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 bool inline_notify(vmIntrinsics::ID id); 229 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y); 230 // This returns Type::AnyPtr, RawPtr, or OopPtr. 231 int classify_unsafe_addr(Node* &base, Node* &offset); 232 Node* make_unsafe_address(Node* base, Node* offset); 233 // Helper for inline_unsafe_access. 234 // Generates the guards that check whether the result of 235 // Unsafe.getObject should be recorded in an SATB log buffer. 236 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar); 237 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile); 238 static bool klass_needs_init_guard(Node* kls); 239 bool inline_unsafe_allocate(); 240 bool inline_unsafe_copyMemory(); 241 bool inline_native_currentThread(); 242 #ifdef TRACE_HAVE_INTRINSICS 243 bool inline_native_classID(); 244 bool inline_native_threadID(); 245 #endif 246 bool inline_native_time_funcs(address method, const char* funcName); 247 bool inline_native_isInterrupted(); 248 bool inline_native_Class_query(vmIntrinsics::ID id); 249 bool inline_native_subtype_check(); 250 251 bool inline_native_newArray(); 252 bool inline_native_getLength(); 253 bool inline_array_copyOf(bool is_copyOfRange); 254 bool inline_array_equals(); 255 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark); 256 bool inline_native_clone(bool is_virtual); 257 bool inline_native_Reflection_getCallerClass(); 258 // Helper function for inlining native object hash method 259 bool inline_native_hashcode(bool is_virtual, bool is_static); 260 bool inline_native_getClass(); 261 262 // Helper functions for inlining arraycopy 263 bool inline_arraycopy(); 264 AllocateArrayNode* tightly_coupled_allocation(Node* ptr, 265 RegionNode* slow_region); 266 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp); 267 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp); 268 269 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind; 270 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind); 271 bool inline_unsafe_ordered_store(BasicType type); 272 bool inline_unsafe_fence(vmIntrinsics::ID id); 273 bool inline_fp_conversions(vmIntrinsics::ID id); 274 bool inline_number_methods(vmIntrinsics::ID id); 275 bool inline_reference_get(); 276 bool inline_Class_cast(); 277 bool inline_aescrypt_Block(vmIntrinsics::ID id); 278 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 279 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 280 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 281 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 282 bool inline_ghash_processBlocks(); 283 bool inline_sha_implCompress(vmIntrinsics::ID id); 284 bool inline_digestBase_implCompressMB(int predicate); 285 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 286 bool long_state, address stubAddr, const char *stubName, 287 Node* src_start, Node* ofs, Node* limit); 288 Node* get_state_from_sha_object(Node *sha_object); 289 Node* get_state_from_sha5_object(Node *sha_object); 290 Node* inline_digestBase_implCompressMB_predicate(int predicate); 291 bool inline_encodeISOArray(); 292 bool inline_updateCRC32(); 293 bool inline_updateBytesCRC32(); 294 bool inline_updateByteBufferCRC32(); 295 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class); 296 bool inline_updateBytesCRC32C(); 297 bool inline_updateDirectByteBufferCRC32C(); 298 bool inline_multiplyToLen(); 299 bool inline_squareToLen(); 300 bool inline_mulAdd(); 301 302 bool inline_profileBoolean(); 303 bool inline_isCompileConstant(); 304 }; 305 306 307 //---------------------------make_vm_intrinsic---------------------------- 308 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 309 vmIntrinsics::ID id = m->intrinsic_id(); 310 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 311 312 ccstr disable_intr = NULL; 313 314 if ((DisableIntrinsic[0] != '\0' 315 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 316 (method_has_option_value("DisableIntrinsic", disable_intr) 317 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 318 // disabled by a user request on the command line: 319 // example: -XX:DisableIntrinsic=_hashCode,_getClass 320 return NULL; 321 } 322 323 if (!m->is_loaded()) { 324 // do not attempt to inline unloaded methods 325 return NULL; 326 } 327 328 // Only a few intrinsics implement a virtual dispatch. 329 // They are expensive calls which are also frequently overridden. 330 if (is_virtual) { 331 switch (id) { 332 case vmIntrinsics::_hashCode: 333 case vmIntrinsics::_clone: 334 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 335 break; 336 default: 337 return NULL; 338 } 339 } 340 341 // -XX:-InlineNatives disables nearly all intrinsics: 342 if (!InlineNatives) { 343 switch (id) { 344 case vmIntrinsics::_indexOf: 345 case vmIntrinsics::_compareTo: 346 case vmIntrinsics::_equals: 347 case vmIntrinsics::_equalsC: 348 case vmIntrinsics::_getAndAddInt: 349 case vmIntrinsics::_getAndAddLong: 350 case vmIntrinsics::_getAndSetInt: 351 case vmIntrinsics::_getAndSetLong: 352 case vmIntrinsics::_getAndSetObject: 353 case vmIntrinsics::_loadFence: 354 case vmIntrinsics::_storeFence: 355 case vmIntrinsics::_fullFence: 356 break; // InlineNatives does not control String.compareTo 357 case vmIntrinsics::_Reference_get: 358 break; // InlineNatives does not control Reference.get 359 default: 360 return NULL; 361 } 362 } 363 364 int predicates = 0; 365 bool does_virtual_dispatch = false; 366 367 switch (id) { 368 case vmIntrinsics::_compareTo: 369 if (!SpecialStringCompareTo) return NULL; 370 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 371 break; 372 case vmIntrinsics::_indexOf: 373 if (!SpecialStringIndexOf) return NULL; 374 break; 375 case vmIntrinsics::_equals: 376 if (!SpecialStringEquals) return NULL; 377 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 378 break; 379 case vmIntrinsics::_equalsC: 380 if (!SpecialArraysEquals) return NULL; 381 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 382 break; 383 case vmIntrinsics::_arraycopy: 384 if (!InlineArrayCopy) return NULL; 385 break; 386 case vmIntrinsics::_copyMemory: 387 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 388 if (!InlineArrayCopy) return NULL; 389 break; 390 case vmIntrinsics::_hashCode: 391 if (!InlineObjectHash) return NULL; 392 does_virtual_dispatch = true; 393 break; 394 case vmIntrinsics::_clone: 395 does_virtual_dispatch = true; 396 case vmIntrinsics::_copyOf: 397 case vmIntrinsics::_copyOfRange: 398 if (!InlineObjectCopy) return NULL; 399 // These also use the arraycopy intrinsic mechanism: 400 if (!InlineArrayCopy) return NULL; 401 break; 402 case vmIntrinsics::_encodeISOArray: 403 if (!SpecialEncodeISOArray) return NULL; 404 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 405 break; 406 case vmIntrinsics::_checkIndex: 407 // We do not intrinsify this. The optimizer does fine with it. 408 return NULL; 409 410 case vmIntrinsics::_getCallerClass: 411 if (!InlineReflectionGetCallerClass) return NULL; 412 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 413 break; 414 415 case vmIntrinsics::_bitCount_i: 416 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 417 break; 418 419 case vmIntrinsics::_bitCount_l: 420 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 421 break; 422 423 case vmIntrinsics::_numberOfLeadingZeros_i: 424 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 425 break; 426 427 case vmIntrinsics::_numberOfLeadingZeros_l: 428 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 429 break; 430 431 case vmIntrinsics::_numberOfTrailingZeros_i: 432 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 433 break; 434 435 case vmIntrinsics::_numberOfTrailingZeros_l: 436 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 437 break; 438 439 case vmIntrinsics::_reverseBytes_c: 440 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 441 break; 442 case vmIntrinsics::_reverseBytes_s: 443 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 444 break; 445 case vmIntrinsics::_reverseBytes_i: 446 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 447 break; 448 case vmIntrinsics::_reverseBytes_l: 449 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 450 break; 451 452 case vmIntrinsics::_Reference_get: 453 // Use the intrinsic version of Reference.get() so that the value in 454 // the referent field can be registered by the G1 pre-barrier code. 455 // Also add memory barrier to prevent commoning reads from this field 456 // across safepoint since GC can change it value. 457 break; 458 459 case vmIntrinsics::_compareAndSwapObject: 460 #ifdef _LP64 461 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 462 #endif 463 break; 464 465 case vmIntrinsics::_compareAndSwapLong: 466 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 467 break; 468 469 case vmIntrinsics::_getAndAddInt: 470 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 471 break; 472 473 case vmIntrinsics::_getAndAddLong: 474 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 475 break; 476 477 case vmIntrinsics::_getAndSetInt: 478 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 479 break; 480 481 case vmIntrinsics::_getAndSetLong: 482 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 483 break; 484 485 case vmIntrinsics::_getAndSetObject: 486 #ifdef _LP64 487 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 488 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 489 break; 490 #else 491 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 492 break; 493 #endif 494 495 case vmIntrinsics::_aescrypt_encryptBlock: 496 case vmIntrinsics::_aescrypt_decryptBlock: 497 if (!UseAESIntrinsics) return NULL; 498 break; 499 500 case vmIntrinsics::_multiplyToLen: 501 if (!UseMultiplyToLenIntrinsic) return NULL; 502 break; 503 504 case vmIntrinsics::_squareToLen: 505 if (!UseSquareToLenIntrinsic) return NULL; 506 break; 507 508 case vmIntrinsics::_mulAdd: 509 if (!UseMulAddIntrinsic) return NULL; 510 break; 511 512 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 513 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 514 if (!UseAESIntrinsics) return NULL; 515 // these two require the predicated logic 516 predicates = 1; 517 break; 518 519 case vmIntrinsics::_sha_implCompress: 520 if (!UseSHA1Intrinsics) return NULL; 521 break; 522 523 case vmIntrinsics::_sha2_implCompress: 524 if (!UseSHA256Intrinsics) return NULL; 525 break; 526 527 case vmIntrinsics::_sha5_implCompress: 528 if (!UseSHA512Intrinsics) return NULL; 529 break; 530 531 case vmIntrinsics::_digestBase_implCompressMB: 532 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 533 predicates = 3; 534 break; 535 536 case vmIntrinsics::_ghash_processBlocks: 537 if (!UseGHASHIntrinsics) return NULL; 538 break; 539 540 case vmIntrinsics::_updateCRC32: 541 case vmIntrinsics::_updateBytesCRC32: 542 case vmIntrinsics::_updateByteBufferCRC32: 543 if (!UseCRC32Intrinsics) return NULL; 544 break; 545 546 case vmIntrinsics::_updateBytesCRC32C: 547 case vmIntrinsics::_updateDirectByteBufferCRC32C: 548 if (!UseCRC32CIntrinsics) return NULL; 549 break; 550 551 case vmIntrinsics::_incrementExactI: 552 case vmIntrinsics::_addExactI: 553 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 554 break; 555 case vmIntrinsics::_incrementExactL: 556 case vmIntrinsics::_addExactL: 557 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 558 break; 559 case vmIntrinsics::_decrementExactI: 560 case vmIntrinsics::_subtractExactI: 561 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 562 break; 563 case vmIntrinsics::_decrementExactL: 564 case vmIntrinsics::_subtractExactL: 565 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 566 break; 567 case vmIntrinsics::_negateExactI: 568 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 569 break; 570 case vmIntrinsics::_negateExactL: 571 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 572 break; 573 case vmIntrinsics::_multiplyExactI: 574 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 575 break; 576 case vmIntrinsics::_multiplyExactL: 577 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 578 break; 579 580 case vmIntrinsics::_getShortUnaligned: 581 case vmIntrinsics::_getCharUnaligned: 582 case vmIntrinsics::_getIntUnaligned: 583 case vmIntrinsics::_getLongUnaligned: 584 case vmIntrinsics::_putShortUnaligned: 585 case vmIntrinsics::_putCharUnaligned: 586 case vmIntrinsics::_putIntUnaligned: 587 case vmIntrinsics::_putLongUnaligned: 588 if (!UseUnalignedAccesses) return NULL; 589 break; 590 591 default: 592 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 593 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 594 break; 595 } 596 597 // -XX:-InlineClassNatives disables natives from the Class class. 598 // The flag applies to all reflective calls, notably Array.newArray 599 // (visible to Java programmers as Array.newInstance). 600 if (m->holder()->name() == ciSymbol::java_lang_Class() || 601 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 602 if (!InlineClassNatives) return NULL; 603 } 604 605 // -XX:-InlineThreadNatives disables natives from the Thread class. 606 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 607 if (!InlineThreadNatives) return NULL; 608 } 609 610 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 611 if (m->holder()->name() == ciSymbol::java_lang_Math() || 612 m->holder()->name() == ciSymbol::java_lang_Float() || 613 m->holder()->name() == ciSymbol::java_lang_Double()) { 614 if (!InlineMathNatives) return NULL; 615 } 616 617 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 618 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 619 if (!InlineUnsafeOps) return NULL; 620 } 621 622 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 623 } 624 625 //----------------------register_library_intrinsics----------------------- 626 // Initialize this file's data structures, for each Compile instance. 627 void Compile::register_library_intrinsics() { 628 // Nothing to do here. 629 } 630 631 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 632 LibraryCallKit kit(jvms, this); 633 Compile* C = kit.C; 634 int nodes = C->unique(); 635 #ifndef PRODUCT 636 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 637 char buf[1000]; 638 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 639 tty->print_cr("Intrinsic %s", str); 640 } 641 #endif 642 ciMethod* callee = kit.callee(); 643 const int bci = kit.bci(); 644 645 // Try to inline the intrinsic. 646 if (kit.try_to_inline(_last_predicate)) { 647 if (C->print_intrinsics() || C->print_inlining()) { 648 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 649 } 650 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 651 if (C->log()) { 652 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 653 vmIntrinsics::name_at(intrinsic_id()), 654 (is_virtual() ? " virtual='1'" : ""), 655 C->unique() - nodes); 656 } 657 // Push the result from the inlined method onto the stack. 658 kit.push_result(); 659 C->print_inlining_update(this); 660 return kit.transfer_exceptions_into_jvms(); 661 } 662 663 // The intrinsic bailed out 664 if (C->print_intrinsics() || C->print_inlining()) { 665 if (jvms->has_method()) { 666 // Not a root compile. 667 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 668 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 669 } else { 670 // Root compile 671 tty->print("Did not generate intrinsic %s%s at bci:%d in", 672 vmIntrinsics::name_at(intrinsic_id()), 673 (is_virtual() ? " (virtual)" : ""), bci); 674 } 675 } 676 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 677 C->print_inlining_update(this); 678 return NULL; 679 } 680 681 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 682 LibraryCallKit kit(jvms, this); 683 Compile* C = kit.C; 684 int nodes = C->unique(); 685 _last_predicate = predicate; 686 #ifndef PRODUCT 687 assert(is_predicated() && predicate < predicates_count(), "sanity"); 688 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 689 char buf[1000]; 690 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 691 tty->print_cr("Predicate for intrinsic %s", str); 692 } 693 #endif 694 ciMethod* callee = kit.callee(); 695 const int bci = kit.bci(); 696 697 Node* slow_ctl = kit.try_to_predicate(predicate); 698 if (!kit.failing()) { 699 if (C->print_intrinsics() || C->print_inlining()) { 700 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 701 } 702 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 703 if (C->log()) { 704 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 705 vmIntrinsics::name_at(intrinsic_id()), 706 (is_virtual() ? " virtual='1'" : ""), 707 C->unique() - nodes); 708 } 709 return slow_ctl; // Could be NULL if the check folds. 710 } 711 712 // The intrinsic bailed out 713 if (C->print_intrinsics() || C->print_inlining()) { 714 if (jvms->has_method()) { 715 // Not a root compile. 716 const char* msg = "failed to generate predicate for intrinsic"; 717 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 718 } else { 719 // Root compile 720 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 721 vmIntrinsics::name_at(intrinsic_id()), 722 (is_virtual() ? " (virtual)" : ""), bci); 723 } 724 } 725 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 726 return NULL; 727 } 728 729 bool LibraryCallKit::try_to_inline(int predicate) { 730 // Handle symbolic names for otherwise undistinguished boolean switches: 731 const bool is_store = true; 732 const bool is_native_ptr = true; 733 const bool is_static = true; 734 const bool is_volatile = true; 735 736 if (!jvms()->has_method()) { 737 // Root JVMState has a null method. 738 assert(map()->memory()->Opcode() == Op_Parm, ""); 739 // Insert the memory aliasing node 740 set_all_memory(reset_memory()); 741 } 742 assert(merged_memory(), ""); 743 744 745 switch (intrinsic_id()) { 746 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 747 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 748 case vmIntrinsics::_getClass: return inline_native_getClass(); 749 750 case vmIntrinsics::_dsin: 751 case vmIntrinsics::_dcos: 752 case vmIntrinsics::_dtan: 753 case vmIntrinsics::_dabs: 754 case vmIntrinsics::_datan2: 755 case vmIntrinsics::_dsqrt: 756 case vmIntrinsics::_dexp: 757 case vmIntrinsics::_dlog: 758 case vmIntrinsics::_dlog10: 759 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 760 761 case vmIntrinsics::_min: 762 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 763 764 case vmIntrinsics::_notify: 765 case vmIntrinsics::_notifyAll: 766 if (InlineNotify) { 767 return inline_notify(intrinsic_id()); 768 } 769 return false; 770 771 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 772 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 773 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 774 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 775 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 776 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 777 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 778 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 779 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 780 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 781 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 782 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 783 784 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 785 786 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 787 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 788 case vmIntrinsics::_equals: return inline_string_equals(); 789 790 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 791 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 792 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 793 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 794 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 795 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 796 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 797 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 798 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 799 800 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 801 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 802 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 803 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 804 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 805 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 806 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 807 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 808 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 809 810 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 811 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 812 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 813 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 814 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 815 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 816 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 817 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 818 819 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 820 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 821 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 822 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 823 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 824 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 825 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 826 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 827 828 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 829 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 830 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 831 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 832 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 833 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 834 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 835 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 836 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 837 838 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 839 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 840 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 841 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 842 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 843 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 844 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 845 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 846 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 847 848 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 849 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 850 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 851 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 852 853 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 854 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 855 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 856 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 857 858 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 859 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 860 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 861 862 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 863 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 864 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 865 866 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 867 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 868 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 869 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 870 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 871 872 case vmIntrinsics::_loadFence: 873 case vmIntrinsics::_storeFence: 874 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 875 876 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 877 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 878 879 #ifdef TRACE_HAVE_INTRINSICS 880 case vmIntrinsics::_classID: return inline_native_classID(); 881 case vmIntrinsics::_threadID: return inline_native_threadID(); 882 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 883 #endif 884 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 885 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 886 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 887 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 888 case vmIntrinsics::_newArray: return inline_native_newArray(); 889 case vmIntrinsics::_getLength: return inline_native_getLength(); 890 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 891 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 892 case vmIntrinsics::_equalsC: return inline_array_equals(); 893 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 894 895 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 896 897 case vmIntrinsics::_isInstance: 898 case vmIntrinsics::_getModifiers: 899 case vmIntrinsics::_isInterface: 900 case vmIntrinsics::_isArray: 901 case vmIntrinsics::_isPrimitive: 902 case vmIntrinsics::_getSuperclass: 903 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 904 905 case vmIntrinsics::_floatToRawIntBits: 906 case vmIntrinsics::_floatToIntBits: 907 case vmIntrinsics::_intBitsToFloat: 908 case vmIntrinsics::_doubleToRawLongBits: 909 case vmIntrinsics::_doubleToLongBits: 910 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 911 912 case vmIntrinsics::_numberOfLeadingZeros_i: 913 case vmIntrinsics::_numberOfLeadingZeros_l: 914 case vmIntrinsics::_numberOfTrailingZeros_i: 915 case vmIntrinsics::_numberOfTrailingZeros_l: 916 case vmIntrinsics::_bitCount_i: 917 case vmIntrinsics::_bitCount_l: 918 case vmIntrinsics::_reverseBytes_i: 919 case vmIntrinsics::_reverseBytes_l: 920 case vmIntrinsics::_reverseBytes_s: 921 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 922 923 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 924 925 case vmIntrinsics::_Reference_get: return inline_reference_get(); 926 927 case vmIntrinsics::_Class_cast: return inline_Class_cast(); 928 929 case vmIntrinsics::_aescrypt_encryptBlock: 930 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 931 932 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 933 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 934 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 935 936 case vmIntrinsics::_sha_implCompress: 937 case vmIntrinsics::_sha2_implCompress: 938 case vmIntrinsics::_sha5_implCompress: 939 return inline_sha_implCompress(intrinsic_id()); 940 941 case vmIntrinsics::_digestBase_implCompressMB: 942 return inline_digestBase_implCompressMB(predicate); 943 944 case vmIntrinsics::_multiplyToLen: 945 return inline_multiplyToLen(); 946 947 case vmIntrinsics::_squareToLen: 948 return inline_squareToLen(); 949 950 case vmIntrinsics::_mulAdd: 951 return inline_mulAdd(); 952 953 case vmIntrinsics::_ghash_processBlocks: 954 return inline_ghash_processBlocks(); 955 956 case vmIntrinsics::_encodeISOArray: 957 return inline_encodeISOArray(); 958 959 case vmIntrinsics::_updateCRC32: 960 return inline_updateCRC32(); 961 case vmIntrinsics::_updateBytesCRC32: 962 return inline_updateBytesCRC32(); 963 case vmIntrinsics::_updateByteBufferCRC32: 964 return inline_updateByteBufferCRC32(); 965 966 case vmIntrinsics::_updateBytesCRC32C: 967 return inline_updateBytesCRC32C(); 968 case vmIntrinsics::_updateDirectByteBufferCRC32C: 969 return inline_updateDirectByteBufferCRC32C(); 970 971 case vmIntrinsics::_profileBoolean: 972 return inline_profileBoolean(); 973 case vmIntrinsics::_isCompileConstant: 974 return inline_isCompileConstant(); 975 976 default: 977 // If you get here, it may be that someone has added a new intrinsic 978 // to the list in vmSymbols.hpp without implementing it here. 979 #ifndef PRODUCT 980 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 981 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 982 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 983 } 984 #endif 985 return false; 986 } 987 } 988 989 Node* LibraryCallKit::try_to_predicate(int predicate) { 990 if (!jvms()->has_method()) { 991 // Root JVMState has a null method. 992 assert(map()->memory()->Opcode() == Op_Parm, ""); 993 // Insert the memory aliasing node 994 set_all_memory(reset_memory()); 995 } 996 assert(merged_memory(), ""); 997 998 switch (intrinsic_id()) { 999 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 1000 return inline_cipherBlockChaining_AESCrypt_predicate(false); 1001 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 1002 return inline_cipherBlockChaining_AESCrypt_predicate(true); 1003 case vmIntrinsics::_digestBase_implCompressMB: 1004 return inline_digestBase_implCompressMB_predicate(predicate); 1005 1006 default: 1007 // If you get here, it may be that someone has added a new intrinsic 1008 // to the list in vmSymbols.hpp without implementing it here. 1009 #ifndef PRODUCT 1010 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 1011 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 1012 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 1013 } 1014 #endif 1015 Node* slow_ctl = control(); 1016 set_control(top()); // No fast path instrinsic 1017 return slow_ctl; 1018 } 1019 } 1020 1021 //------------------------------set_result------------------------------- 1022 // Helper function for finishing intrinsics. 1023 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 1024 record_for_igvn(region); 1025 set_control(_gvn.transform(region)); 1026 set_result( _gvn.transform(value)); 1027 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 1028 } 1029 1030 //------------------------------generate_guard--------------------------- 1031 // Helper function for generating guarded fast-slow graph structures. 1032 // The given 'test', if true, guards a slow path. If the test fails 1033 // then a fast path can be taken. (We generally hope it fails.) 1034 // In all cases, GraphKit::control() is updated to the fast path. 1035 // The returned value represents the control for the slow path. 1036 // The return value is never 'top'; it is either a valid control 1037 // or NULL if it is obvious that the slow path can never be taken. 1038 // Also, if region and the slow control are not NULL, the slow edge 1039 // is appended to the region. 1040 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 1041 if (stopped()) { 1042 // Already short circuited. 1043 return NULL; 1044 } 1045 1046 // Build an if node and its projections. 1047 // If test is true we take the slow path, which we assume is uncommon. 1048 if (_gvn.type(test) == TypeInt::ZERO) { 1049 // The slow branch is never taken. No need to build this guard. 1050 return NULL; 1051 } 1052 1053 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 1054 1055 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 1056 if (if_slow == top()) { 1057 // The slow branch is never taken. No need to build this guard. 1058 return NULL; 1059 } 1060 1061 if (region != NULL) 1062 region->add_req(if_slow); 1063 1064 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 1065 set_control(if_fast); 1066 1067 return if_slow; 1068 } 1069 1070 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 1071 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 1072 } 1073 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 1074 return generate_guard(test, region, PROB_FAIR); 1075 } 1076 1077 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1078 Node* *pos_index) { 1079 if (stopped()) 1080 return NULL; // already stopped 1081 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1082 return NULL; // index is already adequately typed 1083 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1084 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1085 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1086 if (is_neg != NULL && pos_index != NULL) { 1087 // Emulate effect of Parse::adjust_map_after_if. 1088 Node* ccast = new CastIINode(index, TypeInt::POS); 1089 ccast->set_req(0, control()); 1090 (*pos_index) = _gvn.transform(ccast); 1091 } 1092 return is_neg; 1093 } 1094 1095 // Make sure that 'position' is a valid limit index, in [0..length]. 1096 // There are two equivalent plans for checking this: 1097 // A. (offset + copyLength) unsigned<= arrayLength 1098 // B. offset <= (arrayLength - copyLength) 1099 // We require that all of the values above, except for the sum and 1100 // difference, are already known to be non-negative. 1101 // Plan A is robust in the face of overflow, if offset and copyLength 1102 // are both hugely positive. 1103 // 1104 // Plan B is less direct and intuitive, but it does not overflow at 1105 // all, since the difference of two non-negatives is always 1106 // representable. Whenever Java methods must perform the equivalent 1107 // check they generally use Plan B instead of Plan A. 1108 // For the moment we use Plan A. 1109 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1110 Node* subseq_length, 1111 Node* array_length, 1112 RegionNode* region) { 1113 if (stopped()) 1114 return NULL; // already stopped 1115 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1116 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1117 return NULL; // common case of whole-array copy 1118 Node* last = subseq_length; 1119 if (!zero_offset) // last += offset 1120 last = _gvn.transform(new AddINode(last, offset)); 1121 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1122 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1123 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1124 return is_over; 1125 } 1126 1127 1128 //--------------------------generate_current_thread-------------------- 1129 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1130 ciKlass* thread_klass = env()->Thread_klass(); 1131 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1132 Node* thread = _gvn.transform(new ThreadLocalNode()); 1133 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1134 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1135 tls_output = thread; 1136 return threadObj; 1137 } 1138 1139 1140 //------------------------------make_string_method_node------------------------ 1141 // Helper method for String intrinsic functions. This version is called 1142 // with str1 and str2 pointing to String object nodes. 1143 // 1144 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1145 Node* no_ctrl = NULL; 1146 1147 // Get start addr of string 1148 Node* str1_value = load_String_value(no_ctrl, str1); 1149 Node* str1_offset = load_String_offset(no_ctrl, str1); 1150 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1151 1152 // Get length of string 1 1153 Node* str1_len = load_String_length(no_ctrl, str1); 1154 1155 Node* str2_value = load_String_value(no_ctrl, str2); 1156 Node* str2_offset = load_String_offset(no_ctrl, str2); 1157 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1158 1159 Node* str2_len = NULL; 1160 Node* result = NULL; 1161 1162 switch (opcode) { 1163 case Op_StrIndexOf: 1164 // Get length of string 2 1165 str2_len = load_String_length(no_ctrl, str2); 1166 1167 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1168 str1_start, str1_len, str2_start, str2_len); 1169 break; 1170 case Op_StrComp: 1171 // Get length of string 2 1172 str2_len = load_String_length(no_ctrl, str2); 1173 1174 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1175 str1_start, str1_len, str2_start, str2_len); 1176 break; 1177 case Op_StrEquals: 1178 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1179 str1_start, str2_start, str1_len); 1180 break; 1181 default: 1182 ShouldNotReachHere(); 1183 return NULL; 1184 } 1185 1186 // All these intrinsics have checks. 1187 C->set_has_split_ifs(true); // Has chance for split-if optimization 1188 1189 return _gvn.transform(result); 1190 } 1191 1192 // Helper method for String intrinsic functions. This version is called 1193 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1194 // to Int nodes containing the lenghts of str1 and str2. 1195 // 1196 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1197 Node* result = NULL; 1198 switch (opcode) { 1199 case Op_StrIndexOf: 1200 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1201 str1_start, cnt1, str2_start, cnt2); 1202 break; 1203 case Op_StrComp: 1204 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1205 str1_start, cnt1, str2_start, cnt2); 1206 break; 1207 case Op_StrEquals: 1208 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1209 str1_start, str2_start, cnt1); 1210 break; 1211 default: 1212 ShouldNotReachHere(); 1213 return NULL; 1214 } 1215 1216 // All these intrinsics have checks. 1217 C->set_has_split_ifs(true); // Has chance for split-if optimization 1218 1219 return _gvn.transform(result); 1220 } 1221 1222 //------------------------------inline_string_compareTo------------------------ 1223 // public int java.lang.String.compareTo(String anotherString); 1224 bool LibraryCallKit::inline_string_compareTo() { 1225 Node* receiver = null_check(argument(0)); 1226 Node* arg = null_check(argument(1)); 1227 if (stopped()) { 1228 return true; 1229 } 1230 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1231 return true; 1232 } 1233 1234 //------------------------------inline_string_equals------------------------ 1235 bool LibraryCallKit::inline_string_equals() { 1236 Node* receiver = null_check_receiver(); 1237 // NOTE: Do not null check argument for String.equals() because spec 1238 // allows to specify NULL as argument. 1239 Node* argument = this->argument(1); 1240 if (stopped()) { 1241 return true; 1242 } 1243 1244 // paths (plus control) merge 1245 RegionNode* region = new RegionNode(5); 1246 Node* phi = new PhiNode(region, TypeInt::BOOL); 1247 1248 // does source == target string? 1249 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1250 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1251 1252 Node* if_eq = generate_slow_guard(bol, NULL); 1253 if (if_eq != NULL) { 1254 // receiver == argument 1255 phi->init_req(2, intcon(1)); 1256 region->init_req(2, if_eq); 1257 } 1258 1259 // get String klass for instanceOf 1260 ciInstanceKlass* klass = env()->String_klass(); 1261 1262 if (!stopped()) { 1263 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1264 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1265 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1266 1267 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1268 //instanceOf == true, fallthrough 1269 1270 if (inst_false != NULL) { 1271 phi->init_req(3, intcon(0)); 1272 region->init_req(3, inst_false); 1273 } 1274 } 1275 1276 if (!stopped()) { 1277 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1278 1279 // Properly cast the argument to String 1280 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1281 // This path is taken only when argument's type is String:NotNull. 1282 argument = cast_not_null(argument, false); 1283 1284 Node* no_ctrl = NULL; 1285 1286 // Get start addr of receiver 1287 Node* receiver_val = load_String_value(no_ctrl, receiver); 1288 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1289 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1290 1291 // Get length of receiver 1292 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1293 1294 // Get start addr of argument 1295 Node* argument_val = load_String_value(no_ctrl, argument); 1296 Node* argument_offset = load_String_offset(no_ctrl, argument); 1297 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1298 1299 // Get length of argument 1300 Node* argument_cnt = load_String_length(no_ctrl, argument); 1301 1302 // Check for receiver count != argument count 1303 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1304 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1305 Node* if_ne = generate_slow_guard(bol, NULL); 1306 if (if_ne != NULL) { 1307 phi->init_req(4, intcon(0)); 1308 region->init_req(4, if_ne); 1309 } 1310 1311 // Check for count == 0 is done by assembler code for StrEquals. 1312 1313 if (!stopped()) { 1314 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1315 phi->init_req(1, equals); 1316 region->init_req(1, control()); 1317 } 1318 } 1319 1320 // post merge 1321 set_control(_gvn.transform(region)); 1322 record_for_igvn(region); 1323 1324 set_result(_gvn.transform(phi)); 1325 return true; 1326 } 1327 1328 //------------------------------inline_array_equals---------------------------- 1329 bool LibraryCallKit::inline_array_equals() { 1330 Node* arg1 = argument(0); 1331 Node* arg2 = argument(1); 1332 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1333 return true; 1334 } 1335 1336 // Java version of String.indexOf(constant string) 1337 // class StringDecl { 1338 // StringDecl(char[] ca) { 1339 // offset = 0; 1340 // count = ca.length; 1341 // value = ca; 1342 // } 1343 // int offset; 1344 // int count; 1345 // char[] value; 1346 // } 1347 // 1348 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1349 // int targetOffset, int cache_i, int md2) { 1350 // int cache = cache_i; 1351 // int sourceOffset = string_object.offset; 1352 // int sourceCount = string_object.count; 1353 // int targetCount = target_object.length; 1354 // 1355 // int targetCountLess1 = targetCount - 1; 1356 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1357 // 1358 // char[] source = string_object.value; 1359 // char[] target = target_object; 1360 // int lastChar = target[targetCountLess1]; 1361 // 1362 // outer_loop: 1363 // for (int i = sourceOffset; i < sourceEnd; ) { 1364 // int src = source[i + targetCountLess1]; 1365 // if (src == lastChar) { 1366 // // With random strings and a 4-character alphabet, 1367 // // reverse matching at this point sets up 0.8% fewer 1368 // // frames, but (paradoxically) makes 0.3% more probes. 1369 // // Since those probes are nearer the lastChar probe, 1370 // // there is may be a net D$ win with reverse matching. 1371 // // But, reversing loop inhibits unroll of inner loop 1372 // // for unknown reason. So, does running outer loop from 1373 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1374 // for (int j = 0; j < targetCountLess1; j++) { 1375 // if (target[targetOffset + j] != source[i+j]) { 1376 // if ((cache & (1 << source[i+j])) == 0) { 1377 // if (md2 < j+1) { 1378 // i += j+1; 1379 // continue outer_loop; 1380 // } 1381 // } 1382 // i += md2; 1383 // continue outer_loop; 1384 // } 1385 // } 1386 // return i - sourceOffset; 1387 // } 1388 // if ((cache & (1 << src)) == 0) { 1389 // i += targetCountLess1; 1390 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1391 // i++; 1392 // } 1393 // return -1; 1394 // } 1395 1396 //------------------------------string_indexOf------------------------ 1397 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1398 jint cache_i, jint md2_i) { 1399 1400 Node* no_ctrl = NULL; 1401 float likely = PROB_LIKELY(0.9); 1402 float unlikely = PROB_UNLIKELY(0.9); 1403 1404 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1405 1406 Node* source = load_String_value(no_ctrl, string_object); 1407 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1408 Node* sourceCount = load_String_length(no_ctrl, string_object); 1409 1410 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1411 jint target_length = target_array->length(); 1412 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1413 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1414 1415 // String.value field is known to be @Stable. 1416 if (UseImplicitStableValues) { 1417 target = cast_array_to_stable(target, target_type); 1418 } 1419 1420 IdealKit kit(this, false, true); 1421 #define __ kit. 1422 Node* zero = __ ConI(0); 1423 Node* one = __ ConI(1); 1424 Node* cache = __ ConI(cache_i); 1425 Node* md2 = __ ConI(md2_i); 1426 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1427 Node* targetCountLess1 = __ ConI(target_length - 1); 1428 Node* targetOffset = __ ConI(targetOffset_i); 1429 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1430 1431 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1432 Node* outer_loop = __ make_label(2 /* goto */); 1433 Node* return_ = __ make_label(1); 1434 1435 __ set(rtn,__ ConI(-1)); 1436 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1437 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1438 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1439 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1440 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1441 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1442 Node* tpj = __ AddI(targetOffset, __ value(j)); 1443 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1444 Node* ipj = __ AddI(__ value(i), __ value(j)); 1445 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1446 __ if_then(targ, BoolTest::ne, src2); { 1447 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1448 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1449 __ increment(i, __ AddI(__ value(j), one)); 1450 __ goto_(outer_loop); 1451 } __ end_if(); __ dead(j); 1452 }__ end_if(); __ dead(j); 1453 __ increment(i, md2); 1454 __ goto_(outer_loop); 1455 }__ end_if(); 1456 __ increment(j, one); 1457 }__ end_loop(); __ dead(j); 1458 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1459 __ goto_(return_); 1460 }__ end_if(); 1461 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1462 __ increment(i, targetCountLess1); 1463 }__ end_if(); 1464 __ increment(i, one); 1465 __ bind(outer_loop); 1466 }__ end_loop(); __ dead(i); 1467 __ bind(return_); 1468 1469 // Final sync IdealKit and GraphKit. 1470 final_sync(kit); 1471 Node* result = __ value(rtn); 1472 #undef __ 1473 C->set_has_loops(true); 1474 return result; 1475 } 1476 1477 //------------------------------inline_string_indexOf------------------------ 1478 bool LibraryCallKit::inline_string_indexOf() { 1479 Node* receiver = argument(0); 1480 Node* arg = argument(1); 1481 1482 Node* result; 1483 if (Matcher::has_match_rule(Op_StrIndexOf) && 1484 UseSSE42Intrinsics) { 1485 // Generate SSE4.2 version of indexOf 1486 // We currently only have match rules that use SSE4.2 1487 1488 receiver = null_check(receiver); 1489 arg = null_check(arg); 1490 if (stopped()) { 1491 return true; 1492 } 1493 1494 // Make the merge point 1495 RegionNode* result_rgn = new RegionNode(4); 1496 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1497 Node* no_ctrl = NULL; 1498 1499 // Get start addr of source string 1500 Node* source = load_String_value(no_ctrl, receiver); 1501 Node* source_offset = load_String_offset(no_ctrl, receiver); 1502 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1503 1504 // Get length of source string 1505 Node* source_cnt = load_String_length(no_ctrl, receiver); 1506 1507 // Get start addr of substring 1508 Node* substr = load_String_value(no_ctrl, arg); 1509 Node* substr_offset = load_String_offset(no_ctrl, arg); 1510 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1511 1512 // Get length of source string 1513 Node* substr_cnt = load_String_length(no_ctrl, arg); 1514 1515 // Check for substr count > string count 1516 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1517 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1518 Node* if_gt = generate_slow_guard(bol, NULL); 1519 if (if_gt != NULL) { 1520 result_phi->init_req(2, intcon(-1)); 1521 result_rgn->init_req(2, if_gt); 1522 } 1523 1524 if (!stopped()) { 1525 // Check for substr count == 0 1526 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1527 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1528 Node* if_zero = generate_slow_guard(bol, NULL); 1529 if (if_zero != NULL) { 1530 result_phi->init_req(3, intcon(0)); 1531 result_rgn->init_req(3, if_zero); 1532 } 1533 } 1534 1535 if (!stopped()) { 1536 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1537 result_phi->init_req(1, result); 1538 result_rgn->init_req(1, control()); 1539 } 1540 set_control(_gvn.transform(result_rgn)); 1541 record_for_igvn(result_rgn); 1542 result = _gvn.transform(result_phi); 1543 1544 } else { // Use LibraryCallKit::string_indexOf 1545 // don't intrinsify if argument isn't a constant string. 1546 if (!arg->is_Con()) { 1547 return false; 1548 } 1549 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1550 if (str_type == NULL) { 1551 return false; 1552 } 1553 ciInstanceKlass* klass = env()->String_klass(); 1554 ciObject* str_const = str_type->const_oop(); 1555 if (str_const == NULL || str_const->klass() != klass) { 1556 return false; 1557 } 1558 ciInstance* str = str_const->as_instance(); 1559 assert(str != NULL, "must be instance"); 1560 1561 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1562 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1563 1564 int o; 1565 int c; 1566 if (java_lang_String::has_offset_field()) { 1567 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1568 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1569 } else { 1570 o = 0; 1571 c = pat->length(); 1572 } 1573 1574 // constant strings have no offset and count == length which 1575 // simplifies the resulting code somewhat so lets optimize for that. 1576 if (o != 0 || c != pat->length()) { 1577 return false; 1578 } 1579 1580 receiver = null_check(receiver, T_OBJECT); 1581 // NOTE: No null check on the argument is needed since it's a constant String oop. 1582 if (stopped()) { 1583 return true; 1584 } 1585 1586 // The null string as a pattern always returns 0 (match at beginning of string) 1587 if (c == 0) { 1588 set_result(intcon(0)); 1589 return true; 1590 } 1591 1592 // Generate default indexOf 1593 jchar lastChar = pat->char_at(o + (c - 1)); 1594 int cache = 0; 1595 int i; 1596 for (i = 0; i < c - 1; i++) { 1597 assert(i < pat->length(), "out of range"); 1598 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1599 } 1600 1601 int md2 = c; 1602 for (i = 0; i < c - 1; i++) { 1603 assert(i < pat->length(), "out of range"); 1604 if (pat->char_at(o + i) == lastChar) { 1605 md2 = (c - 1) - i; 1606 } 1607 } 1608 1609 result = string_indexOf(receiver, pat, o, cache, md2); 1610 } 1611 set_result(result); 1612 return true; 1613 } 1614 1615 //--------------------------round_double_node-------------------------------- 1616 // Round a double node if necessary. 1617 Node* LibraryCallKit::round_double_node(Node* n) { 1618 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1619 n = _gvn.transform(new RoundDoubleNode(0, n)); 1620 return n; 1621 } 1622 1623 //------------------------------inline_math----------------------------------- 1624 // public static double Math.abs(double) 1625 // public static double Math.sqrt(double) 1626 // public static double Math.log(double) 1627 // public static double Math.log10(double) 1628 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1629 Node* arg = round_double_node(argument(0)); 1630 Node* n; 1631 switch (id) { 1632 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1633 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1634 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1635 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1636 default: fatal_unexpected_iid(id); break; 1637 } 1638 set_result(_gvn.transform(n)); 1639 return true; 1640 } 1641 1642 //------------------------------inline_trig---------------------------------- 1643 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1644 // argument reduction which will turn into a fast/slow diamond. 1645 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1646 Node* arg = round_double_node(argument(0)); 1647 Node* n = NULL; 1648 1649 switch (id) { 1650 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1651 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1652 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1653 default: fatal_unexpected_iid(id); break; 1654 } 1655 n = _gvn.transform(n); 1656 1657 // Rounding required? Check for argument reduction! 1658 if (Matcher::strict_fp_requires_explicit_rounding) { 1659 static const double pi_4 = 0.7853981633974483; 1660 static const double neg_pi_4 = -0.7853981633974483; 1661 // pi/2 in 80-bit extended precision 1662 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1663 // -pi/2 in 80-bit extended precision 1664 // 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}; 1665 // Cutoff value for using this argument reduction technique 1666 //static const double pi_2_minus_epsilon = 1.564660403643354; 1667 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1668 1669 // Pseudocode for sin: 1670 // if (x <= Math.PI / 4.0) { 1671 // if (x >= -Math.PI / 4.0) return fsin(x); 1672 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1673 // } else { 1674 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1675 // } 1676 // return StrictMath.sin(x); 1677 1678 // Pseudocode for cos: 1679 // if (x <= Math.PI / 4.0) { 1680 // if (x >= -Math.PI / 4.0) return fcos(x); 1681 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1682 // } else { 1683 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1684 // } 1685 // return StrictMath.cos(x); 1686 1687 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1688 // requires a special machine instruction to load it. Instead we'll try 1689 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1690 // probably do the math inside the SIN encoding. 1691 1692 // Make the merge point 1693 RegionNode* r = new RegionNode(3); 1694 Node* phi = new PhiNode(r, Type::DOUBLE); 1695 1696 // Flatten arg so we need only 1 test 1697 Node *abs = _gvn.transform(new AbsDNode(arg)); 1698 // Node for PI/4 constant 1699 Node *pi4 = makecon(TypeD::make(pi_4)); 1700 // Check PI/4 : abs(arg) 1701 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1702 // Check: If PI/4 < abs(arg) then go slow 1703 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1704 // Branch either way 1705 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1706 set_control(opt_iff(r,iff)); 1707 1708 // Set fast path result 1709 phi->init_req(2, n); 1710 1711 // Slow path - non-blocking leaf call 1712 Node* call = NULL; 1713 switch (id) { 1714 case vmIntrinsics::_dsin: 1715 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1716 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1717 "Sin", NULL, arg, top()); 1718 break; 1719 case vmIntrinsics::_dcos: 1720 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1721 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1722 "Cos", NULL, arg, top()); 1723 break; 1724 case vmIntrinsics::_dtan: 1725 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1726 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1727 "Tan", NULL, arg, top()); 1728 break; 1729 } 1730 assert(control()->in(0) == call, ""); 1731 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1732 r->init_req(1, control()); 1733 phi->init_req(1, slow_result); 1734 1735 // Post-merge 1736 set_control(_gvn.transform(r)); 1737 record_for_igvn(r); 1738 n = _gvn.transform(phi); 1739 1740 C->set_has_split_ifs(true); // Has chance for split-if optimization 1741 } 1742 set_result(n); 1743 return true; 1744 } 1745 1746 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1747 //------------------- 1748 //result=(result.isNaN())? funcAddr():result; 1749 // Check: If isNaN() by checking result!=result? then either trap 1750 // or go to runtime 1751 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1752 // Build the boolean node 1753 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1754 1755 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1756 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1757 // The pow or exp intrinsic returned a NaN, which requires a call 1758 // to the runtime. Recompile with the runtime call. 1759 uncommon_trap(Deoptimization::Reason_intrinsic, 1760 Deoptimization::Action_make_not_entrant); 1761 } 1762 return result; 1763 } else { 1764 // If this inlining ever returned NaN in the past, we compile a call 1765 // to the runtime to properly handle corner cases 1766 1767 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1768 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1769 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1770 1771 if (!if_slow->is_top()) { 1772 RegionNode* result_region = new RegionNode(3); 1773 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1774 1775 result_region->init_req(1, if_fast); 1776 result_val->init_req(1, result); 1777 1778 set_control(if_slow); 1779 1780 const TypePtr* no_memory_effects = NULL; 1781 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1782 no_memory_effects, 1783 x, top(), y, y ? top() : NULL); 1784 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1785 #ifdef ASSERT 1786 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1787 assert(value_top == top(), "second value must be top"); 1788 #endif 1789 1790 result_region->init_req(2, control()); 1791 result_val->init_req(2, value); 1792 set_control(_gvn.transform(result_region)); 1793 return _gvn.transform(result_val); 1794 } else { 1795 return result; 1796 } 1797 } 1798 } 1799 1800 //------------------------------inline_exp------------------------------------- 1801 // Inline exp instructions, if possible. The Intel hardware only misses 1802 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1803 bool LibraryCallKit::inline_exp() { 1804 Node* arg = round_double_node(argument(0)); 1805 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1806 1807 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1808 set_result(n); 1809 1810 C->set_has_split_ifs(true); // Has chance for split-if optimization 1811 return true; 1812 } 1813 1814 //------------------------------inline_pow------------------------------------- 1815 // Inline power instructions, if possible. 1816 bool LibraryCallKit::inline_pow() { 1817 // Pseudocode for pow 1818 // if (y == 2) { 1819 // return x * x; 1820 // } else { 1821 // if (x <= 0.0) { 1822 // long longy = (long)y; 1823 // if ((double)longy == y) { // if y is long 1824 // if (y + 1 == y) longy = 0; // huge number: even 1825 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1826 // } else { 1827 // result = NaN; 1828 // } 1829 // } else { 1830 // result = DPow(x,y); 1831 // } 1832 // if (result != result)? { 1833 // result = uncommon_trap() or runtime_call(); 1834 // } 1835 // return result; 1836 // } 1837 1838 Node* x = round_double_node(argument(0)); 1839 Node* y = round_double_node(argument(2)); 1840 1841 Node* result = NULL; 1842 1843 Node* const_two_node = makecon(TypeD::make(2.0)); 1844 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1845 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1846 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1847 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1848 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1849 1850 RegionNode* region_node = new RegionNode(3); 1851 region_node->init_req(1, if_true); 1852 1853 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1854 // special case for x^y where y == 2, we can convert it to x * x 1855 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1856 1857 // set control to if_false since we will now process the false branch 1858 set_control(if_false); 1859 1860 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1861 // Short form: skip the fancy tests and just check for NaN result. 1862 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1863 } else { 1864 // If this inlining ever returned NaN in the past, include all 1865 // checks + call to the runtime. 1866 1867 // Set the merge point for If node with condition of (x <= 0.0) 1868 // There are four possible paths to region node and phi node 1869 RegionNode *r = new RegionNode(4); 1870 Node *phi = new PhiNode(r, Type::DOUBLE); 1871 1872 // Build the first if node: if (x <= 0.0) 1873 // Node for 0 constant 1874 Node *zeronode = makecon(TypeD::ZERO); 1875 // Check x:0 1876 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1877 // Check: If (x<=0) then go complex path 1878 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1879 // Branch either way 1880 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1881 // Fast path taken; set region slot 3 1882 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1883 r->init_req(3,fast_taken); // Capture fast-control 1884 1885 // Fast path not-taken, i.e. slow path 1886 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1887 1888 // Set fast path result 1889 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1890 phi->init_req(3, fast_result); 1891 1892 // Complex path 1893 // Build the second if node (if y is long) 1894 // Node for (long)y 1895 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1896 // Node for (double)((long) y) 1897 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1898 // Check (double)((long) y) : y 1899 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1900 // Check if (y isn't long) then go to slow path 1901 1902 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1903 // Branch either way 1904 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1905 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1906 1907 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1908 1909 // Calculate DPow(abs(x), y)*(1 & (long)y) 1910 // Node for constant 1 1911 Node *conone = longcon(1); 1912 // 1& (long)y 1913 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1914 1915 // A huge number is always even. Detect a huge number by checking 1916 // if y + 1 == y and set integer to be tested for parity to 0. 1917 // Required for corner case: 1918 // (long)9.223372036854776E18 = max_jlong 1919 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1920 // max_jlong is odd but 9.223372036854776E18 is even 1921 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1922 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1923 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1924 Node* correctedsign = NULL; 1925 if (ConditionalMoveLimit != 0) { 1926 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1927 } else { 1928 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1929 RegionNode *r = new RegionNode(3); 1930 Node *phi = new PhiNode(r, TypeLong::LONG); 1931 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1932 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1933 phi->init_req(1, signnode); 1934 phi->init_req(2, longcon(0)); 1935 correctedsign = _gvn.transform(phi); 1936 ylong_path = _gvn.transform(r); 1937 record_for_igvn(r); 1938 } 1939 1940 // zero node 1941 Node *conzero = longcon(0); 1942 // Check (1&(long)y)==0? 1943 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1944 // Check if (1&(long)y)!=0?, if so the result is negative 1945 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1946 // abs(x) 1947 Node *absx=_gvn.transform(new AbsDNode(x)); 1948 // abs(x)^y 1949 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1950 // -abs(x)^y 1951 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1952 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1953 Node *signresult = NULL; 1954 if (ConditionalMoveLimit != 0) { 1955 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1956 } else { 1957 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1958 RegionNode *r = new RegionNode(3); 1959 Node *phi = new PhiNode(r, Type::DOUBLE); 1960 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1961 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1962 phi->init_req(1, absxpowy); 1963 phi->init_req(2, negabsxpowy); 1964 signresult = _gvn.transform(phi); 1965 ylong_path = _gvn.transform(r); 1966 record_for_igvn(r); 1967 } 1968 // Set complex path fast result 1969 r->init_req(2, ylong_path); 1970 phi->init_req(2, signresult); 1971 1972 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1973 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1974 r->init_req(1,slow_path); 1975 phi->init_req(1,slow_result); 1976 1977 // Post merge 1978 set_control(_gvn.transform(r)); 1979 record_for_igvn(r); 1980 result = _gvn.transform(phi); 1981 } 1982 1983 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1984 1985 // control from finish_pow_exp is now input to the region node 1986 region_node->set_req(2, control()); 1987 // the result from finish_pow_exp is now input to the phi node 1988 phi_node->init_req(2, result); 1989 set_control(_gvn.transform(region_node)); 1990 record_for_igvn(region_node); 1991 set_result(_gvn.transform(phi_node)); 1992 1993 C->set_has_split_ifs(true); // Has chance for split-if optimization 1994 return true; 1995 } 1996 1997 //------------------------------runtime_math----------------------------- 1998 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1999 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 2000 "must be (DD)D or (D)D type"); 2001 2002 // Inputs 2003 Node* a = round_double_node(argument(0)); 2004 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 2005 2006 const TypePtr* no_memory_effects = NULL; 2007 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 2008 no_memory_effects, 2009 a, top(), b, b ? top() : NULL); 2010 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 2011 #ifdef ASSERT 2012 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 2013 assert(value_top == top(), "second value must be top"); 2014 #endif 2015 2016 set_result(value); 2017 return true; 2018 } 2019 2020 //------------------------------inline_math_native----------------------------- 2021 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 2022 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 2023 switch (id) { 2024 // These intrinsics are not properly supported on all hardware 2025 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 2026 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 2027 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 2028 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 2029 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 2030 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 2031 2032 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 2033 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 2034 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 2035 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 2036 2037 // These intrinsics are supported on all hardware 2038 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 2039 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 2040 2041 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 2042 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 2043 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 2044 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 2045 #undef FN_PTR 2046 2047 // These intrinsics are not yet correctly implemented 2048 case vmIntrinsics::_datan2: 2049 return false; 2050 2051 default: 2052 fatal_unexpected_iid(id); 2053 return false; 2054 } 2055 } 2056 2057 static bool is_simple_name(Node* n) { 2058 return (n->req() == 1 // constant 2059 || (n->is_Type() && n->as_Type()->type()->singleton()) 2060 || n->is_Proj() // parameter or return value 2061 || n->is_Phi() // local of some sort 2062 ); 2063 } 2064 2065 //----------------------------inline_notify-----------------------------------* 2066 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) { 2067 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type(); 2068 address func; 2069 if (id == vmIntrinsics::_notify) { 2070 func = OptoRuntime::monitor_notify_Java(); 2071 } else { 2072 func = OptoRuntime::monitor_notifyAll_Java(); 2073 } 2074 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0)); 2075 make_slow_call_ex(call, env()->Throwable_klass(), false); 2076 return true; 2077 } 2078 2079 2080 //----------------------------inline_min_max----------------------------------- 2081 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 2082 set_result(generate_min_max(id, argument(0), argument(1))); 2083 return true; 2084 } 2085 2086 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2087 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2088 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2089 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2090 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2091 2092 { 2093 PreserveJVMState pjvms(this); 2094 PreserveReexecuteState preexecs(this); 2095 jvms()->set_should_reexecute(true); 2096 2097 set_control(slow_path); 2098 set_i_o(i_o()); 2099 2100 uncommon_trap(Deoptimization::Reason_intrinsic, 2101 Deoptimization::Action_none); 2102 } 2103 2104 set_control(fast_path); 2105 set_result(math); 2106 } 2107 2108 template <typename OverflowOp> 2109 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2110 typedef typename OverflowOp::MathOp MathOp; 2111 2112 MathOp* mathOp = new MathOp(arg1, arg2); 2113 Node* operation = _gvn.transform( mathOp ); 2114 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2115 inline_math_mathExact(operation, ofcheck); 2116 return true; 2117 } 2118 2119 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2120 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2121 } 2122 2123 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2124 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2125 } 2126 2127 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2128 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2129 } 2130 2131 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2132 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2133 } 2134 2135 bool LibraryCallKit::inline_math_negateExactI() { 2136 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2137 } 2138 2139 bool LibraryCallKit::inline_math_negateExactL() { 2140 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2141 } 2142 2143 bool LibraryCallKit::inline_math_multiplyExactI() { 2144 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2145 } 2146 2147 bool LibraryCallKit::inline_math_multiplyExactL() { 2148 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2149 } 2150 2151 Node* 2152 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2153 // These are the candidate return value: 2154 Node* xvalue = x0; 2155 Node* yvalue = y0; 2156 2157 if (xvalue == yvalue) { 2158 return xvalue; 2159 } 2160 2161 bool want_max = (id == vmIntrinsics::_max); 2162 2163 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2164 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2165 if (txvalue == NULL || tyvalue == NULL) return top(); 2166 // This is not really necessary, but it is consistent with a 2167 // hypothetical MaxINode::Value method: 2168 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2169 2170 // %%% This folding logic should (ideally) be in a different place. 2171 // Some should be inside IfNode, and there to be a more reliable 2172 // transformation of ?: style patterns into cmoves. We also want 2173 // more powerful optimizations around cmove and min/max. 2174 2175 // Try to find a dominating comparison of these guys. 2176 // It can simplify the index computation for Arrays.copyOf 2177 // and similar uses of System.arraycopy. 2178 // First, compute the normalized version of CmpI(x, y). 2179 int cmp_op = Op_CmpI; 2180 Node* xkey = xvalue; 2181 Node* ykey = yvalue; 2182 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2183 if (ideal_cmpxy->is_Cmp()) { 2184 // E.g., if we have CmpI(length - offset, count), 2185 // it might idealize to CmpI(length, count + offset) 2186 cmp_op = ideal_cmpxy->Opcode(); 2187 xkey = ideal_cmpxy->in(1); 2188 ykey = ideal_cmpxy->in(2); 2189 } 2190 2191 // Start by locating any relevant comparisons. 2192 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2193 Node* cmpxy = NULL; 2194 Node* cmpyx = NULL; 2195 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2196 Node* cmp = start_from->fast_out(k); 2197 if (cmp->outcnt() > 0 && // must have prior uses 2198 cmp->in(0) == NULL && // must be context-independent 2199 cmp->Opcode() == cmp_op) { // right kind of compare 2200 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2201 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2202 } 2203 } 2204 2205 const int NCMPS = 2; 2206 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2207 int cmpn; 2208 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2209 if (cmps[cmpn] != NULL) break; // find a result 2210 } 2211 if (cmpn < NCMPS) { 2212 // Look for a dominating test that tells us the min and max. 2213 int depth = 0; // Limit search depth for speed 2214 Node* dom = control(); 2215 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2216 if (++depth >= 100) break; 2217 Node* ifproj = dom; 2218 if (!ifproj->is_Proj()) continue; 2219 Node* iff = ifproj->in(0); 2220 if (!iff->is_If()) continue; 2221 Node* bol = iff->in(1); 2222 if (!bol->is_Bool()) continue; 2223 Node* cmp = bol->in(1); 2224 if (cmp == NULL) continue; 2225 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2226 if (cmps[cmpn] == cmp) break; 2227 if (cmpn == NCMPS) continue; 2228 BoolTest::mask btest = bol->as_Bool()->_test._test; 2229 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2230 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2231 // At this point, we know that 'x btest y' is true. 2232 switch (btest) { 2233 case BoolTest::eq: 2234 // They are proven equal, so we can collapse the min/max. 2235 // Either value is the answer. Choose the simpler. 2236 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2237 return yvalue; 2238 return xvalue; 2239 case BoolTest::lt: // x < y 2240 case BoolTest::le: // x <= y 2241 return (want_max ? yvalue : xvalue); 2242 case BoolTest::gt: // x > y 2243 case BoolTest::ge: // x >= y 2244 return (want_max ? xvalue : yvalue); 2245 } 2246 } 2247 } 2248 2249 // We failed to find a dominating test. 2250 // Let's pick a test that might GVN with prior tests. 2251 Node* best_bol = NULL; 2252 BoolTest::mask best_btest = BoolTest::illegal; 2253 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2254 Node* cmp = cmps[cmpn]; 2255 if (cmp == NULL) continue; 2256 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2257 Node* bol = cmp->fast_out(j); 2258 if (!bol->is_Bool()) continue; 2259 BoolTest::mask btest = bol->as_Bool()->_test._test; 2260 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2261 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2262 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2263 best_bol = bol->as_Bool(); 2264 best_btest = btest; 2265 } 2266 } 2267 } 2268 2269 Node* answer_if_true = NULL; 2270 Node* answer_if_false = NULL; 2271 switch (best_btest) { 2272 default: 2273 if (cmpxy == NULL) 2274 cmpxy = ideal_cmpxy; 2275 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2276 // and fall through: 2277 case BoolTest::lt: // x < y 2278 case BoolTest::le: // x <= y 2279 answer_if_true = (want_max ? yvalue : xvalue); 2280 answer_if_false = (want_max ? xvalue : yvalue); 2281 break; 2282 case BoolTest::gt: // x > y 2283 case BoolTest::ge: // x >= y 2284 answer_if_true = (want_max ? xvalue : yvalue); 2285 answer_if_false = (want_max ? yvalue : xvalue); 2286 break; 2287 } 2288 2289 jint hi, lo; 2290 if (want_max) { 2291 // We can sharpen the minimum. 2292 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2293 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2294 } else { 2295 // We can sharpen the maximum. 2296 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2297 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2298 } 2299 2300 // Use a flow-free graph structure, to avoid creating excess control edges 2301 // which could hinder other optimizations. 2302 // Since Math.min/max is often used with arraycopy, we want 2303 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2304 Node* cmov = CMoveNode::make(NULL, best_bol, 2305 answer_if_false, answer_if_true, 2306 TypeInt::make(lo, hi, widen)); 2307 2308 return _gvn.transform(cmov); 2309 2310 /* 2311 // This is not as desirable as it may seem, since Min and Max 2312 // nodes do not have a full set of optimizations. 2313 // And they would interfere, anyway, with 'if' optimizations 2314 // and with CMoveI canonical forms. 2315 switch (id) { 2316 case vmIntrinsics::_min: 2317 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2318 case vmIntrinsics::_max: 2319 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2320 default: 2321 ShouldNotReachHere(); 2322 } 2323 */ 2324 } 2325 2326 inline int 2327 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2328 const TypePtr* base_type = TypePtr::NULL_PTR; 2329 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2330 if (base_type == NULL) { 2331 // Unknown type. 2332 return Type::AnyPtr; 2333 } else if (base_type == TypePtr::NULL_PTR) { 2334 // Since this is a NULL+long form, we have to switch to a rawptr. 2335 base = _gvn.transform(new CastX2PNode(offset)); 2336 offset = MakeConX(0); 2337 return Type::RawPtr; 2338 } else if (base_type->base() == Type::RawPtr) { 2339 return Type::RawPtr; 2340 } else if (base_type->isa_oopptr()) { 2341 // Base is never null => always a heap address. 2342 if (base_type->ptr() == TypePtr::NotNull) { 2343 return Type::OopPtr; 2344 } 2345 // Offset is small => always a heap address. 2346 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2347 if (offset_type != NULL && 2348 base_type->offset() == 0 && // (should always be?) 2349 offset_type->_lo >= 0 && 2350 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2351 return Type::OopPtr; 2352 } 2353 // Otherwise, it might either be oop+off or NULL+addr. 2354 return Type::AnyPtr; 2355 } else { 2356 // No information: 2357 return Type::AnyPtr; 2358 } 2359 } 2360 2361 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2362 int kind = classify_unsafe_addr(base, offset); 2363 if (kind == Type::RawPtr) { 2364 return basic_plus_adr(top(), base, offset); 2365 } else { 2366 return basic_plus_adr(base, offset); 2367 } 2368 } 2369 2370 //--------------------------inline_number_methods----------------------------- 2371 // inline int Integer.numberOfLeadingZeros(int) 2372 // inline int Long.numberOfLeadingZeros(long) 2373 // 2374 // inline int Integer.numberOfTrailingZeros(int) 2375 // inline int Long.numberOfTrailingZeros(long) 2376 // 2377 // inline int Integer.bitCount(int) 2378 // inline int Long.bitCount(long) 2379 // 2380 // inline char Character.reverseBytes(char) 2381 // inline short Short.reverseBytes(short) 2382 // inline int Integer.reverseBytes(int) 2383 // inline long Long.reverseBytes(long) 2384 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2385 Node* arg = argument(0); 2386 Node* n; 2387 switch (id) { 2388 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2389 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2390 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2391 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2392 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2393 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2394 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2395 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2396 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2397 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2398 default: fatal_unexpected_iid(id); break; 2399 } 2400 set_result(_gvn.transform(n)); 2401 return true; 2402 } 2403 2404 //----------------------------inline_unsafe_access---------------------------- 2405 2406 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2407 2408 // Helper that guards and inserts a pre-barrier. 2409 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2410 Node* pre_val, bool need_mem_bar) { 2411 // We could be accessing the referent field of a reference object. If so, when G1 2412 // is enabled, we need to log the value in the referent field in an SATB buffer. 2413 // This routine performs some compile time filters and generates suitable 2414 // runtime filters that guard the pre-barrier code. 2415 // Also add memory barrier for non volatile load from the referent field 2416 // to prevent commoning of loads across safepoint. 2417 if (!UseG1GC && !need_mem_bar) 2418 return; 2419 2420 // Some compile time checks. 2421 2422 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2423 const TypeX* otype = offset->find_intptr_t_type(); 2424 if (otype != NULL && otype->is_con() && 2425 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2426 // Constant offset but not the reference_offset so just return 2427 return; 2428 } 2429 2430 // We only need to generate the runtime guards for instances. 2431 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2432 if (btype != NULL) { 2433 if (btype->isa_aryptr()) { 2434 // Array type so nothing to do 2435 return; 2436 } 2437 2438 const TypeInstPtr* itype = btype->isa_instptr(); 2439 if (itype != NULL) { 2440 // Can the klass of base_oop be statically determined to be 2441 // _not_ a sub-class of Reference and _not_ Object? 2442 ciKlass* klass = itype->klass(); 2443 if ( klass->is_loaded() && 2444 !klass->is_subtype_of(env()->Reference_klass()) && 2445 !env()->Object_klass()->is_subtype_of(klass)) { 2446 return; 2447 } 2448 } 2449 } 2450 2451 // The compile time filters did not reject base_oop/offset so 2452 // we need to generate the following runtime filters 2453 // 2454 // if (offset == java_lang_ref_Reference::_reference_offset) { 2455 // if (instance_of(base, java.lang.ref.Reference)) { 2456 // pre_barrier(_, pre_val, ...); 2457 // } 2458 // } 2459 2460 float likely = PROB_LIKELY( 0.999); 2461 float unlikely = PROB_UNLIKELY(0.999); 2462 2463 IdealKit ideal(this); 2464 #define __ ideal. 2465 2466 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2467 2468 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2469 // Update graphKit memory and control from IdealKit. 2470 sync_kit(ideal); 2471 2472 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2473 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2474 2475 // Update IdealKit memory and control from graphKit. 2476 __ sync_kit(this); 2477 2478 Node* one = __ ConI(1); 2479 // is_instof == 0 if base_oop == NULL 2480 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2481 2482 // Update graphKit from IdeakKit. 2483 sync_kit(ideal); 2484 2485 // Use the pre-barrier to record the value in the referent field 2486 pre_barrier(false /* do_load */, 2487 __ ctrl(), 2488 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2489 pre_val /* pre_val */, 2490 T_OBJECT); 2491 if (need_mem_bar) { 2492 // Add memory barrier to prevent commoning reads from this field 2493 // across safepoint since GC can change its value. 2494 insert_mem_bar(Op_MemBarCPUOrder); 2495 } 2496 // Update IdealKit from graphKit. 2497 __ sync_kit(this); 2498 2499 } __ end_if(); // _ref_type != ref_none 2500 } __ end_if(); // offset == referent_offset 2501 2502 // Final sync IdealKit and GraphKit. 2503 final_sync(ideal); 2504 #undef __ 2505 } 2506 2507 2508 // Interpret Unsafe.fieldOffset cookies correctly: 2509 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2510 2511 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2512 // Attempt to infer a sharper value type from the offset and base type. 2513 ciKlass* sharpened_klass = NULL; 2514 2515 // See if it is an instance field, with an object type. 2516 if (alias_type->field() != NULL) { 2517 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2518 if (alias_type->field()->type()->is_klass()) { 2519 sharpened_klass = alias_type->field()->type()->as_klass(); 2520 } 2521 } 2522 2523 // See if it is a narrow oop array. 2524 if (adr_type->isa_aryptr()) { 2525 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2526 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2527 if (elem_type != NULL) { 2528 sharpened_klass = elem_type->klass(); 2529 } 2530 } 2531 } 2532 2533 // The sharpened class might be unloaded if there is no class loader 2534 // contraint in place. 2535 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2536 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2537 2538 #ifndef PRODUCT 2539 if (C->print_intrinsics() || C->print_inlining()) { 2540 tty->print(" from base type: "); adr_type->dump(); 2541 tty->print(" sharpened value: "); tjp->dump(); 2542 } 2543 #endif 2544 // Sharpen the value type. 2545 return tjp; 2546 } 2547 return NULL; 2548 } 2549 2550 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2551 if (callee()->is_static()) return false; // caller must have the capability! 2552 2553 #ifndef PRODUCT 2554 { 2555 ResourceMark rm; 2556 // Check the signatures. 2557 ciSignature* sig = callee()->signature(); 2558 #ifdef ASSERT 2559 if (!is_store) { 2560 // Object getObject(Object base, int/long offset), etc. 2561 BasicType rtype = sig->return_type()->basic_type(); 2562 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2563 rtype = T_ADDRESS; // it is really a C void* 2564 assert(rtype == type, "getter must return the expected value"); 2565 if (!is_native_ptr) { 2566 assert(sig->count() == 2, "oop getter has 2 arguments"); 2567 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2568 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2569 } else { 2570 assert(sig->count() == 1, "native getter has 1 argument"); 2571 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2572 } 2573 } else { 2574 // void putObject(Object base, int/long offset, Object x), etc. 2575 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2576 if (!is_native_ptr) { 2577 assert(sig->count() == 3, "oop putter has 3 arguments"); 2578 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2579 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2580 } else { 2581 assert(sig->count() == 2, "native putter has 2 arguments"); 2582 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2583 } 2584 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2585 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2586 vtype = T_ADDRESS; // it is really a C void* 2587 assert(vtype == type, "putter must accept the expected value"); 2588 } 2589 #endif // ASSERT 2590 } 2591 #endif //PRODUCT 2592 2593 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2594 2595 Node* receiver = argument(0); // type: oop 2596 2597 // Build address expression. 2598 Node* adr; 2599 Node* heap_base_oop = top(); 2600 Node* offset = top(); 2601 Node* val; 2602 2603 if (!is_native_ptr) { 2604 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2605 Node* base = argument(1); // type: oop 2606 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2607 offset = argument(2); // type: long 2608 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2609 // to be plain byte offsets, which are also the same as those accepted 2610 // by oopDesc::field_base. 2611 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2612 "fieldOffset must be byte-scaled"); 2613 // 32-bit machines ignore the high half! 2614 offset = ConvL2X(offset); 2615 adr = make_unsafe_address(base, offset); 2616 heap_base_oop = base; 2617 val = is_store ? argument(4) : NULL; 2618 } else { 2619 Node* ptr = argument(1); // type: long 2620 ptr = ConvL2X(ptr); // adjust Java long to machine word 2621 adr = make_unsafe_address(NULL, ptr); 2622 val = is_store ? argument(3) : NULL; 2623 } 2624 2625 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2626 2627 // First guess at the value type. 2628 const Type *value_type = Type::get_const_basic_type(type); 2629 2630 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2631 // there was not enough information to nail it down. 2632 Compile::AliasType* alias_type = C->alias_type(adr_type); 2633 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2634 2635 // We will need memory barriers unless we can determine a unique 2636 // alias category for this reference. (Note: If for some reason 2637 // the barriers get omitted and the unsafe reference begins to "pollute" 2638 // the alias analysis of the rest of the graph, either Compile::can_alias 2639 // or Compile::must_alias will throw a diagnostic assert.) 2640 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2641 2642 // If we are reading the value of the referent field of a Reference 2643 // object (either by using Unsafe directly or through reflection) 2644 // then, if G1 is enabled, we need to record the referent in an 2645 // SATB log buffer using the pre-barrier mechanism. 2646 // Also we need to add memory barrier to prevent commoning reads 2647 // from this field across safepoint since GC can change its value. 2648 bool need_read_barrier = !is_native_ptr && !is_store && 2649 offset != top() && heap_base_oop != top(); 2650 2651 if (!is_store && type == T_OBJECT) { 2652 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2653 if (tjp != NULL) { 2654 value_type = tjp; 2655 } 2656 } 2657 2658 receiver = null_check(receiver); 2659 if (stopped()) { 2660 return true; 2661 } 2662 // Heap pointers get a null-check from the interpreter, 2663 // as a courtesy. However, this is not guaranteed by Unsafe, 2664 // and it is not possible to fully distinguish unintended nulls 2665 // from intended ones in this API. 2666 2667 if (is_volatile) { 2668 // We need to emit leading and trailing CPU membars (see below) in 2669 // addition to memory membars when is_volatile. This is a little 2670 // too strong, but avoids the need to insert per-alias-type 2671 // volatile membars (for stores; compare Parse::do_put_xxx), which 2672 // we cannot do effectively here because we probably only have a 2673 // rough approximation of type. 2674 need_mem_bar = true; 2675 // For Stores, place a memory ordering barrier now. 2676 if (is_store) { 2677 insert_mem_bar(Op_MemBarRelease); 2678 } else { 2679 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2680 insert_mem_bar(Op_MemBarVolatile); 2681 } 2682 } 2683 } 2684 2685 // Memory barrier to prevent normal and 'unsafe' accesses from 2686 // bypassing each other. Happens after null checks, so the 2687 // exception paths do not take memory state from the memory barrier, 2688 // so there's no problems making a strong assert about mixing users 2689 // of safe & unsafe memory. 2690 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2691 2692 if (!is_store) { 2693 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2694 // To be valid, unsafe loads may depend on other conditions than 2695 // the one that guards them: pin the Load node 2696 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile); 2697 // load value 2698 switch (type) { 2699 case T_BOOLEAN: 2700 case T_CHAR: 2701 case T_BYTE: 2702 case T_SHORT: 2703 case T_INT: 2704 case T_LONG: 2705 case T_FLOAT: 2706 case T_DOUBLE: 2707 break; 2708 case T_OBJECT: 2709 if (need_read_barrier) { 2710 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2711 } 2712 break; 2713 case T_ADDRESS: 2714 // Cast to an int type. 2715 p = _gvn.transform(new CastP2XNode(NULL, p)); 2716 p = ConvX2UL(p); 2717 break; 2718 default: 2719 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2720 break; 2721 } 2722 // The load node has the control of the preceding MemBarCPUOrder. All 2723 // following nodes will have the control of the MemBarCPUOrder inserted at 2724 // the end of this method. So, pushing the load onto the stack at a later 2725 // point is fine. 2726 set_result(p); 2727 } else { 2728 // place effect of store into memory 2729 switch (type) { 2730 case T_DOUBLE: 2731 val = dstore_rounding(val); 2732 break; 2733 case T_ADDRESS: 2734 // Repackage the long as a pointer. 2735 val = ConvL2X(val); 2736 val = _gvn.transform(new CastX2PNode(val)); 2737 break; 2738 } 2739 2740 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2741 if (type != T_OBJECT ) { 2742 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2743 } else { 2744 // Possibly an oop being stored to Java heap or native memory 2745 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2746 // oop to Java heap. 2747 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2748 } else { 2749 // We can't tell at compile time if we are storing in the Java heap or outside 2750 // of it. So we need to emit code to conditionally do the proper type of 2751 // store. 2752 2753 IdealKit ideal(this); 2754 #define __ ideal. 2755 // QQQ who knows what probability is here?? 2756 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2757 // Sync IdealKit and graphKit. 2758 sync_kit(ideal); 2759 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2760 // Update IdealKit memory. 2761 __ sync_kit(this); 2762 } __ else_(); { 2763 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2764 } __ end_if(); 2765 // Final sync IdealKit and GraphKit. 2766 final_sync(ideal); 2767 #undef __ 2768 } 2769 } 2770 } 2771 2772 if (is_volatile) { 2773 if (!is_store) { 2774 insert_mem_bar(Op_MemBarAcquire); 2775 } else { 2776 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2777 insert_mem_bar(Op_MemBarVolatile); 2778 } 2779 } 2780 } 2781 2782 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2783 2784 return true; 2785 } 2786 2787 //----------------------------inline_unsafe_load_store---------------------------- 2788 // This method serves a couple of different customers (depending on LoadStoreKind): 2789 // 2790 // LS_cmpxchg: 2791 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2792 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2793 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2794 // 2795 // LS_xadd: 2796 // public int getAndAddInt( Object o, long offset, int delta) 2797 // public long getAndAddLong(Object o, long offset, long delta) 2798 // 2799 // LS_xchg: 2800 // int getAndSet(Object o, long offset, int newValue) 2801 // long getAndSet(Object o, long offset, long newValue) 2802 // Object getAndSet(Object o, long offset, Object newValue) 2803 // 2804 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2805 // This basic scheme here is the same as inline_unsafe_access, but 2806 // differs in enough details that combining them would make the code 2807 // overly confusing. (This is a true fact! I originally combined 2808 // them, but even I was confused by it!) As much code/comments as 2809 // possible are retained from inline_unsafe_access though to make 2810 // the correspondences clearer. - dl 2811 2812 if (callee()->is_static()) return false; // caller must have the capability! 2813 2814 #ifndef PRODUCT 2815 BasicType rtype; 2816 { 2817 ResourceMark rm; 2818 // Check the signatures. 2819 ciSignature* sig = callee()->signature(); 2820 rtype = sig->return_type()->basic_type(); 2821 if (kind == LS_xadd || kind == LS_xchg) { 2822 // Check the signatures. 2823 #ifdef ASSERT 2824 assert(rtype == type, "get and set must return the expected type"); 2825 assert(sig->count() == 3, "get and set has 3 arguments"); 2826 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2827 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2828 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2829 #endif // ASSERT 2830 } else if (kind == LS_cmpxchg) { 2831 // Check the signatures. 2832 #ifdef ASSERT 2833 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2834 assert(sig->count() == 4, "CAS has 4 arguments"); 2835 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2836 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2837 #endif // ASSERT 2838 } else { 2839 ShouldNotReachHere(); 2840 } 2841 } 2842 #endif //PRODUCT 2843 2844 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2845 2846 // Get arguments: 2847 Node* receiver = NULL; 2848 Node* base = NULL; 2849 Node* offset = NULL; 2850 Node* oldval = NULL; 2851 Node* newval = NULL; 2852 if (kind == LS_cmpxchg) { 2853 const bool two_slot_type = type2size[type] == 2; 2854 receiver = argument(0); // type: oop 2855 base = argument(1); // type: oop 2856 offset = argument(2); // type: long 2857 oldval = argument(4); // type: oop, int, or long 2858 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2859 } else if (kind == LS_xadd || kind == LS_xchg){ 2860 receiver = argument(0); // type: oop 2861 base = argument(1); // type: oop 2862 offset = argument(2); // type: long 2863 oldval = NULL; 2864 newval = argument(4); // type: oop, int, or long 2865 } 2866 2867 // Null check receiver. 2868 receiver = null_check(receiver); 2869 if (stopped()) { 2870 return true; 2871 } 2872 2873 // Build field offset expression. 2874 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2875 // to be plain byte offsets, which are also the same as those accepted 2876 // by oopDesc::field_base. 2877 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2878 // 32-bit machines ignore the high half of long offsets 2879 offset = ConvL2X(offset); 2880 Node* adr = make_unsafe_address(base, offset); 2881 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2882 2883 // For CAS, unlike inline_unsafe_access, there seems no point in 2884 // trying to refine types. Just use the coarse types here. 2885 const Type *value_type = Type::get_const_basic_type(type); 2886 Compile::AliasType* alias_type = C->alias_type(adr_type); 2887 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2888 2889 if (kind == LS_xchg && type == T_OBJECT) { 2890 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2891 if (tjp != NULL) { 2892 value_type = tjp; 2893 } 2894 } 2895 2896 int alias_idx = C->get_alias_index(adr_type); 2897 2898 // Memory-model-wise, a LoadStore acts like a little synchronized 2899 // block, so needs barriers on each side. These don't translate 2900 // into actual barriers on most machines, but we still need rest of 2901 // compiler to respect ordering. 2902 2903 insert_mem_bar(Op_MemBarRelease); 2904 insert_mem_bar(Op_MemBarCPUOrder); 2905 2906 // 4984716: MemBars must be inserted before this 2907 // memory node in order to avoid a false 2908 // dependency which will confuse the scheduler. 2909 Node *mem = memory(alias_idx); 2910 2911 // For now, we handle only those cases that actually exist: ints, 2912 // longs, and Object. Adding others should be straightforward. 2913 Node* load_store; 2914 switch(type) { 2915 case T_INT: 2916 if (kind == LS_xadd) { 2917 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2918 } else if (kind == LS_xchg) { 2919 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2920 } else if (kind == LS_cmpxchg) { 2921 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2922 } else { 2923 ShouldNotReachHere(); 2924 } 2925 break; 2926 case T_LONG: 2927 if (kind == LS_xadd) { 2928 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2929 } else if (kind == LS_xchg) { 2930 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2931 } else if (kind == LS_cmpxchg) { 2932 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2933 } else { 2934 ShouldNotReachHere(); 2935 } 2936 break; 2937 case T_OBJECT: 2938 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2939 // could be delayed during Parse (for example, in adjust_map_after_if()). 2940 // Execute transformation here to avoid barrier generation in such case. 2941 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2942 newval = _gvn.makecon(TypePtr::NULL_PTR); 2943 2944 // Reference stores need a store barrier. 2945 if (kind == LS_xchg) { 2946 // If pre-barrier must execute before the oop store, old value will require do_load here. 2947 if (!can_move_pre_barrier()) { 2948 pre_barrier(true /* do_load*/, 2949 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2950 NULL /* pre_val*/, 2951 T_OBJECT); 2952 } // Else move pre_barrier to use load_store value, see below. 2953 } else if (kind == LS_cmpxchg) { 2954 // Same as for newval above: 2955 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2956 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2957 } 2958 // The only known value which might get overwritten is oldval. 2959 pre_barrier(false /* do_load */, 2960 control(), NULL, NULL, max_juint, NULL, NULL, 2961 oldval /* pre_val */, 2962 T_OBJECT); 2963 } else { 2964 ShouldNotReachHere(); 2965 } 2966 2967 #ifdef _LP64 2968 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2969 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2970 if (kind == LS_xchg) { 2971 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2972 newval_enc, adr_type, value_type->make_narrowoop())); 2973 } else { 2974 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2975 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2976 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2977 newval_enc, oldval_enc)); 2978 } 2979 } else 2980 #endif 2981 { 2982 if (kind == LS_xchg) { 2983 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2984 } else { 2985 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2986 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2987 } 2988 } 2989 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2990 break; 2991 default: 2992 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2993 break; 2994 } 2995 2996 // SCMemProjNodes represent the memory state of a LoadStore. Their 2997 // main role is to prevent LoadStore nodes from being optimized away 2998 // when their results aren't used. 2999 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 3000 set_memory(proj, alias_idx); 3001 3002 if (type == T_OBJECT && kind == LS_xchg) { 3003 #ifdef _LP64 3004 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 3005 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 3006 } 3007 #endif 3008 if (can_move_pre_barrier()) { 3009 // Don't need to load pre_val. The old value is returned by load_store. 3010 // The pre_barrier can execute after the xchg as long as no safepoint 3011 // gets inserted between them. 3012 pre_barrier(false /* do_load */, 3013 control(), NULL, NULL, max_juint, NULL, NULL, 3014 load_store /* pre_val */, 3015 T_OBJECT); 3016 } 3017 } 3018 3019 // Add the trailing membar surrounding the access 3020 insert_mem_bar(Op_MemBarCPUOrder); 3021 insert_mem_bar(Op_MemBarAcquire); 3022 3023 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3024 set_result(load_store); 3025 return true; 3026 } 3027 3028 //----------------------------inline_unsafe_ordered_store---------------------- 3029 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3030 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3031 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3032 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3033 // This is another variant of inline_unsafe_access, differing in 3034 // that it always issues store-store ("release") barrier and ensures 3035 // store-atomicity (which only matters for "long"). 3036 3037 if (callee()->is_static()) return false; // caller must have the capability! 3038 3039 #ifndef PRODUCT 3040 { 3041 ResourceMark rm; 3042 // Check the signatures. 3043 ciSignature* sig = callee()->signature(); 3044 #ifdef ASSERT 3045 BasicType rtype = sig->return_type()->basic_type(); 3046 assert(rtype == T_VOID, "must return void"); 3047 assert(sig->count() == 3, "has 3 arguments"); 3048 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3049 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3050 #endif // ASSERT 3051 } 3052 #endif //PRODUCT 3053 3054 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3055 3056 // Get arguments: 3057 Node* receiver = argument(0); // type: oop 3058 Node* base = argument(1); // type: oop 3059 Node* offset = argument(2); // type: long 3060 Node* val = argument(4); // type: oop, int, or long 3061 3062 // Null check receiver. 3063 receiver = null_check(receiver); 3064 if (stopped()) { 3065 return true; 3066 } 3067 3068 // Build field offset expression. 3069 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3070 // 32-bit machines ignore the high half of long offsets 3071 offset = ConvL2X(offset); 3072 Node* adr = make_unsafe_address(base, offset); 3073 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3074 const Type *value_type = Type::get_const_basic_type(type); 3075 Compile::AliasType* alias_type = C->alias_type(adr_type); 3076 3077 insert_mem_bar(Op_MemBarRelease); 3078 insert_mem_bar(Op_MemBarCPUOrder); 3079 // Ensure that the store is atomic for longs: 3080 const bool require_atomic_access = true; 3081 Node* store; 3082 if (type == T_OBJECT) // reference stores need a store barrier. 3083 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3084 else { 3085 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3086 } 3087 insert_mem_bar(Op_MemBarCPUOrder); 3088 return true; 3089 } 3090 3091 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3092 // Regardless of form, don't allow previous ld/st to move down, 3093 // then issue acquire, release, or volatile mem_bar. 3094 insert_mem_bar(Op_MemBarCPUOrder); 3095 switch(id) { 3096 case vmIntrinsics::_loadFence: 3097 insert_mem_bar(Op_LoadFence); 3098 return true; 3099 case vmIntrinsics::_storeFence: 3100 insert_mem_bar(Op_StoreFence); 3101 return true; 3102 case vmIntrinsics::_fullFence: 3103 insert_mem_bar(Op_MemBarVolatile); 3104 return true; 3105 default: 3106 fatal_unexpected_iid(id); 3107 return false; 3108 } 3109 } 3110 3111 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3112 if (!kls->is_Con()) { 3113 return true; 3114 } 3115 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3116 if (klsptr == NULL) { 3117 return true; 3118 } 3119 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3120 // don't need a guard for a klass that is already initialized 3121 return !ik->is_initialized(); 3122 } 3123 3124 //----------------------------inline_unsafe_allocate--------------------------- 3125 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3126 bool LibraryCallKit::inline_unsafe_allocate() { 3127 if (callee()->is_static()) return false; // caller must have the capability! 3128 3129 null_check_receiver(); // null-check, then ignore 3130 Node* cls = null_check(argument(1)); 3131 if (stopped()) return true; 3132 3133 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3134 kls = null_check(kls); 3135 if (stopped()) return true; // argument was like int.class 3136 3137 Node* test = NULL; 3138 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3139 // Note: The argument might still be an illegal value like 3140 // Serializable.class or Object[].class. The runtime will handle it. 3141 // But we must make an explicit check for initialization. 3142 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3143 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3144 // can generate code to load it as unsigned byte. 3145 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3146 Node* bits = intcon(InstanceKlass::fully_initialized); 3147 test = _gvn.transform(new SubINode(inst, bits)); 3148 // The 'test' is non-zero if we need to take a slow path. 3149 } 3150 3151 Node* obj = new_instance(kls, test); 3152 set_result(obj); 3153 return true; 3154 } 3155 3156 #ifdef TRACE_HAVE_INTRINSICS 3157 /* 3158 * oop -> myklass 3159 * myklass->trace_id |= USED 3160 * return myklass->trace_id & ~0x3 3161 */ 3162 bool LibraryCallKit::inline_native_classID() { 3163 null_check_receiver(); // null-check, then ignore 3164 Node* cls = null_check(argument(1), T_OBJECT); 3165 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3166 kls = null_check(kls, T_OBJECT); 3167 ByteSize offset = TRACE_ID_OFFSET; 3168 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3169 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3170 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3171 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3172 Node* clsused = longcon(0x01l); // set the class bit 3173 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3174 3175 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3176 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3177 set_result(andl); 3178 return true; 3179 } 3180 3181 bool LibraryCallKit::inline_native_threadID() { 3182 Node* tls_ptr = NULL; 3183 Node* cur_thr = generate_current_thread(tls_ptr); 3184 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3185 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3186 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3187 3188 Node* threadid = NULL; 3189 size_t thread_id_size = OSThread::thread_id_size(); 3190 if (thread_id_size == (size_t) BytesPerLong) { 3191 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3192 } else if (thread_id_size == (size_t) BytesPerInt) { 3193 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3194 } else { 3195 ShouldNotReachHere(); 3196 } 3197 set_result(threadid); 3198 return true; 3199 } 3200 #endif 3201 3202 //------------------------inline_native_time_funcs-------------- 3203 // inline code for System.currentTimeMillis() and System.nanoTime() 3204 // these have the same type and signature 3205 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3206 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3207 const TypePtr* no_memory_effects = NULL; 3208 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3209 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3210 #ifdef ASSERT 3211 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3212 assert(value_top == top(), "second value must be top"); 3213 #endif 3214 set_result(value); 3215 return true; 3216 } 3217 3218 //------------------------inline_native_currentThread------------------ 3219 bool LibraryCallKit::inline_native_currentThread() { 3220 Node* junk = NULL; 3221 set_result(generate_current_thread(junk)); 3222 return true; 3223 } 3224 3225 //------------------------inline_native_isInterrupted------------------ 3226 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3227 bool LibraryCallKit::inline_native_isInterrupted() { 3228 // Add a fast path to t.isInterrupted(clear_int): 3229 // (t == Thread.current() && 3230 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3231 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3232 // So, in the common case that the interrupt bit is false, 3233 // we avoid making a call into the VM. Even if the interrupt bit 3234 // is true, if the clear_int argument is false, we avoid the VM call. 3235 // However, if the receiver is not currentThread, we must call the VM, 3236 // because there must be some locking done around the operation. 3237 3238 // We only go to the fast case code if we pass two guards. 3239 // Paths which do not pass are accumulated in the slow_region. 3240 3241 enum { 3242 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3243 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3244 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3245 PATH_LIMIT 3246 }; 3247 3248 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3249 // out of the function. 3250 insert_mem_bar(Op_MemBarCPUOrder); 3251 3252 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3253 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3254 3255 RegionNode* slow_region = new RegionNode(1); 3256 record_for_igvn(slow_region); 3257 3258 // (a) Receiving thread must be the current thread. 3259 Node* rec_thr = argument(0); 3260 Node* tls_ptr = NULL; 3261 Node* cur_thr = generate_current_thread(tls_ptr); 3262 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3263 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3264 3265 generate_slow_guard(bol_thr, slow_region); 3266 3267 // (b) Interrupt bit on TLS must be false. 3268 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3269 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3270 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3271 3272 // Set the control input on the field _interrupted read to prevent it floating up. 3273 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3274 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3275 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3276 3277 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3278 3279 // First fast path: if (!TLS._interrupted) return false; 3280 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3281 result_rgn->init_req(no_int_result_path, false_bit); 3282 result_val->init_req(no_int_result_path, intcon(0)); 3283 3284 // drop through to next case 3285 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3286 3287 #ifndef TARGET_OS_FAMILY_windows 3288 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3289 Node* clr_arg = argument(1); 3290 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3291 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3292 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3293 3294 // Second fast path: ... else if (!clear_int) return true; 3295 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3296 result_rgn->init_req(no_clear_result_path, false_arg); 3297 result_val->init_req(no_clear_result_path, intcon(1)); 3298 3299 // drop through to next case 3300 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3301 #else 3302 // To return true on Windows you must read the _interrupted field 3303 // and check the the event state i.e. take the slow path. 3304 #endif // TARGET_OS_FAMILY_windows 3305 3306 // (d) Otherwise, go to the slow path. 3307 slow_region->add_req(control()); 3308 set_control( _gvn.transform(slow_region)); 3309 3310 if (stopped()) { 3311 // There is no slow path. 3312 result_rgn->init_req(slow_result_path, top()); 3313 result_val->init_req(slow_result_path, top()); 3314 } else { 3315 // non-virtual because it is a private non-static 3316 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3317 3318 Node* slow_val = set_results_for_java_call(slow_call); 3319 // this->control() comes from set_results_for_java_call 3320 3321 Node* fast_io = slow_call->in(TypeFunc::I_O); 3322 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3323 3324 // These two phis are pre-filled with copies of of the fast IO and Memory 3325 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3326 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3327 3328 result_rgn->init_req(slow_result_path, control()); 3329 result_io ->init_req(slow_result_path, i_o()); 3330 result_mem->init_req(slow_result_path, reset_memory()); 3331 result_val->init_req(slow_result_path, slow_val); 3332 3333 set_all_memory(_gvn.transform(result_mem)); 3334 set_i_o( _gvn.transform(result_io)); 3335 } 3336 3337 C->set_has_split_ifs(true); // Has chance for split-if optimization 3338 set_result(result_rgn, result_val); 3339 return true; 3340 } 3341 3342 //---------------------------load_mirror_from_klass---------------------------- 3343 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3344 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3345 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3346 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3347 } 3348 3349 //-----------------------load_klass_from_mirror_common------------------------- 3350 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3351 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3352 // and branch to the given path on the region. 3353 // If never_see_null, take an uncommon trap on null, so we can optimistically 3354 // compile for the non-null case. 3355 // If the region is NULL, force never_see_null = true. 3356 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3357 bool never_see_null, 3358 RegionNode* region, 3359 int null_path, 3360 int offset) { 3361 if (region == NULL) never_see_null = true; 3362 Node* p = basic_plus_adr(mirror, offset); 3363 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3364 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3365 Node* null_ctl = top(); 3366 kls = null_check_oop(kls, &null_ctl, never_see_null); 3367 if (region != NULL) { 3368 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3369 region->init_req(null_path, null_ctl); 3370 } else { 3371 assert(null_ctl == top(), "no loose ends"); 3372 } 3373 return kls; 3374 } 3375 3376 //--------------------(inline_native_Class_query helpers)--------------------- 3377 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3378 // Fall through if (mods & mask) == bits, take the guard otherwise. 3379 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3380 // Branch around if the given klass has the given modifier bit set. 3381 // Like generate_guard, adds a new path onto the region. 3382 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3383 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3384 Node* mask = intcon(modifier_mask); 3385 Node* bits = intcon(modifier_bits); 3386 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3387 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3388 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3389 return generate_fair_guard(bol, region); 3390 } 3391 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3392 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3393 } 3394 3395 //-------------------------inline_native_Class_query------------------- 3396 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3397 const Type* return_type = TypeInt::BOOL; 3398 Node* prim_return_value = top(); // what happens if it's a primitive class? 3399 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3400 bool expect_prim = false; // most of these guys expect to work on refs 3401 3402 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3403 3404 Node* mirror = argument(0); 3405 Node* obj = top(); 3406 3407 switch (id) { 3408 case vmIntrinsics::_isInstance: 3409 // nothing is an instance of a primitive type 3410 prim_return_value = intcon(0); 3411 obj = argument(1); 3412 break; 3413 case vmIntrinsics::_getModifiers: 3414 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3415 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3416 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3417 break; 3418 case vmIntrinsics::_isInterface: 3419 prim_return_value = intcon(0); 3420 break; 3421 case vmIntrinsics::_isArray: 3422 prim_return_value = intcon(0); 3423 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3424 break; 3425 case vmIntrinsics::_isPrimitive: 3426 prim_return_value = intcon(1); 3427 expect_prim = true; // obviously 3428 break; 3429 case vmIntrinsics::_getSuperclass: 3430 prim_return_value = null(); 3431 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3432 break; 3433 case vmIntrinsics::_getClassAccessFlags: 3434 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3435 return_type = TypeInt::INT; // not bool! 6297094 3436 break; 3437 default: 3438 fatal_unexpected_iid(id); 3439 break; 3440 } 3441 3442 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3443 if (mirror_con == NULL) return false; // cannot happen? 3444 3445 #ifndef PRODUCT 3446 if (C->print_intrinsics() || C->print_inlining()) { 3447 ciType* k = mirror_con->java_mirror_type(); 3448 if (k) { 3449 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3450 k->print_name(); 3451 tty->cr(); 3452 } 3453 } 3454 #endif 3455 3456 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3457 RegionNode* region = new RegionNode(PATH_LIMIT); 3458 record_for_igvn(region); 3459 PhiNode* phi = new PhiNode(region, return_type); 3460 3461 // The mirror will never be null of Reflection.getClassAccessFlags, however 3462 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3463 // if it is. See bug 4774291. 3464 3465 // For Reflection.getClassAccessFlags(), the null check occurs in 3466 // the wrong place; see inline_unsafe_access(), above, for a similar 3467 // situation. 3468 mirror = null_check(mirror); 3469 // If mirror or obj is dead, only null-path is taken. 3470 if (stopped()) return true; 3471 3472 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3473 3474 // Now load the mirror's klass metaobject, and null-check it. 3475 // Side-effects region with the control path if the klass is null. 3476 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3477 // If kls is null, we have a primitive mirror. 3478 phi->init_req(_prim_path, prim_return_value); 3479 if (stopped()) { set_result(region, phi); return true; } 3480 bool safe_for_replace = (region->in(_prim_path) == top()); 3481 3482 Node* p; // handy temp 3483 Node* null_ctl; 3484 3485 // Now that we have the non-null klass, we can perform the real query. 3486 // For constant classes, the query will constant-fold in LoadNode::Value. 3487 Node* query_value = top(); 3488 switch (id) { 3489 case vmIntrinsics::_isInstance: 3490 // nothing is an instance of a primitive type 3491 query_value = gen_instanceof(obj, kls, safe_for_replace); 3492 break; 3493 3494 case vmIntrinsics::_getModifiers: 3495 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3496 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3497 break; 3498 3499 case vmIntrinsics::_isInterface: 3500 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3501 if (generate_interface_guard(kls, region) != NULL) 3502 // A guard was added. If the guard is taken, it was an interface. 3503 phi->add_req(intcon(1)); 3504 // If we fall through, it's a plain class. 3505 query_value = intcon(0); 3506 break; 3507 3508 case vmIntrinsics::_isArray: 3509 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3510 if (generate_array_guard(kls, region) != NULL) 3511 // A guard was added. If the guard is taken, it was an array. 3512 phi->add_req(intcon(1)); 3513 // If we fall through, it's a plain class. 3514 query_value = intcon(0); 3515 break; 3516 3517 case vmIntrinsics::_isPrimitive: 3518 query_value = intcon(0); // "normal" path produces false 3519 break; 3520 3521 case vmIntrinsics::_getSuperclass: 3522 // The rules here are somewhat unfortunate, but we can still do better 3523 // with random logic than with a JNI call. 3524 // Interfaces store null or Object as _super, but must report null. 3525 // Arrays store an intermediate super as _super, but must report Object. 3526 // Other types can report the actual _super. 3527 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3528 if (generate_interface_guard(kls, region) != NULL) 3529 // A guard was added. If the guard is taken, it was an interface. 3530 phi->add_req(null()); 3531 if (generate_array_guard(kls, region) != NULL) 3532 // A guard was added. If the guard is taken, it was an array. 3533 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3534 // If we fall through, it's a plain class. Get its _super. 3535 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3536 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3537 null_ctl = top(); 3538 kls = null_check_oop(kls, &null_ctl); 3539 if (null_ctl != top()) { 3540 // If the guard is taken, Object.superClass is null (both klass and mirror). 3541 region->add_req(null_ctl); 3542 phi ->add_req(null()); 3543 } 3544 if (!stopped()) { 3545 query_value = load_mirror_from_klass(kls); 3546 } 3547 break; 3548 3549 case vmIntrinsics::_getClassAccessFlags: 3550 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3551 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3552 break; 3553 3554 default: 3555 fatal_unexpected_iid(id); 3556 break; 3557 } 3558 3559 // Fall-through is the normal case of a query to a real class. 3560 phi->init_req(1, query_value); 3561 region->init_req(1, control()); 3562 3563 C->set_has_split_ifs(true); // Has chance for split-if optimization 3564 set_result(region, phi); 3565 return true; 3566 } 3567 3568 //-------------------------inline_Class_cast------------------- 3569 bool LibraryCallKit::inline_Class_cast() { 3570 Node* mirror = argument(0); // Class 3571 Node* obj = argument(1); 3572 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3573 if (mirror_con == NULL) { 3574 return false; // dead path (mirror->is_top()). 3575 } 3576 if (obj == NULL || obj->is_top()) { 3577 return false; // dead path 3578 } 3579 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr(); 3580 3581 // First, see if Class.cast() can be folded statically. 3582 // java_mirror_type() returns non-null for compile-time Class constants. 3583 ciType* tm = mirror_con->java_mirror_type(); 3584 if (tm != NULL && tm->is_klass() && 3585 tp != NULL && tp->klass() != NULL) { 3586 if (!tp->klass()->is_loaded()) { 3587 // Don't use intrinsic when class is not loaded. 3588 return false; 3589 } else { 3590 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass()); 3591 if (static_res == Compile::SSC_always_true) { 3592 // isInstance() is true - fold the code. 3593 set_result(obj); 3594 return true; 3595 } else if (static_res == Compile::SSC_always_false) { 3596 // Don't use intrinsic, have to throw ClassCastException. 3597 // If the reference is null, the non-intrinsic bytecode will 3598 // be optimized appropriately. 3599 return false; 3600 } 3601 } 3602 } 3603 3604 // Bailout intrinsic and do normal inlining if exception path is frequent. 3605 if (too_many_traps(Deoptimization::Reason_intrinsic)) { 3606 return false; 3607 } 3608 3609 // Generate dynamic checks. 3610 // Class.cast() is java implementation of _checkcast bytecode. 3611 // Do checkcast (Parse::do_checkcast()) optimizations here. 3612 3613 mirror = null_check(mirror); 3614 // If mirror is dead, only null-path is taken. 3615 if (stopped()) { 3616 return true; 3617 } 3618 3619 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive). 3620 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT }; 3621 RegionNode* region = new RegionNode(PATH_LIMIT); 3622 record_for_igvn(region); 3623 3624 // Now load the mirror's klass metaobject, and null-check it. 3625 // If kls is null, we have a primitive mirror and 3626 // nothing is an instance of a primitive type. 3627 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path); 3628 3629 Node* res = top(); 3630 if (!stopped()) { 3631 Node* bad_type_ctrl = top(); 3632 // Do checkcast optimizations. 3633 res = gen_checkcast(obj, kls, &bad_type_ctrl); 3634 region->init_req(_bad_type_path, bad_type_ctrl); 3635 } 3636 if (region->in(_prim_path) != top() || 3637 region->in(_bad_type_path) != top()) { 3638 // Let Interpreter throw ClassCastException. 3639 PreserveJVMState pjvms(this); 3640 set_control(_gvn.transform(region)); 3641 uncommon_trap(Deoptimization::Reason_intrinsic, 3642 Deoptimization::Action_maybe_recompile); 3643 } 3644 if (!stopped()) { 3645 set_result(res); 3646 } 3647 return true; 3648 } 3649 3650 3651 //--------------------------inline_native_subtype_check------------------------ 3652 // This intrinsic takes the JNI calls out of the heart of 3653 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3654 bool LibraryCallKit::inline_native_subtype_check() { 3655 // Pull both arguments off the stack. 3656 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3657 args[0] = argument(0); 3658 args[1] = argument(1); 3659 Node* klasses[2]; // corresponding Klasses: superk, subk 3660 klasses[0] = klasses[1] = top(); 3661 3662 enum { 3663 // A full decision tree on {superc is prim, subc is prim}: 3664 _prim_0_path = 1, // {P,N} => false 3665 // {P,P} & superc!=subc => false 3666 _prim_same_path, // {P,P} & superc==subc => true 3667 _prim_1_path, // {N,P} => false 3668 _ref_subtype_path, // {N,N} & subtype check wins => true 3669 _both_ref_path, // {N,N} & subtype check loses => false 3670 PATH_LIMIT 3671 }; 3672 3673 RegionNode* region = new RegionNode(PATH_LIMIT); 3674 Node* phi = new PhiNode(region, TypeInt::BOOL); 3675 record_for_igvn(region); 3676 3677 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3678 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3679 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3680 3681 // First null-check both mirrors and load each mirror's klass metaobject. 3682 int which_arg; 3683 for (which_arg = 0; which_arg <= 1; which_arg++) { 3684 Node* arg = args[which_arg]; 3685 arg = null_check(arg); 3686 if (stopped()) break; 3687 args[which_arg] = arg; 3688 3689 Node* p = basic_plus_adr(arg, class_klass_offset); 3690 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type); 3691 klasses[which_arg] = _gvn.transform(kls); 3692 } 3693 3694 // Having loaded both klasses, test each for null. 3695 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3696 for (which_arg = 0; which_arg <= 1; which_arg++) { 3697 Node* kls = klasses[which_arg]; 3698 Node* null_ctl = top(); 3699 kls = null_check_oop(kls, &null_ctl, never_see_null); 3700 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3701 region->init_req(prim_path, null_ctl); 3702 if (stopped()) break; 3703 klasses[which_arg] = kls; 3704 } 3705 3706 if (!stopped()) { 3707 // now we have two reference types, in klasses[0..1] 3708 Node* subk = klasses[1]; // the argument to isAssignableFrom 3709 Node* superk = klasses[0]; // the receiver 3710 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3711 // now we have a successful reference subtype check 3712 region->set_req(_ref_subtype_path, control()); 3713 } 3714 3715 // If both operands are primitive (both klasses null), then 3716 // we must return true when they are identical primitives. 3717 // It is convenient to test this after the first null klass check. 3718 set_control(region->in(_prim_0_path)); // go back to first null check 3719 if (!stopped()) { 3720 // Since superc is primitive, make a guard for the superc==subc case. 3721 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3722 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3723 generate_guard(bol_eq, region, PROB_FAIR); 3724 if (region->req() == PATH_LIMIT+1) { 3725 // A guard was added. If the added guard is taken, superc==subc. 3726 region->swap_edges(PATH_LIMIT, _prim_same_path); 3727 region->del_req(PATH_LIMIT); 3728 } 3729 region->set_req(_prim_0_path, control()); // Not equal after all. 3730 } 3731 3732 // these are the only paths that produce 'true': 3733 phi->set_req(_prim_same_path, intcon(1)); 3734 phi->set_req(_ref_subtype_path, intcon(1)); 3735 3736 // pull together the cases: 3737 assert(region->req() == PATH_LIMIT, "sane region"); 3738 for (uint i = 1; i < region->req(); i++) { 3739 Node* ctl = region->in(i); 3740 if (ctl == NULL || ctl == top()) { 3741 region->set_req(i, top()); 3742 phi ->set_req(i, top()); 3743 } else if (phi->in(i) == NULL) { 3744 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3745 } 3746 } 3747 3748 set_control(_gvn.transform(region)); 3749 set_result(_gvn.transform(phi)); 3750 return true; 3751 } 3752 3753 //---------------------generate_array_guard_common------------------------ 3754 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3755 bool obj_array, bool not_array) { 3756 3757 if (stopped()) { 3758 return NULL; 3759 } 3760 3761 // If obj_array/non_array==false/false: 3762 // Branch around if the given klass is in fact an array (either obj or prim). 3763 // If obj_array/non_array==false/true: 3764 // Branch around if the given klass is not an array klass of any kind. 3765 // If obj_array/non_array==true/true: 3766 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3767 // If obj_array/non_array==true/false: 3768 // Branch around if the kls is an oop array (Object[] or subtype) 3769 // 3770 // Like generate_guard, adds a new path onto the region. 3771 jint layout_con = 0; 3772 Node* layout_val = get_layout_helper(kls, layout_con); 3773 if (layout_val == NULL) { 3774 bool query = (obj_array 3775 ? Klass::layout_helper_is_objArray(layout_con) 3776 : Klass::layout_helper_is_array(layout_con)); 3777 if (query == not_array) { 3778 return NULL; // never a branch 3779 } else { // always a branch 3780 Node* always_branch = control(); 3781 if (region != NULL) 3782 region->add_req(always_branch); 3783 set_control(top()); 3784 return always_branch; 3785 } 3786 } 3787 // Now test the correct condition. 3788 jint nval = (obj_array 3789 ? ((jint)Klass::_lh_array_tag_type_value 3790 << Klass::_lh_array_tag_shift) 3791 : Klass::_lh_neutral_value); 3792 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3793 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3794 // invert the test if we are looking for a non-array 3795 if (not_array) btest = BoolTest(btest).negate(); 3796 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3797 return generate_fair_guard(bol, region); 3798 } 3799 3800 3801 //-----------------------inline_native_newArray-------------------------- 3802 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3803 bool LibraryCallKit::inline_native_newArray() { 3804 Node* mirror = argument(0); 3805 Node* count_val = argument(1); 3806 3807 mirror = null_check(mirror); 3808 // If mirror or obj is dead, only null-path is taken. 3809 if (stopped()) return true; 3810 3811 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3812 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3813 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3814 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3815 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3816 3817 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3818 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3819 result_reg, _slow_path); 3820 Node* normal_ctl = control(); 3821 Node* no_array_ctl = result_reg->in(_slow_path); 3822 3823 // Generate code for the slow case. We make a call to newArray(). 3824 set_control(no_array_ctl); 3825 if (!stopped()) { 3826 // Either the input type is void.class, or else the 3827 // array klass has not yet been cached. Either the 3828 // ensuing call will throw an exception, or else it 3829 // will cache the array klass for next time. 3830 PreserveJVMState pjvms(this); 3831 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3832 Node* slow_result = set_results_for_java_call(slow_call); 3833 // this->control() comes from set_results_for_java_call 3834 result_reg->set_req(_slow_path, control()); 3835 result_val->set_req(_slow_path, slow_result); 3836 result_io ->set_req(_slow_path, i_o()); 3837 result_mem->set_req(_slow_path, reset_memory()); 3838 } 3839 3840 set_control(normal_ctl); 3841 if (!stopped()) { 3842 // Normal case: The array type has been cached in the java.lang.Class. 3843 // The following call works fine even if the array type is polymorphic. 3844 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3845 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3846 result_reg->init_req(_normal_path, control()); 3847 result_val->init_req(_normal_path, obj); 3848 result_io ->init_req(_normal_path, i_o()); 3849 result_mem->init_req(_normal_path, reset_memory()); 3850 } 3851 3852 // Return the combined state. 3853 set_i_o( _gvn.transform(result_io) ); 3854 set_all_memory( _gvn.transform(result_mem)); 3855 3856 C->set_has_split_ifs(true); // Has chance for split-if optimization 3857 set_result(result_reg, result_val); 3858 return true; 3859 } 3860 3861 //----------------------inline_native_getLength-------------------------- 3862 // public static native int java.lang.reflect.Array.getLength(Object array); 3863 bool LibraryCallKit::inline_native_getLength() { 3864 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3865 3866 Node* array = null_check(argument(0)); 3867 // If array is dead, only null-path is taken. 3868 if (stopped()) return true; 3869 3870 // Deoptimize if it is a non-array. 3871 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3872 3873 if (non_array != NULL) { 3874 PreserveJVMState pjvms(this); 3875 set_control(non_array); 3876 uncommon_trap(Deoptimization::Reason_intrinsic, 3877 Deoptimization::Action_maybe_recompile); 3878 } 3879 3880 // If control is dead, only non-array-path is taken. 3881 if (stopped()) return true; 3882 3883 // The works fine even if the array type is polymorphic. 3884 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3885 Node* result = load_array_length(array); 3886 3887 C->set_has_split_ifs(true); // Has chance for split-if optimization 3888 set_result(result); 3889 return true; 3890 } 3891 3892 //------------------------inline_array_copyOf---------------------------- 3893 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3894 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3895 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3896 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3897 3898 // Get the arguments. 3899 Node* original = argument(0); 3900 Node* start = is_copyOfRange? argument(1): intcon(0); 3901 Node* end = is_copyOfRange? argument(2): argument(1); 3902 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3903 3904 Node* newcopy; 3905 3906 // Set the original stack and the reexecute bit for the interpreter to reexecute 3907 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3908 { PreserveReexecuteState preexecs(this); 3909 jvms()->set_should_reexecute(true); 3910 3911 array_type_mirror = null_check(array_type_mirror); 3912 original = null_check(original); 3913 3914 // Check if a null path was taken unconditionally. 3915 if (stopped()) return true; 3916 3917 Node* orig_length = load_array_length(original); 3918 3919 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3920 klass_node = null_check(klass_node); 3921 3922 RegionNode* bailout = new RegionNode(1); 3923 record_for_igvn(bailout); 3924 3925 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3926 // Bail out if that is so. 3927 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3928 if (not_objArray != NULL) { 3929 // Improve the klass node's type from the new optimistic assumption: 3930 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3931 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3932 Node* cast = new CastPPNode(klass_node, akls); 3933 cast->init_req(0, control()); 3934 klass_node = _gvn.transform(cast); 3935 } 3936 3937 // Bail out if either start or end is negative. 3938 generate_negative_guard(start, bailout, &start); 3939 generate_negative_guard(end, bailout, &end); 3940 3941 Node* length = end; 3942 if (_gvn.type(start) != TypeInt::ZERO) { 3943 length = _gvn.transform(new SubINode(end, start)); 3944 } 3945 3946 // Bail out if length is negative. 3947 // Without this the new_array would throw 3948 // NegativeArraySizeException but IllegalArgumentException is what 3949 // should be thrown 3950 generate_negative_guard(length, bailout, &length); 3951 3952 if (bailout->req() > 1) { 3953 PreserveJVMState pjvms(this); 3954 set_control(_gvn.transform(bailout)); 3955 uncommon_trap(Deoptimization::Reason_intrinsic, 3956 Deoptimization::Action_maybe_recompile); 3957 } 3958 3959 if (!stopped()) { 3960 // How many elements will we copy from the original? 3961 // The answer is MinI(orig_length - start, length). 3962 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3963 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3964 3965 // Generate a direct call to the right arraycopy function(s). 3966 // We know the copy is disjoint but we might not know if the 3967 // oop stores need checking. 3968 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3969 // This will fail a store-check if x contains any non-nulls. 3970 3971 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to 3972 // loads/stores but it is legal only if we're sure the 3973 // Arrays.copyOf would succeed. So we need all input arguments 3974 // to the copyOf to be validated, including that the copy to the 3975 // new array won't trigger an ArrayStoreException. That subtype 3976 // check can be optimized if we know something on the type of 3977 // the input array from type speculation. 3978 if (_gvn.type(klass_node)->singleton()) { 3979 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass(); 3980 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass(); 3981 3982 int test = C->static_subtype_check(superk, subk); 3983 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) { 3984 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr(); 3985 if (t_original->speculative_type() != NULL) { 3986 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true); 3987 } 3988 } 3989 } 3990 3991 bool validated = false; 3992 // Reason_class_check rather than Reason_intrinsic because we 3993 // want to intrinsify even if this traps. 3994 if (!too_many_traps(Deoptimization::Reason_class_check)) { 3995 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original), 3996 klass_node); 3997 3998 if (not_subtype_ctrl != top()) { 3999 PreserveJVMState pjvms(this); 4000 set_control(not_subtype_ctrl); 4001 uncommon_trap(Deoptimization::Reason_class_check, 4002 Deoptimization::Action_make_not_entrant); 4003 assert(stopped(), "Should be stopped"); 4004 } 4005 validated = true; 4006 } 4007 4008 if (!stopped()) { 4009 newcopy = new_array(klass_node, length, 0); // no arguments to push 4010 4011 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, 4012 load_object_klass(original), klass_node); 4013 if (!is_copyOfRange) { 4014 ac->set_copyof(validated); 4015 } else { 4016 ac->set_copyofrange(validated); 4017 } 4018 Node* n = _gvn.transform(ac); 4019 if (n == ac) { 4020 ac->connect_outputs(this); 4021 } else { 4022 assert(validated, "shouldn't transform if all arguments not validated"); 4023 set_all_memory(n); 4024 } 4025 } 4026 } 4027 } // original reexecute is set back here 4028 4029 C->set_has_split_ifs(true); // Has chance for split-if optimization 4030 if (!stopped()) { 4031 set_result(newcopy); 4032 } 4033 return true; 4034 } 4035 4036 4037 //----------------------generate_virtual_guard--------------------------- 4038 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 4039 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 4040 RegionNode* slow_region) { 4041 ciMethod* method = callee(); 4042 int vtable_index = method->vtable_index(); 4043 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4044 err_msg_res("bad index %d", vtable_index)); 4045 // Get the Method* out of the appropriate vtable entry. 4046 int entry_offset = (InstanceKlass::vtable_start_offset() + 4047 vtable_index*vtableEntry::size()) * wordSize + 4048 vtableEntry::method_offset_in_bytes(); 4049 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 4050 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 4051 4052 // Compare the target method with the expected method (e.g., Object.hashCode). 4053 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 4054 4055 Node* native_call = makecon(native_call_addr); 4056 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 4057 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 4058 4059 return generate_slow_guard(test_native, slow_region); 4060 } 4061 4062 //-----------------------generate_method_call---------------------------- 4063 // Use generate_method_call to make a slow-call to the real 4064 // method if the fast path fails. An alternative would be to 4065 // use a stub like OptoRuntime::slow_arraycopy_Java. 4066 // This only works for expanding the current library call, 4067 // not another intrinsic. (E.g., don't use this for making an 4068 // arraycopy call inside of the copyOf intrinsic.) 4069 CallJavaNode* 4070 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 4071 // When compiling the intrinsic method itself, do not use this technique. 4072 guarantee(callee() != C->method(), "cannot make slow-call to self"); 4073 4074 ciMethod* method = callee(); 4075 // ensure the JVMS we have will be correct for this call 4076 guarantee(method_id == method->intrinsic_id(), "must match"); 4077 4078 const TypeFunc* tf = TypeFunc::make(method); 4079 CallJavaNode* slow_call; 4080 if (is_static) { 4081 assert(!is_virtual, ""); 4082 slow_call = new CallStaticJavaNode(C, tf, 4083 SharedRuntime::get_resolve_static_call_stub(), 4084 method, bci()); 4085 } else if (is_virtual) { 4086 null_check_receiver(); 4087 int vtable_index = Method::invalid_vtable_index; 4088 if (UseInlineCaches) { 4089 // Suppress the vtable call 4090 } else { 4091 // hashCode and clone are not a miranda methods, 4092 // so the vtable index is fixed. 4093 // No need to use the linkResolver to get it. 4094 vtable_index = method->vtable_index(); 4095 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 4096 err_msg_res("bad index %d", vtable_index)); 4097 } 4098 slow_call = new CallDynamicJavaNode(tf, 4099 SharedRuntime::get_resolve_virtual_call_stub(), 4100 method, vtable_index, bci()); 4101 } else { // neither virtual nor static: opt_virtual 4102 null_check_receiver(); 4103 slow_call = new CallStaticJavaNode(C, tf, 4104 SharedRuntime::get_resolve_opt_virtual_call_stub(), 4105 method, bci()); 4106 slow_call->set_optimized_virtual(true); 4107 } 4108 set_arguments_for_java_call(slow_call); 4109 set_edges_for_java_call(slow_call); 4110 return slow_call; 4111 } 4112 4113 4114 /** 4115 * Build special case code for calls to hashCode on an object. This call may 4116 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 4117 * slightly different code. 4118 */ 4119 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 4120 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 4121 assert(!(is_virtual && is_static), "either virtual, special, or static"); 4122 4123 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 4124 4125 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4126 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 4127 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 4128 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4129 Node* obj = NULL; 4130 if (!is_static) { 4131 // Check for hashing null object 4132 obj = null_check_receiver(); 4133 if (stopped()) return true; // unconditionally null 4134 result_reg->init_req(_null_path, top()); 4135 result_val->init_req(_null_path, top()); 4136 } else { 4137 // Do a null check, and return zero if null. 4138 // System.identityHashCode(null) == 0 4139 obj = argument(0); 4140 Node* null_ctl = top(); 4141 obj = null_check_oop(obj, &null_ctl); 4142 result_reg->init_req(_null_path, null_ctl); 4143 result_val->init_req(_null_path, _gvn.intcon(0)); 4144 } 4145 4146 // Unconditionally null? Then return right away. 4147 if (stopped()) { 4148 set_control( result_reg->in(_null_path)); 4149 if (!stopped()) 4150 set_result(result_val->in(_null_path)); 4151 return true; 4152 } 4153 4154 // We only go to the fast case code if we pass a number of guards. The 4155 // paths which do not pass are accumulated in the slow_region. 4156 RegionNode* slow_region = new RegionNode(1); 4157 record_for_igvn(slow_region); 4158 4159 // If this is a virtual call, we generate a funny guard. We pull out 4160 // the vtable entry corresponding to hashCode() from the target object. 4161 // If the target method which we are calling happens to be the native 4162 // Object hashCode() method, we pass the guard. We do not need this 4163 // guard for non-virtual calls -- the caller is known to be the native 4164 // Object hashCode(). 4165 if (is_virtual) { 4166 // After null check, get the object's klass. 4167 Node* obj_klass = load_object_klass(obj); 4168 generate_virtual_guard(obj_klass, slow_region); 4169 } 4170 4171 // Get the header out of the object, use LoadMarkNode when available 4172 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4173 // The control of the load must be NULL. Otherwise, the load can move before 4174 // the null check after castPP removal. 4175 Node* no_ctrl = NULL; 4176 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4177 4178 // Test the header to see if it is unlocked. 4179 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4180 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4181 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4182 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4183 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4184 4185 generate_slow_guard(test_unlocked, slow_region); 4186 4187 // Get the hash value and check to see that it has been properly assigned. 4188 // We depend on hash_mask being at most 32 bits and avoid the use of 4189 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4190 // vm: see markOop.hpp. 4191 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4192 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4193 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4194 // This hack lets the hash bits live anywhere in the mark object now, as long 4195 // as the shift drops the relevant bits into the low 32 bits. Note that 4196 // Java spec says that HashCode is an int so there's no point in capturing 4197 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4198 hshifted_header = ConvX2I(hshifted_header); 4199 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4200 4201 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4202 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4203 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4204 4205 generate_slow_guard(test_assigned, slow_region); 4206 4207 Node* init_mem = reset_memory(); 4208 // fill in the rest of the null path: 4209 result_io ->init_req(_null_path, i_o()); 4210 result_mem->init_req(_null_path, init_mem); 4211 4212 result_val->init_req(_fast_path, hash_val); 4213 result_reg->init_req(_fast_path, control()); 4214 result_io ->init_req(_fast_path, i_o()); 4215 result_mem->init_req(_fast_path, init_mem); 4216 4217 // Generate code for the slow case. We make a call to hashCode(). 4218 set_control(_gvn.transform(slow_region)); 4219 if (!stopped()) { 4220 // No need for PreserveJVMState, because we're using up the present state. 4221 set_all_memory(init_mem); 4222 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4223 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4224 Node* slow_result = set_results_for_java_call(slow_call); 4225 // this->control() comes from set_results_for_java_call 4226 result_reg->init_req(_slow_path, control()); 4227 result_val->init_req(_slow_path, slow_result); 4228 result_io ->set_req(_slow_path, i_o()); 4229 result_mem ->set_req(_slow_path, reset_memory()); 4230 } 4231 4232 // Return the combined state. 4233 set_i_o( _gvn.transform(result_io) ); 4234 set_all_memory( _gvn.transform(result_mem)); 4235 4236 set_result(result_reg, result_val); 4237 return true; 4238 } 4239 4240 //---------------------------inline_native_getClass---------------------------- 4241 // public final native Class<?> java.lang.Object.getClass(); 4242 // 4243 // Build special case code for calls to getClass on an object. 4244 bool LibraryCallKit::inline_native_getClass() { 4245 Node* obj = null_check_receiver(); 4246 if (stopped()) return true; 4247 set_result(load_mirror_from_klass(load_object_klass(obj))); 4248 return true; 4249 } 4250 4251 //-----------------inline_native_Reflection_getCallerClass--------------------- 4252 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4253 // 4254 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4255 // 4256 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4257 // in that it must skip particular security frames and checks for 4258 // caller sensitive methods. 4259 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4260 #ifndef PRODUCT 4261 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4262 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4263 } 4264 #endif 4265 4266 if (!jvms()->has_method()) { 4267 #ifndef PRODUCT 4268 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4269 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4270 } 4271 #endif 4272 return false; 4273 } 4274 4275 // Walk back up the JVM state to find the caller at the required 4276 // depth. 4277 JVMState* caller_jvms = jvms(); 4278 4279 // Cf. JVM_GetCallerClass 4280 // NOTE: Start the loop at depth 1 because the current JVM state does 4281 // not include the Reflection.getCallerClass() frame. 4282 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4283 ciMethod* m = caller_jvms->method(); 4284 switch (n) { 4285 case 0: 4286 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4287 break; 4288 case 1: 4289 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4290 if (!m->caller_sensitive()) { 4291 #ifndef PRODUCT 4292 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4293 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4294 } 4295 #endif 4296 return false; // bail-out; let JVM_GetCallerClass do the work 4297 } 4298 break; 4299 default: 4300 if (!m->is_ignored_by_security_stack_walk()) { 4301 // We have reached the desired frame; return the holder class. 4302 // Acquire method holder as java.lang.Class and push as constant. 4303 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4304 ciInstance* caller_mirror = caller_klass->java_mirror(); 4305 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4306 4307 #ifndef PRODUCT 4308 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4309 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()); 4310 tty->print_cr(" JVM state at this point:"); 4311 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4312 ciMethod* m = jvms()->of_depth(i)->method(); 4313 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4314 } 4315 } 4316 #endif 4317 return true; 4318 } 4319 break; 4320 } 4321 } 4322 4323 #ifndef PRODUCT 4324 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4325 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4326 tty->print_cr(" JVM state at this point:"); 4327 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4328 ciMethod* m = jvms()->of_depth(i)->method(); 4329 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4330 } 4331 } 4332 #endif 4333 4334 return false; // bail-out; let JVM_GetCallerClass do the work 4335 } 4336 4337 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4338 Node* arg = argument(0); 4339 Node* result; 4340 4341 switch (id) { 4342 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4343 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4344 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4345 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4346 4347 case vmIntrinsics::_doubleToLongBits: { 4348 // two paths (plus control) merge in a wood 4349 RegionNode *r = new RegionNode(3); 4350 Node *phi = new PhiNode(r, TypeLong::LONG); 4351 4352 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4353 // Build the boolean node 4354 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4355 4356 // Branch either way. 4357 // NaN case is less traveled, which makes all the difference. 4358 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4359 Node *opt_isnan = _gvn.transform(ifisnan); 4360 assert( opt_isnan->is_If(), "Expect an IfNode"); 4361 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4362 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4363 4364 set_control(iftrue); 4365 4366 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4367 Node *slow_result = longcon(nan_bits); // return NaN 4368 phi->init_req(1, _gvn.transform( slow_result )); 4369 r->init_req(1, iftrue); 4370 4371 // Else fall through 4372 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4373 set_control(iffalse); 4374 4375 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4376 r->init_req(2, iffalse); 4377 4378 // Post merge 4379 set_control(_gvn.transform(r)); 4380 record_for_igvn(r); 4381 4382 C->set_has_split_ifs(true); // Has chance for split-if optimization 4383 result = phi; 4384 assert(result->bottom_type()->isa_long(), "must be"); 4385 break; 4386 } 4387 4388 case vmIntrinsics::_floatToIntBits: { 4389 // two paths (plus control) merge in a wood 4390 RegionNode *r = new RegionNode(3); 4391 Node *phi = new PhiNode(r, TypeInt::INT); 4392 4393 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4394 // Build the boolean node 4395 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4396 4397 // Branch either way. 4398 // NaN case is less traveled, which makes all the difference. 4399 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4400 Node *opt_isnan = _gvn.transform(ifisnan); 4401 assert( opt_isnan->is_If(), "Expect an IfNode"); 4402 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4403 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4404 4405 set_control(iftrue); 4406 4407 static const jint nan_bits = 0x7fc00000; 4408 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4409 phi->init_req(1, _gvn.transform( slow_result )); 4410 r->init_req(1, iftrue); 4411 4412 // Else fall through 4413 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4414 set_control(iffalse); 4415 4416 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4417 r->init_req(2, iffalse); 4418 4419 // Post merge 4420 set_control(_gvn.transform(r)); 4421 record_for_igvn(r); 4422 4423 C->set_has_split_ifs(true); // Has chance for split-if optimization 4424 result = phi; 4425 assert(result->bottom_type()->isa_int(), "must be"); 4426 break; 4427 } 4428 4429 default: 4430 fatal_unexpected_iid(id); 4431 break; 4432 } 4433 set_result(_gvn.transform(result)); 4434 return true; 4435 } 4436 4437 #ifdef _LP64 4438 #define XTOP ,top() /*additional argument*/ 4439 #else //_LP64 4440 #define XTOP /*no additional argument*/ 4441 #endif //_LP64 4442 4443 //----------------------inline_unsafe_copyMemory------------------------- 4444 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4445 bool LibraryCallKit::inline_unsafe_copyMemory() { 4446 if (callee()->is_static()) return false; // caller must have the capability! 4447 null_check_receiver(); // null-check receiver 4448 if (stopped()) return true; 4449 4450 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4451 4452 Node* src_ptr = argument(1); // type: oop 4453 Node* src_off = ConvL2X(argument(2)); // type: long 4454 Node* dst_ptr = argument(4); // type: oop 4455 Node* dst_off = ConvL2X(argument(5)); // type: long 4456 Node* size = ConvL2X(argument(7)); // type: long 4457 4458 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4459 "fieldOffset must be byte-scaled"); 4460 4461 Node* src = make_unsafe_address(src_ptr, src_off); 4462 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4463 4464 // Conservatively insert a memory barrier on all memory slices. 4465 // Do not let writes of the copy source or destination float below the copy. 4466 insert_mem_bar(Op_MemBarCPUOrder); 4467 4468 // Call it. Note that the length argument is not scaled. 4469 make_runtime_call(RC_LEAF|RC_NO_FP, 4470 OptoRuntime::fast_arraycopy_Type(), 4471 StubRoutines::unsafe_arraycopy(), 4472 "unsafe_arraycopy", 4473 TypeRawPtr::BOTTOM, 4474 src, dst, size XTOP); 4475 4476 // Do not let reads of the copy destination float above the copy. 4477 insert_mem_bar(Op_MemBarCPUOrder); 4478 4479 return true; 4480 } 4481 4482 //------------------------clone_coping----------------------------------- 4483 // Helper function for inline_native_clone. 4484 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4485 assert(obj_size != NULL, ""); 4486 Node* raw_obj = alloc_obj->in(1); 4487 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4488 4489 AllocateNode* alloc = NULL; 4490 if (ReduceBulkZeroing) { 4491 // We will be completely responsible for initializing this object - 4492 // mark Initialize node as complete. 4493 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4494 // The object was just allocated - there should be no any stores! 4495 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4496 // Mark as complete_with_arraycopy so that on AllocateNode 4497 // expansion, we know this AllocateNode is initialized by an array 4498 // copy and a StoreStore barrier exists after the array copy. 4499 alloc->initialization()->set_complete_with_arraycopy(); 4500 } 4501 4502 // Copy the fastest available way. 4503 // TODO: generate fields copies for small objects instead. 4504 Node* src = obj; 4505 Node* dest = alloc_obj; 4506 Node* size = _gvn.transform(obj_size); 4507 4508 // Exclude the header but include array length to copy by 8 bytes words. 4509 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4510 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4511 instanceOopDesc::base_offset_in_bytes(); 4512 // base_off: 4513 // 8 - 32-bit VM 4514 // 12 - 64-bit VM, compressed klass 4515 // 16 - 64-bit VM, normal klass 4516 if (base_off % BytesPerLong != 0) { 4517 assert(UseCompressedClassPointers, ""); 4518 if (is_array) { 4519 // Exclude length to copy by 8 bytes words. 4520 base_off += sizeof(int); 4521 } else { 4522 // Include klass to copy by 8 bytes words. 4523 base_off = instanceOopDesc::klass_offset_in_bytes(); 4524 } 4525 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4526 } 4527 src = basic_plus_adr(src, base_off); 4528 dest = basic_plus_adr(dest, base_off); 4529 4530 // Compute the length also, if needed: 4531 Node* countx = size; 4532 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4533 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4534 4535 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4536 4537 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4538 ac->set_clonebasic(); 4539 Node* n = _gvn.transform(ac); 4540 if (n == ac) { 4541 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4542 } else { 4543 set_all_memory(n); 4544 } 4545 4546 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4547 if (card_mark) { 4548 assert(!is_array, ""); 4549 // Put in store barrier for any and all oops we are sticking 4550 // into this object. (We could avoid this if we could prove 4551 // that the object type contains no oop fields at all.) 4552 Node* no_particular_value = NULL; 4553 Node* no_particular_field = NULL; 4554 int raw_adr_idx = Compile::AliasIdxRaw; 4555 post_barrier(control(), 4556 memory(raw_adr_type), 4557 alloc_obj, 4558 no_particular_field, 4559 raw_adr_idx, 4560 no_particular_value, 4561 T_OBJECT, 4562 false); 4563 } 4564 4565 // Do not let reads from the cloned object float above the arraycopy. 4566 if (alloc != NULL) { 4567 // Do not let stores that initialize this object be reordered with 4568 // a subsequent store that would make this object accessible by 4569 // other threads. 4570 // Record what AllocateNode this StoreStore protects so that 4571 // escape analysis can go from the MemBarStoreStoreNode to the 4572 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4573 // based on the escape status of the AllocateNode. 4574 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4575 } else { 4576 insert_mem_bar(Op_MemBarCPUOrder); 4577 } 4578 } 4579 4580 //------------------------inline_native_clone---------------------------- 4581 // protected native Object java.lang.Object.clone(); 4582 // 4583 // Here are the simple edge cases: 4584 // null receiver => normal trap 4585 // virtual and clone was overridden => slow path to out-of-line clone 4586 // not cloneable or finalizer => slow path to out-of-line Object.clone 4587 // 4588 // The general case has two steps, allocation and copying. 4589 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4590 // 4591 // Copying also has two cases, oop arrays and everything else. 4592 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4593 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4594 // 4595 // These steps fold up nicely if and when the cloned object's klass 4596 // can be sharply typed as an object array, a type array, or an instance. 4597 // 4598 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4599 PhiNode* result_val; 4600 4601 // Set the reexecute bit for the interpreter to reexecute 4602 // the bytecode that invokes Object.clone if deoptimization happens. 4603 { PreserveReexecuteState preexecs(this); 4604 jvms()->set_should_reexecute(true); 4605 4606 Node* obj = null_check_receiver(); 4607 if (stopped()) return true; 4608 4609 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 4610 4611 // If we are going to clone an instance, we need its exact type to 4612 // know the number and types of fields to convert the clone to 4613 // loads/stores. Maybe a speculative type can help us. 4614 if (!obj_type->klass_is_exact() && 4615 obj_type->speculative_type() != NULL && 4616 obj_type->speculative_type()->is_instance_klass()) { 4617 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass(); 4618 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem && 4619 !spec_ik->has_injected_fields()) { 4620 ciKlass* k = obj_type->klass(); 4621 if (!k->is_instance_klass() || 4622 k->as_instance_klass()->is_interface() || 4623 k->as_instance_klass()->has_subklass()) { 4624 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false); 4625 } 4626 } 4627 } 4628 4629 Node* obj_klass = load_object_klass(obj); 4630 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4631 const TypeOopPtr* toop = ((tklass != NULL) 4632 ? tklass->as_instance_type() 4633 : TypeInstPtr::NOTNULL); 4634 4635 // Conservatively insert a memory barrier on all memory slices. 4636 // Do not let writes into the original float below the clone. 4637 insert_mem_bar(Op_MemBarCPUOrder); 4638 4639 // paths into result_reg: 4640 enum { 4641 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4642 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4643 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4644 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4645 PATH_LIMIT 4646 }; 4647 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4648 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4649 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4650 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4651 record_for_igvn(result_reg); 4652 4653 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4654 int raw_adr_idx = Compile::AliasIdxRaw; 4655 4656 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4657 if (array_ctl != NULL) { 4658 // It's an array. 4659 PreserveJVMState pjvms(this); 4660 set_control(array_ctl); 4661 Node* obj_length = load_array_length(obj); 4662 Node* obj_size = NULL; 4663 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4664 4665 if (!use_ReduceInitialCardMarks()) { 4666 // If it is an oop array, it requires very special treatment, 4667 // because card marking is required on each card of the array. 4668 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4669 if (is_obja != NULL) { 4670 PreserveJVMState pjvms2(this); 4671 set_control(is_obja); 4672 // Generate a direct call to the right arraycopy function(s). 4673 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4674 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4675 ac->set_cloneoop(); 4676 Node* n = _gvn.transform(ac); 4677 assert(n == ac, "cannot disappear"); 4678 ac->connect_outputs(this); 4679 4680 result_reg->init_req(_objArray_path, control()); 4681 result_val->init_req(_objArray_path, alloc_obj); 4682 result_i_o ->set_req(_objArray_path, i_o()); 4683 result_mem ->set_req(_objArray_path, reset_memory()); 4684 } 4685 } 4686 // Otherwise, there are no card marks to worry about. 4687 // (We can dispense with card marks if we know the allocation 4688 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4689 // causes the non-eden paths to take compensating steps to 4690 // simulate a fresh allocation, so that no further 4691 // card marks are required in compiled code to initialize 4692 // the object.) 4693 4694 if (!stopped()) { 4695 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4696 4697 // Present the results of the copy. 4698 result_reg->init_req(_array_path, control()); 4699 result_val->init_req(_array_path, alloc_obj); 4700 result_i_o ->set_req(_array_path, i_o()); 4701 result_mem ->set_req(_array_path, reset_memory()); 4702 } 4703 } 4704 4705 // We only go to the instance fast case code if we pass a number of guards. 4706 // The paths which do not pass are accumulated in the slow_region. 4707 RegionNode* slow_region = new RegionNode(1); 4708 record_for_igvn(slow_region); 4709 if (!stopped()) { 4710 // It's an instance (we did array above). Make the slow-path tests. 4711 // If this is a virtual call, we generate a funny guard. We grab 4712 // the vtable entry corresponding to clone() from the target object. 4713 // If the target method which we are calling happens to be the 4714 // Object clone() method, we pass the guard. We do not need this 4715 // guard for non-virtual calls; the caller is known to be the native 4716 // Object clone(). 4717 if (is_virtual) { 4718 generate_virtual_guard(obj_klass, slow_region); 4719 } 4720 4721 // The object must be cloneable and must not have a finalizer. 4722 // Both of these conditions may be checked in a single test. 4723 // We could optimize the cloneable test further, but we don't care. 4724 generate_access_flags_guard(obj_klass, 4725 // Test both conditions: 4726 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4727 // Must be cloneable but not finalizer: 4728 JVM_ACC_IS_CLONEABLE, 4729 slow_region); 4730 } 4731 4732 if (!stopped()) { 4733 // It's an instance, and it passed the slow-path tests. 4734 PreserveJVMState pjvms(this); 4735 Node* obj_size = NULL; 4736 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4737 // is reexecuted if deoptimization occurs and there could be problems when merging 4738 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4739 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4740 4741 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4742 4743 // Present the results of the slow call. 4744 result_reg->init_req(_instance_path, control()); 4745 result_val->init_req(_instance_path, alloc_obj); 4746 result_i_o ->set_req(_instance_path, i_o()); 4747 result_mem ->set_req(_instance_path, reset_memory()); 4748 } 4749 4750 // Generate code for the slow case. We make a call to clone(). 4751 set_control(_gvn.transform(slow_region)); 4752 if (!stopped()) { 4753 PreserveJVMState pjvms(this); 4754 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4755 Node* slow_result = set_results_for_java_call(slow_call); 4756 // this->control() comes from set_results_for_java_call 4757 result_reg->init_req(_slow_path, control()); 4758 result_val->init_req(_slow_path, slow_result); 4759 result_i_o ->set_req(_slow_path, i_o()); 4760 result_mem ->set_req(_slow_path, reset_memory()); 4761 } 4762 4763 // Return the combined state. 4764 set_control( _gvn.transform(result_reg)); 4765 set_i_o( _gvn.transform(result_i_o)); 4766 set_all_memory( _gvn.transform(result_mem)); 4767 } // original reexecute is set back here 4768 4769 set_result(_gvn.transform(result_val)); 4770 return true; 4771 } 4772 4773 // If we have a tighly coupled allocation, the arraycopy may take care 4774 // of the array initialization. If one of the guards we insert between 4775 // the allocation and the arraycopy causes a deoptimization, an 4776 // unitialized array will escape the compiled method. To prevent that 4777 // we set the JVM state for uncommon traps between the allocation and 4778 // the arraycopy to the state before the allocation so, in case of 4779 // deoptimization, we'll reexecute the allocation and the 4780 // initialization. 4781 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) { 4782 if (alloc != NULL) { 4783 ciMethod* trap_method = alloc->jvms()->method(); 4784 int trap_bci = alloc->jvms()->bci(); 4785 4786 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) & 4787 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) { 4788 // Make sure there's no store between the allocation and the 4789 // arraycopy otherwise visible side effects could be rexecuted 4790 // in case of deoptimization and cause incorrect execution. 4791 bool no_interfering_store = true; 4792 Node* mem = alloc->in(TypeFunc::Memory); 4793 if (mem->is_MergeMem()) { 4794 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) { 4795 Node* n = mms.memory(); 4796 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4797 assert(n->is_Store(), "what else?"); 4798 no_interfering_store = false; 4799 break; 4800 } 4801 } 4802 } else { 4803 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) { 4804 Node* n = mms.memory(); 4805 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) { 4806 assert(n->is_Store(), "what else?"); 4807 no_interfering_store = false; 4808 break; 4809 } 4810 } 4811 } 4812 4813 if (no_interfering_store) { 4814 JVMState* old_jvms = alloc->jvms()->clone_shallow(C); 4815 uint size = alloc->req(); 4816 SafePointNode* sfpt = new SafePointNode(size, old_jvms); 4817 old_jvms->set_map(sfpt); 4818 for (uint i = 0; i < size; i++) { 4819 sfpt->init_req(i, alloc->in(i)); 4820 } 4821 // re-push array length for deoptimization 4822 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength)); 4823 old_jvms->set_sp(old_jvms->sp()+1); 4824 old_jvms->set_monoff(old_jvms->monoff()+1); 4825 old_jvms->set_scloff(old_jvms->scloff()+1); 4826 old_jvms->set_endoff(old_jvms->endoff()+1); 4827 old_jvms->set_should_reexecute(true); 4828 4829 sfpt->set_i_o(map()->i_o()); 4830 sfpt->set_memory(map()->memory()); 4831 sfpt->set_control(map()->control()); 4832 4833 JVMState* saved_jvms = jvms(); 4834 saved_reexecute_sp = _reexecute_sp; 4835 4836 set_jvms(sfpt->jvms()); 4837 _reexecute_sp = jvms()->sp(); 4838 4839 return saved_jvms; 4840 } 4841 } 4842 } 4843 return NULL; 4844 } 4845 4846 // In case of a deoptimization, we restart execution at the 4847 // allocation, allocating a new array. We would leave an uninitialized 4848 // array in the heap that GCs wouldn't expect. Move the allocation 4849 // after the traps so we don't allocate the array if we 4850 // deoptimize. This is possible because tightly_coupled_allocation() 4851 // guarantees there's no observer of the allocated array at this point 4852 // and the control flow is simple enough. 4853 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) { 4854 if (saved_jvms != NULL && !stopped()) { 4855 assert(alloc != NULL, "only with a tightly coupled allocation"); 4856 // restore JVM state to the state at the arraycopy 4857 saved_jvms->map()->set_control(map()->control()); 4858 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?"); 4859 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?"); 4860 // If we've improved the types of some nodes (null check) while 4861 // emitting the guards, propagate them to the current state 4862 map()->replaced_nodes().apply(saved_jvms->map()); 4863 set_jvms(saved_jvms); 4864 _reexecute_sp = saved_reexecute_sp; 4865 4866 // Remove the allocation from above the guards 4867 CallProjections callprojs; 4868 alloc->extract_projections(&callprojs, true); 4869 InitializeNode* init = alloc->initialization(); 4870 Node* alloc_mem = alloc->in(TypeFunc::Memory); 4871 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O)); 4872 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem); 4873 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0)); 4874 4875 // move the allocation here (after the guards) 4876 _gvn.hash_delete(alloc); 4877 alloc->set_req(TypeFunc::Control, control()); 4878 alloc->set_req(TypeFunc::I_O, i_o()); 4879 Node *mem = reset_memory(); 4880 set_all_memory(mem); 4881 alloc->set_req(TypeFunc::Memory, mem); 4882 set_control(init->proj_out(TypeFunc::Control)); 4883 set_i_o(callprojs.fallthrough_ioproj); 4884 4885 // Update memory as done in GraphKit::set_output_for_allocation() 4886 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength)); 4887 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type(); 4888 if (ary_type->isa_aryptr() && length_type != NULL) { 4889 ary_type = ary_type->is_aryptr()->cast_to_size(length_type); 4890 } 4891 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot); 4892 int elemidx = C->get_alias_index(telemref); 4893 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw); 4894 set_memory(init->proj_out(TypeFunc::Memory), elemidx); 4895 4896 Node* allocx = _gvn.transform(alloc); 4897 assert(allocx == alloc, "where has the allocation gone?"); 4898 assert(dest->is_CheckCastPP(), "not an allocation result?"); 4899 4900 _gvn.hash_delete(dest); 4901 dest->set_req(0, control()); 4902 Node* destx = _gvn.transform(dest); 4903 assert(destx == dest, "where has the allocation result gone?"); 4904 } 4905 } 4906 4907 4908 //------------------------------inline_arraycopy----------------------- 4909 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4910 // Object dest, int destPos, 4911 // int length); 4912 bool LibraryCallKit::inline_arraycopy() { 4913 // Get the arguments. 4914 Node* src = argument(0); // type: oop 4915 Node* src_offset = argument(1); // type: int 4916 Node* dest = argument(2); // type: oop 4917 Node* dest_offset = argument(3); // type: int 4918 Node* length = argument(4); // type: int 4919 4920 4921 // Check for allocation before we add nodes that would confuse 4922 // tightly_coupled_allocation() 4923 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4924 4925 int saved_reexecute_sp = -1; 4926 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp); 4927 // See arraycopy_restore_alloc_state() comment 4928 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards 4929 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation 4930 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards 4931 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL); 4932 4933 // The following tests must be performed 4934 // (1) src and dest are arrays. 4935 // (2) src and dest arrays must have elements of the same BasicType 4936 // (3) src and dest must not be null. 4937 // (4) src_offset must not be negative. 4938 // (5) dest_offset must not be negative. 4939 // (6) length must not be negative. 4940 // (7) src_offset + length must not exceed length of src. 4941 // (8) dest_offset + length must not exceed length of dest. 4942 // (9) each element of an oop array must be assignable 4943 4944 // (3) src and dest must not be null. 4945 // always do this here because we need the JVM state for uncommon traps 4946 Node* null_ctl = top(); 4947 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY); 4948 assert(null_ctl->is_top(), "no null control here"); 4949 dest = null_check(dest, T_ARRAY); 4950 4951 if (!can_emit_guards) { 4952 // if saved_jvms == NULL and alloc != NULL, we don't emit any 4953 // guards but the arraycopy node could still take advantage of a 4954 // tightly allocated allocation. tightly_coupled_allocation() is 4955 // called again to make sure it takes the null check above into 4956 // account: the null check is mandatory and if it caused an 4957 // uncommon trap to be emitted then the allocation can't be 4958 // considered tightly coupled in this context. 4959 alloc = tightly_coupled_allocation(dest, NULL); 4960 } 4961 4962 bool validated = false; 4963 4964 const Type* src_type = _gvn.type(src); 4965 const Type* dest_type = _gvn.type(dest); 4966 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4967 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4968 4969 // Do we have the type of src? 4970 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4971 // Do we have the type of dest? 4972 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4973 // Is the type for src from speculation? 4974 bool src_spec = false; 4975 // Is the type for dest from speculation? 4976 bool dest_spec = false; 4977 4978 if ((!has_src || !has_dest) && can_emit_guards) { 4979 // We don't have sufficient type information, let's see if 4980 // speculative types can help. We need to have types for both src 4981 // and dest so that it pays off. 4982 4983 // Do we already have or could we have type information for src 4984 bool could_have_src = has_src; 4985 // Do we already have or could we have type information for dest 4986 bool could_have_dest = has_dest; 4987 4988 ciKlass* src_k = NULL; 4989 if (!has_src) { 4990 src_k = src_type->speculative_type_not_null(); 4991 if (src_k != NULL && src_k->is_array_klass()) { 4992 could_have_src = true; 4993 } 4994 } 4995 4996 ciKlass* dest_k = NULL; 4997 if (!has_dest) { 4998 dest_k = dest_type->speculative_type_not_null(); 4999 if (dest_k != NULL && dest_k->is_array_klass()) { 5000 could_have_dest = true; 5001 } 5002 } 5003 5004 if (could_have_src && could_have_dest) { 5005 // This is going to pay off so emit the required guards 5006 if (!has_src) { 5007 src = maybe_cast_profiled_obj(src, src_k, true); 5008 src_type = _gvn.type(src); 5009 top_src = src_type->isa_aryptr(); 5010 has_src = (top_src != NULL && top_src->klass() != NULL); 5011 src_spec = true; 5012 } 5013 if (!has_dest) { 5014 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5015 dest_type = _gvn.type(dest); 5016 top_dest = dest_type->isa_aryptr(); 5017 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 5018 dest_spec = true; 5019 } 5020 } 5021 } 5022 5023 if (has_src && has_dest && can_emit_guards) { 5024 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 5025 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 5026 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 5027 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 5028 5029 if (src_elem == dest_elem && src_elem == T_OBJECT) { 5030 // If both arrays are object arrays then having the exact types 5031 // for both will remove the need for a subtype check at runtime 5032 // before the call and may make it possible to pick a faster copy 5033 // routine (without a subtype check on every element) 5034 // Do we have the exact type of src? 5035 bool could_have_src = src_spec; 5036 // Do we have the exact type of dest? 5037 bool could_have_dest = dest_spec; 5038 ciKlass* src_k = top_src->klass(); 5039 ciKlass* dest_k = top_dest->klass(); 5040 if (!src_spec) { 5041 src_k = src_type->speculative_type_not_null(); 5042 if (src_k != NULL && src_k->is_array_klass()) { 5043 could_have_src = true; 5044 } 5045 } 5046 if (!dest_spec) { 5047 dest_k = dest_type->speculative_type_not_null(); 5048 if (dest_k != NULL && dest_k->is_array_klass()) { 5049 could_have_dest = true; 5050 } 5051 } 5052 if (could_have_src && could_have_dest) { 5053 // If we can have both exact types, emit the missing guards 5054 if (could_have_src && !src_spec) { 5055 src = maybe_cast_profiled_obj(src, src_k, true); 5056 } 5057 if (could_have_dest && !dest_spec) { 5058 dest = maybe_cast_profiled_obj(dest, dest_k, true); 5059 } 5060 } 5061 } 5062 } 5063 5064 ciMethod* trap_method = method(); 5065 int trap_bci = bci(); 5066 if (saved_jvms != NULL) { 5067 trap_method = alloc->jvms()->method(); 5068 trap_bci = alloc->jvms()->bci(); 5069 } 5070 5071 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) && 5072 can_emit_guards && 5073 !src->is_top() && !dest->is_top()) { 5074 // validate arguments: enables transformation the ArrayCopyNode 5075 validated = true; 5076 5077 RegionNode* slow_region = new RegionNode(1); 5078 record_for_igvn(slow_region); 5079 5080 // (1) src and dest are arrays. 5081 generate_non_array_guard(load_object_klass(src), slow_region); 5082 generate_non_array_guard(load_object_klass(dest), slow_region); 5083 5084 // (2) src and dest arrays must have elements of the same BasicType 5085 // done at macro expansion or at Ideal transformation time 5086 5087 // (4) src_offset must not be negative. 5088 generate_negative_guard(src_offset, slow_region); 5089 5090 // (5) dest_offset must not be negative. 5091 generate_negative_guard(dest_offset, slow_region); 5092 5093 // (7) src_offset + length must not exceed length of src. 5094 generate_limit_guard(src_offset, length, 5095 load_array_length(src), 5096 slow_region); 5097 5098 // (8) dest_offset + length must not exceed length of dest. 5099 generate_limit_guard(dest_offset, length, 5100 load_array_length(dest), 5101 slow_region); 5102 5103 // (9) each element of an oop array must be assignable 5104 Node* src_klass = load_object_klass(src); 5105 Node* dest_klass = load_object_klass(dest); 5106 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 5107 5108 if (not_subtype_ctrl != top()) { 5109 PreserveJVMState pjvms(this); 5110 set_control(not_subtype_ctrl); 5111 uncommon_trap(Deoptimization::Reason_intrinsic, 5112 Deoptimization::Action_make_not_entrant); 5113 assert(stopped(), "Should be stopped"); 5114 } 5115 { 5116 PreserveJVMState pjvms(this); 5117 set_control(_gvn.transform(slow_region)); 5118 uncommon_trap(Deoptimization::Reason_intrinsic, 5119 Deoptimization::Action_make_not_entrant); 5120 assert(stopped(), "Should be stopped"); 5121 } 5122 } 5123 5124 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp); 5125 5126 if (stopped()) { 5127 return true; 5128 } 5129 5130 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 5131 // Create LoadRange and LoadKlass nodes for use during macro expansion here 5132 // so the compiler has a chance to eliminate them: during macro expansion, 5133 // we have to set their control (CastPP nodes are eliminated). 5134 load_object_klass(src), load_object_klass(dest), 5135 load_array_length(src), load_array_length(dest)); 5136 5137 ac->set_arraycopy(validated); 5138 5139 Node* n = _gvn.transform(ac); 5140 if (n == ac) { 5141 ac->connect_outputs(this); 5142 } else { 5143 assert(validated, "shouldn't transform if all arguments not validated"); 5144 set_all_memory(n); 5145 } 5146 5147 return true; 5148 } 5149 5150 5151 // Helper function which determines if an arraycopy immediately follows 5152 // an allocation, with no intervening tests or other escapes for the object. 5153 AllocateArrayNode* 5154 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 5155 RegionNode* slow_region) { 5156 if (stopped()) return NULL; // no fast path 5157 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 5158 5159 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 5160 if (alloc == NULL) return NULL; 5161 5162 Node* rawmem = memory(Compile::AliasIdxRaw); 5163 // Is the allocation's memory state untouched? 5164 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 5165 // Bail out if there have been raw-memory effects since the allocation. 5166 // (Example: There might have been a call or safepoint.) 5167 return NULL; 5168 } 5169 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 5170 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 5171 return NULL; 5172 } 5173 5174 // There must be no unexpected observers of this allocation. 5175 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 5176 Node* obs = ptr->fast_out(i); 5177 if (obs != this->map()) { 5178 return NULL; 5179 } 5180 } 5181 5182 // This arraycopy must unconditionally follow the allocation of the ptr. 5183 Node* alloc_ctl = ptr->in(0); 5184 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 5185 5186 Node* ctl = control(); 5187 while (ctl != alloc_ctl) { 5188 // There may be guards which feed into the slow_region. 5189 // Any other control flow means that we might not get a chance 5190 // to finish initializing the allocated object. 5191 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 5192 IfNode* iff = ctl->in(0)->as_If(); 5193 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 5194 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 5195 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 5196 ctl = iff->in(0); // This test feeds the known slow_region. 5197 continue; 5198 } 5199 // One more try: Various low-level checks bottom out in 5200 // uncommon traps. If the debug-info of the trap omits 5201 // any reference to the allocation, as we've already 5202 // observed, then there can be no objection to the trap. 5203 bool found_trap = false; 5204 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 5205 Node* obs = not_ctl->fast_out(j); 5206 if (obs->in(0) == not_ctl && obs->is_Call() && 5207 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 5208 found_trap = true; break; 5209 } 5210 } 5211 if (found_trap) { 5212 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 5213 continue; 5214 } 5215 } 5216 return NULL; 5217 } 5218 5219 // If we get this far, we have an allocation which immediately 5220 // precedes the arraycopy, and we can take over zeroing the new object. 5221 // The arraycopy will finish the initialization, and provide 5222 // a new control state to which we will anchor the destination pointer. 5223 5224 return alloc; 5225 } 5226 5227 //-------------inline_encodeISOArray----------------------------------- 5228 // encode char[] to byte[] in ISO_8859_1 5229 bool LibraryCallKit::inline_encodeISOArray() { 5230 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 5231 // no receiver since it is static method 5232 Node *src = argument(0); 5233 Node *src_offset = argument(1); 5234 Node *dst = argument(2); 5235 Node *dst_offset = argument(3); 5236 Node *length = argument(4); 5237 5238 const Type* src_type = src->Value(&_gvn); 5239 const Type* dst_type = dst->Value(&_gvn); 5240 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5241 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 5242 if (top_src == NULL || top_src->klass() == NULL || 5243 top_dest == NULL || top_dest->klass() == NULL) { 5244 // failed array check 5245 return false; 5246 } 5247 5248 // Figure out the size and type of the elements we will be copying. 5249 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5250 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5251 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 5252 return false; 5253 } 5254 Node* src_start = array_element_address(src, src_offset, src_elem); 5255 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 5256 // 'src_start' points to src array + scaled offset 5257 // 'dst_start' points to dst array + scaled offset 5258 5259 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 5260 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 5261 enc = _gvn.transform(enc); 5262 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 5263 set_memory(res_mem, mtype); 5264 set_result(enc); 5265 return true; 5266 } 5267 5268 //-------------inline_multiplyToLen----------------------------------- 5269 bool LibraryCallKit::inline_multiplyToLen() { 5270 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 5271 5272 address stubAddr = StubRoutines::multiplyToLen(); 5273 if (stubAddr == NULL) { 5274 return false; // Intrinsic's stub is not implemented on this platform 5275 } 5276 const char* stubName = "multiplyToLen"; 5277 5278 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 5279 5280 Node* x = argument(1); 5281 Node* xlen = argument(2); 5282 Node* y = argument(3); 5283 Node* ylen = argument(4); 5284 Node* z = argument(5); 5285 5286 const Type* x_type = x->Value(&_gvn); 5287 const Type* y_type = y->Value(&_gvn); 5288 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5289 const TypeAryPtr* top_y = y_type->isa_aryptr(); 5290 if (top_x == NULL || top_x->klass() == NULL || 5291 top_y == NULL || top_y->klass() == NULL) { 5292 // failed array check 5293 return false; 5294 } 5295 5296 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5297 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5298 if (x_elem != T_INT || y_elem != T_INT) { 5299 return false; 5300 } 5301 5302 // Set the original stack and the reexecute bit for the interpreter to reexecute 5303 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 5304 // on the return from z array allocation in runtime. 5305 { PreserveReexecuteState preexecs(this); 5306 jvms()->set_should_reexecute(true); 5307 5308 Node* x_start = array_element_address(x, intcon(0), x_elem); 5309 Node* y_start = array_element_address(y, intcon(0), y_elem); 5310 // 'x_start' points to x array + scaled xlen 5311 // 'y_start' points to y array + scaled ylen 5312 5313 // Allocate the result array 5314 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 5315 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 5316 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 5317 5318 IdealKit ideal(this); 5319 5320 #define __ ideal. 5321 Node* one = __ ConI(1); 5322 Node* zero = __ ConI(0); 5323 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 5324 __ set(need_alloc, zero); 5325 __ set(z_alloc, z); 5326 __ if_then(z, BoolTest::eq, null()); { 5327 __ increment (need_alloc, one); 5328 } __ else_(); { 5329 // Update graphKit memory and control from IdealKit. 5330 sync_kit(ideal); 5331 Node* zlen_arg = load_array_length(z); 5332 // Update IdealKit memory and control from graphKit. 5333 __ sync_kit(this); 5334 __ if_then(zlen_arg, BoolTest::lt, zlen); { 5335 __ increment (need_alloc, one); 5336 } __ end_if(); 5337 } __ end_if(); 5338 5339 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5340 // Update graphKit memory and control from IdealKit. 5341 sync_kit(ideal); 5342 Node * narr = new_array(klass_node, zlen, 1); 5343 // Update IdealKit memory and control from graphKit. 5344 __ sync_kit(this); 5345 __ set(z_alloc, narr); 5346 } __ end_if(); 5347 5348 sync_kit(ideal); 5349 z = __ value(z_alloc); 5350 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5351 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5352 // Final sync IdealKit and GraphKit. 5353 final_sync(ideal); 5354 #undef __ 5355 5356 Node* z_start = array_element_address(z, intcon(0), T_INT); 5357 5358 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5359 OptoRuntime::multiplyToLen_Type(), 5360 stubAddr, stubName, TypePtr::BOTTOM, 5361 x_start, xlen, y_start, ylen, z_start, zlen); 5362 } // original reexecute is set back here 5363 5364 C->set_has_split_ifs(true); // Has chance for split-if optimization 5365 set_result(z); 5366 return true; 5367 } 5368 5369 //-------------inline_squareToLen------------------------------------ 5370 bool LibraryCallKit::inline_squareToLen() { 5371 assert(UseSquareToLenIntrinsic, "not implementated on this platform"); 5372 5373 address stubAddr = StubRoutines::squareToLen(); 5374 if (stubAddr == NULL) { 5375 return false; // Intrinsic's stub is not implemented on this platform 5376 } 5377 const char* stubName = "squareToLen"; 5378 5379 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters"); 5380 5381 Node* x = argument(0); 5382 Node* len = argument(1); 5383 Node* z = argument(2); 5384 Node* zlen = argument(3); 5385 5386 const Type* x_type = x->Value(&_gvn); 5387 const Type* z_type = z->Value(&_gvn); 5388 const TypeAryPtr* top_x = x_type->isa_aryptr(); 5389 const TypeAryPtr* top_z = z_type->isa_aryptr(); 5390 if (top_x == NULL || top_x->klass() == NULL || 5391 top_z == NULL || top_z->klass() == NULL) { 5392 // failed array check 5393 return false; 5394 } 5395 5396 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5397 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5398 if (x_elem != T_INT || z_elem != T_INT) { 5399 return false; 5400 } 5401 5402 5403 Node* x_start = array_element_address(x, intcon(0), x_elem); 5404 Node* z_start = array_element_address(z, intcon(0), z_elem); 5405 5406 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5407 OptoRuntime::squareToLen_Type(), 5408 stubAddr, stubName, TypePtr::BOTTOM, 5409 x_start, len, z_start, zlen); 5410 5411 set_result(z); 5412 return true; 5413 } 5414 5415 //-------------inline_mulAdd------------------------------------------ 5416 bool LibraryCallKit::inline_mulAdd() { 5417 assert(UseMulAddIntrinsic, "not implementated on this platform"); 5418 5419 address stubAddr = StubRoutines::mulAdd(); 5420 if (stubAddr == NULL) { 5421 return false; // Intrinsic's stub is not implemented on this platform 5422 } 5423 const char* stubName = "mulAdd"; 5424 5425 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters"); 5426 5427 Node* out = argument(0); 5428 Node* in = argument(1); 5429 Node* offset = argument(2); 5430 Node* len = argument(3); 5431 Node* k = argument(4); 5432 5433 const Type* out_type = out->Value(&_gvn); 5434 const Type* in_type = in->Value(&_gvn); 5435 const TypeAryPtr* top_out = out_type->isa_aryptr(); 5436 const TypeAryPtr* top_in = in_type->isa_aryptr(); 5437 if (top_out == NULL || top_out->klass() == NULL || 5438 top_in == NULL || top_in->klass() == NULL) { 5439 // failed array check 5440 return false; 5441 } 5442 5443 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5444 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5445 if (out_elem != T_INT || in_elem != T_INT) { 5446 return false; 5447 } 5448 5449 Node* outlen = load_array_length(out); 5450 Node* new_offset = _gvn.transform(new SubINode(outlen, offset)); 5451 Node* out_start = array_element_address(out, intcon(0), out_elem); 5452 Node* in_start = array_element_address(in, intcon(0), in_elem); 5453 5454 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5455 OptoRuntime::mulAdd_Type(), 5456 stubAddr, stubName, TypePtr::BOTTOM, 5457 out_start,in_start, new_offset, len, k); 5458 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5459 set_result(result); 5460 return true; 5461 } 5462 5463 5464 /** 5465 * Calculate CRC32 for byte. 5466 * int java.util.zip.CRC32.update(int crc, int b) 5467 */ 5468 bool LibraryCallKit::inline_updateCRC32() { 5469 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5470 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5471 // no receiver since it is static method 5472 Node* crc = argument(0); // type: int 5473 Node* b = argument(1); // type: int 5474 5475 /* 5476 * int c = ~ crc; 5477 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5478 * b = b ^ (c >>> 8); 5479 * crc = ~b; 5480 */ 5481 5482 Node* M1 = intcon(-1); 5483 crc = _gvn.transform(new XorINode(crc, M1)); 5484 Node* result = _gvn.transform(new XorINode(crc, b)); 5485 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5486 5487 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5488 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5489 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5490 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5491 5492 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5493 result = _gvn.transform(new XorINode(crc, result)); 5494 result = _gvn.transform(new XorINode(result, M1)); 5495 set_result(result); 5496 return true; 5497 } 5498 5499 /** 5500 * Calculate CRC32 for byte[] array. 5501 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5502 */ 5503 bool LibraryCallKit::inline_updateBytesCRC32() { 5504 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5505 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5506 // no receiver since it is static method 5507 Node* crc = argument(0); // type: int 5508 Node* src = argument(1); // type: oop 5509 Node* offset = argument(2); // type: int 5510 Node* length = argument(3); // type: int 5511 5512 const Type* src_type = src->Value(&_gvn); 5513 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5514 if (top_src == NULL || top_src->klass() == NULL) { 5515 // failed array check 5516 return false; 5517 } 5518 5519 // Figure out the size and type of the elements we will be copying. 5520 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5521 if (src_elem != T_BYTE) { 5522 return false; 5523 } 5524 5525 // 'src_start' points to src array + scaled offset 5526 Node* src_start = array_element_address(src, offset, src_elem); 5527 5528 // We assume that range check is done by caller. 5529 // TODO: generate range check (offset+length < src.length) in debug VM. 5530 5531 // Call the stub. 5532 address stubAddr = StubRoutines::updateBytesCRC32(); 5533 const char *stubName = "updateBytesCRC32"; 5534 5535 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5536 stubAddr, stubName, TypePtr::BOTTOM, 5537 crc, src_start, length); 5538 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5539 set_result(result); 5540 return true; 5541 } 5542 5543 /** 5544 * Calculate CRC32 for ByteBuffer. 5545 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5546 */ 5547 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5548 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5549 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5550 // no receiver since it is static method 5551 Node* crc = argument(0); // type: int 5552 Node* src = argument(1); // type: long 5553 Node* offset = argument(3); // type: int 5554 Node* length = argument(4); // type: int 5555 5556 src = ConvL2X(src); // adjust Java long to machine word 5557 Node* base = _gvn.transform(new CastX2PNode(src)); 5558 offset = ConvI2X(offset); 5559 5560 // 'src_start' points to src array + scaled offset 5561 Node* src_start = basic_plus_adr(top(), base, offset); 5562 5563 // Call the stub. 5564 address stubAddr = StubRoutines::updateBytesCRC32(); 5565 const char *stubName = "updateBytesCRC32"; 5566 5567 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5568 stubAddr, stubName, TypePtr::BOTTOM, 5569 crc, src_start, length); 5570 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5571 set_result(result); 5572 return true; 5573 } 5574 5575 //------------------------------get_table_from_crc32c_class----------------------- 5576 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) { 5577 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class); 5578 assert (table != NULL, "wrong version of java.util.zip.CRC32C"); 5579 5580 return table; 5581 } 5582 5583 //------------------------------inline_updateBytesCRC32C----------------------- 5584 // 5585 // Calculate CRC32C for byte[] array. 5586 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end) 5587 // 5588 bool LibraryCallKit::inline_updateBytesCRC32C() { 5589 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5590 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5591 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5592 // no receiver since it is a static method 5593 Node* crc = argument(0); // type: int 5594 Node* src = argument(1); // type: oop 5595 Node* offset = argument(2); // type: int 5596 Node* end = argument(3); // type: int 5597 5598 Node* length = _gvn.transform(new SubINode(end, offset)); 5599 5600 const Type* src_type = src->Value(&_gvn); 5601 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5602 if (top_src == NULL || top_src->klass() == NULL) { 5603 // failed array check 5604 return false; 5605 } 5606 5607 // Figure out the size and type of the elements we will be copying. 5608 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5609 if (src_elem != T_BYTE) { 5610 return false; 5611 } 5612 5613 // 'src_start' points to src array + scaled offset 5614 Node* src_start = array_element_address(src, offset, src_elem); 5615 5616 // static final int[] byteTable in class CRC32C 5617 Node* table = get_table_from_crc32c_class(callee()->holder()); 5618 Node* table_start = array_element_address(table, intcon(0), T_INT); 5619 5620 // We assume that range check is done by caller. 5621 // TODO: generate range check (offset+length < src.length) in debug VM. 5622 5623 // Call the stub. 5624 address stubAddr = StubRoutines::updateBytesCRC32C(); 5625 const char *stubName = "updateBytesCRC32C"; 5626 5627 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5628 stubAddr, stubName, TypePtr::BOTTOM, 5629 crc, src_start, length, table_start); 5630 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5631 set_result(result); 5632 return true; 5633 } 5634 5635 //------------------------------inline_updateDirectByteBufferCRC32C----------------------- 5636 // 5637 // Calculate CRC32C for DirectByteBuffer. 5638 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end) 5639 // 5640 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() { 5641 assert(UseCRC32CIntrinsics, "need CRC32C instruction support"); 5642 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long"); 5643 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded"); 5644 // no receiver since it is a static method 5645 Node* crc = argument(0); // type: int 5646 Node* src = argument(1); // type: long 5647 Node* offset = argument(3); // type: int 5648 Node* end = argument(4); // type: int 5649 5650 Node* length = _gvn.transform(new SubINode(end, offset)); 5651 5652 src = ConvL2X(src); // adjust Java long to machine word 5653 Node* base = _gvn.transform(new CastX2PNode(src)); 5654 offset = ConvI2X(offset); 5655 5656 // 'src_start' points to src array + scaled offset 5657 Node* src_start = basic_plus_adr(top(), base, offset); 5658 5659 // static final int[] byteTable in class CRC32C 5660 Node* table = get_table_from_crc32c_class(callee()->holder()); 5661 Node* table_start = array_element_address(table, intcon(0), T_INT); 5662 5663 // Call the stub. 5664 address stubAddr = StubRoutines::updateBytesCRC32C(); 5665 const char *stubName = "updateBytesCRC32C"; 5666 5667 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(), 5668 stubAddr, stubName, TypePtr::BOTTOM, 5669 crc, src_start, length, table_start); 5670 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5671 set_result(result); 5672 return true; 5673 } 5674 5675 //----------------------------inline_reference_get---------------------------- 5676 // public T java.lang.ref.Reference.get(); 5677 bool LibraryCallKit::inline_reference_get() { 5678 const int referent_offset = java_lang_ref_Reference::referent_offset; 5679 guarantee(referent_offset > 0, "should have already been set"); 5680 5681 // Get the argument: 5682 Node* reference_obj = null_check_receiver(); 5683 if (stopped()) return true; 5684 5685 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5686 5687 ciInstanceKlass* klass = env()->Object_klass(); 5688 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5689 5690 Node* no_ctrl = NULL; 5691 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5692 5693 // Use the pre-barrier to record the value in the referent field 5694 pre_barrier(false /* do_load */, 5695 control(), 5696 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5697 result /* pre_val */, 5698 T_OBJECT); 5699 5700 // Add memory barrier to prevent commoning reads from this field 5701 // across safepoint since GC can change its value. 5702 insert_mem_bar(Op_MemBarCPUOrder); 5703 5704 set_result(result); 5705 return true; 5706 } 5707 5708 5709 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5710 bool is_exact=true, bool is_static=false, 5711 ciInstanceKlass * fromKls=NULL) { 5712 if (fromKls == NULL) { 5713 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5714 assert(tinst != NULL, "obj is null"); 5715 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5716 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5717 fromKls = tinst->klass()->as_instance_klass(); 5718 } else { 5719 assert(is_static, "only for static field access"); 5720 } 5721 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName), 5722 ciSymbol::make(fieldTypeString), 5723 is_static); 5724 5725 assert (field != NULL, "undefined field"); 5726 if (field == NULL) return (Node *) NULL; 5727 5728 if (is_static) { 5729 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror()); 5730 fromObj = makecon(tip); 5731 } 5732 5733 // Next code copied from Parse::do_get_xxx(): 5734 5735 // Compute address and memory type. 5736 int offset = field->offset_in_bytes(); 5737 bool is_vol = field->is_volatile(); 5738 ciType* field_klass = field->type(); 5739 assert(field_klass->is_loaded(), "should be loaded"); 5740 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5741 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5742 BasicType bt = field->layout_type(); 5743 5744 // Build the resultant type of the load 5745 const Type *type; 5746 if (bt == T_OBJECT) { 5747 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5748 } else { 5749 type = Type::get_const_basic_type(bt); 5750 } 5751 5752 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5753 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5754 } 5755 // Build the load. 5756 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5757 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol); 5758 // If reference is volatile, prevent following memory ops from 5759 // floating up past the volatile read. Also prevents commoning 5760 // another volatile read. 5761 if (is_vol) { 5762 // Memory barrier includes bogus read of value to force load BEFORE membar 5763 insert_mem_bar(Op_MemBarAcquire, loadedField); 5764 } 5765 return loadedField; 5766 } 5767 5768 5769 //------------------------------inline_aescrypt_Block----------------------- 5770 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5771 address stubAddr; 5772 const char *stubName; 5773 assert(UseAES, "need AES instruction support"); 5774 5775 switch(id) { 5776 case vmIntrinsics::_aescrypt_encryptBlock: 5777 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5778 stubName = "aescrypt_encryptBlock"; 5779 break; 5780 case vmIntrinsics::_aescrypt_decryptBlock: 5781 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5782 stubName = "aescrypt_decryptBlock"; 5783 break; 5784 } 5785 if (stubAddr == NULL) return false; 5786 5787 Node* aescrypt_object = argument(0); 5788 Node* src = argument(1); 5789 Node* src_offset = argument(2); 5790 Node* dest = argument(3); 5791 Node* dest_offset = argument(4); 5792 5793 // (1) src and dest are arrays. 5794 const Type* src_type = src->Value(&_gvn); 5795 const Type* dest_type = dest->Value(&_gvn); 5796 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5797 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5798 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5799 5800 // for the quick and dirty code we will skip all the checks. 5801 // we are just trying to get the call to be generated. 5802 Node* src_start = src; 5803 Node* dest_start = dest; 5804 if (src_offset != NULL || dest_offset != NULL) { 5805 assert(src_offset != NULL && dest_offset != NULL, ""); 5806 src_start = array_element_address(src, src_offset, T_BYTE); 5807 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5808 } 5809 5810 // now need to get the start of its expanded key array 5811 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5812 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5813 if (k_start == NULL) return false; 5814 5815 if (Matcher::pass_original_key_for_aes()) { 5816 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5817 // compatibility issues between Java key expansion and SPARC crypto instructions 5818 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5819 if (original_k_start == NULL) return false; 5820 5821 // Call the stub. 5822 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5823 stubAddr, stubName, TypePtr::BOTTOM, 5824 src_start, dest_start, k_start, original_k_start); 5825 } else { 5826 // Call the stub. 5827 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5828 stubAddr, stubName, TypePtr::BOTTOM, 5829 src_start, dest_start, k_start); 5830 } 5831 5832 return true; 5833 } 5834 5835 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5836 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5837 address stubAddr; 5838 const char *stubName; 5839 5840 assert(UseAES, "need AES instruction support"); 5841 5842 switch(id) { 5843 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5844 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5845 stubName = "cipherBlockChaining_encryptAESCrypt"; 5846 break; 5847 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5848 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5849 stubName = "cipherBlockChaining_decryptAESCrypt"; 5850 break; 5851 } 5852 if (stubAddr == NULL) return false; 5853 5854 Node* cipherBlockChaining_object = argument(0); 5855 Node* src = argument(1); 5856 Node* src_offset = argument(2); 5857 Node* len = argument(3); 5858 Node* dest = argument(4); 5859 Node* dest_offset = argument(5); 5860 5861 // (1) src and dest are arrays. 5862 const Type* src_type = src->Value(&_gvn); 5863 const Type* dest_type = dest->Value(&_gvn); 5864 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5865 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5866 assert (top_src != NULL && top_src->klass() != NULL 5867 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5868 5869 // checks are the responsibility of the caller 5870 Node* src_start = src; 5871 Node* dest_start = dest; 5872 if (src_offset != NULL || dest_offset != NULL) { 5873 assert(src_offset != NULL && dest_offset != NULL, ""); 5874 src_start = array_element_address(src, src_offset, T_BYTE); 5875 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5876 } 5877 5878 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5879 // (because of the predicated logic executed earlier). 5880 // so we cast it here safely. 5881 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5882 5883 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5884 if (embeddedCipherObj == NULL) return false; 5885 5886 // cast it to what we know it will be at runtime 5887 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5888 assert(tinst != NULL, "CBC obj is null"); 5889 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5890 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5891 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5892 5893 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5894 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5895 const TypeOopPtr* xtype = aklass->as_instance_type(); 5896 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5897 aescrypt_object = _gvn.transform(aescrypt_object); 5898 5899 // we need to get the start of the aescrypt_object's expanded key array 5900 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5901 if (k_start == NULL) return false; 5902 5903 // similarly, get the start address of the r vector 5904 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5905 if (objRvec == NULL) return false; 5906 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5907 5908 Node* cbcCrypt; 5909 if (Matcher::pass_original_key_for_aes()) { 5910 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5911 // compatibility issues between Java key expansion and SPARC crypto instructions 5912 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5913 if (original_k_start == NULL) return false; 5914 5915 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5916 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5917 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5918 stubAddr, stubName, TypePtr::BOTTOM, 5919 src_start, dest_start, k_start, r_start, len, original_k_start); 5920 } else { 5921 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5922 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5923 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5924 stubAddr, stubName, TypePtr::BOTTOM, 5925 src_start, dest_start, k_start, r_start, len); 5926 } 5927 5928 // return cipher length (int) 5929 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5930 set_result(retvalue); 5931 return true; 5932 } 5933 5934 //------------------------------get_key_start_from_aescrypt_object----------------------- 5935 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5936 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5937 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5938 if (objAESCryptKey == NULL) return (Node *) NULL; 5939 5940 // now have the array, need to get the start address of the K array 5941 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5942 return k_start; 5943 } 5944 5945 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5946 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5947 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5948 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5949 if (objAESCryptKey == NULL) return (Node *) NULL; 5950 5951 // now have the array, need to get the start address of the lastKey array 5952 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5953 return original_k_start; 5954 } 5955 5956 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5957 // Return node representing slow path of predicate check. 5958 // the pseudo code we want to emulate with this predicate is: 5959 // for encryption: 5960 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5961 // for decryption: 5962 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5963 // note cipher==plain is more conservative than the original java code but that's OK 5964 // 5965 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5966 // The receiver was checked for NULL already. 5967 Node* objCBC = argument(0); 5968 5969 // Load embeddedCipher field of CipherBlockChaining object. 5970 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5971 5972 // get AESCrypt klass for instanceOf check 5973 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5974 // will have same classloader as CipherBlockChaining object 5975 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5976 assert(tinst != NULL, "CBCobj is null"); 5977 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5978 5979 // we want to do an instanceof comparison against the AESCrypt class 5980 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5981 if (!klass_AESCrypt->is_loaded()) { 5982 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5983 Node* ctrl = control(); 5984 set_control(top()); // no regular fast path 5985 return ctrl; 5986 } 5987 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5988 5989 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5990 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5991 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5992 5993 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5994 5995 // for encryption, we are done 5996 if (!decrypting) 5997 return instof_false; // even if it is NULL 5998 5999 // for decryption, we need to add a further check to avoid 6000 // taking the intrinsic path when cipher and plain are the same 6001 // see the original java code for why. 6002 RegionNode* region = new RegionNode(3); 6003 region->init_req(1, instof_false); 6004 Node* src = argument(1); 6005 Node* dest = argument(4); 6006 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 6007 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 6008 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 6009 region->init_req(2, src_dest_conjoint); 6010 6011 record_for_igvn(region); 6012 return _gvn.transform(region); 6013 } 6014 6015 //------------------------------inline_ghash_processBlocks 6016 bool LibraryCallKit::inline_ghash_processBlocks() { 6017 address stubAddr; 6018 const char *stubName; 6019 assert(UseGHASHIntrinsics, "need GHASH intrinsics support"); 6020 6021 stubAddr = StubRoutines::ghash_processBlocks(); 6022 stubName = "ghash_processBlocks"; 6023 6024 Node* data = argument(0); 6025 Node* offset = argument(1); 6026 Node* len = argument(2); 6027 Node* state = argument(3); 6028 Node* subkeyH = argument(4); 6029 6030 Node* state_start = array_element_address(state, intcon(0), T_LONG); 6031 assert(state_start, "state is NULL"); 6032 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG); 6033 assert(subkeyH_start, "subkeyH is NULL"); 6034 Node* data_start = array_element_address(data, offset, T_BYTE); 6035 assert(data_start, "data is NULL"); 6036 6037 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP, 6038 OptoRuntime::ghash_processBlocks_Type(), 6039 stubAddr, stubName, TypePtr::BOTTOM, 6040 state_start, subkeyH_start, data_start, len); 6041 return true; 6042 } 6043 6044 //------------------------------inline_sha_implCompress----------------------- 6045 // 6046 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 6047 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 6048 // 6049 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 6050 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 6051 // 6052 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 6053 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 6054 // 6055 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 6056 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 6057 6058 Node* sha_obj = argument(0); 6059 Node* src = argument(1); // type oop 6060 Node* ofs = argument(2); // type int 6061 6062 const Type* src_type = src->Value(&_gvn); 6063 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6064 if (top_src == NULL || top_src->klass() == NULL) { 6065 // failed array check 6066 return false; 6067 } 6068 // Figure out the size and type of the elements we will be copying. 6069 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6070 if (src_elem != T_BYTE) { 6071 return false; 6072 } 6073 // 'src_start' points to src array + offset 6074 Node* src_start = array_element_address(src, ofs, src_elem); 6075 Node* state = NULL; 6076 address stubAddr; 6077 const char *stubName; 6078 6079 switch(id) { 6080 case vmIntrinsics::_sha_implCompress: 6081 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 6082 state = get_state_from_sha_object(sha_obj); 6083 stubAddr = StubRoutines::sha1_implCompress(); 6084 stubName = "sha1_implCompress"; 6085 break; 6086 case vmIntrinsics::_sha2_implCompress: 6087 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 6088 state = get_state_from_sha_object(sha_obj); 6089 stubAddr = StubRoutines::sha256_implCompress(); 6090 stubName = "sha256_implCompress"; 6091 break; 6092 case vmIntrinsics::_sha5_implCompress: 6093 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 6094 state = get_state_from_sha5_object(sha_obj); 6095 stubAddr = StubRoutines::sha512_implCompress(); 6096 stubName = "sha512_implCompress"; 6097 break; 6098 default: 6099 fatal_unexpected_iid(id); 6100 return false; 6101 } 6102 if (state == NULL) return false; 6103 6104 // Call the stub. 6105 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 6106 stubAddr, stubName, TypePtr::BOTTOM, 6107 src_start, state); 6108 6109 return true; 6110 } 6111 6112 //------------------------------inline_digestBase_implCompressMB----------------------- 6113 // 6114 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 6115 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 6116 // 6117 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 6118 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6119 "need SHA1/SHA256/SHA512 instruction support"); 6120 assert((uint)predicate < 3, "sanity"); 6121 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 6122 6123 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 6124 Node* src = argument(1); // byte[] array 6125 Node* ofs = argument(2); // type int 6126 Node* limit = argument(3); // type int 6127 6128 const Type* src_type = src->Value(&_gvn); 6129 const TypeAryPtr* top_src = src_type->isa_aryptr(); 6130 if (top_src == NULL || top_src->klass() == NULL) { 6131 // failed array check 6132 return false; 6133 } 6134 // Figure out the size and type of the elements we will be copying. 6135 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 6136 if (src_elem != T_BYTE) { 6137 return false; 6138 } 6139 // 'src_start' points to src array + offset 6140 Node* src_start = array_element_address(src, ofs, src_elem); 6141 6142 const char* klass_SHA_name = NULL; 6143 const char* stub_name = NULL; 6144 address stub_addr = NULL; 6145 bool long_state = false; 6146 6147 switch (predicate) { 6148 case 0: 6149 if (UseSHA1Intrinsics) { 6150 klass_SHA_name = "sun/security/provider/SHA"; 6151 stub_name = "sha1_implCompressMB"; 6152 stub_addr = StubRoutines::sha1_implCompressMB(); 6153 } 6154 break; 6155 case 1: 6156 if (UseSHA256Intrinsics) { 6157 klass_SHA_name = "sun/security/provider/SHA2"; 6158 stub_name = "sha256_implCompressMB"; 6159 stub_addr = StubRoutines::sha256_implCompressMB(); 6160 } 6161 break; 6162 case 2: 6163 if (UseSHA512Intrinsics) { 6164 klass_SHA_name = "sun/security/provider/SHA5"; 6165 stub_name = "sha512_implCompressMB"; 6166 stub_addr = StubRoutines::sha512_implCompressMB(); 6167 long_state = true; 6168 } 6169 break; 6170 default: 6171 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6172 } 6173 if (klass_SHA_name != NULL) { 6174 // get DigestBase klass to lookup for SHA klass 6175 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 6176 assert(tinst != NULL, "digestBase_obj is not instance???"); 6177 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6178 6179 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6180 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 6181 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6182 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 6183 } 6184 return false; 6185 } 6186 //------------------------------inline_sha_implCompressMB----------------------- 6187 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 6188 bool long_state, address stubAddr, const char *stubName, 6189 Node* src_start, Node* ofs, Node* limit) { 6190 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 6191 const TypeOopPtr* xtype = aklass->as_instance_type(); 6192 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 6193 sha_obj = _gvn.transform(sha_obj); 6194 6195 Node* state; 6196 if (long_state) { 6197 state = get_state_from_sha5_object(sha_obj); 6198 } else { 6199 state = get_state_from_sha_object(sha_obj); 6200 } 6201 if (state == NULL) return false; 6202 6203 // Call the stub. 6204 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 6205 OptoRuntime::digestBase_implCompressMB_Type(), 6206 stubAddr, stubName, TypePtr::BOTTOM, 6207 src_start, state, ofs, limit); 6208 // return ofs (int) 6209 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 6210 set_result(result); 6211 6212 return true; 6213 } 6214 6215 //------------------------------get_state_from_sha_object----------------------- 6216 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 6217 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 6218 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 6219 if (sha_state == NULL) return (Node *) NULL; 6220 6221 // now have the array, need to get the start address of the state array 6222 Node* state = array_element_address(sha_state, intcon(0), T_INT); 6223 return state; 6224 } 6225 6226 //------------------------------get_state_from_sha5_object----------------------- 6227 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 6228 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 6229 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 6230 if (sha_state == NULL) return (Node *) NULL; 6231 6232 // now have the array, need to get the start address of the state array 6233 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 6234 return state; 6235 } 6236 6237 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 6238 // Return node representing slow path of predicate check. 6239 // the pseudo code we want to emulate with this predicate is: 6240 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 6241 // 6242 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 6243 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 6244 "need SHA1/SHA256/SHA512 instruction support"); 6245 assert((uint)predicate < 3, "sanity"); 6246 6247 // The receiver was checked for NULL already. 6248 Node* digestBaseObj = argument(0); 6249 6250 // get DigestBase klass for instanceOf check 6251 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 6252 assert(tinst != NULL, "digestBaseObj is null"); 6253 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 6254 6255 const char* klass_SHA_name = NULL; 6256 switch (predicate) { 6257 case 0: 6258 if (UseSHA1Intrinsics) { 6259 // we want to do an instanceof comparison against the SHA class 6260 klass_SHA_name = "sun/security/provider/SHA"; 6261 } 6262 break; 6263 case 1: 6264 if (UseSHA256Intrinsics) { 6265 // we want to do an instanceof comparison against the SHA2 class 6266 klass_SHA_name = "sun/security/provider/SHA2"; 6267 } 6268 break; 6269 case 2: 6270 if (UseSHA512Intrinsics) { 6271 // we want to do an instanceof comparison against the SHA5 class 6272 klass_SHA_name = "sun/security/provider/SHA5"; 6273 } 6274 break; 6275 default: 6276 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 6277 } 6278 6279 ciKlass* klass_SHA = NULL; 6280 if (klass_SHA_name != NULL) { 6281 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 6282 } 6283 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 6284 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 6285 Node* ctrl = control(); 6286 set_control(top()); // no intrinsic path 6287 return ctrl; 6288 } 6289 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 6290 6291 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 6292 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 6293 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 6294 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 6295 6296 return instof_false; // even if it is NULL 6297 } 6298 6299 bool LibraryCallKit::inline_profileBoolean() { 6300 Node* counts = argument(1); 6301 const TypeAryPtr* ary = NULL; 6302 ciArray* aobj = NULL; 6303 if (counts->is_Con() 6304 && (ary = counts->bottom_type()->isa_aryptr()) != NULL 6305 && (aobj = ary->const_oop()->as_array()) != NULL 6306 && (aobj->length() == 2)) { 6307 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively. 6308 jint false_cnt = aobj->element_value(0).as_int(); 6309 jint true_cnt = aobj->element_value(1).as_int(); 6310 6311 if (C->log() != NULL) { 6312 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'", 6313 false_cnt, true_cnt); 6314 } 6315 6316 if (false_cnt + true_cnt == 0) { 6317 // According to profile, never executed. 6318 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6319 Deoptimization::Action_reinterpret); 6320 return true; 6321 } 6322 6323 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt) 6324 // is a number of each value occurrences. 6325 Node* result = argument(0); 6326 if (false_cnt == 0 || true_cnt == 0) { 6327 // According to profile, one value has been never seen. 6328 int expected_val = (false_cnt == 0) ? 1 : 0; 6329 6330 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val))); 6331 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 6332 6333 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN); 6334 Node* fast_path = _gvn.transform(new IfTrueNode(check)); 6335 Node* slow_path = _gvn.transform(new IfFalseNode(check)); 6336 6337 { // Slow path: uncommon trap for never seen value and then reexecute 6338 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows 6339 // the value has been seen at least once. 6340 PreserveJVMState pjvms(this); 6341 PreserveReexecuteState preexecs(this); 6342 jvms()->set_should_reexecute(true); 6343 6344 set_control(slow_path); 6345 set_i_o(i_o()); 6346 6347 uncommon_trap_exact(Deoptimization::Reason_intrinsic, 6348 Deoptimization::Action_reinterpret); 6349 } 6350 // The guard for never seen value enables sharpening of the result and 6351 // returning a constant. It allows to eliminate branches on the same value 6352 // later on. 6353 set_control(fast_path); 6354 result = intcon(expected_val); 6355 } 6356 // Stop profiling. 6357 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode. 6358 // By replacing method body with profile data (represented as ProfileBooleanNode 6359 // on IR level) we effectively disable profiling. 6360 // It enables full speed execution once optimized code is generated. 6361 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt)); 6362 C->record_for_igvn(profile); 6363 set_result(profile); 6364 return true; 6365 } else { 6366 // Continue profiling. 6367 // Profile data isn't available at the moment. So, execute method's bytecode version. 6368 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod 6369 // is compiled and counters aren't available since corresponding MethodHandle 6370 // isn't a compile-time constant. 6371 return false; 6372 } 6373 } 6374 6375 bool LibraryCallKit::inline_isCompileConstant() { 6376 Node* n = argument(0); 6377 set_result(n->is_Con() ? intcon(1) : intcon(0)); 6378 return true; 6379 }