1 /* 2 * Copyright (c) 1998, 2025, 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 "ci/ciInlineKlass.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "ci/ciSymbols.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "interpreter/linkResolver.hpp" 31 #include "jvm_io.h" 32 #include "memory/resourceArea.hpp" 33 #include "memory/universe.hpp" 34 #include "oops/oop.inline.hpp" 35 #include "opto/addnode.hpp" 36 #include "opto/castnode.hpp" 37 #include "opto/convertnode.hpp" 38 #include "opto/divnode.hpp" 39 #include "opto/idealGraphPrinter.hpp" 40 #include "opto/idealKit.hpp" 41 #include "opto/inlinetypenode.hpp" 42 #include "opto/matcher.hpp" 43 #include "opto/memnode.hpp" 44 #include "opto/mulnode.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/parse.hpp" 47 #include "opto/runtime.hpp" 48 #include "runtime/deoptimization.hpp" 49 #include "runtime/sharedRuntime.hpp" 50 51 #ifndef PRODUCT 52 extern uint explicit_null_checks_inserted, 53 explicit_null_checks_elided; 54 #endif 55 56 Node* Parse::record_profile_for_speculation_at_array_load(Node* ld) { 57 // Feed unused profile data to type speculation 58 if (UseTypeSpeculation && UseArrayLoadStoreProfile) { 59 ciKlass* array_type = nullptr; 60 ciKlass* element_type = nullptr; 61 ProfilePtrKind element_ptr = ProfileMaybeNull; 62 bool flat_array = true; 63 bool null_free_array = true; 64 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 65 if (element_type != nullptr || element_ptr != ProfileMaybeNull) { 66 ld = record_profile_for_speculation(ld, element_type, element_ptr); 67 } 68 } 69 return ld; 70 } 71 72 73 //---------------------------------array_load---------------------------------- 74 void Parse::array_load(BasicType bt) { 75 const Type* elemtype = Type::TOP; 76 Node* adr = array_addressing(bt, 0, elemtype); 77 if (stopped()) return; // guaranteed null or range check 78 79 Node* array_index = pop(); 80 Node* array = pop(); 81 82 // Handle inline type arrays 83 const TypeOopPtr* element_ptr = elemtype->make_oopptr(); 84 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr(); 85 86 if (!array_type->is_not_flat()) { 87 // Cannot statically determine if array is a flat array, emit runtime check 88 assert(UseArrayFlattening && is_reference_type(bt) && element_ptr->can_be_inline_type() && 89 (!element_ptr->is_inlinetypeptr() || element_ptr->inline_klass()->maybe_flat_in_array()), "array can't be flat"); 90 IdealKit ideal(this); 91 IdealVariable res(ideal); 92 ideal.declarations_done(); 93 ideal.if_then(flat_array_test(array, /* flat = */ false)); { 94 // Non-flat array 95 sync_kit(ideal); 96 if (!array_type->is_flat()) { 97 assert(array_type->is_flat() || control()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found"); 98 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 99 DecoratorSet decorator_set = IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD; 100 if (needs_range_check(array_type->size(), array_index)) { 101 // We've emitted a RangeCheck but now insert an additional check between the range check and the actual load. 102 // We cannot pin the load to two separate nodes. Instead, we pin it conservatively here such that it cannot 103 // possibly float above the range check at any point. 104 decorator_set |= C2_UNKNOWN_CONTROL_LOAD; 105 } 106 Node* ld = access_load_at(array, adr, adr_type, element_ptr, bt, decorator_set); 107 if (element_ptr->is_inlinetypeptr()) { 108 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass()); 109 } 110 ideal.set(res, ld); 111 } 112 ideal.sync_kit(this); 113 } ideal.else_(); { 114 // Flat array 115 sync_kit(ideal); 116 if (!array_type->is_not_flat()) { 117 if (element_ptr->is_inlinetypeptr()) { 118 ciInlineKlass* vk = element_ptr->inline_klass(); 119 Node* flat_array = cast_to_flat_array(array, vk, false, false, false); 120 Node* vt = InlineTypeNode::make_from_flat_array(this, vk, flat_array, array_index); 121 ideal.set(res, vt); 122 } else { 123 // Element type is unknown, and thus we cannot statically determine the exact flat array layout. Emit a 124 // runtime call to correctly load the inline type element from the flat array. 125 Node* inline_type = load_from_unknown_flat_array(array, array_index, element_ptr); 126 bool is_null_free = array_type->is_null_free() || !UseNullableValueFlattening; 127 if (is_null_free) { 128 inline_type = cast_not_null(inline_type); 129 } 130 ideal.set(res, inline_type); 131 } 132 } 133 ideal.sync_kit(this); 134 } ideal.end_if(); 135 sync_kit(ideal); 136 Node* ld = _gvn.transform(ideal.value(res)); 137 ld = record_profile_for_speculation_at_array_load(ld); 138 push_node(bt, ld); 139 return; 140 } 141 142 if (elemtype == TypeInt::BOOL) { 143 bt = T_BOOLEAN; 144 } 145 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 146 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt, 147 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD); 148 ld = record_profile_for_speculation_at_array_load(ld); 149 // Loading an inline type from a non-flat array 150 if (element_ptr != nullptr && element_ptr->is_inlinetypeptr()) { 151 assert(!array_type->is_null_free() || !element_ptr->maybe_null(), "inline type array elements should never be null"); 152 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass()); 153 } 154 push_node(bt, ld); 155 } 156 157 Node* Parse::load_from_unknown_flat_array(Node* array, Node* array_index, const TypeOopPtr* element_ptr) { 158 // Below membars keep this access to an unknown flat array correctly 159 // ordered with other unknown and known flat array accesses. 160 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES)); 161 162 Node* call = nullptr; 163 { 164 // Re-execute flat array load if runtime call triggers deoptimization 165 PreserveReexecuteState preexecs(this); 166 jvms()->set_bci(_bci); 167 jvms()->set_should_reexecute(true); 168 inc_sp(2); 169 kill_dead_locals(); 170 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO, 171 OptoRuntime::load_unknown_inline_Type(), 172 OptoRuntime::load_unknown_inline_Java(), 173 nullptr, TypeRawPtr::BOTTOM, 174 array, array_index); 175 } 176 make_slow_call_ex(call, env()->Throwable_klass(), false); 177 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 178 179 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES)); 180 181 // Keep track of the information that the inline type is in flat arrays 182 const Type* unknown_value = element_ptr->is_instptr()->cast_to_flat_in_array(); 183 return _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value)); 184 } 185 186 //--------------------------------array_store---------------------------------- 187 void Parse::array_store(BasicType bt) { 188 const Type* elemtype = Type::TOP; 189 Node* adr = array_addressing(bt, type2size[bt], elemtype); 190 if (stopped()) return; // guaranteed null or range check 191 Node* stored_value_casted = nullptr; 192 if (bt == T_OBJECT) { 193 stored_value_casted = array_store_check(adr, elemtype); 194 if (stopped()) { 195 return; 196 } 197 } 198 Node* const stored_value = pop_node(bt); // Value to store 199 Node* const array_index = pop(); // Index in the array 200 Node* array = pop(); // The array itself 201 202 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr(); 203 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt); 204 205 if (elemtype == TypeInt::BOOL) { 206 bt = T_BOOLEAN; 207 } else if (bt == T_OBJECT) { 208 elemtype = elemtype->make_oopptr(); 209 const Type* stored_value_casted_type = _gvn.type(stored_value_casted); 210 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat. 211 // This is only legal for non-null stores because the array_store_check always passes for null, even 212 // if the array is null-free. Null stores are handled in GraphKit::inline_array_null_guard(). 213 bool not_inline = !stored_value_casted_type->maybe_null() && !stored_value_casted_type->is_oopptr()->can_be_inline_type(); 214 bool not_null_free = not_inline; 215 bool not_flat = not_inline || ( stored_value_casted_type->is_inlinetypeptr() && 216 !stored_value_casted_type->inline_klass()->maybe_flat_in_array()); 217 if (!array_type->is_not_null_free() && not_null_free) { 218 // Storing a non-inline type, mark array as not null-free. 219 array_type = array_type->cast_to_not_null_free(); 220 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type)); 221 replace_in_map(array, cast); 222 array = cast; 223 } 224 if (!array_type->is_not_flat() && not_flat) { 225 // Storing to a non-flat array, mark array as not flat. 226 array_type = array_type->cast_to_not_flat(); 227 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type)); 228 replace_in_map(array, cast); 229 array = cast; 230 } 231 232 if (!array_type->is_flat() && array_type->is_null_free()) { 233 // Store to non-flat null-free inline type array (elements can never be null) 234 assert(!stored_value_casted_type->maybe_null(), "should be guaranteed by array store check"); 235 if (elemtype->is_inlinetypeptr() && elemtype->inline_klass()->is_empty()) { 236 // Ignore empty inline stores, array is already initialized. 237 return; 238 } 239 } else if (!array_type->is_not_flat()) { 240 // Array might be a flat array, emit runtime checks (for nullptr, a simple inline_array_null_guard is sufficient). 241 assert(UseArrayFlattening && !not_flat && elemtype->is_oopptr()->can_be_inline_type() && 242 (!array_type->klass_is_exact() || array_type->is_flat()), "array can't be a flat array"); 243 // TODO 8350865 Depending on the available layouts, we can avoid this check in below flat/not-flat branches. Also the safe_for_replace arg is now always true. 244 array = inline_array_null_guard(array, stored_value_casted, 3, true); 245 IdealKit ideal(this); 246 ideal.if_then(flat_array_test(array, /* flat = */ false)); { 247 // Non-flat array 248 if (!array_type->is_flat()) { 249 sync_kit(ideal); 250 assert(array_type->is_flat() || ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found"); 251 inc_sp(3); 252 access_store_at(array, adr, adr_type, stored_value_casted, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false); 253 dec_sp(3); 254 ideal.sync_kit(this); 255 } 256 } ideal.else_(); { 257 // Flat array 258 sync_kit(ideal); 259 if (!array_type->is_not_flat()) { 260 // Try to determine the inline klass type of the stored value 261 ciInlineKlass* vk = nullptr; 262 if (stored_value_casted_type->is_inlinetypeptr()) { 263 vk = stored_value_casted_type->inline_klass(); 264 } else if (elemtype->is_inlinetypeptr()) { 265 vk = elemtype->inline_klass(); 266 } 267 268 if (vk != nullptr) { 269 // Element type is known, cast and store to flat array layout. 270 Node* flat_array = cast_to_flat_array(array, vk, false, false, false); 271 272 // Re-execute flat array store if buffering triggers deoptimization 273 PreserveReexecuteState preexecs(this); 274 jvms()->set_should_reexecute(true); 275 inc_sp(3); 276 277 if (!stored_value_casted->is_InlineType()) { 278 assert(_gvn.type(stored_value_casted) == TypePtr::NULL_PTR, "Unexpected value"); 279 stored_value_casted = InlineTypeNode::make_null(_gvn, vk); 280 } 281 282 stored_value_casted->as_InlineType()->store_flat_array(this, flat_array, array_index); 283 } else { 284 // Element type is unknown, emit a runtime call since the flat array layout is not statically known. 285 store_to_unknown_flat_array(array, array_index, stored_value_casted); 286 } 287 } 288 ideal.sync_kit(this); 289 } 290 ideal.end_if(); 291 sync_kit(ideal); 292 return; 293 } else if (!array_type->is_not_null_free()) { 294 // Array is not flat but may be null free 295 assert(elemtype->is_oopptr()->can_be_inline_type(), "array can't be null-free"); 296 array = inline_array_null_guard(array, stored_value_casted, 3, true); 297 } 298 } 299 inc_sp(3); 300 access_store_at(array, adr, adr_type, stored_value, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY); 301 dec_sp(3); 302 } 303 304 // Emit a runtime call to store to a flat array whose element type is either unknown (i.e. we do not know the flat 305 // array layout) or not exact (could have different flat array layouts at runtime). 306 void Parse::store_to_unknown_flat_array(Node* array, Node* const idx, Node* non_null_stored_value) { 307 // Below membars keep this access to an unknown flat array correctly 308 // ordered with other unknown and known flat array accesses. 309 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES)); 310 311 Node* call = nullptr; 312 { 313 // Re-execute flat array store if runtime call triggers deoptimization 314 PreserveReexecuteState preexecs(this); 315 jvms()->set_bci(_bci); 316 jvms()->set_should_reexecute(true); 317 inc_sp(3); 318 kill_dead_locals(); 319 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO, 320 OptoRuntime::store_unknown_inline_Type(), 321 OptoRuntime::store_unknown_inline_Java(), 322 nullptr, TypeRawPtr::BOTTOM, 323 non_null_stored_value, array, idx); 324 } 325 make_slow_call_ex(call, env()->Throwable_klass(), false); 326 327 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES)); 328 } 329 330 //------------------------------array_addressing------------------------------- 331 // Pull array and index from the stack. Compute pointer-to-element. 332 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) { 333 Node *idx = peek(0+vals); // Get from stack without popping 334 Node *ary = peek(1+vals); // in case of exception 335 336 // Null check the array base, with correct stack contents 337 ary = null_check(ary, T_ARRAY); 338 // Compile-time detect of null-exception? 339 if (stopped()) return top(); 340 341 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 342 const TypeInt* sizetype = arytype->size(); 343 elemtype = arytype->elem(); 344 345 if (UseUniqueSubclasses) { 346 const Type* el = elemtype->make_ptr(); 347 if (el && el->isa_instptr()) { 348 const TypeInstPtr* toop = el->is_instptr(); 349 if (toop->instance_klass()->unique_concrete_subklass()) { 350 // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 351 const Type* subklass = Type::get_const_type(toop->instance_klass()); 352 elemtype = subklass->join_speculative(el); 353 } 354 } 355 } 356 357 if (!arytype->is_loaded()) { 358 // Only fails for some -Xcomp runs 359 // The class is unloaded. We have to run this bytecode in the interpreter. 360 ciKlass* klass = arytype->unloaded_klass(); 361 362 uncommon_trap(Deoptimization::Reason_unloaded, 363 Deoptimization::Action_reinterpret, 364 klass, "!loaded array"); 365 return top(); 366 } 367 368 ary = create_speculative_inline_type_array_checks(ary, arytype, elemtype); 369 370 if (needs_range_check(sizetype, idx)) { 371 create_range_check(idx, ary, sizetype); 372 } else if (C->log() != nullptr) { 373 C->log()->elem("observe that='!need_range_check'"); 374 } 375 376 // Check for always knowing you are throwing a range-check exception 377 if (stopped()) return top(); 378 379 // Make array address computation control dependent to prevent it 380 // from floating above the range check during loop optimizations. 381 Node* ptr = array_element_address(ary, idx, type, sizetype, control()); 382 assert(ptr != top(), "top should go hand-in-hand with stopped"); 383 384 return ptr; 385 } 386 387 // Check if we need a range check for an array access. This is the case if the index is either negative or if it could 388 // be greater or equal the smallest possible array size (i.e. out-of-bounds). 389 bool Parse::needs_range_check(const TypeInt* size_type, const Node* index) const { 390 const TypeInt* index_type = _gvn.type(index)->is_int(); 391 return index_type->_hi >= size_type->_lo || index_type->_lo < 0; 392 } 393 394 void Parse::create_range_check(Node* idx, Node* ary, const TypeInt* sizetype) { 395 Node* tst; 396 if (sizetype->_hi <= 0) { 397 // The greatest array bound is negative, so we can conclude that we're 398 // compiling unreachable code, but the unsigned compare trick used below 399 // only works with non-negative lengths. Instead, hack "tst" to be zero so 400 // the uncommon_trap path will always be taken. 401 tst = _gvn.intcon(0); 402 } else { 403 // Range is constant in array-oop, so we can use the original state of mem 404 Node* len = load_array_length(ary); 405 406 // Test length vs index (standard trick using unsigned compare) 407 Node* chk = _gvn.transform(new CmpUNode(idx, len) ); 408 BoolTest::mask btest = BoolTest::lt; 409 tst = _gvn.transform(new BoolNode(chk, btest) ); 410 } 411 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN); 412 _gvn.set_type(rc, rc->Value(&_gvn)); 413 if (!tst->is_Con()) { 414 record_for_igvn(rc); 415 } 416 set_control(_gvn.transform(new IfTrueNode(rc))); 417 // Branch to failure if out of bounds 418 { 419 PreserveJVMState pjvms(this); 420 set_control(_gvn.transform(new IfFalseNode(rc))); 421 if (C->allow_range_check_smearing()) { 422 // Do not use builtin_throw, since range checks are sometimes 423 // made more stringent by an optimistic transformation. 424 // This creates "tentative" range checks at this point, 425 // which are not guaranteed to throw exceptions. 426 // See IfNode::Ideal, is_range_check, adjust_check. 427 uncommon_trap(Deoptimization::Reason_range_check, 428 Deoptimization::Action_make_not_entrant, 429 nullptr, "range_check"); 430 } else { 431 // If we have already recompiled with the range-check-widening 432 // heroic optimization turned off, then we must really be throwing 433 // range check exceptions. 434 builtin_throw(Deoptimization::Reason_range_check); 435 } 436 } 437 } 438 439 // For inline type arrays, we can use the profiling information for array accesses to speculate on the type, flatness, 440 // and null-freeness. We can either prepare the speculative type for later uses or emit explicit speculative checks with 441 // traps now. In the latter case, the speculative type guarantees can avoid additional runtime checks later (e.g. 442 // non-null-free implies non-flat which allows us to remove flatness checks). This makes the graph simpler. 443 Node* Parse::create_speculative_inline_type_array_checks(Node* array, const TypeAryPtr* array_type, 444 const Type*& element_type) { 445 if (!array_type->is_flat() && !array_type->is_not_flat()) { 446 // For arrays that might be flat, speculate that the array has the exact type reported in the profile data such that 447 // we can rely on a fixed memory layout (i.e. either a flat layout or not). 448 array = cast_to_speculative_array_type(array, array_type, element_type); 449 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) { 450 // Array is known to be either flat or not flat. If possible, update the speculative type by using the profile data 451 // at this bci. 452 array = cast_to_profiled_array_type(array); 453 } 454 455 // Even though the type does not tell us whether we have an inline type array or not, we can still check the profile data 456 // whether we have a non-null-free or non-flat array. Speculating on a non-null-free array doesn't help aaload but could 457 // be profitable for a subsequent aastore. 458 if (!array_type->is_null_free() && !array_type->is_not_null_free()) { 459 array = speculate_non_null_free_array(array, array_type); 460 } 461 if (!array_type->is_flat() && !array_type->is_not_flat()) { 462 array = speculate_non_flat_array(array, array_type); 463 } 464 return array; 465 } 466 467 // Speculate that the array has the exact type reported in the profile data. We emit a trap when this turns out to be 468 // wrong. On the fast path, we add a CheckCastPP to use the exact type. 469 Node* Parse::cast_to_speculative_array_type(Node* const array, const TypeAryPtr*& array_type, const Type*& element_type) { 470 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check; 471 ciKlass* speculative_array_type = array_type->speculative_type(); 472 if (too_many_traps_or_recompiles(reason) || speculative_array_type == nullptr) { 473 // No speculative type, check profile data at this bci 474 speculative_array_type = nullptr; 475 reason = Deoptimization::Reason_class_check; 476 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) { 477 ciKlass* profiled_element_type = nullptr; 478 ProfilePtrKind element_ptr = ProfileMaybeNull; 479 bool flat_array = true; 480 bool null_free_array = true; 481 method()->array_access_profiled_type(bci(), speculative_array_type, profiled_element_type, element_ptr, flat_array, 482 null_free_array); 483 } 484 } 485 if (speculative_array_type != nullptr) { 486 // Speculate that this array has the exact type reported by profile data 487 Node* casted_array = nullptr; 488 DEBUG_ONLY(Node* old_control = control();) 489 Node* slow_ctl = type_check_receiver(array, speculative_array_type, 1.0, &casted_array); 490 if (stopped()) { 491 // The check always fails and therefore profile information is incorrect. Don't use it. 492 assert(old_control == slow_ctl, "type check should have been removed"); 493 set_control(slow_ctl); 494 } else if (!slow_ctl->is_top()) { 495 { PreserveJVMState pjvms(this); 496 set_control(slow_ctl); 497 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 498 } 499 replace_in_map(array, casted_array); 500 array_type = _gvn.type(casted_array)->is_aryptr(); 501 element_type = array_type->elem(); 502 return casted_array; 503 } 504 } 505 return array; 506 } 507 508 // Create a CheckCastPP when the speculative type can improve the current type. 509 Node* Parse::cast_to_profiled_array_type(Node* const array) { 510 ciKlass* array_type = nullptr; 511 ciKlass* element_type = nullptr; 512 ProfilePtrKind element_ptr = ProfileMaybeNull; 513 bool flat_array = true; 514 bool null_free_array = true; 515 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array); 516 if (array_type != nullptr) { 517 return record_profile_for_speculation(array, array_type, ProfileMaybeNull); 518 } 519 return array; 520 } 521 522 // Speculate that the array is non-null-free. We emit a trap when this turns out to be 523 // wrong. On the fast path, we add a CheckCastPP to use the non-null-free type. 524 Node* Parse::speculate_non_null_free_array(Node* const array, const TypeAryPtr*& array_type) { 525 bool null_free_array = true; 526 Deoptimization::DeoptReason reason = Deoptimization::Reason_none; 527 if (array_type->speculative() != nullptr && 528 array_type->speculative()->is_aryptr()->is_not_null_free() && 529 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) { 530 null_free_array = false; 531 reason = Deoptimization::Reason_speculate_class_check; 532 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) { 533 ciKlass* profiled_array_type = nullptr; 534 ciKlass* profiled_element_type = nullptr; 535 ProfilePtrKind element_ptr = ProfileMaybeNull; 536 bool flat_array = true; 537 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array, 538 null_free_array); 539 reason = Deoptimization::Reason_class_check; 540 } 541 if (!null_free_array) { 542 { // Deoptimize if null-free array 543 BuildCutout unless(this, null_free_array_test(array, /* null_free = */ false), PROB_MAX); 544 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 545 } 546 assert(!stopped(), "null-free array should have been caught earlier"); 547 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_null_free())); 548 replace_in_map(array, casted_array); 549 array_type = _gvn.type(casted_array)->is_aryptr(); 550 return casted_array; 551 } 552 return array; 553 } 554 555 // Speculate that the array is non-flat. We emit a trap when this turns out to be wrong. 556 // On the fast path, we add a CheckCastPP to use the non-flat type. 557 Node* Parse::speculate_non_flat_array(Node* const array, const TypeAryPtr* const array_type) { 558 bool flat_array = true; 559 Deoptimization::DeoptReason reason = Deoptimization::Reason_none; 560 if (array_type->speculative() != nullptr && 561 array_type->speculative()->is_aryptr()->is_not_flat() && 562 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) { 563 flat_array = false; 564 reason = Deoptimization::Reason_speculate_class_check; 565 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) { 566 ciKlass* profiled_array_type = nullptr; 567 ciKlass* profiled_element_type = nullptr; 568 ProfilePtrKind element_ptr = ProfileMaybeNull; 569 bool null_free_array = true; 570 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array, 571 null_free_array); 572 reason = Deoptimization::Reason_class_check; 573 } 574 if (!flat_array) { 575 { // Deoptimize if flat array 576 BuildCutout unless(this, flat_array_test(array, /* flat = */ false), PROB_MAX); 577 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 578 } 579 assert(!stopped(), "flat array should have been caught earlier"); 580 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_flat())); 581 replace_in_map(array, casted_array); 582 return casted_array; 583 } 584 return array; 585 } 586 587 // returns IfNode 588 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) { 589 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 590 Node *tst = _gvn.transform(new BoolNode(cmp, mask)); 591 IfNode *iff = create_and_map_if(control(), tst, prob, cnt); 592 return iff; 593 } 594 595 596 // sentinel value for the target bci to mark never taken branches 597 // (according to profiling) 598 static const int never_reached = INT_MAX; 599 600 //------------------------------helper for tableswitch------------------------- 601 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) { 602 // True branch, use existing map info 603 { PreserveJVMState pjvms(this); 604 Node *iftrue = _gvn.transform( new IfTrueNode (iff) ); 605 set_control( iftrue ); 606 if (unc) { 607 repush_if_args(); 608 uncommon_trap(Deoptimization::Reason_unstable_if, 609 Deoptimization::Action_reinterpret, 610 nullptr, 611 "taken always"); 612 } else { 613 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 614 merge_new_path(dest_bci_if_true); 615 } 616 } 617 618 // False branch 619 Node *iffalse = _gvn.transform( new IfFalseNode(iff) ); 620 set_control( iffalse ); 621 } 622 623 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) { 624 // True branch, use existing map info 625 { PreserveJVMState pjvms(this); 626 Node *iffalse = _gvn.transform( new IfFalseNode (iff) ); 627 set_control( iffalse ); 628 if (unc) { 629 repush_if_args(); 630 uncommon_trap(Deoptimization::Reason_unstable_if, 631 Deoptimization::Action_reinterpret, 632 nullptr, 633 "taken never"); 634 } else { 635 assert(dest_bci_if_true != never_reached, "inconsistent dest"); 636 merge_new_path(dest_bci_if_true); 637 } 638 } 639 640 // False branch 641 Node *iftrue = _gvn.transform( new IfTrueNode(iff) ); 642 set_control( iftrue ); 643 } 644 645 void Parse::jump_if_always_fork(int dest_bci, bool unc) { 646 // False branch, use existing map and control() 647 if (unc) { 648 repush_if_args(); 649 uncommon_trap(Deoptimization::Reason_unstable_if, 650 Deoptimization::Action_reinterpret, 651 nullptr, 652 "taken never"); 653 } else { 654 assert(dest_bci != never_reached, "inconsistent dest"); 655 merge_new_path(dest_bci); 656 } 657 } 658 659 660 extern "C" { 661 static int jint_cmp(const void *i, const void *j) { 662 int a = *(jint *)i; 663 int b = *(jint *)j; 664 return a > b ? 1 : a < b ? -1 : 0; 665 } 666 } 667 668 669 class SwitchRange : public StackObj { 670 // a range of integers coupled with a bci destination 671 jint _lo; // inclusive lower limit 672 jint _hi; // inclusive upper limit 673 int _dest; 674 float _cnt; // how many times this range was hit according to profiling 675 676 public: 677 jint lo() const { return _lo; } 678 jint hi() const { return _hi; } 679 int dest() const { return _dest; } 680 bool is_singleton() const { return _lo == _hi; } 681 float cnt() const { return _cnt; } 682 683 void setRange(jint lo, jint hi, int dest, float cnt) { 684 assert(lo <= hi, "must be a non-empty range"); 685 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt; 686 assert(_cnt >= 0, ""); 687 } 688 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) { 689 assert(lo <= hi, "must be a non-empty range"); 690 if (lo == _hi+1) { 691 // see merge_ranges() comment below 692 if (trim_ranges) { 693 if (cnt == 0) { 694 if (_cnt != 0) { 695 return false; 696 } 697 if (dest != _dest) { 698 _dest = never_reached; 699 } 700 } else { 701 if (_cnt == 0) { 702 return false; 703 } 704 if (dest != _dest) { 705 return false; 706 } 707 } 708 } else { 709 if (dest != _dest) { 710 return false; 711 } 712 } 713 _hi = hi; 714 _cnt += cnt; 715 return true; 716 } 717 return false; 718 } 719 720 void set (jint value, int dest, float cnt) { 721 setRange(value, value, dest, cnt); 722 } 723 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) { 724 return adjoinRange(value, value, dest, cnt, trim_ranges); 725 } 726 bool adjoin(SwitchRange& other) { 727 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false); 728 } 729 730 void print() { 731 if (is_singleton()) 732 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt()); 733 else if (lo() == min_jint) 734 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt()); 735 else if (hi() == max_jint) 736 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt()); 737 else 738 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt()); 739 } 740 }; 741 742 // We try to minimize the number of ranges and the size of the taken 743 // ones using profiling data. When ranges are created, 744 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge 745 // if both were never hit or both were hit to build longer unreached 746 // ranges. Here, we now merge adjoining ranges with the same 747 // destination and finally set destination of unreached ranges to the 748 // special value never_reached because it can help minimize the number 749 // of tests that are necessary. 750 // 751 // For instance: 752 // [0, 1] to target1 sometimes taken 753 // [1, 2] to target1 never taken 754 // [2, 3] to target2 never taken 755 // would lead to: 756 // [0, 1] to target1 sometimes taken 757 // [1, 3] never taken 758 // 759 // (first 2 ranges to target1 are not merged) 760 static void merge_ranges(SwitchRange* ranges, int& rp) { 761 if (rp == 0) { 762 return; 763 } 764 int shift = 0; 765 for (int j = 0; j < rp; j++) { 766 SwitchRange& r1 = ranges[j-shift]; 767 SwitchRange& r2 = ranges[j+1]; 768 if (r1.adjoin(r2)) { 769 shift++; 770 } else if (shift > 0) { 771 ranges[j+1-shift] = r2; 772 } 773 } 774 rp -= shift; 775 for (int j = 0; j <= rp; j++) { 776 SwitchRange& r = ranges[j]; 777 if (r.cnt() == 0 && r.dest() != never_reached) { 778 r.setRange(r.lo(), r.hi(), never_reached, r.cnt()); 779 } 780 } 781 } 782 783 //-------------------------------do_tableswitch-------------------------------- 784 void Parse::do_tableswitch() { 785 // Get information about tableswitch 786 int default_dest = iter().get_dest_table(0); 787 jint lo_index = iter().get_int_table(1); 788 jint hi_index = iter().get_int_table(2); 789 int len = hi_index - lo_index + 1; 790 791 if (len < 1) { 792 // If this is a backward branch, add safepoint 793 maybe_add_safepoint(default_dest); 794 pop(); // the effect of the instruction execution on the operand stack 795 merge(default_dest); 796 return; 797 } 798 799 ciMethodData* methodData = method()->method_data(); 800 ciMultiBranchData* profile = nullptr; 801 if (methodData->is_mature() && UseSwitchProfiling) { 802 ciProfileData* data = methodData->bci_to_data(bci()); 803 if (data != nullptr && data->is_MultiBranchData()) { 804 profile = (ciMultiBranchData*)data; 805 } 806 } 807 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 808 809 // generate decision tree, using trichotomy when possible 810 int rnum = len+2; 811 bool makes_backward_branch = (default_dest <= bci()); 812 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 813 int rp = -1; 814 if (lo_index != min_jint) { 815 float cnt = 1.0F; 816 if (profile != nullptr) { 817 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F); 818 } 819 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt); 820 } 821 for (int j = 0; j < len; j++) { 822 jint match_int = lo_index+j; 823 int dest = iter().get_dest_table(j+3); 824 makes_backward_branch |= (dest <= bci()); 825 float cnt = 1.0F; 826 if (profile != nullptr) { 827 cnt = (float)profile->count_at(j); 828 } 829 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) { 830 ranges[++rp].set(match_int, dest, cnt); 831 } 832 } 833 jint highest = lo_index+(len-1); 834 assert(ranges[rp].hi() == highest, ""); 835 if (highest != max_jint) { 836 float cnt = 1.0F; 837 if (profile != nullptr) { 838 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F); 839 } 840 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) { 841 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt); 842 } 843 } 844 assert(rp < len+2, "not too many ranges"); 845 846 if (trim_ranges) { 847 merge_ranges(ranges, rp); 848 } 849 850 // Safepoint in case if backward branch observed 851 if (makes_backward_branch) { 852 add_safepoint(); 853 } 854 855 Node* lookup = pop(); // lookup value 856 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 857 } 858 859 860 //------------------------------do_lookupswitch-------------------------------- 861 void Parse::do_lookupswitch() { 862 // Get information about lookupswitch 863 int default_dest = iter().get_dest_table(0); 864 jint len = iter().get_int_table(1); 865 866 if (len < 1) { // If this is a backward branch, add safepoint 867 maybe_add_safepoint(default_dest); 868 pop(); // the effect of the instruction execution on the operand stack 869 merge(default_dest); 870 return; 871 } 872 873 ciMethodData* methodData = method()->method_data(); 874 ciMultiBranchData* profile = nullptr; 875 if (methodData->is_mature() && UseSwitchProfiling) { 876 ciProfileData* data = methodData->bci_to_data(bci()); 877 if (data != nullptr && data->is_MultiBranchData()) { 878 profile = (ciMultiBranchData*)data; 879 } 880 } 881 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 882 883 // generate decision tree, using trichotomy when possible 884 jint* table = NEW_RESOURCE_ARRAY(jint, len*3); 885 { 886 for (int j = 0; j < len; j++) { 887 table[3*j+0] = iter().get_int_table(2+2*j); 888 table[3*j+1] = iter().get_dest_table(2+2*j+1); 889 // Handle overflow when converting from uint to jint 890 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j)); 891 } 892 qsort(table, len, 3*sizeof(table[0]), jint_cmp); 893 } 894 895 float default_cnt = 1.0F; 896 if (profile != nullptr) { 897 juint defaults = max_juint - len; 898 default_cnt = (float)profile->default_count()/(float)defaults; 899 } 900 901 int rnum = len*2+1; 902 bool makes_backward_branch = (default_dest <= bci()); 903 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 904 int rp = -1; 905 for (int j = 0; j < len; j++) { 906 jint match_int = table[3*j+0]; 907 jint dest = table[3*j+1]; 908 jint cnt = table[3*j+2]; 909 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 910 makes_backward_branch |= (dest <= bci()); 911 float c = default_cnt * ((float)match_int - (float)next_lo); 912 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) { 913 assert(default_dest != never_reached, "sentinel value for dead destinations"); 914 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c); 915 } 916 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) { 917 assert(dest != never_reached, "sentinel value for dead destinations"); 918 ranges[++rp].set(match_int, dest, (float)cnt); 919 } 920 } 921 jint highest = table[3*(len-1)]; 922 assert(ranges[rp].hi() == highest, ""); 923 if (highest != max_jint && 924 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) { 925 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest)); 926 } 927 assert(rp < rnum, "not too many ranges"); 928 929 if (trim_ranges) { 930 merge_ranges(ranges, rp); 931 } 932 933 // Safepoint in case backward branch observed 934 if (makes_backward_branch) { 935 add_safepoint(); 936 } 937 938 Node *lookup = pop(); // lookup value 939 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 940 } 941 942 static float if_prob(float taken_cnt, float total_cnt) { 943 assert(taken_cnt <= total_cnt, ""); 944 if (total_cnt == 0) { 945 return PROB_FAIR; 946 } 947 float p = taken_cnt / total_cnt; 948 return clamp(p, PROB_MIN, PROB_MAX); 949 } 950 951 static float if_cnt(float cnt) { 952 if (cnt == 0) { 953 return COUNT_UNKNOWN; 954 } 955 return cnt; 956 } 957 958 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) { 959 float total_cnt = 0; 960 for (SwitchRange* sr = lo; sr <= hi; sr++) { 961 total_cnt += sr->cnt(); 962 } 963 return total_cnt; 964 } 965 966 class SwitchRanges : public ResourceObj { 967 public: 968 SwitchRange* _lo; 969 SwitchRange* _hi; 970 SwitchRange* _mid; 971 float _cost; 972 973 enum { 974 Start, 975 LeftDone, 976 RightDone, 977 Done 978 } _state; 979 980 SwitchRanges(SwitchRange *lo, SwitchRange *hi) 981 : _lo(lo), _hi(hi), _mid(nullptr), 982 _cost(0), _state(Start) { 983 } 984 985 SwitchRanges() 986 : _lo(nullptr), _hi(nullptr), _mid(nullptr), 987 _cost(0), _state(Start) {} 988 }; 989 990 // Estimate cost of performing a binary search on lo..hi 991 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) { 992 GrowableArray<SwitchRanges> tree; 993 SwitchRanges root(lo, hi); 994 tree.push(root); 995 996 float cost = 0; 997 do { 998 SwitchRanges& r = *tree.adr_at(tree.length()-1); 999 if (r._hi != r._lo) { 1000 if (r._mid == nullptr) { 1001 float r_cnt = sum_of_cnts(r._lo, r._hi); 1002 1003 if (r_cnt == 0) { 1004 tree.pop(); 1005 cost = 0; 1006 continue; 1007 } 1008 1009 SwitchRange* mid = nullptr; 1010 mid = r._lo; 1011 for (float cnt = 0; ; ) { 1012 assert(mid <= r._hi, "out of bounds"); 1013 cnt += mid->cnt(); 1014 if (cnt > r_cnt / 2) { 1015 break; 1016 } 1017 mid++; 1018 } 1019 assert(mid <= r._hi, "out of bounds"); 1020 r._mid = mid; 1021 r._cost = r_cnt / total_cnt; 1022 } 1023 r._cost += cost; 1024 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) { 1025 cost = 0; 1026 r._state = SwitchRanges::LeftDone; 1027 tree.push(SwitchRanges(r._lo, r._mid-1)); 1028 } else if (r._state < SwitchRanges::RightDone) { 1029 cost = 0; 1030 r._state = SwitchRanges::RightDone; 1031 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi)); 1032 } else { 1033 tree.pop(); 1034 cost = r._cost; 1035 } 1036 } else { 1037 tree.pop(); 1038 cost = r._cost; 1039 } 1040 } while (tree.length() > 0); 1041 1042 1043 return cost; 1044 } 1045 1046 // It sometimes pays off to test most common ranges before the binary search 1047 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) { 1048 uint nr = hi - lo + 1; 1049 float total_cnt = sum_of_cnts(lo, hi); 1050 1051 float min = compute_tree_cost(lo, hi, total_cnt); 1052 float extra = 1; 1053 float sub = 0; 1054 1055 SwitchRange* array1 = lo; 1056 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr); 1057 1058 SwitchRange* ranges = nullptr; 1059 1060 while (nr >= 2) { 1061 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays"); 1062 ranges = (lo == array1) ? array2 : array1; 1063 1064 // Find highest frequency range 1065 SwitchRange* candidate = lo; 1066 for (SwitchRange* sr = lo+1; sr <= hi; sr++) { 1067 if (sr->cnt() > candidate->cnt()) { 1068 candidate = sr; 1069 } 1070 } 1071 SwitchRange most_freq = *candidate; 1072 if (most_freq.cnt() == 0) { 1073 break; 1074 } 1075 1076 // Copy remaining ranges into another array 1077 int shift = 0; 1078 for (uint i = 0; i < nr; i++) { 1079 SwitchRange* sr = &lo[i]; 1080 if (sr != candidate) { 1081 ranges[i-shift] = *sr; 1082 } else { 1083 shift++; 1084 if (i > 0 && i < nr-1) { 1085 SwitchRange prev = lo[i-1]; 1086 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt()); 1087 if (prev.adjoin(lo[i+1])) { 1088 shift++; 1089 i++; 1090 } 1091 ranges[i-shift] = prev; 1092 } 1093 } 1094 } 1095 nr -= shift; 1096 1097 // Evaluate cost of testing the most common range and performing a 1098 // binary search on the other ranges 1099 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt); 1100 if (cost >= min) { 1101 break; 1102 } 1103 // swap arrays 1104 lo = &ranges[0]; 1105 hi = &ranges[nr-1]; 1106 1107 // It pays off: emit the test for the most common range 1108 assert(most_freq.cnt() > 0, "must be taken"); 1109 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo()))); 1110 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo())))); 1111 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le)); 1112 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt())); 1113 jump_if_true_fork(iff, most_freq.dest(), false); 1114 1115 sub += most_freq.cnt() / total_cnt; 1116 extra += 1 - sub; 1117 min = cost; 1118 } 1119 } 1120 1121 //----------------------------create_jump_tables------------------------------- 1122 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 1123 // Are jumptables enabled 1124 if (!UseJumpTables) return false; 1125 1126 // Are jumptables supported 1127 if (!Matcher::has_match_rule(Op_Jump)) return false; 1128 1129 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 1130 1131 // Decide if a guard is needed to lop off big ranges at either (or 1132 // both) end(s) of the input set. We'll call this the default target 1133 // even though we can't be sure that it is the true "default". 1134 1135 bool needs_guard = false; 1136 int default_dest; 1137 int64_t total_outlier_size = 0; 1138 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1; 1139 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1; 1140 1141 if (lo->dest() == hi->dest()) { 1142 total_outlier_size = hi_size + lo_size; 1143 default_dest = lo->dest(); 1144 } else if (lo_size > hi_size) { 1145 total_outlier_size = lo_size; 1146 default_dest = lo->dest(); 1147 } else { 1148 total_outlier_size = hi_size; 1149 default_dest = hi->dest(); 1150 } 1151 1152 float total = sum_of_cnts(lo, hi); 1153 float cost = compute_tree_cost(lo, hi, total); 1154 1155 // If a guard test will eliminate very sparse end ranges, then 1156 // it is worth the cost of an extra jump. 1157 float trimmed_cnt = 0; 1158 if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 1159 needs_guard = true; 1160 if (default_dest == lo->dest()) { 1161 trimmed_cnt += lo->cnt(); 1162 lo++; 1163 } 1164 if (default_dest == hi->dest()) { 1165 trimmed_cnt += hi->cnt(); 1166 hi--; 1167 } 1168 } 1169 1170 // Find the total number of cases and ranges 1171 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1; 1172 int num_range = hi - lo + 1; 1173 1174 // Don't create table if: too large, too small, or too sparse. 1175 if (num_cases > MaxJumpTableSize) 1176 return false; 1177 if (UseSwitchProfiling) { 1178 // MinJumpTableSize is set so with a well balanced binary tree, 1179 // when the number of ranges is MinJumpTableSize, it's cheaper to 1180 // go through a JumpNode that a tree of IfNodes. Average cost of a 1181 // tree of IfNodes with MinJumpTableSize is 1182 // log2f(MinJumpTableSize) comparisons. So if the cost computed 1183 // from profile data is less than log2f(MinJumpTableSize) then 1184 // going with the binary search is cheaper. 1185 if (cost < log2f(MinJumpTableSize)) { 1186 return false; 1187 } 1188 } else { 1189 if (num_cases < MinJumpTableSize) 1190 return false; 1191 } 1192 if (num_cases > (MaxJumpTableSparseness * num_range)) 1193 return false; 1194 1195 // Normalize table lookups to zero 1196 int lowval = lo->lo(); 1197 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) ); 1198 1199 // Generate a guard to protect against input keyvals that aren't 1200 // in the switch domain. 1201 if (needs_guard) { 1202 Node* size = _gvn.intcon(num_cases); 1203 Node* cmp = _gvn.transform(new CmpUNode(key_val, size)); 1204 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge)); 1205 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt)); 1206 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0); 1207 1208 total -= trimmed_cnt; 1209 } 1210 1211 // Create an ideal node JumpTable that has projections 1212 // of all possible ranges for a switch statement 1213 // The key_val input must be converted to a pointer offset and scaled. 1214 // Compare Parse::array_addressing above. 1215 1216 // Clean the 32-bit int into a real 64-bit offset. 1217 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 1218 // Make I2L conversion control dependent to prevent it from 1219 // floating above the range check during loop optimizations. 1220 // Do not use a narrow int type here to prevent the data path from dying 1221 // while the control path is not removed. This can happen if the type of key_val 1222 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast 1223 // would be replaced by TOP while C2 is not able to fold the corresponding range checks. 1224 // Set _carry_dependency for the cast to avoid being removed by IGVN. 1225 #ifdef _LP64 1226 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */); 1227 #endif 1228 1229 // Shift the value by wordsize so we have an index into the table, rather 1230 // than a switch value 1231 Node *shiftWord = _gvn.MakeConX(wordSize); 1232 key_val = _gvn.transform( new MulXNode( key_val, shiftWord)); 1233 1234 // Create the JumpNode 1235 Arena* arena = C->comp_arena(); 1236 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases); 1237 int i = 0; 1238 if (total == 0) { 1239 for (SwitchRange* r = lo; r <= hi; r++) { 1240 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 1241 probs[i] = 1.0F / num_cases; 1242 } 1243 } 1244 } else { 1245 for (SwitchRange* r = lo; r <= hi; r++) { 1246 float prob = r->cnt()/total; 1247 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 1248 probs[i] = prob / (r->hi() - r->lo() + 1); 1249 } 1250 } 1251 } 1252 1253 ciMethodData* methodData = method()->method_data(); 1254 ciMultiBranchData* profile = nullptr; 1255 if (methodData->is_mature()) { 1256 ciProfileData* data = methodData->bci_to_data(bci()); 1257 if (data != nullptr && data->is_MultiBranchData()) { 1258 profile = (ciMultiBranchData*)data; 1259 } 1260 } 1261 1262 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total)); 1263 1264 // These are the switch destinations hanging off the jumpnode 1265 i = 0; 1266 for (SwitchRange* r = lo; r <= hi; r++) { 1267 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) { 1268 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 1269 { 1270 PreserveJVMState pjvms(this); 1271 set_control(input); 1272 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0); 1273 } 1274 } 1275 } 1276 assert(i == num_cases, "miscount of cases"); 1277 stop_and_kill_map(); // no more uses for this JVMS 1278 return true; 1279 } 1280 1281 //----------------------------jump_switch_ranges------------------------------- 1282 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 1283 Block* switch_block = block(); 1284 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 1285 1286 if (switch_depth == 0) { 1287 // Do special processing for the top-level call. 1288 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 1289 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 1290 1291 // Decrement pred-numbers for the unique set of nodes. 1292 #ifdef ASSERT 1293 if (!trim_ranges) { 1294 // Ensure that the block's successors are a (duplicate-free) set. 1295 int successors_counted = 0; // block occurrences in [hi..lo] 1296 int unique_successors = switch_block->num_successors(); 1297 for (int i = 0; i < unique_successors; i++) { 1298 Block* target = switch_block->successor_at(i); 1299 1300 // Check that the set of successors is the same in both places. 1301 int successors_found = 0; 1302 for (SwitchRange* p = lo; p <= hi; p++) { 1303 if (p->dest() == target->start()) successors_found++; 1304 } 1305 assert(successors_found > 0, "successor must be known"); 1306 successors_counted += successors_found; 1307 } 1308 assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 1309 } 1310 #endif 1311 1312 // Maybe prune the inputs, based on the type of key_val. 1313 jint min_val = min_jint; 1314 jint max_val = max_jint; 1315 const TypeInt* ti = key_val->bottom_type()->isa_int(); 1316 if (ti != nullptr) { 1317 min_val = ti->_lo; 1318 max_val = ti->_hi; 1319 assert(min_val <= max_val, "invalid int type"); 1320 } 1321 while (lo->hi() < min_val) { 1322 lo++; 1323 } 1324 if (lo->lo() < min_val) { 1325 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt()); 1326 } 1327 while (hi->lo() > max_val) { 1328 hi--; 1329 } 1330 if (hi->hi() > max_val) { 1331 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt()); 1332 } 1333 1334 linear_search_switch_ranges(key_val, lo, hi); 1335 } 1336 1337 #ifndef PRODUCT 1338 if (switch_depth == 0) { 1339 _max_switch_depth = 0; 1340 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1; 1341 } 1342 #endif 1343 1344 assert(lo <= hi, "must be a non-empty set of ranges"); 1345 if (lo == hi) { 1346 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0); 1347 } else { 1348 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 1349 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 1350 1351 if (create_jump_tables(key_val, lo, hi)) return; 1352 1353 SwitchRange* mid = nullptr; 1354 float total_cnt = sum_of_cnts(lo, hi); 1355 1356 int nr = hi - lo + 1; 1357 if (UseSwitchProfiling) { 1358 // Don't keep the binary search tree balanced: pick up mid point 1359 // that split frequencies in half. 1360 float cnt = 0; 1361 for (SwitchRange* sr = lo; sr <= hi; sr++) { 1362 cnt += sr->cnt(); 1363 if (cnt >= total_cnt / 2) { 1364 mid = sr; 1365 break; 1366 } 1367 } 1368 } else { 1369 mid = lo + nr/2; 1370 1371 // if there is an easy choice, pivot at a singleton: 1372 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1373 1374 assert(lo < mid && mid <= hi, "good pivot choice"); 1375 assert(nr != 2 || mid == hi, "should pick higher of 2"); 1376 assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1377 } 1378 1379 1380 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo()); 1381 1382 if (mid->is_singleton()) { 1383 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt())); 1384 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0); 1385 1386 // Special Case: If there are exactly three ranges, and the high 1387 // and low range each go to the same place, omit the "gt" test, 1388 // since it will not discriminate anything. 1389 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo; 1390 1391 // if there is a higher range, test for it and process it: 1392 if (mid < hi && !eq_test_only) { 1393 // two comparisons of same values--should enable 1 test for 2 branches 1394 // Use BoolTest::lt instead of BoolTest::gt 1395 float cnt = sum_of_cnts(lo, mid-1); 1396 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt)); 1397 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) ); 1398 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) ); 1399 { PreserveJVMState pjvms(this); 1400 set_control(iffalse); 1401 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1402 } 1403 set_control(iftrue); 1404 } 1405 1406 } else { 1407 // mid is a range, not a singleton, so treat mid..hi as a unit 1408 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi); 1409 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)); 1410 1411 // if there is a higher range, test for it and process it: 1412 if (mid == hi) { 1413 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0); 1414 } else { 1415 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) ); 1416 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) ); 1417 { PreserveJVMState pjvms(this); 1418 set_control(iftrue); 1419 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1); 1420 } 1421 set_control(iffalse); 1422 } 1423 } 1424 1425 // in any case, process the lower range 1426 if (mid == lo) { 1427 if (mid->is_singleton()) { 1428 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1); 1429 } else { 1430 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0); 1431 } 1432 } else { 1433 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1434 } 1435 } 1436 1437 // Decrease pred_count for each successor after all is done. 1438 if (switch_depth == 0) { 1439 int unique_successors = switch_block->num_successors(); 1440 for (int i = 0; i < unique_successors; i++) { 1441 Block* target = switch_block->successor_at(i); 1442 // Throw away the pre-allocated path for each unique successor. 1443 target->next_path_num(); 1444 } 1445 } 1446 1447 #ifndef PRODUCT 1448 _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1449 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1450 SwitchRange* r; 1451 int nsing = 0; 1452 for( r = lo; r <= hi; r++ ) { 1453 if( r->is_singleton() ) nsing++; 1454 } 1455 tty->print(">>> "); 1456 _method->print_short_name(); 1457 tty->print_cr(" switch decision tree"); 1458 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1459 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1460 if (_max_switch_depth > _est_switch_depth) { 1461 tty->print_cr("******** BAD SWITCH DEPTH ********"); 1462 } 1463 tty->print(" "); 1464 for( r = lo; r <= hi; r++ ) { 1465 r->print(); 1466 } 1467 tty->cr(); 1468 } 1469 #endif 1470 } 1471 1472 Node* Parse::floating_point_mod(Node* a, Node* b, BasicType type) { 1473 assert(type == BasicType::T_FLOAT || type == BasicType::T_DOUBLE, "only float and double are floating points"); 1474 CallNode* mod = type == BasicType::T_DOUBLE ? static_cast<CallNode*>(new ModDNode(C, a, b)) : new ModFNode(C, a, b); 1475 1476 Node* prev_mem = set_predefined_input_for_runtime_call(mod); 1477 mod = _gvn.transform(mod)->as_Call(); 1478 set_predefined_output_for_runtime_call(mod, prev_mem, TypeRawPtr::BOTTOM); 1479 Node* result = _gvn.transform(new ProjNode(mod, TypeFunc::Parms + 0)); 1480 record_for_igvn(mod); 1481 return result; 1482 } 1483 1484 void Parse::l2f() { 1485 Node* f2 = pop(); 1486 Node* f1 = pop(); 1487 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1488 CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1489 "l2f", nullptr, //no memory effects 1490 f1, f2); 1491 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0)); 1492 1493 push(res); 1494 } 1495 1496 // Handle jsr and jsr_w bytecode 1497 void Parse::do_jsr() { 1498 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1499 1500 // Store information about current state, tagged with new _jsr_bci 1501 int return_bci = iter().next_bci(); 1502 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1503 1504 // The way we do things now, there is only one successor block 1505 // for the jsr, because the target code is cloned by ciTypeFlow. 1506 Block* target = successor_for_bci(jsr_bci); 1507 1508 // What got pushed? 1509 const Type* ret_addr = target->peek(); 1510 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1511 1512 // Effect on jsr on stack 1513 push(_gvn.makecon(ret_addr)); 1514 1515 // Flow to the jsr. 1516 merge(jsr_bci); 1517 } 1518 1519 // Handle ret bytecode 1520 void Parse::do_ret() { 1521 // Find to whom we return. 1522 assert(block()->num_successors() == 1, "a ret can only go one place now"); 1523 Block* target = block()->successor_at(0); 1524 assert(!target->is_ready(), "our arrival must be expected"); 1525 int pnum = target->next_path_num(); 1526 merge_common(target, pnum); 1527 } 1528 1529 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) { 1530 if (btest != BoolTest::eq && btest != BoolTest::ne) { 1531 // Only ::eq and ::ne are supported for profile injection. 1532 return false; 1533 } 1534 if (test->is_Cmp() && 1535 test->in(1)->Opcode() == Op_ProfileBoolean) { 1536 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1); 1537 int false_cnt = profile->false_count(); 1538 int true_cnt = profile->true_count(); 1539 1540 // Counts matching depends on the actual test operation (::eq or ::ne). 1541 // No need to scale the counts because profile injection was designed 1542 // to feed exact counts into VM. 1543 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt; 1544 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt; 1545 1546 profile->consume(); 1547 return true; 1548 } 1549 return false; 1550 } 1551 1552 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1553 // We also check that individual counters are positive first, otherwise the sum can become positive. 1554 // (check for saturation, integer overflow, and immature counts) 1555 static bool counters_are_meaningful(int counter1, int counter2, int min) { 1556 // check for saturation, including "uint" values too big to fit in "int" 1557 if (counter1 < 0 || counter2 < 0) { 1558 return false; 1559 } 1560 // check for integer overflow of the sum 1561 int64_t sum = (int64_t)counter1 + (int64_t)counter2; 1562 STATIC_ASSERT(sizeof(counter1) < sizeof(sum)); 1563 if (sum > INT_MAX) { 1564 return false; 1565 } 1566 // check if mature 1567 return (counter1 + counter2) >= min; 1568 } 1569 1570 //--------------------------dynamic_branch_prediction-------------------------- 1571 // Try to gather dynamic branch prediction behavior. Return a probability 1572 // of the branch being taken and set the "cnt" field. Returns a -1.0 1573 // if we need to use static prediction for some reason. 1574 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) { 1575 ResourceMark rm; 1576 1577 cnt = COUNT_UNKNOWN; 1578 1579 int taken = 0; 1580 int not_taken = 0; 1581 1582 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken); 1583 1584 if (use_mdo) { 1585 // Use MethodData information if it is available 1586 // FIXME: free the ProfileData structure 1587 ciMethodData* methodData = method()->method_data(); 1588 if (!methodData->is_mature()) return PROB_UNKNOWN; 1589 ciProfileData* data = methodData->bci_to_data(bci()); 1590 if (data == nullptr) { 1591 return PROB_UNKNOWN; 1592 } 1593 if (!data->is_JumpData()) return PROB_UNKNOWN; 1594 1595 // get taken and not taken values 1596 // NOTE: saturated UINT_MAX values become negative, 1597 // as do counts above INT_MAX. 1598 taken = data->as_JumpData()->taken(); 1599 not_taken = 0; 1600 if (data->is_BranchData()) { 1601 not_taken = data->as_BranchData()->not_taken(); 1602 } 1603 1604 // scale the counts to be commensurate with invocation counts: 1605 // NOTE: overflow for positive values is clamped at INT_MAX 1606 taken = method()->scale_count(taken); 1607 not_taken = method()->scale_count(not_taken); 1608 } 1609 // At this point, saturation or overflow is indicated by INT_MAX 1610 // or a negative value. 1611 1612 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1613 // We also check that individual counters are positive first, otherwise the sum can become positive. 1614 if (!counters_are_meaningful(taken, not_taken, 40)) { 1615 if (C->log() != nullptr) { 1616 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1617 } 1618 return PROB_UNKNOWN; 1619 } 1620 1621 // Compute frequency that we arrive here 1622 float sum = taken + not_taken; 1623 // Adjust, if this block is a cloned private block but the 1624 // Jump counts are shared. Taken the private counts for 1625 // just this path instead of the shared counts. 1626 if( block()->count() > 0 ) 1627 sum = block()->count(); 1628 cnt = sum / FreqCountInvocations; 1629 1630 // Pin probability to sane limits 1631 float prob; 1632 if( !taken ) 1633 prob = (0+PROB_MIN) / 2; 1634 else if( !not_taken ) 1635 prob = (1+PROB_MAX) / 2; 1636 else { // Compute probability of true path 1637 prob = (float)taken / (float)(taken + not_taken); 1638 if (prob > PROB_MAX) prob = PROB_MAX; 1639 if (prob < PROB_MIN) prob = PROB_MIN; 1640 } 1641 1642 assert((cnt > 0.0f) && (prob > 0.0f), 1643 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken); 1644 1645 if (C->log() != nullptr) { 1646 const char* prob_str = nullptr; 1647 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1648 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1649 char prob_str_buf[30]; 1650 if (prob_str == nullptr) { 1651 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob); 1652 prob_str = prob_str_buf; 1653 } 1654 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'", 1655 iter().get_dest(), taken, not_taken, cnt, prob_str); 1656 } 1657 return prob; 1658 } 1659 1660 //-----------------------------branch_prediction------------------------------- 1661 float Parse::branch_prediction(float& cnt, 1662 BoolTest::mask btest, 1663 int target_bci, 1664 Node* test) { 1665 float prob = dynamic_branch_prediction(cnt, btest, test); 1666 // If prob is unknown, switch to static prediction 1667 if (prob != PROB_UNKNOWN) return prob; 1668 1669 prob = PROB_FAIR; // Set default value 1670 if (btest == BoolTest::eq) // Exactly equal test? 1671 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1672 else if (btest == BoolTest::ne) 1673 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1674 1675 // If this is a conditional test guarding a backwards branch, 1676 // assume its a loop-back edge. Make it a likely taken branch. 1677 if (target_bci < bci()) { 1678 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1679 // Since it's an OSR, we probably have profile data, but since 1680 // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1681 // Let's make a special check here for completely zero counts. 1682 ciMethodData* methodData = method()->method_data(); 1683 if (!methodData->is_empty()) { 1684 ciProfileData* data = methodData->bci_to_data(bci()); 1685 // Only stop for truly zero counts, which mean an unknown part 1686 // of the OSR-ed method, and we want to deopt to gather more stats. 1687 // If you have ANY counts, then this loop is simply 'cold' relative 1688 // to the OSR loop. 1689 if (data == nullptr || 1690 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) { 1691 // This is the only way to return PROB_UNKNOWN: 1692 return PROB_UNKNOWN; 1693 } 1694 } 1695 } 1696 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1697 } 1698 1699 assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1700 return prob; 1701 } 1702 1703 // The magic constants are chosen so as to match the output of 1704 // branch_prediction() when the profile reports a zero taken count. 1705 // It is important to distinguish zero counts unambiguously, because 1706 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1707 // very small but nonzero probabilities, which if confused with zero 1708 // counts would keep the program recompiling indefinitely. 1709 bool Parse::seems_never_taken(float prob) const { 1710 return prob < PROB_MIN; 1711 } 1712 1713 //-------------------------------repush_if_args-------------------------------- 1714 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1715 inline int Parse::repush_if_args() { 1716 if (PrintOpto && WizardMode) { 1717 tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1718 Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1719 method()->print_name(); tty->cr(); 1720 } 1721 int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1722 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1723 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1724 assert(argument(0) != nullptr, "must exist"); 1725 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist"); 1726 inc_sp(bc_depth); 1727 return bc_depth; 1728 } 1729 1730 // Used by StressUnstableIfTraps 1731 static volatile int _trap_stress_counter = 0; 1732 1733 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) { 1734 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter)); 1735 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, MemNode::unordered); 1736 counter = _gvn.transform(new AddINode(counter, intcon(1))); 1737 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, MemNode::unordered); 1738 } 1739 1740 //----------------------------------do_ifnull---------------------------------- 1741 void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1742 int target_bci = iter().get_dest(); 1743 1744 Node* counter = nullptr; 1745 Node* incr_store = nullptr; 1746 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0); 1747 if (do_stress_trap) { 1748 increment_trap_stress_counter(counter, incr_store); 1749 } 1750 1751 Block* branch_block = successor_for_bci(target_bci); 1752 Block* next_block = successor_for_bci(iter().next_bci()); 1753 1754 float cnt; 1755 float prob = branch_prediction(cnt, btest, target_bci, c); 1756 if (prob == PROB_UNKNOWN) { 1757 // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1758 if (PrintOpto && Verbose) { 1759 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1760 } 1761 repush_if_args(); // to gather stats on loop 1762 uncommon_trap(Deoptimization::Reason_unreached, 1763 Deoptimization::Action_reinterpret, 1764 nullptr, "cold"); 1765 if (C->eliminate_boxing()) { 1766 // Mark the successor blocks as parsed 1767 branch_block->next_path_num(); 1768 next_block->next_path_num(); 1769 } 1770 return; 1771 } 1772 1773 NOT_PRODUCT(explicit_null_checks_inserted++); 1774 1775 // Generate real control flow 1776 Node *tst = _gvn.transform( new BoolNode( c, btest ) ); 1777 1778 // Sanity check the probability value 1779 assert(prob > 0.0f,"Bad probability in Parser"); 1780 // Need xform to put node in hash table 1781 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1782 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1783 // True branch 1784 { PreserveJVMState pjvms(this); 1785 Node* iftrue = _gvn.transform( new IfTrueNode (iff) ); 1786 set_control(iftrue); 1787 1788 if (stopped()) { // Path is dead? 1789 NOT_PRODUCT(explicit_null_checks_elided++); 1790 if (C->eliminate_boxing()) { 1791 // Mark the successor block as parsed 1792 branch_block->next_path_num(); 1793 } 1794 } else { // Path is live. 1795 adjust_map_after_if(btest, c, prob, branch_block); 1796 if (!stopped()) { 1797 merge(target_bci); 1798 } 1799 } 1800 } 1801 1802 // False branch 1803 Node* iffalse = _gvn.transform( new IfFalseNode(iff) ); 1804 set_control(iffalse); 1805 1806 if (stopped()) { // Path is dead? 1807 NOT_PRODUCT(explicit_null_checks_elided++); 1808 if (C->eliminate_boxing()) { 1809 // Mark the successor block as parsed 1810 next_block->next_path_num(); 1811 } 1812 } else { // Path is live. 1813 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block); 1814 } 1815 1816 if (do_stress_trap) { 1817 stress_trap(iff, counter, incr_store); 1818 } 1819 } 1820 1821 //------------------------------------do_if------------------------------------ 1822 void Parse::do_if(BoolTest::mask btest, Node* c, bool can_trap, bool new_path, Node** ctrl_taken) { 1823 int target_bci = iter().get_dest(); 1824 1825 Block* branch_block = successor_for_bci(target_bci); 1826 Block* next_block = successor_for_bci(iter().next_bci()); 1827 1828 float cnt; 1829 float prob = branch_prediction(cnt, btest, target_bci, c); 1830 float untaken_prob = 1.0 - prob; 1831 1832 if (prob == PROB_UNKNOWN) { 1833 if (PrintOpto && Verbose) { 1834 tty->print_cr("Never-taken edge stops compilation at bci %d", bci()); 1835 } 1836 repush_if_args(); // to gather stats on loop 1837 uncommon_trap(Deoptimization::Reason_unreached, 1838 Deoptimization::Action_reinterpret, 1839 nullptr, "cold"); 1840 if (C->eliminate_boxing()) { 1841 // Mark the successor blocks as parsed 1842 branch_block->next_path_num(); 1843 next_block->next_path_num(); 1844 } 1845 return; 1846 } 1847 1848 Node* counter = nullptr; 1849 Node* incr_store = nullptr; 1850 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0); 1851 if (do_stress_trap) { 1852 increment_trap_stress_counter(counter, incr_store); 1853 } 1854 1855 // Sanity check the probability value 1856 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1857 1858 bool taken_if_true = true; 1859 // Convert BoolTest to canonical form: 1860 if (!BoolTest(btest).is_canonical()) { 1861 btest = BoolTest(btest).negate(); 1862 taken_if_true = false; 1863 // prob is NOT updated here; it remains the probability of the taken 1864 // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1865 } 1866 assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1867 1868 Node* tst0 = new BoolNode(c, btest); 1869 Node* tst = _gvn.transform(tst0); 1870 BoolTest::mask taken_btest = BoolTest::illegal; 1871 BoolTest::mask untaken_btest = BoolTest::illegal; 1872 1873 if (tst->is_Bool()) { 1874 // Refresh c from the transformed bool node, since it may be 1875 // simpler than the original c. Also re-canonicalize btest. 1876 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)). 1877 // That can arise from statements like: if (x instanceof C) ... 1878 if (tst != tst0) { 1879 // Canonicalize one more time since transform can change it. 1880 btest = tst->as_Bool()->_test._test; 1881 if (!BoolTest(btest).is_canonical()) { 1882 // Reverse edges one more time... 1883 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1884 btest = tst->as_Bool()->_test._test; 1885 assert(BoolTest(btest).is_canonical(), "sanity"); 1886 taken_if_true = !taken_if_true; 1887 } 1888 c = tst->in(1); 1889 } 1890 BoolTest::mask neg_btest = BoolTest(btest).negate(); 1891 taken_btest = taken_if_true ? btest : neg_btest; 1892 untaken_btest = taken_if_true ? neg_btest : btest; 1893 } 1894 1895 // Generate real control flow 1896 float true_prob = (taken_if_true ? prob : untaken_prob); 1897 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1898 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1899 Node* taken_branch = new IfTrueNode(iff); 1900 Node* untaken_branch = new IfFalseNode(iff); 1901 if (!taken_if_true) { // Finish conversion to canonical form 1902 Node* tmp = taken_branch; 1903 taken_branch = untaken_branch; 1904 untaken_branch = tmp; 1905 } 1906 1907 // Branch is taken: 1908 { PreserveJVMState pjvms(this); 1909 taken_branch = _gvn.transform(taken_branch); 1910 set_control(taken_branch); 1911 1912 if (stopped()) { 1913 if (C->eliminate_boxing() && !new_path) { 1914 // Mark the successor block as parsed (if we haven't created a new path) 1915 branch_block->next_path_num(); 1916 } 1917 } else { 1918 adjust_map_after_if(taken_btest, c, prob, branch_block, can_trap); 1919 if (!stopped()) { 1920 if (new_path) { 1921 // Merge by using a new path 1922 merge_new_path(target_bci); 1923 } else if (ctrl_taken != nullptr) { 1924 // Don't merge but save taken branch to be wired by caller 1925 *ctrl_taken = control(); 1926 } else { 1927 merge(target_bci); 1928 } 1929 } 1930 } 1931 } 1932 1933 untaken_branch = _gvn.transform(untaken_branch); 1934 set_control(untaken_branch); 1935 1936 // Branch not taken. 1937 if (stopped() && ctrl_taken == nullptr) { 1938 if (C->eliminate_boxing()) { 1939 // Mark the successor block as parsed (if caller does not re-wire control flow) 1940 next_block->next_path_num(); 1941 } 1942 } else { 1943 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block, can_trap); 1944 } 1945 1946 if (do_stress_trap) { 1947 stress_trap(iff, counter, incr_store); 1948 } 1949 } 1950 1951 1952 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) { 1953 if (t->speculative() == nullptr) { 1954 return ProfileUnknownNull; 1955 } 1956 if (t->speculative_always_null()) { 1957 return ProfileAlwaysNull; 1958 } 1959 if (t->speculative_maybe_null()) { 1960 return ProfileMaybeNull; 1961 } 1962 return ProfileNeverNull; 1963 } 1964 1965 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) { 1966 inc_sp(2); 1967 Node* cast = null_check_common(input, T_OBJECT, true, nullptr, 1968 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) && 1969 speculative_ptr_kind(tinput) == ProfileAlwaysNull); 1970 dec_sp(2); 1971 if (btest == BoolTest::ne) { 1972 { 1973 PreserveJVMState pjvms(this); 1974 replace_in_map(input, cast); 1975 int target_bci = iter().get_dest(); 1976 merge(target_bci); 1977 } 1978 record_for_igvn(eq_region); 1979 set_control(_gvn.transform(eq_region)); 1980 } else { 1981 replace_in_map(input, cast); 1982 } 1983 } 1984 1985 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) { 1986 inc_sp(2); 1987 null_ctl = top(); 1988 Node* cast = null_check_oop(input, &null_ctl, 1989 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)), 1990 false, 1991 speculative_ptr_kind(tinput) == ProfileNeverNull && 1992 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)); 1993 dec_sp(2); 1994 assert(!stopped(), "null input should have been caught earlier"); 1995 return cast; 1996 } 1997 1998 void Parse::acmp_known_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, ciKlass* input_type, BoolTest::mask btest, Node* eq_region) { 1999 Node* ne_region = new RegionNode(1); 2000 Node* null_ctl; 2001 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl); 2002 ne_region->add_req(null_ctl); 2003 2004 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast); 2005 { 2006 PreserveJVMState pjvms(this); 2007 inc_sp(2); 2008 set_control(slow_ctl); 2009 Deoptimization::DeoptReason reason; 2010 if (tinput->speculative_type() != nullptr && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) { 2011 reason = Deoptimization::Reason_speculate_class_check; 2012 } else { 2013 reason = Deoptimization::Reason_class_check; 2014 } 2015 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile); 2016 } 2017 ne_region->add_req(control()); 2018 2019 record_for_igvn(ne_region); 2020 set_control(_gvn.transform(ne_region)); 2021 if (btest == BoolTest::ne) { 2022 { 2023 PreserveJVMState pjvms(this); 2024 if (null_ctl == top()) { 2025 replace_in_map(input, cast); 2026 } 2027 int target_bci = iter().get_dest(); 2028 merge(target_bci); 2029 } 2030 record_for_igvn(eq_region); 2031 set_control(_gvn.transform(eq_region)); 2032 } else { 2033 if (null_ctl == top()) { 2034 replace_in_map(input, cast); 2035 } 2036 set_control(_gvn.transform(ne_region)); 2037 } 2038 } 2039 2040 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) { 2041 Node* ne_region = new RegionNode(1); 2042 Node* null_ctl; 2043 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl); 2044 ne_region->add_req(null_ctl); 2045 2046 { 2047 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX); 2048 inc_sp(2); 2049 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile); 2050 } 2051 2052 ne_region->add_req(control()); 2053 2054 record_for_igvn(ne_region); 2055 set_control(_gvn.transform(ne_region)); 2056 if (btest == BoolTest::ne) { 2057 { 2058 PreserveJVMState pjvms(this); 2059 if (null_ctl == top()) { 2060 replace_in_map(input, cast); 2061 } 2062 int target_bci = iter().get_dest(); 2063 merge(target_bci); 2064 } 2065 record_for_igvn(eq_region); 2066 set_control(_gvn.transform(eq_region)); 2067 } else { 2068 if (null_ctl == top()) { 2069 replace_in_map(input, cast); 2070 } 2071 set_control(_gvn.transform(ne_region)); 2072 } 2073 } 2074 2075 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) { 2076 ciKlass* left_type = nullptr; 2077 ciKlass* right_type = nullptr; 2078 ProfilePtrKind left_ptr = ProfileUnknownNull; 2079 ProfilePtrKind right_ptr = ProfileUnknownNull; 2080 bool left_inline_type = true; 2081 bool right_inline_type = true; 2082 2083 // Leverage profiling at acmp 2084 if (UseACmpProfile) { 2085 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type); 2086 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) { 2087 left_type = nullptr; 2088 right_type = nullptr; 2089 left_inline_type = true; 2090 right_inline_type = true; 2091 } 2092 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) { 2093 left_ptr = ProfileUnknownNull; 2094 right_ptr = ProfileUnknownNull; 2095 } 2096 } 2097 2098 if (UseTypeSpeculation) { 2099 record_profile_for_speculation(left, left_type, left_ptr); 2100 record_profile_for_speculation(right, right_type, right_ptr); 2101 } 2102 2103 if (!EnableValhalla) { 2104 Node* cmp = CmpP(left, right); 2105 cmp = optimize_cmp_with_klass(cmp); 2106 do_if(btest, cmp); 2107 return; 2108 } 2109 2110 // Check for equality before potentially allocating 2111 if (left == right) { 2112 do_if(btest, makecon(TypeInt::CC_EQ)); 2113 return; 2114 } 2115 2116 // Allocate inline type operands and re-execute on deoptimization 2117 if (left->is_InlineType()) { 2118 if (_gvn.type(right)->is_zero_type() || 2119 (right->is_InlineType() && _gvn.type(right->as_InlineType()->get_is_init())->is_zero_type())) { 2120 // Null checking a scalarized but nullable inline type. Check the IsInit 2121 // input instead of the oop input to avoid keeping buffer allocations alive. 2122 Node* cmp = CmpI(left->as_InlineType()->get_is_init(), intcon(0)); 2123 do_if(btest, cmp); 2124 return; 2125 } else { 2126 PreserveReexecuteState preexecs(this); 2127 inc_sp(2); 2128 jvms()->set_should_reexecute(true); 2129 left = left->as_InlineType()->buffer(this)->get_oop(); 2130 } 2131 } 2132 if (right->is_InlineType()) { 2133 PreserveReexecuteState preexecs(this); 2134 inc_sp(2); 2135 jvms()->set_should_reexecute(true); 2136 right = right->as_InlineType()->buffer(this)->get_oop(); 2137 } 2138 2139 // First, do a normal pointer comparison 2140 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr(); 2141 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr(); 2142 Node* cmp = CmpP(left, right); 2143 cmp = optimize_cmp_with_klass(cmp); 2144 if (tleft == nullptr || !tleft->can_be_inline_type() || 2145 tright == nullptr || !tright->can_be_inline_type()) { 2146 // This is sufficient, if one of the operands can't be an inline type 2147 do_if(btest, cmp); 2148 return; 2149 } 2150 2151 // Don't add traps to unstable if branches because additional checks are required to 2152 // decide if the operands are equal/substitutable and we therefore shouldn't prune 2153 // branches for one if based on the profiling of the acmp branches. 2154 // Also, OptimizeUnstableIf would set an incorrect re-rexecution state because it 2155 // assumes that there is a 1-1 mapping between the if and the acmp branches and that 2156 // hitting a trap means that we will take the corresponding acmp branch on re-execution. 2157 const bool can_trap = true; 2158 2159 Node* eq_region = nullptr; 2160 if (btest == BoolTest::eq) { 2161 do_if(btest, cmp, !can_trap, true); 2162 if (stopped()) { 2163 // Pointers are equal, operands must be equal 2164 return; 2165 } 2166 } else { 2167 assert(btest == BoolTest::ne, "only eq or ne"); 2168 Node* is_not_equal = nullptr; 2169 eq_region = new RegionNode(3); 2170 { 2171 PreserveJVMState pjvms(this); 2172 // Pointers are not equal, but more checks are needed to determine if the operands are (not) substitutable 2173 do_if(btest, cmp, !can_trap, false, &is_not_equal); 2174 if (!stopped()) { 2175 eq_region->init_req(1, control()); 2176 } 2177 } 2178 if (is_not_equal == nullptr || is_not_equal->is_top()) { 2179 record_for_igvn(eq_region); 2180 set_control(_gvn.transform(eq_region)); 2181 return; 2182 } 2183 set_control(is_not_equal); 2184 } 2185 2186 // Prefer speculative types if available 2187 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) { 2188 if (tleft->speculative_type() != nullptr) { 2189 left_type = tleft->speculative_type(); 2190 } 2191 if (tright->speculative_type() != nullptr) { 2192 right_type = tright->speculative_type(); 2193 } 2194 } 2195 2196 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) { 2197 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft); 2198 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) { 2199 left_ptr = speculative_left_ptr; 2200 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) { 2201 left_ptr = speculative_left_ptr; 2202 } 2203 } 2204 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) { 2205 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright); 2206 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) { 2207 right_ptr = speculative_right_ptr; 2208 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) { 2209 right_ptr = speculative_right_ptr; 2210 } 2211 } 2212 2213 if (left_ptr == ProfileAlwaysNull) { 2214 // Comparison with null. Assert the input is indeed null and we're done. 2215 acmp_always_null_input(left, tleft, btest, eq_region); 2216 return; 2217 } 2218 if (right_ptr == ProfileAlwaysNull) { 2219 // Comparison with null. Assert the input is indeed null and we're done. 2220 acmp_always_null_input(right, tright, btest, eq_region); 2221 return; 2222 } 2223 if (left_type != nullptr && !left_type->is_inlinetype()) { 2224 // Comparison with an object of known type 2225 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region); 2226 return; 2227 } 2228 if (right_type != nullptr && !right_type->is_inlinetype()) { 2229 // Comparison with an object of known type 2230 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region); 2231 return; 2232 } 2233 if (!left_inline_type) { 2234 // Comparison with an object known not to be an inline type 2235 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region); 2236 return; 2237 } 2238 if (!right_inline_type) { 2239 // Comparison with an object known not to be an inline type 2240 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region); 2241 return; 2242 } 2243 2244 // Pointers are not equal, check if first operand is non-null 2245 Node* ne_region = new RegionNode(6); 2246 Node* null_ctl; 2247 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl); 2248 ne_region->init_req(1, null_ctl); 2249 2250 // First operand is non-null, check if it is an inline type 2251 Node* is_value = inline_type_test(not_null_right); 2252 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN); 2253 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff)); 2254 ne_region->init_req(2, not_value); 2255 set_control(_gvn.transform(new IfTrueNode(is_value_iff))); 2256 2257 // The first operand is an inline type, check if the second operand is non-null 2258 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl); 2259 ne_region->init_req(3, null_ctl); 2260 2261 // Check if both operands are of the same class. 2262 Node* kls_left = load_object_klass(not_null_left); 2263 Node* kls_right = load_object_klass(not_null_right); 2264 Node* kls_cmp = CmpP(kls_left, kls_right); 2265 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne)); 2266 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN); 2267 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff)); 2268 set_control(_gvn.transform(new IfFalseNode(kls_iff))); 2269 ne_region->init_req(4, kls_ne); 2270 2271 if (stopped()) { 2272 record_for_igvn(ne_region); 2273 set_control(_gvn.transform(ne_region)); 2274 if (btest == BoolTest::ne) { 2275 { 2276 PreserveJVMState pjvms(this); 2277 int target_bci = iter().get_dest(); 2278 merge(target_bci); 2279 } 2280 record_for_igvn(eq_region); 2281 set_control(_gvn.transform(eq_region)); 2282 } 2283 return; 2284 } 2285 2286 // Both operands are values types of the same class, we need to perform a 2287 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable(). 2288 Node* ne_io_phi = PhiNode::make(ne_region, i_o()); 2289 Node* mem = reset_memory(); 2290 Node* ne_mem_phi = PhiNode::make(ne_region, mem); 2291 2292 Node* eq_io_phi = nullptr; 2293 Node* eq_mem_phi = nullptr; 2294 if (eq_region != nullptr) { 2295 eq_io_phi = PhiNode::make(eq_region, i_o()); 2296 eq_mem_phi = PhiNode::make(eq_region, mem); 2297 } 2298 2299 set_all_memory(mem); 2300 2301 kill_dead_locals(); 2302 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature()); 2303 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method); 2304 call->set_override_symbolic_info(true); 2305 call->init_req(TypeFunc::Parms, not_null_left); 2306 call->init_req(TypeFunc::Parms+1, not_null_right); 2307 inc_sp(2); 2308 set_edges_for_java_call(call, false, false); 2309 Node* ret = set_results_for_java_call(call, false, true); 2310 dec_sp(2); 2311 2312 // Test the return value of ValueObjectMethods::isSubstitutable() 2313 // This is the last check, do_if can emit traps now. 2314 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1))); 2315 Node* ctl = C->top(); 2316 if (btest == BoolTest::eq) { 2317 PreserveJVMState pjvms(this); 2318 do_if(btest, subst_cmp, can_trap); 2319 if (!stopped()) { 2320 ctl = control(); 2321 } 2322 } else { 2323 assert(btest == BoolTest::ne, "only eq or ne"); 2324 PreserveJVMState pjvms(this); 2325 do_if(btest, subst_cmp, can_trap, false, &ctl); 2326 if (!stopped()) { 2327 eq_region->init_req(2, control()); 2328 eq_io_phi->init_req(2, i_o()); 2329 eq_mem_phi->init_req(2, reset_memory()); 2330 } 2331 } 2332 ne_region->init_req(5, ctl); 2333 ne_io_phi->init_req(5, i_o()); 2334 ne_mem_phi->init_req(5, reset_memory()); 2335 2336 record_for_igvn(ne_region); 2337 set_control(_gvn.transform(ne_region)); 2338 set_i_o(_gvn.transform(ne_io_phi)); 2339 set_all_memory(_gvn.transform(ne_mem_phi)); 2340 2341 if (btest == BoolTest::ne) { 2342 { 2343 PreserveJVMState pjvms(this); 2344 int target_bci = iter().get_dest(); 2345 merge(target_bci); 2346 } 2347 2348 record_for_igvn(eq_region); 2349 set_control(_gvn.transform(eq_region)); 2350 set_i_o(_gvn.transform(eq_io_phi)); 2351 set_all_memory(_gvn.transform(eq_mem_phi)); 2352 } 2353 } 2354 2355 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information. 2356 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and 2357 // then either takes the trap or executes the original, unstable if. 2358 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) { 2359 // Search for an unstable if trap 2360 CallStaticJavaNode* trap = nullptr; 2361 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if"); 2362 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if); 2363 if (trap == nullptr || !trap->jvms()->should_reexecute()) { 2364 // No suitable trap found. Remove unused counter load and increment. 2365 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory)); 2366 return; 2367 } 2368 2369 // Remove trap from optimization list since we add another path to the trap. 2370 bool success = C->remove_unstable_if_trap(trap, true); 2371 assert(success, "Trap already modified"); 2372 2373 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise 2374 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31] 2375 Node* mask = intcon(right_n_bits(freq_log)); 2376 counter = _gvn.transform(new AndINode(counter, mask)); 2377 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0))); 2378 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq)); 2379 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If(); 2380 Node* if_true = _gvn.transform(new IfTrueNode(iff)); 2381 Node* if_false = _gvn.transform(new IfFalseNode(iff)); 2382 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken"); 2383 2384 // Trap 2385 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection"); 2386 2387 Node* trap_region = new RegionNode(3); 2388 trap_region->set_req(1, trap_proj); 2389 trap_region->set_req(2, if_true); 2390 trap->set_req(0, _gvn.transform(trap_region)); 2391 2392 // Don't trap, execute original if 2393 orig_iff->set_req(0, if_false); 2394 } 2395 2396 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const { 2397 // Randomly skip emitting an uncommon trap 2398 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) { 2399 return false; 2400 } 2401 // Don't want to speculate on uncommon traps when running with -Xcomp 2402 if (!UseInterpreter) { 2403 return false; 2404 } 2405 return seems_never_taken(prob) && 2406 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if); 2407 } 2408 2409 void Parse::maybe_add_predicate_after_if(Block* path) { 2410 if (path->is_SEL_head() && path->preds_parsed() == 0) { 2411 // Add predicates at bci of if dominating the loop so traps can be 2412 // recorded on the if's profile data 2413 int bc_depth = repush_if_args(); 2414 add_parse_predicates(); 2415 dec_sp(bc_depth); 2416 path->set_has_predicates(); 2417 } 2418 } 2419 2420 2421 //----------------------------adjust_map_after_if------------------------------ 2422 // Adjust the JVM state to reflect the result of taking this path. 2423 // Basically, it means inspecting the CmpNode controlling this 2424 // branch, seeing how it constrains a tested value, and then 2425 // deciding if it's worth our while to encode this constraint 2426 // as graph nodes in the current abstract interpretation map. 2427 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path, bool can_trap) { 2428 if (!c->is_Cmp()) { 2429 maybe_add_predicate_after_if(path); 2430 return; 2431 } 2432 2433 if (stopped() || btest == BoolTest::illegal) { 2434 return; // nothing to do 2435 } 2436 2437 bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 2438 2439 if (can_trap && path_is_suitable_for_uncommon_trap(prob)) { 2440 repush_if_args(); 2441 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if, 2442 Deoptimization::Action_reinterpret, 2443 nullptr, 2444 (is_fallthrough ? "taken always" : "taken never")); 2445 2446 if (call != nullptr) { 2447 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path)); 2448 } 2449 return; 2450 } 2451 2452 Node* val = c->in(1); 2453 Node* con = c->in(2); 2454 const Type* tcon = _gvn.type(con); 2455 const Type* tval = _gvn.type(val); 2456 bool have_con = tcon->singleton(); 2457 if (tval->singleton()) { 2458 if (!have_con) { 2459 // Swap, so constant is in con. 2460 con = val; 2461 tcon = tval; 2462 val = c->in(2); 2463 tval = _gvn.type(val); 2464 btest = BoolTest(btest).commute(); 2465 have_con = true; 2466 } else { 2467 // Do we have two constants? Then leave well enough alone. 2468 have_con = false; 2469 } 2470 } 2471 if (!have_con) { // remaining adjustments need a con 2472 maybe_add_predicate_after_if(path); 2473 return; 2474 } 2475 2476 sharpen_type_after_if(btest, con, tcon, val, tval); 2477 maybe_add_predicate_after_if(path); 2478 } 2479 2480 2481 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 2482 Node* ldk; 2483 if (n->is_DecodeNKlass()) { 2484 if (n->in(1)->Opcode() != Op_LoadNKlass) { 2485 return nullptr; 2486 } else { 2487 ldk = n->in(1); 2488 } 2489 } else if (n->Opcode() != Op_LoadKlass) { 2490 return nullptr; 2491 } else { 2492 ldk = n; 2493 } 2494 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 2495 2496 Node* adr = ldk->in(MemNode::Address); 2497 intptr_t off = 0; 2498 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 2499 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 2500 return nullptr; 2501 const TypePtr* tp = gvn->type(obj)->is_ptr(); 2502 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 2503 return nullptr; 2504 2505 return obj; 2506 } 2507 2508 void Parse::sharpen_type_after_if(BoolTest::mask btest, 2509 Node* con, const Type* tcon, 2510 Node* val, const Type* tval) { 2511 // Look for opportunities to sharpen the type of a node 2512 // whose klass is compared with a constant klass. 2513 if (btest == BoolTest::eq && tcon->isa_klassptr()) { 2514 Node* obj = extract_obj_from_klass_load(&_gvn, val); 2515 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 2516 if (obj != nullptr && (con_type->isa_instptr() || con_type->isa_aryptr())) { 2517 // Found: 2518 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 2519 // or the narrowOop equivalent. 2520 const Type* obj_type = _gvn.type(obj); 2521 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 2522 if (tboth != nullptr && tboth->klass_is_exact() && tboth != obj_type && 2523 tboth->higher_equal(obj_type)) { 2524 // obj has to be of the exact type Foo if the CmpP succeeds. 2525 int obj_in_map = map()->find_edge(obj); 2526 JVMState* jvms = this->jvms(); 2527 if (obj_in_map >= 0 && 2528 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 2529 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth); 2530 const Type* tcc = ccast->as_Type()->type(); 2531 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve"); 2532 // Delay transform() call to allow recovery of pre-cast value 2533 // at the control merge. 2534 _gvn.set_type_bottom(ccast); 2535 record_for_igvn(ccast); 2536 if (tboth->is_inlinetypeptr()) { 2537 ccast = InlineTypeNode::make_from_oop(this, ccast, tboth->exact_klass(true)->as_inline_klass()); 2538 } 2539 // Here's the payoff. 2540 replace_in_map(obj, ccast); 2541 } 2542 } 2543 } 2544 } 2545 2546 int val_in_map = map()->find_edge(val); 2547 if (val_in_map < 0) return; // replace_in_map would be useless 2548 { 2549 JVMState* jvms = this->jvms(); 2550 if (!(jvms->is_loc(val_in_map) || 2551 jvms->is_stk(val_in_map))) 2552 return; // again, it would be useless 2553 } 2554 2555 // Check for a comparison to a constant, and "know" that the compared 2556 // value is constrained on this path. 2557 assert(tcon->singleton(), ""); 2558 ConstraintCastNode* ccast = nullptr; 2559 Node* cast = nullptr; 2560 2561 switch (btest) { 2562 case BoolTest::eq: // Constant test? 2563 { 2564 const Type* tboth = tcon->join_speculative(tval); 2565 if (tboth == tval) break; // Nothing to gain. 2566 if (tcon->isa_int()) { 2567 ccast = new CastIINode(control(), val, tboth); 2568 } else if (tcon == TypePtr::NULL_PTR) { 2569 // Cast to null, but keep the pointer identity temporarily live. 2570 ccast = new CastPPNode(control(), val, tboth); 2571 } else { 2572 const TypeF* tf = tcon->isa_float_constant(); 2573 const TypeD* td = tcon->isa_double_constant(); 2574 // Exclude tests vs float/double 0 as these could be 2575 // either +0 or -0. Just because you are equal to +0 2576 // doesn't mean you ARE +0! 2577 // Note, following code also replaces Long and Oop values. 2578 if ((!tf || tf->_f != 0.0) && 2579 (!td || td->_d != 0.0)) 2580 cast = con; // Replace non-constant val by con. 2581 } 2582 } 2583 break; 2584 2585 case BoolTest::ne: 2586 if (tcon == TypePtr::NULL_PTR) { 2587 cast = cast_not_null(val, false); 2588 } 2589 break; 2590 2591 default: 2592 // (At this point we could record int range types with CastII.) 2593 break; 2594 } 2595 2596 if (ccast != nullptr) { 2597 const Type* tcc = ccast->as_Type()->type(); 2598 assert(tcc != tval && tcc->higher_equal(tval), "must improve"); 2599 // Delay transform() call to allow recovery of pre-cast value 2600 // at the control merge. 2601 _gvn.set_type_bottom(ccast); 2602 record_for_igvn(ccast); 2603 cast = ccast; 2604 } 2605 2606 if (cast != nullptr) { // Here's the payoff. 2607 replace_in_map(val, cast); 2608 } 2609 } 2610 2611 /** 2612 * Use speculative type to optimize CmpP node: if comparison is 2613 * against the low level class, cast the object to the speculative 2614 * type if any. CmpP should then go away. 2615 * 2616 * @param c expected CmpP node 2617 * @return result of CmpP on object casted to speculative type 2618 * 2619 */ 2620 Node* Parse::optimize_cmp_with_klass(Node* c) { 2621 // If this is transformed by the _gvn to a comparison with the low 2622 // level klass then we may be able to use speculation 2623 if (c->Opcode() == Op_CmpP && 2624 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 2625 c->in(2)->is_Con()) { 2626 Node* load_klass = nullptr; 2627 Node* decode = nullptr; 2628 if (c->in(1)->Opcode() == Op_DecodeNKlass) { 2629 decode = c->in(1); 2630 load_klass = c->in(1)->in(1); 2631 } else { 2632 load_klass = c->in(1); 2633 } 2634 if (load_klass->in(2)->is_AddP()) { 2635 Node* addp = load_klass->in(2); 2636 Node* obj = addp->in(AddPNode::Address); 2637 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 2638 if (obj_type->speculative_type_not_null() != nullptr) { 2639 ciKlass* k = obj_type->speculative_type(); 2640 inc_sp(2); 2641 obj = maybe_cast_profiled_obj(obj, k); 2642 dec_sp(2); 2643 if (obj->is_InlineType()) { 2644 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated"); 2645 obj = obj->as_InlineType()->get_oop(); 2646 } 2647 // Make the CmpP use the casted obj 2648 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 2649 load_klass = load_klass->clone(); 2650 load_klass->set_req(2, addp); 2651 load_klass = _gvn.transform(load_klass); 2652 if (decode != nullptr) { 2653 decode = decode->clone(); 2654 decode->set_req(1, load_klass); 2655 load_klass = _gvn.transform(decode); 2656 } 2657 c = c->clone(); 2658 c->set_req(1, load_klass); 2659 c = _gvn.transform(c); 2660 } 2661 } 2662 } 2663 return c; 2664 } 2665 2666 //------------------------------do_one_bytecode-------------------------------- 2667 // Parse this bytecode, and alter the Parsers JVM->Node mapping 2668 void Parse::do_one_bytecode() { 2669 Node *a, *b, *c, *d; // Handy temps 2670 BoolTest::mask btest; 2671 int i; 2672 2673 assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 2674 2675 if (C->check_node_count(NodeLimitFudgeFactor * 5, 2676 "out of nodes parsing method")) { 2677 return; 2678 } 2679 2680 #ifdef ASSERT 2681 // for setting breakpoints 2682 if (TraceOptoParse) { 2683 tty->print(" @"); 2684 dump_bci(bci()); 2685 tty->print(" %s", Bytecodes::name(bc())); 2686 tty->cr(); 2687 } 2688 #endif 2689 2690 switch (bc()) { 2691 case Bytecodes::_nop: 2692 // do nothing 2693 break; 2694 case Bytecodes::_lconst_0: 2695 push_pair(longcon(0)); 2696 break; 2697 2698 case Bytecodes::_lconst_1: 2699 push_pair(longcon(1)); 2700 break; 2701 2702 case Bytecodes::_fconst_0: 2703 push(zerocon(T_FLOAT)); 2704 break; 2705 2706 case Bytecodes::_fconst_1: 2707 push(makecon(TypeF::ONE)); 2708 break; 2709 2710 case Bytecodes::_fconst_2: 2711 push(makecon(TypeF::make(2.0f))); 2712 break; 2713 2714 case Bytecodes::_dconst_0: 2715 push_pair(zerocon(T_DOUBLE)); 2716 break; 2717 2718 case Bytecodes::_dconst_1: 2719 push_pair(makecon(TypeD::ONE)); 2720 break; 2721 2722 case Bytecodes::_iconst_m1:push(intcon(-1)); break; 2723 case Bytecodes::_iconst_0: push(intcon( 0)); break; 2724 case Bytecodes::_iconst_1: push(intcon( 1)); break; 2725 case Bytecodes::_iconst_2: push(intcon( 2)); break; 2726 case Bytecodes::_iconst_3: push(intcon( 3)); break; 2727 case Bytecodes::_iconst_4: push(intcon( 4)); break; 2728 case Bytecodes::_iconst_5: push(intcon( 5)); break; 2729 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 2730 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 2731 case Bytecodes::_aconst_null: push(null()); break; 2732 2733 case Bytecodes::_ldc: 2734 case Bytecodes::_ldc_w: 2735 case Bytecodes::_ldc2_w: { 2736 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded 2737 assert(!iter().is_in_error(), "ldc is in error state"); 2738 ciConstant constant = iter().get_constant(); 2739 assert(constant.is_loaded(), "constant is not loaded"); 2740 const Type* con_type = Type::make_from_constant(constant); 2741 if (con_type != nullptr) { 2742 push_node(con_type->basic_type(), makecon(con_type)); 2743 } 2744 break; 2745 } 2746 2747 case Bytecodes::_aload_0: 2748 push( local(0) ); 2749 break; 2750 case Bytecodes::_aload_1: 2751 push( local(1) ); 2752 break; 2753 case Bytecodes::_aload_2: 2754 push( local(2) ); 2755 break; 2756 case Bytecodes::_aload_3: 2757 push( local(3) ); 2758 break; 2759 case Bytecodes::_aload: 2760 push( local(iter().get_index()) ); 2761 break; 2762 2763 case Bytecodes::_fload_0: 2764 case Bytecodes::_iload_0: 2765 push( local(0) ); 2766 break; 2767 case Bytecodes::_fload_1: 2768 case Bytecodes::_iload_1: 2769 push( local(1) ); 2770 break; 2771 case Bytecodes::_fload_2: 2772 case Bytecodes::_iload_2: 2773 push( local(2) ); 2774 break; 2775 case Bytecodes::_fload_3: 2776 case Bytecodes::_iload_3: 2777 push( local(3) ); 2778 break; 2779 case Bytecodes::_fload: 2780 case Bytecodes::_iload: 2781 push( local(iter().get_index()) ); 2782 break; 2783 case Bytecodes::_lload_0: 2784 push_pair_local( 0 ); 2785 break; 2786 case Bytecodes::_lload_1: 2787 push_pair_local( 1 ); 2788 break; 2789 case Bytecodes::_lload_2: 2790 push_pair_local( 2 ); 2791 break; 2792 case Bytecodes::_lload_3: 2793 push_pair_local( 3 ); 2794 break; 2795 case Bytecodes::_lload: 2796 push_pair_local( iter().get_index() ); 2797 break; 2798 2799 case Bytecodes::_dload_0: 2800 push_pair_local(0); 2801 break; 2802 case Bytecodes::_dload_1: 2803 push_pair_local(1); 2804 break; 2805 case Bytecodes::_dload_2: 2806 push_pair_local(2); 2807 break; 2808 case Bytecodes::_dload_3: 2809 push_pair_local(3); 2810 break; 2811 case Bytecodes::_dload: 2812 push_pair_local(iter().get_index()); 2813 break; 2814 case Bytecodes::_fstore_0: 2815 case Bytecodes::_istore_0: 2816 case Bytecodes::_astore_0: 2817 set_local( 0, pop() ); 2818 break; 2819 case Bytecodes::_fstore_1: 2820 case Bytecodes::_istore_1: 2821 case Bytecodes::_astore_1: 2822 set_local( 1, pop() ); 2823 break; 2824 case Bytecodes::_fstore_2: 2825 case Bytecodes::_istore_2: 2826 case Bytecodes::_astore_2: 2827 set_local( 2, pop() ); 2828 break; 2829 case Bytecodes::_fstore_3: 2830 case Bytecodes::_istore_3: 2831 case Bytecodes::_astore_3: 2832 set_local( 3, pop() ); 2833 break; 2834 case Bytecodes::_fstore: 2835 case Bytecodes::_istore: 2836 case Bytecodes::_astore: 2837 set_local( iter().get_index(), pop() ); 2838 break; 2839 // long stores 2840 case Bytecodes::_lstore_0: 2841 set_pair_local( 0, pop_pair() ); 2842 break; 2843 case Bytecodes::_lstore_1: 2844 set_pair_local( 1, pop_pair() ); 2845 break; 2846 case Bytecodes::_lstore_2: 2847 set_pair_local( 2, pop_pair() ); 2848 break; 2849 case Bytecodes::_lstore_3: 2850 set_pair_local( 3, pop_pair() ); 2851 break; 2852 case Bytecodes::_lstore: 2853 set_pair_local( iter().get_index(), pop_pair() ); 2854 break; 2855 2856 // double stores 2857 case Bytecodes::_dstore_0: 2858 set_pair_local( 0, pop_pair() ); 2859 break; 2860 case Bytecodes::_dstore_1: 2861 set_pair_local( 1, pop_pair() ); 2862 break; 2863 case Bytecodes::_dstore_2: 2864 set_pair_local( 2, pop_pair() ); 2865 break; 2866 case Bytecodes::_dstore_3: 2867 set_pair_local( 3, pop_pair() ); 2868 break; 2869 case Bytecodes::_dstore: 2870 set_pair_local( iter().get_index(), pop_pair() ); 2871 break; 2872 2873 case Bytecodes::_pop: dec_sp(1); break; 2874 case Bytecodes::_pop2: dec_sp(2); break; 2875 case Bytecodes::_swap: 2876 a = pop(); 2877 b = pop(); 2878 push(a); 2879 push(b); 2880 break; 2881 case Bytecodes::_dup: 2882 a = pop(); 2883 push(a); 2884 push(a); 2885 break; 2886 case Bytecodes::_dup_x1: 2887 a = pop(); 2888 b = pop(); 2889 push( a ); 2890 push( b ); 2891 push( a ); 2892 break; 2893 case Bytecodes::_dup_x2: 2894 a = pop(); 2895 b = pop(); 2896 c = pop(); 2897 push( a ); 2898 push( c ); 2899 push( b ); 2900 push( a ); 2901 break; 2902 case Bytecodes::_dup2: 2903 a = pop(); 2904 b = pop(); 2905 push( b ); 2906 push( a ); 2907 push( b ); 2908 push( a ); 2909 break; 2910 2911 case Bytecodes::_dup2_x1: 2912 // before: .. c, b, a 2913 // after: .. b, a, c, b, a 2914 // not tested 2915 a = pop(); 2916 b = pop(); 2917 c = pop(); 2918 push( b ); 2919 push( a ); 2920 push( c ); 2921 push( b ); 2922 push( a ); 2923 break; 2924 case Bytecodes::_dup2_x2: 2925 // before: .. d, c, b, a 2926 // after: .. b, a, d, c, b, a 2927 // not tested 2928 a = pop(); 2929 b = pop(); 2930 c = pop(); 2931 d = pop(); 2932 push( b ); 2933 push( a ); 2934 push( d ); 2935 push( c ); 2936 push( b ); 2937 push( a ); 2938 break; 2939 2940 case Bytecodes::_arraylength: { 2941 // Must do null-check with value on expression stack 2942 Node *ary = null_check(peek(), T_ARRAY); 2943 // Compile-time detect of null-exception? 2944 if (stopped()) return; 2945 a = pop(); 2946 push(load_array_length(a)); 2947 break; 2948 } 2949 2950 case Bytecodes::_baload: array_load(T_BYTE); break; 2951 case Bytecodes::_caload: array_load(T_CHAR); break; 2952 case Bytecodes::_iaload: array_load(T_INT); break; 2953 case Bytecodes::_saload: array_load(T_SHORT); break; 2954 case Bytecodes::_faload: array_load(T_FLOAT); break; 2955 case Bytecodes::_aaload: array_load(T_OBJECT); break; 2956 case Bytecodes::_laload: array_load(T_LONG); break; 2957 case Bytecodes::_daload: array_load(T_DOUBLE); break; 2958 case Bytecodes::_bastore: array_store(T_BYTE); break; 2959 case Bytecodes::_castore: array_store(T_CHAR); break; 2960 case Bytecodes::_iastore: array_store(T_INT); break; 2961 case Bytecodes::_sastore: array_store(T_SHORT); break; 2962 case Bytecodes::_fastore: array_store(T_FLOAT); break; 2963 case Bytecodes::_aastore: array_store(T_OBJECT); break; 2964 case Bytecodes::_lastore: array_store(T_LONG); break; 2965 case Bytecodes::_dastore: array_store(T_DOUBLE); break; 2966 2967 case Bytecodes::_getfield: 2968 do_getfield(); 2969 break; 2970 2971 case Bytecodes::_getstatic: 2972 do_getstatic(); 2973 break; 2974 2975 case Bytecodes::_putfield: 2976 do_putfield(); 2977 break; 2978 2979 case Bytecodes::_putstatic: 2980 do_putstatic(); 2981 break; 2982 2983 case Bytecodes::_irem: 2984 // Must keep both values on the expression-stack during null-check 2985 zero_check_int(peek()); 2986 // Compile-time detect of null-exception? 2987 if (stopped()) return; 2988 b = pop(); 2989 a = pop(); 2990 push(_gvn.transform(new ModINode(control(), a, b))); 2991 break; 2992 case Bytecodes::_idiv: 2993 // Must keep both values on the expression-stack during null-check 2994 zero_check_int(peek()); 2995 // Compile-time detect of null-exception? 2996 if (stopped()) return; 2997 b = pop(); 2998 a = pop(); 2999 push( _gvn.transform( new DivINode(control(),a,b) ) ); 3000 break; 3001 case Bytecodes::_imul: 3002 b = pop(); a = pop(); 3003 push( _gvn.transform( new MulINode(a,b) ) ); 3004 break; 3005 case Bytecodes::_iadd: 3006 b = pop(); a = pop(); 3007 push( _gvn.transform( new AddINode(a,b) ) ); 3008 break; 3009 case Bytecodes::_ineg: 3010 a = pop(); 3011 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) ); 3012 break; 3013 case Bytecodes::_isub: 3014 b = pop(); a = pop(); 3015 push( _gvn.transform( new SubINode(a,b) ) ); 3016 break; 3017 case Bytecodes::_iand: 3018 b = pop(); a = pop(); 3019 push( _gvn.transform( new AndINode(a,b) ) ); 3020 break; 3021 case Bytecodes::_ior: 3022 b = pop(); a = pop(); 3023 push( _gvn.transform( new OrINode(a,b) ) ); 3024 break; 3025 case Bytecodes::_ixor: 3026 b = pop(); a = pop(); 3027 push( _gvn.transform( new XorINode(a,b) ) ); 3028 break; 3029 case Bytecodes::_ishl: 3030 b = pop(); a = pop(); 3031 push( _gvn.transform( new LShiftINode(a,b) ) ); 3032 break; 3033 case Bytecodes::_ishr: 3034 b = pop(); a = pop(); 3035 push( _gvn.transform( new RShiftINode(a,b) ) ); 3036 break; 3037 case Bytecodes::_iushr: 3038 b = pop(); a = pop(); 3039 push( _gvn.transform( new URShiftINode(a,b) ) ); 3040 break; 3041 3042 case Bytecodes::_fneg: 3043 a = pop(); 3044 b = _gvn.transform(new NegFNode (a)); 3045 push(b); 3046 break; 3047 3048 case Bytecodes::_fsub: 3049 b = pop(); 3050 a = pop(); 3051 c = _gvn.transform( new SubFNode(a,b) ); 3052 push(c); 3053 break; 3054 3055 case Bytecodes::_fadd: 3056 b = pop(); 3057 a = pop(); 3058 c = _gvn.transform( new AddFNode(a,b) ); 3059 push(c); 3060 break; 3061 3062 case Bytecodes::_fmul: 3063 b = pop(); 3064 a = pop(); 3065 c = _gvn.transform( new MulFNode(a,b) ); 3066 push(c); 3067 break; 3068 3069 case Bytecodes::_fdiv: 3070 b = pop(); 3071 a = pop(); 3072 c = _gvn.transform( new DivFNode(nullptr,a,b) ); 3073 push(c); 3074 break; 3075 3076 case Bytecodes::_frem: 3077 // Generate a ModF node. 3078 b = pop(); 3079 a = pop(); 3080 push(floating_point_mod(a, b, BasicType::T_FLOAT)); 3081 break; 3082 3083 case Bytecodes::_fcmpl: 3084 b = pop(); 3085 a = pop(); 3086 c = _gvn.transform( new CmpF3Node( a, b)); 3087 push(c); 3088 break; 3089 case Bytecodes::_fcmpg: 3090 b = pop(); 3091 a = pop(); 3092 3093 // Same as fcmpl but need to flip the unordered case. Swap the inputs, 3094 // which negates the result sign except for unordered. Flip the unordered 3095 // as well by using CmpF3 which implements unordered-lesser instead of 3096 // unordered-greater semantics. Finally, commute the result bits. Result 3097 // is same as using a CmpF3Greater except we did it with CmpF3 alone. 3098 c = _gvn.transform( new CmpF3Node( b, a)); 3099 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 3100 push(c); 3101 break; 3102 3103 case Bytecodes::_f2i: 3104 a = pop(); 3105 push(_gvn.transform(new ConvF2INode(a))); 3106 break; 3107 3108 case Bytecodes::_d2i: 3109 a = pop_pair(); 3110 b = _gvn.transform(new ConvD2INode(a)); 3111 push( b ); 3112 break; 3113 3114 case Bytecodes::_f2d: 3115 a = pop(); 3116 b = _gvn.transform( new ConvF2DNode(a)); 3117 push_pair( b ); 3118 break; 3119 3120 case Bytecodes::_d2f: 3121 a = pop_pair(); 3122 b = _gvn.transform( new ConvD2FNode(a)); 3123 push( b ); 3124 break; 3125 3126 case Bytecodes::_l2f: 3127 if (Matcher::convL2FSupported()) { 3128 a = pop_pair(); 3129 b = _gvn.transform( new ConvL2FNode(a)); 3130 push(b); 3131 } else { 3132 l2f(); 3133 } 3134 break; 3135 3136 case Bytecodes::_l2d: 3137 a = pop_pair(); 3138 b = _gvn.transform( new ConvL2DNode(a)); 3139 push_pair(b); 3140 break; 3141 3142 case Bytecodes::_f2l: 3143 a = pop(); 3144 b = _gvn.transform( new ConvF2LNode(a)); 3145 push_pair(b); 3146 break; 3147 3148 case Bytecodes::_d2l: 3149 a = pop_pair(); 3150 b = _gvn.transform( new ConvD2LNode(a)); 3151 push_pair(b); 3152 break; 3153 3154 case Bytecodes::_dsub: 3155 b = pop_pair(); 3156 a = pop_pair(); 3157 c = _gvn.transform( new SubDNode(a,b) ); 3158 push_pair(c); 3159 break; 3160 3161 case Bytecodes::_dadd: 3162 b = pop_pair(); 3163 a = pop_pair(); 3164 c = _gvn.transform( new AddDNode(a,b) ); 3165 push_pair(c); 3166 break; 3167 3168 case Bytecodes::_dmul: 3169 b = pop_pair(); 3170 a = pop_pair(); 3171 c = _gvn.transform( new MulDNode(a,b) ); 3172 push_pair(c); 3173 break; 3174 3175 case Bytecodes::_ddiv: 3176 b = pop_pair(); 3177 a = pop_pair(); 3178 c = _gvn.transform( new DivDNode(nullptr,a,b) ); 3179 push_pair(c); 3180 break; 3181 3182 case Bytecodes::_dneg: 3183 a = pop_pair(); 3184 b = _gvn.transform(new NegDNode (a)); 3185 push_pair(b); 3186 break; 3187 3188 case Bytecodes::_drem: 3189 // Generate a ModD node. 3190 b = pop_pair(); 3191 a = pop_pair(); 3192 push_pair(floating_point_mod(a, b, BasicType::T_DOUBLE)); 3193 break; 3194 3195 case Bytecodes::_dcmpl: 3196 b = pop_pair(); 3197 a = pop_pair(); 3198 c = _gvn.transform( new CmpD3Node( a, b)); 3199 push(c); 3200 break; 3201 3202 case Bytecodes::_dcmpg: 3203 b = pop_pair(); 3204 a = pop_pair(); 3205 // Same as dcmpl but need to flip the unordered case. 3206 // Commute the inputs, which negates the result sign except for unordered. 3207 // Flip the unordered as well by using CmpD3 which implements 3208 // unordered-lesser instead of unordered-greater semantics. 3209 // Finally, negate the result bits. Result is same as using a 3210 // CmpD3Greater except we did it with CmpD3 alone. 3211 c = _gvn.transform( new CmpD3Node( b, a)); 3212 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) ); 3213 push(c); 3214 break; 3215 3216 3217 // Note for longs -> lo word is on TOS, hi word is on TOS - 1 3218 case Bytecodes::_land: 3219 b = pop_pair(); 3220 a = pop_pair(); 3221 c = _gvn.transform( new AndLNode(a,b) ); 3222 push_pair(c); 3223 break; 3224 case Bytecodes::_lor: 3225 b = pop_pair(); 3226 a = pop_pair(); 3227 c = _gvn.transform( new OrLNode(a,b) ); 3228 push_pair(c); 3229 break; 3230 case Bytecodes::_lxor: 3231 b = pop_pair(); 3232 a = pop_pair(); 3233 c = _gvn.transform( new XorLNode(a,b) ); 3234 push_pair(c); 3235 break; 3236 3237 case Bytecodes::_lshl: 3238 b = pop(); // the shift count 3239 a = pop_pair(); // value to be shifted 3240 c = _gvn.transform( new LShiftLNode(a,b) ); 3241 push_pair(c); 3242 break; 3243 case Bytecodes::_lshr: 3244 b = pop(); // the shift count 3245 a = pop_pair(); // value to be shifted 3246 c = _gvn.transform( new RShiftLNode(a,b) ); 3247 push_pair(c); 3248 break; 3249 case Bytecodes::_lushr: 3250 b = pop(); // the shift count 3251 a = pop_pair(); // value to be shifted 3252 c = _gvn.transform( new URShiftLNode(a,b) ); 3253 push_pair(c); 3254 break; 3255 case Bytecodes::_lmul: 3256 b = pop_pair(); 3257 a = pop_pair(); 3258 c = _gvn.transform( new MulLNode(a,b) ); 3259 push_pair(c); 3260 break; 3261 3262 case Bytecodes::_lrem: 3263 // Must keep both values on the expression-stack during null-check 3264 assert(peek(0) == top(), "long word order"); 3265 zero_check_long(peek(1)); 3266 // Compile-time detect of null-exception? 3267 if (stopped()) return; 3268 b = pop_pair(); 3269 a = pop_pair(); 3270 c = _gvn.transform( new ModLNode(control(),a,b) ); 3271 push_pair(c); 3272 break; 3273 3274 case Bytecodes::_ldiv: 3275 // Must keep both values on the expression-stack during null-check 3276 assert(peek(0) == top(), "long word order"); 3277 zero_check_long(peek(1)); 3278 // Compile-time detect of null-exception? 3279 if (stopped()) return; 3280 b = pop_pair(); 3281 a = pop_pair(); 3282 c = _gvn.transform( new DivLNode(control(),a,b) ); 3283 push_pair(c); 3284 break; 3285 3286 case Bytecodes::_ladd: 3287 b = pop_pair(); 3288 a = pop_pair(); 3289 c = _gvn.transform( new AddLNode(a,b) ); 3290 push_pair(c); 3291 break; 3292 case Bytecodes::_lsub: 3293 b = pop_pair(); 3294 a = pop_pair(); 3295 c = _gvn.transform( new SubLNode(a,b) ); 3296 push_pair(c); 3297 break; 3298 case Bytecodes::_lcmp: 3299 // Safepoints are now inserted _before_ branches. The long-compare 3300 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 3301 // slew of control flow. These are usually followed by a CmpI vs zero and 3302 // a branch; this pattern then optimizes to the obvious long-compare and 3303 // branch. However, if the branch is backwards there's a Safepoint 3304 // inserted. The inserted Safepoint captures the JVM state at the 3305 // pre-branch point, i.e. it captures the 3-way value. Thus if a 3306 // long-compare is used to control a loop the debug info will force 3307 // computation of the 3-way value, even though the generated code uses a 3308 // long-compare and branch. We try to rectify the situation by inserting 3309 // a SafePoint here and have it dominate and kill the safepoint added at a 3310 // following backwards branch. At this point the JVM state merely holds 2 3311 // longs but not the 3-way value. 3312 switch (iter().next_bc()) { 3313 case Bytecodes::_ifgt: 3314 case Bytecodes::_iflt: 3315 case Bytecodes::_ifge: 3316 case Bytecodes::_ifle: 3317 case Bytecodes::_ifne: 3318 case Bytecodes::_ifeq: 3319 // If this is a backwards branch in the bytecodes, add Safepoint 3320 maybe_add_safepoint(iter().next_get_dest()); 3321 default: 3322 break; 3323 } 3324 b = pop_pair(); 3325 a = pop_pair(); 3326 c = _gvn.transform( new CmpL3Node( a, b )); 3327 push(c); 3328 break; 3329 3330 case Bytecodes::_lneg: 3331 a = pop_pair(); 3332 b = _gvn.transform( new SubLNode(longcon(0),a)); 3333 push_pair(b); 3334 break; 3335 case Bytecodes::_l2i: 3336 a = pop_pair(); 3337 push( _gvn.transform( new ConvL2INode(a))); 3338 break; 3339 case Bytecodes::_i2l: 3340 a = pop(); 3341 b = _gvn.transform( new ConvI2LNode(a)); 3342 push_pair(b); 3343 break; 3344 case Bytecodes::_i2b: 3345 // Sign extend 3346 a = pop(); 3347 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true); 3348 push(a); 3349 break; 3350 case Bytecodes::_i2s: 3351 a = pop(); 3352 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true); 3353 push(a); 3354 break; 3355 case Bytecodes::_i2c: 3356 a = pop(); 3357 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true); 3358 push(a); 3359 break; 3360 3361 case Bytecodes::_i2f: 3362 a = pop(); 3363 b = _gvn.transform( new ConvI2FNode(a) ) ; 3364 push(b); 3365 break; 3366 3367 case Bytecodes::_i2d: 3368 a = pop(); 3369 b = _gvn.transform( new ConvI2DNode(a)); 3370 push_pair(b); 3371 break; 3372 3373 case Bytecodes::_iinc: // Increment local 3374 i = iter().get_index(); // Get local index 3375 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 3376 break; 3377 3378 // Exit points of synchronized methods must have an unlock node 3379 case Bytecodes::_return: 3380 return_current(nullptr); 3381 break; 3382 3383 case Bytecodes::_ireturn: 3384 case Bytecodes::_areturn: 3385 case Bytecodes::_freturn: 3386 return_current(cast_to_non_larval(pop())); 3387 break; 3388 case Bytecodes::_lreturn: 3389 case Bytecodes::_dreturn: 3390 return_current(pop_pair()); 3391 break; 3392 3393 case Bytecodes::_athrow: 3394 // null exception oop throws null pointer exception 3395 null_check(peek()); 3396 if (stopped()) return; 3397 // Hook the thrown exception directly to subsequent handlers. 3398 if (BailoutToInterpreterForThrows) { 3399 // Keep method interpreted from now on. 3400 uncommon_trap(Deoptimization::Reason_unhandled, 3401 Deoptimization::Action_make_not_compilable); 3402 return; 3403 } 3404 if (env()->jvmti_can_post_on_exceptions()) { 3405 // check if we must post exception events, take uncommon trap if so (with must_throw = false) 3406 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 3407 } 3408 // Here if either can_post_on_exceptions or should_post_on_exceptions is false 3409 add_exception_state(make_exception_state(peek())); 3410 break; 3411 3412 case Bytecodes::_goto: // fall through 3413 case Bytecodes::_goto_w: { 3414 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 3415 3416 // If this is a backwards branch in the bytecodes, add Safepoint 3417 maybe_add_safepoint(target_bci); 3418 3419 // Merge the current control into the target basic block 3420 merge(target_bci); 3421 3422 // See if we can get some profile data and hand it off to the next block 3423 Block *target_block = block()->successor_for_bci(target_bci); 3424 if (target_block->pred_count() != 1) break; 3425 ciMethodData* methodData = method()->method_data(); 3426 if (!methodData->is_mature()) break; 3427 ciProfileData* data = methodData->bci_to_data(bci()); 3428 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch"); 3429 int taken = ((ciJumpData*)data)->taken(); 3430 taken = method()->scale_count(taken); 3431 target_block->set_count(taken); 3432 break; 3433 } 3434 3435 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 3436 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 3437 handle_if_null: 3438 // If this is a backwards branch in the bytecodes, add Safepoint 3439 maybe_add_safepoint(iter().get_dest()); 3440 a = null(); 3441 b = cast_to_non_larval(pop()); 3442 if (b->is_InlineType()) { 3443 // Null checking a scalarized but nullable inline type. Check the IsInit 3444 // input instead of the oop input to avoid keeping buffer allocations alive 3445 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_is_init(), zerocon(T_INT))); 3446 } else { 3447 if (!_gvn.type(b)->speculative_maybe_null() && 3448 !too_many_traps(Deoptimization::Reason_speculate_null_check)) { 3449 inc_sp(1); 3450 Node* null_ctl = top(); 3451 b = null_check_oop(b, &null_ctl, true, true, true); 3452 assert(null_ctl->is_top(), "no null control here"); 3453 dec_sp(1); 3454 } else if (_gvn.type(b)->speculative_always_null() && 3455 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) { 3456 inc_sp(1); 3457 b = null_assert(b); 3458 dec_sp(1); 3459 } 3460 c = _gvn.transform( new CmpPNode(b, a) ); 3461 } 3462 do_ifnull(btest, c); 3463 break; 3464 3465 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 3466 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 3467 handle_if_acmp: 3468 // If this is a backwards branch in the bytecodes, add Safepoint 3469 maybe_add_safepoint(iter().get_dest()); 3470 a = cast_to_non_larval(pop()); 3471 b = cast_to_non_larval(pop()); 3472 do_acmp(btest, b, a); 3473 break; 3474 3475 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 3476 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 3477 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 3478 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 3479 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 3480 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 3481 handle_ifxx: 3482 // If this is a backwards branch in the bytecodes, add Safepoint 3483 maybe_add_safepoint(iter().get_dest()); 3484 a = _gvn.intcon(0); 3485 b = pop(); 3486 c = _gvn.transform( new CmpINode(b, a) ); 3487 do_if(btest, c); 3488 break; 3489 3490 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 3491 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 3492 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 3493 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 3494 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 3495 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 3496 handle_if_icmp: 3497 // If this is a backwards branch in the bytecodes, add Safepoint 3498 maybe_add_safepoint(iter().get_dest()); 3499 a = pop(); 3500 b = pop(); 3501 c = _gvn.transform( new CmpINode( b, a ) ); 3502 do_if(btest, c); 3503 break; 3504 3505 case Bytecodes::_tableswitch: 3506 do_tableswitch(); 3507 break; 3508 3509 case Bytecodes::_lookupswitch: 3510 do_lookupswitch(); 3511 break; 3512 3513 case Bytecodes::_invokestatic: 3514 case Bytecodes::_invokedynamic: 3515 case Bytecodes::_invokespecial: 3516 case Bytecodes::_invokevirtual: 3517 case Bytecodes::_invokeinterface: 3518 do_call(); 3519 break; 3520 case Bytecodes::_checkcast: 3521 do_checkcast(); 3522 break; 3523 case Bytecodes::_instanceof: 3524 do_instanceof(); 3525 break; 3526 case Bytecodes::_anewarray: 3527 do_newarray(); 3528 break; 3529 case Bytecodes::_newarray: 3530 do_newarray((BasicType)iter().get_index()); 3531 break; 3532 case Bytecodes::_multianewarray: 3533 do_multianewarray(); 3534 break; 3535 case Bytecodes::_new: 3536 do_new(); 3537 break; 3538 3539 case Bytecodes::_jsr: 3540 case Bytecodes::_jsr_w: 3541 do_jsr(); 3542 break; 3543 3544 case Bytecodes::_ret: 3545 do_ret(); 3546 break; 3547 3548 3549 case Bytecodes::_monitorenter: 3550 do_monitor_enter(); 3551 break; 3552 3553 case Bytecodes::_monitorexit: 3554 do_monitor_exit(); 3555 break; 3556 3557 case Bytecodes::_breakpoint: 3558 // Breakpoint set concurrently to compile 3559 // %%% use an uncommon trap? 3560 C->record_failure("breakpoint in method"); 3561 return; 3562 3563 default: 3564 #ifndef PRODUCT 3565 map()->dump(99); 3566 #endif 3567 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 3568 ShouldNotReachHere(); 3569 } 3570 3571 #ifndef PRODUCT 3572 if (failing()) { return; } 3573 constexpr int perBytecode = 6; 3574 if (C->should_print_igv(perBytecode)) { 3575 IdealGraphPrinter* printer = C->igv_printer(); 3576 char buffer[256]; 3577 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s, map: %d", bci(), Bytecodes::name(bc()), map() == nullptr ? -1 : map()->_idx); 3578 bool old = printer->traverse_outs(); 3579 printer->set_traverse_outs(true); 3580 printer->print_graph(buffer); 3581 printer->set_traverse_outs(old); 3582 } 3583 #endif 3584 }