1 /* 2 * Copyright (c) 1998, 2014, 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 "opto/addnode.hpp" 33 #include "opto/castnode.hpp" 34 #include "opto/convertnode.hpp" 35 #include "opto/divnode.hpp" 36 #include "opto/idealGraphPrinter.hpp" 37 #include "opto/matcher.hpp" 38 #include "opto/memnode.hpp" 39 #include "opto/mulnode.hpp" 40 #include "opto/opaquenode.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/runtime.hpp" 43 #include "runtime/deoptimization.hpp" 44 #include "runtime/sharedRuntime.hpp" 45 46 extern int explicit_null_checks_inserted, 47 explicit_null_checks_elided; 48 49 //---------------------------------array_load---------------------------------- 50 void Parse::array_load(BasicType elem_type) { 51 const Type* elem = Type::TOP; 52 Node* adr = array_addressing(elem_type, 0, &elem); 53 if (stopped()) return; // guaranteed null or range check 54 dec_sp(2); // Pop array and index 55 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 56 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 57 push(ld); 58 } 59 60 61 //--------------------------------array_store---------------------------------- 62 void Parse::array_store(BasicType elem_type) { 63 Node* adr = array_addressing(elem_type, 1); 64 if (stopped()) return; // guaranteed null or range check 65 Node* val = pop(); 66 dec_sp(2); // Pop array and index 67 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 68 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 69 } 70 71 72 //------------------------------array_addressing------------------------------- 73 // Pull array and index from the stack. Compute pointer-to-element. 74 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 75 Node *idx = peek(0+vals); // Get from stack without popping 76 Node *ary = peek(1+vals); // in case of exception 77 78 // Null check the array base, with correct stack contents 79 ary = null_check(ary, T_ARRAY); 80 // Compile-time detect of null-exception? 81 if (stopped()) return top(); 82 83 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 84 const TypeInt* sizetype = arytype->size(); 85 const Type* elemtype = arytype->elem(); 86 87 if (UseUniqueSubclasses && result2 != NULL) { 88 const Type* el = elemtype->make_ptr(); 89 if (el && el->isa_instptr()) { 90 const TypeInstPtr* toop = el->is_instptr(); 91 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 92 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 93 const Type* subklass = Type::get_const_type(toop->klass()); 94 elemtype = subklass->join_speculative(el); 95 } 96 } 97 } 98 99 // Check for big class initializers with all constant offsets 100 // feeding into a known-size array. 101 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 102 // See if the highest idx value is less than the lowest array bound, 103 // and if the idx value cannot be negative: 104 bool need_range_check = true; 105 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 106 need_range_check = false; 107 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 108 } 109 110 ciKlass * arytype_klass = arytype->klass(); 111 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 112 // Only fails for some -Xcomp runs 113 // The class is unloaded. We have to run this bytecode in the interpreter. 114 uncommon_trap(Deoptimization::Reason_unloaded, 115 Deoptimization::Action_reinterpret, 116 arytype->klass(), "!loaded array"); 117 return top(); 118 } 119 120 // Do the range check 121 if (GenerateRangeChecks && need_range_check) { 122 Node* tst; 123 if (sizetype->_hi <= 0) { 124 // The greatest array bound is negative, so we can conclude that we're 125 // compiling unreachable code, but the unsigned compare trick used below 126 // only works with non-negative lengths. Instead, hack "tst" to be zero so 127 // the uncommon_trap path will always be taken. 128 tst = _gvn.intcon(0); 129 } else { 130 // Range is constant in array-oop, so we can use the original state of mem 131 Node* len = load_array_length(ary); 132 133 // Test length vs index (standard trick using unsigned compare) 134 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 135 BoolTest::mask btest = BoolTest::lt; 136 tst = _gvn.transform( new BoolNode(chk, btest) ); 137 } 138 // Branch to failure if out of bounds 139 { BuildCutout unless(this, tst, PROB_MAX); 140 if (C->allow_range_check_smearing()) { 141 // Do not use builtin_throw, since range checks are sometimes 142 // made more stringent by an optimistic transformation. 143 // This creates "tentative" range checks at this point, 144 // which are not guaranteed to throw exceptions. 145 // See IfNode::Ideal, is_range_check, adjust_check. 146 uncommon_trap(Deoptimization::Reason_range_check, 147 Deoptimization::Action_make_not_entrant, 148 NULL, "range_check"); 149 } else { 150 // If we have already recompiled with the range-check-widening 151 // heroic optimization turned off, then we must really be throwing 152 // range check exceptions. 153 builtin_throw(Deoptimization::Reason_range_check, idx); 154 } 155 } 156 } 157 // Check for always knowing you are throwing a range-check exception 158 if (stopped()) return top(); 159 160 Node* ptr = array_element_address(ary, idx, type, sizetype); 161 162 if (result2 != NULL) *result2 = elemtype; 163 164 assert(ptr != top(), "top should go hand-in-hand with stopped"); 165 166 return ptr; 167 } 168 169 170 // returns IfNode 171 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 172 Node *cmp = _gvn.transform( new CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 173 Node *tst = _gvn.transform( new BoolNode( cmp, mask)); 174 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 175 return iff; 176 } 177 178 // return Region node 179 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 180 Node *region = new RegionNode(3); // 2 results 181 record_for_igvn(region); 182 region->init_req(1, iffalse); 183 region->init_req(2, iftrue ); 184 _gvn.set_type(region, Type::CONTROL); 185 region = _gvn.transform(region); 186 set_control (region); 187 return region; 188 } 189 190 191 //------------------------------helper for tableswitch------------------------- 192 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 193 // True branch, use existing map info 194 { PreserveJVMState pjvms(this); 195 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 196 set_control( iftrue ); 197 profile_switch_case(prof_table_index); 198 merge_new_path(dest_bci_if_true); 199 } 200 201 // False branch 202 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 203 set_control( iffalse ); 204 } 205 206 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 207 // True branch, use existing map info 208 { PreserveJVMState pjvms(this); 209 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 210 set_control( iffalse ); 211 profile_switch_case(prof_table_index); 212 merge_new_path(dest_bci_if_true); 213 } 214 215 // False branch 216 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 217 set_control( iftrue ); 218 } 219 220 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 221 // False branch, use existing map and control() 222 profile_switch_case(prof_table_index); 223 merge_new_path(dest_bci); 224 } 225 226 227 extern "C" { 228 static int jint_cmp(const void *i, const void *j) { 229 int a = *(jint *)i; 230 int b = *(jint *)j; 231 return a > b ? 1 : a < b ? -1 : 0; 232 } 233 } 234 235 236 // Default value for methodData switch indexing. Must be a negative value to avoid 237 // conflict with any legal switch index. 238 #define NullTableIndex -1 239 240 class SwitchRange : public StackObj { 241 // a range of integers coupled with a bci destination 242 jint _lo; // inclusive lower limit 243 jint _hi; // inclusive upper limit 244 int _dest; 245 int _table_index; // index into method data table 246 247 public: 248 jint lo() const { return _lo; } 249 jint hi() const { return _hi; } 250 int dest() const { return _dest; } 251 int table_index() const { return _table_index; } 252 bool is_singleton() const { return _lo == _hi; } 253 254 void setRange(jint lo, jint hi, int dest, int table_index) { 255 assert(lo <= hi, "must be a non-empty range"); 256 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 257 } 258 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 259 assert(lo <= hi, "must be a non-empty range"); 260 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 261 _hi = hi; 262 return true; 263 } 264 return false; 265 } 266 267 void set (jint value, int dest, int table_index) { 268 setRange(value, value, dest, table_index); 269 } 270 bool adjoin(jint value, int dest, int table_index) { 271 return adjoinRange(value, value, dest, table_index); 272 } 273 274 void print() { 275 if (is_singleton()) 276 tty->print(" {%d}=>%d", lo(), dest()); 277 else if (lo() == min_jint) 278 tty->print(" {..%d}=>%d", hi(), dest()); 279 else if (hi() == max_jint) 280 tty->print(" {%d..}=>%d", lo(), dest()); 281 else 282 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 283 } 284 }; 285 286 287 //-------------------------------do_tableswitch-------------------------------- 288 void Parse::do_tableswitch() { 289 Node* lookup = pop(); 290 291 // Get information about tableswitch 292 int default_dest = iter().get_dest_table(0); 293 int lo_index = iter().get_int_table(1); 294 int hi_index = iter().get_int_table(2); 295 int len = hi_index - lo_index + 1; 296 297 if (len < 1) { 298 // If this is a backward branch, add safepoint 299 maybe_add_safepoint(default_dest); 300 merge(default_dest); 301 return; 302 } 303 304 // generate decision tree, using trichotomy when possible 305 int rnum = len+2; 306 bool makes_backward_branch = false; 307 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 308 int rp = -1; 309 if (lo_index != min_jint) { 310 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 311 } 312 for (int j = 0; j < len; j++) { 313 jint match_int = lo_index+j; 314 int dest = iter().get_dest_table(j+3); 315 makes_backward_branch |= (dest <= bci()); 316 int table_index = method_data_update() ? j : NullTableIndex; 317 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 318 ranges[++rp].set(match_int, dest, table_index); 319 } 320 } 321 jint highest = lo_index+(len-1); 322 assert(ranges[rp].hi() == highest, ""); 323 if (highest != max_jint 324 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 325 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 326 } 327 assert(rp < len+2, "not too many ranges"); 328 329 // Safepoint in case if backward branch observed 330 if( makes_backward_branch && UseLoopSafepoints ) 331 add_safepoint(); 332 333 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 334 } 335 336 337 //------------------------------do_lookupswitch-------------------------------- 338 void Parse::do_lookupswitch() { 339 Node *lookup = pop(); // lookup value 340 // Get information about lookupswitch 341 int default_dest = iter().get_dest_table(0); 342 int len = iter().get_int_table(1); 343 344 if (len < 1) { // If this is a backward branch, add safepoint 345 maybe_add_safepoint(default_dest); 346 merge(default_dest); 347 return; 348 } 349 350 // generate decision tree, using trichotomy when possible 351 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 352 { 353 for( int j = 0; j < len; j++ ) { 354 table[j+j+0] = iter().get_int_table(2+j+j); 355 table[j+j+1] = iter().get_dest_table(2+j+j+1); 356 } 357 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 358 } 359 360 int rnum = len*2+1; 361 bool makes_backward_branch = false; 362 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 363 int rp = -1; 364 for( int j = 0; j < len; j++ ) { 365 jint match_int = table[j+j+0]; 366 int dest = table[j+j+1]; 367 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 368 int table_index = method_data_update() ? j : NullTableIndex; 369 makes_backward_branch |= (dest <= bci()); 370 if( match_int != next_lo ) { 371 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 372 } 373 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 374 ranges[++rp].set(match_int, dest, table_index); 375 } 376 } 377 jint highest = table[2*(len-1)]; 378 assert(ranges[rp].hi() == highest, ""); 379 if( highest != max_jint 380 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 381 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 382 } 383 assert(rp < rnum, "not too many ranges"); 384 385 // Safepoint in case backward branch observed 386 if( makes_backward_branch && UseLoopSafepoints ) 387 add_safepoint(); 388 389 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 390 } 391 392 //----------------------------create_jump_tables------------------------------- 393 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 394 // Are jumptables enabled 395 if (!UseJumpTables) return false; 396 397 // Are jumptables supported 398 if (!Matcher::has_match_rule(Op_Jump)) return false; 399 400 // Don't make jump table if profiling 401 if (method_data_update()) return false; 402 403 // Decide if a guard is needed to lop off big ranges at either (or 404 // both) end(s) of the input set. We'll call this the default target 405 // even though we can't be sure that it is the true "default". 406 407 bool needs_guard = false; 408 int default_dest; 409 int64_t total_outlier_size = 0; 410 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 411 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 412 413 if (lo->dest() == hi->dest()) { 414 total_outlier_size = hi_size + lo_size; 415 default_dest = lo->dest(); 416 } else if (lo_size > hi_size) { 417 total_outlier_size = lo_size; 418 default_dest = lo->dest(); 419 } else { 420 total_outlier_size = hi_size; 421 default_dest = hi->dest(); 422 } 423 424 // If a guard test will eliminate very sparse end ranges, then 425 // it is worth the cost of an extra jump. 426 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 427 needs_guard = true; 428 if (default_dest == lo->dest()) lo++; 429 if (default_dest == hi->dest()) hi--; 430 } 431 432 // Find the total number of cases and ranges 433 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 434 int num_range = hi - lo + 1; 435 436 // Don't create table if: too large, too small, or too sparse. 437 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 438 return false; 439 if (num_cases > (MaxJumpTableSparseness * num_range)) 440 return false; 441 442 // Normalize table lookups to zero 443 int lowval = lo->lo(); 444 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 445 446 // Generate a guard to protect against input keyvals that aren't 447 // in the switch domain. 448 if (needs_guard) { 449 Node* size = _gvn.intcon(num_cases); 450 Node* cmp = _gvn.transform( new CmpUNode(key_val, size) ); 451 Node* tst = _gvn.transform( new BoolNode(cmp, BoolTest::ge) ); 452 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 453 jump_if_true_fork(iff, default_dest, NullTableIndex); 454 } 455 456 // Create an ideal node JumpTable that has projections 457 // of all possible ranges for a switch statement 458 // The key_val input must be converted to a pointer offset and scaled. 459 // Compare Parse::array_addressing above. 460 #ifdef _LP64 461 // Clean the 32-bit int into a real 64-bit offset. 462 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 463 const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin); 464 key_val = _gvn.transform( new ConvI2LNode(key_val, lkeytype) ); 465 #endif 466 // Shift the value by wordsize so we have an index into the table, rather 467 // than a switch value 468 Node *shiftWord = _gvn.MakeConX(wordSize); 469 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 470 471 // Create the JumpNode 472 Node* jtn = _gvn.transform( new JumpNode(control(), key_val, num_cases) ); 473 474 // These are the switch destinations hanging off the jumpnode 475 int i = 0; 476 for (SwitchRange* r = lo; r <= hi; r++) { 477 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 478 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 479 { 480 PreserveJVMState pjvms(this); 481 set_control(input); 482 jump_if_always_fork(r->dest(), r->table_index()); 483 } 484 } 485 } 486 assert(i == num_cases, "miscount of cases"); 487 stop_and_kill_map(); // no more uses for this JVMS 488 return true; 489 } 490 491 //----------------------------jump_switch_ranges------------------------------- 492 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 493 Block* switch_block = block(); 494 495 if (switch_depth == 0) { 496 // Do special processing for the top-level call. 497 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 498 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 499 500 // Decrement pred-numbers for the unique set of nodes. 501 #ifdef ASSERT 502 // Ensure that the block's successors are a (duplicate-free) set. 503 int successors_counted = 0; // block occurrences in [hi..lo] 504 int unique_successors = switch_block->num_successors(); 505 for (int i = 0; i < unique_successors; i++) { 506 Block* target = switch_block->successor_at(i); 507 508 // Check that the set of successors is the same in both places. 509 int successors_found = 0; 510 for (SwitchRange* p = lo; p <= hi; p++) { 511 if (p->dest() == target->start()) successors_found++; 512 } 513 assert(successors_found > 0, "successor must be known"); 514 successors_counted += successors_found; 515 } 516 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 517 #endif 518 519 // Maybe prune the inputs, based on the type of key_val. 520 jint min_val = min_jint; 521 jint max_val = max_jint; 522 const TypeInt* ti = key_val->bottom_type()->isa_int(); 523 if (ti != NULL) { 524 min_val = ti->_lo; 525 max_val = ti->_hi; 526 assert(min_val <= max_val, "invalid int type"); 527 } 528 while (lo->hi() < min_val) lo++; 529 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 530 while (hi->lo() > max_val) hi--; 531 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 532 } 533 534 #ifndef PRODUCT 535 if (switch_depth == 0) { 536 _max_switch_depth = 0; 537 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 538 } 539 #endif 540 541 assert(lo <= hi, "must be a non-empty set of ranges"); 542 if (lo == hi) { 543 jump_if_always_fork(lo->dest(), lo->table_index()); 544 } else { 545 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 546 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 547 548 if (create_jump_tables(key_val, lo, hi)) return; 549 550 int nr = hi - lo + 1; 551 552 SwitchRange* mid = lo + nr/2; 553 // if there is an easy choice, pivot at a singleton: 554 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 555 556 assert(lo < mid && mid <= hi, "good pivot choice"); 557 assert(nr != 2 || mid == hi, "should pick higher of 2"); 558 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 559 560 Node *test_val = _gvn.intcon(mid->lo()); 561 562 if (mid->is_singleton()) { 563 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 564 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 565 566 // Special Case: If there are exactly three ranges, and the high 567 // and low range each go to the same place, omit the "gt" test, 568 // since it will not discriminate anything. 569 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 570 if (eq_test_only) { 571 assert(mid == hi-1, ""); 572 } 573 574 // if there is a higher range, test for it and process it: 575 if (mid < hi && !eq_test_only) { 576 // two comparisons of same values--should enable 1 test for 2 branches 577 // Use BoolTest::le instead of BoolTest::gt 578 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 579 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 580 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 581 { PreserveJVMState pjvms(this); 582 set_control(iffalse); 583 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 584 } 585 set_control(iftrue); 586 } 587 588 } else { 589 // mid is a range, not a singleton, so treat mid..hi as a unit 590 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 591 592 // if there is a higher range, test for it and process it: 593 if (mid == hi) { 594 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 595 } else { 596 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 597 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 598 { PreserveJVMState pjvms(this); 599 set_control(iftrue); 600 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 601 } 602 set_control(iffalse); 603 } 604 } 605 606 // in any case, process the lower range 607 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 608 } 609 610 // Decrease pred_count for each successor after all is done. 611 if (switch_depth == 0) { 612 int unique_successors = switch_block->num_successors(); 613 for (int i = 0; i < unique_successors; i++) { 614 Block* target = switch_block->successor_at(i); 615 // Throw away the pre-allocated path for each unique successor. 616 target->next_path_num(); 617 } 618 } 619 620 #ifndef PRODUCT 621 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 622 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 623 SwitchRange* r; 624 int nsing = 0; 625 for( r = lo; r <= hi; r++ ) { 626 if( r->is_singleton() ) nsing++; 627 } 628 tty->print(">>> "); 629 _method->print_short_name(); 630 tty->print_cr(" switch decision tree"); 631 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 632 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 633 if (_max_switch_depth > _est_switch_depth) { 634 tty->print_cr("******** BAD SWITCH DEPTH ********"); 635 } 636 tty->print(" "); 637 for( r = lo; r <= hi; r++ ) { 638 r->print(); 639 } 640 tty->cr(); 641 } 642 #endif 643 } 644 645 void Parse::modf() { 646 Node *f2 = pop(); 647 Node *f1 = pop(); 648 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 649 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 650 "frem", NULL, //no memory effects 651 f1, f2); 652 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 653 654 push(res); 655 } 656 657 void Parse::modd() { 658 Node *d2 = pop_pair(); 659 Node *d1 = pop_pair(); 660 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 661 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 662 "drem", NULL, //no memory effects 663 d1, top(), d2, top()); 664 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 665 666 #ifdef ASSERT 667 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 668 assert(res_top == top(), "second value must be top"); 669 #endif 670 671 push_pair(res_d); 672 } 673 674 void Parse::l2f() { 675 Node* f2 = pop(); 676 Node* f1 = pop(); 677 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 678 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 679 "l2f", NULL, //no memory effects 680 f1, f2); 681 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 682 683 push(res); 684 } 685 686 void Parse::do_irem() { 687 // Must keep both values on the expression-stack during null-check 688 zero_check_int(peek()); 689 // Compile-time detect of null-exception? 690 if (stopped()) return; 691 692 Node* b = pop(); 693 Node* a = pop(); 694 695 const Type *t = _gvn.type(b); 696 if (t != Type::TOP) { 697 const TypeInt *ti = t->is_int(); 698 if (ti->is_con()) { 699 int divisor = ti->get_con(); 700 // check for positive power of 2 701 if (divisor > 0 && 702 (divisor & ~(divisor-1)) == divisor) { 703 // yes ! 704 Node *mask = _gvn.intcon((divisor - 1)); 705 // Sigh, must handle negative dividends 706 Node *zero = _gvn.intcon(0); 707 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 708 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 709 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 710 Node *reg = jump_if_join(ift, iff); 711 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 712 // Negative path; negate/and/negate 713 Node *neg = _gvn.transform( new SubINode(zero, a) ); 714 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 715 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 716 phi->init_req(1, negn); 717 // Fast positive case 718 Node *andx = _gvn.transform( new AndINode(a, mask) ); 719 phi->init_req(2, andx); 720 // Push the merge 721 push( _gvn.transform(phi) ); 722 return; 723 } 724 } 725 } 726 // Default case 727 push( _gvn.transform( new ModINode(control(),a,b) ) ); 728 } 729 730 // Handle jsr and jsr_w bytecode 731 void Parse::do_jsr() { 732 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 733 734 // Store information about current state, tagged with new _jsr_bci 735 int return_bci = iter().next_bci(); 736 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 737 738 // Update method data 739 profile_taken_branch(jsr_bci); 740 741 // The way we do things now, there is only one successor block 742 // for the jsr, because the target code is cloned by ciTypeFlow. 743 Block* target = successor_for_bci(jsr_bci); 744 745 // What got pushed? 746 const Type* ret_addr = target->peek(); 747 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 748 749 // Effect on jsr on stack 750 push(_gvn.makecon(ret_addr)); 751 752 // Flow to the jsr. 753 merge(jsr_bci); 754 } 755 756 // Handle ret bytecode 757 void Parse::do_ret() { 758 // Find to whom we return. 759 assert(block()->num_successors() == 1, "a ret can only go one place now"); 760 Block* target = block()->successor_at(0); 761 assert(!target->is_ready(), "our arrival must be expected"); 762 profile_ret(target->flow()->start()); 763 int pnum = target->next_path_num(); 764 merge_common(target, pnum); 765 } 766 767 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 768 if (btest != BoolTest::eq && btest != BoolTest::ne) { 769 // Only ::eq and ::ne are supported for profile injection. 770 return false; 771 } 772 if (test->is_Cmp() && 773 test->in(1)->Opcode() == Op_ProfileBoolean) { 774 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 775 int false_cnt = profile->false_count(); 776 int true_cnt = profile->true_count(); 777 778 // Counts matching depends on the actual test operation (::eq or ::ne). 779 // No need to scale the counts because profile injection was designed 780 // to feed exact counts into VM. 781 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 782 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 783 784 profile->consume(); 785 return true; 786 } 787 return false; 788 } 789 //--------------------------dynamic_branch_prediction-------------------------- 790 // Try to gather dynamic branch prediction behavior. Return a probability 791 // of the branch being taken and set the "cnt" field. Returns a -1.0 792 // if we need to use static prediction for some reason. 793 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 794 ResourceMark rm; 795 796 cnt = COUNT_UNKNOWN; 797 798 int taken = 0; 799 int not_taken = 0; 800 801 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 802 803 if (use_mdo) { 804 // Use MethodData information if it is available 805 // FIXME: free the ProfileData structure 806 ciMethodData* methodData = method()->method_data(); 807 if (!methodData->is_mature()) return PROB_UNKNOWN; 808 ciProfileData* data = methodData->bci_to_data(bci()); 809 if (!data->is_JumpData()) return PROB_UNKNOWN; 810 811 // get taken and not taken values 812 taken = data->as_JumpData()->taken(); 813 not_taken = 0; 814 if (data->is_BranchData()) { 815 not_taken = data->as_BranchData()->not_taken(); 816 } 817 818 // scale the counts to be commensurate with invocation counts: 819 taken = method()->scale_count(taken); 820 not_taken = method()->scale_count(not_taken); 821 } 822 823 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 824 // We also check that individual counters are positive first, otherwise the sum can become positive. 825 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 826 if (C->log() != NULL) { 827 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 828 } 829 return PROB_UNKNOWN; 830 } 831 832 // Compute frequency that we arrive here 833 float sum = taken + not_taken; 834 // Adjust, if this block is a cloned private block but the 835 // Jump counts are shared. Taken the private counts for 836 // just this path instead of the shared counts. 837 if( block()->count() > 0 ) 838 sum = block()->count(); 839 cnt = sum / FreqCountInvocations; 840 841 // Pin probability to sane limits 842 float prob; 843 if( !taken ) 844 prob = (0+PROB_MIN) / 2; 845 else if( !not_taken ) 846 prob = (1+PROB_MAX) / 2; 847 else { // Compute probability of true path 848 prob = (float)taken / (float)(taken + not_taken); 849 if (prob > PROB_MAX) prob = PROB_MAX; 850 if (prob < PROB_MIN) prob = PROB_MIN; 851 } 852 853 assert((cnt > 0.0f) && (prob > 0.0f), 854 "Bad frequency assignment in if"); 855 856 if (C->log() != NULL) { 857 const char* prob_str = NULL; 858 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 859 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 860 char prob_str_buf[30]; 861 if (prob_str == NULL) { 862 sprintf(prob_str_buf, "%g", prob); 863 prob_str = prob_str_buf; 864 } 865 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'", 866 iter().get_dest(), taken, not_taken, cnt, prob_str); 867 } 868 return prob; 869 } 870 871 //-----------------------------branch_prediction------------------------------- 872 float Parse::branch_prediction(float& cnt, 873 BoolTest::mask btest, 874 int target_bci, 875 Node* test) { 876 float prob = dynamic_branch_prediction(cnt, btest, test); 877 // If prob is unknown, switch to static prediction 878 if (prob != PROB_UNKNOWN) return prob; 879 880 prob = PROB_FAIR; // Set default value 881 if (btest == BoolTest::eq) // Exactly equal test? 882 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 883 else if (btest == BoolTest::ne) 884 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 885 886 // If this is a conditional test guarding a backwards branch, 887 // assume its a loop-back edge. Make it a likely taken branch. 888 if (target_bci < bci()) { 889 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 890 // Since it's an OSR, we probably have profile data, but since 891 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 892 // Let's make a special check here for completely zero counts. 893 ciMethodData* methodData = method()->method_data(); 894 if (!methodData->is_empty()) { 895 ciProfileData* data = methodData->bci_to_data(bci()); 896 // Only stop for truly zero counts, which mean an unknown part 897 // of the OSR-ed method, and we want to deopt to gather more stats. 898 // If you have ANY counts, then this loop is simply 'cold' relative 899 // to the OSR loop. 900 if (data->as_BranchData()->taken() + 901 data->as_BranchData()->not_taken() == 0 ) { 902 // This is the only way to return PROB_UNKNOWN: 903 return PROB_UNKNOWN; 904 } 905 } 906 } 907 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 908 } 909 910 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 911 return prob; 912 } 913 914 // The magic constants are chosen so as to match the output of 915 // branch_prediction() when the profile reports a zero taken count. 916 // It is important to distinguish zero counts unambiguously, because 917 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 918 // very small but nonzero probabilities, which if confused with zero 919 // counts would keep the program recompiling indefinitely. 920 bool Parse::seems_never_taken(float prob) const { 921 return prob < PROB_MIN; 922 } 923 924 // True if the comparison seems to be the kind that will not change its 925 // statistics from true to false. See comments in adjust_map_after_if. 926 // This question is only asked along paths which are already 927 // classifed as untaken (by seems_never_taken), so really, 928 // if a path is never taken, its controlling comparison is 929 // already acting in a stable fashion. If the comparison 930 // seems stable, we will put an expensive uncommon trap 931 // on the untaken path. 932 bool Parse::seems_stable_comparison() const { 933 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 934 return false; 935 } 936 return true; 937 } 938 939 //-------------------------------repush_if_args-------------------------------- 940 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 941 inline int Parse::repush_if_args() { 942 #ifndef PRODUCT 943 if (PrintOpto && WizardMode) { 944 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 945 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 946 method()->print_name(); tty->cr(); 947 } 948 #endif 949 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 950 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 951 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 952 assert(argument(0) != NULL, "must exist"); 953 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 954 inc_sp(bc_depth); 955 return bc_depth; 956 } 957 958 //----------------------------------do_ifnull---------------------------------- 959 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 960 int target_bci = iter().get_dest(); 961 962 Block* branch_block = successor_for_bci(target_bci); 963 Block* next_block = successor_for_bci(iter().next_bci()); 964 965 float cnt; 966 float prob = branch_prediction(cnt, btest, target_bci, c); 967 if (prob == PROB_UNKNOWN) { 968 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 969 #ifndef PRODUCT 970 if (PrintOpto && Verbose) 971 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 972 #endif 973 repush_if_args(); // to gather stats on loop 974 // We need to mark this branch as taken so that if we recompile we will 975 // see that it is possible. In the tiered system the interpreter doesn't 976 // do profiling and by the time we get to the lower tier from the interpreter 977 // the path may be cold again. Make sure it doesn't look untaken 978 profile_taken_branch(target_bci, !ProfileInterpreter); 979 uncommon_trap(Deoptimization::Reason_unreached, 980 Deoptimization::Action_reinterpret, 981 NULL, "cold"); 982 if (C->eliminate_boxing()) { 983 // Mark the successor blocks as parsed 984 branch_block->next_path_num(); 985 next_block->next_path_num(); 986 } 987 return; 988 } 989 990 explicit_null_checks_inserted++; 991 992 // Generate real control flow 993 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 994 995 // Sanity check the probability value 996 assert(prob > 0.0f,"Bad probability in Parser"); 997 // Need xform to put node in hash table 998 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 999 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1000 // True branch 1001 { PreserveJVMState pjvms(this); 1002 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1003 set_control(iftrue); 1004 1005 if (stopped()) { // Path is dead? 1006 explicit_null_checks_elided++; 1007 if (C->eliminate_boxing()) { 1008 // Mark the successor block as parsed 1009 branch_block->next_path_num(); 1010 } 1011 } else { // Path is live. 1012 // Update method data 1013 profile_taken_branch(target_bci); 1014 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1015 if (!stopped()) { 1016 merge(target_bci); 1017 } 1018 } 1019 } 1020 1021 // False branch 1022 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1023 set_control(iffalse); 1024 1025 if (stopped()) { // Path is dead? 1026 explicit_null_checks_elided++; 1027 if (C->eliminate_boxing()) { 1028 // Mark the successor block as parsed 1029 next_block->next_path_num(); 1030 } 1031 } else { // Path is live. 1032 // Update method data 1033 profile_not_taken_branch(); 1034 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1035 next_block, branch_block); 1036 } 1037 } 1038 1039 //------------------------------------do_if------------------------------------ 1040 void Parse::do_if(BoolTest::mask btest, Node* c) { 1041 int target_bci = iter().get_dest(); 1042 1043 Block* branch_block = successor_for_bci(target_bci); 1044 Block* next_block = successor_for_bci(iter().next_bci()); 1045 1046 float cnt; 1047 float prob = branch_prediction(cnt, btest, target_bci, c); 1048 float untaken_prob = 1.0 - prob; 1049 1050 if (prob == PROB_UNKNOWN) { 1051 #ifndef PRODUCT 1052 if (PrintOpto && Verbose) 1053 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1054 #endif 1055 repush_if_args(); // to gather stats on loop 1056 // We need to mark this branch as taken so that if we recompile we will 1057 // see that it is possible. In the tiered system the interpreter doesn't 1058 // do profiling and by the time we get to the lower tier from the interpreter 1059 // the path may be cold again. Make sure it doesn't look untaken 1060 profile_taken_branch(target_bci, !ProfileInterpreter); 1061 uncommon_trap(Deoptimization::Reason_unreached, 1062 Deoptimization::Action_reinterpret, 1063 NULL, "cold"); 1064 if (C->eliminate_boxing()) { 1065 // Mark the successor blocks as parsed 1066 branch_block->next_path_num(); 1067 next_block->next_path_num(); 1068 } 1069 return; 1070 } 1071 1072 // Sanity check the probability value 1073 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1074 1075 bool taken_if_true = true; 1076 // Convert BoolTest to canonical form: 1077 if (!BoolTest(btest).is_canonical()) { 1078 btest = BoolTest(btest).negate(); 1079 taken_if_true = false; 1080 // prob is NOT updated here; it remains the probability of the taken 1081 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1082 } 1083 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1084 1085 Node* tst0 = new BoolNode(c, btest); 1086 Node* tst = _gvn.transform(tst0); 1087 BoolTest::mask taken_btest = BoolTest::illegal; 1088 BoolTest::mask untaken_btest = BoolTest::illegal; 1089 1090 if (tst->is_Bool()) { 1091 // Refresh c from the transformed bool node, since it may be 1092 // simpler than the original c. Also re-canonicalize btest. 1093 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1094 // That can arise from statements like: if (x instanceof C) ... 1095 if (tst != tst0) { 1096 // Canonicalize one more time since transform can change it. 1097 btest = tst->as_Bool()->_test._test; 1098 if (!BoolTest(btest).is_canonical()) { 1099 // Reverse edges one more time... 1100 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1101 btest = tst->as_Bool()->_test._test; 1102 assert(BoolTest(btest).is_canonical(), "sanity"); 1103 taken_if_true = !taken_if_true; 1104 } 1105 c = tst->in(1); 1106 } 1107 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1108 taken_btest = taken_if_true ? btest : neg_btest; 1109 untaken_btest = taken_if_true ? neg_btest : btest; 1110 } 1111 1112 // Generate real control flow 1113 float true_prob = (taken_if_true ? prob : untaken_prob); 1114 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1115 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1116 Node* taken_branch = new IfTrueNode(iff); 1117 Node* untaken_branch = new IfFalseNode(iff); 1118 if (!taken_if_true) { // Finish conversion to canonical form 1119 Node* tmp = taken_branch; 1120 taken_branch = untaken_branch; 1121 untaken_branch = tmp; 1122 } 1123 1124 // Branch is taken: 1125 { PreserveJVMState pjvms(this); 1126 taken_branch = _gvn.transform(taken_branch); 1127 set_control(taken_branch); 1128 1129 if (stopped()) { 1130 if (C->eliminate_boxing()) { 1131 // Mark the successor block as parsed 1132 branch_block->next_path_num(); 1133 } 1134 } else { 1135 // Update method data 1136 profile_taken_branch(target_bci); 1137 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1138 if (!stopped()) { 1139 merge(target_bci); 1140 } 1141 } 1142 } 1143 1144 untaken_branch = _gvn.transform(untaken_branch); 1145 set_control(untaken_branch); 1146 1147 // Branch not taken. 1148 if (stopped()) { 1149 if (C->eliminate_boxing()) { 1150 // Mark the successor block as parsed 1151 next_block->next_path_num(); 1152 } 1153 } else { 1154 // Update method data 1155 profile_not_taken_branch(); 1156 adjust_map_after_if(untaken_btest, c, untaken_prob, 1157 next_block, branch_block); 1158 } 1159 } 1160 1161 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1162 // Don't want to speculate on uncommon traps when running with -Xcomp 1163 if (!UseInterpreter) { 1164 return false; 1165 } 1166 return (seems_never_taken(prob) && seems_stable_comparison()); 1167 } 1168 1169 //----------------------------adjust_map_after_if------------------------------ 1170 // Adjust the JVM state to reflect the result of taking this path. 1171 // Basically, it means inspecting the CmpNode controlling this 1172 // branch, seeing how it constrains a tested value, and then 1173 // deciding if it's worth our while to encode this constraint 1174 // as graph nodes in the current abstract interpretation map. 1175 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1176 Block* path, Block* other_path) { 1177 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1178 return; // nothing to do 1179 1180 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1181 1182 if (path_is_suitable_for_uncommon_trap(prob)) { 1183 repush_if_args(); 1184 uncommon_trap(Deoptimization::Reason_unstable_if, 1185 Deoptimization::Action_reinterpret, 1186 NULL, 1187 (is_fallthrough ? "taken always" : "taken never")); 1188 return; 1189 } 1190 1191 Node* val = c->in(1); 1192 Node* con = c->in(2); 1193 const Type* tcon = _gvn.type(con); 1194 const Type* tval = _gvn.type(val); 1195 bool have_con = tcon->singleton(); 1196 if (tval->singleton()) { 1197 if (!have_con) { 1198 // Swap, so constant is in con. 1199 con = val; 1200 tcon = tval; 1201 val = c->in(2); 1202 tval = _gvn.type(val); 1203 btest = BoolTest(btest).commute(); 1204 have_con = true; 1205 } else { 1206 // Do we have two constants? Then leave well enough alone. 1207 have_con = false; 1208 } 1209 } 1210 if (!have_con) // remaining adjustments need a con 1211 return; 1212 1213 sharpen_type_after_if(btest, con, tcon, val, tval); 1214 } 1215 1216 1217 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1218 Node* ldk; 1219 if (n->is_DecodeNKlass()) { 1220 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1221 return NULL; 1222 } else { 1223 ldk = n->in(1); 1224 } 1225 } else if (n->Opcode() != Op_LoadKlass) { 1226 return NULL; 1227 } else { 1228 ldk = n; 1229 } 1230 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1231 1232 Node* adr = ldk->in(MemNode::Address); 1233 intptr_t off = 0; 1234 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1235 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1236 return NULL; 1237 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1238 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1239 return NULL; 1240 1241 return obj; 1242 } 1243 1244 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1245 Node* con, const Type* tcon, 1246 Node* val, const Type* tval) { 1247 // Look for opportunities to sharpen the type of a node 1248 // whose klass is compared with a constant klass. 1249 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1250 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1251 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1252 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1253 // Found: 1254 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1255 // or the narrowOop equivalent. 1256 const Type* obj_type = _gvn.type(obj); 1257 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1258 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1259 tboth->higher_equal(obj_type)) { 1260 // obj has to be of the exact type Foo if the CmpP succeeds. 1261 int obj_in_map = map()->find_edge(obj); 1262 JVMState* jvms = this->jvms(); 1263 if (obj_in_map >= 0 && 1264 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1265 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1266 const Type* tcc = ccast->as_Type()->type(); 1267 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1268 // Delay transform() call to allow recovery of pre-cast value 1269 // at the control merge. 1270 _gvn.set_type_bottom(ccast); 1271 record_for_igvn(ccast); 1272 // Here's the payoff. 1273 replace_in_map(obj, ccast); 1274 } 1275 } 1276 } 1277 } 1278 1279 int val_in_map = map()->find_edge(val); 1280 if (val_in_map < 0) return; // replace_in_map would be useless 1281 { 1282 JVMState* jvms = this->jvms(); 1283 if (!(jvms->is_loc(val_in_map) || 1284 jvms->is_stk(val_in_map))) 1285 return; // again, it would be useless 1286 } 1287 1288 // Check for a comparison to a constant, and "know" that the compared 1289 // value is constrained on this path. 1290 assert(tcon->singleton(), ""); 1291 ConstraintCastNode* ccast = NULL; 1292 Node* cast = NULL; 1293 1294 switch (btest) { 1295 case BoolTest::eq: // Constant test? 1296 { 1297 const Type* tboth = tcon->join_speculative(tval); 1298 if (tboth == tval) break; // Nothing to gain. 1299 if (tcon->isa_int()) { 1300 ccast = new CastIINode(val, tboth); 1301 } else if (tcon == TypePtr::NULL_PTR) { 1302 // Cast to null, but keep the pointer identity temporarily live. 1303 ccast = new CastPPNode(val, tboth); 1304 } else { 1305 const TypeF* tf = tcon->isa_float_constant(); 1306 const TypeD* td = tcon->isa_double_constant(); 1307 // Exclude tests vs float/double 0 as these could be 1308 // either +0 or -0. Just because you are equal to +0 1309 // doesn't mean you ARE +0! 1310 // Note, following code also replaces Long and Oop values. 1311 if ((!tf || tf->_f != 0.0) && 1312 (!td || td->_d != 0.0)) 1313 cast = con; // Replace non-constant val by con. 1314 } 1315 } 1316 break; 1317 1318 case BoolTest::ne: 1319 if (tcon == TypePtr::NULL_PTR) { 1320 cast = cast_not_null(val, false); 1321 } 1322 break; 1323 1324 default: 1325 // (At this point we could record int range types with CastII.) 1326 break; 1327 } 1328 1329 if (ccast != NULL) { 1330 const Type* tcc = ccast->as_Type()->type(); 1331 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1332 // Delay transform() call to allow recovery of pre-cast value 1333 // at the control merge. 1334 ccast->set_req(0, control()); 1335 _gvn.set_type_bottom(ccast); 1336 record_for_igvn(ccast); 1337 cast = ccast; 1338 } 1339 1340 if (cast != NULL) { // Here's the payoff. 1341 replace_in_map(val, cast); 1342 } 1343 } 1344 1345 /** 1346 * Use speculative type to optimize CmpP node: if comparison is 1347 * against the low level class, cast the object to the speculative 1348 * type if any. CmpP should then go away. 1349 * 1350 * @param c expected CmpP node 1351 * @return result of CmpP on object casted to speculative type 1352 * 1353 */ 1354 Node* Parse::optimize_cmp_with_klass(Node* c) { 1355 // If this is transformed by the _gvn to a comparison with the low 1356 // level klass then we may be able to use speculation 1357 if (c->Opcode() == Op_CmpP && 1358 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1359 c->in(2)->is_Con()) { 1360 Node* load_klass = NULL; 1361 Node* decode = NULL; 1362 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1363 decode = c->in(1); 1364 load_klass = c->in(1)->in(1); 1365 } else { 1366 load_klass = c->in(1); 1367 } 1368 if (load_klass->in(2)->is_AddP()) { 1369 Node* addp = load_klass->in(2); 1370 Node* obj = addp->in(AddPNode::Address); 1371 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1372 if (obj_type->speculative_type_not_null() != NULL) { 1373 ciKlass* k = obj_type->speculative_type(); 1374 inc_sp(2); 1375 obj = maybe_cast_profiled_obj(obj, k); 1376 dec_sp(2); 1377 // Make the CmpP use the casted obj 1378 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1379 load_klass = load_klass->clone(); 1380 load_klass->set_req(2, addp); 1381 load_klass = _gvn.transform(load_klass); 1382 if (decode != NULL) { 1383 decode = decode->clone(); 1384 decode->set_req(1, load_klass); 1385 load_klass = _gvn.transform(decode); 1386 } 1387 c = c->clone(); 1388 c->set_req(1, load_klass); 1389 c = _gvn.transform(c); 1390 } 1391 } 1392 } 1393 return c; 1394 } 1395 1396 //------------------------------do_one_bytecode-------------------------------- 1397 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1398 void Parse::do_one_bytecode() { 1399 Node *a, *b, *c, *d; // Handy temps 1400 BoolTest::mask btest; 1401 int i; 1402 1403 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1404 1405 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1406 "out of nodes parsing method")) { 1407 return; 1408 } 1409 1410 #ifdef ASSERT 1411 // for setting breakpoints 1412 if (TraceOptoParse) { 1413 tty->print(" @"); 1414 dump_bci(bci()); 1415 tty->cr(); 1416 } 1417 #endif 1418 1419 switch (bc()) { 1420 case Bytecodes::_nop: 1421 // do nothing 1422 break; 1423 case Bytecodes::_lconst_0: 1424 push_pair(longcon(0)); 1425 break; 1426 1427 case Bytecodes::_lconst_1: 1428 push_pair(longcon(1)); 1429 break; 1430 1431 case Bytecodes::_fconst_0: 1432 push(zerocon(T_FLOAT)); 1433 break; 1434 1435 case Bytecodes::_fconst_1: 1436 push(makecon(TypeF::ONE)); 1437 break; 1438 1439 case Bytecodes::_fconst_2: 1440 push(makecon(TypeF::make(2.0f))); 1441 break; 1442 1443 case Bytecodes::_dconst_0: 1444 push_pair(zerocon(T_DOUBLE)); 1445 break; 1446 1447 case Bytecodes::_dconst_1: 1448 push_pair(makecon(TypeD::ONE)); 1449 break; 1450 1451 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1452 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1453 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1454 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1455 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1456 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1457 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1458 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1459 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1460 case Bytecodes::_aconst_null: push(null()); break; 1461 case Bytecodes::_ldc: 1462 case Bytecodes::_ldc_w: 1463 case Bytecodes::_ldc2_w: 1464 // If the constant is unresolved, run this BC once in the interpreter. 1465 { 1466 ciConstant constant = iter().get_constant(); 1467 if (constant.basic_type() == T_OBJECT && 1468 !constant.as_object()->is_loaded()) { 1469 int index = iter().get_constant_pool_index(); 1470 constantTag tag = iter().get_constant_pool_tag(index); 1471 uncommon_trap(Deoptimization::make_trap_request 1472 (Deoptimization::Reason_unloaded, 1473 Deoptimization::Action_reinterpret, 1474 index), 1475 NULL, tag.internal_name()); 1476 break; 1477 } 1478 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1479 "must be java_mirror of klass"); 1480 bool pushed = push_constant(constant, true); 1481 guarantee(pushed, "must be possible to push this constant"); 1482 } 1483 1484 break; 1485 1486 case Bytecodes::_aload_0: 1487 push( local(0) ); 1488 break; 1489 case Bytecodes::_aload_1: 1490 push( local(1) ); 1491 break; 1492 case Bytecodes::_aload_2: 1493 push( local(2) ); 1494 break; 1495 case Bytecodes::_aload_3: 1496 push( local(3) ); 1497 break; 1498 case Bytecodes::_aload: 1499 push( local(iter().get_index()) ); 1500 break; 1501 1502 case Bytecodes::_fload_0: 1503 case Bytecodes::_iload_0: 1504 push( local(0) ); 1505 break; 1506 case Bytecodes::_fload_1: 1507 case Bytecodes::_iload_1: 1508 push( local(1) ); 1509 break; 1510 case Bytecodes::_fload_2: 1511 case Bytecodes::_iload_2: 1512 push( local(2) ); 1513 break; 1514 case Bytecodes::_fload_3: 1515 case Bytecodes::_iload_3: 1516 push( local(3) ); 1517 break; 1518 case Bytecodes::_fload: 1519 case Bytecodes::_iload: 1520 push( local(iter().get_index()) ); 1521 break; 1522 case Bytecodes::_lload_0: 1523 push_pair_local( 0 ); 1524 break; 1525 case Bytecodes::_lload_1: 1526 push_pair_local( 1 ); 1527 break; 1528 case Bytecodes::_lload_2: 1529 push_pair_local( 2 ); 1530 break; 1531 case Bytecodes::_lload_3: 1532 push_pair_local( 3 ); 1533 break; 1534 case Bytecodes::_lload: 1535 push_pair_local( iter().get_index() ); 1536 break; 1537 1538 case Bytecodes::_dload_0: 1539 push_pair_local(0); 1540 break; 1541 case Bytecodes::_dload_1: 1542 push_pair_local(1); 1543 break; 1544 case Bytecodes::_dload_2: 1545 push_pair_local(2); 1546 break; 1547 case Bytecodes::_dload_3: 1548 push_pair_local(3); 1549 break; 1550 case Bytecodes::_dload: 1551 push_pair_local(iter().get_index()); 1552 break; 1553 case Bytecodes::_fstore_0: 1554 case Bytecodes::_istore_0: 1555 case Bytecodes::_astore_0: 1556 set_local( 0, pop() ); 1557 break; 1558 case Bytecodes::_fstore_1: 1559 case Bytecodes::_istore_1: 1560 case Bytecodes::_astore_1: 1561 set_local( 1, pop() ); 1562 break; 1563 case Bytecodes::_fstore_2: 1564 case Bytecodes::_istore_2: 1565 case Bytecodes::_astore_2: 1566 set_local( 2, pop() ); 1567 break; 1568 case Bytecodes::_fstore_3: 1569 case Bytecodes::_istore_3: 1570 case Bytecodes::_astore_3: 1571 set_local( 3, pop() ); 1572 break; 1573 case Bytecodes::_fstore: 1574 case Bytecodes::_istore: 1575 case Bytecodes::_astore: 1576 set_local( iter().get_index(), pop() ); 1577 break; 1578 // long stores 1579 case Bytecodes::_lstore_0: 1580 set_pair_local( 0, pop_pair() ); 1581 break; 1582 case Bytecodes::_lstore_1: 1583 set_pair_local( 1, pop_pair() ); 1584 break; 1585 case Bytecodes::_lstore_2: 1586 set_pair_local( 2, pop_pair() ); 1587 break; 1588 case Bytecodes::_lstore_3: 1589 set_pair_local( 3, pop_pair() ); 1590 break; 1591 case Bytecodes::_lstore: 1592 set_pair_local( iter().get_index(), pop_pair() ); 1593 break; 1594 1595 // double stores 1596 case Bytecodes::_dstore_0: 1597 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1598 break; 1599 case Bytecodes::_dstore_1: 1600 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1601 break; 1602 case Bytecodes::_dstore_2: 1603 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1604 break; 1605 case Bytecodes::_dstore_3: 1606 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1607 break; 1608 case Bytecodes::_dstore: 1609 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1610 break; 1611 1612 case Bytecodes::_pop: dec_sp(1); break; 1613 case Bytecodes::_pop2: dec_sp(2); break; 1614 case Bytecodes::_swap: 1615 a = pop(); 1616 b = pop(); 1617 push(a); 1618 push(b); 1619 break; 1620 case Bytecodes::_dup: 1621 a = pop(); 1622 push(a); 1623 push(a); 1624 break; 1625 case Bytecodes::_dup_x1: 1626 a = pop(); 1627 b = pop(); 1628 push( a ); 1629 push( b ); 1630 push( a ); 1631 break; 1632 case Bytecodes::_dup_x2: 1633 a = pop(); 1634 b = pop(); 1635 c = pop(); 1636 push( a ); 1637 push( c ); 1638 push( b ); 1639 push( a ); 1640 break; 1641 case Bytecodes::_dup2: 1642 a = pop(); 1643 b = pop(); 1644 push( b ); 1645 push( a ); 1646 push( b ); 1647 push( a ); 1648 break; 1649 1650 case Bytecodes::_dup2_x1: 1651 // before: .. c, b, a 1652 // after: .. b, a, c, b, a 1653 // not tested 1654 a = pop(); 1655 b = pop(); 1656 c = pop(); 1657 push( b ); 1658 push( a ); 1659 push( c ); 1660 push( b ); 1661 push( a ); 1662 break; 1663 case Bytecodes::_dup2_x2: 1664 // before: .. d, c, b, a 1665 // after: .. b, a, d, c, b, a 1666 // not tested 1667 a = pop(); 1668 b = pop(); 1669 c = pop(); 1670 d = pop(); 1671 push( b ); 1672 push( a ); 1673 push( d ); 1674 push( c ); 1675 push( b ); 1676 push( a ); 1677 break; 1678 1679 case Bytecodes::_arraylength: { 1680 // Must do null-check with value on expression stack 1681 Node *ary = null_check(peek(), T_ARRAY); 1682 // Compile-time detect of null-exception? 1683 if (stopped()) return; 1684 a = pop(); 1685 push(load_array_length(a)); 1686 break; 1687 } 1688 1689 case Bytecodes::_baload: array_load(T_BYTE); break; 1690 case Bytecodes::_caload: array_load(T_CHAR); break; 1691 case Bytecodes::_iaload: array_load(T_INT); break; 1692 case Bytecodes::_saload: array_load(T_SHORT); break; 1693 case Bytecodes::_faload: array_load(T_FLOAT); break; 1694 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1695 case Bytecodes::_laload: { 1696 a = array_addressing(T_LONG, 0); 1697 if (stopped()) return; // guaranteed null or range check 1698 dec_sp(2); // Pop array and index 1699 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1700 break; 1701 } 1702 case Bytecodes::_daload: { 1703 a = array_addressing(T_DOUBLE, 0); 1704 if (stopped()) return; // guaranteed null or range check 1705 dec_sp(2); // Pop array and index 1706 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1707 break; 1708 } 1709 case Bytecodes::_bastore: array_store(T_BYTE); break; 1710 case Bytecodes::_castore: array_store(T_CHAR); break; 1711 case Bytecodes::_iastore: array_store(T_INT); break; 1712 case Bytecodes::_sastore: array_store(T_SHORT); break; 1713 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1714 case Bytecodes::_aastore: { 1715 d = array_addressing(T_OBJECT, 1); 1716 if (stopped()) return; // guaranteed null or range check 1717 array_store_check(); 1718 c = pop(); // Oop to store 1719 b = pop(); // index (already used) 1720 a = pop(); // the array itself 1721 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1722 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1723 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release); 1724 break; 1725 } 1726 case Bytecodes::_lastore: { 1727 a = array_addressing(T_LONG, 2); 1728 if (stopped()) return; // guaranteed null or range check 1729 c = pop_pair(); 1730 dec_sp(2); // Pop array and index 1731 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1732 break; 1733 } 1734 case Bytecodes::_dastore: { 1735 a = array_addressing(T_DOUBLE, 2); 1736 if (stopped()) return; // guaranteed null or range check 1737 c = pop_pair(); 1738 dec_sp(2); // Pop array and index 1739 c = dstore_rounding(c); 1740 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1741 break; 1742 } 1743 case Bytecodes::_getfield: 1744 do_getfield(); 1745 break; 1746 1747 case Bytecodes::_getstatic: 1748 do_getstatic(); 1749 break; 1750 1751 case Bytecodes::_putfield: 1752 do_putfield(); 1753 break; 1754 1755 case Bytecodes::_putstatic: 1756 do_putstatic(); 1757 break; 1758 1759 case Bytecodes::_irem: 1760 do_irem(); 1761 break; 1762 case Bytecodes::_idiv: 1763 // Must keep both values on the expression-stack during null-check 1764 zero_check_int(peek()); 1765 // Compile-time detect of null-exception? 1766 if (stopped()) return; 1767 b = pop(); 1768 a = pop(); 1769 push( _gvn.transform( new DivINode(control(),a,b) ) ); 1770 break; 1771 case Bytecodes::_imul: 1772 b = pop(); a = pop(); 1773 push( _gvn.transform( new MulINode(a,b) ) ); 1774 break; 1775 case Bytecodes::_iadd: 1776 b = pop(); a = pop(); 1777 push( _gvn.transform( new AddINode(a,b) ) ); 1778 break; 1779 case Bytecodes::_ineg: 1780 a = pop(); 1781 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 1782 break; 1783 case Bytecodes::_isub: 1784 b = pop(); a = pop(); 1785 push( _gvn.transform( new SubINode(a,b) ) ); 1786 break; 1787 case Bytecodes::_iand: 1788 b = pop(); a = pop(); 1789 push( _gvn.transform( new AndINode(a,b) ) ); 1790 break; 1791 case Bytecodes::_ior: 1792 b = pop(); a = pop(); 1793 push( _gvn.transform( new OrINode(a,b) ) ); 1794 break; 1795 case Bytecodes::_ixor: 1796 b = pop(); a = pop(); 1797 push( _gvn.transform( new XorINode(a,b) ) ); 1798 break; 1799 case Bytecodes::_ishl: 1800 b = pop(); a = pop(); 1801 push( _gvn.transform( new LShiftINode(a,b) ) ); 1802 break; 1803 case Bytecodes::_ishr: 1804 b = pop(); a = pop(); 1805 push( _gvn.transform( new RShiftINode(a,b) ) ); 1806 break; 1807 case Bytecodes::_iushr: 1808 b = pop(); a = pop(); 1809 push( _gvn.transform( new URShiftINode(a,b) ) ); 1810 break; 1811 1812 case Bytecodes::_fneg: 1813 a = pop(); 1814 b = _gvn.transform(new NegFNode (a)); 1815 push(b); 1816 break; 1817 1818 case Bytecodes::_fsub: 1819 b = pop(); 1820 a = pop(); 1821 c = _gvn.transform( new SubFNode(a,b) ); 1822 d = precision_rounding(c); 1823 push( d ); 1824 break; 1825 1826 case Bytecodes::_fadd: 1827 b = pop(); 1828 a = pop(); 1829 c = _gvn.transform( new AddFNode(a,b) ); 1830 d = precision_rounding(c); 1831 push( d ); 1832 break; 1833 1834 case Bytecodes::_fmul: 1835 b = pop(); 1836 a = pop(); 1837 c = _gvn.transform( new MulFNode(a,b) ); 1838 d = precision_rounding(c); 1839 push( d ); 1840 break; 1841 1842 case Bytecodes::_fdiv: 1843 b = pop(); 1844 a = pop(); 1845 c = _gvn.transform( new DivFNode(0,a,b) ); 1846 d = precision_rounding(c); 1847 push( d ); 1848 break; 1849 1850 case Bytecodes::_frem: 1851 if (Matcher::has_match_rule(Op_ModF)) { 1852 // Generate a ModF node. 1853 b = pop(); 1854 a = pop(); 1855 c = _gvn.transform( new ModFNode(0,a,b) ); 1856 d = precision_rounding(c); 1857 push( d ); 1858 } 1859 else { 1860 // Generate a call. 1861 modf(); 1862 } 1863 break; 1864 1865 case Bytecodes::_fcmpl: 1866 b = pop(); 1867 a = pop(); 1868 c = _gvn.transform( new CmpF3Node( a, b)); 1869 push(c); 1870 break; 1871 case Bytecodes::_fcmpg: 1872 b = pop(); 1873 a = pop(); 1874 1875 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1876 // which negates the result sign except for unordered. Flip the unordered 1877 // as well by using CmpF3 which implements unordered-lesser instead of 1878 // unordered-greater semantics. Finally, commute the result bits. Result 1879 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1880 c = _gvn.transform( new CmpF3Node( b, a)); 1881 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 1882 push(c); 1883 break; 1884 1885 case Bytecodes::_f2i: 1886 a = pop(); 1887 push(_gvn.transform(new ConvF2INode(a))); 1888 break; 1889 1890 case Bytecodes::_d2i: 1891 a = pop_pair(); 1892 b = _gvn.transform(new ConvD2INode(a)); 1893 push( b ); 1894 break; 1895 1896 case Bytecodes::_f2d: 1897 a = pop(); 1898 b = _gvn.transform( new ConvF2DNode(a)); 1899 push_pair( b ); 1900 break; 1901 1902 case Bytecodes::_d2f: 1903 a = pop_pair(); 1904 b = _gvn.transform( new ConvD2FNode(a)); 1905 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1906 //b = _gvn.transform(new RoundFloatNode(0, b) ); 1907 push( b ); 1908 break; 1909 1910 case Bytecodes::_l2f: 1911 if (Matcher::convL2FSupported()) { 1912 a = pop_pair(); 1913 b = _gvn.transform( new ConvL2FNode(a)); 1914 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1915 // Rather than storing the result into an FP register then pushing 1916 // out to memory to round, the machine instruction that implements 1917 // ConvL2D is responsible for rounding. 1918 // c = precision_rounding(b); 1919 c = _gvn.transform(b); 1920 push(c); 1921 } else { 1922 l2f(); 1923 } 1924 break; 1925 1926 case Bytecodes::_l2d: 1927 a = pop_pair(); 1928 b = _gvn.transform( new ConvL2DNode(a)); 1929 // For i486.ad, rounding is always necessary (see _l2f above). 1930 // c = dprecision_rounding(b); 1931 c = _gvn.transform(b); 1932 push_pair(c); 1933 break; 1934 1935 case Bytecodes::_f2l: 1936 a = pop(); 1937 b = _gvn.transform( new ConvF2LNode(a)); 1938 push_pair(b); 1939 break; 1940 1941 case Bytecodes::_d2l: 1942 a = pop_pair(); 1943 b = _gvn.transform( new ConvD2LNode(a)); 1944 push_pair(b); 1945 break; 1946 1947 case Bytecodes::_dsub: 1948 b = pop_pair(); 1949 a = pop_pair(); 1950 c = _gvn.transform( new SubDNode(a,b) ); 1951 d = dprecision_rounding(c); 1952 push_pair( d ); 1953 break; 1954 1955 case Bytecodes::_dadd: 1956 b = pop_pair(); 1957 a = pop_pair(); 1958 c = _gvn.transform( new AddDNode(a,b) ); 1959 d = dprecision_rounding(c); 1960 push_pair( d ); 1961 break; 1962 1963 case Bytecodes::_dmul: 1964 b = pop_pair(); 1965 a = pop_pair(); 1966 c = _gvn.transform( new MulDNode(a,b) ); 1967 d = dprecision_rounding(c); 1968 push_pair( d ); 1969 break; 1970 1971 case Bytecodes::_ddiv: 1972 b = pop_pair(); 1973 a = pop_pair(); 1974 c = _gvn.transform( new DivDNode(0,a,b) ); 1975 d = dprecision_rounding(c); 1976 push_pair( d ); 1977 break; 1978 1979 case Bytecodes::_dneg: 1980 a = pop_pair(); 1981 b = _gvn.transform(new NegDNode (a)); 1982 push_pair(b); 1983 break; 1984 1985 case Bytecodes::_drem: 1986 if (Matcher::has_match_rule(Op_ModD)) { 1987 // Generate a ModD node. 1988 b = pop_pair(); 1989 a = pop_pair(); 1990 // a % b 1991 1992 c = _gvn.transform( new ModDNode(0,a,b) ); 1993 d = dprecision_rounding(c); 1994 push_pair( d ); 1995 } 1996 else { 1997 // Generate a call. 1998 modd(); 1999 } 2000 break; 2001 2002 case Bytecodes::_dcmpl: 2003 b = pop_pair(); 2004 a = pop_pair(); 2005 c = _gvn.transform( new CmpD3Node( a, b)); 2006 push(c); 2007 break; 2008 2009 case Bytecodes::_dcmpg: 2010 b = pop_pair(); 2011 a = pop_pair(); 2012 // Same as dcmpl but need to flip the unordered case. 2013 // Commute the inputs, which negates the result sign except for unordered. 2014 // Flip the unordered as well by using CmpD3 which implements 2015 // unordered-lesser instead of unordered-greater semantics. 2016 // Finally, negate the result bits. Result is same as using a 2017 // CmpD3Greater except we did it with CmpD3 alone. 2018 c = _gvn.transform( new CmpD3Node( b, a)); 2019 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2020 push(c); 2021 break; 2022 2023 2024 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2025 case Bytecodes::_land: 2026 b = pop_pair(); 2027 a = pop_pair(); 2028 c = _gvn.transform( new AndLNode(a,b) ); 2029 push_pair(c); 2030 break; 2031 case Bytecodes::_lor: 2032 b = pop_pair(); 2033 a = pop_pair(); 2034 c = _gvn.transform( new OrLNode(a,b) ); 2035 push_pair(c); 2036 break; 2037 case Bytecodes::_lxor: 2038 b = pop_pair(); 2039 a = pop_pair(); 2040 c = _gvn.transform( new XorLNode(a,b) ); 2041 push_pair(c); 2042 break; 2043 2044 case Bytecodes::_lshl: 2045 b = pop(); // the shift count 2046 a = pop_pair(); // value to be shifted 2047 c = _gvn.transform( new LShiftLNode(a,b) ); 2048 push_pair(c); 2049 break; 2050 case Bytecodes::_lshr: 2051 b = pop(); // the shift count 2052 a = pop_pair(); // value to be shifted 2053 c = _gvn.transform( new RShiftLNode(a,b) ); 2054 push_pair(c); 2055 break; 2056 case Bytecodes::_lushr: 2057 b = pop(); // the shift count 2058 a = pop_pair(); // value to be shifted 2059 c = _gvn.transform( new URShiftLNode(a,b) ); 2060 push_pair(c); 2061 break; 2062 case Bytecodes::_lmul: 2063 b = pop_pair(); 2064 a = pop_pair(); 2065 c = _gvn.transform( new MulLNode(a,b) ); 2066 push_pair(c); 2067 break; 2068 2069 case Bytecodes::_lrem: 2070 // Must keep both values on the expression-stack during null-check 2071 assert(peek(0) == top(), "long word order"); 2072 zero_check_long(peek(1)); 2073 // Compile-time detect of null-exception? 2074 if (stopped()) return; 2075 b = pop_pair(); 2076 a = pop_pair(); 2077 c = _gvn.transform( new ModLNode(control(),a,b) ); 2078 push_pair(c); 2079 break; 2080 2081 case Bytecodes::_ldiv: 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 DivLNode(control(),a,b) ); 2090 push_pair(c); 2091 break; 2092 2093 case Bytecodes::_ladd: 2094 b = pop_pair(); 2095 a = pop_pair(); 2096 c = _gvn.transform( new AddLNode(a,b) ); 2097 push_pair(c); 2098 break; 2099 case Bytecodes::_lsub: 2100 b = pop_pair(); 2101 a = pop_pair(); 2102 c = _gvn.transform( new SubLNode(a,b) ); 2103 push_pair(c); 2104 break; 2105 case Bytecodes::_lcmp: 2106 // Safepoints are now inserted _before_ branches. The long-compare 2107 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2108 // slew of control flow. These are usually followed by a CmpI vs zero and 2109 // a branch; this pattern then optimizes to the obvious long-compare and 2110 // branch. However, if the branch is backwards there's a Safepoint 2111 // inserted. The inserted Safepoint captures the JVM state at the 2112 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2113 // long-compare is used to control a loop the debug info will force 2114 // computation of the 3-way value, even though the generated code uses a 2115 // long-compare and branch. We try to rectify the situation by inserting 2116 // a SafePoint here and have it dominate and kill the safepoint added at a 2117 // following backwards branch. At this point the JVM state merely holds 2 2118 // longs but not the 3-way value. 2119 if( UseLoopSafepoints ) { 2120 switch( iter().next_bc() ) { 2121 case Bytecodes::_ifgt: 2122 case Bytecodes::_iflt: 2123 case Bytecodes::_ifge: 2124 case Bytecodes::_ifle: 2125 case Bytecodes::_ifne: 2126 case Bytecodes::_ifeq: 2127 // If this is a backwards branch in the bytecodes, add Safepoint 2128 maybe_add_safepoint(iter().next_get_dest()); 2129 } 2130 } 2131 b = pop_pair(); 2132 a = pop_pair(); 2133 c = _gvn.transform( new CmpL3Node( a, b )); 2134 push(c); 2135 break; 2136 2137 case Bytecodes::_lneg: 2138 a = pop_pair(); 2139 b = _gvn.transform( new SubLNode(longcon(0),a)); 2140 push_pair(b); 2141 break; 2142 case Bytecodes::_l2i: 2143 a = pop_pair(); 2144 push( _gvn.transform( new ConvL2INode(a))); 2145 break; 2146 case Bytecodes::_i2l: 2147 a = pop(); 2148 b = _gvn.transform( new ConvI2LNode(a)); 2149 push_pair(b); 2150 break; 2151 case Bytecodes::_i2b: 2152 // Sign extend 2153 a = pop(); 2154 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2155 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2156 push( a ); 2157 break; 2158 case Bytecodes::_i2s: 2159 a = pop(); 2160 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2161 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2162 push( a ); 2163 break; 2164 case Bytecodes::_i2c: 2165 a = pop(); 2166 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2167 break; 2168 2169 case Bytecodes::_i2f: 2170 a = pop(); 2171 b = _gvn.transform( new ConvI2FNode(a) ) ; 2172 c = precision_rounding(b); 2173 push (b); 2174 break; 2175 2176 case Bytecodes::_i2d: 2177 a = pop(); 2178 b = _gvn.transform( new ConvI2DNode(a)); 2179 push_pair(b); 2180 break; 2181 2182 case Bytecodes::_iinc: // Increment local 2183 i = iter().get_index(); // Get local index 2184 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2185 break; 2186 2187 // Exit points of synchronized methods must have an unlock node 2188 case Bytecodes::_return: 2189 return_current(NULL); 2190 break; 2191 2192 case Bytecodes::_ireturn: 2193 case Bytecodes::_areturn: 2194 case Bytecodes::_freturn: 2195 return_current(pop()); 2196 break; 2197 case Bytecodes::_lreturn: 2198 return_current(pop_pair()); 2199 break; 2200 case Bytecodes::_dreturn: 2201 return_current(pop_pair()); 2202 break; 2203 2204 case Bytecodes::_athrow: 2205 // null exception oop throws NULL pointer exception 2206 null_check(peek()); 2207 if (stopped()) return; 2208 // Hook the thrown exception directly to subsequent handlers. 2209 if (BailoutToInterpreterForThrows) { 2210 // Keep method interpreted from now on. 2211 uncommon_trap(Deoptimization::Reason_unhandled, 2212 Deoptimization::Action_make_not_compilable); 2213 return; 2214 } 2215 if (env()->jvmti_can_post_on_exceptions()) { 2216 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2217 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2218 } 2219 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2220 add_exception_state(make_exception_state(peek())); 2221 break; 2222 2223 case Bytecodes::_goto: // fall through 2224 case Bytecodes::_goto_w: { 2225 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2226 2227 // If this is a backwards branch in the bytecodes, add Safepoint 2228 maybe_add_safepoint(target_bci); 2229 2230 // Update method data 2231 profile_taken_branch(target_bci); 2232 2233 // Merge the current control into the target basic block 2234 merge(target_bci); 2235 2236 // See if we can get some profile data and hand it off to the next block 2237 Block *target_block = block()->successor_for_bci(target_bci); 2238 if (target_block->pred_count() != 1) break; 2239 ciMethodData* methodData = method()->method_data(); 2240 if (!methodData->is_mature()) break; 2241 ciProfileData* data = methodData->bci_to_data(bci()); 2242 assert( data->is_JumpData(), "" ); 2243 int taken = ((ciJumpData*)data)->taken(); 2244 taken = method()->scale_count(taken); 2245 target_block->set_count(taken); 2246 break; 2247 } 2248 2249 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2250 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2251 handle_if_null: 2252 // If this is a backwards branch in the bytecodes, add Safepoint 2253 maybe_add_safepoint(iter().get_dest()); 2254 a = null(); 2255 b = pop(); 2256 if (!_gvn.type(b)->speculative_maybe_null() && 2257 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2258 inc_sp(1); 2259 Node* null_ctl = top(); 2260 b = null_check_oop(b, &null_ctl, true, true, true); 2261 assert(null_ctl->is_top(), "no null control here"); 2262 dec_sp(1); 2263 } 2264 c = _gvn.transform( new CmpPNode(b, a) ); 2265 do_ifnull(btest, c); 2266 break; 2267 2268 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2269 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2270 handle_if_acmp: 2271 // If this is a backwards branch in the bytecodes, add Safepoint 2272 maybe_add_safepoint(iter().get_dest()); 2273 a = pop(); 2274 b = pop(); 2275 c = _gvn.transform( new CmpPNode(b, a) ); 2276 c = optimize_cmp_with_klass(c); 2277 do_if(btest, c); 2278 break; 2279 2280 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2281 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2282 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2283 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2284 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2285 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2286 handle_ifxx: 2287 // If this is a backwards branch in the bytecodes, add Safepoint 2288 maybe_add_safepoint(iter().get_dest()); 2289 a = _gvn.intcon(0); 2290 b = pop(); 2291 c = _gvn.transform( new CmpINode(b, a) ); 2292 do_if(btest, c); 2293 break; 2294 2295 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2296 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2297 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2298 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2299 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2300 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2301 handle_if_icmp: 2302 // If this is a backwards branch in the bytecodes, add Safepoint 2303 maybe_add_safepoint(iter().get_dest()); 2304 a = pop(); 2305 b = pop(); 2306 c = _gvn.transform( new CmpINode( b, a ) ); 2307 do_if(btest, c); 2308 break; 2309 2310 case Bytecodes::_tableswitch: 2311 do_tableswitch(); 2312 break; 2313 2314 case Bytecodes::_lookupswitch: 2315 do_lookupswitch(); 2316 break; 2317 2318 case Bytecodes::_invokestatic: 2319 case Bytecodes::_invokedynamic: 2320 case Bytecodes::_invokespecial: 2321 case Bytecodes::_invokevirtual: 2322 case Bytecodes::_invokeinterface: 2323 do_call(); 2324 break; 2325 case Bytecodes::_checkcast: 2326 do_checkcast(); 2327 break; 2328 case Bytecodes::_instanceof: 2329 do_instanceof(); 2330 break; 2331 case Bytecodes::_anewarray: 2332 do_anewarray(); 2333 break; 2334 case Bytecodes::_newarray: 2335 do_newarray((BasicType)iter().get_index()); 2336 break; 2337 case Bytecodes::_multianewarray: 2338 do_multianewarray(); 2339 break; 2340 case Bytecodes::_new: 2341 do_new(); 2342 break; 2343 2344 case Bytecodes::_jsr: 2345 case Bytecodes::_jsr_w: 2346 do_jsr(); 2347 break; 2348 2349 case Bytecodes::_ret: 2350 do_ret(); 2351 break; 2352 2353 2354 case Bytecodes::_monitorenter: 2355 do_monitor_enter(); 2356 break; 2357 2358 case Bytecodes::_monitorexit: 2359 do_monitor_exit(); 2360 break; 2361 2362 case Bytecodes::_breakpoint: 2363 // Breakpoint set concurrently to compile 2364 // %%% use an uncommon trap? 2365 C->record_failure("breakpoint in method"); 2366 return; 2367 2368 default: 2369 #ifndef PRODUCT 2370 map()->dump(99); 2371 #endif 2372 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2373 ShouldNotReachHere(); 2374 } 2375 2376 #ifndef PRODUCT 2377 IdealGraphPrinter *printer = IdealGraphPrinter::printer(); 2378 if (printer && printer->should_print(_method)) { 2379 char buffer[256]; 2380 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2381 bool old = printer->traverse_outs(); 2382 printer->set_traverse_outs(true); 2383 printer->print_method(C, buffer, 4); 2384 printer->set_traverse_outs(old); 2385 } 2386 #endif 2387 }