1 /* 2 * Copyright (c) 1999, 2014, 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_aescrypt_Block(vmIntrinsics::ID id); 272 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id); 273 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting); 274 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object); 275 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object); 276 bool inline_sha_implCompress(vmIntrinsics::ID id); 277 bool inline_digestBase_implCompressMB(int predicate); 278 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA, 279 bool long_state, address stubAddr, const char *stubName, 280 Node* src_start, Node* ofs, Node* limit); 281 Node* get_state_from_sha_object(Node *sha_object); 282 Node* get_state_from_sha5_object(Node *sha_object); 283 Node* inline_digestBase_implCompressMB_predicate(int predicate); 284 bool inline_encodeISOArray(); 285 bool inline_updateCRC32(); 286 bool inline_updateBytesCRC32(); 287 bool inline_updateByteBufferCRC32(); 288 bool inline_multiplyToLen(); 289 }; 290 291 292 //---------------------------make_vm_intrinsic---------------------------- 293 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) { 294 vmIntrinsics::ID id = m->intrinsic_id(); 295 assert(id != vmIntrinsics::_none, "must be a VM intrinsic"); 296 297 ccstr disable_intr = NULL; 298 299 if ((DisableIntrinsic[0] != '\0' 300 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) || 301 (method_has_option_value("DisableIntrinsic", disable_intr) 302 && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) { 303 // disabled by a user request on the command line: 304 // example: -XX:DisableIntrinsic=_hashCode,_getClass 305 return NULL; 306 } 307 308 if (!m->is_loaded()) { 309 // do not attempt to inline unloaded methods 310 return NULL; 311 } 312 313 // Only a few intrinsics implement a virtual dispatch. 314 // They are expensive calls which are also frequently overridden. 315 if (is_virtual) { 316 switch (id) { 317 case vmIntrinsics::_hashCode: 318 case vmIntrinsics::_clone: 319 // OK, Object.hashCode and Object.clone intrinsics come in both flavors 320 break; 321 default: 322 return NULL; 323 } 324 } 325 326 // -XX:-InlineNatives disables nearly all intrinsics: 327 if (!InlineNatives) { 328 switch (id) { 329 case vmIntrinsics::_indexOf: 330 case vmIntrinsics::_compareTo: 331 case vmIntrinsics::_equals: 332 case vmIntrinsics::_equalsC: 333 case vmIntrinsics::_getAndAddInt: 334 case vmIntrinsics::_getAndAddLong: 335 case vmIntrinsics::_getAndSetInt: 336 case vmIntrinsics::_getAndSetLong: 337 case vmIntrinsics::_getAndSetObject: 338 case vmIntrinsics::_loadFence: 339 case vmIntrinsics::_storeFence: 340 case vmIntrinsics::_fullFence: 341 break; // InlineNatives does not control String.compareTo 342 case vmIntrinsics::_Reference_get: 343 break; // InlineNatives does not control Reference.get 344 default: 345 return NULL; 346 } 347 } 348 349 int predicates = 0; 350 bool does_virtual_dispatch = false; 351 352 switch (id) { 353 case vmIntrinsics::_compareTo: 354 if (!SpecialStringCompareTo) return NULL; 355 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL; 356 break; 357 case vmIntrinsics::_indexOf: 358 if (!SpecialStringIndexOf) return NULL; 359 break; 360 case vmIntrinsics::_equals: 361 if (!SpecialStringEquals) return NULL; 362 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL; 363 break; 364 case vmIntrinsics::_equalsC: 365 if (!SpecialArraysEquals) return NULL; 366 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL; 367 break; 368 case vmIntrinsics::_arraycopy: 369 if (!InlineArrayCopy) return NULL; 370 break; 371 case vmIntrinsics::_copyMemory: 372 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL; 373 if (!InlineArrayCopy) return NULL; 374 break; 375 case vmIntrinsics::_hashCode: 376 if (!InlineObjectHash) return NULL; 377 does_virtual_dispatch = true; 378 break; 379 case vmIntrinsics::_clone: 380 does_virtual_dispatch = true; 381 case vmIntrinsics::_copyOf: 382 case vmIntrinsics::_copyOfRange: 383 if (!InlineObjectCopy) return NULL; 384 // These also use the arraycopy intrinsic mechanism: 385 if (!InlineArrayCopy) return NULL; 386 break; 387 case vmIntrinsics::_encodeISOArray: 388 if (!SpecialEncodeISOArray) return NULL; 389 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL; 390 break; 391 case vmIntrinsics::_checkIndex: 392 // We do not intrinsify this. The optimizer does fine with it. 393 return NULL; 394 395 case vmIntrinsics::_getCallerClass: 396 if (!InlineReflectionGetCallerClass) return NULL; 397 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL; 398 break; 399 400 case vmIntrinsics::_bitCount_i: 401 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL; 402 break; 403 404 case vmIntrinsics::_bitCount_l: 405 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL; 406 break; 407 408 case vmIntrinsics::_numberOfLeadingZeros_i: 409 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL; 410 break; 411 412 case vmIntrinsics::_numberOfLeadingZeros_l: 413 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL; 414 break; 415 416 case vmIntrinsics::_numberOfTrailingZeros_i: 417 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL; 418 break; 419 420 case vmIntrinsics::_numberOfTrailingZeros_l: 421 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL; 422 break; 423 424 case vmIntrinsics::_reverseBytes_c: 425 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL; 426 break; 427 case vmIntrinsics::_reverseBytes_s: 428 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL; 429 break; 430 case vmIntrinsics::_reverseBytes_i: 431 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL; 432 break; 433 case vmIntrinsics::_reverseBytes_l: 434 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL; 435 break; 436 437 case vmIntrinsics::_Reference_get: 438 // Use the intrinsic version of Reference.get() so that the value in 439 // the referent field can be registered by the G1 pre-barrier code. 440 // Also add memory barrier to prevent commoning reads from this field 441 // across safepoint since GC can change it value. 442 break; 443 444 case vmIntrinsics::_compareAndSwapObject: 445 #ifdef _LP64 446 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL; 447 #endif 448 break; 449 450 case vmIntrinsics::_compareAndSwapLong: 451 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL; 452 break; 453 454 case vmIntrinsics::_getAndAddInt: 455 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL; 456 break; 457 458 case vmIntrinsics::_getAndAddLong: 459 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL; 460 break; 461 462 case vmIntrinsics::_getAndSetInt: 463 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL; 464 break; 465 466 case vmIntrinsics::_getAndSetLong: 467 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL; 468 break; 469 470 case vmIntrinsics::_getAndSetObject: 471 #ifdef _LP64 472 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 473 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL; 474 break; 475 #else 476 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL; 477 break; 478 #endif 479 480 case vmIntrinsics::_aescrypt_encryptBlock: 481 case vmIntrinsics::_aescrypt_decryptBlock: 482 if (!UseAESIntrinsics) return NULL; 483 break; 484 485 case vmIntrinsics::_multiplyToLen: 486 if (!UseMultiplyToLenIntrinsic) return NULL; 487 break; 488 489 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 490 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 491 if (!UseAESIntrinsics) return NULL; 492 // these two require the predicated logic 493 predicates = 1; 494 break; 495 496 case vmIntrinsics::_sha_implCompress: 497 if (!UseSHA1Intrinsics) return NULL; 498 break; 499 500 case vmIntrinsics::_sha2_implCompress: 501 if (!UseSHA256Intrinsics) return NULL; 502 break; 503 504 case vmIntrinsics::_sha5_implCompress: 505 if (!UseSHA512Intrinsics) return NULL; 506 break; 507 508 case vmIntrinsics::_digestBase_implCompressMB: 509 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL; 510 predicates = 3; 511 break; 512 513 case vmIntrinsics::_updateCRC32: 514 case vmIntrinsics::_updateBytesCRC32: 515 case vmIntrinsics::_updateByteBufferCRC32: 516 if (!UseCRC32Intrinsics) return NULL; 517 break; 518 519 case vmIntrinsics::_incrementExactI: 520 case vmIntrinsics::_addExactI: 521 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL; 522 break; 523 case vmIntrinsics::_incrementExactL: 524 case vmIntrinsics::_addExactL: 525 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL; 526 break; 527 case vmIntrinsics::_decrementExactI: 528 case vmIntrinsics::_subtractExactI: 529 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 530 break; 531 case vmIntrinsics::_decrementExactL: 532 case vmIntrinsics::_subtractExactL: 533 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 534 break; 535 case vmIntrinsics::_negateExactI: 536 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL; 537 break; 538 case vmIntrinsics::_negateExactL: 539 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL; 540 break; 541 case vmIntrinsics::_multiplyExactI: 542 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL; 543 break; 544 case vmIntrinsics::_multiplyExactL: 545 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL; 546 break; 547 548 default: 549 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility"); 550 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?"); 551 break; 552 } 553 554 // -XX:-InlineClassNatives disables natives from the Class class. 555 // The flag applies to all reflective calls, notably Array.newArray 556 // (visible to Java programmers as Array.newInstance). 557 if (m->holder()->name() == ciSymbol::java_lang_Class() || 558 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) { 559 if (!InlineClassNatives) return NULL; 560 } 561 562 // -XX:-InlineThreadNatives disables natives from the Thread class. 563 if (m->holder()->name() == ciSymbol::java_lang_Thread()) { 564 if (!InlineThreadNatives) return NULL; 565 } 566 567 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes. 568 if (m->holder()->name() == ciSymbol::java_lang_Math() || 569 m->holder()->name() == ciSymbol::java_lang_Float() || 570 m->holder()->name() == ciSymbol::java_lang_Double()) { 571 if (!InlineMathNatives) return NULL; 572 } 573 574 // -XX:-InlineUnsafeOps disables natives from the Unsafe class. 575 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) { 576 if (!InlineUnsafeOps) return NULL; 577 } 578 579 return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id); 580 } 581 582 //----------------------register_library_intrinsics----------------------- 583 // Initialize this file's data structures, for each Compile instance. 584 void Compile::register_library_intrinsics() { 585 // Nothing to do here. 586 } 587 588 JVMState* LibraryIntrinsic::generate(JVMState* jvms) { 589 LibraryCallKit kit(jvms, this); 590 Compile* C = kit.C; 591 int nodes = C->unique(); 592 #ifndef PRODUCT 593 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 594 char buf[1000]; 595 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 596 tty->print_cr("Intrinsic %s", str); 597 } 598 #endif 599 ciMethod* callee = kit.callee(); 600 const int bci = kit.bci(); 601 602 // Try to inline the intrinsic. 603 if (kit.try_to_inline(_last_predicate)) { 604 if (C->print_intrinsics() || C->print_inlining()) { 605 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)"); 606 } 607 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 608 if (C->log()) { 609 C->log()->elem("intrinsic id='%s'%s nodes='%d'", 610 vmIntrinsics::name_at(intrinsic_id()), 611 (is_virtual() ? " virtual='1'" : ""), 612 C->unique() - nodes); 613 } 614 // Push the result from the inlined method onto the stack. 615 kit.push_result(); 616 C->print_inlining_update(this); 617 return kit.transfer_exceptions_into_jvms(); 618 } 619 620 // The intrinsic bailed out 621 if (C->print_intrinsics() || C->print_inlining()) { 622 if (jvms->has_method()) { 623 // Not a root compile. 624 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)"; 625 C->print_inlining(callee, jvms->depth() - 1, bci, msg); 626 } else { 627 // Root compile 628 tty->print("Did not generate intrinsic %s%s at bci:%d in", 629 vmIntrinsics::name_at(intrinsic_id()), 630 (is_virtual() ? " (virtual)" : ""), bci); 631 } 632 } 633 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 634 C->print_inlining_update(this); 635 return NULL; 636 } 637 638 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) { 639 LibraryCallKit kit(jvms, this); 640 Compile* C = kit.C; 641 int nodes = C->unique(); 642 _last_predicate = predicate; 643 #ifndef PRODUCT 644 assert(is_predicated() && predicate < predicates_count(), "sanity"); 645 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 646 char buf[1000]; 647 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf)); 648 tty->print_cr("Predicate for intrinsic %s", str); 649 } 650 #endif 651 ciMethod* callee = kit.callee(); 652 const int bci = kit.bci(); 653 654 Node* slow_ctl = kit.try_to_predicate(predicate); 655 if (!kit.failing()) { 656 if (C->print_intrinsics() || C->print_inlining()) { 657 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)"); 658 } 659 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked); 660 if (C->log()) { 661 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'", 662 vmIntrinsics::name_at(intrinsic_id()), 663 (is_virtual() ? " virtual='1'" : ""), 664 C->unique() - nodes); 665 } 666 return slow_ctl; // Could be NULL if the check folds. 667 } 668 669 // The intrinsic bailed out 670 if (C->print_intrinsics() || C->print_inlining()) { 671 if (jvms->has_method()) { 672 // Not a root compile. 673 const char* msg = "failed to generate predicate for intrinsic"; 674 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg); 675 } else { 676 // Root compile 677 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in", 678 vmIntrinsics::name_at(intrinsic_id()), 679 (is_virtual() ? " (virtual)" : ""), bci); 680 } 681 } 682 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed); 683 return NULL; 684 } 685 686 bool LibraryCallKit::try_to_inline(int predicate) { 687 // Handle symbolic names for otherwise undistinguished boolean switches: 688 const bool is_store = true; 689 const bool is_native_ptr = true; 690 const bool is_static = true; 691 const bool is_volatile = true; 692 693 if (!jvms()->has_method()) { 694 // Root JVMState has a null method. 695 assert(map()->memory()->Opcode() == Op_Parm, ""); 696 // Insert the memory aliasing node 697 set_all_memory(reset_memory()); 698 } 699 assert(merged_memory(), ""); 700 701 702 switch (intrinsic_id()) { 703 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static); 704 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static); 705 case vmIntrinsics::_getClass: return inline_native_getClass(); 706 707 case vmIntrinsics::_dsin: 708 case vmIntrinsics::_dcos: 709 case vmIntrinsics::_dtan: 710 case vmIntrinsics::_dabs: 711 case vmIntrinsics::_datan2: 712 case vmIntrinsics::_dsqrt: 713 case vmIntrinsics::_dexp: 714 case vmIntrinsics::_dlog: 715 case vmIntrinsics::_dlog10: 716 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id()); 717 718 case vmIntrinsics::_min: 719 case vmIntrinsics::_max: return inline_min_max(intrinsic_id()); 720 721 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */); 722 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */); 723 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */); 724 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */); 725 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */); 726 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */); 727 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI(); 728 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL(); 729 case vmIntrinsics::_negateExactI: return inline_math_negateExactI(); 730 case vmIntrinsics::_negateExactL: return inline_math_negateExactL(); 731 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */); 732 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */); 733 734 case vmIntrinsics::_arraycopy: return inline_arraycopy(); 735 736 case vmIntrinsics::_compareTo: return inline_string_compareTo(); 737 case vmIntrinsics::_indexOf: return inline_string_indexOf(); 738 case vmIntrinsics::_equals: return inline_string_equals(); 739 740 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile); 741 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile); 742 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile); 743 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile); 744 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile); 745 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile); 746 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile); 747 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile); 748 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 749 750 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile); 751 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile); 752 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile); 753 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile); 754 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile); 755 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile); 756 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile); 757 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile); 758 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile); 759 760 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile); 761 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile); 762 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile); 763 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile); 764 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile); 765 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile); 766 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile); 767 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile); 768 769 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile); 770 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile); 771 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile); 772 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile); 773 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile); 774 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile); 775 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile); 776 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile); 777 778 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile); 779 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile); 780 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile); 781 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile); 782 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile); 783 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile); 784 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile); 785 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile); 786 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile); 787 788 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile); 789 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile); 790 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile); 791 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile); 792 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile); 793 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile); 794 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile); 795 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile); 796 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile); 797 798 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static); 799 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static); 800 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static); 801 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static); 802 803 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg); 804 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg); 805 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg); 806 807 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT); 808 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT); 809 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG); 810 811 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd); 812 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd); 813 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg); 814 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg); 815 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg); 816 817 case vmIntrinsics::_loadFence: 818 case vmIntrinsics::_storeFence: 819 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id()); 820 821 case vmIntrinsics::_currentThread: return inline_native_currentThread(); 822 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted(); 823 824 #ifdef TRACE_HAVE_INTRINSICS 825 case vmIntrinsics::_classID: return inline_native_classID(); 826 case vmIntrinsics::_threadID: return inline_native_threadID(); 827 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime"); 828 #endif 829 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis"); 830 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime"); 831 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate(); 832 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory(); 833 case vmIntrinsics::_newArray: return inline_native_newArray(); 834 case vmIntrinsics::_getLength: return inline_native_getLength(); 835 case vmIntrinsics::_copyOf: return inline_array_copyOf(false); 836 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true); 837 case vmIntrinsics::_equalsC: return inline_array_equals(); 838 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual()); 839 840 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check(); 841 842 case vmIntrinsics::_isInstance: 843 case vmIntrinsics::_getModifiers: 844 case vmIntrinsics::_isInterface: 845 case vmIntrinsics::_isArray: 846 case vmIntrinsics::_isPrimitive: 847 case vmIntrinsics::_getSuperclass: 848 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id()); 849 850 case vmIntrinsics::_floatToRawIntBits: 851 case vmIntrinsics::_floatToIntBits: 852 case vmIntrinsics::_intBitsToFloat: 853 case vmIntrinsics::_doubleToRawLongBits: 854 case vmIntrinsics::_doubleToLongBits: 855 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id()); 856 857 case vmIntrinsics::_numberOfLeadingZeros_i: 858 case vmIntrinsics::_numberOfLeadingZeros_l: 859 case vmIntrinsics::_numberOfTrailingZeros_i: 860 case vmIntrinsics::_numberOfTrailingZeros_l: 861 case vmIntrinsics::_bitCount_i: 862 case vmIntrinsics::_bitCount_l: 863 case vmIntrinsics::_reverseBytes_i: 864 case vmIntrinsics::_reverseBytes_l: 865 case vmIntrinsics::_reverseBytes_s: 866 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id()); 867 868 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass(); 869 870 case vmIntrinsics::_Reference_get: return inline_reference_get(); 871 872 case vmIntrinsics::_aescrypt_encryptBlock: 873 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id()); 874 875 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 876 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 877 return inline_cipherBlockChaining_AESCrypt(intrinsic_id()); 878 879 case vmIntrinsics::_sha_implCompress: 880 case vmIntrinsics::_sha2_implCompress: 881 case vmIntrinsics::_sha5_implCompress: 882 return inline_sha_implCompress(intrinsic_id()); 883 884 case vmIntrinsics::_digestBase_implCompressMB: 885 return inline_digestBase_implCompressMB(predicate); 886 887 case vmIntrinsics::_multiplyToLen: 888 return inline_multiplyToLen(); 889 890 case vmIntrinsics::_encodeISOArray: 891 return inline_encodeISOArray(); 892 893 case vmIntrinsics::_updateCRC32: 894 return inline_updateCRC32(); 895 case vmIntrinsics::_updateBytesCRC32: 896 return inline_updateBytesCRC32(); 897 case vmIntrinsics::_updateByteBufferCRC32: 898 return inline_updateByteBufferCRC32(); 899 900 default: 901 // If you get here, it may be that someone has added a new intrinsic 902 // to the list in vmSymbols.hpp without implementing it here. 903 #ifndef PRODUCT 904 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 905 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)", 906 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 907 } 908 #endif 909 return false; 910 } 911 } 912 913 Node* LibraryCallKit::try_to_predicate(int predicate) { 914 if (!jvms()->has_method()) { 915 // Root JVMState has a null method. 916 assert(map()->memory()->Opcode() == Op_Parm, ""); 917 // Insert the memory aliasing node 918 set_all_memory(reset_memory()); 919 } 920 assert(merged_memory(), ""); 921 922 switch (intrinsic_id()) { 923 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 924 return inline_cipherBlockChaining_AESCrypt_predicate(false); 925 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 926 return inline_cipherBlockChaining_AESCrypt_predicate(true); 927 case vmIntrinsics::_digestBase_implCompressMB: 928 return inline_digestBase_implCompressMB_predicate(predicate); 929 930 default: 931 // If you get here, it may be that someone has added a new intrinsic 932 // to the list in vmSymbols.hpp without implementing it here. 933 #ifndef PRODUCT 934 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) { 935 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)", 936 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id()); 937 } 938 #endif 939 Node* slow_ctl = control(); 940 set_control(top()); // No fast path instrinsic 941 return slow_ctl; 942 } 943 } 944 945 //------------------------------set_result------------------------------- 946 // Helper function for finishing intrinsics. 947 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) { 948 record_for_igvn(region); 949 set_control(_gvn.transform(region)); 950 set_result( _gvn.transform(value)); 951 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity"); 952 } 953 954 //------------------------------generate_guard--------------------------- 955 // Helper function for generating guarded fast-slow graph structures. 956 // The given 'test', if true, guards a slow path. If the test fails 957 // then a fast path can be taken. (We generally hope it fails.) 958 // In all cases, GraphKit::control() is updated to the fast path. 959 // The returned value represents the control for the slow path. 960 // The return value is never 'top'; it is either a valid control 961 // or NULL if it is obvious that the slow path can never be taken. 962 // Also, if region and the slow control are not NULL, the slow edge 963 // is appended to the region. 964 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) { 965 if (stopped()) { 966 // Already short circuited. 967 return NULL; 968 } 969 970 // Build an if node and its projections. 971 // If test is true we take the slow path, which we assume is uncommon. 972 if (_gvn.type(test) == TypeInt::ZERO) { 973 // The slow branch is never taken. No need to build this guard. 974 return NULL; 975 } 976 977 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN); 978 979 Node* if_slow = _gvn.transform(new IfTrueNode(iff)); 980 if (if_slow == top()) { 981 // The slow branch is never taken. No need to build this guard. 982 return NULL; 983 } 984 985 if (region != NULL) 986 region->add_req(if_slow); 987 988 Node* if_fast = _gvn.transform(new IfFalseNode(iff)); 989 set_control(if_fast); 990 991 return if_slow; 992 } 993 994 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) { 995 return generate_guard(test, region, PROB_UNLIKELY_MAG(3)); 996 } 997 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) { 998 return generate_guard(test, region, PROB_FAIR); 999 } 1000 1001 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region, 1002 Node* *pos_index) { 1003 if (stopped()) 1004 return NULL; // already stopped 1005 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint] 1006 return NULL; // index is already adequately typed 1007 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0))); 1008 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1009 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN); 1010 if (is_neg != NULL && pos_index != NULL) { 1011 // Emulate effect of Parse::adjust_map_after_if. 1012 Node* ccast = new CastIINode(index, TypeInt::POS); 1013 ccast->set_req(0, control()); 1014 (*pos_index) = _gvn.transform(ccast); 1015 } 1016 return is_neg; 1017 } 1018 1019 // Make sure that 'position' is a valid limit index, in [0..length]. 1020 // There are two equivalent plans for checking this: 1021 // A. (offset + copyLength) unsigned<= arrayLength 1022 // B. offset <= (arrayLength - copyLength) 1023 // We require that all of the values above, except for the sum and 1024 // difference, are already known to be non-negative. 1025 // Plan A is robust in the face of overflow, if offset and copyLength 1026 // are both hugely positive. 1027 // 1028 // Plan B is less direct and intuitive, but it does not overflow at 1029 // all, since the difference of two non-negatives is always 1030 // representable. Whenever Java methods must perform the equivalent 1031 // check they generally use Plan B instead of Plan A. 1032 // For the moment we use Plan A. 1033 inline Node* LibraryCallKit::generate_limit_guard(Node* offset, 1034 Node* subseq_length, 1035 Node* array_length, 1036 RegionNode* region) { 1037 if (stopped()) 1038 return NULL; // already stopped 1039 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO; 1040 if (zero_offset && subseq_length->eqv_uncast(array_length)) 1041 return NULL; // common case of whole-array copy 1042 Node* last = subseq_length; 1043 if (!zero_offset) // last += offset 1044 last = _gvn.transform(new AddINode(last, offset)); 1045 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last)); 1046 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt)); 1047 Node* is_over = generate_guard(bol_lt, region, PROB_MIN); 1048 return is_over; 1049 } 1050 1051 1052 //--------------------------generate_current_thread-------------------- 1053 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) { 1054 ciKlass* thread_klass = env()->Thread_klass(); 1055 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull); 1056 Node* thread = _gvn.transform(new ThreadLocalNode()); 1057 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset())); 1058 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered); 1059 tls_output = thread; 1060 return threadObj; 1061 } 1062 1063 1064 //------------------------------make_string_method_node------------------------ 1065 // Helper method for String intrinsic functions. This version is called 1066 // with str1 and str2 pointing to String object nodes. 1067 // 1068 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) { 1069 Node* no_ctrl = NULL; 1070 1071 // Get start addr of string 1072 Node* str1_value = load_String_value(no_ctrl, str1); 1073 Node* str1_offset = load_String_offset(no_ctrl, str1); 1074 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR); 1075 1076 // Get length of string 1 1077 Node* str1_len = load_String_length(no_ctrl, str1); 1078 1079 Node* str2_value = load_String_value(no_ctrl, str2); 1080 Node* str2_offset = load_String_offset(no_ctrl, str2); 1081 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR); 1082 1083 Node* str2_len = NULL; 1084 Node* result = NULL; 1085 1086 switch (opcode) { 1087 case Op_StrIndexOf: 1088 // Get length of string 2 1089 str2_len = load_String_length(no_ctrl, str2); 1090 1091 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1092 str1_start, str1_len, str2_start, str2_len); 1093 break; 1094 case Op_StrComp: 1095 // Get length of string 2 1096 str2_len = load_String_length(no_ctrl, str2); 1097 1098 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1099 str1_start, str1_len, str2_start, str2_len); 1100 break; 1101 case Op_StrEquals: 1102 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1103 str1_start, str2_start, str1_len); 1104 break; 1105 default: 1106 ShouldNotReachHere(); 1107 return NULL; 1108 } 1109 1110 // All these intrinsics have checks. 1111 C->set_has_split_ifs(true); // Has chance for split-if optimization 1112 1113 return _gvn.transform(result); 1114 } 1115 1116 // Helper method for String intrinsic functions. This version is called 1117 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing 1118 // to Int nodes containing the lenghts of str1 and str2. 1119 // 1120 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) { 1121 Node* result = NULL; 1122 switch (opcode) { 1123 case Op_StrIndexOf: 1124 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS), 1125 str1_start, cnt1, str2_start, cnt2); 1126 break; 1127 case Op_StrComp: 1128 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS), 1129 str1_start, cnt1, str2_start, cnt2); 1130 break; 1131 case Op_StrEquals: 1132 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS), 1133 str1_start, str2_start, cnt1); 1134 break; 1135 default: 1136 ShouldNotReachHere(); 1137 return NULL; 1138 } 1139 1140 // All these intrinsics have checks. 1141 C->set_has_split_ifs(true); // Has chance for split-if optimization 1142 1143 return _gvn.transform(result); 1144 } 1145 1146 //------------------------------inline_string_compareTo------------------------ 1147 // public int java.lang.String.compareTo(String anotherString); 1148 bool LibraryCallKit::inline_string_compareTo() { 1149 Node* receiver = null_check(argument(0)); 1150 Node* arg = null_check(argument(1)); 1151 if (stopped()) { 1152 return true; 1153 } 1154 set_result(make_string_method_node(Op_StrComp, receiver, arg)); 1155 return true; 1156 } 1157 1158 //------------------------------inline_string_equals------------------------ 1159 bool LibraryCallKit::inline_string_equals() { 1160 Node* receiver = null_check_receiver(); 1161 // NOTE: Do not null check argument for String.equals() because spec 1162 // allows to specify NULL as argument. 1163 Node* argument = this->argument(1); 1164 if (stopped()) { 1165 return true; 1166 } 1167 1168 // paths (plus control) merge 1169 RegionNode* region = new RegionNode(5); 1170 Node* phi = new PhiNode(region, TypeInt::BOOL); 1171 1172 // does source == target string? 1173 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument)); 1174 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1175 1176 Node* if_eq = generate_slow_guard(bol, NULL); 1177 if (if_eq != NULL) { 1178 // receiver == argument 1179 phi->init_req(2, intcon(1)); 1180 region->init_req(2, if_eq); 1181 } 1182 1183 // get String klass for instanceOf 1184 ciInstanceKlass* klass = env()->String_klass(); 1185 1186 if (!stopped()) { 1187 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass))); 1188 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1))); 1189 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1190 1191 Node* inst_false = generate_guard(bol, NULL, PROB_MIN); 1192 //instanceOf == true, fallthrough 1193 1194 if (inst_false != NULL) { 1195 phi->init_req(3, intcon(0)); 1196 region->init_req(3, inst_false); 1197 } 1198 } 1199 1200 if (!stopped()) { 1201 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass); 1202 1203 // Properly cast the argument to String 1204 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type)); 1205 // This path is taken only when argument's type is String:NotNull. 1206 argument = cast_not_null(argument, false); 1207 1208 Node* no_ctrl = NULL; 1209 1210 // Get start addr of receiver 1211 Node* receiver_val = load_String_value(no_ctrl, receiver); 1212 Node* receiver_offset = load_String_offset(no_ctrl, receiver); 1213 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR); 1214 1215 // Get length of receiver 1216 Node* receiver_cnt = load_String_length(no_ctrl, receiver); 1217 1218 // Get start addr of argument 1219 Node* argument_val = load_String_value(no_ctrl, argument); 1220 Node* argument_offset = load_String_offset(no_ctrl, argument); 1221 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR); 1222 1223 // Get length of argument 1224 Node* argument_cnt = load_String_length(no_ctrl, argument); 1225 1226 // Check for receiver count != argument count 1227 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt)); 1228 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 1229 Node* if_ne = generate_slow_guard(bol, NULL); 1230 if (if_ne != NULL) { 1231 phi->init_req(4, intcon(0)); 1232 region->init_req(4, if_ne); 1233 } 1234 1235 // Check for count == 0 is done by assembler code for StrEquals. 1236 1237 if (!stopped()) { 1238 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt); 1239 phi->init_req(1, equals); 1240 region->init_req(1, control()); 1241 } 1242 } 1243 1244 // post merge 1245 set_control(_gvn.transform(region)); 1246 record_for_igvn(region); 1247 1248 set_result(_gvn.transform(phi)); 1249 return true; 1250 } 1251 1252 //------------------------------inline_array_equals---------------------------- 1253 bool LibraryCallKit::inline_array_equals() { 1254 Node* arg1 = argument(0); 1255 Node* arg2 = argument(1); 1256 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2))); 1257 return true; 1258 } 1259 1260 // Java version of String.indexOf(constant string) 1261 // class StringDecl { 1262 // StringDecl(char[] ca) { 1263 // offset = 0; 1264 // count = ca.length; 1265 // value = ca; 1266 // } 1267 // int offset; 1268 // int count; 1269 // char[] value; 1270 // } 1271 // 1272 // static int string_indexOf_J(StringDecl string_object, char[] target_object, 1273 // int targetOffset, int cache_i, int md2) { 1274 // int cache = cache_i; 1275 // int sourceOffset = string_object.offset; 1276 // int sourceCount = string_object.count; 1277 // int targetCount = target_object.length; 1278 // 1279 // int targetCountLess1 = targetCount - 1; 1280 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1; 1281 // 1282 // char[] source = string_object.value; 1283 // char[] target = target_object; 1284 // int lastChar = target[targetCountLess1]; 1285 // 1286 // outer_loop: 1287 // for (int i = sourceOffset; i < sourceEnd; ) { 1288 // int src = source[i + targetCountLess1]; 1289 // if (src == lastChar) { 1290 // // With random strings and a 4-character alphabet, 1291 // // reverse matching at this point sets up 0.8% fewer 1292 // // frames, but (paradoxically) makes 0.3% more probes. 1293 // // Since those probes are nearer the lastChar probe, 1294 // // there is may be a net D$ win with reverse matching. 1295 // // But, reversing loop inhibits unroll of inner loop 1296 // // for unknown reason. So, does running outer loop from 1297 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount) 1298 // for (int j = 0; j < targetCountLess1; j++) { 1299 // if (target[targetOffset + j] != source[i+j]) { 1300 // if ((cache & (1 << source[i+j])) == 0) { 1301 // if (md2 < j+1) { 1302 // i += j+1; 1303 // continue outer_loop; 1304 // } 1305 // } 1306 // i += md2; 1307 // continue outer_loop; 1308 // } 1309 // } 1310 // return i - sourceOffset; 1311 // } 1312 // if ((cache & (1 << src)) == 0) { 1313 // i += targetCountLess1; 1314 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump. 1315 // i++; 1316 // } 1317 // return -1; 1318 // } 1319 1320 //------------------------------string_indexOf------------------------ 1321 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i, 1322 jint cache_i, jint md2_i) { 1323 1324 Node* no_ctrl = NULL; 1325 float likely = PROB_LIKELY(0.9); 1326 float unlikely = PROB_UNLIKELY(0.9); 1327 1328 const int nargs = 0; // no arguments to push back for uncommon trap in predicate 1329 1330 Node* source = load_String_value(no_ctrl, string_object); 1331 Node* sourceOffset = load_String_offset(no_ctrl, string_object); 1332 Node* sourceCount = load_String_length(no_ctrl, string_object); 1333 1334 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true))); 1335 jint target_length = target_array->length(); 1336 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin)); 1337 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot); 1338 1339 // String.value field is known to be @Stable. 1340 if (UseImplicitStableValues) { 1341 target = cast_array_to_stable(target, target_type); 1342 } 1343 1344 IdealKit kit(this, false, true); 1345 #define __ kit. 1346 Node* zero = __ ConI(0); 1347 Node* one = __ ConI(1); 1348 Node* cache = __ ConI(cache_i); 1349 Node* md2 = __ ConI(md2_i); 1350 Node* lastChar = __ ConI(target_array->char_at(target_length - 1)); 1351 Node* targetCount = __ ConI(target_length); 1352 Node* targetCountLess1 = __ ConI(target_length - 1); 1353 Node* targetOffset = __ ConI(targetOffset_i); 1354 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1); 1355 1356 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done(); 1357 Node* outer_loop = __ make_label(2 /* goto */); 1358 Node* return_ = __ make_label(1); 1359 1360 __ set(rtn,__ ConI(-1)); 1361 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); { 1362 Node* i2 = __ AddI(__ value(i), targetCountLess1); 1363 // pin to prohibit loading of "next iteration" value which may SEGV (rare) 1364 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS); 1365 __ if_then(src, BoolTest::eq, lastChar, unlikely); { 1366 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); { 1367 Node* tpj = __ AddI(targetOffset, __ value(j)); 1368 Node* targ = load_array_element(no_ctrl, target, tpj, target_type); 1369 Node* ipj = __ AddI(__ value(i), __ value(j)); 1370 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS); 1371 __ if_then(targ, BoolTest::ne, src2); { 1372 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); { 1373 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); { 1374 __ increment(i, __ AddI(__ value(j), one)); 1375 __ goto_(outer_loop); 1376 } __ end_if(); __ dead(j); 1377 }__ end_if(); __ dead(j); 1378 __ increment(i, md2); 1379 __ goto_(outer_loop); 1380 }__ end_if(); 1381 __ increment(j, one); 1382 }__ end_loop(); __ dead(j); 1383 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i); 1384 __ goto_(return_); 1385 }__ end_if(); 1386 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); { 1387 __ increment(i, targetCountLess1); 1388 }__ end_if(); 1389 __ increment(i, one); 1390 __ bind(outer_loop); 1391 }__ end_loop(); __ dead(i); 1392 __ bind(return_); 1393 1394 // Final sync IdealKit and GraphKit. 1395 final_sync(kit); 1396 Node* result = __ value(rtn); 1397 #undef __ 1398 C->set_has_loops(true); 1399 return result; 1400 } 1401 1402 //------------------------------inline_string_indexOf------------------------ 1403 bool LibraryCallKit::inline_string_indexOf() { 1404 Node* receiver = argument(0); 1405 Node* arg = argument(1); 1406 1407 Node* result; 1408 // Disable the use of pcmpestri until it can be guaranteed that 1409 // the load doesn't cross into the uncommited space. 1410 if (Matcher::has_match_rule(Op_StrIndexOf) && 1411 UseSSE42Intrinsics) { 1412 // Generate SSE4.2 version of indexOf 1413 // We currently only have match rules that use SSE4.2 1414 1415 receiver = null_check(receiver); 1416 arg = null_check(arg); 1417 if (stopped()) { 1418 return true; 1419 } 1420 1421 ciInstanceKlass* str_klass = env()->String_klass(); 1422 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass); 1423 1424 // Make the merge point 1425 RegionNode* result_rgn = new RegionNode(4); 1426 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT); 1427 Node* no_ctrl = NULL; 1428 1429 // Get start addr of source string 1430 Node* source = load_String_value(no_ctrl, receiver); 1431 Node* source_offset = load_String_offset(no_ctrl, receiver); 1432 Node* source_start = array_element_address(source, source_offset, T_CHAR); 1433 1434 // Get length of source string 1435 Node* source_cnt = load_String_length(no_ctrl, receiver); 1436 1437 // Get start addr of substring 1438 Node* substr = load_String_value(no_ctrl, arg); 1439 Node* substr_offset = load_String_offset(no_ctrl, arg); 1440 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR); 1441 1442 // Get length of source string 1443 Node* substr_cnt = load_String_length(no_ctrl, arg); 1444 1445 // Check for substr count > string count 1446 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt)); 1447 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt)); 1448 Node* if_gt = generate_slow_guard(bol, NULL); 1449 if (if_gt != NULL) { 1450 result_phi->init_req(2, intcon(-1)); 1451 result_rgn->init_req(2, if_gt); 1452 } 1453 1454 if (!stopped()) { 1455 // Check for substr count == 0 1456 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0))); 1457 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq)); 1458 Node* if_zero = generate_slow_guard(bol, NULL); 1459 if (if_zero != NULL) { 1460 result_phi->init_req(3, intcon(0)); 1461 result_rgn->init_req(3, if_zero); 1462 } 1463 } 1464 1465 if (!stopped()) { 1466 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt); 1467 result_phi->init_req(1, result); 1468 result_rgn->init_req(1, control()); 1469 } 1470 set_control(_gvn.transform(result_rgn)); 1471 record_for_igvn(result_rgn); 1472 result = _gvn.transform(result_phi); 1473 1474 } else { // Use LibraryCallKit::string_indexOf 1475 // don't intrinsify if argument isn't a constant string. 1476 if (!arg->is_Con()) { 1477 return false; 1478 } 1479 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr(); 1480 if (str_type == NULL) { 1481 return false; 1482 } 1483 ciInstanceKlass* klass = env()->String_klass(); 1484 ciObject* str_const = str_type->const_oop(); 1485 if (str_const == NULL || str_const->klass() != klass) { 1486 return false; 1487 } 1488 ciInstance* str = str_const->as_instance(); 1489 assert(str != NULL, "must be instance"); 1490 1491 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object(); 1492 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array 1493 1494 int o; 1495 int c; 1496 if (java_lang_String::has_offset_field()) { 1497 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int(); 1498 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int(); 1499 } else { 1500 o = 0; 1501 c = pat->length(); 1502 } 1503 1504 // constant strings have no offset and count == length which 1505 // simplifies the resulting code somewhat so lets optimize for that. 1506 if (o != 0 || c != pat->length()) { 1507 return false; 1508 } 1509 1510 receiver = null_check(receiver, T_OBJECT); 1511 // NOTE: No null check on the argument is needed since it's a constant String oop. 1512 if (stopped()) { 1513 return true; 1514 } 1515 1516 // The null string as a pattern always returns 0 (match at beginning of string) 1517 if (c == 0) { 1518 set_result(intcon(0)); 1519 return true; 1520 } 1521 1522 // Generate default indexOf 1523 jchar lastChar = pat->char_at(o + (c - 1)); 1524 int cache = 0; 1525 int i; 1526 for (i = 0; i < c - 1; i++) { 1527 assert(i < pat->length(), "out of range"); 1528 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1))); 1529 } 1530 1531 int md2 = c; 1532 for (i = 0; i < c - 1; i++) { 1533 assert(i < pat->length(), "out of range"); 1534 if (pat->char_at(o + i) == lastChar) { 1535 md2 = (c - 1) - i; 1536 } 1537 } 1538 1539 result = string_indexOf(receiver, pat, o, cache, md2); 1540 } 1541 set_result(result); 1542 return true; 1543 } 1544 1545 //--------------------------round_double_node-------------------------------- 1546 // Round a double node if necessary. 1547 Node* LibraryCallKit::round_double_node(Node* n) { 1548 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1) 1549 n = _gvn.transform(new RoundDoubleNode(0, n)); 1550 return n; 1551 } 1552 1553 //------------------------------inline_math----------------------------------- 1554 // public static double Math.abs(double) 1555 // public static double Math.sqrt(double) 1556 // public static double Math.log(double) 1557 // public static double Math.log10(double) 1558 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) { 1559 Node* arg = round_double_node(argument(0)); 1560 Node* n; 1561 switch (id) { 1562 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break; 1563 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break; 1564 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break; 1565 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break; 1566 default: fatal_unexpected_iid(id); break; 1567 } 1568 set_result(_gvn.transform(n)); 1569 return true; 1570 } 1571 1572 //------------------------------inline_trig---------------------------------- 1573 // Inline sin/cos/tan instructions, if possible. If rounding is required, do 1574 // argument reduction which will turn into a fast/slow diamond. 1575 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) { 1576 Node* arg = round_double_node(argument(0)); 1577 Node* n = NULL; 1578 1579 switch (id) { 1580 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break; 1581 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break; 1582 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break; 1583 default: fatal_unexpected_iid(id); break; 1584 } 1585 n = _gvn.transform(n); 1586 1587 // Rounding required? Check for argument reduction! 1588 if (Matcher::strict_fp_requires_explicit_rounding) { 1589 static const double pi_4 = 0.7853981633974483; 1590 static const double neg_pi_4 = -0.7853981633974483; 1591 // pi/2 in 80-bit extended precision 1592 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00}; 1593 // -pi/2 in 80-bit extended precision 1594 // 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}; 1595 // Cutoff value for using this argument reduction technique 1596 //static const double pi_2_minus_epsilon = 1.564660403643354; 1597 //static const double neg_pi_2_plus_epsilon = -1.564660403643354; 1598 1599 // Pseudocode for sin: 1600 // if (x <= Math.PI / 4.0) { 1601 // if (x >= -Math.PI / 4.0) return fsin(x); 1602 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0); 1603 // } else { 1604 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0); 1605 // } 1606 // return StrictMath.sin(x); 1607 1608 // Pseudocode for cos: 1609 // if (x <= Math.PI / 4.0) { 1610 // if (x >= -Math.PI / 4.0) return fcos(x); 1611 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0); 1612 // } else { 1613 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0); 1614 // } 1615 // return StrictMath.cos(x); 1616 1617 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it 1618 // requires a special machine instruction to load it. Instead we'll try 1619 // the 'easy' case. If we really need the extra range +/- PI/2 we'll 1620 // probably do the math inside the SIN encoding. 1621 1622 // Make the merge point 1623 RegionNode* r = new RegionNode(3); 1624 Node* phi = new PhiNode(r, Type::DOUBLE); 1625 1626 // Flatten arg so we need only 1 test 1627 Node *abs = _gvn.transform(new AbsDNode(arg)); 1628 // Node for PI/4 constant 1629 Node *pi4 = makecon(TypeD::make(pi_4)); 1630 // Check PI/4 : abs(arg) 1631 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs)); 1632 // Check: If PI/4 < abs(arg) then go slow 1633 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt )); 1634 // Branch either way 1635 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1636 set_control(opt_iff(r,iff)); 1637 1638 // Set fast path result 1639 phi->init_req(2, n); 1640 1641 // Slow path - non-blocking leaf call 1642 Node* call = NULL; 1643 switch (id) { 1644 case vmIntrinsics::_dsin: 1645 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1646 CAST_FROM_FN_PTR(address, SharedRuntime::dsin), 1647 "Sin", NULL, arg, top()); 1648 break; 1649 case vmIntrinsics::_dcos: 1650 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1651 CAST_FROM_FN_PTR(address, SharedRuntime::dcos), 1652 "Cos", NULL, arg, top()); 1653 break; 1654 case vmIntrinsics::_dtan: 1655 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(), 1656 CAST_FROM_FN_PTR(address, SharedRuntime::dtan), 1657 "Tan", NULL, arg, top()); 1658 break; 1659 } 1660 assert(control()->in(0) == call, ""); 1661 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 1662 r->init_req(1, control()); 1663 phi->init_req(1, slow_result); 1664 1665 // Post-merge 1666 set_control(_gvn.transform(r)); 1667 record_for_igvn(r); 1668 n = _gvn.transform(phi); 1669 1670 C->set_has_split_ifs(true); // Has chance for split-if optimization 1671 } 1672 set_result(n); 1673 return true; 1674 } 1675 1676 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) { 1677 //------------------- 1678 //result=(result.isNaN())? funcAddr():result; 1679 // Check: If isNaN() by checking result!=result? then either trap 1680 // or go to runtime 1681 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result)); 1682 // Build the boolean node 1683 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq)); 1684 1685 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1686 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT); 1687 // The pow or exp intrinsic returned a NaN, which requires a call 1688 // to the runtime. Recompile with the runtime call. 1689 uncommon_trap(Deoptimization::Reason_intrinsic, 1690 Deoptimization::Action_make_not_entrant); 1691 } 1692 return result; 1693 } else { 1694 // If this inlining ever returned NaN in the past, we compile a call 1695 // to the runtime to properly handle corner cases 1696 1697 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 1698 Node* if_slow = _gvn.transform(new IfFalseNode(iff)); 1699 Node* if_fast = _gvn.transform(new IfTrueNode(iff)); 1700 1701 if (!if_slow->is_top()) { 1702 RegionNode* result_region = new RegionNode(3); 1703 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE); 1704 1705 result_region->init_req(1, if_fast); 1706 result_val->init_req(1, result); 1707 1708 set_control(if_slow); 1709 1710 const TypePtr* no_memory_effects = NULL; 1711 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1712 no_memory_effects, 1713 x, top(), y, y ? top() : NULL); 1714 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0)); 1715 #ifdef ASSERT 1716 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1)); 1717 assert(value_top == top(), "second value must be top"); 1718 #endif 1719 1720 result_region->init_req(2, control()); 1721 result_val->init_req(2, value); 1722 set_control(_gvn.transform(result_region)); 1723 return _gvn.transform(result_val); 1724 } else { 1725 return result; 1726 } 1727 } 1728 } 1729 1730 //------------------------------inline_exp------------------------------------- 1731 // Inline exp instructions, if possible. The Intel hardware only misses 1732 // really odd corner cases (+/- Infinity). Just uncommon-trap them. 1733 bool LibraryCallKit::inline_exp() { 1734 Node* arg = round_double_node(argument(0)); 1735 Node* n = _gvn.transform(new ExpDNode(C, control(), arg)); 1736 1737 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP"); 1738 set_result(n); 1739 1740 C->set_has_split_ifs(true); // Has chance for split-if optimization 1741 return true; 1742 } 1743 1744 //------------------------------inline_pow------------------------------------- 1745 // Inline power instructions, if possible. 1746 bool LibraryCallKit::inline_pow() { 1747 // Pseudocode for pow 1748 // if (y == 2) { 1749 // return x * x; 1750 // } else { 1751 // if (x <= 0.0) { 1752 // long longy = (long)y; 1753 // if ((double)longy == y) { // if y is long 1754 // if (y + 1 == y) longy = 0; // huge number: even 1755 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y); 1756 // } else { 1757 // result = NaN; 1758 // } 1759 // } else { 1760 // result = DPow(x,y); 1761 // } 1762 // if (result != result)? { 1763 // result = uncommon_trap() or runtime_call(); 1764 // } 1765 // return result; 1766 // } 1767 1768 Node* x = round_double_node(argument(0)); 1769 Node* y = round_double_node(argument(2)); 1770 1771 Node* result = NULL; 1772 1773 Node* const_two_node = makecon(TypeD::make(2.0)); 1774 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node)); 1775 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq)); 1776 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1777 Node* if_true = _gvn.transform(new IfTrueNode(if_node)); 1778 Node* if_false = _gvn.transform(new IfFalseNode(if_node)); 1779 1780 RegionNode* region_node = new RegionNode(3); 1781 region_node->init_req(1, if_true); 1782 1783 Node* phi_node = new PhiNode(region_node, Type::DOUBLE); 1784 // special case for x^y where y == 2, we can convert it to x * x 1785 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x))); 1786 1787 // set control to if_false since we will now process the false branch 1788 set_control(if_false); 1789 1790 if (!too_many_traps(Deoptimization::Reason_intrinsic)) { 1791 // Short form: skip the fancy tests and just check for NaN result. 1792 result = _gvn.transform(new PowDNode(C, control(), x, y)); 1793 } else { 1794 // If this inlining ever returned NaN in the past, include all 1795 // checks + call to the runtime. 1796 1797 // Set the merge point for If node with condition of (x <= 0.0) 1798 // There are four possible paths to region node and phi node 1799 RegionNode *r = new RegionNode(4); 1800 Node *phi = new PhiNode(r, Type::DOUBLE); 1801 1802 // Build the first if node: if (x <= 0.0) 1803 // Node for 0 constant 1804 Node *zeronode = makecon(TypeD::ZERO); 1805 // Check x:0 1806 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode)); 1807 // Check: If (x<=0) then go complex path 1808 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le )); 1809 // Branch either way 1810 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1811 // Fast path taken; set region slot 3 1812 Node *fast_taken = _gvn.transform(new IfFalseNode(if1)); 1813 r->init_req(3,fast_taken); // Capture fast-control 1814 1815 // Fast path not-taken, i.e. slow path 1816 Node *complex_path = _gvn.transform(new IfTrueNode(if1)); 1817 1818 // Set fast path result 1819 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y)); 1820 phi->init_req(3, fast_result); 1821 1822 // Complex path 1823 // Build the second if node (if y is long) 1824 // Node for (long)y 1825 Node *longy = _gvn.transform(new ConvD2LNode(y)); 1826 // Node for (double)((long) y) 1827 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy)); 1828 // Check (double)((long) y) : y 1829 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y)); 1830 // Check if (y isn't long) then go to slow path 1831 1832 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne )); 1833 // Branch either way 1834 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN); 1835 Node* ylong_path = _gvn.transform(new IfFalseNode(if2)); 1836 1837 Node *slow_path = _gvn.transform(new IfTrueNode(if2)); 1838 1839 // Calculate DPow(abs(x), y)*(1 & (long)y) 1840 // Node for constant 1 1841 Node *conone = longcon(1); 1842 // 1& (long)y 1843 Node *signnode= _gvn.transform(new AndLNode(conone, longy)); 1844 1845 // A huge number is always even. Detect a huge number by checking 1846 // if y + 1 == y and set integer to be tested for parity to 0. 1847 // Required for corner case: 1848 // (long)9.223372036854776E18 = max_jlong 1849 // (double)(long)9.223372036854776E18 = 9.223372036854776E18 1850 // max_jlong is odd but 9.223372036854776E18 is even 1851 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1)))); 1852 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y)); 1853 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq )); 1854 Node* correctedsign = NULL; 1855 if (ConditionalMoveLimit != 0) { 1856 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG)); 1857 } else { 1858 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN); 1859 RegionNode *r = new RegionNode(3); 1860 Node *phi = new PhiNode(r, TypeLong::LONG); 1861 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1))); 1862 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1))); 1863 phi->init_req(1, signnode); 1864 phi->init_req(2, longcon(0)); 1865 correctedsign = _gvn.transform(phi); 1866 ylong_path = _gvn.transform(r); 1867 record_for_igvn(r); 1868 } 1869 1870 // zero node 1871 Node *conzero = longcon(0); 1872 // Check (1&(long)y)==0? 1873 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero)); 1874 // Check if (1&(long)y)!=0?, if so the result is negative 1875 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne )); 1876 // abs(x) 1877 Node *absx=_gvn.transform(new AbsDNode(x)); 1878 // abs(x)^y 1879 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y)); 1880 // -abs(x)^y 1881 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy)); 1882 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y) 1883 Node *signresult = NULL; 1884 if (ConditionalMoveLimit != 0) { 1885 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE)); 1886 } else { 1887 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN); 1888 RegionNode *r = new RegionNode(3); 1889 Node *phi = new PhiNode(r, Type::DOUBLE); 1890 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven))); 1891 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven))); 1892 phi->init_req(1, absxpowy); 1893 phi->init_req(2, negabsxpowy); 1894 signresult = _gvn.transform(phi); 1895 ylong_path = _gvn.transform(r); 1896 record_for_igvn(r); 1897 } 1898 // Set complex path fast result 1899 r->init_req(2, ylong_path); 1900 phi->init_req(2, signresult); 1901 1902 static const jlong nan_bits = CONST64(0x7ff8000000000000); 1903 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN 1904 r->init_req(1,slow_path); 1905 phi->init_req(1,slow_result); 1906 1907 // Post merge 1908 set_control(_gvn.transform(r)); 1909 record_for_igvn(r); 1910 result = _gvn.transform(phi); 1911 } 1912 1913 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW"); 1914 1915 // control from finish_pow_exp is now input to the region node 1916 region_node->set_req(2, control()); 1917 // the result from finish_pow_exp is now input to the phi node 1918 phi_node->init_req(2, result); 1919 set_control(_gvn.transform(region_node)); 1920 record_for_igvn(region_node); 1921 set_result(_gvn.transform(phi_node)); 1922 1923 C->set_has_split_ifs(true); // Has chance for split-if optimization 1924 return true; 1925 } 1926 1927 //------------------------------runtime_math----------------------------- 1928 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) { 1929 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(), 1930 "must be (DD)D or (D)D type"); 1931 1932 // Inputs 1933 Node* a = round_double_node(argument(0)); 1934 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL; 1935 1936 const TypePtr* no_memory_effects = NULL; 1937 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName, 1938 no_memory_effects, 1939 a, top(), b, b ? top() : NULL); 1940 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0)); 1941 #ifdef ASSERT 1942 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1)); 1943 assert(value_top == top(), "second value must be top"); 1944 #endif 1945 1946 set_result(value); 1947 return true; 1948 } 1949 1950 //------------------------------inline_math_native----------------------------- 1951 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) { 1952 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f) 1953 switch (id) { 1954 // These intrinsics are not properly supported on all hardware 1955 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) : 1956 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS"); 1957 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) : 1958 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN"); 1959 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) : 1960 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN"); 1961 1962 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) : 1963 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG"); 1964 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) : 1965 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10"); 1966 1967 // These intrinsics are supported on all hardware 1968 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false; 1969 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false; 1970 1971 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() : 1972 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP"); 1973 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() : 1974 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW"); 1975 #undef FN_PTR 1976 1977 // These intrinsics are not yet correctly implemented 1978 case vmIntrinsics::_datan2: 1979 return false; 1980 1981 default: 1982 fatal_unexpected_iid(id); 1983 return false; 1984 } 1985 } 1986 1987 static bool is_simple_name(Node* n) { 1988 return (n->req() == 1 // constant 1989 || (n->is_Type() && n->as_Type()->type()->singleton()) 1990 || n->is_Proj() // parameter or return value 1991 || n->is_Phi() // local of some sort 1992 ); 1993 } 1994 1995 //----------------------------inline_min_max----------------------------------- 1996 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) { 1997 set_result(generate_min_max(id, argument(0), argument(1))); 1998 return true; 1999 } 2000 2001 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) { 2002 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) ); 2003 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 2004 Node* fast_path = _gvn.transform( new IfFalseNode(check)); 2005 Node* slow_path = _gvn.transform( new IfTrueNode(check) ); 2006 2007 { 2008 PreserveJVMState pjvms(this); 2009 PreserveReexecuteState preexecs(this); 2010 jvms()->set_should_reexecute(true); 2011 2012 set_control(slow_path); 2013 set_i_o(i_o()); 2014 2015 uncommon_trap(Deoptimization::Reason_intrinsic, 2016 Deoptimization::Action_none); 2017 } 2018 2019 set_control(fast_path); 2020 set_result(math); 2021 } 2022 2023 template <typename OverflowOp> 2024 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) { 2025 typedef typename OverflowOp::MathOp MathOp; 2026 2027 MathOp* mathOp = new MathOp(arg1, arg2); 2028 Node* operation = _gvn.transform( mathOp ); 2029 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) ); 2030 inline_math_mathExact(operation, ofcheck); 2031 return true; 2032 } 2033 2034 bool LibraryCallKit::inline_math_addExactI(bool is_increment) { 2035 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1)); 2036 } 2037 2038 bool LibraryCallKit::inline_math_addExactL(bool is_increment) { 2039 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2)); 2040 } 2041 2042 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) { 2043 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1)); 2044 } 2045 2046 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) { 2047 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2)); 2048 } 2049 2050 bool LibraryCallKit::inline_math_negateExactI() { 2051 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0)); 2052 } 2053 2054 bool LibraryCallKit::inline_math_negateExactL() { 2055 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0)); 2056 } 2057 2058 bool LibraryCallKit::inline_math_multiplyExactI() { 2059 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1)); 2060 } 2061 2062 bool LibraryCallKit::inline_math_multiplyExactL() { 2063 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2)); 2064 } 2065 2066 Node* 2067 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) { 2068 // These are the candidate return value: 2069 Node* xvalue = x0; 2070 Node* yvalue = y0; 2071 2072 if (xvalue == yvalue) { 2073 return xvalue; 2074 } 2075 2076 bool want_max = (id == vmIntrinsics::_max); 2077 2078 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int(); 2079 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int(); 2080 if (txvalue == NULL || tyvalue == NULL) return top(); 2081 // This is not really necessary, but it is consistent with a 2082 // hypothetical MaxINode::Value method: 2083 int widen = MAX2(txvalue->_widen, tyvalue->_widen); 2084 2085 // %%% This folding logic should (ideally) be in a different place. 2086 // Some should be inside IfNode, and there to be a more reliable 2087 // transformation of ?: style patterns into cmoves. We also want 2088 // more powerful optimizations around cmove and min/max. 2089 2090 // Try to find a dominating comparison of these guys. 2091 // It can simplify the index computation for Arrays.copyOf 2092 // and similar uses of System.arraycopy. 2093 // First, compute the normalized version of CmpI(x, y). 2094 int cmp_op = Op_CmpI; 2095 Node* xkey = xvalue; 2096 Node* ykey = yvalue; 2097 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey)); 2098 if (ideal_cmpxy->is_Cmp()) { 2099 // E.g., if we have CmpI(length - offset, count), 2100 // it might idealize to CmpI(length, count + offset) 2101 cmp_op = ideal_cmpxy->Opcode(); 2102 xkey = ideal_cmpxy->in(1); 2103 ykey = ideal_cmpxy->in(2); 2104 } 2105 2106 // Start by locating any relevant comparisons. 2107 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey; 2108 Node* cmpxy = NULL; 2109 Node* cmpyx = NULL; 2110 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) { 2111 Node* cmp = start_from->fast_out(k); 2112 if (cmp->outcnt() > 0 && // must have prior uses 2113 cmp->in(0) == NULL && // must be context-independent 2114 cmp->Opcode() == cmp_op) { // right kind of compare 2115 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp; 2116 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp; 2117 } 2118 } 2119 2120 const int NCMPS = 2; 2121 Node* cmps[NCMPS] = { cmpxy, cmpyx }; 2122 int cmpn; 2123 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2124 if (cmps[cmpn] != NULL) break; // find a result 2125 } 2126 if (cmpn < NCMPS) { 2127 // Look for a dominating test that tells us the min and max. 2128 int depth = 0; // Limit search depth for speed 2129 Node* dom = control(); 2130 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) { 2131 if (++depth >= 100) break; 2132 Node* ifproj = dom; 2133 if (!ifproj->is_Proj()) continue; 2134 Node* iff = ifproj->in(0); 2135 if (!iff->is_If()) continue; 2136 Node* bol = iff->in(1); 2137 if (!bol->is_Bool()) continue; 2138 Node* cmp = bol->in(1); 2139 if (cmp == NULL) continue; 2140 for (cmpn = 0; cmpn < NCMPS; cmpn++) 2141 if (cmps[cmpn] == cmp) break; 2142 if (cmpn == NCMPS) continue; 2143 BoolTest::mask btest = bol->as_Bool()->_test._test; 2144 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate(); 2145 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2146 // At this point, we know that 'x btest y' is true. 2147 switch (btest) { 2148 case BoolTest::eq: 2149 // They are proven equal, so we can collapse the min/max. 2150 // Either value is the answer. Choose the simpler. 2151 if (is_simple_name(yvalue) && !is_simple_name(xvalue)) 2152 return yvalue; 2153 return xvalue; 2154 case BoolTest::lt: // x < y 2155 case BoolTest::le: // x <= y 2156 return (want_max ? yvalue : xvalue); 2157 case BoolTest::gt: // x > y 2158 case BoolTest::ge: // x >= y 2159 return (want_max ? xvalue : yvalue); 2160 } 2161 } 2162 } 2163 2164 // We failed to find a dominating test. 2165 // Let's pick a test that might GVN with prior tests. 2166 Node* best_bol = NULL; 2167 BoolTest::mask best_btest = BoolTest::illegal; 2168 for (cmpn = 0; cmpn < NCMPS; cmpn++) { 2169 Node* cmp = cmps[cmpn]; 2170 if (cmp == NULL) continue; 2171 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) { 2172 Node* bol = cmp->fast_out(j); 2173 if (!bol->is_Bool()) continue; 2174 BoolTest::mask btest = bol->as_Bool()->_test._test; 2175 if (btest == BoolTest::eq || btest == BoolTest::ne) continue; 2176 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute(); 2177 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) { 2178 best_bol = bol->as_Bool(); 2179 best_btest = btest; 2180 } 2181 } 2182 } 2183 2184 Node* answer_if_true = NULL; 2185 Node* answer_if_false = NULL; 2186 switch (best_btest) { 2187 default: 2188 if (cmpxy == NULL) 2189 cmpxy = ideal_cmpxy; 2190 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt)); 2191 // and fall through: 2192 case BoolTest::lt: // x < y 2193 case BoolTest::le: // x <= y 2194 answer_if_true = (want_max ? yvalue : xvalue); 2195 answer_if_false = (want_max ? xvalue : yvalue); 2196 break; 2197 case BoolTest::gt: // x > y 2198 case BoolTest::ge: // x >= y 2199 answer_if_true = (want_max ? xvalue : yvalue); 2200 answer_if_false = (want_max ? yvalue : xvalue); 2201 break; 2202 } 2203 2204 jint hi, lo; 2205 if (want_max) { 2206 // We can sharpen the minimum. 2207 hi = MAX2(txvalue->_hi, tyvalue->_hi); 2208 lo = MAX2(txvalue->_lo, tyvalue->_lo); 2209 } else { 2210 // We can sharpen the maximum. 2211 hi = MIN2(txvalue->_hi, tyvalue->_hi); 2212 lo = MIN2(txvalue->_lo, tyvalue->_lo); 2213 } 2214 2215 // Use a flow-free graph structure, to avoid creating excess control edges 2216 // which could hinder other optimizations. 2217 // Since Math.min/max is often used with arraycopy, we want 2218 // tightly_coupled_allocation to be able to see beyond min/max expressions. 2219 Node* cmov = CMoveNode::make(NULL, best_bol, 2220 answer_if_false, answer_if_true, 2221 TypeInt::make(lo, hi, widen)); 2222 2223 return _gvn.transform(cmov); 2224 2225 /* 2226 // This is not as desirable as it may seem, since Min and Max 2227 // nodes do not have a full set of optimizations. 2228 // And they would interfere, anyway, with 'if' optimizations 2229 // and with CMoveI canonical forms. 2230 switch (id) { 2231 case vmIntrinsics::_min: 2232 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break; 2233 case vmIntrinsics::_max: 2234 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break; 2235 default: 2236 ShouldNotReachHere(); 2237 } 2238 */ 2239 } 2240 2241 inline int 2242 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) { 2243 const TypePtr* base_type = TypePtr::NULL_PTR; 2244 if (base != NULL) base_type = _gvn.type(base)->isa_ptr(); 2245 if (base_type == NULL) { 2246 // Unknown type. 2247 return Type::AnyPtr; 2248 } else if (base_type == TypePtr::NULL_PTR) { 2249 // Since this is a NULL+long form, we have to switch to a rawptr. 2250 base = _gvn.transform(new CastX2PNode(offset)); 2251 offset = MakeConX(0); 2252 return Type::RawPtr; 2253 } else if (base_type->base() == Type::RawPtr) { 2254 return Type::RawPtr; 2255 } else if (base_type->isa_oopptr()) { 2256 // Base is never null => always a heap address. 2257 if (base_type->ptr() == TypePtr::NotNull) { 2258 return Type::OopPtr; 2259 } 2260 // Offset is small => always a heap address. 2261 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t(); 2262 if (offset_type != NULL && 2263 base_type->offset() == 0 && // (should always be?) 2264 offset_type->_lo >= 0 && 2265 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) { 2266 return Type::OopPtr; 2267 } 2268 // Otherwise, it might either be oop+off or NULL+addr. 2269 return Type::AnyPtr; 2270 } else { 2271 // No information: 2272 return Type::AnyPtr; 2273 } 2274 } 2275 2276 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) { 2277 int kind = classify_unsafe_addr(base, offset); 2278 if (kind == Type::RawPtr) { 2279 return basic_plus_adr(top(), base, offset); 2280 } else { 2281 return basic_plus_adr(base, offset); 2282 } 2283 } 2284 2285 //--------------------------inline_number_methods----------------------------- 2286 // inline int Integer.numberOfLeadingZeros(int) 2287 // inline int Long.numberOfLeadingZeros(long) 2288 // 2289 // inline int Integer.numberOfTrailingZeros(int) 2290 // inline int Long.numberOfTrailingZeros(long) 2291 // 2292 // inline int Integer.bitCount(int) 2293 // inline int Long.bitCount(long) 2294 // 2295 // inline char Character.reverseBytes(char) 2296 // inline short Short.reverseBytes(short) 2297 // inline int Integer.reverseBytes(int) 2298 // inline long Long.reverseBytes(long) 2299 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) { 2300 Node* arg = argument(0); 2301 Node* n; 2302 switch (id) { 2303 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break; 2304 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break; 2305 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break; 2306 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break; 2307 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break; 2308 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break; 2309 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break; 2310 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break; 2311 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break; 2312 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break; 2313 default: fatal_unexpected_iid(id); break; 2314 } 2315 set_result(_gvn.transform(n)); 2316 return true; 2317 } 2318 2319 //----------------------------inline_unsafe_access---------------------------- 2320 2321 const static BasicType T_ADDRESS_HOLDER = T_LONG; 2322 2323 // Helper that guards and inserts a pre-barrier. 2324 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset, 2325 Node* pre_val, bool need_mem_bar) { 2326 // We could be accessing the referent field of a reference object. If so, when G1 2327 // is enabled, we need to log the value in the referent field in an SATB buffer. 2328 // This routine performs some compile time filters and generates suitable 2329 // runtime filters that guard the pre-barrier code. 2330 // Also add memory barrier for non volatile load from the referent field 2331 // to prevent commoning of loads across safepoint. 2332 if (!UseG1GC && !need_mem_bar) 2333 return; 2334 2335 // Some compile time checks. 2336 2337 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset? 2338 const TypeX* otype = offset->find_intptr_t_type(); 2339 if (otype != NULL && otype->is_con() && 2340 otype->get_con() != java_lang_ref_Reference::referent_offset) { 2341 // Constant offset but not the reference_offset so just return 2342 return; 2343 } 2344 2345 // We only need to generate the runtime guards for instances. 2346 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr(); 2347 if (btype != NULL) { 2348 if (btype->isa_aryptr()) { 2349 // Array type so nothing to do 2350 return; 2351 } 2352 2353 const TypeInstPtr* itype = btype->isa_instptr(); 2354 if (itype != NULL) { 2355 // Can the klass of base_oop be statically determined to be 2356 // _not_ a sub-class of Reference and _not_ Object? 2357 ciKlass* klass = itype->klass(); 2358 if ( klass->is_loaded() && 2359 !klass->is_subtype_of(env()->Reference_klass()) && 2360 !env()->Object_klass()->is_subtype_of(klass)) { 2361 return; 2362 } 2363 } 2364 } 2365 2366 // The compile time filters did not reject base_oop/offset so 2367 // we need to generate the following runtime filters 2368 // 2369 // if (offset == java_lang_ref_Reference::_reference_offset) { 2370 // if (instance_of(base, java.lang.ref.Reference)) { 2371 // pre_barrier(_, pre_val, ...); 2372 // } 2373 // } 2374 2375 float likely = PROB_LIKELY( 0.999); 2376 float unlikely = PROB_UNLIKELY(0.999); 2377 2378 IdealKit ideal(this); 2379 #define __ ideal. 2380 2381 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset); 2382 2383 __ if_then(offset, BoolTest::eq, referent_off, unlikely); { 2384 // Update graphKit memory and control from IdealKit. 2385 sync_kit(ideal); 2386 2387 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass())); 2388 Node* is_instof = gen_instanceof(base_oop, ref_klass_con); 2389 2390 // Update IdealKit memory and control from graphKit. 2391 __ sync_kit(this); 2392 2393 Node* one = __ ConI(1); 2394 // is_instof == 0 if base_oop == NULL 2395 __ if_then(is_instof, BoolTest::eq, one, unlikely); { 2396 2397 // Update graphKit from IdeakKit. 2398 sync_kit(ideal); 2399 2400 // Use the pre-barrier to record the value in the referent field 2401 pre_barrier(false /* do_load */, 2402 __ ctrl(), 2403 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 2404 pre_val /* pre_val */, 2405 T_OBJECT); 2406 if (need_mem_bar) { 2407 // Add memory barrier to prevent commoning reads from this field 2408 // across safepoint since GC can change its value. 2409 insert_mem_bar(Op_MemBarCPUOrder); 2410 } 2411 // Update IdealKit from graphKit. 2412 __ sync_kit(this); 2413 2414 } __ end_if(); // _ref_type != ref_none 2415 } __ end_if(); // offset == referent_offset 2416 2417 // Final sync IdealKit and GraphKit. 2418 final_sync(ideal); 2419 #undef __ 2420 } 2421 2422 2423 // Interpret Unsafe.fieldOffset cookies correctly: 2424 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset); 2425 2426 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) { 2427 // Attempt to infer a sharper value type from the offset and base type. 2428 ciKlass* sharpened_klass = NULL; 2429 2430 // See if it is an instance field, with an object type. 2431 if (alias_type->field() != NULL) { 2432 assert(!is_native_ptr, "native pointer op cannot use a java address"); 2433 if (alias_type->field()->type()->is_klass()) { 2434 sharpened_klass = alias_type->field()->type()->as_klass(); 2435 } 2436 } 2437 2438 // See if it is a narrow oop array. 2439 if (adr_type->isa_aryptr()) { 2440 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) { 2441 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr(); 2442 if (elem_type != NULL) { 2443 sharpened_klass = elem_type->klass(); 2444 } 2445 } 2446 } 2447 2448 // The sharpened class might be unloaded if there is no class loader 2449 // contraint in place. 2450 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) { 2451 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass); 2452 2453 #ifndef PRODUCT 2454 if (C->print_intrinsics() || C->print_inlining()) { 2455 tty->print(" from base type: "); adr_type->dump(); 2456 tty->print(" sharpened value: "); tjp->dump(); 2457 } 2458 #endif 2459 // Sharpen the value type. 2460 return tjp; 2461 } 2462 return NULL; 2463 } 2464 2465 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) { 2466 if (callee()->is_static()) return false; // caller must have the capability! 2467 2468 #ifndef PRODUCT 2469 { 2470 ResourceMark rm; 2471 // Check the signatures. 2472 ciSignature* sig = callee()->signature(); 2473 #ifdef ASSERT 2474 if (!is_store) { 2475 // Object getObject(Object base, int/long offset), etc. 2476 BasicType rtype = sig->return_type()->basic_type(); 2477 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name()) 2478 rtype = T_ADDRESS; // it is really a C void* 2479 assert(rtype == type, "getter must return the expected value"); 2480 if (!is_native_ptr) { 2481 assert(sig->count() == 2, "oop getter has 2 arguments"); 2482 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object"); 2483 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct"); 2484 } else { 2485 assert(sig->count() == 1, "native getter has 1 argument"); 2486 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long"); 2487 } 2488 } else { 2489 // void putObject(Object base, int/long offset, Object x), etc. 2490 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value"); 2491 if (!is_native_ptr) { 2492 assert(sig->count() == 3, "oop putter has 3 arguments"); 2493 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object"); 2494 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct"); 2495 } else { 2496 assert(sig->count() == 2, "native putter has 2 arguments"); 2497 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long"); 2498 } 2499 BasicType vtype = sig->type_at(sig->count()-1)->basic_type(); 2500 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name()) 2501 vtype = T_ADDRESS; // it is really a C void* 2502 assert(vtype == type, "putter must accept the expected value"); 2503 } 2504 #endif // ASSERT 2505 } 2506 #endif //PRODUCT 2507 2508 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2509 2510 Node* receiver = argument(0); // type: oop 2511 2512 // Build address expression. See the code in inline_unsafe_prefetch. 2513 Node* adr; 2514 Node* heap_base_oop = top(); 2515 Node* offset = top(); 2516 Node* val; 2517 2518 if (!is_native_ptr) { 2519 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2520 Node* base = argument(1); // type: oop 2521 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2522 offset = argument(2); // type: long 2523 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2524 // to be plain byte offsets, which are also the same as those accepted 2525 // by oopDesc::field_base. 2526 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2527 "fieldOffset must be byte-scaled"); 2528 // 32-bit machines ignore the high half! 2529 offset = ConvL2X(offset); 2530 adr = make_unsafe_address(base, offset); 2531 heap_base_oop = base; 2532 val = is_store ? argument(4) : NULL; 2533 } else { 2534 Node* ptr = argument(1); // type: long 2535 ptr = ConvL2X(ptr); // adjust Java long to machine word 2536 adr = make_unsafe_address(NULL, ptr); 2537 val = is_store ? argument(3) : NULL; 2538 } 2539 2540 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2541 2542 // First guess at the value type. 2543 const Type *value_type = Type::get_const_basic_type(type); 2544 2545 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM, 2546 // there was not enough information to nail it down. 2547 Compile::AliasType* alias_type = C->alias_type(adr_type); 2548 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2549 2550 // We will need memory barriers unless we can determine a unique 2551 // alias category for this reference. (Note: If for some reason 2552 // the barriers get omitted and the unsafe reference begins to "pollute" 2553 // the alias analysis of the rest of the graph, either Compile::can_alias 2554 // or Compile::must_alias will throw a diagnostic assert.) 2555 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM); 2556 2557 // If we are reading the value of the referent field of a Reference 2558 // object (either by using Unsafe directly or through reflection) 2559 // then, if G1 is enabled, we need to record the referent in an 2560 // SATB log buffer using the pre-barrier mechanism. 2561 // Also we need to add memory barrier to prevent commoning reads 2562 // from this field across safepoint since GC can change its value. 2563 bool need_read_barrier = !is_native_ptr && !is_store && 2564 offset != top() && heap_base_oop != top(); 2565 2566 if (!is_store && type == T_OBJECT) { 2567 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr); 2568 if (tjp != NULL) { 2569 value_type = tjp; 2570 } 2571 } 2572 2573 receiver = null_check(receiver); 2574 if (stopped()) { 2575 return true; 2576 } 2577 // Heap pointers get a null-check from the interpreter, 2578 // as a courtesy. However, this is not guaranteed by Unsafe, 2579 // and it is not possible to fully distinguish unintended nulls 2580 // from intended ones in this API. 2581 2582 if (is_volatile) { 2583 // We need to emit leading and trailing CPU membars (see below) in 2584 // addition to memory membars when is_volatile. This is a little 2585 // too strong, but avoids the need to insert per-alias-type 2586 // volatile membars (for stores; compare Parse::do_put_xxx), which 2587 // we cannot do effectively here because we probably only have a 2588 // rough approximation of type. 2589 need_mem_bar = true; 2590 // For Stores, place a memory ordering barrier now. 2591 #ifndef AARCH64 2592 if (is_store) { 2593 insert_mem_bar(Op_MemBarRelease); 2594 } else { 2595 if (support_IRIW_for_not_multiple_copy_atomic_cpu) { 2596 insert_mem_bar(Op_MemBarVolatile); 2597 } 2598 } 2599 #endif // AARCH64 2600 } 2601 2602 // Memory barrier to prevent normal and 'unsafe' accesses from 2603 // bypassing each other. Happens after null checks, so the 2604 // exception paths do not take memory state from the memory barrier, 2605 // so there's no problems making a strong assert about mixing users 2606 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar 2607 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl. 2608 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2609 2610 if (!is_store) { 2611 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered; 2612 Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile); 2613 // load value 2614 switch (type) { 2615 case T_BOOLEAN: 2616 case T_CHAR: 2617 case T_BYTE: 2618 case T_SHORT: 2619 case T_INT: 2620 case T_LONG: 2621 case T_FLOAT: 2622 case T_DOUBLE: 2623 break; 2624 case T_OBJECT: 2625 if (need_read_barrier) { 2626 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar)); 2627 } 2628 break; 2629 case T_ADDRESS: 2630 // Cast to an int type. 2631 p = _gvn.transform(new CastP2XNode(NULL, p)); 2632 p = ConvX2UL(p); 2633 break; 2634 default: 2635 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2636 break; 2637 } 2638 // The load node has the control of the preceding MemBarCPUOrder. All 2639 // following nodes will have the control of the MemBarCPUOrder inserted at 2640 // the end of this method. So, pushing the load onto the stack at a later 2641 // point is fine. 2642 set_result(p); 2643 } else { 2644 // place effect of store into memory 2645 switch (type) { 2646 case T_DOUBLE: 2647 val = dstore_rounding(val); 2648 break; 2649 case T_ADDRESS: 2650 // Repackage the long as a pointer. 2651 val = ConvL2X(val); 2652 val = _gvn.transform(new CastX2PNode(val)); 2653 break; 2654 } 2655 2656 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered; 2657 if (type != T_OBJECT ) { 2658 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile); 2659 } else { 2660 // Possibly an oop being stored to Java heap or native memory 2661 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) { 2662 // oop to Java heap. 2663 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2664 } else { 2665 // We can't tell at compile time if we are storing in the Java heap or outside 2666 // of it. So we need to emit code to conditionally do the proper type of 2667 // store. 2668 2669 IdealKit ideal(this); 2670 #define __ ideal. 2671 // QQQ who knows what probability is here?? 2672 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); { 2673 // Sync IdealKit and graphKit. 2674 sync_kit(ideal); 2675 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo); 2676 // Update IdealKit memory. 2677 __ sync_kit(this); 2678 } __ else_(); { 2679 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile); 2680 } __ end_if(); 2681 // Final sync IdealKit and GraphKit. 2682 final_sync(ideal); 2683 #undef __ 2684 } 2685 } 2686 } 2687 2688 #ifndef AARCH64 2689 if (is_volatile) { 2690 if (!is_store) { 2691 insert_mem_bar(Op_MemBarAcquire); 2692 } else { 2693 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) { 2694 insert_mem_bar(Op_MemBarVolatile); 2695 } 2696 } 2697 } 2698 #endif // AARCH64 2699 2700 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder); 2701 2702 return true; 2703 } 2704 2705 //----------------------------inline_unsafe_prefetch---------------------------- 2706 2707 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) { 2708 #ifndef PRODUCT 2709 { 2710 ResourceMark rm; 2711 // Check the signatures. 2712 ciSignature* sig = callee()->signature(); 2713 #ifdef ASSERT 2714 // Object getObject(Object base, int/long offset), etc. 2715 BasicType rtype = sig->return_type()->basic_type(); 2716 if (!is_native_ptr) { 2717 assert(sig->count() == 2, "oop prefetch has 2 arguments"); 2718 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object"); 2719 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct"); 2720 } else { 2721 assert(sig->count() == 1, "native prefetch has 1 argument"); 2722 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long"); 2723 } 2724 #endif // ASSERT 2725 } 2726 #endif // !PRODUCT 2727 2728 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2729 2730 const int idx = is_static ? 0 : 1; 2731 if (!is_static) { 2732 null_check_receiver(); 2733 if (stopped()) { 2734 return true; 2735 } 2736 } 2737 2738 // Build address expression. See the code in inline_unsafe_access. 2739 Node *adr; 2740 if (!is_native_ptr) { 2741 // The base is either a Java object or a value produced by Unsafe.staticFieldBase 2742 Node* base = argument(idx + 0); // type: oop 2743 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset 2744 Node* offset = argument(idx + 1); // type: long 2745 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2746 // to be plain byte offsets, which are also the same as those accepted 2747 // by oopDesc::field_base. 2748 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 2749 "fieldOffset must be byte-scaled"); 2750 // 32-bit machines ignore the high half! 2751 offset = ConvL2X(offset); 2752 adr = make_unsafe_address(base, offset); 2753 } else { 2754 Node* ptr = argument(idx + 0); // type: long 2755 ptr = ConvL2X(ptr); // adjust Java long to machine word 2756 adr = make_unsafe_address(NULL, ptr); 2757 } 2758 2759 // Generate the read or write prefetch 2760 Node *prefetch; 2761 if (is_store) { 2762 prefetch = new PrefetchWriteNode(i_o(), adr); 2763 } else { 2764 prefetch = new PrefetchReadNode(i_o(), adr); 2765 } 2766 prefetch->init_req(0, control()); 2767 set_i_o(_gvn.transform(prefetch)); 2768 2769 return true; 2770 } 2771 2772 //----------------------------inline_unsafe_load_store---------------------------- 2773 // This method serves a couple of different customers (depending on LoadStoreKind): 2774 // 2775 // LS_cmpxchg: 2776 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x); 2777 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x); 2778 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x); 2779 // 2780 // LS_xadd: 2781 // public int getAndAddInt( Object o, long offset, int delta) 2782 // public long getAndAddLong(Object o, long offset, long delta) 2783 // 2784 // LS_xchg: 2785 // int getAndSet(Object o, long offset, int newValue) 2786 // long getAndSet(Object o, long offset, long newValue) 2787 // Object getAndSet(Object o, long offset, Object newValue) 2788 // 2789 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) { 2790 // This basic scheme here is the same as inline_unsafe_access, but 2791 // differs in enough details that combining them would make the code 2792 // overly confusing. (This is a true fact! I originally combined 2793 // them, but even I was confused by it!) As much code/comments as 2794 // possible are retained from inline_unsafe_access though to make 2795 // the correspondences clearer. - dl 2796 2797 if (callee()->is_static()) return false; // caller must have the capability! 2798 2799 #ifndef PRODUCT 2800 BasicType rtype; 2801 { 2802 ResourceMark rm; 2803 // Check the signatures. 2804 ciSignature* sig = callee()->signature(); 2805 rtype = sig->return_type()->basic_type(); 2806 if (kind == LS_xadd || kind == LS_xchg) { 2807 // Check the signatures. 2808 #ifdef ASSERT 2809 assert(rtype == type, "get and set must return the expected type"); 2810 assert(sig->count() == 3, "get and set has 3 arguments"); 2811 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object"); 2812 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long"); 2813 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta"); 2814 #endif // ASSERT 2815 } else if (kind == LS_cmpxchg) { 2816 // Check the signatures. 2817 #ifdef ASSERT 2818 assert(rtype == T_BOOLEAN, "CAS must return boolean"); 2819 assert(sig->count() == 4, "CAS has 4 arguments"); 2820 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object"); 2821 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long"); 2822 #endif // ASSERT 2823 } else { 2824 ShouldNotReachHere(); 2825 } 2826 } 2827 #endif //PRODUCT 2828 2829 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 2830 2831 // Get arguments: 2832 Node* receiver = NULL; 2833 Node* base = NULL; 2834 Node* offset = NULL; 2835 Node* oldval = NULL; 2836 Node* newval = NULL; 2837 if (kind == LS_cmpxchg) { 2838 const bool two_slot_type = type2size[type] == 2; 2839 receiver = argument(0); // type: oop 2840 base = argument(1); // type: oop 2841 offset = argument(2); // type: long 2842 oldval = argument(4); // type: oop, int, or long 2843 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long 2844 } else if (kind == LS_xadd || kind == LS_xchg){ 2845 receiver = argument(0); // type: oop 2846 base = argument(1); // type: oop 2847 offset = argument(2); // type: long 2848 oldval = NULL; 2849 newval = argument(4); // type: oop, int, or long 2850 } 2851 2852 // Null check receiver. 2853 receiver = null_check(receiver); 2854 if (stopped()) { 2855 return true; 2856 } 2857 2858 // Build field offset expression. 2859 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset 2860 // to be plain byte offsets, which are also the same as those accepted 2861 // by oopDesc::field_base. 2862 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 2863 // 32-bit machines ignore the high half of long offsets 2864 offset = ConvL2X(offset); 2865 Node* adr = make_unsafe_address(base, offset); 2866 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 2867 2868 // For CAS, unlike inline_unsafe_access, there seems no point in 2869 // trying to refine types. Just use the coarse types here. 2870 const Type *value_type = Type::get_const_basic_type(type); 2871 Compile::AliasType* alias_type = C->alias_type(adr_type); 2872 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here"); 2873 2874 if (kind == LS_xchg && type == T_OBJECT) { 2875 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type); 2876 if (tjp != NULL) { 2877 value_type = tjp; 2878 } 2879 } 2880 2881 int alias_idx = C->get_alias_index(adr_type); 2882 2883 // Memory-model-wise, a LoadStore acts like a little synchronized 2884 // block, so needs barriers on each side. These don't translate 2885 // into actual barriers on most machines, but we still need rest of 2886 // compiler to respect ordering. 2887 2888 insert_mem_bar(Op_MemBarRelease); 2889 insert_mem_bar(Op_MemBarCPUOrder); 2890 2891 // 4984716: MemBars must be inserted before this 2892 // memory node in order to avoid a false 2893 // dependency which will confuse the scheduler. 2894 Node *mem = memory(alias_idx); 2895 2896 // For now, we handle only those cases that actually exist: ints, 2897 // longs, and Object. Adding others should be straightforward. 2898 Node* load_store; 2899 switch(type) { 2900 case T_INT: 2901 if (kind == LS_xadd) { 2902 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type)); 2903 } else if (kind == LS_xchg) { 2904 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type)); 2905 } else if (kind == LS_cmpxchg) { 2906 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval)); 2907 } else { 2908 ShouldNotReachHere(); 2909 } 2910 break; 2911 case T_LONG: 2912 if (kind == LS_xadd) { 2913 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type)); 2914 } else if (kind == LS_xchg) { 2915 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type)); 2916 } else if (kind == LS_cmpxchg) { 2917 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval)); 2918 } else { 2919 ShouldNotReachHere(); 2920 } 2921 break; 2922 case T_OBJECT: 2923 // Transformation of a value which could be NULL pointer (CastPP #NULL) 2924 // could be delayed during Parse (for example, in adjust_map_after_if()). 2925 // Execute transformation here to avoid barrier generation in such case. 2926 if (_gvn.type(newval) == TypePtr::NULL_PTR) 2927 newval = _gvn.makecon(TypePtr::NULL_PTR); 2928 2929 // Reference stores need a store barrier. 2930 if (kind == LS_xchg) { 2931 // If pre-barrier must execute before the oop store, old value will require do_load here. 2932 if (!can_move_pre_barrier()) { 2933 pre_barrier(true /* do_load*/, 2934 control(), base, adr, alias_idx, newval, value_type->make_oopptr(), 2935 NULL /* pre_val*/, 2936 T_OBJECT); 2937 } // Else move pre_barrier to use load_store value, see below. 2938 } else if (kind == LS_cmpxchg) { 2939 // Same as for newval above: 2940 if (_gvn.type(oldval) == TypePtr::NULL_PTR) { 2941 oldval = _gvn.makecon(TypePtr::NULL_PTR); 2942 } 2943 // The only known value which might get overwritten is oldval. 2944 pre_barrier(false /* do_load */, 2945 control(), NULL, NULL, max_juint, NULL, NULL, 2946 oldval /* pre_val */, 2947 T_OBJECT); 2948 } else { 2949 ShouldNotReachHere(); 2950 } 2951 2952 #ifdef _LP64 2953 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2954 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop())); 2955 if (kind == LS_xchg) { 2956 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, 2957 newval_enc, adr_type, value_type->make_narrowoop())); 2958 } else { 2959 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2960 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop())); 2961 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, 2962 newval_enc, oldval_enc)); 2963 } 2964 } else 2965 #endif 2966 { 2967 if (kind == LS_xchg) { 2968 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr())); 2969 } else { 2970 assert(kind == LS_cmpxchg, "wrong LoadStore operation"); 2971 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval)); 2972 } 2973 } 2974 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true); 2975 break; 2976 default: 2977 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type))); 2978 break; 2979 } 2980 2981 // SCMemProjNodes represent the memory state of a LoadStore. Their 2982 // main role is to prevent LoadStore nodes from being optimized away 2983 // when their results aren't used. 2984 Node* proj = _gvn.transform(new SCMemProjNode(load_store)); 2985 set_memory(proj, alias_idx); 2986 2987 if (type == T_OBJECT && kind == LS_xchg) { 2988 #ifdef _LP64 2989 if (adr->bottom_type()->is_ptr_to_narrowoop()) { 2990 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type())); 2991 } 2992 #endif 2993 if (can_move_pre_barrier()) { 2994 // Don't need to load pre_val. The old value is returned by load_store. 2995 // The pre_barrier can execute after the xchg as long as no safepoint 2996 // gets inserted between them. 2997 pre_barrier(false /* do_load */, 2998 control(), NULL, NULL, max_juint, NULL, NULL, 2999 load_store /* pre_val */, 3000 T_OBJECT); 3001 } 3002 } 3003 3004 // Add the trailing membar surrounding the access 3005 insert_mem_bar(Op_MemBarCPUOrder); 3006 insert_mem_bar(Op_MemBarAcquire); 3007 3008 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match"); 3009 set_result(load_store); 3010 return true; 3011 } 3012 3013 //----------------------------inline_unsafe_ordered_store---------------------- 3014 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x); 3015 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x); 3016 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x); 3017 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) { 3018 // This is another variant of inline_unsafe_access, differing in 3019 // that it always issues store-store ("release") barrier and ensures 3020 // store-atomicity (which only matters for "long"). 3021 3022 if (callee()->is_static()) return false; // caller must have the capability! 3023 3024 #ifndef PRODUCT 3025 { 3026 ResourceMark rm; 3027 // Check the signatures. 3028 ciSignature* sig = callee()->signature(); 3029 #ifdef ASSERT 3030 BasicType rtype = sig->return_type()->basic_type(); 3031 assert(rtype == T_VOID, "must return void"); 3032 assert(sig->count() == 3, "has 3 arguments"); 3033 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object"); 3034 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long"); 3035 #endif // ASSERT 3036 } 3037 #endif //PRODUCT 3038 3039 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 3040 3041 // Get arguments: 3042 Node* receiver = argument(0); // type: oop 3043 Node* base = argument(1); // type: oop 3044 Node* offset = argument(2); // type: long 3045 Node* val = argument(4); // type: oop, int, or long 3046 3047 // Null check receiver. 3048 receiver = null_check(receiver); 3049 if (stopped()) { 3050 return true; 3051 } 3052 3053 // Build field offset expression. 3054 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled"); 3055 // 32-bit machines ignore the high half of long offsets 3056 offset = ConvL2X(offset); 3057 Node* adr = make_unsafe_address(base, offset); 3058 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr(); 3059 const Type *value_type = Type::get_const_basic_type(type); 3060 Compile::AliasType* alias_type = C->alias_type(adr_type); 3061 3062 insert_mem_bar(Op_MemBarRelease); 3063 insert_mem_bar(Op_MemBarCPUOrder); 3064 // Ensure that the store is atomic for longs: 3065 const bool require_atomic_access = true; 3066 Node* store; 3067 if (type == T_OBJECT) // reference stores need a store barrier. 3068 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release); 3069 else { 3070 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access); 3071 } 3072 insert_mem_bar(Op_MemBarCPUOrder); 3073 return true; 3074 } 3075 3076 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) { 3077 // Regardless of form, don't allow previous ld/st to move down, 3078 // then issue acquire, release, or volatile mem_bar. 3079 insert_mem_bar(Op_MemBarCPUOrder); 3080 switch(id) { 3081 case vmIntrinsics::_loadFence: 3082 insert_mem_bar(Op_LoadFence); 3083 return true; 3084 case vmIntrinsics::_storeFence: 3085 insert_mem_bar(Op_StoreFence); 3086 return true; 3087 case vmIntrinsics::_fullFence: 3088 insert_mem_bar(Op_MemBarVolatile); 3089 return true; 3090 default: 3091 fatal_unexpected_iid(id); 3092 return false; 3093 } 3094 } 3095 3096 bool LibraryCallKit::klass_needs_init_guard(Node* kls) { 3097 if (!kls->is_Con()) { 3098 return true; 3099 } 3100 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr(); 3101 if (klsptr == NULL) { 3102 return true; 3103 } 3104 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass(); 3105 // don't need a guard for a klass that is already initialized 3106 return !ik->is_initialized(); 3107 } 3108 3109 //----------------------------inline_unsafe_allocate--------------------------- 3110 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls); 3111 bool LibraryCallKit::inline_unsafe_allocate() { 3112 if (callee()->is_static()) return false; // caller must have the capability! 3113 3114 null_check_receiver(); // null-check, then ignore 3115 Node* cls = null_check(argument(1)); 3116 if (stopped()) return true; 3117 3118 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3119 kls = null_check(kls); 3120 if (stopped()) return true; // argument was like int.class 3121 3122 Node* test = NULL; 3123 if (LibraryCallKit::klass_needs_init_guard(kls)) { 3124 // Note: The argument might still be an illegal value like 3125 // Serializable.class or Object[].class. The runtime will handle it. 3126 // But we must make an explicit check for initialization. 3127 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset())); 3128 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler 3129 // can generate code to load it as unsigned byte. 3130 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered); 3131 Node* bits = intcon(InstanceKlass::fully_initialized); 3132 test = _gvn.transform(new SubINode(inst, bits)); 3133 // The 'test' is non-zero if we need to take a slow path. 3134 } 3135 3136 Node* obj = new_instance(kls, test); 3137 set_result(obj); 3138 return true; 3139 } 3140 3141 #ifdef TRACE_HAVE_INTRINSICS 3142 /* 3143 * oop -> myklass 3144 * myklass->trace_id |= USED 3145 * return myklass->trace_id & ~0x3 3146 */ 3147 bool LibraryCallKit::inline_native_classID() { 3148 null_check_receiver(); // null-check, then ignore 3149 Node* cls = null_check(argument(1), T_OBJECT); 3150 Node* kls = load_klass_from_mirror(cls, false, NULL, 0); 3151 kls = null_check(kls, T_OBJECT); 3152 ByteSize offset = TRACE_ID_OFFSET; 3153 Node* insp = basic_plus_adr(kls, in_bytes(offset)); 3154 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered); 3155 Node* bits = longcon(~0x03l); // ignore bit 0 & 1 3156 Node* andl = _gvn.transform(new AndLNode(tvalue, bits)); 3157 Node* clsused = longcon(0x01l); // set the class bit 3158 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused)); 3159 3160 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr(); 3161 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered); 3162 set_result(andl); 3163 return true; 3164 } 3165 3166 bool LibraryCallKit::inline_native_threadID() { 3167 Node* tls_ptr = NULL; 3168 Node* cur_thr = generate_current_thread(tls_ptr); 3169 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3170 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3171 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset())); 3172 3173 Node* threadid = NULL; 3174 size_t thread_id_size = OSThread::thread_id_size(); 3175 if (thread_id_size == (size_t) BytesPerLong) { 3176 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered)); 3177 } else if (thread_id_size == (size_t) BytesPerInt) { 3178 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered); 3179 } else { 3180 ShouldNotReachHere(); 3181 } 3182 set_result(threadid); 3183 return true; 3184 } 3185 #endif 3186 3187 //------------------------inline_native_time_funcs-------------- 3188 // inline code for System.currentTimeMillis() and System.nanoTime() 3189 // these have the same type and signature 3190 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) { 3191 const TypeFunc* tf = OptoRuntime::void_long_Type(); 3192 const TypePtr* no_memory_effects = NULL; 3193 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects); 3194 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0)); 3195 #ifdef ASSERT 3196 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1)); 3197 assert(value_top == top(), "second value must be top"); 3198 #endif 3199 set_result(value); 3200 return true; 3201 } 3202 3203 //------------------------inline_native_currentThread------------------ 3204 bool LibraryCallKit::inline_native_currentThread() { 3205 Node* junk = NULL; 3206 set_result(generate_current_thread(junk)); 3207 return true; 3208 } 3209 3210 //------------------------inline_native_isInterrupted------------------ 3211 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted); 3212 bool LibraryCallKit::inline_native_isInterrupted() { 3213 // Add a fast path to t.isInterrupted(clear_int): 3214 // (t == Thread.current() && 3215 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int))) 3216 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int) 3217 // So, in the common case that the interrupt bit is false, 3218 // we avoid making a call into the VM. Even if the interrupt bit 3219 // is true, if the clear_int argument is false, we avoid the VM call. 3220 // However, if the receiver is not currentThread, we must call the VM, 3221 // because there must be some locking done around the operation. 3222 3223 // We only go to the fast case code if we pass two guards. 3224 // Paths which do not pass are accumulated in the slow_region. 3225 3226 enum { 3227 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted 3228 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int 3229 slow_result_path = 3, // slow path: t.isInterrupted(clear_int) 3230 PATH_LIMIT 3231 }; 3232 3233 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag 3234 // out of the function. 3235 insert_mem_bar(Op_MemBarCPUOrder); 3236 3237 RegionNode* result_rgn = new RegionNode(PATH_LIMIT); 3238 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL); 3239 3240 RegionNode* slow_region = new RegionNode(1); 3241 record_for_igvn(slow_region); 3242 3243 // (a) Receiving thread must be the current thread. 3244 Node* rec_thr = argument(0); 3245 Node* tls_ptr = NULL; 3246 Node* cur_thr = generate_current_thread(tls_ptr); 3247 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr)); 3248 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne)); 3249 3250 generate_slow_guard(bol_thr, slow_region); 3251 3252 // (b) Interrupt bit on TLS must be false. 3253 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset())); 3254 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3255 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset())); 3256 3257 // Set the control input on the field _interrupted read to prevent it floating up. 3258 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered); 3259 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0))); 3260 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne)); 3261 3262 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN); 3263 3264 // First fast path: if (!TLS._interrupted) return false; 3265 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit)); 3266 result_rgn->init_req(no_int_result_path, false_bit); 3267 result_val->init_req(no_int_result_path, intcon(0)); 3268 3269 // drop through to next case 3270 set_control( _gvn.transform(new IfTrueNode(iff_bit))); 3271 3272 #ifndef TARGET_OS_FAMILY_windows 3273 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path. 3274 Node* clr_arg = argument(1); 3275 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0))); 3276 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne)); 3277 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN); 3278 3279 // Second fast path: ... else if (!clear_int) return true; 3280 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg)); 3281 result_rgn->init_req(no_clear_result_path, false_arg); 3282 result_val->init_req(no_clear_result_path, intcon(1)); 3283 3284 // drop through to next case 3285 set_control( _gvn.transform(new IfTrueNode(iff_arg))); 3286 #else 3287 // To return true on Windows you must read the _interrupted field 3288 // and check the the event state i.e. take the slow path. 3289 #endif // TARGET_OS_FAMILY_windows 3290 3291 // (d) Otherwise, go to the slow path. 3292 slow_region->add_req(control()); 3293 set_control( _gvn.transform(slow_region)); 3294 3295 if (stopped()) { 3296 // There is no slow path. 3297 result_rgn->init_req(slow_result_path, top()); 3298 result_val->init_req(slow_result_path, top()); 3299 } else { 3300 // non-virtual because it is a private non-static 3301 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted); 3302 3303 Node* slow_val = set_results_for_java_call(slow_call); 3304 // this->control() comes from set_results_for_java_call 3305 3306 Node* fast_io = slow_call->in(TypeFunc::I_O); 3307 Node* fast_mem = slow_call->in(TypeFunc::Memory); 3308 3309 // These two phis are pre-filled with copies of of the fast IO and Memory 3310 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM); 3311 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO); 3312 3313 result_rgn->init_req(slow_result_path, control()); 3314 result_io ->init_req(slow_result_path, i_o()); 3315 result_mem->init_req(slow_result_path, reset_memory()); 3316 result_val->init_req(slow_result_path, slow_val); 3317 3318 set_all_memory(_gvn.transform(result_mem)); 3319 set_i_o( _gvn.transform(result_io)); 3320 } 3321 3322 C->set_has_split_ifs(true); // Has chance for split-if optimization 3323 set_result(result_rgn, result_val); 3324 return true; 3325 } 3326 3327 //---------------------------load_mirror_from_klass---------------------------- 3328 // Given a klass oop, load its java mirror (a java.lang.Class oop). 3329 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) { 3330 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset())); 3331 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered); 3332 } 3333 3334 //-----------------------load_klass_from_mirror_common------------------------- 3335 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop. 3336 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE), 3337 // and branch to the given path on the region. 3338 // If never_see_null, take an uncommon trap on null, so we can optimistically 3339 // compile for the non-null case. 3340 // If the region is NULL, force never_see_null = true. 3341 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror, 3342 bool never_see_null, 3343 RegionNode* region, 3344 int null_path, 3345 int offset) { 3346 if (region == NULL) never_see_null = true; 3347 Node* p = basic_plus_adr(mirror, offset); 3348 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3349 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type)); 3350 Node* null_ctl = top(); 3351 kls = null_check_oop(kls, &null_ctl, never_see_null); 3352 if (region != NULL) { 3353 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class). 3354 region->init_req(null_path, null_ctl); 3355 } else { 3356 assert(null_ctl == top(), "no loose ends"); 3357 } 3358 return kls; 3359 } 3360 3361 //--------------------(inline_native_Class_query helpers)--------------------- 3362 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER. 3363 // Fall through if (mods & mask) == bits, take the guard otherwise. 3364 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) { 3365 // Branch around if the given klass has the given modifier bit set. 3366 // Like generate_guard, adds a new path onto the region. 3367 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3368 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered); 3369 Node* mask = intcon(modifier_mask); 3370 Node* bits = intcon(modifier_bits); 3371 Node* mbit = _gvn.transform(new AndINode(mods, mask)); 3372 Node* cmp = _gvn.transform(new CmpINode(mbit, bits)); 3373 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne)); 3374 return generate_fair_guard(bol, region); 3375 } 3376 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) { 3377 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region); 3378 } 3379 3380 //-------------------------inline_native_Class_query------------------- 3381 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) { 3382 const Type* return_type = TypeInt::BOOL; 3383 Node* prim_return_value = top(); // what happens if it's a primitive class? 3384 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3385 bool expect_prim = false; // most of these guys expect to work on refs 3386 3387 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT }; 3388 3389 Node* mirror = argument(0); 3390 Node* obj = top(); 3391 3392 switch (id) { 3393 case vmIntrinsics::_isInstance: 3394 // nothing is an instance of a primitive type 3395 prim_return_value = intcon(0); 3396 obj = argument(1); 3397 break; 3398 case vmIntrinsics::_getModifiers: 3399 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3400 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line"); 3401 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin); 3402 break; 3403 case vmIntrinsics::_isInterface: 3404 prim_return_value = intcon(0); 3405 break; 3406 case vmIntrinsics::_isArray: 3407 prim_return_value = intcon(0); 3408 expect_prim = true; // cf. ObjectStreamClass.getClassSignature 3409 break; 3410 case vmIntrinsics::_isPrimitive: 3411 prim_return_value = intcon(1); 3412 expect_prim = true; // obviously 3413 break; 3414 case vmIntrinsics::_getSuperclass: 3415 prim_return_value = null(); 3416 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR); 3417 break; 3418 case vmIntrinsics::_getClassAccessFlags: 3419 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC); 3420 return_type = TypeInt::INT; // not bool! 6297094 3421 break; 3422 default: 3423 fatal_unexpected_iid(id); 3424 break; 3425 } 3426 3427 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr(); 3428 if (mirror_con == NULL) return false; // cannot happen? 3429 3430 #ifndef PRODUCT 3431 if (C->print_intrinsics() || C->print_inlining()) { 3432 ciType* k = mirror_con->java_mirror_type(); 3433 if (k) { 3434 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id())); 3435 k->print_name(); 3436 tty->cr(); 3437 } 3438 } 3439 #endif 3440 3441 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive). 3442 RegionNode* region = new RegionNode(PATH_LIMIT); 3443 record_for_igvn(region); 3444 PhiNode* phi = new PhiNode(region, return_type); 3445 3446 // The mirror will never be null of Reflection.getClassAccessFlags, however 3447 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE 3448 // if it is. See bug 4774291. 3449 3450 // For Reflection.getClassAccessFlags(), the null check occurs in 3451 // the wrong place; see inline_unsafe_access(), above, for a similar 3452 // situation. 3453 mirror = null_check(mirror); 3454 // If mirror or obj is dead, only null-path is taken. 3455 if (stopped()) return true; 3456 3457 if (expect_prim) never_see_null = false; // expect nulls (meaning prims) 3458 3459 // Now load the mirror's klass metaobject, and null-check it. 3460 // Side-effects region with the control path if the klass is null. 3461 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path); 3462 // If kls is null, we have a primitive mirror. 3463 phi->init_req(_prim_path, prim_return_value); 3464 if (stopped()) { set_result(region, phi); return true; } 3465 bool safe_for_replace = (region->in(_prim_path) == top()); 3466 3467 Node* p; // handy temp 3468 Node* null_ctl; 3469 3470 // Now that we have the non-null klass, we can perform the real query. 3471 // For constant classes, the query will constant-fold in LoadNode::Value. 3472 Node* query_value = top(); 3473 switch (id) { 3474 case vmIntrinsics::_isInstance: 3475 // nothing is an instance of a primitive type 3476 query_value = gen_instanceof(obj, kls, safe_for_replace); 3477 break; 3478 3479 case vmIntrinsics::_getModifiers: 3480 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset())); 3481 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3482 break; 3483 3484 case vmIntrinsics::_isInterface: 3485 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3486 if (generate_interface_guard(kls, region) != NULL) 3487 // A guard was added. If the guard is taken, it was an interface. 3488 phi->add_req(intcon(1)); 3489 // If we fall through, it's a plain class. 3490 query_value = intcon(0); 3491 break; 3492 3493 case vmIntrinsics::_isArray: 3494 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.) 3495 if (generate_array_guard(kls, region) != NULL) 3496 // A guard was added. If the guard is taken, it was an array. 3497 phi->add_req(intcon(1)); 3498 // If we fall through, it's a plain class. 3499 query_value = intcon(0); 3500 break; 3501 3502 case vmIntrinsics::_isPrimitive: 3503 query_value = intcon(0); // "normal" path produces false 3504 break; 3505 3506 case vmIntrinsics::_getSuperclass: 3507 // The rules here are somewhat unfortunate, but we can still do better 3508 // with random logic than with a JNI call. 3509 // Interfaces store null or Object as _super, but must report null. 3510 // Arrays store an intermediate super as _super, but must report Object. 3511 // Other types can report the actual _super. 3512 // (To verify this code sequence, check the asserts in JVM_IsInterface.) 3513 if (generate_interface_guard(kls, region) != NULL) 3514 // A guard was added. If the guard is taken, it was an interface. 3515 phi->add_req(null()); 3516 if (generate_array_guard(kls, region) != NULL) 3517 // A guard was added. If the guard is taken, it was an array. 3518 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror()))); 3519 // If we fall through, it's a plain class. Get its _super. 3520 p = basic_plus_adr(kls, in_bytes(Klass::super_offset())); 3521 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL)); 3522 null_ctl = top(); 3523 kls = null_check_oop(kls, &null_ctl); 3524 if (null_ctl != top()) { 3525 // If the guard is taken, Object.superClass is null (both klass and mirror). 3526 region->add_req(null_ctl); 3527 phi ->add_req(null()); 3528 } 3529 if (!stopped()) { 3530 query_value = load_mirror_from_klass(kls); 3531 } 3532 break; 3533 3534 case vmIntrinsics::_getClassAccessFlags: 3535 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset())); 3536 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered); 3537 break; 3538 3539 default: 3540 fatal_unexpected_iid(id); 3541 break; 3542 } 3543 3544 // Fall-through is the normal case of a query to a real class. 3545 phi->init_req(1, query_value); 3546 region->init_req(1, control()); 3547 3548 C->set_has_split_ifs(true); // Has chance for split-if optimization 3549 set_result(region, phi); 3550 return true; 3551 } 3552 3553 //--------------------------inline_native_subtype_check------------------------ 3554 // This intrinsic takes the JNI calls out of the heart of 3555 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc. 3556 bool LibraryCallKit::inline_native_subtype_check() { 3557 // Pull both arguments off the stack. 3558 Node* args[2]; // two java.lang.Class mirrors: superc, subc 3559 args[0] = argument(0); 3560 args[1] = argument(1); 3561 Node* klasses[2]; // corresponding Klasses: superk, subk 3562 klasses[0] = klasses[1] = top(); 3563 3564 enum { 3565 // A full decision tree on {superc is prim, subc is prim}: 3566 _prim_0_path = 1, // {P,N} => false 3567 // {P,P} & superc!=subc => false 3568 _prim_same_path, // {P,P} & superc==subc => true 3569 _prim_1_path, // {N,P} => false 3570 _ref_subtype_path, // {N,N} & subtype check wins => true 3571 _both_ref_path, // {N,N} & subtype check loses => false 3572 PATH_LIMIT 3573 }; 3574 3575 RegionNode* region = new RegionNode(PATH_LIMIT); 3576 Node* phi = new PhiNode(region, TypeInt::BOOL); 3577 record_for_igvn(region); 3578 3579 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads 3580 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL; 3581 int class_klass_offset = java_lang_Class::klass_offset_in_bytes(); 3582 3583 // First null-check both mirrors and load each mirror's klass metaobject. 3584 int which_arg; 3585 for (which_arg = 0; which_arg <= 1; which_arg++) { 3586 Node* arg = args[which_arg]; 3587 arg = null_check(arg); 3588 if (stopped()) break; 3589 args[which_arg] = arg; 3590 3591 Node* p = basic_plus_adr(arg, class_klass_offset); 3592 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type); 3593 klasses[which_arg] = _gvn.transform(kls); 3594 } 3595 3596 // Having loaded both klasses, test each for null. 3597 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3598 for (which_arg = 0; which_arg <= 1; which_arg++) { 3599 Node* kls = klasses[which_arg]; 3600 Node* null_ctl = top(); 3601 kls = null_check_oop(kls, &null_ctl, never_see_null); 3602 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path); 3603 region->init_req(prim_path, null_ctl); 3604 if (stopped()) break; 3605 klasses[which_arg] = kls; 3606 } 3607 3608 if (!stopped()) { 3609 // now we have two reference types, in klasses[0..1] 3610 Node* subk = klasses[1]; // the argument to isAssignableFrom 3611 Node* superk = klasses[0]; // the receiver 3612 region->set_req(_both_ref_path, gen_subtype_check(subk, superk)); 3613 // now we have a successful reference subtype check 3614 region->set_req(_ref_subtype_path, control()); 3615 } 3616 3617 // If both operands are primitive (both klasses null), then 3618 // we must return true when they are identical primitives. 3619 // It is convenient to test this after the first null klass check. 3620 set_control(region->in(_prim_0_path)); // go back to first null check 3621 if (!stopped()) { 3622 // Since superc is primitive, make a guard for the superc==subc case. 3623 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1])); 3624 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq)); 3625 generate_guard(bol_eq, region, PROB_FAIR); 3626 if (region->req() == PATH_LIMIT+1) { 3627 // A guard was added. If the added guard is taken, superc==subc. 3628 region->swap_edges(PATH_LIMIT, _prim_same_path); 3629 region->del_req(PATH_LIMIT); 3630 } 3631 region->set_req(_prim_0_path, control()); // Not equal after all. 3632 } 3633 3634 // these are the only paths that produce 'true': 3635 phi->set_req(_prim_same_path, intcon(1)); 3636 phi->set_req(_ref_subtype_path, intcon(1)); 3637 3638 // pull together the cases: 3639 assert(region->req() == PATH_LIMIT, "sane region"); 3640 for (uint i = 1; i < region->req(); i++) { 3641 Node* ctl = region->in(i); 3642 if (ctl == NULL || ctl == top()) { 3643 region->set_req(i, top()); 3644 phi ->set_req(i, top()); 3645 } else if (phi->in(i) == NULL) { 3646 phi->set_req(i, intcon(0)); // all other paths produce 'false' 3647 } 3648 } 3649 3650 set_control(_gvn.transform(region)); 3651 set_result(_gvn.transform(phi)); 3652 return true; 3653 } 3654 3655 //---------------------generate_array_guard_common------------------------ 3656 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, 3657 bool obj_array, bool not_array) { 3658 // If obj_array/non_array==false/false: 3659 // Branch around if the given klass is in fact an array (either obj or prim). 3660 // If obj_array/non_array==false/true: 3661 // Branch around if the given klass is not an array klass of any kind. 3662 // If obj_array/non_array==true/true: 3663 // Branch around if the kls is not an oop array (kls is int[], String, etc.) 3664 // If obj_array/non_array==true/false: 3665 // Branch around if the kls is an oop array (Object[] or subtype) 3666 // 3667 // Like generate_guard, adds a new path onto the region. 3668 jint layout_con = 0; 3669 Node* layout_val = get_layout_helper(kls, layout_con); 3670 if (layout_val == NULL) { 3671 bool query = (obj_array 3672 ? Klass::layout_helper_is_objArray(layout_con) 3673 : Klass::layout_helper_is_array(layout_con)); 3674 if (query == not_array) { 3675 return NULL; // never a branch 3676 } else { // always a branch 3677 Node* always_branch = control(); 3678 if (region != NULL) 3679 region->add_req(always_branch); 3680 set_control(top()); 3681 return always_branch; 3682 } 3683 } 3684 // Now test the correct condition. 3685 jint nval = (obj_array 3686 ? ((jint)Klass::_lh_array_tag_type_value 3687 << Klass::_lh_array_tag_shift) 3688 : Klass::_lh_neutral_value); 3689 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval))); 3690 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array 3691 // invert the test if we are looking for a non-array 3692 if (not_array) btest = BoolTest(btest).negate(); 3693 Node* bol = _gvn.transform(new BoolNode(cmp, btest)); 3694 return generate_fair_guard(bol, region); 3695 } 3696 3697 3698 //-----------------------inline_native_newArray-------------------------- 3699 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length); 3700 bool LibraryCallKit::inline_native_newArray() { 3701 Node* mirror = argument(0); 3702 Node* count_val = argument(1); 3703 3704 mirror = null_check(mirror); 3705 // If mirror or obj is dead, only null-path is taken. 3706 if (stopped()) return true; 3707 3708 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT }; 3709 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3710 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 3711 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3712 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3713 3714 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check); 3715 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null, 3716 result_reg, _slow_path); 3717 Node* normal_ctl = control(); 3718 Node* no_array_ctl = result_reg->in(_slow_path); 3719 3720 // Generate code for the slow case. We make a call to newArray(). 3721 set_control(no_array_ctl); 3722 if (!stopped()) { 3723 // Either the input type is void.class, or else the 3724 // array klass has not yet been cached. Either the 3725 // ensuing call will throw an exception, or else it 3726 // will cache the array klass for next time. 3727 PreserveJVMState pjvms(this); 3728 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray); 3729 Node* slow_result = set_results_for_java_call(slow_call); 3730 // this->control() comes from set_results_for_java_call 3731 result_reg->set_req(_slow_path, control()); 3732 result_val->set_req(_slow_path, slow_result); 3733 result_io ->set_req(_slow_path, i_o()); 3734 result_mem->set_req(_slow_path, reset_memory()); 3735 } 3736 3737 set_control(normal_ctl); 3738 if (!stopped()) { 3739 // Normal case: The array type has been cached in the java.lang.Class. 3740 // The following call works fine even if the array type is polymorphic. 3741 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3742 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push 3743 result_reg->init_req(_normal_path, control()); 3744 result_val->init_req(_normal_path, obj); 3745 result_io ->init_req(_normal_path, i_o()); 3746 result_mem->init_req(_normal_path, reset_memory()); 3747 } 3748 3749 // Return the combined state. 3750 set_i_o( _gvn.transform(result_io) ); 3751 set_all_memory( _gvn.transform(result_mem)); 3752 3753 C->set_has_split_ifs(true); // Has chance for split-if optimization 3754 set_result(result_reg, result_val); 3755 return true; 3756 } 3757 3758 //----------------------inline_native_getLength-------------------------- 3759 // public static native int java.lang.reflect.Array.getLength(Object array); 3760 bool LibraryCallKit::inline_native_getLength() { 3761 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3762 3763 Node* array = null_check(argument(0)); 3764 // If array is dead, only null-path is taken. 3765 if (stopped()) return true; 3766 3767 // Deoptimize if it is a non-array. 3768 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL); 3769 3770 if (non_array != NULL) { 3771 PreserveJVMState pjvms(this); 3772 set_control(non_array); 3773 uncommon_trap(Deoptimization::Reason_intrinsic, 3774 Deoptimization::Action_maybe_recompile); 3775 } 3776 3777 // If control is dead, only non-array-path is taken. 3778 if (stopped()) return true; 3779 3780 // The works fine even if the array type is polymorphic. 3781 // It could be a dynamic mix of int[], boolean[], Object[], etc. 3782 Node* result = load_array_length(array); 3783 3784 C->set_has_split_ifs(true); // Has chance for split-if optimization 3785 set_result(result); 3786 return true; 3787 } 3788 3789 //------------------------inline_array_copyOf---------------------------- 3790 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType); 3791 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType); 3792 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) { 3793 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false; 3794 3795 // Get the arguments. 3796 Node* original = argument(0); 3797 Node* start = is_copyOfRange? argument(1): intcon(0); 3798 Node* end = is_copyOfRange? argument(2): argument(1); 3799 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2); 3800 3801 Node* newcopy; 3802 3803 // Set the original stack and the reexecute bit for the interpreter to reexecute 3804 // the bytecode that invokes Arrays.copyOf if deoptimization happens. 3805 { PreserveReexecuteState preexecs(this); 3806 jvms()->set_should_reexecute(true); 3807 3808 array_type_mirror = null_check(array_type_mirror); 3809 original = null_check(original); 3810 3811 // Check if a null path was taken unconditionally. 3812 if (stopped()) return true; 3813 3814 Node* orig_length = load_array_length(original); 3815 3816 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0); 3817 klass_node = null_check(klass_node); 3818 3819 RegionNode* bailout = new RegionNode(1); 3820 record_for_igvn(bailout); 3821 3822 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc. 3823 // Bail out if that is so. 3824 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout); 3825 if (not_objArray != NULL) { 3826 // Improve the klass node's type from the new optimistic assumption: 3827 ciKlass* ak = ciArrayKlass::make(env()->Object_klass()); 3828 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); 3829 Node* cast = new CastPPNode(klass_node, akls); 3830 cast->init_req(0, control()); 3831 klass_node = _gvn.transform(cast); 3832 } 3833 3834 // Bail out if either start or end is negative. 3835 generate_negative_guard(start, bailout, &start); 3836 generate_negative_guard(end, bailout, &end); 3837 3838 Node* length = end; 3839 if (_gvn.type(start) != TypeInt::ZERO) { 3840 length = _gvn.transform(new SubINode(end, start)); 3841 } 3842 3843 // Bail out if length is negative. 3844 // Without this the new_array would throw 3845 // NegativeArraySizeException but IllegalArgumentException is what 3846 // should be thrown 3847 generate_negative_guard(length, bailout, &length); 3848 3849 if (bailout->req() > 1) { 3850 PreserveJVMState pjvms(this); 3851 set_control(_gvn.transform(bailout)); 3852 uncommon_trap(Deoptimization::Reason_intrinsic, 3853 Deoptimization::Action_maybe_recompile); 3854 } 3855 3856 if (!stopped()) { 3857 // How many elements will we copy from the original? 3858 // The answer is MinI(orig_length - start, length). 3859 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start)); 3860 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length); 3861 3862 newcopy = new_array(klass_node, length, 0); // no arguments to push 3863 3864 // Generate a direct call to the right arraycopy function(s). 3865 // We know the copy is disjoint but we might not know if the 3866 // oop stores need checking. 3867 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class). 3868 // This will fail a store-check if x contains any non-nulls. 3869 3870 Node* alloc = tightly_coupled_allocation(newcopy, NULL); 3871 3872 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL, 3873 load_object_klass(original), klass_node); 3874 if (!is_copyOfRange) { 3875 ac->set_copyof(); 3876 } else { 3877 ac->set_copyofrange(); 3878 } 3879 Node* n = _gvn.transform(ac); 3880 assert(n == ac, "cannot disappear"); 3881 ac->connect_outputs(this); 3882 } 3883 } // original reexecute is set back here 3884 3885 C->set_has_split_ifs(true); // Has chance for split-if optimization 3886 if (!stopped()) { 3887 set_result(newcopy); 3888 } 3889 return true; 3890 } 3891 3892 3893 //----------------------generate_virtual_guard--------------------------- 3894 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call. 3895 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass, 3896 RegionNode* slow_region) { 3897 ciMethod* method = callee(); 3898 int vtable_index = method->vtable_index(); 3899 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3900 err_msg_res("bad index %d", vtable_index)); 3901 // Get the Method* out of the appropriate vtable entry. 3902 int entry_offset = (InstanceKlass::vtable_start_offset() + 3903 vtable_index*vtableEntry::size()) * wordSize + 3904 vtableEntry::method_offset_in_bytes(); 3905 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset); 3906 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered); 3907 3908 // Compare the target method with the expected method (e.g., Object.hashCode). 3909 const TypePtr* native_call_addr = TypeMetadataPtr::make(method); 3910 3911 Node* native_call = makecon(native_call_addr); 3912 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call)); 3913 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne)); 3914 3915 return generate_slow_guard(test_native, slow_region); 3916 } 3917 3918 //-----------------------generate_method_call---------------------------- 3919 // Use generate_method_call to make a slow-call to the real 3920 // method if the fast path fails. An alternative would be to 3921 // use a stub like OptoRuntime::slow_arraycopy_Java. 3922 // This only works for expanding the current library call, 3923 // not another intrinsic. (E.g., don't use this for making an 3924 // arraycopy call inside of the copyOf intrinsic.) 3925 CallJavaNode* 3926 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) { 3927 // When compiling the intrinsic method itself, do not use this technique. 3928 guarantee(callee() != C->method(), "cannot make slow-call to self"); 3929 3930 ciMethod* method = callee(); 3931 // ensure the JVMS we have will be correct for this call 3932 guarantee(method_id == method->intrinsic_id(), "must match"); 3933 3934 const TypeFunc* tf = TypeFunc::make(method); 3935 CallJavaNode* slow_call; 3936 if (is_static) { 3937 assert(!is_virtual, ""); 3938 slow_call = new CallStaticJavaNode(C, tf, 3939 SharedRuntime::get_resolve_static_call_stub(), 3940 method, bci()); 3941 } else if (is_virtual) { 3942 null_check_receiver(); 3943 int vtable_index = Method::invalid_vtable_index; 3944 if (UseInlineCaches) { 3945 // Suppress the vtable call 3946 } else { 3947 // hashCode and clone are not a miranda methods, 3948 // so the vtable index is fixed. 3949 // No need to use the linkResolver to get it. 3950 vtable_index = method->vtable_index(); 3951 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index, 3952 err_msg_res("bad index %d", vtable_index)); 3953 } 3954 slow_call = new CallDynamicJavaNode(tf, 3955 SharedRuntime::get_resolve_virtual_call_stub(), 3956 method, vtable_index, bci()); 3957 } else { // neither virtual nor static: opt_virtual 3958 null_check_receiver(); 3959 slow_call = new CallStaticJavaNode(C, tf, 3960 SharedRuntime::get_resolve_opt_virtual_call_stub(), 3961 method, bci()); 3962 slow_call->set_optimized_virtual(true); 3963 } 3964 set_arguments_for_java_call(slow_call); 3965 set_edges_for_java_call(slow_call); 3966 return slow_call; 3967 } 3968 3969 3970 /** 3971 * Build special case code for calls to hashCode on an object. This call may 3972 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate 3973 * slightly different code. 3974 */ 3975 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) { 3976 assert(is_static == callee()->is_static(), "correct intrinsic selection"); 3977 assert(!(is_virtual && is_static), "either virtual, special, or static"); 3978 3979 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT }; 3980 3981 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 3982 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT); 3983 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO); 3984 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 3985 Node* obj = NULL; 3986 if (!is_static) { 3987 // Check for hashing null object 3988 obj = null_check_receiver(); 3989 if (stopped()) return true; // unconditionally null 3990 result_reg->init_req(_null_path, top()); 3991 result_val->init_req(_null_path, top()); 3992 } else { 3993 // Do a null check, and return zero if null. 3994 // System.identityHashCode(null) == 0 3995 obj = argument(0); 3996 Node* null_ctl = top(); 3997 obj = null_check_oop(obj, &null_ctl); 3998 result_reg->init_req(_null_path, null_ctl); 3999 result_val->init_req(_null_path, _gvn.intcon(0)); 4000 } 4001 4002 // Unconditionally null? Then return right away. 4003 if (stopped()) { 4004 set_control( result_reg->in(_null_path)); 4005 if (!stopped()) 4006 set_result(result_val->in(_null_path)); 4007 return true; 4008 } 4009 4010 // We only go to the fast case code if we pass a number of guards. The 4011 // paths which do not pass are accumulated in the slow_region. 4012 RegionNode* slow_region = new RegionNode(1); 4013 record_for_igvn(slow_region); 4014 4015 // If this is a virtual call, we generate a funny guard. We pull out 4016 // the vtable entry corresponding to hashCode() from the target object. 4017 // If the target method which we are calling happens to be the native 4018 // Object hashCode() method, we pass the guard. We do not need this 4019 // guard for non-virtual calls -- the caller is known to be the native 4020 // Object hashCode(). 4021 if (is_virtual) { 4022 // After null check, get the object's klass. 4023 Node* obj_klass = load_object_klass(obj); 4024 generate_virtual_guard(obj_klass, slow_region); 4025 } 4026 4027 // Get the header out of the object, use LoadMarkNode when available 4028 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); 4029 // The control of the load must be NULL. Otherwise, the load can move before 4030 // the null check after castPP removal. 4031 Node* no_ctrl = NULL; 4032 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 4033 4034 // Test the header to see if it is unlocked. 4035 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place); 4036 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask)); 4037 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value); 4038 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val)); 4039 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne)); 4040 4041 generate_slow_guard(test_unlocked, slow_region); 4042 4043 // Get the hash value and check to see that it has been properly assigned. 4044 // We depend on hash_mask being at most 32 bits and avoid the use of 4045 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit 4046 // vm: see markOop.hpp. 4047 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask); 4048 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift); 4049 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift)); 4050 // This hack lets the hash bits live anywhere in the mark object now, as long 4051 // as the shift drops the relevant bits into the low 32 bits. Note that 4052 // Java spec says that HashCode is an int so there's no point in capturing 4053 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build). 4054 hshifted_header = ConvX2I(hshifted_header); 4055 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask)); 4056 4057 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash); 4058 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val)); 4059 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq)); 4060 4061 generate_slow_guard(test_assigned, slow_region); 4062 4063 Node* init_mem = reset_memory(); 4064 // fill in the rest of the null path: 4065 result_io ->init_req(_null_path, i_o()); 4066 result_mem->init_req(_null_path, init_mem); 4067 4068 result_val->init_req(_fast_path, hash_val); 4069 result_reg->init_req(_fast_path, control()); 4070 result_io ->init_req(_fast_path, i_o()); 4071 result_mem->init_req(_fast_path, init_mem); 4072 4073 // Generate code for the slow case. We make a call to hashCode(). 4074 set_control(_gvn.transform(slow_region)); 4075 if (!stopped()) { 4076 // No need for PreserveJVMState, because we're using up the present state. 4077 set_all_memory(init_mem); 4078 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode; 4079 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static); 4080 Node* slow_result = set_results_for_java_call(slow_call); 4081 // this->control() comes from set_results_for_java_call 4082 result_reg->init_req(_slow_path, control()); 4083 result_val->init_req(_slow_path, slow_result); 4084 result_io ->set_req(_slow_path, i_o()); 4085 result_mem ->set_req(_slow_path, reset_memory()); 4086 } 4087 4088 // Return the combined state. 4089 set_i_o( _gvn.transform(result_io) ); 4090 set_all_memory( _gvn.transform(result_mem)); 4091 4092 set_result(result_reg, result_val); 4093 return true; 4094 } 4095 4096 //---------------------------inline_native_getClass---------------------------- 4097 // public final native Class<?> java.lang.Object.getClass(); 4098 // 4099 // Build special case code for calls to getClass on an object. 4100 bool LibraryCallKit::inline_native_getClass() { 4101 Node* obj = null_check_receiver(); 4102 if (stopped()) return true; 4103 set_result(load_mirror_from_klass(load_object_klass(obj))); 4104 return true; 4105 } 4106 4107 //-----------------inline_native_Reflection_getCallerClass--------------------- 4108 // public static native Class<?> sun.reflect.Reflection.getCallerClass(); 4109 // 4110 // In the presence of deep enough inlining, getCallerClass() becomes a no-op. 4111 // 4112 // NOTE: This code must perform the same logic as JVM_GetCallerClass 4113 // in that it must skip particular security frames and checks for 4114 // caller sensitive methods. 4115 bool LibraryCallKit::inline_native_Reflection_getCallerClass() { 4116 #ifndef PRODUCT 4117 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4118 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass"); 4119 } 4120 #endif 4121 4122 if (!jvms()->has_method()) { 4123 #ifndef PRODUCT 4124 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4125 tty->print_cr(" Bailing out because intrinsic was inlined at top level"); 4126 } 4127 #endif 4128 return false; 4129 } 4130 4131 // Walk back up the JVM state to find the caller at the required 4132 // depth. 4133 JVMState* caller_jvms = jvms(); 4134 4135 // Cf. JVM_GetCallerClass 4136 // NOTE: Start the loop at depth 1 because the current JVM state does 4137 // not include the Reflection.getCallerClass() frame. 4138 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) { 4139 ciMethod* m = caller_jvms->method(); 4140 switch (n) { 4141 case 0: 4142 fatal("current JVM state does not include the Reflection.getCallerClass frame"); 4143 break; 4144 case 1: 4145 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass). 4146 if (!m->caller_sensitive()) { 4147 #ifndef PRODUCT 4148 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4149 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n); 4150 } 4151 #endif 4152 return false; // bail-out; let JVM_GetCallerClass do the work 4153 } 4154 break; 4155 default: 4156 if (!m->is_ignored_by_security_stack_walk()) { 4157 // We have reached the desired frame; return the holder class. 4158 // Acquire method holder as java.lang.Class and push as constant. 4159 ciInstanceKlass* caller_klass = caller_jvms->method()->holder(); 4160 ciInstance* caller_mirror = caller_klass->java_mirror(); 4161 set_result(makecon(TypeInstPtr::make(caller_mirror))); 4162 4163 #ifndef PRODUCT 4164 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4165 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()); 4166 tty->print_cr(" JVM state at this point:"); 4167 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4168 ciMethod* m = jvms()->of_depth(i)->method(); 4169 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4170 } 4171 } 4172 #endif 4173 return true; 4174 } 4175 break; 4176 } 4177 } 4178 4179 #ifndef PRODUCT 4180 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) { 4181 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth()); 4182 tty->print_cr(" JVM state at this point:"); 4183 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) { 4184 ciMethod* m = jvms()->of_depth(i)->method(); 4185 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8()); 4186 } 4187 } 4188 #endif 4189 4190 return false; // bail-out; let JVM_GetCallerClass do the work 4191 } 4192 4193 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) { 4194 Node* arg = argument(0); 4195 Node* result; 4196 4197 switch (id) { 4198 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break; 4199 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break; 4200 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break; 4201 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break; 4202 4203 case vmIntrinsics::_doubleToLongBits: { 4204 // two paths (plus control) merge in a wood 4205 RegionNode *r = new RegionNode(3); 4206 Node *phi = new PhiNode(r, TypeLong::LONG); 4207 4208 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg)); 4209 // Build the boolean node 4210 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4211 4212 // Branch either way. 4213 // NaN case is less traveled, which makes all the difference. 4214 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4215 Node *opt_isnan = _gvn.transform(ifisnan); 4216 assert( opt_isnan->is_If(), "Expect an IfNode"); 4217 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4218 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4219 4220 set_control(iftrue); 4221 4222 static const jlong nan_bits = CONST64(0x7ff8000000000000); 4223 Node *slow_result = longcon(nan_bits); // return NaN 4224 phi->init_req(1, _gvn.transform( slow_result )); 4225 r->init_req(1, iftrue); 4226 4227 // Else fall through 4228 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4229 set_control(iffalse); 4230 4231 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg))); 4232 r->init_req(2, iffalse); 4233 4234 // Post merge 4235 set_control(_gvn.transform(r)); 4236 record_for_igvn(r); 4237 4238 C->set_has_split_ifs(true); // Has chance for split-if optimization 4239 result = phi; 4240 assert(result->bottom_type()->isa_long(), "must be"); 4241 break; 4242 } 4243 4244 case vmIntrinsics::_floatToIntBits: { 4245 // two paths (plus control) merge in a wood 4246 RegionNode *r = new RegionNode(3); 4247 Node *phi = new PhiNode(r, TypeInt::INT); 4248 4249 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg)); 4250 // Build the boolean node 4251 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne)); 4252 4253 // Branch either way. 4254 // NaN case is less traveled, which makes all the difference. 4255 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN); 4256 Node *opt_isnan = _gvn.transform(ifisnan); 4257 assert( opt_isnan->is_If(), "Expect an IfNode"); 4258 IfNode *opt_ifisnan = (IfNode*)opt_isnan; 4259 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan)); 4260 4261 set_control(iftrue); 4262 4263 static const jint nan_bits = 0x7fc00000; 4264 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN 4265 phi->init_req(1, _gvn.transform( slow_result )); 4266 r->init_req(1, iftrue); 4267 4268 // Else fall through 4269 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan)); 4270 set_control(iffalse); 4271 4272 phi->init_req(2, _gvn.transform(new MoveF2INode(arg))); 4273 r->init_req(2, iffalse); 4274 4275 // Post merge 4276 set_control(_gvn.transform(r)); 4277 record_for_igvn(r); 4278 4279 C->set_has_split_ifs(true); // Has chance for split-if optimization 4280 result = phi; 4281 assert(result->bottom_type()->isa_int(), "must be"); 4282 break; 4283 } 4284 4285 default: 4286 fatal_unexpected_iid(id); 4287 break; 4288 } 4289 set_result(_gvn.transform(result)); 4290 return true; 4291 } 4292 4293 #ifdef _LP64 4294 #define XTOP ,top() /*additional argument*/ 4295 #else //_LP64 4296 #define XTOP /*no additional argument*/ 4297 #endif //_LP64 4298 4299 //----------------------inline_unsafe_copyMemory------------------------- 4300 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes); 4301 bool LibraryCallKit::inline_unsafe_copyMemory() { 4302 if (callee()->is_static()) return false; // caller must have the capability! 4303 null_check_receiver(); // null-check receiver 4304 if (stopped()) return true; 4305 4306 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe". 4307 4308 Node* src_ptr = argument(1); // type: oop 4309 Node* src_off = ConvL2X(argument(2)); // type: long 4310 Node* dst_ptr = argument(4); // type: oop 4311 Node* dst_off = ConvL2X(argument(5)); // type: long 4312 Node* size = ConvL2X(argument(7)); // type: long 4313 4314 assert(Unsafe_field_offset_to_byte_offset(11) == 11, 4315 "fieldOffset must be byte-scaled"); 4316 4317 Node* src = make_unsafe_address(src_ptr, src_off); 4318 Node* dst = make_unsafe_address(dst_ptr, dst_off); 4319 4320 // Conservatively insert a memory barrier on all memory slices. 4321 // Do not let writes of the copy source or destination float below the copy. 4322 insert_mem_bar(Op_MemBarCPUOrder); 4323 4324 // Call it. Note that the length argument is not scaled. 4325 make_runtime_call(RC_LEAF|RC_NO_FP, 4326 OptoRuntime::fast_arraycopy_Type(), 4327 StubRoutines::unsafe_arraycopy(), 4328 "unsafe_arraycopy", 4329 TypeRawPtr::BOTTOM, 4330 src, dst, size XTOP); 4331 4332 // Do not let reads of the copy destination float above the copy. 4333 insert_mem_bar(Op_MemBarCPUOrder); 4334 4335 return true; 4336 } 4337 4338 //------------------------clone_coping----------------------------------- 4339 // Helper function for inline_native_clone. 4340 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) { 4341 assert(obj_size != NULL, ""); 4342 Node* raw_obj = alloc_obj->in(1); 4343 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), ""); 4344 4345 AllocateNode* alloc = NULL; 4346 if (ReduceBulkZeroing) { 4347 // We will be completely responsible for initializing this object - 4348 // mark Initialize node as complete. 4349 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn); 4350 // The object was just allocated - there should be no any stores! 4351 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), ""); 4352 // Mark as complete_with_arraycopy so that on AllocateNode 4353 // expansion, we know this AllocateNode is initialized by an array 4354 // copy and a StoreStore barrier exists after the array copy. 4355 alloc->initialization()->set_complete_with_arraycopy(); 4356 } 4357 4358 // Copy the fastest available way. 4359 // TODO: generate fields copies for small objects instead. 4360 Node* src = obj; 4361 Node* dest = alloc_obj; 4362 Node* size = _gvn.transform(obj_size); 4363 4364 // Exclude the header but include array length to copy by 8 bytes words. 4365 // Can't use base_offset_in_bytes(bt) since basic type is unknown. 4366 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() : 4367 instanceOopDesc::base_offset_in_bytes(); 4368 // base_off: 4369 // 8 - 32-bit VM 4370 // 12 - 64-bit VM, compressed klass 4371 // 16 - 64-bit VM, normal klass 4372 if (base_off % BytesPerLong != 0) { 4373 assert(UseCompressedClassPointers, ""); 4374 if (is_array) { 4375 // Exclude length to copy by 8 bytes words. 4376 base_off += sizeof(int); 4377 } else { 4378 // Include klass to copy by 8 bytes words. 4379 base_off = instanceOopDesc::klass_offset_in_bytes(); 4380 } 4381 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment"); 4382 } 4383 src = basic_plus_adr(src, base_off); 4384 dest = basic_plus_adr(dest, base_off); 4385 4386 // Compute the length also, if needed: 4387 Node* countx = size; 4388 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off))); 4389 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) )); 4390 4391 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4392 4393 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false); 4394 ac->set_clonebasic(); 4395 Node* n = _gvn.transform(ac); 4396 assert(n == ac, "cannot disappear"); 4397 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type); 4398 4399 // If necessary, emit some card marks afterwards. (Non-arrays only.) 4400 if (card_mark) { 4401 assert(!is_array, ""); 4402 // Put in store barrier for any and all oops we are sticking 4403 // into this object. (We could avoid this if we could prove 4404 // that the object type contains no oop fields at all.) 4405 Node* no_particular_value = NULL; 4406 Node* no_particular_field = NULL; 4407 int raw_adr_idx = Compile::AliasIdxRaw; 4408 post_barrier(control(), 4409 memory(raw_adr_type), 4410 alloc_obj, 4411 no_particular_field, 4412 raw_adr_idx, 4413 no_particular_value, 4414 T_OBJECT, 4415 false); 4416 } 4417 4418 // Do not let reads from the cloned object float above the arraycopy. 4419 if (alloc != NULL) { 4420 // Do not let stores that initialize this object be reordered with 4421 // a subsequent store that would make this object accessible by 4422 // other threads. 4423 // Record what AllocateNode this StoreStore protects so that 4424 // escape analysis can go from the MemBarStoreStoreNode to the 4425 // AllocateNode and eliminate the MemBarStoreStoreNode if possible 4426 // based on the escape status of the AllocateNode. 4427 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress)); 4428 } else { 4429 insert_mem_bar(Op_MemBarCPUOrder); 4430 } 4431 } 4432 4433 //------------------------inline_native_clone---------------------------- 4434 // protected native Object java.lang.Object.clone(); 4435 // 4436 // Here are the simple edge cases: 4437 // null receiver => normal trap 4438 // virtual and clone was overridden => slow path to out-of-line clone 4439 // not cloneable or finalizer => slow path to out-of-line Object.clone 4440 // 4441 // The general case has two steps, allocation and copying. 4442 // Allocation has two cases, and uses GraphKit::new_instance or new_array. 4443 // 4444 // Copying also has two cases, oop arrays and everything else. 4445 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy). 4446 // Everything else uses the tight inline loop supplied by CopyArrayNode. 4447 // 4448 // These steps fold up nicely if and when the cloned object's klass 4449 // can be sharply typed as an object array, a type array, or an instance. 4450 // 4451 bool LibraryCallKit::inline_native_clone(bool is_virtual) { 4452 PhiNode* result_val; 4453 4454 // Set the reexecute bit for the interpreter to reexecute 4455 // the bytecode that invokes Object.clone if deoptimization happens. 4456 { PreserveReexecuteState preexecs(this); 4457 jvms()->set_should_reexecute(true); 4458 4459 Node* obj = null_check_receiver(); 4460 if (stopped()) return true; 4461 4462 Node* obj_klass = load_object_klass(obj); 4463 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr(); 4464 const TypeOopPtr* toop = ((tklass != NULL) 4465 ? tklass->as_instance_type() 4466 : TypeInstPtr::NOTNULL); 4467 4468 // Conservatively insert a memory barrier on all memory slices. 4469 // Do not let writes into the original float below the clone. 4470 insert_mem_bar(Op_MemBarCPUOrder); 4471 4472 // paths into result_reg: 4473 enum { 4474 _slow_path = 1, // out-of-line call to clone method (virtual or not) 4475 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy 4476 _array_path, // plain array allocation, plus arrayof_long_arraycopy 4477 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy 4478 PATH_LIMIT 4479 }; 4480 RegionNode* result_reg = new RegionNode(PATH_LIMIT); 4481 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL); 4482 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO); 4483 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM); 4484 record_for_igvn(result_reg); 4485 4486 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM; 4487 int raw_adr_idx = Compile::AliasIdxRaw; 4488 4489 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL); 4490 if (array_ctl != NULL) { 4491 // It's an array. 4492 PreserveJVMState pjvms(this); 4493 set_control(array_ctl); 4494 Node* obj_length = load_array_length(obj); 4495 Node* obj_size = NULL; 4496 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push 4497 4498 if (!use_ReduceInitialCardMarks()) { 4499 // If it is an oop array, it requires very special treatment, 4500 // because card marking is required on each card of the array. 4501 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL); 4502 if (is_obja != NULL) { 4503 PreserveJVMState pjvms2(this); 4504 set_control(is_obja); 4505 // Generate a direct call to the right arraycopy function(s). 4506 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL); 4507 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL); 4508 ac->set_cloneoop(); 4509 Node* n = _gvn.transform(ac); 4510 assert(n == ac, "cannot disappear"); 4511 ac->connect_outputs(this); 4512 4513 result_reg->init_req(_objArray_path, control()); 4514 result_val->init_req(_objArray_path, alloc_obj); 4515 result_i_o ->set_req(_objArray_path, i_o()); 4516 result_mem ->set_req(_objArray_path, reset_memory()); 4517 } 4518 } 4519 // Otherwise, there are no card marks to worry about. 4520 // (We can dispense with card marks if we know the allocation 4521 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks 4522 // causes the non-eden paths to take compensating steps to 4523 // simulate a fresh allocation, so that no further 4524 // card marks are required in compiled code to initialize 4525 // the object.) 4526 4527 if (!stopped()) { 4528 copy_to_clone(obj, alloc_obj, obj_size, true, false); 4529 4530 // Present the results of the copy. 4531 result_reg->init_req(_array_path, control()); 4532 result_val->init_req(_array_path, alloc_obj); 4533 result_i_o ->set_req(_array_path, i_o()); 4534 result_mem ->set_req(_array_path, reset_memory()); 4535 } 4536 } 4537 4538 // We only go to the instance fast case code if we pass a number of guards. 4539 // The paths which do not pass are accumulated in the slow_region. 4540 RegionNode* slow_region = new RegionNode(1); 4541 record_for_igvn(slow_region); 4542 if (!stopped()) { 4543 // It's an instance (we did array above). Make the slow-path tests. 4544 // If this is a virtual call, we generate a funny guard. We grab 4545 // the vtable entry corresponding to clone() from the target object. 4546 // If the target method which we are calling happens to be the 4547 // Object clone() method, we pass the guard. We do not need this 4548 // guard for non-virtual calls; the caller is known to be the native 4549 // Object clone(). 4550 if (is_virtual) { 4551 generate_virtual_guard(obj_klass, slow_region); 4552 } 4553 4554 // The object must be cloneable and must not have a finalizer. 4555 // Both of these conditions may be checked in a single test. 4556 // We could optimize the cloneable test further, but we don't care. 4557 generate_access_flags_guard(obj_klass, 4558 // Test both conditions: 4559 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER, 4560 // Must be cloneable but not finalizer: 4561 JVM_ACC_IS_CLONEABLE, 4562 slow_region); 4563 } 4564 4565 if (!stopped()) { 4566 // It's an instance, and it passed the slow-path tests. 4567 PreserveJVMState pjvms(this); 4568 Node* obj_size = NULL; 4569 // Need to deoptimize on exception from allocation since Object.clone intrinsic 4570 // is reexecuted if deoptimization occurs and there could be problems when merging 4571 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false). 4572 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true); 4573 4574 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks()); 4575 4576 // Present the results of the slow call. 4577 result_reg->init_req(_instance_path, control()); 4578 result_val->init_req(_instance_path, alloc_obj); 4579 result_i_o ->set_req(_instance_path, i_o()); 4580 result_mem ->set_req(_instance_path, reset_memory()); 4581 } 4582 4583 // Generate code for the slow case. We make a call to clone(). 4584 set_control(_gvn.transform(slow_region)); 4585 if (!stopped()) { 4586 PreserveJVMState pjvms(this); 4587 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual); 4588 Node* slow_result = set_results_for_java_call(slow_call); 4589 // this->control() comes from set_results_for_java_call 4590 result_reg->init_req(_slow_path, control()); 4591 result_val->init_req(_slow_path, slow_result); 4592 result_i_o ->set_req(_slow_path, i_o()); 4593 result_mem ->set_req(_slow_path, reset_memory()); 4594 } 4595 4596 // Return the combined state. 4597 set_control( _gvn.transform(result_reg)); 4598 set_i_o( _gvn.transform(result_i_o)); 4599 set_all_memory( _gvn.transform(result_mem)); 4600 } // original reexecute is set back here 4601 4602 set_result(_gvn.transform(result_val)); 4603 return true; 4604 } 4605 4606 //------------------------------inline_arraycopy----------------------- 4607 // public static native void java.lang.System.arraycopy(Object src, int srcPos, 4608 // Object dest, int destPos, 4609 // int length); 4610 bool LibraryCallKit::inline_arraycopy() { 4611 // Get the arguments. 4612 Node* src = argument(0); // type: oop 4613 Node* src_offset = argument(1); // type: int 4614 Node* dest = argument(2); // type: oop 4615 Node* dest_offset = argument(3); // type: int 4616 Node* length = argument(4); // type: int 4617 4618 // The following tests must be performed 4619 // (1) src and dest are arrays. 4620 // (2) src and dest arrays must have elements of the same BasicType 4621 // (3) src and dest must not be null. 4622 // (4) src_offset must not be negative. 4623 // (5) dest_offset must not be negative. 4624 // (6) length must not be negative. 4625 // (7) src_offset + length must not exceed length of src. 4626 // (8) dest_offset + length must not exceed length of dest. 4627 // (9) each element of an oop array must be assignable 4628 4629 // (3) src and dest must not be null. 4630 // always do this here because we need the JVM state for uncommon traps 4631 src = null_check(src, T_ARRAY); 4632 dest = null_check(dest, T_ARRAY); 4633 4634 bool notest = false; 4635 4636 const Type* src_type = _gvn.type(src); 4637 const Type* dest_type = _gvn.type(dest); 4638 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4639 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 4640 4641 // Do we have the type of src? 4642 bool has_src = (top_src != NULL && top_src->klass() != NULL); 4643 // Do we have the type of dest? 4644 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4645 // Is the type for src from speculation? 4646 bool src_spec = false; 4647 // Is the type for dest from speculation? 4648 bool dest_spec = false; 4649 4650 if (!has_src || !has_dest) { 4651 // We don't have sufficient type information, let's see if 4652 // speculative types can help. We need to have types for both src 4653 // and dest so that it pays off. 4654 4655 // Do we already have or could we have type information for src 4656 bool could_have_src = has_src; 4657 // Do we already have or could we have type information for dest 4658 bool could_have_dest = has_dest; 4659 4660 ciKlass* src_k = NULL; 4661 if (!has_src) { 4662 src_k = src_type->speculative_type_not_null(); 4663 if (src_k != NULL && src_k->is_array_klass()) { 4664 could_have_src = true; 4665 } 4666 } 4667 4668 ciKlass* dest_k = NULL; 4669 if (!has_dest) { 4670 dest_k = dest_type->speculative_type_not_null(); 4671 if (dest_k != NULL && dest_k->is_array_klass()) { 4672 could_have_dest = true; 4673 } 4674 } 4675 4676 if (could_have_src && could_have_dest) { 4677 // This is going to pay off so emit the required guards 4678 if (!has_src) { 4679 src = maybe_cast_profiled_obj(src, src_k); 4680 src_type = _gvn.type(src); 4681 top_src = src_type->isa_aryptr(); 4682 has_src = (top_src != NULL && top_src->klass() != NULL); 4683 src_spec = true; 4684 } 4685 if (!has_dest) { 4686 dest = maybe_cast_profiled_obj(dest, dest_k); 4687 dest_type = _gvn.type(dest); 4688 top_dest = dest_type->isa_aryptr(); 4689 has_dest = (top_dest != NULL && top_dest->klass() != NULL); 4690 dest_spec = true; 4691 } 4692 } 4693 } 4694 4695 if (has_src && has_dest) { 4696 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type(); 4697 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type(); 4698 if (src_elem == T_ARRAY) src_elem = T_OBJECT; 4699 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT; 4700 4701 if (src_elem == dest_elem && src_elem == T_OBJECT) { 4702 // If both arrays are object arrays then having the exact types 4703 // for both will remove the need for a subtype check at runtime 4704 // before the call and may make it possible to pick a faster copy 4705 // routine (without a subtype check on every element) 4706 // Do we have the exact type of src? 4707 bool could_have_src = src_spec; 4708 // Do we have the exact type of dest? 4709 bool could_have_dest = dest_spec; 4710 ciKlass* src_k = top_src->klass(); 4711 ciKlass* dest_k = top_dest->klass(); 4712 if (!src_spec) { 4713 src_k = src_type->speculative_type_not_null(); 4714 if (src_k != NULL && src_k->is_array_klass()) { 4715 could_have_src = true; 4716 } 4717 } 4718 if (!dest_spec) { 4719 dest_k = dest_type->speculative_type_not_null(); 4720 if (dest_k != NULL && dest_k->is_array_klass()) { 4721 could_have_dest = true; 4722 } 4723 } 4724 if (could_have_src && could_have_dest) { 4725 // If we can have both exact types, emit the missing guards 4726 if (could_have_src && !src_spec) { 4727 src = maybe_cast_profiled_obj(src, src_k); 4728 } 4729 if (could_have_dest && !dest_spec) { 4730 dest = maybe_cast_profiled_obj(dest, dest_k); 4731 } 4732 } 4733 } 4734 } 4735 4736 if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) { 4737 // validate arguments: enables transformation the ArrayCopyNode 4738 notest = true; 4739 4740 RegionNode* slow_region = new RegionNode(1); 4741 record_for_igvn(slow_region); 4742 4743 // (1) src and dest are arrays. 4744 generate_non_array_guard(load_object_klass(src), slow_region); 4745 generate_non_array_guard(load_object_klass(dest), slow_region); 4746 4747 // (2) src and dest arrays must have elements of the same BasicType 4748 // done at macro expansion or at Ideal transformation time 4749 4750 // (4) src_offset must not be negative. 4751 generate_negative_guard(src_offset, slow_region); 4752 4753 // (5) dest_offset must not be negative. 4754 generate_negative_guard(dest_offset, slow_region); 4755 4756 // (7) src_offset + length must not exceed length of src. 4757 generate_limit_guard(src_offset, length, 4758 load_array_length(src), 4759 slow_region); 4760 4761 // (8) dest_offset + length must not exceed length of dest. 4762 generate_limit_guard(dest_offset, length, 4763 load_array_length(dest), 4764 slow_region); 4765 4766 // (9) each element of an oop array must be assignable 4767 Node* src_klass = load_object_klass(src); 4768 Node* dest_klass = load_object_klass(dest); 4769 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass); 4770 4771 if (not_subtype_ctrl != top()) { 4772 PreserveJVMState pjvms(this); 4773 set_control(not_subtype_ctrl); 4774 uncommon_trap(Deoptimization::Reason_intrinsic, 4775 Deoptimization::Action_make_not_entrant); 4776 assert(stopped(), "Should be stopped"); 4777 } 4778 { 4779 PreserveJVMState pjvms(this); 4780 set_control(_gvn.transform(slow_region)); 4781 uncommon_trap(Deoptimization::Reason_intrinsic, 4782 Deoptimization::Action_make_not_entrant); 4783 assert(stopped(), "Should be stopped"); 4784 } 4785 } 4786 4787 if (stopped()) { 4788 return true; 4789 } 4790 4791 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL); 4792 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, 4793 // Create LoadRange and LoadKlass nodes for use during macro expansion here 4794 // so the compiler has a chance to eliminate them: during macro expansion, 4795 // we have to set their control (CastPP nodes are eliminated). 4796 load_object_klass(src), load_object_klass(dest), 4797 load_array_length(src), load_array_length(dest)); 4798 4799 if (notest) { 4800 ac->set_arraycopy_notest(); 4801 } 4802 4803 Node* n = _gvn.transform(ac); 4804 assert(n == ac, "cannot disappear"); 4805 ac->connect_outputs(this); 4806 4807 return true; 4808 } 4809 4810 4811 // Helper function which determines if an arraycopy immediately follows 4812 // an allocation, with no intervening tests or other escapes for the object. 4813 AllocateArrayNode* 4814 LibraryCallKit::tightly_coupled_allocation(Node* ptr, 4815 RegionNode* slow_region) { 4816 if (stopped()) return NULL; // no fast path 4817 if (C->AliasLevel() == 0) return NULL; // no MergeMems around 4818 4819 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn); 4820 if (alloc == NULL) return NULL; 4821 4822 Node* rawmem = memory(Compile::AliasIdxRaw); 4823 // Is the allocation's memory state untouched? 4824 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) { 4825 // Bail out if there have been raw-memory effects since the allocation. 4826 // (Example: There might have been a call or safepoint.) 4827 return NULL; 4828 } 4829 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw); 4830 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) { 4831 return NULL; 4832 } 4833 4834 // There must be no unexpected observers of this allocation. 4835 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) { 4836 Node* obs = ptr->fast_out(i); 4837 if (obs != this->map()) { 4838 return NULL; 4839 } 4840 } 4841 4842 // This arraycopy must unconditionally follow the allocation of the ptr. 4843 Node* alloc_ctl = ptr->in(0); 4844 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo"); 4845 4846 Node* ctl = control(); 4847 while (ctl != alloc_ctl) { 4848 // There may be guards which feed into the slow_region. 4849 // Any other control flow means that we might not get a chance 4850 // to finish initializing the allocated object. 4851 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) { 4852 IfNode* iff = ctl->in(0)->as_If(); 4853 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con); 4854 assert(not_ctl != NULL && not_ctl != ctl, "found alternate"); 4855 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) { 4856 ctl = iff->in(0); // This test feeds the known slow_region. 4857 continue; 4858 } 4859 // One more try: Various low-level checks bottom out in 4860 // uncommon traps. If the debug-info of the trap omits 4861 // any reference to the allocation, as we've already 4862 // observed, then there can be no objection to the trap. 4863 bool found_trap = false; 4864 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) { 4865 Node* obs = not_ctl->fast_out(j); 4866 if (obs->in(0) == not_ctl && obs->is_Call() && 4867 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) { 4868 found_trap = true; break; 4869 } 4870 } 4871 if (found_trap) { 4872 ctl = iff->in(0); // This test feeds a harmless uncommon trap. 4873 continue; 4874 } 4875 } 4876 return NULL; 4877 } 4878 4879 // If we get this far, we have an allocation which immediately 4880 // precedes the arraycopy, and we can take over zeroing the new object. 4881 // The arraycopy will finish the initialization, and provide 4882 // a new control state to which we will anchor the destination pointer. 4883 4884 return alloc; 4885 } 4886 4887 //-------------inline_encodeISOArray----------------------------------- 4888 // encode char[] to byte[] in ISO_8859_1 4889 bool LibraryCallKit::inline_encodeISOArray() { 4890 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters"); 4891 // no receiver since it is static method 4892 Node *src = argument(0); 4893 Node *src_offset = argument(1); 4894 Node *dst = argument(2); 4895 Node *dst_offset = argument(3); 4896 Node *length = argument(4); 4897 4898 const Type* src_type = src->Value(&_gvn); 4899 const Type* dst_type = dst->Value(&_gvn); 4900 const TypeAryPtr* top_src = src_type->isa_aryptr(); 4901 const TypeAryPtr* top_dest = dst_type->isa_aryptr(); 4902 if (top_src == NULL || top_src->klass() == NULL || 4903 top_dest == NULL || top_dest->klass() == NULL) { 4904 // failed array check 4905 return false; 4906 } 4907 4908 // Figure out the size and type of the elements we will be copying. 4909 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4910 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4911 if (src_elem != T_CHAR || dst_elem != T_BYTE) { 4912 return false; 4913 } 4914 Node* src_start = array_element_address(src, src_offset, src_elem); 4915 Node* dst_start = array_element_address(dst, dst_offset, dst_elem); 4916 // 'src_start' points to src array + scaled offset 4917 // 'dst_start' points to dst array + scaled offset 4918 4919 const TypeAryPtr* mtype = TypeAryPtr::BYTES; 4920 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length); 4921 enc = _gvn.transform(enc); 4922 Node* res_mem = _gvn.transform(new SCMemProjNode(enc)); 4923 set_memory(res_mem, mtype); 4924 set_result(enc); 4925 return true; 4926 } 4927 4928 //-------------inline_multiplyToLen----------------------------------- 4929 bool LibraryCallKit::inline_multiplyToLen() { 4930 assert(UseMultiplyToLenIntrinsic, "not implementated on this platform"); 4931 4932 address stubAddr = StubRoutines::multiplyToLen(); 4933 if (stubAddr == NULL) { 4934 return false; // Intrinsic's stub is not implemented on this platform 4935 } 4936 const char* stubName = "multiplyToLen"; 4937 4938 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters"); 4939 4940 Node* x = argument(1); 4941 Node* xlen = argument(2); 4942 Node* y = argument(3); 4943 Node* ylen = argument(4); 4944 Node* z = argument(5); 4945 4946 const Type* x_type = x->Value(&_gvn); 4947 const Type* y_type = y->Value(&_gvn); 4948 const TypeAryPtr* top_x = x_type->isa_aryptr(); 4949 const TypeAryPtr* top_y = y_type->isa_aryptr(); 4950 if (top_x == NULL || top_x->klass() == NULL || 4951 top_y == NULL || top_y->klass() == NULL) { 4952 // failed array check 4953 return false; 4954 } 4955 4956 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4957 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 4958 if (x_elem != T_INT || y_elem != T_INT) { 4959 return false; 4960 } 4961 4962 // Set the original stack and the reexecute bit for the interpreter to reexecute 4963 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens 4964 // on the return from z array allocation in runtime. 4965 { PreserveReexecuteState preexecs(this); 4966 jvms()->set_should_reexecute(true); 4967 4968 Node* x_start = array_element_address(x, intcon(0), x_elem); 4969 Node* y_start = array_element_address(y, intcon(0), y_elem); 4970 // 'x_start' points to x array + scaled xlen 4971 // 'y_start' points to y array + scaled ylen 4972 4973 // Allocate the result array 4974 Node* zlen = _gvn.transform(new AddINode(xlen, ylen)); 4975 ciKlass* klass = ciTypeArrayKlass::make(T_INT); 4976 Node* klass_node = makecon(TypeKlassPtr::make(klass)); 4977 4978 IdealKit ideal(this); 4979 4980 #define __ ideal. 4981 Node* one = __ ConI(1); 4982 Node* zero = __ ConI(0); 4983 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done(); 4984 __ set(need_alloc, zero); 4985 __ set(z_alloc, z); 4986 __ if_then(z, BoolTest::eq, null()); { 4987 __ increment (need_alloc, one); 4988 } __ else_(); { 4989 // Update graphKit memory and control from IdealKit. 4990 sync_kit(ideal); 4991 Node* zlen_arg = load_array_length(z); 4992 // Update IdealKit memory and control from graphKit. 4993 __ sync_kit(this); 4994 __ if_then(zlen_arg, BoolTest::lt, zlen); { 4995 __ increment (need_alloc, one); 4996 } __ end_if(); 4997 } __ end_if(); 4998 4999 __ if_then(__ value(need_alloc), BoolTest::ne, zero); { 5000 // Update graphKit memory and control from IdealKit. 5001 sync_kit(ideal); 5002 Node * narr = new_array(klass_node, zlen, 1); 5003 // Update IdealKit memory and control from graphKit. 5004 __ sync_kit(this); 5005 __ set(z_alloc, narr); 5006 } __ end_if(); 5007 5008 sync_kit(ideal); 5009 z = __ value(z_alloc); 5010 // Can't use TypeAryPtr::INTS which uses Bottom offset. 5011 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass)); 5012 // Final sync IdealKit and GraphKit. 5013 final_sync(ideal); 5014 #undef __ 5015 5016 Node* z_start = array_element_address(z, intcon(0), T_INT); 5017 5018 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5019 OptoRuntime::multiplyToLen_Type(), 5020 stubAddr, stubName, TypePtr::BOTTOM, 5021 x_start, xlen, y_start, ylen, z_start, zlen); 5022 } // original reexecute is set back here 5023 5024 C->set_has_split_ifs(true); // Has chance for split-if optimization 5025 set_result(z); 5026 return true; 5027 } 5028 5029 5030 /** 5031 * Calculate CRC32 for byte. 5032 * int java.util.zip.CRC32.update(int crc, int b) 5033 */ 5034 bool LibraryCallKit::inline_updateCRC32() { 5035 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5036 assert(callee()->signature()->size() == 2, "update has 2 parameters"); 5037 // no receiver since it is static method 5038 Node* crc = argument(0); // type: int 5039 Node* b = argument(1); // type: int 5040 5041 /* 5042 * int c = ~ crc; 5043 * b = timesXtoThe32[(b ^ c) & 0xFF]; 5044 * b = b ^ (c >>> 8); 5045 * crc = ~b; 5046 */ 5047 5048 Node* M1 = intcon(-1); 5049 crc = _gvn.transform(new XorINode(crc, M1)); 5050 Node* result = _gvn.transform(new XorINode(crc, b)); 5051 result = _gvn.transform(new AndINode(result, intcon(0xFF))); 5052 5053 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr())); 5054 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2))); 5055 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset)); 5056 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered); 5057 5058 crc = _gvn.transform(new URShiftINode(crc, intcon(8))); 5059 result = _gvn.transform(new XorINode(crc, result)); 5060 result = _gvn.transform(new XorINode(result, M1)); 5061 set_result(result); 5062 return true; 5063 } 5064 5065 /** 5066 * Calculate CRC32 for byte[] array. 5067 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len) 5068 */ 5069 bool LibraryCallKit::inline_updateBytesCRC32() { 5070 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5071 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters"); 5072 // no receiver since it is static method 5073 Node* crc = argument(0); // type: int 5074 Node* src = argument(1); // type: oop 5075 Node* offset = argument(2); // type: int 5076 Node* length = argument(3); // type: int 5077 5078 const Type* src_type = src->Value(&_gvn); 5079 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5080 if (top_src == NULL || top_src->klass() == NULL) { 5081 // failed array check 5082 return false; 5083 } 5084 5085 // Figure out the size and type of the elements we will be copying. 5086 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5087 if (src_elem != T_BYTE) { 5088 return false; 5089 } 5090 5091 // 'src_start' points to src array + scaled offset 5092 Node* src_start = array_element_address(src, offset, src_elem); 5093 5094 // We assume that range check is done by caller. 5095 // TODO: generate range check (offset+length < src.length) in debug VM. 5096 5097 // Call the stub. 5098 address stubAddr = StubRoutines::updateBytesCRC32(); 5099 const char *stubName = "updateBytesCRC32"; 5100 5101 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5102 stubAddr, stubName, TypePtr::BOTTOM, 5103 crc, src_start, length); 5104 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5105 set_result(result); 5106 return true; 5107 } 5108 5109 /** 5110 * Calculate CRC32 for ByteBuffer. 5111 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len) 5112 */ 5113 bool LibraryCallKit::inline_updateByteBufferCRC32() { 5114 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support"); 5115 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long"); 5116 // no receiver since it is static method 5117 Node* crc = argument(0); // type: int 5118 Node* src = argument(1); // type: long 5119 Node* offset = argument(3); // type: int 5120 Node* length = argument(4); // type: int 5121 5122 src = ConvL2X(src); // adjust Java long to machine word 5123 Node* base = _gvn.transform(new CastX2PNode(src)); 5124 offset = ConvI2X(offset); 5125 5126 // 'src_start' points to src array + scaled offset 5127 Node* src_start = basic_plus_adr(top(), base, offset); 5128 5129 // Call the stub. 5130 address stubAddr = StubRoutines::updateBytesCRC32(); 5131 const char *stubName = "updateBytesCRC32"; 5132 5133 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(), 5134 stubAddr, stubName, TypePtr::BOTTOM, 5135 crc, src_start, length); 5136 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5137 set_result(result); 5138 return true; 5139 } 5140 5141 //----------------------------inline_reference_get---------------------------- 5142 // public T java.lang.ref.Reference.get(); 5143 bool LibraryCallKit::inline_reference_get() { 5144 const int referent_offset = java_lang_ref_Reference::referent_offset; 5145 guarantee(referent_offset > 0, "should have already been set"); 5146 5147 // Get the argument: 5148 Node* reference_obj = null_check_receiver(); 5149 if (stopped()) return true; 5150 5151 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset); 5152 5153 ciInstanceKlass* klass = env()->Object_klass(); 5154 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass); 5155 5156 Node* no_ctrl = NULL; 5157 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered); 5158 5159 // Use the pre-barrier to record the value in the referent field 5160 pre_barrier(false /* do_load */, 5161 control(), 5162 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */, 5163 result /* pre_val */, 5164 T_OBJECT); 5165 5166 // Add memory barrier to prevent commoning reads from this field 5167 // across safepoint since GC can change its value. 5168 insert_mem_bar(Op_MemBarCPUOrder); 5169 5170 set_result(result); 5171 return true; 5172 } 5173 5174 5175 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, 5176 bool is_exact=true, bool is_static=false) { 5177 5178 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr(); 5179 assert(tinst != NULL, "obj is null"); 5180 assert(tinst->klass()->is_loaded(), "obj is not loaded"); 5181 assert(!is_exact || tinst->klass_is_exact(), "klass not exact"); 5182 5183 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName), 5184 ciSymbol::make(fieldTypeString), 5185 is_static); 5186 if (field == NULL) return (Node *) NULL; 5187 assert (field != NULL, "undefined field"); 5188 5189 // Next code copied from Parse::do_get_xxx(): 5190 5191 // Compute address and memory type. 5192 int offset = field->offset_in_bytes(); 5193 bool is_vol = field->is_volatile(); 5194 ciType* field_klass = field->type(); 5195 assert(field_klass->is_loaded(), "should be loaded"); 5196 const TypePtr* adr_type = C->alias_type(field)->adr_type(); 5197 Node *adr = basic_plus_adr(fromObj, fromObj, offset); 5198 BasicType bt = field->layout_type(); 5199 5200 // Build the resultant type of the load 5201 const Type *type; 5202 if (bt == T_OBJECT) { 5203 type = TypeOopPtr::make_from_klass(field_klass->as_klass()); 5204 } else { 5205 type = Type::get_const_basic_type(bt); 5206 } 5207 5208 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) { 5209 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier 5210 } 5211 // Build the load. 5212 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered; 5213 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol); 5214 // If reference is volatile, prevent following memory ops from 5215 // floating up past the volatile read. Also prevents commoning 5216 // another volatile read. 5217 if (is_vol) { 5218 // Memory barrier includes bogus read of value to force load BEFORE membar 5219 insert_mem_bar(Op_MemBarAcquire, loadedField); 5220 } 5221 return loadedField; 5222 } 5223 5224 5225 //------------------------------inline_aescrypt_Block----------------------- 5226 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) { 5227 address stubAddr; 5228 const char *stubName; 5229 assert(UseAES, "need AES instruction support"); 5230 5231 switch(id) { 5232 case vmIntrinsics::_aescrypt_encryptBlock: 5233 stubAddr = StubRoutines::aescrypt_encryptBlock(); 5234 stubName = "aescrypt_encryptBlock"; 5235 break; 5236 case vmIntrinsics::_aescrypt_decryptBlock: 5237 stubAddr = StubRoutines::aescrypt_decryptBlock(); 5238 stubName = "aescrypt_decryptBlock"; 5239 break; 5240 } 5241 if (stubAddr == NULL) return false; 5242 5243 Node* aescrypt_object = argument(0); 5244 Node* src = argument(1); 5245 Node* src_offset = argument(2); 5246 Node* dest = argument(3); 5247 Node* dest_offset = argument(4); 5248 5249 // (1) src and dest are arrays. 5250 const Type* src_type = src->Value(&_gvn); 5251 const Type* dest_type = dest->Value(&_gvn); 5252 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5253 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5254 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5255 5256 // for the quick and dirty code we will skip all the checks. 5257 // we are just trying to get the call to be generated. 5258 Node* src_start = src; 5259 Node* dest_start = dest; 5260 if (src_offset != NULL || dest_offset != NULL) { 5261 assert(src_offset != NULL && dest_offset != NULL, ""); 5262 src_start = array_element_address(src, src_offset, T_BYTE); 5263 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5264 } 5265 5266 // now need to get the start of its expanded key array 5267 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5268 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5269 if (k_start == NULL) return false; 5270 5271 if (Matcher::pass_original_key_for_aes()) { 5272 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5273 // compatibility issues between Java key expansion and SPARC crypto instructions 5274 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5275 if (original_k_start == NULL) return false; 5276 5277 // Call the stub. 5278 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5279 stubAddr, stubName, TypePtr::BOTTOM, 5280 src_start, dest_start, k_start, original_k_start); 5281 } else { 5282 // Call the stub. 5283 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(), 5284 stubAddr, stubName, TypePtr::BOTTOM, 5285 src_start, dest_start, k_start); 5286 } 5287 5288 return true; 5289 } 5290 5291 //------------------------------inline_cipherBlockChaining_AESCrypt----------------------- 5292 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) { 5293 address stubAddr; 5294 const char *stubName; 5295 5296 assert(UseAES, "need AES instruction support"); 5297 5298 switch(id) { 5299 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt: 5300 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt(); 5301 stubName = "cipherBlockChaining_encryptAESCrypt"; 5302 break; 5303 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt: 5304 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt(); 5305 stubName = "cipherBlockChaining_decryptAESCrypt"; 5306 break; 5307 } 5308 if (stubAddr == NULL) return false; 5309 5310 Node* cipherBlockChaining_object = argument(0); 5311 Node* src = argument(1); 5312 Node* src_offset = argument(2); 5313 Node* len = argument(3); 5314 Node* dest = argument(4); 5315 Node* dest_offset = argument(5); 5316 5317 // (1) src and dest are arrays. 5318 const Type* src_type = src->Value(&_gvn); 5319 const Type* dest_type = dest->Value(&_gvn); 5320 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5321 const TypeAryPtr* top_dest = dest_type->isa_aryptr(); 5322 assert (top_src != NULL && top_src->klass() != NULL 5323 && top_dest != NULL && top_dest->klass() != NULL, "args are strange"); 5324 5325 // checks are the responsibility of the caller 5326 Node* src_start = src; 5327 Node* dest_start = dest; 5328 if (src_offset != NULL || dest_offset != NULL) { 5329 assert(src_offset != NULL && dest_offset != NULL, ""); 5330 src_start = array_element_address(src, src_offset, T_BYTE); 5331 dest_start = array_element_address(dest, dest_offset, T_BYTE); 5332 } 5333 5334 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object 5335 // (because of the predicated logic executed earlier). 5336 // so we cast it here safely. 5337 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java 5338 5339 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5340 if (embeddedCipherObj == NULL) return false; 5341 5342 // cast it to what we know it will be at runtime 5343 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr(); 5344 assert(tinst != NULL, "CBC obj is null"); 5345 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded"); 5346 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5347 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded"); 5348 5349 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5350 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt); 5351 const TypeOopPtr* xtype = aklass->as_instance_type(); 5352 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype); 5353 aescrypt_object = _gvn.transform(aescrypt_object); 5354 5355 // we need to get the start of the aescrypt_object's expanded key array 5356 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object); 5357 if (k_start == NULL) return false; 5358 5359 // similarly, get the start address of the r vector 5360 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false); 5361 if (objRvec == NULL) return false; 5362 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE); 5363 5364 Node* cbcCrypt; 5365 if (Matcher::pass_original_key_for_aes()) { 5366 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to 5367 // compatibility issues between Java key expansion and SPARC crypto instructions 5368 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object); 5369 if (original_k_start == NULL) return false; 5370 5371 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start 5372 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5373 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5374 stubAddr, stubName, TypePtr::BOTTOM, 5375 src_start, dest_start, k_start, r_start, len, original_k_start); 5376 } else { 5377 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len 5378 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP, 5379 OptoRuntime::cipherBlockChaining_aescrypt_Type(), 5380 stubAddr, stubName, TypePtr::BOTTOM, 5381 src_start, dest_start, k_start, r_start, len); 5382 } 5383 5384 // return cipher length (int) 5385 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms)); 5386 set_result(retvalue); 5387 return true; 5388 } 5389 5390 //------------------------------get_key_start_from_aescrypt_object----------------------- 5391 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) { 5392 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false); 5393 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5394 if (objAESCryptKey == NULL) return (Node *) NULL; 5395 5396 // now have the array, need to get the start address of the K array 5397 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT); 5398 return k_start; 5399 } 5400 5401 //------------------------------get_original_key_start_from_aescrypt_object----------------------- 5402 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) { 5403 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false); 5404 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt"); 5405 if (objAESCryptKey == NULL) return (Node *) NULL; 5406 5407 // now have the array, need to get the start address of the lastKey array 5408 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE); 5409 return original_k_start; 5410 } 5411 5412 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate---------------------------- 5413 // Return node representing slow path of predicate check. 5414 // the pseudo code we want to emulate with this predicate is: 5415 // for encryption: 5416 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath 5417 // for decryption: 5418 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath 5419 // note cipher==plain is more conservative than the original java code but that's OK 5420 // 5421 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) { 5422 // The receiver was checked for NULL already. 5423 Node* objCBC = argument(0); 5424 5425 // Load embeddedCipher field of CipherBlockChaining object. 5426 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false); 5427 5428 // get AESCrypt klass for instanceOf check 5429 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point 5430 // will have same classloader as CipherBlockChaining object 5431 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr(); 5432 assert(tinst != NULL, "CBCobj is null"); 5433 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded"); 5434 5435 // we want to do an instanceof comparison against the AESCrypt class 5436 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt")); 5437 if (!klass_AESCrypt->is_loaded()) { 5438 // if AESCrypt is not even loaded, we never take the intrinsic fast path 5439 Node* ctrl = control(); 5440 set_control(top()); // no regular fast path 5441 return ctrl; 5442 } 5443 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass(); 5444 5445 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt))); 5446 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1))); 5447 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5448 5449 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5450 5451 // for encryption, we are done 5452 if (!decrypting) 5453 return instof_false; // even if it is NULL 5454 5455 // for decryption, we need to add a further check to avoid 5456 // taking the intrinsic path when cipher and plain are the same 5457 // see the original java code for why. 5458 RegionNode* region = new RegionNode(3); 5459 region->init_req(1, instof_false); 5460 Node* src = argument(1); 5461 Node* dest = argument(4); 5462 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest)); 5463 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq)); 5464 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN); 5465 region->init_req(2, src_dest_conjoint); 5466 5467 record_for_igvn(region); 5468 return _gvn.transform(region); 5469 } 5470 5471 //------------------------------inline_sha_implCompress----------------------- 5472 // 5473 // Calculate SHA (i.e., SHA-1) for single-block byte[] array. 5474 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs) 5475 // 5476 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array. 5477 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs) 5478 // 5479 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array. 5480 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs) 5481 // 5482 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) { 5483 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters"); 5484 5485 Node* sha_obj = argument(0); 5486 Node* src = argument(1); // type oop 5487 Node* ofs = argument(2); // type int 5488 5489 const Type* src_type = src->Value(&_gvn); 5490 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5491 if (top_src == NULL || top_src->klass() == NULL) { 5492 // failed array check 5493 return false; 5494 } 5495 // Figure out the size and type of the elements we will be copying. 5496 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5497 if (src_elem != T_BYTE) { 5498 return false; 5499 } 5500 // 'src_start' points to src array + offset 5501 Node* src_start = array_element_address(src, ofs, src_elem); 5502 Node* state = NULL; 5503 address stubAddr; 5504 const char *stubName; 5505 5506 switch(id) { 5507 case vmIntrinsics::_sha_implCompress: 5508 assert(UseSHA1Intrinsics, "need SHA1 instruction support"); 5509 state = get_state_from_sha_object(sha_obj); 5510 stubAddr = StubRoutines::sha1_implCompress(); 5511 stubName = "sha1_implCompress"; 5512 break; 5513 case vmIntrinsics::_sha2_implCompress: 5514 assert(UseSHA256Intrinsics, "need SHA256 instruction support"); 5515 state = get_state_from_sha_object(sha_obj); 5516 stubAddr = StubRoutines::sha256_implCompress(); 5517 stubName = "sha256_implCompress"; 5518 break; 5519 case vmIntrinsics::_sha5_implCompress: 5520 assert(UseSHA512Intrinsics, "need SHA512 instruction support"); 5521 state = get_state_from_sha5_object(sha_obj); 5522 stubAddr = StubRoutines::sha512_implCompress(); 5523 stubName = "sha512_implCompress"; 5524 break; 5525 default: 5526 fatal_unexpected_iid(id); 5527 return false; 5528 } 5529 if (state == NULL) return false; 5530 5531 // Call the stub. 5532 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(), 5533 stubAddr, stubName, TypePtr::BOTTOM, 5534 src_start, state); 5535 5536 return true; 5537 } 5538 5539 //------------------------------inline_digestBase_implCompressMB----------------------- 5540 // 5541 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array. 5542 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit) 5543 // 5544 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) { 5545 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5546 "need SHA1/SHA256/SHA512 instruction support"); 5547 assert((uint)predicate < 3, "sanity"); 5548 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters"); 5549 5550 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already. 5551 Node* src = argument(1); // byte[] array 5552 Node* ofs = argument(2); // type int 5553 Node* limit = argument(3); // type int 5554 5555 const Type* src_type = src->Value(&_gvn); 5556 const TypeAryPtr* top_src = src_type->isa_aryptr(); 5557 if (top_src == NULL || top_src->klass() == NULL) { 5558 // failed array check 5559 return false; 5560 } 5561 // Figure out the size and type of the elements we will be copying. 5562 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type(); 5563 if (src_elem != T_BYTE) { 5564 return false; 5565 } 5566 // 'src_start' points to src array + offset 5567 Node* src_start = array_element_address(src, ofs, src_elem); 5568 5569 const char* klass_SHA_name = NULL; 5570 const char* stub_name = NULL; 5571 address stub_addr = NULL; 5572 bool long_state = false; 5573 5574 switch (predicate) { 5575 case 0: 5576 if (UseSHA1Intrinsics) { 5577 klass_SHA_name = "sun/security/provider/SHA"; 5578 stub_name = "sha1_implCompressMB"; 5579 stub_addr = StubRoutines::sha1_implCompressMB(); 5580 } 5581 break; 5582 case 1: 5583 if (UseSHA256Intrinsics) { 5584 klass_SHA_name = "sun/security/provider/SHA2"; 5585 stub_name = "sha256_implCompressMB"; 5586 stub_addr = StubRoutines::sha256_implCompressMB(); 5587 } 5588 break; 5589 case 2: 5590 if (UseSHA512Intrinsics) { 5591 klass_SHA_name = "sun/security/provider/SHA5"; 5592 stub_name = "sha512_implCompressMB"; 5593 stub_addr = StubRoutines::sha512_implCompressMB(); 5594 long_state = true; 5595 } 5596 break; 5597 default: 5598 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5599 } 5600 if (klass_SHA_name != NULL) { 5601 // get DigestBase klass to lookup for SHA klass 5602 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr(); 5603 assert(tinst != NULL, "digestBase_obj is not instance???"); 5604 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5605 5606 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5607 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded"); 5608 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5609 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit); 5610 } 5611 return false; 5612 } 5613 //------------------------------inline_sha_implCompressMB----------------------- 5614 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA, 5615 bool long_state, address stubAddr, const char *stubName, 5616 Node* src_start, Node* ofs, Node* limit) { 5617 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA); 5618 const TypeOopPtr* xtype = aklass->as_instance_type(); 5619 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype); 5620 sha_obj = _gvn.transform(sha_obj); 5621 5622 Node* state; 5623 if (long_state) { 5624 state = get_state_from_sha5_object(sha_obj); 5625 } else { 5626 state = get_state_from_sha_object(sha_obj); 5627 } 5628 if (state == NULL) return false; 5629 5630 // Call the stub. 5631 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, 5632 OptoRuntime::digestBase_implCompressMB_Type(), 5633 stubAddr, stubName, TypePtr::BOTTOM, 5634 src_start, state, ofs, limit); 5635 // return ofs (int) 5636 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms)); 5637 set_result(result); 5638 5639 return true; 5640 } 5641 5642 //------------------------------get_state_from_sha_object----------------------- 5643 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) { 5644 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false); 5645 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2"); 5646 if (sha_state == NULL) return (Node *) NULL; 5647 5648 // now have the array, need to get the start address of the state array 5649 Node* state = array_element_address(sha_state, intcon(0), T_INT); 5650 return state; 5651 } 5652 5653 //------------------------------get_state_from_sha5_object----------------------- 5654 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) { 5655 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false); 5656 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5"); 5657 if (sha_state == NULL) return (Node *) NULL; 5658 5659 // now have the array, need to get the start address of the state array 5660 Node* state = array_element_address(sha_state, intcon(0), T_LONG); 5661 return state; 5662 } 5663 5664 //----------------------------inline_digestBase_implCompressMB_predicate---------------------------- 5665 // Return node representing slow path of predicate check. 5666 // the pseudo code we want to emulate with this predicate is: 5667 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath 5668 // 5669 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) { 5670 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics, 5671 "need SHA1/SHA256/SHA512 instruction support"); 5672 assert((uint)predicate < 3, "sanity"); 5673 5674 // The receiver was checked for NULL already. 5675 Node* digestBaseObj = argument(0); 5676 5677 // get DigestBase klass for instanceOf check 5678 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr(); 5679 assert(tinst != NULL, "digestBaseObj is null"); 5680 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded"); 5681 5682 const char* klass_SHA_name = NULL; 5683 switch (predicate) { 5684 case 0: 5685 if (UseSHA1Intrinsics) { 5686 // we want to do an instanceof comparison against the SHA class 5687 klass_SHA_name = "sun/security/provider/SHA"; 5688 } 5689 break; 5690 case 1: 5691 if (UseSHA256Intrinsics) { 5692 // we want to do an instanceof comparison against the SHA2 class 5693 klass_SHA_name = "sun/security/provider/SHA2"; 5694 } 5695 break; 5696 case 2: 5697 if (UseSHA512Intrinsics) { 5698 // we want to do an instanceof comparison against the SHA5 class 5699 klass_SHA_name = "sun/security/provider/SHA5"; 5700 } 5701 break; 5702 default: 5703 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate)); 5704 } 5705 5706 ciKlass* klass_SHA = NULL; 5707 if (klass_SHA_name != NULL) { 5708 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name)); 5709 } 5710 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) { 5711 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path 5712 Node* ctrl = control(); 5713 set_control(top()); // no intrinsic path 5714 return ctrl; 5715 } 5716 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass(); 5717 5718 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA))); 5719 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1))); 5720 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne)); 5721 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN); 5722 5723 return instof_false; // even if it is NULL 5724 }