1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "interpreter/linkResolver.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "memory/universe.hpp" 33 #include "oops/oop.inline.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/divnode.hpp" 38 #include "opto/idealGraphPrinter.hpp" 39 #include "opto/matcher.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/mulnode.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/parse.hpp" 44 #include "opto/runtime.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 #ifndef PRODUCT 49 extern int explicit_null_checks_inserted, 50 explicit_null_checks_elided; 51 #endif 52 53 //---------------------------------array_load---------------------------------- 54 void Parse::array_load(BasicType bt) { 55 const Type* elemtype = Type::TOP; 56 bool big_val = bt == T_DOUBLE || bt == T_LONG; 57 Node* adr = array_addressing(bt, 0, false, &elemtype); 58 if (stopped()) return; // guaranteed null or range check 59 60 pop(); // index (already used) 61 Node* array = pop(); // the array itself 62 63 if (elemtype == TypeInt::BOOL) { 64 bt = T_BOOLEAN; 65 } else if (bt == T_OBJECT) { 66 elemtype = _gvn.type(array)->is_aryptr()->elem()->make_oopptr(); 67 } 68 69 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 70 71 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt, 72 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD); 73 if (big_val) { 74 push_pair(ld); 75 } else { 76 push(ld); 77 } 78 } 79 80 81 //--------------------------------array_store---------------------------------- 82 void Parse::array_store(BasicType bt) { 83 const Type* elemtype = Type::TOP; 84 bool big_val = bt == T_DOUBLE || bt == T_LONG; 85 Node* adr = array_addressing(bt, big_val ? 2 : 1, true, &elemtype); 86 if (stopped()) return; // guaranteed null or range check 87 if (bt == T_OBJECT) { 88 array_store_check(); 89 } 90 Node* val; // Oop to store 91 if (big_val) { 92 val = pop_pair(); 93 } else { 94 val = pop(); 95 } 96 pop(); // index (already used) 97 Node* array = pop(); // the array itself 98 99 if (elemtype == TypeInt::BOOL) { 100 bt = T_BOOLEAN; 101 } else if (bt == T_OBJECT) { 102 elemtype = _gvn.type(array)->is_aryptr()->elem()->make_oopptr(); 103 } 104 105 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 106 107 access_store_at(control(), array, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY); 108 } 109 110 111 //------------------------------array_addressing------------------------------- 112 // Pull array and index from the stack. Compute pointer-to-element. 113 Node* Parse::array_addressing(BasicType type, int vals, bool is_store, const Type* *result2) { 114 Node *idx = peek(0+vals); // Get from stack without popping 115 Node *ary = peek(1+vals); // in case of exception 116 117 // Null check the array base, with correct stack contents 118 ary = null_check(ary, T_ARRAY); 119 // Compile-time detect of null-exception? 120 if (stopped()) return top(); 121 122 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 123 const TypeInt* sizetype = arytype->size(); 124 const Type* elemtype = arytype->elem(); 125 126 if (UseUniqueSubclasses && result2 != NULL) { 127 const Type* el = elemtype->make_ptr(); 128 if (el && el->isa_instptr()) { 129 const TypeInstPtr* toop = el->is_instptr(); 130 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 131 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 132 const Type* subklass = Type::get_const_type(toop->klass()); 133 elemtype = subklass->join_speculative(el); 134 } 135 } 136 } 137 138 // Check for big class initializers with all constant offsets 139 // feeding into a known-size array. 140 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 141 // See if the highest idx value is less than the lowest array bound, 142 // and if the idx value cannot be negative: 143 bool need_range_check = true; 144 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 145 need_range_check = false; 146 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 147 } 148 149 ciKlass * arytype_klass = arytype->klass(); 150 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 151 // Only fails for some -Xcomp runs 152 // The class is unloaded. We have to run this bytecode in the interpreter. 153 uncommon_trap(Deoptimization::Reason_unloaded, 154 Deoptimization::Action_reinterpret, 155 arytype->klass(), "!loaded array"); 156 return top(); 157 } 158 159 // Do the range check 160 if (GenerateRangeChecks && need_range_check) { 161 Node* tst; 162 if (sizetype->_hi <= 0) { 163 // The greatest array bound is negative, so we can conclude that we're 164 // compiling unreachable code, but the unsigned compare trick used below 165 // only works with non-negative lengths. Instead, hack "tst" to be zero so 166 // the uncommon_trap path will always be taken. 167 tst = _gvn.intcon(0); 168 } else { 169 // Range is constant in array-oop, so we can use the original state of mem 170 Node* len = load_array_length(ary); 171 172 // Test length vs index (standard trick using unsigned compare) 173 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 174 BoolTest::mask btest = BoolTest::lt; 175 tst = _gvn.transform( new BoolNode(chk, btest) ); 176 } 177 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 178 _gvn.set_type(rc, rc->Value(&_gvn)); 179 if (!tst->is_Con()) { 180 record_for_igvn(rc); 181 } 182 set_control(_gvn.transform(new IfTrueNode(rc))); 183 // Branch to failure if out of bounds 184 { 185 PreserveJVMState pjvms(this); 186 set_control(_gvn.transform(new IfFalseNode(rc))); 187 if (C->allow_range_check_smearing()) { 188 // Do not use builtin_throw, since range checks are sometimes 189 // made more stringent by an optimistic transformation. 190 // This creates "tentative" range checks at this point, 191 // which are not guaranteed to throw exceptions. 192 // See IfNode::Ideal, is_range_check, adjust_check. 193 uncommon_trap(Deoptimization::Reason_range_check, 194 Deoptimization::Action_make_not_entrant, 195 NULL, "range_check"); 196 } else { 197 // If we have already recompiled with the range-check-widening 198 // heroic optimization turned off, then we must really be throwing 199 // range check exceptions. 200 builtin_throw(Deoptimization::Reason_range_check, idx); 201 } 202 } 203 } 204 // Check for always knowing you are throwing a range-check exception 205 if (stopped()) return top(); 206 207 if (is_store) { 208 ary = access_resolve_for_write(ary); 209 } else { 210 ary = access_resolve_for_read(ary); 211 } 212 213 // Make array address computation control dependent to prevent it 214 // from floating above the range check during loop optimizations. 215 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 216 217 if (result2 != NULL) *result2 = elemtype; 218 219 assert(ptr != top(), "top should go hand-in-hand with stopped"); 220 221 return ptr; 222 } 223 224 225 // returns IfNode 226 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 227 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 228 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 229 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 230 return iff; 231 } 232 233 // return Region node 234 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 235 Node *region = new RegionNode(3); // 2 results 236 record_for_igvn(region); 237 region->init_req(1, iffalse); 238 region->init_req(2, iftrue ); 239 _gvn.set_type(region, Type::CONTROL); 240 region = _gvn.transform(region); 241 set_control (region); 242 return region; 243 } 244 245 // sentinel value for the target bci to mark never taken branches 246 // (according to profiling) 247 static const int never_reached = INT_MAX; 248 249 //------------------------------helper for tableswitch------------------------- 250 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 251 // True branch, use existing map info 252 { PreserveJVMState pjvms(this); 253 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 254 set_control( iftrue ); 255 if (unc) { 256 repush_if_args(); 257 uncommon_trap(Deoptimization::Reason_unstable_if, 258 Deoptimization::Action_reinterpret, 259 NULL, 260 "taken always"); 261 } else { 262 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 263 profile_switch_case(prof_table_index); 264 merge_new_path(dest_bci_if_true); 265 } 266 } 267 268 // False branch 269 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 270 set_control( iffalse ); 271 } 272 273 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 274 // True branch, use existing map info 275 { PreserveJVMState pjvms(this); 276 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 277 set_control( iffalse ); 278 if (unc) { 279 repush_if_args(); 280 uncommon_trap(Deoptimization::Reason_unstable_if, 281 Deoptimization::Action_reinterpret, 282 NULL, 283 "taken never"); 284 } else { 285 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 286 profile_switch_case(prof_table_index); 287 merge_new_path(dest_bci_if_true); 288 } 289 } 290 291 // False branch 292 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 293 set_control( iftrue ); 294 } 295 296 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index, bool unc) { 297 // False branch, use existing map and control() 298 if (unc) { 299 repush_if_args(); 300 uncommon_trap(Deoptimization::Reason_unstable_if, 301 Deoptimization::Action_reinterpret, 302 NULL, 303 "taken never"); 304 } else { 305 assert(dest_bci != never_reached, "inconsistent dest"); 306 profile_switch_case(prof_table_index); 307 merge_new_path(dest_bci); 308 } 309 } 310 311 312 extern "C" { 313 static int jint_cmp(const void *i, const void *j) { 314 int a = *(jint *)i; 315 int b = *(jint *)j; 316 return a > b ? 1 : a < b ? -1 : 0; 317 } 318 } 319 320 321 // Default value for methodData switch indexing. Must be a negative value to avoid 322 // conflict with any legal switch index. 323 #define NullTableIndex -1 324 325 class SwitchRange : public StackObj { 326 // a range of integers coupled with a bci destination 327 jint _lo; // inclusive lower limit 328 jint _hi; // inclusive upper limit 329 int _dest; 330 int _table_index; // index into method data table 331 float _cnt; // how many times this range was hit according to profiling 332 333 public: 334 jint lo() const { return _lo; } 335 jint hi() const { return _hi; } 336 int dest() const { return _dest; } 337 int table_index() const { return _table_index; } 338 bool is_singleton() const { return _lo == _hi; } 339 float cnt() const { return _cnt; } 340 341 void setRange(jint lo, jint hi, int dest, int table_index, float cnt) { 342 assert(lo <= hi, "must be a non-empty range"); 343 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; _cnt = cnt; 344 assert(_cnt >= 0, ""); 345 } 346 bool adjoinRange(jint lo, jint hi, int dest, int table_index, float cnt, bool trim_ranges) { 347 assert(lo <= hi, "must be a non-empty range"); 348 if (lo == _hi+1 && table_index == _table_index) { 349 // see merge_ranges() comment below 350 if (trim_ranges) { 351 if (cnt == 0) { 352 if (_cnt != 0) { 353 return false; 354 } 355 if (dest != _dest) { 356 _dest = never_reached; 357 } 358 } else { 359 if (_cnt == 0) { 360 return false; 361 } 362 if (dest != _dest) { 363 return false; 364 } 365 } 366 } else { 367 if (dest != _dest) { 368 return false; 369 } 370 } 371 _hi = hi; 372 _cnt += cnt; 373 return true; 374 } 375 return false; 376 } 377 378 void set (jint value, int dest, int table_index, float cnt) { 379 setRange(value, value, dest, table_index, cnt); 380 } 381 bool adjoin(jint value, int dest, int table_index, float cnt, bool trim_ranges) { 382 return adjoinRange(value, value, dest, table_index, cnt, trim_ranges); 383 } 384 bool adjoin(SwitchRange& other) { 385 return adjoinRange(other._lo, other._hi, other._dest, other._table_index, other._cnt, false); 386 } 387 388 void print() { 389 if (is_singleton()) 390 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 391 else if (lo() == min_jint) 392 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 393 else if (hi() == max_jint) 394 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 395 else 396 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 397 } 398 }; 399 400 // We try to minimize the number of ranges and the size of the taken 401 // ones using profiling data. When ranges are created, 402 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 403 // if both were never hit or both were hit to build longer unreached 404 // ranges. Here, we now merge adjoining ranges with the same 405 // destination and finally set destination of unreached ranges to the 406 // special value never_reached because it can help minimize the number 407 // of tests that are necessary. 408 // 409 // For instance: 410 // [0, 1] to target1 sometimes taken 411 // [1, 2] to target1 never taken 412 // [2, 3] to target2 never taken 413 // would lead to: 414 // [0, 1] to target1 sometimes taken 415 // [1, 3] never taken 416 // 417 // (first 2 ranges to target1 are not merged) 418 static void merge_ranges(SwitchRange* ranges, int& rp) { 419 if (rp == 0) { 420 return; 421 } 422 int shift = 0; 423 for (int j = 0; j < rp; j++) { 424 SwitchRange& r1 = ranges[j-shift]; 425 SwitchRange& r2 = ranges[j+1]; 426 if (r1.adjoin(r2)) { 427 shift++; 428 } else if (shift > 0) { 429 ranges[j+1-shift] = r2; 430 } 431 } 432 rp -= shift; 433 for (int j = 0; j <= rp; j++) { 434 SwitchRange& r = ranges[j]; 435 if (r.cnt() == 0 && r.dest() != never_reached) { 436 r.setRange(r.lo(), r.hi(), never_reached, r.table_index(), r.cnt()); 437 } 438 } 439 } 440 441 //-------------------------------do_tableswitch-------------------------------- 442 void Parse::do_tableswitch() { 443 Node* lookup = pop(); 444 // Get information about tableswitch 445 int default_dest = iter().get_dest_table(0); 446 int lo_index = iter().get_int_table(1); 447 int hi_index = iter().get_int_table(2); 448 int len = hi_index - lo_index + 1; 449 450 if (len < 1) { 451 // If this is a backward branch, add safepoint 452 maybe_add_safepoint(default_dest); 453 merge(default_dest); 454 return; 455 } 456 457 ciMethodData* methodData = method()->method_data(); 458 ciMultiBranchData* profile = NULL; 459 if (methodData->is_mature() && UseSwitchProfiling) { 460 ciProfileData* data = methodData->bci_to_data(bci()); 461 if (data != NULL && data->is_MultiBranchData()) { 462 profile = (ciMultiBranchData*)data; 463 } 464 } 465 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 466 467 // generate decision tree, using trichotomy when possible 468 int rnum = len+2; 469 bool makes_backward_branch = false; 470 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 471 int rp = -1; 472 if (lo_index != min_jint) { 473 uint cnt = 1; 474 if (profile != NULL) { 475 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1); 476 } 477 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex, cnt); 478 } 479 for (int j = 0; j < len; j++) { 480 jint match_int = lo_index+j; 481 int dest = iter().get_dest_table(j+3); 482 makes_backward_branch |= (dest <= bci()); 483 int table_index = method_data_update() ? j : NullTableIndex; 484 uint cnt = 1; 485 if (profile != NULL) { 486 cnt = profile->count_at(j); 487 } 488 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 489 ranges[++rp].set(match_int, dest, table_index, cnt); 490 } 491 } 492 jint highest = lo_index+(len-1); 493 assert(ranges[rp].hi() == highest, ""); 494 if (highest != max_jint) { 495 uint cnt = 1; 496 if (profile != NULL) { 497 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1); 498 } 499 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, cnt, trim_ranges)) { 500 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, cnt); 501 } 502 } 503 assert(rp < len+2, "not too many ranges"); 504 505 if (trim_ranges) { 506 merge_ranges(ranges, rp); 507 } 508 509 // Safepoint in case if backward branch observed 510 if( makes_backward_branch && UseLoopSafepoints ) 511 add_safepoint(); 512 513 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 514 } 515 516 517 //------------------------------do_lookupswitch-------------------------------- 518 void Parse::do_lookupswitch() { 519 Node *lookup = pop(); // lookup value 520 // Get information about lookupswitch 521 int default_dest = iter().get_dest_table(0); 522 int len = iter().get_int_table(1); 523 524 if (len < 1) { // If this is a backward branch, add safepoint 525 maybe_add_safepoint(default_dest); 526 merge(default_dest); 527 return; 528 } 529 530 ciMethodData* methodData = method()->method_data(); 531 ciMultiBranchData* profile = NULL; 532 if (methodData->is_mature() && UseSwitchProfiling) { 533 ciProfileData* data = methodData->bci_to_data(bci()); 534 if (data != NULL && data->is_MultiBranchData()) { 535 profile = (ciMultiBranchData*)data; 536 } 537 } 538 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 539 540 // generate decision tree, using trichotomy when possible 541 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 542 { 543 for (int j = 0; j < len; j++) { 544 table[3*j+0] = iter().get_int_table(2+2*j); 545 table[3*j+1] = iter().get_dest_table(2+2*j+1); 546 table[3*j+2] = profile == NULL ? 1 : profile->count_at(j); 547 } 548 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 549 } 550 551 float defaults = 0; 552 jint prev = min_jint; 553 for (int j = 0; j < len; j++) { 554 jint match_int = table[3*j+0]; 555 if (match_int != prev) { 556 defaults += (float)match_int - prev; 557 } 558 prev = match_int+1; 559 } 560 if (prev-1 != max_jint) { 561 defaults += (float)max_jint - prev + 1; 562 } 563 float default_cnt = 1; 564 if (profile != NULL) { 565 default_cnt = profile->default_count()/defaults; 566 } 567 568 int rnum = len*2+1; 569 bool makes_backward_branch = false; 570 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 571 int rp = -1; 572 for (int j = 0; j < len; j++) { 573 jint match_int = table[3*j+0]; 574 int dest = table[3*j+1]; 575 int cnt = table[3*j+2]; 576 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 577 int table_index = method_data_update() ? j : NullTableIndex; 578 makes_backward_branch |= (dest <= bci()); 579 float c = default_cnt * ((float)match_int - next_lo); 580 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, NullTableIndex, c, trim_ranges))) { 581 assert(default_dest != never_reached, "sentinel value for dead destinations"); 582 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex, c); 583 } 584 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 585 assert(dest != never_reached, "sentinel value for dead destinations"); 586 ranges[++rp].set(match_int, dest, table_index, cnt); 587 } 588 } 589 jint highest = table[3*(len-1)]; 590 assert(ranges[rp].hi() == highest, ""); 591 if (highest != max_jint && 592 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest), trim_ranges)) { 593 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest)); 594 } 595 assert(rp < rnum, "not too many ranges"); 596 597 if (trim_ranges) { 598 merge_ranges(ranges, rp); 599 } 600 601 // Safepoint in case backward branch observed 602 if (makes_backward_branch && UseLoopSafepoints) 603 add_safepoint(); 604 605 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 606 } 607 608 static float if_prob(float taken_cnt, float total_cnt) { 609 assert(taken_cnt <= total_cnt, ""); 610 if (total_cnt == 0) { 611 return PROB_FAIR; 612 } 613 float p = taken_cnt / total_cnt; 614 return MIN2(MAX2(p, PROB_MIN), PROB_MAX); 615 } 616 617 static float if_cnt(float cnt) { 618 if (cnt == 0) { 619 return COUNT_UNKNOWN; 620 } 621 return cnt; 622 } 623 624 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 625 float total_cnt = 0; 626 for (SwitchRange* sr = lo; sr <= hi; sr++) { 627 total_cnt += sr->cnt(); 628 } 629 return total_cnt; 630 } 631 632 class SwitchRanges : public ResourceObj { 633 public: 634 SwitchRange* _lo; 635 SwitchRange* _hi; 636 SwitchRange* _mid; 637 float _cost; 638 639 enum { 640 Start, 641 LeftDone, 642 RightDone, 643 Done 644 } _state; 645 646 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 647 : _lo(lo), _hi(hi), _mid(NULL), 648 _cost(0), _state(Start) { 649 } 650 651 SwitchRanges() 652 : _lo(NULL), _hi(NULL), _mid(NULL), 653 _cost(0), _state(Start) {} 654 }; 655 656 // Estimate cost of performing a binary search on lo..hi 657 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 658 GrowableArray<SwitchRanges> tree; 659 SwitchRanges root(lo, hi); 660 tree.push(root); 661 662 float cost = 0; 663 do { 664 SwitchRanges& r = *tree.adr_at(tree.length()-1); 665 if (r._hi != r._lo) { 666 if (r._mid == NULL) { 667 float r_cnt = sum_of_cnts(r._lo, r._hi); 668 669 if (r_cnt == 0) { 670 tree.pop(); 671 cost = 0; 672 continue; 673 } 674 675 SwitchRange* mid = NULL; 676 mid = r._lo; 677 for (float cnt = 0; ; ) { 678 assert(mid <= r._hi, "out of bounds"); 679 cnt += mid->cnt(); 680 if (cnt > r_cnt / 2) { 681 break; 682 } 683 mid++; 684 } 685 assert(mid <= r._hi, "out of bounds"); 686 r._mid = mid; 687 r._cost = r_cnt / total_cnt; 688 } 689 r._cost += cost; 690 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 691 cost = 0; 692 r._state = SwitchRanges::LeftDone; 693 tree.push(SwitchRanges(r._lo, r._mid-1)); 694 } else if (r._state < SwitchRanges::RightDone) { 695 cost = 0; 696 r._state = SwitchRanges::RightDone; 697 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 698 } else { 699 tree.pop(); 700 cost = r._cost; 701 } 702 } else { 703 tree.pop(); 704 cost = r._cost; 705 } 706 } while (tree.length() > 0); 707 708 709 return cost; 710 } 711 712 // It sometimes pays off to test most common ranges before the binary search 713 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 714 uint nr = hi - lo + 1; 715 float total_cnt = sum_of_cnts(lo, hi); 716 717 float min = compute_tree_cost(lo, hi, total_cnt); 718 float extra = 1; 719 float sub = 0; 720 721 SwitchRange* array1 = lo; 722 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 723 724 SwitchRange* ranges = NULL; 725 726 while (nr >= 2) { 727 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 728 ranges = (lo == array1) ? array2 : array1; 729 730 // Find highest frequency range 731 SwitchRange* candidate = lo; 732 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 733 if (sr->cnt() > candidate->cnt()) { 734 candidate = sr; 735 } 736 } 737 SwitchRange most_freq = *candidate; 738 if (most_freq.cnt() == 0) { 739 break; 740 } 741 742 // Copy remaining ranges into another array 743 int shift = 0; 744 for (uint i = 0; i < nr; i++) { 745 SwitchRange* sr = &lo[i]; 746 if (sr != candidate) { 747 ranges[i-shift] = *sr; 748 } else { 749 shift++; 750 if (i > 0 && i < nr-1) { 751 SwitchRange prev = lo[i-1]; 752 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.table_index(), prev.cnt()); 753 if (prev.adjoin(lo[i+1])) { 754 shift++; 755 i++; 756 } 757 ranges[i-shift] = prev; 758 } 759 } 760 } 761 nr -= shift; 762 763 // Evaluate cost of testing the most common range and performing a 764 // binary search on the other ranges 765 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 766 if (cost >= min) { 767 break; 768 } 769 // swap arrays 770 lo = &ranges[0]; 771 hi = &ranges[nr-1]; 772 773 // It pays off: emit the test for the most common range 774 assert(most_freq.cnt() > 0, "must be taken"); 775 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 776 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo()))); 777 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 778 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 779 jump_if_true_fork(iff, most_freq.dest(), most_freq.table_index(), false); 780 781 sub += most_freq.cnt() / total_cnt; 782 extra += 1 - sub; 783 min = cost; 784 } 785 } 786 787 //----------------------------create_jump_tables------------------------------- 788 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 789 // Are jumptables enabled 790 if (!UseJumpTables) return false; 791 792 // Are jumptables supported 793 if (!Matcher::has_match_rule(Op_Jump)) return false; 794 795 // Don't make jump table if profiling 796 if (method_data_update()) return false; 797 798 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 799 800 // Decide if a guard is needed to lop off big ranges at either (or 801 // both) end(s) of the input set. We'll call this the default target 802 // even though we can't be sure that it is the true "default". 803 804 bool needs_guard = false; 805 int default_dest; 806 int64_t total_outlier_size = 0; 807 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 808 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 809 810 if (lo->dest() == hi->dest()) { 811 total_outlier_size = hi_size + lo_size; 812 default_dest = lo->dest(); 813 } else if (lo_size > hi_size) { 814 total_outlier_size = lo_size; 815 default_dest = lo->dest(); 816 } else { 817 total_outlier_size = hi_size; 818 default_dest = hi->dest(); 819 } 820 821 float total = sum_of_cnts(lo, hi); 822 float cost = compute_tree_cost(lo, hi, total); 823 824 // If a guard test will eliminate very sparse end ranges, then 825 // it is worth the cost of an extra jump. 826 float trimmed_cnt = 0; 827 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 828 needs_guard = true; 829 if (default_dest == lo->dest()) { 830 trimmed_cnt += lo->cnt(); 831 lo++; 832 } 833 if (default_dest == hi->dest()) { 834 trimmed_cnt += hi->cnt(); 835 hi--; 836 } 837 } 838 839 // Find the total number of cases and ranges 840 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 841 int num_range = hi - lo + 1; 842 843 // Don't create table if: too large, too small, or too sparse. 844 if (num_cases > MaxJumpTableSize) 845 return false; 846 if (UseSwitchProfiling) { 847 // MinJumpTableSize is set so with a well balanced binary tree, 848 // when the number of ranges is MinJumpTableSize, it's cheaper to 849 // go through a JumpNode that a tree of IfNodes. Average cost of a 850 // tree of IfNodes with MinJumpTableSize is 851 // log2f(MinJumpTableSize) comparisons. So if the cost computed 852 // from profile data is less than log2f(MinJumpTableSize) then 853 // going with the binary search is cheaper. 854 if (cost < log2f(MinJumpTableSize)) { 855 return false; 856 } 857 } else { 858 if (num_cases < MinJumpTableSize) 859 return false; 860 } 861 if (num_cases > (MaxJumpTableSparseness * num_range)) 862 return false; 863 864 // Normalize table lookups to zero 865 int lowval = lo->lo(); 866 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 867 868 // Generate a guard to protect against input keyvals that aren't 869 // in the switch domain. 870 if (needs_guard) { 871 Node* size = _gvn.intcon(num_cases); 872 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 873 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 874 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 875 jump_if_true_fork(iff, default_dest, NullTableIndex, trim_ranges && trimmed_cnt == 0); 876 877 total -= trimmed_cnt; 878 } 879 880 // Create an ideal node JumpTable that has projections 881 // of all possible ranges for a switch statement 882 // The key_val input must be converted to a pointer offset and scaled. 883 // Compare Parse::array_addressing above. 884 885 // Clean the 32-bit int into a real 64-bit offset. 886 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 887 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 888 // Make I2L conversion control dependent to prevent it from 889 // floating above the range check during loop optimizations. 890 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 891 892 // Shift the value by wordsize so we have an index into the table, rather 893 // than a switch value 894 Node *shiftWord = _gvn.MakeConX(wordSize); 895 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 896 897 // Create the JumpNode 898 Arena* arena = C->comp_arena(); 899 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 900 int i = 0; 901 if (total == 0) { 902 for (SwitchRange* r = lo; r <= hi; r++) { 903 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 904 probs[i] = 1.0F / num_cases; 905 } 906 } 907 } else { 908 for (SwitchRange* r = lo; r <= hi; r++) { 909 float prob = r->cnt()/total; 910 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 911 probs[i] = prob / (r->hi() - r->lo() + 1); 912 } 913 } 914 } 915 916 ciMethodData* methodData = method()->method_data(); 917 ciMultiBranchData* profile = NULL; 918 if (methodData->is_mature()) { 919 ciProfileData* data = methodData->bci_to_data(bci()); 920 if (data != NULL && data->is_MultiBranchData()) { 921 profile = (ciMultiBranchData*)data; 922 } 923 } 924 925 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total)); 926 927 // These are the switch destinations hanging off the jumpnode 928 i = 0; 929 for (SwitchRange* r = lo; r <= hi; r++) { 930 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 931 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 932 { 933 PreserveJVMState pjvms(this); 934 set_control(input); 935 jump_if_always_fork(r->dest(), r->table_index(), trim_ranges && r->cnt() == 0); 936 } 937 } 938 } 939 assert(i == num_cases, "miscount of cases"); 940 stop_and_kill_map(); // no more uses for this JVMS 941 return true; 942 } 943 944 //----------------------------jump_switch_ranges------------------------------- 945 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 946 Block* switch_block = block(); 947 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 948 949 if (switch_depth == 0) { 950 // Do special processing for the top-level call. 951 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 952 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 953 954 // Decrement pred-numbers for the unique set of nodes. 955 #ifdef ASSERT 956 if (!trim_ranges) { 957 // Ensure that the block's successors are a (duplicate-free) set. 958 int successors_counted = 0; // block occurrences in [hi..lo] 959 int unique_successors = switch_block->num_successors(); 960 for (int i = 0; i < unique_successors; i++) { 961 Block* target = switch_block->successor_at(i); 962 963 // Check that the set of successors is the same in both places. 964 int successors_found = 0; 965 for (SwitchRange* p = lo; p <= hi; p++) { 966 if (p->dest() == target->start()) successors_found++; 967 } 968 assert(successors_found > 0, "successor must be known"); 969 successors_counted += successors_found; 970 } 971 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 972 } 973 #endif 974 975 // Maybe prune the inputs, based on the type of key_val. 976 jint min_val = min_jint; 977 jint max_val = max_jint; 978 const TypeInt* ti = key_val->bottom_type()->isa_int(); 979 if (ti != NULL) { 980 min_val = ti->_lo; 981 max_val = ti->_hi; 982 assert(min_val <= max_val, "invalid int type"); 983 } 984 while (lo->hi() < min_val) { 985 lo++; 986 } 987 if (lo->lo() < min_val) { 988 lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index(), lo->cnt()); 989 } 990 while (hi->lo() > max_val) { 991 hi--; 992 } 993 if (hi->hi() > max_val) { 994 hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index(), hi->cnt()); 995 } 996 997 linear_search_switch_ranges(key_val, lo, hi); 998 } 999 1000 #ifndef PRODUCT 1001 if (switch_depth == 0) { 1002 _max_switch_depth = 0; 1003 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 1004 } 1005 #endif 1006 1007 assert(lo <= hi, "must be a non-empty set of ranges"); 1008 if (lo == hi) { 1009 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1010 } else { 1011 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 1012 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 1013 1014 if (create_jump_tables(key_val, lo, hi)) return; 1015 1016 SwitchRange* mid = NULL; 1017 float total_cnt = sum_of_cnts(lo, hi); 1018 1019 int nr = hi - lo + 1; 1020 if (UseSwitchProfiling) { 1021 // Don't keep the binary search tree balanced: pick up mid point 1022 // that split frequencies in half. 1023 float cnt = 0; 1024 for (SwitchRange* sr = lo; sr <= hi; sr++) { 1025 cnt += sr->cnt(); 1026 if (cnt >= total_cnt / 2) { 1027 mid = sr; 1028 break; 1029 } 1030 } 1031 } else { 1032 mid = lo + nr/2; 1033 1034 // if there is an easy choice, pivot at a singleton: 1035 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1036 1037 assert(lo < mid && mid <= hi, "good pivot choice"); 1038 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1039 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1040 } 1041 1042 1043 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1044 1045 if (mid->is_singleton()) { 1046 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1047 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index(), trim_ranges && mid->cnt() == 0); 1048 1049 // Special Case: If there are exactly three ranges, and the high 1050 // and low range each go to the same place, omit the "gt" test, 1051 // since it will not discriminate anything. 1052 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1053 1054 // if there is a higher range, test for it and process it: 1055 if (mid < hi && !eq_test_only) { 1056 // two comparisons of same values--should enable 1 test for 2 branches 1057 // Use BoolTest::le instead of BoolTest::gt 1058 float cnt = sum_of_cnts(lo, mid-1); 1059 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le, if_prob(cnt, total_cnt), if_cnt(cnt)); 1060 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 1061 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 1062 { PreserveJVMState pjvms(this); 1063 set_control(iffalse); 1064 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1065 } 1066 set_control(iftrue); 1067 } 1068 1069 } else { 1070 // mid is a range, not a singleton, so treat mid..hi as a unit 1071 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1072 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, mid == lo ? BoolTest::gt : BoolTest::ge, if_prob(cnt, total_cnt), if_cnt(cnt)); 1073 1074 // if there is a higher range, test for it and process it: 1075 if (mid == hi) { 1076 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index(), trim_ranges && cnt == 0); 1077 } else { 1078 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1079 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1080 { PreserveJVMState pjvms(this); 1081 set_control(iftrue); 1082 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1083 } 1084 set_control(iffalse); 1085 } 1086 } 1087 1088 // in any case, process the lower range 1089 if (mid == lo) { 1090 if (mid->is_singleton()) { 1091 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1092 } else { 1093 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1094 } 1095 } else { 1096 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1097 } 1098 } 1099 1100 // Decrease pred_count for each successor after all is done. 1101 if (switch_depth == 0) { 1102 int unique_successors = switch_block->num_successors(); 1103 for (int i = 0; i < unique_successors; i++) { 1104 Block* target = switch_block->successor_at(i); 1105 // Throw away the pre-allocated path for each unique successor. 1106 target->next_path_num(); 1107 } 1108 } 1109 1110 #ifndef PRODUCT 1111 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1112 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1113 SwitchRange* r; 1114 int nsing = 0; 1115 for( r = lo; r <= hi; r++ ) { 1116 if( r->is_singleton() ) nsing++; 1117 } 1118 tty->print(">>> "); 1119 _method->print_short_name(); 1120 tty->print_cr(" switch decision tree"); 1121 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1122 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1123 if (_max_switch_depth > _est_switch_depth) { 1124 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1125 } 1126 tty->print(" "); 1127 for( r = lo; r <= hi; r++ ) { 1128 r->print(); 1129 } 1130 tty->cr(); 1131 } 1132 #endif 1133 } 1134 1135 void Parse::modf() { 1136 Node *f2 = pop(); 1137 Node *f1 = pop(); 1138 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1139 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1140 "frem", NULL, //no memory effects 1141 f1, f2); 1142 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1143 1144 push(res); 1145 } 1146 1147 void Parse::modd() { 1148 Node *d2 = pop_pair(); 1149 Node *d1 = pop_pair(); 1150 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1151 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1152 "drem", NULL, //no memory effects 1153 d1, top(), d2, top()); 1154 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1155 1156 #ifdef ASSERT 1157 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 1158 assert(res_top == top(), "second value must be top"); 1159 #endif 1160 1161 push_pair(res_d); 1162 } 1163 1164 void Parse::l2f() { 1165 Node* f2 = pop(); 1166 Node* f1 = pop(); 1167 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1168 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1169 "l2f", NULL, //no memory effects 1170 f1, f2); 1171 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1172 1173 push(res); 1174 } 1175 1176 void Parse::do_irem() { 1177 // Must keep both values on the expression-stack during null-check 1178 zero_check_int(peek()); 1179 // Compile-time detect of null-exception? 1180 if (stopped()) return; 1181 1182 Node* b = pop(); 1183 Node* a = pop(); 1184 1185 const Type *t = _gvn.type(b); 1186 if (t != Type::TOP) { 1187 const TypeInt *ti = t->is_int(); 1188 if (ti->is_con()) { 1189 int divisor = ti->get_con(); 1190 // check for positive power of 2 1191 if (divisor > 0 && 1192 (divisor & ~(divisor-1)) == divisor) { 1193 // yes ! 1194 Node *mask = _gvn.intcon((divisor - 1)); 1195 // Sigh, must handle negative dividends 1196 Node *zero = _gvn.intcon(0); 1197 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN); 1198 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 1199 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 1200 Node *reg = jump_if_join(ift, iff); 1201 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 1202 // Negative path; negate/and/negate 1203 Node *neg = _gvn.transform( new SubINode(zero, a) ); 1204 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 1205 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 1206 phi->init_req(1, negn); 1207 // Fast positive case 1208 Node *andx = _gvn.transform( new AndINode(a, mask) ); 1209 phi->init_req(2, andx); 1210 // Push the merge 1211 push( _gvn.transform(phi) ); 1212 return; 1213 } 1214 } 1215 } 1216 // Default case 1217 push( _gvn.transform( new ModINode(control(),a,b) ) ); 1218 } 1219 1220 // Handle jsr and jsr_w bytecode 1221 void Parse::do_jsr() { 1222 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1223 1224 // Store information about current state, tagged with new _jsr_bci 1225 int return_bci = iter().next_bci(); 1226 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1227 1228 // Update method data 1229 profile_taken_branch(jsr_bci); 1230 1231 // The way we do things now, there is only one successor block 1232 // for the jsr, because the target code is cloned by ciTypeFlow. 1233 Block* target = successor_for_bci(jsr_bci); 1234 1235 // What got pushed? 1236 const Type* ret_addr = target->peek(); 1237 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1238 1239 // Effect on jsr on stack 1240 push(_gvn.makecon(ret_addr)); 1241 1242 // Flow to the jsr. 1243 merge(jsr_bci); 1244 } 1245 1246 // Handle ret bytecode 1247 void Parse::do_ret() { 1248 // Find to whom we return. 1249 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1250 Block* target = block()->successor_at(0); 1251 assert(!target->is_ready(), "our arrival must be expected"); 1252 profile_ret(target->flow()->start()); 1253 int pnum = target->next_path_num(); 1254 merge_common(target, pnum); 1255 } 1256 1257 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1258 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1259 // Only ::eq and ::ne are supported for profile injection. 1260 return false; 1261 } 1262 if (test->is_Cmp() && 1263 test->in(1)->Opcode() == Op_ProfileBoolean) { 1264 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1265 int false_cnt = profile->false_count(); 1266 int true_cnt = profile->true_count(); 1267 1268 // Counts matching depends on the actual test operation (::eq or ::ne). 1269 // No need to scale the counts because profile injection was designed 1270 // to feed exact counts into VM. 1271 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1272 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1273 1274 profile->consume(); 1275 return true; 1276 } 1277 return false; 1278 } 1279 //--------------------------dynamic_branch_prediction-------------------------- 1280 // Try to gather dynamic branch prediction behavior. Return a probability 1281 // of the branch being taken and set the "cnt" field. Returns a -1.0 1282 // if we need to use static prediction for some reason. 1283 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1284 ResourceMark rm; 1285 1286 cnt = COUNT_UNKNOWN; 1287 1288 int taken = 0; 1289 int not_taken = 0; 1290 1291 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1292 1293 if (use_mdo) { 1294 // Use MethodData information if it is available 1295 // FIXME: free the ProfileData structure 1296 ciMethodData* methodData = method()->method_data(); 1297 if (!methodData->is_mature()) return PROB_UNKNOWN; 1298 ciProfileData* data = methodData->bci_to_data(bci()); 1299 if (data == NULL) { 1300 return PROB_UNKNOWN; 1301 } 1302 if (!data->is_JumpData()) return PROB_UNKNOWN; 1303 1304 // get taken and not taken values 1305 taken = data->as_JumpData()->taken(); 1306 not_taken = 0; 1307 if (data->is_BranchData()) { 1308 not_taken = data->as_BranchData()->not_taken(); 1309 } 1310 1311 // scale the counts to be commensurate with invocation counts: 1312 taken = method()->scale_count(taken); 1313 not_taken = method()->scale_count(not_taken); 1314 } 1315 1316 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1317 // We also check that individual counters are positive first, otherwise the sum can become positive. 1318 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 1319 if (C->log() != NULL) { 1320 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1321 } 1322 return PROB_UNKNOWN; 1323 } 1324 1325 // Compute frequency that we arrive here 1326 float sum = taken + not_taken; 1327 // Adjust, if this block is a cloned private block but the 1328 // Jump counts are shared. Taken the private counts for 1329 // just this path instead of the shared counts. 1330 if( block()->count() > 0 ) 1331 sum = block()->count(); 1332 cnt = sum / FreqCountInvocations; 1333 1334 // Pin probability to sane limits 1335 float prob; 1336 if( !taken ) 1337 prob = (0+PROB_MIN) / 2; 1338 else if( !not_taken ) 1339 prob = (1+PROB_MAX) / 2; 1340 else { // Compute probability of true path 1341 prob = (float)taken / (float)(taken + not_taken); 1342 if (prob > PROB_MAX) prob = PROB_MAX; 1343 if (prob < PROB_MIN) prob = PROB_MIN; 1344 } 1345 1346 assert((cnt > 0.0f) && (prob > 0.0f), 1347 "Bad frequency assignment in if"); 1348 1349 if (C->log() != NULL) { 1350 const char* prob_str = NULL; 1351 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1352 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1353 char prob_str_buf[30]; 1354 if (prob_str == NULL) { 1355 sprintf(prob_str_buf, "%g", prob); 1356 prob_str = prob_str_buf; 1357 } 1358 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1359 iter().get_dest(), taken, not_taken, cnt, prob_str); 1360 } 1361 return prob; 1362 } 1363 1364 //-----------------------------branch_prediction------------------------------- 1365 float Parse::branch_prediction(float& cnt, 1366 BoolTest::mask btest, 1367 int target_bci, 1368 Node* test) { 1369 float prob = dynamic_branch_prediction(cnt, btest, test); 1370 // If prob is unknown, switch to static prediction 1371 if (prob != PROB_UNKNOWN) return prob; 1372 1373 prob = PROB_FAIR; // Set default value 1374 if (btest == BoolTest::eq) // Exactly equal test? 1375 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1376 else if (btest == BoolTest::ne) 1377 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1378 1379 // If this is a conditional test guarding a backwards branch, 1380 // assume its a loop-back edge. Make it a likely taken branch. 1381 if (target_bci < bci()) { 1382 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1383 // Since it's an OSR, we probably have profile data, but since 1384 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1385 // Let's make a special check here for completely zero counts. 1386 ciMethodData* methodData = method()->method_data(); 1387 if (!methodData->is_empty()) { 1388 ciProfileData* data = methodData->bci_to_data(bci()); 1389 // Only stop for truly zero counts, which mean an unknown part 1390 // of the OSR-ed method, and we want to deopt to gather more stats. 1391 // If you have ANY counts, then this loop is simply 'cold' relative 1392 // to the OSR loop. 1393 if (data == NULL || 1394 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1395 // This is the only way to return PROB_UNKNOWN: 1396 return PROB_UNKNOWN; 1397 } 1398 } 1399 } 1400 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1401 } 1402 1403 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1404 return prob; 1405 } 1406 1407 // The magic constants are chosen so as to match the output of 1408 // branch_prediction() when the profile reports a zero taken count. 1409 // It is important to distinguish zero counts unambiguously, because 1410 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1411 // very small but nonzero probabilities, which if confused with zero 1412 // counts would keep the program recompiling indefinitely. 1413 bool Parse::seems_never_taken(float prob) const { 1414 return prob < PROB_MIN; 1415 } 1416 1417 // True if the comparison seems to be the kind that will not change its 1418 // statistics from true to false. See comments in adjust_map_after_if. 1419 // This question is only asked along paths which are already 1420 // classifed as untaken (by seems_never_taken), so really, 1421 // if a path is never taken, its controlling comparison is 1422 // already acting in a stable fashion. If the comparison 1423 // seems stable, we will put an expensive uncommon trap 1424 // on the untaken path. 1425 bool Parse::seems_stable_comparison() const { 1426 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 1427 return false; 1428 } 1429 return true; 1430 } 1431 1432 //-------------------------------repush_if_args-------------------------------- 1433 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1434 inline int Parse::repush_if_args() { 1435 if (PrintOpto && WizardMode) { 1436 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1437 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1438 method()->print_name(); tty->cr(); 1439 } 1440 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1441 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1442 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1443 assert(argument(0) != NULL, "must exist"); 1444 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 1445 inc_sp(bc_depth); 1446 return bc_depth; 1447 } 1448 1449 //----------------------------------do_ifnull---------------------------------- 1450 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1451 int target_bci = iter().get_dest(); 1452 1453 Block* branch_block = successor_for_bci(target_bci); 1454 Block* next_block = successor_for_bci(iter().next_bci()); 1455 1456 float cnt; 1457 float prob = branch_prediction(cnt, btest, target_bci, c); 1458 if (prob == PROB_UNKNOWN) { 1459 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1460 if (PrintOpto && Verbose) { 1461 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1462 } 1463 repush_if_args(); // to gather stats on loop 1464 // We need to mark this branch as taken so that if we recompile we will 1465 // see that it is possible. In the tiered system the interpreter doesn't 1466 // do profiling and by the time we get to the lower tier from the interpreter 1467 // the path may be cold again. Make sure it doesn't look untaken 1468 profile_taken_branch(target_bci, !ProfileInterpreter); 1469 uncommon_trap(Deoptimization::Reason_unreached, 1470 Deoptimization::Action_reinterpret, 1471 NULL, "cold"); 1472 if (C->eliminate_boxing()) { 1473 // Mark the successor blocks as parsed 1474 branch_block->next_path_num(); 1475 next_block->next_path_num(); 1476 } 1477 return; 1478 } 1479 1480 NOT_PRODUCT(explicit_null_checks_inserted++); 1481 1482 // Generate real control flow 1483 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1484 1485 // Sanity check the probability value 1486 assert(prob > 0.0f,"Bad probability in Parser"); 1487 // Need xform to put node in hash table 1488 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1489 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1490 // True branch 1491 { PreserveJVMState pjvms(this); 1492 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1493 set_control(iftrue); 1494 1495 if (stopped()) { // Path is dead? 1496 NOT_PRODUCT(explicit_null_checks_elided++); 1497 if (C->eliminate_boxing()) { 1498 // Mark the successor block as parsed 1499 branch_block->next_path_num(); 1500 } 1501 } else { // Path is live. 1502 // Update method data 1503 profile_taken_branch(target_bci); 1504 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1505 if (!stopped()) { 1506 merge(target_bci); 1507 } 1508 } 1509 } 1510 1511 // False branch 1512 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1513 set_control(iffalse); 1514 1515 if (stopped()) { // Path is dead? 1516 NOT_PRODUCT(explicit_null_checks_elided++); 1517 if (C->eliminate_boxing()) { 1518 // Mark the successor block as parsed 1519 next_block->next_path_num(); 1520 } 1521 } else { // Path is live. 1522 // Update method data 1523 profile_not_taken_branch(); 1524 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1525 next_block, branch_block); 1526 } 1527 } 1528 1529 //------------------------------------do_if------------------------------------ 1530 void Parse::do_if(BoolTest::mask btest, Node* c) { 1531 int target_bci = iter().get_dest(); 1532 1533 Block* branch_block = successor_for_bci(target_bci); 1534 Block* next_block = successor_for_bci(iter().next_bci()); 1535 1536 float cnt; 1537 float prob = branch_prediction(cnt, btest, target_bci, c); 1538 float untaken_prob = 1.0 - prob; 1539 1540 if (prob == PROB_UNKNOWN) { 1541 if (PrintOpto && Verbose) { 1542 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1543 } 1544 repush_if_args(); // to gather stats on loop 1545 // We need to mark this branch as taken so that if we recompile we will 1546 // see that it is possible. In the tiered system the interpreter doesn't 1547 // do profiling and by the time we get to the lower tier from the interpreter 1548 // the path may be cold again. Make sure it doesn't look untaken 1549 profile_taken_branch(target_bci, !ProfileInterpreter); 1550 uncommon_trap(Deoptimization::Reason_unreached, 1551 Deoptimization::Action_reinterpret, 1552 NULL, "cold"); 1553 if (C->eliminate_boxing()) { 1554 // Mark the successor blocks as parsed 1555 branch_block->next_path_num(); 1556 next_block->next_path_num(); 1557 } 1558 return; 1559 } 1560 1561 // Sanity check the probability value 1562 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1563 1564 bool taken_if_true = true; 1565 // Convert BoolTest to canonical form: 1566 if (!BoolTest(btest).is_canonical()) { 1567 btest = BoolTest(btest).negate(); 1568 taken_if_true = false; 1569 // prob is NOT updated here; it remains the probability of the taken 1570 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1571 } 1572 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1573 1574 Node* tst0 = new BoolNode(c, btest); 1575 Node* tst = _gvn.transform(tst0); 1576 BoolTest::mask taken_btest = BoolTest::illegal; 1577 BoolTest::mask untaken_btest = BoolTest::illegal; 1578 1579 if (tst->is_Bool()) { 1580 // Refresh c from the transformed bool node, since it may be 1581 // simpler than the original c. Also re-canonicalize btest. 1582 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1583 // That can arise from statements like: if (x instanceof C) ... 1584 if (tst != tst0) { 1585 // Canonicalize one more time since transform can change it. 1586 btest = tst->as_Bool()->_test._test; 1587 if (!BoolTest(btest).is_canonical()) { 1588 // Reverse edges one more time... 1589 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1590 btest = tst->as_Bool()->_test._test; 1591 assert(BoolTest(btest).is_canonical(), "sanity"); 1592 taken_if_true = !taken_if_true; 1593 } 1594 c = tst->in(1); 1595 } 1596 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1597 taken_btest = taken_if_true ? btest : neg_btest; 1598 untaken_btest = taken_if_true ? neg_btest : btest; 1599 } 1600 1601 // Generate real control flow 1602 float true_prob = (taken_if_true ? prob : untaken_prob); 1603 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1604 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1605 Node* taken_branch = new IfTrueNode(iff); 1606 Node* untaken_branch = new IfFalseNode(iff); 1607 if (!taken_if_true) { // Finish conversion to canonical form 1608 Node* tmp = taken_branch; 1609 taken_branch = untaken_branch; 1610 untaken_branch = tmp; 1611 } 1612 1613 // Branch is taken: 1614 { PreserveJVMState pjvms(this); 1615 taken_branch = _gvn.transform(taken_branch); 1616 set_control(taken_branch); 1617 1618 if (stopped()) { 1619 if (C->eliminate_boxing()) { 1620 // Mark the successor block as parsed 1621 branch_block->next_path_num(); 1622 } 1623 } else { 1624 // Update method data 1625 profile_taken_branch(target_bci); 1626 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1627 if (!stopped()) { 1628 merge(target_bci); 1629 } 1630 } 1631 } 1632 1633 untaken_branch = _gvn.transform(untaken_branch); 1634 set_control(untaken_branch); 1635 1636 // Branch not taken. 1637 if (stopped()) { 1638 if (C->eliminate_boxing()) { 1639 // Mark the successor block as parsed 1640 next_block->next_path_num(); 1641 } 1642 } else { 1643 // Update method data 1644 profile_not_taken_branch(); 1645 adjust_map_after_if(untaken_btest, c, untaken_prob, 1646 next_block, branch_block); 1647 } 1648 } 1649 1650 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1651 // Don't want to speculate on uncommon traps when running with -Xcomp 1652 if (!UseInterpreter) { 1653 return false; 1654 } 1655 return (seems_never_taken(prob) && seems_stable_comparison()); 1656 } 1657 1658 void Parse::maybe_add_predicate_after_if(Block* path) { 1659 if (path->is_SEL_head() && path->preds_parsed() == 0) { 1660 // Add predicates at bci of if dominating the loop so traps can be 1661 // recorded on the if's profile data 1662 int bc_depth = repush_if_args(); 1663 add_predicate(); 1664 dec_sp(bc_depth); 1665 path->set_has_predicates(); 1666 } 1667 } 1668 1669 1670 //----------------------------adjust_map_after_if------------------------------ 1671 // Adjust the JVM state to reflect the result of taking this path. 1672 // Basically, it means inspecting the CmpNode controlling this 1673 // branch, seeing how it constrains a tested value, and then 1674 // deciding if it's worth our while to encode this constraint 1675 // as graph nodes in the current abstract interpretation map. 1676 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1677 Block* path, Block* other_path) { 1678 if (!c->is_Cmp()) { 1679 maybe_add_predicate_after_if(path); 1680 return; 1681 } 1682 1683 if (stopped() || btest == BoolTest::illegal) { 1684 return; // nothing to do 1685 } 1686 1687 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1688 1689 if (path_is_suitable_for_uncommon_trap(prob)) { 1690 repush_if_args(); 1691 uncommon_trap(Deoptimization::Reason_unstable_if, 1692 Deoptimization::Action_reinterpret, 1693 NULL, 1694 (is_fallthrough ? "taken always" : "taken never")); 1695 return; 1696 } 1697 1698 Node* val = c->in(1); 1699 Node* con = c->in(2); 1700 const Type* tcon = _gvn.type(con); 1701 const Type* tval = _gvn.type(val); 1702 bool have_con = tcon->singleton(); 1703 if (tval->singleton()) { 1704 if (!have_con) { 1705 // Swap, so constant is in con. 1706 con = val; 1707 tcon = tval; 1708 val = c->in(2); 1709 tval = _gvn.type(val); 1710 btest = BoolTest(btest).commute(); 1711 have_con = true; 1712 } else { 1713 // Do we have two constants? Then leave well enough alone. 1714 have_con = false; 1715 } 1716 } 1717 if (!have_con) { // remaining adjustments need a con 1718 maybe_add_predicate_after_if(path); 1719 return; 1720 } 1721 1722 sharpen_type_after_if(btest, con, tcon, val, tval); 1723 maybe_add_predicate_after_if(path); 1724 } 1725 1726 1727 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1728 Node* ldk; 1729 if (n->is_DecodeNKlass()) { 1730 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1731 return NULL; 1732 } else { 1733 ldk = n->in(1); 1734 } 1735 } else if (n->Opcode() != Op_LoadKlass) { 1736 return NULL; 1737 } else { 1738 ldk = n; 1739 } 1740 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1741 1742 Node* adr = ldk->in(MemNode::Address); 1743 intptr_t off = 0; 1744 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1745 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1746 return NULL; 1747 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1748 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1749 return NULL; 1750 1751 return obj; 1752 } 1753 1754 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1755 Node* con, const Type* tcon, 1756 Node* val, const Type* tval) { 1757 // Look for opportunities to sharpen the type of a node 1758 // whose klass is compared with a constant klass. 1759 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1760 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1761 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1762 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1763 // Found: 1764 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1765 // or the narrowOop equivalent. 1766 const Type* obj_type = _gvn.type(obj); 1767 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1768 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1769 tboth->higher_equal(obj_type)) { 1770 // obj has to be of the exact type Foo if the CmpP succeeds. 1771 int obj_in_map = map()->find_edge(obj); 1772 JVMState* jvms = this->jvms(); 1773 if (obj_in_map >= 0 && 1774 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1775 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1776 const Type* tcc = ccast->as_Type()->type(); 1777 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1778 // Delay transform() call to allow recovery of pre-cast value 1779 // at the control merge. 1780 _gvn.set_type_bottom(ccast); 1781 record_for_igvn(ccast); 1782 // Here's the payoff. 1783 replace_in_map(obj, ccast); 1784 } 1785 } 1786 } 1787 } 1788 1789 int val_in_map = map()->find_edge(val); 1790 if (val_in_map < 0) return; // replace_in_map would be useless 1791 { 1792 JVMState* jvms = this->jvms(); 1793 if (!(jvms->is_loc(val_in_map) || 1794 jvms->is_stk(val_in_map))) 1795 return; // again, it would be useless 1796 } 1797 1798 // Check for a comparison to a constant, and "know" that the compared 1799 // value is constrained on this path. 1800 assert(tcon->singleton(), ""); 1801 ConstraintCastNode* ccast = NULL; 1802 Node* cast = NULL; 1803 1804 switch (btest) { 1805 case BoolTest::eq: // Constant test? 1806 { 1807 const Type* tboth = tcon->join_speculative(tval); 1808 if (tboth == tval) break; // Nothing to gain. 1809 if (tcon->isa_int()) { 1810 ccast = new CastIINode(val, tboth); 1811 } else if (tcon == TypePtr::NULL_PTR) { 1812 // Cast to null, but keep the pointer identity temporarily live. 1813 ccast = new CastPPNode(val, tboth); 1814 } else { 1815 const TypeF* tf = tcon->isa_float_constant(); 1816 const TypeD* td = tcon->isa_double_constant(); 1817 // Exclude tests vs float/double 0 as these could be 1818 // either +0 or -0. Just because you are equal to +0 1819 // doesn't mean you ARE +0! 1820 // Note, following code also replaces Long and Oop values. 1821 if ((!tf || tf->_f != 0.0) && 1822 (!td || td->_d != 0.0)) 1823 cast = con; // Replace non-constant val by con. 1824 } 1825 } 1826 break; 1827 1828 case BoolTest::ne: 1829 if (tcon == TypePtr::NULL_PTR) { 1830 cast = cast_not_null(val, false); 1831 } 1832 break; 1833 1834 default: 1835 // (At this point we could record int range types with CastII.) 1836 break; 1837 } 1838 1839 if (ccast != NULL) { 1840 const Type* tcc = ccast->as_Type()->type(); 1841 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1842 // Delay transform() call to allow recovery of pre-cast value 1843 // at the control merge. 1844 ccast->set_req(0, control()); 1845 _gvn.set_type_bottom(ccast); 1846 record_for_igvn(ccast); 1847 cast = ccast; 1848 } 1849 1850 if (cast != NULL) { // Here's the payoff. 1851 replace_in_map(val, cast); 1852 } 1853 } 1854 1855 /** 1856 * Use speculative type to optimize CmpP node: if comparison is 1857 * against the low level class, cast the object to the speculative 1858 * type if any. CmpP should then go away. 1859 * 1860 * @param c expected CmpP node 1861 * @return result of CmpP on object casted to speculative type 1862 * 1863 */ 1864 Node* Parse::optimize_cmp_with_klass(Node* c) { 1865 // If this is transformed by the _gvn to a comparison with the low 1866 // level klass then we may be able to use speculation 1867 if (c->Opcode() == Op_CmpP && 1868 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1869 c->in(2)->is_Con()) { 1870 Node* load_klass = NULL; 1871 Node* decode = NULL; 1872 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1873 decode = c->in(1); 1874 load_klass = c->in(1)->in(1); 1875 } else { 1876 load_klass = c->in(1); 1877 } 1878 if (load_klass->in(2)->is_AddP()) { 1879 Node* addp = load_klass->in(2); 1880 Node* obj = addp->in(AddPNode::Address); 1881 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1882 if (obj_type->speculative_type_not_null() != NULL) { 1883 ciKlass* k = obj_type->speculative_type(); 1884 inc_sp(2); 1885 obj = maybe_cast_profiled_obj(obj, k); 1886 dec_sp(2); 1887 // Make the CmpP use the casted obj 1888 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1889 load_klass = load_klass->clone(); 1890 load_klass->set_req(2, addp); 1891 load_klass = _gvn.transform(load_klass); 1892 if (decode != NULL) { 1893 decode = decode->clone(); 1894 decode->set_req(1, load_klass); 1895 load_klass = _gvn.transform(decode); 1896 } 1897 c = c->clone(); 1898 c->set_req(1, load_klass); 1899 c = _gvn.transform(c); 1900 } 1901 } 1902 } 1903 return c; 1904 } 1905 1906 //------------------------------do_one_bytecode-------------------------------- 1907 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1908 void Parse::do_one_bytecode() { 1909 Node *a, *b, *c, *d; // Handy temps 1910 BoolTest::mask btest; 1911 int i; 1912 1913 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1914 1915 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1916 "out of nodes parsing method")) { 1917 return; 1918 } 1919 1920 #ifdef ASSERT 1921 // for setting breakpoints 1922 if (TraceOptoParse) { 1923 tty->print(" @"); 1924 dump_bci(bci()); 1925 tty->cr(); 1926 } 1927 #endif 1928 1929 switch (bc()) { 1930 case Bytecodes::_nop: 1931 // do nothing 1932 break; 1933 case Bytecodes::_lconst_0: 1934 push_pair(longcon(0)); 1935 break; 1936 1937 case Bytecodes::_lconst_1: 1938 push_pair(longcon(1)); 1939 break; 1940 1941 case Bytecodes::_fconst_0: 1942 push(zerocon(T_FLOAT)); 1943 break; 1944 1945 case Bytecodes::_fconst_1: 1946 push(makecon(TypeF::ONE)); 1947 break; 1948 1949 case Bytecodes::_fconst_2: 1950 push(makecon(TypeF::make(2.0f))); 1951 break; 1952 1953 case Bytecodes::_dconst_0: 1954 push_pair(zerocon(T_DOUBLE)); 1955 break; 1956 1957 case Bytecodes::_dconst_1: 1958 push_pair(makecon(TypeD::ONE)); 1959 break; 1960 1961 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1962 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1963 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1964 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1965 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1966 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1967 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1968 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1969 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1970 case Bytecodes::_aconst_null: push(null()); break; 1971 case Bytecodes::_ldc: 1972 case Bytecodes::_ldc_w: 1973 case Bytecodes::_ldc2_w: 1974 // If the constant is unresolved, run this BC once in the interpreter. 1975 { 1976 ciConstant constant = iter().get_constant(); 1977 if (!constant.is_valid() || 1978 (constant.basic_type() == T_OBJECT && 1979 !constant.as_object()->is_loaded())) { 1980 int index = iter().get_constant_pool_index(); 1981 constantTag tag = iter().get_constant_pool_tag(index); 1982 uncommon_trap(Deoptimization::make_trap_request 1983 (Deoptimization::Reason_unloaded, 1984 Deoptimization::Action_reinterpret, 1985 index), 1986 NULL, tag.internal_name()); 1987 break; 1988 } 1989 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1990 "must be java_mirror of klass"); 1991 const Type* con_type = Type::make_from_constant(constant); 1992 if (con_type != NULL) { 1993 push_node(con_type->basic_type(), makecon(con_type)); 1994 } 1995 } 1996 1997 break; 1998 1999 case Bytecodes::_aload_0: 2000 push( local(0) ); 2001 break; 2002 case Bytecodes::_aload_1: 2003 push( local(1) ); 2004 break; 2005 case Bytecodes::_aload_2: 2006 push( local(2) ); 2007 break; 2008 case Bytecodes::_aload_3: 2009 push( local(3) ); 2010 break; 2011 case Bytecodes::_aload: 2012 push( local(iter().get_index()) ); 2013 break; 2014 2015 case Bytecodes::_fload_0: 2016 case Bytecodes::_iload_0: 2017 push( local(0) ); 2018 break; 2019 case Bytecodes::_fload_1: 2020 case Bytecodes::_iload_1: 2021 push( local(1) ); 2022 break; 2023 case Bytecodes::_fload_2: 2024 case Bytecodes::_iload_2: 2025 push( local(2) ); 2026 break; 2027 case Bytecodes::_fload_3: 2028 case Bytecodes::_iload_3: 2029 push( local(3) ); 2030 break; 2031 case Bytecodes::_fload: 2032 case Bytecodes::_iload: 2033 push( local(iter().get_index()) ); 2034 break; 2035 case Bytecodes::_lload_0: 2036 push_pair_local( 0 ); 2037 break; 2038 case Bytecodes::_lload_1: 2039 push_pair_local( 1 ); 2040 break; 2041 case Bytecodes::_lload_2: 2042 push_pair_local( 2 ); 2043 break; 2044 case Bytecodes::_lload_3: 2045 push_pair_local( 3 ); 2046 break; 2047 case Bytecodes::_lload: 2048 push_pair_local( iter().get_index() ); 2049 break; 2050 2051 case Bytecodes::_dload_0: 2052 push_pair_local(0); 2053 break; 2054 case Bytecodes::_dload_1: 2055 push_pair_local(1); 2056 break; 2057 case Bytecodes::_dload_2: 2058 push_pair_local(2); 2059 break; 2060 case Bytecodes::_dload_3: 2061 push_pair_local(3); 2062 break; 2063 case Bytecodes::_dload: 2064 push_pair_local(iter().get_index()); 2065 break; 2066 case Bytecodes::_fstore_0: 2067 case Bytecodes::_istore_0: 2068 case Bytecodes::_astore_0: 2069 set_local( 0, pop() ); 2070 break; 2071 case Bytecodes::_fstore_1: 2072 case Bytecodes::_istore_1: 2073 case Bytecodes::_astore_1: 2074 set_local( 1, pop() ); 2075 break; 2076 case Bytecodes::_fstore_2: 2077 case Bytecodes::_istore_2: 2078 case Bytecodes::_astore_2: 2079 set_local( 2, pop() ); 2080 break; 2081 case Bytecodes::_fstore_3: 2082 case Bytecodes::_istore_3: 2083 case Bytecodes::_astore_3: 2084 set_local( 3, pop() ); 2085 break; 2086 case Bytecodes::_fstore: 2087 case Bytecodes::_istore: 2088 case Bytecodes::_astore: 2089 set_local( iter().get_index(), pop() ); 2090 break; 2091 // long stores 2092 case Bytecodes::_lstore_0: 2093 set_pair_local( 0, pop_pair() ); 2094 break; 2095 case Bytecodes::_lstore_1: 2096 set_pair_local( 1, pop_pair() ); 2097 break; 2098 case Bytecodes::_lstore_2: 2099 set_pair_local( 2, pop_pair() ); 2100 break; 2101 case Bytecodes::_lstore_3: 2102 set_pair_local( 3, pop_pair() ); 2103 break; 2104 case Bytecodes::_lstore: 2105 set_pair_local( iter().get_index(), pop_pair() ); 2106 break; 2107 2108 // double stores 2109 case Bytecodes::_dstore_0: 2110 set_pair_local( 0, dstore_rounding(pop_pair()) ); 2111 break; 2112 case Bytecodes::_dstore_1: 2113 set_pair_local( 1, dstore_rounding(pop_pair()) ); 2114 break; 2115 case Bytecodes::_dstore_2: 2116 set_pair_local( 2, dstore_rounding(pop_pair()) ); 2117 break; 2118 case Bytecodes::_dstore_3: 2119 set_pair_local( 3, dstore_rounding(pop_pair()) ); 2120 break; 2121 case Bytecodes::_dstore: 2122 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 2123 break; 2124 2125 case Bytecodes::_pop: dec_sp(1); break; 2126 case Bytecodes::_pop2: dec_sp(2); break; 2127 case Bytecodes::_swap: 2128 a = pop(); 2129 b = pop(); 2130 push(a); 2131 push(b); 2132 break; 2133 case Bytecodes::_dup: 2134 a = pop(); 2135 push(a); 2136 push(a); 2137 break; 2138 case Bytecodes::_dup_x1: 2139 a = pop(); 2140 b = pop(); 2141 push( a ); 2142 push( b ); 2143 push( a ); 2144 break; 2145 case Bytecodes::_dup_x2: 2146 a = pop(); 2147 b = pop(); 2148 c = pop(); 2149 push( a ); 2150 push( c ); 2151 push( b ); 2152 push( a ); 2153 break; 2154 case Bytecodes::_dup2: 2155 a = pop(); 2156 b = pop(); 2157 push( b ); 2158 push( a ); 2159 push( b ); 2160 push( a ); 2161 break; 2162 2163 case Bytecodes::_dup2_x1: 2164 // before: .. c, b, a 2165 // after: .. b, a, c, b, a 2166 // not tested 2167 a = pop(); 2168 b = pop(); 2169 c = pop(); 2170 push( b ); 2171 push( a ); 2172 push( c ); 2173 push( b ); 2174 push( a ); 2175 break; 2176 case Bytecodes::_dup2_x2: 2177 // before: .. d, c, b, a 2178 // after: .. b, a, d, c, b, a 2179 // not tested 2180 a = pop(); 2181 b = pop(); 2182 c = pop(); 2183 d = pop(); 2184 push( b ); 2185 push( a ); 2186 push( d ); 2187 push( c ); 2188 push( b ); 2189 push( a ); 2190 break; 2191 2192 case Bytecodes::_arraylength: { 2193 // Must do null-check with value on expression stack 2194 Node *ary = null_check(peek(), T_ARRAY); 2195 // Compile-time detect of null-exception? 2196 if (stopped()) return; 2197 a = pop(); 2198 push(load_array_length(a)); 2199 break; 2200 } 2201 2202 case Bytecodes::_baload: array_load(T_BYTE); break; 2203 case Bytecodes::_caload: array_load(T_CHAR); break; 2204 case Bytecodes::_iaload: array_load(T_INT); break; 2205 case Bytecodes::_saload: array_load(T_SHORT); break; 2206 case Bytecodes::_faload: array_load(T_FLOAT); break; 2207 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2208 case Bytecodes::_laload: array_load(T_LONG); break; 2209 case Bytecodes::_daload: array_load(T_DOUBLE); break; 2210 case Bytecodes::_bastore: array_store(T_BYTE); break; 2211 case Bytecodes::_castore: array_store(T_CHAR); break; 2212 case Bytecodes::_iastore: array_store(T_INT); break; 2213 case Bytecodes::_sastore: array_store(T_SHORT); break; 2214 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2215 case Bytecodes::_aastore: array_store(T_OBJECT); break; 2216 case Bytecodes::_lastore: array_store(T_LONG); break; 2217 case Bytecodes::_dastore: array_store(T_DOUBLE); break; 2218 2219 case Bytecodes::_getfield: 2220 do_getfield(); 2221 break; 2222 2223 case Bytecodes::_getstatic: 2224 do_getstatic(); 2225 break; 2226 2227 case Bytecodes::_putfield: 2228 do_putfield(); 2229 break; 2230 2231 case Bytecodes::_putstatic: 2232 do_putstatic(); 2233 break; 2234 2235 case Bytecodes::_irem: 2236 do_irem(); 2237 break; 2238 case Bytecodes::_idiv: 2239 // Must keep both values on the expression-stack during null-check 2240 zero_check_int(peek()); 2241 // Compile-time detect of null-exception? 2242 if (stopped()) return; 2243 b = pop(); 2244 a = pop(); 2245 push( _gvn.transform( new DivINode(control(),a,b) ) ); 2246 break; 2247 case Bytecodes::_imul: 2248 b = pop(); a = pop(); 2249 push( _gvn.transform( new MulINode(a,b) ) ); 2250 break; 2251 case Bytecodes::_iadd: 2252 b = pop(); a = pop(); 2253 push( _gvn.transform( new AddINode(a,b) ) ); 2254 break; 2255 case Bytecodes::_ineg: 2256 a = pop(); 2257 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 2258 break; 2259 case Bytecodes::_isub: 2260 b = pop(); a = pop(); 2261 push( _gvn.transform( new SubINode(a,b) ) ); 2262 break; 2263 case Bytecodes::_iand: 2264 b = pop(); a = pop(); 2265 push( _gvn.transform( new AndINode(a,b) ) ); 2266 break; 2267 case Bytecodes::_ior: 2268 b = pop(); a = pop(); 2269 push( _gvn.transform( new OrINode(a,b) ) ); 2270 break; 2271 case Bytecodes::_ixor: 2272 b = pop(); a = pop(); 2273 push( _gvn.transform( new XorINode(a,b) ) ); 2274 break; 2275 case Bytecodes::_ishl: 2276 b = pop(); a = pop(); 2277 push( _gvn.transform( new LShiftINode(a,b) ) ); 2278 break; 2279 case Bytecodes::_ishr: 2280 b = pop(); a = pop(); 2281 push( _gvn.transform( new RShiftINode(a,b) ) ); 2282 break; 2283 case Bytecodes::_iushr: 2284 b = pop(); a = pop(); 2285 push( _gvn.transform( new URShiftINode(a,b) ) ); 2286 break; 2287 2288 case Bytecodes::_fneg: 2289 a = pop(); 2290 b = _gvn.transform(new NegFNode (a)); 2291 push(b); 2292 break; 2293 2294 case Bytecodes::_fsub: 2295 b = pop(); 2296 a = pop(); 2297 c = _gvn.transform( new SubFNode(a,b) ); 2298 d = precision_rounding(c); 2299 push( d ); 2300 break; 2301 2302 case Bytecodes::_fadd: 2303 b = pop(); 2304 a = pop(); 2305 c = _gvn.transform( new AddFNode(a,b) ); 2306 d = precision_rounding(c); 2307 push( d ); 2308 break; 2309 2310 case Bytecodes::_fmul: 2311 b = pop(); 2312 a = pop(); 2313 c = _gvn.transform( new MulFNode(a,b) ); 2314 d = precision_rounding(c); 2315 push( d ); 2316 break; 2317 2318 case Bytecodes::_fdiv: 2319 b = pop(); 2320 a = pop(); 2321 c = _gvn.transform( new DivFNode(0,a,b) ); 2322 d = precision_rounding(c); 2323 push( d ); 2324 break; 2325 2326 case Bytecodes::_frem: 2327 if (Matcher::has_match_rule(Op_ModF)) { 2328 // Generate a ModF node. 2329 b = pop(); 2330 a = pop(); 2331 c = _gvn.transform( new ModFNode(0,a,b) ); 2332 d = precision_rounding(c); 2333 push( d ); 2334 } 2335 else { 2336 // Generate a call. 2337 modf(); 2338 } 2339 break; 2340 2341 case Bytecodes::_fcmpl: 2342 b = pop(); 2343 a = pop(); 2344 c = _gvn.transform( new CmpF3Node( a, b)); 2345 push(c); 2346 break; 2347 case Bytecodes::_fcmpg: 2348 b = pop(); 2349 a = pop(); 2350 2351 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 2352 // which negates the result sign except for unordered. Flip the unordered 2353 // as well by using CmpF3 which implements unordered-lesser instead of 2354 // unordered-greater semantics. Finally, commute the result bits. Result 2355 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 2356 c = _gvn.transform( new CmpF3Node( b, a)); 2357 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2358 push(c); 2359 break; 2360 2361 case Bytecodes::_f2i: 2362 a = pop(); 2363 push(_gvn.transform(new ConvF2INode(a))); 2364 break; 2365 2366 case Bytecodes::_d2i: 2367 a = pop_pair(); 2368 b = _gvn.transform(new ConvD2INode(a)); 2369 push( b ); 2370 break; 2371 2372 case Bytecodes::_f2d: 2373 a = pop(); 2374 b = _gvn.transform( new ConvF2DNode(a)); 2375 push_pair( b ); 2376 break; 2377 2378 case Bytecodes::_d2f: 2379 a = pop_pair(); 2380 b = _gvn.transform( new ConvD2FNode(a)); 2381 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 2382 //b = _gvn.transform(new RoundFloatNode(0, b) ); 2383 push( b ); 2384 break; 2385 2386 case Bytecodes::_l2f: 2387 if (Matcher::convL2FSupported()) { 2388 a = pop_pair(); 2389 b = _gvn.transform( new ConvL2FNode(a)); 2390 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 2391 // Rather than storing the result into an FP register then pushing 2392 // out to memory to round, the machine instruction that implements 2393 // ConvL2D is responsible for rounding. 2394 // c = precision_rounding(b); 2395 c = _gvn.transform(b); 2396 push(c); 2397 } else { 2398 l2f(); 2399 } 2400 break; 2401 2402 case Bytecodes::_l2d: 2403 a = pop_pair(); 2404 b = _gvn.transform( new ConvL2DNode(a)); 2405 // For i486.ad, rounding is always necessary (see _l2f above). 2406 // c = dprecision_rounding(b); 2407 c = _gvn.transform(b); 2408 push_pair(c); 2409 break; 2410 2411 case Bytecodes::_f2l: 2412 a = pop(); 2413 b = _gvn.transform( new ConvF2LNode(a)); 2414 push_pair(b); 2415 break; 2416 2417 case Bytecodes::_d2l: 2418 a = pop_pair(); 2419 b = _gvn.transform( new ConvD2LNode(a)); 2420 push_pair(b); 2421 break; 2422 2423 case Bytecodes::_dsub: 2424 b = pop_pair(); 2425 a = pop_pair(); 2426 c = _gvn.transform( new SubDNode(a,b) ); 2427 d = dprecision_rounding(c); 2428 push_pair( d ); 2429 break; 2430 2431 case Bytecodes::_dadd: 2432 b = pop_pair(); 2433 a = pop_pair(); 2434 c = _gvn.transform( new AddDNode(a,b) ); 2435 d = dprecision_rounding(c); 2436 push_pair( d ); 2437 break; 2438 2439 case Bytecodes::_dmul: 2440 b = pop_pair(); 2441 a = pop_pair(); 2442 c = _gvn.transform( new MulDNode(a,b) ); 2443 d = dprecision_rounding(c); 2444 push_pair( d ); 2445 break; 2446 2447 case Bytecodes::_ddiv: 2448 b = pop_pair(); 2449 a = pop_pair(); 2450 c = _gvn.transform( new DivDNode(0,a,b) ); 2451 d = dprecision_rounding(c); 2452 push_pair( d ); 2453 break; 2454 2455 case Bytecodes::_dneg: 2456 a = pop_pair(); 2457 b = _gvn.transform(new NegDNode (a)); 2458 push_pair(b); 2459 break; 2460 2461 case Bytecodes::_drem: 2462 if (Matcher::has_match_rule(Op_ModD)) { 2463 // Generate a ModD node. 2464 b = pop_pair(); 2465 a = pop_pair(); 2466 // a % b 2467 2468 c = _gvn.transform( new ModDNode(0,a,b) ); 2469 d = dprecision_rounding(c); 2470 push_pair( d ); 2471 } 2472 else { 2473 // Generate a call. 2474 modd(); 2475 } 2476 break; 2477 2478 case Bytecodes::_dcmpl: 2479 b = pop_pair(); 2480 a = pop_pair(); 2481 c = _gvn.transform( new CmpD3Node( a, b)); 2482 push(c); 2483 break; 2484 2485 case Bytecodes::_dcmpg: 2486 b = pop_pair(); 2487 a = pop_pair(); 2488 // Same as dcmpl but need to flip the unordered case. 2489 // Commute the inputs, which negates the result sign except for unordered. 2490 // Flip the unordered as well by using CmpD3 which implements 2491 // unordered-lesser instead of unordered-greater semantics. 2492 // Finally, negate the result bits. Result is same as using a 2493 // CmpD3Greater except we did it with CmpD3 alone. 2494 c = _gvn.transform( new CmpD3Node( b, a)); 2495 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2496 push(c); 2497 break; 2498 2499 2500 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2501 case Bytecodes::_land: 2502 b = pop_pair(); 2503 a = pop_pair(); 2504 c = _gvn.transform( new AndLNode(a,b) ); 2505 push_pair(c); 2506 break; 2507 case Bytecodes::_lor: 2508 b = pop_pair(); 2509 a = pop_pair(); 2510 c = _gvn.transform( new OrLNode(a,b) ); 2511 push_pair(c); 2512 break; 2513 case Bytecodes::_lxor: 2514 b = pop_pair(); 2515 a = pop_pair(); 2516 c = _gvn.transform( new XorLNode(a,b) ); 2517 push_pair(c); 2518 break; 2519 2520 case Bytecodes::_lshl: 2521 b = pop(); // the shift count 2522 a = pop_pair(); // value to be shifted 2523 c = _gvn.transform( new LShiftLNode(a,b) ); 2524 push_pair(c); 2525 break; 2526 case Bytecodes::_lshr: 2527 b = pop(); // the shift count 2528 a = pop_pair(); // value to be shifted 2529 c = _gvn.transform( new RShiftLNode(a,b) ); 2530 push_pair(c); 2531 break; 2532 case Bytecodes::_lushr: 2533 b = pop(); // the shift count 2534 a = pop_pair(); // value to be shifted 2535 c = _gvn.transform( new URShiftLNode(a,b) ); 2536 push_pair(c); 2537 break; 2538 case Bytecodes::_lmul: 2539 b = pop_pair(); 2540 a = pop_pair(); 2541 c = _gvn.transform( new MulLNode(a,b) ); 2542 push_pair(c); 2543 break; 2544 2545 case Bytecodes::_lrem: 2546 // Must keep both values on the expression-stack during null-check 2547 assert(peek(0) == top(), "long word order"); 2548 zero_check_long(peek(1)); 2549 // Compile-time detect of null-exception? 2550 if (stopped()) return; 2551 b = pop_pair(); 2552 a = pop_pair(); 2553 c = _gvn.transform( new ModLNode(control(),a,b) ); 2554 push_pair(c); 2555 break; 2556 2557 case Bytecodes::_ldiv: 2558 // Must keep both values on the expression-stack during null-check 2559 assert(peek(0) == top(), "long word order"); 2560 zero_check_long(peek(1)); 2561 // Compile-time detect of null-exception? 2562 if (stopped()) return; 2563 b = pop_pair(); 2564 a = pop_pair(); 2565 c = _gvn.transform( new DivLNode(control(),a,b) ); 2566 push_pair(c); 2567 break; 2568 2569 case Bytecodes::_ladd: 2570 b = pop_pair(); 2571 a = pop_pair(); 2572 c = _gvn.transform( new AddLNode(a,b) ); 2573 push_pair(c); 2574 break; 2575 case Bytecodes::_lsub: 2576 b = pop_pair(); 2577 a = pop_pair(); 2578 c = _gvn.transform( new SubLNode(a,b) ); 2579 push_pair(c); 2580 break; 2581 case Bytecodes::_lcmp: 2582 // Safepoints are now inserted _before_ branches. The long-compare 2583 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2584 // slew of control flow. These are usually followed by a CmpI vs zero and 2585 // a branch; this pattern then optimizes to the obvious long-compare and 2586 // branch. However, if the branch is backwards there's a Safepoint 2587 // inserted. The inserted Safepoint captures the JVM state at the 2588 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2589 // long-compare is used to control a loop the debug info will force 2590 // computation of the 3-way value, even though the generated code uses a 2591 // long-compare and branch. We try to rectify the situation by inserting 2592 // a SafePoint here and have it dominate and kill the safepoint added at a 2593 // following backwards branch. At this point the JVM state merely holds 2 2594 // longs but not the 3-way value. 2595 if( UseLoopSafepoints ) { 2596 switch( iter().next_bc() ) { 2597 case Bytecodes::_ifgt: 2598 case Bytecodes::_iflt: 2599 case Bytecodes::_ifge: 2600 case Bytecodes::_ifle: 2601 case Bytecodes::_ifne: 2602 case Bytecodes::_ifeq: 2603 // If this is a backwards branch in the bytecodes, add Safepoint 2604 maybe_add_safepoint(iter().next_get_dest()); 2605 default: 2606 break; 2607 } 2608 } 2609 b = pop_pair(); 2610 a = pop_pair(); 2611 c = _gvn.transform( new CmpL3Node( a, b )); 2612 push(c); 2613 break; 2614 2615 case Bytecodes::_lneg: 2616 a = pop_pair(); 2617 b = _gvn.transform( new SubLNode(longcon(0),a)); 2618 push_pair(b); 2619 break; 2620 case Bytecodes::_l2i: 2621 a = pop_pair(); 2622 push( _gvn.transform( new ConvL2INode(a))); 2623 break; 2624 case Bytecodes::_i2l: 2625 a = pop(); 2626 b = _gvn.transform( new ConvI2LNode(a)); 2627 push_pair(b); 2628 break; 2629 case Bytecodes::_i2b: 2630 // Sign extend 2631 a = pop(); 2632 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2633 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2634 push( a ); 2635 break; 2636 case Bytecodes::_i2s: 2637 a = pop(); 2638 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2639 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2640 push( a ); 2641 break; 2642 case Bytecodes::_i2c: 2643 a = pop(); 2644 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2645 break; 2646 2647 case Bytecodes::_i2f: 2648 a = pop(); 2649 b = _gvn.transform( new ConvI2FNode(a) ) ; 2650 c = precision_rounding(b); 2651 push (b); 2652 break; 2653 2654 case Bytecodes::_i2d: 2655 a = pop(); 2656 b = _gvn.transform( new ConvI2DNode(a)); 2657 push_pair(b); 2658 break; 2659 2660 case Bytecodes::_iinc: // Increment local 2661 i = iter().get_index(); // Get local index 2662 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2663 break; 2664 2665 // Exit points of synchronized methods must have an unlock node 2666 case Bytecodes::_return: 2667 return_current(NULL); 2668 break; 2669 2670 case Bytecodes::_ireturn: 2671 case Bytecodes::_areturn: 2672 case Bytecodes::_freturn: 2673 return_current(pop()); 2674 break; 2675 case Bytecodes::_lreturn: 2676 return_current(pop_pair()); 2677 break; 2678 case Bytecodes::_dreturn: 2679 return_current(pop_pair()); 2680 break; 2681 2682 case Bytecodes::_athrow: 2683 // null exception oop throws NULL pointer exception 2684 null_check(peek()); 2685 if (stopped()) return; 2686 // Hook the thrown exception directly to subsequent handlers. 2687 if (BailoutToInterpreterForThrows) { 2688 // Keep method interpreted from now on. 2689 uncommon_trap(Deoptimization::Reason_unhandled, 2690 Deoptimization::Action_make_not_compilable); 2691 return; 2692 } 2693 if (env()->jvmti_can_post_on_exceptions()) { 2694 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2695 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2696 } 2697 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2698 add_exception_state(make_exception_state(peek())); 2699 break; 2700 2701 case Bytecodes::_goto: // fall through 2702 case Bytecodes::_goto_w: { 2703 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2704 2705 // If this is a backwards branch in the bytecodes, add Safepoint 2706 maybe_add_safepoint(target_bci); 2707 2708 // Update method data 2709 profile_taken_branch(target_bci); 2710 2711 // Merge the current control into the target basic block 2712 merge(target_bci); 2713 2714 // See if we can get some profile data and hand it off to the next block 2715 Block *target_block = block()->successor_for_bci(target_bci); 2716 if (target_block->pred_count() != 1) break; 2717 ciMethodData* methodData = method()->method_data(); 2718 if (!methodData->is_mature()) break; 2719 ciProfileData* data = methodData->bci_to_data(bci()); 2720 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch"); 2721 int taken = ((ciJumpData*)data)->taken(); 2722 taken = method()->scale_count(taken); 2723 target_block->set_count(taken); 2724 break; 2725 } 2726 2727 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2728 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2729 handle_if_null: 2730 // If this is a backwards branch in the bytecodes, add Safepoint 2731 maybe_add_safepoint(iter().get_dest()); 2732 a = null(); 2733 b = pop(); 2734 if (!_gvn.type(b)->speculative_maybe_null() && 2735 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2736 inc_sp(1); 2737 Node* null_ctl = top(); 2738 b = null_check_oop(b, &null_ctl, true, true, true); 2739 assert(null_ctl->is_top(), "no null control here"); 2740 dec_sp(1); 2741 } else if (_gvn.type(b)->speculative_always_null() && 2742 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2743 inc_sp(1); 2744 b = null_assert(b); 2745 dec_sp(1); 2746 } 2747 c = _gvn.transform( new CmpPNode(b, a) ); 2748 do_ifnull(btest, c); 2749 break; 2750 2751 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2752 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2753 handle_if_acmp: 2754 // If this is a backwards branch in the bytecodes, add Safepoint 2755 maybe_add_safepoint(iter().get_dest()); 2756 a = pop(); 2757 b = pop(); 2758 #if INCLUDE_SHENANDOAHGC 2759 if (UseShenandoahGC && ShenandoahAcmpBarrier) { 2760 a = access_resolve_for_write(a); 2761 b = access_resolve_for_write(b); 2762 } 2763 #endif 2764 c = _gvn.transform( new CmpPNode(b, a) ); 2765 c = optimize_cmp_with_klass(c); 2766 do_if(btest, c); 2767 break; 2768 2769 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2770 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2771 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2772 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2773 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2774 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2775 handle_ifxx: 2776 // If this is a backwards branch in the bytecodes, add Safepoint 2777 maybe_add_safepoint(iter().get_dest()); 2778 a = _gvn.intcon(0); 2779 b = pop(); 2780 c = _gvn.transform( new CmpINode(b, a) ); 2781 do_if(btest, c); 2782 break; 2783 2784 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2785 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2786 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2787 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2788 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2789 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2790 handle_if_icmp: 2791 // If this is a backwards branch in the bytecodes, add Safepoint 2792 maybe_add_safepoint(iter().get_dest()); 2793 a = pop(); 2794 b = pop(); 2795 c = _gvn.transform( new CmpINode( b, a ) ); 2796 do_if(btest, c); 2797 break; 2798 2799 case Bytecodes::_tableswitch: 2800 do_tableswitch(); 2801 break; 2802 2803 case Bytecodes::_lookupswitch: 2804 do_lookupswitch(); 2805 break; 2806 2807 case Bytecodes::_invokestatic: 2808 case Bytecodes::_invokedynamic: 2809 case Bytecodes::_invokespecial: 2810 case Bytecodes::_invokevirtual: 2811 case Bytecodes::_invokeinterface: 2812 do_call(); 2813 break; 2814 case Bytecodes::_checkcast: 2815 do_checkcast(); 2816 break; 2817 case Bytecodes::_instanceof: 2818 do_instanceof(); 2819 break; 2820 case Bytecodes::_anewarray: 2821 do_anewarray(); 2822 break; 2823 case Bytecodes::_newarray: 2824 do_newarray((BasicType)iter().get_index()); 2825 break; 2826 case Bytecodes::_multianewarray: 2827 do_multianewarray(); 2828 break; 2829 case Bytecodes::_new: 2830 do_new(); 2831 break; 2832 2833 case Bytecodes::_jsr: 2834 case Bytecodes::_jsr_w: 2835 do_jsr(); 2836 break; 2837 2838 case Bytecodes::_ret: 2839 do_ret(); 2840 break; 2841 2842 2843 case Bytecodes::_monitorenter: 2844 do_monitor_enter(); 2845 break; 2846 2847 case Bytecodes::_monitorexit: 2848 do_monitor_exit(); 2849 break; 2850 2851 case Bytecodes::_breakpoint: 2852 // Breakpoint set concurrently to compile 2853 // %%% use an uncommon trap? 2854 C->record_failure("breakpoint in method"); 2855 return; 2856 2857 default: 2858 #ifndef PRODUCT 2859 map()->dump(99); 2860 #endif 2861 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2862 ShouldNotReachHere(); 2863 } 2864 2865 #ifndef PRODUCT 2866 IdealGraphPrinter *printer = C->printer(); 2867 if (printer && printer->should_print(1)) { 2868 char buffer[256]; 2869 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2870 bool old = printer->traverse_outs(); 2871 printer->set_traverse_outs(true); 2872 printer->print_method(buffer, 4); 2873 printer->set_traverse_outs(old); 2874 } 2875 #endif 2876 }