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(); 300 301 bool inline_profileBoolean(); 302 bool inline_isCompileConstant(); 303 }; 304 305 306 //---------------------------make_vm_intrinsic---------------------------- 307 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 308 vmIntrinsics::ID id = m->intrinsic_id(); 309 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 310 311 ccstr disable_intr = NULL; 312 313 if ((DisableIntrinsic[0] != '\0' 314 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 315 (method_has_option_value("DisableIntrinsic", disable_intr) 316 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 317 // disabled by a user request on the command line: 318 // example: -XX:DisableIntrinsic=_hashCode,_getClass 319 return NULL; 320 } 321 322 if (!m->is_loaded()) { 323 // do not attempt to inline unloaded methods 324 return NULL; 325 } 326 327 // Only a few intrinsics implement a virtual dispatch. 328 // They are expensive calls which are also frequently overridden. 329 if (is_virtual) { 330 switch (id) { 331 case vmIntrinsics::_hashCode: 332 case vmIntrinsics::_clone: 333 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 334 break; 335 default: 336 return NULL; 337 } 338 } 339 340 // -XX:-InlineNatives disables nearly all intrinsics: 341 if (!InlineNatives) { 342 switch (id) { 343 case vmIntrinsics::_indexOf: 344 case vmIntrinsics::_compareTo: 345 case vmIntrinsics::_equals: 346 case vmIntrinsics::_equalsC: 347 case vmIntrinsics::_getAndAddInt: 348 case vmIntrinsics::_getAndAddLong: 349 case vmIntrinsics::_getAndSetInt: 350 case vmIntrinsics::_getAndSetLong: 351 case vmIntrinsics::_getAndSetObject: 352 case vmIntrinsics::_loadFence: 353 case vmIntrinsics::_storeFence: 354 case vmIntrinsics::_fullFence: 355 break; // InlineNatives does not control String.compareTo 356 case vmIntrinsics::_Reference_get: 357 break; // InlineNatives does not control Reference.get 358 default: 359 return NULL; 360 } 361 } 362 363 int predicates = 0; 364 bool does_virtual_dispatch = false; 365 366 switch (id) { 367 case vmIntrinsics::_compareTo: 368 if (!SpecialStringCompareTo) return NULL; 369 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 370 break; 371 case vmIntrinsics::_indexOf: 372 if (!SpecialStringIndexOf) return NULL; 373 break; 374 case vmIntrinsics::_equals: 375 if (!SpecialStringEquals) return NULL; 376 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 377 break; 378 case vmIntrinsics::_equalsC: 379 if (!SpecialArraysEquals) return NULL; 380 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 381 break; 382 case vmIntrinsics::_arraycopy: 383 if (!InlineArrayCopy) return NULL; 384 break; 385 case vmIntrinsics::_copyMemory: 386 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 387 if (!InlineArrayCopy) return NULL; 388 break; 389 case vmIntrinsics::_hashCode: 390 if (!InlineObjectHash) return NULL; 391 does_virtual_dispatch = true; 392 break; 393 case vmIntrinsics::_clone: 394 does_virtual_dispatch = true; 395 case vmIntrinsics::_copyOf: 396 case vmIntrinsics::_copyOfRange: 397 if (!InlineObjectCopy) return NULL; 398 // These also use the arraycopy intrinsic mechanism: 399 if (!InlineArrayCopy) return NULL; 400 break; 401 case vmIntrinsics::_encodeISOArray: 402 if (!SpecialEncodeISOArray) return NULL; 403 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 404 break; 405 case vmIntrinsics::_checkIndex: 406 // We do not intrinsify this. The optimizer does fine with it. 407 return NULL; 408 409 case vmIntrinsics::_getCallerClass: 410 if (!InlineReflectionGetCallerClass) return NULL; 411 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 412 break; 413 414 case vmIntrinsics::_bitCount_i: 415 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 416 break; 417 418 case vmIntrinsics::_bitCount_l: 419 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 420 break; 421 422 case vmIntrinsics::_numberOfLeadingZeros_i: 423 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 424 break; 425 426 case vmIntrinsics::_numberOfLeadingZeros_l: 427 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 428 break; 429 430 case vmIntrinsics::_numberOfTrailingZeros_i: 431 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 432 break; 433 434 case vmIntrinsics::_numberOfTrailingZeros_l: 435 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 436 break; 437 438 case vmIntrinsics::_reverseBytes_c: 439 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 440 break; 441 case vmIntrinsics::_reverseBytes_s: 442 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 443 break; 444 case vmIntrinsics::_reverseBytes_i: 445 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 446 break; 447 case vmIntrinsics::_reverseBytes_l: 448 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 449 break; 450 451 case vmIntrinsics::_Reference_get: 452 // Use the intrinsic version of Reference.get() so that the value in 453 // the referent field can be registered by the G1 pre-barrier code. 454 // Also add memory barrier to prevent commoning reads from this field 455 // across safepoint since GC can change it value. 456 break; 457 458 case vmIntrinsics::_compareAndSwapObject: 459 #ifdef _LP64 460 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 461 #endif 462 break; 463 464 case vmIntrinsics::_compareAndSwapLong: 465 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 466 break; 467 468 case vmIntrinsics::_getAndAddInt: 469 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 470 break; 471 472 case vmIntrinsics::_getAndAddLong: 473 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 474 break; 475 476 case vmIntrinsics::_getAndSetInt: 477 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 478 break; 479 480 case vmIntrinsics::_getAndSetLong: 481 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 482 break; 483 484 case vmIntrinsics::_getAndSetObject: 485 #ifdef _LP64 486 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 487 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 488 break; 489 #else 490 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 491 break; 492 #endif 493 494 case vmIntrinsics::_aescrypt_encryptBlock: 495 case vmIntrinsics::_aescrypt_decryptBlock: 496 if (!UseAESIntrinsics) return NULL; 497 break; 498 499 case vmIntrinsics::_multiplyToLen: 500 if (!UseMultiplyToLenIntrinsic) return NULL; 501 break; 502 503 case vmIntrinsics::_squareToLen: 504 if (!UseSquareToLenIntrinsic) return NULL; 505 break; 506 507 case vmIntrinsics::_mulAdd: 508 if (!UseMulAddIntrinsic) return NULL; 509 break; 510 511 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 512 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 513 if (!UseAESIntrinsics) return NULL; 514 // these two require the predicated logic 515 predicates = 1; 516 break; 517 518 case vmIntrinsics::_sha_implCompress: 519 if (!UseSHA1Intrinsics) return NULL; 520 break; 521 522 case vmIntrinsics::_sha2_implCompress: 523 if (!UseSHA256Intrinsics) return NULL; 524 break; 525 526 case vmIntrinsics::_sha5_implCompress: 527 if (!UseSHA512Intrinsics) return NULL; 528 break; 529 530 case vmIntrinsics::_digestBase_implCompressMB: 531 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 532 predicates = 3; 533 break; 534 535 case vmIntrinsics::_ghash_processBlocks: 536 if (!UseGHASHIntrinsics) return NULL; 537 break; 538 539 case vmIntrinsics::_updateCRC32: 540 case vmIntrinsics::_updateBytesCRC32: 541 case vmIntrinsics::_updateByteBufferCRC32: 542 if (!UseCRC32Intrinsics) return NULL; 543 break; 544 545 case vmIntrinsics::_updateBytesCRC32C: 546 case vmIntrinsics::_updateDirectByteBufferCRC32C: 547 if (!UseCRC32CIntrinsics) return NULL; 548 break; 549 550 case vmIntrinsics::_incrementExactI: 551 case vmIntrinsics::_addExactI: 552 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 553 break; 554 case vmIntrinsics::_incrementExactL: 555 case vmIntrinsics::_addExactL: 556 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 557 break; 558 case vmIntrinsics::_decrementExactI: 559 case vmIntrinsics::_subtractExactI: 560 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 561 break; 562 case vmIntrinsics::_decrementExactL: 563 case vmIntrinsics::_subtractExactL: 564 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 565 break; 566 case vmIntrinsics::_negateExactI: 567 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 568 break; 569 case vmIntrinsics::_negateExactL: 570 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 571 break; 572 case vmIntrinsics::_multiplyExactI: 573 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 574 break; 575 case vmIntrinsics::_multiplyExactL: 576 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 577 break; 578 579 case vmIntrinsics::_getShortUnaligned: 580 case vmIntrinsics::_getCharUnaligned: 581 case vmIntrinsics::_getIntUnaligned: 582 case vmIntrinsics::_getLongUnaligned: 583 case vmIntrinsics::_putShortUnaligned: 584 case vmIntrinsics::_putCharUnaligned: 585 case vmIntrinsics::_putIntUnaligned: 586 case vmIntrinsics::_putLongUnaligned: 587 if (!UseUnalignedAccesses) return NULL; 588 break; 589 590 default: 591 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 592 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 593 break; 594 } 595 596 // -XX:-InlineClassNatives disables natives from the Class class. 597 // The flag applies to all reflective calls, notably Array.newArray 598 // (visible to Java programmers as Array.newInstance). 599 if (m->holder()->name() == ciSymbol::java_lang_Class() || 600 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 601 if (!InlineClassNatives) return NULL; 602 } 603 604 // -XX:-InlineThreadNatives disables natives from the Thread class. 605 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 606 if (!InlineThreadNatives) return NULL; 607 } 608 609 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 610 if (m->holder()->name() == ciSymbol::java_lang_Math() || 611 m->holder()->name() == ciSymbol::java_lang_Float() || 612 m->holder()->name() == ciSymbol::java_lang_Double()) { 613 if (!InlineMathNatives) return NULL; 614 } 615 616 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 617 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 618 if (!InlineUnsafeOps) return NULL; 619 } 620 621 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 622 } 623 624 //----------------------register_library_intrinsics----------------------- 625 // Initialize this file's data structures, for each Compile instance. 626 void Compile::register_library_intrinsics() { 627 // Nothing to do here. 628 } 629 630 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 631 LibraryCallKit kit(jvms, this); 632 Compile* C = kit.C; 633 int nodes = C->unique(); 634 #ifndef PRODUCT 635 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 636 char buf[1000]; 637 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 638 tty->print_cr("Intrinsic %s", str); 639 } 640 #endif 641 ciMethod* callee = kit.callee(); 642 const int bci = kit.bci(); 643 644 // Try to inline the intrinsic. 645 if (kit.try_to_inline(_last_predicate)) { 646 if (C->print_intrinsics() || C->print_inlining()) { 647 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 648 } 649 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 650 if (C->log()) { 651 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 652 vmIntrinsics::name_at(intrinsic_id()), 653 (is_virtual() ? " virtual='1'" : ""), 654 C->unique() - nodes); 655 } 656 // Push the result from the inlined method onto the stack. 657 kit.push_result(); 658 C->print_inlining_update(this); 659 return kit.transfer_exceptions_into_jvms(); 660 } 661 662 // The intrinsic bailed out 663 if (C->print_intrinsics() || C->print_inlining()) { 664 if (jvms->has_method()) { 665 // Not a root compile. 666 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 667 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 668 } else { 669 // Root compile 670 tty->print("Did not generate intrinsic %s%s at bci:%d in", 671 vmIntrinsics::name_at(intrinsic_id()), 672 (is_virtual() ? " (virtual)" : ""), bci); 673 } 674 } 675 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 676 C->print_inlining_update(this); 677 return NULL; 678 } 679 680 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 681 LibraryCallKit kit(jvms, this); 682 Compile* C = kit.C; 683 int nodes = C->unique(); 684 _last_predicate = predicate; 685 #ifndef PRODUCT 686 assert(is_predicated() && predicate < predicates_count(), "sanity"); 687 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 688 char buf[1000]; 689 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 690 tty->print_cr("Predicate for intrinsic %s", str); 691 } 692 #endif 693 ciMethod* callee = kit.callee(); 694 const int bci = kit.bci(); 695 696 Node* slow_ctl = kit.try_to_predicate(predicate); 697 if (!kit.failing()) { 698 if (C->print_intrinsics() || C->print_inlining()) { 699 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 700 } 701 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 702 if (C->log()) { 703 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 704 vmIntrinsics::name_at(intrinsic_id()), 705 (is_virtual() ? " virtual='1'" : ""), 706 C->unique() - nodes); 707 } 708 return slow_ctl; // Could be NULL if the check folds. 709 } 710 711 // The intrinsic bailed out 712 if (C->print_intrinsics() || C->print_inlining()) { 713 if (jvms->has_method()) { 714 // Not a root compile. 715 const char* msg = "failed to generate predicate for intrinsic"; 716 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 717 } else { 718 // Root compile 719 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 720 vmIntrinsics::name_at(intrinsic_id()), 721 (is_virtual() ? " (virtual)" : ""), bci); 722 } 723 } 724 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 725 return NULL; 726 } 727 728 bool LibraryCallKit::try_to_inline(int predicate) { 729 // Handle symbolic names for otherwise undistinguished boolean switches: 730 const bool is_store = true; 731 const bool is_native_ptr = true; 732 const bool is_static = true; 733 const bool is_volatile = true; 734 735 if (!jvms()->has_method()) { 736 // Root JVMState has a null method. 737 assert(map()->memory()->Opcode() == Op_Parm, ""); 738 // Insert the memory aliasing node 739 set_all_memory(reset_memory()); 740 } 741 assert(merged_memory(), ""); 742 743 744 switch (intrinsic_id()) { 745 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 746 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 747 case vmIntrinsics::_getClass: return inline_native_getClass(); 748 749 case vmIntrinsics::_dsin: 750 case vmIntrinsics::_dcos: 751 case vmIntrinsics::_dtan: 752 case vmIntrinsics::_dabs: 753 case vmIntrinsics::_datan2: 754 case vmIntrinsics::_dsqrt: 755 case vmIntrinsics::_dexp: 756 case vmIntrinsics::_dlog: 757 case vmIntrinsics::_dlog10: 758 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 759 760 case vmIntrinsics::_min: 761 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 762 763 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 764 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 765 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 766 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 767 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 768 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 769 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 770 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 771 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 772 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 773 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 774 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 775 776 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 777 778 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 779 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 780 case vmIntrinsics::_equals: return inline_string_equals(); 781 782 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 783 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 784 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 785 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 786 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 787 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 788 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 789 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 790 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 791 792 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 793 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 794 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 795 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 796 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 797 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 798 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 799 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 800 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 801 802 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 803 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 804 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 805 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 806 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 807 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 808 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 809 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 810 811 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 812 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 813 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 814 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 815 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 816 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 817 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 818 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 819 820 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 821 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 822 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 823 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 824 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 825 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 826 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 827 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 828 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 829 830 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 831 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 832 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 833 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 834 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 835 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 836 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 837 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 838 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 839 840 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 841 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 842 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 843 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 844 845 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 846 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 847 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 848 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 849 850 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 851 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 852 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 853 854 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 855 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 856 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 857 858 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 859 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 860 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 861 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 862 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 863 864 case vmIntrinsics::_loadFence: 865 case vmIntrinsics::_storeFence: 866 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 867 868 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 869 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 870 871 #ifdef TRACE_HAVE_INTRINSICS 872 case vmIntrinsics::_classID: return inline_native_classID(); 873 case vmIntrinsics::_threadID: return inline_native_threadID(); 874 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 875 #endif 876 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 877 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 878 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 879 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 880 case vmIntrinsics::_newArray: return inline_native_newArray(); 881 case vmIntrinsics::_getLength: return inline_native_getLength(); 882 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 883 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 884 case vmIntrinsics::_equalsC: return inline_array_equals(); 885 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 886 887 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 888 889 case vmIntrinsics::_isInstance: 890 case vmIntrinsics::_getModifiers: 891 case vmIntrinsics::_isInterface: 892 case vmIntrinsics::_isArray: 893 case vmIntrinsics::_isPrimitive: 894 case vmIntrinsics::_getSuperclass: 895 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 896 897 case vmIntrinsics::_floatToRawIntBits: 898 case vmIntrinsics::_floatToIntBits: 899 case vmIntrinsics::_intBitsToFloat: 900 case vmIntrinsics::_doubleToRawLongBits: 901 case vmIntrinsics::_doubleToLongBits: 902 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 903 904 case vmIntrinsics::_numberOfLeadingZeros_i: 905 case vmIntrinsics::_numberOfLeadingZeros_l: 906 case vmIntrinsics::_numberOfTrailingZeros_i: 907 case vmIntrinsics::_numberOfTrailingZeros_l: 908 case vmIntrinsics::_bitCount_i: 909 case vmIntrinsics::_bitCount_l: 910 case vmIntrinsics::_reverseBytes_i: 911 case vmIntrinsics::_reverseBytes_l: 912 case vmIntrinsics::_reverseBytes_s: 913 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 914 915 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 916 917 case vmIntrinsics::_Reference_get: return inline_reference_get(); 918 919 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 920 921 case vmIntrinsics::_aescrypt_encryptBlock: 922 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 923 924 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 925 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 926 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 927 928 case vmIntrinsics::_sha_implCompress: 929 case vmIntrinsics::_sha2_implCompress: 930 case vmIntrinsics::_sha5_implCompress: 931 return inline_sha_implCompress(intrinsic_id()); 932 933 case vmIntrinsics::_digestBase_implCompressMB: 934 return inline_digestBase_implCompressMB(predicate); 935 936 case vmIntrinsics::_multiplyToLen: 937 return inline_multiplyToLen(); 938 939 case vmIntrinsics::_squareToLen: 940 return inline_squareToLen(); 941 942 case vmIntrinsics::_mulAdd: 943 return inline_mulAdd(); 944 945 case vmIntrinsics::_ghash_processBlocks: 946 return inline_ghash_processBlocks(); 947 948 case vmIntrinsics::_encodeISOArray: 949 return inline_encodeISOArray(); 950 951 case vmIntrinsics::_updateCRC32: 952 return inline_updateCRC32(); 953 case vmIntrinsics::_updateBytesCRC32: 954 return inline_updateBytesCRC32(); 955 case vmIntrinsics::_updateByteBufferCRC32: 956 return inline_updateByteBufferCRC32(); 957 958 case vmIntrinsics::_updateBytesCRC32C: 959 return inline_updateBytesCRC32C(); 960 case vmIntrinsics::_updateDirectByteBufferCRC32C: 961 return inline_updateDirectByteBufferCRC32C(); 962 963 case vmIntrinsics::_profileBoolean: 964 return inline_profileBoolean(); 965 case vmIntrinsics::_isCompileConstant: 966 return inline_isCompileConstant(); 967 968 default: 969 // If you get here, it may be that someone has added a new intrinsic 970 // to the list in vmSymbols.hpp without implementing it here. 971 #ifndef PRODUCT 972 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 973 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 974 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 975 } 976 #endif 977 return false; 978 } 979 } 980 981 Node* LibraryCallKit::try_to_predicate(int predicate) { 982 if (!jvms()->has_method()) { 983 // Root JVMState has a null method. 984 assert(map()->memory()->Opcode() == Op_Parm, ""); 985 // Insert the memory aliasing node 986 set_all_memory(reset_memory()); 987 } 988 assert(merged_memory(), ""); 989 990 switch (intrinsic_id()) { 991 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 992 return inline_cipherBlockChaining_AESCrypt_predicate(false); 993 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 994 return inline_cipherBlockChaining_AESCrypt_predicate(true); 995 case vmIntrinsics::_digestBase_implCompressMB: 996 return inline_digestBase_implCompressMB_predicate(predicate); 997 998 default: 999 // If you get here, it may be that someone has added a new intrinsic 1000 // to the list in vmSymbols.hpp without implementing it here. 1001 #ifndef PRODUCT 1002 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 1003 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 1004 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 1005 } 1006 #endif 1007 Node* slow_ctl = control(); 1008 set_control(top()); // No fast path instrinsic 1009 return slow_ctl; 1010 } 1011 } 1012 1013 //------------------------------set_result------------------------------- 1014 // Helper function for finishing intrinsics. 1015 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1016 record_for_igvn(region); 1017 set_control(_gvn.transform(region)); 1018 set_result( _gvn.transform(value)); 1019 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1020 } 1021 1022 //------------------------------generate_guard--------------------------- 1023 // Helper function for generating guarded fast-slow graph structures. 1024 // The given 'test', if true, guards a slow path. If the test fails 1025 // then a fast path can be taken. (We generally hope it fails.) 1026 // In all cases, GraphKit::control() is updated to the fast path. 1027 // The returned value represents the control for the slow path. 1028 // The return value is never 'top'; it is either a valid control 1029 // or NULL if it is obvious that the slow path can never be taken. 1030 // Also, if region and the slow control are not NULL, the slow edge 1031 // is appended to the region. 1032 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1033 if (stopped()) { 1034 // Already short circuited. 1035 return NULL; 1036 } 1037 1038 // Build an if node and its projections. 1039 // If test is true we take the slow path, which we assume is uncommon. 1040 if (_gvn.type(test) == TypeInt::ZERO) { 1041 // The slow branch is never taken. No need to build this guard. 1042 return NULL; 1043 } 1044 1045 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1046 1047 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1048 if (if_slow == top()) { 1049 // The slow branch is never taken. No need to build this guard. 1050 return NULL; 1051 } 1052 1053 if (region != NULL) 1054 region->add_req(if_slow); 1055 1056 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1057 set_control(if_fast); 1058 1059 return if_slow; 1060 } 1061 1062 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1063 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1064 } 1065 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1066 return generate_guard(test, region, PROB_FAIR); 1067 } 1068 1069 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1070 Node* *pos_index) { 1071 if (stopped()) 1072 return NULL; // already stopped 1073 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1074 return NULL; // index is already adequately typed 1075 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1076 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1077 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1078 if (is_neg != NULL && pos_index != NULL) { 1079 // Emulate effect of Parse::adjust_map_after_if. 1080 Node* ccast = new CastIINode(index, TypeInt::POS); 1081 ccast->set_req(0, control()); 1082 (*pos_index) = _gvn.transform(ccast); 1083 } 1084 return is_neg; 1085 } 1086 1087 // Make sure that 'position' is a valid limit index, in [0..length]. 1088 // There are two equivalent plans for checking this: 1089 // A. (offset + copyLength) unsigned<= arrayLength 1090 // B. offset <= (arrayLength - copyLength) 1091 // We require that all of the values above, except for the sum and 1092 // difference, are already known to be non-negative. 1093 // Plan A is robust in the face of overflow, if offset and copyLength 1094 // are both hugely positive. 1095 // 1096 // Plan B is less direct and intuitive, but it does not overflow at 1097 // all, since the difference of two non-negatives is always 1098 // representable. Whenever Java methods must perform the equivalent 1099 // check they generally use Plan B instead of Plan A. 1100 // For the moment we use Plan A. 1101 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1102 Node* subseq_length, 1103 Node* array_length, 1104 RegionNode* region) { 1105 if (stopped()) 1106 return NULL; // already stopped 1107 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1108 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1109 return NULL; // common case of whole-array copy 1110 Node* last = subseq_length; 1111 if (!zero_offset) // last += offset 1112 last = _gvn.transform(new AddINode(last, offset)); 1113 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1114 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1115 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1116 return is_over; 1117 } 1118 1119 1120 //--------------------------generate_current_thread-------------------- 1121 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1122 ciKlass* thread_klass = env()->Thread_klass(); 1123 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1124 Node* thread = _gvn.transform(new ThreadLocalNode()); 1125 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1126 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1127 tls_output = thread; 1128 return threadObj; 1129 } 1130 1131 1132 //------------------------------make_string_method_node------------------------ 1133 // Helper method for String intrinsic functions. This version is called 1134 // with str1 and str2 pointing to String object nodes. 1135 // 1136 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1137 Node* no_ctrl = NULL; 1138 1139 // Get start addr of string 1140 Node* str1_value = load_String_value(no_ctrl, str1); 1141 Node* str1_offset = load_String_offset(no_ctrl, str1); 1142 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1143 1144 // Get length of string 1 1145 Node* str1_len = load_String_length(no_ctrl, str1); 1146 1147 Node* str2_value = load_String_value(no_ctrl, str2); 1148 Node* str2_offset = load_String_offset(no_ctrl, str2); 1149 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1150 1151 Node* str2_len = NULL; 1152 Node* result = NULL; 1153 1154 switch (opcode) { 1155 case Op_StrIndexOf: 1156 // Get length of string 2 1157 str2_len = load_String_length(no_ctrl, str2); 1158 1159 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1160 str1_start, str1_len, str2_start, str2_len); 1161 break; 1162 case Op_StrComp: 1163 // Get length of string 2 1164 str2_len = load_String_length(no_ctrl, str2); 1165 1166 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1167 str1_start, str1_len, str2_start, str2_len); 1168 break; 1169 case Op_StrEquals: 1170 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1171 str1_start, str2_start, str1_len); 1172 break; 1173 default: 1174 ShouldNotReachHere(); 1175 return NULL; 1176 } 1177 1178 // All these intrinsics have checks. 1179 C->set_has_split_ifs(true); // Has chance for split-if optimization 1180 1181 return _gvn.transform(result); 1182 } 1183 1184 // Helper method for String intrinsic functions. This version is called 1185 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1186 // to Int nodes containing the lenghts of str1 and str2. 1187 // 1188 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1189 Node* result = NULL; 1190 switch (opcode) { 1191 case Op_StrIndexOf: 1192 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1193 str1_start, cnt1, str2_start, cnt2); 1194 break; 1195 case Op_StrComp: 1196 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1197 str1_start, cnt1, str2_start, cnt2); 1198 break; 1199 case Op_StrEquals: 1200 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1201 str1_start, str2_start, cnt1); 1202 break; 1203 default: 1204 ShouldNotReachHere(); 1205 return NULL; 1206 } 1207 1208 // All these intrinsics have checks. 1209 C->set_has_split_ifs(true); // Has chance for split-if optimization 1210 1211 return _gvn.transform(result); 1212 } 1213 1214 //------------------------------inline_string_compareTo------------------------ 1215 // public int java.lang.String.compareTo(String anotherString); 1216 bool LibraryCallKit::inline_string_compareTo() { 1217 Node* receiver = null_check(argument(0)); 1218 Node* arg = null_check(argument(1)); 1219 if (stopped()) { 1220 return true; 1221 } 1222 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1223 return true; 1224 } 1225 1226 //------------------------------inline_string_equals------------------------ 1227 bool LibraryCallKit::inline_string_equals() { 1228 Node* receiver = null_check_receiver(); 1229 // NOTE: Do not null check argument for String.equals() because spec 1230 // allows to specify NULL as argument. 1231 Node* argument = this->argument(1); 1232 if (stopped()) { 1233 return true; 1234 } 1235 1236 // paths (plus control) merge 1237 RegionNode* region = new RegionNode(5); 1238 Node* phi = new PhiNode(region, TypeInt::BOOL); 1239 1240 // does source == target string? 1241 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1242 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1243 1244 Node* if_eq = generate_slow_guard(bol, NULL); 1245 if (if_eq != NULL) { 1246 // receiver == argument 1247 phi->init_req(2, intcon(1)); 1248 region->init_req(2, if_eq); 1249 } 1250 1251 // get String klass for instanceOf 1252 ciInstanceKlass* klass = env()->String_klass(); 1253 1254 if (!stopped()) { 1255 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1256 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1257 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1258 1259 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1260 //instanceOf == true, fallthrough 1261 1262 if (inst_false != NULL) { 1263 phi->init_req(3, intcon(0)); 1264 region->init_req(3, inst_false); 1265 } 1266 } 1267 1268 if (!stopped()) { 1269 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1270 1271 // Properly cast the argument to String 1272 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1273 // This path is taken only when argument's type is String:NotNull. 1274 argument = cast_not_null(argument, false); 1275 1276 Node* no_ctrl = NULL; 1277 1278 // Get start addr of receiver 1279 Node* receiver_val = load_String_value(no_ctrl, receiver); 1280 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1281 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1282 1283 // Get length of receiver 1284 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1285 1286 // Get start addr of argument 1287 Node* argument_val = load_String_value(no_ctrl, argument); 1288 Node* argument_offset = load_String_offset(no_ctrl, argument); 1289 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1290 1291 // Get length of argument 1292 Node* argument_cnt = load_String_length(no_ctrl, argument); 1293 1294 // Check for receiver count != argument count 1295 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1296 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1297 Node* if_ne = generate_slow_guard(bol, NULL); 1298 if (if_ne != NULL) { 1299 phi->init_req(4, intcon(0)); 1300 region->init_req(4, if_ne); 1301 } 1302 1303 // Check for count == 0 is done by assembler code for StrEquals. 1304 1305 if (!stopped()) { 1306 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1307 phi->init_req(1, equals); 1308 region->init_req(1, control()); 1309 } 1310 } 1311 1312 // post merge 1313 set_control(_gvn.transform(region)); 1314 record_for_igvn(region); 1315 1316 set_result(_gvn.transform(phi)); 1317 return true; 1318 } 1319 1320 //------------------------------inline_array_equals---------------------------- 1321 bool LibraryCallKit::inline_array_equals() { 1322 Node* arg1 = argument(0); 1323 Node* arg2 = argument(1); 1324 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1325 return true; 1326 } 1327 1328 // Java version of String.indexOf(constant string) 1329 // class StringDecl { 1330 // StringDecl(char[] ca) { 1331 // offset = 0; 1332 // count = ca.length; 1333 // value = ca; 1334 // } 1335 // int offset; 1336 // int count; 1337 // char[] value; 1338 // } 1339 // 1340 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1341 // int targetOffset, int cache_i, int md2) { 1342 // int cache = cache_i; 1343 // int sourceOffset = string_object.offset; 1344 // int sourceCount = string_object.count; 1345 // int targetCount = target_object.length; 1346 // 1347 // int targetCountLess1 = targetCount - 1; 1348 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1349 // 1350 // char[] source = string_object.value; 1351 // char[] target = target_object; 1352 // int lastChar = target[targetCountLess1]; 1353 // 1354 // outer_loop: 1355 // for (int i = sourceOffset; i < sourceEnd; ) { 1356 // int src = source[i + targetCountLess1]; 1357 // if (src == lastChar) { 1358 // // With random strings and a 4-character alphabet, 1359 // // reverse matching at this point sets up 0.8% fewer 1360 // // frames, but (paradoxically) makes 0.3% more probes. 1361 // // Since those probes are nearer the lastChar probe, 1362 // // there is may be a net D$ win with reverse matching. 1363 // // But, reversing loop inhibits unroll of inner loop 1364 // // for unknown reason. So, does running outer loop from 1365 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1366 // for (int j = 0; j < targetCountLess1; j++) { 1367 // if (target[targetOffset + j] != source[i+j]) { 1368 // if ((cache & (1 << source[i+j])) == 0) { 1369 // if (md2 < j+1) { 1370 // i += j+1; 1371 // continue outer_loop; 1372 // } 1373 // } 1374 // i += md2; 1375 // continue outer_loop; 1376 // } 1377 // } 1378 // return i - sourceOffset; 1379 // } 1380 // if ((cache & (1 << src)) == 0) { 1381 // i += targetCountLess1; 1382 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1383 // i++; 1384 // } 1385 // return -1; 1386 // } 1387 1388 //------------------------------string_indexOf------------------------ 1389 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1390 jint cache_i, jint md2_i) { 1391 1392 Node* no_ctrl = NULL; 1393 float likely = PROB_LIKELY(0.9); 1394 float unlikely = PROB_UNLIKELY(0.9); 1395 1396 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1397 1398 Node* source = load_String_value(no_ctrl, string_object); 1399 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1400 Node* sourceCount = load_String_length(no_ctrl, string_object); 1401 1402 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1403 jint target_length = target_array->length(); 1404 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1405 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1406 1407 // String.value field is known to be @Stable. 1408 if (UseImplicitStableValues) { 1409 target = cast_array_to_stable(target, target_type); 1410 } 1411 1412 IdealKit kit(this, false, true); 1413 #define __ kit. 1414 Node* zero = __ ConI(0); 1415 Node* one = __ ConI(1); 1416 Node* cache = __ ConI(cache_i); 1417 Node* md2 = __ ConI(md2_i); 1418 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1419 Node* targetCountLess1 = __ ConI(target_length - 1); 1420 Node* targetOffset = __ ConI(targetOffset_i); 1421 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1422 1423 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1424 Node* outer_loop = __ make_label(2 /* goto */); 1425 Node* return_ = __ make_label(1); 1426 1427 __ set(rtn,__ ConI(-1)); 1428 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1429 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1430 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1431 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1432 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1433 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1434 Node* tpj = __ AddI(targetOffset, __ value(j)); 1435 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1436 Node* ipj = __ AddI(__ value(i), __ value(j)); 1437 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1438 __ if_then(targ, BoolTest::ne, src2); { 1439 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1440 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1441 __ increment(i, __ AddI(__ value(j), one)); 1442 __ goto_(outer_loop); 1443 } __ end_if(); __ dead(j); 1444 }__ end_if(); __ dead(j); 1445 __ increment(i, md2); 1446 __ goto_(outer_loop); 1447 }__ end_if(); 1448 __ increment(j, one); 1449 }__ end_loop(); __ dead(j); 1450 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1451 __ goto_(return_); 1452 }__ end_if(); 1453 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1454 __ increment(i, targetCountLess1); 1455 }__ end_if(); 1456 __ increment(i, one); 1457 __ bind(outer_loop); 1458 }__ end_loop(); __ dead(i); 1459 __ bind(return_); 1460 1461 // Final sync IdealKit and GraphKit. 1462 final_sync(kit); 1463 Node* result = __ value(rtn); 1464 #undef __ 1465 C->set_has_loops(true); 1466 return result; 1467 } 1468 1469 //------------------------------inline_string_indexOf------------------------ 1470 bool LibraryCallKit::inline_string_indexOf() { 1471 Node* receiver = argument(0); 1472 Node* arg = argument(1); 1473 1474 Node* result; 1475 if (Matcher::has_match_rule(Op_StrIndexOf) && 1476 UseSSE42Intrinsics) { 1477 // Generate SSE4.2 version of indexOf 1478 // We currently only have match rules that use SSE4.2 1479 1480 receiver = null_check(receiver); 1481 arg = null_check(arg); 1482 if (stopped()) { 1483 return true; 1484 } 1485 1486 // Make the merge point 1487 RegionNode* result_rgn = new RegionNode(4); 1488 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1489 Node* no_ctrl = NULL; 1490 1491 // Get start addr of source string 1492 Node* source = load_String_value(no_ctrl, receiver); 1493 Node* source_offset = load_String_offset(no_ctrl, receiver); 1494 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1495 1496 // Get length of source string 1497 Node* source_cnt = load_String_length(no_ctrl, receiver); 1498 1499 // Get start addr of substring 1500 Node* substr = load_String_value(no_ctrl, arg); 1501 Node* substr_offset = load_String_offset(no_ctrl, arg); 1502 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1503 1504 // Get length of source string 1505 Node* substr_cnt = load_String_length(no_ctrl, arg); 1506 1507 // Check for substr count > string count 1508 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1509 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1510 Node* if_gt = generate_slow_guard(bol, NULL); 1511 if (if_gt != NULL) { 1512 result_phi->init_req(2, intcon(-1)); 1513 result_rgn->init_req(2, if_gt); 1514 } 1515 1516 if (!stopped()) { 1517 // Check for substr count == 0 1518 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1519 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1520 Node* if_zero = generate_slow_guard(bol, NULL); 1521 if (if_zero != NULL) { 1522 result_phi->init_req(3, intcon(0)); 1523 result_rgn->init_req(3, if_zero); 1524 } 1525 } 1526 1527 if (!stopped()) { 1528 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1529 result_phi->init_req(1, result); 1530 result_rgn->init_req(1, control()); 1531 } 1532 set_control(_gvn.transform(result_rgn)); 1533 record_for_igvn(result_rgn); 1534 result = _gvn.transform(result_phi); 1535 1536 } else { // Use LibraryCallKit::string_indexOf 1537 // don't intrinsify if argument isn't a constant string. 1538 if (!arg->is_Con()) { 1539 return false; 1540 } 1541 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1542 if (str_type == NULL) { 1543 return false; 1544 } 1545 ciInstanceKlass* klass = env()->String_klass(); 1546 ciObject* str_const = str_type->const_oop(); 1547 if (str_const == NULL || str_const->klass() != klass) { 1548 return false; 1549 } 1550 ciInstance* str = str_const->as_instance(); 1551 assert(str != NULL, "must be instance"); 1552 1553 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1554 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1555 1556 int o; 1557 int c; 1558 if (java_lang_String::has_offset_field()) { 1559 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1560 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1561 } else { 1562 o = 0; 1563 c = pat->length(); 1564 } 1565 1566 // constant strings have no offset and count == length which 1567 // simplifies the resulting code somewhat so lets optimize for that. 1568 if (o != 0 || c != pat->length()) { 1569 return false; 1570 } 1571 1572 receiver = null_check(receiver, T_OBJECT); 1573 // NOTE: No null check on the argument is needed since it's a constant String oop. 1574 if (stopped()) { 1575 return true; 1576 } 1577 1578 // The null string as a pattern always returns 0 (match at beginning of string) 1579 if (c == 0) { 1580 set_result(intcon(0)); 1581 return true; 1582 } 1583 1584 // Generate default indexOf 1585 jchar lastChar = pat->char_at(o + (c - 1)); 1586 int cache = 0; 1587 int i; 1588 for (i = 0; i < c - 1; i++) { 1589 assert(i < pat->length(), "out of range"); 1590 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1591 } 1592 1593 int md2 = c; 1594 for (i = 0; i < c - 1; i++) { 1595 assert(i < pat->length(), "out of range"); 1596 if (pat->char_at(o + i) == lastChar) { 1597 md2 = (c - 1) - i; 1598 } 1599 } 1600 1601 result = string_indexOf(receiver, pat, o, cache, md2); 1602 } 1603 set_result(result); 1604 return true; 1605 } 1606 1607 //--------------------------round_double_node-------------------------------- 1608 // Round a double node if necessary. 1609 Node* LibraryCallKit::round_double_node(Node* n) { 1610 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1611 n = _gvn.transform(new RoundDoubleNode(0, n)); 1612 return n; 1613 } 1614 1615 //------------------------------inline_math----------------------------------- 1616 // public static double Math.abs(double) 1617 // public static double Math.sqrt(double) 1618 // public static double Math.log(double) 1619 // public static double Math.log10(double) 1620 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1621 Node* arg = round_double_node(argument(0)); 1622 Node* n; 1623 switch (id) { 1624 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1625 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1626 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1627 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1628 default: fatal_unexpected_iid(id); break; 1629 } 1630 set_result(_gvn.transform(n)); 1631 return true; 1632 } 1633 1634 //------------------------------inline_trig---------------------------------- 1635 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1636 // argument reduction which will turn into a fast/slow diamond. 1637 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1638 Node* arg = round_double_node(argument(0)); 1639 Node* n = NULL; 1640 1641 switch (id) { 1642 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1643 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1644 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1645 default: fatal_unexpected_iid(id); break; 1646 } 1647 n = _gvn.transform(n); 1648 1649 // Rounding required? Check for argument reduction! 1650 if (Matcher::strict_fp_requires_explicit_rounding) { 1651 static const double pi_4 = 0.7853981633974483; 1652 static const double neg_pi_4 = -0.7853981633974483; 1653 // pi/2 in 80-bit extended precision 1654 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1655 // -pi/2 in 80-bit extended precision 1656 // 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}; 1657 // Cutoff value for using this argument reduction technique 1658 //static const double pi_2_minus_epsilon = 1.564660403643354; 1659 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1660 1661 // Pseudocode for sin: 1662 // if (x <= Math.PI / 4.0) { 1663 // if (x >= -Math.PI / 4.0) return fsin(x); 1664 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1665 // } else { 1666 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1667 // } 1668 // return StrictMath.sin(x); 1669 1670 // Pseudocode for cos: 1671 // if (x <= Math.PI / 4.0) { 1672 // if (x >= -Math.PI / 4.0) return fcos(x); 1673 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1674 // } else { 1675 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1676 // } 1677 // return StrictMath.cos(x); 1678 1679 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1680 // requires a special machine instruction to load it. Instead we'll try 1681 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1682 // probably do the math inside the SIN encoding. 1683 1684 // Make the merge point 1685 RegionNode* r = new RegionNode(3); 1686 Node* phi = new PhiNode(r, Type::DOUBLE); 1687 1688 // Flatten arg so we need only 1 test 1689 Node *abs = _gvn.transform(new AbsDNode(arg)); 1690 // Node for PI/4 constant 1691 Node *pi4 = makecon(TypeD::make(pi_4)); 1692 // Check PI/4 : abs(arg) 1693 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1694 // Check: If PI/4 < abs(arg) then go slow 1695 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1696 // Branch either way 1697 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1698 set_control(opt_iff(r,iff)); 1699 1700 // Set fast path result 1701 phi->init_req(2, n); 1702 1703 // Slow path - non-blocking leaf call 1704 Node* call = NULL; 1705 switch (id) { 1706 case vmIntrinsics::_dsin: 1707 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1708 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1709 "Sin", NULL, arg, top()); 1710 break; 1711 case vmIntrinsics::_dcos: 1712 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1713 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1714 "Cos", NULL, arg, top()); 1715 break; 1716 case vmIntrinsics::_dtan: 1717 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1718 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1719 "Tan", NULL, arg, top()); 1720 break; 1721 } 1722 assert(control()->in(0) == call, ""); 1723 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1724 r->init_req(1, control()); 1725 phi->init_req(1, slow_result); 1726 1727 // Post-merge 1728 set_control(_gvn.transform(r)); 1729 record_for_igvn(r); 1730 n = _gvn.transform(phi); 1731 1732 C->set_has_split_ifs(true); // Has chance for split-if optimization 1733 } 1734 set_result(n); 1735 return true; 1736 } 1737 1738 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1739 //------------------- 1740 //result=(result.isNaN())? funcAddr():result; 1741 // Check: If isNaN() by checking result!=result? then either trap 1742 // or go to runtime 1743 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1744 // Build the boolean node 1745 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1746 1747 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1748 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1749 // The pow or exp intrinsic returned a NaN, which requires a call 1750 // to the runtime. Recompile with the runtime call. 1751 uncommon_trap(Deoptimization::Reason_intrinsic, 1752 Deoptimization::Action_make_not_entrant); 1753 } 1754 return result; 1755 } else { 1756 // If this inlining ever returned NaN in the past, we compile a call 1757 // to the runtime to properly handle corner cases 1758 1759 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1760 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1761 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1762 1763 if (!if_slow->is_top()) { 1764 RegionNode* result_region = new RegionNode(3); 1765 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1766 1767 result_region->init_req(1, if_fast); 1768 result_val->init_req(1, result); 1769 1770 set_control(if_slow); 1771 1772 const TypePtr* no_memory_effects = NULL; 1773 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1774 no_memory_effects, 1775 x, top(), y, y ? top() : NULL); 1776 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1777 #ifdef ASSERT 1778 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1779 assert(value_top == top(), "second value must be top"); 1780 #endif 1781 1782 result_region->init_req(2, control()); 1783 result_val->init_req(2, value); 1784 set_control(_gvn.transform(result_region)); 1785 return _gvn.transform(result_val); 1786 } else { 1787 return result; 1788 } 1789 } 1790 } 1791 1792 //------------------------------inline_exp------------------------------------- 1793 // Inline exp instructions, if possible. The Intel hardware only misses 1794 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1795 bool LibraryCallKit::inline_exp() { 1796 Node* arg = round_double_node(argument(0)); 1797 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1798 1799 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1800 set_result(n); 1801 1802 C->set_has_split_ifs(true); // Has chance for split-if optimization 1803 return true; 1804 } 1805 1806 //------------------------------inline_pow------------------------------------- 1807 // Inline power instructions, if possible. 1808 bool LibraryCallKit::inline_pow() { 1809 // Pseudocode for pow 1810 // if (y == 2) { 1811 // return x * x; 1812 // } else { 1813 // if (x <= 0.0) { 1814 // long longy = (long)y; 1815 // if ((double)longy == y) { // if y is long 1816 // if (y + 1 == y) longy = 0; // huge number: even 1817 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1818 // } else { 1819 // result = NaN; 1820 // } 1821 // } else { 1822 // result = DPow(x,y); 1823 // } 1824 // if (result != result)? { 1825 // result = uncommon_trap() or runtime_call(); 1826 // } 1827 // return result; 1828 // } 1829 1830 Node* x = round_double_node(argument(0)); 1831 Node* y = round_double_node(argument(2)); 1832 1833 Node* result = NULL; 1834 1835 Node* const_two_node = makecon(TypeD::make(2.0)); 1836 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1837 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1838 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1839 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1840 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1841 1842 RegionNode* region_node = new RegionNode(3); 1843 region_node->init_req(1, if_true); 1844 1845 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1846 // special case for x^y where y == 2, we can convert it to x * x 1847 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1848 1849 // set control to if_false since we will now process the false branch 1850 set_control(if_false); 1851 1852 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1853 // Short form: skip the fancy tests and just check for NaN result. 1854 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1855 } else { 1856 // If this inlining ever returned NaN in the past, include all 1857 // checks + call to the runtime. 1858 1859 // Set the merge point for If node with condition of (x <= 0.0) 1860 // There are four possible paths to region node and phi node 1861 RegionNode *r = new RegionNode(4); 1862 Node *phi = new PhiNode(r, Type::DOUBLE); 1863 1864 // Build the first if node: if (x <= 0.0) 1865 // Node for 0 constant 1866 Node *zeronode = makecon(TypeD::ZERO); 1867 // Check x:0 1868 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1869 // Check: If (x<=0) then go complex path 1870 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1871 // Branch either way 1872 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1873 // Fast path taken; set region slot 3 1874 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1875 r->init_req(3,fast_taken); // Capture fast-control 1876 1877 // Fast path not-taken, i.e. slow path 1878 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1879 1880 // Set fast path result 1881 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1882 phi->init_req(3, fast_result); 1883 1884 // Complex path 1885 // Build the second if node (if y is long) 1886 // Node for (long)y 1887 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1888 // Node for (double)((long) y) 1889 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1890 // Check (double)((long) y) : y 1891 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1892 // Check if (y isn't long) then go to slow path 1893 1894 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1895 // Branch either way 1896 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1897 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1898 1899 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1900 1901 // Calculate DPow(abs(x), y)*(1 & (long)y) 1902 // Node for constant 1 1903 Node *conone = longcon(1); 1904 // 1& (long)y 1905 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1906 1907 // A huge number is always even. Detect a huge number by checking 1908 // if y + 1 == y and set integer to be tested for parity to 0. 1909 // Required for corner case: 1910 // (long)9.223372036854776E18 = max_jlong 1911 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1912 // max_jlong is odd but 9.223372036854776E18 is even 1913 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1914 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1915 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1916 Node* correctedsign = NULL; 1917 if (ConditionalMoveLimit != 0) { 1918 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1919 } else { 1920 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1921 RegionNode *r = new RegionNode(3); 1922 Node *phi = new PhiNode(r, TypeLong::LONG); 1923 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1924 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1925 phi->init_req(1, signnode); 1926 phi->init_req(2, longcon(0)); 1927 correctedsign = _gvn.transform(phi); 1928 ylong_path = _gvn.transform(r); 1929 record_for_igvn(r); 1930 } 1931 1932 // zero node 1933 Node *conzero = longcon(0); 1934 // Check (1&(long)y)==0? 1935 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1936 // Check if (1&(long)y)!=0?, if so the result is negative 1937 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1938 // abs(x) 1939 Node *absx=_gvn.transform(new AbsDNode(x)); 1940 // abs(x)^y 1941 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1942 // -abs(x)^y 1943 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1944 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1945 Node *signresult = NULL; 1946 if (ConditionalMoveLimit != 0) { 1947 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1948 } else { 1949 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1950 RegionNode *r = new RegionNode(3); 1951 Node *phi = new PhiNode(r, Type::DOUBLE); 1952 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1953 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1954 phi->init_req(1, absxpowy); 1955 phi->init_req(2, negabsxpowy); 1956 signresult = _gvn.transform(phi); 1957 ylong_path = _gvn.transform(r); 1958 record_for_igvn(r); 1959 } 1960 // Set complex path fast result 1961 r->init_req(2, ylong_path); 1962 phi->init_req(2, signresult); 1963 1964 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1965 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1966 r->init_req(1,slow_path); 1967 phi->init_req(1,slow_result); 1968 1969 // Post merge 1970 set_control(_gvn.transform(r)); 1971 record_for_igvn(r); 1972 result = _gvn.transform(phi); 1973 } 1974 1975 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1976 1977 // control from finish_pow_exp is now input to the region node 1978 region_node->set_req(2, control()); 1979 // the result from finish_pow_exp is now input to the phi node 1980 phi_node->init_req(2, result); 1981 set_control(_gvn.transform(region_node)); 1982 record_for_igvn(region_node); 1983 set_result(_gvn.transform(phi_node)); 1984 1985 C->set_has_split_ifs(true); // Has chance for split-if optimization 1986 return true; 1987 } 1988 1989 //------------------------------runtime_math----------------------------- 1990 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1991 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1992 "must be (DD)D or (D)D type"); 1993 1994 // Inputs 1995 Node* a = round_double_node(argument(0)); 1996 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1997 1998 const TypePtr* no_memory_effects = NULL; 1999 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 2000 no_memory_effects, 2001 a, top(), b, b ? top() : NULL); 2002 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 2003 #ifdef ASSERT 2004 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 2005 assert(value_top == top(), "second value must be top"); 2006 #endif 2007 2008 set_result(value); 2009 return true; 2010 } 2011 2012 //------------------------------inline_math_native----------------------------- 2013 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2014 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2015 switch (id) { 2016 // These intrinsics are not properly supported on all hardware 2017 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2018 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2019 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2020 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2021 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2022 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2023 2024 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2025 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2026 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2027 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2028 2029 // These intrinsics are supported on all hardware 2030 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2031 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2032 2033 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2034 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2035 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2036 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2037 #undef FN_PTR 2038 2039 // These intrinsics are not yet correctly implemented 2040 case vmIntrinsics::_datan2: 2041 return false; 2042 2043 default: 2044 fatal_unexpected_iid(id); 2045 return false; 2046 } 2047 } 2048 2049 static bool is_simple_name(Node* n) { 2050 return (n->req() == 1 // constant 2051 || (n->is_Type() && n->as_Type()->type()->singleton()) 2052 || n->is_Proj() // parameter or return value 2053 || n->is_Phi() // local of some sort 2054 ); 2055 } 2056 2057 //----------------------------inline_min_max----------------------------------- 2058 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2059 set_result(generate_min_max(id, argument(0), argument(1))); 2060 return true; 2061 } 2062 2063 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2064 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2065 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2066 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2067 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2068 2069 { 2070 PreserveJVMState pjvms(this); 2071 PreserveReexecuteState preexecs(this); 2072 jvms()->set_should_reexecute(true); 2073 2074 set_control(slow_path); 2075 set_i_o(i_o()); 2076 2077 uncommon_trap(Deoptimization::Reason_intrinsic, 2078 Deoptimization::Action_none); 2079 } 2080 2081 set_control(fast_path); 2082 set_result(math); 2083 } 2084 2085 template <typename OverflowOp> 2086 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2087 typedef typename OverflowOp::MathOp MathOp; 2088 2089 MathOp* mathOp = new MathOp(arg1, arg2); 2090 Node* operation = _gvn.transform( mathOp ); 2091 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2092 inline_math_mathExact(operation, ofcheck); 2093 return true; 2094 } 2095 2096 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2097 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2098 } 2099 2100 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2101 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2102 } 2103 2104 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2105 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2106 } 2107 2108 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2109 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2110 } 2111 2112 bool LibraryCallKit::inline_math_negateExactI() { 2113 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2114 } 2115 2116 bool LibraryCallKit::inline_math_negateExactL() { 2117 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2118 } 2119 2120 bool LibraryCallKit::inline_math_multiplyExactI() { 2121 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2122 } 2123 2124 bool LibraryCallKit::inline_math_multiplyExactL() { 2125 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2126 } 2127 2128 Node* 2129 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2130 // These are the candidate return value: 2131 Node* xvalue = x0; 2132 Node* yvalue = y0; 2133 2134 if (xvalue == yvalue) { 2135 return xvalue; 2136 } 2137 2138 bool want_max = (id == vmIntrinsics::_max); 2139 2140 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2141 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2142 if (txvalue == NULL || tyvalue == NULL) return top(); 2143 // This is not really necessary, but it is consistent with a 2144 // hypothetical MaxINode::Value method: 2145 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2146 2147 // %%% This folding logic should (ideally) be in a different place. 2148 // Some should be inside IfNode, and there to be a more reliable 2149 // transformation of ?: style patterns into cmoves. We also want 2150 // more powerful optimizations around cmove and min/max. 2151 2152 // Try to find a dominating comparison of these guys. 2153 // It can simplify the index computation for Arrays.copyOf 2154 // and similar uses of System.arraycopy. 2155 // First, compute the normalized version of CmpI(x, y). 2156 int cmp_op = Op_CmpI; 2157 Node* xkey = xvalue; 2158 Node* ykey = yvalue; 2159 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2160 if (ideal_cmpxy->is_Cmp()) { 2161 // E.g., if we have CmpI(length - offset, count), 2162 // it might idealize to CmpI(length, count + offset) 2163 cmp_op = ideal_cmpxy->Opcode(); 2164 xkey = ideal_cmpxy->in(1); 2165 ykey = ideal_cmpxy->in(2); 2166 } 2167 2168 // Start by locating any relevant comparisons. 2169 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2170 Node* cmpxy = NULL; 2171 Node* cmpyx = NULL; 2172 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2173 Node* cmp = start_from->fast_out(k); 2174 if (cmp->outcnt() > 0 && // must have prior uses 2175 cmp->in(0) == NULL && // must be context-independent 2176 cmp->Opcode() == cmp_op) { // right kind of compare 2177 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2178 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2179 } 2180 } 2181 2182 const int NCMPS = 2; 2183 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2184 int cmpn; 2185 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2186 if (cmps[cmpn] != NULL) break; // find a result 2187 } 2188 if (cmpn < NCMPS) { 2189 // Look for a dominating test that tells us the min and max. 2190 int depth = 0; // Limit search depth for speed 2191 Node* dom = control(); 2192 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2193 if (++depth >= 100) break; 2194 Node* ifproj = dom; 2195 if (!ifproj->is_Proj()) continue; 2196 Node* iff = ifproj->in(0); 2197 if (!iff->is_If()) continue; 2198 Node* bol = iff->in(1); 2199 if (!bol->is_Bool()) continue; 2200 Node* cmp = bol->in(1); 2201 if (cmp == NULL) continue; 2202 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2203 if (cmps[cmpn] == cmp) break; 2204 if (cmpn == NCMPS) continue; 2205 BoolTest::mask btest = bol->as_Bool()->_test._test; 2206 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2207 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2208 // At this point, we know that 'x btest y' is true. 2209 switch (btest) { 2210 case BoolTest::eq: 2211 // They are proven equal, so we can collapse the min/max. 2212 // Either value is the answer. Choose the simpler. 2213 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2214 return yvalue; 2215 return xvalue; 2216 case BoolTest::lt: // x < y 2217 case BoolTest::le: // x <= y 2218 return (want_max ? yvalue : xvalue); 2219 case BoolTest::gt: // x > y 2220 case BoolTest::ge: // x >= y 2221 return (want_max ? xvalue : yvalue); 2222 } 2223 } 2224 } 2225 2226 // We failed to find a dominating test. 2227 // Let's pick a test that might GVN with prior tests. 2228 Node* best_bol = NULL; 2229 BoolTest::mask best_btest = BoolTest::illegal; 2230 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2231 Node* cmp = cmps[cmpn]; 2232 if (cmp == NULL) continue; 2233 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2234 Node* bol = cmp->fast_out(j); 2235 if (!bol->is_Bool()) continue; 2236 BoolTest::mask btest = bol->as_Bool()->_test._test; 2237 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2238 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2239 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2240 best_bol = bol->as_Bool(); 2241 best_btest = btest; 2242 } 2243 } 2244 } 2245 2246 Node* answer_if_true = NULL; 2247 Node* answer_if_false = NULL; 2248 switch (best_btest) { 2249 default: 2250 if (cmpxy == NULL) 2251 cmpxy = ideal_cmpxy; 2252 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2253 // and fall through: 2254 case BoolTest::lt: // x < y 2255 case BoolTest::le: // x <= y 2256 answer_if_true = (want_max ? yvalue : xvalue); 2257 answer_if_false = (want_max ? xvalue : yvalue); 2258 break; 2259 case BoolTest::gt: // x > y 2260 case BoolTest::ge: // x >= y 2261 answer_if_true = (want_max ? xvalue : yvalue); 2262 answer_if_false = (want_max ? yvalue : xvalue); 2263 break; 2264 } 2265 2266 jint hi, lo; 2267 if (want_max) { 2268 // We can sharpen the minimum. 2269 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2270 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2271 } else { 2272 // We can sharpen the maximum. 2273 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2274 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2275 } 2276 2277 // Use a flow-free graph structure, to avoid creating excess control edges 2278 // which could hinder other optimizations. 2279 // Since Math.min/max is often used with arraycopy, we want 2280 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2281 Node* cmov = CMoveNode::make(NULL, best_bol, 2282 answer_if_false, answer_if_true, 2283 TypeInt::make(lo, hi, widen)); 2284 2285 return _gvn.transform(cmov); 2286 2287 /* 2288 // This is not as desirable as it may seem, since Min and Max 2289 // nodes do not have a full set of optimizations. 2290 // And they would interfere, anyway, with 'if' optimizations 2291 // and with CMoveI canonical forms. 2292 switch (id) { 2293 case vmIntrinsics::_min: 2294 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2295 case vmIntrinsics::_max: 2296 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2297 default: 2298 ShouldNotReachHere(); 2299 } 2300 */ 2301 } 2302 2303 inline int 2304 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2305 const TypePtr* base_type = TypePtr::NULL_PTR; 2306 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2307 if (base_type == NULL) { 2308 // Unknown type. 2309 return Type::AnyPtr; 2310 } else if (base_type == TypePtr::NULL_PTR) { 2311 // Since this is a NULL+long form, we have to switch to a rawptr. 2312 base = _gvn.transform(new CastX2PNode(offset)); 2313 offset = MakeConX(0); 2314 return Type::RawPtr; 2315 } else if (base_type->base() == Type::RawPtr) { 2316 return Type::RawPtr; 2317 } else if (base_type->isa_oopptr()) { 2318 // Base is never null => always a heap address. 2319 if (base_type->ptr() == TypePtr::NotNull) { 2320 return Type::OopPtr; 2321 } 2322 // Offset is small => always a heap address. 2323 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2324 if (offset_type != NULL && 2325 base_type->offset() == 0 && // (should always be?) 2326 offset_type->_lo >= 0 && 2327 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2328 return Type::OopPtr; 2329 } 2330 // Otherwise, it might either be oop+off or NULL+addr. 2331 return Type::AnyPtr; 2332 } else { 2333 // No information: 2334 return Type::AnyPtr; 2335 } 2336 } 2337 2338 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2339 int kind = classify_unsafe_addr(base, offset); 2340 if (kind == Type::RawPtr) { 2341 return basic_plus_adr(top(), base, offset); 2342 } else { 2343 return basic_plus_adr(base, offset); 2344 } 2345 } 2346 2347 //--------------------------inline_number_methods----------------------------- 2348 // inline int Integer.numberOfLeadingZeros(int) 2349 // inline int Long.numberOfLeadingZeros(long) 2350 // 2351 // inline int Integer.numberOfTrailingZeros(int) 2352 // inline int Long.numberOfTrailingZeros(long) 2353 // 2354 // inline int Integer.bitCount(int) 2355 // inline int Long.bitCount(long) 2356 // 2357 // inline char Character.reverseBytes(char) 2358 // inline short Short.reverseBytes(short) 2359 // inline int Integer.reverseBytes(int) 2360 // inline long Long.reverseBytes(long) 2361 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2362 Node* arg = argument(0); 2363 Node* n; 2364 switch (id) { 2365 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2366 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2367 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2368 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2369 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2370 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2371 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2372 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2373 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2374 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2375 default: fatal_unexpected_iid(id); break; 2376 } 2377 set_result(_gvn.transform(n)); 2378 return true; 2379 } 2380 2381 //----------------------------inline_unsafe_access---------------------------- 2382 2383 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2384 2385 // Helper that guards and inserts a pre-barrier. 2386 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2387 Node* pre_val, bool need_mem_bar) { 2388 // We could be accessing the referent field of a reference object. If so, when G1 2389 // is enabled, we need to log the value in the referent field in an SATB buffer. 2390 // This routine performs some compile time filters and generates suitable 2391 // runtime filters that guard the pre-barrier code. 2392 // Also add memory barrier for non volatile load from the referent field 2393 // to prevent commoning of loads across safepoint. 2394 if (!UseG1GC && !need_mem_bar) 2395 return; 2396 2397 // Some compile time checks. 2398 2399 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2400 const TypeX* otype = offset->find_intptr_t_type(); 2401 if (otype != NULL && otype->is_con() && 2402 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2403 // Constant offset but not the reference_offset so just return 2404 return; 2405 } 2406 2407 // We only need to generate the runtime guards for instances. 2408 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2409 if (btype != NULL) { 2410 if (btype->isa_aryptr()) { 2411 // Array type so nothing to do 2412 return; 2413 } 2414 2415 const TypeInstPtr* itype = btype->isa_instptr(); 2416 if (itype != NULL) { 2417 // Can the klass of base_oop be statically determined to be 2418 // _not_ a sub-class of Reference and _not_ Object? 2419 ciKlass* klass = itype->klass(); 2420 if ( klass->is_loaded() && 2421 !klass->is_subtype_of(env()->Reference_klass()) && 2422 !env()->Object_klass()->is_subtype_of(klass)) { 2423 return; 2424 } 2425 } 2426 } 2427 2428 // The compile time filters did not reject base_oop/offset so 2429 // we need to generate the following runtime filters 2430 // 2431 // if (offset == java_lang_ref_Reference::_reference_offset) { 2432 // if (instance_of(base, java.lang.ref.Reference)) { 2433 // pre_barrier(_, pre_val, ...); 2434 // } 2435 // } 2436 2437 float likely = PROB_LIKELY( 0.999); 2438 float unlikely = PROB_UNLIKELY(0.999); 2439 2440 IdealKit ideal(this); 2441 #define __ ideal. 2442 2443 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2444 2445 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2446 // Update graphKit memory and control from IdealKit. 2447 sync_kit(ideal); 2448 2449 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2450 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2451 2452 // Update IdealKit memory and control from graphKit. 2453 __ sync_kit(this); 2454 2455 Node* one = __ ConI(1); 2456 // is_instof == 0 if base_oop == NULL 2457 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2458 2459 // Update graphKit from IdeakKit. 2460 sync_kit(ideal); 2461 2462 // Use the pre-barrier to record the value in the referent field 2463 pre_barrier(false /* do_load */, 2464 __ ctrl(), 2465 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2466 pre_val /* pre_val */, 2467 T_OBJECT); 2468 if (need_mem_bar) { 2469 // Add memory barrier to prevent commoning reads from this field 2470 // across safepoint since GC can change its value. 2471 insert_mem_bar(Op_MemBarCPUOrder); 2472 } 2473 // Update IdealKit from graphKit. 2474 __ sync_kit(this); 2475 2476 } __ end_if(); // _ref_type != ref_none 2477 } __ end_if(); // offset == referent_offset 2478 2479 // Final sync IdealKit and GraphKit. 2480 final_sync(ideal); 2481 #undef __ 2482 } 2483 2484 2485 // Interpret Unsafe.fieldOffset cookies correctly: 2486 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2487 2488 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2489 // Attempt to infer a sharper value type from the offset and base type. 2490 ciKlass* sharpened_klass = NULL; 2491 2492 // See if it is an instance field, with an object type. 2493 if (alias_type->field() != NULL) { 2494 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2495 if (alias_type->field()->type()->is_klass()) { 2496 sharpened_klass = alias_type->field()->type()->as_klass(); 2497 } 2498 } 2499 2500 // See if it is a narrow oop array. 2501 if (adr_type->isa_aryptr()) { 2502 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2503 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2504 if (elem_type != NULL) { 2505 sharpened_klass = elem_type->klass(); 2506 } 2507 } 2508 } 2509 2510 // The sharpened class might be unloaded if there is no class loader 2511 // contraint in place. 2512 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2513 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2514 2515 #ifndef PRODUCT 2516 if (C->print_intrinsics() || C->print_inlining()) { 2517 tty->print(" from base type: "); adr_type->dump(); 2518 tty->print(" sharpened value: "); tjp->dump(); 2519 } 2520 #endif 2521 // Sharpen the value type. 2522 return tjp; 2523 } 2524 return NULL; 2525 } 2526 2527 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2528 if (callee()->is_static()) return false; // caller must have the capability! 2529 2530 #ifndef PRODUCT 2531 { 2532 ResourceMark rm; 2533 // Check the signatures. 2534 ciSignature* sig = callee()->signature(); 2535 #ifdef ASSERT 2536 if (!is_store) { 2537 // Object getObject(Object base, int/long offset), etc. 2538 BasicType rtype = sig->return_type()->basic_type(); 2539 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2540 rtype = T_ADDRESS; // it is really a C void* 2541 assert(rtype == type, "getter must return the expected value"); 2542 if (!is_native_ptr) { 2543 assert(sig->count() == 2, "oop getter has 2 arguments"); 2544 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2545 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2546 } else { 2547 assert(sig->count() == 1, "native getter has 1 argument"); 2548 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2549 } 2550 } else { 2551 // void putObject(Object base, int/long offset, Object x), etc. 2552 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2553 if (!is_native_ptr) { 2554 assert(sig->count() == 3, "oop putter has 3 arguments"); 2555 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2556 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2557 } else { 2558 assert(sig->count() == 2, "native putter has 2 arguments"); 2559 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2560 } 2561 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2562 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2563 vtype = T_ADDRESS; // it is really a C void* 2564 assert(vtype == type, "putter must accept the expected value"); 2565 } 2566 #endif // ASSERT 2567 } 2568 #endif //PRODUCT 2569 2570 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2571 2572 Node* receiver = argument(0); // type: oop 2573 2574 // Build address expression. 2575 Node* adr; 2576 Node* heap_base_oop = top(); 2577 Node* offset = top(); 2578 Node* val; 2579 2580 if (!is_native_ptr) { 2581 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2582 Node* base = argument(1); // type: oop 2583 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2584 offset = argument(2); // type: long 2585 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2586 // to be plain byte offsets, which are also the same as those accepted 2587 // by oopDesc::field_base. 2588 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2589 "fieldOffset must be byte-scaled"); 2590 // 32-bit machines ignore the high half! 2591 offset = ConvL2X(offset); 2592 adr = make_unsafe_address(base, offset); 2593 heap_base_oop = base; 2594 val = is_store ? argument(4) : NULL; 2595 } else { 2596 Node* ptr = argument(1); // type: long 2597 ptr = ConvL2X(ptr); // adjust Java long to machine word 2598 adr = make_unsafe_address(NULL, ptr); 2599 val = is_store ? argument(3) : NULL; 2600 } 2601 2602 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2603 2604 // First guess at the value type. 2605 const Type *value_type = Type::get_const_basic_type(type); 2606 2607 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2608 // there was not enough information to nail it down. 2609 Compile::AliasType* alias_type = C->alias_type(adr_type); 2610 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2611 2612 // We will need memory barriers unless we can determine a unique 2613 // alias category for this reference. (Note: If for some reason 2614 // the barriers get omitted and the unsafe reference begins to "pollute" 2615 // the alias analysis of the rest of the graph, either Compile::can_alias 2616 // or Compile::must_alias will throw a diagnostic assert.) 2617 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2618 2619 // If we are reading the value of the referent field of a Reference 2620 // object (either by using Unsafe directly or through reflection) 2621 // then, if G1 is enabled, we need to record the referent in an 2622 // SATB log buffer using the pre-barrier mechanism. 2623 // Also we need to add memory barrier to prevent commoning reads 2624 // from this field across safepoint since GC can change its value. 2625 bool need_read_barrier = !is_native_ptr && !is_store && 2626 offset != top() && heap_base_oop != top(); 2627 2628 if (!is_store && type == T_OBJECT) { 2629 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2630 if (tjp != NULL) { 2631 value_type = tjp; 2632 } 2633 } 2634 2635 receiver = null_check(receiver); 2636 if (stopped()) { 2637 return true; 2638 } 2639 // Heap pointers get a null-check from the interpreter, 2640 // as a courtesy. However, this is not guaranteed by Unsafe, 2641 // and it is not possible to fully distinguish unintended nulls 2642 // from intended ones in this API. 2643 2644 if (is_volatile) { 2645 // We need to emit leading and trailing CPU membars (see below) in 2646 // addition to memory membars when is_volatile. This is a little 2647 // too strong, but avoids the need to insert per-alias-type 2648 // volatile membars (for stores; compare Parse::do_put_xxx), which 2649 // we cannot do effectively here because we probably only have a 2650 // rough approximation of type. 2651 need_mem_bar = true; 2652 // For Stores, place a memory ordering barrier now. 2653 if (is_store) { 2654 insert_mem_bar(Op_MemBarRelease); 2655 } else { 2656 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2657 insert_mem_bar(Op_MemBarVolatile); 2658 } 2659 } 2660 } 2661 2662 // Memory barrier to prevent normal and 'unsafe' accesses from 2663 // bypassing each other. Happens after null checks, so the 2664 // exception paths do not take memory state from the memory barrier, 2665 // so there's no problems making a strong assert about mixing users 2666 // of safe & unsafe memory. 2667 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2668 2669 if (!is_store) { 2670 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2671 // To be valid, unsafe loads may depend on other conditions than 2672 // the one that guards them: pin the Load node 2673 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2674 // load value 2675 switch (type) { 2676 case T_BOOLEAN: 2677 case T_CHAR: 2678 case T_BYTE: 2679 case T_SHORT: 2680 case T_INT: 2681 case T_LONG: 2682 case T_FLOAT: 2683 case T_DOUBLE: 2684 break; 2685 case T_OBJECT: 2686 if (need_read_barrier) { 2687 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2688 } 2689 break; 2690 case T_ADDRESS: 2691 // Cast to an int type. 2692 p = _gvn.transform(new CastP2XNode(NULL, p)); 2693 p = ConvX2UL(p); 2694 break; 2695 default: 2696 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2697 break; 2698 } 2699 // The load node has the control of the preceding MemBarCPUOrder. All 2700 // following nodes will have the control of the MemBarCPUOrder inserted at 2701 // the end of this method. So, pushing the load onto the stack at a later 2702 // point is fine. 2703 set_result(p); 2704 } else { 2705 // place effect of store into memory 2706 switch (type) { 2707 case T_DOUBLE: 2708 val = dstore_rounding(val); 2709 break; 2710 case T_ADDRESS: 2711 // Repackage the long as a pointer. 2712 val = ConvL2X(val); 2713 val = _gvn.transform(new CastX2PNode(val)); 2714 break; 2715 } 2716 2717 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2718 if (type != T_OBJECT ) { 2719 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2720 } else { 2721 // Possibly an oop being stored to Java heap or native memory 2722 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2723 // oop to Java heap. 2724 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2725 } else { 2726 // We can't tell at compile time if we are storing in the Java heap or outside 2727 // of it. So we need to emit code to conditionally do the proper type of 2728 // store. 2729 2730 IdealKit ideal(this); 2731 #define __ ideal. 2732 // QQQ who knows what probability is here?? 2733 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2734 // Sync IdealKit and graphKit. 2735 sync_kit(ideal); 2736 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2737 // Update IdealKit memory. 2738 __ sync_kit(this); 2739 } __ else_(); { 2740 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2741 } __ end_if(); 2742 // Final sync IdealKit and GraphKit. 2743 final_sync(ideal); 2744 #undef __ 2745 } 2746 } 2747 } 2748 2749 if (is_volatile) { 2750 if (!is_store) { 2751 insert_mem_bar(Op_MemBarAcquire); 2752 } else { 2753 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2754 insert_mem_bar(Op_MemBarVolatile); 2755 } 2756 } 2757 } 2758 2759 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2760 2761 return true; 2762 } 2763 2764 //----------------------------inline_unsafe_load_store---------------------------- 2765 // This method serves a couple of different customers (depending on LoadStoreKind): 2766 // 2767 // LS_cmpxchg: 2768 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2769 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2770 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2771 // 2772 // LS_xadd: 2773 // public int getAndAddInt( Object o, long offset, int delta) 2774 // public long getAndAddLong(Object o, long offset, long delta) 2775 // 2776 // LS_xchg: 2777 // int getAndSet(Object o, long offset, int newValue) 2778 // long getAndSet(Object o, long offset, long newValue) 2779 // Object getAndSet(Object o, long offset, Object newValue) 2780 // 2781 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2782 // This basic scheme here is the same as inline_unsafe_access, but 2783 // differs in enough details that combining them would make the code 2784 // overly confusing. (This is a true fact! I originally combined 2785 // them, but even I was confused by it!) As much code/comments as 2786 // possible are retained from inline_unsafe_access though to make 2787 // the correspondences clearer. - dl 2788 2789 if (callee()->is_static()) return false; // caller must have the capability! 2790 2791 #ifndef PRODUCT 2792 BasicType rtype; 2793 { 2794 ResourceMark rm; 2795 // Check the signatures. 2796 ciSignature* sig = callee()->signature(); 2797 rtype = sig->return_type()->basic_type(); 2798 if (kind == LS_xadd || kind == LS_xchg) { 2799 // Check the signatures. 2800 #ifdef ASSERT 2801 assert(rtype == type, "get and set must return the expected type"); 2802 assert(sig->count() == 3, "get and set has 3 arguments"); 2803 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2804 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2805 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2806 #endif // ASSERT 2807 } else if (kind == LS_cmpxchg) { 2808 // Check the signatures. 2809 #ifdef ASSERT 2810 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2811 assert(sig->count() == 4, "CAS has 4 arguments"); 2812 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2813 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2814 #endif // ASSERT 2815 } else { 2816 ShouldNotReachHere(); 2817 } 2818 } 2819 #endif //PRODUCT 2820 2821 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2822 2823 // Get arguments: 2824 Node* receiver = NULL; 2825 Node* base = NULL; 2826 Node* offset = NULL; 2827 Node* oldval = NULL; 2828 Node* newval = NULL; 2829 if (kind == LS_cmpxchg) { 2830 const bool two_slot_type = type2size[type] == 2; 2831 receiver = argument(0); // type: oop 2832 base = argument(1); // type: oop 2833 offset = argument(2); // type: long 2834 oldval = argument(4); // type: oop, int, or long 2835 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2836 } else if (kind == LS_xadd || kind == LS_xchg){ 2837 receiver = argument(0); // type: oop 2838 base = argument(1); // type: oop 2839 offset = argument(2); // type: long 2840 oldval = NULL; 2841 newval = argument(4); // type: oop, int, or long 2842 } 2843 2844 // Null check receiver. 2845 receiver = null_check(receiver); 2846 if (stopped()) { 2847 return true; 2848 } 2849 2850 // Build field offset expression. 2851 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2852 // to be plain byte offsets, which are also the same as those accepted 2853 // by oopDesc::field_base. 2854 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2855 // 32-bit machines ignore the high half of long offsets 2856 offset = ConvL2X(offset); 2857 Node* adr = make_unsafe_address(base, offset); 2858 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2859 2860 // For CAS, unlike inline_unsafe_access, there seems no point in 2861 // trying to refine types. Just use the coarse types here. 2862 const Type *value_type = Type::get_const_basic_type(type); 2863 Compile::AliasType* alias_type = C->alias_type(adr_type); 2864 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2865 2866 if (kind == LS_xchg && type == T_OBJECT) { 2867 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2868 if (tjp != NULL) { 2869 value_type = tjp; 2870 } 2871 } 2872 2873 int alias_idx = C->get_alias_index(adr_type); 2874 2875 // Memory-model-wise, a LoadStore acts like a little synchronized 2876 // block, so needs barriers on each side. These don't translate 2877 // into actual barriers on most machines, but we still need rest of 2878 // compiler to respect ordering. 2879 2880 insert_mem_bar(Op_MemBarRelease); 2881 insert_mem_bar(Op_MemBarCPUOrder); 2882 2883 // 4984716: MemBars must be inserted before this 2884 // memory node in order to avoid a false 2885 // dependency which will confuse the scheduler. 2886 Node *mem = memory(alias_idx); 2887 2888 // For now, we handle only those cases that actually exist: ints, 2889 // longs, and Object. Adding others should be straightforward. 2890 Node* load_store; 2891 switch(type) { 2892 case T_INT: 2893 if (kind == LS_xadd) { 2894 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2895 } else if (kind == LS_xchg) { 2896 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2897 } else if (kind == LS_cmpxchg) { 2898 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2899 } else { 2900 ShouldNotReachHere(); 2901 } 2902 break; 2903 case T_LONG: 2904 if (kind == LS_xadd) { 2905 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2906 } else if (kind == LS_xchg) { 2907 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2908 } else if (kind == LS_cmpxchg) { 2909 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2910 } else { 2911 ShouldNotReachHere(); 2912 } 2913 break; 2914 case T_OBJECT: 2915 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2916 // could be delayed during Parse (for example, in adjust_map_after_if()). 2917 // Execute transformation here to avoid barrier generation in such case. 2918 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2919 newval = _gvn.makecon(TypePtr::NULL_PTR); 2920 2921 // Reference stores need a store barrier. 2922 if (kind == LS_xchg) { 2923 // If pre-barrier must execute before the oop store, old value will require do_load here. 2924 if (!can_move_pre_barrier()) { 2925 pre_barrier(true /* do_load*/, 2926 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2927 NULL /* pre_val*/, 2928 T_OBJECT); 2929 } // Else move pre_barrier to use load_store value, see below. 2930 } else if (kind == LS_cmpxchg) { 2931 // Same as for newval above: 2932 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2933 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2934 } 2935 // The only known value which might get overwritten is oldval. 2936 pre_barrier(false /* do_load */, 2937 control(), NULL, NULL, max_juint, NULL, NULL, 2938 oldval /* pre_val */, 2939 T_OBJECT); 2940 } else { 2941 ShouldNotReachHere(); 2942 } 2943 2944 #ifdef _LP64 2945 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2946 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2947 if (kind == LS_xchg) { 2948 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2949 newval_enc, adr_type, value_type->make_narrowoop())); 2950 } else { 2951 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2952 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2953 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2954 newval_enc, oldval_enc)); 2955 } 2956 } else 2957 #endif 2958 { 2959 if (kind == LS_xchg) { 2960 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2961 } else { 2962 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2963 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2964 } 2965 } 2966 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2967 break; 2968 default: 2969 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2970 break; 2971 } 2972 2973 // SCMemProjNodes represent the memory state of a LoadStore. Their 2974 // main role is to prevent LoadStore nodes from being optimized away 2975 // when their results aren't used. 2976 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2977 set_memory(proj, alias_idx); 2978 2979 if (type == T_OBJECT && kind == LS_xchg) { 2980 #ifdef _LP64 2981 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2982 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2983 } 2984 #endif 2985 if (can_move_pre_barrier()) { 2986 // Don't need to load pre_val. The old value is returned by load_store. 2987 // The pre_barrier can execute after the xchg as long as no safepoint 2988 // gets inserted between them. 2989 pre_barrier(false /* do_load */, 2990 control(), NULL, NULL, max_juint, NULL, NULL, 2991 load_store /* pre_val */, 2992 T_OBJECT); 2993 } 2994 } 2995 2996 // Add the trailing membar surrounding the access 2997 insert_mem_bar(Op_MemBarCPUOrder); 2998 insert_mem_bar(Op_MemBarAcquire); 2999 3000 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3001 set_result(load_store); 3002 return true; 3003 } 3004 3005 //----------------------------inline_unsafe_ordered_store---------------------- 3006 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3007 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3008 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3009 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3010 // This is another variant of inline_unsafe_access, differing in 3011 // that it always issues store-store ("release") barrier and ensures 3012 // store-atomicity (which only matters for "long"). 3013 3014 if (callee()->is_static()) return false; // caller must have the capability! 3015 3016 #ifndef PRODUCT 3017 { 3018 ResourceMark rm; 3019 // Check the signatures. 3020 ciSignature* sig = callee()->signature(); 3021 #ifdef ASSERT 3022 BasicType rtype = sig->return_type()->basic_type(); 3023 assert(rtype == T_VOID, "must return void"); 3024 assert(sig->count() == 3, "has 3 arguments"); 3025 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3026 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3027 #endif // ASSERT 3028 } 3029 #endif //PRODUCT 3030 3031 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3032 3033 // Get arguments: 3034 Node* receiver = argument(0); // type: oop 3035 Node* base = argument(1); // type: oop 3036 Node* offset = argument(2); // type: long 3037 Node* val = argument(4); // type: oop, int, or long 3038 3039 // Null check receiver. 3040 receiver = null_check(receiver); 3041 if (stopped()) { 3042 return true; 3043 } 3044 3045 // Build field offset expression. 3046 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3047 // 32-bit machines ignore the high half of long offsets 3048 offset = ConvL2X(offset); 3049 Node* adr = make_unsafe_address(base, offset); 3050 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3051 const Type *value_type = Type::get_const_basic_type(type); 3052 Compile::AliasType* alias_type = C->alias_type(adr_type); 3053 3054 insert_mem_bar(Op_MemBarRelease); 3055 insert_mem_bar(Op_MemBarCPUOrder); 3056 // Ensure that the store is atomic for longs: 3057 const bool require_atomic_access = true; 3058 Node* store; 3059 if (type == T_OBJECT) // reference stores need a store barrier. 3060 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3061 else { 3062 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3063 } 3064 insert_mem_bar(Op_MemBarCPUOrder); 3065 return true; 3066 } 3067 3068 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3069 // Regardless of form, don't allow previous ld/st to move down, 3070 // then issue acquire, release, or volatile mem_bar. 3071 insert_mem_bar(Op_MemBarCPUOrder); 3072 switch(id) { 3073 case vmIntrinsics::_loadFence: 3074 insert_mem_bar(Op_LoadFence); 3075 return true; 3076 case vmIntrinsics::_storeFence: 3077 insert_mem_bar(Op_StoreFence); 3078 return true; 3079 case vmIntrinsics::_fullFence: 3080 insert_mem_bar(Op_MemBarVolatile); 3081 return true; 3082 default: 3083 fatal_unexpected_iid(id); 3084 return false; 3085 } 3086 } 3087 3088 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3089 if (!kls->is_Con()) { 3090 return true; 3091 } 3092 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3093 if (klsptr == NULL) { 3094 return true; 3095 } 3096 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3097 // don't need a guard for a klass that is already initialized 3098 return !ik->is_initialized(); 3099 } 3100 3101 //----------------------------inline_unsafe_allocate--------------------------- 3102 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3103 bool LibraryCallKit::inline_unsafe_allocate() { 3104 if (callee()->is_static()) return false; // caller must have the capability! 3105 3106 null_check_receiver(); // null-check, then ignore 3107 Node* cls = null_check(argument(1)); 3108 if (stopped()) return true; 3109 3110 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3111 kls = null_check(kls); 3112 if (stopped()) return true; // argument was like int.class 3113 3114 Node* test = NULL; 3115 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3116 // Note: The argument might still be an illegal value like 3117 // Serializable.class or Object[].class. The runtime will handle it. 3118 // But we must make an explicit check for initialization. 3119 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3120 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3121 // can generate code to load it as unsigned byte. 3122 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3123 Node* bits = intcon(InstanceKlass::fully_initialized); 3124 test = _gvn.transform(new SubINode(inst, bits)); 3125 // The 'test' is non-zero if we need to take a slow path. 3126 } 3127 3128 Node* obj = new_instance(kls, test); 3129 set_result(obj); 3130 return true; 3131 } 3132 3133 #ifdef TRACE_HAVE_INTRINSICS 3134 /* 3135 * oop -> myklass 3136 * myklass->trace_id |= USED 3137 * return myklass->trace_id & ~0x3 3138 */ 3139 bool LibraryCallKit::inline_native_classID() { 3140 null_check_receiver(); // null-check, then ignore 3141 Node* cls = null_check(argument(1), T_OBJECT); 3142 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3143 kls = null_check(kls, T_OBJECT); 3144 ByteSize offset = TRACE_ID_OFFSET; 3145 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3146 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3147 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3148 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3149 Node* clsused = longcon(0x01l); // set the class bit 3150 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3151 3152 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3153 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3154 set_result(andl); 3155 return true; 3156 } 3157 3158 bool LibraryCallKit::inline_native_threadID() { 3159 Node* tls_ptr = NULL; 3160 Node* cur_thr = generate_current_thread(tls_ptr); 3161 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3162 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3163 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3164 3165 Node* threadid = NULL; 3166 size_t thread_id_size = OSThread::thread_id_size(); 3167 if (thread_id_size == (size_t) BytesPerLong) { 3168 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3169 } else if (thread_id_size == (size_t) BytesPerInt) { 3170 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3171 } else { 3172 ShouldNotReachHere(); 3173 } 3174 set_result(threadid); 3175 return true; 3176 } 3177 #endif 3178 3179 //------------------------inline_native_time_funcs-------------- 3180 // inline code for System.currentTimeMillis() and System.nanoTime() 3181 // these have the same type and signature 3182 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3183 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3184 const TypePtr* no_memory_effects = NULL; 3185 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3186 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3187 #ifdef ASSERT 3188 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3189 assert(value_top == top(), "second value must be top"); 3190 #endif 3191 set_result(value); 3192 return true; 3193 } 3194 3195 //------------------------inline_native_currentThread------------------ 3196 bool LibraryCallKit::inline_native_currentThread() { 3197 Node* junk = NULL; 3198 set_result(generate_current_thread(junk)); 3199 return true; 3200 } 3201 3202 //------------------------inline_native_isInterrupted------------------ 3203 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3204 bool LibraryCallKit::inline_native_isInterrupted() { 3205 // Add a fast path to t.isInterrupted(clear_int): 3206 // (t == Thread.current() && 3207 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3208 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3209 // So, in the common case that the interrupt bit is false, 3210 // we avoid making a call into the VM. Even if the interrupt bit 3211 // is true, if the clear_int argument is false, we avoid the VM call. 3212 // However, if the receiver is not currentThread, we must call the VM, 3213 // because there must be some locking done around the operation. 3214 3215 // We only go to the fast case code if we pass two guards. 3216 // Paths which do not pass are accumulated in the slow_region. 3217 3218 enum { 3219 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3220 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3221 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3222 PATH_LIMIT 3223 }; 3224 3225 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3226 // out of the function. 3227 insert_mem_bar(Op_MemBarCPUOrder); 3228 3229 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3230 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3231 3232 RegionNode* slow_region = new RegionNode(1); 3233 record_for_igvn(slow_region); 3234 3235 // (a) Receiving thread must be the current thread. 3236 Node* rec_thr = argument(0); 3237 Node* tls_ptr = NULL; 3238 Node* cur_thr = generate_current_thread(tls_ptr); 3239 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3240 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3241 3242 generate_slow_guard(bol_thr, slow_region); 3243 3244 // (b) Interrupt bit on TLS must be false. 3245 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3246 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3247 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3248 3249 // Set the control input on the field _interrupted read to prevent it floating up. 3250 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3251 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3252 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3253 3254 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3255 3256 // First fast path: if (!TLS._interrupted) return false; 3257 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3258 result_rgn->init_req(no_int_result_path, false_bit); 3259 result_val->init_req(no_int_result_path, intcon(0)); 3260 3261 // drop through to next case 3262 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3263 3264 #ifndef TARGET_OS_FAMILY_windows 3265 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3266 Node* clr_arg = argument(1); 3267 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3268 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3269 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3270 3271 // Second fast path: ... else if (!clear_int) return true; 3272 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3273 result_rgn->init_req(no_clear_result_path, false_arg); 3274 result_val->init_req(no_clear_result_path, intcon(1)); 3275 3276 // drop through to next case 3277 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3278 #else 3279 // To return true on Windows you must read the _interrupted field 3280 // and check the the event state i.e. take the slow path. 3281 #endif // TARGET_OS_FAMILY_windows 3282 3283 // (d) Otherwise, go to the slow path. 3284 slow_region->add_req(control()); 3285 set_control( _gvn.transform(slow_region)); 3286 3287 if (stopped()) { 3288 // There is no slow path. 3289 result_rgn->init_req(slow_result_path, top()); 3290 result_val->init_req(slow_result_path, top()); 3291 } else { 3292 // non-virtual because it is a private non-static 3293 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3294 3295 Node* slow_val = set_results_for_java_call(slow_call); 3296 // this->control() comes from set_results_for_java_call 3297 3298 Node* fast_io = slow_call->in(TypeFunc::I_O); 3299 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3300 3301 // These two phis are pre-filled with copies of of the fast IO and Memory 3302 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3303 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3304 3305 result_rgn->init_req(slow_result_path, control()); 3306 result_io ->init_req(slow_result_path, i_o()); 3307 result_mem->init_req(slow_result_path, reset_memory()); 3308 result_val->init_req(slow_result_path, slow_val); 3309 3310 set_all_memory(_gvn.transform(result_mem)); 3311 set_i_o( _gvn.transform(result_io)); 3312 } 3313 3314 C->set_has_split_ifs(true); // Has chance for split-if optimization 3315 set_result(result_rgn, result_val); 3316 return true; 3317 } 3318 3319 //---------------------------load_mirror_from_klass---------------------------- 3320 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3321 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3322 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3323 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3324 } 3325 3326 //-----------------------load_klass_from_mirror_common------------------------- 3327 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3328 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3329 // and branch to the given path on the region. 3330 // If never_see_null, take an uncommon trap on null, so we can optimistically 3331 // compile for the non-null case. 3332 // If the region is NULL, force never_see_null = true. 3333 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3334 bool never_see_null, 3335 RegionNode* region, 3336 int null_path, 3337 int offset) { 3338 if (region == NULL) never_see_null = true; 3339 Node* p = basic_plus_adr(mirror, offset); 3340 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3341 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3342 Node* null_ctl = top(); 3343 kls = null_check_oop(kls, &null_ctl, never_see_null); 3344 if (region != NULL) { 3345 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3346 region->init_req(null_path, null_ctl); 3347 } else { 3348 assert(null_ctl == top(), "no loose ends"); 3349 } 3350 return kls; 3351 } 3352 3353 //--------------------(inline_native_Class_query helpers)--------------------- 3354 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3355 // Fall through if (mods & mask) == bits, take the guard otherwise. 3356 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3357 // Branch around if the given klass has the given modifier bit set. 3358 // Like generate_guard, adds a new path onto the region. 3359 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3360 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3361 Node* mask = intcon(modifier_mask); 3362 Node* bits = intcon(modifier_bits); 3363 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3364 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3365 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3366 return generate_fair_guard(bol, region); 3367 } 3368 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3369 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3370 } 3371 3372 //-------------------------inline_native_Class_query------------------- 3373 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3374 const Type* return_type = TypeInt::BOOL; 3375 Node* prim_return_value = top(); // what happens if it's a primitive class? 3376 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3377 bool expect_prim = false; // most of these guys expect to work on refs 3378 3379 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3380 3381 Node* mirror = argument(0); 3382 Node* obj = top(); 3383 3384 switch (id) { 3385 case vmIntrinsics::_isInstance: 3386 // nothing is an instance of a primitive type 3387 prim_return_value = intcon(0); 3388 obj = argument(1); 3389 break; 3390 case vmIntrinsics::_getModifiers: 3391 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3392 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3393 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3394 break; 3395 case vmIntrinsics::_isInterface: 3396 prim_return_value = intcon(0); 3397 break; 3398 case vmIntrinsics::_isArray: 3399 prim_return_value = intcon(0); 3400 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3401 break; 3402 case vmIntrinsics::_isPrimitive: 3403 prim_return_value = intcon(1); 3404 expect_prim = true; // obviously 3405 break; 3406 case vmIntrinsics::_getSuperclass: 3407 prim_return_value = null(); 3408 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3409 break; 3410 case vmIntrinsics::_getClassAccessFlags: 3411 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3412 return_type = TypeInt::INT; // not bool! 6297094 3413 break; 3414 default: 3415 fatal_unexpected_iid(id); 3416 break; 3417 } 3418 3419 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3420 if (mirror_con == NULL) return false; // cannot happen? 3421 3422 #ifndef PRODUCT 3423 if (C->print_intrinsics() || C->print_inlining()) { 3424 ciType* k = mirror_con->java_mirror_type(); 3425 if (k) { 3426 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3427 k->print_name(); 3428 tty->cr(); 3429 } 3430 } 3431 #endif 3432 3433 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3434 RegionNode* region = new RegionNode(PATH_LIMIT); 3435 record_for_igvn(region); 3436 PhiNode* phi = new PhiNode(region, return_type); 3437 3438 // The mirror will never be null of Reflection.getClassAccessFlags, however 3439 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3440 // if it is. See bug 4774291. 3441 3442 // For Reflection.getClassAccessFlags(), the null check occurs in 3443 // the wrong place; see inline_unsafe_access(), above, for a similar 3444 // situation. 3445 mirror = null_check(mirror); 3446 // If mirror or obj is dead, only null-path is taken. 3447 if (stopped()) return true; 3448 3449 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3450 3451 // Now load the mirror's klass metaobject, and null-check it. 3452 // Side-effects region with the control path if the klass is null. 3453 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3454 // If kls is null, we have a primitive mirror. 3455 phi->init_req(_prim_path, prim_return_value); 3456 if (stopped()) { set_result(region, phi); return true; } 3457 bool safe_for_replace = (region->in(_prim_path) == top()); 3458 3459 Node* p; // handy temp 3460 Node* null_ctl; 3461 3462 // Now that we have the non-null klass, we can perform the real query. 3463 // For constant classes, the query will constant-fold in LoadNode::Value. 3464 Node* query_value = top(); 3465 switch (id) { 3466 case vmIntrinsics::_isInstance: 3467 // nothing is an instance of a primitive type 3468 query_value = gen_instanceof(obj, kls, safe_for_replace); 3469 break; 3470 3471 case vmIntrinsics::_getModifiers: 3472 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3473 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3474 break; 3475 3476 case vmIntrinsics::_isInterface: 3477 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3478 if (generate_interface_guard(kls, region) != NULL) 3479 // A guard was added. If the guard is taken, it was an interface. 3480 phi->add_req(intcon(1)); 3481 // If we fall through, it's a plain class. 3482 query_value = intcon(0); 3483 break; 3484 3485 case vmIntrinsics::_isArray: 3486 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3487 if (generate_array_guard(kls, region) != NULL) 3488 // A guard was added. If the guard is taken, it was an array. 3489 phi->add_req(intcon(1)); 3490 // If we fall through, it's a plain class. 3491 query_value = intcon(0); 3492 break; 3493 3494 case vmIntrinsics::_isPrimitive: 3495 query_value = intcon(0); // "normal" path produces false 3496 break; 3497 3498 case vmIntrinsics::_getSuperclass: 3499 // The rules here are somewhat unfortunate, but we can still do better 3500 // with random logic than with a JNI call. 3501 // Interfaces store null or Object as _super, but must report null. 3502 // Arrays store an intermediate super as _super, but must report Object. 3503 // Other types can report the actual _super. 3504 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3505 if (generate_interface_guard(kls, region) != NULL) 3506 // A guard was added. If the guard is taken, it was an interface. 3507 phi->add_req(null()); 3508 if (generate_array_guard(kls, region) != NULL) 3509 // A guard was added. If the guard is taken, it was an array. 3510 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3511 // If we fall through, it's a plain class. Get its _super. 3512 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3513 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3514 null_ctl = top(); 3515 kls = null_check_oop(kls, &null_ctl); 3516 if (null_ctl != top()) { 3517 // If the guard is taken, Object.superClass is null (both klass and mirror). 3518 region->add_req(null_ctl); 3519 phi ->add_req(null()); 3520 } 3521 if (!stopped()) { 3522 query_value = load_mirror_from_klass(kls); 3523 } 3524 break; 3525 3526 case vmIntrinsics::_getClassAccessFlags: 3527 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3528 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3529 break; 3530 3531 default: 3532 fatal_unexpected_iid(id); 3533 break; 3534 } 3535 3536 // Fall-through is the normal case of a query to a real class. 3537 phi->init_req(1, query_value); 3538 region->init_req(1, control()); 3539 3540 C->set_has_split_ifs(true); // Has chance for split-if optimization 3541 set_result(region, phi); 3542 return true; 3543 } 3544 3545 //-------------------------inline_Class_cast------------------- 3546 bool LibraryCallKit::inline_Class_cast() { 3547 Node* mirror = argument(0); // Class 3548 Node* obj = argument(1); 3549 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3550 if (mirror_con == NULL) { 3551 return false; // dead path (mirror->is_top()). 3552 } 3553 if (obj == NULL || obj->is_top()) { 3554 return false; // dead path 3555 } 3556 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3557 3558 // First, see if Class.cast() can be folded statically. 3559 // java_mirror_type() returns non-null for compile-time Class constants. 3560 ciType* tm = mirror_con->java_mirror_type(); 3561 if (tm != NULL && tm->is_klass() && 3562 tp != NULL && tp->klass() != NULL) { 3563 if (!tp->klass()->is_loaded()) { 3564 // Don't use intrinsic when class is not loaded. 3565 return false; 3566 } else { 3567 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3568 if (static_res == Compile::SSC_always_true) { 3569 // isInstance() is true - fold the code. 3570 set_result(obj); 3571 return true; 3572 } else if (static_res == Compile::SSC_always_false) { 3573 // Don't use intrinsic, have to throw ClassCastException. 3574 // If the reference is null, the non-intrinsic bytecode will 3575 // be optimized appropriately. 3576 return false; 3577 } 3578 } 3579 } 3580 3581 // Bailout intrinsic and do normal inlining if exception path is frequent. 3582 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3583 return false; 3584 } 3585 3586 // Generate dynamic checks. 3587 // Class.cast() is java implementation of _checkcast bytecode. 3588 // Do checkcast (Parse::do_checkcast()) optimizations here. 3589 3590 mirror = null_check(mirror); 3591 // If mirror is dead, only null-path is taken. 3592 if (stopped()) { 3593 return true; 3594 } 3595 3596 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3597 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3598 RegionNode* region = new RegionNode(PATH_LIMIT); 3599 record_for_igvn(region); 3600 3601 // Now load the mirror's klass metaobject, and null-check it. 3602 // If kls is null, we have a primitive mirror and 3603 // nothing is an instance of a primitive type. 3604 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3605 3606 Node* res = top(); 3607 if (!stopped()) { 3608 Node* bad_type_ctrl = top(); 3609 // Do checkcast optimizations. 3610 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3611 region->init_req(_bad_type_path, bad_type_ctrl); 3612 } 3613 if (region->in(_prim_path) != top() || 3614 region->in(_bad_type_path) != top()) { 3615 // Let Interpreter throw ClassCastException. 3616 PreserveJVMState pjvms(this); 3617 set_control(_gvn.transform(region)); 3618 uncommon_trap(Deoptimization::Reason_intrinsic, 3619 Deoptimization::Action_maybe_recompile); 3620 } 3621 if (!stopped()) { 3622 set_result(res); 3623 } 3624 return true; 3625 } 3626 3627 3628 //--------------------------inline_native_subtype_check------------------------ 3629 // This intrinsic takes the JNI calls out of the heart of 3630 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3631 bool LibraryCallKit::inline_native_subtype_check() { 3632 // Pull both arguments off the stack. 3633 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3634 args[0] = argument(0); 3635 args[1] = argument(1); 3636 Node* klasses[2]; // corresponding Klasses: superk, subk 3637 klasses[0] = klasses[1] = top(); 3638 3639 enum { 3640 // A full decision tree on {superc is prim, subc is prim}: 3641 _prim_0_path = 1, // {P,N} => false 3642 // {P,P} & superc!=subc => false 3643 _prim_same_path, // {P,P} & superc==subc => true 3644 _prim_1_path, // {N,P} => false 3645 _ref_subtype_path, // {N,N} & subtype check wins => true 3646 _both_ref_path, // {N,N} & subtype check loses => false 3647 PATH_LIMIT 3648 }; 3649 3650 RegionNode* region = new RegionNode(PATH_LIMIT); 3651 Node* phi = new PhiNode(region, TypeInt::BOOL); 3652 record_for_igvn(region); 3653 3654 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3655 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3656 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3657 3658 // First null-check both mirrors and load each mirror's klass metaobject. 3659 int which_arg; 3660 for (which_arg = 0; which_arg <= 1; which_arg++) { 3661 Node* arg = args[which_arg]; 3662 arg = null_check(arg); 3663 if (stopped()) break; 3664 args[which_arg] = arg; 3665 3666 Node* p = basic_plus_adr(arg, class_klass_offset); 3667 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3668 klasses[which_arg] = _gvn.transform(kls); 3669 } 3670 3671 // Having loaded both klasses, test each for null. 3672 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3673 for (which_arg = 0; which_arg <= 1; which_arg++) { 3674 Node* kls = klasses[which_arg]; 3675 Node* null_ctl = top(); 3676 kls = null_check_oop(kls, &null_ctl, never_see_null); 3677 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3678 region->init_req(prim_path, null_ctl); 3679 if (stopped()) break; 3680 klasses[which_arg] = kls; 3681 } 3682 3683 if (!stopped()) { 3684 // now we have two reference types, in klasses[0..1] 3685 Node* subk = klasses[1]; // the argument to isAssignableFrom 3686 Node* superk = klasses[0]; // the receiver 3687 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3688 // now we have a successful reference subtype check 3689 region->set_req(_ref_subtype_path, control()); 3690 } 3691 3692 // If both operands are primitive (both klasses null), then 3693 // we must return true when they are identical primitives. 3694 // It is convenient to test this after the first null klass check. 3695 set_control(region->in(_prim_0_path)); // go back to first null check 3696 if (!stopped()) { 3697 // Since superc is primitive, make a guard for the superc==subc case. 3698 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3699 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3700 generate_guard(bol_eq, region, PROB_FAIR); 3701 if (region->req() == PATH_LIMIT+1) { 3702 // A guard was added. If the added guard is taken, superc==subc. 3703 region->swap_edges(PATH_LIMIT, _prim_same_path); 3704 region->del_req(PATH_LIMIT); 3705 } 3706 region->set_req(_prim_0_path, control()); // Not equal after all. 3707 } 3708 3709 // these are the only paths that produce 'true': 3710 phi->set_req(_prim_same_path, intcon(1)); 3711 phi->set_req(_ref_subtype_path, intcon(1)); 3712 3713 // pull together the cases: 3714 assert(region->req() == PATH_LIMIT, "sane region"); 3715 for (uint i = 1; i < region->req(); i++) { 3716 Node* ctl = region->in(i); 3717 if (ctl == NULL || ctl == top()) { 3718 region->set_req(i, top()); 3719 phi ->set_req(i, top()); 3720 } else if (phi->in(i) == NULL) { 3721 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3722 } 3723 } 3724 3725 set_control(_gvn.transform(region)); 3726 set_result(_gvn.transform(phi)); 3727 return true; 3728 } 3729 3730 //---------------------generate_array_guard_common------------------------ 3731 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3732 bool obj_array, bool not_array) { 3733 3734 if (stopped()) { 3735 return NULL; 3736 } 3737 3738 // If obj_array/non_array==false/false: 3739 // Branch around if the given klass is in fact an array (either obj or prim). 3740 // If obj_array/non_array==false/true: 3741 // Branch around if the given klass is not an array klass of any kind. 3742 // If obj_array/non_array==true/true: 3743 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3744 // If obj_array/non_array==true/false: 3745 // Branch around if the kls is an oop array (Object[] or subtype) 3746 // 3747 // Like generate_guard, adds a new path onto the region. 3748 jint layout_con = 0; 3749 Node* layout_val = get_layout_helper(kls, layout_con); 3750 if (layout_val == NULL) { 3751 bool query = (obj_array 3752 ? Klass::layout_helper_is_objArray(layout_con) 3753 : Klass::layout_helper_is_array(layout_con)); 3754 if (query == not_array) { 3755 return NULL; // never a branch 3756 } else { // always a branch 3757 Node* always_branch = control(); 3758 if (region != NULL) 3759 region->add_req(always_branch); 3760 set_control(top()); 3761 return always_branch; 3762 } 3763 } 3764 // Now test the correct condition. 3765 jint nval = (obj_array 3766 ? ((jint)Klass::_lh_array_tag_type_value 3767 << Klass::_lh_array_tag_shift) 3768 : Klass::_lh_neutral_value); 3769 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3770 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3771 // invert the test if we are looking for a non-array 3772 if (not_array) btest = BoolTest(btest).negate(); 3773 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3774 return generate_fair_guard(bol, region); 3775 } 3776 3777 3778 //-----------------------inline_native_newArray-------------------------- 3779 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3780 bool LibraryCallKit::inline_native_newArray() { 3781 Node* mirror = argument(0); 3782 Node* count_val = argument(1); 3783 3784 mirror = null_check(mirror); 3785 // If mirror or obj is dead, only null-path is taken. 3786 if (stopped()) return true; 3787 3788 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3789 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3790 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3791 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3792 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3793 3794 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3795 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3796 result_reg, _slow_path); 3797 Node* normal_ctl = control(); 3798 Node* no_array_ctl = result_reg->in(_slow_path); 3799 3800 // Generate code for the slow case. We make a call to newArray(). 3801 set_control(no_array_ctl); 3802 if (!stopped()) { 3803 // Either the input type is void.class, or else the 3804 // array klass has not yet been cached. Either the 3805 // ensuing call will throw an exception, or else it 3806 // will cache the array klass for next time. 3807 PreserveJVMState pjvms(this); 3808 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3809 Node* slow_result = set_results_for_java_call(slow_call); 3810 // this->control() comes from set_results_for_java_call 3811 result_reg->set_req(_slow_path, control()); 3812 result_val->set_req(_slow_path, slow_result); 3813 result_io ->set_req(_slow_path, i_o()); 3814 result_mem->set_req(_slow_path, reset_memory()); 3815 } 3816 3817 set_control(normal_ctl); 3818 if (!stopped()) { 3819 // Normal case: The array type has been cached in the java.lang.Class. 3820 // The following call works fine even if the array type is polymorphic. 3821 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3822 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3823 result_reg->init_req(_normal_path, control()); 3824 result_val->init_req(_normal_path, obj); 3825 result_io ->init_req(_normal_path, i_o()); 3826 result_mem->init_req(_normal_path, reset_memory()); 3827 } 3828 3829 // Return the combined state. 3830 set_i_o( _gvn.transform(result_io) ); 3831 set_all_memory( _gvn.transform(result_mem)); 3832 3833 C->set_has_split_ifs(true); // Has chance for split-if optimization 3834 set_result(result_reg, result_val); 3835 return true; 3836 } 3837 3838 //----------------------inline_native_getLength-------------------------- 3839 // public static native int java.lang.reflect.Array.getLength(Object array); 3840 bool LibraryCallKit::inline_native_getLength() { 3841 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3842 3843 Node* array = null_check(argument(0)); 3844 // If array is dead, only null-path is taken. 3845 if (stopped()) return true; 3846 3847 // Deoptimize if it is a non-array. 3848 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3849 3850 if (non_array != NULL) { 3851 PreserveJVMState pjvms(this); 3852 set_control(non_array); 3853 uncommon_trap(Deoptimization::Reason_intrinsic, 3854 Deoptimization::Action_maybe_recompile); 3855 } 3856 3857 // If control is dead, only non-array-path is taken. 3858 if (stopped()) return true; 3859 3860 // The works fine even if the array type is polymorphic. 3861 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3862 Node* result = load_array_length(array); 3863 3864 C->set_has_split_ifs(true); // Has chance for split-if optimization 3865 set_result(result); 3866 return true; 3867 } 3868 3869 //------------------------inline_array_copyOf---------------------------- 3870 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3871 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3872 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3873 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3874 3875 // Get the arguments. 3876 Node* original = argument(0); 3877 Node* start = is_copyOfRange? argument(1): intcon(0); 3878 Node* end = is_copyOfRange? argument(2): argument(1); 3879 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3880 3881 Node* newcopy; 3882 3883 // Set the original stack and the reexecute bit for the interpreter to reexecute 3884 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3885 { PreserveReexecuteState preexecs(this); 3886 jvms()->set_should_reexecute(true); 3887 3888 array_type_mirror = null_check(array_type_mirror); 3889 original = null_check(original); 3890 3891 // Check if a null path was taken unconditionally. 3892 if (stopped()) return true; 3893 3894 Node* orig_length = load_array_length(original); 3895 3896 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3897 klass_node = null_check(klass_node); 3898 3899 RegionNode* bailout = new RegionNode(1); 3900 record_for_igvn(bailout); 3901 3902 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3903 // Bail out if that is so. 3904 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3905 if (not_objArray != NULL) { 3906 // Improve the klass node's type from the new optimistic assumption: 3907 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3908 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3909 Node* cast = new CastPPNode(klass_node, akls); 3910 cast->init_req(0, control()); 3911 klass_node = _gvn.transform(cast); 3912 } 3913 3914 // Bail out if either start or end is negative. 3915 generate_negative_guard(start, bailout, &start); 3916 generate_negative_guard(end, bailout, &end); 3917 3918 Node* length = end; 3919 if (_gvn.type(start) != TypeInt::ZERO) { 3920 length = _gvn.transform(new SubINode(end, start)); 3921 } 3922 3923 // Bail out if length is negative. 3924 // Without this the new_array would throw 3925 // NegativeArraySizeException but IllegalArgumentException is what 3926 // should be thrown 3927 generate_negative_guard(length, bailout, &length); 3928 3929 if (bailout->req() > 1) { 3930 PreserveJVMState pjvms(this); 3931 set_control(_gvn.transform(bailout)); 3932 uncommon_trap(Deoptimization::Reason_intrinsic, 3933 Deoptimization::Action_maybe_recompile); 3934 } 3935 3936 if (!stopped()) { 3937 // How many elements will we copy from the original? 3938 // The answer is MinI(orig_length - start, length). 3939 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3940 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3941 3942 // Generate a direct call to the right arraycopy function(s). 3943 // We know the copy is disjoint but we might not know if the 3944 // oop stores need checking. 3945 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3946 // This will fail a store-check if x contains any non-nulls. 3947 3948 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3949 // loads/stores but it is legal only if we're sure the 3950 // Arrays.copyOf would succeed. So we need all input arguments 3951 // to the copyOf to be validated, including that the copy to the 3952 // new array won't trigger an ArrayStoreException. That subtype 3953 // check can be optimized if we know something on the type of 3954 // the input array from type speculation. 3955 if (_gvn.type(klass_node)->singleton()) { 3956 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3957 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3958 3959 int test = C->static_subtype_check(superk, subk); 3960 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3961 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3962 if (t_original->speculative_type() != NULL) { 3963 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3964 } 3965 } 3966 } 3967 3968 bool validated = false; 3969 // Reason_class_check rather than Reason_intrinsic because we 3970 // want to intrinsify even if this traps. 3971 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3972 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3973 klass_node); 3974 3975 if (not_subtype_ctrl != top()) { 3976 PreserveJVMState pjvms(this); 3977 set_control(not_subtype_ctrl); 3978 uncommon_trap(Deoptimization::Reason_class_check, 3979 Deoptimization::Action_make_not_entrant); 3980 assert(stopped(), "Should be stopped"); 3981 } 3982 validated = true; 3983 } 3984 3985 if (!stopped()) { 3986 newcopy = new_array(klass_node, length, 0); // no arguments to push 3987 3988 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 3989 load_object_klass(original), klass_node); 3990 if (!is_copyOfRange) { 3991 ac->set_copyof(validated); 3992 } else { 3993 ac->set_copyofrange(validated); 3994 } 3995 Node* n = _gvn.transform(ac); 3996 if (n == ac) { 3997 ac->connect_outputs(this); 3998 } else { 3999 assert(validated, "shouldn't transform if all arguments not validated"); 4000 set_all_memory(n); 4001 } 4002 } 4003 } 4004 } // original reexecute is set back here 4005 4006 C->set_has_split_ifs(true); // Has chance for split-if optimization 4007 if (!stopped()) { 4008 set_result(newcopy); 4009 } 4010 return true; 4011 } 4012 4013 4014 //----------------------generate_virtual_guard--------------------------- 4015 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4016 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4017 RegionNode* slow_region) { 4018 ciMethod* method = callee(); 4019 int vtable_index = method->vtable_index(); 4020 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4021 err_msg_res("bad index %d", vtable_index)); 4022 // Get the Method* out of the appropriate vtable entry. 4023 int entry_offset = (InstanceKlass::vtable_start_offset() + 4024 vtable_index*vtableEntry::size()) * wordSize + 4025 vtableEntry::method_offset_in_bytes(); 4026 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4027 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4028 4029 // Compare the target method with the expected method (e.g., Object.hashCode). 4030 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4031 4032 Node* native_call = makecon(native_call_addr); 4033 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4034 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4035 4036 return generate_slow_guard(test_native, slow_region); 4037 } 4038 4039 //-----------------------generate_method_call---------------------------- 4040 // Use generate_method_call to make a slow-call to the real 4041 // method if the fast path fails. An alternative would be to 4042 // use a stub like OptoRuntime::slow_arraycopy_Java. 4043 // This only works for expanding the current library call, 4044 // not another intrinsic. (E.g., don't use this for making an 4045 // arraycopy call inside of the copyOf intrinsic.) 4046 CallJavaNode* 4047 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4048 // When compiling the intrinsic method itself, do not use this technique. 4049 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4050 4051 ciMethod* method = callee(); 4052 // ensure the JVMS we have will be correct for this call 4053 guarantee(method_id == method->intrinsic_id(), "must match"); 4054 4055 const TypeFunc* tf = TypeFunc::make(method); 4056 CallJavaNode* slow_call; 4057 if (is_static) { 4058 assert(!is_virtual, ""); 4059 slow_call = new CallStaticJavaNode(C, tf, 4060 SharedRuntime::get_resolve_static_call_stub(), 4061 method, bci()); 4062 } else if (is_virtual) { 4063 null_check_receiver(); 4064 int vtable_index = Method::invalid_vtable_index; 4065 if (UseInlineCaches) { 4066 // Suppress the vtable call 4067 } else { 4068 // hashCode and clone are not a miranda methods, 4069 // so the vtable index is fixed. 4070 // No need to use the linkResolver to get it. 4071 vtable_index = method->vtable_index(); 4072 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4073 err_msg_res("bad index %d", vtable_index)); 4074 } 4075 slow_call = new CallDynamicJavaNode(tf, 4076 SharedRuntime::get_resolve_virtual_call_stub(), 4077 method, vtable_index, bci()); 4078 } else { // neither virtual nor static: opt_virtual 4079 null_check_receiver(); 4080 slow_call = new CallStaticJavaNode(C, tf, 4081 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4082 method, bci()); 4083 slow_call->set_optimized_virtual(true); 4084 } 4085 set_arguments_for_java_call(slow_call); 4086 set_edges_for_java_call(slow_call); 4087 return slow_call; 4088 } 4089 4090 4091 /** 4092 * Build special case code for calls to hashCode on an object. This call may 4093 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4094 * slightly different code. 4095 */ 4096 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4097 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4098 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4099 4100 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4101 4102 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4103 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4104 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4105 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4106 Node* obj = NULL; 4107 if (!is_static) { 4108 // Check for hashing null object 4109 obj = null_check_receiver(); 4110 if (stopped()) return true; // unconditionally null 4111 result_reg->init_req(_null_path, top()); 4112 result_val->init_req(_null_path, top()); 4113 } else { 4114 // Do a null check, and return zero if null. 4115 // System.identityHashCode(null) == 0 4116 obj = argument(0); 4117 Node* null_ctl = top(); 4118 obj = null_check_oop(obj, &null_ctl); 4119 result_reg->init_req(_null_path, null_ctl); 4120 result_val->init_req(_null_path, _gvn.intcon(0)); 4121 } 4122 4123 // Unconditionally null? Then return right away. 4124 if (stopped()) { 4125 set_control( result_reg->in(_null_path)); 4126 if (!stopped()) 4127 set_result(result_val->in(_null_path)); 4128 return true; 4129 } 4130 4131 // We only go to the fast case code if we pass a number of guards. The 4132 // paths which do not pass are accumulated in the slow_region. 4133 RegionNode* slow_region = new RegionNode(1); 4134 record_for_igvn(slow_region); 4135 4136 // If this is a virtual call, we generate a funny guard. We pull out 4137 // the vtable entry corresponding to hashCode() from the target object. 4138 // If the target method which we are calling happens to be the native 4139 // Object hashCode() method, we pass the guard. We do not need this 4140 // guard for non-virtual calls -- the caller is known to be the native 4141 // Object hashCode(). 4142 if (is_virtual) { 4143 // After null check, get the object's klass. 4144 Node* obj_klass = load_object_klass(obj); 4145 generate_virtual_guard(obj_klass, slow_region); 4146 } 4147 4148 // Get the header out of the object, use LoadMarkNode when available 4149 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4150 // The control of the load must be NULL. Otherwise, the load can move before 4151 // the null check after castPP removal. 4152 Node* no_ctrl = NULL; 4153 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4154 4155 // Test the header to see if it is unlocked. 4156 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4157 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4158 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4159 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4160 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4161 4162 generate_slow_guard(test_unlocked, slow_region); 4163 4164 // Get the hash value and check to see that it has been properly assigned. 4165 // We depend on hash_mask being at most 32 bits and avoid the use of 4166 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4167 // vm: see markOop.hpp. 4168 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4169 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4170 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4171 // This hack lets the hash bits live anywhere in the mark object now, as long 4172 // as the shift drops the relevant bits into the low 32 bits. Note that 4173 // Java spec says that HashCode is an int so there's no point in capturing 4174 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4175 hshifted_header = ConvX2I(hshifted_header); 4176 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4177 4178 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4179 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4180 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4181 4182 generate_slow_guard(test_assigned, slow_region); 4183 4184 Node* init_mem = reset_memory(); 4185 // fill in the rest of the null path: 4186 result_io ->init_req(_null_path, i_o()); 4187 result_mem->init_req(_null_path, init_mem); 4188 4189 result_val->init_req(_fast_path, hash_val); 4190 result_reg->init_req(_fast_path, control()); 4191 result_io ->init_req(_fast_path, i_o()); 4192 result_mem->init_req(_fast_path, init_mem); 4193 4194 // Generate code for the slow case. We make a call to hashCode(). 4195 set_control(_gvn.transform(slow_region)); 4196 if (!stopped()) { 4197 // No need for PreserveJVMState, because we're using up the present state. 4198 set_all_memory(init_mem); 4199 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4200 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4201 Node* slow_result = set_results_for_java_call(slow_call); 4202 // this->control() comes from set_results_for_java_call 4203 result_reg->init_req(_slow_path, control()); 4204 result_val->init_req(_slow_path, slow_result); 4205 result_io ->set_req(_slow_path, i_o()); 4206 result_mem ->set_req(_slow_path, reset_memory()); 4207 } 4208 4209 // Return the combined state. 4210 set_i_o( _gvn.transform(result_io) ); 4211 set_all_memory( _gvn.transform(result_mem)); 4212 4213 set_result(result_reg, result_val); 4214 return true; 4215 } 4216 4217 //---------------------------inline_native_getClass---------------------------- 4218 // public final native Class<?> java.lang.Object.getClass(); 4219 // 4220 // Build special case code for calls to getClass on an object. 4221 bool LibraryCallKit::inline_native_getClass() { 4222 Node* obj = null_check_receiver(); 4223 if (stopped()) return true; 4224 set_result(load_mirror_from_klass(load_object_klass(obj))); 4225 return true; 4226 } 4227 4228 //-----------------inline_native_Reflection_getCallerClass--------------------- 4229 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4230 // 4231 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4232 // 4233 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4234 // in that it must skip particular security frames and checks for 4235 // caller sensitive methods. 4236 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4237 #ifndef PRODUCT 4238 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4239 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4240 } 4241 #endif 4242 4243 if (!jvms()->has_method()) { 4244 #ifndef PRODUCT 4245 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4246 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4247 } 4248 #endif 4249 return false; 4250 } 4251 4252 // Walk back up the JVM state to find the caller at the required 4253 // depth. 4254 JVMState* caller_jvms = jvms(); 4255 4256 // Cf. JVM_GetCallerClass 4257 // NOTE: Start the loop at depth 1 because the current JVM state does 4258 // not include the Reflection.getCallerClass() frame. 4259 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4260 ciMethod* m = caller_jvms->method(); 4261 switch (n) { 4262 case 0: 4263 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4264 break; 4265 case 1: 4266 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4267 if (!m->caller_sensitive()) { 4268 #ifndef PRODUCT 4269 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4270 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4271 } 4272 #endif 4273 return false; // bail-out; let JVM_GetCallerClass do the work 4274 } 4275 break; 4276 default: 4277 if (!m->is_ignored_by_security_stack_walk()) { 4278 // We have reached the desired frame; return the holder class. 4279 // Acquire method holder as java.lang.Class and push as constant. 4280 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4281 ciInstance* caller_mirror = caller_klass->java_mirror(); 4282 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4283 4284 #ifndef PRODUCT 4285 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4286 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()); 4287 tty->print_cr(" JVM state at this point:"); 4288 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4289 ciMethod* m = jvms()->of_depth(i)->method(); 4290 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4291 } 4292 } 4293 #endif 4294 return true; 4295 } 4296 break; 4297 } 4298 } 4299 4300 #ifndef PRODUCT 4301 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4302 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4303 tty->print_cr(" JVM state at this point:"); 4304 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4305 ciMethod* m = jvms()->of_depth(i)->method(); 4306 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4307 } 4308 } 4309 #endif 4310 4311 return false; // bail-out; let JVM_GetCallerClass do the work 4312 } 4313 4314 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4315 Node* arg = argument(0); 4316 Node* result; 4317 4318 switch (id) { 4319 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4320 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4321 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4322 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4323 4324 case vmIntrinsics::_doubleToLongBits: { 4325 // two paths (plus control) merge in a wood 4326 RegionNode *r = new RegionNode(3); 4327 Node *phi = new PhiNode(r, TypeLong::LONG); 4328 4329 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4330 // Build the boolean node 4331 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4332 4333 // Branch either way. 4334 // NaN case is less traveled, which makes all the difference. 4335 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4336 Node *opt_isnan = _gvn.transform(ifisnan); 4337 assert( opt_isnan->is_If(), "Expect an IfNode"); 4338 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4339 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4340 4341 set_control(iftrue); 4342 4343 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4344 Node *slow_result = longcon(nan_bits); // return NaN 4345 phi->init_req(1, _gvn.transform( slow_result )); 4346 r->init_req(1, iftrue); 4347 4348 // Else fall through 4349 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4350 set_control(iffalse); 4351 4352 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4353 r->init_req(2, iffalse); 4354 4355 // Post merge 4356 set_control(_gvn.transform(r)); 4357 record_for_igvn(r); 4358 4359 C->set_has_split_ifs(true); // Has chance for split-if optimization 4360 result = phi; 4361 assert(result->bottom_type()->isa_long(), "must be"); 4362 break; 4363 } 4364 4365 case vmIntrinsics::_floatToIntBits: { 4366 // two paths (plus control) merge in a wood 4367 RegionNode *r = new RegionNode(3); 4368 Node *phi = new PhiNode(r, TypeInt::INT); 4369 4370 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4371 // Build the boolean node 4372 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4373 4374 // Branch either way. 4375 // NaN case is less traveled, which makes all the difference. 4376 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4377 Node *opt_isnan = _gvn.transform(ifisnan); 4378 assert( opt_isnan->is_If(), "Expect an IfNode"); 4379 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4380 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4381 4382 set_control(iftrue); 4383 4384 static const jint nan_bits = 0x7fc00000; 4385 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4386 phi->init_req(1, _gvn.transform( slow_result )); 4387 r->init_req(1, iftrue); 4388 4389 // Else fall through 4390 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4391 set_control(iffalse); 4392 4393 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4394 r->init_req(2, iffalse); 4395 4396 // Post merge 4397 set_control(_gvn.transform(r)); 4398 record_for_igvn(r); 4399 4400 C->set_has_split_ifs(true); // Has chance for split-if optimization 4401 result = phi; 4402 assert(result->bottom_type()->isa_int(), "must be"); 4403 break; 4404 } 4405 4406 default: 4407 fatal_unexpected_iid(id); 4408 break; 4409 } 4410 set_result(_gvn.transform(result)); 4411 return true; 4412 } 4413 4414 #ifdef _LP64 4415 #define XTOP ,top() /*additional argument*/ 4416 #else //_LP64 4417 #define XTOP /*no additional argument*/ 4418 #endif //_LP64 4419 4420 //----------------------inline_unsafe_copyMemory------------------------- 4421 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4422 bool LibraryCallKit::inline_unsafe_copyMemory() { 4423 if (callee()->is_static()) return false; // caller must have the capability! 4424 null_check_receiver(); // null-check receiver 4425 if (stopped()) return true; 4426 4427 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4428 4429 Node* src_ptr = argument(1); // type: oop 4430 Node* src_off = ConvL2X(argument(2)); // type: long 4431 Node* dst_ptr = argument(4); // type: oop 4432 Node* dst_off = ConvL2X(argument(5)); // type: long 4433 Node* size = ConvL2X(argument(7)); // type: long 4434 4435 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4436 "fieldOffset must be byte-scaled"); 4437 4438 Node* src = make_unsafe_address(src_ptr, src_off); 4439 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4440 4441 // Conservatively insert a memory barrier on all memory slices. 4442 // Do not let writes of the copy source or destination float below the copy. 4443 insert_mem_bar(Op_MemBarCPUOrder); 4444 4445 // Call it. Note that the length argument is not scaled. 4446 make_runtime_call(RC_LEAF|RC_NO_FP, 4447 OptoRuntime::fast_arraycopy_Type(), 4448 StubRoutines::unsafe_arraycopy(), 4449 "unsafe_arraycopy", 4450 TypeRawPtr::BOTTOM, 4451 src, dst, size XTOP); 4452 4453 // Do not let reads of the copy destination float above the copy. 4454 insert_mem_bar(Op_MemBarCPUOrder); 4455 4456 return true; 4457 } 4458 4459 //------------------------clone_coping----------------------------------- 4460 // Helper function for inline_native_clone. 4461 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4462 assert(obj_size != NULL, ""); 4463 Node* raw_obj = alloc_obj->in(1); 4464 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4465 4466 AllocateNode* alloc = NULL; 4467 if (ReduceBulkZeroing) { 4468 // We will be completely responsible for initializing this object - 4469 // mark Initialize node as complete. 4470 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4471 // The object was just allocated - there should be no any stores! 4472 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4473 // Mark as complete_with_arraycopy so that on AllocateNode 4474 // expansion, we know this AllocateNode is initialized by an array 4475 // copy and a StoreStore barrier exists after the array copy. 4476 alloc->initialization()->set_complete_with_arraycopy(); 4477 } 4478 4479 // Copy the fastest available way. 4480 // TODO: generate fields copies for small objects instead. 4481 Node* src = obj; 4482 Node* dest = alloc_obj; 4483 Node* size = _gvn.transform(obj_size); 4484 4485 // Exclude the header but include array length to copy by 8 bytes words. 4486 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4487 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4488 instanceOopDesc::base_offset_in_bytes(); 4489 // base_off: 4490 // 8 - 32-bit VM 4491 // 12 - 64-bit VM, compressed klass 4492 // 16 - 64-bit VM, normal klass 4493 if (base_off % BytesPerLong != 0) { 4494 assert(UseCompressedClassPointers, ""); 4495 if (is_array) { 4496 // Exclude length to copy by 8 bytes words. 4497 base_off += sizeof(int); 4498 } else { 4499 // Include klass to copy by 8 bytes words. 4500 base_off = instanceOopDesc::klass_offset_in_bytes(); 4501 } 4502 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4503 } 4504 src = basic_plus_adr(src, base_off); 4505 dest = basic_plus_adr(dest, base_off); 4506 4507 // Compute the length also, if needed: 4508 Node* countx = size; 4509 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4510 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4511 4512 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4513 4514 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4515 ac->set_clonebasic(); 4516 Node* n = _gvn.transform(ac); 4517 if (n == ac) { 4518 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4519 } else { 4520 set_all_memory(n); 4521 } 4522 4523 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4524 if (card_mark) { 4525 assert(!is_array, ""); 4526 // Put in store barrier for any and all oops we are sticking 4527 // into this object. (We could avoid this if we could prove 4528 // that the object type contains no oop fields at all.) 4529 Node* no_particular_value = NULL; 4530 Node* no_particular_field = NULL; 4531 int raw_adr_idx = Compile::AliasIdxRaw; 4532 post_barrier(control(), 4533 memory(raw_adr_type), 4534 alloc_obj, 4535 no_particular_field, 4536 raw_adr_idx, 4537 no_particular_value, 4538 T_OBJECT, 4539 false); 4540 } 4541 4542 // Do not let reads from the cloned object float above the arraycopy. 4543 if (alloc != NULL) { 4544 // Do not let stores that initialize this object be reordered with 4545 // a subsequent store that would make this object accessible by 4546 // other threads. 4547 // Record what AllocateNode this StoreStore protects so that 4548 // escape analysis can go from the MemBarStoreStoreNode to the 4549 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4550 // based on the escape status of the AllocateNode. 4551 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4552 } else { 4553 insert_mem_bar(Op_MemBarCPUOrder); 4554 } 4555 } 4556 4557 //------------------------inline_native_clone---------------------------- 4558 // protected native Object java.lang.Object.clone(); 4559 // 4560 // Here are the simple edge cases: 4561 // null receiver => normal trap 4562 // virtual and clone was overridden => slow path to out-of-line clone 4563 // not cloneable or finalizer => slow path to out-of-line Object.clone 4564 // 4565 // The general case has two steps, allocation and copying. 4566 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4567 // 4568 // Copying also has two cases, oop arrays and everything else. 4569 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4570 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4571 // 4572 // These steps fold up nicely if and when the cloned object's klass 4573 // can be sharply typed as an object array, a type array, or an instance. 4574 // 4575 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4576 PhiNode* result_val; 4577 4578 // Set the reexecute bit for the interpreter to reexecute 4579 // the bytecode that invokes Object.clone if deoptimization happens. 4580 { PreserveReexecuteState preexecs(this); 4581 jvms()->set_should_reexecute(true); 4582 4583 Node* obj = null_check_receiver(); 4584 if (stopped()) return true; 4585 4586 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4587 4588 // If we are going to clone an instance, we need its exact type to 4589 // know the number and types of fields to convert the clone to 4590 // loads/stores. Maybe a speculative type can help us. 4591 if (!obj_type->klass_is_exact() && 4592 obj_type->speculative_type() != NULL && 4593 obj_type->speculative_type()->is_instance_klass()) { 4594 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4595 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4596 !spec_ik->has_injected_fields()) { 4597 ciKlass* k = obj_type->klass(); 4598 if (!k->is_instance_klass() || 4599 k->as_instance_klass()->is_interface() || 4600 k->as_instance_klass()->has_subklass()) { 4601 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4602 } 4603 } 4604 } 4605 4606 Node* obj_klass = load_object_klass(obj); 4607 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4608 const TypeOopPtr* toop = ((tklass != NULL) 4609 ? tklass->as_instance_type() 4610 : TypeInstPtr::NOTNULL); 4611 4612 // Conservatively insert a memory barrier on all memory slices. 4613 // Do not let writes into the original float below the clone. 4614 insert_mem_bar(Op_MemBarCPUOrder); 4615 4616 // paths into result_reg: 4617 enum { 4618 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4619 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4620 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4621 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4622 PATH_LIMIT 4623 }; 4624 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4625 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4626 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4627 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4628 record_for_igvn(result_reg); 4629 4630 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4631 int raw_adr_idx = Compile::AliasIdxRaw; 4632 4633 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4634 if (array_ctl != NULL) { 4635 // It's an array. 4636 PreserveJVMState pjvms(this); 4637 set_control(array_ctl); 4638 Node* obj_length = load_array_length(obj); 4639 Node* obj_size = NULL; 4640 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4641 4642 if (!use_ReduceInitialCardMarks()) { 4643 // If it is an oop array, it requires very special treatment, 4644 // because card marking is required on each card of the array. 4645 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4646 if (is_obja != NULL) { 4647 PreserveJVMState pjvms2(this); 4648 set_control(is_obja); 4649 // Generate a direct call to the right arraycopy function(s). 4650 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4651 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4652 ac->set_cloneoop(); 4653 Node* n = _gvn.transform(ac); 4654 assert(n == ac, "cannot disappear"); 4655 ac->connect_outputs(this); 4656 4657 result_reg->init_req(_objArray_path, control()); 4658 result_val->init_req(_objArray_path, alloc_obj); 4659 result_i_o ->set_req(_objArray_path, i_o()); 4660 result_mem ->set_req(_objArray_path, reset_memory()); 4661 } 4662 } 4663 // Otherwise, there are no card marks to worry about. 4664 // (We can dispense with card marks if we know the allocation 4665 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4666 // causes the non-eden paths to take compensating steps to 4667 // simulate a fresh allocation, so that no further 4668 // card marks are required in compiled code to initialize 4669 // the object.) 4670 4671 if (!stopped()) { 4672 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4673 4674 // Present the results of the copy. 4675 result_reg->init_req(_array_path, control()); 4676 result_val->init_req(_array_path, alloc_obj); 4677 result_i_o ->set_req(_array_path, i_o()); 4678 result_mem ->set_req(_array_path, reset_memory()); 4679 } 4680 } 4681 4682 // We only go to the instance fast case code if we pass a number of guards. 4683 // The paths which do not pass are accumulated in the slow_region. 4684 RegionNode* slow_region = new RegionNode(1); 4685 record_for_igvn(slow_region); 4686 if (!stopped()) { 4687 // It's an instance (we did array above). Make the slow-path tests. 4688 // If this is a virtual call, we generate a funny guard. We grab 4689 // the vtable entry corresponding to clone() from the target object. 4690 // If the target method which we are calling happens to be the 4691 // Object clone() method, we pass the guard. We do not need this 4692 // guard for non-virtual calls; the caller is known to be the native 4693 // Object clone(). 4694 if (is_virtual) { 4695 generate_virtual_guard(obj_klass, slow_region); 4696 } 4697 4698 // The object must be cloneable and must not have a finalizer. 4699 // Both of these conditions may be checked in a single test. 4700 // We could optimize the cloneable test further, but we don't care. 4701 generate_access_flags_guard(obj_klass, 4702 // Test both conditions: 4703 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4704 // Must be cloneable but not finalizer: 4705 JVM_ACC_IS_CLONEABLE, 4706 slow_region); 4707 } 4708 4709 if (!stopped()) { 4710 // It's an instance, and it passed the slow-path tests. 4711 PreserveJVMState pjvms(this); 4712 Node* obj_size = NULL; 4713 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4714 // is reexecuted if deoptimization occurs and there could be problems when merging 4715 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4716 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4717 4718 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4719 4720 // Present the results of the slow call. 4721 result_reg->init_req(_instance_path, control()); 4722 result_val->init_req(_instance_path, alloc_obj); 4723 result_i_o ->set_req(_instance_path, i_o()); 4724 result_mem ->set_req(_instance_path, reset_memory()); 4725 } 4726 4727 // Generate code for the slow case. We make a call to clone(). 4728 set_control(_gvn.transform(slow_region)); 4729 if (!stopped()) { 4730 PreserveJVMState pjvms(this); 4731 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4732 Node* slow_result = set_results_for_java_call(slow_call); 4733 // this->control() comes from set_results_for_java_call 4734 result_reg->init_req(_slow_path, control()); 4735 result_val->init_req(_slow_path, slow_result); 4736 result_i_o ->set_req(_slow_path, i_o()); 4737 result_mem ->set_req(_slow_path, reset_memory()); 4738 } 4739 4740 // Return the combined state. 4741 set_control( _gvn.transform(result_reg)); 4742 set_i_o( _gvn.transform(result_i_o)); 4743 set_all_memory( _gvn.transform(result_mem)); 4744 } // original reexecute is set back here 4745 4746 set_result(_gvn.transform(result_val)); 4747 return true; 4748 } 4749 4750 // If we have a tighly coupled allocation, the arraycopy may take care 4751 // of the array initialization. If one of the guards we insert between 4752 // the allocation and the arraycopy causes a deoptimization, an 4753 // unitialized array will escape the compiled method. To prevent that 4754 // we set the JVM state for uncommon traps between the allocation and 4755 // the arraycopy to the state before the allocation so, in case of 4756 // deoptimization, we'll reexecute the allocation and the 4757 // initialization. 4758 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4759 if (alloc != NULL) { 4760 ciMethod* trap_method = alloc->jvms()->method(); 4761 int trap_bci = alloc->jvms()->bci(); 4762 4763 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4764 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4765 // Make sure there's no store between the allocation and the 4766 // arraycopy otherwise visible side effects could be rexecuted 4767 // in case of deoptimization and cause incorrect execution. 4768 bool no_interfering_store = true; 4769 Node* mem = alloc->in(TypeFunc::Memory); 4770 if (mem->is_MergeMem()) { 4771 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4772 Node* n = mms.memory(); 4773 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4774 assert(n->is_Store(), "what else?"); 4775 no_interfering_store = false; 4776 break; 4777 } 4778 } 4779 } else { 4780 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4781 Node* n = mms.memory(); 4782 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4783 assert(n->is_Store(), "what else?"); 4784 no_interfering_store = false; 4785 break; 4786 } 4787 } 4788 } 4789 4790 if (no_interfering_store) { 4791 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4792 uint size = alloc->req(); 4793 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4794 old_jvms->set_map(sfpt); 4795 for (uint i = 0; i < size; i++) { 4796 sfpt->init_req(i, alloc->in(i)); 4797 } 4798 // re-push array length for deoptimization 4799 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4800 old_jvms->set_sp(old_jvms->sp()+1); 4801 old_jvms->set_monoff(old_jvms->monoff()+1); 4802 old_jvms->set_scloff(old_jvms->scloff()+1); 4803 old_jvms->set_endoff(old_jvms->endoff()+1); 4804 old_jvms->set_should_reexecute(true); 4805 4806 sfpt->set_i_o(map()->i_o()); 4807 sfpt->set_memory(map()->memory()); 4808 sfpt->set_control(map()->control()); 4809 4810 JVMState* saved_jvms = jvms(); 4811 saved_reexecute_sp = _reexecute_sp; 4812 4813 set_jvms(sfpt->jvms()); 4814 _reexecute_sp = jvms()->sp(); 4815 4816 return saved_jvms; 4817 } 4818 } 4819 } 4820 return NULL; 4821 } 4822 4823 // In case of a deoptimization, we restart execution at the 4824 // allocation, allocating a new array. We would leave an uninitialized 4825 // array in the heap that GCs wouldn't expect. Move the allocation 4826 // after the traps so we don't allocate the array if we 4827 // deoptimize. This is possible because tightly_coupled_allocation() 4828 // guarantees there's no observer of the allocated array at this point 4829 // and the control flow is simple enough. 4830 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4831 if (saved_jvms != NULL && !stopped()) { 4832 assert(alloc != NULL, "only with a tightly coupled allocation"); 4833 // restore JVM state to the state at the arraycopy 4834 saved_jvms->map()->set_control(map()->control()); 4835 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4836 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4837 // If we've improved the types of some nodes (null check) while 4838 // emitting the guards, propagate them to the current state 4839 map()->replaced_nodes().apply(saved_jvms->map()); 4840 set_jvms(saved_jvms); 4841 _reexecute_sp = saved_reexecute_sp; 4842 4843 // Remove the allocation from above the guards 4844 CallProjections callprojs; 4845 alloc->extract_projections(&callprojs, true); 4846 InitializeNode* init = alloc->initialization(); 4847 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4848 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4849 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4850 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4851 4852 // move the allocation here (after the guards) 4853 _gvn.hash_delete(alloc); 4854 alloc->set_req(TypeFunc::Control, control()); 4855 alloc->set_req(TypeFunc::I_O, i_o()); 4856 Node *mem = reset_memory(); 4857 set_all_memory(mem); 4858 alloc->set_req(TypeFunc::Memory, mem); 4859 set_control(init->proj_out(TypeFunc::Control)); 4860 set_i_o(callprojs.fallthrough_ioproj); 4861 4862 // Update memory as done in GraphKit::set_output_for_allocation() 4863 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4864 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4865 if (ary_type->isa_aryptr() && length_type != NULL) { 4866 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4867 } 4868 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4869 int elemidx = C->get_alias_index(telemref); 4870 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4871 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4872 4873 Node* allocx = _gvn.transform(alloc); 4874 assert(allocx == alloc, "where has the allocation gone?"); 4875 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4876 4877 _gvn.hash_delete(dest); 4878 dest->set_req(0, control()); 4879 Node* destx = _gvn.transform(dest); 4880 assert(destx == dest, "where has the allocation result gone?"); 4881 } 4882 } 4883 4884 4885 //------------------------------inline_arraycopy----------------------- 4886 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4887 // Object dest, int destPos, 4888 // int length); 4889 bool LibraryCallKit::inline_arraycopy() { 4890 // Get the arguments. 4891 Node* src = argument(0); // type: oop 4892 Node* src_offset = argument(1); // type: int 4893 Node* dest = argument(2); // type: oop 4894 Node* dest_offset = argument(3); // type: int 4895 Node* length = argument(4); // type: int 4896 4897 4898 // Check for allocation before we add nodes that would confuse 4899 // tightly_coupled_allocation() 4900 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4901 4902 int saved_reexecute_sp = -1; 4903 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4904 // See arraycopy_restore_alloc_state() comment 4905 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4906 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4907 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4908 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4909 4910 // The following tests must be performed 4911 // (1) src and dest are arrays. 4912 // (2) src and dest arrays must have elements of the same BasicType 4913 // (3) src and dest must not be null. 4914 // (4) src_offset must not be negative. 4915 // (5) dest_offset must not be negative. 4916 // (6) length must not be negative. 4917 // (7) src_offset + length must not exceed length of src. 4918 // (8) dest_offset + length must not exceed length of dest. 4919 // (9) each element of an oop array must be assignable 4920 4921 // (3) src and dest must not be null. 4922 // always do this here because we need the JVM state for uncommon traps 4923 Node* null_ctl = top(); 4924 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4925 assert(null_ctl->is_top(), "no null control here"); 4926 dest = null_check(dest, T_ARRAY); 4927 4928 if (!can_emit_guards) { 4929 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4930 // guards but the arraycopy node could still take advantage of a 4931 // tightly allocated allocation. tightly_coupled_allocation() is 4932 // called again to make sure it takes the null check above into 4933 // account: the null check is mandatory and if it caused an 4934 // uncommon trap to be emitted then the allocation can't be 4935 // considered tightly coupled in this context. 4936 alloc = tightly_coupled_allocation(dest, NULL); 4937 } 4938 4939 bool validated = false; 4940 4941 const Type* src_type = _gvn.type(src); 4942 const Type* dest_type = _gvn.type(dest); 4943 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4944 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4945 4946 // Do we have the type of src? 4947 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4948 // Do we have the type of dest? 4949 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4950 // Is the type for src from speculation? 4951 bool src_spec = false; 4952 // Is the type for dest from speculation? 4953 bool dest_spec = false; 4954 4955 if ((!has_src || !has_dest) && can_emit_guards) { 4956 // We don't have sufficient type information, let's see if 4957 // speculative types can help. We need to have types for both src 4958 // and dest so that it pays off. 4959 4960 // Do we already have or could we have type information for src 4961 bool could_have_src = has_src; 4962 // Do we already have or could we have type information for dest 4963 bool could_have_dest = has_dest; 4964 4965 ciKlass* src_k = NULL; 4966 if (!has_src) { 4967 src_k = src_type->speculative_type_not_null(); 4968 if (src_k != NULL && src_k->is_array_klass()) { 4969 could_have_src = true; 4970 } 4971 } 4972 4973 ciKlass* dest_k = NULL; 4974 if (!has_dest) { 4975 dest_k = dest_type->speculative_type_not_null(); 4976 if (dest_k != NULL && dest_k->is_array_klass()) { 4977 could_have_dest = true; 4978 } 4979 } 4980 4981 if (could_have_src && could_have_dest) { 4982 // This is going to pay off so emit the required guards 4983 if (!has_src) { 4984 src = maybe_cast_profiled_obj(src, src_k, true); 4985 src_type = _gvn.type(src); 4986 top_src = src_type->isa_aryptr(); 4987 has_src = (top_src != NULL && top_src->klass() != NULL); 4988 src_spec = true; 4989 } 4990 if (!has_dest) { 4991 dest = maybe_cast_profiled_obj(dest, dest_k, true); 4992 dest_type = _gvn.type(dest); 4993 top_dest = dest_type->isa_aryptr(); 4994 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4995 dest_spec = true; 4996 } 4997 } 4998 } 4999 5000 if (has_src && has_dest && can_emit_guards) { 5001 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 5002 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 5003 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 5004 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 5005 5006 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5007 // If both arrays are object arrays then having the exact types 5008 // for both will remove the need for a subtype check at runtime 5009 // before the call and may make it possible to pick a faster copy 5010 // routine (without a subtype check on every element) 5011 // Do we have the exact type of src? 5012 bool could_have_src = src_spec; 5013 // Do we have the exact type of dest? 5014 bool could_have_dest = dest_spec; 5015 ciKlass* src_k = top_src->klass(); 5016 ciKlass* dest_k = top_dest->klass(); 5017 if (!src_spec) { 5018 src_k = src_type->speculative_type_not_null(); 5019 if (src_k != NULL && src_k->is_array_klass()) { 5020 could_have_src = true; 5021 } 5022 } 5023 if (!dest_spec) { 5024 dest_k = dest_type->speculative_type_not_null(); 5025 if (dest_k != NULL && dest_k->is_array_klass()) { 5026 could_have_dest = true; 5027 } 5028 } 5029 if (could_have_src && could_have_dest) { 5030 // If we can have both exact types, emit the missing guards 5031 if (could_have_src && !src_spec) { 5032 src = maybe_cast_profiled_obj(src, src_k, true); 5033 } 5034 if (could_have_dest && !dest_spec) { 5035 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5036 } 5037 } 5038 } 5039 } 5040 5041 ciMethod* trap_method = method(); 5042 int trap_bci = bci(); 5043 if (saved_jvms != NULL) { 5044 trap_method = alloc->jvms()->method(); 5045 trap_bci = alloc->jvms()->bci(); 5046 } 5047 5048 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5049 can_emit_guards && 5050 !src->is_top() && !dest->is_top()) { 5051 // validate arguments: enables transformation the ArrayCopyNode 5052 validated = true; 5053 5054 RegionNode* slow_region = new RegionNode(1); 5055 record_for_igvn(slow_region); 5056 5057 // (1) src and dest are arrays. 5058 generate_non_array_guard(load_object_klass(src), slow_region); 5059 generate_non_array_guard(load_object_klass(dest), slow_region); 5060 5061 // (2) src and dest arrays must have elements of the same BasicType 5062 // done at macro expansion or at Ideal transformation time 5063 5064 // (4) src_offset must not be negative. 5065 generate_negative_guard(src_offset, slow_region); 5066 5067 // (5) dest_offset must not be negative. 5068 generate_negative_guard(dest_offset, slow_region); 5069 5070 // (7) src_offset + length must not exceed length of src. 5071 generate_limit_guard(src_offset, length, 5072 load_array_length(src), 5073 slow_region); 5074 5075 // (8) dest_offset + length must not exceed length of dest. 5076 generate_limit_guard(dest_offset, length, 5077 load_array_length(dest), 5078 slow_region); 5079 5080 // (9) each element of an oop array must be assignable 5081 Node* src_klass = load_object_klass(src); 5082 Node* dest_klass = load_object_klass(dest); 5083 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5084 5085 if (not_subtype_ctrl != top()) { 5086 PreserveJVMState pjvms(this); 5087 set_control(not_subtype_ctrl); 5088 uncommon_trap(Deoptimization::Reason_intrinsic, 5089 Deoptimization::Action_make_not_entrant); 5090 assert(stopped(), "Should be stopped"); 5091 } 5092 { 5093 PreserveJVMState pjvms(this); 5094 set_control(_gvn.transform(slow_region)); 5095 uncommon_trap(Deoptimization::Reason_intrinsic, 5096 Deoptimization::Action_make_not_entrant); 5097 assert(stopped(), "Should be stopped"); 5098 } 5099 } 5100 5101 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5102 5103 if (stopped()) { 5104 return true; 5105 } 5106 5107 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5108 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5109 // so the compiler has a chance to eliminate them: during macro expansion, 5110 // we have to set their control (CastPP nodes are eliminated). 5111 load_object_klass(src), load_object_klass(dest), 5112 load_array_length(src), load_array_length(dest)); 5113 5114 ac->set_arraycopy(validated); 5115 5116 Node* n = _gvn.transform(ac); 5117 if (n == ac) { 5118 ac->connect_outputs(this); 5119 } else { 5120 assert(validated, "shouldn't transform if all arguments not validated"); 5121 set_all_memory(n); 5122 } 5123 5124 return true; 5125 } 5126 5127 5128 // Helper function which determines if an arraycopy immediately follows 5129 // an allocation, with no intervening tests or other escapes for the object. 5130 AllocateArrayNode* 5131 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5132 RegionNode* slow_region) { 5133 if (stopped()) return NULL; // no fast path 5134 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5135 5136 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5137 if (alloc == NULL) return NULL; 5138 5139 Node* rawmem = memory(Compile::AliasIdxRaw); 5140 // Is the allocation's memory state untouched? 5141 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5142 // Bail out if there have been raw-memory effects since the allocation. 5143 // (Example: There might have been a call or safepoint.) 5144 return NULL; 5145 } 5146 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5147 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5148 return NULL; 5149 } 5150 5151 // There must be no unexpected observers of this allocation. 5152 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5153 Node* obs = ptr->fast_out(i); 5154 if (obs != this->map()) { 5155 return NULL; 5156 } 5157 } 5158 5159 // This arraycopy must unconditionally follow the allocation of the ptr. 5160 Node* alloc_ctl = ptr->in(0); 5161 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5162 5163 Node* ctl = control(); 5164 while (ctl != alloc_ctl) { 5165 // There may be guards which feed into the slow_region. 5166 // Any other control flow means that we might not get a chance 5167 // to finish initializing the allocated object. 5168 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5169 IfNode* iff = ctl->in(0)->as_If(); 5170 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5171 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5172 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5173 ctl = iff->in(0); // This test feeds the known slow_region. 5174 continue; 5175 } 5176 // One more try: Various low-level checks bottom out in 5177 // uncommon traps. If the debug-info of the trap omits 5178 // any reference to the allocation, as we've already 5179 // observed, then there can be no objection to the trap. 5180 bool found_trap = false; 5181 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5182 Node* obs = not_ctl->fast_out(j); 5183 if (obs->in(0) == not_ctl && obs->is_Call() && 5184 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5185 found_trap = true; break; 5186 } 5187 } 5188 if (found_trap) { 5189 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5190 continue; 5191 } 5192 } 5193 return NULL; 5194 } 5195 5196 // If we get this far, we have an allocation which immediately 5197 // precedes the arraycopy, and we can take over zeroing the new object. 5198 // The arraycopy will finish the initialization, and provide 5199 // a new control state to which we will anchor the destination pointer. 5200 5201 return alloc; 5202 } 5203 5204 //-------------inline_encodeISOArray----------------------------------- 5205 // encode char[] to byte[] in ISO_8859_1 5206 bool LibraryCallKit::inline_encodeISOArray() { 5207 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5208 // no receiver since it is static method 5209 Node *src = argument(0); 5210 Node *src_offset = argument(1); 5211 Node *dst = argument(2); 5212 Node *dst_offset = argument(3); 5213 Node *length = argument(4); 5214 5215 const Type* src_type = src->Value(&_gvn); 5216 const Type* dst_type = dst->Value(&_gvn); 5217 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5218 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5219 if (top_src == NULL || top_src->klass() == NULL || 5220 top_dest == NULL || top_dest->klass() == NULL) { 5221 // failed array check 5222 return false; 5223 } 5224 5225 // Figure out the size and type of the elements we will be copying. 5226 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5227 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5228 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5229 return false; 5230 } 5231 Node* src_start = array_element_address(src, src_offset, src_elem); 5232 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5233 // 'src_start' points to src array + scaled offset 5234 // 'dst_start' points to dst array + scaled offset 5235 5236 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5237 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5238 enc = _gvn.transform(enc); 5239 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5240 set_memory(res_mem, mtype); 5241 set_result(enc); 5242 return true; 5243 } 5244 5245 //-------------inline_multiplyToLen----------------------------------- 5246 bool LibraryCallKit::inline_multiplyToLen() { 5247 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5248 5249 address stubAddr = StubRoutines::multiplyToLen(); 5250 if (stubAddr == NULL) { 5251 return false; // Intrinsic's stub is not implemented on this platform 5252 } 5253 const char* stubName = "multiplyToLen"; 5254 5255 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5256 5257 Node* x = argument(1); 5258 Node* xlen = argument(2); 5259 Node* y = argument(3); 5260 Node* ylen = argument(4); 5261 Node* z = argument(5); 5262 5263 const Type* x_type = x->Value(&_gvn); 5264 const Type* y_type = y->Value(&_gvn); 5265 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5266 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5267 if (top_x == NULL || top_x->klass() == NULL || 5268 top_y == NULL || top_y->klass() == NULL) { 5269 // failed array check 5270 return false; 5271 } 5272 5273 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5274 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5275 if (x_elem != T_INT || y_elem != T_INT) { 5276 return false; 5277 } 5278 5279 // Set the original stack and the reexecute bit for the interpreter to reexecute 5280 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5281 // on the return from z array allocation in runtime. 5282 { PreserveReexecuteState preexecs(this); 5283 jvms()->set_should_reexecute(true); 5284 5285 Node* x_start = array_element_address(x, intcon(0), x_elem); 5286 Node* y_start = array_element_address(y, intcon(0), y_elem); 5287 // 'x_start' points to x array + scaled xlen 5288 // 'y_start' points to y array + scaled ylen 5289 5290 // Allocate the result array 5291 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5292 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5293 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5294 5295 IdealKit ideal(this); 5296 5297 #define __ ideal. 5298 Node* one = __ ConI(1); 5299 Node* zero = __ ConI(0); 5300 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5301 __ set(need_alloc, zero); 5302 __ set(z_alloc, z); 5303 __ if_then(z, BoolTest::eq, null()); { 5304 __ increment (need_alloc, one); 5305 } __ else_(); { 5306 // Update graphKit memory and control from IdealKit. 5307 sync_kit(ideal); 5308 Node* zlen_arg = load_array_length(z); 5309 // Update IdealKit memory and control from graphKit. 5310 __ sync_kit(this); 5311 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5312 __ increment (need_alloc, one); 5313 } __ end_if(); 5314 } __ end_if(); 5315 5316 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5317 // Update graphKit memory and control from IdealKit. 5318 sync_kit(ideal); 5319 Node * narr = new_array(klass_node, zlen, 1); 5320 // Update IdealKit memory and control from graphKit. 5321 __ sync_kit(this); 5322 __ set(z_alloc, narr); 5323 } __ end_if(); 5324 5325 sync_kit(ideal); 5326 z = __ value(z_alloc); 5327 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5328 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5329 // Final sync IdealKit and GraphKit. 5330 final_sync(ideal); 5331 #undef __ 5332 5333 Node* z_start = array_element_address(z, intcon(0), T_INT); 5334 5335 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5336 OptoRuntime::multiplyToLen_Type(), 5337 stubAddr, stubName, TypePtr::BOTTOM, 5338 x_start, xlen, y_start, ylen, z_start, zlen); 5339 } // original reexecute is set back here 5340 5341 C->set_has_split_ifs(true); // Has chance for split-if optimization 5342 set_result(z); 5343 return true; 5344 } 5345 5346 //-------------inline_squareToLen------------------------------------ 5347 bool LibraryCallKit::inline_squareToLen() { 5348 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5349 5350 address stubAddr = StubRoutines::squareToLen(); 5351 if (stubAddr == NULL) { 5352 return false; // Intrinsic's stub is not implemented on this platform 5353 } 5354 const char* stubName = "squareToLen"; 5355 5356 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5357 5358 Node* x = argument(0); 5359 Node* len = argument(1); 5360 Node* z = argument(2); 5361 Node* zlen = argument(3); 5362 5363 const Type* x_type = x->Value(&_gvn); 5364 const Type* z_type = z->Value(&_gvn); 5365 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5366 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5367 if (top_x == NULL || top_x->klass() == NULL || 5368 top_z == NULL || top_z->klass() == NULL) { 5369 // failed array check 5370 return false; 5371 } 5372 5373 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5374 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5375 if (x_elem != T_INT || z_elem != T_INT) { 5376 return false; 5377 } 5378 5379 5380 Node* x_start = array_element_address(x, intcon(0), x_elem); 5381 Node* z_start = array_element_address(z, intcon(0), z_elem); 5382 5383 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5384 OptoRuntime::squareToLen_Type(), 5385 stubAddr, stubName, TypePtr::BOTTOM, 5386 x_start, len, z_start, zlen); 5387 5388 set_result(z); 5389 return true; 5390 } 5391 5392 //-------------inline_mulAdd------------------------------------------ 5393 bool LibraryCallKit::inline_mulAdd() { 5394 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5395 5396 address stubAddr = StubRoutines::mulAdd(); 5397 if (stubAddr == NULL) { 5398 return false; // Intrinsic's stub is not implemented on this platform 5399 } 5400 const char* stubName = "mulAdd"; 5401 5402 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5403 5404 Node* out = argument(0); 5405 Node* in = argument(1); 5406 Node* offset = argument(2); 5407 Node* len = argument(3); 5408 Node* k = argument(4); 5409 5410 const Type* out_type = out->Value(&_gvn); 5411 const Type* in_type = in->Value(&_gvn); 5412 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5413 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5414 if (top_out == NULL || top_out->klass() == NULL || 5415 top_in == NULL || top_in->klass() == NULL) { 5416 // failed array check 5417 return false; 5418 } 5419 5420 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5421 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5422 if (out_elem != T_INT || in_elem != T_INT) { 5423 return false; 5424 } 5425 5426 Node* outlen = load_array_length(out); 5427 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5428 Node* out_start = array_element_address(out, intcon(0), out_elem); 5429 Node* in_start = array_element_address(in, intcon(0), in_elem); 5430 5431 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5432 OptoRuntime::mulAdd_Type(), 5433 stubAddr, stubName, TypePtr::BOTTOM, 5434 out_start,in_start, new_offset, len, k); 5435 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5436 set_result(result); 5437 return true; 5438 } 5439 5440 5441 /** 5442 * Calculate CRC32 for byte. 5443 * int java.util.zip.CRC32.update(int crc, int b) 5444 */ 5445 bool LibraryCallKit::inline_updateCRC32() { 5446 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5447 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5448 // no receiver since it is static method 5449 Node* crc = argument(0); // type: int 5450 Node* b = argument(1); // type: int 5451 5452 /* 5453 * int c = ~ crc; 5454 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5455 * b = b ^ (c >>> 8); 5456 * crc = ~b; 5457 */ 5458 5459 Node* M1 = intcon(-1); 5460 crc = _gvn.transform(new XorINode(crc, M1)); 5461 Node* result = _gvn.transform(new XorINode(crc, b)); 5462 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5463 5464 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5465 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5466 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5467 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5468 5469 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5470 result = _gvn.transform(new XorINode(crc, result)); 5471 result = _gvn.transform(new XorINode(result, M1)); 5472 set_result(result); 5473 return true; 5474 } 5475 5476 /** 5477 * Calculate CRC32 for byte[] array. 5478 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5479 */ 5480 bool LibraryCallKit::inline_updateBytesCRC32() { 5481 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5482 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5483 // no receiver since it is static method 5484 Node* crc = argument(0); // type: int 5485 Node* src = argument(1); // type: oop 5486 Node* offset = argument(2); // type: int 5487 Node* length = argument(3); // type: int 5488 5489 const Type* src_type = src->Value(&_gvn); 5490 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5491 if (top_src == NULL || top_src->klass() == NULL) { 5492 // failed array check 5493 return false; 5494 } 5495 5496 // Figure out the size and type of the elements we will be copying. 5497 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5498 if (src_elem != T_BYTE) { 5499 return false; 5500 } 5501 5502 // 'src_start' points to src array + scaled offset 5503 Node* src_start = array_element_address(src, offset, src_elem); 5504 5505 // We assume that range check is done by caller. 5506 // TODO: generate range check (offset+length < src.length) in debug VM. 5507 5508 // Call the stub. 5509 address stubAddr = StubRoutines::updateBytesCRC32(); 5510 const char *stubName = "updateBytesCRC32"; 5511 5512 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5513 stubAddr, stubName, TypePtr::BOTTOM, 5514 crc, src_start, length); 5515 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5516 set_result(result); 5517 return true; 5518 } 5519 5520 /** 5521 * Calculate CRC32 for ByteBuffer. 5522 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5523 */ 5524 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5525 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5526 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5527 // no receiver since it is static method 5528 Node* crc = argument(0); // type: int 5529 Node* src = argument(1); // type: long 5530 Node* offset = argument(3); // type: int 5531 Node* length = argument(4); // type: int 5532 5533 src = ConvL2X(src); // adjust Java long to machine word 5534 Node* base = _gvn.transform(new CastX2PNode(src)); 5535 offset = ConvI2X(offset); 5536 5537 // 'src_start' points to src array + scaled offset 5538 Node* src_start = basic_plus_adr(top(), base, offset); 5539 5540 // Call the stub. 5541 address stubAddr = StubRoutines::updateBytesCRC32(); 5542 const char *stubName = "updateBytesCRC32"; 5543 5544 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5545 stubAddr, stubName, TypePtr::BOTTOM, 5546 crc, src_start, length); 5547 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5548 set_result(result); 5549 return true; 5550 } 5551 5552 //------------------------------get_table_from_crc32c_class----------------------- 5553 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5554 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5555 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5556 5557 return table; 5558 } 5559 5560 //------------------------------inline_updateBytesCRC32C----------------------- 5561 // 5562 // Calculate CRC32C for byte[] array. 5563 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5564 // 5565 bool LibraryCallKit::inline_updateBytesCRC32C() { 5566 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5567 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5568 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5569 // no receiver since it is a static method 5570 Node* crc = argument(0); // type: int 5571 Node* src = argument(1); // type: oop 5572 Node* offset = argument(2); // type: int 5573 Node* end = argument(3); // type: int 5574 5575 Node* length = _gvn.transform(new SubINode(end, offset)); 5576 5577 const Type* src_type = src->Value(&_gvn); 5578 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5579 if (top_src == NULL || top_src->klass() == NULL) { 5580 // failed array check 5581 return false; 5582 } 5583 5584 // Figure out the size and type of the elements we will be copying. 5585 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5586 if (src_elem != T_BYTE) { 5587 return false; 5588 } 5589 5590 // 'src_start' points to src array + scaled offset 5591 Node* src_start = array_element_address(src, offset, src_elem); 5592 5593 // static final int[] byteTable in class CRC32C 5594 Node* table = get_table_from_crc32c_class(callee()->holder()); 5595 Node* table_start = array_element_address(table, intcon(0), T_INT); 5596 5597 // We assume that range check is done by caller. 5598 // TODO: generate range check (offset+length < src.length) in debug VM. 5599 5600 // Call the stub. 5601 address stubAddr = StubRoutines::updateBytesCRC32C(); 5602 const char *stubName = "updateBytesCRC32C"; 5603 5604 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5605 stubAddr, stubName, TypePtr::BOTTOM, 5606 crc, src_start, length, table_start); 5607 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5608 set_result(result); 5609 return true; 5610 } 5611 5612 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5613 // 5614 // Calculate CRC32C for DirectByteBuffer. 5615 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5616 // 5617 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5618 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5619 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5620 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5621 // no receiver since it is a static method 5622 Node* crc = argument(0); // type: int 5623 Node* src = argument(1); // type: long 5624 Node* offset = argument(3); // type: int 5625 Node* end = argument(4); // type: int 5626 5627 Node* length = _gvn.transform(new SubINode(end, offset)); 5628 5629 src = ConvL2X(src); // adjust Java long to machine word 5630 Node* base = _gvn.transform(new CastX2PNode(src)); 5631 offset = ConvI2X(offset); 5632 5633 // 'src_start' points to src array + scaled offset 5634 Node* src_start = basic_plus_adr(top(), base, offset); 5635 5636 // static final int[] byteTable in class CRC32C 5637 Node* table = get_table_from_crc32c_class(callee()->holder()); 5638 Node* table_start = array_element_address(table, intcon(0), T_INT); 5639 5640 // Call the stub. 5641 address stubAddr = StubRoutines::updateBytesCRC32C(); 5642 const char *stubName = "updateBytesCRC32C"; 5643 5644 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5645 stubAddr, stubName, TypePtr::BOTTOM, 5646 crc, src_start, length, table_start); 5647 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5648 set_result(result); 5649 return true; 5650 } 5651 5652 //----------------------------inline_reference_get---------------------------- 5653 // public T java.lang.ref.Reference.get(); 5654 bool LibraryCallKit::inline_reference_get() { 5655 const int referent_offset = java_lang_ref_Reference::referent_offset; 5656 guarantee(referent_offset > 0, "should have already been set"); 5657 5658 // Get the argument: 5659 Node* reference_obj = null_check_receiver(); 5660 if (stopped()) return true; 5661 5662 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5663 5664 ciInstanceKlass* klass = env()->Object_klass(); 5665 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5666 5667 Node* no_ctrl = NULL; 5668 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5669 5670 // Use the pre-barrier to record the value in the referent field 5671 pre_barrier(false /* do_load */, 5672 control(), 5673 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5674 result /* pre_val */, 5675 T_OBJECT); 5676 5677 // Add memory barrier to prevent commoning reads from this field 5678 // across safepoint since GC can change its value. 5679 insert_mem_bar(Op_MemBarCPUOrder); 5680 5681 set_result(result); 5682 return true; 5683 } 5684 5685 5686 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5687 bool is_exact=true, bool is_static=false, 5688 ciInstanceKlass * fromKls=NULL) { 5689 if (fromKls == NULL) { 5690 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5691 assert(tinst != NULL, "obj is null"); 5692 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5693 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5694 fromKls = tinst->klass()->as_instance_klass(); 5695 } else { 5696 assert(is_static, "only for static field access"); 5697 } 5698 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5699 ciSymbol::make(fieldTypeString), 5700 is_static); 5701 5702 assert (field != NULL, "undefined field"); 5703 if (field == NULL) return (Node *) NULL; 5704 5705 if (is_static) { 5706 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5707 fromObj = makecon(tip); 5708 } 5709 5710 // Next code copied from Parse::do_get_xxx(): 5711 5712 // Compute address and memory type. 5713 int offset = field->offset_in_bytes(); 5714 bool is_vol = field->is_volatile(); 5715 ciType* field_klass = field->type(); 5716 assert(field_klass->is_loaded(), "should be loaded"); 5717 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5718 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5719 BasicType bt = field->layout_type(); 5720 5721 // Build the resultant type of the load 5722 const Type *type; 5723 if (bt == T_OBJECT) { 5724 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5725 } else { 5726 type = Type::get_const_basic_type(bt); 5727 } 5728 5729 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5730 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5731 } 5732 // Build the load. 5733 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5734 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5735 // If reference is volatile, prevent following memory ops from 5736 // floating up past the volatile read. Also prevents commoning 5737 // another volatile read. 5738 if (is_vol) { 5739 // Memory barrier includes bogus read of value to force load BEFORE membar 5740 insert_mem_bar(Op_MemBarAcquire, loadedField); 5741 } 5742 return loadedField; 5743 } 5744 5745 5746 //------------------------------inline_aescrypt_Block----------------------- 5747 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5748 address stubAddr; 5749 const char *stubName; 5750 assert(UseAES, "need AES instruction support"); 5751 5752 switch(id) { 5753 case vmIntrinsics::_aescrypt_encryptBlock: 5754 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5755 stubName = "aescrypt_encryptBlock"; 5756 break; 5757 case vmIntrinsics::_aescrypt_decryptBlock: 5758 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5759 stubName = "aescrypt_decryptBlock"; 5760 break; 5761 } 5762 if (stubAddr == NULL) return false; 5763 5764 Node* aescrypt_object = argument(0); 5765 Node* src = argument(1); 5766 Node* src_offset = argument(2); 5767 Node* dest = argument(3); 5768 Node* dest_offset = argument(4); 5769 5770 // (1) src and dest are arrays. 5771 const Type* src_type = src->Value(&_gvn); 5772 const Type* dest_type = dest->Value(&_gvn); 5773 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5774 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5775 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5776 5777 // for the quick and dirty code we will skip all the checks. 5778 // we are just trying to get the call to be generated. 5779 Node* src_start = src; 5780 Node* dest_start = dest; 5781 if (src_offset != NULL || dest_offset != NULL) { 5782 assert(src_offset != NULL && dest_offset != NULL, ""); 5783 src_start = array_element_address(src, src_offset, T_BYTE); 5784 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5785 } 5786 5787 // now need to get the start of its expanded key array 5788 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5789 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5790 if (k_start == NULL) return false; 5791 5792 if (Matcher::pass_original_key_for_aes()) { 5793 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5794 // compatibility issues between Java key expansion and SPARC crypto instructions 5795 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5796 if (original_k_start == NULL) return false; 5797 5798 // Call the stub. 5799 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5800 stubAddr, stubName, TypePtr::BOTTOM, 5801 src_start, dest_start, k_start, original_k_start); 5802 } else { 5803 // Call the stub. 5804 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5805 stubAddr, stubName, TypePtr::BOTTOM, 5806 src_start, dest_start, k_start); 5807 } 5808 5809 return true; 5810 } 5811 5812 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5813 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5814 address stubAddr; 5815 const char *stubName; 5816 5817 assert(UseAES, "need AES instruction support"); 5818 5819 switch(id) { 5820 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5821 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5822 stubName = "cipherBlockChaining_encryptAESCrypt"; 5823 break; 5824 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5825 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5826 stubName = "cipherBlockChaining_decryptAESCrypt"; 5827 break; 5828 } 5829 if (stubAddr == NULL) return false; 5830 5831 Node* cipherBlockChaining_object = argument(0); 5832 Node* src = argument(1); 5833 Node* src_offset = argument(2); 5834 Node* len = argument(3); 5835 Node* dest = argument(4); 5836 Node* dest_offset = argument(5); 5837 5838 // (1) src and dest are arrays. 5839 const Type* src_type = src->Value(&_gvn); 5840 const Type* dest_type = dest->Value(&_gvn); 5841 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5842 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5843 assert (top_src != NULL && top_src->klass() != NULL 5844 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5845 5846 // checks are the responsibility of the caller 5847 Node* src_start = src; 5848 Node* dest_start = dest; 5849 if (src_offset != NULL || dest_offset != NULL) { 5850 assert(src_offset != NULL && dest_offset != NULL, ""); 5851 src_start = array_element_address(src, src_offset, T_BYTE); 5852 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5853 } 5854 5855 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5856 // (because of the predicated logic executed earlier). 5857 // so we cast it here safely. 5858 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5859 5860 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5861 if (embeddedCipherObj == NULL) return false; 5862 5863 // cast it to what we know it will be at runtime 5864 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5865 assert(tinst != NULL, "CBC obj is null"); 5866 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5867 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5868 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5869 5870 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5871 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5872 const TypeOopPtr* xtype = aklass->as_instance_type(); 5873 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5874 aescrypt_object = _gvn.transform(aescrypt_object); 5875 5876 // we need to get the start of the aescrypt_object's expanded key array 5877 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5878 if (k_start == NULL) return false; 5879 5880 // similarly, get the start address of the r vector 5881 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5882 if (objRvec == NULL) return false; 5883 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5884 5885 Node* cbcCrypt; 5886 if (Matcher::pass_original_key_for_aes()) { 5887 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5888 // compatibility issues between Java key expansion and SPARC crypto instructions 5889 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5890 if (original_k_start == NULL) return false; 5891 5892 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5893 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5894 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5895 stubAddr, stubName, TypePtr::BOTTOM, 5896 src_start, dest_start, k_start, r_start, len, original_k_start); 5897 } else { 5898 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5899 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5900 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5901 stubAddr, stubName, TypePtr::BOTTOM, 5902 src_start, dest_start, k_start, r_start, len); 5903 } 5904 5905 // return cipher length (int) 5906 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5907 set_result(retvalue); 5908 return true; 5909 } 5910 5911 //------------------------------get_key_start_from_aescrypt_object----------------------- 5912 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5913 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5914 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5915 if (objAESCryptKey == NULL) return (Node *) NULL; 5916 5917 // now have the array, need to get the start address of the K array 5918 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5919 return k_start; 5920 } 5921 5922 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5923 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5924 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5925 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5926 if (objAESCryptKey == NULL) return (Node *) NULL; 5927 5928 // now have the array, need to get the start address of the lastKey array 5929 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5930 return original_k_start; 5931 } 5932 5933 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5934 // Return node representing slow path of predicate check. 5935 // the pseudo code we want to emulate with this predicate is: 5936 // for encryption: 5937 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5938 // for decryption: 5939 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5940 // note cipher==plain is more conservative than the original java code but that's OK 5941 // 5942 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5943 // The receiver was checked for NULL already. 5944 Node* objCBC = argument(0); 5945 5946 // Load embeddedCipher field of CipherBlockChaining object. 5947 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5948 5949 // get AESCrypt klass for instanceOf check 5950 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5951 // will have same classloader as CipherBlockChaining object 5952 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5953 assert(tinst != NULL, "CBCobj is null"); 5954 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5955 5956 // we want to do an instanceof comparison against the AESCrypt class 5957 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5958 if (!klass_AESCrypt->is_loaded()) { 5959 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5960 Node* ctrl = control(); 5961 set_control(top()); // no regular fast path 5962 return ctrl; 5963 } 5964 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5965 5966 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5967 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5968 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5969 5970 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5971 5972 // for encryption, we are done 5973 if (!decrypting) 5974 return instof_false; // even if it is NULL 5975 5976 // for decryption, we need to add a further check to avoid 5977 // taking the intrinsic path when cipher and plain are the same 5978 // see the original java code for why. 5979 RegionNode* region = new RegionNode(3); 5980 region->init_req(1, instof_false); 5981 Node* src = argument(1); 5982 Node* dest = argument(4); 5983 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5984 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5985 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5986 region->init_req(2, src_dest_conjoint); 5987 5988 record_for_igvn(region); 5989 return _gvn.transform(region); 5990 } 5991 5992 //------------------------------inline_ghash_processBlocks 5993 bool LibraryCallKit::inline_ghash_processBlocks() { 5994 address stubAddr; 5995 const char *stubName; 5996 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 5997 5998 stubAddr = StubRoutines::ghash_processBlocks(); 5999 stubName = "ghash_processBlocks"; 6000 6001 Node* data = argument(0); 6002 Node* offset = argument(1); 6003 Node* len = argument(2); 6004 Node* state = argument(3); 6005 Node* subkeyH = argument(4); 6006 6007 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6008 assert(state_start, "state is NULL"); 6009 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6010 assert(subkeyH_start, "subkeyH is NULL"); 6011 Node* data_start = array_element_address(data, offset, T_BYTE); 6012 assert(data_start, "data is NULL"); 6013 6014 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6015 OptoRuntime::ghash_processBlocks_Type(), 6016 stubAddr, stubName, TypePtr::BOTTOM, 6017 state_start, subkeyH_start, data_start, len); 6018 return true; 6019 } 6020 6021 //------------------------------inline_sha_implCompress----------------------- 6022 // 6023 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6024 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6025 // 6026 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6027 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6028 // 6029 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6030 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6031 // 6032 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6033 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6034 6035 Node* sha_obj = argument(0); 6036 Node* src = argument(1); // type oop 6037 Node* ofs = argument(2); // type int 6038 6039 const Type* src_type = src->Value(&_gvn); 6040 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6041 if (top_src == NULL || top_src->klass() == NULL) { 6042 // failed array check 6043 return false; 6044 } 6045 // Figure out the size and type of the elements we will be copying. 6046 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6047 if (src_elem != T_BYTE) { 6048 return false; 6049 } 6050 // 'src_start' points to src array + offset 6051 Node* src_start = array_element_address(src, ofs, src_elem); 6052 Node* state = NULL; 6053 address stubAddr; 6054 const char *stubName; 6055 6056 switch(id) { 6057 case vmIntrinsics::_sha_implCompress: 6058 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6059 state = get_state_from_sha_object(sha_obj); 6060 stubAddr = StubRoutines::sha1_implCompress(); 6061 stubName = "sha1_implCompress"; 6062 break; 6063 case vmIntrinsics::_sha2_implCompress: 6064 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6065 state = get_state_from_sha_object(sha_obj); 6066 stubAddr = StubRoutines::sha256_implCompress(); 6067 stubName = "sha256_implCompress"; 6068 break; 6069 case vmIntrinsics::_sha5_implCompress: 6070 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6071 state = get_state_from_sha5_object(sha_obj); 6072 stubAddr = StubRoutines::sha512_implCompress(); 6073 stubName = "sha512_implCompress"; 6074 break; 6075 default: 6076 fatal_unexpected_iid(id); 6077 return false; 6078 } 6079 if (state == NULL) return false; 6080 6081 // Call the stub. 6082 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6083 stubAddr, stubName, TypePtr::BOTTOM, 6084 src_start, state); 6085 6086 return true; 6087 } 6088 6089 //------------------------------inline_digestBase_implCompressMB----------------------- 6090 // 6091 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6092 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6093 // 6094 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6095 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6096 "need SHA1/SHA256/SHA512 instruction support"); 6097 assert((uint)predicate < 3, "sanity"); 6098 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6099 6100 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6101 Node* src = argument(1); // byte[] array 6102 Node* ofs = argument(2); // type int 6103 Node* limit = argument(3); // type int 6104 6105 const Type* src_type = src->Value(&_gvn); 6106 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6107 if (top_src == NULL || top_src->klass() == NULL) { 6108 // failed array check 6109 return false; 6110 } 6111 // Figure out the size and type of the elements we will be copying. 6112 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6113 if (src_elem != T_BYTE) { 6114 return false; 6115 } 6116 // 'src_start' points to src array + offset 6117 Node* src_start = array_element_address(src, ofs, src_elem); 6118 6119 const char* klass_SHA_name = NULL; 6120 const char* stub_name = NULL; 6121 address stub_addr = NULL; 6122 bool long_state = false; 6123 6124 switch (predicate) { 6125 case 0: 6126 if (UseSHA1Intrinsics) { 6127 klass_SHA_name = "sun/security/provider/SHA"; 6128 stub_name = "sha1_implCompressMB"; 6129 stub_addr = StubRoutines::sha1_implCompressMB(); 6130 } 6131 break; 6132 case 1: 6133 if (UseSHA256Intrinsics) { 6134 klass_SHA_name = "sun/security/provider/SHA2"; 6135 stub_name = "sha256_implCompressMB"; 6136 stub_addr = StubRoutines::sha256_implCompressMB(); 6137 } 6138 break; 6139 case 2: 6140 if (UseSHA512Intrinsics) { 6141 klass_SHA_name = "sun/security/provider/SHA5"; 6142 stub_name = "sha512_implCompressMB"; 6143 stub_addr = StubRoutines::sha512_implCompressMB(); 6144 long_state = true; 6145 } 6146 break; 6147 default: 6148 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6149 } 6150 if (klass_SHA_name != NULL) { 6151 // get DigestBase klass to lookup for SHA klass 6152 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6153 assert(tinst != NULL, "digestBase_obj is not instance???"); 6154 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6155 6156 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6157 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6158 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6159 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6160 } 6161 return false; 6162 } 6163 //------------------------------inline_sha_implCompressMB----------------------- 6164 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6165 bool long_state, address stubAddr, const char *stubName, 6166 Node* src_start, Node* ofs, Node* limit) { 6167 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6168 const TypeOopPtr* xtype = aklass->as_instance_type(); 6169 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6170 sha_obj = _gvn.transform(sha_obj); 6171 6172 Node* state; 6173 if (long_state) { 6174 state = get_state_from_sha5_object(sha_obj); 6175 } else { 6176 state = get_state_from_sha_object(sha_obj); 6177 } 6178 if (state == NULL) return false; 6179 6180 // Call the stub. 6181 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6182 OptoRuntime::digestBase_implCompressMB_Type(), 6183 stubAddr, stubName, TypePtr::BOTTOM, 6184 src_start, state, ofs, limit); 6185 // return ofs (int) 6186 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6187 set_result(result); 6188 6189 return true; 6190 } 6191 6192 //------------------------------get_state_from_sha_object----------------------- 6193 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6194 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6195 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6196 if (sha_state == NULL) return (Node *) NULL; 6197 6198 // now have the array, need to get the start address of the state array 6199 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6200 return state; 6201 } 6202 6203 //------------------------------get_state_from_sha5_object----------------------- 6204 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6205 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6206 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6207 if (sha_state == NULL) return (Node *) NULL; 6208 6209 // now have the array, need to get the start address of the state array 6210 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6211 return state; 6212 } 6213 6214 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6215 // Return node representing slow path of predicate check. 6216 // the pseudo code we want to emulate with this predicate is: 6217 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6218 // 6219 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6220 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6221 "need SHA1/SHA256/SHA512 instruction support"); 6222 assert((uint)predicate < 3, "sanity"); 6223 6224 // The receiver was checked for NULL already. 6225 Node* digestBaseObj = argument(0); 6226 6227 // get DigestBase klass for instanceOf check 6228 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6229 assert(tinst != NULL, "digestBaseObj is null"); 6230 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6231 6232 const char* klass_SHA_name = NULL; 6233 switch (predicate) { 6234 case 0: 6235 if (UseSHA1Intrinsics) { 6236 // we want to do an instanceof comparison against the SHA class 6237 klass_SHA_name = "sun/security/provider/SHA"; 6238 } 6239 break; 6240 case 1: 6241 if (UseSHA256Intrinsics) { 6242 // we want to do an instanceof comparison against the SHA2 class 6243 klass_SHA_name = "sun/security/provider/SHA2"; 6244 } 6245 break; 6246 case 2: 6247 if (UseSHA512Intrinsics) { 6248 // we want to do an instanceof comparison against the SHA5 class 6249 klass_SHA_name = "sun/security/provider/SHA5"; 6250 } 6251 break; 6252 default: 6253 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6254 } 6255 6256 ciKlass* klass_SHA = NULL; 6257 if (klass_SHA_name != NULL) { 6258 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6259 } 6260 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6261 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6262 Node* ctrl = control(); 6263 set_control(top()); // no intrinsic path 6264 return ctrl; 6265 } 6266 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6267 6268 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6269 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6270 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6271 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6272 6273 return instof_false; // even if it is NULL 6274 } 6275 6276 bool LibraryCallKit::inline_profileBoolean() { 6277 Node* counts = argument(1); 6278 const TypeAryPtr* ary = NULL; 6279 ciArray* aobj = NULL; 6280 if (counts->is_Con() 6281 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6282 && (aobj = ary->const_oop()->as_array()) != NULL 6283 && (aobj->length() == 2)) { 6284 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6285 jint false_cnt = aobj->element_value(0).as_int(); 6286 jint true_cnt = aobj->element_value(1).as_int(); 6287 6288 if (C->log() != NULL) { 6289 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6290 false_cnt, true_cnt); 6291 } 6292 6293 if (false_cnt + true_cnt == 0) { 6294 // According to profile, never executed. 6295 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6296 Deoptimization::Action_reinterpret); 6297 return true; 6298 } 6299 6300 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6301 // is a number of each value occurrences. 6302 Node* result = argument(0); 6303 if (false_cnt == 0 || true_cnt == 0) { 6304 // According to profile, one value has been never seen. 6305 int expected_val = (false_cnt == 0) ? 1 : 0; 6306 6307 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6308 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6309 6310 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6311 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6312 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6313 6314 { // Slow path: uncommon trap for never seen value and then reexecute 6315 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6316 // the value has been seen at least once. 6317 PreserveJVMState pjvms(this); 6318 PreserveReexecuteState preexecs(this); 6319 jvms()->set_should_reexecute(true); 6320 6321 set_control(slow_path); 6322 set_i_o(i_o()); 6323 6324 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6325 Deoptimization::Action_reinterpret); 6326 } 6327 // The guard for never seen value enables sharpening of the result and 6328 // returning a constant. It allows to eliminate branches on the same value 6329 // later on. 6330 set_control(fast_path); 6331 result = intcon(expected_val); 6332 } 6333 // Stop profiling. 6334 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6335 // By replacing method body with profile data (represented as ProfileBooleanNode 6336 // on IR level) we effectively disable profiling. 6337 // It enables full speed execution once optimized code is generated. 6338 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6339 C->record_for_igvn(profile); 6340 set_result(profile); 6341 return true; 6342 } else { 6343 // Continue profiling. 6344 // Profile data isn't available at the moment. So, execute method's bytecode version. 6345 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6346 // is compiled and counters aren't available since corresponding MethodHandle 6347 // isn't a compile-time constant. 6348 return false; 6349 } 6350 } 6351 6352 bool LibraryCallKit::inline_isCompileConstant() { 6353 Node* n = argument(0); 6354 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6355 return true; 6356 }