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 // Branch to failure if out of bounds 140 { BuildCutout unless(this, tst, PROB_MAX); 141 if (C->allow_range_check_smearing()) { 142 // Do not use builtin_throw, since range checks are sometimes 143 // made more stringent by an optimistic transformation. 144 // This creates "tentative" range checks at this point, 145 // which are not guaranteed to throw exceptions. 146 // See IfNode::Ideal, is_range_check, adjust_check. 147 uncommon_trap(Deoptimization::Reason_range_check, 148 Deoptimization::Action_make_not_entrant, 149 NULL, "range_check"); 150 } else { 151 // If we have already recompiled with the range-check-widening 152 // heroic optimization turned off, then we must really be throwing 153 // range check exceptions. 154 builtin_throw(Deoptimization::Reason_range_check, idx); 155 } 156 } 157 } 158 // Check for always knowing you are throwing a range-check exception 159 if (stopped()) return top(); 160 161 Node* ptr = array_element_address(ary, idx, type, sizetype); 162 163 if (result2 != NULL) *result2 = elemtype; 164 165 assert(ptr != top(), "top should go hand-in-hand with stopped"); 166 167 return ptr; 168 } 169 170 171 // returns IfNode 172 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 173 Node *cmp = _gvn.transform( new CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 174 Node *tst = _gvn.transform( new BoolNode( cmp, mask)); 175 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 176 return iff; 177 } 178 179 // return Region node 180 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 181 Node *region = new RegionNode(3); // 2 results 182 record_for_igvn(region); 183 region->init_req(1, iffalse); 184 region->init_req(2, iftrue ); 185 _gvn.set_type(region, Type::CONTROL); 186 region = _gvn.transform(region); 187 set_control (region); 188 return region; 189 } 190 191 192 //------------------------------helper for tableswitch------------------------- 193 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 194 // True branch, use existing map info 195 { PreserveJVMState pjvms(this); 196 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 197 set_control( iftrue ); 198 profile_switch_case(prof_table_index); 199 merge_new_path(dest_bci_if_true); 200 } 201 202 // False branch 203 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 204 set_control( iffalse ); 205 } 206 207 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 208 // True branch, use existing map info 209 { PreserveJVMState pjvms(this); 210 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 211 set_control( iffalse ); 212 profile_switch_case(prof_table_index); 213 merge_new_path(dest_bci_if_true); 214 } 215 216 // False branch 217 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 218 set_control( iftrue ); 219 } 220 221 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 222 // False branch, use existing map and control() 223 profile_switch_case(prof_table_index); 224 merge_new_path(dest_bci); 225 } 226 227 228 extern "C" { 229 static int jint_cmp(const void *i, const void *j) { 230 int a = *(jint *)i; 231 int b = *(jint *)j; 232 return a > b ? 1 : a < b ? -1 : 0; 233 } 234 } 235 236 237 // Default value for methodData switch indexing. Must be a negative value to avoid 238 // conflict with any legal switch index. 239 #define NullTableIndex -1 240 241 class SwitchRange : public StackObj { 242 // a range of integers coupled with a bci destination 243 jint _lo; // inclusive lower limit 244 jint _hi; // inclusive upper limit 245 int _dest; 246 int _table_index; // index into method data table 247 248 public: 249 jint lo() const { return _lo; } 250 jint hi() const { return _hi; } 251 int dest() const { return _dest; } 252 int table_index() const { return _table_index; } 253 bool is_singleton() const { return _lo == _hi; } 254 255 void setRange(jint lo, jint hi, int dest, int table_index) { 256 assert(lo <= hi, "must be a non-empty range"); 257 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 258 } 259 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 260 assert(lo <= hi, "must be a non-empty range"); 261 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 262 _hi = hi; 263 return true; 264 } 265 return false; 266 } 267 268 void set (jint value, int dest, int table_index) { 269 setRange(value, value, dest, table_index); 270 } 271 bool adjoin(jint value, int dest, int table_index) { 272 return adjoinRange(value, value, dest, table_index); 273 } 274 275 void print() { 276 if (is_singleton()) 277 tty->print(" {%d}=>%d", lo(), dest()); 278 else if (lo() == min_jint) 279 tty->print(" {..%d}=>%d", hi(), dest()); 280 else if (hi() == max_jint) 281 tty->print(" {%d..}=>%d", lo(), dest()); 282 else 283 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 284 } 285 }; 286 287 288 //-------------------------------do_tableswitch-------------------------------- 289 void Parse::do_tableswitch() { 290 Node* lookup = pop(); 291 292 // Get information about tableswitch 293 int default_dest = iter().get_dest_table(0); 294 int lo_index = iter().get_int_table(1); 295 int hi_index = iter().get_int_table(2); 296 int len = hi_index - lo_index + 1; 297 298 if (len < 1) { 299 // If this is a backward branch, add safepoint 300 maybe_add_safepoint(default_dest); 301 merge(default_dest); 302 return; 303 } 304 305 // generate decision tree, using trichotomy when possible 306 int rnum = len+2; 307 bool makes_backward_branch = false; 308 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 309 int rp = -1; 310 if (lo_index != min_jint) { 311 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 312 } 313 for (int j = 0; j < len; j++) { 314 jint match_int = lo_index+j; 315 int dest = iter().get_dest_table(j+3); 316 makes_backward_branch |= (dest <= bci()); 317 int table_index = method_data_update() ? j : NullTableIndex; 318 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 319 ranges[++rp].set(match_int, dest, table_index); 320 } 321 } 322 jint highest = lo_index+(len-1); 323 assert(ranges[rp].hi() == highest, ""); 324 if (highest != max_jint 325 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 326 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 327 } 328 assert(rp < len+2, "not too many ranges"); 329 330 // Safepoint in case if backward branch observed 331 if( makes_backward_branch && UseLoopSafepoints ) 332 add_safepoint(); 333 334 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 335 } 336 337 338 //------------------------------do_lookupswitch-------------------------------- 339 void Parse::do_lookupswitch() { 340 Node *lookup = pop(); // lookup value 341 // Get information about lookupswitch 342 int default_dest = iter().get_dest_table(0); 343 int len = iter().get_int_table(1); 344 345 if (len < 1) { // If this is a backward branch, add safepoint 346 maybe_add_safepoint(default_dest); 347 merge(default_dest); 348 return; 349 } 350 351 // generate decision tree, using trichotomy when possible 352 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 353 { 354 for( int j = 0; j < len; j++ ) { 355 table[j+j+0] = iter().get_int_table(2+j+j); 356 table[j+j+1] = iter().get_dest_table(2+j+j+1); 357 } 358 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 359 } 360 361 int rnum = len*2+1; 362 bool makes_backward_branch = false; 363 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 364 int rp = -1; 365 for( int j = 0; j < len; j++ ) { 366 jint match_int = table[j+j+0]; 367 int dest = table[j+j+1]; 368 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 369 int table_index = method_data_update() ? j : NullTableIndex; 370 makes_backward_branch |= (dest <= bci()); 371 if( match_int != next_lo ) { 372 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 373 } 374 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 375 ranges[++rp].set(match_int, dest, table_index); 376 } 377 } 378 jint highest = table[2*(len-1)]; 379 assert(ranges[rp].hi() == highest, ""); 380 if( highest != max_jint 381 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 382 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 383 } 384 assert(rp < rnum, "not too many ranges"); 385 386 // Safepoint in case backward branch observed 387 if( makes_backward_branch && UseLoopSafepoints ) 388 add_safepoint(); 389 390 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 391 } 392 393 //----------------------------create_jump_tables------------------------------- 394 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 395 // Are jumptables enabled 396 if (!UseJumpTables) return false; 397 398 // Are jumptables supported 399 if (!Matcher::has_match_rule(Op_Jump)) return false; 400 401 // Don't make jump table if profiling 402 if (method_data_update()) return false; 403 404 // Decide if a guard is needed to lop off big ranges at either (or 405 // both) end(s) of the input set. We'll call this the default target 406 // even though we can't be sure that it is the true "default". 407 408 bool needs_guard = false; 409 int default_dest; 410 int64_t total_outlier_size = 0; 411 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 412 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 413 414 if (lo->dest() == hi->dest()) { 415 total_outlier_size = hi_size + lo_size; 416 default_dest = lo->dest(); 417 } else if (lo_size > hi_size) { 418 total_outlier_size = lo_size; 419 default_dest = lo->dest(); 420 } else { 421 total_outlier_size = hi_size; 422 default_dest = hi->dest(); 423 } 424 425 // If a guard test will eliminate very sparse end ranges, then 426 // it is worth the cost of an extra jump. 427 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 428 needs_guard = true; 429 if (default_dest == lo->dest()) lo++; 430 if (default_dest == hi->dest()) hi--; 431 } 432 433 // Find the total number of cases and ranges 434 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 435 int num_range = hi - lo + 1; 436 437 // Don't create table if: too large, too small, or too sparse. 438 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 439 return false; 440 if (num_cases > (MaxJumpTableSparseness * num_range)) 441 return false; 442 443 // Normalize table lookups to zero 444 int lowval = lo->lo(); 445 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 446 447 // Generate a guard to protect against input keyvals that aren't 448 // in the switch domain. 449 if (needs_guard) { 450 Node* size = _gvn.intcon(num_cases); 451 Node* cmp = _gvn.transform( new CmpUNode(key_val, size) ); 452 Node* tst = _gvn.transform( new BoolNode(cmp, BoolTest::ge) ); 453 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 454 jump_if_true_fork(iff, default_dest, NullTableIndex); 455 } 456 457 // Create an ideal node JumpTable that has projections 458 // of all possible ranges for a switch statement 459 // The key_val input must be converted to a pointer offset and scaled. 460 // Compare Parse::array_addressing above. 461 #ifdef _LP64 462 // Clean the 32-bit int into a real 64-bit offset. 463 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 464 const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin); 465 key_val = _gvn.transform( new ConvI2LNode(key_val, lkeytype) ); 466 #endif 467 // Shift the value by wordsize so we have an index into the table, rather 468 // than a switch value 469 Node *shiftWord = _gvn.MakeConX(wordSize); 470 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 471 472 // Create the JumpNode 473 Node* jtn = _gvn.transform( new JumpNode(control(), key_val, num_cases) ); 474 475 // These are the switch destinations hanging off the jumpnode 476 int i = 0; 477 for (SwitchRange* r = lo; r <= hi; r++) { 478 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 479 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 480 { 481 PreserveJVMState pjvms(this); 482 set_control(input); 483 jump_if_always_fork(r->dest(), r->table_index()); 484 } 485 } 486 } 487 assert(i == num_cases, "miscount of cases"); 488 stop_and_kill_map(); // no more uses for this JVMS 489 return true; 490 } 491 492 //----------------------------jump_switch_ranges------------------------------- 493 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 494 Block* switch_block = block(); 495 496 if (switch_depth == 0) { 497 // Do special processing for the top-level call. 498 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 499 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 500 501 // Decrement pred-numbers for the unique set of nodes. 502 #ifdef ASSERT 503 // Ensure that the block's successors are a (duplicate-free) set. 504 int successors_counted = 0; // block occurrences in [hi..lo] 505 int unique_successors = switch_block->num_successors(); 506 for (int i = 0; i < unique_successors; i++) { 507 Block* target = switch_block->successor_at(i); 508 509 // Check that the set of successors is the same in both places. 510 int successors_found = 0; 511 for (SwitchRange* p = lo; p <= hi; p++) { 512 if (p->dest() == target->start()) successors_found++; 513 } 514 assert(successors_found > 0, "successor must be known"); 515 successors_counted += successors_found; 516 } 517 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 518 #endif 519 520 // Maybe prune the inputs, based on the type of key_val. 521 jint min_val = min_jint; 522 jint max_val = max_jint; 523 const TypeInt* ti = key_val->bottom_type()->isa_int(); 524 if (ti != NULL) { 525 min_val = ti->_lo; 526 max_val = ti->_hi; 527 assert(min_val <= max_val, "invalid int type"); 528 } 529 while (lo->hi() < min_val) lo++; 530 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 531 while (hi->lo() > max_val) hi--; 532 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 533 } 534 535 #ifndef PRODUCT 536 if (switch_depth == 0) { 537 _max_switch_depth = 0; 538 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 539 } 540 #endif 541 542 assert(lo <= hi, "must be a non-empty set of ranges"); 543 if (lo == hi) { 544 jump_if_always_fork(lo->dest(), lo->table_index()); 545 } else { 546 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 547 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 548 549 if (create_jump_tables(key_val, lo, hi)) return; 550 551 int nr = hi - lo + 1; 552 553 SwitchRange* mid = lo + nr/2; 554 // if there is an easy choice, pivot at a singleton: 555 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 556 557 assert(lo < mid && mid <= hi, "good pivot choice"); 558 assert(nr != 2 || mid == hi, "should pick higher of 2"); 559 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 560 561 Node *test_val = _gvn.intcon(mid->lo()); 562 563 if (mid->is_singleton()) { 564 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 565 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 566 567 // Special Case: If there are exactly three ranges, and the high 568 // and low range each go to the same place, omit the "gt" test, 569 // since it will not discriminate anything. 570 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 571 if (eq_test_only) { 572 assert(mid == hi-1, ""); 573 } 574 575 // if there is a higher range, test for it and process it: 576 if (mid < hi && !eq_test_only) { 577 // two comparisons of same values--should enable 1 test for 2 branches 578 // Use BoolTest::le instead of BoolTest::gt 579 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 580 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 581 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 582 { PreserveJVMState pjvms(this); 583 set_control(iffalse); 584 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 585 } 586 set_control(iftrue); 587 } 588 589 } else { 590 // mid is a range, not a singleton, so treat mid..hi as a unit 591 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 592 593 // if there is a higher range, test for it and process it: 594 if (mid == hi) { 595 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 596 } else { 597 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 598 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 599 { PreserveJVMState pjvms(this); 600 set_control(iftrue); 601 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 602 } 603 set_control(iffalse); 604 } 605 } 606 607 // in any case, process the lower range 608 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 609 } 610 611 // Decrease pred_count for each successor after all is done. 612 if (switch_depth == 0) { 613 int unique_successors = switch_block->num_successors(); 614 for (int i = 0; i < unique_successors; i++) { 615 Block* target = switch_block->successor_at(i); 616 // Throw away the pre-allocated path for each unique successor. 617 target->next_path_num(); 618 } 619 } 620 621 #ifndef PRODUCT 622 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 623 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 624 SwitchRange* r; 625 int nsing = 0; 626 for( r = lo; r <= hi; r++ ) { 627 if( r->is_singleton() ) nsing++; 628 } 629 tty->print(">>> "); 630 _method->print_short_name(); 631 tty->print_cr(" switch decision tree"); 632 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 633 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 634 if (_max_switch_depth > _est_switch_depth) { 635 tty->print_cr("******** BAD SWITCH DEPTH ********"); 636 } 637 tty->print(" "); 638 for( r = lo; r <= hi; r++ ) { 639 r->print(); 640 } 641 tty->cr(); 642 } 643 #endif 644 } 645 646 void Parse::modf() { 647 Node *f2 = pop(); 648 Node *f1 = pop(); 649 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 650 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 651 "frem", NULL, //no memory effects 652 f1, f2); 653 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 654 655 push(res); 656 } 657 658 void Parse::modd() { 659 Node *d2 = pop_pair(); 660 Node *d1 = pop_pair(); 661 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 662 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 663 "drem", NULL, //no memory effects 664 d1, top(), d2, top()); 665 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 666 667 #ifdef ASSERT 668 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 669 assert(res_top == top(), "second value must be top"); 670 #endif 671 672 push_pair(res_d); 673 } 674 675 void Parse::l2f() { 676 Node* f2 = pop(); 677 Node* f1 = pop(); 678 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 679 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 680 "l2f", NULL, //no memory effects 681 f1, f2); 682 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 683 684 push(res); 685 } 686 687 void Parse::do_irem() { 688 // Must keep both values on the expression-stack during null-check 689 zero_check_int(peek()); 690 // Compile-time detect of null-exception? 691 if (stopped()) return; 692 693 Node* b = pop(); 694 Node* a = pop(); 695 696 const Type *t = _gvn.type(b); 697 if (t != Type::TOP) { 698 const TypeInt *ti = t->is_int(); 699 if (ti->is_con()) { 700 int divisor = ti->get_con(); 701 // check for positive power of 2 702 if (divisor > 0 && 703 (divisor & ~(divisor-1)) == divisor) { 704 // yes ! 705 Node *mask = _gvn.intcon((divisor - 1)); 706 // Sigh, must handle negative dividends 707 Node *zero = _gvn.intcon(0); 708 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 709 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 710 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 711 Node *reg = jump_if_join(ift, iff); 712 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 713 // Negative path; negate/and/negate 714 Node *neg = _gvn.transform( new SubINode(zero, a) ); 715 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 716 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 717 phi->init_req(1, negn); 718 // Fast positive case 719 Node *andx = _gvn.transform( new AndINode(a, mask) ); 720 phi->init_req(2, andx); 721 // Push the merge 722 push( _gvn.transform(phi) ); 723 return; 724 } 725 } 726 } 727 // Default case 728 push( _gvn.transform( new ModINode(control(),a,b) ) ); 729 } 730 731 // Handle jsr and jsr_w bytecode 732 void Parse::do_jsr() { 733 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 734 735 // Store information about current state, tagged with new _jsr_bci 736 int return_bci = iter().next_bci(); 737 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 738 739 // Update method data 740 profile_taken_branch(jsr_bci); 741 742 // The way we do things now, there is only one successor block 743 // for the jsr, because the target code is cloned by ciTypeFlow. 744 Block* target = successor_for_bci(jsr_bci); 745 746 // What got pushed? 747 const Type* ret_addr = target->peek(); 748 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 749 750 // Effect on jsr on stack 751 push(_gvn.makecon(ret_addr)); 752 753 // Flow to the jsr. 754 merge(jsr_bci); 755 } 756 757 // Handle ret bytecode 758 void Parse::do_ret() { 759 // Find to whom we return. 760 assert(block()->num_successors() == 1, "a ret can only go one place now"); 761 Block* target = block()->successor_at(0); 762 assert(!target->is_ready(), "our arrival must be expected"); 763 profile_ret(target->flow()->start()); 764 int pnum = target->next_path_num(); 765 merge_common(target, pnum); 766 } 767 768 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 769 if (btest != BoolTest::eq && btest != BoolTest::ne) { 770 // Only ::eq and ::ne are supported for profile injection. 771 return false; 772 } 773 if (test->is_Cmp() && 774 test->in(1)->Opcode() == Op_ProfileBoolean) { 775 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 776 int false_cnt = profile->false_count(); 777 int true_cnt = profile->true_count(); 778 779 // Counts matching depends on the actual test operation (::eq or ::ne). 780 // No need to scale the counts because profile injection was designed 781 // to feed exact counts into VM. 782 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 783 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 784 785 profile->consume(); 786 return true; 787 } 788 return false; 789 } 790 //--------------------------dynamic_branch_prediction-------------------------- 791 // Try to gather dynamic branch prediction behavior. Return a probability 792 // of the branch being taken and set the "cnt" field. Returns a -1.0 793 // if we need to use static prediction for some reason. 794 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 795 ResourceMark rm; 796 797 cnt = COUNT_UNKNOWN; 798 799 int taken = 0; 800 int not_taken = 0; 801 802 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 803 804 if (use_mdo) { 805 // Use MethodData information if it is available 806 // FIXME: free the ProfileData structure 807 ciMethodData* methodData = method()->method_data(); 808 if (!methodData->is_mature()) return PROB_UNKNOWN; 809 ciProfileData* data = methodData->bci_to_data(bci()); 810 if (!data->is_JumpData()) return PROB_UNKNOWN; 811 812 // get taken and not taken values 813 taken = data->as_JumpData()->taken(); 814 not_taken = 0; 815 if (data->is_BranchData()) { 816 not_taken = data->as_BranchData()->not_taken(); 817 } 818 819 // scale the counts to be commensurate with invocation counts: 820 taken = method()->scale_count(taken); 821 not_taken = method()->scale_count(not_taken); 822 } 823 824 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 825 // We also check that individual counters are positive first, otherwise the sum can become positive. 826 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 827 if (C->log() != NULL) { 828 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 829 } 830 return PROB_UNKNOWN; 831 } 832 833 // Compute frequency that we arrive here 834 float sum = taken + not_taken; 835 // Adjust, if this block is a cloned private block but the 836 // Jump counts are shared. Taken the private counts for 837 // just this path instead of the shared counts. 838 if( block()->count() > 0 ) 839 sum = block()->count(); 840 cnt = sum / FreqCountInvocations; 841 842 // Pin probability to sane limits 843 float prob; 844 if( !taken ) 845 prob = (0+PROB_MIN) / 2; 846 else if( !not_taken ) 847 prob = (1+PROB_MAX) / 2; 848 else { // Compute probability of true path 849 prob = (float)taken / (float)(taken + not_taken); 850 if (prob > PROB_MAX) prob = PROB_MAX; 851 if (prob < PROB_MIN) prob = PROB_MIN; 852 } 853 854 assert((cnt > 0.0f) && (prob > 0.0f), 855 "Bad frequency assignment in if"); 856 857 if (C->log() != NULL) { 858 const char* prob_str = NULL; 859 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 860 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 861 char prob_str_buf[30]; 862 if (prob_str == NULL) { 863 sprintf(prob_str_buf, "%g", prob); 864 prob_str = prob_str_buf; 865 } 866 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 867 iter().get_dest(), taken, not_taken, cnt, prob_str); 868 } 869 return prob; 870 } 871 872 //-----------------------------branch_prediction------------------------------- 873 float Parse::branch_prediction(float& cnt, 874 BoolTest::mask btest, 875 int target_bci, 876 Node* test) { 877 float prob = dynamic_branch_prediction(cnt, btest, test); 878 // If prob is unknown, switch to static prediction 879 if (prob != PROB_UNKNOWN) return prob; 880 881 prob = PROB_FAIR; // Set default value 882 if (btest == BoolTest::eq) // Exactly equal test? 883 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 884 else if (btest == BoolTest::ne) 885 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 886 887 // If this is a conditional test guarding a backwards branch, 888 // assume its a loop-back edge. Make it a likely taken branch. 889 if (target_bci < bci()) { 890 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 891 // Since it's an OSR, we probably have profile data, but since 892 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 893 // Let's make a special check here for completely zero counts. 894 ciMethodData* methodData = method()->method_data(); 895 if (!methodData->is_empty()) { 896 ciProfileData* data = methodData->bci_to_data(bci()); 897 // Only stop for truly zero counts, which mean an unknown part 898 // of the OSR-ed method, and we want to deopt to gather more stats. 899 // If you have ANY counts, then this loop is simply 'cold' relative 900 // to the OSR loop. 901 if (data->as_BranchData()->taken() + 902 data->as_BranchData()->not_taken() == 0 ) { 903 // This is the only way to return PROB_UNKNOWN: 904 return PROB_UNKNOWN; 905 } 906 } 907 } 908 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 909 } 910 911 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 912 return prob; 913 } 914 915 // The magic constants are chosen so as to match the output of 916 // branch_prediction() when the profile reports a zero taken count. 917 // It is important to distinguish zero counts unambiguously, because 918 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 919 // very small but nonzero probabilities, which if confused with zero 920 // counts would keep the program recompiling indefinitely. 921 bool Parse::seems_never_taken(float prob) const { 922 return prob < PROB_MIN; 923 } 924 925 // True if the comparison seems to be the kind that will not change its 926 // statistics from true to false. See comments in adjust_map_after_if. 927 // This question is only asked along paths which are already 928 // classifed as untaken (by seems_never_taken), so really, 929 // if a path is never taken, its controlling comparison is 930 // already acting in a stable fashion. If the comparison 931 // seems stable, we will put an expensive uncommon trap 932 // on the untaken path. 933 bool Parse::seems_stable_comparison() const { 934 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 935 return false; 936 } 937 return true; 938 } 939 940 //-------------------------------repush_if_args-------------------------------- 941 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 942 inline int Parse::repush_if_args() { 943 #ifndef PRODUCT 944 if (PrintOpto && WizardMode) { 945 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 946 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 947 method()->print_name(); tty->cr(); 948 } 949 #endif 950 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 951 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 952 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 953 assert(argument(0) != NULL, "must exist"); 954 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 955 inc_sp(bc_depth); 956 return bc_depth; 957 } 958 959 //----------------------------------do_ifnull---------------------------------- 960 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 961 int target_bci = iter().get_dest(); 962 963 Block* branch_block = successor_for_bci(target_bci); 964 Block* next_block = successor_for_bci(iter().next_bci()); 965 966 float cnt; 967 float prob = branch_prediction(cnt, btest, target_bci, c); 968 if (prob == PROB_UNKNOWN) { 969 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 970 #ifndef PRODUCT 971 if (PrintOpto && Verbose) 972 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 973 #endif 974 repush_if_args(); // to gather stats on loop 975 // We need to mark this branch as taken so that if we recompile we will 976 // see that it is possible. In the tiered system the interpreter doesn't 977 // do profiling and by the time we get to the lower tier from the interpreter 978 // the path may be cold again. Make sure it doesn't look untaken 979 profile_taken_branch(target_bci, !ProfileInterpreter); 980 uncommon_trap(Deoptimization::Reason_unreached, 981 Deoptimization::Action_reinterpret, 982 NULL, "cold"); 983 if (C->eliminate_boxing()) { 984 // Mark the successor blocks as parsed 985 branch_block->next_path_num(); 986 next_block->next_path_num(); 987 } 988 return; 989 } 990 991 explicit_null_checks_inserted++; 992 993 // Generate real control flow 994 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 995 996 // Sanity check the probability value 997 assert(prob > 0.0f,"Bad probability in Parser"); 998 // Need xform to put node in hash table 999 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1000 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1001 // True branch 1002 { PreserveJVMState pjvms(this); 1003 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1004 set_control(iftrue); 1005 1006 if (stopped()) { // Path is dead? 1007 explicit_null_checks_elided++; 1008 if (C->eliminate_boxing()) { 1009 // Mark the successor block as parsed 1010 branch_block->next_path_num(); 1011 } 1012 } else { // Path is live. 1013 // Update method data 1014 profile_taken_branch(target_bci); 1015 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1016 if (!stopped()) { 1017 merge(target_bci); 1018 } 1019 } 1020 } 1021 1022 // False branch 1023 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1024 set_control(iffalse); 1025 1026 if (stopped()) { // Path is dead? 1027 explicit_null_checks_elided++; 1028 if (C->eliminate_boxing()) { 1029 // Mark the successor block as parsed 1030 next_block->next_path_num(); 1031 } 1032 } else { // Path is live. 1033 // Update method data 1034 profile_not_taken_branch(); 1035 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1036 next_block, branch_block); 1037 } 1038 } 1039 1040 //------------------------------------do_if------------------------------------ 1041 void Parse::do_if(BoolTest::mask btest, Node* c) { 1042 int target_bci = iter().get_dest(); 1043 1044 Block* branch_block = successor_for_bci(target_bci); 1045 Block* next_block = successor_for_bci(iter().next_bci()); 1046 1047 float cnt; 1048 float prob = branch_prediction(cnt, btest, target_bci, c); 1049 float untaken_prob = 1.0 - prob; 1050 1051 if (prob == PROB_UNKNOWN) { 1052 #ifndef PRODUCT 1053 if (PrintOpto && Verbose) 1054 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1055 #endif 1056 repush_if_args(); // to gather stats on loop 1057 // We need to mark this branch as taken so that if we recompile we will 1058 // see that it is possible. In the tiered system the interpreter doesn't 1059 // do profiling and by the time we get to the lower tier from the interpreter 1060 // the path may be cold again. Make sure it doesn't look untaken 1061 profile_taken_branch(target_bci, !ProfileInterpreter); 1062 uncommon_trap(Deoptimization::Reason_unreached, 1063 Deoptimization::Action_reinterpret, 1064 NULL, "cold"); 1065 if (C->eliminate_boxing()) { 1066 // Mark the successor blocks as parsed 1067 branch_block->next_path_num(); 1068 next_block->next_path_num(); 1069 } 1070 return; 1071 } 1072 1073 // Sanity check the probability value 1074 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1075 1076 bool taken_if_true = true; 1077 // Convert BoolTest to canonical form: 1078 if (!BoolTest(btest).is_canonical()) { 1079 btest = BoolTest(btest).negate(); 1080 taken_if_true = false; 1081 // prob is NOT updated here; it remains the probability of the taken 1082 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1083 } 1084 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1085 1086 Node* tst0 = new BoolNode(c, btest); 1087 Node* tst = _gvn.transform(tst0); 1088 BoolTest::mask taken_btest = BoolTest::illegal; 1089 BoolTest::mask untaken_btest = BoolTest::illegal; 1090 1091 if (tst->is_Bool()) { 1092 // Refresh c from the transformed bool node, since it may be 1093 // simpler than the original c. Also re-canonicalize btest. 1094 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1095 // That can arise from statements like: if (x instanceof C) ... 1096 if (tst != tst0) { 1097 // Canonicalize one more time since transform can change it. 1098 btest = tst->as_Bool()->_test._test; 1099 if (!BoolTest(btest).is_canonical()) { 1100 // Reverse edges one more time... 1101 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1102 btest = tst->as_Bool()->_test._test; 1103 assert(BoolTest(btest).is_canonical(), "sanity"); 1104 taken_if_true = !taken_if_true; 1105 } 1106 c = tst->in(1); 1107 } 1108 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1109 taken_btest = taken_if_true ? btest : neg_btest; 1110 untaken_btest = taken_if_true ? neg_btest : btest; 1111 } 1112 1113 // Generate real control flow 1114 float true_prob = (taken_if_true ? prob : untaken_prob); 1115 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1116 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1117 Node* taken_branch = new IfTrueNode(iff); 1118 Node* untaken_branch = new IfFalseNode(iff); 1119 if (!taken_if_true) { // Finish conversion to canonical form 1120 Node* tmp = taken_branch; 1121 taken_branch = untaken_branch; 1122 untaken_branch = tmp; 1123 } 1124 1125 // Branch is taken: 1126 { PreserveJVMState pjvms(this); 1127 taken_branch = _gvn.transform(taken_branch); 1128 set_control(taken_branch); 1129 1130 if (stopped()) { 1131 if (C->eliminate_boxing()) { 1132 // Mark the successor block as parsed 1133 branch_block->next_path_num(); 1134 } 1135 } else { 1136 // Update method data 1137 profile_taken_branch(target_bci); 1138 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1139 if (!stopped()) { 1140 merge(target_bci); 1141 } 1142 } 1143 } 1144 1145 untaken_branch = _gvn.transform(untaken_branch); 1146 set_control(untaken_branch); 1147 1148 // Branch not taken. 1149 if (stopped()) { 1150 if (C->eliminate_boxing()) { 1151 // Mark the successor block as parsed 1152 next_block->next_path_num(); 1153 } 1154 } else { 1155 // Update method data 1156 profile_not_taken_branch(); 1157 adjust_map_after_if(untaken_btest, c, untaken_prob, 1158 next_block, branch_block); 1159 } 1160 } 1161 1162 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1163 // Don't want to speculate on uncommon traps when running with -Xcomp 1164 if (!UseInterpreter) { 1165 return false; 1166 } 1167 return (seems_never_taken(prob) && seems_stable_comparison()); 1168 } 1169 1170 //----------------------------adjust_map_after_if------------------------------ 1171 // Adjust the JVM state to reflect the result of taking this path. 1172 // Basically, it means inspecting the CmpNode controlling this 1173 // branch, seeing how it constrains a tested value, and then 1174 // deciding if it's worth our while to encode this constraint 1175 // as graph nodes in the current abstract interpretation map. 1176 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1177 Block* path, Block* other_path) { 1178 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1179 return; // nothing to do 1180 1181 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1182 1183 if (path_is_suitable_for_uncommon_trap(prob)) { 1184 repush_if_args(); 1185 uncommon_trap(Deoptimization::Reason_unstable_if, 1186 Deoptimization::Action_reinterpret, 1187 NULL, 1188 (is_fallthrough ? "taken always" : "taken never")); 1189 return; 1190 } 1191 1192 Node* val = c->in(1); 1193 Node* con = c->in(2); 1194 const Type* tcon = _gvn.type(con); 1195 const Type* tval = _gvn.type(val); 1196 bool have_con = tcon->singleton(); 1197 if (tval->singleton()) { 1198 if (!have_con) { 1199 // Swap, so constant is in con. 1200 con = val; 1201 tcon = tval; 1202 val = c->in(2); 1203 tval = _gvn.type(val); 1204 btest = BoolTest(btest).commute(); 1205 have_con = true; 1206 } else { 1207 // Do we have two constants? Then leave well enough alone. 1208 have_con = false; 1209 } 1210 } 1211 if (!have_con) // remaining adjustments need a con 1212 return; 1213 1214 sharpen_type_after_if(btest, con, tcon, val, tval); 1215 } 1216 1217 1218 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1219 Node* ldk; 1220 if (n->is_DecodeNKlass()) { 1221 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1222 return NULL; 1223 } else { 1224 ldk = n->in(1); 1225 } 1226 } else if (n->Opcode() != Op_LoadKlass) { 1227 return NULL; 1228 } else { 1229 ldk = n; 1230 } 1231 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1232 1233 Node* adr = ldk->in(MemNode::Address); 1234 intptr_t off = 0; 1235 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1236 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1237 return NULL; 1238 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1239 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1240 return NULL; 1241 1242 return obj; 1243 } 1244 1245 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1246 Node* con, const Type* tcon, 1247 Node* val, const Type* tval) { 1248 // Look for opportunities to sharpen the type of a node 1249 // whose klass is compared with a constant klass. 1250 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1251 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1252 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1253 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1254 // Found: 1255 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1256 // or the narrowOop equivalent. 1257 const Type* obj_type = _gvn.type(obj); 1258 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1259 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1260 tboth->higher_equal(obj_type)) { 1261 // obj has to be of the exact type Foo if the CmpP succeeds. 1262 int obj_in_map = map()->find_edge(obj); 1263 JVMState* jvms = this->jvms(); 1264 if (obj_in_map >= 0 && 1265 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1266 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1267 const Type* tcc = ccast->as_Type()->type(); 1268 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1269 // Delay transform() call to allow recovery of pre-cast value 1270 // at the control merge. 1271 _gvn.set_type_bottom(ccast); 1272 record_for_igvn(ccast); 1273 // Here's the payoff. 1274 replace_in_map(obj, ccast); 1275 } 1276 } 1277 } 1278 } 1279 1280 int val_in_map = map()->find_edge(val); 1281 if (val_in_map < 0) return; // replace_in_map would be useless 1282 { 1283 JVMState* jvms = this->jvms(); 1284 if (!(jvms->is_loc(val_in_map) || 1285 jvms->is_stk(val_in_map))) 1286 return; // again, it would be useless 1287 } 1288 1289 // Check for a comparison to a constant, and "know" that the compared 1290 // value is constrained on this path. 1291 assert(tcon->singleton(), ""); 1292 ConstraintCastNode* ccast = NULL; 1293 Node* cast = NULL; 1294 1295 switch (btest) { 1296 case BoolTest::eq: // Constant test? 1297 { 1298 const Type* tboth = tcon->join_speculative(tval); 1299 if (tboth == tval) break; // Nothing to gain. 1300 if (tcon->isa_int()) { 1301 ccast = new CastIINode(val, tboth); 1302 } else if (tcon == TypePtr::NULL_PTR) { 1303 // Cast to null, but keep the pointer identity temporarily live. 1304 ccast = new CastPPNode(val, tboth); 1305 } else { 1306 const TypeF* tf = tcon->isa_float_constant(); 1307 const TypeD* td = tcon->isa_double_constant(); 1308 // Exclude tests vs float/double 0 as these could be 1309 // either +0 or -0. Just because you are equal to +0 1310 // doesn't mean you ARE +0! 1311 // Note, following code also replaces Long and Oop values. 1312 if ((!tf || tf->_f != 0.0) && 1313 (!td || td->_d != 0.0)) 1314 cast = con; // Replace non-constant val by con. 1315 } 1316 } 1317 break; 1318 1319 case BoolTest::ne: 1320 if (tcon == TypePtr::NULL_PTR) { 1321 cast = cast_not_null(val, false); 1322 } 1323 break; 1324 1325 default: 1326 // (At this point we could record int range types with CastII.) 1327 break; 1328 } 1329 1330 if (ccast != NULL) { 1331 const Type* tcc = ccast->as_Type()->type(); 1332 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1333 // Delay transform() call to allow recovery of pre-cast value 1334 // at the control merge. 1335 ccast->set_req(0, control()); 1336 _gvn.set_type_bottom(ccast); 1337 record_for_igvn(ccast); 1338 cast = ccast; 1339 } 1340 1341 if (cast != NULL) { // Here's the payoff. 1342 replace_in_map(val, cast); 1343 } 1344 } 1345 1346 /** 1347 * Use speculative type to optimize CmpP node: if comparison is 1348 * against the low level class, cast the object to the speculative 1349 * type if any. CmpP should then go away. 1350 * 1351 * @param c expected CmpP node 1352 * @return result of CmpP on object casted to speculative type 1353 * 1354 */ 1355 Node* Parse::optimize_cmp_with_klass(Node* c) { 1356 // If this is transformed by the _gvn to a comparison with the low 1357 // level klass then we may be able to use speculation 1358 if (c->Opcode() == Op_CmpP && 1359 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1360 c->in(2)->is_Con()) { 1361 Node* load_klass = NULL; 1362 Node* decode = NULL; 1363 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1364 decode = c->in(1); 1365 load_klass = c->in(1)->in(1); 1366 } else { 1367 load_klass = c->in(1); 1368 } 1369 if (load_klass->in(2)->is_AddP()) { 1370 Node* addp = load_klass->in(2); 1371 Node* obj = addp->in(AddPNode::Address); 1372 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1373 if (obj_type->speculative_type_not_null() != NULL) { 1374 ciKlass* k = obj_type->speculative_type(); 1375 inc_sp(2); 1376 obj = maybe_cast_profiled_obj(obj, k); 1377 dec_sp(2); 1378 // Make the CmpP use the casted obj 1379 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1380 load_klass = load_klass->clone(); 1381 load_klass->set_req(2, addp); 1382 load_klass = _gvn.transform(load_klass); 1383 if (decode != NULL) { 1384 decode = decode->clone(); 1385 decode->set_req(1, load_klass); 1386 load_klass = _gvn.transform(decode); 1387 } 1388 c = c->clone(); 1389 c->set_req(1, load_klass); 1390 c = _gvn.transform(c); 1391 } 1392 } 1393 } 1394 return c; 1395 } 1396 1397 //------------------------------do_one_bytecode-------------------------------- 1398 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1399 void Parse::do_one_bytecode() { 1400 Node *a, *b, *c, *d; // Handy temps 1401 BoolTest::mask btest; 1402 int i; 1403 1404 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1405 1406 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1407 "out of nodes parsing method")) { 1408 return; 1409 } 1410 1411 #ifdef ASSERT 1412 // for setting breakpoints 1413 if (TraceOptoParse) { 1414 tty->print(" @"); 1415 dump_bci(bci()); 1416 tty->cr(); 1417 } 1418 #endif 1419 1420 switch (bc()) { 1421 case Bytecodes::_nop: 1422 // do nothing 1423 break; 1424 case Bytecodes::_lconst_0: 1425 push_pair(longcon(0)); 1426 break; 1427 1428 case Bytecodes::_lconst_1: 1429 push_pair(longcon(1)); 1430 break; 1431 1432 case Bytecodes::_fconst_0: 1433 push(zerocon(T_FLOAT)); 1434 break; 1435 1436 case Bytecodes::_fconst_1: 1437 push(makecon(TypeF::ONE)); 1438 break; 1439 1440 case Bytecodes::_fconst_2: 1441 push(makecon(TypeF::make(2.0f))); 1442 break; 1443 1444 case Bytecodes::_dconst_0: 1445 push_pair(zerocon(T_DOUBLE)); 1446 break; 1447 1448 case Bytecodes::_dconst_1: 1449 push_pair(makecon(TypeD::ONE)); 1450 break; 1451 1452 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1453 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1454 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1455 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1456 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1457 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1458 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1459 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1460 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1461 case Bytecodes::_aconst_null: push(null()); break; 1462 case Bytecodes::_ldc: 1463 case Bytecodes::_ldc_w: 1464 case Bytecodes::_ldc2_w: 1465 // If the constant is unresolved, run this BC once in the interpreter. 1466 { 1467 ciConstant constant = iter().get_constant(); 1468 if (constant.basic_type() == T_OBJECT && 1469 !constant.as_object()->is_loaded()) { 1470 int index = iter().get_constant_pool_index(); 1471 constantTag tag = iter().get_constant_pool_tag(index); 1472 uncommon_trap(Deoptimization::make_trap_request 1473 (Deoptimization::Reason_unloaded, 1474 Deoptimization::Action_reinterpret, 1475 index), 1476 NULL, tag.internal_name()); 1477 break; 1478 } 1479 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1480 "must be java_mirror of klass"); 1481 const Type* con_type = Type::make_from_constant(constant); 1482 if (con_type != NULL) { 1483 push_node(con_type->basic_type(), makecon(con_type)); 1484 } 1485 } 1486 1487 break; 1488 1489 case Bytecodes::_aload_0: 1490 push( local(0) ); 1491 break; 1492 case Bytecodes::_aload_1: 1493 push( local(1) ); 1494 break; 1495 case Bytecodes::_aload_2: 1496 push( local(2) ); 1497 break; 1498 case Bytecodes::_aload_3: 1499 push( local(3) ); 1500 break; 1501 case Bytecodes::_aload: 1502 push( local(iter().get_index()) ); 1503 break; 1504 1505 case Bytecodes::_fload_0: 1506 case Bytecodes::_iload_0: 1507 push( local(0) ); 1508 break; 1509 case Bytecodes::_fload_1: 1510 case Bytecodes::_iload_1: 1511 push( local(1) ); 1512 break; 1513 case Bytecodes::_fload_2: 1514 case Bytecodes::_iload_2: 1515 push( local(2) ); 1516 break; 1517 case Bytecodes::_fload_3: 1518 case Bytecodes::_iload_3: 1519 push( local(3) ); 1520 break; 1521 case Bytecodes::_fload: 1522 case Bytecodes::_iload: 1523 push( local(iter().get_index()) ); 1524 break; 1525 case Bytecodes::_lload_0: 1526 push_pair_local( 0 ); 1527 break; 1528 case Bytecodes::_lload_1: 1529 push_pair_local( 1 ); 1530 break; 1531 case Bytecodes::_lload_2: 1532 push_pair_local( 2 ); 1533 break; 1534 case Bytecodes::_lload_3: 1535 push_pair_local( 3 ); 1536 break; 1537 case Bytecodes::_lload: 1538 push_pair_local( iter().get_index() ); 1539 break; 1540 1541 case Bytecodes::_dload_0: 1542 push_pair_local(0); 1543 break; 1544 case Bytecodes::_dload_1: 1545 push_pair_local(1); 1546 break; 1547 case Bytecodes::_dload_2: 1548 push_pair_local(2); 1549 break; 1550 case Bytecodes::_dload_3: 1551 push_pair_local(3); 1552 break; 1553 case Bytecodes::_dload: 1554 push_pair_local(iter().get_index()); 1555 break; 1556 case Bytecodes::_fstore_0: 1557 case Bytecodes::_istore_0: 1558 case Bytecodes::_astore_0: 1559 set_local( 0, pop() ); 1560 break; 1561 case Bytecodes::_fstore_1: 1562 case Bytecodes::_istore_1: 1563 case Bytecodes::_astore_1: 1564 set_local( 1, pop() ); 1565 break; 1566 case Bytecodes::_fstore_2: 1567 case Bytecodes::_istore_2: 1568 case Bytecodes::_astore_2: 1569 set_local( 2, pop() ); 1570 break; 1571 case Bytecodes::_fstore_3: 1572 case Bytecodes::_istore_3: 1573 case Bytecodes::_astore_3: 1574 set_local( 3, pop() ); 1575 break; 1576 case Bytecodes::_fstore: 1577 case Bytecodes::_istore: 1578 case Bytecodes::_astore: 1579 set_local( iter().get_index(), pop() ); 1580 break; 1581 // long stores 1582 case Bytecodes::_lstore_0: 1583 set_pair_local( 0, pop_pair() ); 1584 break; 1585 case Bytecodes::_lstore_1: 1586 set_pair_local( 1, pop_pair() ); 1587 break; 1588 case Bytecodes::_lstore_2: 1589 set_pair_local( 2, pop_pair() ); 1590 break; 1591 case Bytecodes::_lstore_3: 1592 set_pair_local( 3, pop_pair() ); 1593 break; 1594 case Bytecodes::_lstore: 1595 set_pair_local( iter().get_index(), pop_pair() ); 1596 break; 1597 1598 // double stores 1599 case Bytecodes::_dstore_0: 1600 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1601 break; 1602 case Bytecodes::_dstore_1: 1603 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1604 break; 1605 case Bytecodes::_dstore_2: 1606 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1607 break; 1608 case Bytecodes::_dstore_3: 1609 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1610 break; 1611 case Bytecodes::_dstore: 1612 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1613 break; 1614 1615 case Bytecodes::_pop: dec_sp(1); break; 1616 case Bytecodes::_pop2: dec_sp(2); break; 1617 case Bytecodes::_swap: 1618 a = pop(); 1619 b = pop(); 1620 push(a); 1621 push(b); 1622 break; 1623 case Bytecodes::_dup: 1624 a = pop(); 1625 push(a); 1626 push(a); 1627 break; 1628 case Bytecodes::_dup_x1: 1629 a = pop(); 1630 b = pop(); 1631 push( a ); 1632 push( b ); 1633 push( a ); 1634 break; 1635 case Bytecodes::_dup_x2: 1636 a = pop(); 1637 b = pop(); 1638 c = pop(); 1639 push( a ); 1640 push( c ); 1641 push( b ); 1642 push( a ); 1643 break; 1644 case Bytecodes::_dup2: 1645 a = pop(); 1646 b = pop(); 1647 push( b ); 1648 push( a ); 1649 push( b ); 1650 push( a ); 1651 break; 1652 1653 case Bytecodes::_dup2_x1: 1654 // before: .. c, b, a 1655 // after: .. b, a, c, b, a 1656 // not tested 1657 a = pop(); 1658 b = pop(); 1659 c = pop(); 1660 push( b ); 1661 push( a ); 1662 push( c ); 1663 push( b ); 1664 push( a ); 1665 break; 1666 case Bytecodes::_dup2_x2: 1667 // before: .. d, c, b, a 1668 // after: .. b, a, d, c, b, a 1669 // not tested 1670 a = pop(); 1671 b = pop(); 1672 c = pop(); 1673 d = pop(); 1674 push( b ); 1675 push( a ); 1676 push( d ); 1677 push( c ); 1678 push( b ); 1679 push( a ); 1680 break; 1681 1682 case Bytecodes::_arraylength: { 1683 // Must do null-check with value on expression stack 1684 Node *ary = null_check(peek(), T_ARRAY); 1685 // Compile-time detect of null-exception? 1686 if (stopped()) return; 1687 a = pop(); 1688 push(load_array_length(a)); 1689 break; 1690 } 1691 1692 case Bytecodes::_baload: array_load(T_BYTE); break; 1693 case Bytecodes::_caload: array_load(T_CHAR); break; 1694 case Bytecodes::_iaload: array_load(T_INT); break; 1695 case Bytecodes::_saload: array_load(T_SHORT); break; 1696 case Bytecodes::_faload: array_load(T_FLOAT); break; 1697 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1698 case Bytecodes::_laload: { 1699 a = array_addressing(T_LONG, 0); 1700 if (stopped()) return; // guaranteed null or range check 1701 dec_sp(2); // Pop array and index 1702 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1703 break; 1704 } 1705 case Bytecodes::_daload: { 1706 a = array_addressing(T_DOUBLE, 0); 1707 if (stopped()) return; // guaranteed null or range check 1708 dec_sp(2); // Pop array and index 1709 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1710 break; 1711 } 1712 case Bytecodes::_bastore: array_store(T_BYTE); break; 1713 case Bytecodes::_castore: array_store(T_CHAR); break; 1714 case Bytecodes::_iastore: array_store(T_INT); break; 1715 case Bytecodes::_sastore: array_store(T_SHORT); break; 1716 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1717 case Bytecodes::_aastore: { 1718 d = array_addressing(T_OBJECT, 1); 1719 if (stopped()) return; // guaranteed null or range check 1720 array_store_check(); 1721 c = pop(); // Oop to store 1722 b = pop(); // index (already used) 1723 a = pop(); // the array itself 1724 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1725 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1726 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, 1727 StoreNode::release_if_reference(T_OBJECT)); 1728 break; 1729 } 1730 case Bytecodes::_lastore: { 1731 a = array_addressing(T_LONG, 2); 1732 if (stopped()) return; // guaranteed null or range check 1733 c = pop_pair(); 1734 dec_sp(2); // Pop array and index 1735 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1736 break; 1737 } 1738 case Bytecodes::_dastore: { 1739 a = array_addressing(T_DOUBLE, 2); 1740 if (stopped()) return; // guaranteed null or range check 1741 c = pop_pair(); 1742 dec_sp(2); // Pop array and index 1743 c = dstore_rounding(c); 1744 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1745 break; 1746 } 1747 case Bytecodes::_getfield: 1748 do_getfield(); 1749 break; 1750 1751 case Bytecodes::_getstatic: 1752 do_getstatic(); 1753 break; 1754 1755 case Bytecodes::_putfield: 1756 do_putfield(); 1757 break; 1758 1759 case Bytecodes::_putstatic: 1760 do_putstatic(); 1761 break; 1762 1763 case Bytecodes::_irem: 1764 do_irem(); 1765 break; 1766 case Bytecodes::_idiv: 1767 // Must keep both values on the expression-stack during null-check 1768 zero_check_int(peek()); 1769 // Compile-time detect of null-exception? 1770 if (stopped()) return; 1771 b = pop(); 1772 a = pop(); 1773 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1774 break; 1775 case Bytecodes::_imul: 1776 b = pop(); a = pop(); 1777 push( _gvn.transform( new MulINode(a,b) ) ); 1778 break; 1779 case Bytecodes::_iadd: 1780 b = pop(); a = pop(); 1781 push( _gvn.transform( new AddINode(a,b) ) ); 1782 break; 1783 case Bytecodes::_ineg: 1784 a = pop(); 1785 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1786 break; 1787 case Bytecodes::_isub: 1788 b = pop(); a = pop(); 1789 push( _gvn.transform( new SubINode(a,b) ) ); 1790 break; 1791 case Bytecodes::_iand: 1792 b = pop(); a = pop(); 1793 push( _gvn.transform( new AndINode(a,b) ) ); 1794 break; 1795 case Bytecodes::_ior: 1796 b = pop(); a = pop(); 1797 push( _gvn.transform( new OrINode(a,b) ) ); 1798 break; 1799 case Bytecodes::_ixor: 1800 b = pop(); a = pop(); 1801 push( _gvn.transform( new XorINode(a,b) ) ); 1802 break; 1803 case Bytecodes::_ishl: 1804 b = pop(); a = pop(); 1805 push( _gvn.transform( new LShiftINode(a,b) ) ); 1806 break; 1807 case Bytecodes::_ishr: 1808 b = pop(); a = pop(); 1809 push( _gvn.transform( new RShiftINode(a,b) ) ); 1810 break; 1811 case Bytecodes::_iushr: 1812 b = pop(); a = pop(); 1813 push( _gvn.transform( new URShiftINode(a,b) ) ); 1814 break; 1815 1816 case Bytecodes::_fneg: 1817 a = pop(); 1818 b = _gvn.transform(new NegFNode (a)); 1819 push(b); 1820 break; 1821 1822 case Bytecodes::_fsub: 1823 b = pop(); 1824 a = pop(); 1825 c = _gvn.transform( new SubFNode(a,b) ); 1826 d = precision_rounding(c); 1827 push( d ); 1828 break; 1829 1830 case Bytecodes::_fadd: 1831 b = pop(); 1832 a = pop(); 1833 c = _gvn.transform( new AddFNode(a,b) ); 1834 d = precision_rounding(c); 1835 push( d ); 1836 break; 1837 1838 case Bytecodes::_fmul: 1839 b = pop(); 1840 a = pop(); 1841 c = _gvn.transform( new MulFNode(a,b) ); 1842 d = precision_rounding(c); 1843 push( d ); 1844 break; 1845 1846 case Bytecodes::_fdiv: 1847 b = pop(); 1848 a = pop(); 1849 c = _gvn.transform( new DivFNode(0,a,b) ); 1850 d = precision_rounding(c); 1851 push( d ); 1852 break; 1853 1854 case Bytecodes::_frem: 1855 if (Matcher::has_match_rule(Op_ModF)) { 1856 // Generate a ModF node. 1857 b = pop(); 1858 a = pop(); 1859 c = _gvn.transform( new ModFNode(0,a,b) ); 1860 d = precision_rounding(c); 1861 push( d ); 1862 } 1863 else { 1864 // Generate a call. 1865 modf(); 1866 } 1867 break; 1868 1869 case Bytecodes::_fcmpl: 1870 b = pop(); 1871 a = pop(); 1872 c = _gvn.transform( new CmpF3Node( a, b)); 1873 push(c); 1874 break; 1875 case Bytecodes::_fcmpg: 1876 b = pop(); 1877 a = pop(); 1878 1879 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1880 // which negates the result sign except for unordered. Flip the unordered 1881 // as well by using CmpF3 which implements unordered-lesser instead of 1882 // unordered-greater semantics. Finally, commute the result bits. Result 1883 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1884 c = _gvn.transform( new CmpF3Node( b, a)); 1885 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1886 push(c); 1887 break; 1888 1889 case Bytecodes::_f2i: 1890 a = pop(); 1891 push(_gvn.transform(new ConvF2INode(a))); 1892 break; 1893 1894 case Bytecodes::_d2i: 1895 a = pop_pair(); 1896 b = _gvn.transform(new ConvD2INode(a)); 1897 push( b ); 1898 break; 1899 1900 case Bytecodes::_f2d: 1901 a = pop(); 1902 b = _gvn.transform( new ConvF2DNode(a)); 1903 push_pair( b ); 1904 break; 1905 1906 case Bytecodes::_d2f: 1907 a = pop_pair(); 1908 b = _gvn.transform( new ConvD2FNode(a)); 1909 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1910 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1911 push( b ); 1912 break; 1913 1914 case Bytecodes::_l2f: 1915 if (Matcher::convL2FSupported()) { 1916 a = pop_pair(); 1917 b = _gvn.transform( new ConvL2FNode(a)); 1918 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1919 // Rather than storing the result into an FP register then pushing 1920 // out to memory to round, the machine instruction that implements 1921 // ConvL2D is responsible for rounding. 1922 // c = precision_rounding(b); 1923 c = _gvn.transform(b); 1924 push(c); 1925 } else { 1926 l2f(); 1927 } 1928 break; 1929 1930 case Bytecodes::_l2d: 1931 a = pop_pair(); 1932 b = _gvn.transform( new ConvL2DNode(a)); 1933 // For i486.ad, rounding is always necessary (see _l2f above). 1934 // c = dprecision_rounding(b); 1935 c = _gvn.transform(b); 1936 push_pair(c); 1937 break; 1938 1939 case Bytecodes::_f2l: 1940 a = pop(); 1941 b = _gvn.transform( new ConvF2LNode(a)); 1942 push_pair(b); 1943 break; 1944 1945 case Bytecodes::_d2l: 1946 a = pop_pair(); 1947 b = _gvn.transform( new ConvD2LNode(a)); 1948 push_pair(b); 1949 break; 1950 1951 case Bytecodes::_dsub: 1952 b = pop_pair(); 1953 a = pop_pair(); 1954 c = _gvn.transform( new SubDNode(a,b) ); 1955 d = dprecision_rounding(c); 1956 push_pair( d ); 1957 break; 1958 1959 case Bytecodes::_dadd: 1960 b = pop_pair(); 1961 a = pop_pair(); 1962 c = _gvn.transform( new AddDNode(a,b) ); 1963 d = dprecision_rounding(c); 1964 push_pair( d ); 1965 break; 1966 1967 case Bytecodes::_dmul: 1968 b = pop_pair(); 1969 a = pop_pair(); 1970 c = _gvn.transform( new MulDNode(a,b) ); 1971 d = dprecision_rounding(c); 1972 push_pair( d ); 1973 break; 1974 1975 case Bytecodes::_ddiv: 1976 b = pop_pair(); 1977 a = pop_pair(); 1978 c = _gvn.transform( new DivDNode(0,a,b) ); 1979 d = dprecision_rounding(c); 1980 push_pair( d ); 1981 break; 1982 1983 case Bytecodes::_dneg: 1984 a = pop_pair(); 1985 b = _gvn.transform(new NegDNode (a)); 1986 push_pair(b); 1987 break; 1988 1989 case Bytecodes::_drem: 1990 if (Matcher::has_match_rule(Op_ModD)) { 1991 // Generate a ModD node. 1992 b = pop_pair(); 1993 a = pop_pair(); 1994 // a % b 1995 1996 c = _gvn.transform( new ModDNode(0,a,b) ); 1997 d = dprecision_rounding(c); 1998 push_pair( d ); 1999 } 2000 else { 2001 // Generate a call. 2002 modd(); 2003 } 2004 break; 2005 2006 case Bytecodes::_dcmpl: 2007 b = pop_pair(); 2008 a = pop_pair(); 2009 c = _gvn.transform( new CmpD3Node( a, b)); 2010 push(c); 2011 break; 2012 2013 case Bytecodes::_dcmpg: 2014 b = pop_pair(); 2015 a = pop_pair(); 2016 // Same as dcmpl but need to flip the unordered case. 2017 // Commute the inputs, which negates the result sign except for unordered. 2018 // Flip the unordered as well by using CmpD3 which implements 2019 // unordered-lesser instead of unordered-greater semantics. 2020 // Finally, negate the result bits. Result is same as using a 2021 // CmpD3Greater except we did it with CmpD3 alone. 2022 c = _gvn.transform( new CmpD3Node( b, a)); 2023 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2024 push(c); 2025 break; 2026 2027 2028 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2029 case Bytecodes::_land: 2030 b = pop_pair(); 2031 a = pop_pair(); 2032 c = _gvn.transform( new AndLNode(a,b) ); 2033 push_pair(c); 2034 break; 2035 case Bytecodes::_lor: 2036 b = pop_pair(); 2037 a = pop_pair(); 2038 c = _gvn.transform( new OrLNode(a,b) ); 2039 push_pair(c); 2040 break; 2041 case Bytecodes::_lxor: 2042 b = pop_pair(); 2043 a = pop_pair(); 2044 c = _gvn.transform( new XorLNode(a,b) ); 2045 push_pair(c); 2046 break; 2047 2048 case Bytecodes::_lshl: 2049 b = pop(); // the shift count 2050 a = pop_pair(); // value to be shifted 2051 c = _gvn.transform( new LShiftLNode(a,b) ); 2052 push_pair(c); 2053 break; 2054 case Bytecodes::_lshr: 2055 b = pop(); // the shift count 2056 a = pop_pair(); // value to be shifted 2057 c = _gvn.transform( new RShiftLNode(a,b) ); 2058 push_pair(c); 2059 break; 2060 case Bytecodes::_lushr: 2061 b = pop(); // the shift count 2062 a = pop_pair(); // value to be shifted 2063 c = _gvn.transform( new URShiftLNode(a,b) ); 2064 push_pair(c); 2065 break; 2066 case Bytecodes::_lmul: 2067 b = pop_pair(); 2068 a = pop_pair(); 2069 c = _gvn.transform( new MulLNode(a,b) ); 2070 push_pair(c); 2071 break; 2072 2073 case Bytecodes::_lrem: 2074 // Must keep both values on the expression-stack during null-check 2075 assert(peek(0) == top(), "long word order"); 2076 zero_check_long(peek(1)); 2077 // Compile-time detect of null-exception? 2078 if (stopped()) return; 2079 b = pop_pair(); 2080 a = pop_pair(); 2081 c = _gvn.transform( new ModLNode(control(),a,b) ); 2082 push_pair(c); 2083 break; 2084 2085 case Bytecodes::_ldiv: 2086 // Must keep both values on the expression-stack during null-check 2087 assert(peek(0) == top(), "long word order"); 2088 zero_check_long(peek(1)); 2089 // Compile-time detect of null-exception? 2090 if (stopped()) return; 2091 b = pop_pair(); 2092 a = pop_pair(); 2093 c = _gvn.transform( new DivLNode(control(),a,b) ); 2094 push_pair(c); 2095 break; 2096 2097 case Bytecodes::_ladd: 2098 b = pop_pair(); 2099 a = pop_pair(); 2100 c = _gvn.transform( new AddLNode(a,b) ); 2101 push_pair(c); 2102 break; 2103 case Bytecodes::_lsub: 2104 b = pop_pair(); 2105 a = pop_pair(); 2106 c = _gvn.transform( new SubLNode(a,b) ); 2107 push_pair(c); 2108 break; 2109 case Bytecodes::_lcmp: 2110 // Safepoints are now inserted _before_ branches. The long-compare 2111 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2112 // slew of control flow. These are usually followed by a CmpI vs zero and 2113 // a branch; this pattern then optimizes to the obvious long-compare and 2114 // branch. However, if the branch is backwards there's a Safepoint 2115 // inserted. The inserted Safepoint captures the JVM state at the 2116 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2117 // long-compare is used to control a loop the debug info will force 2118 // computation of the 3-way value, even though the generated code uses a 2119 // long-compare and branch. We try to rectify the situation by inserting 2120 // a SafePoint here and have it dominate and kill the safepoint added at a 2121 // following backwards branch. At this point the JVM state merely holds 2 2122 // longs but not the 3-way value. 2123 if( UseLoopSafepoints ) { 2124 switch( iter().next_bc() ) { 2125 case Bytecodes::_ifgt: 2126 case Bytecodes::_iflt: 2127 case Bytecodes::_ifge: 2128 case Bytecodes::_ifle: 2129 case Bytecodes::_ifne: 2130 case Bytecodes::_ifeq: 2131 // If this is a backwards branch in the bytecodes, add Safepoint 2132 maybe_add_safepoint(iter().next_get_dest()); 2133 } 2134 } 2135 b = pop_pair(); 2136 a = pop_pair(); 2137 c = _gvn.transform( new CmpL3Node( a, b )); 2138 push(c); 2139 break; 2140 2141 case Bytecodes::_lneg: 2142 a = pop_pair(); 2143 b = _gvn.transform( new SubLNode(longcon(0),a)); 2144 push_pair(b); 2145 break; 2146 case Bytecodes::_l2i: 2147 a = pop_pair(); 2148 push( _gvn.transform( new ConvL2INode(a))); 2149 break; 2150 case Bytecodes::_i2l: 2151 a = pop(); 2152 b = _gvn.transform( new ConvI2LNode(a)); 2153 push_pair(b); 2154 break; 2155 case Bytecodes::_i2b: 2156 // Sign extend 2157 a = pop(); 2158 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2159 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2160 push( a ); 2161 break; 2162 case Bytecodes::_i2s: 2163 a = pop(); 2164 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2165 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2166 push( a ); 2167 break; 2168 case Bytecodes::_i2c: 2169 a = pop(); 2170 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2171 break; 2172 2173 case Bytecodes::_i2f: 2174 a = pop(); 2175 b = _gvn.transform( new ConvI2FNode(a) ) ; 2176 c = precision_rounding(b); 2177 push (b); 2178 break; 2179 2180 case Bytecodes::_i2d: 2181 a = pop(); 2182 b = _gvn.transform( new ConvI2DNode(a)); 2183 push_pair(b); 2184 break; 2185 2186 case Bytecodes::_iinc: // Increment local 2187 i = iter().get_index(); // Get local index 2188 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2189 break; 2190 2191 // Exit points of synchronized methods must have an unlock node 2192 case Bytecodes::_return: 2193 return_current(NULL); 2194 break; 2195 2196 case Bytecodes::_ireturn: 2197 case Bytecodes::_areturn: 2198 case Bytecodes::_freturn: 2199 return_current(pop()); 2200 break; 2201 case Bytecodes::_lreturn: 2202 return_current(pop_pair()); 2203 break; 2204 case Bytecodes::_dreturn: 2205 return_current(pop_pair()); 2206 break; 2207 2208 case Bytecodes::_athrow: 2209 // null exception oop throws NULL pointer exception 2210 null_check(peek()); 2211 if (stopped()) return; 2212 // Hook the thrown exception directly to subsequent handlers. 2213 if (BailoutToInterpreterForThrows) { 2214 // Keep method interpreted from now on. 2215 uncommon_trap(Deoptimization::Reason_unhandled, 2216 Deoptimization::Action_make_not_compilable); 2217 return; 2218 } 2219 if (env()->jvmti_can_post_on_exceptions()) { 2220 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2221 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2222 } 2223 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2224 add_exception_state(make_exception_state(peek())); 2225 break; 2226 2227 case Bytecodes::_goto: // fall through 2228 case Bytecodes::_goto_w: { 2229 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2230 2231 // If this is a backwards branch in the bytecodes, add Safepoint 2232 maybe_add_safepoint(target_bci); 2233 2234 // Update method data 2235 profile_taken_branch(target_bci); 2236 2237 // Merge the current control into the target basic block 2238 merge(target_bci); 2239 2240 // See if we can get some profile data and hand it off to the next block 2241 Block *target_block = block()->successor_for_bci(target_bci); 2242 if (target_block->pred_count() != 1) break; 2243 ciMethodData* methodData = method()->method_data(); 2244 if (!methodData->is_mature()) break; 2245 ciProfileData* data = methodData->bci_to_data(bci()); 2246 assert( data->is_JumpData(), "" ); 2247 int taken = ((ciJumpData*)data)->taken(); 2248 taken = method()->scale_count(taken); 2249 target_block->set_count(taken); 2250 break; 2251 } 2252 2253 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2254 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2255 handle_if_null: 2256 // If this is a backwards branch in the bytecodes, add Safepoint 2257 maybe_add_safepoint(iter().get_dest()); 2258 a = null(); 2259 b = pop(); 2260 if (!_gvn.type(b)->speculative_maybe_null() && 2261 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2262 inc_sp(1); 2263 Node* null_ctl = top(); 2264 b = null_check_oop(b, &null_ctl, true, true, true); 2265 assert(null_ctl->is_top(), "no null control here"); 2266 dec_sp(1); 2267 } 2268 c = _gvn.transform( new CmpPNode(b, a) ); 2269 do_ifnull(btest, c); 2270 break; 2271 2272 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2273 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2274 handle_if_acmp: 2275 // If this is a backwards branch in the bytecodes, add Safepoint 2276 maybe_add_safepoint(iter().get_dest()); 2277 a = pop(); 2278 b = pop(); 2279 c = _gvn.transform( new CmpPNode(b, a) ); 2280 c = optimize_cmp_with_klass(c); 2281 do_if(btest, c); 2282 break; 2283 2284 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2285 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2286 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2287 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2288 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2289 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2290 handle_ifxx: 2291 // If this is a backwards branch in the bytecodes, add Safepoint 2292 maybe_add_safepoint(iter().get_dest()); 2293 a = _gvn.intcon(0); 2294 b = pop(); 2295 c = _gvn.transform( new CmpINode(b, a) ); 2296 do_if(btest, c); 2297 break; 2298 2299 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2300 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2301 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2302 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2303 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2304 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2305 handle_if_icmp: 2306 // If this is a backwards branch in the bytecodes, add Safepoint 2307 maybe_add_safepoint(iter().get_dest()); 2308 a = pop(); 2309 b = pop(); 2310 c = _gvn.transform( new CmpINode( b, a ) ); 2311 do_if(btest, c); 2312 break; 2313 2314 case Bytecodes::_tableswitch: 2315 do_tableswitch(); 2316 break; 2317 2318 case Bytecodes::_lookupswitch: 2319 do_lookupswitch(); 2320 break; 2321 2322 case Bytecodes::_invokestatic: 2323 case Bytecodes::_invokedynamic: 2324 case Bytecodes::_invokespecial: 2325 case Bytecodes::_invokevirtual: 2326 case Bytecodes::_invokeinterface: 2327 do_call(); 2328 break; 2329 case Bytecodes::_checkcast: 2330 do_checkcast(); 2331 break; 2332 case Bytecodes::_instanceof: 2333 do_instanceof(); 2334 break; 2335 case Bytecodes::_anewarray: 2336 do_anewarray(); 2337 break; 2338 case Bytecodes::_newarray: 2339 do_newarray((BasicType)iter().get_index()); 2340 break; 2341 case Bytecodes::_multianewarray: 2342 do_multianewarray(); 2343 break; 2344 case Bytecodes::_new: 2345 do_new(); 2346 break; 2347 2348 case Bytecodes::_jsr: 2349 case Bytecodes::_jsr_w: 2350 do_jsr(); 2351 break; 2352 2353 case Bytecodes::_ret: 2354 do_ret(); 2355 break; 2356 2357 2358 case Bytecodes::_monitorenter: 2359 do_monitor_enter(); 2360 break; 2361 2362 case Bytecodes::_monitorexit: 2363 do_monitor_exit(); 2364 break; 2365 2366 case Bytecodes::_breakpoint: 2367 // Breakpoint set concurrently to compile 2368 // %%% use an uncommon trap? 2369 C->record_failure("breakpoint in method"); 2370 return; 2371 2372 default: 2373 #ifndef PRODUCT 2374 map()->dump(99); 2375 #endif 2376 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2377 ShouldNotReachHere(); 2378 } 2379 2380 #ifndef PRODUCT 2381 IdealGraphPrinter *printer = C->printer(); 2382 if (printer && printer->should_print(1)) { 2383 char buffer[256]; 2384 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2385 bool old = printer->traverse_outs(); 2386 printer->set_traverse_outs(true); 2387 printer->print_method(buffer, 4); 2388 printer->set_traverse_outs(old); 2389 } 2390 #endif 2391 }