1 /* 2 * Copyright (c) 1999, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/systemDictionary.hpp" 27 #include "classfile/vmSymbols.hpp" 28 #include "compiler/compileBroker.hpp" 29 #include "compiler/compileLog.hpp" 30 #include "oops/objArrayKlass.hpp" 31 #include "opto/addnode.hpp" 32 #include "opto/callGenerator.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/connode.hpp" 35 #include "opto/idealKit.hpp" 36 #include "opto/mathexactnode.hpp" 37 #include "opto/mulnode.hpp" 38 #include "opto/parse.hpp" 39 #include "opto/runtime.hpp" 40 #include "opto/subnode.hpp" 41 #include "prims/nativeLookup.hpp" 42 #include "runtime/sharedRuntime.hpp" 43 #include "trace/traceMacros.hpp" 44 #include "utilities/macros.hpp" 45 #if INCLUDE_ALL_GCS 46 #include "gc_implementation/shenandoah/shenandoahBarrierSetC2.hpp" 47 #include "gc_implementation/shenandoah/shenandoahSupport.hpp" 48 #endif 49 50 51 class LibraryIntrinsic : public InlineCallGenerator { 52 // Extend the set of intrinsics known to the runtime: 53 public: 54 private: 55 bool _is_virtual; 56 bool _does_virtual_dispatch; 57 int8_t _predicates_count; // Intrinsic is predicated by several conditions 58 int8_t _last_predicate; // Last generated predicate 59 vmIntrinsics::ID _intrinsic_id; 60 61 public: 62 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 63 : InlineCallGenerator(m), 64 _is_virtual(is_virtual), 65 _does_virtual_dispatch(does_virtual_dispatch), 66 _predicates_count((int8_t)predicates_count), 67 _last_predicate((int8_t)-1), 68 _intrinsic_id(id) 69 { 70 } 71 virtual bool is_intrinsic() const { return true; } 72 virtual bool is_virtual() const { return _is_virtual; } 73 virtual bool is_predicated() const { return _predicates_count > 0; } 74 virtual int predicates_count() const { return _predicates_count; } 75 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 76 virtual JVMState* generate(JVMState* jvms); 77 virtual Node* generate_predicate(JVMState* jvms, int predicate); 78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 79 }; 80 81 82 // Local helper class for LibraryIntrinsic: 83 class LibraryCallKit : public GraphKit { 84 private: 85 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 86 Node* _result; // the result node, if any 87 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 88 89 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 90 91 public: 92 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 93 : GraphKit(jvms), 94 _intrinsic(intrinsic), 95 _result(NULL) 96 { 97 // Check if this is a root compile. In that case we don't have a caller. 98 if (!jvms->has_method()) { 99 _reexecute_sp = sp(); 100 } else { 101 // Find out how many arguments the interpreter needs when deoptimizing 102 // and save the stack pointer value so it can used by uncommon_trap. 103 // We find the argument count by looking at the declared signature. 104 bool ignored_will_link; 105 ciSignature* declared_signature = NULL; 106 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 107 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 108 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 109 } 110 } 111 112 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 113 114 ciMethod* caller() const { return jvms()->method(); } 115 int bci() const { return jvms()->bci(); } 116 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 117 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 118 ciMethod* callee() const { return _intrinsic->method(); } 119 120 bool try_to_inline(int predicate); 121 Node* try_to_predicate(int predicate); 122 123 void push_result() { 124 // Push the result onto the stack. 125 if (!stopped() && result() != NULL) { 126 BasicType bt = result()->bottom_type()->basic_type(); 127 push_node(bt, result()); 128 } 129 } 130 131 private: 132 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 133 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 134 } 135 136 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 137 void set_result(RegionNode* region, PhiNode* value); 138 Node* result() { return _result; } 139 140 virtual int reexecute_sp() { return _reexecute_sp; } 141 142 // Helper functions to inline natives 143 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 144 Node* generate_slow_guard(Node* test, RegionNode* region); 145 Node* generate_fair_guard(Node* test, RegionNode* region); 146 Node* generate_negative_guard(Node* index, RegionNode* region, 147 // resulting CastII of index: 148 Node* *pos_index = NULL); 149 Node* generate_nonpositive_guard(Node* index, bool never_negative, 150 // resulting CastII of index: 151 Node* *pos_index = NULL); 152 Node* generate_limit_guard(Node* offset, Node* subseq_length, 153 Node* array_length, 154 RegionNode* region); 155 Node* generate_current_thread(Node* &tls_output); 156 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset, 157 bool disjoint_bases, const char* &name, bool dest_uninitialized); 158 Node* load_mirror_from_klass(Node* klass); 159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path, 161 int offset); 162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 163 RegionNode* region, int null_path) { 164 int offset = java_lang_Class::klass_offset_in_bytes(); 165 return load_klass_from_mirror_common(mirror, never_see_null, 166 region, null_path, 167 offset); 168 } 169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 170 RegionNode* region, int null_path) { 171 int offset = java_lang_Class::array_klass_offset_in_bytes(); 172 return load_klass_from_mirror_common(mirror, never_see_null, 173 region, null_path, 174 offset); 175 } 176 Node* generate_access_flags_guard(Node* kls, 177 int modifier_mask, int modifier_bits, 178 RegionNode* region); 179 Node* generate_interface_guard(Node* kls, RegionNode* region); 180 Node* generate_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, false); 182 } 183 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, false, true); 185 } 186 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, false); 188 } 189 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 190 return generate_array_guard_common(kls, region, true, true); 191 } 192 Node* generate_array_guard_common(Node* kls, RegionNode* region, 193 bool obj_array, bool not_array); 194 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 195 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 196 bool is_virtual = false, bool is_static = false); 197 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, false, true); 199 } 200 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 201 return generate_method_call(method_id, true, false); 202 } 203 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static); 204 205 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 206 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 207 bool inline_string_compareTo(); 208 bool inline_string_indexOf(); 209 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 210 bool inline_string_equals(); 211 Node* round_double_node(Node* n); 212 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 213 bool inline_math_native(vmIntrinsics::ID id); 214 bool inline_trig(vmIntrinsics::ID id); 215 bool inline_math(vmIntrinsics::ID id); 216 template <typename OverflowOp> 217 bool inline_math_overflow(Node* arg1, Node* arg2); 218 void inline_math_mathExact(Node* math, Node* test); 219 bool inline_math_addExactI(bool is_increment); 220 bool inline_math_addExactL(bool is_increment); 221 bool inline_math_multiplyExactI(); 222 bool inline_math_multiplyExactL(); 223 bool inline_math_negateExactI(); 224 bool inline_math_negateExactL(); 225 bool inline_math_subtractExactI(bool is_decrement); 226 bool inline_math_subtractExactL(bool is_decrement); 227 bool inline_exp(); 228 bool inline_pow(); 229 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 230 bool inline_min_max(vmIntrinsics::ID id); 231 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 232 // This returns Type::AnyPtr, RawPtr, or OopPtr. 233 int classify_unsafe_addr(Node* &base, Node* &offset); 234 Node* make_unsafe_address(Node* base, Node* offset); 235 // Helper for inline_unsafe_access. 236 // Generates the guards that check whether the result of 237 // Unsafe.getObject should be recorded in an SATB log buffer. 238 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 239 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned); 240 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static); 241 static bool klass_needs_init_guard(Node* kls); 242 bool inline_unsafe_allocate(); 243 bool inline_unsafe_copyMemory(); 244 bool inline_native_currentThread(); 245 #ifdef TRACE_HAVE_INTRINSICS 246 bool inline_native_classID(); 247 bool inline_native_threadID(); 248 #endif 249 bool inline_native_time_funcs(address method, const char* funcName); 250 bool inline_native_isInterrupted(); 251 bool inline_native_Class_query(vmIntrinsics::ID id); 252 bool inline_native_subtype_check(); 253 254 bool inline_native_newArray(); 255 bool inline_native_getLength(); 256 bool inline_array_copyOf(bool is_copyOfRange); 257 bool inline_array_equals(); 258 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 259 bool inline_native_clone(bool is_virtual); 260 bool inline_native_Reflection_getCallerClass(); 261 // Helper function for inlining native object hash method 262 bool inline_native_hashcode(bool is_virtual, bool is_static); 263 bool inline_native_getClass(); 264 265 // Helper functions for inlining arraycopy 266 bool inline_arraycopy(); 267 void generate_arraycopy(const TypePtr* adr_type, 268 BasicType basic_elem_type, 269 Node* src, Node* src_offset, 270 Node* dest, Node* dest_offset, 271 Node* copy_length, 272 bool disjoint_bases = false, 273 bool length_never_negative = false, 274 RegionNode* slow_region = NULL); 275 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 276 RegionNode* slow_region); 277 void generate_clear_array(const TypePtr* adr_type, 278 Node* dest, 279 BasicType basic_elem_type, 280 Node* slice_off, 281 Node* slice_len, 282 Node* slice_end); 283 bool generate_block_arraycopy(const TypePtr* adr_type, 284 BasicType basic_elem_type, 285 AllocateNode* alloc, 286 Node* src, Node* src_offset, 287 Node* dest, Node* dest_offset, 288 Node* dest_size, bool dest_uninitialized); 289 void generate_slow_arraycopy(const TypePtr* adr_type, 290 Node* src, Node* src_offset, 291 Node* dest, Node* dest_offset, 292 Node* copy_length, bool dest_uninitialized); 293 Node* generate_checkcast_arraycopy(const TypePtr* adr_type, 294 Node* dest_elem_klass, 295 Node* src, Node* src_offset, 296 Node* dest, Node* dest_offset, 297 Node* copy_length, bool dest_uninitialized); 298 Node* generate_generic_arraycopy(const TypePtr* adr_type, 299 Node* src, Node* src_offset, 300 Node* dest, Node* dest_offset, 301 Node* copy_length, bool dest_uninitialized); 302 void generate_unchecked_arraycopy(const TypePtr* adr_type, 303 BasicType basic_elem_type, 304 bool disjoint_bases, 305 Node* src, Node* src_offset, 306 Node* dest, Node* dest_offset, 307 Node* copy_length, bool dest_uninitialized); 308 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 309 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 310 bool inline_unsafe_ordered_store(BasicType type); 311 bool inline_unsafe_fence(vmIntrinsics::ID id); 312 bool inline_fp_conversions(vmIntrinsics::ID id); 313 bool inline_number_methods(vmIntrinsics::ID id); 314 bool inline_reference_get(); 315 bool inline_aescrypt_Block(vmIntrinsics::ID id); 316 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 317 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 318 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 319 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 320 bool inline_sha_implCompress(vmIntrinsics::ID id); 321 bool inline_digestBase_implCompressMB(int predicate); 322 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 323 bool long_state, address stubAddr, const char *stubName, 324 Node* src_start, Node* ofs, Node* limit); 325 Node* get_state_from_sha_object(Node *sha_object); 326 Node* get_state_from_sha5_object(Node *sha_object); 327 Node* inline_digestBase_implCompressMB_predicate(int predicate); 328 bool inline_encodeISOArray(); 329 bool inline_updateCRC32(); 330 bool inline_updateBytesCRC32(); 331 bool inline_updateByteBufferCRC32(); 332 bool inline_multiplyToLen(); 333 bool inline_squareToLen(); 334 bool inline_mulAdd(); 335 bool inline_montgomeryMultiply(); 336 bool inline_montgomerySquare(); 337 338 bool inline_profileBoolean(); 339 }; 340 341 342 //---------------------------make_vm_intrinsic---------------------------- 343 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 344 vmIntrinsics::ID id = m->intrinsic_id(); 345 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 346 347 ccstr disable_intr = NULL; 348 349 if ((DisableIntrinsic[0] != '\0' 350 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 351 (method_has_option_value("DisableIntrinsic", disable_intr) 352 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 353 // disabled by a user request on the command line: 354 // example: -XX:DisableIntrinsic=_hashCode,_getClass 355 return NULL; 356 } 357 358 if (!m->is_loaded()) { 359 // do not attempt to inline unloaded methods 360 return NULL; 361 } 362 363 // Only a few intrinsics implement a virtual dispatch. 364 // They are expensive calls which are also frequently overridden. 365 if (is_virtual) { 366 switch (id) { 367 case vmIntrinsics::_hashCode: 368 case vmIntrinsics::_clone: 369 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 370 break; 371 default: 372 return NULL; 373 } 374 } 375 376 // -XX:-InlineNatives disables nearly all intrinsics: 377 if (!InlineNatives) { 378 switch (id) { 379 case vmIntrinsics::_indexOf: 380 case vmIntrinsics::_compareTo: 381 case vmIntrinsics::_equals: 382 case vmIntrinsics::_equalsC: 383 case vmIntrinsics::_getAndAddInt: 384 case vmIntrinsics::_getAndAddLong: 385 case vmIntrinsics::_getAndSetInt: 386 case vmIntrinsics::_getAndSetLong: 387 case vmIntrinsics::_getAndSetObject: 388 case vmIntrinsics::_loadFence: 389 case vmIntrinsics::_storeFence: 390 case vmIntrinsics::_fullFence: 391 break; // InlineNatives does not control String.compareTo 392 case vmIntrinsics::_Reference_get: 393 break; // InlineNatives does not control Reference.get 394 default: 395 return NULL; 396 } 397 } 398 399 int predicates = 0; 400 bool does_virtual_dispatch = false; 401 402 switch (id) { 403 case vmIntrinsics::_compareTo: 404 if (!SpecialStringCompareTo) return NULL; 405 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 406 break; 407 case vmIntrinsics::_indexOf: 408 if (!SpecialStringIndexOf) return NULL; 409 break; 410 case vmIntrinsics::_equals: 411 if (!SpecialStringEquals) return NULL; 412 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 413 break; 414 case vmIntrinsics::_equalsC: 415 if (!SpecialArraysEquals) return NULL; 416 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 417 break; 418 case vmIntrinsics::_arraycopy: 419 if (!InlineArrayCopy) return NULL; 420 break; 421 case vmIntrinsics::_copyMemory: 422 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 423 if (!InlineArrayCopy) return NULL; 424 break; 425 case vmIntrinsics::_hashCode: 426 if (!InlineObjectHash) return NULL; 427 does_virtual_dispatch = true; 428 break; 429 case vmIntrinsics::_clone: 430 does_virtual_dispatch = true; 431 case vmIntrinsics::_copyOf: 432 case vmIntrinsics::_copyOfRange: 433 if (!InlineObjectCopy) return NULL; 434 // These also use the arraycopy intrinsic mechanism: 435 if (!InlineArrayCopy) return NULL; 436 break; 437 case vmIntrinsics::_encodeISOArray: 438 if (!SpecialEncodeISOArray) return NULL; 439 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 440 break; 441 case vmIntrinsics::_checkIndex: 442 // We do not intrinsify this. The optimizer does fine with it. 443 return NULL; 444 445 case vmIntrinsics::_getCallerClass: 446 if (!UseNewReflection) return NULL; 447 if (!InlineReflectionGetCallerClass) return NULL; 448 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 449 break; 450 451 case vmIntrinsics::_bitCount_i: 452 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 453 break; 454 455 case vmIntrinsics::_bitCount_l: 456 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 457 break; 458 459 case vmIntrinsics::_numberOfLeadingZeros_i: 460 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 461 break; 462 463 case vmIntrinsics::_numberOfLeadingZeros_l: 464 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 465 break; 466 467 case vmIntrinsics::_numberOfTrailingZeros_i: 468 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 469 break; 470 471 case vmIntrinsics::_numberOfTrailingZeros_l: 472 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 473 break; 474 475 case vmIntrinsics::_reverseBytes_c: 476 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 477 break; 478 case vmIntrinsics::_reverseBytes_s: 479 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 480 break; 481 case vmIntrinsics::_reverseBytes_i: 482 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 483 break; 484 case vmIntrinsics::_reverseBytes_l: 485 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 486 break; 487 488 case vmIntrinsics::_Reference_get: 489 // Use the intrinsic version of Reference.get() so that the value in 490 // the referent field can be registered by the G1 pre-barrier code. 491 // Also add memory barrier to prevent commoning reads from this field 492 // across safepoint since GC can change it value. 493 break; 494 495 case vmIntrinsics::_compareAndSwapObject: 496 #ifdef _LP64 497 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 498 #endif 499 break; 500 501 case vmIntrinsics::_compareAndSwapLong: 502 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 503 break; 504 505 case vmIntrinsics::_getAndAddInt: 506 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 507 break; 508 509 case vmIntrinsics::_getAndAddLong: 510 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 511 break; 512 513 case vmIntrinsics::_getAndSetInt: 514 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 515 break; 516 517 case vmIntrinsics::_getAndSetLong: 518 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 519 break; 520 521 case vmIntrinsics::_getAndSetObject: 522 #ifdef _LP64 523 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 524 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 525 break; 526 #else 527 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 528 break; 529 #endif 530 531 case vmIntrinsics::_aescrypt_encryptBlock: 532 case vmIntrinsics::_aescrypt_decryptBlock: 533 if (!UseAESIntrinsics) return NULL; 534 break; 535 536 case vmIntrinsics::_multiplyToLen: 537 if (!UseMultiplyToLenIntrinsic) return NULL; 538 break; 539 540 case vmIntrinsics::_squareToLen: 541 if (!UseSquareToLenIntrinsic) return NULL; 542 break; 543 544 case vmIntrinsics::_mulAdd: 545 if (!UseMulAddIntrinsic) return NULL; 546 break; 547 548 case vmIntrinsics::_montgomeryMultiply: 549 if (!UseMontgomeryMultiplyIntrinsic) return NULL; 550 break; 551 case vmIntrinsics::_montgomerySquare: 552 if (!UseMontgomerySquareIntrinsic) return NULL; 553 break; 554 555 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 556 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 557 if (!UseAESIntrinsics) return NULL; 558 // these two require the predicated logic 559 predicates = 1; 560 break; 561 562 case vmIntrinsics::_sha_implCompress: 563 if (!UseSHA1Intrinsics) return NULL; 564 break; 565 566 case vmIntrinsics::_sha2_implCompress: 567 if (!UseSHA256Intrinsics) return NULL; 568 break; 569 570 case vmIntrinsics::_sha5_implCompress: 571 if (!UseSHA512Intrinsics) return NULL; 572 break; 573 574 case vmIntrinsics::_digestBase_implCompressMB: 575 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 576 predicates = 3; 577 break; 578 579 case vmIntrinsics::_updateCRC32: 580 case vmIntrinsics::_updateBytesCRC32: 581 case vmIntrinsics::_updateByteBufferCRC32: 582 if (!UseCRC32Intrinsics) return NULL; 583 break; 584 585 case vmIntrinsics::_incrementExactI: 586 case vmIntrinsics::_addExactI: 587 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 588 break; 589 case vmIntrinsics::_incrementExactL: 590 case vmIntrinsics::_addExactL: 591 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 592 break; 593 case vmIntrinsics::_decrementExactI: 594 case vmIntrinsics::_subtractExactI: 595 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 596 break; 597 case vmIntrinsics::_decrementExactL: 598 case vmIntrinsics::_subtractExactL: 599 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 600 break; 601 case vmIntrinsics::_negateExactI: 602 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 603 break; 604 case vmIntrinsics::_negateExactL: 605 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 606 break; 607 case vmIntrinsics::_multiplyExactI: 608 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 609 break; 610 case vmIntrinsics::_multiplyExactL: 611 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 612 break; 613 614 default: 615 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 616 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 617 break; 618 } 619 620 // -XX:-InlineClassNatives disables natives from the Class class. 621 // The flag applies to all reflective calls, notably Array.newArray 622 // (visible to Java programmers as Array.newInstance). 623 if (m->holder()->name() == ciSymbol::java_lang_Class() || 624 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 625 if (!InlineClassNatives) return NULL; 626 } 627 628 // -XX:-InlineThreadNatives disables natives from the Thread class. 629 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 630 if (!InlineThreadNatives) return NULL; 631 } 632 633 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 634 if (m->holder()->name() == ciSymbol::java_lang_Math() || 635 m->holder()->name() == ciSymbol::java_lang_Float() || 636 m->holder()->name() == ciSymbol::java_lang_Double()) { 637 if (!InlineMathNatives) return NULL; 638 } 639 640 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 641 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 642 if (!InlineUnsafeOps) return NULL; 643 } 644 645 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 646 } 647 648 //----------------------register_library_intrinsics----------------------- 649 // Initialize this file's data structures, for each Compile instance. 650 void Compile::register_library_intrinsics() { 651 // Nothing to do here. 652 } 653 654 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 655 LibraryCallKit kit(jvms, this); 656 Compile* C = kit.C; 657 int nodes = C->unique(); 658 #ifndef PRODUCT 659 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 660 char buf[1000]; 661 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 662 tty->print_cr("Intrinsic %s", str); 663 } 664 #endif 665 ciMethod* callee = kit.callee(); 666 const int bci = kit.bci(); 667 668 // Try to inline the intrinsic. 669 if (kit.try_to_inline(_last_predicate)) { 670 if (C->print_intrinsics() || C->print_inlining()) { 671 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 672 } 673 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 674 if (C->log()) { 675 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 676 vmIntrinsics::name_at(intrinsic_id()), 677 (is_virtual() ? " virtual='1'" : ""), 678 C->unique() - nodes); 679 } 680 // Push the result from the inlined method onto the stack. 681 kit.push_result(); 682 return kit.transfer_exceptions_into_jvms(); 683 } 684 685 // The intrinsic bailed out 686 if (C->print_intrinsics() || C->print_inlining()) { 687 if (jvms->has_method()) { 688 // Not a root compile. 689 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 690 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 691 } else { 692 // Root compile 693 tty->print("Did not generate intrinsic %s%s at bci:%d in", 694 vmIntrinsics::name_at(intrinsic_id()), 695 (is_virtual() ? " (virtual)" : ""), bci); 696 } 697 } 698 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 699 return NULL; 700 } 701 702 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 703 LibraryCallKit kit(jvms, this); 704 Compile* C = kit.C; 705 int nodes = C->unique(); 706 _last_predicate = predicate; 707 #ifndef PRODUCT 708 assert(is_predicated() && predicate < predicates_count(), "sanity"); 709 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 710 char buf[1000]; 711 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 712 tty->print_cr("Predicate for intrinsic %s", str); 713 } 714 #endif 715 ciMethod* callee = kit.callee(); 716 const int bci = kit.bci(); 717 718 Node* slow_ctl = kit.try_to_predicate(predicate); 719 if (!kit.failing()) { 720 if (C->print_intrinsics() || C->print_inlining()) { 721 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 722 } 723 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 724 if (C->log()) { 725 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 726 vmIntrinsics::name_at(intrinsic_id()), 727 (is_virtual() ? " virtual='1'" : ""), 728 C->unique() - nodes); 729 } 730 return slow_ctl; // Could be NULL if the check folds. 731 } 732 733 // The intrinsic bailed out 734 if (C->print_intrinsics() || C->print_inlining()) { 735 if (jvms->has_method()) { 736 // Not a root compile. 737 const char* msg = "failed to generate predicate for intrinsic"; 738 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 739 } else { 740 // Root compile 741 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 742 vmIntrinsics::name_at(intrinsic_id()), 743 (is_virtual() ? " (virtual)" : ""), bci); 744 } 745 } 746 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 747 return NULL; 748 } 749 750 bool LibraryCallKit::try_to_inline(int predicate) { 751 // Handle symbolic names for otherwise undistinguished boolean switches: 752 const bool is_store = true; 753 const bool is_native_ptr = true; 754 const bool is_static = true; 755 const bool is_volatile = true; 756 757 if (!jvms()->has_method()) { 758 // Root JVMState has a null method. 759 assert(map()->memory()->Opcode() == Op_Parm, ""); 760 // Insert the memory aliasing node 761 set_all_memory(reset_memory()); 762 } 763 assert(merged_memory(), ""); 764 765 766 switch (intrinsic_id()) { 767 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 768 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 769 case vmIntrinsics::_getClass: return inline_native_getClass(); 770 771 case vmIntrinsics::_dsin: 772 case vmIntrinsics::_dcos: 773 case vmIntrinsics::_dtan: 774 case vmIntrinsics::_dabs: 775 case vmIntrinsics::_datan2: 776 case vmIntrinsics::_dsqrt: 777 case vmIntrinsics::_dexp: 778 case vmIntrinsics::_dlog: 779 case vmIntrinsics::_dlog10: 780 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 781 782 case vmIntrinsics::_min: 783 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 784 785 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 786 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 787 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 788 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 789 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 790 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 791 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 792 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 793 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 794 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 795 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 796 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 797 798 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 799 800 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 801 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 802 case vmIntrinsics::_equals: return inline_string_equals(); 803 804 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false); 805 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false); 806 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 807 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 808 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 809 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false); 810 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false); 811 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 812 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 813 814 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false); 815 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false); 816 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false); 817 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false); 818 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false); 819 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false); 820 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false); 821 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 822 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 823 824 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false); 825 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false); 826 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false); 827 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false); 828 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false); 829 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false); 830 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false); 831 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false); 832 833 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false); 834 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false); 835 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false); 836 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false); 837 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false); 838 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false); 839 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false); 840 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false); 841 842 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false); 843 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false); 844 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false); 845 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false); 846 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false); 847 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false); 848 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false); 849 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false); 850 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false); 851 852 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false); 853 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false); 854 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false); 855 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false); 856 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false); 857 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false); 858 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false); 859 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false); 860 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false); 861 862 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 863 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 864 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 865 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 866 867 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 868 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 869 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 870 871 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 872 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 873 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 874 875 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 876 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 877 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 878 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 879 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 880 881 case vmIntrinsics::_loadFence: 882 case vmIntrinsics::_storeFence: 883 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 884 885 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 886 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 887 888 #ifdef TRACE_HAVE_INTRINSICS 889 case vmIntrinsics::_classID: return inline_native_classID(); 890 case vmIntrinsics::_threadID: return inline_native_threadID(); 891 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 892 #endif 893 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 894 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 895 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 896 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 897 case vmIntrinsics::_newArray: return inline_native_newArray(); 898 case vmIntrinsics::_getLength: return inline_native_getLength(); 899 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 900 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 901 case vmIntrinsics::_equalsC: return inline_array_equals(); 902 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 903 904 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 905 906 case vmIntrinsics::_isInstance: 907 case vmIntrinsics::_getModifiers: 908 case vmIntrinsics::_isInterface: 909 case vmIntrinsics::_isArray: 910 case vmIntrinsics::_isPrimitive: 911 case vmIntrinsics::_getSuperclass: 912 case vmIntrinsics::_getComponentType: 913 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 914 915 case vmIntrinsics::_floatToRawIntBits: 916 case vmIntrinsics::_floatToIntBits: 917 case vmIntrinsics::_intBitsToFloat: 918 case vmIntrinsics::_doubleToRawLongBits: 919 case vmIntrinsics::_doubleToLongBits: 920 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 921 922 case vmIntrinsics::_numberOfLeadingZeros_i: 923 case vmIntrinsics::_numberOfLeadingZeros_l: 924 case vmIntrinsics::_numberOfTrailingZeros_i: 925 case vmIntrinsics::_numberOfTrailingZeros_l: 926 case vmIntrinsics::_bitCount_i: 927 case vmIntrinsics::_bitCount_l: 928 case vmIntrinsics::_reverseBytes_i: 929 case vmIntrinsics::_reverseBytes_l: 930 case vmIntrinsics::_reverseBytes_s: 931 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 932 933 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 934 935 case vmIntrinsics::_Reference_get: return inline_reference_get(); 936 937 case vmIntrinsics::_aescrypt_encryptBlock: 938 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 939 940 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 941 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 942 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 943 944 case vmIntrinsics::_sha_implCompress: 945 case vmIntrinsics::_sha2_implCompress: 946 case vmIntrinsics::_sha5_implCompress: 947 return inline_sha_implCompress(intrinsic_id()); 948 949 case vmIntrinsics::_digestBase_implCompressMB: 950 return inline_digestBase_implCompressMB(predicate); 951 952 case vmIntrinsics::_multiplyToLen: 953 return inline_multiplyToLen(); 954 955 case vmIntrinsics::_squareToLen: 956 return inline_squareToLen(); 957 958 case vmIntrinsics::_mulAdd: 959 return inline_mulAdd(); 960 961 case vmIntrinsics::_montgomeryMultiply: 962 return inline_montgomeryMultiply(); 963 case vmIntrinsics::_montgomerySquare: 964 return inline_montgomerySquare(); 965 966 case vmIntrinsics::_encodeISOArray: 967 return inline_encodeISOArray(); 968 969 case vmIntrinsics::_updateCRC32: 970 return inline_updateCRC32(); 971 case vmIntrinsics::_updateBytesCRC32: 972 return inline_updateBytesCRC32(); 973 case vmIntrinsics::_updateByteBufferCRC32: 974 return inline_updateByteBufferCRC32(); 975 976 case vmIntrinsics::_profileBoolean: 977 return inline_profileBoolean(); 978 979 default: 980 // If you get here, it may be that someone has added a new intrinsic 981 // to the list in vmSymbols.hpp without implementing it here. 982 #ifndef PRODUCT 983 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 984 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 985 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 986 } 987 #endif 988 return false; 989 } 990 } 991 992 Node* LibraryCallKit::try_to_predicate(int predicate) { 993 if (!jvms()->has_method()) { 994 // Root JVMState has a null method. 995 assert(map()->memory()->Opcode() == Op_Parm, ""); 996 // Insert the memory aliasing node 997 set_all_memory(reset_memory()); 998 } 999 assert(merged_memory(), ""); 1000 1001 switch (intrinsic_id()) { 1002 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 1003 return inline_cipherBlockChaining_AESCrypt_predicate(false); 1004 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 1005 return inline_cipherBlockChaining_AESCrypt_predicate(true); 1006 case vmIntrinsics::_digestBase_implCompressMB: 1007 return inline_digestBase_implCompressMB_predicate(predicate); 1008 1009 default: 1010 // If you get here, it may be that someone has added a new intrinsic 1011 // to the list in vmSymbols.hpp without implementing it here. 1012 #ifndef PRODUCT 1013 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 1014 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 1015 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 1016 } 1017 #endif 1018 Node* slow_ctl = control(); 1019 set_control(top()); // No fast path instrinsic 1020 return slow_ctl; 1021 } 1022 } 1023 1024 //------------------------------set_result------------------------------- 1025 // Helper function for finishing intrinsics. 1026 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1027 record_for_igvn(region); 1028 set_control(_gvn.transform(region)); 1029 set_result( _gvn.transform(value)); 1030 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1031 } 1032 1033 //------------------------------generate_guard--------------------------- 1034 // Helper function for generating guarded fast-slow graph structures. 1035 // The given 'test', if true, guards a slow path. If the test fails 1036 // then a fast path can be taken. (We generally hope it fails.) 1037 // In all cases, GraphKit::control() is updated to the fast path. 1038 // The returned value represents the control for the slow path. 1039 // The return value is never 'top'; it is either a valid control 1040 // or NULL if it is obvious that the slow path can never be taken. 1041 // Also, if region and the slow control are not NULL, the slow edge 1042 // is appended to the region. 1043 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1044 if (stopped()) { 1045 // Already short circuited. 1046 return NULL; 1047 } 1048 1049 // Build an if node and its projections. 1050 // If test is true we take the slow path, which we assume is uncommon. 1051 if (_gvn.type(test) == TypeInt::ZERO) { 1052 // The slow branch is never taken. No need to build this guard. 1053 return NULL; 1054 } 1055 1056 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1057 1058 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff)); 1059 if (if_slow == top()) { 1060 // The slow branch is never taken. No need to build this guard. 1061 return NULL; 1062 } 1063 1064 if (region != NULL) 1065 region->add_req(if_slow); 1066 1067 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff)); 1068 set_control(if_fast); 1069 1070 return if_slow; 1071 } 1072 1073 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1074 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1075 } 1076 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1077 return generate_guard(test, region, PROB_FAIR); 1078 } 1079 1080 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1081 Node* *pos_index) { 1082 if (stopped()) 1083 return NULL; // already stopped 1084 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1085 return NULL; // index is already adequately typed 1086 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0))); 1087 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 1088 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1089 if (is_neg != NULL && pos_index != NULL) { 1090 // Emulate effect of Parse::adjust_map_after_if. 1091 Node* ccast = new (C) CastIINode(index, TypeInt::POS); 1092 ccast->set_req(0, control()); 1093 (*pos_index) = _gvn.transform(ccast); 1094 } 1095 return is_neg; 1096 } 1097 1098 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative, 1099 Node* *pos_index) { 1100 if (stopped()) 1101 return NULL; // already stopped 1102 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint] 1103 return NULL; // index is already adequately typed 1104 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0))); 1105 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le); 1106 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq)); 1107 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN); 1108 if (is_notp != NULL && pos_index != NULL) { 1109 // Emulate effect of Parse::adjust_map_after_if. 1110 Node* ccast = new (C) CastIINode(index, TypeInt::POS1); 1111 ccast->set_req(0, control()); 1112 (*pos_index) = _gvn.transform(ccast); 1113 } 1114 return is_notp; 1115 } 1116 1117 // Make sure that 'position' is a valid limit index, in [0..length]. 1118 // There are two equivalent plans for checking this: 1119 // A. (offset + copyLength) unsigned<= arrayLength 1120 // B. offset <= (arrayLength - copyLength) 1121 // We require that all of the values above, except for the sum and 1122 // difference, are already known to be non-negative. 1123 // Plan A is robust in the face of overflow, if offset and copyLength 1124 // are both hugely positive. 1125 // 1126 // Plan B is less direct and intuitive, but it does not overflow at 1127 // all, since the difference of two non-negatives is always 1128 // representable. Whenever Java methods must perform the equivalent 1129 // check they generally use Plan B instead of Plan A. 1130 // For the moment we use Plan A. 1131 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1132 Node* subseq_length, 1133 Node* array_length, 1134 RegionNode* region) { 1135 if (stopped()) 1136 return NULL; // already stopped 1137 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1138 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1139 return NULL; // common case of whole-array copy 1140 Node* last = subseq_length; 1141 if (!zero_offset) // last += offset 1142 last = _gvn.transform(new (C) AddINode(last, offset)); 1143 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last)); 1144 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt)); 1145 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1146 return is_over; 1147 } 1148 1149 1150 //--------------------------generate_current_thread-------------------- 1151 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1152 ciKlass* thread_klass = env()->Thread_klass(); 1153 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1154 Node* thread = _gvn.transform(new (C) ThreadLocalNode()); 1155 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1156 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1157 tls_output = thread; 1158 return threadObj; 1159 } 1160 1161 1162 //------------------------------make_string_method_node------------------------ 1163 // Helper method for String intrinsic functions. This version is called 1164 // with str1 and str2 pointing to String object nodes. 1165 // 1166 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1167 Node* no_ctrl = NULL; 1168 1169 // Get start addr of string 1170 Node* str1_value = load_String_value(no_ctrl, str1); 1171 Node* str1_offset = load_String_offset(no_ctrl, str1); 1172 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1173 1174 // Get length of string 1 1175 Node* str1_len = load_String_length(no_ctrl, str1); 1176 1177 Node* str2_value = load_String_value(no_ctrl, str2); 1178 Node* str2_offset = load_String_offset(no_ctrl, str2); 1179 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1180 1181 Node* str2_len = NULL; 1182 Node* result = NULL; 1183 1184 switch (opcode) { 1185 case Op_StrIndexOf: 1186 // Get length of string 2 1187 str2_len = load_String_length(no_ctrl, str2); 1188 1189 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1190 str1_start, str1_len, str2_start, str2_len); 1191 break; 1192 case Op_StrComp: 1193 // Get length of string 2 1194 str2_len = load_String_length(no_ctrl, str2); 1195 1196 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1197 str1_start, str1_len, str2_start, str2_len); 1198 break; 1199 case Op_StrEquals: 1200 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1201 str1_start, str2_start, str1_len); 1202 break; 1203 default: 1204 ShouldNotReachHere(); 1205 return NULL; 1206 } 1207 1208 // All these intrinsics have checks. 1209 C->set_has_split_ifs(true); // Has chance for split-if optimization 1210 1211 return _gvn.transform(result); 1212 } 1213 1214 // Helper method for String intrinsic functions. This version is called 1215 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1216 // to Int nodes containing the lenghts of str1 and str2. 1217 // 1218 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1219 Node* result = NULL; 1220 switch (opcode) { 1221 case Op_StrIndexOf: 1222 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1223 str1_start, cnt1, str2_start, cnt2); 1224 break; 1225 case Op_StrComp: 1226 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS), 1227 str1_start, cnt1, str2_start, cnt2); 1228 break; 1229 case Op_StrEquals: 1230 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1231 str1_start, str2_start, cnt1); 1232 break; 1233 default: 1234 ShouldNotReachHere(); 1235 return NULL; 1236 } 1237 1238 // All these intrinsics have checks. 1239 C->set_has_split_ifs(true); // Has chance for split-if optimization 1240 1241 return _gvn.transform(result); 1242 } 1243 1244 //------------------------------inline_string_compareTo------------------------ 1245 // public int java.lang.String.compareTo(String anotherString); 1246 bool LibraryCallKit::inline_string_compareTo() { 1247 Node* receiver = null_check(argument(0)); 1248 Node* arg = null_check(argument(1)); 1249 if (stopped()) { 1250 return true; 1251 } 1252 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1253 return true; 1254 } 1255 1256 //------------------------------inline_string_equals------------------------ 1257 bool LibraryCallKit::inline_string_equals() { 1258 Node* receiver = null_check_receiver(); 1259 // NOTE: Do not null check argument for String.equals() because spec 1260 // allows to specify NULL as argument. 1261 Node* argument = this->argument(1); 1262 if (stopped()) { 1263 return true; 1264 } 1265 1266 // paths (plus control) merge 1267 RegionNode* region = new (C) RegionNode(5); 1268 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 1269 1270 // does source == target string? 1271 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument)); 1272 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 1273 1274 Node* if_eq = generate_slow_guard(bol, NULL); 1275 if (if_eq != NULL) { 1276 // receiver == argument 1277 phi->init_req(2, intcon(1)); 1278 region->init_req(2, if_eq); 1279 } 1280 1281 // get String klass for instanceOf 1282 ciInstanceKlass* klass = env()->String_klass(); 1283 1284 if (!stopped()) { 1285 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1286 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1))); 1287 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 1288 1289 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1290 //instanceOf == true, fallthrough 1291 1292 if (inst_false != NULL) { 1293 phi->init_req(3, intcon(0)); 1294 region->init_req(3, inst_false); 1295 } 1296 } 1297 1298 if (!stopped()) { 1299 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1300 1301 // Properly cast the argument to String 1302 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type)); 1303 // This path is taken only when argument's type is String:NotNull. 1304 argument = cast_not_null(argument, false); 1305 1306 Node* no_ctrl = NULL; 1307 1308 // Get start addr of receiver 1309 Node* receiver_val = load_String_value(no_ctrl, receiver); 1310 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1311 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1312 1313 // Get length of receiver 1314 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1315 1316 // Get start addr of argument 1317 Node* argument_val = load_String_value(no_ctrl, argument); 1318 Node* argument_offset = load_String_offset(no_ctrl, argument); 1319 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1320 1321 // Get length of argument 1322 Node* argument_cnt = load_String_length(no_ctrl, argument); 1323 1324 // Check for receiver count != argument count 1325 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt)); 1326 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne)); 1327 Node* if_ne = generate_slow_guard(bol, NULL); 1328 if (if_ne != NULL) { 1329 phi->init_req(4, intcon(0)); 1330 region->init_req(4, if_ne); 1331 } 1332 1333 // Check for count == 0 is done by assembler code for StrEquals. 1334 1335 if (!stopped()) { 1336 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1337 phi->init_req(1, equals); 1338 region->init_req(1, control()); 1339 } 1340 } 1341 1342 // post merge 1343 set_control(_gvn.transform(region)); 1344 record_for_igvn(region); 1345 1346 set_result(_gvn.transform(phi)); 1347 return true; 1348 } 1349 1350 //------------------------------inline_array_equals---------------------------- 1351 bool LibraryCallKit::inline_array_equals() { 1352 Node* arg1 = argument(0); 1353 Node* arg2 = argument(1); 1354 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1355 return true; 1356 } 1357 1358 // Java version of String.indexOf(constant string) 1359 // class StringDecl { 1360 // StringDecl(char[] ca) { 1361 // offset = 0; 1362 // count = ca.length; 1363 // value = ca; 1364 // } 1365 // int offset; 1366 // int count; 1367 // char[] value; 1368 // } 1369 // 1370 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1371 // int targetOffset, int cache_i, int md2) { 1372 // int cache = cache_i; 1373 // int sourceOffset = string_object.offset; 1374 // int sourceCount = string_object.count; 1375 // int targetCount = target_object.length; 1376 // 1377 // int targetCountLess1 = targetCount - 1; 1378 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1379 // 1380 // char[] source = string_object.value; 1381 // char[] target = target_object; 1382 // int lastChar = target[targetCountLess1]; 1383 // 1384 // outer_loop: 1385 // for (int i = sourceOffset; i < sourceEnd; ) { 1386 // int src = source[i + targetCountLess1]; 1387 // if (src == lastChar) { 1388 // // With random strings and a 4-character alphabet, 1389 // // reverse matching at this point sets up 0.8% fewer 1390 // // frames, but (paradoxically) makes 0.3% more probes. 1391 // // Since those probes are nearer the lastChar probe, 1392 // // there is may be a net D$ win with reverse matching. 1393 // // But, reversing loop inhibits unroll of inner loop 1394 // // for unknown reason. So, does running outer loop from 1395 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1396 // for (int j = 0; j < targetCountLess1; j++) { 1397 // if (target[targetOffset + j] != source[i+j]) { 1398 // if ((cache & (1 << source[i+j])) == 0) { 1399 // if (md2 < j+1) { 1400 // i += j+1; 1401 // continue outer_loop; 1402 // } 1403 // } 1404 // i += md2; 1405 // continue outer_loop; 1406 // } 1407 // } 1408 // return i - sourceOffset; 1409 // } 1410 // if ((cache & (1 << src)) == 0) { 1411 // i += targetCountLess1; 1412 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1413 // i++; 1414 // } 1415 // return -1; 1416 // } 1417 1418 //------------------------------string_indexOf------------------------ 1419 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1420 jint cache_i, jint md2_i) { 1421 1422 Node* no_ctrl = NULL; 1423 float likely = PROB_LIKELY(0.9); 1424 float unlikely = PROB_UNLIKELY(0.9); 1425 1426 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1427 1428 Node* source = load_String_value(no_ctrl, string_object); 1429 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1430 Node* sourceCount = load_String_length(no_ctrl, string_object); 1431 1432 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1433 jint target_length = target_array->length(); 1434 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1435 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1436 1437 // String.value field is known to be @Stable. 1438 if (UseImplicitStableValues) { 1439 target = cast_array_to_stable(target, target_type); 1440 } 1441 1442 IdealKit kit(this, false, true); 1443 #define __ kit. 1444 Node* zero = __ ConI(0); 1445 Node* one = __ ConI(1); 1446 Node* cache = __ ConI(cache_i); 1447 Node* md2 = __ ConI(md2_i); 1448 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1449 Node* targetCount = __ ConI(target_length); 1450 Node* targetCountLess1 = __ ConI(target_length - 1); 1451 Node* targetOffset = __ ConI(targetOffset_i); 1452 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1453 1454 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1455 Node* outer_loop = __ make_label(2 /* goto */); 1456 Node* return_ = __ make_label(1); 1457 1458 __ set(rtn,__ ConI(-1)); 1459 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1460 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1461 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1462 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1463 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1464 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1465 Node* tpj = __ AddI(targetOffset, __ value(j)); 1466 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1467 Node* ipj = __ AddI(__ value(i), __ value(j)); 1468 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1469 __ if_then(targ, BoolTest::ne, src2); { 1470 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1471 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1472 __ increment(i, __ AddI(__ value(j), one)); 1473 __ goto_(outer_loop); 1474 } __ end_if(); __ dead(j); 1475 }__ end_if(); __ dead(j); 1476 __ increment(i, md2); 1477 __ goto_(outer_loop); 1478 }__ end_if(); 1479 __ increment(j, one); 1480 }__ end_loop(); __ dead(j); 1481 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1482 __ goto_(return_); 1483 }__ end_if(); 1484 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1485 __ increment(i, targetCountLess1); 1486 }__ end_if(); 1487 __ increment(i, one); 1488 __ bind(outer_loop); 1489 }__ end_loop(); __ dead(i); 1490 __ bind(return_); 1491 1492 // Final sync IdealKit and GraphKit. 1493 final_sync(kit); 1494 Node* result = __ value(rtn); 1495 #undef __ 1496 C->set_has_loops(true); 1497 return result; 1498 } 1499 1500 //------------------------------inline_string_indexOf------------------------ 1501 bool LibraryCallKit::inline_string_indexOf() { 1502 Node* receiver = argument(0); 1503 Node* arg = argument(1); 1504 1505 Node* result; 1506 // Disable the use of pcmpestri until it can be guaranteed that 1507 // the load doesn't cross into the uncommited space. 1508 if (Matcher::has_match_rule(Op_StrIndexOf) && 1509 UseSSE42Intrinsics) { 1510 // Generate SSE4.2 version of indexOf 1511 // We currently only have match rules that use SSE4.2 1512 1513 receiver = null_check(receiver); 1514 arg = null_check(arg); 1515 if (stopped()) { 1516 return true; 1517 } 1518 1519 ciInstanceKlass* str_klass = env()->String_klass(); 1520 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1521 1522 // Make the merge point 1523 RegionNode* result_rgn = new (C) RegionNode(4); 1524 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT); 1525 Node* no_ctrl = NULL; 1526 1527 // Get start addr of source string 1528 Node* source = load_String_value(no_ctrl, receiver); 1529 Node* source_offset = load_String_offset(no_ctrl, receiver); 1530 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1531 1532 // Get length of source string 1533 Node* source_cnt = load_String_length(no_ctrl, receiver); 1534 1535 // Get start addr of substring 1536 Node* substr = load_String_value(no_ctrl, arg); 1537 Node* substr_offset = load_String_offset(no_ctrl, arg); 1538 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1539 1540 // Get length of source string 1541 Node* substr_cnt = load_String_length(no_ctrl, arg); 1542 1543 // Check for substr count > string count 1544 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt)); 1545 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt)); 1546 Node* if_gt = generate_slow_guard(bol, NULL); 1547 if (if_gt != NULL) { 1548 result_phi->init_req(2, intcon(-1)); 1549 result_rgn->init_req(2, if_gt); 1550 } 1551 1552 if (!stopped()) { 1553 // Check for substr count == 0 1554 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0))); 1555 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 1556 Node* if_zero = generate_slow_guard(bol, NULL); 1557 if (if_zero != NULL) { 1558 result_phi->init_req(3, intcon(0)); 1559 result_rgn->init_req(3, if_zero); 1560 } 1561 } 1562 1563 if (!stopped()) { 1564 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1565 result_phi->init_req(1, result); 1566 result_rgn->init_req(1, control()); 1567 } 1568 set_control(_gvn.transform(result_rgn)); 1569 record_for_igvn(result_rgn); 1570 result = _gvn.transform(result_phi); 1571 1572 } else { // Use LibraryCallKit::string_indexOf 1573 // don't intrinsify if argument isn't a constant string. 1574 if (!arg->is_Con()) { 1575 return false; 1576 } 1577 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1578 if (str_type == NULL) { 1579 return false; 1580 } 1581 ciInstanceKlass* klass = env()->String_klass(); 1582 ciObject* str_const = str_type->const_oop(); 1583 if (str_const == NULL || str_const->klass() != klass) { 1584 return false; 1585 } 1586 ciInstance* str = str_const->as_instance(); 1587 assert(str != NULL, "must be instance"); 1588 1589 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1590 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1591 1592 int o; 1593 int c; 1594 if (java_lang_String::has_offset_field()) { 1595 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1596 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1597 } else { 1598 o = 0; 1599 c = pat->length(); 1600 } 1601 1602 // constant strings have no offset and count == length which 1603 // simplifies the resulting code somewhat so lets optimize for that. 1604 if (o != 0 || c != pat->length()) { 1605 return false; 1606 } 1607 1608 receiver = null_check(receiver, T_OBJECT); 1609 // NOTE: No null check on the argument is needed since it's a constant String oop. 1610 if (stopped()) { 1611 return true; 1612 } 1613 1614 // The null string as a pattern always returns 0 (match at beginning of string) 1615 if (c == 0) { 1616 set_result(intcon(0)); 1617 return true; 1618 } 1619 1620 // Generate default indexOf 1621 jchar lastChar = pat->char_at(o + (c - 1)); 1622 int cache = 0; 1623 int i; 1624 for (i = 0; i < c - 1; i++) { 1625 assert(i < pat->length(), "out of range"); 1626 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1627 } 1628 1629 int md2 = c; 1630 for (i = 0; i < c - 1; i++) { 1631 assert(i < pat->length(), "out of range"); 1632 if (pat->char_at(o + i) == lastChar) { 1633 md2 = (c - 1) - i; 1634 } 1635 } 1636 1637 result = string_indexOf(receiver, pat, o, cache, md2); 1638 } 1639 set_result(result); 1640 return true; 1641 } 1642 1643 //--------------------------round_double_node-------------------------------- 1644 // Round a double node if necessary. 1645 Node* LibraryCallKit::round_double_node(Node* n) { 1646 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1647 n = _gvn.transform(new (C) RoundDoubleNode(0, n)); 1648 return n; 1649 } 1650 1651 //------------------------------inline_math----------------------------------- 1652 // public static double Math.abs(double) 1653 // public static double Math.sqrt(double) 1654 // public static double Math.log(double) 1655 // public static double Math.log10(double) 1656 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1657 Node* arg = round_double_node(argument(0)); 1658 Node* n = NULL; 1659 switch (id) { 1660 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break; 1661 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break; 1662 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break; 1663 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break; 1664 default: fatal_unexpected_iid(id); break; 1665 } 1666 set_result(_gvn.transform(n)); 1667 return true; 1668 } 1669 1670 //------------------------------inline_trig---------------------------------- 1671 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1672 // argument reduction which will turn into a fast/slow diamond. 1673 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1674 Node* arg = round_double_node(argument(0)); 1675 Node* n = NULL; 1676 1677 switch (id) { 1678 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break; 1679 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break; 1680 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break; 1681 default: fatal_unexpected_iid(id); break; 1682 } 1683 n = _gvn.transform(n); 1684 1685 // Rounding required? Check for argument reduction! 1686 if (Matcher::strict_fp_requires_explicit_rounding) { 1687 static const double pi_4 = 0.7853981633974483; 1688 static const double neg_pi_4 = -0.7853981633974483; 1689 // pi/2 in 80-bit extended precision 1690 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1691 // -pi/2 in 80-bit extended precision 1692 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00}; 1693 // Cutoff value for using this argument reduction technique 1694 //static const double pi_2_minus_epsilon = 1.564660403643354; 1695 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1696 1697 // Pseudocode for sin: 1698 // if (x <= Math.PI / 4.0) { 1699 // if (x >= -Math.PI / 4.0) return fsin(x); 1700 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1701 // } else { 1702 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1703 // } 1704 // return StrictMath.sin(x); 1705 1706 // Pseudocode for cos: 1707 // if (x <= Math.PI / 4.0) { 1708 // if (x >= -Math.PI / 4.0) return fcos(x); 1709 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1710 // } else { 1711 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1712 // } 1713 // return StrictMath.cos(x); 1714 1715 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1716 // requires a special machine instruction to load it. Instead we'll try 1717 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1718 // probably do the math inside the SIN encoding. 1719 1720 // Make the merge point 1721 RegionNode* r = new (C) RegionNode(3); 1722 Node* phi = new (C) PhiNode(r, Type::DOUBLE); 1723 1724 // Flatten arg so we need only 1 test 1725 Node *abs = _gvn.transform(new (C) AbsDNode(arg)); 1726 // Node for PI/4 constant 1727 Node *pi4 = makecon(TypeD::make(pi_4)); 1728 // Check PI/4 : abs(arg) 1729 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs)); 1730 // Check: If PI/4 < abs(arg) then go slow 1731 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt )); 1732 // Branch either way 1733 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1734 set_control(opt_iff(r,iff)); 1735 1736 // Set fast path result 1737 phi->init_req(2, n); 1738 1739 // Slow path - non-blocking leaf call 1740 Node* call = NULL; 1741 switch (id) { 1742 case vmIntrinsics::_dsin: 1743 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1744 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1745 "Sin", NULL, arg, top()); 1746 break; 1747 case vmIntrinsics::_dcos: 1748 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1749 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1750 "Cos", NULL, arg, top()); 1751 break; 1752 case vmIntrinsics::_dtan: 1753 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1754 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1755 "Tan", NULL, arg, top()); 1756 break; 1757 } 1758 assert(control()->in(0) == call, ""); 1759 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 1760 r->init_req(1, control()); 1761 phi->init_req(1, slow_result); 1762 1763 // Post-merge 1764 set_control(_gvn.transform(r)); 1765 record_for_igvn(r); 1766 n = _gvn.transform(phi); 1767 1768 C->set_has_split_ifs(true); // Has chance for split-if optimization 1769 } 1770 set_result(n); 1771 return true; 1772 } 1773 1774 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1775 //------------------- 1776 //result=(result.isNaN())? funcAddr():result; 1777 // Check: If isNaN() by checking result!=result? then either trap 1778 // or go to runtime 1779 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result)); 1780 // Build the boolean node 1781 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq)); 1782 1783 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1784 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1785 // The pow or exp intrinsic returned a NaN, which requires a call 1786 // to the runtime. Recompile with the runtime call. 1787 uncommon_trap(Deoptimization::Reason_intrinsic, 1788 Deoptimization::Action_make_not_entrant); 1789 } 1790 return result; 1791 } else { 1792 // If this inlining ever returned NaN in the past, we compile a call 1793 // to the runtime to properly handle corner cases 1794 1795 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1796 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff)); 1797 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff)); 1798 1799 if (!if_slow->is_top()) { 1800 RegionNode* result_region = new (C) RegionNode(3); 1801 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE); 1802 1803 result_region->init_req(1, if_fast); 1804 result_val->init_req(1, result); 1805 1806 set_control(if_slow); 1807 1808 const TypePtr* no_memory_effects = NULL; 1809 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1810 no_memory_effects, 1811 x, top(), y, y ? top() : NULL); 1812 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0)); 1813 #ifdef ASSERT 1814 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1)); 1815 assert(value_top == top(), "second value must be top"); 1816 #endif 1817 1818 result_region->init_req(2, control()); 1819 result_val->init_req(2, value); 1820 set_control(_gvn.transform(result_region)); 1821 return _gvn.transform(result_val); 1822 } else { 1823 return result; 1824 } 1825 } 1826 } 1827 1828 //------------------------------inline_exp------------------------------------- 1829 // Inline exp instructions, if possible. The Intel hardware only misses 1830 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1831 bool LibraryCallKit::inline_exp() { 1832 Node* arg = round_double_node(argument(0)); 1833 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg)); 1834 1835 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1836 set_result(n); 1837 1838 C->set_has_split_ifs(true); // Has chance for split-if optimization 1839 return true; 1840 } 1841 1842 //------------------------------inline_pow------------------------------------- 1843 // Inline power instructions, if possible. 1844 bool LibraryCallKit::inline_pow() { 1845 // Pseudocode for pow 1846 // if (y == 2) { 1847 // return x * x; 1848 // } else { 1849 // if (x <= 0.0) { 1850 // long longy = (long)y; 1851 // if ((double)longy == y) { // if y is long 1852 // if (y + 1 == y) longy = 0; // huge number: even 1853 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1854 // } else { 1855 // result = NaN; 1856 // } 1857 // } else { 1858 // result = DPow(x,y); 1859 // } 1860 // if (result != result)? { 1861 // result = uncommon_trap() or runtime_call(); 1862 // } 1863 // return result; 1864 // } 1865 1866 Node* x = round_double_node(argument(0)); 1867 Node* y = round_double_node(argument(2)); 1868 1869 Node* result = NULL; 1870 1871 Node* const_two_node = makecon(TypeD::make(2.0)); 1872 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node)); 1873 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq)); 1874 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1875 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node)); 1876 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node)); 1877 1878 RegionNode* region_node = new (C) RegionNode(3); 1879 region_node->init_req(1, if_true); 1880 1881 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE); 1882 // special case for x^y where y == 2, we can convert it to x * x 1883 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x))); 1884 1885 // set control to if_false since we will now process the false branch 1886 set_control(if_false); 1887 1888 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1889 // Short form: skip the fancy tests and just check for NaN result. 1890 result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1891 } else { 1892 // If this inlining ever returned NaN in the past, include all 1893 // checks + call to the runtime. 1894 1895 // Set the merge point for If node with condition of (x <= 0.0) 1896 // There are four possible paths to region node and phi node 1897 RegionNode *r = new (C) RegionNode(4); 1898 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1899 1900 // Build the first if node: if (x <= 0.0) 1901 // Node for 0 constant 1902 Node *zeronode = makecon(TypeD::ZERO); 1903 // Check x:0 1904 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode)); 1905 // Check: If (x<=0) then go complex path 1906 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le )); 1907 // Branch either way 1908 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1909 // Fast path taken; set region slot 3 1910 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1)); 1911 r->init_req(3,fast_taken); // Capture fast-control 1912 1913 // Fast path not-taken, i.e. slow path 1914 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1)); 1915 1916 // Set fast path result 1917 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y)); 1918 phi->init_req(3, fast_result); 1919 1920 // Complex path 1921 // Build the second if node (if y is long) 1922 // Node for (long)y 1923 Node *longy = _gvn.transform(new (C) ConvD2LNode(y)); 1924 // Node for (double)((long) y) 1925 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy)); 1926 // Check (double)((long) y) : y 1927 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y)); 1928 // Check if (y isn't long) then go to slow path 1929 1930 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne )); 1931 // Branch either way 1932 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1933 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2)); 1934 1935 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2)); 1936 1937 // Calculate DPow(abs(x), y)*(1 & (long)y) 1938 // Node for constant 1 1939 Node *conone = longcon(1); 1940 // 1& (long)y 1941 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy)); 1942 1943 // A huge number is always even. Detect a huge number by checking 1944 // if y + 1 == y and set integer to be tested for parity to 0. 1945 // Required for corner case: 1946 // (long)9.223372036854776E18 = max_jlong 1947 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1948 // max_jlong is odd but 9.223372036854776E18 is even 1949 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1)))); 1950 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y)); 1951 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq )); 1952 Node* correctedsign = NULL; 1953 if (ConditionalMoveLimit != 0) { 1954 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1955 } else { 1956 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1957 RegionNode *r = new (C) RegionNode(3); 1958 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 1959 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1))); 1960 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1))); 1961 phi->init_req(1, signnode); 1962 phi->init_req(2, longcon(0)); 1963 correctedsign = _gvn.transform(phi); 1964 ylong_path = _gvn.transform(r); 1965 record_for_igvn(r); 1966 } 1967 1968 // zero node 1969 Node *conzero = longcon(0); 1970 // Check (1&(long)y)==0? 1971 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero)); 1972 // Check if (1&(long)y)!=0?, if so the result is negative 1973 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne )); 1974 // abs(x) 1975 Node *absx=_gvn.transform(new (C) AbsDNode(x)); 1976 // abs(x)^y 1977 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y)); 1978 // -abs(x)^y 1979 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy)); 1980 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1981 Node *signresult = NULL; 1982 if (ConditionalMoveLimit != 0) { 1983 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1984 } else { 1985 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1986 RegionNode *r = new (C) RegionNode(3); 1987 Node *phi = new (C) PhiNode(r, Type::DOUBLE); 1988 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven))); 1989 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven))); 1990 phi->init_req(1, absxpowy); 1991 phi->init_req(2, negabsxpowy); 1992 signresult = _gvn.transform(phi); 1993 ylong_path = _gvn.transform(r); 1994 record_for_igvn(r); 1995 } 1996 // Set complex path fast result 1997 r->init_req(2, ylong_path); 1998 phi->init_req(2, signresult); 1999 2000 static const jlong nan_bits = CONST64(0x7ff8000000000000); 2001 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 2002 r->init_req(1,slow_path); 2003 phi->init_req(1,slow_result); 2004 2005 // Post merge 2006 set_control(_gvn.transform(r)); 2007 record_for_igvn(r); 2008 result = _gvn.transform(phi); 2009 } 2010 2011 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 2012 2013 // control from finish_pow_exp is now input to the region node 2014 region_node->set_req(2, control()); 2015 // the result from finish_pow_exp is now input to the phi node 2016 phi_node->init_req(2, result); 2017 set_control(_gvn.transform(region_node)); 2018 record_for_igvn(region_node); 2019 set_result(_gvn.transform(phi_node)); 2020 2021 C->set_has_split_ifs(true); // Has chance for split-if optimization 2022 return true; 2023 } 2024 2025 //------------------------------runtime_math----------------------------- 2026 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 2027 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 2028 "must be (DD)D or (D)D type"); 2029 2030 // Inputs 2031 Node* a = round_double_node(argument(0)); 2032 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 2033 2034 const TypePtr* no_memory_effects = NULL; 2035 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 2036 no_memory_effects, 2037 a, top(), b, b ? top() : NULL); 2038 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0)); 2039 #ifdef ASSERT 2040 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1)); 2041 assert(value_top == top(), "second value must be top"); 2042 #endif 2043 2044 set_result(value); 2045 return true; 2046 } 2047 2048 //------------------------------inline_math_native----------------------------- 2049 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2050 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2051 switch (id) { 2052 // These intrinsics are not properly supported on all hardware 2053 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2054 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2055 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2056 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2057 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2058 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2059 2060 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2061 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2062 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2063 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2064 2065 // These intrinsics are supported on all hardware 2066 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2067 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2068 2069 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2070 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2071 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2072 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2073 #undef FN_PTR 2074 2075 // These intrinsics are not yet correctly implemented 2076 case vmIntrinsics::_datan2: 2077 return false; 2078 2079 default: 2080 fatal_unexpected_iid(id); 2081 return false; 2082 } 2083 } 2084 2085 static bool is_simple_name(Node* n) { 2086 return (n->req() == 1 // constant 2087 || (n->is_Type() && n->as_Type()->type()->singleton()) 2088 || n->is_Proj() // parameter or return value 2089 || n->is_Phi() // local of some sort 2090 ); 2091 } 2092 2093 //----------------------------inline_min_max----------------------------------- 2094 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2095 set_result(generate_min_max(id, argument(0), argument(1))); 2096 return true; 2097 } 2098 2099 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2100 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) ); 2101 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2102 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check)); 2103 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) ); 2104 2105 { 2106 PreserveJVMState pjvms(this); 2107 PreserveReexecuteState preexecs(this); 2108 jvms()->set_should_reexecute(true); 2109 2110 set_control(slow_path); 2111 set_i_o(i_o()); 2112 2113 uncommon_trap(Deoptimization::Reason_intrinsic, 2114 Deoptimization::Action_none); 2115 } 2116 2117 set_control(fast_path); 2118 set_result(math); 2119 } 2120 2121 template <typename OverflowOp> 2122 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2123 typedef typename OverflowOp::MathOp MathOp; 2124 2125 MathOp* mathOp = new(C) MathOp(arg1, arg2); 2126 Node* operation = _gvn.transform( mathOp ); 2127 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) ); 2128 inline_math_mathExact(operation, ofcheck); 2129 return true; 2130 } 2131 2132 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2133 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2134 } 2135 2136 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2137 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2138 } 2139 2140 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2141 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2142 } 2143 2144 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2145 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2146 } 2147 2148 bool LibraryCallKit::inline_math_negateExactI() { 2149 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2150 } 2151 2152 bool LibraryCallKit::inline_math_negateExactL() { 2153 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2154 } 2155 2156 bool LibraryCallKit::inline_math_multiplyExactI() { 2157 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2158 } 2159 2160 bool LibraryCallKit::inline_math_multiplyExactL() { 2161 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2162 } 2163 2164 Node* 2165 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2166 // These are the candidate return value: 2167 Node* xvalue = x0; 2168 Node* yvalue = y0; 2169 2170 if (xvalue == yvalue) { 2171 return xvalue; 2172 } 2173 2174 bool want_max = (id == vmIntrinsics::_max); 2175 2176 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2177 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2178 if (txvalue == NULL || tyvalue == NULL) return top(); 2179 // This is not really necessary, but it is consistent with a 2180 // hypothetical MaxINode::Value method: 2181 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2182 2183 // %%% This folding logic should (ideally) be in a different place. 2184 // Some should be inside IfNode, and there to be a more reliable 2185 // transformation of ?: style patterns into cmoves. We also want 2186 // more powerful optimizations around cmove and min/max. 2187 2188 // Try to find a dominating comparison of these guys. 2189 // It can simplify the index computation for Arrays.copyOf 2190 // and similar uses of System.arraycopy. 2191 // First, compute the normalized version of CmpI(x, y). 2192 int cmp_op = Op_CmpI; 2193 Node* xkey = xvalue; 2194 Node* ykey = yvalue; 2195 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey)); 2196 if (ideal_cmpxy->is_Cmp()) { 2197 // E.g., if we have CmpI(length - offset, count), 2198 // it might idealize to CmpI(length, count + offset) 2199 cmp_op = ideal_cmpxy->Opcode(); 2200 xkey = ideal_cmpxy->in(1); 2201 ykey = ideal_cmpxy->in(2); 2202 } 2203 2204 // Start by locating any relevant comparisons. 2205 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2206 Node* cmpxy = NULL; 2207 Node* cmpyx = NULL; 2208 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2209 Node* cmp = start_from->fast_out(k); 2210 if (cmp->outcnt() > 0 && // must have prior uses 2211 cmp->in(0) == NULL && // must be context-independent 2212 cmp->Opcode() == cmp_op) { // right kind of compare 2213 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2214 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2215 } 2216 } 2217 2218 const int NCMPS = 2; 2219 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2220 int cmpn; 2221 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2222 if (cmps[cmpn] != NULL) break; // find a result 2223 } 2224 if (cmpn < NCMPS) { 2225 // Look for a dominating test that tells us the min and max. 2226 int depth = 0; // Limit search depth for speed 2227 Node* dom = control(); 2228 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2229 if (++depth >= 100) break; 2230 Node* ifproj = dom; 2231 if (!ifproj->is_Proj()) continue; 2232 Node* iff = ifproj->in(0); 2233 if (!iff->is_If()) continue; 2234 Node* bol = iff->in(1); 2235 if (!bol->is_Bool()) continue; 2236 Node* cmp = bol->in(1); 2237 if (cmp == NULL) continue; 2238 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2239 if (cmps[cmpn] == cmp) break; 2240 if (cmpn == NCMPS) continue; 2241 BoolTest::mask btest = bol->as_Bool()->_test._test; 2242 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2243 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2244 // At this point, we know that 'x btest y' is true. 2245 switch (btest) { 2246 case BoolTest::eq: 2247 // They are proven equal, so we can collapse the min/max. 2248 // Either value is the answer. Choose the simpler. 2249 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2250 return yvalue; 2251 return xvalue; 2252 case BoolTest::lt: // x < y 2253 case BoolTest::le: // x <= y 2254 return (want_max ? yvalue : xvalue); 2255 case BoolTest::gt: // x > y 2256 case BoolTest::ge: // x >= y 2257 return (want_max ? xvalue : yvalue); 2258 } 2259 } 2260 } 2261 2262 // We failed to find a dominating test. 2263 // Let's pick a test that might GVN with prior tests. 2264 Node* best_bol = NULL; 2265 BoolTest::mask best_btest = BoolTest::illegal; 2266 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2267 Node* cmp = cmps[cmpn]; 2268 if (cmp == NULL) continue; 2269 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2270 Node* bol = cmp->fast_out(j); 2271 if (!bol->is_Bool()) continue; 2272 BoolTest::mask btest = bol->as_Bool()->_test._test; 2273 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2274 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2275 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2276 best_bol = bol->as_Bool(); 2277 best_btest = btest; 2278 } 2279 } 2280 } 2281 2282 Node* answer_if_true = NULL; 2283 Node* answer_if_false = NULL; 2284 switch (best_btest) { 2285 default: 2286 if (cmpxy == NULL) 2287 cmpxy = ideal_cmpxy; 2288 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt)); 2289 // and fall through: 2290 case BoolTest::lt: // x < y 2291 case BoolTest::le: // x <= y 2292 answer_if_true = (want_max ? yvalue : xvalue); 2293 answer_if_false = (want_max ? xvalue : yvalue); 2294 break; 2295 case BoolTest::gt: // x > y 2296 case BoolTest::ge: // x >= y 2297 answer_if_true = (want_max ? xvalue : yvalue); 2298 answer_if_false = (want_max ? yvalue : xvalue); 2299 break; 2300 } 2301 2302 jint hi, lo; 2303 if (want_max) { 2304 // We can sharpen the minimum. 2305 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2306 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2307 } else { 2308 // We can sharpen the maximum. 2309 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2310 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2311 } 2312 2313 // Use a flow-free graph structure, to avoid creating excess control edges 2314 // which could hinder other optimizations. 2315 // Since Math.min/max is often used with arraycopy, we want 2316 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2317 Node* cmov = CMoveNode::make(C, NULL, best_bol, 2318 answer_if_false, answer_if_true, 2319 TypeInt::make(lo, hi, widen)); 2320 2321 return _gvn.transform(cmov); 2322 2323 /* 2324 // This is not as desirable as it may seem, since Min and Max 2325 // nodes do not have a full set of optimizations. 2326 // And they would interfere, anyway, with 'if' optimizations 2327 // and with CMoveI canonical forms. 2328 switch (id) { 2329 case vmIntrinsics::_min: 2330 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2331 case vmIntrinsics::_max: 2332 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2333 default: 2334 ShouldNotReachHere(); 2335 } 2336 */ 2337 } 2338 2339 inline int 2340 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2341 const TypePtr* base_type = TypePtr::NULL_PTR; 2342 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2343 if (base_type == NULL) { 2344 // Unknown type. 2345 return Type::AnyPtr; 2346 } else if (base_type == TypePtr::NULL_PTR) { 2347 // Since this is a NULL+long form, we have to switch to a rawptr. 2348 base = _gvn.transform(new (C) CastX2PNode(offset)); 2349 offset = MakeConX(0); 2350 return Type::RawPtr; 2351 } else if (base_type->base() == Type::RawPtr) { 2352 return Type::RawPtr; 2353 } else if (base_type->isa_oopptr()) { 2354 // Base is never null => always a heap address. 2355 if (base_type->ptr() == TypePtr::NotNull) { 2356 return Type::OopPtr; 2357 } 2358 // Offset is small => always a heap address. 2359 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2360 if (offset_type != NULL && 2361 base_type->offset() == 0 && // (should always be?) 2362 offset_type->_lo >= 0 && 2363 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2364 return Type::OopPtr; 2365 } 2366 // Otherwise, it might either be oop+off or NULL+addr. 2367 return Type::AnyPtr; 2368 } else { 2369 // No information: 2370 return Type::AnyPtr; 2371 } 2372 } 2373 2374 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2375 int kind = classify_unsafe_addr(base, offset); 2376 if (kind == Type::RawPtr) { 2377 return basic_plus_adr(top(), base, offset); 2378 } else { 2379 return basic_plus_adr(base, offset); 2380 } 2381 } 2382 2383 //--------------------------inline_number_methods----------------------------- 2384 // inline int Integer.numberOfLeadingZeros(int) 2385 // inline int Long.numberOfLeadingZeros(long) 2386 // 2387 // inline int Integer.numberOfTrailingZeros(int) 2388 // inline int Long.numberOfTrailingZeros(long) 2389 // 2390 // inline int Integer.bitCount(int) 2391 // inline int Long.bitCount(long) 2392 // 2393 // inline char Character.reverseBytes(char) 2394 // inline short Short.reverseBytes(short) 2395 // inline int Integer.reverseBytes(int) 2396 // inline long Long.reverseBytes(long) 2397 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2398 Node* arg = argument(0); 2399 Node* n = NULL; 2400 switch (id) { 2401 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break; 2402 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break; 2403 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break; 2404 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break; 2405 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break; 2406 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break; 2407 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break; 2408 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break; 2409 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break; 2410 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break; 2411 default: fatal_unexpected_iid(id); break; 2412 } 2413 set_result(_gvn.transform(n)); 2414 return true; 2415 } 2416 2417 //----------------------------inline_unsafe_access---------------------------- 2418 2419 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2420 2421 // Helper that guards and inserts a pre-barrier. 2422 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2423 Node* pre_val, bool need_mem_bar) { 2424 // We could be accessing the referent field of a reference object. If so, when G1 2425 // is enabled, we need to log the value in the referent field in an SATB buffer. 2426 // This routine performs some compile time filters and generates suitable 2427 // runtime filters that guard the pre-barrier code. 2428 // Also add memory barrier for non volatile load from the referent field 2429 // to prevent commoning of loads across safepoint. 2430 if (!(UseG1GC || UseShenandoahGC) && !need_mem_bar) 2431 return; 2432 2433 // Some compile time checks. 2434 2435 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2436 const TypeX* otype = offset->find_intptr_t_type(); 2437 if (otype != NULL && otype->is_con() && 2438 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2439 // Constant offset but not the reference_offset so just return 2440 return; 2441 } 2442 2443 // We only need to generate the runtime guards for instances. 2444 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2445 if (btype != NULL) { 2446 if (btype->isa_aryptr()) { 2447 // Array type so nothing to do 2448 return; 2449 } 2450 2451 const TypeInstPtr* itype = btype->isa_instptr(); 2452 if (itype != NULL) { 2453 // Can the klass of base_oop be statically determined to be 2454 // _not_ a sub-class of Reference and _not_ Object? 2455 ciKlass* klass = itype->klass(); 2456 if ( klass->is_loaded() && 2457 !klass->is_subtype_of(env()->Reference_klass()) && 2458 !env()->Object_klass()->is_subtype_of(klass)) { 2459 return; 2460 } 2461 } 2462 } 2463 2464 // The compile time filters did not reject base_oop/offset so 2465 // we need to generate the following runtime filters 2466 // 2467 // if (offset == java_lang_ref_Reference::_reference_offset) { 2468 // if (instance_of(base, java.lang.ref.Reference)) { 2469 // pre_barrier(_, pre_val, ...); 2470 // } 2471 // } 2472 2473 float likely = PROB_LIKELY( 0.999); 2474 float unlikely = PROB_UNLIKELY(0.999); 2475 2476 IdealKit ideal(this); 2477 #define __ ideal. 2478 2479 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2480 2481 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2482 // Update graphKit memory and control from IdealKit. 2483 sync_kit(ideal); 2484 2485 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2486 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2487 2488 // Update IdealKit memory and control from graphKit. 2489 __ sync_kit(this); 2490 2491 Node* one = __ ConI(1); 2492 // is_instof == 0 if base_oop == NULL 2493 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2494 2495 // Update graphKit from IdeakKit. 2496 sync_kit(ideal); 2497 2498 // Use the pre-barrier to record the value in the referent field 2499 pre_barrier(false /* do_load */, 2500 __ ctrl(), 2501 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2502 pre_val /* pre_val */, 2503 T_OBJECT); 2504 if (need_mem_bar) { 2505 // Add memory barrier to prevent commoning reads from this field 2506 // across safepoint since GC can change its value. 2507 insert_mem_bar(Op_MemBarCPUOrder); 2508 } 2509 // Update IdealKit from graphKit. 2510 __ sync_kit(this); 2511 2512 } __ end_if(); // _ref_type != ref_none 2513 } __ end_if(); // offset == referent_offset 2514 2515 // Final sync IdealKit and GraphKit. 2516 final_sync(ideal); 2517 #undef __ 2518 } 2519 2520 2521 // Interpret Unsafe.fieldOffset cookies correctly: 2522 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2523 2524 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2525 // Attempt to infer a sharper value type from the offset and base type. 2526 ciKlass* sharpened_klass = NULL; 2527 2528 // See if it is an instance field, with an object type. 2529 if (alias_type->field() != NULL) { 2530 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2531 if (alias_type->field()->type()->is_klass()) { 2532 sharpened_klass = alias_type->field()->type()->as_klass(); 2533 } 2534 } 2535 2536 // See if it is a narrow oop array. 2537 if (adr_type->isa_aryptr()) { 2538 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2539 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2540 if (elem_type != NULL) { 2541 sharpened_klass = elem_type->klass(); 2542 } 2543 } 2544 } 2545 2546 // The sharpened class might be unloaded if there is no class loader 2547 // contraint in place. 2548 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2549 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2550 2551 #ifndef PRODUCT 2552 if (C->print_intrinsics() || C->print_inlining()) { 2553 tty->print(" from base type: "); adr_type->dump(); tty->cr(); 2554 tty->print(" sharpened value: "); tjp->dump(); tty->cr(); 2555 } 2556 #endif 2557 // Sharpen the value type. 2558 return tjp; 2559 } 2560 return NULL; 2561 } 2562 2563 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) { 2564 if (callee()->is_static()) return false; // caller must have the capability! 2565 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type"); 2566 2567 #ifndef PRODUCT 2568 { 2569 ResourceMark rm; 2570 // Check the signatures. 2571 ciSignature* sig = callee()->signature(); 2572 #ifdef ASSERT 2573 if (!is_store) { 2574 // Object getObject(Object base, int/long offset), etc. 2575 BasicType rtype = sig->return_type()->basic_type(); 2576 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2577 rtype = T_ADDRESS; // it is really a C void* 2578 assert(rtype == type, "getter must return the expected value"); 2579 if (!is_native_ptr) { 2580 assert(sig->count() == 2, "oop getter has 2 arguments"); 2581 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2582 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2583 } else { 2584 assert(sig->count() == 1, "native getter has 1 argument"); 2585 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2586 } 2587 } else { 2588 // void putObject(Object base, int/long offset, Object x), etc. 2589 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2590 if (!is_native_ptr) { 2591 assert(sig->count() == 3, "oop putter has 3 arguments"); 2592 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2593 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2594 } else { 2595 assert(sig->count() == 2, "native putter has 2 arguments"); 2596 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2597 } 2598 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2599 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2600 vtype = T_ADDRESS; // it is really a C void* 2601 assert(vtype == type, "putter must accept the expected value"); 2602 } 2603 #endif // ASSERT 2604 } 2605 #endif //PRODUCT 2606 2607 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2608 2609 Node* receiver = argument(0); // type: oop 2610 2611 // Build address expression. See the code in inline_unsafe_prefetch. 2612 Node* adr; 2613 Node* heap_base_oop = top(); 2614 Node* offset = top(); 2615 Node* val; 2616 2617 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2618 Node* base = argument(1); // type: oop 2619 2620 if (!is_native_ptr) { 2621 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2622 offset = argument(2); // type: long 2623 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2624 // to be plain byte offsets, which are also the same as those accepted 2625 // by oopDesc::field_base. 2626 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2627 "fieldOffset must be byte-scaled"); 2628 // 32-bit machines ignore the high half! 2629 offset = ConvL2X(offset); 2630 adr = make_unsafe_address(base, offset); 2631 heap_base_oop = base; 2632 val = is_store ? argument(4) : NULL; 2633 } else { 2634 Node* ptr = argument(1); // type: long 2635 ptr = ConvL2X(ptr); // adjust Java long to machine word 2636 adr = make_unsafe_address(NULL, ptr); 2637 val = is_store ? argument(3) : NULL; 2638 } 2639 2640 if ((_gvn.type(base)->isa_ptr() == TypePtr::NULL_PTR) && type == T_OBJECT) { 2641 return false; // off-heap oop accesses are not supported 2642 } 2643 2644 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2645 2646 // Try to categorize the address. 2647 Compile::AliasType* alias_type = C->alias_type(adr_type); 2648 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2649 2650 if (alias_type->adr_type() == TypeInstPtr::KLASS || 2651 alias_type->adr_type() == TypeAryPtr::RANGE) { 2652 return false; // not supported 2653 } 2654 2655 bool mismatched = false; 2656 BasicType bt = alias_type->basic_type(); 2657 if (bt != T_ILLEGAL) { 2658 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access"); 2659 if (bt == T_BYTE && adr_type->isa_aryptr()) { 2660 // Alias type doesn't differentiate between byte[] and boolean[]). 2661 // Use address type to get the element type. 2662 bt = adr_type->is_aryptr()->elem()->array_element_basic_type(); 2663 } 2664 if (bt == T_ARRAY || bt == T_NARROWOOP) { 2665 // accessing an array field with getObject is not a mismatch 2666 bt = T_OBJECT; 2667 } 2668 if ((bt == T_OBJECT) != (type == T_OBJECT)) { 2669 // Don't intrinsify mismatched object accesses 2670 return false; 2671 } 2672 mismatched = (bt != type); 2673 } 2674 2675 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched"); 2676 2677 // First guess at the value type. 2678 const Type *value_type = Type::get_const_basic_type(type); 2679 2680 // We will need memory barriers unless we can determine a unique 2681 // alias category for this reference. (Note: If for some reason 2682 // the barriers get omitted and the unsafe reference begins to "pollute" 2683 // the alias analysis of the rest of the graph, either Compile::can_alias 2684 // or Compile::must_alias will throw a diagnostic assert.) 2685 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2686 2687 #if INCLUDE_ALL_GCS 2688 // Work around JDK-8220714 bug. This is done for Shenandoah only, until 2689 // the shared code fix is upstreamed and properly tested there. 2690 if (UseShenandoahGC) { 2691 need_mem_bar |= is_native_ptr; 2692 } 2693 #endif 2694 2695 // If we are reading the value of the referent field of a Reference 2696 // object (either by using Unsafe directly or through reflection) 2697 // then, if G1 is enabled, we need to record the referent in an 2698 // SATB log buffer using the pre-barrier mechanism. 2699 // Also we need to add memory barrier to prevent commoning reads 2700 // from this field across safepoint since GC can change its value. 2701 bool need_read_barrier = !is_native_ptr && !is_store && 2702 offset != top() && heap_base_oop != top(); 2703 2704 if (!is_store && type == T_OBJECT) { 2705 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2706 if (tjp != NULL) { 2707 value_type = tjp; 2708 } 2709 } 2710 2711 receiver = null_check(receiver); 2712 if (stopped()) { 2713 return true; 2714 } 2715 // Heap pointers get a null-check from the interpreter, 2716 // as a courtesy. However, this is not guaranteed by Unsafe, 2717 // and it is not possible to fully distinguish unintended nulls 2718 // from intended ones in this API. 2719 2720 Node* load = NULL; 2721 Node* store = NULL; 2722 Node* leading_membar = NULL; 2723 if (is_volatile) { 2724 // We need to emit leading and trailing CPU membars (see below) in 2725 // addition to memory membars when is_volatile. This is a little 2726 // too strong, but avoids the need to insert per-alias-type 2727 // volatile membars (for stores; compare Parse::do_put_xxx), which 2728 // we cannot do effectively here because we probably only have a 2729 // rough approximation of type. 2730 need_mem_bar = true; 2731 // For Stores, place a memory ordering barrier now. 2732 if (is_store) { 2733 leading_membar = insert_mem_bar(Op_MemBarRelease); 2734 } else { 2735 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2736 leading_membar = insert_mem_bar(Op_MemBarVolatile); 2737 } 2738 } 2739 } 2740 2741 // Memory barrier to prevent normal and 'unsafe' accesses from 2742 // bypassing each other. Happens after null checks, so the 2743 // exception paths do not take memory state from the memory barrier, 2744 // so there's no problems making a strong assert about mixing users 2745 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2746 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2747 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2748 2749 if (!is_store) { 2750 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2751 // To be valid, unsafe loads may depend on other conditions than 2752 // the one that guards them: pin the Load node 2753 load = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched); 2754 // load value 2755 switch (type) { 2756 case T_BOOLEAN: 2757 case T_CHAR: 2758 case T_BYTE: 2759 case T_SHORT: 2760 case T_INT: 2761 case T_LONG: 2762 case T_FLOAT: 2763 case T_DOUBLE: 2764 break; 2765 case T_OBJECT: 2766 if (need_read_barrier) { 2767 insert_pre_barrier(heap_base_oop, offset, load, !(is_volatile || need_mem_bar)); 2768 } 2769 break; 2770 case T_ADDRESS: 2771 // Cast to an int type. 2772 load = _gvn.transform(new (C) CastP2XNode(NULL, load)); 2773 load = ConvX2UL(load); 2774 break; 2775 default: 2776 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2777 break; 2778 } 2779 Node* ld = load; 2780 #if INCLUDE_ALL_GCS 2781 if (UseShenandoahGC && (type == T_OBJECT || type == T_ARRAY)) { 2782 ld = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, ld); 2783 } 2784 #endif 2785 // The load node has the control of the preceding MemBarCPUOrder. All 2786 // following nodes will have the control of the MemBarCPUOrder inserted at 2787 // the end of this method. So, pushing the load onto the stack at a later 2788 // point is fine. 2789 set_result(ld); 2790 } else { 2791 // place effect of store into memory 2792 switch (type) { 2793 case T_DOUBLE: 2794 val = dstore_rounding(val); 2795 break; 2796 case T_ADDRESS: 2797 // Repackage the long as a pointer. 2798 val = ConvL2X(val); 2799 val = _gvn.transform(new (C) CastX2PNode(val)); 2800 break; 2801 } 2802 2803 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2804 if (type == T_OBJECT ) { 2805 store = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched); 2806 } else { 2807 store = store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched); 2808 } 2809 } 2810 2811 if (is_volatile) { 2812 if (!is_store) { 2813 Node* mb = insert_mem_bar(Op_MemBarAcquire, load); 2814 mb->as_MemBar()->set_trailing_load(); 2815 } else { 2816 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2817 Node* mb = insert_mem_bar(Op_MemBarVolatile, store); 2818 MemBarNode::set_store_pair(leading_membar->as_MemBar(), mb->as_MemBar()); 2819 } 2820 } 2821 } 2822 2823 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2824 2825 return true; 2826 } 2827 2828 //----------------------------inline_unsafe_prefetch---------------------------- 2829 2830 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2831 #ifndef PRODUCT 2832 { 2833 ResourceMark rm; 2834 // Check the signatures. 2835 ciSignature* sig = callee()->signature(); 2836 #ifdef ASSERT 2837 // Object getObject(Object base, int/long offset), etc. 2838 BasicType rtype = sig->return_type()->basic_type(); 2839 if (!is_native_ptr) { 2840 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2841 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2842 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2843 } else { 2844 assert(sig->count() == 1, "native prefetch has 1 argument"); 2845 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2846 } 2847 #endif // ASSERT 2848 } 2849 #endif // !PRODUCT 2850 2851 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2852 2853 const int idx = is_static ? 0 : 1; 2854 if (!is_static) { 2855 null_check_receiver(); 2856 if (stopped()) { 2857 return true; 2858 } 2859 } 2860 2861 // Build address expression. See the code in inline_unsafe_access. 2862 Node *adr; 2863 if (!is_native_ptr) { 2864 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2865 Node* base = argument(idx + 0); // type: oop 2866 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2867 Node* offset = argument(idx + 1); // type: long 2868 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2869 // to be plain byte offsets, which are also the same as those accepted 2870 // by oopDesc::field_base. 2871 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2872 "fieldOffset must be byte-scaled"); 2873 // 32-bit machines ignore the high half! 2874 offset = ConvL2X(offset); 2875 adr = make_unsafe_address(base, offset); 2876 } else { 2877 Node* ptr = argument(idx + 0); // type: long 2878 ptr = ConvL2X(ptr); // adjust Java long to machine word 2879 adr = make_unsafe_address(NULL, ptr); 2880 } 2881 2882 // Generate the read or write prefetch 2883 Node *prefetch; 2884 if (is_store) { 2885 prefetch = new (C) PrefetchWriteNode(i_o(), adr); 2886 } else { 2887 prefetch = new (C) PrefetchReadNode(i_o(), adr); 2888 } 2889 prefetch->init_req(0, control()); 2890 set_i_o(_gvn.transform(prefetch)); 2891 2892 return true; 2893 } 2894 2895 //----------------------------inline_unsafe_load_store---------------------------- 2896 // This method serves a couple of different customers (depending on LoadStoreKind): 2897 // 2898 // LS_cmpxchg: 2899 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2900 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2901 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2902 // 2903 // LS_xadd: 2904 // public int getAndAddInt( Object o, long offset, int delta) 2905 // public long getAndAddLong(Object o, long offset, long delta) 2906 // 2907 // LS_xchg: 2908 // int getAndSet(Object o, long offset, int newValue) 2909 // long getAndSet(Object o, long offset, long newValue) 2910 // Object getAndSet(Object o, long offset, Object newValue) 2911 // 2912 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2913 // This basic scheme here is the same as inline_unsafe_access, but 2914 // differs in enough details that combining them would make the code 2915 // overly confusing. (This is a true fact! I originally combined 2916 // them, but even I was confused by it!) As much code/comments as 2917 // possible are retained from inline_unsafe_access though to make 2918 // the correspondences clearer. - dl 2919 2920 if (callee()->is_static()) return false; // caller must have the capability! 2921 2922 #ifndef PRODUCT 2923 BasicType rtype; 2924 { 2925 ResourceMark rm; 2926 // Check the signatures. 2927 ciSignature* sig = callee()->signature(); 2928 rtype = sig->return_type()->basic_type(); 2929 if (kind == LS_xadd || kind == LS_xchg) { 2930 // Check the signatures. 2931 #ifdef ASSERT 2932 assert(rtype == type, "get and set must return the expected type"); 2933 assert(sig->count() == 3, "get and set has 3 arguments"); 2934 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2935 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2936 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2937 #endif // ASSERT 2938 } else if (kind == LS_cmpxchg) { 2939 // Check the signatures. 2940 #ifdef ASSERT 2941 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2942 assert(sig->count() == 4, "CAS has 4 arguments"); 2943 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2944 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2945 #endif // ASSERT 2946 } else { 2947 ShouldNotReachHere(); 2948 } 2949 } 2950 #endif //PRODUCT 2951 2952 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2953 2954 // Get arguments: 2955 Node* receiver = NULL; 2956 Node* base = NULL; 2957 Node* offset = NULL; 2958 Node* oldval = NULL; 2959 Node* newval = NULL; 2960 if (kind == LS_cmpxchg) { 2961 const bool two_slot_type = type2size[type] == 2; 2962 receiver = argument(0); // type: oop 2963 base = argument(1); // type: oop 2964 offset = argument(2); // type: long 2965 oldval = argument(4); // type: oop, int, or long 2966 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2967 } else if (kind == LS_xadd || kind == LS_xchg){ 2968 receiver = argument(0); // type: oop 2969 base = argument(1); // type: oop 2970 offset = argument(2); // type: long 2971 oldval = NULL; 2972 newval = argument(4); // type: oop, int, or long 2973 } 2974 2975 // Build field offset expression. 2976 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2977 // to be plain byte offsets, which are also the same as those accepted 2978 // by oopDesc::field_base. 2979 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2980 // 32-bit machines ignore the high half of long offsets 2981 offset = ConvL2X(offset); 2982 Node* adr = make_unsafe_address(base, offset); 2983 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2984 2985 Compile::AliasType* alias_type = C->alias_type(adr_type); 2986 BasicType bt = alias_type->basic_type(); 2987 if (bt != T_ILLEGAL && 2988 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) { 2989 // Don't intrinsify mismatched object accesses. 2990 return false; 2991 } 2992 2993 // For CAS, unlike inline_unsafe_access, there seems no point in 2994 // trying to refine types. Just use the coarse types here. 2995 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2996 const Type *value_type = Type::get_const_basic_type(type); 2997 2998 if (kind == LS_xchg && type == T_OBJECT) { 2999 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 3000 if (tjp != NULL) { 3001 value_type = tjp; 3002 } 3003 } 3004 3005 // Null check receiver. 3006 receiver = null_check(receiver); 3007 if (stopped()) { 3008 return true; 3009 } 3010 3011 int alias_idx = C->get_alias_index(adr_type); 3012 3013 // Memory-model-wise, a LoadStore acts like a little synchronized 3014 // block, so needs barriers on each side. These don't translate 3015 // into actual barriers on most machines, but we still need rest of 3016 // compiler to respect ordering. 3017 3018 Node* leading_membar = insert_mem_bar(Op_MemBarRelease); 3019 insert_mem_bar(Op_MemBarCPUOrder); 3020 3021 // 4984716: MemBars must be inserted before this 3022 // memory node in order to avoid a false 3023 // dependency which will confuse the scheduler. 3024 Node *mem = memory(alias_idx); 3025 3026 // For now, we handle only those cases that actually exist: ints, 3027 // longs, and Object. Adding others should be straightforward. 3028 Node* load_store = NULL; 3029 switch(type) { 3030 case T_INT: 3031 if (kind == LS_xadd) { 3032 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type)); 3033 } else if (kind == LS_xchg) { 3034 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type)); 3035 } else if (kind == LS_cmpxchg) { 3036 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval)); 3037 } else { 3038 ShouldNotReachHere(); 3039 } 3040 break; 3041 case T_LONG: 3042 if (kind == LS_xadd) { 3043 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type)); 3044 } else if (kind == LS_xchg) { 3045 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type)); 3046 } else if (kind == LS_cmpxchg) { 3047 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 3048 } else { 3049 ShouldNotReachHere(); 3050 } 3051 break; 3052 case T_OBJECT: 3053 // Transformation of a value which could be NULL pointer (CastPP #NULL) 3054 // could be delayed during Parse (for example, in adjust_map_after_if()). 3055 // Execute transformation here to avoid barrier generation in such case. 3056 if (_gvn.type(newval) == TypePtr::NULL_PTR) 3057 newval = _gvn.makecon(TypePtr::NULL_PTR); 3058 3059 // Reference stores need a store barrier. 3060 if (kind == LS_xchg) { 3061 // If pre-barrier must execute before the oop store, old value will require do_load here. 3062 if (!can_move_pre_barrier()) { 3063 pre_barrier(true /* do_load*/, 3064 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 3065 NULL /* pre_val*/, 3066 T_OBJECT); 3067 } // Else move pre_barrier to use load_store value, see below. 3068 } else if (kind == LS_cmpxchg) { 3069 // Same as for newval above: 3070 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 3071 oldval = _gvn.makecon(TypePtr::NULL_PTR); 3072 } 3073 // The only known value which might get overwritten is oldval. 3074 pre_barrier(false /* do_load */, 3075 control(), NULL, NULL, max_juint, NULL, NULL, 3076 oldval /* pre_val */, 3077 T_OBJECT); 3078 } else { 3079 ShouldNotReachHere(); 3080 } 3081 3082 #ifdef _LP64 3083 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3084 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 3085 if (kind == LS_xchg) { 3086 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr, 3087 newval_enc, adr_type, value_type->make_narrowoop())); 3088 } else { 3089 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 3090 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 3091 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr, 3092 newval_enc, oldval_enc)); 3093 } 3094 } else 3095 #endif 3096 { 3097 if (kind == LS_xchg) { 3098 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 3099 } else { 3100 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 3101 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 3102 } 3103 } 3104 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 3105 break; 3106 default: 3107 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 3108 break; 3109 } 3110 3111 // SCMemProjNodes represent the memory state of a LoadStore. Their 3112 // main role is to prevent LoadStore nodes from being optimized away 3113 // when their results aren't used. 3114 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store)); 3115 set_memory(proj, alias_idx); 3116 3117 Node* access = load_store; 3118 3119 if (type == T_OBJECT && kind == LS_xchg) { 3120 #ifdef _LP64 3121 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3122 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type())); 3123 } 3124 #endif 3125 #if INCLUDE_ALL_GCS 3126 if (UseShenandoahGC) { 3127 load_store = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, load_store); 3128 } 3129 #endif 3130 if (can_move_pre_barrier()) { 3131 // Don't need to load pre_val. The old value is returned by load_store. 3132 // The pre_barrier can execute after the xchg as long as no safepoint 3133 // gets inserted between them. 3134 pre_barrier(false /* do_load */, 3135 control(), NULL, NULL, max_juint, NULL, NULL, 3136 load_store /* pre_val */, 3137 T_OBJECT); 3138 } 3139 } 3140 3141 // Add the trailing membar surrounding the access 3142 insert_mem_bar(Op_MemBarCPUOrder); 3143 Node* mb = insert_mem_bar(Op_MemBarAcquire, access); 3144 MemBarNode::set_load_store_pair(leading_membar->as_MemBar(), mb->as_MemBar()); 3145 3146 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3147 set_result(load_store); 3148 return true; 3149 } 3150 3151 //----------------------------inline_unsafe_ordered_store---------------------- 3152 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3153 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3154 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3155 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3156 // This is another variant of inline_unsafe_access, differing in 3157 // that it always issues store-store ("release") barrier and ensures 3158 // store-atomicity (which only matters for "long"). 3159 3160 if (callee()->is_static()) return false; // caller must have the capability! 3161 3162 #ifndef PRODUCT 3163 { 3164 ResourceMark rm; 3165 // Check the signatures. 3166 ciSignature* sig = callee()->signature(); 3167 #ifdef ASSERT 3168 BasicType rtype = sig->return_type()->basic_type(); 3169 assert(rtype == T_VOID, "must return void"); 3170 assert(sig->count() == 3, "has 3 arguments"); 3171 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3172 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3173 #endif // ASSERT 3174 } 3175 #endif //PRODUCT 3176 3177 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3178 3179 // Get arguments: 3180 Node* receiver = argument(0); // type: oop 3181 Node* base = argument(1); // type: oop 3182 Node* offset = argument(2); // type: long 3183 Node* val = argument(4); // type: oop, int, or long 3184 3185 // Null check receiver. 3186 receiver = null_check(receiver); 3187 if (stopped()) { 3188 return true; 3189 } 3190 3191 // Build field offset expression. 3192 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3193 // 32-bit machines ignore the high half of long offsets 3194 offset = ConvL2X(offset); 3195 Node* adr = make_unsafe_address(base, offset); 3196 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3197 const Type *value_type = Type::get_const_basic_type(type); 3198 Compile::AliasType* alias_type = C->alias_type(adr_type); 3199 3200 insert_mem_bar(Op_MemBarRelease); 3201 insert_mem_bar(Op_MemBarCPUOrder); 3202 // Ensure that the store is atomic for longs: 3203 const bool require_atomic_access = true; 3204 Node* store; 3205 if (type == T_OBJECT) // reference stores need a store barrier. 3206 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3207 else { 3208 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3209 } 3210 insert_mem_bar(Op_MemBarCPUOrder); 3211 return true; 3212 } 3213 3214 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3215 // Regardless of form, don't allow previous ld/st to move down, 3216 // then issue acquire, release, or volatile mem_bar. 3217 insert_mem_bar(Op_MemBarCPUOrder); 3218 switch(id) { 3219 case vmIntrinsics::_loadFence: 3220 insert_mem_bar(Op_LoadFence); 3221 return true; 3222 case vmIntrinsics::_storeFence: 3223 insert_mem_bar(Op_StoreFence); 3224 return true; 3225 case vmIntrinsics::_fullFence: 3226 insert_mem_bar(Op_MemBarVolatile); 3227 return true; 3228 default: 3229 fatal_unexpected_iid(id); 3230 return false; 3231 } 3232 } 3233 3234 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3235 if (!kls->is_Con()) { 3236 return true; 3237 } 3238 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3239 if (klsptr == NULL) { 3240 return true; 3241 } 3242 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3243 // don't need a guard for a klass that is already initialized 3244 return !ik->is_initialized(); 3245 } 3246 3247 //----------------------------inline_unsafe_allocate--------------------------- 3248 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3249 bool LibraryCallKit::inline_unsafe_allocate() { 3250 if (callee()->is_static()) return false; // caller must have the capability! 3251 3252 null_check_receiver(); // null-check, then ignore 3253 Node* cls = null_check(argument(1)); 3254 if (stopped()) return true; 3255 3256 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3257 kls = null_check(kls); 3258 if (stopped()) return true; // argument was like int.class 3259 3260 Node* test = NULL; 3261 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3262 // Note: The argument might still be an illegal value like 3263 // Serializable.class or Object[].class. The runtime will handle it. 3264 // But we must make an explicit check for initialization. 3265 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3266 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3267 // can generate code to load it as unsigned byte. 3268 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3269 Node* bits = intcon(InstanceKlass::fully_initialized); 3270 test = _gvn.transform(new (C) SubINode(inst, bits)); 3271 // The 'test' is non-zero if we need to take a slow path. 3272 } 3273 3274 Node* obj = new_instance(kls, test); 3275 set_result(obj); 3276 return true; 3277 } 3278 3279 #ifdef TRACE_HAVE_INTRINSICS 3280 /* 3281 * oop -> myklass 3282 * myklass->trace_id |= USED 3283 * return myklass->trace_id & ~0x3 3284 */ 3285 bool LibraryCallKit::inline_native_classID() { 3286 null_check_receiver(); // null-check, then ignore 3287 Node* cls = null_check(argument(1), T_OBJECT); 3288 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3289 kls = null_check(kls, T_OBJECT); 3290 ByteSize offset = TRACE_ID_OFFSET; 3291 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3292 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3293 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3294 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits)); 3295 Node* clsused = longcon(0x01l); // set the class bit 3296 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused)); 3297 3298 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3299 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3300 set_result(andl); 3301 return true; 3302 } 3303 3304 bool LibraryCallKit::inline_native_threadID() { 3305 Node* tls_ptr = NULL; 3306 Node* cur_thr = generate_current_thread(tls_ptr); 3307 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3308 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3309 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3310 3311 Node* threadid = NULL; 3312 size_t thread_id_size = OSThread::thread_id_size(); 3313 if (thread_id_size == (size_t) BytesPerLong) { 3314 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3315 } else if (thread_id_size == (size_t) BytesPerInt) { 3316 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3317 } else { 3318 ShouldNotReachHere(); 3319 } 3320 set_result(threadid); 3321 return true; 3322 } 3323 #endif 3324 3325 //------------------------inline_native_time_funcs-------------- 3326 // inline code for System.currentTimeMillis() and System.nanoTime() 3327 // these have the same type and signature 3328 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3329 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3330 const TypePtr* no_memory_effects = NULL; 3331 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3332 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0)); 3333 #ifdef ASSERT 3334 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1)); 3335 assert(value_top == top(), "second value must be top"); 3336 #endif 3337 set_result(value); 3338 return true; 3339 } 3340 3341 //------------------------inline_native_currentThread------------------ 3342 bool LibraryCallKit::inline_native_currentThread() { 3343 Node* junk = NULL; 3344 set_result(generate_current_thread(junk)); 3345 return true; 3346 } 3347 3348 //------------------------inline_native_isInterrupted------------------ 3349 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3350 bool LibraryCallKit::inline_native_isInterrupted() { 3351 // Add a fast path to t.isInterrupted(clear_int): 3352 // (t == Thread.current() && 3353 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3354 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3355 // So, in the common case that the interrupt bit is false, 3356 // we avoid making a call into the VM. Even if the interrupt bit 3357 // is true, if the clear_int argument is false, we avoid the VM call. 3358 // However, if the receiver is not currentThread, we must call the VM, 3359 // because there must be some locking done around the operation. 3360 3361 // We only go to the fast case code if we pass two guards. 3362 // Paths which do not pass are accumulated in the slow_region. 3363 3364 enum { 3365 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3366 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3367 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3368 PATH_LIMIT 3369 }; 3370 3371 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3372 // out of the function. 3373 insert_mem_bar(Op_MemBarCPUOrder); 3374 3375 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT); 3376 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL); 3377 3378 RegionNode* slow_region = new (C) RegionNode(1); 3379 record_for_igvn(slow_region); 3380 3381 // (a) Receiving thread must be the current thread. 3382 Node* rec_thr = argument(0); 3383 Node* tls_ptr = NULL; 3384 Node* cur_thr = generate_current_thread(tls_ptr); 3385 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr)); 3386 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne)); 3387 3388 generate_slow_guard(bol_thr, slow_region); 3389 3390 // (b) Interrupt bit on TLS must be false. 3391 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3392 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3393 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3394 3395 // Set the control input on the field _interrupted read to prevent it floating up. 3396 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3397 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0))); 3398 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne)); 3399 3400 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3401 3402 // First fast path: if (!TLS._interrupted) return false; 3403 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit)); 3404 result_rgn->init_req(no_int_result_path, false_bit); 3405 result_val->init_req(no_int_result_path, intcon(0)); 3406 3407 // drop through to next case 3408 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit))); 3409 3410 #ifndef TARGET_OS_FAMILY_windows 3411 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3412 Node* clr_arg = argument(1); 3413 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0))); 3414 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne)); 3415 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3416 3417 // Second fast path: ... else if (!clear_int) return true; 3418 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg)); 3419 result_rgn->init_req(no_clear_result_path, false_arg); 3420 result_val->init_req(no_clear_result_path, intcon(1)); 3421 3422 // drop through to next case 3423 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg))); 3424 #else 3425 // To return true on Windows you must read the _interrupted field 3426 // and check the the event state i.e. take the slow path. 3427 #endif // TARGET_OS_FAMILY_windows 3428 3429 // (d) Otherwise, go to the slow path. 3430 slow_region->add_req(control()); 3431 set_control( _gvn.transform(slow_region)); 3432 3433 if (stopped()) { 3434 // There is no slow path. 3435 result_rgn->init_req(slow_result_path, top()); 3436 result_val->init_req(slow_result_path, top()); 3437 } else { 3438 // non-virtual because it is a private non-static 3439 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3440 3441 Node* slow_val = set_results_for_java_call(slow_call); 3442 // this->control() comes from set_results_for_java_call 3443 3444 Node* fast_io = slow_call->in(TypeFunc::I_O); 3445 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3446 3447 // These two phis are pre-filled with copies of of the fast IO and Memory 3448 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3449 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3450 3451 result_rgn->init_req(slow_result_path, control()); 3452 result_io ->init_req(slow_result_path, i_o()); 3453 result_mem->init_req(slow_result_path, reset_memory()); 3454 result_val->init_req(slow_result_path, slow_val); 3455 3456 set_all_memory(_gvn.transform(result_mem)); 3457 set_i_o( _gvn.transform(result_io)); 3458 } 3459 3460 C->set_has_split_ifs(true); // Has chance for split-if optimization 3461 set_result(result_rgn, result_val); 3462 return true; 3463 } 3464 3465 //---------------------------load_mirror_from_klass---------------------------- 3466 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3467 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3468 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3469 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3470 } 3471 3472 //-----------------------load_klass_from_mirror_common------------------------- 3473 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3474 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3475 // and branch to the given path on the region. 3476 // If never_see_null, take an uncommon trap on null, so we can optimistically 3477 // compile for the non-null case. 3478 // If the region is NULL, force never_see_null = true. 3479 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3480 bool never_see_null, 3481 RegionNode* region, 3482 int null_path, 3483 int offset) { 3484 if (region == NULL) never_see_null = true; 3485 Node* p = basic_plus_adr(mirror, offset); 3486 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3487 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3488 Node* null_ctl = top(); 3489 kls = null_check_oop(kls, &null_ctl, never_see_null); 3490 if (region != NULL) { 3491 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3492 region->init_req(null_path, null_ctl); 3493 } else { 3494 assert(null_ctl == top(), "no loose ends"); 3495 } 3496 return kls; 3497 } 3498 3499 //--------------------(inline_native_Class_query helpers)--------------------- 3500 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3501 // Fall through if (mods & mask) == bits, take the guard otherwise. 3502 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3503 // Branch around if the given klass has the given modifier bit set. 3504 // Like generate_guard, adds a new path onto the region. 3505 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3506 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3507 Node* mask = intcon(modifier_mask); 3508 Node* bits = intcon(modifier_bits); 3509 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask)); 3510 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits)); 3511 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne)); 3512 return generate_fair_guard(bol, region); 3513 } 3514 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3515 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3516 } 3517 3518 //-------------------------inline_native_Class_query------------------- 3519 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3520 const Type* return_type = TypeInt::BOOL; 3521 Node* prim_return_value = top(); // what happens if it's a primitive class? 3522 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3523 bool expect_prim = false; // most of these guys expect to work on refs 3524 3525 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3526 3527 Node* mirror = argument(0); 3528 Node* obj = top(); 3529 3530 switch (id) { 3531 case vmIntrinsics::_isInstance: 3532 // nothing is an instance of a primitive type 3533 prim_return_value = intcon(0); 3534 obj = argument(1); 3535 break; 3536 case vmIntrinsics::_getModifiers: 3537 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3538 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3539 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3540 break; 3541 case vmIntrinsics::_isInterface: 3542 prim_return_value = intcon(0); 3543 break; 3544 case vmIntrinsics::_isArray: 3545 prim_return_value = intcon(0); 3546 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3547 break; 3548 case vmIntrinsics::_isPrimitive: 3549 prim_return_value = intcon(1); 3550 expect_prim = true; // obviously 3551 break; 3552 case vmIntrinsics::_getSuperclass: 3553 prim_return_value = null(); 3554 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3555 break; 3556 case vmIntrinsics::_getComponentType: 3557 prim_return_value = null(); 3558 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3559 break; 3560 case vmIntrinsics::_getClassAccessFlags: 3561 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3562 return_type = TypeInt::INT; // not bool! 6297094 3563 break; 3564 default: 3565 fatal_unexpected_iid(id); 3566 break; 3567 } 3568 3569 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3570 if (mirror_con == NULL) return false; // cannot happen? 3571 3572 #ifndef PRODUCT 3573 if (C->print_intrinsics() || C->print_inlining()) { 3574 ciType* k = mirror_con->java_mirror_type(); 3575 if (k) { 3576 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3577 k->print_name(); 3578 tty->cr(); 3579 } 3580 } 3581 #endif 3582 3583 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3584 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3585 record_for_igvn(region); 3586 PhiNode* phi = new (C) PhiNode(region, return_type); 3587 3588 // The mirror will never be null of Reflection.getClassAccessFlags, however 3589 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3590 // if it is. See bug 4774291. 3591 3592 // For Reflection.getClassAccessFlags(), the null check occurs in 3593 // the wrong place; see inline_unsafe_access(), above, for a similar 3594 // situation. 3595 mirror = null_check(mirror); 3596 // If mirror or obj is dead, only null-path is taken. 3597 if (stopped()) return true; 3598 3599 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3600 3601 // Now load the mirror's klass metaobject, and null-check it. 3602 // Side-effects region with the control path if the klass is null. 3603 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3604 // If kls is null, we have a primitive mirror. 3605 phi->init_req(_prim_path, prim_return_value); 3606 if (stopped()) { set_result(region, phi); return true; } 3607 bool safe_for_replace = (region->in(_prim_path) == top()); 3608 3609 Node* p; // handy temp 3610 Node* null_ctl; 3611 3612 // Now that we have the non-null klass, we can perform the real query. 3613 // For constant classes, the query will constant-fold in LoadNode::Value. 3614 Node* query_value = top(); 3615 switch (id) { 3616 case vmIntrinsics::_isInstance: 3617 // nothing is an instance of a primitive type 3618 query_value = gen_instanceof(obj, kls, safe_for_replace); 3619 break; 3620 3621 case vmIntrinsics::_getModifiers: 3622 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3623 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3624 break; 3625 3626 case vmIntrinsics::_isInterface: 3627 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3628 if (generate_interface_guard(kls, region) != NULL) 3629 // A guard was added. If the guard is taken, it was an interface. 3630 phi->add_req(intcon(1)); 3631 // If we fall through, it's a plain class. 3632 query_value = intcon(0); 3633 break; 3634 3635 case vmIntrinsics::_isArray: 3636 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3637 if (generate_array_guard(kls, region) != NULL) 3638 // A guard was added. If the guard is taken, it was an array. 3639 phi->add_req(intcon(1)); 3640 // If we fall through, it's a plain class. 3641 query_value = intcon(0); 3642 break; 3643 3644 case vmIntrinsics::_isPrimitive: 3645 query_value = intcon(0); // "normal" path produces false 3646 break; 3647 3648 case vmIntrinsics::_getSuperclass: 3649 // The rules here are somewhat unfortunate, but we can still do better 3650 // with random logic than with a JNI call. 3651 // Interfaces store null or Object as _super, but must report null. 3652 // Arrays store an intermediate super as _super, but must report Object. 3653 // Other types can report the actual _super. 3654 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3655 if (generate_interface_guard(kls, region) != NULL) 3656 // A guard was added. If the guard is taken, it was an interface. 3657 phi->add_req(null()); 3658 if (generate_array_guard(kls, region) != NULL) 3659 // A guard was added. If the guard is taken, it was an array. 3660 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3661 // If we fall through, it's a plain class. Get its _super. 3662 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3663 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3664 null_ctl = top(); 3665 kls = null_check_oop(kls, &null_ctl); 3666 if (null_ctl != top()) { 3667 // If the guard is taken, Object.superClass is null (both klass and mirror). 3668 region->add_req(null_ctl); 3669 phi ->add_req(null()); 3670 } 3671 if (!stopped()) { 3672 query_value = load_mirror_from_klass(kls); 3673 } 3674 break; 3675 3676 case vmIntrinsics::_getComponentType: 3677 if (generate_array_guard(kls, region) != NULL) { 3678 // Be sure to pin the oop load to the guard edge just created: 3679 Node* is_array_ctrl = region->in(region->req()-1); 3680 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset())); 3681 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3682 phi->add_req(cmo); 3683 } 3684 query_value = null(); // non-array case is null 3685 break; 3686 3687 case vmIntrinsics::_getClassAccessFlags: 3688 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3689 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3690 break; 3691 3692 default: 3693 fatal_unexpected_iid(id); 3694 break; 3695 } 3696 3697 // Fall-through is the normal case of a query to a real class. 3698 phi->init_req(1, query_value); 3699 region->init_req(1, control()); 3700 3701 C->set_has_split_ifs(true); // Has chance for split-if optimization 3702 set_result(region, phi); 3703 return true; 3704 } 3705 3706 //--------------------------inline_native_subtype_check------------------------ 3707 // This intrinsic takes the JNI calls out of the heart of 3708 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3709 bool LibraryCallKit::inline_native_subtype_check() { 3710 // Pull both arguments off the stack. 3711 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3712 args[0] = argument(0); 3713 args[1] = argument(1); 3714 Node* klasses[2]; // corresponding Klasses: superk, subk 3715 klasses[0] = klasses[1] = top(); 3716 3717 enum { 3718 // A full decision tree on {superc is prim, subc is prim}: 3719 _prim_0_path = 1, // {P,N} => false 3720 // {P,P} & superc!=subc => false 3721 _prim_same_path, // {P,P} & superc==subc => true 3722 _prim_1_path, // {N,P} => false 3723 _ref_subtype_path, // {N,N} & subtype check wins => true 3724 _both_ref_path, // {N,N} & subtype check loses => false 3725 PATH_LIMIT 3726 }; 3727 3728 RegionNode* region = new (C) RegionNode(PATH_LIMIT); 3729 Node* phi = new (C) PhiNode(region, TypeInt::BOOL); 3730 record_for_igvn(region); 3731 3732 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3733 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3734 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3735 3736 // First null-check both mirrors and load each mirror's klass metaobject. 3737 int which_arg; 3738 for (which_arg = 0; which_arg <= 1; which_arg++) { 3739 Node* arg = args[which_arg]; 3740 arg = null_check(arg); 3741 if (stopped()) break; 3742 args[which_arg] = arg; 3743 3744 Node* p = basic_plus_adr(arg, class_klass_offset); 3745 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3746 klasses[which_arg] = _gvn.transform(kls); 3747 } 3748 3749 // Having loaded both klasses, test each for null. 3750 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3751 for (which_arg = 0; which_arg <= 1; which_arg++) { 3752 Node* kls = klasses[which_arg]; 3753 Node* null_ctl = top(); 3754 kls = null_check_oop(kls, &null_ctl, never_see_null); 3755 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3756 region->init_req(prim_path, null_ctl); 3757 if (stopped()) break; 3758 klasses[which_arg] = kls; 3759 } 3760 3761 if (!stopped()) { 3762 // now we have two reference types, in klasses[0..1] 3763 Node* subk = klasses[1]; // the argument to isAssignableFrom 3764 Node* superk = klasses[0]; // the receiver 3765 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3766 // now we have a successful reference subtype check 3767 region->set_req(_ref_subtype_path, control()); 3768 } 3769 3770 // If both operands are primitive (both klasses null), then 3771 // we must return true when they are identical primitives. 3772 // It is convenient to test this after the first null klass check. 3773 set_control(region->in(_prim_0_path)); // go back to first null check 3774 if (!stopped()) { 3775 // Since superc is primitive, make a guard for the superc==subc case. 3776 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1])); 3777 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq)); 3778 generate_guard(bol_eq, region, PROB_FAIR); 3779 if (region->req() == PATH_LIMIT+1) { 3780 // A guard was added. If the added guard is taken, superc==subc. 3781 region->swap_edges(PATH_LIMIT, _prim_same_path); 3782 region->del_req(PATH_LIMIT); 3783 } 3784 region->set_req(_prim_0_path, control()); // Not equal after all. 3785 } 3786 3787 // these are the only paths that produce 'true': 3788 phi->set_req(_prim_same_path, intcon(1)); 3789 phi->set_req(_ref_subtype_path, intcon(1)); 3790 3791 // pull together the cases: 3792 assert(region->req() == PATH_LIMIT, "sane region"); 3793 for (uint i = 1; i < region->req(); i++) { 3794 Node* ctl = region->in(i); 3795 if (ctl == NULL || ctl == top()) { 3796 region->set_req(i, top()); 3797 phi ->set_req(i, top()); 3798 } else if (phi->in(i) == NULL) { 3799 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3800 } 3801 } 3802 3803 set_control(_gvn.transform(region)); 3804 set_result(_gvn.transform(phi)); 3805 return true; 3806 } 3807 3808 //---------------------generate_array_guard_common------------------------ 3809 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3810 bool obj_array, bool not_array) { 3811 // If obj_array/non_array==false/false: 3812 // Branch around if the given klass is in fact an array (either obj or prim). 3813 // If obj_array/non_array==false/true: 3814 // Branch around if the given klass is not an array klass of any kind. 3815 // If obj_array/non_array==true/true: 3816 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3817 // If obj_array/non_array==true/false: 3818 // Branch around if the kls is an oop array (Object[] or subtype) 3819 // 3820 // Like generate_guard, adds a new path onto the region. 3821 jint layout_con = 0; 3822 Node* layout_val = get_layout_helper(kls, layout_con); 3823 if (layout_val == NULL) { 3824 bool query = (obj_array 3825 ? Klass::layout_helper_is_objArray(layout_con) 3826 : Klass::layout_helper_is_array(layout_con)); 3827 if (query == not_array) { 3828 return NULL; // never a branch 3829 } else { // always a branch 3830 Node* always_branch = control(); 3831 if (region != NULL) 3832 region->add_req(always_branch); 3833 set_control(top()); 3834 return always_branch; 3835 } 3836 } 3837 // Now test the correct condition. 3838 jint nval = (obj_array 3839 ? (jint)(Klass::_lh_array_tag_type_value 3840 << Klass::_lh_array_tag_shift) 3841 : Klass::_lh_neutral_value); 3842 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval))); 3843 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3844 // invert the test if we are looking for a non-array 3845 if (not_array) btest = BoolTest(btest).negate(); 3846 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest)); 3847 return generate_fair_guard(bol, region); 3848 } 3849 3850 3851 //-----------------------inline_native_newArray-------------------------- 3852 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3853 bool LibraryCallKit::inline_native_newArray() { 3854 Node* mirror = argument(0); 3855 Node* count_val = argument(1); 3856 3857 mirror = null_check(mirror); 3858 // If mirror or obj is dead, only null-path is taken. 3859 if (stopped()) return true; 3860 3861 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3862 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 3863 PhiNode* result_val = new(C) PhiNode(result_reg, 3864 TypeInstPtr::NOTNULL); 3865 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 3866 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 3867 TypePtr::BOTTOM); 3868 3869 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3870 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3871 result_reg, _slow_path); 3872 Node* normal_ctl = control(); 3873 Node* no_array_ctl = result_reg->in(_slow_path); 3874 3875 // Generate code for the slow case. We make a call to newArray(). 3876 set_control(no_array_ctl); 3877 if (!stopped()) { 3878 // Either the input type is void.class, or else the 3879 // array klass has not yet been cached. Either the 3880 // ensuing call will throw an exception, or else it 3881 // will cache the array klass for next time. 3882 PreserveJVMState pjvms(this); 3883 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3884 Node* slow_result = set_results_for_java_call(slow_call); 3885 // this->control() comes from set_results_for_java_call 3886 result_reg->set_req(_slow_path, control()); 3887 result_val->set_req(_slow_path, slow_result); 3888 result_io ->set_req(_slow_path, i_o()); 3889 result_mem->set_req(_slow_path, reset_memory()); 3890 } 3891 3892 set_control(normal_ctl); 3893 if (!stopped()) { 3894 // Normal case: The array type has been cached in the java.lang.Class. 3895 // The following call works fine even if the array type is polymorphic. 3896 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3897 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3898 result_reg->init_req(_normal_path, control()); 3899 result_val->init_req(_normal_path, obj); 3900 result_io ->init_req(_normal_path, i_o()); 3901 result_mem->init_req(_normal_path, reset_memory()); 3902 } 3903 3904 // Return the combined state. 3905 set_i_o( _gvn.transform(result_io) ); 3906 set_all_memory( _gvn.transform(result_mem)); 3907 3908 C->set_has_split_ifs(true); // Has chance for split-if optimization 3909 set_result(result_reg, result_val); 3910 return true; 3911 } 3912 3913 //----------------------inline_native_getLength-------------------------- 3914 // public static native int java.lang.reflect.Array.getLength(Object array); 3915 bool LibraryCallKit::inline_native_getLength() { 3916 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3917 3918 Node* array = null_check(argument(0)); 3919 // If array is dead, only null-path is taken. 3920 if (stopped()) return true; 3921 3922 // Deoptimize if it is a non-array. 3923 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3924 3925 if (non_array != NULL) { 3926 PreserveJVMState pjvms(this); 3927 set_control(non_array); 3928 uncommon_trap(Deoptimization::Reason_intrinsic, 3929 Deoptimization::Action_maybe_recompile); 3930 } 3931 3932 // If control is dead, only non-array-path is taken. 3933 if (stopped()) return true; 3934 3935 // The works fine even if the array type is polymorphic. 3936 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3937 Node* result = load_array_length(array); 3938 3939 C->set_has_split_ifs(true); // Has chance for split-if optimization 3940 set_result(result); 3941 return true; 3942 } 3943 3944 //------------------------inline_array_copyOf---------------------------- 3945 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3946 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3947 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3948 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3949 3950 // Get the arguments. 3951 Node* original = argument(0); 3952 Node* start = is_copyOfRange? argument(1): intcon(0); 3953 Node* end = is_copyOfRange? argument(2): argument(1); 3954 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3955 3956 Node* newcopy = NULL; 3957 3958 // Set the original stack and the reexecute bit for the interpreter to reexecute 3959 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3960 { PreserveReexecuteState preexecs(this); 3961 jvms()->set_should_reexecute(true); 3962 3963 array_type_mirror = null_check(array_type_mirror); 3964 original = null_check(original); 3965 3966 // Check if a null path was taken unconditionally. 3967 if (stopped()) return true; 3968 3969 Node* orig_length = load_array_length(original); 3970 3971 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3972 klass_node = null_check(klass_node); 3973 3974 RegionNode* bailout = new (C) RegionNode(1); 3975 record_for_igvn(bailout); 3976 3977 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3978 // Bail out if that is so. 3979 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3980 if (not_objArray != NULL) { 3981 // Improve the klass node's type from the new optimistic assumption: 3982 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3983 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3984 Node* cast = new (C) CastPPNode(klass_node, akls); 3985 cast->init_req(0, control()); 3986 klass_node = _gvn.transform(cast); 3987 } 3988 3989 // Bail out if either start or end is negative. 3990 generate_negative_guard(start, bailout, &start); 3991 generate_negative_guard(end, bailout, &end); 3992 3993 Node* length = end; 3994 if (_gvn.type(start) != TypeInt::ZERO) { 3995 length = _gvn.transform(new (C) SubINode(end, start)); 3996 } 3997 3998 // Bail out if length is negative. 3999 // Without this the new_array would throw 4000 // NegativeArraySizeException but IllegalArgumentException is what 4001 // should be thrown 4002 generate_negative_guard(length, bailout, &length); 4003 4004 if (bailout->req() > 1) { 4005 PreserveJVMState pjvms(this); 4006 set_control(_gvn.transform(bailout)); 4007 uncommon_trap(Deoptimization::Reason_intrinsic, 4008 Deoptimization::Action_maybe_recompile); 4009 } 4010 4011 if (!stopped()) { 4012 // How many elements will we copy from the original? 4013 // The answer is MinI(orig_length - start, length). 4014 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start)); 4015 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 4016 4017 newcopy = new_array(klass_node, length, 0); // no argments to push 4018 4019 // Generate a direct call to the right arraycopy function(s). 4020 // We know the copy is disjoint but we might not know if the 4021 // oop stores need checking. 4022 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 4023 // This will fail a store-check if x contains any non-nulls. 4024 bool disjoint_bases = true; 4025 // if start > orig_length then the length of the copy may be 4026 // negative. 4027 bool length_never_negative = !is_copyOfRange; 4028 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4029 original, start, newcopy, intcon(0), moved, 4030 disjoint_bases, length_never_negative); 4031 } 4032 } // original reexecute is set back here 4033 4034 C->set_has_split_ifs(true); // Has chance for split-if optimization 4035 if (!stopped()) { 4036 set_result(newcopy); 4037 } 4038 return true; 4039 } 4040 4041 4042 //----------------------generate_virtual_guard--------------------------- 4043 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4044 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4045 RegionNode* slow_region) { 4046 ciMethod* method = callee(); 4047 int vtable_index = method->vtable_index(); 4048 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4049 err_msg_res("bad index %d", vtable_index)); 4050 // Get the Method* out of the appropriate vtable entry. 4051 int entry_offset = (InstanceKlass::vtable_start_offset() + 4052 vtable_index*vtableEntry::size()) * wordSize + 4053 vtableEntry::method_offset_in_bytes(); 4054 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4055 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4056 4057 // Compare the target method with the expected method (e.g., Object.hashCode). 4058 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4059 4060 Node* native_call = makecon(native_call_addr); 4061 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call)); 4062 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne)); 4063 4064 return generate_slow_guard(test_native, slow_region); 4065 } 4066 4067 //-----------------------generate_method_call---------------------------- 4068 // Use generate_method_call to make a slow-call to the real 4069 // method if the fast path fails. An alternative would be to 4070 // use a stub like OptoRuntime::slow_arraycopy_Java. 4071 // This only works for expanding the current library call, 4072 // not another intrinsic. (E.g., don't use this for making an 4073 // arraycopy call inside of the copyOf intrinsic.) 4074 CallJavaNode* 4075 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4076 // When compiling the intrinsic method itself, do not use this technique. 4077 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4078 4079 ciMethod* method = callee(); 4080 // ensure the JVMS we have will be correct for this call 4081 guarantee(method_id == method->intrinsic_id(), "must match"); 4082 4083 const TypeFunc* tf = TypeFunc::make(method); 4084 CallJavaNode* slow_call; 4085 if (is_static) { 4086 assert(!is_virtual, ""); 4087 slow_call = new(C) CallStaticJavaNode(C, tf, 4088 SharedRuntime::get_resolve_static_call_stub(), 4089 method, bci()); 4090 } else if (is_virtual) { 4091 null_check_receiver(); 4092 int vtable_index = Method::invalid_vtable_index; 4093 if (UseInlineCaches) { 4094 // Suppress the vtable call 4095 } else { 4096 // hashCode and clone are not a miranda methods, 4097 // so the vtable index is fixed. 4098 // No need to use the linkResolver to get it. 4099 vtable_index = method->vtable_index(); 4100 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4101 err_msg_res("bad index %d", vtable_index)); 4102 } 4103 slow_call = new(C) CallDynamicJavaNode(tf, 4104 SharedRuntime::get_resolve_virtual_call_stub(), 4105 method, vtable_index, bci()); 4106 } else { // neither virtual nor static: opt_virtual 4107 null_check_receiver(); 4108 slow_call = new(C) CallStaticJavaNode(C, tf, 4109 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4110 method, bci()); 4111 slow_call->set_optimized_virtual(true); 4112 } 4113 set_arguments_for_java_call(slow_call); 4114 set_edges_for_java_call(slow_call); 4115 return slow_call; 4116 } 4117 4118 4119 /** 4120 * Build special case code for calls to hashCode on an object. This call may 4121 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4122 * slightly different code. 4123 */ 4124 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4125 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4126 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4127 4128 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4129 4130 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4131 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT); 4132 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO); 4133 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4134 Node* obj = NULL; 4135 if (!is_static) { 4136 // Check for hashing null object 4137 obj = null_check_receiver(); 4138 if (stopped()) return true; // unconditionally null 4139 result_reg->init_req(_null_path, top()); 4140 result_val->init_req(_null_path, top()); 4141 } else { 4142 // Do a null check, and return zero if null. 4143 // System.identityHashCode(null) == 0 4144 obj = argument(0); 4145 Node* null_ctl = top(); 4146 obj = null_check_oop(obj, &null_ctl); 4147 result_reg->init_req(_null_path, null_ctl); 4148 result_val->init_req(_null_path, _gvn.intcon(0)); 4149 } 4150 4151 // Unconditionally null? Then return right away. 4152 if (stopped()) { 4153 set_control( result_reg->in(_null_path)); 4154 if (!stopped()) 4155 set_result(result_val->in(_null_path)); 4156 return true; 4157 } 4158 4159 // We only go to the fast case code if we pass a number of guards. The 4160 // paths which do not pass are accumulated in the slow_region. 4161 RegionNode* slow_region = new (C) RegionNode(1); 4162 record_for_igvn(slow_region); 4163 4164 // If this is a virtual call, we generate a funny guard. We pull out 4165 // the vtable entry corresponding to hashCode() from the target object. 4166 // If the target method which we are calling happens to be the native 4167 // Object hashCode() method, we pass the guard. We do not need this 4168 // guard for non-virtual calls -- the caller is known to be the native 4169 // Object hashCode(). 4170 if (is_virtual) { 4171 // After null check, get the object's klass. 4172 Node* obj_klass = load_object_klass(obj); 4173 generate_virtual_guard(obj_klass, slow_region); 4174 } 4175 4176 // Get the header out of the object, use LoadMarkNode when available 4177 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4178 // The control of the load must be NULL. Otherwise, the load can move before 4179 // the null check after castPP removal. 4180 Node* no_ctrl = NULL; 4181 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4182 4183 // Test the header to see if it is unlocked. 4184 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4185 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask)); 4186 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4187 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val)); 4188 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne)); 4189 4190 generate_slow_guard(test_unlocked, slow_region); 4191 4192 // Get the hash value and check to see that it has been properly assigned. 4193 // We depend on hash_mask being at most 32 bits and avoid the use of 4194 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4195 // vm: see markOop.hpp. 4196 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4197 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4198 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift)); 4199 // This hack lets the hash bits live anywhere in the mark object now, as long 4200 // as the shift drops the relevant bits into the low 32 bits. Note that 4201 // Java spec says that HashCode is an int so there's no point in capturing 4202 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4203 hshifted_header = ConvX2I(hshifted_header); 4204 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask)); 4205 4206 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4207 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val)); 4208 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq)); 4209 4210 generate_slow_guard(test_assigned, slow_region); 4211 4212 Node* init_mem = reset_memory(); 4213 // fill in the rest of the null path: 4214 result_io ->init_req(_null_path, i_o()); 4215 result_mem->init_req(_null_path, init_mem); 4216 4217 result_val->init_req(_fast_path, hash_val); 4218 result_reg->init_req(_fast_path, control()); 4219 result_io ->init_req(_fast_path, i_o()); 4220 result_mem->init_req(_fast_path, init_mem); 4221 4222 // Generate code for the slow case. We make a call to hashCode(). 4223 set_control(_gvn.transform(slow_region)); 4224 if (!stopped()) { 4225 // No need for PreserveJVMState, because we're using up the present state. 4226 set_all_memory(init_mem); 4227 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4228 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4229 Node* slow_result = set_results_for_java_call(slow_call); 4230 // this->control() comes from set_results_for_java_call 4231 result_reg->init_req(_slow_path, control()); 4232 result_val->init_req(_slow_path, slow_result); 4233 result_io ->set_req(_slow_path, i_o()); 4234 result_mem ->set_req(_slow_path, reset_memory()); 4235 } 4236 4237 // Return the combined state. 4238 set_i_o( _gvn.transform(result_io) ); 4239 set_all_memory( _gvn.transform(result_mem)); 4240 4241 set_result(result_reg, result_val); 4242 return true; 4243 } 4244 4245 //---------------------------inline_native_getClass---------------------------- 4246 // public final native Class<?> java.lang.Object.getClass(); 4247 // 4248 // Build special case code for calls to getClass on an object. 4249 bool LibraryCallKit::inline_native_getClass() { 4250 Node* obj = null_check_receiver(); 4251 if (stopped()) return true; 4252 set_result(load_mirror_from_klass(load_object_klass(obj))); 4253 return true; 4254 } 4255 4256 //-----------------inline_native_Reflection_getCallerClass--------------------- 4257 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4258 // 4259 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4260 // 4261 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4262 // in that it must skip particular security frames and checks for 4263 // caller sensitive methods. 4264 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4265 #ifndef PRODUCT 4266 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4267 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4268 } 4269 #endif 4270 4271 if (!jvms()->has_method()) { 4272 #ifndef PRODUCT 4273 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4274 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4275 } 4276 #endif 4277 return false; 4278 } 4279 4280 // Walk back up the JVM state to find the caller at the required 4281 // depth. 4282 JVMState* caller_jvms = jvms(); 4283 4284 // Cf. JVM_GetCallerClass 4285 // NOTE: Start the loop at depth 1 because the current JVM state does 4286 // not include the Reflection.getCallerClass() frame. 4287 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4288 ciMethod* m = caller_jvms->method(); 4289 switch (n) { 4290 case 0: 4291 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4292 break; 4293 case 1: 4294 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4295 if (!m->caller_sensitive()) { 4296 #ifndef PRODUCT 4297 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4298 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4299 } 4300 #endif 4301 return false; // bail-out; let JVM_GetCallerClass do the work 4302 } 4303 break; 4304 default: 4305 if (!m->is_ignored_by_security_stack_walk()) { 4306 // We have reached the desired frame; return the holder class. 4307 // Acquire method holder as java.lang.Class and push as constant. 4308 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4309 ciInstance* caller_mirror = caller_klass->java_mirror(); 4310 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4311 4312 #ifndef PRODUCT 4313 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4314 tty->print_cr(" Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth()); 4315 tty->print_cr(" JVM state at this point:"); 4316 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4317 ciMethod* m = jvms()->of_depth(i)->method(); 4318 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4319 } 4320 } 4321 #endif 4322 return true; 4323 } 4324 break; 4325 } 4326 } 4327 4328 #ifndef PRODUCT 4329 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4330 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4331 tty->print_cr(" JVM state at this point:"); 4332 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4333 ciMethod* m = jvms()->of_depth(i)->method(); 4334 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4335 } 4336 } 4337 #endif 4338 4339 return false; // bail-out; let JVM_GetCallerClass do the work 4340 } 4341 4342 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4343 Node* arg = argument(0); 4344 Node* result = NULL; 4345 4346 switch (id) { 4347 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break; 4348 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break; 4349 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break; 4350 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break; 4351 4352 case vmIntrinsics::_doubleToLongBits: { 4353 // two paths (plus control) merge in a wood 4354 RegionNode *r = new (C) RegionNode(3); 4355 Node *phi = new (C) PhiNode(r, TypeLong::LONG); 4356 4357 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg)); 4358 // Build the boolean node 4359 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4360 4361 // Branch either way. 4362 // NaN case is less traveled, which makes all the difference. 4363 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4364 Node *opt_isnan = _gvn.transform(ifisnan); 4365 assert( opt_isnan->is_If(), "Expect an IfNode"); 4366 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4367 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4368 4369 set_control(iftrue); 4370 4371 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4372 Node *slow_result = longcon(nan_bits); // return NaN 4373 phi->init_req(1, _gvn.transform( slow_result )); 4374 r->init_req(1, iftrue); 4375 4376 // Else fall through 4377 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4378 set_control(iffalse); 4379 4380 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg))); 4381 r->init_req(2, iffalse); 4382 4383 // Post merge 4384 set_control(_gvn.transform(r)); 4385 record_for_igvn(r); 4386 4387 C->set_has_split_ifs(true); // Has chance for split-if optimization 4388 result = phi; 4389 assert(result->bottom_type()->isa_long(), "must be"); 4390 break; 4391 } 4392 4393 case vmIntrinsics::_floatToIntBits: { 4394 // two paths (plus control) merge in a wood 4395 RegionNode *r = new (C) RegionNode(3); 4396 Node *phi = new (C) PhiNode(r, TypeInt::INT); 4397 4398 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg)); 4399 // Build the boolean node 4400 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne)); 4401 4402 // Branch either way. 4403 // NaN case is less traveled, which makes all the difference. 4404 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4405 Node *opt_isnan = _gvn.transform(ifisnan); 4406 assert( opt_isnan->is_If(), "Expect an IfNode"); 4407 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4408 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan)); 4409 4410 set_control(iftrue); 4411 4412 static const jint nan_bits = 0x7fc00000; 4413 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4414 phi->init_req(1, _gvn.transform( slow_result )); 4415 r->init_req(1, iftrue); 4416 4417 // Else fall through 4418 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan)); 4419 set_control(iffalse); 4420 4421 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg))); 4422 r->init_req(2, iffalse); 4423 4424 // Post merge 4425 set_control(_gvn.transform(r)); 4426 record_for_igvn(r); 4427 4428 C->set_has_split_ifs(true); // Has chance for split-if optimization 4429 result = phi; 4430 assert(result->bottom_type()->isa_int(), "must be"); 4431 break; 4432 } 4433 4434 default: 4435 fatal_unexpected_iid(id); 4436 break; 4437 } 4438 set_result(_gvn.transform(result)); 4439 return true; 4440 } 4441 4442 #ifdef _LP64 4443 #define XTOP ,top() /*additional argument*/ 4444 #else //_LP64 4445 #define XTOP /*no additional argument*/ 4446 #endif //_LP64 4447 4448 //----------------------inline_unsafe_copyMemory------------------------- 4449 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4450 bool LibraryCallKit::inline_unsafe_copyMemory() { 4451 if (callee()->is_static()) return false; // caller must have the capability! 4452 null_check_receiver(); // null-check receiver 4453 if (stopped()) return true; 4454 4455 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4456 4457 Node* src_ptr = argument(1); // type: oop 4458 Node* src_off = ConvL2X(argument(2)); // type: long 4459 Node* dst_ptr = argument(4); // type: oop 4460 Node* dst_off = ConvL2X(argument(5)); // type: long 4461 Node* size = ConvL2X(argument(7)); // type: long 4462 4463 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4464 "fieldOffset must be byte-scaled"); 4465 4466 Node* src = make_unsafe_address(src_ptr, src_off); 4467 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4468 4469 // Conservatively insert a memory barrier on all memory slices. 4470 // Do not let writes of the copy source or destination float below the copy. 4471 insert_mem_bar(Op_MemBarCPUOrder); 4472 4473 // Call it. Note that the length argument is not scaled. 4474 make_runtime_call(RC_LEAF|RC_NO_FP, 4475 OptoRuntime::fast_arraycopy_Type(), 4476 StubRoutines::unsafe_arraycopy(), 4477 "unsafe_arraycopy", 4478 TypeRawPtr::BOTTOM, 4479 src, dst, size XTOP); 4480 4481 // Do not let reads of the copy destination float above the copy. 4482 insert_mem_bar(Op_MemBarCPUOrder); 4483 4484 return true; 4485 } 4486 4487 //------------------------clone_coping----------------------------------- 4488 // Helper function for inline_native_clone. 4489 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4490 assert(obj_size != NULL, ""); 4491 Node* raw_obj = alloc_obj->in(1); 4492 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4493 4494 AllocateNode* alloc = NULL; 4495 if (ReduceBulkZeroing) { 4496 // We will be completely responsible for initializing this object - 4497 // mark Initialize node as complete. 4498 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4499 // The object was just allocated - there should be no any stores! 4500 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4501 // Mark as complete_with_arraycopy so that on AllocateNode 4502 // expansion, we know this AllocateNode is initialized by an array 4503 // copy and a StoreStore barrier exists after the array copy. 4504 alloc->initialization()->set_complete_with_arraycopy(); 4505 } 4506 4507 // Copy the fastest available way. 4508 // TODO: generate fields copies for small objects instead. 4509 Node* src = obj; 4510 Node* dest = alloc_obj; 4511 Node* size = _gvn.transform(obj_size); 4512 4513 // Exclude the header but include array length to copy by 8 bytes words. 4514 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4515 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4516 instanceOopDesc::base_offset_in_bytes(); 4517 // base_off: 4518 // 8 - 32-bit VM 4519 // 12 - 64-bit VM, compressed klass 4520 // 16 - 64-bit VM, normal klass 4521 if (base_off % BytesPerLong != 0) { 4522 assert(UseCompressedClassPointers, ""); 4523 if (is_array) { 4524 // Exclude length to copy by 8 bytes words. 4525 base_off += sizeof(int); 4526 } else { 4527 // Include klass to copy by 8 bytes words. 4528 base_off = instanceOopDesc::klass_offset_in_bytes(); 4529 } 4530 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4531 } 4532 src = basic_plus_adr(src, base_off); 4533 dest = basic_plus_adr(dest, base_off); 4534 4535 // Compute the length also, if needed: 4536 Node* countx = size; 4537 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off))); 4538 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) )); 4539 4540 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4541 bool disjoint_bases = true; 4542 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases, 4543 src, NULL, dest, NULL, countx, 4544 /*dest_uninitialized*/true); 4545 4546 if (UseShenandoahGC && ShenandoahCloneBarrier) { 4547 // Make sure that references in the cloned object are updated for Shenandoah. 4548 make_runtime_call(RC_LEAF|RC_NO_FP, 4549 OptoRuntime::shenandoah_clone_barrier_Type(), 4550 CAST_FROM_FN_PTR(address, SharedRuntime::shenandoah_clone_barrier), 4551 "shenandoah_clone_barrier", TypePtr::BOTTOM, 4552 alloc_obj); 4553 } 4554 4555 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4556 if (card_mark) { 4557 assert(!is_array, ""); 4558 // Put in store barrier for any and all oops we are sticking 4559 // into this object. (We could avoid this if we could prove 4560 // that the object type contains no oop fields at all.) 4561 Node* no_particular_value = NULL; 4562 Node* no_particular_field = NULL; 4563 int raw_adr_idx = Compile::AliasIdxRaw; 4564 post_barrier(control(), 4565 memory(raw_adr_type), 4566 alloc_obj, 4567 no_particular_field, 4568 raw_adr_idx, 4569 no_particular_value, 4570 T_OBJECT, 4571 false); 4572 } 4573 4574 // Do not let reads from the cloned object float above the arraycopy. 4575 if (alloc != NULL) { 4576 // Do not let stores that initialize this object be reordered with 4577 // a subsequent store that would make this object accessible by 4578 // other threads. 4579 // Record what AllocateNode this StoreStore protects so that 4580 // escape analysis can go from the MemBarStoreStoreNode to the 4581 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4582 // based on the escape status of the AllocateNode. 4583 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4584 } else { 4585 insert_mem_bar(Op_MemBarCPUOrder); 4586 } 4587 } 4588 4589 //------------------------inline_native_clone---------------------------- 4590 // protected native Object java.lang.Object.clone(); 4591 // 4592 // Here are the simple edge cases: 4593 // null receiver => normal trap 4594 // virtual and clone was overridden => slow path to out-of-line clone 4595 // not cloneable or finalizer => slow path to out-of-line Object.clone 4596 // 4597 // The general case has two steps, allocation and copying. 4598 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4599 // 4600 // Copying also has two cases, oop arrays and everything else. 4601 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4602 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4603 // 4604 // These steps fold up nicely if and when the cloned object's klass 4605 // can be sharply typed as an object array, a type array, or an instance. 4606 // 4607 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4608 PhiNode* result_val; 4609 4610 // Set the reexecute bit for the interpreter to reexecute 4611 // the bytecode that invokes Object.clone if deoptimization happens. 4612 { PreserveReexecuteState preexecs(this); 4613 jvms()->set_should_reexecute(true); 4614 4615 Node* obj = null_check_receiver(); 4616 if (stopped()) return true; 4617 4618 Node* obj_klass = load_object_klass(obj); 4619 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4620 const TypeOopPtr* toop = ((tklass != NULL) 4621 ? tklass->as_instance_type() 4622 : TypeInstPtr::NOTNULL); 4623 4624 // Conservatively insert a memory barrier on all memory slices. 4625 // Do not let writes into the original float below the clone. 4626 insert_mem_bar(Op_MemBarCPUOrder); 4627 4628 // paths into result_reg: 4629 enum { 4630 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4631 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4632 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4633 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4634 PATH_LIMIT 4635 }; 4636 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT); 4637 result_val = new(C) PhiNode(result_reg, 4638 TypeInstPtr::NOTNULL); 4639 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO); 4640 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, 4641 TypePtr::BOTTOM); 4642 record_for_igvn(result_reg); 4643 4644 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4645 int raw_adr_idx = Compile::AliasIdxRaw; 4646 4647 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4648 if (array_ctl != NULL) { 4649 // It's an array. 4650 PreserveJVMState pjvms(this); 4651 set_control(array_ctl); 4652 Node* obj_length = load_array_length(obj); 4653 Node* obj_size = NULL; 4654 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4655 4656 if (!use_ReduceInitialCardMarks()) { 4657 // If it is an oop array, it requires very special treatment, 4658 // because card marking is required on each card of the array. 4659 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4660 if (is_obja != NULL) { 4661 PreserveJVMState pjvms2(this); 4662 set_control(is_obja); 4663 // Generate a direct call to the right arraycopy function(s). 4664 bool disjoint_bases = true; 4665 bool length_never_negative = true; 4666 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT, 4667 obj, intcon(0), alloc_obj, intcon(0), 4668 obj_length, 4669 disjoint_bases, length_never_negative); 4670 result_reg->init_req(_objArray_path, control()); 4671 result_val->init_req(_objArray_path, alloc_obj); 4672 result_i_o ->set_req(_objArray_path, i_o()); 4673 result_mem ->set_req(_objArray_path, reset_memory()); 4674 } 4675 } 4676 // Otherwise, there are no card marks to worry about. 4677 // (We can dispense with card marks if we know the allocation 4678 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4679 // causes the non-eden paths to take compensating steps to 4680 // simulate a fresh allocation, so that no further 4681 // card marks are required in compiled code to initialize 4682 // the object.) 4683 4684 if (!stopped()) { 4685 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4686 4687 // Present the results of the copy. 4688 result_reg->init_req(_array_path, control()); 4689 result_val->init_req(_array_path, alloc_obj); 4690 result_i_o ->set_req(_array_path, i_o()); 4691 result_mem ->set_req(_array_path, reset_memory()); 4692 } 4693 } 4694 4695 // We only go to the instance fast case code if we pass a number of guards. 4696 // The paths which do not pass are accumulated in the slow_region. 4697 RegionNode* slow_region = new (C) RegionNode(1); 4698 record_for_igvn(slow_region); 4699 if (!stopped()) { 4700 // It's an instance (we did array above). Make the slow-path tests. 4701 // If this is a virtual call, we generate a funny guard. We grab 4702 // the vtable entry corresponding to clone() from the target object. 4703 // If the target method which we are calling happens to be the 4704 // Object clone() method, we pass the guard. We do not need this 4705 // guard for non-virtual calls; the caller is known to be the native 4706 // Object clone(). 4707 if (is_virtual) { 4708 generate_virtual_guard(obj_klass, slow_region); 4709 } 4710 4711 // The object must be cloneable and must not have a finalizer. 4712 // Both of these conditions may be checked in a single test. 4713 // We could optimize the cloneable test further, but we don't care. 4714 generate_access_flags_guard(obj_klass, 4715 // Test both conditions: 4716 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4717 // Must be cloneable but not finalizer: 4718 JVM_ACC_IS_CLONEABLE, 4719 slow_region); 4720 } 4721 4722 if (!stopped()) { 4723 // It's an instance, and it passed the slow-path tests. 4724 PreserveJVMState pjvms(this); 4725 Node* obj_size = NULL; 4726 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4727 // is reexecuted if deoptimization occurs and there could be problems when merging 4728 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4729 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4730 4731 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4732 4733 // Present the results of the slow call. 4734 result_reg->init_req(_instance_path, control()); 4735 result_val->init_req(_instance_path, alloc_obj); 4736 result_i_o ->set_req(_instance_path, i_o()); 4737 result_mem ->set_req(_instance_path, reset_memory()); 4738 } 4739 4740 // Generate code for the slow case. We make a call to clone(). 4741 set_control(_gvn.transform(slow_region)); 4742 if (!stopped()) { 4743 PreserveJVMState pjvms(this); 4744 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4745 Node* slow_result = set_results_for_java_call(slow_call); 4746 // this->control() comes from set_results_for_java_call 4747 result_reg->init_req(_slow_path, control()); 4748 result_val->init_req(_slow_path, slow_result); 4749 result_i_o ->set_req(_slow_path, i_o()); 4750 result_mem ->set_req(_slow_path, reset_memory()); 4751 } 4752 4753 // Return the combined state. 4754 set_control( _gvn.transform(result_reg)); 4755 set_i_o( _gvn.transform(result_i_o)); 4756 set_all_memory( _gvn.transform(result_mem)); 4757 } // original reexecute is set back here 4758 4759 set_result(_gvn.transform(result_val)); 4760 return true; 4761 } 4762 4763 //------------------------------basictype2arraycopy---------------------------- 4764 address LibraryCallKit::basictype2arraycopy(BasicType t, 4765 Node* src_offset, 4766 Node* dest_offset, 4767 bool disjoint_bases, 4768 const char* &name, 4769 bool dest_uninitialized) { 4770 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);; 4771 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);; 4772 4773 bool aligned = false; 4774 bool disjoint = disjoint_bases; 4775 4776 // if the offsets are the same, we can treat the memory regions as 4777 // disjoint, because either the memory regions are in different arrays, 4778 // or they are identical (which we can treat as disjoint.) We can also 4779 // treat a copy with a destination index less that the source index 4780 // as disjoint since a low->high copy will work correctly in this case. 4781 if (src_offset_inttype != NULL && src_offset_inttype->is_con() && 4782 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) { 4783 // both indices are constants 4784 int s_offs = src_offset_inttype->get_con(); 4785 int d_offs = dest_offset_inttype->get_con(); 4786 int element_size = type2aelembytes(t); 4787 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) && 4788 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0); 4789 if (s_offs >= d_offs) disjoint = true; 4790 } else if (src_offset == dest_offset && src_offset != NULL) { 4791 // This can occur if the offsets are identical non-constants. 4792 disjoint = true; 4793 } 4794 4795 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized); 4796 } 4797 4798 4799 //------------------------------inline_arraycopy----------------------- 4800 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4801 // Object dest, int destPos, 4802 // int length); 4803 bool LibraryCallKit::inline_arraycopy() { 4804 // Get the arguments. 4805 Node* src = argument(0); // type: oop 4806 Node* src_offset = argument(1); // type: int 4807 Node* dest = argument(2); // type: oop 4808 Node* dest_offset = argument(3); // type: int 4809 Node* length = argument(4); // type: int 4810 4811 // Compile time checks. If any of these checks cannot be verified at compile time, 4812 // we do not make a fast path for this call. Instead, we let the call remain as it 4813 // is. The checks we choose to mandate at compile time are: 4814 // 4815 // (1) src and dest are arrays. 4816 const Type* src_type = src->Value(&_gvn); 4817 const Type* dest_type = dest->Value(&_gvn); 4818 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4819 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4820 4821 // Do we have the type of src? 4822 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4823 // Do we have the type of dest? 4824 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4825 // Is the type for src from speculation? 4826 bool src_spec = false; 4827 // Is the type for dest from speculation? 4828 bool dest_spec = false; 4829 4830 if (!has_src || !has_dest) { 4831 // We don't have sufficient type information, let's see if 4832 // speculative types can help. We need to have types for both src 4833 // and dest so that it pays off. 4834 4835 // Do we already have or could we have type information for src 4836 bool could_have_src = has_src; 4837 // Do we already have or could we have type information for dest 4838 bool could_have_dest = has_dest; 4839 4840 ciKlass* src_k = NULL; 4841 if (!has_src) { 4842 src_k = src_type->speculative_type(); 4843 if (src_k != NULL && src_k->is_array_klass()) { 4844 could_have_src = true; 4845 } 4846 } 4847 4848 ciKlass* dest_k = NULL; 4849 if (!has_dest) { 4850 dest_k = dest_type->speculative_type(); 4851 if (dest_k != NULL && dest_k->is_array_klass()) { 4852 could_have_dest = true; 4853 } 4854 } 4855 4856 if (could_have_src && could_have_dest) { 4857 // This is going to pay off so emit the required guards 4858 if (!has_src) { 4859 src = maybe_cast_profiled_obj(src, src_k); 4860 src_type = _gvn.type(src); 4861 top_src = src_type->isa_aryptr(); 4862 has_src = (top_src != NULL && top_src->klass() != NULL); 4863 src_spec = true; 4864 } 4865 if (!has_dest) { 4866 dest = maybe_cast_profiled_obj(dest, dest_k); 4867 dest_type = _gvn.type(dest); 4868 top_dest = dest_type->isa_aryptr(); 4869 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4870 dest_spec = true; 4871 } 4872 } 4873 } 4874 4875 if (!has_src || !has_dest) { 4876 // Conservatively insert a memory barrier on all memory slices. 4877 // Do not let writes into the source float below the arraycopy. 4878 insert_mem_bar(Op_MemBarCPUOrder); 4879 4880 // Call StubRoutines::generic_arraycopy stub. 4881 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT, 4882 src, src_offset, dest, dest_offset, length); 4883 4884 // Do not let reads from the destination float above the arraycopy. 4885 // Since we cannot type the arrays, we don't know which slices 4886 // might be affected. We could restrict this barrier only to those 4887 // memory slices which pertain to array elements--but don't bother. 4888 if (!InsertMemBarAfterArraycopy) 4889 // (If InsertMemBarAfterArraycopy, there is already one in place.) 4890 insert_mem_bar(Op_MemBarCPUOrder); 4891 return true; 4892 } 4893 4894 // (2) src and dest arrays must have elements of the same BasicType 4895 // Figure out the size and type of the elements we will be copying. 4896 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4897 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4898 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4899 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4900 4901 if (src_elem != dest_elem || dest_elem == T_VOID) { 4902 // The component types are not the same or are not recognized. Punt. 4903 // (But, avoid the native method wrapper to JVM_ArrayCopy.) 4904 generate_slow_arraycopy(TypePtr::BOTTOM, 4905 src, src_offset, dest, dest_offset, length, 4906 /*dest_uninitialized*/false); 4907 return true; 4908 } 4909 4910 if (src_elem == T_OBJECT) { 4911 // If both arrays are object arrays then having the exact types 4912 // for both will remove the need for a subtype check at runtime 4913 // before the call and may make it possible to pick a faster copy 4914 // routine (without a subtype check on every element) 4915 // Do we have the exact type of src? 4916 bool could_have_src = src_spec; 4917 // Do we have the exact type of dest? 4918 bool could_have_dest = dest_spec; 4919 ciKlass* src_k = top_src->klass(); 4920 ciKlass* dest_k = top_dest->klass(); 4921 if (!src_spec) { 4922 src_k = src_type->speculative_type(); 4923 if (src_k != NULL && src_k->is_array_klass()) { 4924 could_have_src = true; 4925 } 4926 } 4927 if (!dest_spec) { 4928 dest_k = dest_type->speculative_type(); 4929 if (dest_k != NULL && dest_k->is_array_klass()) { 4930 could_have_dest = true; 4931 } 4932 } 4933 if (could_have_src && could_have_dest) { 4934 // If we can have both exact types, emit the missing guards 4935 if (could_have_src && !src_spec) { 4936 src = maybe_cast_profiled_obj(src, src_k); 4937 } 4938 if (could_have_dest && !dest_spec) { 4939 dest = maybe_cast_profiled_obj(dest, dest_k); 4940 } 4941 } 4942 } 4943 4944 //--------------------------------------------------------------------------- 4945 // We will make a fast path for this call to arraycopy. 4946 4947 // We have the following tests left to perform: 4948 // 4949 // (3) src and dest must not be null. 4950 // (4) src_offset must not be negative. 4951 // (5) dest_offset must not be negative. 4952 // (6) length must not be negative. 4953 // (7) src_offset + length must not exceed length of src. 4954 // (8) dest_offset + length must not exceed length of dest. 4955 // (9) each element of an oop array must be assignable 4956 4957 RegionNode* slow_region = new (C) RegionNode(1); 4958 record_for_igvn(slow_region); 4959 4960 // (3) operands must not be null 4961 // We currently perform our null checks with the null_check routine. 4962 // This means that the null exceptions will be reported in the caller 4963 // rather than (correctly) reported inside of the native arraycopy call. 4964 // This should be corrected, given time. We do our null check with the 4965 // stack pointer restored. 4966 src = null_check(src, T_ARRAY); 4967 dest = null_check(dest, T_ARRAY); 4968 4969 // (4) src_offset must not be negative. 4970 generate_negative_guard(src_offset, slow_region); 4971 4972 // (5) dest_offset must not be negative. 4973 generate_negative_guard(dest_offset, slow_region); 4974 4975 // (6) length must not be negative (moved to generate_arraycopy()). 4976 // generate_negative_guard(length, slow_region); 4977 4978 // (7) src_offset + length must not exceed length of src. 4979 generate_limit_guard(src_offset, length, 4980 load_array_length(src), 4981 slow_region); 4982 4983 // (8) dest_offset + length must not exceed length of dest. 4984 generate_limit_guard(dest_offset, length, 4985 load_array_length(dest), 4986 slow_region); 4987 4988 // (9) each element of an oop array must be assignable 4989 // The generate_arraycopy subroutine checks this. 4990 4991 // This is where the memory effects are placed: 4992 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem); 4993 generate_arraycopy(adr_type, dest_elem, 4994 src, src_offset, dest, dest_offset, length, 4995 false, false, slow_region); 4996 4997 return true; 4998 } 4999 5000 //-----------------------------generate_arraycopy---------------------- 5001 // Generate an optimized call to arraycopy. 5002 // Caller must guard against non-arrays. 5003 // Caller must determine a common array basic-type for both arrays. 5004 // Caller must validate offsets against array bounds. 5005 // The slow_region has already collected guard failure paths 5006 // (such as out of bounds length or non-conformable array types). 5007 // The generated code has this shape, in general: 5008 // 5009 // if (length == 0) return // via zero_path 5010 // slowval = -1 5011 // if (types unknown) { 5012 // slowval = call generic copy loop 5013 // if (slowval == 0) return // via checked_path 5014 // } else if (indexes in bounds) { 5015 // if ((is object array) && !(array type check)) { 5016 // slowval = call checked copy loop 5017 // if (slowval == 0) return // via checked_path 5018 // } else { 5019 // call bulk copy loop 5020 // return // via fast_path 5021 // } 5022 // } 5023 // // adjust params for remaining work: 5024 // if (slowval != -1) { 5025 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n 5026 // } 5027 // slow_region: 5028 // call slow arraycopy(src, src_offset, dest, dest_offset, length) 5029 // return // via slow_call_path 5030 // 5031 // This routine is used from several intrinsics: System.arraycopy, 5032 // Object.clone (the array subcase), and Arrays.copyOf[Range]. 5033 // 5034 void 5035 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type, 5036 BasicType basic_elem_type, 5037 Node* src, Node* src_offset, 5038 Node* dest, Node* dest_offset, 5039 Node* copy_length, 5040 bool disjoint_bases, 5041 bool length_never_negative, 5042 RegionNode* slow_region) { 5043 5044 if (slow_region == NULL) { 5045 slow_region = new(C) RegionNode(1); 5046 record_for_igvn(slow_region); 5047 } 5048 5049 Node* original_dest = dest; 5050 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed 5051 bool dest_uninitialized = false; 5052 5053 // See if this is the initialization of a newly-allocated array. 5054 // If so, we will take responsibility here for initializing it to zero. 5055 // (Note: Because tightly_coupled_allocation performs checks on the 5056 // out-edges of the dest, we need to avoid making derived pointers 5057 // from it until we have checked its uses.) 5058 if (ReduceBulkZeroing 5059 && !ZeroTLAB // pointless if already zeroed 5060 && basic_elem_type != T_CONFLICT // avoid corner case 5061 && !src->eqv_uncast(dest) 5062 && ((alloc = tightly_coupled_allocation(dest, slow_region)) 5063 != NULL) 5064 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0 5065 && alloc->maybe_set_complete(&_gvn)) { 5066 // "You break it, you buy it." 5067 InitializeNode* init = alloc->initialization(); 5068 assert(init->is_complete(), "we just did this"); 5069 init->set_complete_with_arraycopy(); 5070 assert(dest->is_CheckCastPP(), "sanity"); 5071 assert(dest->in(0)->in(0) == init, "dest pinned"); 5072 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory 5073 // From this point on, every exit path is responsible for 5074 // initializing any non-copied parts of the object to zero. 5075 // Also, if this flag is set we make sure that arraycopy interacts properly 5076 // with G1, eliding pre-barriers. See CR 6627983. 5077 dest_uninitialized = true; 5078 } else { 5079 // No zeroing elimination here. 5080 alloc = NULL; 5081 //original_dest = dest; 5082 //dest_uninitialized = false; 5083 } 5084 5085 // Results are placed here: 5086 enum { fast_path = 1, // normal void-returning assembly stub 5087 checked_path = 2, // special assembly stub with cleanup 5088 slow_call_path = 3, // something went wrong; call the VM 5089 zero_path = 4, // bypass when length of copy is zero 5090 bcopy_path = 5, // copy primitive array by 64-bit blocks 5091 PATH_LIMIT = 6 5092 }; 5093 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT); 5094 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO); 5095 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type); 5096 record_for_igvn(result_region); 5097 _gvn.set_type_bottom(result_i_o); 5098 _gvn.set_type_bottom(result_memory); 5099 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice"); 5100 5101 // The slow_control path: 5102 Node* slow_control; 5103 Node* slow_i_o = i_o(); 5104 Node* slow_mem = memory(adr_type); 5105 debug_only(slow_control = (Node*) badAddress); 5106 5107 // Checked control path: 5108 Node* checked_control = top(); 5109 Node* checked_mem = NULL; 5110 Node* checked_i_o = NULL; 5111 Node* checked_value = NULL; 5112 5113 if (basic_elem_type == T_CONFLICT) { 5114 assert(!dest_uninitialized, ""); 5115 Node* cv = generate_generic_arraycopy(adr_type, 5116 src, src_offset, dest, dest_offset, 5117 copy_length, dest_uninitialized); 5118 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5119 checked_control = control(); 5120 checked_i_o = i_o(); 5121 checked_mem = memory(adr_type); 5122 checked_value = cv; 5123 set_control(top()); // no fast path 5124 } 5125 5126 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative); 5127 if (not_pos != NULL) { 5128 PreserveJVMState pjvms(this); 5129 set_control(not_pos); 5130 5131 // (6) length must not be negative. 5132 if (!length_never_negative) { 5133 generate_negative_guard(copy_length, slow_region); 5134 } 5135 5136 // copy_length is 0. 5137 if (!stopped() && dest_uninitialized) { 5138 Node* dest_length = alloc->in(AllocateNode::ALength); 5139 if (copy_length->eqv_uncast(dest_length) 5140 || _gvn.find_int_con(dest_length, 1) <= 0) { 5141 // There is no zeroing to do. No need for a secondary raw memory barrier. 5142 } else { 5143 // Clear the whole thing since there are no source elements to copy. 5144 generate_clear_array(adr_type, dest, basic_elem_type, 5145 intcon(0), NULL, 5146 alloc->in(AllocateNode::AllocSize)); 5147 // Use a secondary InitializeNode as raw memory barrier. 5148 // Currently it is needed only on this path since other 5149 // paths have stub or runtime calls as raw memory barriers. 5150 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, 5151 Compile::AliasIdxRaw, 5152 top())->as_Initialize(); 5153 init->set_complete(&_gvn); // (there is no corresponding AllocateNode) 5154 } 5155 } 5156 5157 // Present the results of the fast call. 5158 result_region->init_req(zero_path, control()); 5159 result_i_o ->init_req(zero_path, i_o()); 5160 result_memory->init_req(zero_path, memory(adr_type)); 5161 } 5162 5163 if (!stopped() && dest_uninitialized) { 5164 // We have to initialize the *uncopied* part of the array to zero. 5165 // The copy destination is the slice dest[off..off+len]. The other slices 5166 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length]. 5167 Node* dest_size = alloc->in(AllocateNode::AllocSize); 5168 Node* dest_length = alloc->in(AllocateNode::ALength); 5169 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset, 5170 copy_length)); 5171 5172 // If there is a head section that needs zeroing, do it now. 5173 if (find_int_con(dest_offset, -1) != 0) { 5174 generate_clear_array(adr_type, dest, basic_elem_type, 5175 intcon(0), dest_offset, 5176 NULL); 5177 } 5178 5179 // Next, perform a dynamic check on the tail length. 5180 // It is often zero, and we can win big if we prove this. 5181 // There are two wins: Avoid generating the ClearArray 5182 // with its attendant messy index arithmetic, and upgrade 5183 // the copy to a more hardware-friendly word size of 64 bits. 5184 Node* tail_ctl = NULL; 5185 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) { 5186 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length)); 5187 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt)); 5188 tail_ctl = generate_slow_guard(bol_lt, NULL); 5189 assert(tail_ctl != NULL || !stopped(), "must be an outcome"); 5190 } 5191 5192 // At this point, let's assume there is no tail. 5193 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) { 5194 // There is no tail. Try an upgrade to a 64-bit copy. 5195 bool didit = false; 5196 { PreserveJVMState pjvms(this); 5197 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc, 5198 src, src_offset, dest, dest_offset, 5199 dest_size, dest_uninitialized); 5200 if (didit) { 5201 // Present the results of the block-copying fast call. 5202 result_region->init_req(bcopy_path, control()); 5203 result_i_o ->init_req(bcopy_path, i_o()); 5204 result_memory->init_req(bcopy_path, memory(adr_type)); 5205 } 5206 } 5207 if (didit) 5208 set_control(top()); // no regular fast path 5209 } 5210 5211 // Clear the tail, if any. 5212 if (tail_ctl != NULL) { 5213 Node* notail_ctl = stopped() ? NULL : control(); 5214 set_control(tail_ctl); 5215 if (notail_ctl == NULL) { 5216 generate_clear_array(adr_type, dest, basic_elem_type, 5217 dest_tail, NULL, 5218 dest_size); 5219 } else { 5220 // Make a local merge. 5221 Node* done_ctl = new(C) RegionNode(3); 5222 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type); 5223 done_ctl->init_req(1, notail_ctl); 5224 done_mem->init_req(1, memory(adr_type)); 5225 generate_clear_array(adr_type, dest, basic_elem_type, 5226 dest_tail, NULL, 5227 dest_size); 5228 done_ctl->init_req(2, control()); 5229 done_mem->init_req(2, memory(adr_type)); 5230 set_control( _gvn.transform(done_ctl)); 5231 set_memory( _gvn.transform(done_mem), adr_type ); 5232 } 5233 } 5234 } 5235 5236 BasicType copy_type = basic_elem_type; 5237 assert(basic_elem_type != T_ARRAY, "caller must fix this"); 5238 if (!stopped() && copy_type == T_OBJECT) { 5239 // If src and dest have compatible element types, we can copy bits. 5240 // Types S[] and D[] are compatible if D is a supertype of S. 5241 // 5242 // If they are not, we will use checked_oop_disjoint_arraycopy, 5243 // which performs a fast optimistic per-oop check, and backs off 5244 // further to JVM_ArrayCopy on the first per-oop check that fails. 5245 // (Actually, we don't move raw bits only; the GC requires card marks.) 5246 5247 // Get the Klass* for both src and dest 5248 Node* src_klass = load_object_klass(src); 5249 Node* dest_klass = load_object_klass(dest); 5250 5251 // Generate the subtype check. 5252 // This might fold up statically, or then again it might not. 5253 // 5254 // Non-static example: Copying List<String>.elements to a new String[]. 5255 // The backing store for a List<String> is always an Object[], 5256 // but its elements are always type String, if the generic types 5257 // are correct at the source level. 5258 // 5259 // Test S[] against D[], not S against D, because (probably) 5260 // the secondary supertype cache is less busy for S[] than S. 5261 // This usually only matters when D is an interface. 5262 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5263 // Plug failing path into checked_oop_disjoint_arraycopy 5264 if (not_subtype_ctrl != top()) { 5265 PreserveJVMState pjvms(this); 5266 set_control(not_subtype_ctrl); 5267 // (At this point we can assume disjoint_bases, since types differ.) 5268 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset()); 5269 Node* p1 = basic_plus_adr(dest_klass, ek_offset); 5270 Node* n1 = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p1, TypeRawPtr::BOTTOM); 5271 Node* dest_elem_klass = _gvn.transform(n1); 5272 Node* cv = generate_checkcast_arraycopy(adr_type, 5273 dest_elem_klass, 5274 src, src_offset, dest, dest_offset, 5275 ConvI2X(copy_length), dest_uninitialized); 5276 if (cv == NULL) cv = intcon(-1); // failure (no stub available) 5277 checked_control = control(); 5278 checked_i_o = i_o(); 5279 checked_mem = memory(adr_type); 5280 checked_value = cv; 5281 } 5282 // At this point we know we do not need type checks on oop stores. 5283 5284 // Let's see if we need card marks: 5285 if (alloc != NULL && use_ReduceInitialCardMarks() && ! UseShenandoahGC) { 5286 // If we do not need card marks, copy using the jint or jlong stub. 5287 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT); 5288 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type), 5289 "sizes agree"); 5290 } 5291 } 5292 5293 if (!stopped()) { 5294 // Generate the fast path, if possible. 5295 PreserveJVMState pjvms(this); 5296 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases, 5297 src, src_offset, dest, dest_offset, 5298 ConvI2X(copy_length), dest_uninitialized); 5299 5300 // Present the results of the fast call. 5301 result_region->init_req(fast_path, control()); 5302 result_i_o ->init_req(fast_path, i_o()); 5303 result_memory->init_req(fast_path, memory(adr_type)); 5304 } 5305 5306 // Here are all the slow paths up to this point, in one bundle: 5307 slow_control = top(); 5308 if (slow_region != NULL) 5309 slow_control = _gvn.transform(slow_region); 5310 DEBUG_ONLY(slow_region = (RegionNode*)badAddress); 5311 5312 set_control(checked_control); 5313 if (!stopped()) { 5314 // Clean up after the checked call. 5315 // The returned value is either 0 or -1^K, 5316 // where K = number of partially transferred array elements. 5317 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0))); 5318 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq)); 5319 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN); 5320 5321 // If it is 0, we are done, so transfer to the end. 5322 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff)); 5323 result_region->init_req(checked_path, checks_done); 5324 result_i_o ->init_req(checked_path, checked_i_o); 5325 result_memory->init_req(checked_path, checked_mem); 5326 5327 // If it is not zero, merge into the slow call. 5328 set_control( _gvn.transform(new(C) IfFalseNode(iff) )); 5329 RegionNode* slow_reg2 = new(C) RegionNode(3); 5330 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO); 5331 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type); 5332 record_for_igvn(slow_reg2); 5333 slow_reg2 ->init_req(1, slow_control); 5334 slow_i_o2 ->init_req(1, slow_i_o); 5335 slow_mem2 ->init_req(1, slow_mem); 5336 slow_reg2 ->init_req(2, control()); 5337 slow_i_o2 ->init_req(2, checked_i_o); 5338 slow_mem2 ->init_req(2, checked_mem); 5339 5340 slow_control = _gvn.transform(slow_reg2); 5341 slow_i_o = _gvn.transform(slow_i_o2); 5342 slow_mem = _gvn.transform(slow_mem2); 5343 5344 if (alloc != NULL) { 5345 // We'll restart from the very beginning, after zeroing the whole thing. 5346 // This can cause double writes, but that's OK since dest is brand new. 5347 // So we ignore the low 31 bits of the value returned from the stub. 5348 } else { 5349 // We must continue the copy exactly where it failed, or else 5350 // another thread might see the wrong number of writes to dest. 5351 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1))); 5352 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT); 5353 slow_offset->init_req(1, intcon(0)); 5354 slow_offset->init_req(2, checked_offset); 5355 slow_offset = _gvn.transform(slow_offset); 5356 5357 // Adjust the arguments by the conditionally incoming offset. 5358 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset)); 5359 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset)); 5360 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset)); 5361 5362 // Tweak the node variables to adjust the code produced below: 5363 src_offset = src_off_plus; 5364 dest_offset = dest_off_plus; 5365 copy_length = length_minus; 5366 } 5367 } 5368 5369 set_control(slow_control); 5370 if (!stopped()) { 5371 // Generate the slow path, if needed. 5372 PreserveJVMState pjvms(this); // replace_in_map may trash the map 5373 5374 set_memory(slow_mem, adr_type); 5375 set_i_o(slow_i_o); 5376 5377 if (dest_uninitialized) { 5378 generate_clear_array(adr_type, dest, basic_elem_type, 5379 intcon(0), NULL, 5380 alloc->in(AllocateNode::AllocSize)); 5381 } 5382 5383 generate_slow_arraycopy(adr_type, 5384 src, src_offset, dest, dest_offset, 5385 copy_length, /*dest_uninitialized*/false); 5386 5387 result_region->init_req(slow_call_path, control()); 5388 result_i_o ->init_req(slow_call_path, i_o()); 5389 result_memory->init_req(slow_call_path, memory(adr_type)); 5390 } 5391 5392 // Remove unused edges. 5393 for (uint i = 1; i < result_region->req(); i++) { 5394 if (result_region->in(i) == NULL) 5395 result_region->init_req(i, top()); 5396 } 5397 5398 // Finished; return the combined state. 5399 set_control( _gvn.transform(result_region)); 5400 set_i_o( _gvn.transform(result_i_o) ); 5401 set_memory( _gvn.transform(result_memory), adr_type ); 5402 5403 // The memory edges above are precise in order to model effects around 5404 // array copies accurately to allow value numbering of field loads around 5405 // arraycopy. Such field loads, both before and after, are common in Java 5406 // collections and similar classes involving header/array data structures. 5407 // 5408 // But with low number of register or when some registers are used or killed 5409 // by arraycopy calls it causes registers spilling on stack. See 6544710. 5410 // The next memory barrier is added to avoid it. If the arraycopy can be 5411 // optimized away (which it can, sometimes) then we can manually remove 5412 // the membar also. 5413 // 5414 // Do not let reads from the cloned object float above the arraycopy. 5415 if (alloc != NULL) { 5416 // Do not let stores that initialize this object be reordered with 5417 // a subsequent store that would make this object accessible by 5418 // other threads. 5419 // Record what AllocateNode this StoreStore protects so that 5420 // escape analysis can go from the MemBarStoreStoreNode to the 5421 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 5422 // based on the escape status of the AllocateNode. 5423 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 5424 } else if (InsertMemBarAfterArraycopy) 5425 insert_mem_bar(Op_MemBarCPUOrder); 5426 } 5427 5428 5429 // Helper function which determines if an arraycopy immediately follows 5430 // an allocation, with no intervening tests or other escapes for the object. 5431 AllocateArrayNode* 5432 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5433 RegionNode* slow_region) { 5434 if (stopped()) return NULL; // no fast path 5435 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5436 5437 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5438 if (alloc == NULL) return NULL; 5439 5440 Node* rawmem = memory(Compile::AliasIdxRaw); 5441 // Is the allocation's memory state untouched? 5442 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5443 // Bail out if there have been raw-memory effects since the allocation. 5444 // (Example: There might have been a call or safepoint.) 5445 return NULL; 5446 } 5447 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5448 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5449 return NULL; 5450 } 5451 5452 // There must be no unexpected observers of this allocation. 5453 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5454 Node* obs = ptr->fast_out(i); 5455 if (obs != this->map()) { 5456 return NULL; 5457 } 5458 } 5459 5460 // This arraycopy must unconditionally follow the allocation of the ptr. 5461 Node* alloc_ctl = ptr->in(0); 5462 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5463 5464 Node* ctl = control(); 5465 while (ctl != alloc_ctl) { 5466 // There may be guards which feed into the slow_region. 5467 // Any other control flow means that we might not get a chance 5468 // to finish initializing the allocated object. 5469 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5470 IfNode* iff = ctl->in(0)->as_If(); 5471 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5472 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5473 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5474 ctl = iff->in(0); // This test feeds the known slow_region. 5475 continue; 5476 } 5477 // One more try: Various low-level checks bottom out in 5478 // uncommon traps. If the debug-info of the trap omits 5479 // any reference to the allocation, as we've already 5480 // observed, then there can be no objection to the trap. 5481 bool found_trap = false; 5482 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5483 Node* obs = not_ctl->fast_out(j); 5484 if (obs->in(0) == not_ctl && obs->is_Call() && 5485 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5486 found_trap = true; break; 5487 } 5488 } 5489 if (found_trap) { 5490 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5491 continue; 5492 } 5493 } 5494 return NULL; 5495 } 5496 5497 // If we get this far, we have an allocation which immediately 5498 // precedes the arraycopy, and we can take over zeroing the new object. 5499 // The arraycopy will finish the initialization, and provide 5500 // a new control state to which we will anchor the destination pointer. 5501 5502 return alloc; 5503 } 5504 5505 // Helper for initialization of arrays, creating a ClearArray. 5506 // It writes zero bits in [start..end), within the body of an array object. 5507 // The memory effects are all chained onto the 'adr_type' alias category. 5508 // 5509 // Since the object is otherwise uninitialized, we are free 5510 // to put a little "slop" around the edges of the cleared area, 5511 // as long as it does not go back into the array's header, 5512 // or beyond the array end within the heap. 5513 // 5514 // The lower edge can be rounded down to the nearest jint and the 5515 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes. 5516 // 5517 // Arguments: 5518 // adr_type memory slice where writes are generated 5519 // dest oop of the destination array 5520 // basic_elem_type element type of the destination 5521 // slice_idx array index of first element to store 5522 // slice_len number of elements to store (or NULL) 5523 // dest_size total size in bytes of the array object 5524 // 5525 // Exactly one of slice_len or dest_size must be non-NULL. 5526 // If dest_size is non-NULL, zeroing extends to the end of the object. 5527 // If slice_len is non-NULL, the slice_idx value must be a constant. 5528 void 5529 LibraryCallKit::generate_clear_array(const TypePtr* adr_type, 5530 Node* dest, 5531 BasicType basic_elem_type, 5532 Node* slice_idx, 5533 Node* slice_len, 5534 Node* dest_size) { 5535 // one or the other but not both of slice_len and dest_size: 5536 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, ""); 5537 if (slice_len == NULL) slice_len = top(); 5538 if (dest_size == NULL) dest_size = top(); 5539 5540 // operate on this memory slice: 5541 Node* mem = memory(adr_type); // memory slice to operate on 5542 5543 // scaling and rounding of indexes: 5544 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5545 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5546 int clear_low = (-1 << scale) & (BytesPerInt - 1); 5547 int bump_bit = (-1 << scale) & BytesPerInt; 5548 5549 // determine constant starts and ends 5550 const intptr_t BIG_NEG = -128; 5551 assert(BIG_NEG + 2*abase < 0, "neg enough"); 5552 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG); 5553 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG); 5554 if (slice_len_con == 0) { 5555 return; // nothing to do here 5556 } 5557 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low; 5558 intptr_t end_con = find_intptr_t_con(dest_size, -1); 5559 if (slice_idx_con >= 0 && slice_len_con >= 0) { 5560 assert(end_con < 0, "not two cons"); 5561 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale), 5562 BytesPerLong); 5563 } 5564 5565 if (start_con >= 0 && end_con >= 0) { 5566 // Constant start and end. Simple. 5567 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5568 start_con, end_con, &_gvn); 5569 } else if (start_con >= 0 && dest_size != top()) { 5570 // Constant start, pre-rounded end after the tail of the array. 5571 Node* end = dest_size; 5572 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5573 start_con, end, &_gvn); 5574 } else if (start_con >= 0 && slice_len != top()) { 5575 // Constant start, non-constant end. End needs rounding up. 5576 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8) 5577 intptr_t end_base = abase + (slice_idx_con << scale); 5578 int end_round = (-1 << scale) & (BytesPerLong - 1); 5579 Node* end = ConvI2X(slice_len); 5580 if (scale != 0) 5581 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) )); 5582 end_base += end_round; 5583 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base))); 5584 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round))); 5585 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5586 start_con, end, &_gvn); 5587 } else if (start_con < 0 && dest_size != top()) { 5588 // Non-constant start, pre-rounded end after the tail of the array. 5589 // This is almost certainly a "round-to-end" operation. 5590 Node* start = slice_idx; 5591 start = ConvI2X(start); 5592 if (scale != 0) 5593 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) )); 5594 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase))); 5595 if ((bump_bit | clear_low) != 0) { 5596 int to_clear = (bump_bit | clear_low); 5597 // Align up mod 8, then store a jint zero unconditionally 5598 // just before the mod-8 boundary. 5599 if (((abase + bump_bit) & ~to_clear) - bump_bit 5600 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) { 5601 bump_bit = 0; 5602 assert((abase & to_clear) == 0, "array base must be long-aligned"); 5603 } else { 5604 // Bump 'start' up to (or past) the next jint boundary: 5605 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit))); 5606 assert((abase & clear_low) == 0, "array base must be int-aligned"); 5607 } 5608 // Round bumped 'start' down to jlong boundary in body of array. 5609 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear))); 5610 if (bump_bit != 0) { 5611 // Store a zero to the immediately preceding jint: 5612 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit))); 5613 Node* p1 = basic_plus_adr(dest, x1); 5614 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered); 5615 mem = _gvn.transform(mem); 5616 } 5617 } 5618 Node* end = dest_size; // pre-rounded 5619 mem = ClearArrayNode::clear_memory(control(), mem, dest, 5620 start, end, &_gvn); 5621 } else { 5622 // Non-constant start, unrounded non-constant end. 5623 // (Nobody zeroes a random midsection of an array using this routine.) 5624 ShouldNotReachHere(); // fix caller 5625 } 5626 5627 // Done. 5628 set_memory(mem, adr_type); 5629 } 5630 5631 5632 bool 5633 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type, 5634 BasicType basic_elem_type, 5635 AllocateNode* alloc, 5636 Node* src, Node* src_offset, 5637 Node* dest, Node* dest_offset, 5638 Node* dest_size, bool dest_uninitialized) { 5639 // See if there is an advantage from block transfer. 5640 int scale = exact_log2(type2aelembytes(basic_elem_type)); 5641 if (scale >= LogBytesPerLong) 5642 return false; // it is already a block transfer 5643 5644 // Look at the alignment of the starting offsets. 5645 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type); 5646 5647 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1); 5648 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1); 5649 if (src_off_con < 0 || dest_off_con < 0) 5650 // At present, we can only understand constants. 5651 return false; 5652 5653 intptr_t src_off = abase + (src_off_con << scale); 5654 intptr_t dest_off = abase + (dest_off_con << scale); 5655 5656 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) { 5657 // Non-aligned; too bad. 5658 // One more chance: Pick off an initial 32-bit word. 5659 // This is a common case, since abase can be odd mod 8. 5660 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt && 5661 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) { 5662 Node* sptr = basic_plus_adr(src, src_off); 5663 Node* dptr = basic_plus_adr(dest, dest_off); 5664 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered); 5665 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered); 5666 src_off += BytesPerInt; 5667 dest_off += BytesPerInt; 5668 } else { 5669 return false; 5670 } 5671 } 5672 assert(src_off % BytesPerLong == 0, ""); 5673 assert(dest_off % BytesPerLong == 0, ""); 5674 5675 // Do this copy by giant steps. 5676 Node* sptr = basic_plus_adr(src, src_off); 5677 Node* dptr = basic_plus_adr(dest, dest_off); 5678 Node* countx = dest_size; 5679 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off))); 5680 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong))); 5681 5682 bool disjoint_bases = true; // since alloc != NULL 5683 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases, 5684 sptr, NULL, dptr, NULL, countx, dest_uninitialized); 5685 5686 return true; 5687 } 5688 5689 5690 // Helper function; generates code for the slow case. 5691 // We make a call to a runtime method which emulates the native method, 5692 // but without the native wrapper overhead. 5693 void 5694 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type, 5695 Node* src, Node* src_offset, 5696 Node* dest, Node* dest_offset, 5697 Node* copy_length, bool dest_uninitialized) { 5698 assert(!dest_uninitialized, "Invariant"); 5699 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON, 5700 OptoRuntime::slow_arraycopy_Type(), 5701 OptoRuntime::slow_arraycopy_Java(), 5702 "slow_arraycopy", adr_type, 5703 src, src_offset, dest, dest_offset, 5704 copy_length); 5705 5706 // Handle exceptions thrown by this fellow: 5707 make_slow_call_ex(call, env()->Throwable_klass(), false); 5708 } 5709 5710 // Helper function; generates code for cases requiring runtime checks. 5711 Node* 5712 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type, 5713 Node* dest_elem_klass, 5714 Node* src, Node* src_offset, 5715 Node* dest, Node* dest_offset, 5716 Node* copy_length, bool dest_uninitialized) { 5717 if (stopped()) return NULL; 5718 5719 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized); 5720 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5721 return NULL; 5722 } 5723 5724 // Pick out the parameters required to perform a store-check 5725 // for the target array. This is an optimistic check. It will 5726 // look in each non-null element's class, at the desired klass's 5727 // super_check_offset, for the desired klass. 5728 int sco_offset = in_bytes(Klass::super_check_offset_offset()); 5729 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset); 5730 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered); 5731 Node* check_offset = ConvI2X(_gvn.transform(n3)); 5732 Node* check_value = dest_elem_klass; 5733 5734 Node* src_start = array_element_address(src, src_offset, T_OBJECT); 5735 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT); 5736 5737 // (We know the arrays are never conjoint, because their types differ.) 5738 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5739 OptoRuntime::checkcast_arraycopy_Type(), 5740 copyfunc_addr, "checkcast_arraycopy", adr_type, 5741 // five arguments, of which two are 5742 // intptr_t (jlong in LP64) 5743 src_start, dest_start, 5744 copy_length XTOP, 5745 check_offset XTOP, 5746 check_value); 5747 5748 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5749 } 5750 5751 5752 // Helper function; generates code for cases requiring runtime checks. 5753 Node* 5754 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type, 5755 Node* src, Node* src_offset, 5756 Node* dest, Node* dest_offset, 5757 Node* copy_length, bool dest_uninitialized) { 5758 assert(!dest_uninitialized, "Invariant"); 5759 if (stopped()) return NULL; 5760 address copyfunc_addr = StubRoutines::generic_arraycopy(); 5761 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path. 5762 return NULL; 5763 } 5764 5765 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5766 OptoRuntime::generic_arraycopy_Type(), 5767 copyfunc_addr, "generic_arraycopy", adr_type, 5768 src, src_offset, dest, dest_offset, copy_length); 5769 5770 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 5771 } 5772 5773 // Helper function; generates the fast out-of-line call to an arraycopy stub. 5774 void 5775 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type, 5776 BasicType basic_elem_type, 5777 bool disjoint_bases, 5778 Node* src, Node* src_offset, 5779 Node* dest, Node* dest_offset, 5780 Node* copy_length, bool dest_uninitialized) { 5781 if (stopped()) return; // nothing to do 5782 5783 Node* src_start = src; 5784 Node* dest_start = dest; 5785 if (src_offset != NULL || dest_offset != NULL) { 5786 assert(src_offset != NULL && dest_offset != NULL, ""); 5787 src_start = array_element_address(src, src_offset, basic_elem_type); 5788 dest_start = array_element_address(dest, dest_offset, basic_elem_type); 5789 } 5790 5791 // Figure out which arraycopy runtime method to call. 5792 const char* copyfunc_name = "arraycopy"; 5793 address copyfunc_addr = 5794 basictype2arraycopy(basic_elem_type, src_offset, dest_offset, 5795 disjoint_bases, copyfunc_name, dest_uninitialized); 5796 5797 // Call it. Note that the count_ix value is not scaled to a byte-size. 5798 make_runtime_call(RC_LEAF|RC_NO_FP, 5799 OptoRuntime::fast_arraycopy_Type(), 5800 copyfunc_addr, copyfunc_name, adr_type, 5801 src_start, dest_start, copy_length XTOP); 5802 } 5803 5804 //-------------inline_encodeISOArray----------------------------------- 5805 // encode char[] to byte[] in ISO_8859_1 5806 bool LibraryCallKit::inline_encodeISOArray() { 5807 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5808 // no receiver since it is static method 5809 Node *src = argument(0); 5810 Node *src_offset = argument(1); 5811 Node *dst = argument(2); 5812 Node *dst_offset = argument(3); 5813 Node *length = argument(4); 5814 5815 const Type* src_type = src->Value(&_gvn); 5816 const Type* dst_type = dst->Value(&_gvn); 5817 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5818 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5819 if (top_src == NULL || top_src->klass() == NULL || 5820 top_dest == NULL || top_dest->klass() == NULL) { 5821 // failed array check 5822 return false; 5823 } 5824 5825 // Figure out the size and type of the elements we will be copying. 5826 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5827 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5828 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5829 return false; 5830 } 5831 Node* src_start = array_element_address(src, src_offset, src_elem); 5832 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5833 // 'src_start' points to src array + scaled offset 5834 // 'dst_start' points to dst array + scaled offset 5835 5836 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5837 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5838 enc = _gvn.transform(enc); 5839 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc)); 5840 set_memory(res_mem, mtype); 5841 set_result(enc); 5842 return true; 5843 } 5844 5845 //-------------inline_multiplyToLen----------------------------------- 5846 bool LibraryCallKit::inline_multiplyToLen() { 5847 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5848 5849 address stubAddr = StubRoutines::multiplyToLen(); 5850 if (stubAddr == NULL) { 5851 return false; // Intrinsic's stub is not implemented on this platform 5852 } 5853 const char* stubName = "multiplyToLen"; 5854 5855 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5856 5857 // no receiver because it is a static method 5858 Node* x = argument(0); 5859 Node* xlen = argument(1); 5860 Node* y = argument(2); 5861 Node* ylen = argument(3); 5862 Node* z = argument(4); 5863 5864 const Type* x_type = x->Value(&_gvn); 5865 const Type* y_type = y->Value(&_gvn); 5866 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5867 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5868 if (top_x == NULL || top_x->klass() == NULL || 5869 top_y == NULL || top_y->klass() == NULL) { 5870 // failed array check 5871 return false; 5872 } 5873 5874 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5875 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5876 if (x_elem != T_INT || y_elem != T_INT) { 5877 return false; 5878 } 5879 5880 // Set the original stack and the reexecute bit for the interpreter to reexecute 5881 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5882 // on the return from z array allocation in runtime. 5883 { PreserveReexecuteState preexecs(this); 5884 jvms()->set_should_reexecute(true); 5885 5886 Node* x_start = array_element_address(x, intcon(0), x_elem); 5887 Node* y_start = array_element_address(y, intcon(0), y_elem); 5888 // 'x_start' points to x array + scaled xlen 5889 // 'y_start' points to y array + scaled ylen 5890 5891 // Allocate the result array 5892 Node* zlen = _gvn.transform(new(C) AddINode(xlen, ylen)); 5893 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5894 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5895 5896 IdealKit ideal(this); 5897 5898 #define __ ideal. 5899 Node* one = __ ConI(1); 5900 Node* zero = __ ConI(0); 5901 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5902 __ set(need_alloc, zero); 5903 __ set(z_alloc, z); 5904 __ if_then(z, BoolTest::eq, null()); { 5905 __ increment (need_alloc, one); 5906 } __ else_(); { 5907 // Update graphKit memory and control from IdealKit. 5908 sync_kit(ideal); 5909 Node* zlen_arg = load_array_length(z); 5910 // Update IdealKit memory and control from graphKit. 5911 __ sync_kit(this); 5912 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5913 __ increment (need_alloc, one); 5914 } __ end_if(); 5915 } __ end_if(); 5916 5917 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5918 // Update graphKit memory and control from IdealKit. 5919 sync_kit(ideal); 5920 Node * narr = new_array(klass_node, zlen, 1); 5921 // Update IdealKit memory and control from graphKit. 5922 __ sync_kit(this); 5923 __ set(z_alloc, narr); 5924 } __ end_if(); 5925 5926 sync_kit(ideal); 5927 z = __ value(z_alloc); 5928 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5929 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5930 // Final sync IdealKit and GraphKit. 5931 final_sync(ideal); 5932 #undef __ 5933 5934 Node* z_start = array_element_address(z, intcon(0), T_INT); 5935 5936 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5937 OptoRuntime::multiplyToLen_Type(), 5938 stubAddr, stubName, TypePtr::BOTTOM, 5939 x_start, xlen, y_start, ylen, z_start, zlen); 5940 } // original reexecute is set back here 5941 5942 C->set_has_split_ifs(true); // Has chance for split-if optimization 5943 set_result(z); 5944 return true; 5945 } 5946 5947 //-------------inline_squareToLen------------------------------------ 5948 bool LibraryCallKit::inline_squareToLen() { 5949 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5950 5951 address stubAddr = StubRoutines::squareToLen(); 5952 if (stubAddr == NULL) { 5953 return false; // Intrinsic's stub is not implemented on this platform 5954 } 5955 const char* stubName = "squareToLen"; 5956 5957 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5958 5959 Node* x = argument(0); 5960 Node* len = argument(1); 5961 Node* z = argument(2); 5962 Node* zlen = argument(3); 5963 5964 const Type* x_type = x->Value(&_gvn); 5965 const Type* z_type = z->Value(&_gvn); 5966 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5967 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5968 if (top_x == NULL || top_x->klass() == NULL || 5969 top_z == NULL || top_z->klass() == NULL) { 5970 // failed array check 5971 return false; 5972 } 5973 5974 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5975 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5976 if (x_elem != T_INT || z_elem != T_INT) { 5977 return false; 5978 } 5979 5980 5981 Node* x_start = array_element_address(x, intcon(0), x_elem); 5982 Node* z_start = array_element_address(z, intcon(0), z_elem); 5983 5984 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5985 OptoRuntime::squareToLen_Type(), 5986 stubAddr, stubName, TypePtr::BOTTOM, 5987 x_start, len, z_start, zlen); 5988 5989 set_result(z); 5990 return true; 5991 } 5992 5993 //-------------inline_mulAdd------------------------------------------ 5994 bool LibraryCallKit::inline_mulAdd() { 5995 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5996 5997 address stubAddr = StubRoutines::mulAdd(); 5998 if (stubAddr == NULL) { 5999 return false; // Intrinsic's stub is not implemented on this platform 6000 } 6001 const char* stubName = "mulAdd"; 6002 6003 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 6004 6005 Node* out = argument(0); 6006 Node* in = argument(1); 6007 Node* offset = argument(2); 6008 Node* len = argument(3); 6009 Node* k = argument(4); 6010 6011 const Type* out_type = out->Value(&_gvn); 6012 const Type* in_type = in->Value(&_gvn); 6013 const TypeAryPtr* top_out = out_type->isa_aryptr(); 6014 const TypeAryPtr* top_in = in_type->isa_aryptr(); 6015 if (top_out == NULL || top_out->klass() == NULL || 6016 top_in == NULL || top_in->klass() == NULL) { 6017 // failed array check 6018 return false; 6019 } 6020 6021 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6022 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6023 if (out_elem != T_INT || in_elem != T_INT) { 6024 return false; 6025 } 6026 6027 Node* outlen = load_array_length(out); 6028 Node* new_offset = _gvn.transform(new (C) SubINode(outlen, offset)); 6029 Node* out_start = array_element_address(out, intcon(0), out_elem); 6030 Node* in_start = array_element_address(in, intcon(0), in_elem); 6031 6032 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6033 OptoRuntime::mulAdd_Type(), 6034 stubAddr, stubName, TypePtr::BOTTOM, 6035 out_start,in_start, new_offset, len, k); 6036 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6037 set_result(result); 6038 return true; 6039 } 6040 6041 //-------------inline_montgomeryMultiply----------------------------------- 6042 bool LibraryCallKit::inline_montgomeryMultiply() { 6043 address stubAddr = StubRoutines::montgomeryMultiply(); 6044 if (stubAddr == NULL) { 6045 return false; // Intrinsic's stub is not implemented on this platform 6046 } 6047 6048 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 6049 const char* stubName = "montgomery_multiply"; 6050 6051 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 6052 6053 Node* a = argument(0); 6054 Node* b = argument(1); 6055 Node* n = argument(2); 6056 Node* len = argument(3); 6057 Node* inv = argument(4); 6058 Node* m = argument(6); 6059 6060 const Type* a_type = a->Value(&_gvn); 6061 const TypeAryPtr* top_a = a_type->isa_aryptr(); 6062 const Type* b_type = b->Value(&_gvn); 6063 const TypeAryPtr* top_b = b_type->isa_aryptr(); 6064 const Type* n_type = a->Value(&_gvn); 6065 const TypeAryPtr* top_n = n_type->isa_aryptr(); 6066 const Type* m_type = a->Value(&_gvn); 6067 const TypeAryPtr* top_m = m_type->isa_aryptr(); 6068 if (top_a == NULL || top_a->klass() == NULL || 6069 top_b == NULL || top_b->klass() == NULL || 6070 top_n == NULL || top_n->klass() == NULL || 6071 top_m == NULL || top_m->klass() == NULL) { 6072 // failed array check 6073 return false; 6074 } 6075 6076 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6077 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6078 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6079 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6080 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 6081 return false; 6082 } 6083 6084 // Make the call 6085 { 6086 Node* a_start = array_element_address(a, intcon(0), a_elem); 6087 Node* b_start = array_element_address(b, intcon(0), b_elem); 6088 Node* n_start = array_element_address(n, intcon(0), n_elem); 6089 Node* m_start = array_element_address(m, intcon(0), m_elem); 6090 6091 Node* call = NULL; 6092 if (CCallingConventionRequiresIntsAsLongs) { 6093 Node* len_I2L = ConvI2L(len); 6094 call = make_runtime_call(RC_LEAF, 6095 OptoRuntime::montgomeryMultiply_Type(), 6096 stubAddr, stubName, TypePtr::BOTTOM, 6097 a_start, b_start, n_start, len_I2L XTOP, inv, 6098 top(), m_start); 6099 } else { 6100 call = make_runtime_call(RC_LEAF, 6101 OptoRuntime::montgomeryMultiply_Type(), 6102 stubAddr, stubName, TypePtr::BOTTOM, 6103 a_start, b_start, n_start, len, inv, top(), 6104 m_start); 6105 } 6106 set_result(m); 6107 } 6108 6109 return true; 6110 } 6111 6112 bool LibraryCallKit::inline_montgomerySquare() { 6113 address stubAddr = StubRoutines::montgomerySquare(); 6114 if (stubAddr == NULL) { 6115 return false; // Intrinsic's stub is not implemented on this platform 6116 } 6117 6118 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 6119 const char* stubName = "montgomery_square"; 6120 6121 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 6122 6123 Node* a = argument(0); 6124 Node* n = argument(1); 6125 Node* len = argument(2); 6126 Node* inv = argument(3); 6127 Node* m = argument(5); 6128 6129 const Type* a_type = a->Value(&_gvn); 6130 const TypeAryPtr* top_a = a_type->isa_aryptr(); 6131 const Type* n_type = a->Value(&_gvn); 6132 const TypeAryPtr* top_n = n_type->isa_aryptr(); 6133 const Type* m_type = a->Value(&_gvn); 6134 const TypeAryPtr* top_m = m_type->isa_aryptr(); 6135 if (top_a == NULL || top_a->klass() == NULL || 6136 top_n == NULL || top_n->klass() == NULL || 6137 top_m == NULL || top_m->klass() == NULL) { 6138 // failed array check 6139 return false; 6140 } 6141 6142 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6143 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6144 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6145 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 6146 return false; 6147 } 6148 6149 // Make the call 6150 { 6151 Node* a_start = array_element_address(a, intcon(0), a_elem); 6152 Node* n_start = array_element_address(n, intcon(0), n_elem); 6153 Node* m_start = array_element_address(m, intcon(0), m_elem); 6154 6155 Node* call = NULL; 6156 if (CCallingConventionRequiresIntsAsLongs) { 6157 Node* len_I2L = ConvI2L(len); 6158 call = make_runtime_call(RC_LEAF, 6159 OptoRuntime::montgomerySquare_Type(), 6160 stubAddr, stubName, TypePtr::BOTTOM, 6161 a_start, n_start, len_I2L XTOP, inv, top(), 6162 m_start); 6163 } else { 6164 call = make_runtime_call(RC_LEAF, 6165 OptoRuntime::montgomerySquare_Type(), 6166 stubAddr, stubName, TypePtr::BOTTOM, 6167 a_start, n_start, len, inv, top(), 6168 m_start); 6169 } 6170 6171 set_result(m); 6172 } 6173 6174 return true; 6175 } 6176 6177 6178 /** 6179 * Calculate CRC32 for byte. 6180 * int java.util.zip.CRC32.update(int crc, int b) 6181 */ 6182 bool LibraryCallKit::inline_updateCRC32() { 6183 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6184 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 6185 // no receiver since it is static method 6186 Node* crc = argument(0); // type: int 6187 Node* b = argument(1); // type: int 6188 6189 /* 6190 * int c = ~ crc; 6191 * b = timesXtoThe32[(b ^ c) & 0xFF]; 6192 * b = b ^ (c >>> 8); 6193 * crc = ~b; 6194 */ 6195 6196 Node* M1 = intcon(-1); 6197 crc = _gvn.transform(new (C) XorINode(crc, M1)); 6198 Node* result = _gvn.transform(new (C) XorINode(crc, b)); 6199 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF))); 6200 6201 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 6202 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2))); 6203 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 6204 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 6205 6206 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8))); 6207 result = _gvn.transform(new (C) XorINode(crc, result)); 6208 result = _gvn.transform(new (C) XorINode(result, M1)); 6209 set_result(result); 6210 return true; 6211 } 6212 6213 /** 6214 * Calculate CRC32 for byte[] array. 6215 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 6216 */ 6217 bool LibraryCallKit::inline_updateBytesCRC32() { 6218 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6219 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 6220 // no receiver since it is static method 6221 Node* crc = argument(0); // type: int 6222 Node* src = argument(1); // type: oop 6223 Node* offset = argument(2); // type: int 6224 Node* length = argument(3); // type: int 6225 6226 const Type* src_type = src->Value(&_gvn); 6227 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6228 if (top_src == NULL || top_src->klass() == NULL) { 6229 // failed array check 6230 return false; 6231 } 6232 6233 // Figure out the size and type of the elements we will be copying. 6234 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6235 if (src_elem != T_BYTE) { 6236 return false; 6237 } 6238 6239 // 'src_start' points to src array + scaled offset 6240 Node* src_start = array_element_address(src, offset, src_elem); 6241 6242 // We assume that range check is done by caller. 6243 // TODO: generate range check (offset+length < src.length) in debug VM. 6244 6245 // Call the stub. 6246 address stubAddr = StubRoutines::updateBytesCRC32(); 6247 const char *stubName = "updateBytesCRC32"; 6248 Node* call; 6249 if (CCallingConventionRequiresIntsAsLongs) { 6250 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6251 stubAddr, stubName, TypePtr::BOTTOM, 6252 crc XTOP, src_start, length XTOP); 6253 } else { 6254 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6255 stubAddr, stubName, TypePtr::BOTTOM, 6256 crc, src_start, length); 6257 } 6258 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6259 set_result(result); 6260 return true; 6261 } 6262 6263 /** 6264 * Calculate CRC32 for ByteBuffer. 6265 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 6266 */ 6267 bool LibraryCallKit::inline_updateByteBufferCRC32() { 6268 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 6269 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 6270 // no receiver since it is static method 6271 Node* crc = argument(0); // type: int 6272 Node* src = argument(1); // type: long 6273 Node* offset = argument(3); // type: int 6274 Node* length = argument(4); // type: int 6275 6276 src = ConvL2X(src); // adjust Java long to machine word 6277 Node* base = _gvn.transform(new (C) CastX2PNode(src)); 6278 offset = ConvI2X(offset); 6279 6280 // 'src_start' points to src array + scaled offset 6281 Node* src_start = basic_plus_adr(top(), base, offset); 6282 6283 // Call the stub. 6284 address stubAddr = StubRoutines::updateBytesCRC32(); 6285 const char *stubName = "updateBytesCRC32"; 6286 Node* call; 6287 if (CCallingConventionRequiresIntsAsLongs) { 6288 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6289 stubAddr, stubName, TypePtr::BOTTOM, 6290 crc XTOP, src_start, length XTOP); 6291 } else { 6292 call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 6293 stubAddr, stubName, TypePtr::BOTTOM, 6294 crc, src_start, length); 6295 } 6296 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6297 set_result(result); 6298 return true; 6299 } 6300 6301 //----------------------------inline_reference_get---------------------------- 6302 // public T java.lang.ref.Reference.get(); 6303 bool LibraryCallKit::inline_reference_get() { 6304 const int referent_offset = java_lang_ref_Reference::referent_offset; 6305 guarantee(referent_offset > 0, "should have already been set"); 6306 6307 // Get the argument: 6308 Node* reference_obj = null_check_receiver(); 6309 if (stopped()) return true; 6310 6311 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 6312 6313 ciInstanceKlass* klass = env()->Object_klass(); 6314 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 6315 6316 Node* no_ctrl = NULL; 6317 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 6318 6319 #if INCLUDE_ALL_GCS 6320 if (UseShenandoahGC) { 6321 result = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, result); 6322 } 6323 #endif 6324 6325 // Use the pre-barrier to record the value in the referent field 6326 pre_barrier(false /* do_load */, 6327 control(), 6328 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 6329 result /* pre_val */, 6330 T_OBJECT); 6331 6332 // Add memory barrier to prevent commoning reads from this field 6333 // across safepoint since GC can change its value. 6334 insert_mem_bar(Op_MemBarCPUOrder); 6335 6336 set_result(result); 6337 return true; 6338 } 6339 6340 6341 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 6342 bool is_exact=true, bool is_static=false) { 6343 6344 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 6345 assert(tinst != NULL, "obj is null"); 6346 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 6347 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 6348 6349 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 6350 ciSymbol::make(fieldTypeString), 6351 is_static); 6352 if (field == NULL) return (Node *) NULL; 6353 assert (field != NULL, "undefined field"); 6354 6355 // Next code copied from Parse::do_get_xxx(): 6356 6357 // Compute address and memory type. 6358 int offset = field->offset_in_bytes(); 6359 bool is_vol = field->is_volatile(); 6360 ciType* field_klass = field->type(); 6361 assert(field_klass->is_loaded(), "should be loaded"); 6362 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 6363 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 6364 BasicType bt = field->layout_type(); 6365 6366 // Build the resultant type of the load 6367 const Type *type; 6368 if (bt == T_OBJECT) { 6369 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 6370 } else { 6371 type = Type::get_const_basic_type(bt); 6372 } 6373 6374 Node* leading_membar = NULL; 6375 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 6376 leading_membar = insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 6377 } 6378 // Build the load. 6379 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 6380 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 6381 #if INCLUDE_ALL_GCS 6382 if (UseShenandoahGC && (bt == T_OBJECT || bt == T_ARRAY)) { 6383 loadedField = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, loadedField); 6384 } 6385 #endif 6386 6387 // If reference is volatile, prevent following memory ops from 6388 // floating up past the volatile read. Also prevents commoning 6389 // another volatile read. 6390 if (is_vol) { 6391 // Memory barrier includes bogus read of value to force load BEFORE membar 6392 Node* mb = insert_mem_bar(Op_MemBarAcquire, loadedField); 6393 mb->as_MemBar()->set_trailing_load(); 6394 } 6395 return loadedField; 6396 } 6397 6398 6399 //------------------------------inline_aescrypt_Block----------------------- 6400 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 6401 address stubAddr = NULL; 6402 const char *stubName; 6403 assert(UseAES, "need AES instruction support"); 6404 6405 switch(id) { 6406 case vmIntrinsics::_aescrypt_encryptBlock: 6407 stubAddr = StubRoutines::aescrypt_encryptBlock(); 6408 stubName = "aescrypt_encryptBlock"; 6409 break; 6410 case vmIntrinsics::_aescrypt_decryptBlock: 6411 stubAddr = StubRoutines::aescrypt_decryptBlock(); 6412 stubName = "aescrypt_decryptBlock"; 6413 break; 6414 } 6415 if (stubAddr == NULL) return false; 6416 6417 Node* aescrypt_object = argument(0); 6418 Node* src = argument(1); 6419 Node* src_offset = argument(2); 6420 Node* dest = argument(3); 6421 Node* dest_offset = argument(4); 6422 6423 // (1) src and dest are arrays. 6424 const Type* src_type = src->Value(&_gvn); 6425 const Type* dest_type = dest->Value(&_gvn); 6426 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6427 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6428 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6429 6430 // for the quick and dirty code we will skip all the checks. 6431 // we are just trying to get the call to be generated. 6432 Node* src_start = src; 6433 Node* dest_start = dest; 6434 if (src_offset != NULL || dest_offset != NULL) { 6435 assert(src_offset != NULL && dest_offset != NULL, ""); 6436 src_start = array_element_address(src, src_offset, T_BYTE); 6437 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6438 } 6439 6440 // now need to get the start of its expanded key array 6441 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6442 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6443 if (k_start == NULL) return false; 6444 6445 if (Matcher::pass_original_key_for_aes()) { 6446 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6447 // compatibility issues between Java key expansion and SPARC crypto instructions 6448 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6449 if (original_k_start == NULL) return false; 6450 6451 // Call the stub. 6452 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6453 stubAddr, stubName, TypePtr::BOTTOM, 6454 src_start, dest_start, k_start, original_k_start); 6455 } else { 6456 // Call the stub. 6457 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 6458 stubAddr, stubName, TypePtr::BOTTOM, 6459 src_start, dest_start, k_start); 6460 } 6461 6462 return true; 6463 } 6464 6465 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 6466 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 6467 address stubAddr = NULL; 6468 const char *stubName = NULL; 6469 6470 assert(UseAES, "need AES instruction support"); 6471 6472 switch(id) { 6473 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 6474 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 6475 stubName = "cipherBlockChaining_encryptAESCrypt"; 6476 break; 6477 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 6478 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 6479 stubName = "cipherBlockChaining_decryptAESCrypt"; 6480 break; 6481 } 6482 if (stubAddr == NULL) return false; 6483 6484 Node* cipherBlockChaining_object = argument(0); 6485 Node* src = argument(1); 6486 Node* src_offset = argument(2); 6487 Node* len = argument(3); 6488 Node* dest = argument(4); 6489 Node* dest_offset = argument(5); 6490 6491 // (1) src and dest are arrays. 6492 const Type* src_type = src->Value(&_gvn); 6493 const Type* dest_type = dest->Value(&_gvn); 6494 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6495 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 6496 assert (top_src != NULL && top_src->klass() != NULL 6497 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 6498 6499 // checks are the responsibility of the caller 6500 Node* src_start = src; 6501 Node* dest_start = dest; 6502 if (src_offset != NULL || dest_offset != NULL) { 6503 assert(src_offset != NULL && dest_offset != NULL, ""); 6504 src_start = array_element_address(src, src_offset, T_BYTE); 6505 dest_start = array_element_address(dest, dest_offset, T_BYTE); 6506 } 6507 6508 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 6509 // (because of the predicated logic executed earlier). 6510 // so we cast it here safely. 6511 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 6512 6513 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6514 if (embeddedCipherObj == NULL) return false; 6515 6516 // cast it to what we know it will be at runtime 6517 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 6518 assert(tinst != NULL, "CBC obj is null"); 6519 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 6520 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6521 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 6522 6523 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6524 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6525 const TypeOopPtr* xtype = aklass->as_instance_type(); 6526 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype); 6527 aescrypt_object = _gvn.transform(aescrypt_object); 6528 6529 // we need to get the start of the aescrypt_object's expanded key array 6530 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6531 if (k_start == NULL) return false; 6532 6533 // similarly, get the start address of the r vector 6534 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6535 if (objRvec == NULL) return false; 6536 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6537 6538 Node* cbcCrypt; 6539 if (Matcher::pass_original_key_for_aes()) { 6540 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6541 // compatibility issues between Java key expansion and SPARC crypto instructions 6542 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6543 if (original_k_start == NULL) return false; 6544 6545 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6546 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6547 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6548 stubAddr, stubName, TypePtr::BOTTOM, 6549 src_start, dest_start, k_start, r_start, len, original_k_start); 6550 } else { 6551 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6552 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6553 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6554 stubAddr, stubName, TypePtr::BOTTOM, 6555 src_start, dest_start, k_start, r_start, len); 6556 } 6557 6558 // return cipher length (int) 6559 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms)); 6560 set_result(retvalue); 6561 return true; 6562 } 6563 6564 //------------------------------get_key_start_from_aescrypt_object----------------------- 6565 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6566 #ifdef PPC64 6567 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys. 6568 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns. 6569 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption. 6570 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]). 6571 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false); 6572 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6573 if (objSessionK == NULL) { 6574 return (Node *) NULL; 6575 } 6576 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS); 6577 #else 6578 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6579 #endif // PPC64 6580 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6581 if (objAESCryptKey == NULL) return (Node *) NULL; 6582 6583 // now have the array, need to get the start address of the K array 6584 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6585 return k_start; 6586 } 6587 6588 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6589 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6590 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6591 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6592 if (objAESCryptKey == NULL) return (Node *) NULL; 6593 6594 // now have the array, need to get the start address of the lastKey array 6595 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6596 return original_k_start; 6597 } 6598 6599 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6600 // Return node representing slow path of predicate check. 6601 // the pseudo code we want to emulate with this predicate is: 6602 // for encryption: 6603 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6604 // for decryption: 6605 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6606 // note cipher==plain is more conservative than the original java code but that's OK 6607 // 6608 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6609 // The receiver was checked for NULL already. 6610 Node* objCBC = argument(0); 6611 6612 // Load embeddedCipher field of CipherBlockChaining object. 6613 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6614 6615 // get AESCrypt klass for instanceOf check 6616 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6617 // will have same classloader as CipherBlockChaining object 6618 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6619 assert(tinst != NULL, "CBCobj is null"); 6620 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6621 6622 // we want to do an instanceof comparison against the AESCrypt class 6623 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6624 if (!klass_AESCrypt->is_loaded()) { 6625 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6626 Node* ctrl = control(); 6627 set_control(top()); // no regular fast path 6628 return ctrl; 6629 } 6630 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6631 6632 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6633 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1))); 6634 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 6635 6636 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6637 6638 // for encryption, we are done 6639 if (!decrypting) 6640 return instof_false; // even if it is NULL 6641 6642 // for decryption, we need to add a further check to avoid 6643 // taking the intrinsic path when cipher and plain are the same 6644 // see the original java code for why. 6645 RegionNode* region = new(C) RegionNode(3); 6646 region->init_req(1, instof_false); 6647 Node* src = argument(1); 6648 Node* dest = argument(4); 6649 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest)); 6650 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq)); 6651 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6652 region->init_req(2, src_dest_conjoint); 6653 6654 record_for_igvn(region); 6655 return _gvn.transform(region); 6656 } 6657 6658 //------------------------------inline_sha_implCompress----------------------- 6659 // 6660 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6661 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6662 // 6663 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6664 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6665 // 6666 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6667 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6668 // 6669 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6670 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6671 6672 Node* sha_obj = argument(0); 6673 Node* src = argument(1); // type oop 6674 Node* ofs = argument(2); // type int 6675 6676 const Type* src_type = src->Value(&_gvn); 6677 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6678 if (top_src == NULL || top_src->klass() == NULL) { 6679 // failed array check 6680 return false; 6681 } 6682 // Figure out the size and type of the elements we will be copying. 6683 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6684 if (src_elem != T_BYTE) { 6685 return false; 6686 } 6687 // 'src_start' points to src array + offset 6688 Node* src_start = array_element_address(src, ofs, src_elem); 6689 Node* state = NULL; 6690 address stubAddr; 6691 const char *stubName; 6692 6693 switch(id) { 6694 case vmIntrinsics::_sha_implCompress: 6695 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6696 state = get_state_from_sha_object(sha_obj); 6697 stubAddr = StubRoutines::sha1_implCompress(); 6698 stubName = "sha1_implCompress"; 6699 break; 6700 case vmIntrinsics::_sha2_implCompress: 6701 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6702 state = get_state_from_sha_object(sha_obj); 6703 stubAddr = StubRoutines::sha256_implCompress(); 6704 stubName = "sha256_implCompress"; 6705 break; 6706 case vmIntrinsics::_sha5_implCompress: 6707 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6708 state = get_state_from_sha5_object(sha_obj); 6709 stubAddr = StubRoutines::sha512_implCompress(); 6710 stubName = "sha512_implCompress"; 6711 break; 6712 default: 6713 fatal_unexpected_iid(id); 6714 return false; 6715 } 6716 if (state == NULL) return false; 6717 6718 // Call the stub. 6719 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6720 stubAddr, stubName, TypePtr::BOTTOM, 6721 src_start, state); 6722 6723 return true; 6724 } 6725 6726 //------------------------------inline_digestBase_implCompressMB----------------------- 6727 // 6728 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6729 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6730 // 6731 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6732 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6733 "need SHA1/SHA256/SHA512 instruction support"); 6734 assert((uint)predicate < 3, "sanity"); 6735 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6736 6737 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6738 Node* src = argument(1); // byte[] array 6739 Node* ofs = argument(2); // type int 6740 Node* limit = argument(3); // type int 6741 6742 const Type* src_type = src->Value(&_gvn); 6743 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6744 if (top_src == NULL || top_src->klass() == NULL) { 6745 // failed array check 6746 return false; 6747 } 6748 // Figure out the size and type of the elements we will be copying. 6749 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6750 if (src_elem != T_BYTE) { 6751 return false; 6752 } 6753 // 'src_start' points to src array + offset 6754 Node* src_start = array_element_address(src, ofs, src_elem); 6755 6756 const char* klass_SHA_name = NULL; 6757 const char* stub_name = NULL; 6758 address stub_addr = NULL; 6759 bool long_state = false; 6760 6761 switch (predicate) { 6762 case 0: 6763 if (UseSHA1Intrinsics) { 6764 klass_SHA_name = "sun/security/provider/SHA"; 6765 stub_name = "sha1_implCompressMB"; 6766 stub_addr = StubRoutines::sha1_implCompressMB(); 6767 } 6768 break; 6769 case 1: 6770 if (UseSHA256Intrinsics) { 6771 klass_SHA_name = "sun/security/provider/SHA2"; 6772 stub_name = "sha256_implCompressMB"; 6773 stub_addr = StubRoutines::sha256_implCompressMB(); 6774 } 6775 break; 6776 case 2: 6777 if (UseSHA512Intrinsics) { 6778 klass_SHA_name = "sun/security/provider/SHA5"; 6779 stub_name = "sha512_implCompressMB"; 6780 stub_addr = StubRoutines::sha512_implCompressMB(); 6781 long_state = true; 6782 } 6783 break; 6784 default: 6785 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6786 } 6787 if (klass_SHA_name != NULL) { 6788 // get DigestBase klass to lookup for SHA klass 6789 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6790 assert(tinst != NULL, "digestBase_obj is not instance???"); 6791 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6792 6793 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6794 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6795 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6796 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6797 } 6798 return false; 6799 } 6800 //------------------------------inline_sha_implCompressMB----------------------- 6801 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6802 bool long_state, address stubAddr, const char *stubName, 6803 Node* src_start, Node* ofs, Node* limit) { 6804 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6805 const TypeOopPtr* xtype = aklass->as_instance_type(); 6806 Node* sha_obj = new (C) CheckCastPPNode(control(), digestBase_obj, xtype); 6807 sha_obj = _gvn.transform(sha_obj); 6808 6809 Node* state; 6810 if (long_state) { 6811 state = get_state_from_sha5_object(sha_obj); 6812 } else { 6813 state = get_state_from_sha_object(sha_obj); 6814 } 6815 if (state == NULL) return false; 6816 6817 // Call the stub. 6818 Node *call; 6819 if (CCallingConventionRequiresIntsAsLongs) { 6820 call = make_runtime_call(RC_LEAF|RC_NO_FP, 6821 OptoRuntime::digestBase_implCompressMB_Type(), 6822 stubAddr, stubName, TypePtr::BOTTOM, 6823 src_start, state, ofs XTOP, limit XTOP); 6824 } else { 6825 call = make_runtime_call(RC_LEAF|RC_NO_FP, 6826 OptoRuntime::digestBase_implCompressMB_Type(), 6827 stubAddr, stubName, TypePtr::BOTTOM, 6828 src_start, state, ofs, limit); 6829 } 6830 // return ofs (int) 6831 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms)); 6832 set_result(result); 6833 6834 return true; 6835 } 6836 6837 //------------------------------get_state_from_sha_object----------------------- 6838 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6839 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6840 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6841 if (sha_state == NULL) return (Node *) NULL; 6842 6843 // now have the array, need to get the start address of the state array 6844 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6845 return state; 6846 } 6847 6848 //------------------------------get_state_from_sha5_object----------------------- 6849 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6850 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6851 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6852 if (sha_state == NULL) return (Node *) NULL; 6853 6854 // now have the array, need to get the start address of the state array 6855 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6856 return state; 6857 } 6858 6859 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6860 // Return node representing slow path of predicate check. 6861 // the pseudo code we want to emulate with this predicate is: 6862 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6863 // 6864 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6865 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6866 "need SHA1/SHA256/SHA512 instruction support"); 6867 assert((uint)predicate < 3, "sanity"); 6868 6869 // The receiver was checked for NULL already. 6870 Node* digestBaseObj = argument(0); 6871 6872 // get DigestBase klass for instanceOf check 6873 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6874 assert(tinst != NULL, "digestBaseObj is null"); 6875 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6876 6877 const char* klass_SHA_name = NULL; 6878 switch (predicate) { 6879 case 0: 6880 if (UseSHA1Intrinsics) { 6881 // we want to do an instanceof comparison against the SHA class 6882 klass_SHA_name = "sun/security/provider/SHA"; 6883 } 6884 break; 6885 case 1: 6886 if (UseSHA256Intrinsics) { 6887 // we want to do an instanceof comparison against the SHA2 class 6888 klass_SHA_name = "sun/security/provider/SHA2"; 6889 } 6890 break; 6891 case 2: 6892 if (UseSHA512Intrinsics) { 6893 // we want to do an instanceof comparison against the SHA5 class 6894 klass_SHA_name = "sun/security/provider/SHA5"; 6895 } 6896 break; 6897 default: 6898 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6899 } 6900 6901 ciKlass* klass_SHA = NULL; 6902 if (klass_SHA_name != NULL) { 6903 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6904 } 6905 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6906 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6907 Node* ctrl = control(); 6908 set_control(top()); // no intrinsic path 6909 return ctrl; 6910 } 6911 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6912 6913 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6914 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instofSHA, intcon(1))); 6915 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne)); 6916 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6917 6918 return instof_false; // even if it is NULL 6919 } 6920 6921 bool LibraryCallKit::inline_profileBoolean() { 6922 Node* counts = argument(1); 6923 const TypeAryPtr* ary = NULL; 6924 ciArray* aobj = NULL; 6925 if (counts->is_Con() 6926 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6927 && (aobj = ary->const_oop()->as_array()) != NULL 6928 && (aobj->length() == 2)) { 6929 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6930 jint false_cnt = aobj->element_value(0).as_int(); 6931 jint true_cnt = aobj->element_value(1).as_int(); 6932 6933 if (C->log() != NULL) { 6934 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6935 false_cnt, true_cnt); 6936 } 6937 6938 if (false_cnt + true_cnt == 0) { 6939 // According to profile, never executed. 6940 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6941 Deoptimization::Action_reinterpret); 6942 return true; 6943 } 6944 6945 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6946 // is a number of each value occurrences. 6947 Node* result = argument(0); 6948 if (false_cnt == 0 || true_cnt == 0) { 6949 // According to profile, one value has been never seen. 6950 int expected_val = (false_cnt == 0) ? 1 : 0; 6951 6952 Node* cmp = _gvn.transform(new (C) CmpINode(result, intcon(expected_val))); 6953 Node* test = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq)); 6954 6955 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6956 Node* fast_path = _gvn.transform(new (C) IfTrueNode(check)); 6957 Node* slow_path = _gvn.transform(new (C) IfFalseNode(check)); 6958 6959 { // Slow path: uncommon trap for never seen value and then reexecute 6960 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6961 // the value has been seen at least once. 6962 PreserveJVMState pjvms(this); 6963 PreserveReexecuteState preexecs(this); 6964 jvms()->set_should_reexecute(true); 6965 6966 set_control(slow_path); 6967 set_i_o(i_o()); 6968 6969 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6970 Deoptimization::Action_reinterpret); 6971 } 6972 // The guard for never seen value enables sharpening of the result and 6973 // returning a constant. It allows to eliminate branches on the same value 6974 // later on. 6975 set_control(fast_path); 6976 result = intcon(expected_val); 6977 } 6978 // Stop profiling. 6979 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6980 // By replacing method body with profile data (represented as ProfileBooleanNode 6981 // on IR level) we effectively disable profiling. 6982 // It enables full speed execution once optimized code is generated. 6983 Node* profile = _gvn.transform(new (C) ProfileBooleanNode(result, false_cnt, true_cnt)); 6984 C->record_for_igvn(profile); 6985 set_result(profile); 6986 return true; 6987 } else { 6988 // Continue profiling. 6989 // Profile data isn't available at the moment. So, execute method's bytecode version. 6990 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6991 // is compiled and counters aren't available since corresponding MethodHandle 6992 // isn't a compile-time constant. 6993 return false; 6994 } 6995 }