1 /* 2 * Copyright (c) 1998, 2023, 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/vmSymbols.hpp" 28 #include "compiler/compileLog.hpp" 29 #include "interpreter/linkResolver.hpp" 30 #include "jvm_io.h" 31 #include "memory/resourceArea.hpp" 32 #include "memory/universe.hpp" 33 #include "oops/oop.inline.hpp" 34 #include "opto/addnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/divnode.hpp" 38 #include "opto/idealGraphPrinter.hpp" 39 #include "opto/matcher.hpp" 40 #include "opto/memnode.hpp" 41 #include "opto/mulnode.hpp" 42 #include "opto/opaquenode.hpp" 43 #include "opto/parse.hpp" 44 #include "opto/runtime.hpp" 45 #include "runtime/deoptimization.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 #ifndef PRODUCT 49 extern uint explicit_null_checks_inserted, 50 explicit_null_checks_elided; 51 #endif 52 53 //---------------------------------array_load---------------------------------- 54 void Parse::array_load(BasicType bt) { 55 const Type* elemtype = Type::TOP; 56 bool big_val = bt == T_DOUBLE || bt == T_LONG; 57 Node* adr = array_addressing(bt, 0, elemtype); 58 if (stopped()) return; // guaranteed null or range check 59 60 pop(); // index (already used) 61 Node* array = pop(); // the array itself 62 63 if (elemtype == TypeInt::BOOL) { 64 bt = T_BOOLEAN; 65 } 66 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 67 68 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt, 69 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD); 70 if (big_val) { 71 push_pair(ld); 72 } else { 73 push(ld); 74 } 75 } 76 77 78 //--------------------------------array_store---------------------------------- 79 void Parse::array_store(BasicType bt) { 80 const Type* elemtype = Type::TOP; 81 bool big_val = bt == T_DOUBLE || bt == T_LONG; 82 Node* adr = array_addressing(bt, big_val ? 2 : 1, elemtype); 83 if (stopped()) return; // guaranteed null or range check 84 if (bt == T_OBJECT) { 85 array_store_check(); 86 if (stopped()) { 87 return; 88 } 89 } 90 Node* val; // Oop to store 91 if (big_val) { 92 val = pop_pair(); 93 } else { 94 val = pop(); 95 } 96 pop(); // index (already used) 97 Node* array = pop(); // the array itself 98 99 if (elemtype == TypeInt::BOOL) { 100 bt = T_BOOLEAN; 101 } 102 103 if (DoPartialEscapeAnalysis && bt == T_OBJECT) { 104 PartialEscapeAnalysis* pea = PEA(); 105 PEAState& as = jvms()->alloc_state(); 106 VirtualState* obj_vs = as.as_virtual(pea, val); // obj is tracked and Virtual 107 VirtualState* ary_vs = as.as_virtual(pea, array); // array is not-tracked or Escaped 108 109 if (ary_vs == nullptr && obj_vs != nullptr) { 110 val = as.materialize(this, val); 111 } 112 } 113 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 114 115 access_store_at(array, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY); 116 } 117 118 119 //------------------------------array_addressing------------------------------- 120 // Pull array and index from the stack. Compute pointer-to-element. 121 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) { 122 Node *idx = peek(0+vals); // Get from stack without popping 123 Node *ary = peek(1+vals); // in case of exception 124 125 // Null check the array base, with correct stack contents 126 ary = null_check(ary, T_ARRAY); 127 // Compile-time detect of null-exception? 128 if (stopped()) return top(); 129 130 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 131 const TypeInt* sizetype = arytype->size(); 132 elemtype = arytype->elem(); 133 134 if (UseUniqueSubclasses) { 135 const Type* el = elemtype->make_ptr(); 136 if (el && el->isa_instptr()) { 137 const TypeInstPtr* toop = el->is_instptr(); 138 if (toop->instance_klass()->unique_concrete_subklass()) { 139 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 140 const Type* subklass = Type::get_const_type(toop->instance_klass()); 141 elemtype = subklass->join_speculative(el); 142 } 143 } 144 } 145 146 // Check for big class initializers with all constant offsets 147 // feeding into a known-size array. 148 const TypeInt* idxtype = _gvn.type(idx)->is_int(); 149 // See if the highest idx value is less than the lowest array bound, 150 // and if the idx value cannot be negative: 151 bool need_range_check = true; 152 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 153 need_range_check = false; 154 if (C->log() != nullptr) C->log()->elem("observe that='!need_range_check'"); 155 } 156 157 if (!arytype->is_loaded()) { 158 // Only fails for some -Xcomp runs 159 // The class is unloaded. We have to run this bytecode in the interpreter. 160 ciKlass* klass = arytype->unloaded_klass(); 161 162 uncommon_trap(Deoptimization::Reason_unloaded, 163 Deoptimization::Action_reinterpret, 164 klass, "!loaded array"); 165 return top(); 166 } 167 168 // Do the range check 169 if (need_range_check) { 170 Node* tst; 171 if (sizetype->_hi <= 0) { 172 // The greatest array bound is negative, so we can conclude that we're 173 // compiling unreachable code, but the unsigned compare trick used below 174 // only works with non-negative lengths. Instead, hack "tst" to be zero so 175 // the uncommon_trap path will always be taken. 176 tst = _gvn.intcon(0); 177 } else { 178 // Range is constant in array-oop, so we can use the original state of mem 179 Node* len = load_array_length(ary); 180 181 // Test length vs index (standard trick using unsigned compare) 182 Node* chk = _gvn.transform( new CmpUNode(idx, len) ); 183 BoolTest::mask btest = BoolTest::lt; 184 tst = _gvn.transform( new BoolNode(chk, btest) ); 185 } 186 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 187 _gvn.set_type(rc, rc->Value(&_gvn)); 188 if (!tst->is_Con()) { 189 record_for_igvn(rc); 190 } 191 set_control(_gvn.transform(new IfTrueNode(rc))); 192 // Branch to failure if out of bounds 193 { 194 PreserveJVMState pjvms(this); 195 set_control(_gvn.transform(new IfFalseNode(rc))); 196 if (C->allow_range_check_smearing()) { 197 // Do not use builtin_throw, since range checks are sometimes 198 // made more stringent by an optimistic transformation. 199 // This creates "tentative" range checks at this point, 200 // which are not guaranteed to throw exceptions. 201 // See IfNode::Ideal, is_range_check, adjust_check. 202 uncommon_trap(Deoptimization::Reason_range_check, 203 Deoptimization::Action_make_not_entrant, 204 nullptr, "range_check"); 205 } else { 206 // If we have already recompiled with the range-check-widening 207 // heroic optimization turned off, then we must really be throwing 208 // range check exceptions. 209 builtin_throw(Deoptimization::Reason_range_check); 210 } 211 } 212 } 213 // Check for always knowing you are throwing a range-check exception 214 if (stopped()) return top(); 215 216 // Make array address computation control dependent to prevent it 217 // from floating above the range check during loop optimizations. 218 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 219 assert(ptr != top(), "top should go hand-in-hand with stopped"); 220 221 return ptr; 222 } 223 224 225 // returns IfNode 226 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 227 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 228 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 229 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 230 return iff; 231 } 232 233 234 // sentinel value for the target bci to mark never taken branches 235 // (according to profiling) 236 static const int never_reached = INT_MAX; 237 238 //------------------------------helper for tableswitch------------------------- 239 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) { 240 // True branch, use existing map info 241 { PreserveJVMState pjvms(this); 242 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 243 set_control( iftrue ); 244 if (unc) { 245 repush_if_args(); 246 uncommon_trap(Deoptimization::Reason_unstable_if, 247 Deoptimization::Action_reinterpret, 248 nullptr, 249 "taken always"); 250 } else { 251 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 252 merge_new_path(dest_bci_if_true); 253 } 254 } 255 256 // False branch 257 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 258 set_control( iffalse ); 259 } 260 261 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) { 262 // True branch, use existing map info 263 { PreserveJVMState pjvms(this); 264 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 265 set_control( iffalse ); 266 if (unc) { 267 repush_if_args(); 268 uncommon_trap(Deoptimization::Reason_unstable_if, 269 Deoptimization::Action_reinterpret, 270 nullptr, 271 "taken never"); 272 } else { 273 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 274 merge_new_path(dest_bci_if_true); 275 } 276 } 277 278 // False branch 279 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 280 set_control( iftrue ); 281 } 282 283 void Parse::jump_if_always_fork(int dest_bci, bool unc) { 284 // False branch, use existing map and control() 285 if (unc) { 286 repush_if_args(); 287 uncommon_trap(Deoptimization::Reason_unstable_if, 288 Deoptimization::Action_reinterpret, 289 nullptr, 290 "taken never"); 291 } else { 292 assert(dest_bci != never_reached, "inconsistent dest"); 293 merge_new_path(dest_bci); 294 } 295 } 296 297 298 extern "C" { 299 static int jint_cmp(const void *i, const void *j) { 300 int a = *(jint *)i; 301 int b = *(jint *)j; 302 return a > b ? 1 : a < b ? -1 : 0; 303 } 304 } 305 306 307 class SwitchRange : public StackObj { 308 // a range of integers coupled with a bci destination 309 jint _lo; // inclusive lower limit 310 jint _hi; // inclusive upper limit 311 int _dest; 312 float _cnt; // how many times this range was hit according to profiling 313 314 public: 315 jint lo() const { return _lo; } 316 jint hi() const { return _hi; } 317 int dest() const { return _dest; } 318 bool is_singleton() const { return _lo == _hi; } 319 float cnt() const { return _cnt; } 320 321 void setRange(jint lo, jint hi, int dest, float cnt) { 322 assert(lo <= hi, "must be a non-empty range"); 323 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt; 324 assert(_cnt >= 0, ""); 325 } 326 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) { 327 assert(lo <= hi, "must be a non-empty range"); 328 if (lo == _hi+1) { 329 // see merge_ranges() comment below 330 if (trim_ranges) { 331 if (cnt == 0) { 332 if (_cnt != 0) { 333 return false; 334 } 335 if (dest != _dest) { 336 _dest = never_reached; 337 } 338 } else { 339 if (_cnt == 0) { 340 return false; 341 } 342 if (dest != _dest) { 343 return false; 344 } 345 } 346 } else { 347 if (dest != _dest) { 348 return false; 349 } 350 } 351 _hi = hi; 352 _cnt += cnt; 353 return true; 354 } 355 return false; 356 } 357 358 void set (jint value, int dest, float cnt) { 359 setRange(value, value, dest, cnt); 360 } 361 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) { 362 return adjoinRange(value, value, dest, cnt, trim_ranges); 363 } 364 bool adjoin(SwitchRange& other) { 365 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false); 366 } 367 368 void print() { 369 if (is_singleton()) 370 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 371 else if (lo() == min_jint) 372 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 373 else if (hi() == max_jint) 374 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 375 else 376 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 377 } 378 }; 379 380 // We try to minimize the number of ranges and the size of the taken 381 // ones using profiling data. When ranges are created, 382 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 383 // if both were never hit or both were hit to build longer unreached 384 // ranges. Here, we now merge adjoining ranges with the same 385 // destination and finally set destination of unreached ranges to the 386 // special value never_reached because it can help minimize the number 387 // of tests that are necessary. 388 // 389 // For instance: 390 // [0, 1] to target1 sometimes taken 391 // [1, 2] to target1 never taken 392 // [2, 3] to target2 never taken 393 // would lead to: 394 // [0, 1] to target1 sometimes taken 395 // [1, 3] never taken 396 // 397 // (first 2 ranges to target1 are not merged) 398 static void merge_ranges(SwitchRange* ranges, int& rp) { 399 if (rp == 0) { 400 return; 401 } 402 int shift = 0; 403 for (int j = 0; j < rp; j++) { 404 SwitchRange& r1 = ranges[j-shift]; 405 SwitchRange& r2 = ranges[j+1]; 406 if (r1.adjoin(r2)) { 407 shift++; 408 } else if (shift > 0) { 409 ranges[j+1-shift] = r2; 410 } 411 } 412 rp -= shift; 413 for (int j = 0; j <= rp; j++) { 414 SwitchRange& r = ranges[j]; 415 if (r.cnt() == 0 && r.dest() != never_reached) { 416 r.setRange(r.lo(), r.hi(), never_reached, r.cnt()); 417 } 418 } 419 } 420 421 //-------------------------------do_tableswitch-------------------------------- 422 void Parse::do_tableswitch() { 423 // Get information about tableswitch 424 int default_dest = iter().get_dest_table(0); 425 jint lo_index = iter().get_int_table(1); 426 jint hi_index = iter().get_int_table(2); 427 int len = hi_index - lo_index + 1; 428 429 if (len < 1) { 430 // If this is a backward branch, add safepoint 431 maybe_add_safepoint(default_dest); 432 pop(); // the effect of the instruction execution on the operand stack 433 merge(default_dest); 434 return; 435 } 436 437 ciMethodData* methodData = method()->method_data(); 438 ciMultiBranchData* profile = nullptr; 439 if (methodData->is_mature() && UseSwitchProfiling) { 440 ciProfileData* data = methodData->bci_to_data(bci()); 441 if (data != nullptr && data->is_MultiBranchData()) { 442 profile = (ciMultiBranchData*)data; 443 } 444 } 445 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 446 447 // generate decision tree, using trichotomy when possible 448 int rnum = len+2; 449 bool makes_backward_branch = (default_dest <= bci()); 450 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 451 int rp = -1; 452 if (lo_index != min_jint) { 453 float cnt = 1.0F; 454 if (profile != nullptr) { 455 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F); 456 } 457 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt); 458 } 459 for (int j = 0; j < len; j++) { 460 jint match_int = lo_index+j; 461 int dest = iter().get_dest_table(j+3); 462 makes_backward_branch |= (dest <= bci()); 463 float cnt = 1.0F; 464 if (profile != nullptr) { 465 cnt = (float)profile->count_at(j); 466 } 467 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) { 468 ranges[++rp].set(match_int, dest, cnt); 469 } 470 } 471 jint highest = lo_index+(len-1); 472 assert(ranges[rp].hi() == highest, ""); 473 if (highest != max_jint) { 474 float cnt = 1.0F; 475 if (profile != nullptr) { 476 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F); 477 } 478 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) { 479 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt); 480 } 481 } 482 assert(rp < len+2, "not too many ranges"); 483 484 if (trim_ranges) { 485 merge_ranges(ranges, rp); 486 } 487 488 // Safepoint in case if backward branch observed 489 if (makes_backward_branch) { 490 add_safepoint(); 491 } 492 493 Node* lookup = pop(); // lookup value 494 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 495 } 496 497 498 //------------------------------do_lookupswitch-------------------------------- 499 void Parse::do_lookupswitch() { 500 // Get information about lookupswitch 501 int default_dest = iter().get_dest_table(0); 502 jint len = iter().get_int_table(1); 503 504 if (len < 1) { // If this is a backward branch, add safepoint 505 maybe_add_safepoint(default_dest); 506 pop(); // the effect of the instruction execution on the operand stack 507 merge(default_dest); 508 return; 509 } 510 511 ciMethodData* methodData = method()->method_data(); 512 ciMultiBranchData* profile = nullptr; 513 if (methodData->is_mature() && UseSwitchProfiling) { 514 ciProfileData* data = methodData->bci_to_data(bci()); 515 if (data != nullptr && data->is_MultiBranchData()) { 516 profile = (ciMultiBranchData*)data; 517 } 518 } 519 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 520 521 // generate decision tree, using trichotomy when possible 522 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 523 { 524 for (int j = 0; j < len; j++) { 525 table[3*j+0] = iter().get_int_table(2+2*j); 526 table[3*j+1] = iter().get_dest_table(2+2*j+1); 527 // Handle overflow when converting from uint to jint 528 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j)); 529 } 530 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 531 } 532 533 float default_cnt = 1.0F; 534 if (profile != nullptr) { 535 juint defaults = max_juint - len; 536 default_cnt = (float)profile->default_count()/(float)defaults; 537 } 538 539 int rnum = len*2+1; 540 bool makes_backward_branch = (default_dest <= bci()); 541 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 542 int rp = -1; 543 for (int j = 0; j < len; j++) { 544 jint match_int = table[3*j+0]; 545 jint dest = table[3*j+1]; 546 jint cnt = table[3*j+2]; 547 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 548 makes_backward_branch |= (dest <= bci()); 549 float c = default_cnt * ((float)match_int - (float)next_lo); 550 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) { 551 assert(default_dest != never_reached, "sentinel value for dead destinations"); 552 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c); 553 } 554 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) { 555 assert(dest != never_reached, "sentinel value for dead destinations"); 556 ranges[++rp].set(match_int, dest, (float)cnt); 557 } 558 } 559 jint highest = table[3*(len-1)]; 560 assert(ranges[rp].hi() == highest, ""); 561 if (highest != max_jint && 562 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) { 563 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest)); 564 } 565 assert(rp < rnum, "not too many ranges"); 566 567 if (trim_ranges) { 568 merge_ranges(ranges, rp); 569 } 570 571 // Safepoint in case backward branch observed 572 if (makes_backward_branch) { 573 add_safepoint(); 574 } 575 576 Node *lookup = pop(); // lookup value 577 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 578 } 579 580 static float if_prob(float taken_cnt, float total_cnt) { 581 assert(taken_cnt <= total_cnt, ""); 582 if (total_cnt == 0) { 583 return PROB_FAIR; 584 } 585 float p = taken_cnt / total_cnt; 586 return clamp(p, PROB_MIN, PROB_MAX); 587 } 588 589 static float if_cnt(float cnt) { 590 if (cnt == 0) { 591 return COUNT_UNKNOWN; 592 } 593 return cnt; 594 } 595 596 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 597 float total_cnt = 0; 598 for (SwitchRange* sr = lo; sr <= hi; sr++) { 599 total_cnt += sr->cnt(); 600 } 601 return total_cnt; 602 } 603 604 class SwitchRanges : public ResourceObj { 605 public: 606 SwitchRange* _lo; 607 SwitchRange* _hi; 608 SwitchRange* _mid; 609 float _cost; 610 611 enum { 612 Start, 613 LeftDone, 614 RightDone, 615 Done 616 } _state; 617 618 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 619 : _lo(lo), _hi(hi), _mid(nullptr), 620 _cost(0), _state(Start) { 621 } 622 623 SwitchRanges() 624 : _lo(nullptr), _hi(nullptr), _mid(nullptr), 625 _cost(0), _state(Start) {} 626 }; 627 628 // Estimate cost of performing a binary search on lo..hi 629 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 630 GrowableArray<SwitchRanges> tree; 631 SwitchRanges root(lo, hi); 632 tree.push(root); 633 634 float cost = 0; 635 do { 636 SwitchRanges& r = *tree.adr_at(tree.length()-1); 637 if (r._hi != r._lo) { 638 if (r._mid == nullptr) { 639 float r_cnt = sum_of_cnts(r._lo, r._hi); 640 641 if (r_cnt == 0) { 642 tree.pop(); 643 cost = 0; 644 continue; 645 } 646 647 SwitchRange* mid = nullptr; 648 mid = r._lo; 649 for (float cnt = 0; ; ) { 650 assert(mid <= r._hi, "out of bounds"); 651 cnt += mid->cnt(); 652 if (cnt > r_cnt / 2) { 653 break; 654 } 655 mid++; 656 } 657 assert(mid <= r._hi, "out of bounds"); 658 r._mid = mid; 659 r._cost = r_cnt / total_cnt; 660 } 661 r._cost += cost; 662 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 663 cost = 0; 664 r._state = SwitchRanges::LeftDone; 665 tree.push(SwitchRanges(r._lo, r._mid-1)); 666 } else if (r._state < SwitchRanges::RightDone) { 667 cost = 0; 668 r._state = SwitchRanges::RightDone; 669 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 670 } else { 671 tree.pop(); 672 cost = r._cost; 673 } 674 } else { 675 tree.pop(); 676 cost = r._cost; 677 } 678 } while (tree.length() > 0); 679 680 681 return cost; 682 } 683 684 // It sometimes pays off to test most common ranges before the binary search 685 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 686 uint nr = hi - lo + 1; 687 float total_cnt = sum_of_cnts(lo, hi); 688 689 float min = compute_tree_cost(lo, hi, total_cnt); 690 float extra = 1; 691 float sub = 0; 692 693 SwitchRange* array1 = lo; 694 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 695 696 SwitchRange* ranges = nullptr; 697 698 while (nr >= 2) { 699 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 700 ranges = (lo == array1) ? array2 : array1; 701 702 // Find highest frequency range 703 SwitchRange* candidate = lo; 704 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 705 if (sr->cnt() > candidate->cnt()) { 706 candidate = sr; 707 } 708 } 709 SwitchRange most_freq = *candidate; 710 if (most_freq.cnt() == 0) { 711 break; 712 } 713 714 // Copy remaining ranges into another array 715 int shift = 0; 716 for (uint i = 0; i < nr; i++) { 717 SwitchRange* sr = &lo[i]; 718 if (sr != candidate) { 719 ranges[i-shift] = *sr; 720 } else { 721 shift++; 722 if (i > 0 && i < nr-1) { 723 SwitchRange prev = lo[i-1]; 724 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt()); 725 if (prev.adjoin(lo[i+1])) { 726 shift++; 727 i++; 728 } 729 ranges[i-shift] = prev; 730 } 731 } 732 } 733 nr -= shift; 734 735 // Evaluate cost of testing the most common range and performing a 736 // binary search on the other ranges 737 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 738 if (cost >= min) { 739 break; 740 } 741 // swap arrays 742 lo = &ranges[0]; 743 hi = &ranges[nr-1]; 744 745 // It pays off: emit the test for the most common range 746 assert(most_freq.cnt() > 0, "must be taken"); 747 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 748 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo())))); 749 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 750 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 751 jump_if_true_fork(iff, most_freq.dest(), false); 752 753 sub += most_freq.cnt() / total_cnt; 754 extra += 1 - sub; 755 min = cost; 756 } 757 } 758 759 //----------------------------create_jump_tables------------------------------- 760 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 761 // Are jumptables enabled 762 if (!UseJumpTables) return false; 763 764 // Are jumptables supported 765 if (!Matcher::has_match_rule(Op_Jump)) return false; 766 767 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 768 769 // Decide if a guard is needed to lop off big ranges at either (or 770 // both) end(s) of the input set. We'll call this the default target 771 // even though we can't be sure that it is the true "default". 772 773 bool needs_guard = false; 774 int default_dest; 775 int64_t total_outlier_size = 0; 776 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 777 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 778 779 if (lo->dest() == hi->dest()) { 780 total_outlier_size = hi_size + lo_size; 781 default_dest = lo->dest(); 782 } else if (lo_size > hi_size) { 783 total_outlier_size = lo_size; 784 default_dest = lo->dest(); 785 } else { 786 total_outlier_size = hi_size; 787 default_dest = hi->dest(); 788 } 789 790 float total = sum_of_cnts(lo, hi); 791 float cost = compute_tree_cost(lo, hi, total); 792 793 // If a guard test will eliminate very sparse end ranges, then 794 // it is worth the cost of an extra jump. 795 float trimmed_cnt = 0; 796 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 797 needs_guard = true; 798 if (default_dest == lo->dest()) { 799 trimmed_cnt += lo->cnt(); 800 lo++; 801 } 802 if (default_dest == hi->dest()) { 803 trimmed_cnt += hi->cnt(); 804 hi--; 805 } 806 } 807 808 // Find the total number of cases and ranges 809 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 810 int num_range = hi - lo + 1; 811 812 // Don't create table if: too large, too small, or too sparse. 813 if (num_cases > MaxJumpTableSize) 814 return false; 815 if (UseSwitchProfiling) { 816 // MinJumpTableSize is set so with a well balanced binary tree, 817 // when the number of ranges is MinJumpTableSize, it's cheaper to 818 // go through a JumpNode that a tree of IfNodes. Average cost of a 819 // tree of IfNodes with MinJumpTableSize is 820 // log2f(MinJumpTableSize) comparisons. So if the cost computed 821 // from profile data is less than log2f(MinJumpTableSize) then 822 // going with the binary search is cheaper. 823 if (cost < log2f(MinJumpTableSize)) { 824 return false; 825 } 826 } else { 827 if (num_cases < MinJumpTableSize) 828 return false; 829 } 830 if (num_cases > (MaxJumpTableSparseness * num_range)) 831 return false; 832 833 // Normalize table lookups to zero 834 int lowval = lo->lo(); 835 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 836 837 // Generate a guard to protect against input keyvals that aren't 838 // in the switch domain. 839 if (needs_guard) { 840 Node* size = _gvn.intcon(num_cases); 841 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 842 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 843 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 844 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0); 845 846 total -= trimmed_cnt; 847 } 848 849 // Create an ideal node JumpTable that has projections 850 // of all possible ranges for a switch statement 851 // The key_val input must be converted to a pointer offset and scaled. 852 // Compare Parse::array_addressing above. 853 854 // Clean the 32-bit int into a real 64-bit offset. 855 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 856 // Make I2L conversion control dependent to prevent it from 857 // floating above the range check during loop optimizations. 858 // Do not use a narrow int type here to prevent the data path from dying 859 // while the control path is not removed. This can happen if the type of key_val 860 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast 861 // would be replaced by TOP while C2 is not able to fold the corresponding range checks. 862 // Set _carry_dependency for the cast to avoid being removed by IGVN. 863 #ifdef _LP64 864 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */); 865 #endif 866 867 // Shift the value by wordsize so we have an index into the table, rather 868 // than a switch value 869 Node *shiftWord = _gvn.MakeConX(wordSize); 870 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 871 872 // Create the JumpNode 873 Arena* arena = C->comp_arena(); 874 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 875 int i = 0; 876 if (total == 0) { 877 for (SwitchRange* r = lo; r <= hi; r++) { 878 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 879 probs[i] = 1.0F / num_cases; 880 } 881 } 882 } else { 883 for (SwitchRange* r = lo; r <= hi; r++) { 884 float prob = r->cnt()/total; 885 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 886 probs[i] = prob / (r->hi() - r->lo() + 1); 887 } 888 } 889 } 890 891 ciMethodData* methodData = method()->method_data(); 892 ciMultiBranchData* profile = nullptr; 893 if (methodData->is_mature()) { 894 ciProfileData* data = methodData->bci_to_data(bci()); 895 if (data != nullptr && data->is_MultiBranchData()) { 896 profile = (ciMultiBranchData*)data; 897 } 898 } 899 900 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total)); 901 902 // These are the switch destinations hanging off the jumpnode 903 i = 0; 904 for (SwitchRange* r = lo; r <= hi; r++) { 905 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 906 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 907 { 908 PreserveJVMState pjvms(this); 909 set_control(input); 910 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0); 911 } 912 } 913 } 914 assert(i == num_cases, "miscount of cases"); 915 stop_and_kill_map(); // no more uses for this JVMS 916 return true; 917 } 918 919 //----------------------------jump_switch_ranges------------------------------- 920 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 921 Block* switch_block = block(); 922 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 923 924 if (switch_depth == 0) { 925 // Do special processing for the top-level call. 926 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 927 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 928 929 // Decrement pred-numbers for the unique set of nodes. 930 #ifdef ASSERT 931 if (!trim_ranges) { 932 // Ensure that the block's successors are a (duplicate-free) set. 933 int successors_counted = 0; // block occurrences in [hi..lo] 934 int unique_successors = switch_block->num_successors(); 935 for (int i = 0; i < unique_successors; i++) { 936 Block* target = switch_block->successor_at(i); 937 938 // Check that the set of successors is the same in both places. 939 int successors_found = 0; 940 for (SwitchRange* p = lo; p <= hi; p++) { 941 if (p->dest() == target->start()) successors_found++; 942 } 943 assert(successors_found > 0, "successor must be known"); 944 successors_counted += successors_found; 945 } 946 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 947 } 948 #endif 949 950 // Maybe prune the inputs, based on the type of key_val. 951 jint min_val = min_jint; 952 jint max_val = max_jint; 953 const TypeInt* ti = key_val->bottom_type()->isa_int(); 954 if (ti != nullptr) { 955 min_val = ti->_lo; 956 max_val = ti->_hi; 957 assert(min_val <= max_val, "invalid int type"); 958 } 959 while (lo->hi() < min_val) { 960 lo++; 961 } 962 if (lo->lo() < min_val) { 963 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt()); 964 } 965 while (hi->lo() > max_val) { 966 hi--; 967 } 968 if (hi->hi() > max_val) { 969 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt()); 970 } 971 972 linear_search_switch_ranges(key_val, lo, hi); 973 } 974 975 #ifndef PRODUCT 976 if (switch_depth == 0) { 977 _max_switch_depth = 0; 978 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1; 979 } 980 #endif 981 982 assert(lo <= hi, "must be a non-empty set of ranges"); 983 if (lo == hi) { 984 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0); 985 } else { 986 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 987 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 988 989 if (create_jump_tables(key_val, lo, hi)) return; 990 991 SwitchRange* mid = nullptr; 992 float total_cnt = sum_of_cnts(lo, hi); 993 994 int nr = hi - lo + 1; 995 if (UseSwitchProfiling) { 996 // Don't keep the binary search tree balanced: pick up mid point 997 // that split frequencies in half. 998 float cnt = 0; 999 for (SwitchRange* sr = lo; sr <= hi; sr++) { 1000 cnt += sr->cnt(); 1001 if (cnt >= total_cnt / 2) { 1002 mid = sr; 1003 break; 1004 } 1005 } 1006 } else { 1007 mid = lo + nr/2; 1008 1009 // if there is an easy choice, pivot at a singleton: 1010 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1011 1012 assert(lo < mid && mid <= hi, "good pivot choice"); 1013 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1014 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1015 } 1016 1017 1018 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1019 1020 if (mid->is_singleton()) { 1021 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1022 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0); 1023 1024 // Special Case: If there are exactly three ranges, and the high 1025 // and low range each go to the same place, omit the "gt" test, 1026 // since it will not discriminate anything. 1027 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1028 1029 // if there is a higher range, test for it and process it: 1030 if (mid < hi && !eq_test_only) { 1031 // two comparisons of same values--should enable 1 test for 2 branches 1032 // Use BoolTest::lt instead of BoolTest::gt 1033 float cnt = sum_of_cnts(lo, mid-1); 1034 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt)); 1035 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) ); 1036 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) ); 1037 { PreserveJVMState pjvms(this); 1038 set_control(iffalse); 1039 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1040 } 1041 set_control(iftrue); 1042 } 1043 1044 } else { 1045 // mid is a range, not a singleton, so treat mid..hi as a unit 1046 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1047 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)); 1048 1049 // if there is a higher range, test for it and process it: 1050 if (mid == hi) { 1051 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0); 1052 } else { 1053 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1054 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1055 { PreserveJVMState pjvms(this); 1056 set_control(iftrue); 1057 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1058 } 1059 set_control(iffalse); 1060 } 1061 } 1062 1063 // in any case, process the lower range 1064 if (mid == lo) { 1065 if (mid->is_singleton()) { 1066 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1067 } else { 1068 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0); 1069 } 1070 } else { 1071 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1072 } 1073 } 1074 1075 // Decrease pred_count for each successor after all is done. 1076 if (switch_depth == 0) { 1077 int unique_successors = switch_block->num_successors(); 1078 for (int i = 0; i < unique_successors; i++) { 1079 Block* target = switch_block->successor_at(i); 1080 // Throw away the pre-allocated path for each unique successor. 1081 target->next_path_num(); 1082 } 1083 } 1084 1085 #ifndef PRODUCT 1086 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1087 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1088 SwitchRange* r; 1089 int nsing = 0; 1090 for( r = lo; r <= hi; r++ ) { 1091 if( r->is_singleton() ) nsing++; 1092 } 1093 tty->print(">>> "); 1094 _method->print_short_name(); 1095 tty->print_cr(" switch decision tree"); 1096 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1097 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1098 if (_max_switch_depth > _est_switch_depth) { 1099 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1100 } 1101 tty->print(" "); 1102 for( r = lo; r <= hi; r++ ) { 1103 r->print(); 1104 } 1105 tty->cr(); 1106 } 1107 #endif 1108 } 1109 1110 void Parse::modf() { 1111 Node *f2 = pop(); 1112 Node *f1 = pop(); 1113 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1114 CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1115 "frem", nullptr, //no memory effects 1116 f1, f2); 1117 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1118 1119 push(res); 1120 } 1121 1122 void Parse::modd() { 1123 Node *d2 = pop_pair(); 1124 Node *d1 = pop_pair(); 1125 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1126 CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1127 "drem", nullptr, //no memory effects 1128 d1, top(), d2, top()); 1129 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1130 1131 #ifdef ASSERT 1132 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1)); 1133 assert(res_top == top(), "second value must be top"); 1134 #endif 1135 1136 push_pair(res_d); 1137 } 1138 1139 void Parse::l2f() { 1140 Node* f2 = pop(); 1141 Node* f1 = pop(); 1142 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1143 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1144 "l2f", nullptr, //no memory effects 1145 f1, f2); 1146 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1147 1148 push(res); 1149 } 1150 1151 // Handle jsr and jsr_w bytecode 1152 void Parse::do_jsr() { 1153 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1154 1155 // Store information about current state, tagged with new _jsr_bci 1156 int return_bci = iter().next_bci(); 1157 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1158 1159 // The way we do things now, there is only one successor block 1160 // for the jsr, because the target code is cloned by ciTypeFlow. 1161 Block* target = successor_for_bci(jsr_bci); 1162 1163 // What got pushed? 1164 const Type* ret_addr = target->peek(); 1165 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1166 1167 // Effect on jsr on stack 1168 push(_gvn.makecon(ret_addr)); 1169 1170 // Flow to the jsr. 1171 merge(jsr_bci); 1172 } 1173 1174 // Handle ret bytecode 1175 void Parse::do_ret() { 1176 // Find to whom we return. 1177 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1178 Block* target = block()->successor_at(0); 1179 assert(!target->is_ready(), "our arrival must be expected"); 1180 int pnum = target->next_path_num(); 1181 merge_common(target, pnum); 1182 } 1183 1184 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1185 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1186 // Only ::eq and ::ne are supported for profile injection. 1187 return false; 1188 } 1189 if (test->is_Cmp() && 1190 test->in(1)->Opcode() == Op_ProfileBoolean) { 1191 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1192 int false_cnt = profile->false_count(); 1193 int true_cnt = profile->true_count(); 1194 1195 // Counts matching depends on the actual test operation (::eq or ::ne). 1196 // No need to scale the counts because profile injection was designed 1197 // to feed exact counts into VM. 1198 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1199 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1200 1201 profile->consume(); 1202 return true; 1203 } 1204 return false; 1205 } 1206 1207 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1208 // We also check that individual counters are positive first, otherwise the sum can become positive. 1209 // (check for saturation, integer overflow, and immature counts) 1210 static bool counters_are_meaningful(int counter1, int counter2, int min) { 1211 // check for saturation, including "uint" values too big to fit in "int" 1212 if (counter1 < 0 || counter2 < 0) { 1213 return false; 1214 } 1215 // check for integer overflow of the sum 1216 int64_t sum = (int64_t)counter1 + (int64_t)counter2; 1217 STATIC_ASSERT(sizeof(counter1) < sizeof(sum)); 1218 if (sum > INT_MAX) { 1219 return false; 1220 } 1221 // check if mature 1222 return (counter1 + counter2) >= min; 1223 } 1224 1225 //--------------------------dynamic_branch_prediction-------------------------- 1226 // Try to gather dynamic branch prediction behavior. Return a probability 1227 // of the branch being taken and set the "cnt" field. Returns a -1.0 1228 // if we need to use static prediction for some reason. 1229 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1230 ResourceMark rm; 1231 1232 cnt = COUNT_UNKNOWN; 1233 1234 int taken = 0; 1235 int not_taken = 0; 1236 1237 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1238 1239 if (use_mdo) { 1240 // Use MethodData information if it is available 1241 // FIXME: free the ProfileData structure 1242 ciMethodData* methodData = method()->method_data(); 1243 if (!methodData->is_mature()) return PROB_UNKNOWN; 1244 ciProfileData* data = methodData->bci_to_data(bci()); 1245 if (data == nullptr) { 1246 return PROB_UNKNOWN; 1247 } 1248 if (!data->is_JumpData()) return PROB_UNKNOWN; 1249 1250 // get taken and not taken values 1251 // NOTE: saturated UINT_MAX values become negative, 1252 // as do counts above INT_MAX. 1253 taken = data->as_JumpData()->taken(); 1254 not_taken = 0; 1255 if (data->is_BranchData()) { 1256 not_taken = data->as_BranchData()->not_taken(); 1257 } 1258 1259 // scale the counts to be commensurate with invocation counts: 1260 // NOTE: overflow for positive values is clamped at INT_MAX 1261 taken = method()->scale_count(taken); 1262 not_taken = method()->scale_count(not_taken); 1263 } 1264 // At this point, saturation or overflow is indicated by INT_MAX 1265 // or a negative value. 1266 1267 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1268 // We also check that individual counters are positive first, otherwise the sum can become positive. 1269 if (!counters_are_meaningful(taken, not_taken, 40)) { 1270 if (C->log() != nullptr) { 1271 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1272 } 1273 return PROB_UNKNOWN; 1274 } 1275 1276 // Compute frequency that we arrive here 1277 float sum = taken + not_taken; 1278 // Adjust, if this block is a cloned private block but the 1279 // Jump counts are shared. Taken the private counts for 1280 // just this path instead of the shared counts. 1281 if( block()->count() > 0 ) 1282 sum = block()->count(); 1283 cnt = sum / FreqCountInvocations; 1284 1285 // Pin probability to sane limits 1286 float prob; 1287 if( !taken ) 1288 prob = (0+PROB_MIN) / 2; 1289 else if( !not_taken ) 1290 prob = (1+PROB_MAX) / 2; 1291 else { // Compute probability of true path 1292 prob = (float)taken / (float)(taken + not_taken); 1293 if (prob > PROB_MAX) prob = PROB_MAX; 1294 if (prob < PROB_MIN) prob = PROB_MIN; 1295 } 1296 1297 assert((cnt > 0.0f) && (prob > 0.0f), 1298 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken); 1299 1300 if (C->log() != nullptr) { 1301 const char* prob_str = nullptr; 1302 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1303 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1304 char prob_str_buf[30]; 1305 if (prob_str == nullptr) { 1306 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob); 1307 prob_str = prob_str_buf; 1308 } 1309 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1310 iter().get_dest(), taken, not_taken, cnt, prob_str); 1311 } 1312 return prob; 1313 } 1314 1315 //-----------------------------branch_prediction------------------------------- 1316 float Parse::branch_prediction(float& cnt, 1317 BoolTest::mask btest, 1318 int target_bci, 1319 Node* test) { 1320 float prob = dynamic_branch_prediction(cnt, btest, test); 1321 // If prob is unknown, switch to static prediction 1322 if (prob != PROB_UNKNOWN) return prob; 1323 1324 prob = PROB_FAIR; // Set default value 1325 if (btest == BoolTest::eq) // Exactly equal test? 1326 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1327 else if (btest == BoolTest::ne) 1328 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1329 1330 // If this is a conditional test guarding a backwards branch, 1331 // assume its a loop-back edge. Make it a likely taken branch. 1332 if (target_bci < bci()) { 1333 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1334 // Since it's an OSR, we probably have profile data, but since 1335 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1336 // Let's make a special check here for completely zero counts. 1337 ciMethodData* methodData = method()->method_data(); 1338 if (!methodData->is_empty()) { 1339 ciProfileData* data = methodData->bci_to_data(bci()); 1340 // Only stop for truly zero counts, which mean an unknown part 1341 // of the OSR-ed method, and we want to deopt to gather more stats. 1342 // If you have ANY counts, then this loop is simply 'cold' relative 1343 // to the OSR loop. 1344 if (data == nullptr || 1345 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1346 // This is the only way to return PROB_UNKNOWN: 1347 return PROB_UNKNOWN; 1348 } 1349 } 1350 } 1351 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1352 } 1353 1354 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1355 return prob; 1356 } 1357 1358 // The magic constants are chosen so as to match the output of 1359 // branch_prediction() when the profile reports a zero taken count. 1360 // It is important to distinguish zero counts unambiguously, because 1361 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1362 // very small but nonzero probabilities, which if confused with zero 1363 // counts would keep the program recompiling indefinitely. 1364 bool Parse::seems_never_taken(float prob) const { 1365 return prob < PROB_MIN; 1366 } 1367 1368 // True if the comparison seems to be the kind that will not change its 1369 // statistics from true to false. See comments in adjust_map_after_if. 1370 // This question is only asked along paths which are already 1371 // classified as untaken (by seems_never_taken), so really, 1372 // if a path is never taken, its controlling comparison is 1373 // already acting in a stable fashion. If the comparison 1374 // seems stable, we will put an expensive uncommon trap 1375 // on the untaken path. 1376 bool Parse::seems_stable_comparison() const { 1377 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) { 1378 return false; 1379 } 1380 return true; 1381 } 1382 1383 //-------------------------------repush_if_args-------------------------------- 1384 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1385 inline int Parse::repush_if_args() { 1386 if (PrintOpto && WizardMode) { 1387 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1388 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1389 method()->print_name(); tty->cr(); 1390 } 1391 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1392 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1393 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1394 assert(argument(0) != nullptr, "must exist"); 1395 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist"); 1396 inc_sp(bc_depth); 1397 return bc_depth; 1398 } 1399 1400 //----------------------------------do_ifnull---------------------------------- 1401 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1402 int target_bci = iter().get_dest(); 1403 1404 Block* branch_block = successor_for_bci(target_bci); 1405 Block* next_block = successor_for_bci(iter().next_bci()); 1406 1407 float cnt; 1408 float prob = branch_prediction(cnt, btest, target_bci, c); 1409 if (prob == PROB_UNKNOWN) { 1410 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1411 if (PrintOpto && Verbose) { 1412 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1413 } 1414 repush_if_args(); // to gather stats on loop 1415 uncommon_trap(Deoptimization::Reason_unreached, 1416 Deoptimization::Action_reinterpret, 1417 nullptr, "cold"); 1418 if (C->eliminate_boxing()) { 1419 // Mark the successor blocks as parsed 1420 branch_block->next_path_num(); 1421 next_block->next_path_num(); 1422 } 1423 return; 1424 } 1425 1426 NOT_PRODUCT(explicit_null_checks_inserted++); 1427 1428 // Generate real control flow 1429 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1430 1431 // Sanity check the probability value 1432 assert(prob > 0.0f,"Bad probability in Parser"); 1433 // Need xform to put node in hash table 1434 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1435 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1436 // True branch 1437 { PreserveJVMState pjvms(this); 1438 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1439 set_control(iftrue); 1440 1441 if (stopped()) { // Path is dead? 1442 NOT_PRODUCT(explicit_null_checks_elided++); 1443 if (C->eliminate_boxing()) { 1444 // Mark the successor block as parsed 1445 branch_block->next_path_num(); 1446 } 1447 } else { // Path is live. 1448 adjust_map_after_if(btest, c, prob, branch_block); 1449 if (!stopped()) { 1450 merge(target_bci); 1451 } 1452 } 1453 } 1454 1455 // False branch 1456 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1457 set_control(iffalse); 1458 1459 if (stopped()) { // Path is dead? 1460 NOT_PRODUCT(explicit_null_checks_elided++); 1461 if (C->eliminate_boxing()) { 1462 // Mark the successor block as parsed 1463 next_block->next_path_num(); 1464 } 1465 } else { // Path is live. 1466 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block); 1467 } 1468 } 1469 1470 //------------------------------------do_if------------------------------------ 1471 void Parse::do_if(BoolTest::mask btest, Node* c) { 1472 int target_bci = iter().get_dest(); 1473 1474 Block* branch_block = successor_for_bci(target_bci); 1475 Block* next_block = successor_for_bci(iter().next_bci()); 1476 1477 float cnt; 1478 float prob = branch_prediction(cnt, btest, target_bci, c); 1479 float untaken_prob = 1.0 - prob; 1480 1481 if (prob == PROB_UNKNOWN) { 1482 if (PrintOpto && Verbose) { 1483 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1484 } 1485 repush_if_args(); // to gather stats on loop 1486 uncommon_trap(Deoptimization::Reason_unreached, 1487 Deoptimization::Action_reinterpret, 1488 nullptr, "cold"); 1489 if (C->eliminate_boxing()) { 1490 // Mark the successor blocks as parsed 1491 branch_block->next_path_num(); 1492 next_block->next_path_num(); 1493 } 1494 return; 1495 } 1496 1497 // Sanity check the probability value 1498 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1499 1500 bool taken_if_true = true; 1501 // Convert BoolTest to canonical form: 1502 if (!BoolTest(btest).is_canonical()) { 1503 btest = BoolTest(btest).negate(); 1504 taken_if_true = false; 1505 // prob is NOT updated here; it remains the probability of the taken 1506 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1507 } 1508 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1509 1510 Node* tst0 = new BoolNode(c, btest); 1511 Node* tst = _gvn.transform(tst0); 1512 BoolTest::mask taken_btest = BoolTest::illegal; 1513 BoolTest::mask untaken_btest = BoolTest::illegal; 1514 1515 if (tst->is_Bool()) { 1516 // Refresh c from the transformed bool node, since it may be 1517 // simpler than the original c. Also re-canonicalize btest. 1518 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)). 1519 // That can arise from statements like: if (x instanceof C) ... 1520 if (tst != tst0) { 1521 // Canonicalize one more time since transform can change it. 1522 btest = tst->as_Bool()->_test._test; 1523 if (!BoolTest(btest).is_canonical()) { 1524 // Reverse edges one more time... 1525 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1526 btest = tst->as_Bool()->_test._test; 1527 assert(BoolTest(btest).is_canonical(), "sanity"); 1528 taken_if_true = !taken_if_true; 1529 } 1530 c = tst->in(1); 1531 } 1532 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1533 taken_btest = taken_if_true ? btest : neg_btest; 1534 untaken_btest = taken_if_true ? neg_btest : btest; 1535 } 1536 1537 // Generate real control flow 1538 float true_prob = (taken_if_true ? prob : untaken_prob); 1539 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1540 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1541 Node* taken_branch = new IfTrueNode(iff); 1542 Node* untaken_branch = new IfFalseNode(iff); 1543 if (!taken_if_true) { // Finish conversion to canonical form 1544 Node* tmp = taken_branch; 1545 taken_branch = untaken_branch; 1546 untaken_branch = tmp; 1547 } 1548 1549 // Branch is taken: 1550 { PreserveJVMState pjvms(this); 1551 taken_branch = _gvn.transform(taken_branch); 1552 set_control(taken_branch); 1553 1554 if (stopped()) { 1555 if (C->eliminate_boxing()) { 1556 // Mark the successor block as parsed 1557 branch_block->next_path_num(); 1558 } 1559 } else { 1560 adjust_map_after_if(taken_btest, c, prob, branch_block); 1561 if (!stopped()) { 1562 merge(target_bci); 1563 } 1564 } 1565 } 1566 1567 untaken_branch = _gvn.transform(untaken_branch); 1568 set_control(untaken_branch); 1569 1570 // Branch not taken. 1571 if (stopped()) { 1572 if (C->eliminate_boxing()) { 1573 // Mark the successor block as parsed 1574 next_block->next_path_num(); 1575 } 1576 } else { 1577 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block); 1578 } 1579 } 1580 1581 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 1582 // Don't want to speculate on uncommon traps when running with -Xcomp 1583 if (!UseInterpreter) { 1584 return false; 1585 } 1586 return (seems_never_taken(prob) && seems_stable_comparison()); 1587 } 1588 1589 void Parse::maybe_add_predicate_after_if(Block* path) { 1590 if (path->is_SEL_head() && path->preds_parsed() == 0) { 1591 // Add predicates at bci of if dominating the loop so traps can be 1592 // recorded on the if's profile data 1593 int bc_depth = repush_if_args(); 1594 add_parse_predicates(); 1595 dec_sp(bc_depth); 1596 path->set_has_predicates(); 1597 } 1598 } 1599 1600 1601 //----------------------------adjust_map_after_if------------------------------ 1602 // Adjust the JVM state to reflect the result of taking this path. 1603 // Basically, it means inspecting the CmpNode controlling this 1604 // branch, seeing how it constrains a tested value, and then 1605 // deciding if it's worth our while to encode this constraint 1606 // as graph nodes in the current abstract interpretation map. 1607 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) { 1608 if (!c->is_Cmp()) { 1609 maybe_add_predicate_after_if(path); 1610 return; 1611 } 1612 1613 if (stopped() || btest == BoolTest::illegal) { 1614 return; // nothing to do 1615 } 1616 1617 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1618 1619 if (path_is_suitable_for_uncommon_trap(prob)) { 1620 repush_if_args(); 1621 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if, 1622 Deoptimization::Action_reinterpret, 1623 nullptr, 1624 (is_fallthrough ? "taken always" : "taken never")); 1625 1626 if (call != nullptr) { 1627 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path)); 1628 } 1629 return; 1630 } 1631 1632 Node* val = c->in(1); 1633 Node* con = c->in(2); 1634 const Type* tcon = _gvn.type(con); 1635 const Type* tval = _gvn.type(val); 1636 bool have_con = tcon->singleton(); 1637 if (tval->singleton()) { 1638 if (!have_con) { 1639 // Swap, so constant is in con. 1640 con = val; 1641 tcon = tval; 1642 val = c->in(2); 1643 tval = _gvn.type(val); 1644 btest = BoolTest(btest).commute(); 1645 have_con = true; 1646 } else { 1647 // Do we have two constants? Then leave well enough alone. 1648 have_con = false; 1649 } 1650 } 1651 if (!have_con) { // remaining adjustments need a con 1652 maybe_add_predicate_after_if(path); 1653 return; 1654 } 1655 1656 sharpen_type_after_if(btest, con, tcon, val, tval); 1657 maybe_add_predicate_after_if(path); 1658 } 1659 1660 1661 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1662 Node* ldk; 1663 if (n->is_DecodeNKlass()) { 1664 if (n->in(1)->Opcode() != Op_LoadNKlass) { 1665 return nullptr; 1666 } else { 1667 ldk = n->in(1); 1668 } 1669 } else if (n->Opcode() != Op_LoadKlass) { 1670 return nullptr; 1671 } else { 1672 ldk = n; 1673 } 1674 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1675 1676 Node* adr = ldk->in(MemNode::Address); 1677 intptr_t off = 0; 1678 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1679 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1680 return nullptr; 1681 const TypePtr* tp = gvn->type(obj)->is_ptr(); 1682 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1683 return nullptr; 1684 1685 return obj; 1686 } 1687 1688 void Parse::sharpen_type_after_if(BoolTest::mask btest, 1689 Node* con, const Type* tcon, 1690 Node* val, const Type* tval) { 1691 // Look for opportunities to sharpen the type of a node 1692 // whose klass is compared with a constant klass. 1693 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1694 Node* obj = extract_obj_from_klass_load(&_gvn, val); 1695 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1696 if (obj != nullptr && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1697 // Found: 1698 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1699 // or the narrowOop equivalent. 1700 const Type* obj_type = _gvn.type(obj); 1701 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1702 if (tboth != nullptr && tboth->klass_is_exact() && tboth != obj_type && 1703 tboth->higher_equal(obj_type)) { 1704 // obj has to be of the exact type Foo if the CmpP succeeds. 1705 int obj_in_map = map()->find_edge(obj); 1706 JVMState* jvms = this->jvms(); 1707 if (obj_in_map >= 0 && 1708 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1709 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 1710 const Type* tcc = ccast->as_Type()->type(); 1711 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 1712 // Delay transform() call to allow recovery of pre-cast value 1713 // at the control merge. 1714 _gvn.set_type_bottom(ccast); 1715 record_for_igvn(ccast); 1716 // Here's the payoff. 1717 replace_in_map(obj, ccast); 1718 } 1719 } 1720 } 1721 } 1722 1723 int val_in_map = map()->find_edge(val); 1724 if (val_in_map < 0) return; // replace_in_map would be useless 1725 { 1726 JVMState* jvms = this->jvms(); 1727 if (!(jvms->is_loc(val_in_map) || 1728 jvms->is_stk(val_in_map))) 1729 return; // again, it would be useless 1730 } 1731 1732 // Check for a comparison to a constant, and "know" that the compared 1733 // value is constrained on this path. 1734 assert(tcon->singleton(), ""); 1735 ConstraintCastNode* ccast = nullptr; 1736 Node* cast = nullptr; 1737 1738 switch (btest) { 1739 case BoolTest::eq: // Constant test? 1740 { 1741 const Type* tboth = tcon->join_speculative(tval); 1742 if (tboth == tval) break; // Nothing to gain. 1743 if (tcon->isa_int()) { 1744 ccast = new CastIINode(control(), val, tboth); 1745 } else if (tcon == TypePtr::NULL_PTR) { 1746 // Cast to null, but keep the pointer identity temporarily live. 1747 ccast = new CastPPNode(control(), val, tboth); 1748 } else { 1749 const TypeF* tf = tcon->isa_float_constant(); 1750 const TypeD* td = tcon->isa_double_constant(); 1751 // Exclude tests vs float/double 0 as these could be 1752 // either +0 or -0. Just because you are equal to +0 1753 // doesn't mean you ARE +0! 1754 // Note, following code also replaces Long and Oop values. 1755 if ((!tf || tf->_f != 0.0) && 1756 (!td || td->_d != 0.0)) 1757 cast = con; // Replace non-constant val by con. 1758 } 1759 } 1760 break; 1761 1762 case BoolTest::ne: 1763 if (tcon == TypePtr::NULL_PTR) { 1764 cast = cast_not_null(val, false); 1765 } 1766 break; 1767 1768 default: 1769 // (At this point we could record int range types with CastII.) 1770 break; 1771 } 1772 1773 if (ccast != nullptr) { 1774 const Type* tcc = ccast->as_Type()->type(); 1775 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 1776 // Delay transform() call to allow recovery of pre-cast value 1777 // at the control merge. 1778 _gvn.set_type_bottom(ccast); 1779 record_for_igvn(ccast); 1780 cast = ccast; 1781 } 1782 1783 if (cast != nullptr) { // Here's the payoff. 1784 replace_in_map(val, cast); 1785 } 1786 } 1787 1788 /** 1789 * Use speculative type to optimize CmpP node: if comparison is 1790 * against the low level class, cast the object to the speculative 1791 * type if any. CmpP should then go away. 1792 * 1793 * @param c expected CmpP node 1794 * @return result of CmpP on object casted to speculative type 1795 * 1796 */ 1797 Node* Parse::optimize_cmp_with_klass(Node* c) { 1798 // If this is transformed by the _gvn to a comparison with the low 1799 // level klass then we may be able to use speculation 1800 if (c->Opcode() == Op_CmpP && 1801 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1802 c->in(2)->is_Con()) { 1803 Node* load_klass = nullptr; 1804 Node* decode = nullptr; 1805 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1806 decode = c->in(1); 1807 load_klass = c->in(1)->in(1); 1808 } else { 1809 load_klass = c->in(1); 1810 } 1811 if (load_klass->in(2)->is_AddP()) { 1812 Node* addp = load_klass->in(2); 1813 Node* obj = addp->in(AddPNode::Address); 1814 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1815 if (obj_type->speculative_type_not_null() != nullptr) { 1816 ciKlass* k = obj_type->speculative_type(); 1817 inc_sp(2); 1818 obj = maybe_cast_profiled_obj(obj, k); 1819 dec_sp(2); 1820 // Make the CmpP use the casted obj 1821 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1822 load_klass = load_klass->clone(); 1823 load_klass->set_req(2, addp); 1824 load_klass = _gvn.transform(load_klass); 1825 if (decode != nullptr) { 1826 decode = decode->clone(); 1827 decode->set_req(1, load_klass); 1828 load_klass = _gvn.transform(decode); 1829 } 1830 c = c->clone(); 1831 c->set_req(1, load_klass); 1832 c = _gvn.transform(c); 1833 } 1834 } 1835 } 1836 return c; 1837 } 1838 1839 //------------------------------do_one_bytecode-------------------------------- 1840 // Parse this bytecode, and alter the Parsers JVM->Node mapping 1841 void Parse::do_one_bytecode() { 1842 Node *a, *b, *c, *d; // Handy temps 1843 BoolTest::mask btest; 1844 int i; 1845 1846 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1847 1848 if (C->check_node_count(NodeLimitFudgeFactor * 5, 1849 "out of nodes parsing method")) { 1850 return; 1851 } 1852 1853 #ifdef ASSERT 1854 // for setting breakpoints 1855 if (TraceOptoParse) { 1856 tty->print(" @"); 1857 dump_bci(bci()); 1858 tty->print(" %s", Bytecodes::name(bc())); 1859 tty->cr(); 1860 } 1861 #endif 1862 1863 switch (bc()) { 1864 case Bytecodes::_nop: 1865 // do nothing 1866 break; 1867 case Bytecodes::_lconst_0: 1868 push_pair(longcon(0)); 1869 break; 1870 1871 case Bytecodes::_lconst_1: 1872 push_pair(longcon(1)); 1873 break; 1874 1875 case Bytecodes::_fconst_0: 1876 push(zerocon(T_FLOAT)); 1877 break; 1878 1879 case Bytecodes::_fconst_1: 1880 push(makecon(TypeF::ONE)); 1881 break; 1882 1883 case Bytecodes::_fconst_2: 1884 push(makecon(TypeF::make(2.0f))); 1885 break; 1886 1887 case Bytecodes::_dconst_0: 1888 push_pair(zerocon(T_DOUBLE)); 1889 break; 1890 1891 case Bytecodes::_dconst_1: 1892 push_pair(makecon(TypeD::ONE)); 1893 break; 1894 1895 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1896 case Bytecodes::_iconst_0: push(intcon( 0)); break; 1897 case Bytecodes::_iconst_1: push(intcon( 1)); break; 1898 case Bytecodes::_iconst_2: push(intcon( 2)); break; 1899 case Bytecodes::_iconst_3: push(intcon( 3)); break; 1900 case Bytecodes::_iconst_4: push(intcon( 4)); break; 1901 case Bytecodes::_iconst_5: push(intcon( 5)); break; 1902 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1903 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1904 case Bytecodes::_aconst_null: push(null()); break; 1905 1906 case Bytecodes::_ldc: 1907 case Bytecodes::_ldc_w: 1908 case Bytecodes::_ldc2_w: { 1909 ciConstant constant = iter().get_constant(); 1910 if (constant.is_loaded()) { 1911 const Type* con_type = Type::make_from_constant(constant); 1912 if (con_type != nullptr) { 1913 push_node(con_type->basic_type(), makecon(con_type)); 1914 } 1915 } else { 1916 // If the constant is unresolved or in error state, run this BC in the interpreter. 1917 if (iter().is_in_error()) { 1918 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unhandled, 1919 Deoptimization::Action_none), 1920 nullptr, "constant in error state", true /* must_throw */); 1921 1922 } else { 1923 int index = iter().get_constant_pool_index(); 1924 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unloaded, 1925 Deoptimization::Action_reinterpret, 1926 index), 1927 nullptr, "unresolved constant", false /* must_throw */); 1928 } 1929 } 1930 break; 1931 } 1932 1933 case Bytecodes::_aload_0: 1934 push( local(0) ); 1935 break; 1936 case Bytecodes::_aload_1: 1937 push( local(1) ); 1938 break; 1939 case Bytecodes::_aload_2: 1940 push( local(2) ); 1941 break; 1942 case Bytecodes::_aload_3: 1943 push( local(3) ); 1944 break; 1945 case Bytecodes::_aload: 1946 push( local(iter().get_index()) ); 1947 break; 1948 1949 case Bytecodes::_fload_0: 1950 case Bytecodes::_iload_0: 1951 push( local(0) ); 1952 break; 1953 case Bytecodes::_fload_1: 1954 case Bytecodes::_iload_1: 1955 push( local(1) ); 1956 break; 1957 case Bytecodes::_fload_2: 1958 case Bytecodes::_iload_2: 1959 push( local(2) ); 1960 break; 1961 case Bytecodes::_fload_3: 1962 case Bytecodes::_iload_3: 1963 push( local(3) ); 1964 break; 1965 case Bytecodes::_fload: 1966 case Bytecodes::_iload: 1967 push( local(iter().get_index()) ); 1968 break; 1969 case Bytecodes::_lload_0: 1970 push_pair_local( 0 ); 1971 break; 1972 case Bytecodes::_lload_1: 1973 push_pair_local( 1 ); 1974 break; 1975 case Bytecodes::_lload_2: 1976 push_pair_local( 2 ); 1977 break; 1978 case Bytecodes::_lload_3: 1979 push_pair_local( 3 ); 1980 break; 1981 case Bytecodes::_lload: 1982 push_pair_local( iter().get_index() ); 1983 break; 1984 1985 case Bytecodes::_dload_0: 1986 push_pair_local(0); 1987 break; 1988 case Bytecodes::_dload_1: 1989 push_pair_local(1); 1990 break; 1991 case Bytecodes::_dload_2: 1992 push_pair_local(2); 1993 break; 1994 case Bytecodes::_dload_3: 1995 push_pair_local(3); 1996 break; 1997 case Bytecodes::_dload: 1998 push_pair_local(iter().get_index()); 1999 break; 2000 case Bytecodes::_fstore_0: 2001 case Bytecodes::_istore_0: 2002 case Bytecodes::_astore_0: 2003 set_local( 0, pop() ); 2004 break; 2005 case Bytecodes::_fstore_1: 2006 case Bytecodes::_istore_1: 2007 case Bytecodes::_astore_1: 2008 set_local( 1, pop() ); 2009 break; 2010 case Bytecodes::_fstore_2: 2011 case Bytecodes::_istore_2: 2012 case Bytecodes::_astore_2: 2013 set_local( 2, pop() ); 2014 break; 2015 case Bytecodes::_fstore_3: 2016 case Bytecodes::_istore_3: 2017 case Bytecodes::_astore_3: 2018 set_local( 3, pop() ); 2019 break; 2020 case Bytecodes::_fstore: 2021 case Bytecodes::_istore: 2022 case Bytecodes::_astore: 2023 set_local( iter().get_index(), pop() ); 2024 break; 2025 // long stores 2026 case Bytecodes::_lstore_0: 2027 set_pair_local( 0, pop_pair() ); 2028 break; 2029 case Bytecodes::_lstore_1: 2030 set_pair_local( 1, pop_pair() ); 2031 break; 2032 case Bytecodes::_lstore_2: 2033 set_pair_local( 2, pop_pair() ); 2034 break; 2035 case Bytecodes::_lstore_3: 2036 set_pair_local( 3, pop_pair() ); 2037 break; 2038 case Bytecodes::_lstore: 2039 set_pair_local( iter().get_index(), pop_pair() ); 2040 break; 2041 2042 // double stores 2043 case Bytecodes::_dstore_0: 2044 set_pair_local( 0, dprecision_rounding(pop_pair()) ); 2045 break; 2046 case Bytecodes::_dstore_1: 2047 set_pair_local( 1, dprecision_rounding(pop_pair()) ); 2048 break; 2049 case Bytecodes::_dstore_2: 2050 set_pair_local( 2, dprecision_rounding(pop_pair()) ); 2051 break; 2052 case Bytecodes::_dstore_3: 2053 set_pair_local( 3, dprecision_rounding(pop_pair()) ); 2054 break; 2055 case Bytecodes::_dstore: 2056 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) ); 2057 break; 2058 2059 case Bytecodes::_pop: dec_sp(1); break; 2060 case Bytecodes::_pop2: dec_sp(2); break; 2061 case Bytecodes::_swap: 2062 a = pop(); 2063 b = pop(); 2064 push(a); 2065 push(b); 2066 break; 2067 case Bytecodes::_dup: 2068 a = pop(); 2069 push(a); 2070 push(a); 2071 break; 2072 case Bytecodes::_dup_x1: 2073 a = pop(); 2074 b = pop(); 2075 push( a ); 2076 push( b ); 2077 push( a ); 2078 break; 2079 case Bytecodes::_dup_x2: 2080 a = pop(); 2081 b = pop(); 2082 c = pop(); 2083 push( a ); 2084 push( c ); 2085 push( b ); 2086 push( a ); 2087 break; 2088 case Bytecodes::_dup2: 2089 a = pop(); 2090 b = pop(); 2091 push( b ); 2092 push( a ); 2093 push( b ); 2094 push( a ); 2095 break; 2096 2097 case Bytecodes::_dup2_x1: 2098 // before: .. c, b, a 2099 // after: .. b, a, c, b, a 2100 // not tested 2101 a = pop(); 2102 b = pop(); 2103 c = pop(); 2104 push( b ); 2105 push( a ); 2106 push( c ); 2107 push( b ); 2108 push( a ); 2109 break; 2110 case Bytecodes::_dup2_x2: 2111 // before: .. d, c, b, a 2112 // after: .. b, a, d, c, b, a 2113 // not tested 2114 a = pop(); 2115 b = pop(); 2116 c = pop(); 2117 d = pop(); 2118 push( b ); 2119 push( a ); 2120 push( d ); 2121 push( c ); 2122 push( b ); 2123 push( a ); 2124 break; 2125 2126 case Bytecodes::_arraylength: { 2127 // Must do null-check with value on expression stack 2128 Node *ary = null_check(peek(), T_ARRAY); 2129 // Compile-time detect of null-exception? 2130 if (stopped()) return; 2131 a = pop(); 2132 push(load_array_length(a)); 2133 break; 2134 } 2135 2136 case Bytecodes::_baload: array_load(T_BYTE); break; 2137 case Bytecodes::_caload: array_load(T_CHAR); break; 2138 case Bytecodes::_iaload: array_load(T_INT); break; 2139 case Bytecodes::_saload: array_load(T_SHORT); break; 2140 case Bytecodes::_faload: array_load(T_FLOAT); break; 2141 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2142 case Bytecodes::_laload: array_load(T_LONG); break; 2143 case Bytecodes::_daload: array_load(T_DOUBLE); break; 2144 case Bytecodes::_bastore: array_store(T_BYTE); break; 2145 case Bytecodes::_castore: array_store(T_CHAR); break; 2146 case Bytecodes::_iastore: array_store(T_INT); break; 2147 case Bytecodes::_sastore: array_store(T_SHORT); break; 2148 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2149 case Bytecodes::_aastore: array_store(T_OBJECT); break; 2150 case Bytecodes::_lastore: array_store(T_LONG); break; 2151 case Bytecodes::_dastore: array_store(T_DOUBLE); break; 2152 2153 case Bytecodes::_getfield: 2154 do_getfield(); 2155 break; 2156 2157 case Bytecodes::_getstatic: 2158 do_getstatic(); 2159 break; 2160 2161 case Bytecodes::_putfield: 2162 do_putfield(); 2163 break; 2164 2165 case Bytecodes::_putstatic: 2166 do_putstatic(); 2167 break; 2168 2169 case Bytecodes::_irem: 2170 // Must keep both values on the expression-stack during null-check 2171 zero_check_int(peek()); 2172 // Compile-time detect of null-exception? 2173 if (stopped()) return; 2174 b = pop(); 2175 a = pop(); 2176 push(_gvn.transform(new ModINode(control(), a, b))); 2177 break; 2178 case Bytecodes::_idiv: 2179 // Must keep both values on the expression-stack during null-check 2180 zero_check_int(peek()); 2181 // Compile-time detect of null-exception? 2182 if (stopped()) return; 2183 b = pop(); 2184 a = pop(); 2185 push( _gvn.transform( new DivINode(control(),a,b) ) ); 2186 break; 2187 case Bytecodes::_imul: 2188 b = pop(); a = pop(); 2189 push( _gvn.transform( new MulINode(a,b) ) ); 2190 break; 2191 case Bytecodes::_iadd: 2192 b = pop(); a = pop(); 2193 push( _gvn.transform( new AddINode(a,b) ) ); 2194 break; 2195 case Bytecodes::_ineg: 2196 a = pop(); 2197 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 2198 break; 2199 case Bytecodes::_isub: 2200 b = pop(); a = pop(); 2201 push( _gvn.transform( new SubINode(a,b) ) ); 2202 break; 2203 case Bytecodes::_iand: 2204 b = pop(); a = pop(); 2205 push( _gvn.transform( new AndINode(a,b) ) ); 2206 break; 2207 case Bytecodes::_ior: 2208 b = pop(); a = pop(); 2209 push( _gvn.transform( new OrINode(a,b) ) ); 2210 break; 2211 case Bytecodes::_ixor: 2212 b = pop(); a = pop(); 2213 push( _gvn.transform( new XorINode(a,b) ) ); 2214 break; 2215 case Bytecodes::_ishl: 2216 b = pop(); a = pop(); 2217 push( _gvn.transform( new LShiftINode(a,b) ) ); 2218 break; 2219 case Bytecodes::_ishr: 2220 b = pop(); a = pop(); 2221 push( _gvn.transform( new RShiftINode(a,b) ) ); 2222 break; 2223 case Bytecodes::_iushr: 2224 b = pop(); a = pop(); 2225 push( _gvn.transform( new URShiftINode(a,b) ) ); 2226 break; 2227 2228 case Bytecodes::_fneg: 2229 a = pop(); 2230 b = _gvn.transform(new NegFNode (a)); 2231 push(b); 2232 break; 2233 2234 case Bytecodes::_fsub: 2235 b = pop(); 2236 a = pop(); 2237 c = _gvn.transform( new SubFNode(a,b) ); 2238 d = precision_rounding(c); 2239 push( d ); 2240 break; 2241 2242 case Bytecodes::_fadd: 2243 b = pop(); 2244 a = pop(); 2245 c = _gvn.transform( new AddFNode(a,b) ); 2246 d = precision_rounding(c); 2247 push( d ); 2248 break; 2249 2250 case Bytecodes::_fmul: 2251 b = pop(); 2252 a = pop(); 2253 c = _gvn.transform( new MulFNode(a,b) ); 2254 d = precision_rounding(c); 2255 push( d ); 2256 break; 2257 2258 case Bytecodes::_fdiv: 2259 b = pop(); 2260 a = pop(); 2261 c = _gvn.transform( new DivFNode(0,a,b) ); 2262 d = precision_rounding(c); 2263 push( d ); 2264 break; 2265 2266 case Bytecodes::_frem: 2267 if (Matcher::has_match_rule(Op_ModF)) { 2268 // Generate a ModF node. 2269 b = pop(); 2270 a = pop(); 2271 c = _gvn.transform( new ModFNode(0,a,b) ); 2272 d = precision_rounding(c); 2273 push( d ); 2274 } 2275 else { 2276 // Generate a call. 2277 modf(); 2278 } 2279 break; 2280 2281 case Bytecodes::_fcmpl: 2282 b = pop(); 2283 a = pop(); 2284 c = _gvn.transform( new CmpF3Node( a, b)); 2285 push(c); 2286 break; 2287 case Bytecodes::_fcmpg: 2288 b = pop(); 2289 a = pop(); 2290 2291 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 2292 // which negates the result sign except for unordered. Flip the unordered 2293 // as well by using CmpF3 which implements unordered-lesser instead of 2294 // unordered-greater semantics. Finally, commute the result bits. Result 2295 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 2296 c = _gvn.transform( new CmpF3Node( b, a)); 2297 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2298 push(c); 2299 break; 2300 2301 case Bytecodes::_f2i: 2302 a = pop(); 2303 push(_gvn.transform(new ConvF2INode(a))); 2304 break; 2305 2306 case Bytecodes::_d2i: 2307 a = pop_pair(); 2308 b = _gvn.transform(new ConvD2INode(a)); 2309 push( b ); 2310 break; 2311 2312 case Bytecodes::_f2d: 2313 a = pop(); 2314 b = _gvn.transform( new ConvF2DNode(a)); 2315 push_pair( b ); 2316 break; 2317 2318 case Bytecodes::_d2f: 2319 a = pop_pair(); 2320 b = _gvn.transform( new ConvD2FNode(a)); 2321 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 2322 //b = _gvn.transform(new RoundFloatNode(0, b) ); 2323 push( b ); 2324 break; 2325 2326 case Bytecodes::_l2f: 2327 if (Matcher::convL2FSupported()) { 2328 a = pop_pair(); 2329 b = _gvn.transform( new ConvL2FNode(a)); 2330 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits. 2331 // Rather than storing the result into an FP register then pushing 2332 // out to memory to round, the machine instruction that implements 2333 // ConvL2D is responsible for rounding. 2334 // c = precision_rounding(b); 2335 push(b); 2336 } else { 2337 l2f(); 2338 } 2339 break; 2340 2341 case Bytecodes::_l2d: 2342 a = pop_pair(); 2343 b = _gvn.transform( new ConvL2DNode(a)); 2344 // For x86_32.ad, rounding is always necessary (see _l2f above). 2345 // c = dprecision_rounding(b); 2346 push_pair(b); 2347 break; 2348 2349 case Bytecodes::_f2l: 2350 a = pop(); 2351 b = _gvn.transform( new ConvF2LNode(a)); 2352 push_pair(b); 2353 break; 2354 2355 case Bytecodes::_d2l: 2356 a = pop_pair(); 2357 b = _gvn.transform( new ConvD2LNode(a)); 2358 push_pair(b); 2359 break; 2360 2361 case Bytecodes::_dsub: 2362 b = pop_pair(); 2363 a = pop_pair(); 2364 c = _gvn.transform( new SubDNode(a,b) ); 2365 d = dprecision_rounding(c); 2366 push_pair( d ); 2367 break; 2368 2369 case Bytecodes::_dadd: 2370 b = pop_pair(); 2371 a = pop_pair(); 2372 c = _gvn.transform( new AddDNode(a,b) ); 2373 d = dprecision_rounding(c); 2374 push_pair( d ); 2375 break; 2376 2377 case Bytecodes::_dmul: 2378 b = pop_pair(); 2379 a = pop_pair(); 2380 c = _gvn.transform( new MulDNode(a,b) ); 2381 d = dprecision_rounding(c); 2382 push_pair( d ); 2383 break; 2384 2385 case Bytecodes::_ddiv: 2386 b = pop_pair(); 2387 a = pop_pair(); 2388 c = _gvn.transform( new DivDNode(0,a,b) ); 2389 d = dprecision_rounding(c); 2390 push_pair( d ); 2391 break; 2392 2393 case Bytecodes::_dneg: 2394 a = pop_pair(); 2395 b = _gvn.transform(new NegDNode (a)); 2396 push_pair(b); 2397 break; 2398 2399 case Bytecodes::_drem: 2400 if (Matcher::has_match_rule(Op_ModD)) { 2401 // Generate a ModD node. 2402 b = pop_pair(); 2403 a = pop_pair(); 2404 // a % b 2405 2406 c = _gvn.transform( new ModDNode(0,a,b) ); 2407 d = dprecision_rounding(c); 2408 push_pair( d ); 2409 } 2410 else { 2411 // Generate a call. 2412 modd(); 2413 } 2414 break; 2415 2416 case Bytecodes::_dcmpl: 2417 b = pop_pair(); 2418 a = pop_pair(); 2419 c = _gvn.transform( new CmpD3Node( a, b)); 2420 push(c); 2421 break; 2422 2423 case Bytecodes::_dcmpg: 2424 b = pop_pair(); 2425 a = pop_pair(); 2426 // Same as dcmpl but need to flip the unordered case. 2427 // Commute the inputs, which negates the result sign except for unordered. 2428 // Flip the unordered as well by using CmpD3 which implements 2429 // unordered-lesser instead of unordered-greater semantics. 2430 // Finally, negate the result bits. Result is same as using a 2431 // CmpD3Greater except we did it with CmpD3 alone. 2432 c = _gvn.transform( new CmpD3Node( b, a)); 2433 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 2434 push(c); 2435 break; 2436 2437 2438 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 2439 case Bytecodes::_land: 2440 b = pop_pair(); 2441 a = pop_pair(); 2442 c = _gvn.transform( new AndLNode(a,b) ); 2443 push_pair(c); 2444 break; 2445 case Bytecodes::_lor: 2446 b = pop_pair(); 2447 a = pop_pair(); 2448 c = _gvn.transform( new OrLNode(a,b) ); 2449 push_pair(c); 2450 break; 2451 case Bytecodes::_lxor: 2452 b = pop_pair(); 2453 a = pop_pair(); 2454 c = _gvn.transform( new XorLNode(a,b) ); 2455 push_pair(c); 2456 break; 2457 2458 case Bytecodes::_lshl: 2459 b = pop(); // the shift count 2460 a = pop_pair(); // value to be shifted 2461 c = _gvn.transform( new LShiftLNode(a,b) ); 2462 push_pair(c); 2463 break; 2464 case Bytecodes::_lshr: 2465 b = pop(); // the shift count 2466 a = pop_pair(); // value to be shifted 2467 c = _gvn.transform( new RShiftLNode(a,b) ); 2468 push_pair(c); 2469 break; 2470 case Bytecodes::_lushr: 2471 b = pop(); // the shift count 2472 a = pop_pair(); // value to be shifted 2473 c = _gvn.transform( new URShiftLNode(a,b) ); 2474 push_pair(c); 2475 break; 2476 case Bytecodes::_lmul: 2477 b = pop_pair(); 2478 a = pop_pair(); 2479 c = _gvn.transform( new MulLNode(a,b) ); 2480 push_pair(c); 2481 break; 2482 2483 case Bytecodes::_lrem: 2484 // Must keep both values on the expression-stack during null-check 2485 assert(peek(0) == top(), "long word order"); 2486 zero_check_long(peek(1)); 2487 // Compile-time detect of null-exception? 2488 if (stopped()) return; 2489 b = pop_pair(); 2490 a = pop_pair(); 2491 c = _gvn.transform( new ModLNode(control(),a,b) ); 2492 push_pair(c); 2493 break; 2494 2495 case Bytecodes::_ldiv: 2496 // Must keep both values on the expression-stack during null-check 2497 assert(peek(0) == top(), "long word order"); 2498 zero_check_long(peek(1)); 2499 // Compile-time detect of null-exception? 2500 if (stopped()) return; 2501 b = pop_pair(); 2502 a = pop_pair(); 2503 c = _gvn.transform( new DivLNode(control(),a,b) ); 2504 push_pair(c); 2505 break; 2506 2507 case Bytecodes::_ladd: 2508 b = pop_pair(); 2509 a = pop_pair(); 2510 c = _gvn.transform( new AddLNode(a,b) ); 2511 push_pair(c); 2512 break; 2513 case Bytecodes::_lsub: 2514 b = pop_pair(); 2515 a = pop_pair(); 2516 c = _gvn.transform( new SubLNode(a,b) ); 2517 push_pair(c); 2518 break; 2519 case Bytecodes::_lcmp: 2520 // Safepoints are now inserted _before_ branches. The long-compare 2521 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 2522 // slew of control flow. These are usually followed by a CmpI vs zero and 2523 // a branch; this pattern then optimizes to the obvious long-compare and 2524 // branch. However, if the branch is backwards there's a Safepoint 2525 // inserted. The inserted Safepoint captures the JVM state at the 2526 // pre-branch point, i.e. it captures the 3-way value. Thus if a 2527 // long-compare is used to control a loop the debug info will force 2528 // computation of the 3-way value, even though the generated code uses a 2529 // long-compare and branch. We try to rectify the situation by inserting 2530 // a SafePoint here and have it dominate and kill the safepoint added at a 2531 // following backwards branch. At this point the JVM state merely holds 2 2532 // longs but not the 3-way value. 2533 switch (iter().next_bc()) { 2534 case Bytecodes::_ifgt: 2535 case Bytecodes::_iflt: 2536 case Bytecodes::_ifge: 2537 case Bytecodes::_ifle: 2538 case Bytecodes::_ifne: 2539 case Bytecodes::_ifeq: 2540 // If this is a backwards branch in the bytecodes, add Safepoint 2541 maybe_add_safepoint(iter().next_get_dest()); 2542 default: 2543 break; 2544 } 2545 b = pop_pair(); 2546 a = pop_pair(); 2547 c = _gvn.transform( new CmpL3Node( a, b )); 2548 push(c); 2549 break; 2550 2551 case Bytecodes::_lneg: 2552 a = pop_pair(); 2553 b = _gvn.transform( new SubLNode(longcon(0),a)); 2554 push_pair(b); 2555 break; 2556 case Bytecodes::_l2i: 2557 a = pop_pair(); 2558 push( _gvn.transform( new ConvL2INode(a))); 2559 break; 2560 case Bytecodes::_i2l: 2561 a = pop(); 2562 b = _gvn.transform( new ConvI2LNode(a)); 2563 push_pair(b); 2564 break; 2565 case Bytecodes::_i2b: 2566 // Sign extend 2567 a = pop(); 2568 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true); 2569 push(a); 2570 break; 2571 case Bytecodes::_i2s: 2572 a = pop(); 2573 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true); 2574 push(a); 2575 break; 2576 case Bytecodes::_i2c: 2577 a = pop(); 2578 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true); 2579 push(a); 2580 break; 2581 2582 case Bytecodes::_i2f: 2583 a = pop(); 2584 b = _gvn.transform( new ConvI2FNode(a) ) ; 2585 c = precision_rounding(b); 2586 push (b); 2587 break; 2588 2589 case Bytecodes::_i2d: 2590 a = pop(); 2591 b = _gvn.transform( new ConvI2DNode(a)); 2592 push_pair(b); 2593 break; 2594 2595 case Bytecodes::_iinc: // Increment local 2596 i = iter().get_index(); // Get local index 2597 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 2598 break; 2599 2600 // Exit points of synchronized methods must have an unlock node 2601 case Bytecodes::_return: 2602 return_current(nullptr); 2603 break; 2604 2605 case Bytecodes::_ireturn: 2606 case Bytecodes::_areturn: 2607 case Bytecodes::_freturn: 2608 return_current(pop()); 2609 break; 2610 case Bytecodes::_lreturn: 2611 return_current(pop_pair()); 2612 break; 2613 case Bytecodes::_dreturn: 2614 return_current(pop_pair()); 2615 break; 2616 2617 case Bytecodes::_athrow: 2618 // null exception oop throws null pointer exception 2619 null_check(peek()); 2620 if (stopped()) return; 2621 // Hook the thrown exception directly to subsequent handlers. 2622 if (BailoutToInterpreterForThrows) { 2623 // Keep method interpreted from now on. 2624 uncommon_trap(Deoptimization::Reason_unhandled, 2625 Deoptimization::Action_make_not_compilable); 2626 return; 2627 } 2628 if (env()->jvmti_can_post_on_exceptions()) { 2629 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 2630 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 2631 } 2632 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 2633 add_exception_state(make_exception_state(peek())); 2634 break; 2635 2636 case Bytecodes::_goto: // fall through 2637 case Bytecodes::_goto_w: { 2638 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 2639 2640 // If this is a backwards branch in the bytecodes, add Safepoint 2641 maybe_add_safepoint(target_bci); 2642 2643 // Merge the current control into the target basic block 2644 merge(target_bci); 2645 2646 // See if we can get some profile data and hand it off to the next block 2647 Block *target_block = block()->successor_for_bci(target_bci); 2648 if (target_block->pred_count() != 1) break; 2649 ciMethodData* methodData = method()->method_data(); 2650 if (!methodData->is_mature()) break; 2651 ciProfileData* data = methodData->bci_to_data(bci()); 2652 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch"); 2653 int taken = ((ciJumpData*)data)->taken(); 2654 taken = method()->scale_count(taken); 2655 target_block->set_count(taken); 2656 break; 2657 } 2658 2659 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 2660 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 2661 handle_if_null: 2662 // If this is a backwards branch in the bytecodes, add Safepoint 2663 maybe_add_safepoint(iter().get_dest()); 2664 a = null(); 2665 b = pop(); 2666 if (!_gvn.type(b)->speculative_maybe_null() && 2667 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 2668 inc_sp(1); 2669 Node* null_ctl = top(); 2670 b = null_check_oop(b, &null_ctl, true, true, true); 2671 assert(null_ctl->is_top(), "no null control here"); 2672 dec_sp(1); 2673 } else if (_gvn.type(b)->speculative_always_null() && 2674 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 2675 inc_sp(1); 2676 b = null_assert(b); 2677 dec_sp(1); 2678 } 2679 c = _gvn.transform( new CmpPNode(b, a) ); 2680 do_ifnull(btest, c); 2681 break; 2682 2683 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 2684 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 2685 handle_if_acmp: 2686 // If this is a backwards branch in the bytecodes, add Safepoint 2687 maybe_add_safepoint(iter().get_dest()); 2688 a = pop(); 2689 b = pop(); 2690 c = _gvn.transform( new CmpPNode(b, a) ); 2691 c = optimize_cmp_with_klass(c); 2692 do_if(btest, c); 2693 break; 2694 2695 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 2696 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 2697 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 2698 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 2699 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 2700 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 2701 handle_ifxx: 2702 // If this is a backwards branch in the bytecodes, add Safepoint 2703 maybe_add_safepoint(iter().get_dest()); 2704 a = _gvn.intcon(0); 2705 b = pop(); 2706 c = _gvn.transform( new CmpINode(b, a) ); 2707 do_if(btest, c); 2708 break; 2709 2710 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 2711 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 2712 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 2713 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 2714 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 2715 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 2716 handle_if_icmp: 2717 // If this is a backwards branch in the bytecodes, add Safepoint 2718 maybe_add_safepoint(iter().get_dest()); 2719 a = pop(); 2720 b = pop(); 2721 c = _gvn.transform( new CmpINode( b, a ) ); 2722 do_if(btest, c); 2723 break; 2724 2725 case Bytecodes::_tableswitch: 2726 do_tableswitch(); 2727 break; 2728 2729 case Bytecodes::_lookupswitch: 2730 do_lookupswitch(); 2731 break; 2732 2733 case Bytecodes::_invokestatic: 2734 case Bytecodes::_invokedynamic: 2735 case Bytecodes::_invokespecial: 2736 case Bytecodes::_invokevirtual: 2737 case Bytecodes::_invokeinterface: 2738 do_call(); 2739 break; 2740 case Bytecodes::_checkcast: 2741 do_checkcast(); 2742 break; 2743 case Bytecodes::_instanceof: 2744 do_instanceof(); 2745 break; 2746 case Bytecodes::_anewarray: 2747 do_anewarray(); 2748 break; 2749 case Bytecodes::_newarray: 2750 do_newarray((BasicType)iter().get_index()); 2751 break; 2752 case Bytecodes::_multianewarray: 2753 do_multianewarray(); 2754 break; 2755 case Bytecodes::_new: 2756 do_new(); 2757 break; 2758 2759 case Bytecodes::_jsr: 2760 case Bytecodes::_jsr_w: 2761 do_jsr(); 2762 break; 2763 2764 case Bytecodes::_ret: 2765 do_ret(); 2766 break; 2767 2768 2769 case Bytecodes::_monitorenter: 2770 do_monitor_enter(); 2771 break; 2772 2773 case Bytecodes::_monitorexit: 2774 do_monitor_exit(); 2775 break; 2776 2777 case Bytecodes::_breakpoint: 2778 // Breakpoint set concurrently to compile 2779 // %%% use an uncommon trap? 2780 C->record_failure("breakpoint in method"); 2781 return; 2782 2783 default: 2784 #ifndef PRODUCT 2785 map()->dump(99); 2786 #endif 2787 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 2788 ShouldNotReachHere(); 2789 } 2790 2791 #ifndef PRODUCT 2792 constexpr int perBytecode = 6; 2793 if (C->should_print_igv(perBytecode)) { 2794 IdealGraphPrinter* printer = C->igv_printer(); 2795 char buffer[256]; 2796 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 2797 bool old = printer->traverse_outs(); 2798 printer->set_traverse_outs(true); 2799 printer->print_method(buffer, perBytecode); 2800 printer->set_traverse_outs(old); 2801 } 2802 #endif 2803 }