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 #include "utilities/macros.hpp" 48 #if INCLUDE_SHENANDOAHGC 49 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp" 50 #endif 51 52 #ifndef PRODUCT 53 extern int explicit_null_checks_inserted, 54 explicit_null_checks_elided; 55 #endif 56 57 //---------------------------------array_load---------------------------------- 58 void Parse::array_load(BasicType bt) { 59 const Type* elemtype = Type::TOP; 60 bool big_val = bt == T_DOUBLE || bt == T_LONG; 61 Node* adr = array_addressing(bt, 0, false, &elemtype); 62 if (stopped()) return; // guaranteed null or range check 63 64 pop(); // index (already used) 65 Node* array = pop(); // the array itself 66 67 if (elemtype == TypeInt::BOOL) { 68 bt = T_BOOLEAN; 69 } else if (bt == T_OBJECT) { 70 elemtype = _gvn.type(array)->is_aryptr()->elem()->make_oopptr(); 71 } 72 73 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 74 75 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt, 76 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD); 77 if (big_val) { 78 push_pair(ld); 79 } else { 80 push(ld); 81 } 82 } 83 84 85 //--------------------------------array_store---------------------------------- 86 void Parse::array_store(BasicType bt) { 87 const Type* elemtype = Type::TOP; 88 bool big_val = bt == T_DOUBLE || bt == T_LONG; 89 Node* adr = array_addressing(bt, big_val ? 2 : 1, true, &elemtype); 90 if (stopped()) return; // guaranteed null or range check 91 if (bt == T_OBJECT) { 92 array_store_check(); 93 } 94 Node* val; // Oop to store 95 if (big_val) { 96 val = pop_pair(); 97 } else { 98 val = pop(); 99 } 100 pop(); // index (already used) 101 Node* array = pop(); // the array itself 102 103 if (elemtype == TypeInt::BOOL) { 104 bt = T_BOOLEAN; 105 } else if (bt == T_OBJECT) { 106 elemtype = _gvn.type(array)->is_aryptr()->elem()->make_oopptr(); 107 } 108 109 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 110 111 access_store_at(control(), array, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY); 112 } 113 114 115 //------------------------------array_addressing------------------------------- 116 // Pull array and index from the stack. Compute pointer-to-element. 117 Node* Parse::array_addressing(BasicType type, int vals, bool is_store, const Type* *result2) { 118 Node *idx = peek(0+vals); // Get from stack without popping 119 Node *ary = peek(1+vals); // in case of exception 120 121 // Null check the array base, with correct stack contents 122 ary = null_check(ary, T_ARRAY); 123 // Compile-time detect of null-exception? 124 if (stopped()) return top(); 125 126 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 127 const TypeInt* sizetype = arytype->size(); 128 const Type* elemtype = arytype->elem(); 129 130 if (UseUniqueSubclasses && result2 != NULL) { 131 const Type* el = elemtype->make_ptr(); 132 if (el && el->isa_instptr()) { 133 const TypeInstPtr* toop = el->is_instptr(); 134 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 135 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 136 const Type* subklass = Type::get_const_type(toop->klass()); 137 elemtype = subklass->join_speculative(el); 138 } 139 } 140 } 141 142 // Check for big class initializers with all constant offsets 143 // feeding into a known-size array. 144 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 145 // See if the highest idx value is less than the lowest array bound, 146 // and if the idx value cannot be negative: 147 bool need_range_check = true; 148 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 149 need_range_check = false; 150 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 151 } 152 153 ciKlass * arytype_klass = arytype->klass(); 154 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 155 // Only fails for some -Xcomp runs 156 // The class is unloaded. We have to run this bytecode in the interpreter. 157 uncommon_trap(Deoptimization::Reason_unloaded, 158 Deoptimization::Action_reinterpret, 159 arytype->klass(), "!loaded array"); 160 return top(); 161 } 162 163 // Do the range check 164 if (GenerateRangeChecks && need_range_check) { 165 Node* tst; 166 if (sizetype->_hi <= 0) { 167 // The greatest array bound is negative, so we can conclude that we're 168 // compiling unreachable code, but the unsigned compare trick used below 169 // only works with non-negative lengths. Instead, hack "tst" to be zero so 170 // the uncommon_trap path will always be taken. 171 tst = _gvn.intcon(0); 172 } else { 173 // Range is constant in array-oop, so we can use the original state of mem 174 Node* len = load_array_length(ary); 175 176 // Test length vs index (standard trick using unsigned compare) 177 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 178 BoolTest::mask btest = BoolTest::lt; 179 tst = _gvn.transform( new BoolNode(chk, btest) ); 180 } 181 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 182 _gvn.set_type(rc, rc->Value(&_gvn)); 183 if (!tst->is_Con()) { 184 record_for_igvn(rc); 185 } 186 set_control(_gvn.transform(new IfTrueNode(rc))); 187 // Branch to failure if out of bounds 188 { 189 PreserveJVMState pjvms(this); 190 set_control(_gvn.transform(new IfFalseNode(rc))); 191 if (C->allow_range_check_smearing()) { 192 // Do not use builtin_throw, since range checks are sometimes 193 // made more stringent by an optimistic transformation. 194 // This creates "tentative" range checks at this point, 195 // which are not guaranteed to throw exceptions. 196 // See IfNode::Ideal, is_range_check, adjust_check. 197 uncommon_trap(Deoptimization::Reason_range_check, 198 Deoptimization::Action_make_not_entrant, 199 NULL, "range_check"); 200 } else { 201 // If we have already recompiled with the range-check-widening 202 // heroic optimization turned off, then we must really be throwing 203 // range check exceptions. 204 builtin_throw(Deoptimization::Reason_range_check, idx); 205 } 206 } 207 } 208 // Check for always knowing you are throwing a range-check exception 209 if (stopped()) return top(); 210 211 if (is_store) { 212 ary = access_resolve_for_write(ary); 213 } else { 214 ary = access_resolve_for_read(ary); 215 } 216 217 // Make array address computation control dependent to prevent it 218 // from floating above the range check during loop optimizations. 219 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 220 221 if (result2 != NULL) *result2 = elemtype; 222 223 assert(ptr != top(), "top should go hand-in-hand with stopped"); 224 225 return ptr; 226 } 227 228 229 // returns IfNode 230 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 231 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 232 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 233 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 234 return iff; 235 } 236 237 // return Region node 238 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 239 Node *region = new RegionNode(3); // 2 results 240 record_for_igvn(region); 241 region->init_req(1, iffalse); 242 region->init_req(2, iftrue ); 243 _gvn.set_type(region, Type::CONTROL); 244 region = _gvn.transform(region); 245 set_control (region); 246 return region; 247 } 248 249 // sentinel value for the target bci to mark never taken branches 250 // (according to profiling) 251 static const int never_reached = INT_MAX; 252 253 //------------------------------helper for tableswitch------------------------- 254 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 255 // True branch, use existing map info 256 { PreserveJVMState pjvms(this); 257 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 258 set_control( iftrue ); 259 if (unc) { 260 repush_if_args(); 261 uncommon_trap(Deoptimization::Reason_unstable_if, 262 Deoptimization::Action_reinterpret, 263 NULL, 264 "taken always"); 265 } else { 266 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 267 profile_switch_case(prof_table_index); 268 merge_new_path(dest_bci_if_true); 269 } 270 } 271 272 // False branch 273 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 274 set_control( iffalse ); 275 } 276 277 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) { 278 // True branch, use existing map info 279 { PreserveJVMState pjvms(this); 280 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 281 set_control( iffalse ); 282 if (unc) { 283 repush_if_args(); 284 uncommon_trap(Deoptimization::Reason_unstable_if, 285 Deoptimization::Action_reinterpret, 286 NULL, 287 "taken never"); 288 } else { 289 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 290 profile_switch_case(prof_table_index); 291 merge_new_path(dest_bci_if_true); 292 } 293 } 294 295 // False branch 296 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 297 set_control( iftrue ); 298 } 299 300 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index, bool unc) { 301 // False branch, use existing map and control() 302 if (unc) { 303 repush_if_args(); 304 uncommon_trap(Deoptimization::Reason_unstable_if, 305 Deoptimization::Action_reinterpret, 306 NULL, 307 "taken never"); 308 } else { 309 assert(dest_bci != never_reached, "inconsistent dest"); 310 profile_switch_case(prof_table_index); 311 merge_new_path(dest_bci); 312 } 313 } 314 315 316 extern "C" { 317 static int jint_cmp(const void *i, const void *j) { 318 int a = *(jint *)i; 319 int b = *(jint *)j; 320 return a > b ? 1 : a < b ? -1 : 0; 321 } 322 } 323 324 325 // Default value for methodData switch indexing. Must be a negative value to avoid 326 // conflict with any legal switch index. 327 #define NullTableIndex -1 328 329 class SwitchRange : public StackObj { 330 // a range of integers coupled with a bci destination 331 jint _lo; // inclusive lower limit 332 jint _hi; // inclusive upper limit 333 int _dest; 334 int _table_index; // index into method data table 335 float _cnt; // how many times this range was hit according to profiling 336 337 public: 338 jint lo() const { return _lo; } 339 jint hi() const { return _hi; } 340 int dest() const { return _dest; } 341 int table_index() const { return _table_index; } 342 bool is_singleton() const { return _lo == _hi; } 343 float cnt() const { return _cnt; } 344 345 void setRange(jint lo, jint hi, int dest, int table_index, float cnt) { 346 assert(lo <= hi, "must be a non-empty range"); 347 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; _cnt = cnt; 348 assert(_cnt >= 0, ""); 349 } 350 bool adjoinRange(jint lo, jint hi, int dest, int table_index, float cnt, bool trim_ranges) { 351 assert(lo <= hi, "must be a non-empty range"); 352 if (lo == _hi+1 && table_index == _table_index) { 353 // see merge_ranges() comment below 354 if (trim_ranges) { 355 if (cnt == 0) { 356 if (_cnt != 0) { 357 return false; 358 } 359 if (dest != _dest) { 360 _dest = never_reached; 361 } 362 } else { 363 if (_cnt == 0) { 364 return false; 365 } 366 if (dest != _dest) { 367 return false; 368 } 369 } 370 } else { 371 if (dest != _dest) { 372 return false; 373 } 374 } 375 _hi = hi; 376 _cnt += cnt; 377 return true; 378 } 379 return false; 380 } 381 382 void set (jint value, int dest, int table_index, float cnt) { 383 setRange(value, value, dest, table_index, cnt); 384 } 385 bool adjoin(jint value, int dest, int table_index, float cnt, bool trim_ranges) { 386 return adjoinRange(value, value, dest, table_index, cnt, trim_ranges); 387 } 388 bool adjoin(SwitchRange& other) { 389 return adjoinRange(other._lo, other._hi, other._dest, other._table_index, other._cnt, false); 390 } 391 392 void print() { 393 if (is_singleton()) 394 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 395 else if (lo() == min_jint) 396 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 397 else if (hi() == max_jint) 398 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 399 else 400 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 401 } 402 }; 403 404 // We try to minimize the number of ranges and the size of the taken 405 // ones using profiling data. When ranges are created, 406 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 407 // if both were never hit or both were hit to build longer unreached 408 // ranges. Here, we now merge adjoining ranges with the same 409 // destination and finally set destination of unreached ranges to the 410 // special value never_reached because it can help minimize the number 411 // of tests that are necessary. 412 // 413 // For instance: 414 // [0, 1] to target1 sometimes taken 415 // [1, 2] to target1 never taken 416 // [2, 3] to target2 never taken 417 // would lead to: 418 // [0, 1] to target1 sometimes taken 419 // [1, 3] never taken 420 // 421 // (first 2 ranges to target1 are not merged) 422 static void merge_ranges(SwitchRange* ranges, int& rp) { 423 if (rp == 0) { 424 return; 425 } 426 int shift = 0; 427 for (int j = 0; j < rp; j++) { 428 SwitchRange& r1 = ranges[j-shift]; 429 SwitchRange& r2 = ranges[j+1]; 430 if (r1.adjoin(r2)) { 431 shift++; 432 } else if (shift > 0) { 433 ranges[j+1-shift] = r2; 434 } 435 } 436 rp -= shift; 437 for (int j = 0; j <= rp; j++) { 438 SwitchRange& r = ranges[j]; 439 if (r.cnt() == 0 && r.dest() != never_reached) { 440 r.setRange(r.lo(), r.hi(), never_reached, r.table_index(), r.cnt()); 441 } 442 } 443 } 444 445 //-------------------------------do_tableswitch-------------------------------- 446 void Parse::do_tableswitch() { 447 Node* lookup = pop(); 448 // Get information about tableswitch 449 int default_dest = iter().get_dest_table(0); 450 int lo_index = iter().get_int_table(1); 451 int hi_index = iter().get_int_table(2); 452 int len = hi_index - lo_index + 1; 453 454 if (len < 1) { 455 // If this is a backward branch, add safepoint 456 maybe_add_safepoint(default_dest); 457 merge(default_dest); 458 return; 459 } 460 461 ciMethodData* methodData = method()->method_data(); 462 ciMultiBranchData* profile = NULL; 463 if (methodData->is_mature() && UseSwitchProfiling) { 464 ciProfileData* data = methodData->bci_to_data(bci()); 465 if (data != NULL && data->is_MultiBranchData()) { 466 profile = (ciMultiBranchData*)data; 467 } 468 } 469 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 470 471 // generate decision tree, using trichotomy when possible 472 int rnum = len+2; 473 bool makes_backward_branch = false; 474 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 475 int rp = -1; 476 if (lo_index != min_jint) { 477 uint cnt = 1; 478 if (profile != NULL) { 479 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1); 480 } 481 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex, cnt); 482 } 483 for (int j = 0; j < len; j++) { 484 jint match_int = lo_index+j; 485 int dest = iter().get_dest_table(j+3); 486 makes_backward_branch |= (dest <= bci()); 487 int table_index = method_data_update() ? j : NullTableIndex; 488 uint cnt = 1; 489 if (profile != NULL) { 490 cnt = profile->count_at(j); 491 } 492 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 493 ranges[++rp].set(match_int, dest, table_index, cnt); 494 } 495 } 496 jint highest = lo_index+(len-1); 497 assert(ranges[rp].hi() == highest, ""); 498 if (highest != max_jint) { 499 uint cnt = 1; 500 if (profile != NULL) { 501 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1); 502 } 503 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, cnt, trim_ranges)) { 504 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, cnt); 505 } 506 } 507 assert(rp < len+2, "not too many ranges"); 508 509 if (trim_ranges) { 510 merge_ranges(ranges, rp); 511 } 512 513 // Safepoint in case if backward branch observed 514 if( makes_backward_branch && UseLoopSafepoints ) 515 add_safepoint(); 516 517 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 518 } 519 520 521 //------------------------------do_lookupswitch-------------------------------- 522 void Parse::do_lookupswitch() { 523 Node *lookup = pop(); // lookup value 524 // Get information about lookupswitch 525 int default_dest = iter().get_dest_table(0); 526 int len = iter().get_int_table(1); 527 528 if (len < 1) { // If this is a backward branch, add safepoint 529 maybe_add_safepoint(default_dest); 530 merge(default_dest); 531 return; 532 } 533 534 ciMethodData* methodData = method()->method_data(); 535 ciMultiBranchData* profile = NULL; 536 if (methodData->is_mature() && UseSwitchProfiling) { 537 ciProfileData* data = methodData->bci_to_data(bci()); 538 if (data != NULL && data->is_MultiBranchData()) { 539 profile = (ciMultiBranchData*)data; 540 } 541 } 542 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 543 544 // generate decision tree, using trichotomy when possible 545 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 546 { 547 for (int j = 0; j < len; j++) { 548 table[3*j+0] = iter().get_int_table(2+2*j); 549 table[3*j+1] = iter().get_dest_table(2+2*j+1); 550 table[3*j+2] = profile == NULL ? 1 : profile->count_at(j); 551 } 552 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 553 } 554 555 float defaults = 0; 556 jint prev = min_jint; 557 for (int j = 0; j < len; j++) { 558 jint match_int = table[3*j+0]; 559 if (match_int != prev) { 560 defaults += (float)match_int - prev; 561 } 562 prev = match_int+1; 563 } 564 if (prev-1 != max_jint) { 565 defaults += (float)max_jint - prev + 1; 566 } 567 float default_cnt = 1; 568 if (profile != NULL) { 569 default_cnt = profile->default_count()/defaults; 570 } 571 572 int rnum = len*2+1; 573 bool makes_backward_branch = false; 574 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 575 int rp = -1; 576 for (int j = 0; j < len; j++) { 577 jint match_int = table[3*j+0]; 578 int dest = table[3*j+1]; 579 int cnt = table[3*j+2]; 580 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 581 int table_index = method_data_update() ? j : NullTableIndex; 582 makes_backward_branch |= (dest <= bci()); 583 float c = default_cnt * ((float)match_int - next_lo); 584 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, NullTableIndex, c, trim_ranges))) { 585 assert(default_dest != never_reached, "sentinel value for dead destinations"); 586 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex, c); 587 } 588 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) { 589 assert(dest != never_reached, "sentinel value for dead destinations"); 590 ranges[++rp].set(match_int, dest, table_index, cnt); 591 } 592 } 593 jint highest = table[3*(len-1)]; 594 assert(ranges[rp].hi() == highest, ""); 595 if (highest != max_jint && 596 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest), trim_ranges)) { 597 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest)); 598 } 599 assert(rp < rnum, "not too many ranges"); 600 601 if (trim_ranges) { 602 merge_ranges(ranges, rp); 603 } 604 605 // Safepoint in case backward branch observed 606 if (makes_backward_branch && UseLoopSafepoints) 607 add_safepoint(); 608 609 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 610 } 611 612 static float if_prob(float taken_cnt, float total_cnt) { 613 assert(taken_cnt <= total_cnt, ""); 614 if (total_cnt == 0) { 615 return PROB_FAIR; 616 } 617 float p = taken_cnt / total_cnt; 618 return MIN2(MAX2(p, PROB_MIN), PROB_MAX); 619 } 620 621 static float if_cnt(float cnt) { 622 if (cnt == 0) { 623 return COUNT_UNKNOWN; 624 } 625 return cnt; 626 } 627 628 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 629 float total_cnt = 0; 630 for (SwitchRange* sr = lo; sr <= hi; sr++) { 631 total_cnt += sr->cnt(); 632 } 633 return total_cnt; 634 } 635 636 class SwitchRanges : public ResourceObj { 637 public: 638 SwitchRange* _lo; 639 SwitchRange* _hi; 640 SwitchRange* _mid; 641 float _cost; 642 643 enum { 644 Start, 645 LeftDone, 646 RightDone, 647 Done 648 } _state; 649 650 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 651 : _lo(lo), _hi(hi), _mid(NULL), 652 _cost(0), _state(Start) { 653 } 654 655 SwitchRanges() 656 : _lo(NULL), _hi(NULL), _mid(NULL), 657 _cost(0), _state(Start) {} 658 }; 659 660 // Estimate cost of performing a binary search on lo..hi 661 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 662 GrowableArray<SwitchRanges> tree; 663 SwitchRanges root(lo, hi); 664 tree.push(root); 665 666 float cost = 0; 667 do { 668 SwitchRanges& r = *tree.adr_at(tree.length()-1); 669 if (r._hi != r._lo) { 670 if (r._mid == NULL) { 671 float r_cnt = sum_of_cnts(r._lo, r._hi); 672 673 if (r_cnt == 0) { 674 tree.pop(); 675 cost = 0; 676 continue; 677 } 678 679 SwitchRange* mid = NULL; 680 mid = r._lo; 681 for (float cnt = 0; ; ) { 682 assert(mid <= r._hi, "out of bounds"); 683 cnt += mid->cnt(); 684 if (cnt > r_cnt / 2) { 685 break; 686 } 687 mid++; 688 } 689 assert(mid <= r._hi, "out of bounds"); 690 r._mid = mid; 691 r._cost = r_cnt / total_cnt; 692 } 693 r._cost += cost; 694 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 695 cost = 0; 696 r._state = SwitchRanges::LeftDone; 697 tree.push(SwitchRanges(r._lo, r._mid-1)); 698 } else if (r._state < SwitchRanges::RightDone) { 699 cost = 0; 700 r._state = SwitchRanges::RightDone; 701 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 702 } else { 703 tree.pop(); 704 cost = r._cost; 705 } 706 } else { 707 tree.pop(); 708 cost = r._cost; 709 } 710 } while (tree.length() > 0); 711 712 713 return cost; 714 } 715 716 // It sometimes pays off to test most common ranges before the binary search 717 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 718 uint nr = hi - lo + 1; 719 float total_cnt = sum_of_cnts(lo, hi); 720 721 float min = compute_tree_cost(lo, hi, total_cnt); 722 float extra = 1; 723 float sub = 0; 724 725 SwitchRange* array1 = lo; 726 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 727 728 SwitchRange* ranges = NULL; 729 730 while (nr >= 2) { 731 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 732 ranges = (lo == array1) ? array2 : array1; 733 734 // Find highest frequency range 735 SwitchRange* candidate = lo; 736 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 737 if (sr->cnt() > candidate->cnt()) { 738 candidate = sr; 739 } 740 } 741 SwitchRange most_freq = *candidate; 742 if (most_freq.cnt() == 0) { 743 break; 744 } 745 746 // Copy remaining ranges into another array 747 int shift = 0; 748 for (uint i = 0; i < nr; i++) { 749 SwitchRange* sr = &lo[i]; 750 if (sr != candidate) { 751 ranges[i-shift] = *sr; 752 } else { 753 shift++; 754 if (i > 0 && i < nr-1) { 755 SwitchRange prev = lo[i-1]; 756 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.table_index(), prev.cnt()); 757 if (prev.adjoin(lo[i+1])) { 758 shift++; 759 i++; 760 } 761 ranges[i-shift] = prev; 762 } 763 } 764 } 765 nr -= shift; 766 767 // Evaluate cost of testing the most common range and performing a 768 // binary search on the other ranges 769 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 770 if (cost >= min) { 771 break; 772 } 773 // swap arrays 774 lo = &ranges[0]; 775 hi = &ranges[nr-1]; 776 777 // It pays off: emit the test for the most common range 778 assert(most_freq.cnt() > 0, "must be taken"); 779 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 780 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo()))); 781 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 782 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 783 jump_if_true_fork(iff, most_freq.dest(), most_freq.table_index(), false); 784 785 sub += most_freq.cnt() / total_cnt; 786 extra += 1 - sub; 787 min = cost; 788 } 789 } 790 791 //----------------------------create_jump_tables------------------------------- 792 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 793 // Are jumptables enabled 794 if (!UseJumpTables) return false; 795 796 // Are jumptables supported 797 if (!Matcher::has_match_rule(Op_Jump)) return false; 798 799 // Don't make jump table if profiling 800 if (method_data_update()) return false; 801 802 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 803 804 // Decide if a guard is needed to lop off big ranges at either (or 805 // both) end(s) of the input set. We'll call this the default target 806 // even though we can't be sure that it is the true "default". 807 808 bool needs_guard = false; 809 int default_dest; 810 int64_t total_outlier_size = 0; 811 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 812 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 813 814 if (lo->dest() == hi->dest()) { 815 total_outlier_size = hi_size + lo_size; 816 default_dest = lo->dest(); 817 } else if (lo_size > hi_size) { 818 total_outlier_size = lo_size; 819 default_dest = lo->dest(); 820 } else { 821 total_outlier_size = hi_size; 822 default_dest = hi->dest(); 823 } 824 825 float total = sum_of_cnts(lo, hi); 826 float cost = compute_tree_cost(lo, hi, total); 827 828 // If a guard test will eliminate very sparse end ranges, then 829 // it is worth the cost of an extra jump. 830 float trimmed_cnt = 0; 831 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 832 needs_guard = true; 833 if (default_dest == lo->dest()) { 834 trimmed_cnt += lo->cnt(); 835 lo++; 836 } 837 if (default_dest == hi->dest()) { 838 trimmed_cnt += hi->cnt(); 839 hi--; 840 } 841 } 842 843 // Find the total number of cases and ranges 844 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 845 int num_range = hi - lo + 1; 846 847 // Don't create table if: too large, too small, or too sparse. 848 if (num_cases > MaxJumpTableSize) 849 return false; 850 if (UseSwitchProfiling) { 851 // MinJumpTableSize is set so with a well balanced binary tree, 852 // when the number of ranges is MinJumpTableSize, it's cheaper to 853 // go through a JumpNode that a tree of IfNodes. Average cost of a 854 // tree of IfNodes with MinJumpTableSize is 855 // log2f(MinJumpTableSize) comparisons. So if the cost computed 856 // from profile data is less than log2f(MinJumpTableSize) then 857 // going with the binary search is cheaper. 858 if (cost < log2f(MinJumpTableSize)) { 859 return false; 860 } 861 } else { 862 if (num_cases < MinJumpTableSize) 863 return false; 864 } 865 if (num_cases > (MaxJumpTableSparseness * num_range)) 866 return false; 867 868 // Normalize table lookups to zero 869 int lowval = lo->lo(); 870 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 871 872 // Generate a guard to protect against input keyvals that aren't 873 // in the switch domain. 874 if (needs_guard) { 875 Node* size = _gvn.intcon(num_cases); 876 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 877 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 878 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 879 jump_if_true_fork(iff, default_dest, NullTableIndex, trim_ranges && trimmed_cnt == 0); 880 881 total -= trimmed_cnt; 882 } 883 884 // Create an ideal node JumpTable that has projections 885 // of all possible ranges for a switch statement 886 // The key_val input must be converted to a pointer offset and scaled. 887 // Compare Parse::array_addressing above. 888 889 // Clean the 32-bit int into a real 64-bit offset. 890 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 891 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin); 892 // Make I2L conversion control dependent to prevent it from 893 // floating above the range check during loop optimizations. 894 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control()); 895 896 // Shift the value by wordsize so we have an index into the table, rather 897 // than a switch value 898 Node *shiftWord = _gvn.MakeConX(wordSize); 899 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 900 901 // Create the JumpNode 902 Arena* arena = C->comp_arena(); 903 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 904 int i = 0; 905 if (total == 0) { 906 for (SwitchRange* r = lo; r <= hi; r++) { 907 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 908 probs[i] = 1.0F / num_cases; 909 } 910 } 911 } else { 912 for (SwitchRange* r = lo; r <= hi; r++) { 913 float prob = r->cnt()/total; 914 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 915 probs[i] = prob / (r->hi() - r->lo() + 1); 916 } 917 } 918 } 919 920 ciMethodData* methodData = method()->method_data(); 921 ciMultiBranchData* profile = NULL; 922 if (methodData->is_mature()) { 923 ciProfileData* data = methodData->bci_to_data(bci()); 924 if (data != NULL && data->is_MultiBranchData()) { 925 profile = (ciMultiBranchData*)data; 926 } 927 } 928 929 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total)); 930 931 // These are the switch destinations hanging off the jumpnode 932 i = 0; 933 for (SwitchRange* r = lo; r <= hi; r++) { 934 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 935 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 936 { 937 PreserveJVMState pjvms(this); 938 set_control(input); 939 jump_if_always_fork(r->dest(), r->table_index(), trim_ranges && r->cnt() == 0); 940 } 941 } 942 } 943 assert(i == num_cases, "miscount of cases"); 944 stop_and_kill_map(); // no more uses for this JVMS 945 return true; 946 } 947 948 //----------------------------jump_switch_ranges------------------------------- 949 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 950 Block* switch_block = block(); 951 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 952 953 if (switch_depth == 0) { 954 // Do special processing for the top-level call. 955 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 956 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 957 958 // Decrement pred-numbers for the unique set of nodes. 959 #ifdef ASSERT 960 if (!trim_ranges) { 961 // Ensure that the block's successors are a (duplicate-free) set. 962 int successors_counted = 0; // block occurrences in [hi..lo] 963 int unique_successors = switch_block->num_successors(); 964 for (int i = 0; i < unique_successors; i++) { 965 Block* target = switch_block->successor_at(i); 966 967 // Check that the set of successors is the same in both places. 968 int successors_found = 0; 969 for (SwitchRange* p = lo; p <= hi; p++) { 970 if (p->dest() == target->start()) successors_found++; 971 } 972 assert(successors_found > 0, "successor must be known"); 973 successors_counted += successors_found; 974 } 975 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 976 } 977 #endif 978 979 // Maybe prune the inputs, based on the type of key_val. 980 jint min_val = min_jint; 981 jint max_val = max_jint; 982 const TypeInt* ti = key_val->bottom_type()->isa_int(); 983 if (ti != NULL) { 984 min_val = ti->_lo; 985 max_val = ti->_hi; 986 assert(min_val <= max_val, "invalid int type"); 987 } 988 while (lo->hi() < min_val) { 989 lo++; 990 } 991 if (lo->lo() < min_val) { 992 lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index(), lo->cnt()); 993 } 994 while (hi->lo() > max_val) { 995 hi--; 996 } 997 if (hi->hi() > max_val) { 998 hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index(), hi->cnt()); 999 } 1000 1001 linear_search_switch_ranges(key_val, lo, hi); 1002 } 1003 1004 #ifndef PRODUCT 1005 if (switch_depth == 0) { 1006 _max_switch_depth = 0; 1007 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 1008 } 1009 #endif 1010 1011 assert(lo <= hi, "must be a non-empty set of ranges"); 1012 if (lo == hi) { 1013 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1014 } else { 1015 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 1016 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 1017 1018 if (create_jump_tables(key_val, lo, hi)) return; 1019 1020 SwitchRange* mid = NULL; 1021 float total_cnt = sum_of_cnts(lo, hi); 1022 1023 int nr = hi - lo + 1; 1024 if (UseSwitchProfiling) { 1025 // Don't keep the binary search tree balanced: pick up mid point 1026 // that split frequencies in half. 1027 float cnt = 0; 1028 for (SwitchRange* sr = lo; sr <= hi; sr++) { 1029 cnt += sr->cnt(); 1030 if (cnt >= total_cnt / 2) { 1031 mid = sr; 1032 break; 1033 } 1034 } 1035 } else { 1036 mid = lo + nr/2; 1037 1038 // if there is an easy choice, pivot at a singleton: 1039 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1040 1041 assert(lo < mid && mid <= hi, "good pivot choice"); 1042 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1043 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1044 } 1045 1046 1047 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1048 1049 if (mid->is_singleton()) { 1050 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1051 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index(), trim_ranges && mid->cnt() == 0); 1052 1053 // Special Case: If there are exactly three ranges, and the high 1054 // and low range each go to the same place, omit the "gt" test, 1055 // since it will not discriminate anything. 1056 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1057 1058 // if there is a higher range, test for it and process it: 1059 if (mid < hi && !eq_test_only) { 1060 // two comparisons of same values--should enable 1 test for 2 branches 1061 // Use BoolTest::le instead of BoolTest::gt 1062 float cnt = sum_of_cnts(lo, mid-1); 1063 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le, if_prob(cnt, total_cnt), if_cnt(cnt)); 1064 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) ); 1065 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) ); 1066 { PreserveJVMState pjvms(this); 1067 set_control(iffalse); 1068 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1069 } 1070 set_control(iftrue); 1071 } 1072 1073 } else { 1074 // mid is a range, not a singleton, so treat mid..hi as a unit 1075 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1076 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)); 1077 1078 // if there is a higher range, test for it and process it: 1079 if (mid == hi) { 1080 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index(), trim_ranges && cnt == 0); 1081 } else { 1082 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1083 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1084 { PreserveJVMState pjvms(this); 1085 set_control(iftrue); 1086 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1087 } 1088 set_control(iffalse); 1089 } 1090 } 1091 1092 // in any case, process the lower range 1093 if (mid == lo) { 1094 if (mid->is_singleton()) { 1095 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1096 } else { 1097 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0); 1098 } 1099 } else { 1100 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1101 } 1102 } 1103 1104 // Decrease pred_count for each successor after all is done. 1105 if (switch_depth == 0) { 1106 int unique_successors = switch_block->num_successors(); 1107 for (int i = 0; i < unique_successors; i++) { 1108 Block* target = switch_block->successor_at(i); 1109 // Throw away the pre-allocated path for each unique successor. 1110 target->next_path_num(); 1111 } 1112 } 1113 1114 #ifndef PRODUCT 1115 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1116 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1117 SwitchRange* r; 1118 int nsing = 0; 1119 for( r = lo; r <= hi; r++ ) { 1120 if( r->is_singleton() ) nsing++; 1121 } 1122 tty->print(">>> "); 1123 _method->print_short_name(); 1124 tty->print_cr(" switch decision tree"); 1125 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1126 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1127 if (_max_switch_depth > _est_switch_depth) { 1128 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1129 } 1130 tty->print(" "); 1131 for( r = lo; r <= hi; r++ ) { 1132 r->print(); 1133 } 1134 tty->cr(); 1135 } 1136 #endif 1137 } 1138 1139 void Parse::modf() { 1140 Node *f2 = pop(); 1141 Node *f1 = pop(); 1142 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1143 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1144 "frem", NULL, //no memory effects 1145 f1, f2); 1146 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1147 1148 push(res); 1149 } 1150 1151 void Parse::modd() { 1152 Node *d2 = pop_pair(); 1153 Node *d1 = pop_pair(); 1154 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1155 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1156 "drem", NULL, //no memory effects 1157 d1, top(), d2, top()); 1158 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1159 1160 #ifdef ASSERT 1161 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 1162 assert(res_top == top(), "second value must be top"); 1163 #endif 1164 1165 push_pair(res_d); 1166 } 1167 1168 void Parse::l2f() { 1169 Node* f2 = pop(); 1170 Node* f1 = pop(); 1171 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1172 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1173 "l2f", NULL, //no memory effects 1174 f1, f2); 1175 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1176 1177 push(res); 1178 } 1179 1180 void Parse::do_irem() { 1181 // Must keep both values on the expression-stack during null-check 1182 zero_check_int(peek()); 1183 // Compile-time detect of null-exception? 1184 if (stopped()) return; 1185 1186 Node* b = pop(); 1187 Node* a = pop(); 1188 1189 const Type *t = _gvn.type(b); 1190 if (t != Type::TOP) { 1191 const TypeInt *ti = t->is_int(); 1192 if (ti->is_con()) { 1193 int divisor = ti->get_con(); 1194 // check for positive power of 2 1195 if (divisor > 0 && 1196 (divisor & ~(divisor-1)) == divisor) { 1197 // yes ! 1198 Node *mask = _gvn.intcon((divisor - 1)); 1199 // Sigh, must handle negative dividends 1200 Node *zero = _gvn.intcon(0); 1201 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN); 1202 Node *iff = _gvn.transform( new IfFalseNode(ifff) ); 1203 Node *ift = _gvn.transform( new IfTrueNode (ifff) ); 1204 Node *reg = jump_if_join(ift, iff); 1205 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 1206 // Negative path; negate/and/negate 1207 Node *neg = _gvn.transform( new SubINode(zero, a) ); 1208 Node *andn= _gvn.transform( new AndINode(neg, mask) ); 1209 Node *negn= _gvn.transform( new SubINode(zero, andn) ); 1210 phi->init_req(1, negn); 1211 // Fast positive case 1212 Node *andx = _gvn.transform( new AndINode(a, mask) ); 1213 phi->init_req(2, andx); 1214 // Push the merge 1215 push( _gvn.transform(phi) ); 1216 return; 1217 } 1218 } 1219 } 1220 // Default case 1221 push( _gvn.transform( new ModINode(control(),a,b) ) ); 1222 } 1223 1224 // Handle jsr and jsr_w bytecode 1225 void Parse::do_jsr() { 1226 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1227 1228 // Store information about current state, tagged with new _jsr_bci 1229 int return_bci = iter().next_bci(); 1230 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1231 1232 // Update method data 1233 profile_taken_branch(jsr_bci); 1234 1235 // The way we do things now, there is only one successor block 1236 // for the jsr, because the target code is cloned by ciTypeFlow. 1237 Block* target = successor_for_bci(jsr_bci); 1238 1239 // What got pushed? 1240 const Type* ret_addr = target->peek(); 1241 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1242 1243 // Effect on jsr on stack 1244 push(_gvn.makecon(ret_addr)); 1245 1246 // Flow to the jsr. 1247 merge(jsr_bci); 1248 } 1249 1250 // Handle ret bytecode 1251 void Parse::do_ret() { 1252 // Find to whom we return. 1253 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1254 Block* target = block()->successor_at(0); 1255 assert(!target->is_ready(), "our arrival must be expected"); 1256 profile_ret(target->flow()->start()); 1257 int pnum = target->next_path_num(); 1258 merge_common(target, pnum); 1259 } 1260 1261 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1262 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1263 // Only ::eq and ::ne are supported for profile injection. 1264 return false; 1265 } 1266 if (test->is_Cmp() && 1267 test->in(1)->Opcode() == Op_ProfileBoolean) { 1268 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1269 int false_cnt = profile->false_count(); 1270 int true_cnt = profile->true_count(); 1271 1272 // Counts matching depends on the actual test operation (::eq or ::ne). 1273 // No need to scale the counts because profile injection was designed 1274 // to feed exact counts into VM. 1275 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1276 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1277 1278 profile->consume(); 1279 return true; 1280 } 1281 return false; 1282 } 1283 //--------------------------dynamic_branch_prediction-------------------------- 1284 // Try to gather dynamic branch prediction behavior. Return a probability 1285 // of the branch being taken and set the "cnt" field. Returns a -1.0 1286 // if we need to use static prediction for some reason. 1287 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1288 ResourceMark rm; 1289 1290 cnt = COUNT_UNKNOWN; 1291 1292 int taken = 0; 1293 int not_taken = 0; 1294 1295 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1296 1297 if (use_mdo) { 1298 // Use MethodData information if it is available 1299 // FIXME: free the ProfileData structure 1300 ciMethodData* methodData = method()->method_data(); 1301 if (!methodData->is_mature()) return PROB_UNKNOWN; 1302 ciProfileData* data = methodData->bci_to_data(bci()); 1303 if (data == NULL) { 1304 return PROB_UNKNOWN; 1305 } 1306 if (!data->is_JumpData()) return PROB_UNKNOWN; 1307 1308 // get taken and not taken values 1309 taken = data->as_JumpData()->taken(); 1310 not_taken = 0; 1311 if (data->is_BranchData()) { 1312 not_taken = data->as_BranchData()->not_taken(); 1313 } 1314 1315 // scale the counts to be commensurate with invocation counts: 1316 taken = method()->scale_count(taken); 1317 not_taken = method()->scale_count(not_taken); 1318 } 1319 1320 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1321 // We also check that individual counters are positive first, otherwise the sum can become positive. 1322 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 1323 if (C->log() != NULL) { 1324 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1325 } 1326 return PROB_UNKNOWN; 1327 } 1328 1329 // Compute frequency that we arrive here 1330 float sum = taken + not_taken; 1331 // Adjust, if this block is a cloned private block but the 1332 // Jump counts are shared. Taken the private counts for 1333 // just this path instead of the shared counts. 1334 if( block()->count() > 0 ) 1335 sum = block()->count(); 1336 cnt = sum / FreqCountInvocations; 1337 1338 // Pin probability to sane limits 1339 float prob; 1340 if( !taken ) 1341 prob = (0+PROB_MIN) / 2; 1342 else if( !not_taken ) 1343 prob = (1+PROB_MAX) / 2; 1344 else { // Compute probability of true path 1345 prob = (float)taken / (float)(taken + not_taken); 1346 if (prob > PROB_MAX) prob = PROB_MAX; 1347 if (prob < PROB_MIN) prob = PROB_MIN; 1348 } 1349 1350 assert((cnt > 0.0f) && (prob > 0.0f), 1351 "Bad frequency assignment in if"); 1352 1353 if (C->log() != NULL) { 1354 const char* prob_str = NULL; 1355 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1356 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1357 char prob_str_buf[30]; 1358 if (prob_str == NULL) { 1359 sprintf(prob_str_buf, "%g", prob); 1360 prob_str = prob_str_buf; 1361 } 1362 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1363 iter().get_dest(), taken, not_taken, cnt, prob_str); 1364 } 1365 return prob; 1366 } 1367 1368 //-----------------------------branch_prediction------------------------------- 1369 float Parse::branch_prediction(float& cnt, 1370 BoolTest::mask btest, 1371 int target_bci, 1372 Node* test) { 1373 float prob = dynamic_branch_prediction(cnt, btest, test); 1374 // If prob is unknown, switch to static prediction 1375 if (prob != PROB_UNKNOWN) return prob; 1376 1377 prob = PROB_FAIR; // Set default value 1378 if (btest == BoolTest::eq) // Exactly equal test? 1379 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1380 else if (btest == BoolTest::ne) 1381 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1382 1383 // If this is a conditional test guarding a backwards branch, 1384 // assume its a loop-back edge. Make it a likely taken branch. 1385 if (target_bci < bci()) { 1386 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1387 // Since it's an OSR, we probably have profile data, but since 1388 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1389 // Let's make a special check here for completely zero counts. 1390 ciMethodData* methodData = method()->method_data(); 1391 if (!methodData->is_empty()) { 1392 ciProfileData* data = methodData->bci_to_data(bci()); 1393 // Only stop for truly zero counts, which mean an unknown part 1394 // of the OSR-ed method, and we want to deopt to gather more stats. 1395 // If you have ANY counts, then this loop is simply 'cold' relative 1396 // to the OSR loop. 1397 if (data == NULL || 1398 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1399 // This is the only way to return PROB_UNKNOWN: 1400 return PROB_UNKNOWN; 1401 } 1402 } 1403 } 1404 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1405 } 1406 1407 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1408 return prob; 1409 } 1410 1411 // The magic constants are chosen so as to match the output of 1412 // branch_prediction() when the profile reports a zero taken count. 1413 // It is important to distinguish zero counts unambiguously, because 1414 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1415 // very small but nonzero probabilities, which if confused with zero 1416 // counts would keep the program recompiling indefinitely. 1417 bool Parse::seems_never_taken(float prob) const { 1418 return prob < PROB_MIN; 1419 } 1420 1421 // True if the comparison seems to be the kind that will not change its 1422 // statistics from true to false. See comments in adjust_map_after_if. 1423 // This question is only asked along paths which are already 1424 // classifed as untaken (by seems_never_taken), so really, 1425 // if a path is never taken, its controlling comparison is 1426 // already acting in a stable fashion. If the comparison 1427 // seems stable, we will put an expensive uncommon trap 1428 // on the untaken path. 1429 bool Parse::seems_stable_comparison() const { 1430 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 1431 return false; 1432 } 1433 return true; 1434 } 1435 1436 //-------------------------------repush_if_args-------------------------------- 1437 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1438 inline int Parse::repush_if_args() { 1439 if (PrintOpto && WizardMode) { 1440 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1441 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1442 method()->print_name(); tty->cr(); 1443 } 1444 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1445 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1446 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1447 assert(argument(0) != NULL, "must exist"); 1448 assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 1449 inc_sp(bc_depth); 1450 return bc_depth; 1451 } 1452 1453 //----------------------------------do_ifnull---------------------------------- 1454 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1455 int target_bci = iter().get_dest(); 1456 1457 Block* branch_block = successor_for_bci(target_bci); 1458 Block* next_block = successor_for_bci(iter().next_bci()); 1459 1460 float cnt; 1461 float prob = branch_prediction(cnt, btest, target_bci, c); 1462 if (prob == PROB_UNKNOWN) { 1463 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1464 if (PrintOpto && Verbose) { 1465 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1466 } 1467 repush_if_args(); // to gather stats on loop 1468 // We need to mark this branch as taken so that if we recompile we will 1469 // see that it is possible. In the tiered system the interpreter doesn't 1470 // do profiling and by the time we get to the lower tier from the interpreter 1471 // the path may be cold again. Make sure it doesn't look untaken 1472 profile_taken_branch(target_bci, !ProfileInterpreter); 1473 uncommon_trap(Deoptimization::Reason_unreached, 1474 Deoptimization::Action_reinterpret, 1475 NULL, "cold"); 1476 if (C->eliminate_boxing()) { 1477 // Mark the successor blocks as parsed 1478 branch_block->next_path_num(); 1479 next_block->next_path_num(); 1480 } 1481 return; 1482 } 1483 1484 NOT_PRODUCT(explicit_null_checks_inserted++); 1485 1486 // Generate real control flow 1487 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1488 1489 // Sanity check the probability value 1490 assert(prob > 0.0f,"Bad probability in Parser"); 1491 // Need xform to put node in hash table 1492 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1493 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1494 // True branch 1495 { PreserveJVMState pjvms(this); 1496 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1497 set_control(iftrue); 1498 1499 if (stopped()) { // Path is dead? 1500 NOT_PRODUCT(explicit_null_checks_elided++); 1501 if (C->eliminate_boxing()) { 1502 // Mark the successor block as parsed 1503 branch_block->next_path_num(); 1504 } 1505 } else { // Path is live. 1506 // Update method data 1507 profile_taken_branch(target_bci); 1508 adjust_map_after_if(btest, c, prob, branch_block, next_block); 1509 if (!stopped()) { 1510 merge(target_bci); 1511 } 1512 } 1513 } 1514 1515 // False branch 1516 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1517 set_control(iffalse); 1518 1519 if (stopped()) { // Path is dead? 1520 NOT_PRODUCT(explicit_null_checks_elided++); 1521 if (C->eliminate_boxing()) { 1522 // Mark the successor block as parsed 1523 next_block->next_path_num(); 1524 } 1525 } else { // Path is live. 1526 // Update method data 1527 profile_not_taken_branch(); 1528 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1529 next_block, branch_block); 1530 } 1531 } 1532 1533 //------------------------------------do_if------------------------------------ 1534 void Parse::do_if(BoolTest::mask btest, Node* c) { 1535 int target_bci = iter().get_dest(); 1536 1537 Block* branch_block = successor_for_bci(target_bci); 1538 Block* next_block = successor_for_bci(iter().next_bci()); 1539 1540 float cnt; 1541 float prob = branch_prediction(cnt, btest, target_bci, c); 1542 float untaken_prob = 1.0 - prob; 1543 1544 if (prob == PROB_UNKNOWN) { 1545 if (PrintOpto && Verbose) { 1546 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1547 } 1548 repush_if_args(); // to gather stats on loop 1549 // We need to mark this branch as taken so that if we recompile we will 1550 // see that it is possible. In the tiered system the interpreter doesn't 1551 // do profiling and by the time we get to the lower tier from the interpreter 1552 // the path may be cold again. Make sure it doesn't look untaken 1553 profile_taken_branch(target_bci, !ProfileInterpreter); 1554 uncommon_trap(Deoptimization::Reason_unreached, 1555 Deoptimization::Action_reinterpret, 1556 NULL, "cold"); 1557 if (C->eliminate_boxing()) { 1558 // Mark the successor blocks as parsed 1559 branch_block->next_path_num(); 1560 next_block->next_path_num(); 1561 } 1562 return; 1563 } 1564 1565 // Sanity check the probability value 1566 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1567 1568 bool taken_if_true = true; 1569 // Convert BoolTest to canonical form: 1570 if (!BoolTest(btest).is_canonical()) { 1571 btest = BoolTest(btest).negate(); 1572 taken_if_true = false; 1573 // prob is NOT updated here; it remains the probability of the taken 1574 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1575 } 1576 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1577 1578 Node* tst0 = new BoolNode(c, btest); 1579 Node* tst = _gvn.transform(tst0); 1580 BoolTest::mask taken_btest = BoolTest::illegal; 1581 BoolTest::mask untaken_btest = BoolTest::illegal; 1582 1583 if (tst->is_Bool()) { 1584 // Refresh c from the transformed bool node, since it may be 1585 // simpler than the original c. Also re-canonicalize btest. 1586 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1587 // That can arise from statements like: if (x instanceof C) ... 1588 if (tst != tst0) { 1589 // Canonicalize one more time since transform can change it. 1590 btest = tst->as_Bool()->_test._test; 1591 if (!BoolTest(btest).is_canonical()) { 1592 // Reverse edges one more time... 1593 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1594 btest = tst->as_Bool()->_test._test; 1595 assert(BoolTest(btest).is_canonical(), "sanity"); 1596 taken_if_true = !taken_if_true; 1597 } 1598 c = tst->in(1); 1599 } 1600 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1601 taken_btest = taken_if_true ? btest : neg_btest; 1602 untaken_btest = taken_if_true ? neg_btest : btest; 1603 } 1604 1605 // Generate real control flow 1606 float true_prob = (taken_if_true ? prob : untaken_prob); 1607 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1608 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1609 Node* taken_branch = new IfTrueNode(iff); 1610 Node* untaken_branch = new IfFalseNode(iff); 1611 if (!taken_if_true) { // Finish conversion to canonical form 1612 Node* tmp = taken_branch; 1613 taken_branch = untaken_branch; 1614 untaken_branch = tmp; 1615 } 1616 1617 taken_branch = _gvn.transform(taken_branch); 1618 untaken_branch = _gvn.transform(untaken_branch); 1619 Node* taken_memory = NULL; 1620 Node* untaken_memory = NULL; 1621 1622 #if INCLUDE_SHENANDOAHGC 1623 ShenandoahBarrierNode::do_cmpp_if(*this, taken_branch, untaken_branch, taken_memory, untaken_memory); 1624 #endif 1625 1626 // Branch is taken: 1627 { PreserveJVMState pjvms(this); 1628 set_control(taken_branch); 1629 if (taken_memory != NULL) { 1630 set_all_memory(taken_memory); 1631 } 1632 1633 if (stopped()) { 1634 if (C->eliminate_boxing()) { 1635 // Mark the successor block as parsed 1636 branch_block->next_path_num(); 1637 } 1638 } else { 1639 // Update method data 1640 profile_taken_branch(target_bci); 1641 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1642 if (!stopped()) { 1643 merge(target_bci); 1644 } 1645 } 1646 } 1647 1648 set_control(untaken_branch); 1649 if (untaken_memory != NULL) { 1650 set_all_memory(untaken_memory); 1651 } 1652 1653 // Branch not taken. 1654 if (stopped()) { 1655 if (C->eliminate_boxing()) { 1656 // Mark the successor block as parsed 1657 next_block->next_path_num(); 1658 } 1659 } else { 1660 // Update method data 1661 profile_not_taken_branch(); 1662 adjust_map_after_if(untaken_btest, c, untaken_prob, 1663 next_block, branch_block); 1664 } 1665 } 1666 1667 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1668 // Don't want to speculate on uncommon traps when running with -Xcomp 1669 if (!UseInterpreter) { 1670 return false; 1671 } 1672 return (seems_never_taken(prob) && seems_stable_comparison()); 1673 } 1674 1675 void Parse::maybe_add_predicate_after_if(Block* path) { 1676 if (path->is_SEL_head() && path->preds_parsed() == 0) { 1677 // Add predicates at bci of if dominating the loop so traps can be 1678 // recorded on the if's profile data 1679 int bc_depth = repush_if_args(); 1680 add_predicate(); 1681 dec_sp(bc_depth); 1682 path->set_has_predicates(); 1683 } 1684 } 1685 1686 1687 //----------------------------adjust_map_after_if------------------------------ 1688 // Adjust the JVM state to reflect the result of taking this path. 1689 // Basically, it means inspecting the CmpNode controlling this 1690 // branch, seeing how it constrains a tested value, and then 1691 // deciding if it's worth our while to encode this constraint 1692 // as graph nodes in the current abstract interpretation map. 1693 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1694 Block* path, Block* other_path) { 1695 if (!c->is_Cmp()) { 1696 maybe_add_predicate_after_if(path); 1697 return; 1698 } 1699 1700 if (stopped() || btest == BoolTest::illegal) { 1701 return; // nothing to do 1702 } 1703 1704 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1705 1706 if (path_is_suitable_for_uncommon_trap(prob)) { 1707 repush_if_args(); 1708 uncommon_trap(Deoptimization::Reason_unstable_if, 1709 Deoptimization::Action_reinterpret, 1710 NULL, 1711 (is_fallthrough ? "taken always" : "taken never")); 1712 return; 1713 } 1714 1715 Node* val = c->in(1); 1716 Node* con = c->in(2); 1717 const Type* tcon = _gvn.type(con); 1718 const Type* tval = _gvn.type(val); 1719 bool have_con = tcon->singleton(); 1720 if (tval->singleton()) { 1721 if (!have_con) { 1722 // Swap, so constant is in con. 1723 con = val; 1724 tcon = tval; 1725 val = c->in(2); 1726 tval = _gvn.type(val); 1727 btest = BoolTest(btest).commute(); 1728 have_con = true; 1729 } else { 1730 // Do we have two constants? Then leave well enough alone. 1731 have_con = false; 1732 } 1733 } 1734 if (!have_con) { // remaining adjustments need a con 1735 maybe_add_predicate_after_if(path); 1736 return; 1737 } 1738 1739 sharpen_type_after_if(btest, con, tcon, val, tval); 1740 maybe_add_predicate_after_if(path); 1741 } 1742 1743 1744 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1745 Node* ldk; 1746 if (n->is_DecodeNKlass()) { 1747 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1748 return NULL; 1749 } else { 1750 ldk = n->in(1); 1751 } 1752 } else if (n->Opcode() != Op_LoadKlass) { 1753 return NULL; 1754 } else { 1755 ldk = n; 1756 } 1757 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1758 1759 Node* adr = ldk->in(MemNode::Address); 1760 intptr_t off = 0; 1761 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1762 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1763 return NULL; 1764 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1765 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1766 return NULL; 1767 1768 return obj; 1769 } 1770 1771 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1772 Node* con, const Type* tcon, 1773 Node* val, const Type* tval) { 1774 // Look for opportunities to sharpen the type of a node 1775 // whose klass is compared with a constant klass. 1776 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1777 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1778 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1779 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1780 // Found: 1781 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1782 // or the narrowOop equivalent. 1783 const Type* obj_type = _gvn.type(obj); 1784 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1785 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1786 tboth->higher_equal(obj_type)) { 1787 // obj has to be of the exact type Foo if the CmpP succeeds. 1788 int obj_in_map = map()->find_edge(obj); 1789 JVMState* jvms = this->jvms(); 1790 if (obj_in_map >= 0 && 1791 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1792 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1793 const Type* tcc = ccast->as_Type()->type(); 1794 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1795 // Delay transform() call to allow recovery of pre-cast value 1796 // at the control merge. 1797 _gvn.set_type_bottom(ccast); 1798 record_for_igvn(ccast); 1799 // Here's the payoff. 1800 replace_in_map(obj, ccast); 1801 } 1802 } 1803 } 1804 } 1805 1806 int val_in_map = map()->find_edge(val); 1807 if (val_in_map < 0) return; // replace_in_map would be useless 1808 { 1809 JVMState* jvms = this->jvms(); 1810 if (!(jvms->is_loc(val_in_map) || 1811 jvms->is_stk(val_in_map))) 1812 return; // again, it would be useless 1813 } 1814 1815 // Check for a comparison to a constant, and "know" that the compared 1816 // value is constrained on this path. 1817 assert(tcon->singleton(), ""); 1818 ConstraintCastNode* ccast = NULL; 1819 Node* cast = NULL; 1820 1821 switch (btest) { 1822 case BoolTest::eq: // Constant test? 1823 { 1824 const Type* tboth = tcon->join_speculative(tval); 1825 if (tboth == tval) break; // Nothing to gain. 1826 if (tcon->isa_int()) { 1827 ccast = new CastIINode(val, tboth); 1828 } else if (tcon == TypePtr::NULL_PTR) { 1829 // Cast to null, but keep the pointer identity temporarily live. 1830 ccast = new CastPPNode(val, tboth); 1831 } else { 1832 const TypeF* tf = tcon->isa_float_constant(); 1833 const TypeD* td = tcon->isa_double_constant(); 1834 // Exclude tests vs float/double 0 as these could be 1835 // either +0 or -0. Just because you are equal to +0 1836 // doesn't mean you ARE +0! 1837 // Note, following code also replaces Long and Oop values. 1838 if ((!tf || tf->_f != 0.0) && 1839 (!td || td->_d != 0.0)) 1840 cast = con; // Replace non-constant val by con. 1841 } 1842 } 1843 break; 1844 1845 case BoolTest::ne: 1846 if (tcon == TypePtr::NULL_PTR) { 1847 cast = cast_not_null(val, false); 1848 } 1849 break; 1850 1851 default: 1852 // (At this point we could record int range types with CastII.) 1853 break; 1854 } 1855 1856 if (ccast != NULL) { 1857 const Type* tcc = ccast->as_Type()->type(); 1858 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1859 // Delay transform() call to allow recovery of pre-cast value 1860 // at the control merge. 1861 ccast->set_req(0, control()); 1862 _gvn.set_type_bottom(ccast); 1863 record_for_igvn(ccast); 1864 cast = ccast; 1865 } 1866 1867 if (cast != NULL) { // Here's the payoff. 1868 replace_in_map(val, cast); 1869 } 1870 } 1871 1872 /** 1873 * Use speculative type to optimize CmpP node: if comparison is 1874 * against the low level class, cast the object to the speculative 1875 * type if any. CmpP should then go away. 1876 * 1877 * @param c expected CmpP node 1878 * @return result of CmpP on object casted to speculative type 1879 * 1880 */ 1881 Node* Parse::optimize_cmp_with_klass(Node* c) { 1882 // If this is transformed by the _gvn to a comparison with the low 1883 // level klass then we may be able to use speculation 1884 if (c->Opcode() == Op_CmpP && 1885 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1886 c->in(2)->is_Con()) { 1887 Node* load_klass = NULL; 1888 Node* decode = NULL; 1889 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1890 decode = c->in(1); 1891 load_klass = c->in(1)->in(1); 1892 } else { 1893 load_klass = c->in(1); 1894 } 1895 if (load_klass->in(2)->is_AddP()) { 1896 Node* addp = load_klass->in(2); 1897 Node* obj = addp->in(AddPNode::Address); 1898 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1899 if (obj_type->speculative_type_not_null() != NULL) { 1900 ciKlass* k = obj_type->speculative_type(); 1901 inc_sp(2); 1902 obj = maybe_cast_profiled_obj(obj, k); 1903 dec_sp(2); 1904 // Make the CmpP use the casted obj 1905 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1906 load_klass = load_klass->clone(); 1907 load_klass->set_req(2, addp); 1908 load_klass = _gvn.transform(load_klass); 1909 if (decode != NULL) { 1910 decode = decode->clone(); 1911 decode->set_req(1, load_klass); 1912 load_klass = _gvn.transform(decode); 1913 } 1914 c = c->clone(); 1915 c->set_req(1, load_klass); 1916 c = _gvn.transform(c); 1917 } 1918 } 1919 } 1920 return c; 1921 } 1922 1923 //------------------------------do_one_bytecode-------------------------------- 1924 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1925 void Parse::do_one_bytecode() { 1926 Node *a, *b, *c, *d; // Handy temps 1927 BoolTest::mask btest; 1928 int i; 1929 1930 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1931 1932 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1933 "out of nodes parsing method")) { 1934 return; 1935 } 1936 1937 #ifdef ASSERT 1938 // for setting breakpoints 1939 if (TraceOptoParse) { 1940 tty->print(" @"); 1941 dump_bci(bci()); 1942 tty->cr(); 1943 } 1944 #endif 1945 1946 switch (bc()) { 1947 case Bytecodes::_nop: 1948 // do nothing 1949 break; 1950 case Bytecodes::_lconst_0: 1951 push_pair(longcon(0)); 1952 break; 1953 1954 case Bytecodes::_lconst_1: 1955 push_pair(longcon(1)); 1956 break; 1957 1958 case Bytecodes::_fconst_0: 1959 push(zerocon(T_FLOAT)); 1960 break; 1961 1962 case Bytecodes::_fconst_1: 1963 push(makecon(TypeF::ONE)); 1964 break; 1965 1966 case Bytecodes::_fconst_2: 1967 push(makecon(TypeF::make(2.0f))); 1968 break; 1969 1970 case Bytecodes::_dconst_0: 1971 push_pair(zerocon(T_DOUBLE)); 1972 break; 1973 1974 case Bytecodes::_dconst_1: 1975 push_pair(makecon(TypeD::ONE)); 1976 break; 1977 1978 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1979 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1980 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1981 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1982 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1983 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1984 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1985 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1986 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1987 case Bytecodes::_aconst_null: push(null()); break; 1988 case Bytecodes::_ldc: 1989 case Bytecodes::_ldc_w: 1990 case Bytecodes::_ldc2_w: 1991 // If the constant is unresolved, run this BC once in the interpreter. 1992 { 1993 ciConstant constant = iter().get_constant(); 1994 if (!constant.is_valid() || 1995 (constant.basic_type() == T_OBJECT && 1996 !constant.as_object()->is_loaded())) { 1997 int index = iter().get_constant_pool_index(); 1998 constantTag tag = iter().get_constant_pool_tag(index); 1999 uncommon_trap(Deoptimization::make_trap_request 2000 (Deoptimization::Reason_unloaded, 2001 Deoptimization::Action_reinterpret, 2002 index), 2003 NULL, tag.internal_name()); 2004 break; 2005 } 2006 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 2007 "must be java_mirror of klass"); 2008 const Type* con_type = Type::make_from_constant(constant); 2009 if (con_type != NULL) { 2010 push_node(con_type->basic_type(), makecon(con_type)); 2011 } 2012 } 2013 2014 break; 2015 2016 case Bytecodes::_aload_0: 2017 push( local(0) ); 2018 break; 2019 case Bytecodes::_aload_1: 2020 push( local(1) ); 2021 break; 2022 case Bytecodes::_aload_2: 2023 push( local(2) ); 2024 break; 2025 case Bytecodes::_aload_3: 2026 push( local(3) ); 2027 break; 2028 case Bytecodes::_aload: 2029 push( local(iter().get_index()) ); 2030 break; 2031 2032 case Bytecodes::_fload_0: 2033 case Bytecodes::_iload_0: 2034 push( local(0) ); 2035 break; 2036 case Bytecodes::_fload_1: 2037 case Bytecodes::_iload_1: 2038 push( local(1) ); 2039 break; 2040 case Bytecodes::_fload_2: 2041 case Bytecodes::_iload_2: 2042 push( local(2) ); 2043 break; 2044 case Bytecodes::_fload_3: 2045 case Bytecodes::_iload_3: 2046 push( local(3) ); 2047 break; 2048 case Bytecodes::_fload: 2049 case Bytecodes::_iload: 2050 push( local(iter().get_index()) ); 2051 break; 2052 case Bytecodes::_lload_0: 2053 push_pair_local( 0 ); 2054 break; 2055 case Bytecodes::_lload_1: 2056 push_pair_local( 1 ); 2057 break; 2058 case Bytecodes::_lload_2: 2059 push_pair_local( 2 ); 2060 break; 2061 case Bytecodes::_lload_3: 2062 push_pair_local( 3 ); 2063 break; 2064 case Bytecodes::_lload: 2065 push_pair_local( iter().get_index() ); 2066 break; 2067 2068 case Bytecodes::_dload_0: 2069 push_pair_local(0); 2070 break; 2071 case Bytecodes::_dload_1: 2072 push_pair_local(1); 2073 break; 2074 case Bytecodes::_dload_2: 2075 push_pair_local(2); 2076 break; 2077 case Bytecodes::_dload_3: 2078 push_pair_local(3); 2079 break; 2080 case Bytecodes::_dload: 2081 push_pair_local(iter().get_index()); 2082 break; 2083 case Bytecodes::_fstore_0: 2084 case Bytecodes::_istore_0: 2085 case Bytecodes::_astore_0: 2086 set_local( 0, pop() ); 2087 break; 2088 case Bytecodes::_fstore_1: 2089 case Bytecodes::_istore_1: 2090 case Bytecodes::_astore_1: 2091 set_local( 1, pop() ); 2092 break; 2093 case Bytecodes::_fstore_2: 2094 case Bytecodes::_istore_2: 2095 case Bytecodes::_astore_2: 2096 set_local( 2, pop() ); 2097 break; 2098 case Bytecodes::_fstore_3: 2099 case Bytecodes::_istore_3: 2100 case Bytecodes::_astore_3: 2101 set_local( 3, pop() ); 2102 break; 2103 case Bytecodes::_fstore: 2104 case Bytecodes::_istore: 2105 case Bytecodes::_astore: 2106 set_local( iter().get_index(), pop() ); 2107 break; 2108 // long stores 2109 case Bytecodes::_lstore_0: 2110 set_pair_local( 0, pop_pair() ); 2111 break; 2112 case Bytecodes::_lstore_1: 2113 set_pair_local( 1, pop_pair() ); 2114 break; 2115 case Bytecodes::_lstore_2: 2116 set_pair_local( 2, pop_pair() ); 2117 break; 2118 case Bytecodes::_lstore_3: 2119 set_pair_local( 3, pop_pair() ); 2120 break; 2121 case Bytecodes::_lstore: 2122 set_pair_local( iter().get_index(), pop_pair() ); 2123 break; 2124 2125 // double stores 2126 case Bytecodes::_dstore_0: 2127 set_pair_local( 0, dstore_rounding(pop_pair()) ); 2128 break; 2129 case Bytecodes::_dstore_1: 2130 set_pair_local( 1, dstore_rounding(pop_pair()) ); 2131 break; 2132 case Bytecodes::_dstore_2: 2133 set_pair_local( 2, dstore_rounding(pop_pair()) ); 2134 break; 2135 case Bytecodes::_dstore_3: 2136 set_pair_local( 3, dstore_rounding(pop_pair()) ); 2137 break; 2138 case Bytecodes::_dstore: 2139 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 2140 break; 2141 2142 case Bytecodes::_pop: dec_sp(1); break; 2143 case Bytecodes::_pop2: dec_sp(2); break; 2144 case Bytecodes::_swap: 2145 a = pop(); 2146 b = pop(); 2147 push(a); 2148 push(b); 2149 break; 2150 case Bytecodes::_dup: 2151 a = pop(); 2152 push(a); 2153 push(a); 2154 break; 2155 case Bytecodes::_dup_x1: 2156 a = pop(); 2157 b = pop(); 2158 push( a ); 2159 push( b ); 2160 push( a ); 2161 break; 2162 case Bytecodes::_dup_x2: 2163 a = pop(); 2164 b = pop(); 2165 c = pop(); 2166 push( a ); 2167 push( c ); 2168 push( b ); 2169 push( a ); 2170 break; 2171 case Bytecodes::_dup2: 2172 a = pop(); 2173 b = pop(); 2174 push( b ); 2175 push( a ); 2176 push( b ); 2177 push( a ); 2178 break; 2179 2180 case Bytecodes::_dup2_x1: 2181 // before: .. c, b, a 2182 // after: .. b, a, c, b, a 2183 // not tested 2184 a = pop(); 2185 b = pop(); 2186 c = pop(); 2187 push( b ); 2188 push( a ); 2189 push( c ); 2190 push( b ); 2191 push( a ); 2192 break; 2193 case Bytecodes::_dup2_x2: 2194 // before: .. d, c, b, a 2195 // after: .. b, a, d, c, b, a 2196 // not tested 2197 a = pop(); 2198 b = pop(); 2199 c = pop(); 2200 d = pop(); 2201 push( b ); 2202 push( a ); 2203 push( d ); 2204 push( c ); 2205 push( b ); 2206 push( a ); 2207 break; 2208 2209 case Bytecodes::_arraylength: { 2210 // Must do null-check with value on expression stack 2211 Node *ary = null_check(peek(), T_ARRAY); 2212 // Compile-time detect of null-exception? 2213 if (stopped()) return; 2214 a = pop(); 2215 push(load_array_length(a)); 2216 break; 2217 } 2218 2219 case Bytecodes::_baload: array_load(T_BYTE); break; 2220 case Bytecodes::_caload: array_load(T_CHAR); break; 2221 case Bytecodes::_iaload: array_load(T_INT); break; 2222 case Bytecodes::_saload: array_load(T_SHORT); break; 2223 case Bytecodes::_faload: array_load(T_FLOAT); break; 2224 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2225 case Bytecodes::_laload: array_load(T_LONG); break; 2226 case Bytecodes::_daload: array_load(T_DOUBLE); break; 2227 case Bytecodes::_bastore: array_store(T_BYTE); break; 2228 case Bytecodes::_castore: array_store(T_CHAR); break; 2229 case Bytecodes::_iastore: array_store(T_INT); break; 2230 case Bytecodes::_sastore: array_store(T_SHORT); break; 2231 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2232 case Bytecodes::_aastore: array_store(T_OBJECT); break; 2233 case Bytecodes::_lastore: array_store(T_LONG); break; 2234 case Bytecodes::_dastore: array_store(T_DOUBLE); break; 2235 2236 case Bytecodes::_getfield: 2237 do_getfield(); 2238 break; 2239 2240 case Bytecodes::_getstatic: 2241 do_getstatic(); 2242 break; 2243 2244 case Bytecodes::_putfield: 2245 do_putfield(); 2246 break; 2247 2248 case Bytecodes::_putstatic: 2249 do_putstatic(); 2250 break; 2251 2252 case Bytecodes::_irem: 2253 do_irem(); 2254 break; 2255 case Bytecodes::_idiv: 2256 // Must keep both values on the expression-stack during null-check 2257 zero_check_int(peek()); 2258 // Compile-time detect of null-exception? 2259 if (stopped()) return; 2260 b = pop(); 2261 a = pop(); 2262 push( _gvn.transform( new DivINode(control(),a,b) ) ); 2263 break; 2264 case Bytecodes::_imul: 2265 b = pop(); a = pop(); 2266 push( _gvn.transform( new MulINode(a,b) ) ); 2267 break; 2268 case Bytecodes::_iadd: 2269 b = pop(); a = pop(); 2270 push( _gvn.transform( new AddINode(a,b) ) ); 2271 break; 2272 case Bytecodes::_ineg: 2273 a = pop(); 2274 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 2275 break; 2276 case Bytecodes::_isub: 2277 b = pop(); a = pop(); 2278 push( _gvn.transform( new SubINode(a,b) ) ); 2279 break; 2280 case Bytecodes::_iand: 2281 b = pop(); a = pop(); 2282 push( _gvn.transform( new AndINode(a,b) ) ); 2283 break; 2284 case Bytecodes::_ior: 2285 b = pop(); a = pop(); 2286 push( _gvn.transform( new OrINode(a,b) ) ); 2287 break; 2288 case Bytecodes::_ixor: 2289 b = pop(); a = pop(); 2290 push( _gvn.transform( new XorINode(a,b) ) ); 2291 break; 2292 case Bytecodes::_ishl: 2293 b = pop(); a = pop(); 2294 push( _gvn.transform( new LShiftINode(a,b) ) ); 2295 break; 2296 case Bytecodes::_ishr: 2297 b = pop(); a = pop(); 2298 push( _gvn.transform( new RShiftINode(a,b) ) ); 2299 break; 2300 case Bytecodes::_iushr: 2301 b = pop(); a = pop(); 2302 push( _gvn.transform( new URShiftINode(a,b) ) ); 2303 break; 2304 2305 case Bytecodes::_fneg: 2306 a = pop(); 2307 b = _gvn.transform(new NegFNode (a)); 2308 push(b); 2309 break; 2310 2311 case Bytecodes::_fsub: 2312 b = pop(); 2313 a = pop(); 2314 c = _gvn.transform( new SubFNode(a,b) ); 2315 d = precision_rounding(c); 2316 push( d ); 2317 break; 2318 2319 case Bytecodes::_fadd: 2320 b = pop(); 2321 a = pop(); 2322 c = _gvn.transform( new AddFNode(a,b) ); 2323 d = precision_rounding(c); 2324 push( d ); 2325 break; 2326 2327 case Bytecodes::_fmul: 2328 b = pop(); 2329 a = pop(); 2330 c = _gvn.transform( new MulFNode(a,b) ); 2331 d = precision_rounding(c); 2332 push( d ); 2333 break; 2334 2335 case Bytecodes::_fdiv: 2336 b = pop(); 2337 a = pop(); 2338 c = _gvn.transform( new DivFNode(0,a,b) ); 2339 d = precision_rounding(c); 2340 push( d ); 2341 break; 2342 2343 case Bytecodes::_frem: 2344 if (Matcher::has_match_rule(Op_ModF)) { 2345 // Generate a ModF node. 2346 b = pop(); 2347 a = pop(); 2348 c = _gvn.transform( new ModFNode(0,a,b) ); 2349 d = precision_rounding(c); 2350 push( d ); 2351 } 2352 else { 2353 // Generate a call. 2354 modf(); 2355 } 2356 break; 2357 2358 case Bytecodes::_fcmpl: 2359 b = pop(); 2360 a = pop(); 2361 c = _gvn.transform( new CmpF3Node( a, b)); 2362 push(c); 2363 break; 2364 case Bytecodes::_fcmpg: 2365 b = pop(); 2366 a = pop(); 2367 2368 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 2369 // which negates the result sign except for unordered. Flip the unordered 2370 // as well by using CmpF3 which implements unordered-lesser instead of 2371 // unordered-greater semantics. Finally, commute the result bits. Result 2372 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 2373 c = _gvn.transform( new CmpF3Node( b, a)); 2374 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2375 push(c); 2376 break; 2377 2378 case Bytecodes::_f2i: 2379 a = pop(); 2380 push(_gvn.transform(new ConvF2INode(a))); 2381 break; 2382 2383 case Bytecodes::_d2i: 2384 a = pop_pair(); 2385 b = _gvn.transform(new ConvD2INode(a)); 2386 push( b ); 2387 break; 2388 2389 case Bytecodes::_f2d: 2390 a = pop(); 2391 b = _gvn.transform( new ConvF2DNode(a)); 2392 push_pair( b ); 2393 break; 2394 2395 case Bytecodes::_d2f: 2396 a = pop_pair(); 2397 b = _gvn.transform( new ConvD2FNode(a)); 2398 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 2399 //b = _gvn.transform(new RoundFloatNode(0, b) ); 2400 push( b ); 2401 break; 2402 2403 case Bytecodes::_l2f: 2404 if (Matcher::convL2FSupported()) { 2405 a = pop_pair(); 2406 b = _gvn.transform( new ConvL2FNode(a)); 2407 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 2408 // Rather than storing the result into an FP register then pushing 2409 // out to memory to round, the machine instruction that implements 2410 // ConvL2D is responsible for rounding. 2411 // c = precision_rounding(b); 2412 c = _gvn.transform(b); 2413 push(c); 2414 } else { 2415 l2f(); 2416 } 2417 break; 2418 2419 case Bytecodes::_l2d: 2420 a = pop_pair(); 2421 b = _gvn.transform( new ConvL2DNode(a)); 2422 // For i486.ad, rounding is always necessary (see _l2f above). 2423 // c = dprecision_rounding(b); 2424 c = _gvn.transform(b); 2425 push_pair(c); 2426 break; 2427 2428 case Bytecodes::_f2l: 2429 a = pop(); 2430 b = _gvn.transform( new ConvF2LNode(a)); 2431 push_pair(b); 2432 break; 2433 2434 case Bytecodes::_d2l: 2435 a = pop_pair(); 2436 b = _gvn.transform( new ConvD2LNode(a)); 2437 push_pair(b); 2438 break; 2439 2440 case Bytecodes::_dsub: 2441 b = pop_pair(); 2442 a = pop_pair(); 2443 c = _gvn.transform( new SubDNode(a,b) ); 2444 d = dprecision_rounding(c); 2445 push_pair( d ); 2446 break; 2447 2448 case Bytecodes::_dadd: 2449 b = pop_pair(); 2450 a = pop_pair(); 2451 c = _gvn.transform( new AddDNode(a,b) ); 2452 d = dprecision_rounding(c); 2453 push_pair( d ); 2454 break; 2455 2456 case Bytecodes::_dmul: 2457 b = pop_pair(); 2458 a = pop_pair(); 2459 c = _gvn.transform( new MulDNode(a,b) ); 2460 d = dprecision_rounding(c); 2461 push_pair( d ); 2462 break; 2463 2464 case Bytecodes::_ddiv: 2465 b = pop_pair(); 2466 a = pop_pair(); 2467 c = _gvn.transform( new DivDNode(0,a,b) ); 2468 d = dprecision_rounding(c); 2469 push_pair( d ); 2470 break; 2471 2472 case Bytecodes::_dneg: 2473 a = pop_pair(); 2474 b = _gvn.transform(new NegDNode (a)); 2475 push_pair(b); 2476 break; 2477 2478 case Bytecodes::_drem: 2479 if (Matcher::has_match_rule(Op_ModD)) { 2480 // Generate a ModD node. 2481 b = pop_pair(); 2482 a = pop_pair(); 2483 // a % b 2484 2485 c = _gvn.transform( new ModDNode(0,a,b) ); 2486 d = dprecision_rounding(c); 2487 push_pair( d ); 2488 } 2489 else { 2490 // Generate a call. 2491 modd(); 2492 } 2493 break; 2494 2495 case Bytecodes::_dcmpl: 2496 b = pop_pair(); 2497 a = pop_pair(); 2498 c = _gvn.transform( new CmpD3Node( a, b)); 2499 push(c); 2500 break; 2501 2502 case Bytecodes::_dcmpg: 2503 b = pop_pair(); 2504 a = pop_pair(); 2505 // Same as dcmpl but need to flip the unordered case. 2506 // Commute the inputs, which negates the result sign except for unordered. 2507 // Flip the unordered as well by using CmpD3 which implements 2508 // unordered-lesser instead of unordered-greater semantics. 2509 // Finally, negate the result bits. Result is same as using a 2510 // CmpD3Greater except we did it with CmpD3 alone. 2511 c = _gvn.transform( new CmpD3Node( b, a)); 2512 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2513 push(c); 2514 break; 2515 2516 2517 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2518 case Bytecodes::_land: 2519 b = pop_pair(); 2520 a = pop_pair(); 2521 c = _gvn.transform( new AndLNode(a,b) ); 2522 push_pair(c); 2523 break; 2524 case Bytecodes::_lor: 2525 b = pop_pair(); 2526 a = pop_pair(); 2527 c = _gvn.transform( new OrLNode(a,b) ); 2528 push_pair(c); 2529 break; 2530 case Bytecodes::_lxor: 2531 b = pop_pair(); 2532 a = pop_pair(); 2533 c = _gvn.transform( new XorLNode(a,b) ); 2534 push_pair(c); 2535 break; 2536 2537 case Bytecodes::_lshl: 2538 b = pop(); // the shift count 2539 a = pop_pair(); // value to be shifted 2540 c = _gvn.transform( new LShiftLNode(a,b) ); 2541 push_pair(c); 2542 break; 2543 case Bytecodes::_lshr: 2544 b = pop(); // the shift count 2545 a = pop_pair(); // value to be shifted 2546 c = _gvn.transform( new RShiftLNode(a,b) ); 2547 push_pair(c); 2548 break; 2549 case Bytecodes::_lushr: 2550 b = pop(); // the shift count 2551 a = pop_pair(); // value to be shifted 2552 c = _gvn.transform( new URShiftLNode(a,b) ); 2553 push_pair(c); 2554 break; 2555 case Bytecodes::_lmul: 2556 b = pop_pair(); 2557 a = pop_pair(); 2558 c = _gvn.transform( new MulLNode(a,b) ); 2559 push_pair(c); 2560 break; 2561 2562 case Bytecodes::_lrem: 2563 // Must keep both values on the expression-stack during null-check 2564 assert(peek(0) == top(), "long word order"); 2565 zero_check_long(peek(1)); 2566 // Compile-time detect of null-exception? 2567 if (stopped()) return; 2568 b = pop_pair(); 2569 a = pop_pair(); 2570 c = _gvn.transform( new ModLNode(control(),a,b) ); 2571 push_pair(c); 2572 break; 2573 2574 case Bytecodes::_ldiv: 2575 // Must keep both values on the expression-stack during null-check 2576 assert(peek(0) == top(), "long word order"); 2577 zero_check_long(peek(1)); 2578 // Compile-time detect of null-exception? 2579 if (stopped()) return; 2580 b = pop_pair(); 2581 a = pop_pair(); 2582 c = _gvn.transform( new DivLNode(control(),a,b) ); 2583 push_pair(c); 2584 break; 2585 2586 case Bytecodes::_ladd: 2587 b = pop_pair(); 2588 a = pop_pair(); 2589 c = _gvn.transform( new AddLNode(a,b) ); 2590 push_pair(c); 2591 break; 2592 case Bytecodes::_lsub: 2593 b = pop_pair(); 2594 a = pop_pair(); 2595 c = _gvn.transform( new SubLNode(a,b) ); 2596 push_pair(c); 2597 break; 2598 case Bytecodes::_lcmp: 2599 // Safepoints are now inserted _before_ branches. The long-compare 2600 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2601 // slew of control flow. These are usually followed by a CmpI vs zero and 2602 // a branch; this pattern then optimizes to the obvious long-compare and 2603 // branch. However, if the branch is backwards there's a Safepoint 2604 // inserted. The inserted Safepoint captures the JVM state at the 2605 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2606 // long-compare is used to control a loop the debug info will force 2607 // computation of the 3-way value, even though the generated code uses a 2608 // long-compare and branch. We try to rectify the situation by inserting 2609 // a SafePoint here and have it dominate and kill the safepoint added at a 2610 // following backwards branch. At this point the JVM state merely holds 2 2611 // longs but not the 3-way value. 2612 if( UseLoopSafepoints ) { 2613 switch( iter().next_bc() ) { 2614 case Bytecodes::_ifgt: 2615 case Bytecodes::_iflt: 2616 case Bytecodes::_ifge: 2617 case Bytecodes::_ifle: 2618 case Bytecodes::_ifne: 2619 case Bytecodes::_ifeq: 2620 // If this is a backwards branch in the bytecodes, add Safepoint 2621 maybe_add_safepoint(iter().next_get_dest()); 2622 default: 2623 break; 2624 } 2625 } 2626 b = pop_pair(); 2627 a = pop_pair(); 2628 c = _gvn.transform( new CmpL3Node( a, b )); 2629 push(c); 2630 break; 2631 2632 case Bytecodes::_lneg: 2633 a = pop_pair(); 2634 b = _gvn.transform( new SubLNode(longcon(0),a)); 2635 push_pair(b); 2636 break; 2637 case Bytecodes::_l2i: 2638 a = pop_pair(); 2639 push( _gvn.transform( new ConvL2INode(a))); 2640 break; 2641 case Bytecodes::_i2l: 2642 a = pop(); 2643 b = _gvn.transform( new ConvI2LNode(a)); 2644 push_pair(b); 2645 break; 2646 case Bytecodes::_i2b: 2647 // Sign extend 2648 a = pop(); 2649 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) ); 2650 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) ); 2651 push( a ); 2652 break; 2653 case Bytecodes::_i2s: 2654 a = pop(); 2655 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) ); 2656 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) ); 2657 push( a ); 2658 break; 2659 case Bytecodes::_i2c: 2660 a = pop(); 2661 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) ); 2662 break; 2663 2664 case Bytecodes::_i2f: 2665 a = pop(); 2666 b = _gvn.transform( new ConvI2FNode(a) ) ; 2667 c = precision_rounding(b); 2668 push (b); 2669 break; 2670 2671 case Bytecodes::_i2d: 2672 a = pop(); 2673 b = _gvn.transform( new ConvI2DNode(a)); 2674 push_pair(b); 2675 break; 2676 2677 case Bytecodes::_iinc: // Increment local 2678 i = iter().get_index(); // Get local index 2679 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2680 break; 2681 2682 // Exit points of synchronized methods must have an unlock node 2683 case Bytecodes::_return: 2684 return_current(NULL); 2685 break; 2686 2687 case Bytecodes::_ireturn: 2688 case Bytecodes::_areturn: 2689 case Bytecodes::_freturn: 2690 return_current(pop()); 2691 break; 2692 case Bytecodes::_lreturn: 2693 return_current(pop_pair()); 2694 break; 2695 case Bytecodes::_dreturn: 2696 return_current(pop_pair()); 2697 break; 2698 2699 case Bytecodes::_athrow: 2700 // null exception oop throws NULL pointer exception 2701 null_check(peek()); 2702 if (stopped()) return; 2703 // Hook the thrown exception directly to subsequent handlers. 2704 if (BailoutToInterpreterForThrows) { 2705 // Keep method interpreted from now on. 2706 uncommon_trap(Deoptimization::Reason_unhandled, 2707 Deoptimization::Action_make_not_compilable); 2708 return; 2709 } 2710 if (env()->jvmti_can_post_on_exceptions()) { 2711 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2712 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2713 } 2714 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2715 add_exception_state(make_exception_state(peek())); 2716 break; 2717 2718 case Bytecodes::_goto: // fall through 2719 case Bytecodes::_goto_w: { 2720 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2721 2722 // If this is a backwards branch in the bytecodes, add Safepoint 2723 maybe_add_safepoint(target_bci); 2724 2725 // Update method data 2726 profile_taken_branch(target_bci); 2727 2728 // Merge the current control into the target basic block 2729 merge(target_bci); 2730 2731 // See if we can get some profile data and hand it off to the next block 2732 Block *target_block = block()->successor_for_bci(target_bci); 2733 if (target_block->pred_count() != 1) break; 2734 ciMethodData* methodData = method()->method_data(); 2735 if (!methodData->is_mature()) break; 2736 ciProfileData* data = methodData->bci_to_data(bci()); 2737 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch"); 2738 int taken = ((ciJumpData*)data)->taken(); 2739 taken = method()->scale_count(taken); 2740 target_block->set_count(taken); 2741 break; 2742 } 2743 2744 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2745 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2746 handle_if_null: 2747 // If this is a backwards branch in the bytecodes, add Safepoint 2748 maybe_add_safepoint(iter().get_dest()); 2749 a = null(); 2750 b = pop(); 2751 if (!_gvn.type(b)->speculative_maybe_null() && 2752 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2753 inc_sp(1); 2754 Node* null_ctl = top(); 2755 b = null_check_oop(b, &null_ctl, true, true, true); 2756 assert(null_ctl->is_top(), "no null control here"); 2757 dec_sp(1); 2758 } else if (_gvn.type(b)->speculative_always_null() && 2759 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2760 inc_sp(1); 2761 b = null_assert(b); 2762 dec_sp(1); 2763 } 2764 c = _gvn.transform( new CmpPNode(b, a) ); 2765 do_ifnull(btest, c); 2766 break; 2767 2768 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2769 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2770 handle_if_acmp: 2771 // If this is a backwards branch in the bytecodes, add Safepoint 2772 maybe_add_safepoint(iter().get_dest()); 2773 a = pop(); 2774 b = pop(); 2775 #if INCLUDE_SHENANDOAHGC 2776 if (UseShenandoahGC && ShenandoahAcmpBarrier && ShenandoahVerifyOptoBarriers) { 2777 a = access_resolve_for_write(a); 2778 b = access_resolve_for_write(b); 2779 } 2780 #endif 2781 c = _gvn.transform( new CmpPNode(b, a) ); 2782 c = optimize_cmp_with_klass(c); 2783 do_if(btest, c); 2784 break; 2785 2786 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2787 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2788 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2789 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2790 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2791 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2792 handle_ifxx: 2793 // If this is a backwards branch in the bytecodes, add Safepoint 2794 maybe_add_safepoint(iter().get_dest()); 2795 a = _gvn.intcon(0); 2796 b = pop(); 2797 c = _gvn.transform( new CmpINode(b, a) ); 2798 do_if(btest, c); 2799 break; 2800 2801 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2802 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2803 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2804 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2805 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2806 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2807 handle_if_icmp: 2808 // If this is a backwards branch in the bytecodes, add Safepoint 2809 maybe_add_safepoint(iter().get_dest()); 2810 a = pop(); 2811 b = pop(); 2812 c = _gvn.transform( new CmpINode( b, a ) ); 2813 do_if(btest, c); 2814 break; 2815 2816 case Bytecodes::_tableswitch: 2817 do_tableswitch(); 2818 break; 2819 2820 case Bytecodes::_lookupswitch: 2821 do_lookupswitch(); 2822 break; 2823 2824 case Bytecodes::_invokestatic: 2825 case Bytecodes::_invokedynamic: 2826 case Bytecodes::_invokespecial: 2827 case Bytecodes::_invokevirtual: 2828 case Bytecodes::_invokeinterface: 2829 do_call(); 2830 break; 2831 case Bytecodes::_checkcast: 2832 do_checkcast(); 2833 break; 2834 case Bytecodes::_instanceof: 2835 do_instanceof(); 2836 break; 2837 case Bytecodes::_anewarray: 2838 do_anewarray(); 2839 break; 2840 case Bytecodes::_newarray: 2841 do_newarray((BasicType)iter().get_index()); 2842 break; 2843 case Bytecodes::_multianewarray: 2844 do_multianewarray(); 2845 break; 2846 case Bytecodes::_new: 2847 do_new(); 2848 break; 2849 2850 case Bytecodes::_jsr: 2851 case Bytecodes::_jsr_w: 2852 do_jsr(); 2853 break; 2854 2855 case Bytecodes::_ret: 2856 do_ret(); 2857 break; 2858 2859 2860 case Bytecodes::_monitorenter: 2861 do_monitor_enter(); 2862 break; 2863 2864 case Bytecodes::_monitorexit: 2865 do_monitor_exit(); 2866 break; 2867 2868 case Bytecodes::_breakpoint: 2869 // Breakpoint set concurrently to compile 2870 // %%% use an uncommon trap? 2871 C->record_failure("breakpoint in method"); 2872 return; 2873 2874 default: 2875 #ifndef PRODUCT 2876 map()->dump(99); 2877 #endif 2878 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2879 ShouldNotReachHere(); 2880 } 2881 2882 #ifndef PRODUCT 2883 IdealGraphPrinter *printer = C->printer(); 2884 if (printer && printer->should_print(1)) { 2885 char buffer[256]; 2886 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2887 bool old = printer->traverse_outs(); 2888 printer->set_traverse_outs(true); 2889 printer->print_method(buffer, 4); 2890 printer->set_traverse_outs(old); 2891 } 2892 #endif 2893 }