1 /* 2 * Copyright (c) 1999, 2015, 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 "asm/macroAssembler.hpp" 27 #include "classfile/systemDictionary.hpp" 28 #include "classfile/vmSymbols.hpp" 29 #include "compiler/compileBroker.hpp" 30 #include "compiler/compileLog.hpp" 31 #include "oops/objArrayKlass.hpp" 32 #include "opto/addnode.hpp" 33 #include "opto/arraycopynode.hpp" 34 #include "opto/callGenerator.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/cfgnode.hpp" 37 #include "opto/convertnode.hpp" 38 #include "opto/countbitsnode.hpp" 39 #include "opto/intrinsicnode.hpp" 40 #include "opto/idealKit.hpp" 41 #include "opto/mathexactnode.hpp" 42 #include "opto/movenode.hpp" 43 #include "opto/mulnode.hpp" 44 #include "opto/narrowptrnode.hpp" 45 #include "opto/opaquenode.hpp" 46 #include "opto/parse.hpp" 47 #include "opto/runtime.hpp" 48 #include "opto/subnode.hpp" 49 #include "prims/nativeLookup.hpp" 50 #include "runtime/sharedRuntime.hpp" 51 #include "trace/traceMacros.hpp" 52 53 class LibraryIntrinsic : public InlineCallGenerator { 54 // Extend the set of intrinsics known to the runtime: 55 public: 56 private: 57 bool _is_virtual; 58 bool _does_virtual_dispatch; 59 int8_t _predicates_count; // Intrinsic is predicated by several conditions 60 int8_t _last_predicate; // Last generated predicate 61 vmIntrinsics::ID _intrinsic_id; 62 63 public: 64 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id) 65 : InlineCallGenerator(m), 66 _is_virtual(is_virtual), 67 _does_virtual_dispatch(does_virtual_dispatch), 68 _predicates_count((int8_t)predicates_count), 69 _last_predicate((int8_t)-1), 70 _intrinsic_id(id) 71 { 72 } 73 virtual bool is_intrinsic() const { return true; } 74 virtual bool is_virtual() const { return _is_virtual; } 75 virtual bool is_predicated() const { return _predicates_count > 0; } 76 virtual int predicates_count() const { return _predicates_count; } 77 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; } 78 virtual JVMState* generate(JVMState* jvms); 79 virtual Node* generate_predicate(JVMState* jvms, int predicate); 80 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; } 81 }; 82 83 84 // Local helper class for LibraryIntrinsic: 85 class LibraryCallKit : public GraphKit { 86 private: 87 LibraryIntrinsic* _intrinsic; // the library intrinsic being called 88 Node* _result; // the result node, if any 89 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted 90 91 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false); 92 93 public: 94 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic) 95 : GraphKit(jvms), 96 _intrinsic(intrinsic), 97 _result(NULL) 98 { 99 // Check if this is a root compile. In that case we don't have a caller. 100 if (!jvms->has_method()) { 101 _reexecute_sp = sp(); 102 } else { 103 // Find out how many arguments the interpreter needs when deoptimizing 104 // and save the stack pointer value so it can used by uncommon_trap. 105 // We find the argument count by looking at the declared signature. 106 bool ignored_will_link; 107 ciSignature* declared_signature = NULL; 108 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature); 109 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci())); 110 _reexecute_sp = sp() + nargs; // "push" arguments back on stack 111 } 112 } 113 114 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; } 115 116 ciMethod* caller() const { return jvms()->method(); } 117 int bci() const { return jvms()->bci(); } 118 LibraryIntrinsic* intrinsic() const { return _intrinsic; } 119 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); } 120 ciMethod* callee() const { return _intrinsic->method(); } 121 122 bool try_to_inline(int predicate); 123 Node* try_to_predicate(int predicate); 124 125 void push_result() { 126 // Push the result onto the stack. 127 if (!stopped() && result() != NULL) { 128 BasicType bt = result()->bottom_type()->basic_type(); 129 push_node(bt, result()); 130 } 131 } 132 133 private: 134 void fatal_unexpected_iid(vmIntrinsics::ID iid) { 135 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid))); 136 } 137 138 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; } 139 void set_result(RegionNode* region, PhiNode* value); 140 Node* result() { return _result; } 141 142 virtual int reexecute_sp() { return _reexecute_sp; } 143 144 // Helper functions to inline natives 145 Node* generate_guard(Node* test, RegionNode* region, float true_prob); 146 Node* generate_slow_guard(Node* test, RegionNode* region); 147 Node* generate_fair_guard(Node* test, RegionNode* region); 148 Node* generate_negative_guard(Node* index, RegionNode* region, 149 // resulting CastII of index: 150 Node* *pos_index = NULL); 151 Node* generate_limit_guard(Node* offset, Node* subseq_length, 152 Node* array_length, 153 RegionNode* region); 154 Node* generate_current_thread(Node* &tls_output); 155 Node* load_mirror_from_klass(Node* klass); 156 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null, 157 RegionNode* region, int null_path, 158 int offset); 159 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, 160 RegionNode* region, int null_path) { 161 int offset = java_lang_Class::klass_offset_in_bytes(); 162 return load_klass_from_mirror_common(mirror, never_see_null, 163 region, null_path, 164 offset); 165 } 166 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null, 167 RegionNode* region, int null_path) { 168 int offset = java_lang_Class::array_klass_offset_in_bytes(); 169 return load_klass_from_mirror_common(mirror, never_see_null, 170 region, null_path, 171 offset); 172 } 173 Node* generate_access_flags_guard(Node* kls, 174 int modifier_mask, int modifier_bits, 175 RegionNode* region); 176 Node* generate_interface_guard(Node* kls, RegionNode* region); 177 Node* generate_array_guard(Node* kls, RegionNode* region) { 178 return generate_array_guard_common(kls, region, false, false); 179 } 180 Node* generate_non_array_guard(Node* kls, RegionNode* region) { 181 return generate_array_guard_common(kls, region, false, true); 182 } 183 Node* generate_objArray_guard(Node* kls, RegionNode* region) { 184 return generate_array_guard_common(kls, region, true, false); 185 } 186 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) { 187 return generate_array_guard_common(kls, region, true, true); 188 } 189 Node* generate_array_guard_common(Node* kls, RegionNode* region, 190 bool obj_array, bool not_array); 191 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region); 192 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id, 193 bool is_virtual = false, bool is_static = false); 194 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) { 195 return generate_method_call(method_id, false, true); 196 } 197 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) { 198 return generate_method_call(method_id, true, false); 199 } 200 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls); 201 202 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2); 203 Node* make_string_method_node(int opcode, Node* str1, Node* str2); 204 bool inline_string_compareTo(); 205 bool inline_string_indexOf(); 206 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i); 207 bool inline_string_equals(); 208 Node* round_double_node(Node* n); 209 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName); 210 bool inline_math_native(vmIntrinsics::ID id); 211 bool inline_trig(vmIntrinsics::ID id); 212 bool inline_math(vmIntrinsics::ID id); 213 template <typename OverflowOp> 214 bool inline_math_overflow(Node* arg1, Node* arg2); 215 void inline_math_mathExact(Node* math, Node* test); 216 bool inline_math_addExactI(bool is_increment); 217 bool inline_math_addExactL(bool is_increment); 218 bool inline_math_multiplyExactI(); 219 bool inline_math_multiplyExactL(); 220 bool inline_math_negateExactI(); 221 bool inline_math_negateExactL(); 222 bool inline_math_subtractExactI(bool is_decrement); 223 bool inline_math_subtractExactL(bool is_decrement); 224 bool inline_exp(); 225 bool inline_pow(); 226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName); 227 bool inline_min_max(vmIntrinsics::ID id); 228 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 229 // This returns Type::AnyPtr, RawPtr, or OopPtr. 230 int classify_unsafe_addr(Node* &base, Node* &offset); 231 Node* make_unsafe_address(Node* base, Node* offset); 232 // Helper for inline_unsafe_access. 233 // Generates the guards that check whether the result of 234 // Unsafe.getObject should be recorded in an SATB log buffer. 235 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 236 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 237 static bool klass_needs_init_guard(Node* kls); 238 bool inline_unsafe_allocate(); 239 bool inline_unsafe_copyMemory(); 240 bool inline_native_currentThread(); 241 #ifdef TRACE_HAVE_INTRINSICS 242 bool inline_native_classID(); 243 bool inline_native_threadID(); 244 #endif 245 bool inline_native_time_funcs(address method, const char* funcName); 246 bool inline_native_isInterrupted(); 247 bool inline_native_Class_query(vmIntrinsics::ID id); 248 bool inline_native_subtype_check(); 249 250 bool inline_native_newArray(); 251 bool inline_native_getLength(); 252 bool inline_array_copyOf(bool is_copyOfRange); 253 bool inline_array_equals(); 254 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 255 bool inline_native_clone(bool is_virtual); 256 bool inline_native_Reflection_getCallerClass(); 257 // Helper function for inlining native object hash method 258 bool inline_native_hashcode(bool is_virtual, bool is_static); 259 bool inline_native_getClass(); 260 261 // Helper functions for inlining arraycopy 262 bool inline_arraycopy(); 263 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 264 RegionNode* slow_region); 265 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 266 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 267 268 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 269 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 270 bool inline_unsafe_ordered_store(BasicType type); 271 bool inline_unsafe_fence(vmIntrinsics::ID id); 272 bool inline_fp_conversions(vmIntrinsics::ID id); 273 bool inline_number_methods(vmIntrinsics::ID id); 274 bool inline_reference_get(); 275 bool inline_Class_cast(); 276 bool inline_aescrypt_Block(vmIntrinsics::ID id); 277 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 278 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 279 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 280 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 281 bool inline_ghash_processBlocks(); 282 bool inline_sha_implCompress(vmIntrinsics::ID id); 283 bool inline_digestBase_implCompressMB(int predicate); 284 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 285 bool long_state, address stubAddr, const char *stubName, 286 Node* src_start, Node* ofs, Node* limit); 287 Node* get_state_from_sha_object(Node *sha_object); 288 Node* get_state_from_sha5_object(Node *sha_object); 289 Node* inline_digestBase_implCompressMB_predicate(int predicate); 290 bool inline_encodeISOArray(); 291 bool inline_updateCRC32(); 292 bool inline_updateBytesCRC32(); 293 bool inline_updateByteBufferCRC32(); 294 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 295 bool inline_updateBytesCRC32C(); 296 bool inline_updateDirectByteBufferCRC32C(); 297 bool inline_multiplyToLen(); 298 bool inline_squareToLen(); 299 bool inline_mulAdd(); 300 bool inline_montgomeryMultiply(); 301 bool inline_montgomerySquare(); 302 303 bool inline_profileBoolean(); 304 bool inline_isCompileConstant(); 305 }; 306 307 308 //---------------------------make_vm_intrinsic---------------------------- 309 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 310 vmIntrinsics::ID id = m->intrinsic_id(); 311 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 312 313 ccstr disable_intr = NULL; 314 315 if ((DisableIntrinsic[0] != '\0' 316 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 317 (method_has_option_value("DisableIntrinsic", disable_intr) 318 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 319 // disabled by a user request on the command line: 320 // example: -XX:DisableIntrinsic=_hashCode,_getClass 321 return NULL; 322 } 323 324 if (!m->is_loaded()) { 325 // do not attempt to inline unloaded methods 326 return NULL; 327 } 328 329 // Only a few intrinsics implement a virtual dispatch. 330 // They are expensive calls which are also frequently overridden. 331 if (is_virtual) { 332 switch (id) { 333 case vmIntrinsics::_hashCode: 334 case vmIntrinsics::_clone: 335 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 336 break; 337 default: 338 return NULL; 339 } 340 } 341 342 // -XX:-InlineNatives disables nearly all intrinsics: 343 if (!InlineNatives) { 344 switch (id) { 345 case vmIntrinsics::_indexOf: 346 case vmIntrinsics::_compareTo: 347 case vmIntrinsics::_equals: 348 case vmIntrinsics::_equalsC: 349 case vmIntrinsics::_getAndAddInt: 350 case vmIntrinsics::_getAndAddLong: 351 case vmIntrinsics::_getAndSetInt: 352 case vmIntrinsics::_getAndSetLong: 353 case vmIntrinsics::_getAndSetObject: 354 case vmIntrinsics::_loadFence: 355 case vmIntrinsics::_storeFence: 356 case vmIntrinsics::_fullFence: 357 break; // InlineNatives does not control String.compareTo 358 case vmIntrinsics::_Reference_get: 359 break; // InlineNatives does not control Reference.get 360 default: 361 return NULL; 362 } 363 } 364 365 int predicates = 0; 366 bool does_virtual_dispatch = false; 367 368 switch (id) { 369 case vmIntrinsics::_compareTo: 370 if (!SpecialStringCompareTo) return NULL; 371 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 372 break; 373 case vmIntrinsics::_indexOf: 374 if (!SpecialStringIndexOf) return NULL; 375 break; 376 case vmIntrinsics::_equals: 377 if (!SpecialStringEquals) return NULL; 378 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 379 break; 380 case vmIntrinsics::_equalsC: 381 if (!SpecialArraysEquals) return NULL; 382 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 383 break; 384 case vmIntrinsics::_arraycopy: 385 if (!InlineArrayCopy) return NULL; 386 break; 387 case vmIntrinsics::_copyMemory: 388 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 389 if (!InlineArrayCopy) return NULL; 390 break; 391 case vmIntrinsics::_hashCode: 392 if (!InlineObjectHash) return NULL; 393 does_virtual_dispatch = true; 394 break; 395 case vmIntrinsics::_clone: 396 does_virtual_dispatch = true; 397 case vmIntrinsics::_copyOf: 398 case vmIntrinsics::_copyOfRange: 399 if (!InlineObjectCopy) return NULL; 400 // These also use the arraycopy intrinsic mechanism: 401 if (!InlineArrayCopy) return NULL; 402 break; 403 case vmIntrinsics::_encodeISOArray: 404 if (!SpecialEncodeISOArray) return NULL; 405 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 406 break; 407 case vmIntrinsics::_checkIndex: 408 // We do not intrinsify this. The optimizer does fine with it. 409 return NULL; 410 411 case vmIntrinsics::_getCallerClass: 412 if (!InlineReflectionGetCallerClass) return NULL; 413 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 414 break; 415 416 case vmIntrinsics::_bitCount_i: 417 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 418 break; 419 420 case vmIntrinsics::_bitCount_l: 421 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 422 break; 423 424 case vmIntrinsics::_numberOfLeadingZeros_i: 425 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 426 break; 427 428 case vmIntrinsics::_numberOfLeadingZeros_l: 429 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 430 break; 431 432 case vmIntrinsics::_numberOfTrailingZeros_i: 433 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 434 break; 435 436 case vmIntrinsics::_numberOfTrailingZeros_l: 437 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 438 break; 439 440 case vmIntrinsics::_reverseBytes_c: 441 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 442 break; 443 case vmIntrinsics::_reverseBytes_s: 444 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 445 break; 446 case vmIntrinsics::_reverseBytes_i: 447 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 448 break; 449 case vmIntrinsics::_reverseBytes_l: 450 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 451 break; 452 453 case vmIntrinsics::_Reference_get: 454 // Use the intrinsic version of Reference.get() so that the value in 455 // the referent field can be registered by the G1 pre-barrier code. 456 // Also add memory barrier to prevent commoning reads from this field 457 // across safepoint since GC can change it value. 458 break; 459 460 case vmIntrinsics::_compareAndSwapObject: 461 #ifdef _LP64 462 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 463 #endif 464 break; 465 466 case vmIntrinsics::_compareAndSwapLong: 467 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 468 break; 469 470 case vmIntrinsics::_getAndAddInt: 471 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 472 break; 473 474 case vmIntrinsics::_getAndAddLong: 475 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 476 break; 477 478 case vmIntrinsics::_getAndSetInt: 479 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 480 break; 481 482 case vmIntrinsics::_getAndSetLong: 483 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 484 break; 485 486 case vmIntrinsics::_getAndSetObject: 487 #ifdef _LP64 488 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 489 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 490 break; 491 #else 492 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 493 break; 494 #endif 495 496 case vmIntrinsics::_aescrypt_encryptBlock: 497 case vmIntrinsics::_aescrypt_decryptBlock: 498 if (!UseAESIntrinsics) return NULL; 499 break; 500 501 case vmIntrinsics::_multiplyToLen: 502 if (!UseMultiplyToLenIntrinsic) return NULL; 503 break; 504 505 case vmIntrinsics::_squareToLen: 506 if (!UseSquareToLenIntrinsic) return NULL; 507 break; 508 509 case vmIntrinsics::_mulAdd: 510 if (!UseMulAddIntrinsic) return NULL; 511 break; 512 513 case vmIntrinsics::_montgomeryMultiply: 514 if (!UseMontgomeryMultiplyIntrinsic) return NULL; 515 break; 516 case vmIntrinsics::_montgomerySquare: 517 if (!UseMontgomerySquareIntrinsic) return NULL; 518 break; 519 520 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 521 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 522 if (!UseAESIntrinsics) return NULL; 523 // these two require the predicated logic 524 predicates = 1; 525 break; 526 527 case vmIntrinsics::_sha_implCompress: 528 if (!UseSHA1Intrinsics) return NULL; 529 break; 530 531 case vmIntrinsics::_sha2_implCompress: 532 if (!UseSHA256Intrinsics) return NULL; 533 break; 534 535 case vmIntrinsics::_sha5_implCompress: 536 if (!UseSHA512Intrinsics) return NULL; 537 break; 538 539 case vmIntrinsics::_digestBase_implCompressMB: 540 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 541 predicates = 3; 542 break; 543 544 case vmIntrinsics::_ghash_processBlocks: 545 if (!UseGHASHIntrinsics) return NULL; 546 break; 547 548 case vmIntrinsics::_updateCRC32: 549 case vmIntrinsics::_updateBytesCRC32: 550 case vmIntrinsics::_updateByteBufferCRC32: 551 if (!UseCRC32Intrinsics) return NULL; 552 break; 553 554 case vmIntrinsics::_updateBytesCRC32C: 555 case vmIntrinsics::_updateDirectByteBufferCRC32C: 556 if (!UseCRC32CIntrinsics) return NULL; 557 break; 558 559 case vmIntrinsics::_incrementExactI: 560 case vmIntrinsics::_addExactI: 561 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 562 break; 563 case vmIntrinsics::_incrementExactL: 564 case vmIntrinsics::_addExactL: 565 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 566 break; 567 case vmIntrinsics::_decrementExactI: 568 case vmIntrinsics::_subtractExactI: 569 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 570 break; 571 case vmIntrinsics::_decrementExactL: 572 case vmIntrinsics::_subtractExactL: 573 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 574 break; 575 case vmIntrinsics::_negateExactI: 576 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 577 break; 578 case vmIntrinsics::_negateExactL: 579 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 580 break; 581 case vmIntrinsics::_multiplyExactI: 582 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 583 break; 584 case vmIntrinsics::_multiplyExactL: 585 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 586 break; 587 588 case vmIntrinsics::_getShortUnaligned: 589 case vmIntrinsics::_getCharUnaligned: 590 case vmIntrinsics::_getIntUnaligned: 591 case vmIntrinsics::_getLongUnaligned: 592 case vmIntrinsics::_putShortUnaligned: 593 case vmIntrinsics::_putCharUnaligned: 594 case vmIntrinsics::_putIntUnaligned: 595 case vmIntrinsics::_putLongUnaligned: 596 if (!UseUnalignedAccesses) return NULL; 597 break; 598 599 default: 600 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 601 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 602 break; 603 } 604 605 // -XX:-InlineClassNatives disables natives from the Class class. 606 // The flag applies to all reflective calls, notably Array.newArray 607 // (visible to Java programmers as Array.newInstance). 608 if (m->holder()->name() == ciSymbol::java_lang_Class() || 609 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 610 if (!InlineClassNatives) return NULL; 611 } 612 613 // -XX:-InlineThreadNatives disables natives from the Thread class. 614 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 615 if (!InlineThreadNatives) return NULL; 616 } 617 618 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 619 if (m->holder()->name() == ciSymbol::java_lang_Math() || 620 m->holder()->name() == ciSymbol::java_lang_Float() || 621 m->holder()->name() == ciSymbol::java_lang_Double()) { 622 if (!InlineMathNatives) return NULL; 623 } 624 625 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 626 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 627 if (!InlineUnsafeOps) return NULL; 628 } 629 630 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 631 } 632 633 //----------------------register_library_intrinsics----------------------- 634 // Initialize this file's data structures, for each Compile instance. 635 void Compile::register_library_intrinsics() { 636 // Nothing to do here. 637 } 638 639 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 640 LibraryCallKit kit(jvms, this); 641 Compile* C = kit.C; 642 int nodes = C->unique(); 643 #ifndef PRODUCT 644 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 645 char buf[1000]; 646 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 647 tty->print_cr("Intrinsic %s", str); 648 } 649 #endif 650 ciMethod* callee = kit.callee(); 651 const int bci = kit.bci(); 652 653 // Try to inline the intrinsic. 654 if (kit.try_to_inline(_last_predicate)) { 655 if (C->print_intrinsics() || C->print_inlining()) { 656 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 657 } 658 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 659 if (C->log()) { 660 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 661 vmIntrinsics::name_at(intrinsic_id()), 662 (is_virtual() ? " virtual='1'" : ""), 663 C->unique() - nodes); 664 } 665 // Push the result from the inlined method onto the stack. 666 kit.push_result(); 667 C->print_inlining_update(this); 668 return kit.transfer_exceptions_into_jvms(); 669 } 670 671 // The intrinsic bailed out 672 if (C->print_intrinsics() || C->print_inlining()) { 673 if (jvms->has_method()) { 674 // Not a root compile. 675 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 676 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 677 } else { 678 // Root compile 679 tty->print("Did not generate intrinsic %s%s at bci:%d in", 680 vmIntrinsics::name_at(intrinsic_id()), 681 (is_virtual() ? " (virtual)" : ""), bci); 682 } 683 } 684 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 685 C->print_inlining_update(this); 686 return NULL; 687 } 688 689 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 690 LibraryCallKit kit(jvms, this); 691 Compile* C = kit.C; 692 int nodes = C->unique(); 693 _last_predicate = predicate; 694 #ifndef PRODUCT 695 assert(is_predicated() && predicate < predicates_count(), "sanity"); 696 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 697 char buf[1000]; 698 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 699 tty->print_cr("Predicate for intrinsic %s", str); 700 } 701 #endif 702 ciMethod* callee = kit.callee(); 703 const int bci = kit.bci(); 704 705 Node* slow_ctl = kit.try_to_predicate(predicate); 706 if (!kit.failing()) { 707 if (C->print_intrinsics() || C->print_inlining()) { 708 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 709 } 710 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 711 if (C->log()) { 712 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 713 vmIntrinsics::name_at(intrinsic_id()), 714 (is_virtual() ? " virtual='1'" : ""), 715 C->unique() - nodes); 716 } 717 return slow_ctl; // Could be NULL if the check folds. 718 } 719 720 // The intrinsic bailed out 721 if (C->print_intrinsics() || C->print_inlining()) { 722 if (jvms->has_method()) { 723 // Not a root compile. 724 const char* msg = "failed to generate predicate for intrinsic"; 725 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 726 } else { 727 // Root compile 728 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 729 vmIntrinsics::name_at(intrinsic_id()), 730 (is_virtual() ? " (virtual)" : ""), bci); 731 } 732 } 733 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 734 return NULL; 735 } 736 737 bool LibraryCallKit::try_to_inline(int predicate) { 738 // Handle symbolic names for otherwise undistinguished boolean switches: 739 const bool is_store = true; 740 const bool is_native_ptr = true; 741 const bool is_static = true; 742 const bool is_volatile = true; 743 744 if (!jvms()->has_method()) { 745 // Root JVMState has a null method. 746 assert(map()->memory()->Opcode() == Op_Parm, ""); 747 // Insert the memory aliasing node 748 set_all_memory(reset_memory()); 749 } 750 assert(merged_memory(), ""); 751 752 753 switch (intrinsic_id()) { 754 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 755 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 756 case vmIntrinsics::_getClass: return inline_native_getClass(); 757 758 case vmIntrinsics::_dsin: 759 case vmIntrinsics::_dcos: 760 case vmIntrinsics::_dtan: 761 case vmIntrinsics::_dabs: 762 case vmIntrinsics::_datan2: 763 case vmIntrinsics::_dsqrt: 764 case vmIntrinsics::_dexp: 765 case vmIntrinsics::_dlog: 766 case vmIntrinsics::_dlog10: 767 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 768 769 case vmIntrinsics::_min: 770 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 771 772 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 773 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 774 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 775 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 776 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 777 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 778 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 779 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 780 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 781 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 782 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 783 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 784 785 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 786 787 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 788 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 789 case vmIntrinsics::_equals: return inline_string_equals(); 790 791 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 792 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 793 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 794 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 795 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 796 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 797 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 798 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 799 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 800 801 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 802 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 803 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 804 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 805 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 806 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 807 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 808 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 809 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 810 811 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 812 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 813 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 814 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 815 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 816 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 817 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 818 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 819 820 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 821 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 822 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 823 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 824 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 825 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 826 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 827 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 828 829 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 830 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 831 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 832 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 833 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 834 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 835 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 836 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 837 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 838 839 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 840 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 841 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 842 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 843 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 844 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 845 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 846 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 847 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 848 849 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 850 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 851 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 852 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 853 854 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 855 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 856 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 857 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 858 859 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 860 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 861 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 862 863 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 864 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 865 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 866 867 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 868 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 869 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 870 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 871 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 872 873 case vmIntrinsics::_loadFence: 874 case vmIntrinsics::_storeFence: 875 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 876 877 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 878 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 879 880 #ifdef TRACE_HAVE_INTRINSICS 881 case vmIntrinsics::_classID: return inline_native_classID(); 882 case vmIntrinsics::_threadID: return inline_native_threadID(); 883 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 884 #endif 885 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 886 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 887 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 888 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 889 case vmIntrinsics::_newArray: return inline_native_newArray(); 890 case vmIntrinsics::_getLength: return inline_native_getLength(); 891 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 892 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 893 case vmIntrinsics::_equalsC: return inline_array_equals(); 894 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 895 896 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 897 898 case vmIntrinsics::_isInstance: 899 case vmIntrinsics::_getModifiers: 900 case vmIntrinsics::_isInterface: 901 case vmIntrinsics::_isArray: 902 case vmIntrinsics::_isPrimitive: 903 case vmIntrinsics::_getSuperclass: 904 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 905 906 case vmIntrinsics::_floatToRawIntBits: 907 case vmIntrinsics::_floatToIntBits: 908 case vmIntrinsics::_intBitsToFloat: 909 case vmIntrinsics::_doubleToRawLongBits: 910 case vmIntrinsics::_doubleToLongBits: 911 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 912 913 case vmIntrinsics::_numberOfLeadingZeros_i: 914 case vmIntrinsics::_numberOfLeadingZeros_l: 915 case vmIntrinsics::_numberOfTrailingZeros_i: 916 case vmIntrinsics::_numberOfTrailingZeros_l: 917 case vmIntrinsics::_bitCount_i: 918 case vmIntrinsics::_bitCount_l: 919 case vmIntrinsics::_reverseBytes_i: 920 case vmIntrinsics::_reverseBytes_l: 921 case vmIntrinsics::_reverseBytes_s: 922 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 923 924 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 925 926 case vmIntrinsics::_Reference_get: return inline_reference_get(); 927 928 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 929 930 case vmIntrinsics::_aescrypt_encryptBlock: 931 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 932 933 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 934 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 935 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 936 937 case vmIntrinsics::_sha_implCompress: 938 case vmIntrinsics::_sha2_implCompress: 939 case vmIntrinsics::_sha5_implCompress: 940 return inline_sha_implCompress(intrinsic_id()); 941 942 case vmIntrinsics::_digestBase_implCompressMB: 943 return inline_digestBase_implCompressMB(predicate); 944 945 case vmIntrinsics::_multiplyToLen: 946 return inline_multiplyToLen(); 947 948 case vmIntrinsics::_squareToLen: 949 return inline_squareToLen(); 950 951 case vmIntrinsics::_mulAdd: 952 return inline_mulAdd(); 953 954 case vmIntrinsics::_montgomeryMultiply: 955 return inline_montgomeryMultiply(); 956 case vmIntrinsics::_montgomerySquare: 957 return inline_montgomerySquare(); 958 959 case vmIntrinsics::_ghash_processBlocks: 960 return inline_ghash_processBlocks(); 961 962 case vmIntrinsics::_encodeISOArray: 963 return inline_encodeISOArray(); 964 965 case vmIntrinsics::_updateCRC32: 966 return inline_updateCRC32(); 967 case vmIntrinsics::_updateBytesCRC32: 968 return inline_updateBytesCRC32(); 969 case vmIntrinsics::_updateByteBufferCRC32: 970 return inline_updateByteBufferCRC32(); 971 972 case vmIntrinsics::_updateBytesCRC32C: 973 return inline_updateBytesCRC32C(); 974 case vmIntrinsics::_updateDirectByteBufferCRC32C: 975 return inline_updateDirectByteBufferCRC32C(); 976 977 case vmIntrinsics::_profileBoolean: 978 return inline_profileBoolean(); 979 case vmIntrinsics::_isCompileConstant: 980 return inline_isCompileConstant(); 981 982 default: 983 // If you get here, it may be that someone has added a new intrinsic 984 // to the list in vmSymbols.hpp without implementing it here. 985 #ifndef PRODUCT 986 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 987 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 988 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 989 } 990 #endif 991 return false; 992 } 993 } 994 995 Node* LibraryCallKit::try_to_predicate(int predicate) { 996 if (!jvms()->has_method()) { 997 // Root JVMState has a null method. 998 assert(map()->memory()->Opcode() == Op_Parm, ""); 999 // Insert the memory aliasing node 1000 set_all_memory(reset_memory()); 1001 } 1002 assert(merged_memory(), ""); 1003 1004 switch (intrinsic_id()) { 1005 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 1006 return inline_cipherBlockChaining_AESCrypt_predicate(false); 1007 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 1008 return inline_cipherBlockChaining_AESCrypt_predicate(true); 1009 case vmIntrinsics::_digestBase_implCompressMB: 1010 return inline_digestBase_implCompressMB_predicate(predicate); 1011 1012 default: 1013 // If you get here, it may be that someone has added a new intrinsic 1014 // to the list in vmSymbols.hpp without implementing it here. 1015 #ifndef PRODUCT 1016 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 1017 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 1018 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 1019 } 1020 #endif 1021 Node* slow_ctl = control(); 1022 set_control(top()); // No fast path instrinsic 1023 return slow_ctl; 1024 } 1025 } 1026 1027 //------------------------------set_result------------------------------- 1028 // Helper function for finishing intrinsics. 1029 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1030 record_for_igvn(region); 1031 set_control(_gvn.transform(region)); 1032 set_result( _gvn.transform(value)); 1033 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1034 } 1035 1036 //------------------------------generate_guard--------------------------- 1037 // Helper function for generating guarded fast-slow graph structures. 1038 // The given 'test', if true, guards a slow path. If the test fails 1039 // then a fast path can be taken. (We generally hope it fails.) 1040 // In all cases, GraphKit::control() is updated to the fast path. 1041 // The returned value represents the control for the slow path. 1042 // The return value is never 'top'; it is either a valid control 1043 // or NULL if it is obvious that the slow path can never be taken. 1044 // Also, if region and the slow control are not NULL, the slow edge 1045 // is appended to the region. 1046 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1047 if (stopped()) { 1048 // Already short circuited. 1049 return NULL; 1050 } 1051 1052 // Build an if node and its projections. 1053 // If test is true we take the slow path, which we assume is uncommon. 1054 if (_gvn.type(test) == TypeInt::ZERO) { 1055 // The slow branch is never taken. No need to build this guard. 1056 return NULL; 1057 } 1058 1059 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1060 1061 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1062 if (if_slow == top()) { 1063 // The slow branch is never taken. No need to build this guard. 1064 return NULL; 1065 } 1066 1067 if (region != NULL) 1068 region->add_req(if_slow); 1069 1070 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1071 set_control(if_fast); 1072 1073 return if_slow; 1074 } 1075 1076 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1077 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1078 } 1079 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1080 return generate_guard(test, region, PROB_FAIR); 1081 } 1082 1083 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1084 Node* *pos_index) { 1085 if (stopped()) 1086 return NULL; // already stopped 1087 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1088 return NULL; // index is already adequately typed 1089 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1090 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1091 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1092 if (is_neg != NULL && pos_index != NULL) { 1093 // Emulate effect of Parse::adjust_map_after_if. 1094 Node* ccast = new CastIINode(index, TypeInt::POS); 1095 ccast->set_req(0, control()); 1096 (*pos_index) = _gvn.transform(ccast); 1097 } 1098 return is_neg; 1099 } 1100 1101 // Make sure that 'position' is a valid limit index, in [0..length]. 1102 // There are two equivalent plans for checking this: 1103 // A. (offset + copyLength) unsigned<= arrayLength 1104 // B. offset <= (arrayLength - copyLength) 1105 // We require that all of the values above, except for the sum and 1106 // difference, are already known to be non-negative. 1107 // Plan A is robust in the face of overflow, if offset and copyLength 1108 // are both hugely positive. 1109 // 1110 // Plan B is less direct and intuitive, but it does not overflow at 1111 // all, since the difference of two non-negatives is always 1112 // representable. Whenever Java methods must perform the equivalent 1113 // check they generally use Plan B instead of Plan A. 1114 // For the moment we use Plan A. 1115 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1116 Node* subseq_length, 1117 Node* array_length, 1118 RegionNode* region) { 1119 if (stopped()) 1120 return NULL; // already stopped 1121 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1122 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1123 return NULL; // common case of whole-array copy 1124 Node* last = subseq_length; 1125 if (!zero_offset) // last += offset 1126 last = _gvn.transform(new AddINode(last, offset)); 1127 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1128 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1129 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1130 return is_over; 1131 } 1132 1133 1134 //--------------------------generate_current_thread-------------------- 1135 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1136 ciKlass* thread_klass = env()->Thread_klass(); 1137 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1138 Node* thread = _gvn.transform(new ThreadLocalNode()); 1139 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1140 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1141 tls_output = thread; 1142 return threadObj; 1143 } 1144 1145 1146 //------------------------------make_string_method_node------------------------ 1147 // Helper method for String intrinsic functions. This version is called 1148 // with str1 and str2 pointing to String object nodes. 1149 // 1150 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1151 Node* no_ctrl = NULL; 1152 1153 // Get start addr of string 1154 Node* str1_value = load_String_value(no_ctrl, str1); 1155 Node* str1_offset = load_String_offset(no_ctrl, str1); 1156 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1157 1158 // Get length of string 1 1159 Node* str1_len = load_String_length(no_ctrl, str1); 1160 1161 Node* str2_value = load_String_value(no_ctrl, str2); 1162 Node* str2_offset = load_String_offset(no_ctrl, str2); 1163 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1164 1165 Node* str2_len = NULL; 1166 Node* result = NULL; 1167 1168 switch (opcode) { 1169 case Op_StrIndexOf: 1170 // Get length of string 2 1171 str2_len = load_String_length(no_ctrl, str2); 1172 1173 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1174 str1_start, str1_len, str2_start, str2_len); 1175 break; 1176 case Op_StrComp: 1177 // Get length of string 2 1178 str2_len = load_String_length(no_ctrl, str2); 1179 1180 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1181 str1_start, str1_len, str2_start, str2_len); 1182 break; 1183 case Op_StrEquals: 1184 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1185 str1_start, str2_start, str1_len); 1186 break; 1187 default: 1188 ShouldNotReachHere(); 1189 return NULL; 1190 } 1191 1192 // All these intrinsics have checks. 1193 C->set_has_split_ifs(true); // Has chance for split-if optimization 1194 1195 return _gvn.transform(result); 1196 } 1197 1198 // Helper method for String intrinsic functions. This version is called 1199 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1200 // to Int nodes containing the lenghts of str1 and str2. 1201 // 1202 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1203 Node* result = NULL; 1204 switch (opcode) { 1205 case Op_StrIndexOf: 1206 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1207 str1_start, cnt1, str2_start, cnt2); 1208 break; 1209 case Op_StrComp: 1210 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1211 str1_start, cnt1, str2_start, cnt2); 1212 break; 1213 case Op_StrEquals: 1214 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1215 str1_start, str2_start, cnt1); 1216 break; 1217 default: 1218 ShouldNotReachHere(); 1219 return NULL; 1220 } 1221 1222 // All these intrinsics have checks. 1223 C->set_has_split_ifs(true); // Has chance for split-if optimization 1224 1225 return _gvn.transform(result); 1226 } 1227 1228 //------------------------------inline_string_compareTo------------------------ 1229 // public int java.lang.String.compareTo(String anotherString); 1230 bool LibraryCallKit::inline_string_compareTo() { 1231 Node* receiver = null_check(argument(0)); 1232 Node* arg = null_check(argument(1)); 1233 if (stopped()) { 1234 return true; 1235 } 1236 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1237 return true; 1238 } 1239 1240 //------------------------------inline_string_equals------------------------ 1241 bool LibraryCallKit::inline_string_equals() { 1242 Node* receiver = null_check_receiver(); 1243 // NOTE: Do not null check argument for String.equals() because spec 1244 // allows to specify NULL as argument. 1245 Node* argument = this->argument(1); 1246 if (stopped()) { 1247 return true; 1248 } 1249 1250 // paths (plus control) merge 1251 RegionNode* region = new RegionNode(5); 1252 Node* phi = new PhiNode(region, TypeInt::BOOL); 1253 1254 // does source == target string? 1255 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1256 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1257 1258 Node* if_eq = generate_slow_guard(bol, NULL); 1259 if (if_eq != NULL) { 1260 // receiver == argument 1261 phi->init_req(2, intcon(1)); 1262 region->init_req(2, if_eq); 1263 } 1264 1265 // get String klass for instanceOf 1266 ciInstanceKlass* klass = env()->String_klass(); 1267 1268 if (!stopped()) { 1269 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1270 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1271 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1272 1273 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1274 //instanceOf == true, fallthrough 1275 1276 if (inst_false != NULL) { 1277 phi->init_req(3, intcon(0)); 1278 region->init_req(3, inst_false); 1279 } 1280 } 1281 1282 if (!stopped()) { 1283 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1284 1285 // Properly cast the argument to String 1286 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1287 // This path is taken only when argument's type is String:NotNull. 1288 argument = cast_not_null(argument, false); 1289 1290 Node* no_ctrl = NULL; 1291 1292 // Get start addr of receiver 1293 Node* receiver_val = load_String_value(no_ctrl, receiver); 1294 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1295 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1296 1297 // Get length of receiver 1298 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1299 1300 // Get start addr of argument 1301 Node* argument_val = load_String_value(no_ctrl, argument); 1302 Node* argument_offset = load_String_offset(no_ctrl, argument); 1303 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1304 1305 // Get length of argument 1306 Node* argument_cnt = load_String_length(no_ctrl, argument); 1307 1308 // Check for receiver count != argument count 1309 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1310 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1311 Node* if_ne = generate_slow_guard(bol, NULL); 1312 if (if_ne != NULL) { 1313 phi->init_req(4, intcon(0)); 1314 region->init_req(4, if_ne); 1315 } 1316 1317 // Check for count == 0 is done by assembler code for StrEquals. 1318 1319 if (!stopped()) { 1320 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1321 phi->init_req(1, equals); 1322 region->init_req(1, control()); 1323 } 1324 } 1325 1326 // post merge 1327 set_control(_gvn.transform(region)); 1328 record_for_igvn(region); 1329 1330 set_result(_gvn.transform(phi)); 1331 return true; 1332 } 1333 1334 //------------------------------inline_array_equals---------------------------- 1335 bool LibraryCallKit::inline_array_equals() { 1336 Node* arg1 = argument(0); 1337 Node* arg2 = argument(1); 1338 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1339 return true; 1340 } 1341 1342 // Java version of String.indexOf(constant string) 1343 // class StringDecl { 1344 // StringDecl(char[] ca) { 1345 // offset = 0; 1346 // count = ca.length; 1347 // value = ca; 1348 // } 1349 // int offset; 1350 // int count; 1351 // char[] value; 1352 // } 1353 // 1354 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1355 // int targetOffset, int cache_i, int md2) { 1356 // int cache = cache_i; 1357 // int sourceOffset = string_object.offset; 1358 // int sourceCount = string_object.count; 1359 // int targetCount = target_object.length; 1360 // 1361 // int targetCountLess1 = targetCount - 1; 1362 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1363 // 1364 // char[] source = string_object.value; 1365 // char[] target = target_object; 1366 // int lastChar = target[targetCountLess1]; 1367 // 1368 // outer_loop: 1369 // for (int i = sourceOffset; i < sourceEnd; ) { 1370 // int src = source[i + targetCountLess1]; 1371 // if (src == lastChar) { 1372 // // With random strings and a 4-character alphabet, 1373 // // reverse matching at this point sets up 0.8% fewer 1374 // // frames, but (paradoxically) makes 0.3% more probes. 1375 // // Since those probes are nearer the lastChar probe, 1376 // // there is may be a net D$ win with reverse matching. 1377 // // But, reversing loop inhibits unroll of inner loop 1378 // // for unknown reason. So, does running outer loop from 1379 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1380 // for (int j = 0; j < targetCountLess1; j++) { 1381 // if (target[targetOffset + j] != source[i+j]) { 1382 // if ((cache & (1 << source[i+j])) == 0) { 1383 // if (md2 < j+1) { 1384 // i += j+1; 1385 // continue outer_loop; 1386 // } 1387 // } 1388 // i += md2; 1389 // continue outer_loop; 1390 // } 1391 // } 1392 // return i - sourceOffset; 1393 // } 1394 // if ((cache & (1 << src)) == 0) { 1395 // i += targetCountLess1; 1396 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1397 // i++; 1398 // } 1399 // return -1; 1400 // } 1401 1402 //------------------------------string_indexOf------------------------ 1403 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1404 jint cache_i, jint md2_i) { 1405 1406 Node* no_ctrl = NULL; 1407 float likely = PROB_LIKELY(0.9); 1408 float unlikely = PROB_UNLIKELY(0.9); 1409 1410 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1411 1412 Node* source = load_String_value(no_ctrl, string_object); 1413 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1414 Node* sourceCount = load_String_length(no_ctrl, string_object); 1415 1416 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1417 jint target_length = target_array->length(); 1418 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1419 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1420 1421 // String.value field is known to be @Stable. 1422 if (UseImplicitStableValues) { 1423 target = cast_array_to_stable(target, target_type); 1424 } 1425 1426 IdealKit kit(this, false, true); 1427 #define __ kit. 1428 Node* zero = __ ConI(0); 1429 Node* one = __ ConI(1); 1430 Node* cache = __ ConI(cache_i); 1431 Node* md2 = __ ConI(md2_i); 1432 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1433 Node* targetCountLess1 = __ ConI(target_length - 1); 1434 Node* targetOffset = __ ConI(targetOffset_i); 1435 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1436 1437 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1438 Node* outer_loop = __ make_label(2 /* goto */); 1439 Node* return_ = __ make_label(1); 1440 1441 __ set(rtn,__ ConI(-1)); 1442 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1443 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1444 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1445 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1446 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1447 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1448 Node* tpj = __ AddI(targetOffset, __ value(j)); 1449 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1450 Node* ipj = __ AddI(__ value(i), __ value(j)); 1451 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1452 __ if_then(targ, BoolTest::ne, src2); { 1453 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1454 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1455 __ increment(i, __ AddI(__ value(j), one)); 1456 __ goto_(outer_loop); 1457 } __ end_if(); __ dead(j); 1458 }__ end_if(); __ dead(j); 1459 __ increment(i, md2); 1460 __ goto_(outer_loop); 1461 }__ end_if(); 1462 __ increment(j, one); 1463 }__ end_loop(); __ dead(j); 1464 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1465 __ goto_(return_); 1466 }__ end_if(); 1467 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1468 __ increment(i, targetCountLess1); 1469 }__ end_if(); 1470 __ increment(i, one); 1471 __ bind(outer_loop); 1472 }__ end_loop(); __ dead(i); 1473 __ bind(return_); 1474 1475 // Final sync IdealKit and GraphKit. 1476 final_sync(kit); 1477 Node* result = __ value(rtn); 1478 #undef __ 1479 C->set_has_loops(true); 1480 return result; 1481 } 1482 1483 //------------------------------inline_string_indexOf------------------------ 1484 bool LibraryCallKit::inline_string_indexOf() { 1485 Node* receiver = argument(0); 1486 Node* arg = argument(1); 1487 1488 Node* result; 1489 if (Matcher::has_match_rule(Op_StrIndexOf) && 1490 UseSSE42Intrinsics) { 1491 // Generate SSE4.2 version of indexOf 1492 // We currently only have match rules that use SSE4.2 1493 1494 receiver = null_check(receiver); 1495 arg = null_check(arg); 1496 if (stopped()) { 1497 return true; 1498 } 1499 1500 // Make the merge point 1501 RegionNode* result_rgn = new RegionNode(4); 1502 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1503 Node* no_ctrl = NULL; 1504 1505 // Get start addr of source string 1506 Node* source = load_String_value(no_ctrl, receiver); 1507 Node* source_offset = load_String_offset(no_ctrl, receiver); 1508 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1509 1510 // Get length of source string 1511 Node* source_cnt = load_String_length(no_ctrl, receiver); 1512 1513 // Get start addr of substring 1514 Node* substr = load_String_value(no_ctrl, arg); 1515 Node* substr_offset = load_String_offset(no_ctrl, arg); 1516 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1517 1518 // Get length of source string 1519 Node* substr_cnt = load_String_length(no_ctrl, arg); 1520 1521 // Check for substr count > string count 1522 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1523 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1524 Node* if_gt = generate_slow_guard(bol, NULL); 1525 if (if_gt != NULL) { 1526 result_phi->init_req(2, intcon(-1)); 1527 result_rgn->init_req(2, if_gt); 1528 } 1529 1530 if (!stopped()) { 1531 // Check for substr count == 0 1532 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1533 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1534 Node* if_zero = generate_slow_guard(bol, NULL); 1535 if (if_zero != NULL) { 1536 result_phi->init_req(3, intcon(0)); 1537 result_rgn->init_req(3, if_zero); 1538 } 1539 } 1540 1541 if (!stopped()) { 1542 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1543 result_phi->init_req(1, result); 1544 result_rgn->init_req(1, control()); 1545 } 1546 set_control(_gvn.transform(result_rgn)); 1547 record_for_igvn(result_rgn); 1548 result = _gvn.transform(result_phi); 1549 1550 } else { // Use LibraryCallKit::string_indexOf 1551 // don't intrinsify if argument isn't a constant string. 1552 if (!arg->is_Con()) { 1553 return false; 1554 } 1555 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1556 if (str_type == NULL) { 1557 return false; 1558 } 1559 ciInstanceKlass* klass = env()->String_klass(); 1560 ciObject* str_const = str_type->const_oop(); 1561 if (str_const == NULL || str_const->klass() != klass) { 1562 return false; 1563 } 1564 ciInstance* str = str_const->as_instance(); 1565 assert(str != NULL, "must be instance"); 1566 1567 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1568 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1569 1570 int o; 1571 int c; 1572 if (java_lang_String::has_offset_field()) { 1573 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1574 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1575 } else { 1576 o = 0; 1577 c = pat->length(); 1578 } 1579 1580 // constant strings have no offset and count == length which 1581 // simplifies the resulting code somewhat so lets optimize for that. 1582 if (o != 0 || c != pat->length()) { 1583 return false; 1584 } 1585 1586 receiver = null_check(receiver, T_OBJECT); 1587 // NOTE: No null check on the argument is needed since it's a constant String oop. 1588 if (stopped()) { 1589 return true; 1590 } 1591 1592 // The null string as a pattern always returns 0 (match at beginning of string) 1593 if (c == 0) { 1594 set_result(intcon(0)); 1595 return true; 1596 } 1597 1598 // Generate default indexOf 1599 jchar lastChar = pat->char_at(o + (c - 1)); 1600 int cache = 0; 1601 int i; 1602 for (i = 0; i < c - 1; i++) { 1603 assert(i < pat->length(), "out of range"); 1604 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1605 } 1606 1607 int md2 = c; 1608 for (i = 0; i < c - 1; i++) { 1609 assert(i < pat->length(), "out of range"); 1610 if (pat->char_at(o + i) == lastChar) { 1611 md2 = (c - 1) - i; 1612 } 1613 } 1614 1615 result = string_indexOf(receiver, pat, o, cache, md2); 1616 } 1617 set_result(result); 1618 return true; 1619 } 1620 1621 //--------------------------round_double_node-------------------------------- 1622 // Round a double node if necessary. 1623 Node* LibraryCallKit::round_double_node(Node* n) { 1624 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1625 n = _gvn.transform(new RoundDoubleNode(0, n)); 1626 return n; 1627 } 1628 1629 //------------------------------inline_math----------------------------------- 1630 // public static double Math.abs(double) 1631 // public static double Math.sqrt(double) 1632 // public static double Math.log(double) 1633 // public static double Math.log10(double) 1634 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1635 Node* arg = round_double_node(argument(0)); 1636 Node* n; 1637 switch (id) { 1638 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1639 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1640 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1641 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1642 default: fatal_unexpected_iid(id); break; 1643 } 1644 set_result(_gvn.transform(n)); 1645 return true; 1646 } 1647 1648 //------------------------------inline_trig---------------------------------- 1649 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1650 // argument reduction which will turn into a fast/slow diamond. 1651 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1652 Node* arg = round_double_node(argument(0)); 1653 Node* n = NULL; 1654 1655 switch (id) { 1656 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1657 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1658 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1659 default: fatal_unexpected_iid(id); break; 1660 } 1661 n = _gvn.transform(n); 1662 1663 // Rounding required? Check for argument reduction! 1664 if (Matcher::strict_fp_requires_explicit_rounding) { 1665 static const double pi_4 = 0.7853981633974483; 1666 static const double neg_pi_4 = -0.7853981633974483; 1667 // pi/2 in 80-bit extended precision 1668 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1669 // -pi/2 in 80-bit extended precision 1670 // 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}; 1671 // Cutoff value for using this argument reduction technique 1672 //static const double pi_2_minus_epsilon = 1.564660403643354; 1673 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1674 1675 // Pseudocode for sin: 1676 // if (x <= Math.PI / 4.0) { 1677 // if (x >= -Math.PI / 4.0) return fsin(x); 1678 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1679 // } else { 1680 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1681 // } 1682 // return StrictMath.sin(x); 1683 1684 // Pseudocode for cos: 1685 // if (x <= Math.PI / 4.0) { 1686 // if (x >= -Math.PI / 4.0) return fcos(x); 1687 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1688 // } else { 1689 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1690 // } 1691 // return StrictMath.cos(x); 1692 1693 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1694 // requires a special machine instruction to load it. Instead we'll try 1695 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1696 // probably do the math inside the SIN encoding. 1697 1698 // Make the merge point 1699 RegionNode* r = new RegionNode(3); 1700 Node* phi = new PhiNode(r, Type::DOUBLE); 1701 1702 // Flatten arg so we need only 1 test 1703 Node *abs = _gvn.transform(new AbsDNode(arg)); 1704 // Node for PI/4 constant 1705 Node *pi4 = makecon(TypeD::make(pi_4)); 1706 // Check PI/4 : abs(arg) 1707 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1708 // Check: If PI/4 < abs(arg) then go slow 1709 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1710 // Branch either way 1711 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1712 set_control(opt_iff(r,iff)); 1713 1714 // Set fast path result 1715 phi->init_req(2, n); 1716 1717 // Slow path - non-blocking leaf call 1718 Node* call = NULL; 1719 switch (id) { 1720 case vmIntrinsics::_dsin: 1721 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1722 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1723 "Sin", NULL, arg, top()); 1724 break; 1725 case vmIntrinsics::_dcos: 1726 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1727 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1728 "Cos", NULL, arg, top()); 1729 break; 1730 case vmIntrinsics::_dtan: 1731 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1732 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1733 "Tan", NULL, arg, top()); 1734 break; 1735 } 1736 assert(control()->in(0) == call, ""); 1737 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1738 r->init_req(1, control()); 1739 phi->init_req(1, slow_result); 1740 1741 // Post-merge 1742 set_control(_gvn.transform(r)); 1743 record_for_igvn(r); 1744 n = _gvn.transform(phi); 1745 1746 C->set_has_split_ifs(true); // Has chance for split-if optimization 1747 } 1748 set_result(n); 1749 return true; 1750 } 1751 1752 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1753 //------------------- 1754 //result=(result.isNaN())? funcAddr():result; 1755 // Check: If isNaN() by checking result!=result? then either trap 1756 // or go to runtime 1757 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1758 // Build the boolean node 1759 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1760 1761 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1762 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1763 // The pow or exp intrinsic returned a NaN, which requires a call 1764 // to the runtime. Recompile with the runtime call. 1765 uncommon_trap(Deoptimization::Reason_intrinsic, 1766 Deoptimization::Action_make_not_entrant); 1767 } 1768 return result; 1769 } else { 1770 // If this inlining ever returned NaN in the past, we compile a call 1771 // to the runtime to properly handle corner cases 1772 1773 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1774 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1775 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1776 1777 if (!if_slow->is_top()) { 1778 RegionNode* result_region = new RegionNode(3); 1779 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1780 1781 result_region->init_req(1, if_fast); 1782 result_val->init_req(1, result); 1783 1784 set_control(if_slow); 1785 1786 const TypePtr* no_memory_effects = NULL; 1787 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1788 no_memory_effects, 1789 x, top(), y, y ? top() : NULL); 1790 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1791 #ifdef ASSERT 1792 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1793 assert(value_top == top(), "second value must be top"); 1794 #endif 1795 1796 result_region->init_req(2, control()); 1797 result_val->init_req(2, value); 1798 set_control(_gvn.transform(result_region)); 1799 return _gvn.transform(result_val); 1800 } else { 1801 return result; 1802 } 1803 } 1804 } 1805 1806 //------------------------------inline_exp------------------------------------- 1807 // Inline exp instructions, if possible. The Intel hardware only misses 1808 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1809 bool LibraryCallKit::inline_exp() { 1810 Node* arg = round_double_node(argument(0)); 1811 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1812 1813 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1814 set_result(n); 1815 1816 C->set_has_split_ifs(true); // Has chance for split-if optimization 1817 return true; 1818 } 1819 1820 //------------------------------inline_pow------------------------------------- 1821 // Inline power instructions, if possible. 1822 bool LibraryCallKit::inline_pow() { 1823 // Pseudocode for pow 1824 // if (y == 2) { 1825 // return x * x; 1826 // } else { 1827 // if (x <= 0.0) { 1828 // long longy = (long)y; 1829 // if ((double)longy == y) { // if y is long 1830 // if (y + 1 == y) longy = 0; // huge number: even 1831 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1832 // } else { 1833 // result = NaN; 1834 // } 1835 // } else { 1836 // result = DPow(x,y); 1837 // } 1838 // if (result != result)? { 1839 // result = uncommon_trap() or runtime_call(); 1840 // } 1841 // return result; 1842 // } 1843 1844 Node* x = round_double_node(argument(0)); 1845 Node* y = round_double_node(argument(2)); 1846 1847 Node* result = NULL; 1848 1849 Node* const_two_node = makecon(TypeD::make(2.0)); 1850 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1851 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1852 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1853 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1854 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1855 1856 RegionNode* region_node = new RegionNode(3); 1857 region_node->init_req(1, if_true); 1858 1859 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1860 // special case for x^y where y == 2, we can convert it to x * x 1861 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1862 1863 // set control to if_false since we will now process the false branch 1864 set_control(if_false); 1865 1866 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1867 // Short form: skip the fancy tests and just check for NaN result. 1868 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1869 } else { 1870 // If this inlining ever returned NaN in the past, include all 1871 // checks + call to the runtime. 1872 1873 // Set the merge point for If node with condition of (x <= 0.0) 1874 // There are four possible paths to region node and phi node 1875 RegionNode *r = new RegionNode(4); 1876 Node *phi = new PhiNode(r, Type::DOUBLE); 1877 1878 // Build the first if node: if (x <= 0.0) 1879 // Node for 0 constant 1880 Node *zeronode = makecon(TypeD::ZERO); 1881 // Check x:0 1882 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1883 // Check: If (x<=0) then go complex path 1884 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1885 // Branch either way 1886 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1887 // Fast path taken; set region slot 3 1888 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1889 r->init_req(3,fast_taken); // Capture fast-control 1890 1891 // Fast path not-taken, i.e. slow path 1892 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1893 1894 // Set fast path result 1895 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1896 phi->init_req(3, fast_result); 1897 1898 // Complex path 1899 // Build the second if node (if y is long) 1900 // Node for (long)y 1901 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1902 // Node for (double)((long) y) 1903 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1904 // Check (double)((long) y) : y 1905 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1906 // Check if (y isn't long) then go to slow path 1907 1908 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1909 // Branch either way 1910 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1911 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1912 1913 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1914 1915 // Calculate DPow(abs(x), y)*(1 & (long)y) 1916 // Node for constant 1 1917 Node *conone = longcon(1); 1918 // 1& (long)y 1919 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1920 1921 // A huge number is always even. Detect a huge number by checking 1922 // if y + 1 == y and set integer to be tested for parity to 0. 1923 // Required for corner case: 1924 // (long)9.223372036854776E18 = max_jlong 1925 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1926 // max_jlong is odd but 9.223372036854776E18 is even 1927 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1928 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1929 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1930 Node* correctedsign = NULL; 1931 if (ConditionalMoveLimit != 0) { 1932 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1933 } else { 1934 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1935 RegionNode *r = new RegionNode(3); 1936 Node *phi = new PhiNode(r, TypeLong::LONG); 1937 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1938 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1939 phi->init_req(1, signnode); 1940 phi->init_req(2, longcon(0)); 1941 correctedsign = _gvn.transform(phi); 1942 ylong_path = _gvn.transform(r); 1943 record_for_igvn(r); 1944 } 1945 1946 // zero node 1947 Node *conzero = longcon(0); 1948 // Check (1&(long)y)==0? 1949 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1950 // Check if (1&(long)y)!=0?, if so the result is negative 1951 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1952 // abs(x) 1953 Node *absx=_gvn.transform(new AbsDNode(x)); 1954 // abs(x)^y 1955 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1956 // -abs(x)^y 1957 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1958 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1959 Node *signresult = NULL; 1960 if (ConditionalMoveLimit != 0) { 1961 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1962 } else { 1963 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1964 RegionNode *r = new RegionNode(3); 1965 Node *phi = new PhiNode(r, Type::DOUBLE); 1966 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1967 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1968 phi->init_req(1, absxpowy); 1969 phi->init_req(2, negabsxpowy); 1970 signresult = _gvn.transform(phi); 1971 ylong_path = _gvn.transform(r); 1972 record_for_igvn(r); 1973 } 1974 // Set complex path fast result 1975 r->init_req(2, ylong_path); 1976 phi->init_req(2, signresult); 1977 1978 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1979 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1980 r->init_req(1,slow_path); 1981 phi->init_req(1,slow_result); 1982 1983 // Post merge 1984 set_control(_gvn.transform(r)); 1985 record_for_igvn(r); 1986 result = _gvn.transform(phi); 1987 } 1988 1989 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1990 1991 // control from finish_pow_exp is now input to the region node 1992 region_node->set_req(2, control()); 1993 // the result from finish_pow_exp is now input to the phi node 1994 phi_node->init_req(2, result); 1995 set_control(_gvn.transform(region_node)); 1996 record_for_igvn(region_node); 1997 set_result(_gvn.transform(phi_node)); 1998 1999 C->set_has_split_ifs(true); // Has chance for split-if optimization 2000 return true; 2001 } 2002 2003 //------------------------------runtime_math----------------------------- 2004 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 2005 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 2006 "must be (DD)D or (D)D type"); 2007 2008 // Inputs 2009 Node* a = round_double_node(argument(0)); 2010 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 2011 2012 const TypePtr* no_memory_effects = NULL; 2013 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 2014 no_memory_effects, 2015 a, top(), b, b ? top() : NULL); 2016 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 2017 #ifdef ASSERT 2018 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 2019 assert(value_top == top(), "second value must be top"); 2020 #endif 2021 2022 set_result(value); 2023 return true; 2024 } 2025 2026 //------------------------------inline_math_native----------------------------- 2027 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2028 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2029 switch (id) { 2030 // These intrinsics are not properly supported on all hardware 2031 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2032 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2033 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2034 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2035 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2036 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2037 2038 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2039 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2040 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2041 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2042 2043 // These intrinsics are supported on all hardware 2044 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2045 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2046 2047 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2048 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2049 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2050 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2051 #undef FN_PTR 2052 2053 // These intrinsics are not yet correctly implemented 2054 case vmIntrinsics::_datan2: 2055 return false; 2056 2057 default: 2058 fatal_unexpected_iid(id); 2059 return false; 2060 } 2061 } 2062 2063 static bool is_simple_name(Node* n) { 2064 return (n->req() == 1 // constant 2065 || (n->is_Type() && n->as_Type()->type()->singleton()) 2066 || n->is_Proj() // parameter or return value 2067 || n->is_Phi() // local of some sort 2068 ); 2069 } 2070 2071 //----------------------------inline_min_max----------------------------------- 2072 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2073 set_result(generate_min_max(id, argument(0), argument(1))); 2074 return true; 2075 } 2076 2077 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2078 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2079 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2080 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2081 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2082 2083 { 2084 PreserveJVMState pjvms(this); 2085 PreserveReexecuteState preexecs(this); 2086 jvms()->set_should_reexecute(true); 2087 2088 set_control(slow_path); 2089 set_i_o(i_o()); 2090 2091 uncommon_trap(Deoptimization::Reason_intrinsic, 2092 Deoptimization::Action_none); 2093 } 2094 2095 set_control(fast_path); 2096 set_result(math); 2097 } 2098 2099 template <typename OverflowOp> 2100 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2101 typedef typename OverflowOp::MathOp MathOp; 2102 2103 MathOp* mathOp = new MathOp(arg1, arg2); 2104 Node* operation = _gvn.transform( mathOp ); 2105 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2106 inline_math_mathExact(operation, ofcheck); 2107 return true; 2108 } 2109 2110 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2111 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2112 } 2113 2114 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2115 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2116 } 2117 2118 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2119 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2120 } 2121 2122 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2123 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2124 } 2125 2126 bool LibraryCallKit::inline_math_negateExactI() { 2127 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2128 } 2129 2130 bool LibraryCallKit::inline_math_negateExactL() { 2131 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2132 } 2133 2134 bool LibraryCallKit::inline_math_multiplyExactI() { 2135 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2136 } 2137 2138 bool LibraryCallKit::inline_math_multiplyExactL() { 2139 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2140 } 2141 2142 Node* 2143 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2144 // These are the candidate return value: 2145 Node* xvalue = x0; 2146 Node* yvalue = y0; 2147 2148 if (xvalue == yvalue) { 2149 return xvalue; 2150 } 2151 2152 bool want_max = (id == vmIntrinsics::_max); 2153 2154 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2155 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2156 if (txvalue == NULL || tyvalue == NULL) return top(); 2157 // This is not really necessary, but it is consistent with a 2158 // hypothetical MaxINode::Value method: 2159 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2160 2161 // %%% This folding logic should (ideally) be in a different place. 2162 // Some should be inside IfNode, and there to be a more reliable 2163 // transformation of ?: style patterns into cmoves. We also want 2164 // more powerful optimizations around cmove and min/max. 2165 2166 // Try to find a dominating comparison of these guys. 2167 // It can simplify the index computation for Arrays.copyOf 2168 // and similar uses of System.arraycopy. 2169 // First, compute the normalized version of CmpI(x, y). 2170 int cmp_op = Op_CmpI; 2171 Node* xkey = xvalue; 2172 Node* ykey = yvalue; 2173 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2174 if (ideal_cmpxy->is_Cmp()) { 2175 // E.g., if we have CmpI(length - offset, count), 2176 // it might idealize to CmpI(length, count + offset) 2177 cmp_op = ideal_cmpxy->Opcode(); 2178 xkey = ideal_cmpxy->in(1); 2179 ykey = ideal_cmpxy->in(2); 2180 } 2181 2182 // Start by locating any relevant comparisons. 2183 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2184 Node* cmpxy = NULL; 2185 Node* cmpyx = NULL; 2186 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2187 Node* cmp = start_from->fast_out(k); 2188 if (cmp->outcnt() > 0 && // must have prior uses 2189 cmp->in(0) == NULL && // must be context-independent 2190 cmp->Opcode() == cmp_op) { // right kind of compare 2191 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2192 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2193 } 2194 } 2195 2196 const int NCMPS = 2; 2197 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2198 int cmpn; 2199 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2200 if (cmps[cmpn] != NULL) break; // find a result 2201 } 2202 if (cmpn < NCMPS) { 2203 // Look for a dominating test that tells us the min and max. 2204 int depth = 0; // Limit search depth for speed 2205 Node* dom = control(); 2206 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2207 if (++depth >= 100) break; 2208 Node* ifproj = dom; 2209 if (!ifproj->is_Proj()) continue; 2210 Node* iff = ifproj->in(0); 2211 if (!iff->is_If()) continue; 2212 Node* bol = iff->in(1); 2213 if (!bol->is_Bool()) continue; 2214 Node* cmp = bol->in(1); 2215 if (cmp == NULL) continue; 2216 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2217 if (cmps[cmpn] == cmp) break; 2218 if (cmpn == NCMPS) continue; 2219 BoolTest::mask btest = bol->as_Bool()->_test._test; 2220 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2221 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2222 // At this point, we know that 'x btest y' is true. 2223 switch (btest) { 2224 case BoolTest::eq: 2225 // They are proven equal, so we can collapse the min/max. 2226 // Either value is the answer. Choose the simpler. 2227 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2228 return yvalue; 2229 return xvalue; 2230 case BoolTest::lt: // x < y 2231 case BoolTest::le: // x <= y 2232 return (want_max ? yvalue : xvalue); 2233 case BoolTest::gt: // x > y 2234 case BoolTest::ge: // x >= y 2235 return (want_max ? xvalue : yvalue); 2236 } 2237 } 2238 } 2239 2240 // We failed to find a dominating test. 2241 // Let's pick a test that might GVN with prior tests. 2242 Node* best_bol = NULL; 2243 BoolTest::mask best_btest = BoolTest::illegal; 2244 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2245 Node* cmp = cmps[cmpn]; 2246 if (cmp == NULL) continue; 2247 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2248 Node* bol = cmp->fast_out(j); 2249 if (!bol->is_Bool()) continue; 2250 BoolTest::mask btest = bol->as_Bool()->_test._test; 2251 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2252 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2253 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2254 best_bol = bol->as_Bool(); 2255 best_btest = btest; 2256 } 2257 } 2258 } 2259 2260 Node* answer_if_true = NULL; 2261 Node* answer_if_false = NULL; 2262 switch (best_btest) { 2263 default: 2264 if (cmpxy == NULL) 2265 cmpxy = ideal_cmpxy; 2266 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2267 // and fall through: 2268 case BoolTest::lt: // x < y 2269 case BoolTest::le: // x <= y 2270 answer_if_true = (want_max ? yvalue : xvalue); 2271 answer_if_false = (want_max ? xvalue : yvalue); 2272 break; 2273 case BoolTest::gt: // x > y 2274 case BoolTest::ge: // x >= y 2275 answer_if_true = (want_max ? xvalue : yvalue); 2276 answer_if_false = (want_max ? yvalue : xvalue); 2277 break; 2278 } 2279 2280 jint hi, lo; 2281 if (want_max) { 2282 // We can sharpen the minimum. 2283 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2284 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2285 } else { 2286 // We can sharpen the maximum. 2287 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2288 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2289 } 2290 2291 // Use a flow-free graph structure, to avoid creating excess control edges 2292 // which could hinder other optimizations. 2293 // Since Math.min/max is often used with arraycopy, we want 2294 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2295 Node* cmov = CMoveNode::make(NULL, best_bol, 2296 answer_if_false, answer_if_true, 2297 TypeInt::make(lo, hi, widen)); 2298 2299 return _gvn.transform(cmov); 2300 2301 /* 2302 // This is not as desirable as it may seem, since Min and Max 2303 // nodes do not have a full set of optimizations. 2304 // And they would interfere, anyway, with 'if' optimizations 2305 // and with CMoveI canonical forms. 2306 switch (id) { 2307 case vmIntrinsics::_min: 2308 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2309 case vmIntrinsics::_max: 2310 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2311 default: 2312 ShouldNotReachHere(); 2313 } 2314 */ 2315 } 2316 2317 inline int 2318 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2319 const TypePtr* base_type = TypePtr::NULL_PTR; 2320 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2321 if (base_type == NULL) { 2322 // Unknown type. 2323 return Type::AnyPtr; 2324 } else if (base_type == TypePtr::NULL_PTR) { 2325 // Since this is a NULL+long form, we have to switch to a rawptr. 2326 base = _gvn.transform(new CastX2PNode(offset)); 2327 offset = MakeConX(0); 2328 return Type::RawPtr; 2329 } else if (base_type->base() == Type::RawPtr) { 2330 return Type::RawPtr; 2331 } else if (base_type->isa_oopptr()) { 2332 // Base is never null => always a heap address. 2333 if (base_type->ptr() == TypePtr::NotNull) { 2334 return Type::OopPtr; 2335 } 2336 // Offset is small => always a heap address. 2337 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2338 if (offset_type != NULL && 2339 base_type->offset() == 0 && // (should always be?) 2340 offset_type->_lo >= 0 && 2341 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2342 return Type::OopPtr; 2343 } 2344 // Otherwise, it might either be oop+off or NULL+addr. 2345 return Type::AnyPtr; 2346 } else { 2347 // No information: 2348 return Type::AnyPtr; 2349 } 2350 } 2351 2352 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2353 int kind = classify_unsafe_addr(base, offset); 2354 if (kind == Type::RawPtr) { 2355 return basic_plus_adr(top(), base, offset); 2356 } else { 2357 return basic_plus_adr(base, offset); 2358 } 2359 } 2360 2361 //--------------------------inline_number_methods----------------------------- 2362 // inline int Integer.numberOfLeadingZeros(int) 2363 // inline int Long.numberOfLeadingZeros(long) 2364 // 2365 // inline int Integer.numberOfTrailingZeros(int) 2366 // inline int Long.numberOfTrailingZeros(long) 2367 // 2368 // inline int Integer.bitCount(int) 2369 // inline int Long.bitCount(long) 2370 // 2371 // inline char Character.reverseBytes(char) 2372 // inline short Short.reverseBytes(short) 2373 // inline int Integer.reverseBytes(int) 2374 // inline long Long.reverseBytes(long) 2375 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2376 Node* arg = argument(0); 2377 Node* n; 2378 switch (id) { 2379 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2380 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2381 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2382 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2383 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2384 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2385 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2386 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2387 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2388 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2389 default: fatal_unexpected_iid(id); break; 2390 } 2391 set_result(_gvn.transform(n)); 2392 return true; 2393 } 2394 2395 //----------------------------inline_unsafe_access---------------------------- 2396 2397 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2398 2399 // Helper that guards and inserts a pre-barrier. 2400 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2401 Node* pre_val, bool need_mem_bar) { 2402 // We could be accessing the referent field of a reference object. If so, when G1 2403 // is enabled, we need to log the value in the referent field in an SATB buffer. 2404 // This routine performs some compile time filters and generates suitable 2405 // runtime filters that guard the pre-barrier code. 2406 // Also add memory barrier for non volatile load from the referent field 2407 // to prevent commoning of loads across safepoint. 2408 if (!UseG1GC && !need_mem_bar) 2409 return; 2410 2411 // Some compile time checks. 2412 2413 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2414 const TypeX* otype = offset->find_intptr_t_type(); 2415 if (otype != NULL && otype->is_con() && 2416 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2417 // Constant offset but not the reference_offset so just return 2418 return; 2419 } 2420 2421 // We only need to generate the runtime guards for instances. 2422 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2423 if (btype != NULL) { 2424 if (btype->isa_aryptr()) { 2425 // Array type so nothing to do 2426 return; 2427 } 2428 2429 const TypeInstPtr* itype = btype->isa_instptr(); 2430 if (itype != NULL) { 2431 // Can the klass of base_oop be statically determined to be 2432 // _not_ a sub-class of Reference and _not_ Object? 2433 ciKlass* klass = itype->klass(); 2434 if ( klass->is_loaded() && 2435 !klass->is_subtype_of(env()->Reference_klass()) && 2436 !env()->Object_klass()->is_subtype_of(klass)) { 2437 return; 2438 } 2439 } 2440 } 2441 2442 // The compile time filters did not reject base_oop/offset so 2443 // we need to generate the following runtime filters 2444 // 2445 // if (offset == java_lang_ref_Reference::_reference_offset) { 2446 // if (instance_of(base, java.lang.ref.Reference)) { 2447 // pre_barrier(_, pre_val, ...); 2448 // } 2449 // } 2450 2451 float likely = PROB_LIKELY( 0.999); 2452 float unlikely = PROB_UNLIKELY(0.999); 2453 2454 IdealKit ideal(this); 2455 #define __ ideal. 2456 2457 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2458 2459 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2460 // Update graphKit memory and control from IdealKit. 2461 sync_kit(ideal); 2462 2463 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2464 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2465 2466 // Update IdealKit memory and control from graphKit. 2467 __ sync_kit(this); 2468 2469 Node* one = __ ConI(1); 2470 // is_instof == 0 if base_oop == NULL 2471 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2472 2473 // Update graphKit from IdeakKit. 2474 sync_kit(ideal); 2475 2476 // Use the pre-barrier to record the value in the referent field 2477 pre_barrier(false /* do_load */, 2478 __ ctrl(), 2479 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2480 pre_val /* pre_val */, 2481 T_OBJECT); 2482 if (need_mem_bar) { 2483 // Add memory barrier to prevent commoning reads from this field 2484 // across safepoint since GC can change its value. 2485 insert_mem_bar(Op_MemBarCPUOrder); 2486 } 2487 // Update IdealKit from graphKit. 2488 __ sync_kit(this); 2489 2490 } __ end_if(); // _ref_type != ref_none 2491 } __ end_if(); // offset == referent_offset 2492 2493 // Final sync IdealKit and GraphKit. 2494 final_sync(ideal); 2495 #undef __ 2496 } 2497 2498 2499 // Interpret Unsafe.fieldOffset cookies correctly: 2500 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2501 2502 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2503 // Attempt to infer a sharper value type from the offset and base type. 2504 ciKlass* sharpened_klass = NULL; 2505 2506 // See if it is an instance field, with an object type. 2507 if (alias_type->field() != NULL) { 2508 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2509 if (alias_type->field()->type()->is_klass()) { 2510 sharpened_klass = alias_type->field()->type()->as_klass(); 2511 } 2512 } 2513 2514 // See if it is a narrow oop array. 2515 if (adr_type->isa_aryptr()) { 2516 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2517 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2518 if (elem_type != NULL) { 2519 sharpened_klass = elem_type->klass(); 2520 } 2521 } 2522 } 2523 2524 // The sharpened class might be unloaded if there is no class loader 2525 // contraint in place. 2526 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2527 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2528 2529 #ifndef PRODUCT 2530 if (C->print_intrinsics() || C->print_inlining()) { 2531 tty->print(" from base type: "); adr_type->dump(); 2532 tty->print(" sharpened value: "); tjp->dump(); 2533 } 2534 #endif 2535 // Sharpen the value type. 2536 return tjp; 2537 } 2538 return NULL; 2539 } 2540 2541 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2542 if (callee()->is_static()) return false; // caller must have the capability! 2543 2544 #ifndef PRODUCT 2545 { 2546 ResourceMark rm; 2547 // Check the signatures. 2548 ciSignature* sig = callee()->signature(); 2549 #ifdef ASSERT 2550 if (!is_store) { 2551 // Object getObject(Object base, int/long offset), etc. 2552 BasicType rtype = sig->return_type()->basic_type(); 2553 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2554 rtype = T_ADDRESS; // it is really a C void* 2555 assert(rtype == type, "getter must return the expected value"); 2556 if (!is_native_ptr) { 2557 assert(sig->count() == 2, "oop getter has 2 arguments"); 2558 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2559 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2560 } else { 2561 assert(sig->count() == 1, "native getter has 1 argument"); 2562 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2563 } 2564 } else { 2565 // void putObject(Object base, int/long offset, Object x), etc. 2566 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2567 if (!is_native_ptr) { 2568 assert(sig->count() == 3, "oop putter has 3 arguments"); 2569 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2570 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2571 } else { 2572 assert(sig->count() == 2, "native putter has 2 arguments"); 2573 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2574 } 2575 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2576 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2577 vtype = T_ADDRESS; // it is really a C void* 2578 assert(vtype == type, "putter must accept the expected value"); 2579 } 2580 #endif // ASSERT 2581 } 2582 #endif //PRODUCT 2583 2584 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2585 2586 Node* receiver = argument(0); // type: oop 2587 2588 // Build address expression. 2589 Node* adr; 2590 Node* heap_base_oop = top(); 2591 Node* offset = top(); 2592 Node* val; 2593 2594 if (!is_native_ptr) { 2595 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2596 Node* base = argument(1); // type: oop 2597 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2598 offset = argument(2); // type: long 2599 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2600 // to be plain byte offsets, which are also the same as those accepted 2601 // by oopDesc::field_base. 2602 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2603 "fieldOffset must be byte-scaled"); 2604 // 32-bit machines ignore the high half! 2605 offset = ConvL2X(offset); 2606 adr = make_unsafe_address(base, offset); 2607 heap_base_oop = base; 2608 val = is_store ? argument(4) : NULL; 2609 } else { 2610 Node* ptr = argument(1); // type: long 2611 ptr = ConvL2X(ptr); // adjust Java long to machine word 2612 adr = make_unsafe_address(NULL, ptr); 2613 val = is_store ? argument(3) : NULL; 2614 } 2615 2616 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2617 2618 // First guess at the value type. 2619 const Type *value_type = Type::get_const_basic_type(type); 2620 2621 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2622 // there was not enough information to nail it down. 2623 Compile::AliasType* alias_type = C->alias_type(adr_type); 2624 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2625 2626 // We will need memory barriers unless we can determine a unique 2627 // alias category for this reference. (Note: If for some reason 2628 // the barriers get omitted and the unsafe reference begins to "pollute" 2629 // the alias analysis of the rest of the graph, either Compile::can_alias 2630 // or Compile::must_alias will throw a diagnostic assert.) 2631 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2632 2633 // If we are reading the value of the referent field of a Reference 2634 // object (either by using Unsafe directly or through reflection) 2635 // then, if G1 is enabled, we need to record the referent in an 2636 // SATB log buffer using the pre-barrier mechanism. 2637 // Also we need to add memory barrier to prevent commoning reads 2638 // from this field across safepoint since GC can change its value. 2639 bool need_read_barrier = !is_native_ptr && !is_store && 2640 offset != top() && heap_base_oop != top(); 2641 2642 if (!is_store && type == T_OBJECT) { 2643 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2644 if (tjp != NULL) { 2645 value_type = tjp; 2646 } 2647 } 2648 2649 receiver = null_check(receiver); 2650 if (stopped()) { 2651 return true; 2652 } 2653 // Heap pointers get a null-check from the interpreter, 2654 // as a courtesy. However, this is not guaranteed by Unsafe, 2655 // and it is not possible to fully distinguish unintended nulls 2656 // from intended ones in this API. 2657 2658 if (is_volatile) { 2659 // We need to emit leading and trailing CPU membars (see below) in 2660 // addition to memory membars when is_volatile. This is a little 2661 // too strong, but avoids the need to insert per-alias-type 2662 // volatile membars (for stores; compare Parse::do_put_xxx), which 2663 // we cannot do effectively here because we probably only have a 2664 // rough approximation of type. 2665 need_mem_bar = true; 2666 // For Stores, place a memory ordering barrier now. 2667 if (is_store) { 2668 insert_mem_bar(Op_MemBarRelease); 2669 } else { 2670 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2671 insert_mem_bar(Op_MemBarVolatile); 2672 } 2673 } 2674 } 2675 2676 // Memory barrier to prevent normal and 'unsafe' accesses from 2677 // bypassing each other. Happens after null checks, so the 2678 // exception paths do not take memory state from the memory barrier, 2679 // so there's no problems making a strong assert about mixing users 2680 // of safe & unsafe memory. 2681 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2682 2683 if (!is_store) { 2684 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2685 // To be valid, unsafe loads may depend on other conditions than 2686 // the one that guards them: pin the Load node 2687 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2688 // load value 2689 switch (type) { 2690 case T_BOOLEAN: 2691 case T_CHAR: 2692 case T_BYTE: 2693 case T_SHORT: 2694 case T_INT: 2695 case T_LONG: 2696 case T_FLOAT: 2697 case T_DOUBLE: 2698 break; 2699 case T_OBJECT: 2700 if (need_read_barrier) { 2701 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2702 } 2703 break; 2704 case T_ADDRESS: 2705 // Cast to an int type. 2706 p = _gvn.transform(new CastP2XNode(NULL, p)); 2707 p = ConvX2UL(p); 2708 break; 2709 default: 2710 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2711 break; 2712 } 2713 // The load node has the control of the preceding MemBarCPUOrder. All 2714 // following nodes will have the control of the MemBarCPUOrder inserted at 2715 // the end of this method. So, pushing the load onto the stack at a later 2716 // point is fine. 2717 set_result(p); 2718 } else { 2719 // place effect of store into memory 2720 switch (type) { 2721 case T_DOUBLE: 2722 val = dstore_rounding(val); 2723 break; 2724 case T_ADDRESS: 2725 // Repackage the long as a pointer. 2726 val = ConvL2X(val); 2727 val = _gvn.transform(new CastX2PNode(val)); 2728 break; 2729 } 2730 2731 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2732 if (type != T_OBJECT ) { 2733 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2734 } else { 2735 // Possibly an oop being stored to Java heap or native memory 2736 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2737 // oop to Java heap. 2738 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2739 } else { 2740 // We can't tell at compile time if we are storing in the Java heap or outside 2741 // of it. So we need to emit code to conditionally do the proper type of 2742 // store. 2743 2744 IdealKit ideal(this); 2745 #define __ ideal. 2746 // QQQ who knows what probability is here?? 2747 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2748 // Sync IdealKit and graphKit. 2749 sync_kit(ideal); 2750 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2751 // Update IdealKit memory. 2752 __ sync_kit(this); 2753 } __ else_(); { 2754 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2755 } __ end_if(); 2756 // Final sync IdealKit and GraphKit. 2757 final_sync(ideal); 2758 #undef __ 2759 } 2760 } 2761 } 2762 2763 if (is_volatile) { 2764 if (!is_store) { 2765 insert_mem_bar(Op_MemBarAcquire); 2766 } else { 2767 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2768 insert_mem_bar(Op_MemBarVolatile); 2769 } 2770 } 2771 } 2772 2773 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2774 2775 return true; 2776 } 2777 2778 //----------------------------inline_unsafe_load_store---------------------------- 2779 // This method serves a couple of different customers (depending on LoadStoreKind): 2780 // 2781 // LS_cmpxchg: 2782 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2783 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2784 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2785 // 2786 // LS_xadd: 2787 // public int getAndAddInt( Object o, long offset, int delta) 2788 // public long getAndAddLong(Object o, long offset, long delta) 2789 // 2790 // LS_xchg: 2791 // int getAndSet(Object o, long offset, int newValue) 2792 // long getAndSet(Object o, long offset, long newValue) 2793 // Object getAndSet(Object o, long offset, Object newValue) 2794 // 2795 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2796 // This basic scheme here is the same as inline_unsafe_access, but 2797 // differs in enough details that combining them would make the code 2798 // overly confusing. (This is a true fact! I originally combined 2799 // them, but even I was confused by it!) As much code/comments as 2800 // possible are retained from inline_unsafe_access though to make 2801 // the correspondences clearer. - dl 2802 2803 if (callee()->is_static()) return false; // caller must have the capability! 2804 2805 #ifndef PRODUCT 2806 BasicType rtype; 2807 { 2808 ResourceMark rm; 2809 // Check the signatures. 2810 ciSignature* sig = callee()->signature(); 2811 rtype = sig->return_type()->basic_type(); 2812 if (kind == LS_xadd || kind == LS_xchg) { 2813 // Check the signatures. 2814 #ifdef ASSERT 2815 assert(rtype == type, "get and set must return the expected type"); 2816 assert(sig->count() == 3, "get and set has 3 arguments"); 2817 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2818 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2819 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2820 #endif // ASSERT 2821 } else if (kind == LS_cmpxchg) { 2822 // Check the signatures. 2823 #ifdef ASSERT 2824 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2825 assert(sig->count() == 4, "CAS has 4 arguments"); 2826 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2827 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2828 #endif // ASSERT 2829 } else { 2830 ShouldNotReachHere(); 2831 } 2832 } 2833 #endif //PRODUCT 2834 2835 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2836 2837 // Get arguments: 2838 Node* receiver = NULL; 2839 Node* base = NULL; 2840 Node* offset = NULL; 2841 Node* oldval = NULL; 2842 Node* newval = NULL; 2843 if (kind == LS_cmpxchg) { 2844 const bool two_slot_type = type2size[type] == 2; 2845 receiver = argument(0); // type: oop 2846 base = argument(1); // type: oop 2847 offset = argument(2); // type: long 2848 oldval = argument(4); // type: oop, int, or long 2849 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2850 } else if (kind == LS_xadd || kind == LS_xchg){ 2851 receiver = argument(0); // type: oop 2852 base = argument(1); // type: oop 2853 offset = argument(2); // type: long 2854 oldval = NULL; 2855 newval = argument(4); // type: oop, int, or long 2856 } 2857 2858 // Null check receiver. 2859 receiver = null_check(receiver); 2860 if (stopped()) { 2861 return true; 2862 } 2863 2864 // Build field offset expression. 2865 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2866 // to be plain byte offsets, which are also the same as those accepted 2867 // by oopDesc::field_base. 2868 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2869 // 32-bit machines ignore the high half of long offsets 2870 offset = ConvL2X(offset); 2871 Node* adr = make_unsafe_address(base, offset); 2872 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2873 2874 // For CAS, unlike inline_unsafe_access, there seems no point in 2875 // trying to refine types. Just use the coarse types here. 2876 const Type *value_type = Type::get_const_basic_type(type); 2877 Compile::AliasType* alias_type = C->alias_type(adr_type); 2878 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2879 2880 if (kind == LS_xchg && type == T_OBJECT) { 2881 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2882 if (tjp != NULL) { 2883 value_type = tjp; 2884 } 2885 } 2886 2887 int alias_idx = C->get_alias_index(adr_type); 2888 2889 // Memory-model-wise, a LoadStore acts like a little synchronized 2890 // block, so needs barriers on each side. These don't translate 2891 // into actual barriers on most machines, but we still need rest of 2892 // compiler to respect ordering. 2893 2894 insert_mem_bar(Op_MemBarRelease); 2895 insert_mem_bar(Op_MemBarCPUOrder); 2896 2897 // 4984716: MemBars must be inserted before this 2898 // memory node in order to avoid a false 2899 // dependency which will confuse the scheduler. 2900 Node *mem = memory(alias_idx); 2901 2902 // For now, we handle only those cases that actually exist: ints, 2903 // longs, and Object. Adding others should be straightforward. 2904 Node* load_store; 2905 switch(type) { 2906 case T_INT: 2907 if (kind == LS_xadd) { 2908 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2909 } else if (kind == LS_xchg) { 2910 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2911 } else if (kind == LS_cmpxchg) { 2912 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2913 } else { 2914 ShouldNotReachHere(); 2915 } 2916 break; 2917 case T_LONG: 2918 if (kind == LS_xadd) { 2919 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2920 } else if (kind == LS_xchg) { 2921 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2922 } else if (kind == LS_cmpxchg) { 2923 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2924 } else { 2925 ShouldNotReachHere(); 2926 } 2927 break; 2928 case T_OBJECT: 2929 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2930 // could be delayed during Parse (for example, in adjust_map_after_if()). 2931 // Execute transformation here to avoid barrier generation in such case. 2932 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2933 newval = _gvn.makecon(TypePtr::NULL_PTR); 2934 2935 // Reference stores need a store barrier. 2936 if (kind == LS_xchg) { 2937 // If pre-barrier must execute before the oop store, old value will require do_load here. 2938 if (!can_move_pre_barrier()) { 2939 pre_barrier(true /* do_load*/, 2940 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2941 NULL /* pre_val*/, 2942 T_OBJECT); 2943 } // Else move pre_barrier to use load_store value, see below. 2944 } else if (kind == LS_cmpxchg) { 2945 // Same as for newval above: 2946 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2947 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2948 } 2949 // The only known value which might get overwritten is oldval. 2950 pre_barrier(false /* do_load */, 2951 control(), NULL, NULL, max_juint, NULL, NULL, 2952 oldval /* pre_val */, 2953 T_OBJECT); 2954 } else { 2955 ShouldNotReachHere(); 2956 } 2957 2958 #ifdef _LP64 2959 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2960 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2961 if (kind == LS_xchg) { 2962 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2963 newval_enc, adr_type, value_type->make_narrowoop())); 2964 } else { 2965 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2966 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2967 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2968 newval_enc, oldval_enc)); 2969 } 2970 } else 2971 #endif 2972 { 2973 if (kind == LS_xchg) { 2974 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2975 } else { 2976 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2977 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2978 } 2979 } 2980 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2981 break; 2982 default: 2983 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2984 break; 2985 } 2986 2987 // SCMemProjNodes represent the memory state of a LoadStore. Their 2988 // main role is to prevent LoadStore nodes from being optimized away 2989 // when their results aren't used. 2990 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2991 set_memory(proj, alias_idx); 2992 2993 if (type == T_OBJECT && kind == LS_xchg) { 2994 #ifdef _LP64 2995 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2996 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2997 } 2998 #endif 2999 if (can_move_pre_barrier()) { 3000 // Don't need to load pre_val. The old value is returned by load_store. 3001 // The pre_barrier can execute after the xchg as long as no safepoint 3002 // gets inserted between them. 3003 pre_barrier(false /* do_load */, 3004 control(), NULL, NULL, max_juint, NULL, NULL, 3005 load_store /* pre_val */, 3006 T_OBJECT); 3007 } 3008 } 3009 3010 // Add the trailing membar surrounding the access 3011 insert_mem_bar(Op_MemBarCPUOrder); 3012 insert_mem_bar(Op_MemBarAcquire); 3013 3014 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3015 set_result(load_store); 3016 return true; 3017 } 3018 3019 //----------------------------inline_unsafe_ordered_store---------------------- 3020 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3021 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3022 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3023 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3024 // This is another variant of inline_unsafe_access, differing in 3025 // that it always issues store-store ("release") barrier and ensures 3026 // store-atomicity (which only matters for "long"). 3027 3028 if (callee()->is_static()) return false; // caller must have the capability! 3029 3030 #ifndef PRODUCT 3031 { 3032 ResourceMark rm; 3033 // Check the signatures. 3034 ciSignature* sig = callee()->signature(); 3035 #ifdef ASSERT 3036 BasicType rtype = sig->return_type()->basic_type(); 3037 assert(rtype == T_VOID, "must return void"); 3038 assert(sig->count() == 3, "has 3 arguments"); 3039 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3040 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3041 #endif // ASSERT 3042 } 3043 #endif //PRODUCT 3044 3045 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3046 3047 // Get arguments: 3048 Node* receiver = argument(0); // type: oop 3049 Node* base = argument(1); // type: oop 3050 Node* offset = argument(2); // type: long 3051 Node* val = argument(4); // type: oop, int, or long 3052 3053 // Null check receiver. 3054 receiver = null_check(receiver); 3055 if (stopped()) { 3056 return true; 3057 } 3058 3059 // Build field offset expression. 3060 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3061 // 32-bit machines ignore the high half of long offsets 3062 offset = ConvL2X(offset); 3063 Node* adr = make_unsafe_address(base, offset); 3064 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3065 const Type *value_type = Type::get_const_basic_type(type); 3066 Compile::AliasType* alias_type = C->alias_type(adr_type); 3067 3068 insert_mem_bar(Op_MemBarRelease); 3069 insert_mem_bar(Op_MemBarCPUOrder); 3070 // Ensure that the store is atomic for longs: 3071 const bool require_atomic_access = true; 3072 Node* store; 3073 if (type == T_OBJECT) // reference stores need a store barrier. 3074 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3075 else { 3076 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3077 } 3078 insert_mem_bar(Op_MemBarCPUOrder); 3079 return true; 3080 } 3081 3082 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3083 // Regardless of form, don't allow previous ld/st to move down, 3084 // then issue acquire, release, or volatile mem_bar. 3085 insert_mem_bar(Op_MemBarCPUOrder); 3086 switch(id) { 3087 case vmIntrinsics::_loadFence: 3088 insert_mem_bar(Op_LoadFence); 3089 return true; 3090 case vmIntrinsics::_storeFence: 3091 insert_mem_bar(Op_StoreFence); 3092 return true; 3093 case vmIntrinsics::_fullFence: 3094 insert_mem_bar(Op_MemBarVolatile); 3095 return true; 3096 default: 3097 fatal_unexpected_iid(id); 3098 return false; 3099 } 3100 } 3101 3102 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3103 if (!kls->is_Con()) { 3104 return true; 3105 } 3106 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3107 if (klsptr == NULL) { 3108 return true; 3109 } 3110 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3111 // don't need a guard for a klass that is already initialized 3112 return !ik->is_initialized(); 3113 } 3114 3115 //----------------------------inline_unsafe_allocate--------------------------- 3116 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3117 bool LibraryCallKit::inline_unsafe_allocate() { 3118 if (callee()->is_static()) return false; // caller must have the capability! 3119 3120 null_check_receiver(); // null-check, then ignore 3121 Node* cls = null_check(argument(1)); 3122 if (stopped()) return true; 3123 3124 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3125 kls = null_check(kls); 3126 if (stopped()) return true; // argument was like int.class 3127 3128 Node* test = NULL; 3129 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3130 // Note: The argument might still be an illegal value like 3131 // Serializable.class or Object[].class. The runtime will handle it. 3132 // But we must make an explicit check for initialization. 3133 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3134 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3135 // can generate code to load it as unsigned byte. 3136 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3137 Node* bits = intcon(InstanceKlass::fully_initialized); 3138 test = _gvn.transform(new SubINode(inst, bits)); 3139 // The 'test' is non-zero if we need to take a slow path. 3140 } 3141 3142 Node* obj = new_instance(kls, test); 3143 set_result(obj); 3144 return true; 3145 } 3146 3147 #ifdef TRACE_HAVE_INTRINSICS 3148 /* 3149 * oop -> myklass 3150 * myklass->trace_id |= USED 3151 * return myklass->trace_id & ~0x3 3152 */ 3153 bool LibraryCallKit::inline_native_classID() { 3154 null_check_receiver(); // null-check, then ignore 3155 Node* cls = null_check(argument(1), T_OBJECT); 3156 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3157 kls = null_check(kls, T_OBJECT); 3158 ByteSize offset = TRACE_ID_OFFSET; 3159 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3160 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3161 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3162 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3163 Node* clsused = longcon(0x01l); // set the class bit 3164 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3165 3166 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3167 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3168 set_result(andl); 3169 return true; 3170 } 3171 3172 bool LibraryCallKit::inline_native_threadID() { 3173 Node* tls_ptr = NULL; 3174 Node* cur_thr = generate_current_thread(tls_ptr); 3175 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3176 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3177 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3178 3179 Node* threadid = NULL; 3180 size_t thread_id_size = OSThread::thread_id_size(); 3181 if (thread_id_size == (size_t) BytesPerLong) { 3182 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3183 } else if (thread_id_size == (size_t) BytesPerInt) { 3184 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3185 } else { 3186 ShouldNotReachHere(); 3187 } 3188 set_result(threadid); 3189 return true; 3190 } 3191 #endif 3192 3193 //------------------------inline_native_time_funcs-------------- 3194 // inline code for System.currentTimeMillis() and System.nanoTime() 3195 // these have the same type and signature 3196 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3197 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3198 const TypePtr* no_memory_effects = NULL; 3199 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3200 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3201 #ifdef ASSERT 3202 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3203 assert(value_top == top(), "second value must be top"); 3204 #endif 3205 set_result(value); 3206 return true; 3207 } 3208 3209 //------------------------inline_native_currentThread------------------ 3210 bool LibraryCallKit::inline_native_currentThread() { 3211 Node* junk = NULL; 3212 set_result(generate_current_thread(junk)); 3213 return true; 3214 } 3215 3216 //------------------------inline_native_isInterrupted------------------ 3217 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3218 bool LibraryCallKit::inline_native_isInterrupted() { 3219 // Add a fast path to t.isInterrupted(clear_int): 3220 // (t == Thread.current() && 3221 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3222 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3223 // So, in the common case that the interrupt bit is false, 3224 // we avoid making a call into the VM. Even if the interrupt bit 3225 // is true, if the clear_int argument is false, we avoid the VM call. 3226 // However, if the receiver is not currentThread, we must call the VM, 3227 // because there must be some locking done around the operation. 3228 3229 // We only go to the fast case code if we pass two guards. 3230 // Paths which do not pass are accumulated in the slow_region. 3231 3232 enum { 3233 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3234 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3235 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3236 PATH_LIMIT 3237 }; 3238 3239 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3240 // out of the function. 3241 insert_mem_bar(Op_MemBarCPUOrder); 3242 3243 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3244 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3245 3246 RegionNode* slow_region = new RegionNode(1); 3247 record_for_igvn(slow_region); 3248 3249 // (a) Receiving thread must be the current thread. 3250 Node* rec_thr = argument(0); 3251 Node* tls_ptr = NULL; 3252 Node* cur_thr = generate_current_thread(tls_ptr); 3253 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3254 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3255 3256 generate_slow_guard(bol_thr, slow_region); 3257 3258 // (b) Interrupt bit on TLS must be false. 3259 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3260 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3261 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3262 3263 // Set the control input on the field _interrupted read to prevent it floating up. 3264 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3265 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3266 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3267 3268 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3269 3270 // First fast path: if (!TLS._interrupted) return false; 3271 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3272 result_rgn->init_req(no_int_result_path, false_bit); 3273 result_val->init_req(no_int_result_path, intcon(0)); 3274 3275 // drop through to next case 3276 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3277 3278 #ifndef TARGET_OS_FAMILY_windows 3279 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3280 Node* clr_arg = argument(1); 3281 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3282 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3283 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3284 3285 // Second fast path: ... else if (!clear_int) return true; 3286 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3287 result_rgn->init_req(no_clear_result_path, false_arg); 3288 result_val->init_req(no_clear_result_path, intcon(1)); 3289 3290 // drop through to next case 3291 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3292 #else 3293 // To return true on Windows you must read the _interrupted field 3294 // and check the the event state i.e. take the slow path. 3295 #endif // TARGET_OS_FAMILY_windows 3296 3297 // (d) Otherwise, go to the slow path. 3298 slow_region->add_req(control()); 3299 set_control( _gvn.transform(slow_region)); 3300 3301 if (stopped()) { 3302 // There is no slow path. 3303 result_rgn->init_req(slow_result_path, top()); 3304 result_val->init_req(slow_result_path, top()); 3305 } else { 3306 // non-virtual because it is a private non-static 3307 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3308 3309 Node* slow_val = set_results_for_java_call(slow_call); 3310 // this->control() comes from set_results_for_java_call 3311 3312 Node* fast_io = slow_call->in(TypeFunc::I_O); 3313 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3314 3315 // These two phis are pre-filled with copies of of the fast IO and Memory 3316 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3317 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3318 3319 result_rgn->init_req(slow_result_path, control()); 3320 result_io ->init_req(slow_result_path, i_o()); 3321 result_mem->init_req(slow_result_path, reset_memory()); 3322 result_val->init_req(slow_result_path, slow_val); 3323 3324 set_all_memory(_gvn.transform(result_mem)); 3325 set_i_o( _gvn.transform(result_io)); 3326 } 3327 3328 C->set_has_split_ifs(true); // Has chance for split-if optimization 3329 set_result(result_rgn, result_val); 3330 return true; 3331 } 3332 3333 //---------------------------load_mirror_from_klass---------------------------- 3334 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3335 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3336 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3337 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3338 } 3339 3340 //-----------------------load_klass_from_mirror_common------------------------- 3341 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3342 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3343 // and branch to the given path on the region. 3344 // If never_see_null, take an uncommon trap on null, so we can optimistically 3345 // compile for the non-null case. 3346 // If the region is NULL, force never_see_null = true. 3347 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3348 bool never_see_null, 3349 RegionNode* region, 3350 int null_path, 3351 int offset) { 3352 if (region == NULL) never_see_null = true; 3353 Node* p = basic_plus_adr(mirror, offset); 3354 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3355 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3356 Node* null_ctl = top(); 3357 kls = null_check_oop(kls, &null_ctl, never_see_null); 3358 if (region != NULL) { 3359 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3360 region->init_req(null_path, null_ctl); 3361 } else { 3362 assert(null_ctl == top(), "no loose ends"); 3363 } 3364 return kls; 3365 } 3366 3367 //--------------------(inline_native_Class_query helpers)--------------------- 3368 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3369 // Fall through if (mods & mask) == bits, take the guard otherwise. 3370 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3371 // Branch around if the given klass has the given modifier bit set. 3372 // Like generate_guard, adds a new path onto the region. 3373 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3374 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3375 Node* mask = intcon(modifier_mask); 3376 Node* bits = intcon(modifier_bits); 3377 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3378 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3379 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3380 return generate_fair_guard(bol, region); 3381 } 3382 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3383 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3384 } 3385 3386 //-------------------------inline_native_Class_query------------------- 3387 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3388 const Type* return_type = TypeInt::BOOL; 3389 Node* prim_return_value = top(); // what happens if it's a primitive class? 3390 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3391 bool expect_prim = false; // most of these guys expect to work on refs 3392 3393 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3394 3395 Node* mirror = argument(0); 3396 Node* obj = top(); 3397 3398 switch (id) { 3399 case vmIntrinsics::_isInstance: 3400 // nothing is an instance of a primitive type 3401 prim_return_value = intcon(0); 3402 obj = argument(1); 3403 break; 3404 case vmIntrinsics::_getModifiers: 3405 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3406 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3407 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3408 break; 3409 case vmIntrinsics::_isInterface: 3410 prim_return_value = intcon(0); 3411 break; 3412 case vmIntrinsics::_isArray: 3413 prim_return_value = intcon(0); 3414 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3415 break; 3416 case vmIntrinsics::_isPrimitive: 3417 prim_return_value = intcon(1); 3418 expect_prim = true; // obviously 3419 break; 3420 case vmIntrinsics::_getSuperclass: 3421 prim_return_value = null(); 3422 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3423 break; 3424 case vmIntrinsics::_getClassAccessFlags: 3425 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3426 return_type = TypeInt::INT; // not bool! 6297094 3427 break; 3428 default: 3429 fatal_unexpected_iid(id); 3430 break; 3431 } 3432 3433 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3434 if (mirror_con == NULL) return false; // cannot happen? 3435 3436 #ifndef PRODUCT 3437 if (C->print_intrinsics() || C->print_inlining()) { 3438 ciType* k = mirror_con->java_mirror_type(); 3439 if (k) { 3440 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3441 k->print_name(); 3442 tty->cr(); 3443 } 3444 } 3445 #endif 3446 3447 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3448 RegionNode* region = new RegionNode(PATH_LIMIT); 3449 record_for_igvn(region); 3450 PhiNode* phi = new PhiNode(region, return_type); 3451 3452 // The mirror will never be null of Reflection.getClassAccessFlags, however 3453 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3454 // if it is. See bug 4774291. 3455 3456 // For Reflection.getClassAccessFlags(), the null check occurs in 3457 // the wrong place; see inline_unsafe_access(), above, for a similar 3458 // situation. 3459 mirror = null_check(mirror); 3460 // If mirror or obj is dead, only null-path is taken. 3461 if (stopped()) return true; 3462 3463 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3464 3465 // Now load the mirror's klass metaobject, and null-check it. 3466 // Side-effects region with the control path if the klass is null. 3467 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3468 // If kls is null, we have a primitive mirror. 3469 phi->init_req(_prim_path, prim_return_value); 3470 if (stopped()) { set_result(region, phi); return true; } 3471 bool safe_for_replace = (region->in(_prim_path) == top()); 3472 3473 Node* p; // handy temp 3474 Node* null_ctl; 3475 3476 // Now that we have the non-null klass, we can perform the real query. 3477 // For constant classes, the query will constant-fold in LoadNode::Value. 3478 Node* query_value = top(); 3479 switch (id) { 3480 case vmIntrinsics::_isInstance: 3481 // nothing is an instance of a primitive type 3482 query_value = gen_instanceof(obj, kls, safe_for_replace); 3483 break; 3484 3485 case vmIntrinsics::_getModifiers: 3486 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3487 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3488 break; 3489 3490 case vmIntrinsics::_isInterface: 3491 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3492 if (generate_interface_guard(kls, region) != NULL) 3493 // A guard was added. If the guard is taken, it was an interface. 3494 phi->add_req(intcon(1)); 3495 // If we fall through, it's a plain class. 3496 query_value = intcon(0); 3497 break; 3498 3499 case vmIntrinsics::_isArray: 3500 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3501 if (generate_array_guard(kls, region) != NULL) 3502 // A guard was added. If the guard is taken, it was an array. 3503 phi->add_req(intcon(1)); 3504 // If we fall through, it's a plain class. 3505 query_value = intcon(0); 3506 break; 3507 3508 case vmIntrinsics::_isPrimitive: 3509 query_value = intcon(0); // "normal" path produces false 3510 break; 3511 3512 case vmIntrinsics::_getSuperclass: 3513 // The rules here are somewhat unfortunate, but we can still do better 3514 // with random logic than with a JNI call. 3515 // Interfaces store null or Object as _super, but must report null. 3516 // Arrays store an intermediate super as _super, but must report Object. 3517 // Other types can report the actual _super. 3518 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3519 if (generate_interface_guard(kls, region) != NULL) 3520 // A guard was added. If the guard is taken, it was an interface. 3521 phi->add_req(null()); 3522 if (generate_array_guard(kls, region) != NULL) 3523 // A guard was added. If the guard is taken, it was an array. 3524 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3525 // If we fall through, it's a plain class. Get its _super. 3526 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3527 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3528 null_ctl = top(); 3529 kls = null_check_oop(kls, &null_ctl); 3530 if (null_ctl != top()) { 3531 // If the guard is taken, Object.superClass is null (both klass and mirror). 3532 region->add_req(null_ctl); 3533 phi ->add_req(null()); 3534 } 3535 if (!stopped()) { 3536 query_value = load_mirror_from_klass(kls); 3537 } 3538 break; 3539 3540 case vmIntrinsics::_getClassAccessFlags: 3541 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3542 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3543 break; 3544 3545 default: 3546 fatal_unexpected_iid(id); 3547 break; 3548 } 3549 3550 // Fall-through is the normal case of a query to a real class. 3551 phi->init_req(1, query_value); 3552 region->init_req(1, control()); 3553 3554 C->set_has_split_ifs(true); // Has chance for split-if optimization 3555 set_result(region, phi); 3556 return true; 3557 } 3558 3559 //-------------------------inline_Class_cast------------------- 3560 bool LibraryCallKit::inline_Class_cast() { 3561 Node* mirror = argument(0); // Class 3562 Node* obj = argument(1); 3563 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3564 if (mirror_con == NULL) { 3565 return false; // dead path (mirror->is_top()). 3566 } 3567 if (obj == NULL || obj->is_top()) { 3568 return false; // dead path 3569 } 3570 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3571 3572 // First, see if Class.cast() can be folded statically. 3573 // java_mirror_type() returns non-null for compile-time Class constants. 3574 ciType* tm = mirror_con->java_mirror_type(); 3575 if (tm != NULL && tm->is_klass() && 3576 tp != NULL && tp->klass() != NULL) { 3577 if (!tp->klass()->is_loaded()) { 3578 // Don't use intrinsic when class is not loaded. 3579 return false; 3580 } else { 3581 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3582 if (static_res == Compile::SSC_always_true) { 3583 // isInstance() is true - fold the code. 3584 set_result(obj); 3585 return true; 3586 } else if (static_res == Compile::SSC_always_false) { 3587 // Don't use intrinsic, have to throw ClassCastException. 3588 // If the reference is null, the non-intrinsic bytecode will 3589 // be optimized appropriately. 3590 return false; 3591 } 3592 } 3593 } 3594 3595 // Bailout intrinsic and do normal inlining if exception path is frequent. 3596 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3597 return false; 3598 } 3599 3600 // Generate dynamic checks. 3601 // Class.cast() is java implementation of _checkcast bytecode. 3602 // Do checkcast (Parse::do_checkcast()) optimizations here. 3603 3604 mirror = null_check(mirror); 3605 // If mirror is dead, only null-path is taken. 3606 if (stopped()) { 3607 return true; 3608 } 3609 3610 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3611 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3612 RegionNode* region = new RegionNode(PATH_LIMIT); 3613 record_for_igvn(region); 3614 3615 // Now load the mirror's klass metaobject, and null-check it. 3616 // If kls is null, we have a primitive mirror and 3617 // nothing is an instance of a primitive type. 3618 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3619 3620 Node* res = top(); 3621 if (!stopped()) { 3622 Node* bad_type_ctrl = top(); 3623 // Do checkcast optimizations. 3624 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3625 region->init_req(_bad_type_path, bad_type_ctrl); 3626 } 3627 if (region->in(_prim_path) != top() || 3628 region->in(_bad_type_path) != top()) { 3629 // Let Interpreter throw ClassCastException. 3630 PreserveJVMState pjvms(this); 3631 set_control(_gvn.transform(region)); 3632 uncommon_trap(Deoptimization::Reason_intrinsic, 3633 Deoptimization::Action_maybe_recompile); 3634 } 3635 if (!stopped()) { 3636 set_result(res); 3637 } 3638 return true; 3639 } 3640 3641 3642 //--------------------------inline_native_subtype_check------------------------ 3643 // This intrinsic takes the JNI calls out of the heart of 3644 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3645 bool LibraryCallKit::inline_native_subtype_check() { 3646 // Pull both arguments off the stack. 3647 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3648 args[0] = argument(0); 3649 args[1] = argument(1); 3650 Node* klasses[2]; // corresponding Klasses: superk, subk 3651 klasses[0] = klasses[1] = top(); 3652 3653 enum { 3654 // A full decision tree on {superc is prim, subc is prim}: 3655 _prim_0_path = 1, // {P,N} => false 3656 // {P,P} & superc!=subc => false 3657 _prim_same_path, // {P,P} & superc==subc => true 3658 _prim_1_path, // {N,P} => false 3659 _ref_subtype_path, // {N,N} & subtype check wins => true 3660 _both_ref_path, // {N,N} & subtype check loses => false 3661 PATH_LIMIT 3662 }; 3663 3664 RegionNode* region = new RegionNode(PATH_LIMIT); 3665 Node* phi = new PhiNode(region, TypeInt::BOOL); 3666 record_for_igvn(region); 3667 3668 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3669 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3670 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3671 3672 // First null-check both mirrors and load each mirror's klass metaobject. 3673 int which_arg; 3674 for (which_arg = 0; which_arg <= 1; which_arg++) { 3675 Node* arg = args[which_arg]; 3676 arg = null_check(arg); 3677 if (stopped()) break; 3678 args[which_arg] = arg; 3679 3680 Node* p = basic_plus_adr(arg, class_klass_offset); 3681 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3682 klasses[which_arg] = _gvn.transform(kls); 3683 } 3684 3685 // Having loaded both klasses, test each for null. 3686 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3687 for (which_arg = 0; which_arg <= 1; which_arg++) { 3688 Node* kls = klasses[which_arg]; 3689 Node* null_ctl = top(); 3690 kls = null_check_oop(kls, &null_ctl, never_see_null); 3691 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3692 region->init_req(prim_path, null_ctl); 3693 if (stopped()) break; 3694 klasses[which_arg] = kls; 3695 } 3696 3697 if (!stopped()) { 3698 // now we have two reference types, in klasses[0..1] 3699 Node* subk = klasses[1]; // the argument to isAssignableFrom 3700 Node* superk = klasses[0]; // the receiver 3701 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3702 // now we have a successful reference subtype check 3703 region->set_req(_ref_subtype_path, control()); 3704 } 3705 3706 // If both operands are primitive (both klasses null), then 3707 // we must return true when they are identical primitives. 3708 // It is convenient to test this after the first null klass check. 3709 set_control(region->in(_prim_0_path)); // go back to first null check 3710 if (!stopped()) { 3711 // Since superc is primitive, make a guard for the superc==subc case. 3712 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3713 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3714 generate_guard(bol_eq, region, PROB_FAIR); 3715 if (region->req() == PATH_LIMIT+1) { 3716 // A guard was added. If the added guard is taken, superc==subc. 3717 region->swap_edges(PATH_LIMIT, _prim_same_path); 3718 region->del_req(PATH_LIMIT); 3719 } 3720 region->set_req(_prim_0_path, control()); // Not equal after all. 3721 } 3722 3723 // these are the only paths that produce 'true': 3724 phi->set_req(_prim_same_path, intcon(1)); 3725 phi->set_req(_ref_subtype_path, intcon(1)); 3726 3727 // pull together the cases: 3728 assert(region->req() == PATH_LIMIT, "sane region"); 3729 for (uint i = 1; i < region->req(); i++) { 3730 Node* ctl = region->in(i); 3731 if (ctl == NULL || ctl == top()) { 3732 region->set_req(i, top()); 3733 phi ->set_req(i, top()); 3734 } else if (phi->in(i) == NULL) { 3735 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3736 } 3737 } 3738 3739 set_control(_gvn.transform(region)); 3740 set_result(_gvn.transform(phi)); 3741 return true; 3742 } 3743 3744 //---------------------generate_array_guard_common------------------------ 3745 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3746 bool obj_array, bool not_array) { 3747 3748 if (stopped()) { 3749 return NULL; 3750 } 3751 3752 // If obj_array/non_array==false/false: 3753 // Branch around if the given klass is in fact an array (either obj or prim). 3754 // If obj_array/non_array==false/true: 3755 // Branch around if the given klass is not an array klass of any kind. 3756 // If obj_array/non_array==true/true: 3757 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3758 // If obj_array/non_array==true/false: 3759 // Branch around if the kls is an oop array (Object[] or subtype) 3760 // 3761 // Like generate_guard, adds a new path onto the region. 3762 jint layout_con = 0; 3763 Node* layout_val = get_layout_helper(kls, layout_con); 3764 if (layout_val == NULL) { 3765 bool query = (obj_array 3766 ? Klass::layout_helper_is_objArray(layout_con) 3767 : Klass::layout_helper_is_array(layout_con)); 3768 if (query == not_array) { 3769 return NULL; // never a branch 3770 } else { // always a branch 3771 Node* always_branch = control(); 3772 if (region != NULL) 3773 region->add_req(always_branch); 3774 set_control(top()); 3775 return always_branch; 3776 } 3777 } 3778 // Now test the correct condition. 3779 jint nval = (obj_array 3780 ? ((jint)Klass::_lh_array_tag_type_value 3781 << Klass::_lh_array_tag_shift) 3782 : Klass::_lh_neutral_value); 3783 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3784 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3785 // invert the test if we are looking for a non-array 3786 if (not_array) btest = BoolTest(btest).negate(); 3787 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3788 return generate_fair_guard(bol, region); 3789 } 3790 3791 3792 //-----------------------inline_native_newArray-------------------------- 3793 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3794 bool LibraryCallKit::inline_native_newArray() { 3795 Node* mirror = argument(0); 3796 Node* count_val = argument(1); 3797 3798 mirror = null_check(mirror); 3799 // If mirror or obj is dead, only null-path is taken. 3800 if (stopped()) return true; 3801 3802 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3803 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3804 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3805 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3806 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3807 3808 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3809 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3810 result_reg, _slow_path); 3811 Node* normal_ctl = control(); 3812 Node* no_array_ctl = result_reg->in(_slow_path); 3813 3814 // Generate code for the slow case. We make a call to newArray(). 3815 set_control(no_array_ctl); 3816 if (!stopped()) { 3817 // Either the input type is void.class, or else the 3818 // array klass has not yet been cached. Either the 3819 // ensuing call will throw an exception, or else it 3820 // will cache the array klass for next time. 3821 PreserveJVMState pjvms(this); 3822 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3823 Node* slow_result = set_results_for_java_call(slow_call); 3824 // this->control() comes from set_results_for_java_call 3825 result_reg->set_req(_slow_path, control()); 3826 result_val->set_req(_slow_path, slow_result); 3827 result_io ->set_req(_slow_path, i_o()); 3828 result_mem->set_req(_slow_path, reset_memory()); 3829 } 3830 3831 set_control(normal_ctl); 3832 if (!stopped()) { 3833 // Normal case: The array type has been cached in the java.lang.Class. 3834 // The following call works fine even if the array type is polymorphic. 3835 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3836 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3837 result_reg->init_req(_normal_path, control()); 3838 result_val->init_req(_normal_path, obj); 3839 result_io ->init_req(_normal_path, i_o()); 3840 result_mem->init_req(_normal_path, reset_memory()); 3841 } 3842 3843 // Return the combined state. 3844 set_i_o( _gvn.transform(result_io) ); 3845 set_all_memory( _gvn.transform(result_mem)); 3846 3847 C->set_has_split_ifs(true); // Has chance for split-if optimization 3848 set_result(result_reg, result_val); 3849 return true; 3850 } 3851 3852 //----------------------inline_native_getLength-------------------------- 3853 // public static native int java.lang.reflect.Array.getLength(Object array); 3854 bool LibraryCallKit::inline_native_getLength() { 3855 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3856 3857 Node* array = null_check(argument(0)); 3858 // If array is dead, only null-path is taken. 3859 if (stopped()) return true; 3860 3861 // Deoptimize if it is a non-array. 3862 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3863 3864 if (non_array != NULL) { 3865 PreserveJVMState pjvms(this); 3866 set_control(non_array); 3867 uncommon_trap(Deoptimization::Reason_intrinsic, 3868 Deoptimization::Action_maybe_recompile); 3869 } 3870 3871 // If control is dead, only non-array-path is taken. 3872 if (stopped()) return true; 3873 3874 // The works fine even if the array type is polymorphic. 3875 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3876 Node* result = load_array_length(array); 3877 3878 C->set_has_split_ifs(true); // Has chance for split-if optimization 3879 set_result(result); 3880 return true; 3881 } 3882 3883 //------------------------inline_array_copyOf---------------------------- 3884 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3885 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3886 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3887 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3888 3889 // Get the arguments. 3890 Node* original = argument(0); 3891 Node* start = is_copyOfRange? argument(1): intcon(0); 3892 Node* end = is_copyOfRange? argument(2): argument(1); 3893 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3894 3895 Node* newcopy; 3896 3897 // Set the original stack and the reexecute bit for the interpreter to reexecute 3898 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3899 { PreserveReexecuteState preexecs(this); 3900 jvms()->set_should_reexecute(true); 3901 3902 array_type_mirror = null_check(array_type_mirror); 3903 original = null_check(original); 3904 3905 // Check if a null path was taken unconditionally. 3906 if (stopped()) return true; 3907 3908 Node* orig_length = load_array_length(original); 3909 3910 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3911 klass_node = null_check(klass_node); 3912 3913 RegionNode* bailout = new RegionNode(1); 3914 record_for_igvn(bailout); 3915 3916 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3917 // Bail out if that is so. 3918 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3919 if (not_objArray != NULL) { 3920 // Improve the klass node's type from the new optimistic assumption: 3921 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3922 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3923 Node* cast = new CastPPNode(klass_node, akls); 3924 cast->init_req(0, control()); 3925 klass_node = _gvn.transform(cast); 3926 } 3927 3928 // Bail out if either start or end is negative. 3929 generate_negative_guard(start, bailout, &start); 3930 generate_negative_guard(end, bailout, &end); 3931 3932 Node* length = end; 3933 if (_gvn.type(start) != TypeInt::ZERO) { 3934 length = _gvn.transform(new SubINode(end, start)); 3935 } 3936 3937 // Bail out if length is negative. 3938 // Without this the new_array would throw 3939 // NegativeArraySizeException but IllegalArgumentException is what 3940 // should be thrown 3941 generate_negative_guard(length, bailout, &length); 3942 3943 if (bailout->req() > 1) { 3944 PreserveJVMState pjvms(this); 3945 set_control(_gvn.transform(bailout)); 3946 uncommon_trap(Deoptimization::Reason_intrinsic, 3947 Deoptimization::Action_maybe_recompile); 3948 } 3949 3950 if (!stopped()) { 3951 // How many elements will we copy from the original? 3952 // The answer is MinI(orig_length - start, length). 3953 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3954 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3955 3956 // Generate a direct call to the right arraycopy function(s). 3957 // We know the copy is disjoint but we might not know if the 3958 // oop stores need checking. 3959 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3960 // This will fail a store-check if x contains any non-nulls. 3961 3962 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3963 // loads/stores but it is legal only if we're sure the 3964 // Arrays.copyOf would succeed. So we need all input arguments 3965 // to the copyOf to be validated, including that the copy to the 3966 // new array won't trigger an ArrayStoreException. That subtype 3967 // check can be optimized if we know something on the type of 3968 // the input array from type speculation. 3969 if (_gvn.type(klass_node)->singleton()) { 3970 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3971 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3972 3973 int test = C->static_subtype_check(superk, subk); 3974 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3975 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3976 if (t_original->speculative_type() != NULL) { 3977 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3978 } 3979 } 3980 } 3981 3982 bool validated = false; 3983 // Reason_class_check rather than Reason_intrinsic because we 3984 // want to intrinsify even if this traps. 3985 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3986 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3987 klass_node); 3988 3989 if (not_subtype_ctrl != top()) { 3990 PreserveJVMState pjvms(this); 3991 set_control(not_subtype_ctrl); 3992 uncommon_trap(Deoptimization::Reason_class_check, 3993 Deoptimization::Action_make_not_entrant); 3994 assert(stopped(), "Should be stopped"); 3995 } 3996 validated = true; 3997 } 3998 3999 if (!stopped()) { 4000 newcopy = new_array(klass_node, length, 0); // no arguments to push 4001 4002 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 4003 load_object_klass(original), klass_node); 4004 if (!is_copyOfRange) { 4005 ac->set_copyof(validated); 4006 } else { 4007 ac->set_copyofrange(validated); 4008 } 4009 Node* n = _gvn.transform(ac); 4010 if (n == ac) { 4011 ac->connect_outputs(this); 4012 } else { 4013 assert(validated, "shouldn't transform if all arguments not validated"); 4014 set_all_memory(n); 4015 } 4016 } 4017 } 4018 } // original reexecute is set back here 4019 4020 C->set_has_split_ifs(true); // Has chance for split-if optimization 4021 if (!stopped()) { 4022 set_result(newcopy); 4023 } 4024 return true; 4025 } 4026 4027 4028 //----------------------generate_virtual_guard--------------------------- 4029 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4030 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4031 RegionNode* slow_region) { 4032 ciMethod* method = callee(); 4033 int vtable_index = method->vtable_index(); 4034 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4035 err_msg_res("bad index %d", vtable_index)); 4036 // Get the Method* out of the appropriate vtable entry. 4037 int entry_offset = (InstanceKlass::vtable_start_offset() + 4038 vtable_index*vtableEntry::size()) * wordSize + 4039 vtableEntry::method_offset_in_bytes(); 4040 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4041 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4042 4043 // Compare the target method with the expected method (e.g., Object.hashCode). 4044 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4045 4046 Node* native_call = makecon(native_call_addr); 4047 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4048 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4049 4050 return generate_slow_guard(test_native, slow_region); 4051 } 4052 4053 //-----------------------generate_method_call---------------------------- 4054 // Use generate_method_call to make a slow-call to the real 4055 // method if the fast path fails. An alternative would be to 4056 // use a stub like OptoRuntime::slow_arraycopy_Java. 4057 // This only works for expanding the current library call, 4058 // not another intrinsic. (E.g., don't use this for making an 4059 // arraycopy call inside of the copyOf intrinsic.) 4060 CallJavaNode* 4061 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4062 // When compiling the intrinsic method itself, do not use this technique. 4063 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4064 4065 ciMethod* method = callee(); 4066 // ensure the JVMS we have will be correct for this call 4067 guarantee(method_id == method->intrinsic_id(), "must match"); 4068 4069 const TypeFunc* tf = TypeFunc::make(method); 4070 CallJavaNode* slow_call; 4071 if (is_static) { 4072 assert(!is_virtual, ""); 4073 slow_call = new CallStaticJavaNode(C, tf, 4074 SharedRuntime::get_resolve_static_call_stub(), 4075 method, bci()); 4076 } else if (is_virtual) { 4077 null_check_receiver(); 4078 int vtable_index = Method::invalid_vtable_index; 4079 if (UseInlineCaches) { 4080 // Suppress the vtable call 4081 } else { 4082 // hashCode and clone are not a miranda methods, 4083 // so the vtable index is fixed. 4084 // No need to use the linkResolver to get it. 4085 vtable_index = method->vtable_index(); 4086 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4087 err_msg_res("bad index %d", vtable_index)); 4088 } 4089 slow_call = new CallDynamicJavaNode(tf, 4090 SharedRuntime::get_resolve_virtual_call_stub(), 4091 method, vtable_index, bci()); 4092 } else { // neither virtual nor static: opt_virtual 4093 null_check_receiver(); 4094 slow_call = new CallStaticJavaNode(C, tf, 4095 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4096 method, bci()); 4097 slow_call->set_optimized_virtual(true); 4098 } 4099 set_arguments_for_java_call(slow_call); 4100 set_edges_for_java_call(slow_call); 4101 return slow_call; 4102 } 4103 4104 4105 /** 4106 * Build special case code for calls to hashCode on an object. This call may 4107 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4108 * slightly different code. 4109 */ 4110 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4111 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4112 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4113 4114 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4115 4116 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4117 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4118 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4119 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4120 Node* obj = NULL; 4121 if (!is_static) { 4122 // Check for hashing null object 4123 obj = null_check_receiver(); 4124 if (stopped()) return true; // unconditionally null 4125 result_reg->init_req(_null_path, top()); 4126 result_val->init_req(_null_path, top()); 4127 } else { 4128 // Do a null check, and return zero if null. 4129 // System.identityHashCode(null) == 0 4130 obj = argument(0); 4131 Node* null_ctl = top(); 4132 obj = null_check_oop(obj, &null_ctl); 4133 result_reg->init_req(_null_path, null_ctl); 4134 result_val->init_req(_null_path, _gvn.intcon(0)); 4135 } 4136 4137 // Unconditionally null? Then return right away. 4138 if (stopped()) { 4139 set_control( result_reg->in(_null_path)); 4140 if (!stopped()) 4141 set_result(result_val->in(_null_path)); 4142 return true; 4143 } 4144 4145 // We only go to the fast case code if we pass a number of guards. The 4146 // paths which do not pass are accumulated in the slow_region. 4147 RegionNode* slow_region = new RegionNode(1); 4148 record_for_igvn(slow_region); 4149 4150 // If this is a virtual call, we generate a funny guard. We pull out 4151 // the vtable entry corresponding to hashCode() from the target object. 4152 // If the target method which we are calling happens to be the native 4153 // Object hashCode() method, we pass the guard. We do not need this 4154 // guard for non-virtual calls -- the caller is known to be the native 4155 // Object hashCode(). 4156 if (is_virtual) { 4157 // After null check, get the object's klass. 4158 Node* obj_klass = load_object_klass(obj); 4159 generate_virtual_guard(obj_klass, slow_region); 4160 } 4161 4162 // Get the header out of the object, use LoadMarkNode when available 4163 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4164 // The control of the load must be NULL. Otherwise, the load can move before 4165 // the null check after castPP removal. 4166 Node* no_ctrl = NULL; 4167 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4168 4169 // Test the header to see if it is unlocked. 4170 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4171 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4172 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4173 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4174 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4175 4176 generate_slow_guard(test_unlocked, slow_region); 4177 4178 // Get the hash value and check to see that it has been properly assigned. 4179 // We depend on hash_mask being at most 32 bits and avoid the use of 4180 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4181 // vm: see markOop.hpp. 4182 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4183 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4184 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4185 // This hack lets the hash bits live anywhere in the mark object now, as long 4186 // as the shift drops the relevant bits into the low 32 bits. Note that 4187 // Java spec says that HashCode is an int so there's no point in capturing 4188 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4189 hshifted_header = ConvX2I(hshifted_header); 4190 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4191 4192 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4193 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4194 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4195 4196 generate_slow_guard(test_assigned, slow_region); 4197 4198 Node* init_mem = reset_memory(); 4199 // fill in the rest of the null path: 4200 result_io ->init_req(_null_path, i_o()); 4201 result_mem->init_req(_null_path, init_mem); 4202 4203 result_val->init_req(_fast_path, hash_val); 4204 result_reg->init_req(_fast_path, control()); 4205 result_io ->init_req(_fast_path, i_o()); 4206 result_mem->init_req(_fast_path, init_mem); 4207 4208 // Generate code for the slow case. We make a call to hashCode(). 4209 set_control(_gvn.transform(slow_region)); 4210 if (!stopped()) { 4211 // No need for PreserveJVMState, because we're using up the present state. 4212 set_all_memory(init_mem); 4213 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4214 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4215 Node* slow_result = set_results_for_java_call(slow_call); 4216 // this->control() comes from set_results_for_java_call 4217 result_reg->init_req(_slow_path, control()); 4218 result_val->init_req(_slow_path, slow_result); 4219 result_io ->set_req(_slow_path, i_o()); 4220 result_mem ->set_req(_slow_path, reset_memory()); 4221 } 4222 4223 // Return the combined state. 4224 set_i_o( _gvn.transform(result_io) ); 4225 set_all_memory( _gvn.transform(result_mem)); 4226 4227 set_result(result_reg, result_val); 4228 return true; 4229 } 4230 4231 //---------------------------inline_native_getClass---------------------------- 4232 // public final native Class<?> java.lang.Object.getClass(); 4233 // 4234 // Build special case code for calls to getClass on an object. 4235 bool LibraryCallKit::inline_native_getClass() { 4236 Node* obj = null_check_receiver(); 4237 if (stopped()) return true; 4238 set_result(load_mirror_from_klass(load_object_klass(obj))); 4239 return true; 4240 } 4241 4242 //-----------------inline_native_Reflection_getCallerClass--------------------- 4243 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4244 // 4245 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4246 // 4247 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4248 // in that it must skip particular security frames and checks for 4249 // caller sensitive methods. 4250 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4251 #ifndef PRODUCT 4252 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4253 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4254 } 4255 #endif 4256 4257 if (!jvms()->has_method()) { 4258 #ifndef PRODUCT 4259 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4260 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4261 } 4262 #endif 4263 return false; 4264 } 4265 4266 // Walk back up the JVM state to find the caller at the required 4267 // depth. 4268 JVMState* caller_jvms = jvms(); 4269 4270 // Cf. JVM_GetCallerClass 4271 // NOTE: Start the loop at depth 1 because the current JVM state does 4272 // not include the Reflection.getCallerClass() frame. 4273 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4274 ciMethod* m = caller_jvms->method(); 4275 switch (n) { 4276 case 0: 4277 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4278 break; 4279 case 1: 4280 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4281 if (!m->caller_sensitive()) { 4282 #ifndef PRODUCT 4283 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4284 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4285 } 4286 #endif 4287 return false; // bail-out; let JVM_GetCallerClass do the work 4288 } 4289 break; 4290 default: 4291 if (!m->is_ignored_by_security_stack_walk()) { 4292 // We have reached the desired frame; return the holder class. 4293 // Acquire method holder as java.lang.Class and push as constant. 4294 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4295 ciInstance* caller_mirror = caller_klass->java_mirror(); 4296 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4297 4298 #ifndef PRODUCT 4299 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4300 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()); 4301 tty->print_cr(" JVM state at this point:"); 4302 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4303 ciMethod* m = jvms()->of_depth(i)->method(); 4304 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4305 } 4306 } 4307 #endif 4308 return true; 4309 } 4310 break; 4311 } 4312 } 4313 4314 #ifndef PRODUCT 4315 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4316 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4317 tty->print_cr(" JVM state at this point:"); 4318 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4319 ciMethod* m = jvms()->of_depth(i)->method(); 4320 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4321 } 4322 } 4323 #endif 4324 4325 return false; // bail-out; let JVM_GetCallerClass do the work 4326 } 4327 4328 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4329 Node* arg = argument(0); 4330 Node* result; 4331 4332 switch (id) { 4333 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4334 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4335 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4336 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4337 4338 case vmIntrinsics::_doubleToLongBits: { 4339 // two paths (plus control) merge in a wood 4340 RegionNode *r = new RegionNode(3); 4341 Node *phi = new PhiNode(r, TypeLong::LONG); 4342 4343 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4344 // Build the boolean node 4345 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4346 4347 // Branch either way. 4348 // NaN case is less traveled, which makes all the difference. 4349 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4350 Node *opt_isnan = _gvn.transform(ifisnan); 4351 assert( opt_isnan->is_If(), "Expect an IfNode"); 4352 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4353 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4354 4355 set_control(iftrue); 4356 4357 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4358 Node *slow_result = longcon(nan_bits); // return NaN 4359 phi->init_req(1, _gvn.transform( slow_result )); 4360 r->init_req(1, iftrue); 4361 4362 // Else fall through 4363 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4364 set_control(iffalse); 4365 4366 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4367 r->init_req(2, iffalse); 4368 4369 // Post merge 4370 set_control(_gvn.transform(r)); 4371 record_for_igvn(r); 4372 4373 C->set_has_split_ifs(true); // Has chance for split-if optimization 4374 result = phi; 4375 assert(result->bottom_type()->isa_long(), "must be"); 4376 break; 4377 } 4378 4379 case vmIntrinsics::_floatToIntBits: { 4380 // two paths (plus control) merge in a wood 4381 RegionNode *r = new RegionNode(3); 4382 Node *phi = new PhiNode(r, TypeInt::INT); 4383 4384 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4385 // Build the boolean node 4386 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4387 4388 // Branch either way. 4389 // NaN case is less traveled, which makes all the difference. 4390 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4391 Node *opt_isnan = _gvn.transform(ifisnan); 4392 assert( opt_isnan->is_If(), "Expect an IfNode"); 4393 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4394 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4395 4396 set_control(iftrue); 4397 4398 static const jint nan_bits = 0x7fc00000; 4399 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4400 phi->init_req(1, _gvn.transform( slow_result )); 4401 r->init_req(1, iftrue); 4402 4403 // Else fall through 4404 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4405 set_control(iffalse); 4406 4407 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4408 r->init_req(2, iffalse); 4409 4410 // Post merge 4411 set_control(_gvn.transform(r)); 4412 record_for_igvn(r); 4413 4414 C->set_has_split_ifs(true); // Has chance for split-if optimization 4415 result = phi; 4416 assert(result->bottom_type()->isa_int(), "must be"); 4417 break; 4418 } 4419 4420 default: 4421 fatal_unexpected_iid(id); 4422 break; 4423 } 4424 set_result(_gvn.transform(result)); 4425 return true; 4426 } 4427 4428 #ifdef _LP64 4429 #define XTOP ,top() /*additional argument*/ 4430 #else //_LP64 4431 #define XTOP /*no additional argument*/ 4432 #endif //_LP64 4433 4434 //----------------------inline_unsafe_copyMemory------------------------- 4435 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4436 bool LibraryCallKit::inline_unsafe_copyMemory() { 4437 if (callee()->is_static()) return false; // caller must have the capability! 4438 null_check_receiver(); // null-check receiver 4439 if (stopped()) return true; 4440 4441 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4442 4443 Node* src_ptr = argument(1); // type: oop 4444 Node* src_off = ConvL2X(argument(2)); // type: long 4445 Node* dst_ptr = argument(4); // type: oop 4446 Node* dst_off = ConvL2X(argument(5)); // type: long 4447 Node* size = ConvL2X(argument(7)); // type: long 4448 4449 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4450 "fieldOffset must be byte-scaled"); 4451 4452 Node* src = make_unsafe_address(src_ptr, src_off); 4453 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4454 4455 // Conservatively insert a memory barrier on all memory slices. 4456 // Do not let writes of the copy source or destination float below the copy. 4457 insert_mem_bar(Op_MemBarCPUOrder); 4458 4459 // Call it. Note that the length argument is not scaled. 4460 make_runtime_call(RC_LEAF|RC_NO_FP, 4461 OptoRuntime::fast_arraycopy_Type(), 4462 StubRoutines::unsafe_arraycopy(), 4463 "unsafe_arraycopy", 4464 TypeRawPtr::BOTTOM, 4465 src, dst, size XTOP); 4466 4467 // Do not let reads of the copy destination float above the copy. 4468 insert_mem_bar(Op_MemBarCPUOrder); 4469 4470 return true; 4471 } 4472 4473 //------------------------clone_coping----------------------------------- 4474 // Helper function for inline_native_clone. 4475 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4476 assert(obj_size != NULL, ""); 4477 Node* raw_obj = alloc_obj->in(1); 4478 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4479 4480 AllocateNode* alloc = NULL; 4481 if (ReduceBulkZeroing) { 4482 // We will be completely responsible for initializing this object - 4483 // mark Initialize node as complete. 4484 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4485 // The object was just allocated - there should be no any stores! 4486 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4487 // Mark as complete_with_arraycopy so that on AllocateNode 4488 // expansion, we know this AllocateNode is initialized by an array 4489 // copy and a StoreStore barrier exists after the array copy. 4490 alloc->initialization()->set_complete_with_arraycopy(); 4491 } 4492 4493 // Copy the fastest available way. 4494 // TODO: generate fields copies for small objects instead. 4495 Node* src = obj; 4496 Node* dest = alloc_obj; 4497 Node* size = _gvn.transform(obj_size); 4498 4499 // Exclude the header but include array length to copy by 8 bytes words. 4500 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4501 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4502 instanceOopDesc::base_offset_in_bytes(); 4503 // base_off: 4504 // 8 - 32-bit VM 4505 // 12 - 64-bit VM, compressed klass 4506 // 16 - 64-bit VM, normal klass 4507 if (base_off % BytesPerLong != 0) { 4508 assert(UseCompressedClassPointers, ""); 4509 if (is_array) { 4510 // Exclude length to copy by 8 bytes words. 4511 base_off += sizeof(int); 4512 } else { 4513 // Include klass to copy by 8 bytes words. 4514 base_off = instanceOopDesc::klass_offset_in_bytes(); 4515 } 4516 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4517 } 4518 src = basic_plus_adr(src, base_off); 4519 dest = basic_plus_adr(dest, base_off); 4520 4521 // Compute the length also, if needed: 4522 Node* countx = size; 4523 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4524 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4525 4526 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4527 4528 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4529 ac->set_clonebasic(); 4530 Node* n = _gvn.transform(ac); 4531 if (n == ac) { 4532 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4533 } else { 4534 set_all_memory(n); 4535 } 4536 4537 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4538 if (card_mark) { 4539 assert(!is_array, ""); 4540 // Put in store barrier for any and all oops we are sticking 4541 // into this object. (We could avoid this if we could prove 4542 // that the object type contains no oop fields at all.) 4543 Node* no_particular_value = NULL; 4544 Node* no_particular_field = NULL; 4545 int raw_adr_idx = Compile::AliasIdxRaw; 4546 post_barrier(control(), 4547 memory(raw_adr_type), 4548 alloc_obj, 4549 no_particular_field, 4550 raw_adr_idx, 4551 no_particular_value, 4552 T_OBJECT, 4553 false); 4554 } 4555 4556 // Do not let reads from the cloned object float above the arraycopy. 4557 if (alloc != NULL) { 4558 // Do not let stores that initialize this object be reordered with 4559 // a subsequent store that would make this object accessible by 4560 // other threads. 4561 // Record what AllocateNode this StoreStore protects so that 4562 // escape analysis can go from the MemBarStoreStoreNode to the 4563 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4564 // based on the escape status of the AllocateNode. 4565 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4566 } else { 4567 insert_mem_bar(Op_MemBarCPUOrder); 4568 } 4569 } 4570 4571 //------------------------inline_native_clone---------------------------- 4572 // protected native Object java.lang.Object.clone(); 4573 // 4574 // Here are the simple edge cases: 4575 // null receiver => normal trap 4576 // virtual and clone was overridden => slow path to out-of-line clone 4577 // not cloneable or finalizer => slow path to out-of-line Object.clone 4578 // 4579 // The general case has two steps, allocation and copying. 4580 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4581 // 4582 // Copying also has two cases, oop arrays and everything else. 4583 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4584 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4585 // 4586 // These steps fold up nicely if and when the cloned object's klass 4587 // can be sharply typed as an object array, a type array, or an instance. 4588 // 4589 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4590 PhiNode* result_val; 4591 4592 // Set the reexecute bit for the interpreter to reexecute 4593 // the bytecode that invokes Object.clone if deoptimization happens. 4594 { PreserveReexecuteState preexecs(this); 4595 jvms()->set_should_reexecute(true); 4596 4597 Node* obj = null_check_receiver(); 4598 if (stopped()) return true; 4599 4600 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4601 4602 // If we are going to clone an instance, we need its exact type to 4603 // know the number and types of fields to convert the clone to 4604 // loads/stores. Maybe a speculative type can help us. 4605 if (!obj_type->klass_is_exact() && 4606 obj_type->speculative_type() != NULL && 4607 obj_type->speculative_type()->is_instance_klass()) { 4608 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4609 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4610 !spec_ik->has_injected_fields()) { 4611 ciKlass* k = obj_type->klass(); 4612 if (!k->is_instance_klass() || 4613 k->as_instance_klass()->is_interface() || 4614 k->as_instance_klass()->has_subklass()) { 4615 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4616 } 4617 } 4618 } 4619 4620 Node* obj_klass = load_object_klass(obj); 4621 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4622 const TypeOopPtr* toop = ((tklass != NULL) 4623 ? tklass->as_instance_type() 4624 : TypeInstPtr::NOTNULL); 4625 4626 // Conservatively insert a memory barrier on all memory slices. 4627 // Do not let writes into the original float below the clone. 4628 insert_mem_bar(Op_MemBarCPUOrder); 4629 4630 // paths into result_reg: 4631 enum { 4632 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4633 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4634 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4635 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4636 PATH_LIMIT 4637 }; 4638 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4639 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4640 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4641 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, 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 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4665 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4666 ac->set_cloneoop(); 4667 Node* n = _gvn.transform(ac); 4668 assert(n == ac, "cannot disappear"); 4669 ac->connect_outputs(this); 4670 4671 result_reg->init_req(_objArray_path, control()); 4672 result_val->init_req(_objArray_path, alloc_obj); 4673 result_i_o ->set_req(_objArray_path, i_o()); 4674 result_mem ->set_req(_objArray_path, reset_memory()); 4675 } 4676 } 4677 // Otherwise, there are no card marks to worry about. 4678 // (We can dispense with card marks if we know the allocation 4679 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4680 // causes the non-eden paths to take compensating steps to 4681 // simulate a fresh allocation, so that no further 4682 // card marks are required in compiled code to initialize 4683 // the object.) 4684 4685 if (!stopped()) { 4686 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4687 4688 // Present the results of the copy. 4689 result_reg->init_req(_array_path, control()); 4690 result_val->init_req(_array_path, alloc_obj); 4691 result_i_o ->set_req(_array_path, i_o()); 4692 result_mem ->set_req(_array_path, reset_memory()); 4693 } 4694 } 4695 4696 // We only go to the instance fast case code if we pass a number of guards. 4697 // The paths which do not pass are accumulated in the slow_region. 4698 RegionNode* slow_region = new RegionNode(1); 4699 record_for_igvn(slow_region); 4700 if (!stopped()) { 4701 // It's an instance (we did array above). Make the slow-path tests. 4702 // If this is a virtual call, we generate a funny guard. We grab 4703 // the vtable entry corresponding to clone() from the target object. 4704 // If the target method which we are calling happens to be the 4705 // Object clone() method, we pass the guard. We do not need this 4706 // guard for non-virtual calls; the caller is known to be the native 4707 // Object clone(). 4708 if (is_virtual) { 4709 generate_virtual_guard(obj_klass, slow_region); 4710 } 4711 4712 // The object must be cloneable and must not have a finalizer. 4713 // Both of these conditions may be checked in a single test. 4714 // We could optimize the cloneable test further, but we don't care. 4715 generate_access_flags_guard(obj_klass, 4716 // Test both conditions: 4717 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4718 // Must be cloneable but not finalizer: 4719 JVM_ACC_IS_CLONEABLE, 4720 slow_region); 4721 } 4722 4723 if (!stopped()) { 4724 // It's an instance, and it passed the slow-path tests. 4725 PreserveJVMState pjvms(this); 4726 Node* obj_size = NULL; 4727 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4728 // is reexecuted if deoptimization occurs and there could be problems when merging 4729 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4730 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4731 4732 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4733 4734 // Present the results of the slow call. 4735 result_reg->init_req(_instance_path, control()); 4736 result_val->init_req(_instance_path, alloc_obj); 4737 result_i_o ->set_req(_instance_path, i_o()); 4738 result_mem ->set_req(_instance_path, reset_memory()); 4739 } 4740 4741 // Generate code for the slow case. We make a call to clone(). 4742 set_control(_gvn.transform(slow_region)); 4743 if (!stopped()) { 4744 PreserveJVMState pjvms(this); 4745 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4746 Node* slow_result = set_results_for_java_call(slow_call); 4747 // this->control() comes from set_results_for_java_call 4748 result_reg->init_req(_slow_path, control()); 4749 result_val->init_req(_slow_path, slow_result); 4750 result_i_o ->set_req(_slow_path, i_o()); 4751 result_mem ->set_req(_slow_path, reset_memory()); 4752 } 4753 4754 // Return the combined state. 4755 set_control( _gvn.transform(result_reg)); 4756 set_i_o( _gvn.transform(result_i_o)); 4757 set_all_memory( _gvn.transform(result_mem)); 4758 } // original reexecute is set back here 4759 4760 set_result(_gvn.transform(result_val)); 4761 return true; 4762 } 4763 4764 // If we have a tighly coupled allocation, the arraycopy may take care 4765 // of the array initialization. If one of the guards we insert between 4766 // the allocation and the arraycopy causes a deoptimization, an 4767 // unitialized array will escape the compiled method. To prevent that 4768 // we set the JVM state for uncommon traps between the allocation and 4769 // the arraycopy to the state before the allocation so, in case of 4770 // deoptimization, we'll reexecute the allocation and the 4771 // initialization. 4772 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4773 if (alloc != NULL) { 4774 ciMethod* trap_method = alloc->jvms()->method(); 4775 int trap_bci = alloc->jvms()->bci(); 4776 4777 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4778 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4779 // Make sure there's no store between the allocation and the 4780 // arraycopy otherwise visible side effects could be rexecuted 4781 // in case of deoptimization and cause incorrect execution. 4782 bool no_interfering_store = true; 4783 Node* mem = alloc->in(TypeFunc::Memory); 4784 if (mem->is_MergeMem()) { 4785 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4786 Node* n = mms.memory(); 4787 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4788 assert(n->is_Store(), "what else?"); 4789 no_interfering_store = false; 4790 break; 4791 } 4792 } 4793 } else { 4794 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4795 Node* n = mms.memory(); 4796 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4797 assert(n->is_Store(), "what else?"); 4798 no_interfering_store = false; 4799 break; 4800 } 4801 } 4802 } 4803 4804 if (no_interfering_store) { 4805 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4806 uint size = alloc->req(); 4807 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4808 old_jvms->set_map(sfpt); 4809 for (uint i = 0; i < size; i++) { 4810 sfpt->init_req(i, alloc->in(i)); 4811 } 4812 // re-push array length for deoptimization 4813 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4814 old_jvms->set_sp(old_jvms->sp()+1); 4815 old_jvms->set_monoff(old_jvms->monoff()+1); 4816 old_jvms->set_scloff(old_jvms->scloff()+1); 4817 old_jvms->set_endoff(old_jvms->endoff()+1); 4818 old_jvms->set_should_reexecute(true); 4819 4820 sfpt->set_i_o(map()->i_o()); 4821 sfpt->set_memory(map()->memory()); 4822 sfpt->set_control(map()->control()); 4823 4824 JVMState* saved_jvms = jvms(); 4825 saved_reexecute_sp = _reexecute_sp; 4826 4827 set_jvms(sfpt->jvms()); 4828 _reexecute_sp = jvms()->sp(); 4829 4830 return saved_jvms; 4831 } 4832 } 4833 } 4834 return NULL; 4835 } 4836 4837 // In case of a deoptimization, we restart execution at the 4838 // allocation, allocating a new array. We would leave an uninitialized 4839 // array in the heap that GCs wouldn't expect. Move the allocation 4840 // after the traps so we don't allocate the array if we 4841 // deoptimize. This is possible because tightly_coupled_allocation() 4842 // guarantees there's no observer of the allocated array at this point 4843 // and the control flow is simple enough. 4844 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4845 if (saved_jvms != NULL && !stopped()) { 4846 assert(alloc != NULL, "only with a tightly coupled allocation"); 4847 // restore JVM state to the state at the arraycopy 4848 saved_jvms->map()->set_control(map()->control()); 4849 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4850 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4851 // If we've improved the types of some nodes (null check) while 4852 // emitting the guards, propagate them to the current state 4853 map()->replaced_nodes().apply(saved_jvms->map()); 4854 set_jvms(saved_jvms); 4855 _reexecute_sp = saved_reexecute_sp; 4856 4857 // Remove the allocation from above the guards 4858 CallProjections callprojs; 4859 alloc->extract_projections(&callprojs, true); 4860 InitializeNode* init = alloc->initialization(); 4861 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4862 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4863 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4864 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4865 4866 // move the allocation here (after the guards) 4867 _gvn.hash_delete(alloc); 4868 alloc->set_req(TypeFunc::Control, control()); 4869 alloc->set_req(TypeFunc::I_O, i_o()); 4870 Node *mem = reset_memory(); 4871 set_all_memory(mem); 4872 alloc->set_req(TypeFunc::Memory, mem); 4873 set_control(init->proj_out(TypeFunc::Control)); 4874 set_i_o(callprojs.fallthrough_ioproj); 4875 4876 // Update memory as done in GraphKit::set_output_for_allocation() 4877 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4878 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4879 if (ary_type->isa_aryptr() && length_type != NULL) { 4880 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4881 } 4882 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4883 int elemidx = C->get_alias_index(telemref); 4884 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4885 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4886 4887 Node* allocx = _gvn.transform(alloc); 4888 assert(allocx == alloc, "where has the allocation gone?"); 4889 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4890 4891 _gvn.hash_delete(dest); 4892 dest->set_req(0, control()); 4893 Node* destx = _gvn.transform(dest); 4894 assert(destx == dest, "where has the allocation result gone?"); 4895 } 4896 } 4897 4898 4899 //------------------------------inline_arraycopy----------------------- 4900 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4901 // Object dest, int destPos, 4902 // int length); 4903 bool LibraryCallKit::inline_arraycopy() { 4904 // Get the arguments. 4905 Node* src = argument(0); // type: oop 4906 Node* src_offset = argument(1); // type: int 4907 Node* dest = argument(2); // type: oop 4908 Node* dest_offset = argument(3); // type: int 4909 Node* length = argument(4); // type: int 4910 4911 4912 // Check for allocation before we add nodes that would confuse 4913 // tightly_coupled_allocation() 4914 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4915 4916 int saved_reexecute_sp = -1; 4917 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4918 // See arraycopy_restore_alloc_state() comment 4919 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4920 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4921 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4922 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4923 4924 // The following tests must be performed 4925 // (1) src and dest are arrays. 4926 // (2) src and dest arrays must have elements of the same BasicType 4927 // (3) src and dest must not be null. 4928 // (4) src_offset must not be negative. 4929 // (5) dest_offset must not be negative. 4930 // (6) length must not be negative. 4931 // (7) src_offset + length must not exceed length of src. 4932 // (8) dest_offset + length must not exceed length of dest. 4933 // (9) each element of an oop array must be assignable 4934 4935 // (3) src and dest must not be null. 4936 // always do this here because we need the JVM state for uncommon traps 4937 Node* null_ctl = top(); 4938 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4939 assert(null_ctl->is_top(), "no null control here"); 4940 dest = null_check(dest, T_ARRAY); 4941 4942 if (!can_emit_guards) { 4943 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4944 // guards but the arraycopy node could still take advantage of a 4945 // tightly allocated allocation. tightly_coupled_allocation() is 4946 // called again to make sure it takes the null check above into 4947 // account: the null check is mandatory and if it caused an 4948 // uncommon trap to be emitted then the allocation can't be 4949 // considered tightly coupled in this context. 4950 alloc = tightly_coupled_allocation(dest, NULL); 4951 } 4952 4953 bool validated = false; 4954 4955 const Type* src_type = _gvn.type(src); 4956 const Type* dest_type = _gvn.type(dest); 4957 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4958 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4959 4960 // Do we have the type of src? 4961 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4962 // Do we have the type of dest? 4963 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4964 // Is the type for src from speculation? 4965 bool src_spec = false; 4966 // Is the type for dest from speculation? 4967 bool dest_spec = false; 4968 4969 if ((!has_src || !has_dest) && can_emit_guards) { 4970 // We don't have sufficient type information, let's see if 4971 // speculative types can help. We need to have types for both src 4972 // and dest so that it pays off. 4973 4974 // Do we already have or could we have type information for src 4975 bool could_have_src = has_src; 4976 // Do we already have or could we have type information for dest 4977 bool could_have_dest = has_dest; 4978 4979 ciKlass* src_k = NULL; 4980 if (!has_src) { 4981 src_k = src_type->speculative_type_not_null(); 4982 if (src_k != NULL && src_k->is_array_klass()) { 4983 could_have_src = true; 4984 } 4985 } 4986 4987 ciKlass* dest_k = NULL; 4988 if (!has_dest) { 4989 dest_k = dest_type->speculative_type_not_null(); 4990 if (dest_k != NULL && dest_k->is_array_klass()) { 4991 could_have_dest = true; 4992 } 4993 } 4994 4995 if (could_have_src && could_have_dest) { 4996 // This is going to pay off so emit the required guards 4997 if (!has_src) { 4998 src = maybe_cast_profiled_obj(src, src_k, true); 4999 src_type = _gvn.type(src); 5000 top_src = src_type->isa_aryptr(); 5001 has_src = (top_src != NULL && top_src->klass() != NULL); 5002 src_spec = true; 5003 } 5004 if (!has_dest) { 5005 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5006 dest_type = _gvn.type(dest); 5007 top_dest = dest_type->isa_aryptr(); 5008 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5009 dest_spec = true; 5010 } 5011 } 5012 } 5013 5014 if (has_src && has_dest && can_emit_guards) { 5015 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 5016 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 5017 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 5018 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 5019 5020 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5021 // If both arrays are object arrays then having the exact types 5022 // for both will remove the need for a subtype check at runtime 5023 // before the call and may make it possible to pick a faster copy 5024 // routine (without a subtype check on every element) 5025 // Do we have the exact type of src? 5026 bool could_have_src = src_spec; 5027 // Do we have the exact type of dest? 5028 bool could_have_dest = dest_spec; 5029 ciKlass* src_k = top_src->klass(); 5030 ciKlass* dest_k = top_dest->klass(); 5031 if (!src_spec) { 5032 src_k = src_type->speculative_type_not_null(); 5033 if (src_k != NULL && src_k->is_array_klass()) { 5034 could_have_src = true; 5035 } 5036 } 5037 if (!dest_spec) { 5038 dest_k = dest_type->speculative_type_not_null(); 5039 if (dest_k != NULL && dest_k->is_array_klass()) { 5040 could_have_dest = true; 5041 } 5042 } 5043 if (could_have_src && could_have_dest) { 5044 // If we can have both exact types, emit the missing guards 5045 if (could_have_src && !src_spec) { 5046 src = maybe_cast_profiled_obj(src, src_k, true); 5047 } 5048 if (could_have_dest && !dest_spec) { 5049 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5050 } 5051 } 5052 } 5053 } 5054 5055 ciMethod* trap_method = method(); 5056 int trap_bci = bci(); 5057 if (saved_jvms != NULL) { 5058 trap_method = alloc->jvms()->method(); 5059 trap_bci = alloc->jvms()->bci(); 5060 } 5061 5062 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5063 can_emit_guards && 5064 !src->is_top() && !dest->is_top()) { 5065 // validate arguments: enables transformation the ArrayCopyNode 5066 validated = true; 5067 5068 RegionNode* slow_region = new RegionNode(1); 5069 record_for_igvn(slow_region); 5070 5071 // (1) src and dest are arrays. 5072 generate_non_array_guard(load_object_klass(src), slow_region); 5073 generate_non_array_guard(load_object_klass(dest), slow_region); 5074 5075 // (2) src and dest arrays must have elements of the same BasicType 5076 // done at macro expansion or at Ideal transformation time 5077 5078 // (4) src_offset must not be negative. 5079 generate_negative_guard(src_offset, slow_region); 5080 5081 // (5) dest_offset must not be negative. 5082 generate_negative_guard(dest_offset, slow_region); 5083 5084 // (7) src_offset + length must not exceed length of src. 5085 generate_limit_guard(src_offset, length, 5086 load_array_length(src), 5087 slow_region); 5088 5089 // (8) dest_offset + length must not exceed length of dest. 5090 generate_limit_guard(dest_offset, length, 5091 load_array_length(dest), 5092 slow_region); 5093 5094 // (9) each element of an oop array must be assignable 5095 Node* src_klass = load_object_klass(src); 5096 Node* dest_klass = load_object_klass(dest); 5097 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5098 5099 if (not_subtype_ctrl != top()) { 5100 PreserveJVMState pjvms(this); 5101 set_control(not_subtype_ctrl); 5102 uncommon_trap(Deoptimization::Reason_intrinsic, 5103 Deoptimization::Action_make_not_entrant); 5104 assert(stopped(), "Should be stopped"); 5105 } 5106 { 5107 PreserveJVMState pjvms(this); 5108 set_control(_gvn.transform(slow_region)); 5109 uncommon_trap(Deoptimization::Reason_intrinsic, 5110 Deoptimization::Action_make_not_entrant); 5111 assert(stopped(), "Should be stopped"); 5112 } 5113 } 5114 5115 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5116 5117 if (stopped()) { 5118 return true; 5119 } 5120 5121 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5122 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5123 // so the compiler has a chance to eliminate them: during macro expansion, 5124 // we have to set their control (CastPP nodes are eliminated). 5125 load_object_klass(src), load_object_klass(dest), 5126 load_array_length(src), load_array_length(dest)); 5127 5128 ac->set_arraycopy(validated); 5129 5130 Node* n = _gvn.transform(ac); 5131 if (n == ac) { 5132 ac->connect_outputs(this); 5133 } else { 5134 assert(validated, "shouldn't transform if all arguments not validated"); 5135 set_all_memory(n); 5136 } 5137 5138 return true; 5139 } 5140 5141 5142 // Helper function which determines if an arraycopy immediately follows 5143 // an allocation, with no intervening tests or other escapes for the object. 5144 AllocateArrayNode* 5145 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5146 RegionNode* slow_region) { 5147 if (stopped()) return NULL; // no fast path 5148 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5149 5150 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5151 if (alloc == NULL) return NULL; 5152 5153 Node* rawmem = memory(Compile::AliasIdxRaw); 5154 // Is the allocation's memory state untouched? 5155 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5156 // Bail out if there have been raw-memory effects since the allocation. 5157 // (Example: There might have been a call or safepoint.) 5158 return NULL; 5159 } 5160 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5161 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5162 return NULL; 5163 } 5164 5165 // There must be no unexpected observers of this allocation. 5166 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5167 Node* obs = ptr->fast_out(i); 5168 if (obs != this->map()) { 5169 return NULL; 5170 } 5171 } 5172 5173 // This arraycopy must unconditionally follow the allocation of the ptr. 5174 Node* alloc_ctl = ptr->in(0); 5175 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5176 5177 Node* ctl = control(); 5178 while (ctl != alloc_ctl) { 5179 // There may be guards which feed into the slow_region. 5180 // Any other control flow means that we might not get a chance 5181 // to finish initializing the allocated object. 5182 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5183 IfNode* iff = ctl->in(0)->as_If(); 5184 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5185 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5186 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5187 ctl = iff->in(0); // This test feeds the known slow_region. 5188 continue; 5189 } 5190 // One more try: Various low-level checks bottom out in 5191 // uncommon traps. If the debug-info of the trap omits 5192 // any reference to the allocation, as we've already 5193 // observed, then there can be no objection to the trap. 5194 bool found_trap = false; 5195 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5196 Node* obs = not_ctl->fast_out(j); 5197 if (obs->in(0) == not_ctl && obs->is_Call() && 5198 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5199 found_trap = true; break; 5200 } 5201 } 5202 if (found_trap) { 5203 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5204 continue; 5205 } 5206 } 5207 return NULL; 5208 } 5209 5210 // If we get this far, we have an allocation which immediately 5211 // precedes the arraycopy, and we can take over zeroing the new object. 5212 // The arraycopy will finish the initialization, and provide 5213 // a new control state to which we will anchor the destination pointer. 5214 5215 return alloc; 5216 } 5217 5218 //-------------inline_encodeISOArray----------------------------------- 5219 // encode char[] to byte[] in ISO_8859_1 5220 bool LibraryCallKit::inline_encodeISOArray() { 5221 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5222 // no receiver since it is static method 5223 Node *src = argument(0); 5224 Node *src_offset = argument(1); 5225 Node *dst = argument(2); 5226 Node *dst_offset = argument(3); 5227 Node *length = argument(4); 5228 5229 const Type* src_type = src->Value(&_gvn); 5230 const Type* dst_type = dst->Value(&_gvn); 5231 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5232 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5233 if (top_src == NULL || top_src->klass() == NULL || 5234 top_dest == NULL || top_dest->klass() == NULL) { 5235 // failed array check 5236 return false; 5237 } 5238 5239 // Figure out the size and type of the elements we will be copying. 5240 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5241 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5242 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5243 return false; 5244 } 5245 Node* src_start = array_element_address(src, src_offset, src_elem); 5246 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5247 // 'src_start' points to src array + scaled offset 5248 // 'dst_start' points to dst array + scaled offset 5249 5250 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5251 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5252 enc = _gvn.transform(enc); 5253 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5254 set_memory(res_mem, mtype); 5255 set_result(enc); 5256 return true; 5257 } 5258 5259 //-------------inline_multiplyToLen----------------------------------- 5260 bool LibraryCallKit::inline_multiplyToLen() { 5261 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5262 5263 address stubAddr = StubRoutines::multiplyToLen(); 5264 if (stubAddr == NULL) { 5265 return false; // Intrinsic's stub is not implemented on this platform 5266 } 5267 const char* stubName = "multiplyToLen"; 5268 5269 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5270 5271 // no receiver because it is a static method 5272 Node* x = argument(0); 5273 Node* xlen = argument(1); 5274 Node* y = argument(2); 5275 Node* ylen = argument(3); 5276 Node* z = argument(4); 5277 5278 const Type* x_type = x->Value(&_gvn); 5279 const Type* y_type = y->Value(&_gvn); 5280 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5281 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5282 if (top_x == NULL || top_x->klass() == NULL || 5283 top_y == NULL || top_y->klass() == NULL) { 5284 // failed array check 5285 return false; 5286 } 5287 5288 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5289 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5290 if (x_elem != T_INT || y_elem != T_INT) { 5291 return false; 5292 } 5293 5294 // Set the original stack and the reexecute bit for the interpreter to reexecute 5295 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5296 // on the return from z array allocation in runtime. 5297 { PreserveReexecuteState preexecs(this); 5298 jvms()->set_should_reexecute(true); 5299 5300 Node* x_start = array_element_address(x, intcon(0), x_elem); 5301 Node* y_start = array_element_address(y, intcon(0), y_elem); 5302 // 'x_start' points to x array + scaled xlen 5303 // 'y_start' points to y array + scaled ylen 5304 5305 // Allocate the result array 5306 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5307 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5308 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5309 5310 IdealKit ideal(this); 5311 5312 #define __ ideal. 5313 Node* one = __ ConI(1); 5314 Node* zero = __ ConI(0); 5315 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5316 __ set(need_alloc, zero); 5317 __ set(z_alloc, z); 5318 __ if_then(z, BoolTest::eq, null()); { 5319 __ increment (need_alloc, one); 5320 } __ else_(); { 5321 // Update graphKit memory and control from IdealKit. 5322 sync_kit(ideal); 5323 Node* zlen_arg = load_array_length(z); 5324 // Update IdealKit memory and control from graphKit. 5325 __ sync_kit(this); 5326 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5327 __ increment (need_alloc, one); 5328 } __ end_if(); 5329 } __ end_if(); 5330 5331 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5332 // Update graphKit memory and control from IdealKit. 5333 sync_kit(ideal); 5334 Node * narr = new_array(klass_node, zlen, 1); 5335 // Update IdealKit memory and control from graphKit. 5336 __ sync_kit(this); 5337 __ set(z_alloc, narr); 5338 } __ end_if(); 5339 5340 sync_kit(ideal); 5341 z = __ value(z_alloc); 5342 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5343 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5344 // Final sync IdealKit and GraphKit. 5345 final_sync(ideal); 5346 #undef __ 5347 5348 Node* z_start = array_element_address(z, intcon(0), T_INT); 5349 5350 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5351 OptoRuntime::multiplyToLen_Type(), 5352 stubAddr, stubName, TypePtr::BOTTOM, 5353 x_start, xlen, y_start, ylen, z_start, zlen); 5354 } // original reexecute is set back here 5355 5356 C->set_has_split_ifs(true); // Has chance for split-if optimization 5357 set_result(z); 5358 return true; 5359 } 5360 5361 //-------------inline_squareToLen------------------------------------ 5362 bool LibraryCallKit::inline_squareToLen() { 5363 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5364 5365 address stubAddr = StubRoutines::squareToLen(); 5366 if (stubAddr == NULL) { 5367 return false; // Intrinsic's stub is not implemented on this platform 5368 } 5369 const char* stubName = "squareToLen"; 5370 5371 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5372 5373 Node* x = argument(0); 5374 Node* len = argument(1); 5375 Node* z = argument(2); 5376 Node* zlen = argument(3); 5377 5378 const Type* x_type = x->Value(&_gvn); 5379 const Type* z_type = z->Value(&_gvn); 5380 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5381 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5382 if (top_x == NULL || top_x->klass() == NULL || 5383 top_z == NULL || top_z->klass() == NULL) { 5384 // failed array check 5385 return false; 5386 } 5387 5388 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5389 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5390 if (x_elem != T_INT || z_elem != T_INT) { 5391 return false; 5392 } 5393 5394 5395 Node* x_start = array_element_address(x, intcon(0), x_elem); 5396 Node* z_start = array_element_address(z, intcon(0), z_elem); 5397 5398 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5399 OptoRuntime::squareToLen_Type(), 5400 stubAddr, stubName, TypePtr::BOTTOM, 5401 x_start, len, z_start, zlen); 5402 5403 set_result(z); 5404 return true; 5405 } 5406 5407 //-------------inline_mulAdd------------------------------------------ 5408 bool LibraryCallKit::inline_mulAdd() { 5409 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5410 5411 address stubAddr = StubRoutines::mulAdd(); 5412 if (stubAddr == NULL) { 5413 return false; // Intrinsic's stub is not implemented on this platform 5414 } 5415 const char* stubName = "mulAdd"; 5416 5417 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5418 5419 Node* out = argument(0); 5420 Node* in = argument(1); 5421 Node* offset = argument(2); 5422 Node* len = argument(3); 5423 Node* k = argument(4); 5424 5425 const Type* out_type = out->Value(&_gvn); 5426 const Type* in_type = in->Value(&_gvn); 5427 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5428 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5429 if (top_out == NULL || top_out->klass() == NULL || 5430 top_in == NULL || top_in->klass() == NULL) { 5431 // failed array check 5432 return false; 5433 } 5434 5435 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5436 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5437 if (out_elem != T_INT || in_elem != T_INT) { 5438 return false; 5439 } 5440 5441 Node* outlen = load_array_length(out); 5442 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5443 Node* out_start = array_element_address(out, intcon(0), out_elem); 5444 Node* in_start = array_element_address(in, intcon(0), in_elem); 5445 5446 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5447 OptoRuntime::mulAdd_Type(), 5448 stubAddr, stubName, TypePtr::BOTTOM, 5449 out_start,in_start, new_offset, len, k); 5450 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5451 set_result(result); 5452 return true; 5453 } 5454 5455 //-------------inline_montgomeryMultiply----------------------------------- 5456 bool LibraryCallKit::inline_montgomeryMultiply() { 5457 address stubAddr = StubRoutines::montgomeryMultiply(); 5458 if (stubAddr == NULL) { 5459 return false; // Intrinsic's stub is not implemented on this platform 5460 } 5461 5462 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform"); 5463 const char* stubName = "montgomery_square"; 5464 5465 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters"); 5466 5467 Node* a = argument(0); 5468 Node* b = argument(1); 5469 Node* n = argument(2); 5470 Node* len = argument(3); 5471 Node* inv = argument(4); 5472 Node* m = argument(6); 5473 5474 const Type* a_type = a->Value(&_gvn); 5475 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5476 const Type* b_type = b->Value(&_gvn); 5477 const TypeAryPtr* top_b = b_type->isa_aryptr(); 5478 const Type* n_type = a->Value(&_gvn); 5479 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5480 const Type* m_type = a->Value(&_gvn); 5481 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5482 if (top_a == NULL || top_a->klass() == NULL || 5483 top_b == NULL || top_b->klass() == NULL || 5484 top_n == NULL || top_n->klass() == NULL || 5485 top_m == NULL || top_m->klass() == NULL) { 5486 // failed array check 5487 return false; 5488 } 5489 5490 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5491 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5492 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5493 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5494 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5495 return false; 5496 } 5497 5498 // Make the call 5499 { 5500 Node* a_start = array_element_address(a, intcon(0), a_elem); 5501 Node* b_start = array_element_address(b, intcon(0), b_elem); 5502 Node* n_start = array_element_address(n, intcon(0), n_elem); 5503 Node* m_start = array_element_address(m, intcon(0), m_elem); 5504 5505 Node* call = make_runtime_call(RC_LEAF, 5506 OptoRuntime::montgomeryMultiply_Type(), 5507 stubAddr, stubName, TypePtr::BOTTOM, 5508 a_start, b_start, n_start, len, inv, top(), 5509 m_start); 5510 set_result(m); 5511 } 5512 5513 return true; 5514 } 5515 5516 bool LibraryCallKit::inline_montgomerySquare() { 5517 address stubAddr = StubRoutines::montgomerySquare(); 5518 if (stubAddr == NULL) { 5519 return false; // Intrinsic's stub is not implemented on this platform 5520 } 5521 5522 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform"); 5523 const char* stubName = "montgomery_square"; 5524 5525 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters"); 5526 5527 Node* a = argument(0); 5528 Node* n = argument(1); 5529 Node* len = argument(2); 5530 Node* inv = argument(3); 5531 Node* m = argument(5); 5532 5533 const Type* a_type = a->Value(&_gvn); 5534 const TypeAryPtr* top_a = a_type->isa_aryptr(); 5535 const Type* n_type = a->Value(&_gvn); 5536 const TypeAryPtr* top_n = n_type->isa_aryptr(); 5537 const Type* m_type = a->Value(&_gvn); 5538 const TypeAryPtr* top_m = m_type->isa_aryptr(); 5539 if (top_a == NULL || top_a->klass() == NULL || 5540 top_n == NULL || top_n->klass() == NULL || 5541 top_m == NULL || top_m->klass() == NULL) { 5542 // failed array check 5543 return false; 5544 } 5545 5546 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5547 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5548 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5549 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) { 5550 return false; 5551 } 5552 5553 // Make the call 5554 { 5555 Node* a_start = array_element_address(a, intcon(0), a_elem); 5556 Node* n_start = array_element_address(n, intcon(0), n_elem); 5557 Node* m_start = array_element_address(m, intcon(0), m_elem); 5558 5559 Node* call = make_runtime_call(RC_LEAF, 5560 OptoRuntime::montgomerySquare_Type(), 5561 stubAddr, stubName, TypePtr::BOTTOM, 5562 a_start, n_start, len, inv, top(), 5563 m_start); 5564 set_result(m); 5565 } 5566 5567 return true; 5568 } 5569 5570 5571 /** 5572 * Calculate CRC32 for byte. 5573 * int java.util.zip.CRC32.update(int crc, int b) 5574 */ 5575 bool LibraryCallKit::inline_updateCRC32() { 5576 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5577 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5578 // no receiver since it is static method 5579 Node* crc = argument(0); // type: int 5580 Node* b = argument(1); // type: int 5581 5582 /* 5583 * int c = ~ crc; 5584 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5585 * b = b ^ (c >>> 8); 5586 * crc = ~b; 5587 */ 5588 5589 Node* M1 = intcon(-1); 5590 crc = _gvn.transform(new XorINode(crc, M1)); 5591 Node* result = _gvn.transform(new XorINode(crc, b)); 5592 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5593 5594 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5595 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5596 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5597 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5598 5599 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5600 result = _gvn.transform(new XorINode(crc, result)); 5601 result = _gvn.transform(new XorINode(result, M1)); 5602 set_result(result); 5603 return true; 5604 } 5605 5606 /** 5607 * Calculate CRC32 for byte[] array. 5608 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5609 */ 5610 bool LibraryCallKit::inline_updateBytesCRC32() { 5611 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5612 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5613 // no receiver since it is static method 5614 Node* crc = argument(0); // type: int 5615 Node* src = argument(1); // type: oop 5616 Node* offset = argument(2); // type: int 5617 Node* length = argument(3); // type: int 5618 5619 const Type* src_type = src->Value(&_gvn); 5620 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5621 if (top_src == NULL || top_src->klass() == NULL) { 5622 // failed array check 5623 return false; 5624 } 5625 5626 // Figure out the size and type of the elements we will be copying. 5627 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5628 if (src_elem != T_BYTE) { 5629 return false; 5630 } 5631 5632 // 'src_start' points to src array + scaled offset 5633 Node* src_start = array_element_address(src, offset, src_elem); 5634 5635 // We assume that range check is done by caller. 5636 // TODO: generate range check (offset+length < src.length) in debug VM. 5637 5638 // Call the stub. 5639 address stubAddr = StubRoutines::updateBytesCRC32(); 5640 const char *stubName = "updateBytesCRC32"; 5641 5642 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5643 stubAddr, stubName, TypePtr::BOTTOM, 5644 crc, src_start, length); 5645 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5646 set_result(result); 5647 return true; 5648 } 5649 5650 /** 5651 * Calculate CRC32 for ByteBuffer. 5652 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5653 */ 5654 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5655 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5656 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5657 // no receiver since it is static method 5658 Node* crc = argument(0); // type: int 5659 Node* src = argument(1); // type: long 5660 Node* offset = argument(3); // type: int 5661 Node* length = argument(4); // type: int 5662 5663 src = ConvL2X(src); // adjust Java long to machine word 5664 Node* base = _gvn.transform(new CastX2PNode(src)); 5665 offset = ConvI2X(offset); 5666 5667 // 'src_start' points to src array + scaled offset 5668 Node* src_start = basic_plus_adr(top(), base, offset); 5669 5670 // Call the stub. 5671 address stubAddr = StubRoutines::updateBytesCRC32(); 5672 const char *stubName = "updateBytesCRC32"; 5673 5674 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5675 stubAddr, stubName, TypePtr::BOTTOM, 5676 crc, src_start, length); 5677 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5678 set_result(result); 5679 return true; 5680 } 5681 5682 //------------------------------get_table_from_crc32c_class----------------------- 5683 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5684 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5685 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5686 5687 return table; 5688 } 5689 5690 //------------------------------inline_updateBytesCRC32C----------------------- 5691 // 5692 // Calculate CRC32C for byte[] array. 5693 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5694 // 5695 bool LibraryCallKit::inline_updateBytesCRC32C() { 5696 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5697 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5698 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5699 // no receiver since it is a static method 5700 Node* crc = argument(0); // type: int 5701 Node* src = argument(1); // type: oop 5702 Node* offset = argument(2); // type: int 5703 Node* end = argument(3); // type: int 5704 5705 Node* length = _gvn.transform(new SubINode(end, offset)); 5706 5707 const Type* src_type = src->Value(&_gvn); 5708 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5709 if (top_src == NULL || top_src->klass() == NULL) { 5710 // failed array check 5711 return false; 5712 } 5713 5714 // Figure out the size and type of the elements we will be copying. 5715 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5716 if (src_elem != T_BYTE) { 5717 return false; 5718 } 5719 5720 // 'src_start' points to src array + scaled offset 5721 Node* src_start = array_element_address(src, offset, src_elem); 5722 5723 // static final int[] byteTable in class CRC32C 5724 Node* table = get_table_from_crc32c_class(callee()->holder()); 5725 Node* table_start = array_element_address(table, intcon(0), T_INT); 5726 5727 // We assume that range check is done by caller. 5728 // TODO: generate range check (offset+length < src.length) in debug VM. 5729 5730 // Call the stub. 5731 address stubAddr = StubRoutines::updateBytesCRC32C(); 5732 const char *stubName = "updateBytesCRC32C"; 5733 5734 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5735 stubAddr, stubName, TypePtr::BOTTOM, 5736 crc, src_start, length, table_start); 5737 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5738 set_result(result); 5739 return true; 5740 } 5741 5742 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5743 // 5744 // Calculate CRC32C for DirectByteBuffer. 5745 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5746 // 5747 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5748 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5749 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5750 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5751 // no receiver since it is a static method 5752 Node* crc = argument(0); // type: int 5753 Node* src = argument(1); // type: long 5754 Node* offset = argument(3); // type: int 5755 Node* end = argument(4); // type: int 5756 5757 Node* length = _gvn.transform(new SubINode(end, offset)); 5758 5759 src = ConvL2X(src); // adjust Java long to machine word 5760 Node* base = _gvn.transform(new CastX2PNode(src)); 5761 offset = ConvI2X(offset); 5762 5763 // 'src_start' points to src array + scaled offset 5764 Node* src_start = basic_plus_adr(top(), base, offset); 5765 5766 // static final int[] byteTable in class CRC32C 5767 Node* table = get_table_from_crc32c_class(callee()->holder()); 5768 Node* table_start = array_element_address(table, intcon(0), T_INT); 5769 5770 // Call the stub. 5771 address stubAddr = StubRoutines::updateBytesCRC32C(); 5772 const char *stubName = "updateBytesCRC32C"; 5773 5774 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5775 stubAddr, stubName, TypePtr::BOTTOM, 5776 crc, src_start, length, table_start); 5777 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5778 set_result(result); 5779 return true; 5780 } 5781 5782 //----------------------------inline_reference_get---------------------------- 5783 // public T java.lang.ref.Reference.get(); 5784 bool LibraryCallKit::inline_reference_get() { 5785 const int referent_offset = java_lang_ref_Reference::referent_offset; 5786 guarantee(referent_offset > 0, "should have already been set"); 5787 5788 // Get the argument: 5789 Node* reference_obj = null_check_receiver(); 5790 if (stopped()) return true; 5791 5792 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5793 5794 ciInstanceKlass* klass = env()->Object_klass(); 5795 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5796 5797 Node* no_ctrl = NULL; 5798 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5799 5800 // Use the pre-barrier to record the value in the referent field 5801 pre_barrier(false /* do_load */, 5802 control(), 5803 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5804 result /* pre_val */, 5805 T_OBJECT); 5806 5807 // Add memory barrier to prevent commoning reads from this field 5808 // across safepoint since GC can change its value. 5809 insert_mem_bar(Op_MemBarCPUOrder); 5810 5811 set_result(result); 5812 return true; 5813 } 5814 5815 5816 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5817 bool is_exact=true, bool is_static=false, 5818 ciInstanceKlass * fromKls=NULL) { 5819 if (fromKls == NULL) { 5820 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5821 assert(tinst != NULL, "obj is null"); 5822 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5823 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5824 fromKls = tinst->klass()->as_instance_klass(); 5825 } else { 5826 assert(is_static, "only for static field access"); 5827 } 5828 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5829 ciSymbol::make(fieldTypeString), 5830 is_static); 5831 5832 assert (field != NULL, "undefined field"); 5833 if (field == NULL) return (Node *) NULL; 5834 5835 if (is_static) { 5836 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5837 fromObj = makecon(tip); 5838 } 5839 5840 // Next code copied from Parse::do_get_xxx(): 5841 5842 // Compute address and memory type. 5843 int offset = field->offset_in_bytes(); 5844 bool is_vol = field->is_volatile(); 5845 ciType* field_klass = field->type(); 5846 assert(field_klass->is_loaded(), "should be loaded"); 5847 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5848 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5849 BasicType bt = field->layout_type(); 5850 5851 // Build the resultant type of the load 5852 const Type *type; 5853 if (bt == T_OBJECT) { 5854 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5855 } else { 5856 type = Type::get_const_basic_type(bt); 5857 } 5858 5859 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5860 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5861 } 5862 // Build the load. 5863 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5864 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5865 // If reference is volatile, prevent following memory ops from 5866 // floating up past the volatile read. Also prevents commoning 5867 // another volatile read. 5868 if (is_vol) { 5869 // Memory barrier includes bogus read of value to force load BEFORE membar 5870 insert_mem_bar(Op_MemBarAcquire, loadedField); 5871 } 5872 return loadedField; 5873 } 5874 5875 5876 //------------------------------inline_aescrypt_Block----------------------- 5877 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5878 address stubAddr; 5879 const char *stubName; 5880 assert(UseAES, "need AES instruction support"); 5881 5882 switch(id) { 5883 case vmIntrinsics::_aescrypt_encryptBlock: 5884 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5885 stubName = "aescrypt_encryptBlock"; 5886 break; 5887 case vmIntrinsics::_aescrypt_decryptBlock: 5888 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5889 stubName = "aescrypt_decryptBlock"; 5890 break; 5891 } 5892 if (stubAddr == NULL) return false; 5893 5894 Node* aescrypt_object = argument(0); 5895 Node* src = argument(1); 5896 Node* src_offset = argument(2); 5897 Node* dest = argument(3); 5898 Node* dest_offset = argument(4); 5899 5900 // (1) src and dest are arrays. 5901 const Type* src_type = src->Value(&_gvn); 5902 const Type* dest_type = dest->Value(&_gvn); 5903 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5904 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5905 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5906 5907 // for the quick and dirty code we will skip all the checks. 5908 // we are just trying to get the call to be generated. 5909 Node* src_start = src; 5910 Node* dest_start = dest; 5911 if (src_offset != NULL || dest_offset != NULL) { 5912 assert(src_offset != NULL && dest_offset != NULL, ""); 5913 src_start = array_element_address(src, src_offset, T_BYTE); 5914 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5915 } 5916 5917 // now need to get the start of its expanded key array 5918 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5919 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5920 if (k_start == NULL) return false; 5921 5922 if (Matcher::pass_original_key_for_aes()) { 5923 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5924 // compatibility issues between Java key expansion and SPARC crypto instructions 5925 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5926 if (original_k_start == NULL) return false; 5927 5928 // Call the stub. 5929 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5930 stubAddr, stubName, TypePtr::BOTTOM, 5931 src_start, dest_start, k_start, original_k_start); 5932 } else { 5933 // Call the stub. 5934 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5935 stubAddr, stubName, TypePtr::BOTTOM, 5936 src_start, dest_start, k_start); 5937 } 5938 5939 return true; 5940 } 5941 5942 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5943 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5944 address stubAddr; 5945 const char *stubName; 5946 5947 assert(UseAES, "need AES instruction support"); 5948 5949 switch(id) { 5950 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5951 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5952 stubName = "cipherBlockChaining_encryptAESCrypt"; 5953 break; 5954 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5955 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5956 stubName = "cipherBlockChaining_decryptAESCrypt"; 5957 break; 5958 } 5959 if (stubAddr == NULL) return false; 5960 5961 Node* cipherBlockChaining_object = argument(0); 5962 Node* src = argument(1); 5963 Node* src_offset = argument(2); 5964 Node* len = argument(3); 5965 Node* dest = argument(4); 5966 Node* dest_offset = argument(5); 5967 5968 // (1) src and dest are arrays. 5969 const Type* src_type = src->Value(&_gvn); 5970 const Type* dest_type = dest->Value(&_gvn); 5971 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5972 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5973 assert (top_src != NULL && top_src->klass() != NULL 5974 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5975 5976 // checks are the responsibility of the caller 5977 Node* src_start = src; 5978 Node* dest_start = dest; 5979 if (src_offset != NULL || dest_offset != NULL) { 5980 assert(src_offset != NULL && dest_offset != NULL, ""); 5981 src_start = array_element_address(src, src_offset, T_BYTE); 5982 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5983 } 5984 5985 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5986 // (because of the predicated logic executed earlier). 5987 // so we cast it here safely. 5988 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5989 5990 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5991 if (embeddedCipherObj == NULL) return false; 5992 5993 // cast it to what we know it will be at runtime 5994 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5995 assert(tinst != NULL, "CBC obj is null"); 5996 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5997 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5998 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5999 6000 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6001 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 6002 const TypeOopPtr* xtype = aklass->as_instance_type(); 6003 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 6004 aescrypt_object = _gvn.transform(aescrypt_object); 6005 6006 // we need to get the start of the aescrypt_object's expanded key array 6007 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 6008 if (k_start == NULL) return false; 6009 6010 // similarly, get the start address of the r vector 6011 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 6012 if (objRvec == NULL) return false; 6013 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 6014 6015 Node* cbcCrypt; 6016 if (Matcher::pass_original_key_for_aes()) { 6017 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 6018 // compatibility issues between Java key expansion and SPARC crypto instructions 6019 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 6020 if (original_k_start == NULL) return false; 6021 6022 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 6023 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6024 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6025 stubAddr, stubName, TypePtr::BOTTOM, 6026 src_start, dest_start, k_start, r_start, len, original_k_start); 6027 } else { 6028 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 6029 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 6030 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 6031 stubAddr, stubName, TypePtr::BOTTOM, 6032 src_start, dest_start, k_start, r_start, len); 6033 } 6034 6035 // return cipher length (int) 6036 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 6037 set_result(retvalue); 6038 return true; 6039 } 6040 6041 //------------------------------get_key_start_from_aescrypt_object----------------------- 6042 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 6043 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 6044 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6045 if (objAESCryptKey == NULL) return (Node *) NULL; 6046 6047 // now have the array, need to get the start address of the K array 6048 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 6049 return k_start; 6050 } 6051 6052 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 6053 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 6054 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 6055 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 6056 if (objAESCryptKey == NULL) return (Node *) NULL; 6057 6058 // now have the array, need to get the start address of the lastKey array 6059 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 6060 return original_k_start; 6061 } 6062 6063 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 6064 // Return node representing slow path of predicate check. 6065 // the pseudo code we want to emulate with this predicate is: 6066 // for encryption: 6067 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 6068 // for decryption: 6069 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 6070 // note cipher==plain is more conservative than the original java code but that's OK 6071 // 6072 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 6073 // The receiver was checked for NULL already. 6074 Node* objCBC = argument(0); 6075 6076 // Load embeddedCipher field of CipherBlockChaining object. 6077 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 6078 6079 // get AESCrypt klass for instanceOf check 6080 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 6081 // will have same classloader as CipherBlockChaining object 6082 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 6083 assert(tinst != NULL, "CBCobj is null"); 6084 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 6085 6086 // we want to do an instanceof comparison against the AESCrypt class 6087 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 6088 if (!klass_AESCrypt->is_loaded()) { 6089 // if AESCrypt is not even loaded, we never take the intrinsic fast path 6090 Node* ctrl = control(); 6091 set_control(top()); // no regular fast path 6092 return ctrl; 6093 } 6094 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 6095 6096 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 6097 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 6098 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6099 6100 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6101 6102 // for encryption, we are done 6103 if (!decrypting) 6104 return instof_false; // even if it is NULL 6105 6106 // for decryption, we need to add a further check to avoid 6107 // taking the intrinsic path when cipher and plain are the same 6108 // see the original java code for why. 6109 RegionNode* region = new RegionNode(3); 6110 region->init_req(1, instof_false); 6111 Node* src = argument(1); 6112 Node* dest = argument(4); 6113 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6114 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6115 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6116 region->init_req(2, src_dest_conjoint); 6117 6118 record_for_igvn(region); 6119 return _gvn.transform(region); 6120 } 6121 6122 //------------------------------inline_ghash_processBlocks 6123 bool LibraryCallKit::inline_ghash_processBlocks() { 6124 address stubAddr; 6125 const char *stubName; 6126 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6127 6128 stubAddr = StubRoutines::ghash_processBlocks(); 6129 stubName = "ghash_processBlocks"; 6130 6131 Node* data = argument(0); 6132 Node* offset = argument(1); 6133 Node* len = argument(2); 6134 Node* state = argument(3); 6135 Node* subkeyH = argument(4); 6136 6137 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6138 assert(state_start, "state is NULL"); 6139 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6140 assert(subkeyH_start, "subkeyH is NULL"); 6141 Node* data_start = array_element_address(data, offset, T_BYTE); 6142 assert(data_start, "data is NULL"); 6143 6144 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6145 OptoRuntime::ghash_processBlocks_Type(), 6146 stubAddr, stubName, TypePtr::BOTTOM, 6147 state_start, subkeyH_start, data_start, len); 6148 return true; 6149 } 6150 6151 //------------------------------inline_sha_implCompress----------------------- 6152 // 6153 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6154 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6155 // 6156 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6157 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6158 // 6159 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6160 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6161 // 6162 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6163 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6164 6165 Node* sha_obj = argument(0); 6166 Node* src = argument(1); // type oop 6167 Node* ofs = argument(2); // type int 6168 6169 const Type* src_type = src->Value(&_gvn); 6170 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6171 if (top_src == NULL || top_src->klass() == NULL) { 6172 // failed array check 6173 return false; 6174 } 6175 // Figure out the size and type of the elements we will be copying. 6176 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6177 if (src_elem != T_BYTE) { 6178 return false; 6179 } 6180 // 'src_start' points to src array + offset 6181 Node* src_start = array_element_address(src, ofs, src_elem); 6182 Node* state = NULL; 6183 address stubAddr; 6184 const char *stubName; 6185 6186 switch(id) { 6187 case vmIntrinsics::_sha_implCompress: 6188 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6189 state = get_state_from_sha_object(sha_obj); 6190 stubAddr = StubRoutines::sha1_implCompress(); 6191 stubName = "sha1_implCompress"; 6192 break; 6193 case vmIntrinsics::_sha2_implCompress: 6194 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6195 state = get_state_from_sha_object(sha_obj); 6196 stubAddr = StubRoutines::sha256_implCompress(); 6197 stubName = "sha256_implCompress"; 6198 break; 6199 case vmIntrinsics::_sha5_implCompress: 6200 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6201 state = get_state_from_sha5_object(sha_obj); 6202 stubAddr = StubRoutines::sha512_implCompress(); 6203 stubName = "sha512_implCompress"; 6204 break; 6205 default: 6206 fatal_unexpected_iid(id); 6207 return false; 6208 } 6209 if (state == NULL) return false; 6210 6211 // Call the stub. 6212 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6213 stubAddr, stubName, TypePtr::BOTTOM, 6214 src_start, state); 6215 6216 return true; 6217 } 6218 6219 //------------------------------inline_digestBase_implCompressMB----------------------- 6220 // 6221 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6222 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6223 // 6224 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6225 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6226 "need SHA1/SHA256/SHA512 instruction support"); 6227 assert((uint)predicate < 3, "sanity"); 6228 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6229 6230 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6231 Node* src = argument(1); // byte[] array 6232 Node* ofs = argument(2); // type int 6233 Node* limit = argument(3); // type int 6234 6235 const Type* src_type = src->Value(&_gvn); 6236 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6237 if (top_src == NULL || top_src->klass() == NULL) { 6238 // failed array check 6239 return false; 6240 } 6241 // Figure out the size and type of the elements we will be copying. 6242 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6243 if (src_elem != T_BYTE) { 6244 return false; 6245 } 6246 // 'src_start' points to src array + offset 6247 Node* src_start = array_element_address(src, ofs, src_elem); 6248 6249 const char* klass_SHA_name = NULL; 6250 const char* stub_name = NULL; 6251 address stub_addr = NULL; 6252 bool long_state = false; 6253 6254 switch (predicate) { 6255 case 0: 6256 if (UseSHA1Intrinsics) { 6257 klass_SHA_name = "sun/security/provider/SHA"; 6258 stub_name = "sha1_implCompressMB"; 6259 stub_addr = StubRoutines::sha1_implCompressMB(); 6260 } 6261 break; 6262 case 1: 6263 if (UseSHA256Intrinsics) { 6264 klass_SHA_name = "sun/security/provider/SHA2"; 6265 stub_name = "sha256_implCompressMB"; 6266 stub_addr = StubRoutines::sha256_implCompressMB(); 6267 } 6268 break; 6269 case 2: 6270 if (UseSHA512Intrinsics) { 6271 klass_SHA_name = "sun/security/provider/SHA5"; 6272 stub_name = "sha512_implCompressMB"; 6273 stub_addr = StubRoutines::sha512_implCompressMB(); 6274 long_state = true; 6275 } 6276 break; 6277 default: 6278 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6279 } 6280 if (klass_SHA_name != NULL) { 6281 // get DigestBase klass to lookup for SHA klass 6282 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6283 assert(tinst != NULL, "digestBase_obj is not instance???"); 6284 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6285 6286 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6287 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6288 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6289 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6290 } 6291 return false; 6292 } 6293 //------------------------------inline_sha_implCompressMB----------------------- 6294 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6295 bool long_state, address stubAddr, const char *stubName, 6296 Node* src_start, Node* ofs, Node* limit) { 6297 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6298 const TypeOopPtr* xtype = aklass->as_instance_type(); 6299 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6300 sha_obj = _gvn.transform(sha_obj); 6301 6302 Node* state; 6303 if (long_state) { 6304 state = get_state_from_sha5_object(sha_obj); 6305 } else { 6306 state = get_state_from_sha_object(sha_obj); 6307 } 6308 if (state == NULL) return false; 6309 6310 // Call the stub. 6311 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6312 OptoRuntime::digestBase_implCompressMB_Type(), 6313 stubAddr, stubName, TypePtr::BOTTOM, 6314 src_start, state, ofs, limit); 6315 // return ofs (int) 6316 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6317 set_result(result); 6318 6319 return true; 6320 } 6321 6322 //------------------------------get_state_from_sha_object----------------------- 6323 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6324 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6325 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6326 if (sha_state == NULL) return (Node *) NULL; 6327 6328 // now have the array, need to get the start address of the state array 6329 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6330 return state; 6331 } 6332 6333 //------------------------------get_state_from_sha5_object----------------------- 6334 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6335 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6336 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6337 if (sha_state == NULL) return (Node *) NULL; 6338 6339 // now have the array, need to get the start address of the state array 6340 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6341 return state; 6342 } 6343 6344 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6345 // Return node representing slow path of predicate check. 6346 // the pseudo code we want to emulate with this predicate is: 6347 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6348 // 6349 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6350 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6351 "need SHA1/SHA256/SHA512 instruction support"); 6352 assert((uint)predicate < 3, "sanity"); 6353 6354 // The receiver was checked for NULL already. 6355 Node* digestBaseObj = argument(0); 6356 6357 // get DigestBase klass for instanceOf check 6358 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6359 assert(tinst != NULL, "digestBaseObj is null"); 6360 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6361 6362 const char* klass_SHA_name = NULL; 6363 switch (predicate) { 6364 case 0: 6365 if (UseSHA1Intrinsics) { 6366 // we want to do an instanceof comparison against the SHA class 6367 klass_SHA_name = "sun/security/provider/SHA"; 6368 } 6369 break; 6370 case 1: 6371 if (UseSHA256Intrinsics) { 6372 // we want to do an instanceof comparison against the SHA2 class 6373 klass_SHA_name = "sun/security/provider/SHA2"; 6374 } 6375 break; 6376 case 2: 6377 if (UseSHA512Intrinsics) { 6378 // we want to do an instanceof comparison against the SHA5 class 6379 klass_SHA_name = "sun/security/provider/SHA5"; 6380 } 6381 break; 6382 default: 6383 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6384 } 6385 6386 ciKlass* klass_SHA = NULL; 6387 if (klass_SHA_name != NULL) { 6388 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6389 } 6390 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6391 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6392 Node* ctrl = control(); 6393 set_control(top()); // no intrinsic path 6394 return ctrl; 6395 } 6396 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6397 6398 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6399 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6400 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6401 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6402 6403 return instof_false; // even if it is NULL 6404 } 6405 6406 bool LibraryCallKit::inline_profileBoolean() { 6407 Node* counts = argument(1); 6408 const TypeAryPtr* ary = NULL; 6409 ciArray* aobj = NULL; 6410 if (counts->is_Con() 6411 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6412 && (aobj = ary->const_oop()->as_array()) != NULL 6413 && (aobj->length() == 2)) { 6414 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6415 jint false_cnt = aobj->element_value(0).as_int(); 6416 jint true_cnt = aobj->element_value(1).as_int(); 6417 6418 if (C->log() != NULL) { 6419 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6420 false_cnt, true_cnt); 6421 } 6422 6423 if (false_cnt + true_cnt == 0) { 6424 // According to profile, never executed. 6425 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6426 Deoptimization::Action_reinterpret); 6427 return true; 6428 } 6429 6430 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6431 // is a number of each value occurrences. 6432 Node* result = argument(0); 6433 if (false_cnt == 0 || true_cnt == 0) { 6434 // According to profile, one value has been never seen. 6435 int expected_val = (false_cnt == 0) ? 1 : 0; 6436 6437 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6438 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6439 6440 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6441 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6442 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6443 6444 { // Slow path: uncommon trap for never seen value and then reexecute 6445 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6446 // the value has been seen at least once. 6447 PreserveJVMState pjvms(this); 6448 PreserveReexecuteState preexecs(this); 6449 jvms()->set_should_reexecute(true); 6450 6451 set_control(slow_path); 6452 set_i_o(i_o()); 6453 6454 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6455 Deoptimization::Action_reinterpret); 6456 } 6457 // The guard for never seen value enables sharpening of the result and 6458 // returning a constant. It allows to eliminate branches on the same value 6459 // later on. 6460 set_control(fast_path); 6461 result = intcon(expected_val); 6462 } 6463 // Stop profiling. 6464 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6465 // By replacing method body with profile data (represented as ProfileBooleanNode 6466 // on IR level) we effectively disable profiling. 6467 // It enables full speed execution once optimized code is generated. 6468 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6469 C->record_for_igvn(profile); 6470 set_result(profile); 6471 return true; 6472 } else { 6473 // Continue profiling. 6474 // Profile data isn't available at the moment. So, execute method's bytecode version. 6475 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6476 // is compiled and counters aren't available since corresponding MethodHandle 6477 // isn't a compile-time constant. 6478 return false; 6479 } 6480 } 6481 6482 bool LibraryCallKit::inline_isCompileConstant() { 6483 Node* n = argument(0); 6484 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6485 return true; 6486 }