1 /* 2 * Copyright (c) 1998, 2013, 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/parse.hpp" 41 #include "opto/runtime.hpp" 42 #include "runtime/deoptimization.hpp" 43 #include "runtime/sharedRuntime.hpp" 44 45 extern int explicit_null_checks_inserted, 46 explicit_null_checks_elided; 47 48 //---------------------------------array_load---------------------------------- 49 void Parse::array_load(BasicType elem_type) { 50 const Type* elem = Type::TOP; 51 Node* adr = array_addressing(elem_type, 0, &elem); 52 if (stopped()) return; // guaranteed null or range check 53 dec_sp(2); // Pop array and index 54 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 55 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 56 push(ld); 57 } 58 59 60 //--------------------------------array_store---------------------------------- 61 void Parse::array_store(BasicType elem_type) { 62 Node* adr = array_addressing(elem_type, 1); 63 if (stopped()) return; // guaranteed null or range check 64 Node* val = pop(); 65 dec_sp(2); // Pop array and index 66 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 67 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 68 } 69 70 71 //------------------------------array_addressing------------------------------- 72 // Pull array and index from the stack. Compute pointer-to-element. 73 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 74 Node *idx = peek(0+vals); // Get from stack without popping 75 Node *ary = peek(1+vals); // in case of exception 76 77 // Null check the array base, with correct stack contents 78 ary = null_check(ary, T_ARRAY); 79 // Compile-time detect of null-exception? 80 if (stopped()) return top(); 81 82 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 83 const TypeInt* sizetype = arytype->size(); 84 const Type* elemtype = arytype->elem(); 85 86 if (UseUniqueSubclasses && result2 != NULL) { 87 const Type* el = elemtype->make_ptr(); 88 if (el && el->isa_instptr()) { 89 const TypeInstPtr* toop = el->is_instptr(); 90 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 91 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 92 const Type* subklass = Type::get_const_type(toop->klass()); 93 elemtype = subklass->join_speculative(el); 94 } 95 } 96 } 97 98 // Check for big class initializers with all constant offsets 99 // feeding into a known-size array. 100 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 101 // See if the highest idx value is less than the lowest array bound, 102 // and if the idx value cannot be negative: 103 bool need_range_check = true; 104 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 105 need_range_check = false; 106 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 107 } 108 109 ciKlass * arytype_klass = arytype->klass(); 110 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 111 // Only fails for some -Xcomp runs 112 // The class is unloaded. We have to run this bytecode in the interpreter. 113 uncommon_trap(Deoptimization::Reason_unloaded, 114 Deoptimization::Action_reinterpret, 115 arytype->klass(), "!loaded array"); 116 return top(); 117 } 118 119 // Do the range check 120 if (GenerateRangeChecks && need_range_check) { 121 Node* tst; 122 if (sizetype->_hi <= 0) { 123 // The greatest array bound is negative, so we can conclude that we're 124 // compiling unreachable code, but the unsigned compare trick used below 125 // only works with non-negative lengths. Instead, hack "tst" to be zero so 126 // the uncommon_trap path will always be taken. 127 tst = _gvn.intcon(0); 128 } else { 129 // Range is constant in array-oop, so we can use the original state of mem 130 Node* len = load_array_length(ary); 131 132 // Test length vs index (standard trick using unsigned compare) 133 Node* chk = _gvn.transform( new (C) CmpUNode(idx, len) ); 134 BoolTest::mask btest = BoolTest::lt; 135 tst = _gvn.transform( new (C) BoolNode(chk, btest) ); 136 } 137 // Branch to failure if out of bounds 138 { BuildCutout unless(this, tst, PROB_MAX); 139 if (C->allow_range_check_smearing()) { 140 // Do not use builtin_throw, since range checks are sometimes 141 // made more stringent by an optimistic transformation. 142 // This creates "tentative" range checks at this point, 143 // which are not guaranteed to throw exceptions. 144 // See IfNode::Ideal, is_range_check, adjust_check. 145 uncommon_trap(Deoptimization::Reason_range_check, 146 Deoptimization::Action_make_not_entrant, 147 NULL, "range_check"); 148 } else { 149 // If we have already recompiled with the range-check-widening 150 // heroic optimization turned off, then we must really be throwing 151 // range check exceptions. 152 builtin_throw(Deoptimization::Reason_range_check, idx); 153 } 154 } 155 } 156 // Check for always knowing you are throwing a range-check exception 157 if (stopped()) return top(); 158 159 Node* ptr = array_element_address(ary, idx, type, sizetype); 160 161 if (result2 != NULL) *result2 = elemtype; 162 163 assert(ptr != top(), "top should go hand-in-hand with stopped"); 164 165 return ptr; 166 } 167 168 169 // returns IfNode 170 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 171 Node *cmp = _gvn.transform( new (C) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 172 Node *tst = _gvn.transform( new (C) BoolNode( cmp, mask)); 173 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 174 return iff; 175 } 176 177 // return Region node 178 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 179 Node *region = new (C) RegionNode(3); // 2 results 180 record_for_igvn(region); 181 region->init_req(1, iffalse); 182 region->init_req(2, iftrue ); 183 _gvn.set_type(region, Type::CONTROL); 184 region = _gvn.transform(region); 185 set_control (region); 186 return region; 187 } 188 189 190 //------------------------------helper for tableswitch------------------------- 191 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 192 // True branch, use existing map info 193 { PreserveJVMState pjvms(this); 194 Node *iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 195 set_control( iftrue ); 196 profile_switch_case(prof_table_index); 197 merge_new_path(dest_bci_if_true); 198 } 199 200 // False branch 201 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 202 set_control( iffalse ); 203 } 204 205 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 206 // True branch, use existing map info 207 { PreserveJVMState pjvms(this); 208 Node *iffalse = _gvn.transform( new (C) IfFalseNode (iff) ); 209 set_control( iffalse ); 210 profile_switch_case(prof_table_index); 211 merge_new_path(dest_bci_if_true); 212 } 213 214 // False branch 215 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff) ); 216 set_control( iftrue ); 217 } 218 219 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 220 // False branch, use existing map and control() 221 profile_switch_case(prof_table_index); 222 merge_new_path(dest_bci); 223 } 224 225 226 extern "C" { 227 static int jint_cmp(const void *i, const void *j) { 228 int a = *(jint *)i; 229 int b = *(jint *)j; 230 return a > b ? 1 : a < b ? -1 : 0; 231 } 232 } 233 234 235 // Default value for methodData switch indexing. Must be a negative value to avoid 236 // conflict with any legal switch index. 237 #define NullTableIndex -1 238 239 class SwitchRange : public StackObj { 240 // a range of integers coupled with a bci destination 241 jint _lo; // inclusive lower limit 242 jint _hi; // inclusive upper limit 243 int _dest; 244 int _table_index; // index into method data table 245 246 public: 247 jint lo() const { return _lo; } 248 jint hi() const { return _hi; } 249 int dest() const { return _dest; } 250 int table_index() const { return _table_index; } 251 bool is_singleton() const { return _lo == _hi; } 252 253 void setRange(jint lo, jint hi, int dest, int table_index) { 254 assert(lo <= hi, "must be a non-empty range"); 255 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 256 } 257 bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 258 assert(lo <= hi, "must be a non-empty range"); 259 if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 260 _hi = hi; 261 return true; 262 } 263 return false; 264 } 265 266 void set (jint value, int dest, int table_index) { 267 setRange(value, value, dest, table_index); 268 } 269 bool adjoin(jint value, int dest, int table_index) { 270 return adjoinRange(value, value, dest, table_index); 271 } 272 273 void print() { 274 if (is_singleton()) 275 tty->print(" {%d}=>%d", lo(), dest()); 276 else if (lo() == min_jint) 277 tty->print(" {..%d}=>%d", hi(), dest()); 278 else if (hi() == max_jint) 279 tty->print(" {%d..}=>%d", lo(), dest()); 280 else 281 tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 282 } 283 }; 284 285 286 //-------------------------------do_tableswitch-------------------------------- 287 void Parse::do_tableswitch() { 288 Node* lookup = pop(); 289 290 // Get information about tableswitch 291 int default_dest = iter().get_dest_table(0); 292 int lo_index = iter().get_int_table(1); 293 int hi_index = iter().get_int_table(2); 294 int len = hi_index - lo_index + 1; 295 296 if (len < 1) { 297 // If this is a backward branch, add safepoint 298 maybe_add_safepoint(default_dest); 299 merge(default_dest); 300 return; 301 } 302 303 // generate decision tree, using trichotomy when possible 304 int rnum = len+2; 305 bool makes_backward_branch = false; 306 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 307 int rp = -1; 308 if (lo_index != min_jint) { 309 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 310 } 311 for (int j = 0; j < len; j++) { 312 jint match_int = lo_index+j; 313 int dest = iter().get_dest_table(j+3); 314 makes_backward_branch |= (dest <= bci()); 315 int table_index = method_data_update() ? j : NullTableIndex; 316 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 317 ranges[++rp].set(match_int, dest, table_index); 318 } 319 } 320 jint highest = lo_index+(len-1); 321 assert(ranges[rp].hi() == highest, ""); 322 if (highest != max_jint 323 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 324 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 325 } 326 assert(rp < len+2, "not too many ranges"); 327 328 // Safepoint in case if backward branch observed 329 if( makes_backward_branch && UseLoopSafepoints ) 330 add_safepoint(); 331 332 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 333 } 334 335 336 //------------------------------do_lookupswitch-------------------------------- 337 void Parse::do_lookupswitch() { 338 Node *lookup = pop(); // lookup value 339 // Get information about lookupswitch 340 int default_dest = iter().get_dest_table(0); 341 int len = iter().get_int_table(1); 342 343 if (len < 1) { // If this is a backward branch, add safepoint 344 maybe_add_safepoint(default_dest); 345 merge(default_dest); 346 return; 347 } 348 349 // generate decision tree, using trichotomy when possible 350 jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 351 { 352 for( int j = 0; j < len; j++ ) { 353 table[j+j+0] = iter().get_int_table(2+j+j); 354 table[j+j+1] = iter().get_dest_table(2+j+j+1); 355 } 356 qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 357 } 358 359 int rnum = len*2+1; 360 bool makes_backward_branch = false; 361 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 362 int rp = -1; 363 for( int j = 0; j < len; j++ ) { 364 jint match_int = table[j+j+0]; 365 int dest = table[j+j+1]; 366 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 367 int table_index = method_data_update() ? j : NullTableIndex; 368 makes_backward_branch |= (dest <= bci()); 369 if( match_int != next_lo ) { 370 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 371 } 372 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 373 ranges[++rp].set(match_int, dest, table_index); 374 } 375 } 376 jint highest = table[2*(len-1)]; 377 assert(ranges[rp].hi() == highest, ""); 378 if( highest != max_jint 379 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 380 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 381 } 382 assert(rp < rnum, "not too many ranges"); 383 384 // Safepoint in case backward branch observed 385 if( makes_backward_branch && UseLoopSafepoints ) 386 add_safepoint(); 387 388 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 389 } 390 391 //----------------------------create_jump_tables------------------------------- 392 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 393 // Are jumptables enabled 394 if (!UseJumpTables) return false; 395 396 // Are jumptables supported 397 if (!Matcher::has_match_rule(Op_Jump)) return false; 398 399 // Don't make jump table if profiling 400 if (method_data_update()) return false; 401 402 // Decide if a guard is needed to lop off big ranges at either (or 403 // both) end(s) of the input set. We'll call this the default target 404 // even though we can't be sure that it is the true "default". 405 406 bool needs_guard = false; 407 int default_dest; 408 int64 total_outlier_size = 0; 409 int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1; 410 int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1; 411 412 if (lo->dest() == hi->dest()) { 413 total_outlier_size = hi_size + lo_size; 414 default_dest = lo->dest(); 415 } else if (lo_size > hi_size) { 416 total_outlier_size = lo_size; 417 default_dest = lo->dest(); 418 } else { 419 total_outlier_size = hi_size; 420 default_dest = hi->dest(); 421 } 422 423 // If a guard test will eliminate very sparse end ranges, then 424 // it is worth the cost of an extra jump. 425 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 426 needs_guard = true; 427 if (default_dest == lo->dest()) lo++; 428 if (default_dest == hi->dest()) hi--; 429 } 430 431 // Find the total number of cases and ranges 432 int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1; 433 int num_range = hi - lo + 1; 434 435 // Don't create table if: too large, too small, or too sparse. 436 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 437 return false; 438 if (num_cases > (MaxJumpTableSparseness * num_range)) 439 return false; 440 441 // Normalize table lookups to zero 442 int lowval = lo->lo(); 443 key_val = _gvn.transform( new (C) SubINode(key_val, _gvn.intcon(lowval)) ); 444 445 // Generate a guard to protect against input keyvals that aren't 446 // in the switch domain. 447 if (needs_guard) { 448 Node* size = _gvn.intcon(num_cases); 449 Node* cmp = _gvn.transform( new (C) CmpUNode(key_val, size) ); 450 Node* tst = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ge) ); 451 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 452 jump_if_true_fork(iff, default_dest, NullTableIndex); 453 } 454 455 // Create an ideal node JumpTable that has projections 456 // of all possible ranges for a switch statement 457 // The key_val input must be converted to a pointer offset and scaled. 458 // Compare Parse::array_addressing above. 459 #ifdef _LP64 460 // Clean the 32-bit int into a real 64-bit offset. 461 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 462 const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin); 463 key_val = _gvn.transform( new (C) ConvI2LNode(key_val, lkeytype) ); 464 #endif 465 // Shift the value by wordsize so we have an index into the table, rather 466 // than a switch value 467 Node *shiftWord = _gvn.MakeConX(wordSize); 468 key_val = _gvn.transform( new (C) MulXNode( key_val, shiftWord)); 469 470 // Create the JumpNode 471 Node* jtn = _gvn.transform( new (C) JumpNode(control(), key_val, num_cases) ); 472 473 // These are the switch destinations hanging off the jumpnode 474 int i = 0; 475 for (SwitchRange* r = lo; r <= hi; r++) { 476 for (int64 j = r->lo(); j <= r->hi(); j++, i++) { 477 Node* input = _gvn.transform(new (C) JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 478 { 479 PreserveJVMState pjvms(this); 480 set_control(input); 481 jump_if_always_fork(r->dest(), r->table_index()); 482 } 483 } 484 } 485 assert(i == num_cases, "miscount of cases"); 486 stop_and_kill_map(); // no more uses for this JVMS 487 return true; 488 } 489 490 //----------------------------jump_switch_ranges------------------------------- 491 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 492 Block* switch_block = block(); 493 494 if (switch_depth == 0) { 495 // Do special processing for the top-level call. 496 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 497 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 498 499 // Decrement pred-numbers for the unique set of nodes. 500 #ifdef ASSERT 501 // Ensure that the block's successors are a (duplicate-free) set. 502 int successors_counted = 0; // block occurrences in [hi..lo] 503 int unique_successors = switch_block->num_successors(); 504 for (int i = 0; i < unique_successors; i++) { 505 Block* target = switch_block->successor_at(i); 506 507 // Check that the set of successors is the same in both places. 508 int successors_found = 0; 509 for (SwitchRange* p = lo; p <= hi; p++) { 510 if (p->dest() == target->start()) successors_found++; 511 } 512 assert(successors_found > 0, "successor must be known"); 513 successors_counted += successors_found; 514 } 515 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 516 #endif 517 518 // Maybe prune the inputs, based on the type of key_val. 519 jint min_val = min_jint; 520 jint max_val = max_jint; 521 const TypeInt* ti = key_val->bottom_type()->isa_int(); 522 if (ti != NULL) { 523 min_val = ti->_lo; 524 max_val = ti->_hi; 525 assert(min_val <= max_val, "invalid int type"); 526 } 527 while (lo->hi() < min_val) lo++; 528 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 529 while (hi->lo() > max_val) hi--; 530 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 531 } 532 533 #ifndef PRODUCT 534 if (switch_depth == 0) { 535 _max_switch_depth = 0; 536 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 537 } 538 #endif 539 540 assert(lo <= hi, "must be a non-empty set of ranges"); 541 if (lo == hi) { 542 jump_if_always_fork(lo->dest(), lo->table_index()); 543 } else { 544 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 545 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 546 547 if (create_jump_tables(key_val, lo, hi)) return; 548 549 int nr = hi - lo + 1; 550 551 SwitchRange* mid = lo + nr/2; 552 // if there is an easy choice, pivot at a singleton: 553 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 554 555 assert(lo < mid && mid <= hi, "good pivot choice"); 556 assert(nr != 2 || mid == hi, "should pick higher of 2"); 557 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 558 559 Node *test_val = _gvn.intcon(mid->lo()); 560 561 if (mid->is_singleton()) { 562 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 563 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 564 565 // Special Case: If there are exactly three ranges, and the high 566 // and low range each go to the same place, omit the "gt" test, 567 // since it will not discriminate anything. 568 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 569 if (eq_test_only) { 570 assert(mid == hi-1, ""); 571 } 572 573 // if there is a higher range, test for it and process it: 574 if (mid < hi && !eq_test_only) { 575 // two comparisons of same values--should enable 1 test for 2 branches 576 // Use BoolTest::le instead of BoolTest::gt 577 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 578 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_le) ); 579 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_le) ); 580 { PreserveJVMState pjvms(this); 581 set_control(iffalse); 582 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 583 } 584 set_control(iftrue); 585 } 586 587 } else { 588 // mid is a range, not a singleton, so treat mid..hi as a unit 589 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 590 591 // if there is a higher range, test for it and process it: 592 if (mid == hi) { 593 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 594 } else { 595 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_ge) ); 596 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_ge) ); 597 { PreserveJVMState pjvms(this); 598 set_control(iftrue); 599 jump_switch_ranges(key_val, mid, hi, switch_depth+1); 600 } 601 set_control(iffalse); 602 } 603 } 604 605 // in any case, process the lower range 606 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 607 } 608 609 // Decrease pred_count for each successor after all is done. 610 if (switch_depth == 0) { 611 int unique_successors = switch_block->num_successors(); 612 for (int i = 0; i < unique_successors; i++) { 613 Block* target = switch_block->successor_at(i); 614 // Throw away the pre-allocated path for each unique successor. 615 target->next_path_num(); 616 } 617 } 618 619 #ifndef PRODUCT 620 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 621 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 622 SwitchRange* r; 623 int nsing = 0; 624 for( r = lo; r <= hi; r++ ) { 625 if( r->is_singleton() ) nsing++; 626 } 627 tty->print(">>> "); 628 _method->print_short_name(); 629 tty->print_cr(" switch decision tree"); 630 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 631 hi-lo+1, nsing, _max_switch_depth, _est_switch_depth); 632 if (_max_switch_depth > _est_switch_depth) { 633 tty->print_cr("******** BAD SWITCH DEPTH ********"); 634 } 635 tty->print(" "); 636 for( r = lo; r <= hi; r++ ) { 637 r->print(); 638 } 639 tty->print_cr(""); 640 } 641 #endif 642 } 643 644 void Parse::modf() { 645 Node *f2 = pop(); 646 Node *f1 = pop(); 647 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 648 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 649 "frem", NULL, //no memory effects 650 f1, f2); 651 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 652 653 push(res); 654 } 655 656 void Parse::modd() { 657 Node *d2 = pop_pair(); 658 Node *d1 = pop_pair(); 659 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 660 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 661 "drem", NULL, //no memory effects 662 d1, top(), d2, top()); 663 Node* res_d = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 664 665 #ifdef ASSERT 666 Node* res_top = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 1)); 667 assert(res_top == top(), "second value must be top"); 668 #endif 669 670 push_pair(res_d); 671 } 672 673 void Parse::l2f() { 674 Node* f2 = pop(); 675 Node* f1 = pop(); 676 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 677 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 678 "l2f", NULL, //no memory effects 679 f1, f2); 680 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 681 682 push(res); 683 } 684 685 void Parse::do_irem() { 686 // Must keep both values on the expression-stack during null-check 687 zero_check_int(peek()); 688 // Compile-time detect of null-exception? 689 if (stopped()) return; 690 691 Node* b = pop(); 692 Node* a = pop(); 693 694 const Type *t = _gvn.type(b); 695 if (t != Type::TOP) { 696 const TypeInt *ti = t->is_int(); 697 if (ti->is_con()) { 698 int divisor = ti->get_con(); 699 // check for positive power of 2 700 if (divisor > 0 && 701 (divisor & ~(divisor-1)) == divisor) { 702 // yes ! 703 Node *mask = _gvn.intcon((divisor - 1)); 704 // Sigh, must handle negative dividends 705 Node *zero = _gvn.intcon(0); 706 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 707 Node *iff = _gvn.transform( new (C) IfFalseNode(ifff) ); 708 Node *ift = _gvn.transform( new (C) IfTrueNode (ifff) ); 709 Node *reg = jump_if_join(ift, iff); 710 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 711 // Negative path; negate/and/negate 712 Node *neg = _gvn.transform( new (C) SubINode(zero, a) ); 713 Node *andn= _gvn.transform( new (C) AndINode(neg, mask) ); 714 Node *negn= _gvn.transform( new (C) SubINode(zero, andn) ); 715 phi->init_req(1, negn); 716 // Fast positive case 717 Node *andx = _gvn.transform( new (C) AndINode(a, mask) ); 718 phi->init_req(2, andx); 719 // Push the merge 720 push( _gvn.transform(phi) ); 721 return; 722 } 723 } 724 } 725 // Default case 726 push( _gvn.transform( new (C) ModINode(control(),a,b) ) ); 727 } 728 729 // Handle jsr and jsr_w bytecode 730 void Parse::do_jsr() { 731 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 732 733 // Store information about current state, tagged with new _jsr_bci 734 int return_bci = iter().next_bci(); 735 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 736 737 // Update method data 738 profile_taken_branch(jsr_bci); 739 740 // The way we do things now, there is only one successor block 741 // for the jsr, because the target code is cloned by ciTypeFlow. 742 Block* target = successor_for_bci(jsr_bci); 743 744 // What got pushed? 745 const Type* ret_addr = target->peek(); 746 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 747 748 // Effect on jsr on stack 749 push(_gvn.makecon(ret_addr)); 750 751 // Flow to the jsr. 752 merge(jsr_bci); 753 } 754 755 // Handle ret bytecode 756 void Parse::do_ret() { 757 // Find to whom we return. 758 assert(block()->num_successors() == 1, "a ret can only go one place now"); 759 Block* target = block()->successor_at(0); 760 assert(!target->is_ready(), "our arrival must be expected"); 761 profile_ret(target->flow()->start()); 762 int pnum = target->next_path_num(); 763 merge_common(target, pnum); 764 } 765 766 //--------------------------dynamic_branch_prediction-------------------------- 767 // Try to gather dynamic branch prediction behavior. Return a probability 768 // of the branch being taken and set the "cnt" field. Returns a -1.0 769 // if we need to use static prediction for some reason. 770 float Parse::dynamic_branch_prediction(float &cnt) { 771 ResourceMark rm; 772 773 cnt = COUNT_UNKNOWN; 774 775 // Use MethodData information if it is available 776 // FIXME: free the ProfileData structure 777 ciMethodData* methodData = method()->method_data(); 778 if (!methodData->is_mature()) return PROB_UNKNOWN; 779 ciProfileData* data = methodData->bci_to_data(bci()); 780 if (!data->is_JumpData()) return PROB_UNKNOWN; 781 782 // get taken and not taken values 783 int taken = data->as_JumpData()->taken(); 784 int not_taken = 0; 785 if (data->is_BranchData()) { 786 not_taken = data->as_BranchData()->not_taken(); 787 } 788 789 // scale the counts to be commensurate with invocation counts: 790 taken = method()->scale_count(taken); 791 not_taken = method()->scale_count(not_taken); 792 793 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 794 // We also check that individual counters are positive first, overwise the sum can become positive. 795 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 796 if (C->log() != NULL) { 797 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 798 } 799 return PROB_UNKNOWN; 800 } 801 802 // Compute frequency that we arrive here 803 float sum = taken + not_taken; 804 // Adjust, if this block is a cloned private block but the 805 // Jump counts are shared. Taken the private counts for 806 // just this path instead of the shared counts. 807 if( block()->count() > 0 ) 808 sum = block()->count(); 809 cnt = sum / FreqCountInvocations; 810 811 // Pin probability to sane limits 812 float prob; 813 if( !taken ) 814 prob = (0+PROB_MIN) / 2; 815 else if( !not_taken ) 816 prob = (1+PROB_MAX) / 2; 817 else { // Compute probability of true path 818 prob = (float)taken / (float)(taken + not_taken); 819 if (prob > PROB_MAX) prob = PROB_MAX; 820 if (prob < PROB_MIN) prob = PROB_MIN; 821 } 822 823 assert((cnt > 0.0f) && (prob > 0.0f), 824 "Bad frequency assignment in if"); 825 826 if (C->log() != NULL) { 827 const char* prob_str = NULL; 828 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 829 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 830 char prob_str_buf[30]; 831 if (prob_str == NULL) { 832 sprintf(prob_str_buf, "%g", prob); 833 prob_str = prob_str_buf; 834 } 835 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'", 836 iter().get_dest(), taken, not_taken, cnt, prob_str); 837 } 838 return prob; 839 } 840 841 //-----------------------------branch_prediction------------------------------- 842 float Parse::branch_prediction(float& cnt, 843 BoolTest::mask btest, 844 int target_bci) { 845 float prob = dynamic_branch_prediction(cnt); 846 // If prob is unknown, switch to static prediction 847 if (prob != PROB_UNKNOWN) return prob; 848 849 prob = PROB_FAIR; // Set default value 850 if (btest == BoolTest::eq) // Exactly equal test? 851 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 852 else if (btest == BoolTest::ne) 853 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 854 855 // If this is a conditional test guarding a backwards branch, 856 // assume its a loop-back edge. Make it a likely taken branch. 857 if (target_bci < bci()) { 858 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 859 // Since it's an OSR, we probably have profile data, but since 860 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 861 // Let's make a special check here for completely zero counts. 862 ciMethodData* methodData = method()->method_data(); 863 if (!methodData->is_empty()) { 864 ciProfileData* data = methodData->bci_to_data(bci()); 865 // Only stop for truly zero counts, which mean an unknown part 866 // of the OSR-ed method, and we want to deopt to gather more stats. 867 // If you have ANY counts, then this loop is simply 'cold' relative 868 // to the OSR loop. 869 if (data->as_BranchData()->taken() + 870 data->as_BranchData()->not_taken() == 0 ) { 871 // This is the only way to return PROB_UNKNOWN: 872 return PROB_UNKNOWN; 873 } 874 } 875 } 876 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 877 } 878 879 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 880 return prob; 881 } 882 883 // The magic constants are chosen so as to match the output of 884 // branch_prediction() when the profile reports a zero taken count. 885 // It is important to distinguish zero counts unambiguously, because 886 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 887 // very small but nonzero probabilities, which if confused with zero 888 // counts would keep the program recompiling indefinitely. 889 bool Parse::seems_never_taken(float prob) { 890 return prob < PROB_MIN; 891 } 892 893 // True if the comparison seems to be the kind that will not change its 894 // statistics from true to false. See comments in adjust_map_after_if. 895 // This question is only asked along paths which are already 896 // classifed as untaken (by seems_never_taken), so really, 897 // if a path is never taken, its controlling comparison is 898 // already acting in a stable fashion. If the comparison 899 // seems stable, we will put an expensive uncommon trap 900 // on the untaken path. To be conservative, and to allow 901 // partially executed counted loops to be compiled fully, 902 // we will plant uncommon traps only after pointer comparisons. 903 bool Parse::seems_stable_comparison(BoolTest::mask btest, Node* cmp) { 904 for (int depth = 4; depth > 0; depth--) { 905 // The following switch can find CmpP here over half the time for 906 // dynamic language code rich with type tests. 907 // Code using counted loops or array manipulations (typical 908 // of benchmarks) will have many (>80%) CmpI instructions. 909 switch (cmp->Opcode()) { 910 case Op_CmpP: 911 // A never-taken null check looks like CmpP/BoolTest::eq. 912 // These certainly should be closed off as uncommon traps. 913 if (btest == BoolTest::eq) 914 return true; 915 // A never-failed type check looks like CmpP/BoolTest::ne. 916 // Let's put traps on those, too, so that we don't have to compile 917 // unused paths with indeterminate dynamic type information. 918 if (ProfileDynamicTypes) 919 return true; 920 return false; 921 922 case Op_CmpI: 923 // A small minority (< 10%) of CmpP are masked as CmpI, 924 // as if by boolean conversion ((p == q? 1: 0) != 0). 925 // Detect that here, even if it hasn't optimized away yet. 926 // Specifically, this covers the 'instanceof' operator. 927 if (btest == BoolTest::ne || btest == BoolTest::eq) { 928 if (_gvn.type(cmp->in(2))->singleton() && 929 cmp->in(1)->is_Phi()) { 930 PhiNode* phi = cmp->in(1)->as_Phi(); 931 int true_path = phi->is_diamond_phi(); 932 if (true_path > 0 && 933 _gvn.type(phi->in(1))->singleton() && 934 _gvn.type(phi->in(2))->singleton()) { 935 // phi->region->if_proj->ifnode->bool->cmp 936 BoolNode* bol = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); 937 btest = bol->_test._test; 938 cmp = bol->in(1); 939 continue; 940 } 941 } 942 } 943 return false; 944 } 945 } 946 return false; 947 } 948 949 //-------------------------------repush_if_args-------------------------------- 950 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 951 inline int Parse::repush_if_args() { 952 #ifndef PRODUCT 953 if (PrintOpto && WizardMode) { 954 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 955 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 956 method()->print_name(); tty->cr(); 957 } 958 #endif 959 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 960 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 961 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 962 assert(argument(0) != NULL, "must exist"); 963 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 964 inc_sp(bc_depth); 965 return bc_depth; 966 } 967 968 //----------------------------------do_ifnull---------------------------------- 969 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 970 int target_bci = iter().get_dest(); 971 972 Block* branch_block = successor_for_bci(target_bci); 973 Block* next_block = successor_for_bci(iter().next_bci()); 974 975 float cnt; 976 float prob = branch_prediction(cnt, btest, target_bci); 977 if (prob == PROB_UNKNOWN) { 978 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 979 #ifndef PRODUCT 980 if (PrintOpto && Verbose) 981 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 982 #endif 983 repush_if_args(); // to gather stats on loop 984 // We need to mark this branch as taken so that if we recompile we will 985 // see that it is possible. In the tiered system the interpreter doesn't 986 // do profiling and by the time we get to the lower tier from the interpreter 987 // the path may be cold again. Make sure it doesn't look untaken 988 profile_taken_branch(target_bci, !ProfileInterpreter); 989 uncommon_trap(Deoptimization::Reason_unreached, 990 Deoptimization::Action_reinterpret, 991 NULL, "cold"); 992 if (C->eliminate_boxing()) { 993 // Mark the successor blocks as parsed 994 branch_block->next_path_num(); 995 next_block->next_path_num(); 996 } 997 return; 998 } 999 1000 explicit_null_checks_inserted++; 1001 1002 // Generate real control flow 1003 Node *tst = _gvn.transform( new (C) BoolNode( c, btest ) ); 1004 1005 // Sanity check the probability value 1006 assert(prob > 0.0f,"Bad probability in Parser"); 1007 // Need xform to put node in hash table 1008 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1009 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1010 // True branch 1011 { PreserveJVMState pjvms(this); 1012 Node* iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 1013 set_control(iftrue); 1014 1015 if (stopped()) { // Path is dead? 1016 explicit_null_checks_elided++; 1017 if (C->eliminate_boxing()) { 1018 // Mark the successor block as parsed 1019 branch_block->next_path_num(); 1020 } 1021 } else { // Path is live. 1022 // Update method data 1023 profile_taken_branch(target_bci); 1024 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1025 if (!stopped()) { 1026 merge(target_bci); 1027 } 1028 } 1029 } 1030 1031 // False branch 1032 Node* iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 1033 set_control(iffalse); 1034 1035 if (stopped()) { // Path is dead? 1036 explicit_null_checks_elided++; 1037 if (C->eliminate_boxing()) { 1038 // Mark the successor block as parsed 1039 next_block->next_path_num(); 1040 } 1041 } else { // Path is live. 1042 // Update method data 1043 profile_not_taken_branch(); 1044 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1045 next_block, branch_block); 1046 } 1047 } 1048 1049 //------------------------------------do_if------------------------------------ 1050 void Parse::do_if(BoolTest::mask btest, Node* c) { 1051 int target_bci = iter().get_dest(); 1052 1053 Block* branch_block = successor_for_bci(target_bci); 1054 Block* next_block = successor_for_bci(iter().next_bci()); 1055 1056 float cnt; 1057 float prob = branch_prediction(cnt, btest, target_bci); 1058 float untaken_prob = 1.0 - prob; 1059 1060 if (prob == PROB_UNKNOWN) { 1061 #ifndef PRODUCT 1062 if (PrintOpto && Verbose) 1063 tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1064 #endif 1065 repush_if_args(); // to gather stats on loop 1066 // We need to mark this branch as taken so that if we recompile we will 1067 // see that it is possible. In the tiered system the interpreter doesn't 1068 // do profiling and by the time we get to the lower tier from the interpreter 1069 // the path may be cold again. Make sure it doesn't look untaken 1070 profile_taken_branch(target_bci, !ProfileInterpreter); 1071 uncommon_trap(Deoptimization::Reason_unreached, 1072 Deoptimization::Action_reinterpret, 1073 NULL, "cold"); 1074 if (C->eliminate_boxing()) { 1075 // Mark the successor blocks as parsed 1076 branch_block->next_path_num(); 1077 next_block->next_path_num(); 1078 } 1079 return; 1080 } 1081 1082 // Sanity check the probability value 1083 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1084 1085 bool taken_if_true = true; 1086 // Convert BoolTest to canonical form: 1087 if (!BoolTest(btest).is_canonical()) { 1088 btest = BoolTest(btest).negate(); 1089 taken_if_true = false; 1090 // prob is NOT updated here; it remains the probability of the taken 1091 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1092 } 1093 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1094 1095 Node* tst0 = new (C) BoolNode(c, btest); 1096 Node* tst = _gvn.transform(tst0); 1097 BoolTest::mask taken_btest = BoolTest::illegal; 1098 BoolTest::mask untaken_btest = BoolTest::illegal; 1099 1100 if (tst->is_Bool()) { 1101 // Refresh c from the transformed bool node, since it may be 1102 // simpler than the original c. Also re-canonicalize btest. 1103 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1104 // That can arise from statements like: if (x instanceof C) ... 1105 if (tst != tst0) { 1106 // Canonicalize one more time since transform can change it. 1107 btest = tst->as_Bool()->_test._test; 1108 if (!BoolTest(btest).is_canonical()) { 1109 // Reverse edges one more time... 1110 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1111 btest = tst->as_Bool()->_test._test; 1112 assert(BoolTest(btest).is_canonical(), "sanity"); 1113 taken_if_true = !taken_if_true; 1114 } 1115 c = tst->in(1); 1116 } 1117 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1118 taken_btest = taken_if_true ? btest : neg_btest; 1119 untaken_btest = taken_if_true ? neg_btest : btest; 1120 } 1121 1122 // Generate real control flow 1123 float true_prob = (taken_if_true ? prob : untaken_prob); 1124 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1125 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1126 Node* taken_branch = new (C) IfTrueNode(iff); 1127 Node* untaken_branch = new (C) IfFalseNode(iff); 1128 if (!taken_if_true) { // Finish conversion to canonical form 1129 Node* tmp = taken_branch; 1130 taken_branch = untaken_branch; 1131 untaken_branch = tmp; 1132 } 1133 1134 // Branch is taken: 1135 { PreserveJVMState pjvms(this); 1136 taken_branch = _gvn.transform(taken_branch); 1137 set_control(taken_branch); 1138 1139 if (stopped()) { 1140 if (C->eliminate_boxing()) { 1141 // Mark the successor block as parsed 1142 branch_block->next_path_num(); 1143 } 1144 } else { 1145 // Update method data 1146 profile_taken_branch(target_bci); 1147 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1148 if (!stopped()) { 1149 merge(target_bci); 1150 } 1151 } 1152 } 1153 1154 untaken_branch = _gvn.transform(untaken_branch); 1155 set_control(untaken_branch); 1156 1157 // Branch not taken. 1158 if (stopped()) { 1159 if (C->eliminate_boxing()) { 1160 // Mark the successor block as parsed 1161 next_block->next_path_num(); 1162 } 1163 } else { 1164 // Update method data 1165 profile_not_taken_branch(); 1166 adjust_map_after_if(untaken_btest, c, untaken_prob, 1167 next_block, branch_block); 1168 } 1169 } 1170 1171 //----------------------------adjust_map_after_if------------------------------ 1172 // Adjust the JVM state to reflect the result of taking this path. 1173 // Basically, it means inspecting the CmpNode controlling this 1174 // branch, seeing how it constrains a tested value, and then 1175 // deciding if it's worth our while to encode this constraint 1176 // as graph nodes in the current abstract interpretation map. 1177 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1178 Block* path, Block* other_path) { 1179 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1180 return; // nothing to do 1181 1182 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1183 1184 if (seems_never_taken(prob) && seems_stable_comparison(btest, c)) { 1185 // If this might possibly turn into an implicit null check, 1186 // and the null has never yet been seen, we need to generate 1187 // an uncommon trap, so as to recompile instead of suffering 1188 // with very slow branches. (We'll get the slow branches if 1189 // the program ever changes phase and starts seeing nulls here.) 1190 // 1191 // We do not inspect for a null constant, since a node may 1192 // optimize to 'null' later on. 1193 // 1194 // Null checks, and other tests which expect inequality, 1195 // show btest == BoolTest::eq along the non-taken branch. 1196 // On the other hand, type tests, must-be-null tests, 1197 // and other tests which expect pointer equality, 1198 // show btest == BoolTest::ne along the non-taken branch. 1199 // We prune both types of branches if they look unused. 1200 repush_if_args(); 1201 // We need to mark this branch as taken so that if we recompile we will 1202 // see that it is possible. In the tiered system the interpreter doesn't 1203 // do profiling and by the time we get to the lower tier from the interpreter 1204 // the path may be cold again. Make sure it doesn't look untaken 1205 if (is_fallthrough) { 1206 profile_not_taken_branch(!ProfileInterpreter); 1207 } else { 1208 profile_taken_branch(iter().get_dest(), !ProfileInterpreter); 1209 } 1210 uncommon_trap(Deoptimization::Reason_unreached, 1211 Deoptimization::Action_reinterpret, 1212 NULL, 1213 (is_fallthrough ? "taken always" : "taken never")); 1214 return; 1215 } 1216 1217 Node* val = c->in(1); 1218 Node* con = c->in(2); 1219 const Type* tcon = _gvn.type(con); 1220 const Type* tval = _gvn.type(val); 1221 bool have_con = tcon->singleton(); 1222 if (tval->singleton()) { 1223 if (!have_con) { 1224 // Swap, so constant is in con. 1225 con = val; 1226 tcon = tval; 1227 val = c->in(2); 1228 tval = _gvn.type(val); 1229 btest = BoolTest(btest).commute(); 1230 have_con = true; 1231 } else { 1232 // Do we have two constants? Then leave well enough alone. 1233 have_con = false; 1234 } 1235 } 1236 if (!have_con) // remaining adjustments need a con 1237 return; 1238 1239 sharpen_type_after_if(btest, con, tcon, val, tval); 1240 } 1241 1242 1243 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1244 Node* ldk; 1245 if (n->is_DecodeNKlass()) { 1246 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1247 return NULL; 1248 } else { 1249 ldk = n->in(1); 1250 } 1251 } else if (n->Opcode() != Op_LoadKlass) { 1252 return NULL; 1253 } else { 1254 ldk = n; 1255 } 1256 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1257 1258 Node* adr = ldk->in(MemNode::Address); 1259 intptr_t off = 0; 1260 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1261 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1262 return NULL; 1263 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1264 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1265 return NULL; 1266 1267 return obj; 1268 } 1269 1270 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1271 Node* con, const Type* tcon, 1272 Node* val, const Type* tval) { 1273 // Look for opportunities to sharpen the type of a node 1274 // whose klass is compared with a constant klass. 1275 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1276 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1277 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1278 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1279 // Found: 1280 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1281 // or the narrowOop equivalent. 1282 const Type* obj_type = _gvn.type(obj); 1283 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1284 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1285 tboth->higher_equal(obj_type)) { 1286 // obj has to be of the exact type Foo if the CmpP succeeds. 1287 int obj_in_map = map()->find_edge(obj); 1288 JVMState* jvms = this->jvms(); 1289 if (obj_in_map >= 0 && 1290 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1291 TypeNode* ccast = new (C) CheckCastPPNode(control(), obj, tboth); 1292 const Type* tcc = ccast->as_Type()->type(); 1293 assert(tcc != obj_type && tcc->higher_equal_speculative(obj_type), "must improve"); 1294 // Delay transform() call to allow recovery of pre-cast value 1295 // at the control merge. 1296 _gvn.set_type_bottom(ccast); 1297 record_for_igvn(ccast); 1298 // Here's the payoff. 1299 replace_in_map(obj, ccast); 1300 } 1301 } 1302 } 1303 } 1304 1305 int val_in_map = map()->find_edge(val); 1306 if (val_in_map < 0) return; // replace_in_map would be useless 1307 { 1308 JVMState* jvms = this->jvms(); 1309 if (!(jvms->is_loc(val_in_map) || 1310 jvms->is_stk(val_in_map))) 1311 return; // again, it would be useless 1312 } 1313 1314 // Check for a comparison to a constant, and "know" that the compared 1315 // value is constrained on this path. 1316 assert(tcon->singleton(), ""); 1317 ConstraintCastNode* ccast = NULL; 1318 Node* cast = NULL; 1319 1320 switch (btest) { 1321 case BoolTest::eq: // Constant test? 1322 { 1323 const Type* tboth = tcon->join_speculative(tval); 1324 if (tboth == tval) break; // Nothing to gain. 1325 if (tcon->isa_int()) { 1326 ccast = new (C) CastIINode(val, tboth); 1327 } else if (tcon == TypePtr::NULL_PTR) { 1328 // Cast to null, but keep the pointer identity temporarily live. 1329 ccast = new (C) CastPPNode(val, tboth); 1330 } else { 1331 const TypeF* tf = tcon->isa_float_constant(); 1332 const TypeD* td = tcon->isa_double_constant(); 1333 // Exclude tests vs float/double 0 as these could be 1334 // either +0 or -0. Just because you are equal to +0 1335 // doesn't mean you ARE +0! 1336 // Note, following code also replaces Long and Oop values. 1337 if ((!tf || tf->_f != 0.0) && 1338 (!td || td->_d != 0.0)) 1339 cast = con; // Replace non-constant val by con. 1340 } 1341 } 1342 break; 1343 1344 case BoolTest::ne: 1345 if (tcon == TypePtr::NULL_PTR) { 1346 cast = cast_not_null(val, false); 1347 } 1348 break; 1349 1350 default: 1351 // (At this point we could record int range types with CastII.) 1352 break; 1353 } 1354 1355 if (ccast != NULL) { 1356 const Type* tcc = ccast->as_Type()->type(); 1357 assert(tcc != tval && tcc->higher_equal_speculative(tval), "must improve"); 1358 // Delay transform() call to allow recovery of pre-cast value 1359 // at the control merge. 1360 ccast->set_req(0, control()); 1361 _gvn.set_type_bottom(ccast); 1362 record_for_igvn(ccast); 1363 cast = ccast; 1364 } 1365 1366 if (cast != NULL) { // Here's the payoff. 1367 replace_in_map(val, cast); 1368 } 1369 } 1370 1371 /** 1372 * Use speculative type to optimize CmpP node: if comparison is 1373 * against the low level class, cast the object to the speculative 1374 * type if any. CmpP should then go away. 1375 * 1376 * @param c expected CmpP node 1377 * @return result of CmpP on object casted to speculative type 1378 * 1379 */ 1380 Node* Parse::optimize_cmp_with_klass(Node* c) { 1381 // If this is transformed by the _gvn to a comparison with the low 1382 // level klass then we may be able to use speculation 1383 if (c->Opcode() == Op_CmpP && 1384 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1385 c->in(2)->is_Con()) { 1386 Node* load_klass = NULL; 1387 Node* decode = NULL; 1388 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1389 decode = c->in(1); 1390 load_klass = c->in(1)->in(1); 1391 } else { 1392 load_klass = c->in(1); 1393 } 1394 if (load_klass->in(2)->is_AddP()) { 1395 Node* addp = load_klass->in(2); 1396 Node* obj = addp->in(AddPNode::Address); 1397 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1398 if (obj_type->speculative_type() != NULL) { 1399 ciKlass* k = obj_type->speculative_type(); 1400 inc_sp(2); 1401 obj = maybe_cast_profiled_obj(obj, k); 1402 dec_sp(2); 1403 // Make the CmpP use the casted obj 1404 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1405 load_klass = load_klass->clone(); 1406 load_klass->set_req(2, addp); 1407 load_klass = _gvn.transform(load_klass); 1408 if (decode != NULL) { 1409 decode = decode->clone(); 1410 decode->set_req(1, load_klass); 1411 load_klass = _gvn.transform(decode); 1412 } 1413 c = c->clone(); 1414 c->set_req(1, load_klass); 1415 c = _gvn.transform(c); 1416 } 1417 } 1418 } 1419 return c; 1420 } 1421 1422 //------------------------------do_one_bytecode-------------------------------- 1423 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1424 void Parse::do_one_bytecode() { 1425 Node *a, *b, *c, *d; // Handy temps 1426 BoolTest::mask btest; 1427 int i; 1428 1429 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1430 1431 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1432 "out of nodes parsing method")) { 1433 return; 1434 } 1435 1436 #ifdef ASSERT 1437 // for setting breakpoints 1438 if (TraceOptoParse) { 1439 tty->print(" @"); 1440 dump_bci(bci()); 1441 tty->cr(); 1442 } 1443 #endif 1444 1445 switch (bc()) { 1446 case Bytecodes::_nop: 1447 // do nothing 1448 break; 1449 case Bytecodes::_lconst_0: 1450 push_pair(longcon(0)); 1451 break; 1452 1453 case Bytecodes::_lconst_1: 1454 push_pair(longcon(1)); 1455 break; 1456 1457 case Bytecodes::_fconst_0: 1458 push(zerocon(T_FLOAT)); 1459 break; 1460 1461 case Bytecodes::_fconst_1: 1462 push(makecon(TypeF::ONE)); 1463 break; 1464 1465 case Bytecodes::_fconst_2: 1466 push(makecon(TypeF::make(2.0f))); 1467 break; 1468 1469 case Bytecodes::_dconst_0: 1470 push_pair(zerocon(T_DOUBLE)); 1471 break; 1472 1473 case Bytecodes::_dconst_1: 1474 push_pair(makecon(TypeD::ONE)); 1475 break; 1476 1477 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1478 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1479 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1480 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1481 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1482 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1483 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1484 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1485 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1486 case Bytecodes::_aconst_null: push(null()); break; 1487 case Bytecodes::_ldc: 1488 case Bytecodes::_ldc_w: 1489 case Bytecodes::_ldc2_w: 1490 // If the constant is unresolved, run this BC once in the interpreter. 1491 { 1492 ciConstant constant = iter().get_constant(); 1493 if (constant.basic_type() == T_OBJECT && 1494 !constant.as_object()->is_loaded()) { 1495 int index = iter().get_constant_pool_index(); 1496 constantTag tag = iter().get_constant_pool_tag(index); 1497 uncommon_trap(Deoptimization::make_trap_request 1498 (Deoptimization::Reason_unloaded, 1499 Deoptimization::Action_reinterpret, 1500 index), 1501 NULL, tag.internal_name()); 1502 break; 1503 } 1504 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1505 "must be java_mirror of klass"); 1506 bool pushed = push_constant(constant, true); 1507 guarantee(pushed, "must be possible to push this constant"); 1508 } 1509 1510 break; 1511 1512 case Bytecodes::_aload_0: 1513 push( local(0) ); 1514 break; 1515 case Bytecodes::_aload_1: 1516 push( local(1) ); 1517 break; 1518 case Bytecodes::_aload_2: 1519 push( local(2) ); 1520 break; 1521 case Bytecodes::_aload_3: 1522 push( local(3) ); 1523 break; 1524 case Bytecodes::_aload: 1525 push( local(iter().get_index()) ); 1526 break; 1527 1528 case Bytecodes::_fload_0: 1529 case Bytecodes::_iload_0: 1530 push( local(0) ); 1531 break; 1532 case Bytecodes::_fload_1: 1533 case Bytecodes::_iload_1: 1534 push( local(1) ); 1535 break; 1536 case Bytecodes::_fload_2: 1537 case Bytecodes::_iload_2: 1538 push( local(2) ); 1539 break; 1540 case Bytecodes::_fload_3: 1541 case Bytecodes::_iload_3: 1542 push( local(3) ); 1543 break; 1544 case Bytecodes::_fload: 1545 case Bytecodes::_iload: 1546 push( local(iter().get_index()) ); 1547 break; 1548 case Bytecodes::_lload_0: 1549 push_pair_local( 0 ); 1550 break; 1551 case Bytecodes::_lload_1: 1552 push_pair_local( 1 ); 1553 break; 1554 case Bytecodes::_lload_2: 1555 push_pair_local( 2 ); 1556 break; 1557 case Bytecodes::_lload_3: 1558 push_pair_local( 3 ); 1559 break; 1560 case Bytecodes::_lload: 1561 push_pair_local( iter().get_index() ); 1562 break; 1563 1564 case Bytecodes::_dload_0: 1565 push_pair_local(0); 1566 break; 1567 case Bytecodes::_dload_1: 1568 push_pair_local(1); 1569 break; 1570 case Bytecodes::_dload_2: 1571 push_pair_local(2); 1572 break; 1573 case Bytecodes::_dload_3: 1574 push_pair_local(3); 1575 break; 1576 case Bytecodes::_dload: 1577 push_pair_local(iter().get_index()); 1578 break; 1579 case Bytecodes::_fstore_0: 1580 case Bytecodes::_istore_0: 1581 case Bytecodes::_astore_0: 1582 set_local( 0, pop() ); 1583 break; 1584 case Bytecodes::_fstore_1: 1585 case Bytecodes::_istore_1: 1586 case Bytecodes::_astore_1: 1587 set_local( 1, pop() ); 1588 break; 1589 case Bytecodes::_fstore_2: 1590 case Bytecodes::_istore_2: 1591 case Bytecodes::_astore_2: 1592 set_local( 2, pop() ); 1593 break; 1594 case Bytecodes::_fstore_3: 1595 case Bytecodes::_istore_3: 1596 case Bytecodes::_astore_3: 1597 set_local( 3, pop() ); 1598 break; 1599 case Bytecodes::_fstore: 1600 case Bytecodes::_istore: 1601 case Bytecodes::_astore: 1602 set_local( iter().get_index(), pop() ); 1603 break; 1604 // long stores 1605 case Bytecodes::_lstore_0: 1606 set_pair_local( 0, pop_pair() ); 1607 break; 1608 case Bytecodes::_lstore_1: 1609 set_pair_local( 1, pop_pair() ); 1610 break; 1611 case Bytecodes::_lstore_2: 1612 set_pair_local( 2, pop_pair() ); 1613 break; 1614 case Bytecodes::_lstore_3: 1615 set_pair_local( 3, pop_pair() ); 1616 break; 1617 case Bytecodes::_lstore: 1618 set_pair_local( iter().get_index(), pop_pair() ); 1619 break; 1620 1621 // double stores 1622 case Bytecodes::_dstore_0: 1623 set_pair_local( 0, dstore_rounding(pop_pair()) ); 1624 break; 1625 case Bytecodes::_dstore_1: 1626 set_pair_local( 1, dstore_rounding(pop_pair()) ); 1627 break; 1628 case Bytecodes::_dstore_2: 1629 set_pair_local( 2, dstore_rounding(pop_pair()) ); 1630 break; 1631 case Bytecodes::_dstore_3: 1632 set_pair_local( 3, dstore_rounding(pop_pair()) ); 1633 break; 1634 case Bytecodes::_dstore: 1635 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1636 break; 1637 1638 case Bytecodes::_pop: dec_sp(1); break; 1639 case Bytecodes::_pop2: dec_sp(2); break; 1640 case Bytecodes::_swap: 1641 a = pop(); 1642 b = pop(); 1643 push(a); 1644 push(b); 1645 break; 1646 case Bytecodes::_dup: 1647 a = pop(); 1648 push(a); 1649 push(a); 1650 break; 1651 case Bytecodes::_dup_x1: 1652 a = pop(); 1653 b = pop(); 1654 push( a ); 1655 push( b ); 1656 push( a ); 1657 break; 1658 case Bytecodes::_dup_x2: 1659 a = pop(); 1660 b = pop(); 1661 c = pop(); 1662 push( a ); 1663 push( c ); 1664 push( b ); 1665 push( a ); 1666 break; 1667 case Bytecodes::_dup2: 1668 a = pop(); 1669 b = pop(); 1670 push( b ); 1671 push( a ); 1672 push( b ); 1673 push( a ); 1674 break; 1675 1676 case Bytecodes::_dup2_x1: 1677 // before: .. c, b, a 1678 // after: .. b, a, c, b, a 1679 // not tested 1680 a = pop(); 1681 b = pop(); 1682 c = pop(); 1683 push( b ); 1684 push( a ); 1685 push( c ); 1686 push( b ); 1687 push( a ); 1688 break; 1689 case Bytecodes::_dup2_x2: 1690 // before: .. d, c, b, a 1691 // after: .. b, a, d, c, b, a 1692 // not tested 1693 a = pop(); 1694 b = pop(); 1695 c = pop(); 1696 d = pop(); 1697 push( b ); 1698 push( a ); 1699 push( d ); 1700 push( c ); 1701 push( b ); 1702 push( a ); 1703 break; 1704 1705 case Bytecodes::_arraylength: { 1706 // Must do null-check with value on expression stack 1707 Node *ary = null_check(peek(), T_ARRAY); 1708 // Compile-time detect of null-exception? 1709 if (stopped()) return; 1710 a = pop(); 1711 push(load_array_length(a)); 1712 break; 1713 } 1714 1715 case Bytecodes::_baload: array_load(T_BYTE); break; 1716 case Bytecodes::_caload: array_load(T_CHAR); break; 1717 case Bytecodes::_iaload: array_load(T_INT); break; 1718 case Bytecodes::_saload: array_load(T_SHORT); break; 1719 case Bytecodes::_faload: array_load(T_FLOAT); break; 1720 case Bytecodes::_aaload: array_load(T_OBJECT); break; 1721 case Bytecodes::_laload: { 1722 a = array_addressing(T_LONG, 0); 1723 if (stopped()) return; // guaranteed null or range check 1724 dec_sp(2); // Pop array and index 1725 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1726 break; 1727 } 1728 case Bytecodes::_daload: { 1729 a = array_addressing(T_DOUBLE, 0); 1730 if (stopped()) return; // guaranteed null or range check 1731 dec_sp(2); // Pop array and index 1732 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1733 break; 1734 } 1735 case Bytecodes::_bastore: array_store(T_BYTE); break; 1736 case Bytecodes::_castore: array_store(T_CHAR); break; 1737 case Bytecodes::_iastore: array_store(T_INT); break; 1738 case Bytecodes::_sastore: array_store(T_SHORT); break; 1739 case Bytecodes::_fastore: array_store(T_FLOAT); break; 1740 case Bytecodes::_aastore: { 1741 d = array_addressing(T_OBJECT, 1); 1742 if (stopped()) return; // guaranteed null or range check 1743 array_store_check(); 1744 c = pop(); // Oop to store 1745 b = pop(); // index (already used) 1746 a = pop(); // the array itself 1747 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1748 const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1749 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release); 1750 break; 1751 } 1752 case Bytecodes::_lastore: { 1753 a = array_addressing(T_LONG, 2); 1754 if (stopped()) return; // guaranteed null or range check 1755 c = pop_pair(); 1756 dec_sp(2); // Pop array and index 1757 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1758 break; 1759 } 1760 case Bytecodes::_dastore: { 1761 a = array_addressing(T_DOUBLE, 2); 1762 if (stopped()) return; // guaranteed null or range check 1763 c = pop_pair(); 1764 dec_sp(2); // Pop array and index 1765 c = dstore_rounding(c); 1766 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1767 break; 1768 } 1769 case Bytecodes::_getfield: 1770 do_getfield(); 1771 break; 1772 1773 case Bytecodes::_getstatic: 1774 do_getstatic(); 1775 break; 1776 1777 case Bytecodes::_putfield: 1778 do_putfield(); 1779 break; 1780 1781 case Bytecodes::_putstatic: 1782 do_putstatic(); 1783 break; 1784 1785 case Bytecodes::_irem: 1786 do_irem(); 1787 break; 1788 case Bytecodes::_idiv: 1789 // Must keep both values on the expression-stack during null-check 1790 zero_check_int(peek()); 1791 // Compile-time detect of null-exception? 1792 if (stopped()) return; 1793 b = pop(); 1794 a = pop(); 1795 push( _gvn.transform( new (C) DivINode(control(),a,b) ) ); 1796 break; 1797 case Bytecodes::_imul: 1798 b = pop(); a = pop(); 1799 push( _gvn.transform( new (C) MulINode(a,b) ) ); 1800 break; 1801 case Bytecodes::_iadd: 1802 b = pop(); a = pop(); 1803 push( _gvn.transform( new (C) AddINode(a,b) ) ); 1804 break; 1805 case Bytecodes::_ineg: 1806 a = pop(); 1807 push( _gvn.transform( new (C) SubINode(_gvn.intcon(0),a)) ); 1808 break; 1809 case Bytecodes::_isub: 1810 b = pop(); a = pop(); 1811 push( _gvn.transform( new (C) SubINode(a,b) ) ); 1812 break; 1813 case Bytecodes::_iand: 1814 b = pop(); a = pop(); 1815 push( _gvn.transform( new (C) AndINode(a,b) ) ); 1816 break; 1817 case Bytecodes::_ior: 1818 b = pop(); a = pop(); 1819 push( _gvn.transform( new (C) OrINode(a,b) ) ); 1820 break; 1821 case Bytecodes::_ixor: 1822 b = pop(); a = pop(); 1823 push( _gvn.transform( new (C) XorINode(a,b) ) ); 1824 break; 1825 case Bytecodes::_ishl: 1826 b = pop(); a = pop(); 1827 push( _gvn.transform( new (C) LShiftINode(a,b) ) ); 1828 break; 1829 case Bytecodes::_ishr: 1830 b = pop(); a = pop(); 1831 push( _gvn.transform( new (C) RShiftINode(a,b) ) ); 1832 break; 1833 case Bytecodes::_iushr: 1834 b = pop(); a = pop(); 1835 push( _gvn.transform( new (C) URShiftINode(a,b) ) ); 1836 break; 1837 1838 case Bytecodes::_fneg: 1839 a = pop(); 1840 b = _gvn.transform(new (C) NegFNode (a)); 1841 push(b); 1842 break; 1843 1844 case Bytecodes::_fsub: 1845 b = pop(); 1846 a = pop(); 1847 c = _gvn.transform( new (C) SubFNode(a,b) ); 1848 d = precision_rounding(c); 1849 push( d ); 1850 break; 1851 1852 case Bytecodes::_fadd: 1853 b = pop(); 1854 a = pop(); 1855 c = _gvn.transform( new (C) AddFNode(a,b) ); 1856 d = precision_rounding(c); 1857 push( d ); 1858 break; 1859 1860 case Bytecodes::_fmul: 1861 b = pop(); 1862 a = pop(); 1863 c = _gvn.transform( new (C) MulFNode(a,b) ); 1864 d = precision_rounding(c); 1865 push( d ); 1866 break; 1867 1868 case Bytecodes::_fdiv: 1869 b = pop(); 1870 a = pop(); 1871 c = _gvn.transform( new (C) DivFNode(0,a,b) ); 1872 d = precision_rounding(c); 1873 push( d ); 1874 break; 1875 1876 case Bytecodes::_frem: 1877 if (Matcher::has_match_rule(Op_ModF)) { 1878 // Generate a ModF node. 1879 b = pop(); 1880 a = pop(); 1881 c = _gvn.transform( new (C) ModFNode(0,a,b) ); 1882 d = precision_rounding(c); 1883 push( d ); 1884 } 1885 else { 1886 // Generate a call. 1887 modf(); 1888 } 1889 break; 1890 1891 case Bytecodes::_fcmpl: 1892 b = pop(); 1893 a = pop(); 1894 c = _gvn.transform( new (C) CmpF3Node( a, b)); 1895 push(c); 1896 break; 1897 case Bytecodes::_fcmpg: 1898 b = pop(); 1899 a = pop(); 1900 1901 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1902 // which negates the result sign except for unordered. Flip the unordered 1903 // as well by using CmpF3 which implements unordered-lesser instead of 1904 // unordered-greater semantics. Finally, commute the result bits. Result 1905 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1906 c = _gvn.transform( new (C) CmpF3Node( b, a)); 1907 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 1908 push(c); 1909 break; 1910 1911 case Bytecodes::_f2i: 1912 a = pop(); 1913 push(_gvn.transform(new (C) ConvF2INode(a))); 1914 break; 1915 1916 case Bytecodes::_d2i: 1917 a = pop_pair(); 1918 b = _gvn.transform(new (C) ConvD2INode(a)); 1919 push( b ); 1920 break; 1921 1922 case Bytecodes::_f2d: 1923 a = pop(); 1924 b = _gvn.transform( new (C) ConvF2DNode(a)); 1925 push_pair( b ); 1926 break; 1927 1928 case Bytecodes::_d2f: 1929 a = pop_pair(); 1930 b = _gvn.transform( new (C) ConvD2FNode(a)); 1931 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1932 //b = _gvn.transform(new (C) RoundFloatNode(0, b) ); 1933 push( b ); 1934 break; 1935 1936 case Bytecodes::_l2f: 1937 if (Matcher::convL2FSupported()) { 1938 a = pop_pair(); 1939 b = _gvn.transform( new (C) ConvL2FNode(a)); 1940 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1941 // Rather than storing the result into an FP register then pushing 1942 // out to memory to round, the machine instruction that implements 1943 // ConvL2D is responsible for rounding. 1944 // c = precision_rounding(b); 1945 c = _gvn.transform(b); 1946 push(c); 1947 } else { 1948 l2f(); 1949 } 1950 break; 1951 1952 case Bytecodes::_l2d: 1953 a = pop_pair(); 1954 b = _gvn.transform( new (C) ConvL2DNode(a)); 1955 // For i486.ad, rounding is always necessary (see _l2f above). 1956 // c = dprecision_rounding(b); 1957 c = _gvn.transform(b); 1958 push_pair(c); 1959 break; 1960 1961 case Bytecodes::_f2l: 1962 a = pop(); 1963 b = _gvn.transform( new (C) ConvF2LNode(a)); 1964 push_pair(b); 1965 break; 1966 1967 case Bytecodes::_d2l: 1968 a = pop_pair(); 1969 b = _gvn.transform( new (C) ConvD2LNode(a)); 1970 push_pair(b); 1971 break; 1972 1973 case Bytecodes::_dsub: 1974 b = pop_pair(); 1975 a = pop_pair(); 1976 c = _gvn.transform( new (C) SubDNode(a,b) ); 1977 d = dprecision_rounding(c); 1978 push_pair( d ); 1979 break; 1980 1981 case Bytecodes::_dadd: 1982 b = pop_pair(); 1983 a = pop_pair(); 1984 c = _gvn.transform( new (C) AddDNode(a,b) ); 1985 d = dprecision_rounding(c); 1986 push_pair( d ); 1987 break; 1988 1989 case Bytecodes::_dmul: 1990 b = pop_pair(); 1991 a = pop_pair(); 1992 c = _gvn.transform( new (C) MulDNode(a,b) ); 1993 d = dprecision_rounding(c); 1994 push_pair( d ); 1995 break; 1996 1997 case Bytecodes::_ddiv: 1998 b = pop_pair(); 1999 a = pop_pair(); 2000 c = _gvn.transform( new (C) DivDNode(0,a,b) ); 2001 d = dprecision_rounding(c); 2002 push_pair( d ); 2003 break; 2004 2005 case Bytecodes::_dneg: 2006 a = pop_pair(); 2007 b = _gvn.transform(new (C) NegDNode (a)); 2008 push_pair(b); 2009 break; 2010 2011 case Bytecodes::_drem: 2012 if (Matcher::has_match_rule(Op_ModD)) { 2013 // Generate a ModD node. 2014 b = pop_pair(); 2015 a = pop_pair(); 2016 // a % b 2017 2018 c = _gvn.transform( new (C) ModDNode(0,a,b) ); 2019 d = dprecision_rounding(c); 2020 push_pair( d ); 2021 } 2022 else { 2023 // Generate a call. 2024 modd(); 2025 } 2026 break; 2027 2028 case Bytecodes::_dcmpl: 2029 b = pop_pair(); 2030 a = pop_pair(); 2031 c = _gvn.transform( new (C) CmpD3Node( a, b)); 2032 push(c); 2033 break; 2034 2035 case Bytecodes::_dcmpg: 2036 b = pop_pair(); 2037 a = pop_pair(); 2038 // Same as dcmpl but need to flip the unordered case. 2039 // Commute the inputs, which negates the result sign except for unordered. 2040 // Flip the unordered as well by using CmpD3 which implements 2041 // unordered-lesser instead of unordered-greater semantics. 2042 // Finally, negate the result bits. Result is same as using a 2043 // CmpD3Greater except we did it with CmpD3 alone. 2044 c = _gvn.transform( new (C) CmpD3Node( b, a)); 2045 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 2046 push(c); 2047 break; 2048 2049 2050 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2051 case Bytecodes::_land: 2052 b = pop_pair(); 2053 a = pop_pair(); 2054 c = _gvn.transform( new (C) AndLNode(a,b) ); 2055 push_pair(c); 2056 break; 2057 case Bytecodes::_lor: 2058 b = pop_pair(); 2059 a = pop_pair(); 2060 c = _gvn.transform( new (C) OrLNode(a,b) ); 2061 push_pair(c); 2062 break; 2063 case Bytecodes::_lxor: 2064 b = pop_pair(); 2065 a = pop_pair(); 2066 c = _gvn.transform( new (C) XorLNode(a,b) ); 2067 push_pair(c); 2068 break; 2069 2070 case Bytecodes::_lshl: 2071 b = pop(); // the shift count 2072 a = pop_pair(); // value to be shifted 2073 c = _gvn.transform( new (C) LShiftLNode(a,b) ); 2074 push_pair(c); 2075 break; 2076 case Bytecodes::_lshr: 2077 b = pop(); // the shift count 2078 a = pop_pair(); // value to be shifted 2079 c = _gvn.transform( new (C) RShiftLNode(a,b) ); 2080 push_pair(c); 2081 break; 2082 case Bytecodes::_lushr: 2083 b = pop(); // the shift count 2084 a = pop_pair(); // value to be shifted 2085 c = _gvn.transform( new (C) URShiftLNode(a,b) ); 2086 push_pair(c); 2087 break; 2088 case Bytecodes::_lmul: 2089 b = pop_pair(); 2090 a = pop_pair(); 2091 c = _gvn.transform( new (C) MulLNode(a,b) ); 2092 push_pair(c); 2093 break; 2094 2095 case Bytecodes::_lrem: 2096 // Must keep both values on the expression-stack during null-check 2097 assert(peek(0) == top(), "long word order"); 2098 zero_check_long(peek(1)); 2099 // Compile-time detect of null-exception? 2100 if (stopped()) return; 2101 b = pop_pair(); 2102 a = pop_pair(); 2103 c = _gvn.transform( new (C) ModLNode(control(),a,b) ); 2104 push_pair(c); 2105 break; 2106 2107 case Bytecodes::_ldiv: 2108 // Must keep both values on the expression-stack during null-check 2109 assert(peek(0) == top(), "long word order"); 2110 zero_check_long(peek(1)); 2111 // Compile-time detect of null-exception? 2112 if (stopped()) return; 2113 b = pop_pair(); 2114 a = pop_pair(); 2115 c = _gvn.transform( new (C) DivLNode(control(),a,b) ); 2116 push_pair(c); 2117 break; 2118 2119 case Bytecodes::_ladd: 2120 b = pop_pair(); 2121 a = pop_pair(); 2122 c = _gvn.transform( new (C) AddLNode(a,b) ); 2123 push_pair(c); 2124 break; 2125 case Bytecodes::_lsub: 2126 b = pop_pair(); 2127 a = pop_pair(); 2128 c = _gvn.transform( new (C) SubLNode(a,b) ); 2129 push_pair(c); 2130 break; 2131 case Bytecodes::_lcmp: 2132 // Safepoints are now inserted _before_ branches. The long-compare 2133 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2134 // slew of control flow. These are usually followed by a CmpI vs zero and 2135 // a branch; this pattern then optimizes to the obvious long-compare and 2136 // branch. However, if the branch is backwards there's a Safepoint 2137 // inserted. The inserted Safepoint captures the JVM state at the 2138 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2139 // long-compare is used to control a loop the debug info will force 2140 // computation of the 3-way value, even though the generated code uses a 2141 // long-compare and branch. We try to rectify the situation by inserting 2142 // a SafePoint here and have it dominate and kill the safepoint added at a 2143 // following backwards branch. At this point the JVM state merely holds 2 2144 // longs but not the 3-way value. 2145 if( UseLoopSafepoints ) { 2146 switch( iter().next_bc() ) { 2147 case Bytecodes::_ifgt: 2148 case Bytecodes::_iflt: 2149 case Bytecodes::_ifge: 2150 case Bytecodes::_ifle: 2151 case Bytecodes::_ifne: 2152 case Bytecodes::_ifeq: 2153 // If this is a backwards branch in the bytecodes, add Safepoint 2154 maybe_add_safepoint(iter().next_get_dest()); 2155 } 2156 } 2157 b = pop_pair(); 2158 a = pop_pair(); 2159 c = _gvn.transform( new (C) CmpL3Node( a, b )); 2160 push(c); 2161 break; 2162 2163 case Bytecodes::_lneg: 2164 a = pop_pair(); 2165 b = _gvn.transform( new (C) SubLNode(longcon(0),a)); 2166 push_pair(b); 2167 break; 2168 case Bytecodes::_l2i: 2169 a = pop_pair(); 2170 push( _gvn.transform( new (C) ConvL2INode(a))); 2171 break; 2172 case Bytecodes::_i2l: 2173 a = pop(); 2174 b = _gvn.transform( new (C) ConvI2LNode(a)); 2175 push_pair(b); 2176 break; 2177 case Bytecodes::_i2b: 2178 // Sign extend 2179 a = pop(); 2180 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(24)) ); 2181 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(24)) ); 2182 push( a ); 2183 break; 2184 case Bytecodes::_i2s: 2185 a = pop(); 2186 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(16)) ); 2187 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(16)) ); 2188 push( a ); 2189 break; 2190 case Bytecodes::_i2c: 2191 a = pop(); 2192 push( _gvn.transform( new (C) AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2193 break; 2194 2195 case Bytecodes::_i2f: 2196 a = pop(); 2197 b = _gvn.transform( new (C) ConvI2FNode(a) ) ; 2198 c = precision_rounding(b); 2199 push (b); 2200 break; 2201 2202 case Bytecodes::_i2d: 2203 a = pop(); 2204 b = _gvn.transform( new (C) ConvI2DNode(a)); 2205 push_pair(b); 2206 break; 2207 2208 case Bytecodes::_iinc: // Increment local 2209 i = iter().get_index(); // Get local index 2210 set_local( i, _gvn.transform( new (C) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2211 break; 2212 2213 // Exit points of synchronized methods must have an unlock node 2214 case Bytecodes::_return: 2215 return_current(NULL); 2216 break; 2217 2218 case Bytecodes::_ireturn: 2219 case Bytecodes::_areturn: 2220 case Bytecodes::_freturn: 2221 return_current(pop()); 2222 break; 2223 case Bytecodes::_lreturn: 2224 return_current(pop_pair()); 2225 break; 2226 case Bytecodes::_dreturn: 2227 return_current(pop_pair()); 2228 break; 2229 2230 case Bytecodes::_athrow: 2231 // null exception oop throws NULL pointer exception 2232 null_check(peek()); 2233 if (stopped()) return; 2234 // Hook the thrown exception directly to subsequent handlers. 2235 if (BailoutToInterpreterForThrows) { 2236 // Keep method interpreted from now on. 2237 uncommon_trap(Deoptimization::Reason_unhandled, 2238 Deoptimization::Action_make_not_compilable); 2239 return; 2240 } 2241 if (env()->jvmti_can_post_on_exceptions()) { 2242 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2243 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2244 } 2245 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2246 add_exception_state(make_exception_state(peek())); 2247 break; 2248 2249 case Bytecodes::_goto: // fall through 2250 case Bytecodes::_goto_w: { 2251 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2252 2253 // If this is a backwards branch in the bytecodes, add Safepoint 2254 maybe_add_safepoint(target_bci); 2255 2256 // Update method data 2257 profile_taken_branch(target_bci); 2258 2259 // Merge the current control into the target basic block 2260 merge(target_bci); 2261 2262 // See if we can get some profile data and hand it off to the next block 2263 Block *target_block = block()->successor_for_bci(target_bci); 2264 if (target_block->pred_count() != 1) break; 2265 ciMethodData* methodData = method()->method_data(); 2266 if (!methodData->is_mature()) break; 2267 ciProfileData* data = methodData->bci_to_data(bci()); 2268 assert( data->is_JumpData(), "" ); 2269 int taken = ((ciJumpData*)data)->taken(); 2270 taken = method()->scale_count(taken); 2271 target_block->set_count(taken); 2272 break; 2273 } 2274 2275 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2276 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2277 handle_if_null: 2278 // If this is a backwards branch in the bytecodes, add Safepoint 2279 maybe_add_safepoint(iter().get_dest()); 2280 a = null(); 2281 b = pop(); 2282 c = _gvn.transform( new (C) CmpPNode(b, a) ); 2283 do_ifnull(btest, c); 2284 break; 2285 2286 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2287 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2288 handle_if_acmp: 2289 // If this is a backwards branch in the bytecodes, add Safepoint 2290 maybe_add_safepoint(iter().get_dest()); 2291 a = pop(); 2292 b = pop(); 2293 c = _gvn.transform( new (C) CmpPNode(b, a) ); 2294 c = optimize_cmp_with_klass(c); 2295 do_if(btest, c); 2296 break; 2297 2298 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2299 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2300 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2301 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2302 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2303 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2304 handle_ifxx: 2305 // If this is a backwards branch in the bytecodes, add Safepoint 2306 maybe_add_safepoint(iter().get_dest()); 2307 a = _gvn.intcon(0); 2308 b = pop(); 2309 c = _gvn.transform( new (C) CmpINode(b, a) ); 2310 do_if(btest, c); 2311 break; 2312 2313 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2314 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2315 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2316 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2317 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2318 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2319 handle_if_icmp: 2320 // If this is a backwards branch in the bytecodes, add Safepoint 2321 maybe_add_safepoint(iter().get_dest()); 2322 a = pop(); 2323 b = pop(); 2324 c = _gvn.transform( new (C) CmpINode( b, a ) ); 2325 do_if(btest, c); 2326 break; 2327 2328 case Bytecodes::_tableswitch: 2329 do_tableswitch(); 2330 break; 2331 2332 case Bytecodes::_lookupswitch: 2333 do_lookupswitch(); 2334 break; 2335 2336 case Bytecodes::_invokestatic: 2337 case Bytecodes::_invokedynamic: 2338 case Bytecodes::_invokespecial: 2339 case Bytecodes::_invokevirtual: 2340 case Bytecodes::_invokeinterface: 2341 do_call(); 2342 break; 2343 case Bytecodes::_checkcast: 2344 do_checkcast(); 2345 break; 2346 case Bytecodes::_instanceof: 2347 do_instanceof(); 2348 break; 2349 case Bytecodes::_anewarray: 2350 do_anewarray(); 2351 break; 2352 case Bytecodes::_newarray: 2353 do_newarray((BasicType)iter().get_index()); 2354 break; 2355 case Bytecodes::_multianewarray: 2356 do_multianewarray(); 2357 break; 2358 case Bytecodes::_new: 2359 do_new(); 2360 break; 2361 2362 case Bytecodes::_jsr: 2363 case Bytecodes::_jsr_w: 2364 do_jsr(); 2365 break; 2366 2367 case Bytecodes::_ret: 2368 do_ret(); 2369 break; 2370 2371 2372 case Bytecodes::_monitorenter: 2373 do_monitor_enter(); 2374 break; 2375 2376 case Bytecodes::_monitorexit: 2377 do_monitor_exit(); 2378 break; 2379 2380 case Bytecodes::_breakpoint: 2381 // Breakpoint set concurrently to compile 2382 // %%% use an uncommon trap? 2383 C->record_failure("breakpoint in method"); 2384 return; 2385 2386 default: 2387 #ifndef PRODUCT 2388 map()->dump(99); 2389 #endif 2390 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2391 ShouldNotReachHere(); 2392 } 2393 2394 #ifndef PRODUCT 2395 IdealGraphPrinter *printer = IdealGraphPrinter::printer(); 2396 if(printer) { 2397 char buffer[256]; 2398 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2399 bool old = printer->traverse_outs(); 2400 printer->set_traverse_outs(true); 2401 printer->print_method(C, buffer, 4); 2402 printer->set_traverse_outs(old); 2403 } 2404 #endif 2405 }