnt = sum / FreqCountInvocations; 853 854 // Pin probability to sane limits 855 float prob; 856 if( !taken ) 857 prob = (0+PROB_MIN) / 2; 858 else if( !not_taken ) 859 prob = (1+PROB_MAX) / 2; 860 else { // Compute probability of true path 861 prob = (float)taken / (float)(taken + not_taken); 862 if (prob > PROB_MAX) prob = PROB_MAX; 863 if (prob < PROB_MIN) prob = PROB_MIN; 864 } 865 866 assert((cnt > 0.0f) && (prob > 0.0f), 867 "Bad frequency assignment in if"); 868 869 if (C->log() != NULL) { 870 const char* prob_str = NULL; 871 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 872 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 873 char prob_str_buf[30]; 874 if (prob_str == NULL) { 875 sprintf(prob_str_buf, "%g", prob); 876 prob_str = prob_str_buf; 877 } 878 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 879 iter().get_dest(), taken, not_taken, cnt, prob_str); 880 } 881 return prob; 882 } 883 884 //-----------------------------branch_prediction------------------------------- 885 float Parse::branch_prediction(float& cnt, 886 BoolTest::mask btest, 887 int target_bci, 888 Node* test) { 889 float prob = dynamic_branch_prediction(cnt, btest, test); 890 // If prob is unknown, switch to static prediction 891 if (prob != PROB_UNKNOWN) return prob; 892 893 prob = PROB_FAIR; // Set default value 894 if (btest == BoolTest::eq) // Exactly equal test? 895 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 896 else if (btest == BoolTest::ne) 897 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 898 899 // If this is a conditional test guarding a backwards branch, 900 // assume its a loop-back edge. Make it a likely taken branch. 901 if (target_bci < bci()) { 902 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 903 // Since it's an OSR, we probably have profile data, but since 904 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 905 // Let's make a special check here for completely zero counts. 906 ciMethodData* methodData = method()->method_data(); 907 if (!methodData->is_empty()) { 908 ciProfileData* data = methodData->bci_to_data(bci()); 909 // Only stop for truly zero counts, which mean an unknown part 910 // of the OSR-ed method, and we want to deopt to gather more stats. 911 // If you have ANY counts, then this loop is simply 'cold' relative 912 // to the OSR loop. 913 if (data->as_BranchData()->taken() + 914 data->as_BranchData()->not_taken() == 0 ) { 915 // This is the only way to return PROB_UNKNOWN: 916 return PROB_UNKNOWN; 917 } 918 } 919 } 920 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 921 } 922 923 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 924 return prob; 925 } 926 927 // The magic constants are chosen so as to match the output of 928 // branch_prediction() when the profile reports a zero taken count. 929 // It is important to distinguish zero counts unambiguously, because 930 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 931 // very small but nonzero probabilities, which if confused with zero 932 // counts would keep the program recompiling indefinitely. 933 bool Parse::seems_never_taken(float prob) const { 934 return prob < PROB_MIN; 935 } 936 937 // True if the comparison seems to be the kind that will not change its 938 // statistics from true to false. See comments in adjust_map_after_if. 939 // This question is only asked along paths which are already 940 // classifed as untaken (by seems_never_taken), so really, 941 // if a path is never taken, its controlling comparison is 942 // already acting in a stable fashion. If the comparison 943 // seems stable, we will put an expensive uncommon trap 944 // on the untaken path. 945 bool Parse::seems_stable_comparison() const { 946 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 947 return false; 948 } 949 return true; 950 } 951 952 //-------------------------------repush_if_args-------------------------------- 953 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 954 inline int Parse::repush_if_args() { 955 if (PrintOpto && WizardMode) { 956 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 957 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 958 method()->print_name(); tty->cr(); 959 } 960 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 961 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 962 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 963 assert(argument(0) != NULL, "must exist"); 964 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 965 inc_sp(bc_depth); 966 return bc_depth; 967 } 968 969 //----------------------------------do_ifnull---------------------------------- 970 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 971 int target_bci = iter().get_dest(); 972 973 Block* branch_block = successor_for_bci(target_bci); 974 Block* next_block = successor_for_bci(iter().next_bci()); 975 976 float cnt; 977 float prob = branch_prediction(cnt, btest, target_bci, c); 978 if (prob == PROB_UNKNOWN) { 979 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 980 if (PrintOpto && Verbose) { 981 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 982 } 983 repush_if_args(); // to gather stats on loop 984 // We need to mark this branch as taken so that if we recompile we will 985 // see that it is possible. In the tiered system the interpreter doesn't 986 // do profiling and by the time we get to the lower tier from the interpreter 987 // the path may be cold again. Make sure it doesn't look untaken 988 profile_taken_branch(target_bci, !ProfileInterpreter); 989 uncommon_trap(Deoptimization::Reason_unreached, 990 Deoptimization::Action_reinterpret, 991 NULL, "cold"); 992 if (C->eliminate_boxing()) { 993 // Mark the successor blocks as parsed 994 branch_block->next_path_num(); 995 next_block->next_path_num(); 996 } 997 return; 998 } 999 1000 explicit_null_checks_inserted++; 1001 1002 // Generate real control flow 1003 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1004 1005 // Sanity check the probability value 1006 assert(prob > 0.0f,"Bad probability in Parser"); 1007 // Need xform to put node in hash table 1008 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1009 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1010 // True branch 1011 { PreserveJVMState pjvms(this); 1012 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1013 set_control(iftrue); 1014 1015 if (stopped()) { // Path is dead? 1016 explicit_null_checks_elided++; 1017 if (C->eliminate_boxing()) { 1018 // Mark the successor block as parsed 1019 branch_block->next_path_num(); 1020 } 1021 } else { // Path is live. 1022 // Update method data 1023 profile_taken_branch(target_bci); 1024 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1025 if (!stopped()) { 1026 merge(target_bci); 1027 } 1028 } 1029 } 1030 1031 // False branch 1032 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1033 set_control(iffalse); 1034 1035 if (stopped()) { // Path is dead? 1036 explicit_null_checks_elided++; 1037 if (C->eliminate_boxing()) { 1038 // Mark the successor block as parsed 1039 next_block->next_path_num(); 1040 } 1041 } else { // Path is live. 1042 // Update method data 1043 profile_not_taken_branch(); 1044 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1045 next_block, branch_block); 1046 } 1047 } 1048 1049 //------------------------------------do_if------------------------------------ 1050 void Parse::do_if(BoolTest::mask btest, Node* c) { 1051 int target_bci = iter().get_dest(); 1052 1053 Block* branch_block = successor_for_bci(target_bci); 1054 Block* next_block = successor_for_bci(iter().next_bci()); 1055 1056 float cnt; 1057 float prob = branch_prediction(cnt, btest, target_bci, c); 1058 float untaken_prob = 1.0 - prob; 1059 1060 if (prob == PROB_UNKNOWN) { 1061 if (PrintOpto && Verbose) { 1062 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1063 } 1064 repush_if_args(); // to gather stats on loop 1065 // We need to mark this branch as taken so that if we recompile we will 1066 // see that it is possible. In the tiered system the interpreter doesn't 1067 // do profiling and by the time we get to the lower tier from the interpreter 1068 // the path may be cold again. Make sure it doesn't look untaken 1069 profile_taken_branch(target_bci, !ProfileInterpreter); 1070 uncommon_trap(Deoptimization::Reason_unreached, 1071 Deoptimization::Action_reinterpret, 1072 NULL, "cold"); 1073 if (C->eliminate_boxing()) { 1074 // Mark the successor blocks as parsed 1075 branch_block->next_path_num(); 1076 next_block->next_path_num(); 1077 } 1078 return; 1079 } 1080 1081 // Sanity check the probability value 1082 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1083 1084 bool taken_if_true = true; 1085 // Convert BoolTest to canonical form: 1086 if (!BoolTest(btest).is_canonical()) { 1087 btest = BoolTest(btest).negate(); 1088 taken_if_true = false; 1089 // prob is NOT updated here; it remains the probability of the taken 1090 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1091 } 1092 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1093 1094 Node* tst0 = new BoolNode(c, btest); 1095 Node* tst = _gvn.transform(tst0); 1096 BoolTest::mask taken_btest = BoolTest::illegal; 1097 BoolTest::mask untaken_btest = BoolTest::illegal; 1098 1099 if (tst->is_Bool()) { 1100 // Refresh c from the transformed bool node, since it may be 1101 // simpler than the original c. Also re-canonicalize btest. 1102 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1103 // That can arise from statements like: if (x instanceof C) ... 1104 if (tst != tst0) { 1105 // Canonicalize one more time since transform can change it. 1106 btest = tst->as_Bool()->_test._test; 1107 if (!BoolTest(btest).is_canonical()) { 1108 // Reverse edges one more time... 1109 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1110 btest = tst->as_Bool()->_test._test; 1111 assert(BoolTest(btest).is_canonical(), "sanity"); 1112 taken_if_true = !taken_if_true; 1113 } 1114 c = tst->in(1); 1115 } 1116 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1117 taken_btest = taken_if_true ? btest : neg_btest; 1118 untaken_btest = taken_if_true ? neg_btest : btest; 1119 } 1120 1121 // Generate real control flow 1122 float true_prob = (taken_if_true ? prob : untaken_prob); 1123 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1124 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1125 Node* taken_branch = new IfTrueNode(iff); 1126 Node* untaken_branch = new IfFalseNode(iff); 1127 if (!taken_if_true) { // Finish conversion to canonical form 1128 Node* tmp = taken_branch; 1129 taken_branch = untaken_branch; 1130 untaken_branch = tmp; 1131 } 1132 1133 // Branch is taken: 1134 { PreserveJVMState pjvms(this); 1135 taken_branch = _gvn.transform(taken_branch); 1136 set_control(taken_branch); 1137 1138 if (stopped()) { 1139 if (C->eliminate_boxing()) { 1140 // Mark the successor block as parsed 1141 branch_block->next_path_num(); 1142 } 1143 } else { 1144 // Update method data 1145 profile_taken_branch(target_bci); 1146 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1147 if (!stopped()) { 1148 merge(target_bci); 1149 } 1150 } 1151 } 1152 1153 untaken_branch = _gvn.transform(untaken_branch); 1154 set_control(untaken_branch); 1155 1156 // Branch not taken. 1157 if (stopped()) { 1158 if (C->eliminate_boxing()) { 1159 // Mark the successor block as parsed 1160 next_block->next_path_num(); 1161 } 1162 } else { 1163 // Update method data 1164 profile_not_taken_branch(); 1165 adjust_map_after_if(untaken_btest, c, untaken_prob, 1166 next_block, branch_block); 1167 } 1168 } 1169 1170 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1171 // Don't want to speculate on uncommon traps when running with -Xcomp 1172 if (!UseInterpreter) { 1173 return false; 1174 } 1175 return (seems_never_taken(prob) && seems_stable_comparison()); 1176 } 1177 1178 //----------------------------adjust_map_after_if------------------------------ 1179 // Adjust the JVM state to reflect the result of taking this path. 1180 // Basically, it means inspecting the CmpNode controlling this 1181 // branch, seeing how it constrains a tested value, and then 1182 // deciding if it's worth our while to encode this constraint 1183 // as graph nodes in the current abstract interpretation map. 1184 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1185 Block* path, Block* other_path) { 1186 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1187 return; // nothing to do 1188 1189 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1190 1191 if (path_is_suitable_for_uncommon_trap(prob)) { 1192 repush_if_args(); 1193 uncommon_trap(Deoptimization::Reason_unstable_if, 1194 Deoptimization::Action_reinterpret, 1195 NULL, 1196 (is_fallthrough ? "taken always" : "taken never")); 1197 return; 1198 } 1199 1200 Node* val = c->in(1); 1201 Node* con = c->in(2); 1202 const Type* tcon = _gvn.type(con); 1203 const Type* tval = _gvn.type(val); 1204 bool have_con = tcon->singleton(); 1205 if (tval->singleton()) { 1206 if (!have_con) { 1207 // Swap, so constant is in con. 1208 con = val; 1209 tcon = tval; 1210 val = c->in(2); 1211 tval = _gvn.type(val); 1212 btest = BoolTest(btest).commute(); 1213 have_con = true; 1214 } else { 1215 // Do we have two constants? Then leave well enough alone. 1216 have_con = false; 1217 } 1218 } 1219 if (!have_con) // remaining adjustments need a con 1220 return; 1221 1222 sharpen_type_after_if(btest, con, tcon, val, tval); 1223 } 1224 1225 1226 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1227 Node* ldk; 1228 if (n->is_DecodeNKlass()) { 1229 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1230 return NULL; 1231 } else { 1232 ldk = n->in(1); 1233 } 1234 } else if (n->Opcode() != Op_LoadKlass) { 1235 return NULL; 1236 } else { 1237 ldk = n; 1238 } 1239 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1240 1241 Node* adr = ldk->in(MemNode::Address); 1242 intptr_t off = 0; 1243 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1244 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1245 return NULL; 1246 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1247 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1248 return NULL; 1249 1250 return obj; 1251 } 1252 1253 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1254 Node* con, const Type* tcon, 1255 Node* val, const Type* tval) { 1256 // Look for opportunities to sharpen the type of a node 1257 // whose klass is compared with a constant klass. 1258 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1259 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1260 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1261 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1262 // Found: 1263 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1264 // or the narrowOop equivalent. 1265 const Type* obj_type = _gvn.type(obj); 1266 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1267 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1268 tboth->higher_equal(obj_type)) { 1269 // obj has to be of the exact type Foo if the CmpP succeeds. 1270 int obj_in_map = map()->find_edge(obj); 1271 JVMState* jvms = this->jvms(); 1272 if (obj_in_map >= 0 && 1273 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1274 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1275 const Type* tcc = ccast->as_Type()->type(); 1276 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1277 // Delay transform() call to allow recovery of pre-cast value 1278 // at the control merge. 1279 _gvn.set_type_bottom(ccast); 1280 record_for_igvn(ccast); 1281 // Here's the payoff. 1282 replace_in_map(obj, ccast); 1283 } 1284 } 1285 } 1286 } 1287 1288 int val_in_map = map()->find_edge(val); 1289 if (val_in_map < 0) return; // replace_in_map would be useless 1290 { 1291 JVMState* jvms = this->jvms(); 1292 if (!(jvms->is_loc(val_in_map) || 1293 jvms->is_stk(val_in_map))) 1294 return; // again, it would be useless 1295 } 1296 1297 // Check for a comparison to a constant, and "know" that the compared 1298 // value is constrained on this path. 1299 assert(tcon->singleton(), ""); 1300 ConstraintCastNode* ccast = NULL; 1301 Node* cast = NULL; 1302 1303 switch (btest) { 1304 case BoolTest::eq: // Constant test? 1305 { 1306 const Type* tboth = tcon->join_speculative(tval); 1307 if (tboth == tval) break; // Nothing to gain. 1308 if (tcon->isa_int()) { 1309 ccast = new CastIINode(val, tboth); 1310 } else if (tcon == TypePtr::NULL_PTR) { 1311 // Cast to null, but keep the pointer identity temporarily live. 1312 ccast = new CastPPNode(val, tboth); 1313 } else { 1314 const TypeF* tf = tcon->isa_float_constant(); 1315 const TypeD* td = tcon->isa_double_constant(); 1316 // Exclude tests vs float/double 0 as these could be 1317 // either +0 or -0. Just because you are equal to +0 1318 // doesn't mean you ARE +0! 1319 // Note, following code also replaces Long and Oop values. 1320 if ((!tf || tf->_f != 0.0) && 1321 (!td || td->_d != 0.0)) 1322 cast = con; // Replace non-constant val by con. 1323 } 1324 } 1325 break; 1326 1327 case BoolTest::ne: 1328 if (tcon == TypePtr::NULL_PTR) { 1329 cast = cast_not_null(val, false); 1330 } 1331 break; 1332 1333 default: 1334 // (At this point we could record int range types with CastII.) 1335 break; 1336 } 1337 1338 if (ccast != NULL) { 1339 const Type* tcc = ccast->as_Type()->type(); 1340 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1341 // Delay transform() call to allow recovery of pre-cast value 1342 // at the control merge. 1343 ccast->set_req(0, control()); 1344 _gvn.set_type_bottom(ccast); 1345 record_for_igvn(ccast); 1346 cast = ccast; 1347 } 1348 1349 if (cast != NULL) { // Here's the payoff. 1350 replace_in_map(val, cast); 1351 } 1352 } 1353 1354 /** 1355 * Use speculative type to optimize CmpP node: if comparison is 1356 * against the low level class, cast the object to the speculative 1357 * type if any. CmpP should then go away. 1358 * 1359 * @param c expected CmpP node 1360 * @return result of CmpP on object casted to speculative type 1361 * 1362 */ 1363 Node* Parse::optimize_cmp_with_klass(Node* c) { 1364 // If this is transformed by the _gvn to a comparison with the low 1365 // level klass then we may be able to use speculation 1366 if (c->Opcode() == Op_CmpP && 1367 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1368 c->in(2)->is_Con()) { 1369 Node* load_klass = NULL; 1370 Node* decode = NULL; 1371 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1372 decode = c->in(1); 1373 load_klass = c->in(1)->in(1); 1374 } else { 1375 load_klass = c->in(1); 1376 } 1377 if (load_klass->in(2)->is_AddP()) { 1378 Node* addp = load_klass->in(2); 1379 Node* obj = addp->in(AddPNode::Address); 1380 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1381 if (obj_type->speculative_type_not_null() != NULL) { 1382 ciKlass* k = obj_type->speculative_type(); 1383 inc_sp(2); 1384 obj = maybe_cast_profiled_obj(obj, k); 1385 dec_sp(2); 1386 // Make the CmpP use the casted obj 1387 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1388 load_klass = load_klass->clone(); 1389 load_klass->set_req(2, addp); 1390 load_klass = _gvn.transform(load_klass); 1391 if (decode != NULL) { 1392 decode = decode->clone(); 1393 decode->set_req(1, load_klass); 1394 load_klass = _gvn.transform(decode); 1395 } 1396 c = c->clone(); 1397 c->set_req(1, load_klass); 1398 c = _gvn.transform(c); 1399 } 1400 } 1401 } 1402 return c; 1403 } 1404 1405 //------------------------------do_one_bytecode-------------------------------- 1406 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1407 void Parse::do_one_bytecode() { 1408 Node *a, *b, *c, *d; // Handy temps 1409 BoolTest::mask btest; 1410 int i; 1411 1412 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1413 1414 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1415 "out of nodes parsing method")) { 1416 return; 1417 } 1418 1419 #ifdef ASSERT 1420 // for setting breakpoints 1421 if (TraceOptoParse) { 1422 tty->print(" @"); 1423 dump_bci(bci()); 1424 tty->cr(); 1425 } 1426 #endif 1427 1428 switch (bc()) { 1429 case Bytecodes::_nop: 1430 // do nothing 1431 break; 1432 case Bytecodes::_lconst_0: 1433 push_pair(longcon(0)); 1434 break; 1435 1436 case Bytecodes::_lconst_1: 1437 push_pair(longcon(1)); 1438 break; 1439 1440 case Bytecodes::_fconst_0: 1441 push(zerocon(T_FLOAT)); 1442 break; 1443 1444 case Bytecodes::_fconst_1: 1445 push(makecon(TypeF::ONE)); 1446 break; 1447 1448 case Bytecodes::_fconst_2: 1449 push(makecon(TypeF::make(2.0f))); 1450 break; 1451 1452 case Bytecodes::_dconst_0: 1453 push_pair(zerocon(T_DOUBLE)); 1454 break; 1455 1456 case Bytecodes::_dconst_1: 1457 push_pair(makecon(TypeD::ONE)); 1458 break; 1459 1460 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1461 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1462 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1463 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1464 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1465 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1466 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1467 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1468 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1469 case Bytecodes::_aconst_null: push(null()); break; 1470 case Bytecodes::_ldc: 1471 case Bytecodes::_ldc_w: 1472 case Bytecodes::_ldc2_w: 1473 // If the constant is unresolved, run this BC once in the interpreter. 1474 { 1475 ciConstant constant = iter().get_constant(); 1476 if (constant.basic_type() == T_OBJECT && 1477 !constant.as_object()->is_loaded()) { 1478 int index = iter().get_constant_pool_index(); 1479 constantTag tag = iter().get_constant_pool_tag(index); 1480 uncommon_trap(Deoptimization::make_trap_request 1481 (Deoptimization::Reason_unloaded, 1482 Deoptimization::Action_reinterpret, 1483 index), 1484 NULL, tag.internal_name()); 1485 break; 1486 } 1487 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1488 "must be java_mirror of klass"); 1489 const Type* con_type = Type::make_from_constant(constant); 1490 if (con_type != NULL) { 1491 push_node(con_type->basic_type(), makecon(con_type)); 1492 } 1493 } 1494 1495 break; 1496 1497 case Bytecodes::_aload_0: 1498 push( local(0) ); 1499 break; 1500 case Bytecodes::_aload_1: 1501 push( local(1) ); 1502 break; 1503 case Bytecodes::_aload_2: 1504 push( local(2) ); 1505 break; 1506 case Bytecodes::_aload_3: 1507 push( local(3) ); 1508 break; 1509 case Bytecodes::_aload: 1510 push( local(iter().get_index()) ); 1511 break; 1512 1513 case Bytecodes::_fload_0: 1514 case Bytecodes::_iload_0: 1515 push( local(0) ); 1516 break; 1517 case Bytecodes::_fload_1: 1518 case Bytecodes::_iload_1: 1519 push( local(1) ); 1520 break; 1521 case Bytecodes::_fload_2: 1522 case Bytecodes::_iload_2: 1523 push( local(2) ); 1524 break; 1525 case Bytecodes::_fload_3: 1526 case Bytecodes::_iload_3: 1527 push( local(3) ); 1528 break; 1529 case Bytecodes::_fload: 1530 case Bytecodes::_iload: 1531 push( local(iter().get_index()) ); 1532 break; 1533 case Bytecodes::_lload_0: 1534 push_pair_local( 0 ); 1535 break; 1536 case Bytecodes::_lload_1: 1537 push_pair_local( 1 ); 1538 break; 1539 case Bytecodes::_lload_2: 1540 push_pair_local( 2 ); 1541 break; 1542 case Bytecodes::_lload_3: 1543 push_pair_local( 3 ); 1544 break; 1545 case Bytecodes::_lload: 1546 push_pair_local( iter().get_index() ); 1547 break; 1548 1549 case Bytecodes::_dload_0: 1550 push_pair_local(0); 1551 break; 1552 case Bytecodes::_dload_1: 1553 push_pair_local(1); 1554 break; 1555 case Bytecodes::_dload_2: 1556 push_pair_local(2); 1557 break; 1558 case Bytecodes::_dload_3: 1559 push_pair_local(3); 1560 break; 1561 case Bytecodes::_dload: 1562 push_pair_local(iter().get_index()); 1563 break; 1564 case Bytecodes::_fstore_0: 1565 case Bytecodes::_istore_0: 1566 case Bytecodes::_astore_0: 1567 set_local( 0, pop() ); 1568 break; 1569 case Bytecodes::_fstore_1: 1570 case Bytecodes::_istore_1: 1571 case Bytecodes::_astore_1: 1572 set_local( 1, pop() ); 1573 break; 1574 case Bytecodes::_fstore_2: 1575 case Bytecodes::_istore_2: 1576 case Bytecodes::_astore_2: 1577 set_local( 2, pop() ); 1578 break; 1579 case Bytecodes::_fstore_3: 1580 case Bytecodes::_istore_3: 1581 case Bytecodes::_astore_3: 1582 set_local( 3, pop() ); 1583 break; 1584 case Bytecodes::_fstore: 1585 case Bytecodes::_istore: 1586 case Bytecodes::_astore: 1587 set_local( iter().get_index(), pop() ); 1588 break; 1589 // long stores 1590 case Bytecodes::_lstore_0: 1591 set_pair_local( 0, pop_pair() ); 1592 break; 1593 case Bytecodes::_lstore_1: 1594 set_pair_local( 1, pop_pair() ); 1595 break; 1596 case Bytecodes::_lstore_2: 1597 set_pair_local( 2, pop_pair() ); 1598 break; 1599 case Bytecodes::_lstore_3: 1600 set_pair_local( 3, pop_pair() ); 1601 break; 1602 case Bytecodes::_lstore: 1603 set_pair_local( iter().get_index(), pop_pair() ); 1604 break; 1605 1606 // double stores 1607 case Bytecodes::_dstore_0: 1608 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1609 break; 1610 case Bytecodes::_dstore_1: 1611 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1612 break; 1613 case Bytecodes::_dstore_2: 1614 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1615 break; 1616 case Bytecodes::_dstore_3: 1617 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1618 break; 1619 case Bytecodes::_dstore: 1620 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1621 break; 1622 1623 case Bytecodes::_pop: dec_sp(1); break; 1624 case Bytecodes::_pop2: dec_sp(2); break; 1625 case Bytecodes::_swap: 1626 a = pop(); 1627 b = pop(); 1628 push(a); 1629 push(b); 1630 break; 1631 case Bytecodes::_dup: 1632 a = pop(); 1633 push(a); 1634 push(a); 1635 break; 1636 case Bytecodes::_dup_x1: 1637 a = pop(); 1638 b = pop(); 1639 push( a ); 1640 push( b ); 1641 push( a ); 1642 break; 1643 case Bytecodes::_dup_x2: 1644 a = pop(); 1645 b = pop(); 1646 c = pop(); 1647 push( a ); 1648 push( c ); 1649 push( b ); 1650 push( a ); 1651 break; 1652 case Bytecodes::_dup2: 1653 a = pop(); 1654 b = pop(); 1655 push( b ); 1656 push( a ); 1657 push( b ); 1658 push( a ); 1659 break; 1660 1661 case Bytecodes::_dup2_x1: 1662 // before: .. c, b, a 1663 // after: .. b, a, c, b, a 1664 // not tested 1665 a = pop(); 1666 b = pop(); 1667 c = pop(); 1668 push( b ); 1669 push( a ); 1670 push( c ); 1671 push( b ); 1672 push( a ); 1673 break; 1674 case Bytecodes::_dup2_x2: 1675 // before: .. d, c, b, a 1676 // after: .. b, a, d, c, b, a 1677 // not tested 1678 a = pop(); 1679 b = pop(); 1680 c = pop(); 1681 d = pop(); 1682 push( b ); 1683 push( a ); 1684 push( d ); 1685 push( c ); 1686 push( b ); 1687 push( a ); 1688 break; 1689 1690 case Bytecodes::_arraylength: { 1691 // Must do null-check with value on expression stack 1692 Node *ary = null_check(peek(), T_ARRAY); 1693 // Compile-time detect of null-exception? 1694 if (stopped()) return; 1695 a = pop(); 1696 push(load_array_length(a)); 1697 break; 1698 } 1699 1700 case Bytecodes::_baload: array_load(T_BYTE); break; 1701 case Bytecodes::_caload: array_load(T_CHAR); break; 1702 case Bytecodes::_iaload: array_load(T_INT); break; 1703 case Bytecodes::_saload: array_load(T_SHORT); break; 1704 case Bytecodes::_faload: array_load(T_FLOAT); break; 1705 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1706 case Bytecodes::_laload: { 1707 a = array_addressing(T_LONG, 0); 1708 if (stopped()) return; // guaranteed null or range check 1709 dec_sp(2); // Pop array and index 1710 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1711 break; 1712 } 1713 case Bytecodes::_daload: { 1714 a = array_addressing(T_DOUBLE, 0); 1715 if (stopped()) return; // guaranteed null or range check 1716 dec_sp(2); // Pop array and index 1717 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1718 break; 1719 } 1720 case Bytecodes::_bastore: array_store(T_BYTE); break; 1721 case Bytecodes::_castore: array_store(T_CHAR); break; 1722 case Bytecodes::_iastore: array_store(T_INT); break; 1723 case Bytecodes::_sastore: array_store(T_SHORT); break; 1724 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1725 case Bytecodes::_aastore: { 1726 d = array_addressing(T_OBJECT, 1); 1727 if (stopped()) return; // guaranteed null or range check 1728 array_store_check(); 1729 c = pop(); // Oop to store 1730 b = pop(); // index (already used) 1731 a = pop(); // the array itself 1732 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1733 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1734 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 1735 StoreNode::release_if_reference(T_OBJECT)); 1736 break; 1737 } 1738 case Bytecodes::_lastore: { 1739 a = array_addressing(T_LONG, 2); 1740 if (stopped()) return; // guaranteed null or range check 1741 c = pop_pair(); 1742 dec_sp(2); // Pop array and index 1743 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1744 break; 1745 } 1746 case Bytecodes::_dastore: { 1747 a = array_addressing(T_DOUBLE, 2); 1748 if (stopped()) return; // guaranteed null or range check 1749 c = pop_pair(); 1750 dec_sp(2); // Pop array and index 1751 c = dstore_rounding(c); 1752 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1753 break; 1754 } 1755 case Bytecodes::_getfield: 1756 do_getfield(); 1757 break; 1758 1759 case Bytecodes::_getstatic: 1760 do_getstatic(); 1761 break; 1762 1763 case Bytecodes::_putfield: 1764 do_putfield(); 1765 break; 1766 1767 case Bytecodes::_putstatic: 1768 do_putstatic(); 1769 break; 1770 1771 case Bytecodes::_irem: 1772 do_irem(); 1773 break; 1774 case Bytecodes::_idiv: 1775 // Must keep both values on the expression-stack during null-check 1776 zero_check_int(peek()); 1777 // Compile-time detect of null-exception? 1778 if (stopped()) return; 1779 b = pop(); 1780 a = pop(); 1781 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1782 break; 1783 case Bytecodes::_imul: 1784 b = pop(); a = pop(); 1785 push( _gvn.transform( new MulINode(a,b) ) ); 1786 break; 1787 case Bytecodes::_iadd: 1788 b = pop(); a = pop(); 1789 push( _gvn.transform( new AddINode(a,b) ) ); 1790 break; 1791 case Bytecodes::_ineg: 1792 a = pop(); 1793 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1794 break; 1795 case Bytecodes::_isub: 1796 b = pop(); a = pop(); 1797 push( _gvn.transform( new SubINode(a,b) ) ); 1798 break; 1799 case Bytecodes::_iand: 1800 b = pop(); a = pop(); 1801 push( _gvn.transform( new AndINode(a,b) ) ); 1802 break; 1803 case Bytecodes::_ior: 1804 b = pop(); a = pop(); 1805 push( _gvn.transform( new OrINode(a,b) ) ); 1806 break; 1807 case Bytecodes::_ixor: 1808 b = pop(); a = pop(); 1809 push( _gvn.transform( new XorINode(a,b) ) ); 1810 break; 1811 case Bytecodes::_ishl: 1812 b = pop(); a = pop(); 1813 push( _gvn.transform( new LShiftINode(a,b) ) ); 1814 break; 1815 case Bytecodes::_ishr: 1816 b = pop(); a = pop(); 1817 push( _gvn.transform( new RShiftINode(a,b) ) ); 1818 break; 1819 case Bytecodes::_iushr: 1820 b = pop(); a = pop(); 1821 push( _gvn.transform( new URShiftINode(a,b) ) ); 1822 break; 1823 1824 case Bytecodes::_fneg: 1825 a = pop(); 1826 b = _gvn.transform(new NegFNode (a)); 1827 push(b); 1828 break; 1829 1830 case Bytecodes::_fsub: 1831 b = pop(); 1832 a = pop(); 1833 c = _gvn.transform( new SubFNode(a,b) ); 1834 d = precision_rounding(c); 1835 push( d ); 1836 break; 1837 1838 case Bytecodes::_fadd: 1839 b = pop(); 1840 a = pop(); 1841 c = _gvn.transform( new AddFNode(a,b) ); 1842 d = precision_rounding(c); 1843 push( d ); 1844 break; 1845 1846 case Bytecodes::_fmul: 1847 b = pop(); 1848 a = pop(); 1849 c = _gvn.transform( new MulFNode(a,b) ); 1850 d = precision_rounding(c); 1851 push( d ); 1852 break; 1853 1854 case Bytecodes::_fdiv: 1855 b = pop(); 1856 a = pop(); 1857 c = _gvn.transform( new DivFNode(0,a,b) ); 1858 d = precision_rounding(c); 1859 push( d ); 1860 break; 1861 1862 case Bytecodes::_frem: 1863 if (Matcher::has_match_rule(Op_ModF)) { 1864 // Generate a ModF node. 1865 b = pop(); 1866 a = pop(); 1867 c = _gvn.transform( new ModFNode(0,a,b) ); 1868 d = precision_rounding(c); 1869 push( d ); 1870 } 1871 else { 1872 // Generate a call. 1873 modf(); 1874 } 1875 break; 1876 1877 case Bytecodes::_fcmpl: 1878 b = pop(); 1879 a = pop(); 1880 c = _gvn.transform( new CmpF3Node( a, b)); 1881 push(c); 1882 break; 1883 case Bytecodes::_fcmpg: 1884 b = pop(); 1885 a = pop(); 1886 1887 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1888 // which negates the result sign except for unordered. Flip the unordered 1889 // as well by using CmpF3 which implements unordered-lesser instead of 1890 // unordered-greater semantics. Finally, commute the result bits. Result 1891 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1892 c = _gvn.transform( new CmpF3Node( b, a)); 1893 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1894 push(c); 1895 break; 1896 1897 case Bytecodes::_f2i: 1898 a = pop(); 1899 push(_gvn.transform(new ConvF2INode(a))); 1900 break; 1901 1902 case Bytecodes::_d2i: 1903 a = pop_pair(); 1904 b = _gvn.transform(new ConvD2INode(a)); 1905 push( b ); 1906 break; 1907 1908 case Bytecodes::_f2d: 1909 a = pop(); 1910 b = _gvn.transform( new ConvF2DNode(a)); 1911 push_pair( b ); 1912 break; 1913 1914 case Bytecodes::_d2f: 1915 a = pop_pair(); 1916 b = _gvn.transform( new ConvD2FNode(a)); 1917 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1918 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1919 push( b ); 1920 break; 1921 1922 case Bytecodes::_l2f: 1923 if (Matcher::convL2FSupported()) { 1924 a = pop_pair(); 1925 b = _gvn.transform( new ConvL2FNode(a)); 1926 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1927 // Rather than storing the result into an FP register then pushing 1928 // out to memory to round, the machine instruction that implements 1929 // ConvL2D is responsible for rounding. 1930 // c = precision_rounding(b); 1931 c = _gvn.transform(b); 1932 push(c); 1933 } else { 1934 l2f(); 1935 } 1936 break; 1937 1938 case Bytecodes::_l2d: 1939 a = pop_pair(); 1940 b = _gvn.transform( new ConvL2DNode(a)); 1941 // For i486.ad, rounding is always necessary (see _l2f above). 1942 // c = dprecision_rounding(b); 1943 c = _gvn.transform(b); 1944 push_pair(c); 1945 break; 1946 1947 case Bytecodes::_f2l: 1948 a = pop(); 1949 b = _gvn.transform( new ConvF2LNode(a)); 1950 push_pair(b); 1951 break; 1952 1953 case Bytecodes::_d2l: 1954 a = pop_pair(); 1955 b = _gvn.transform( new ConvD2LNode(a)); 1956 push_pair(b); 1957 break; 1958 1959 case Bytecodes::_dsub: 1960 b = pop_pair(); 1961 a = pop_pair(); 1962 c = _gvn.transform( new SubDNode(a,b) ); 1963 d = dprecision_rounding(c); 1964 push_pair( d ); 1965 break; 1966 1967 case Bytecodes::_dadd: 1968 b = pop_pair(); 1969 a = pop_pair(); 1970 c = _gvn.transform( new AddDNode(a,b) ); 1971 d = dprecision_rounding(c); 1972 push_pair( d ); 1973 break; 1974 1975 case Bytecodes::_dmul: 1976 b = pop_pair(); 1977 a = pop_pair(); 1978 c = _gvn.transform( new MulDNode(a,b) ); 1979 d = dprecision_rounding(c); 1980 push_pair( d ); 1981 break; 1982 1983 case Bytecodes::_ddiv: 1984 b = pop_pair(); 1985 a = pop_pair(); 1986 c = _gvn.transform( new DivDNode(0,a,b) ); 1987 d = dprecision_rounding(c); 1988 push_pair( d ); 1989 break; 1990 1991 case Bytecodes::_dneg: 1992 a = pop_pair(); 1993 b = _gvn.transform(new NegDNode (a)); 1994 push_pair(b); 1995 break; 1996 1997 case Bytecodes::_drem: 1998 if (Matcher::has_match_rule(Op_ModD)) { 1999 // Generate a ModD node. 2000 b = pop_pair(); 2001 a = pop_pair(); 2002 // a % b 2003 2004 c = _gvn.transform( new ModDNode(0,a,b) ); 2005 d = dprecision_rounding(c); 2006 push_pair( d ); 2007 } 2008 else { 2009 // Generate a call. 2010 modd(); 2011 } 2012 break; 2013 2014 case Bytecodes::_dcmpl: 2015 b = pop_pair(); 2016 a = pop_pair(); 2017 c = _gvn.transform( new CmpD3Node( a, b)); 2018 push(c); 2019 break; 2020 2021 case Bytecodes::_dcmpg: 2022 b = pop_pair(); 2023 a = pop_pair(); 2024 // Same as dcmpl but need to flip the unordered case. 2025 // Commute the inputs, which negates the result sign except for unordered. 2026 // Flip the unordered as well by using CmpD3 which implements 2027 // unordered-lesser instead of unordered-greater semantics. 2028 // Finally, negate the result bits. Result is same as using a 2029 // CmpD3Greater except we did it with CmpD3 alone. 2030 c = _gvn.transform( new CmpD3Node( b, a)); 2031 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2032 push(c); 2033 break; 2034 2035 2036 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2037 case Bytecodes::_land: 2038 b = pop_pair(); 2039 a = pop_pair(); 2040 c = _gvn.transform( new AndLNode(a,b) ); 2041 push_pair(c); 2042 break; 2043 case Bytecodes::_lor: 2044 b = pop_pair(); 2045 a = pop_pair(); 2046 c = _gvn.transform( new OrLNode(a,b) ); 2047 push_pair(c); 2048 break; 2049 case Bytecodes::_lxor: 2050 b = pop_pair(); 2051 a = pop_pair(); 2052 c = _gvn.transform( new XorLNode(a,b) ); 2053 push_pair(c); 2054 break; 2055 2056 case Bytecodes::_lshl: 2057 b = pop(); // the shift count 2058 a = pop_pair(); // value to be shifted 2059 c = _gvn.transform( new LShiftLNode(a,b) ); 2060 push_pair(c); 2061 break; 2062 case Bytecodes::_lshr: 2063 b = pop(); // the shift count 2064 a = pop_pair(); // value to be shifted 2065 c = _gvn.transform( new RShiftLNode(a,b) ); 2066 push_pair(c); 2067 break; 2068 case Bytecodes::_lushr: 2069 b = pop(); // the shift count 2070 a = pop_pair(); // value to be shifted 2071 c = _gvn.transform( new URShiftLNode(a,b) ); 2072 push_pair(c); 2073 break; 2074 case Bytecodes::_lmul: 2075 b = pop_pair(); 2076 a = pop_pair(); 2077 c = _gvn.transform( new MulLNode(a,b) ); 2078 push_pair(c); 2079 break; 2080 2081 case Bytecodes::_lrem: 2082 // Must keep both values on the expression-stack during null-check 2083 assert(peek(0) == top(), "long word order"); 2084 zero_check_long(peek(1)); 2085 // Compile-time detect of null-exception? 2086 if (stopped()) return; 2087 b = pop_pair(); 2088 a = pop_pair(); 2089 c = _gvn.transform( new ModLNode(control(),a,b) ); 2090 push_pair(c); 2091 break; 2092 2093 case Bytecodes::_ldiv: 2094 // Must keep both values on the expression-stack during null-check 2095 assert(peek(0) == top(), "long word order"); 2096 zero_check_long(peek(1)); 2097 // Compile-time detect of null-exception? 2098 if (stopped()) return; 2099 b = pop_pair(); 2100 a = pop_pair(); 2101 c = _gvn.transform( new DivLNode(control(),a,b) ); 2102 push_pair(c); 2103 break; 2104 2105 case Bytecodes::_ladd: 2106 b = pop_pair(); 2107 a = pop_pair(); 2108 c = _gvn.transform( new AddLNode(a,b) ); 2109 push_pair(c); 2110 break; 2111 case Bytecodes::_lsub: 2112 b = pop_pair(); 2113 a = pop_pair(); 2114 c = _gvn.transform( new SubLNode(a,b) ); 2115 push_pair(c); 2116 break; 2117 case Bytecodes::_lcmp: 2118 // Safepoints are now inserted _before_ branches. The long-compare 2119 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2120 // slew of control flow. These are usually followed by a CmpI vs zero and 2121 // a branch; this pattern then optimizes to the obvious long-compare and 2122 // branch. However, if the branch is backwards there's a Safepoint 2123 // inserted. The inserted Safepoint captures the JVM state at the 2124 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2125 // long-compare is used to control a loop the debug info will force 2126 // computation of the 3-way value, even though the generated code uses a 2127 // long-compare and branch. We try to rectify the situation by inserting 2128 // a SafePoint here and have it dominate and kill the safepoint added at a 2129 // following backwards branch. At this point the JVM state merely holds 2 2130 // longs but not the 3-way value. 2131 if( UseLoopSafepoints ) { 2132 switch( iter().next_bc() ) { 2133 case Bytecodes::_ifgt: 2134 case Bytecodes::_iflt: 2135 case Bytecodes::_ifge: 2136 case Bytecodes::_ifle: 2137 case Bytecodes::_ifne: 2138 case Bytecodes::_ifeq: 2139 // If this is a backwards branch in the bytecodes, add Safepoint 2140 maybe_add_safepoint(iter().next_get_dest()); 2141 } 2142 } 2143 b = pop_pair(); 2144 a = pop_pair(); 2145 c = _gvn.transform( new CmpL3Node( a, b )); 2146 push(c); 2147 break; 2148 2149 case Bytecodes::_lneg: 2150 a = pop_pair(); 2151 b = _gvn.transform( new SubLNode(longcon(0),a)); 2152 push_pair(b); 2153 break; 2154 case Bytecodes::_l2i: 2155 a = pop_pair(); 2156 push( _gvn.transform( new ConvL2INode(a))); 2157 break; 2158 case Bytecodes::_i2l: 2159 a = pop(); 2160 b = _gvn.transform( new ConvI2LNode(a)); 2161 push_pair(b); 2162 break; 2163 case Bytecodes::_i2b: 2164 // Sign extend 2165 a = pop(); 2166 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2167 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2168 push( a ); 2169 break; 2170 case Bytecodes::_i2s: 2171 a = pop(); 2172 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2173 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2174 push( a ); 2175 break; 2176 case Bytecodes::_i2c: 2177 a = pop(); 2178 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2179 break; 2180 2181 case Bytecodes::_i2f: 2182 a = pop(); 2183 b = _gvn.transform( new ConvI2FNode(a) ) ; 2184 c = precision_rounding(b); 2185 push (b); 2186 break; 2187 2188 case Bytecodes::_i2d: 2189 a = pop(); 2190 b = _gvn.transform( new ConvI2DNode(a)); 2191 push_pair(b); 2192 break; 2193 2194 case Bytecodes::_iinc: // Increment local 2195 i = iter().get_index(); // Get local index 2196 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2197 break; 2198 2199 // Exit points of synchronized methods must have an unlock node 2200 case Bytecodes::_return: 2201 return_current(NULL); 2202 break; 2203 2204 case Bytecodes::_ireturn: 2205 case Bytecodes::_areturn: 2206 case Bytecodes::_freturn: 2207 return_current(pop()); 2208 break; 2209 case Bytecodes::_lreturn: 2210 return_current(pop_pair()); 2211 break; 2212 case Bytecodes::_dreturn: 2213 return_current(pop_pair()); 2214 break; 2215 2216 case Bytecodes::_athrow: 2217 // null exception oop throws NULL pointer exception 2218 null_check(peek()); 2219 if (stopped()) return; 2220 // Hook the thrown exception directly to subsequent handlers. 2221 if (BailoutToInterpreterForThrows) { 2222 // Keep method interpreted from now on. 2223 uncommon_trap(Deoptimization::Reason_unhandled, 2224 Deoptimization::Action_make_not_compilable); 2225 return; 2226 } 2227 if (env()->jvmti_can_post_on_exceptions()) { 2228 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2229 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2230 } 2231 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2232 add_exception_state(make_exception_state(peek())); 2233 break; 2234 2235 case Bytecodes::_goto: // fall through 2236 case Bytecodes::_goto_w: { 2237 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2238 2239 // If this is a backwards branch in the bytecodes, add Safepoint 2240 maybe_add_safepoint(target_bci); 2241 2242 // Update method data 2243 profile_taken_branch(target_bci); 2244 2245 // Merge the current control into the target basic block 2246 merge(target_bci); 2247 2248 // See if we can get some profile data and hand it off to the next block 2249 Block *target_block = block()->successor_for_bci(target_bci); 2250 if (target_block->pred_count() != 1) break; 2251 ciMethodData* methodData = method()->method_data(); 2252 if (!methodData->is_mature()) break; 2253 ciProfileData* data = methodData->bci_to_data(bci()); 2254 assert( data->is_JumpData(), "" ); 2255 int taken = ((ciJumpData*)data)->taken(); 2256 taken = method()->scale_count(taken); 2257 target_block->set_count(taken); 2258 break; 2259 } 2260 2261 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2262 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2263 handle_if_null: 2264 // If this is a backwards branch in the bytecodes, add Safepoint 2265 maybe_add_safepoint(iter().get_dest()); 2266 a = null(); 2267 b = pop(); 2268 if (!_gvn.type(b)->speculative_maybe_null() && 2269 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2270 inc_sp(1); 2271 Node* null_ctl = top(); 2272 b = null_check_oop(b, &null_ctl, true, true, true); 2273 assert(null_ctl->is_top(), "no null control here"); 2274 dec_sp(1); 2275 } 2276 c = _gvn.transform( new CmpPNode(b, a) ); 2277 do_ifnull(btest, c); 2278 break; 2279 2280 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2281 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2282 handle_if_acmp: 2283 // If this is a backwards branch in the bytecodes, add Safepoint 2284 maybe_add_safepoint(iter().get_dest()); 2285 a = pop(); 2286 b = pop(); 2287 c = _gvn.transform( new CmpPNode(b, a) ); 2288 c = optimize_cmp_with_klass(c); 2289 do_if(btest, c); 2290 break; 2291 2292 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2293 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2294 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2295 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2296 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2297 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2298 handle_ifxx: 2299 // If this is a backwards branch in the bytecodes, add Safepoint 2300 maybe_add_safepoint(iter().get_dest()); 2301 a = _gvn.intcon(0); 2302 b = pop(); 2303 c = _gvn.transform( new CmpINode(b, a) ); 2304 do_if(btest, c); 2305 break; 2306 2307 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2308 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2309 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2310 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2311 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2312 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2313 handle_if_icmp: 2314 // If this is a backwards branch in the bytecodes, add Safepoint 2315 maybe_add_safepoint(iter().get_dest()); 2316 a = pop(); 2317 b = pop(); 2318 c = _gvn.transform( new CmpINode( b, a ) ); 2319 do_if(btest, c); 2320 break; 2321 2322 case Bytecodes::_tableswitch: 2323 do_tableswitch(); 2324 break; 2325 2326 case Bytecodes::_lookupswitch: 2327 do_lookupswitch(); 2328 break; 2329 2330 case Bytecodes::_invokestatic: 2331 case Bytecodes::_invokedynamic: 2332 case Bytecodes::_invokespecial: 2333 case Bytecodes::_invokevirtual: 2334 case Bytecodes::_invokeinterface: 2335 do_call(); 2336 break; 2337 case Bytecodes::_checkcast: 2338 do_checkcast(); 2339 break; 2340 case Bytecodes::_instanceof: 2341 do_instanceof(); 2342 break; 2343 case Bytecodes::_anewarray: 2344 do_anewarray(); 2345 break; 2346 case Bytecodes::_newarray: 2347 do_newarray((BasicType)iter().get_index()); 2348 break; 2349 case Bytecodes::_multianewarray: 2350 do_multianewarray(); 2351 break; 2352 case Bytecodes::_new: 2353 do_new(); 2354 break; 2355 2356 case Bytecodes::_jsr: 2357 case Bytecodes::_jsr_w: 2358 do_jsr(); 2359 break; 2360 2361 case Bytecodes::_ret: 2362 do_ret(); 2363 break; 2364 2365 2366 case Bytecodes::_monitorenter: 2367 do_monitor_enter(); 2368 break; 2369 2370 case Bytecodes::_monitorexit: 2371 do_monitor_exit(); 2372 break; 2373 2374 case Bytecodes::_breakpoint: 2375 // Breakpoint set concurrently to compile 2376 // %%% use an uncommon trap? 2377 C->record_failure("breakpoint in method"); 2378 return; 2379 2380 default: 2381 #ifndef PRODUCT 2382 map()->dump(99); 2383 #endif 2384 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2385 ShouldNotReachHere(); 2386 } 2387 2388 #ifndef PRODUCT 2389 IdealGraphPrinter *printer = C->printer(); 2390 if (printer && printer->should_print(1)) { 2391 char buffer[256]; 2392 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2393 bool old = printer->traverse_outs(); 2394 printer->set_traverse_outs(true); 2395 printer->print_method(buffer, 4); 2396 printer->set_traverse_outs(old); 2397 } 2398 #endif 2399 }