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